drivers/md/raid5.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * raid5.c : Multiple Devices driver for Linux
   4  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   5  *         Copyright (C) 1999, 2000 Ingo Molnar
   6  *         Copyright (C) 2002, 2003 H. Peter Anvin
   7  *
   8  * RAID-4/5/6 management functions.
   9  * Thanks to Penguin Computing for making the RAID-6 development possible
  10  * by donating a test server!
  11  */
  12
  13 /*
  14  * BITMAP UNPLUGGING:
  15  *
  16  * The sequencing for updating the bitmap reliably is a little
  17  * subtle (and I got it wrong the first time) so it deserves some
  18  * explanation.
  19  *
  20  * We group bitmap updates into batches.  Each batch has a number.
  21  * We may write out several batches at once, but that isn't very important.
  22  * conf->seq_write is the number of the last batch successfully written.
  23  * conf->seq_flush is the number of the last batch that was closed to
  24  *    new additions.
  25  * When we discover that we will need to write to any block in a stripe
  26  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
  27  * the number of the batch it will be in. This is seq_flush+1.
  28  * When we are ready to do a write, if that batch hasn't been written yet,
  29  *   we plug the array and queue the stripe for later.
  30  * When an unplug happens, we increment bm_flush, thus closing the current
  31  *   batch.
  32  * When we notice that bm_flush > bm_write, we write out all pending updates
  33  * to the bitmap, and advance bm_write to where bm_flush was.
  34  * This may occasionally write a bit out twice, but is sure never to
  35  * miss any bits.
  36  */
  37
  38 #include <linux/blkdev.h>
  39 #include <linux/kthread.h>
  40 #include <linux/raid/pq.h>
  41 #include <linux/async_tx.h>
  42 #include <linux/module.h>
  43 #include <linux/async.h>
  44 #include <linux/seq_file.h>
  45 #include <linux/cpu.h>
  46 #include <linux/slab.h>
  47 #include <linux/ratelimit.h>
  48 #include <linux/nodemask.h>
  49
  50 #include <trace/events/block.h>
  51 #include <linux/list_sort.h>
  52
  53 #include "md.h"
  54 #include "raid5.h"
  55 #include "raid0.h"
  56 #include "md-bitmap.h"
  57 #include "raid5-log.h"
  58
  59 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
  60
  61 #define cpu_to_group(cpu) cpu_to_node(cpu)
  62 #define ANY_GROUP NUMA_NO_NODE
  63
  64 #define RAID5_MAX_REQ_STRIPES 256
  65
  66 static bool devices_handle_discard_safely = false;
  67 module_param(devices_handle_discard_safely, bool, 0644);
  68 MODULE_PARM_DESC(devices_handle_discard_safely,
  69                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  70 static struct workqueue_struct *raid5_wq;
  71
  72 static void raid5_quiesce(struct mddev *mddev, int quiesce);
  73
  74 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  75 {
  76         int hash = (sect >> RAID5_STRIPE_SHIFT(conf)) & HASH_MASK;
  77         return &conf->stripe_hashtbl[hash];
  78 }
  79
  80 static inline int stripe_hash_locks_hash(struct r5conf *conf, sector_t sect)
  81 {
  82         return (sect >> RAID5_STRIPE_SHIFT(conf)) & STRIPE_HASH_LOCKS_MASK;
  83 }
  84
  85 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  86         __acquires(&conf->device_lock)
  87 {
  88         spin_lock_irq(conf->hash_locks + hash);
  89         spin_lock(&conf->device_lock);
  90 }
  91
  92 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
  93         __releases(&conf->device_lock)
  94 {
  95         spin_unlock(&conf->device_lock);
  96         spin_unlock_irq(conf->hash_locks + hash);
  97 }
  98
  99 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
 100         __acquires(&conf->device_lock)
 101 {
 102         int i;
 103         spin_lock_irq(conf->hash_locks);
 104         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
 105                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
 106         spin_lock(&conf->device_lock);
 107 }
 108
 109 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 110         __releases(&conf->device_lock)
 111 {
 112         int i;
 113         spin_unlock(&conf->device_lock);
 114         for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
 115                 spin_unlock(conf->hash_locks + i);
 116         spin_unlock_irq(conf->hash_locks);
 117 }
 118
 119 /* Find first data disk in a raid6 stripe */
 120 static inline int raid6_d0(struct stripe_head *sh)
 121 {
 122         if (sh->ddf_layout)
 123                 /* ddf always start from first device */
 124                 return 0;
 125         /* md starts just after Q block */
 126         if (sh->qd_idx == sh->disks - 1)
 127                 return 0;
 128         else
 129                 return sh->qd_idx + 1;
 130 }
 131 static inline int raid6_next_disk(int disk, int raid_disks)
 132 {
 133         disk++;
 134         return (disk < raid_disks) ? disk : 0;
 135 }
 136
 137 /* When walking through the disks in a raid5, starting at raid6_d0,
 138  * We need to map each disk to a 'slot', where the data disks are slot
 139  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 140  * is raid_disks-1.  This help does that mapping.
 141  */
 142 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 143                              int *count, int syndrome_disks)
 144 {
 145         int slot = *count;
 146
 147         if (sh->ddf_layout)
 148                 (*count)++;
 149         if (idx == sh->pd_idx)
 150                 return syndrome_disks;
 151         if (idx == sh->qd_idx)
 152                 return syndrome_disks + 1;
 153         if (!sh->ddf_layout)
 154                 (*count)++;
 155         return slot;
 156 }
 157
 158 static void print_raid5_conf(struct r5conf *conf);
 159
 160 static int stripe_operations_active(struct stripe_head *sh)
 161 {
 162         return sh->check_state || sh->reconstruct_state ||
 163                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 164                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 165 }
 166
 167 static bool stripe_is_lowprio(struct stripe_head *sh)
 168 {
 169         return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
 170                 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
 171                !test_bit(STRIPE_R5C_CACHING, &sh->state);
 172 }
 173
 174 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 175         __must_hold(&sh->raid_conf->device_lock)
 176 {
 177         struct r5conf *conf = sh->raid_conf;
 178         struct r5worker_group *group;
 179         int thread_cnt;
 180         int i, cpu = sh->cpu;
 181
 182         if (!cpu_online(cpu)) {
 183                 cpu = cpumask_any(cpu_online_mask);
 184                 sh->cpu = cpu;
 185         }
 186
 187         if (list_empty(&sh->lru)) {
 188                 struct r5worker_group *group;
 189                 group = conf->worker_groups + cpu_to_group(cpu);
 190                 if (stripe_is_lowprio(sh))
 191                         list_add_tail(&sh->lru, &group->loprio_list);
 192                 else
 193                         list_add_tail(&sh->lru, &group->handle_list);
 194                 group->stripes_cnt++;
 195                 sh->group = group;
 196         }
 197
 198         if (conf->worker_cnt_per_group == 0) {
 199                 md_wakeup_thread(conf->mddev->thread);
 200                 return;
 201         }
 202
 203         group = conf->worker_groups + cpu_to_group(sh->cpu);
 204
 205         group->workers[0].working = true;
 206         /* at least one worker should run to avoid race */
 207         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 208
 209         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 210         /* wakeup more workers */
 211         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 212                 if (group->workers[i].working == false) {
 213                         group->workers[i].working = true;
 214                         queue_work_on(sh->cpu, raid5_wq,
 215                                       &group->workers[i].work);
 216                         thread_cnt--;
 217                 }
 218         }
 219 }
 220
 221 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 222                               struct list_head *temp_inactive_list)
 223         __must_hold(&conf->device_lock)
 224 {
 225         int i;
 226         int injournal = 0;      /* number of date pages with R5_InJournal */
 227
 228         BUG_ON(!list_empty(&sh->lru));
 229         BUG_ON(atomic_read(&conf->active_stripes)==0);
 230
 231         if (r5c_is_writeback(conf->log))
 232                 for (i = sh->disks; i--; )
 233                         if (test_bit(R5_InJournal, &sh->dev[i].flags))
 234                                 injournal++;
 235         /*
 236          * In the following cases, the stripe cannot be released to cached
 237          * lists. Therefore, we make the stripe write out and set
 238          * STRIPE_HANDLE:
 239          *   1. when quiesce in r5c write back;
 240          *   2. when resync is requested fot the stripe.
 241          */
 242         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
 243             (conf->quiesce && r5c_is_writeback(conf->log) &&
 244              !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
 245                 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
 246                         r5c_make_stripe_write_out(sh);
 247                 set_bit(STRIPE_HANDLE, &sh->state);
 248         }
 249
 250         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 251                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 252                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 253                         list_add_tail(&sh->lru, &conf->delayed_list);
 254                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 255                            sh->bm_seq - conf->seq_write > 0)
 256                         list_add_tail(&sh->lru, &conf->bitmap_list);
 257                 else {
 258                         clear_bit(STRIPE_DELAYED, &sh->state);
 259                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 260                         if (conf->worker_cnt_per_group == 0) {
 261                                 if (stripe_is_lowprio(sh))
 262                                         list_add_tail(&sh->lru,
 263                                                         &conf->loprio_list);
 264                                 else
 265                                         list_add_tail(&sh->lru,
 266                                                         &conf->handle_list);
 267                         } else {
 268                                 raid5_wakeup_stripe_thread(sh);
 269                                 return;
 270                         }
 271                 }
 272                 md_wakeup_thread(conf->mddev->thread);
 273         } else {
 274                 BUG_ON(stripe_operations_active(sh));
 275                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 276                         if (atomic_dec_return(&conf->preread_active_stripes)
 277                             < IO_THRESHOLD)
 278                                 md_wakeup_thread(conf->mddev->thread);
 279                 atomic_dec(&conf->active_stripes);
 280                 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 281                         if (!r5c_is_writeback(conf->log))
 282                                 list_add_tail(&sh->lru, temp_inactive_list);
 283                         else {
 284                                 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
 285                                 if (injournal == 0)
 286                                         list_add_tail(&sh->lru, temp_inactive_list);
 287                                 else if (injournal == conf->raid_disks - conf->max_degraded) {
 288                                         /* full stripe */
 289                                         if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
 290                                                 atomic_inc(&conf->r5c_cached_full_stripes);
 291                                         if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
 292                                                 atomic_dec(&conf->r5c_cached_partial_stripes);
 293                                         list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
 294                                         r5c_check_cached_full_stripe(conf);
 295                                 } else
 296                                         /*
 297                                          * STRIPE_R5C_PARTIAL_STRIPE is set in
 298                                          * r5c_try_caching_write(). No need to
 299                                          * set it again.
 300                                          */
 301                                         list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
 302                         }
 303                 }
 304         }
 305 }
 306
 307 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 308                              struct list_head *temp_inactive_list)
 309         __must_hold(&conf->device_lock)
 310 {
 311         if (atomic_dec_and_test(&sh->count))
 312                 do_release_stripe(conf, sh, temp_inactive_list);
 313 }
 314
 315 /*
 316  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 317  *
 318  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 319  * given time. Adding stripes only takes device lock, while deleting stripes
 320  * only takes hash lock.
 321  */
 322 static void release_inactive_stripe_list(struct r5conf *conf,
 323                                          struct list_head *temp_inactive_list,
 324                                          int hash)
 325 {
 326         int size;
 327         bool do_wakeup = false;
 328         unsigned long flags;
 329
 330         if (hash == NR_STRIPE_HASH_LOCKS) {
 331                 size = NR_STRIPE_HASH_LOCKS;
 332                 hash = NR_STRIPE_HASH_LOCKS - 1;
 333         } else
 334                 size = 1;
 335         while (size) {
 336                 struct list_head *list = &temp_inactive_list[size - 1];
 337
 338                 /*
 339                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 340                  * remove stripes from the list
 341                  */
 342                 if (!list_empty_careful(list)) {
 343                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 344                         if (list_empty(conf->inactive_list + hash) &&
 345                             !list_empty(list))
 346                                 atomic_dec(&conf->empty_inactive_list_nr);
 347                         list_splice_tail_init(list, conf->inactive_list + hash);
 348                         do_wakeup = true;
 349                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 350                 }
 351                 size--;
 352                 hash--;
 353         }
 354
 355         if (do_wakeup) {
 356                 wake_up(&conf->wait_for_stripe);
 357                 if (atomic_read(&conf->active_stripes) == 0)
 358                         wake_up(&conf->wait_for_quiescent);
 359                 if (conf->retry_read_aligned)
 360                         md_wakeup_thread(conf->mddev->thread);
 361         }
 362 }
 363
 364 static int release_stripe_list(struct r5conf *conf,
 365                                struct list_head *temp_inactive_list)
 366         __must_hold(&conf->device_lock)
 367 {
 368         struct stripe_head *sh, *t;
 369         int count = 0;
 370         struct llist_node *head;
 371
 372         head = llist_del_all(&conf->released_stripes);
 373         head = llist_reverse_order(head);
 374         llist_for_each_entry_safe(sh, t, head, release_list) {
 375                 int hash;
 376
 377                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 378                 smp_mb();
 379                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 380                 /*
 381                  * Don't worry the bit is set here, because if the bit is set
 382                  * again, the count is always > 1. This is true for
 383                  * STRIPE_ON_UNPLUG_LIST bit too.
 384                  */
 385                 hash = sh->hash_lock_index;
 386                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 387                 count++;
 388         }
 389
 390         return count;
 391 }
 392
 393 void raid5_release_stripe(struct stripe_head *sh)
 394 {
 395         struct r5conf *conf = sh->raid_conf;
 396         unsigned long flags;
 397         struct list_head list;
 398         int hash;
 399         bool wakeup;
 400
 401         /* Avoid release_list until the last reference.
 402          */
 403         if (atomic_add_unless(&sh->count, -1, 1))
 404                 return;
 405
 406         if (unlikely(!conf->mddev->thread) ||
 407                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 408                 goto slow_path;
 409         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 410         if (wakeup)
 411                 md_wakeup_thread(conf->mddev->thread);
 412         return;
 413 slow_path:
 414         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 415         if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
 416                 INIT_LIST_HEAD(&list);
 417                 hash = sh->hash_lock_index;
 418                 do_release_stripe(conf, sh, &list);
 419                 spin_unlock_irqrestore(&conf->device_lock, flags);
 420                 release_inactive_stripe_list(conf, &list, hash);
 421         }
 422 }
 423
 424 static inline void remove_hash(struct stripe_head *sh)
 425 {
 426         pr_debug("remove_hash(), stripe %llu\n",
 427                 (unsigned long long)sh->sector);
 428
 429         hlist_del_init(&sh->hash);
 430 }
 431
 432 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 433 {
 434         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 435
 436         pr_debug("insert_hash(), stripe %llu\n",
 437                 (unsigned long long)sh->sector);
 438
 439         hlist_add_head(&sh->hash, hp);
 440 }
 441
 442 /* find an idle stripe, make sure it is unhashed, and return it. */
 443 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 444 {
 445         struct stripe_head *sh = NULL;
 446         struct list_head *first;
 447
 448         if (list_empty(conf->inactive_list + hash))
 449                 goto out;
 450         first = (conf->inactive_list + hash)->next;
 451         sh = list_entry(first, struct stripe_head, lru);
 452         list_del_init(first);
 453         remove_hash(sh);
 454         atomic_inc(&conf->active_stripes);
 455         BUG_ON(hash != sh->hash_lock_index);
 456         if (list_empty(conf->inactive_list + hash))
 457                 atomic_inc(&conf->empty_inactive_list_nr);
 458 out:
 459         return sh;
 460 }
 461
 462 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
 463 static void free_stripe_pages(struct stripe_head *sh)
 464 {
 465         int i;
 466         struct page *p;
 467
 468         /* Have not allocate page pool */
 469         if (!sh->pages)
 470                 return;
 471
 472         for (i = 0; i < sh->nr_pages; i++) {
 473                 p = sh->pages[i];
 474                 if (p)
 475                         put_page(p);
 476                 sh->pages[i] = NULL;
 477         }
 478 }
 479
 480 static int alloc_stripe_pages(struct stripe_head *sh, gfp_t gfp)
 481 {
 482         int i;
 483         struct page *p;
 484
 485         for (i = 0; i < sh->nr_pages; i++) {
 486                 /* The page have allocated. */
 487                 if (sh->pages[i])
 488                         continue;
 489
 490                 p = alloc_page(gfp);
 491                 if (!p) {
 492                         free_stripe_pages(sh);
 493                         return -ENOMEM;
 494                 }
 495                 sh->pages[i] = p;
 496         }
 497         return 0;
 498 }
 499
 500 static int
 501 init_stripe_shared_pages(struct stripe_head *sh, struct r5conf *conf, int disks)
 502 {
 503         int nr_pages, cnt;
 504
 505         if (sh->pages)
 506                 return 0;
 507
 508         /* Each of the sh->dev[i] need one conf->stripe_size */
 509         cnt = PAGE_SIZE / conf->stripe_size;
 510         nr_pages = (disks + cnt - 1) / cnt;
 511
 512         sh->pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
 513         if (!sh->pages)
 514                 return -ENOMEM;
 515         sh->nr_pages = nr_pages;
 516         sh->stripes_per_page = cnt;
 517         return 0;
 518 }
 519 #endif
 520
 521 static void shrink_buffers(struct stripe_head *sh)
 522 {
 523         int i;
 524         int num = sh->raid_conf->pool_size;
 525
 526 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
 527         for (i = 0; i < num ; i++) {
 528                 struct page *p;
 529
 530                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 531                 p = sh->dev[i].page;
 532                 if (!p)
 533                         continue;
 534                 sh->dev[i].page = NULL;
 535                 put_page(p);
 536         }
 537 #else
 538         for (i = 0; i < num; i++)
 539                 sh->dev[i].page = NULL;
 540         free_stripe_pages(sh); /* Free pages */
 541 #endif
 542 }
 543
 544 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 545 {
 546         int i;
 547         int num = sh->raid_conf->pool_size;
 548
 549 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
 550         for (i = 0; i < num; i++) {
 551                 struct page *page;
 552
 553                 if (!(page = alloc_page(gfp))) {
 554                         return 1;
 555                 }
 556                 sh->dev[i].page = page;
 557                 sh->dev[i].orig_page = page;
 558                 sh->dev[i].offset = 0;
 559         }
 560 #else
 561         if (alloc_stripe_pages(sh, gfp))
 562                 return -ENOMEM;
 563
 564         for (i = 0; i < num; i++) {
 565                 sh->dev[i].page = raid5_get_dev_page(sh, i);
 566                 sh->dev[i].orig_page = sh->dev[i].page;
 567                 sh->dev[i].offset = raid5_get_page_offset(sh, i);
 568         }
 569 #endif
 570         return 0;
 571 }
 572
 573 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 574                             struct stripe_head *sh);
 575
 576 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 577 {
 578         struct r5conf *conf = sh->raid_conf;
 579         int i, seq;
 580
 581         BUG_ON(atomic_read(&sh->count) != 0);
 582         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 583         BUG_ON(stripe_operations_active(sh));
 584         BUG_ON(sh->batch_head);
 585
 586         pr_debug("init_stripe called, stripe %llu\n",
 587                 (unsigned long long)sector);
 588 retry:
 589         seq = read_seqcount_begin(&conf->gen_lock);
 590         sh->generation = conf->generation - previous;
 591         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 592         sh->sector = sector;
 593         stripe_set_idx(sector, conf, previous, sh);
 594         sh->state = 0;
 595
 596         for (i = sh->disks; i--; ) {
 597                 struct r5dev *dev = &sh->dev[i];
 598
 599                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 600                     test_bit(R5_LOCKED, &dev->flags)) {
 601                         pr_err("sector=%llx i=%d %p %p %p %p %d\n",
 602                                (unsigned long long)sh->sector, i, dev->toread,
 603                                dev->read, dev->towrite, dev->written,
 604                                test_bit(R5_LOCKED, &dev->flags));
 605                         WARN_ON(1);
 606                 }
 607                 dev->flags = 0;
 608                 dev->sector = raid5_compute_blocknr(sh, i, previous);
 609         }
 610         if (read_seqcount_retry(&conf->gen_lock, seq))
 611                 goto retry;
 612         sh->overwrite_disks = 0;
 613         insert_hash(conf, sh);
 614         sh->cpu = smp_processor_id();
 615         set_bit(STRIPE_BATCH_READY, &sh->state);
 616 }
 617
 618 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 619                                          short generation)
 620 {
 621         struct stripe_head *sh;
 622
 623         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 624         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 625                 if (sh->sector == sector && sh->generation == generation)
 626                         return sh;
 627         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 628         return NULL;
 629 }
 630
 631 static struct stripe_head *find_get_stripe(struct r5conf *conf,
 632                 sector_t sector, short generation, int hash)
 633 {
 634         int inc_empty_inactive_list_flag;
 635         struct stripe_head *sh;
 636
 637         sh = __find_stripe(conf, sector, generation);
 638         if (!sh)
 639                 return NULL;
 640
 641         if (atomic_inc_not_zero(&sh->count))
 642                 return sh;
 643
 644         /*
 645          * Slow path. The reference count is zero which means the stripe must
 646          * be on a list (sh->lru). Must remove the stripe from the list that
 647          * references it with the device_lock held.
 648          */
 649
 650         spin_lock(&conf->device_lock);
 651         if (!atomic_read(&sh->count)) {
 652                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 653                         atomic_inc(&conf->active_stripes);
 654                 BUG_ON(list_empty(&sh->lru) &&
 655                        !test_bit(STRIPE_EXPANDING, &sh->state));
 656                 inc_empty_inactive_list_flag = 0;
 657                 if (!list_empty(conf->inactive_list + hash))
 658                         inc_empty_inactive_list_flag = 1;
 659                 list_del_init(&sh->lru);
 660                 if (list_empty(conf->inactive_list + hash) &&
 661                     inc_empty_inactive_list_flag)
 662                         atomic_inc(&conf->empty_inactive_list_nr);
 663                 if (sh->group) {
 664                         sh->group->stripes_cnt--;
 665                         sh->group = NULL;
 666                 }
 667         }
 668         atomic_inc(&sh->count);
 669         spin_unlock(&conf->device_lock);
 670
 671         return sh;
 672 }
 673
 674 /*
 675  * Need to check if array has failed when deciding whether to:
 676  *  - start an array
 677  *  - remove non-faulty devices
 678  *  - add a spare
 679  *  - allow a reshape
 680  * This determination is simple when no reshape is happening.
 681  * However if there is a reshape, we need to carefully check
 682  * both the before and after sections.
 683  * This is because some failed devices may only affect one
 684  * of the two sections, and some non-in_sync devices may
 685  * be insync in the section most affected by failed devices.
 686  *
 687  * Most calls to this function hold &conf->device_lock. Calls
 688  * in raid5_run() do not require the lock as no other threads
 689  * have been started yet.
 690  */
 691 int raid5_calc_degraded(struct r5conf *conf)
 692 {
 693         int degraded, degraded2;
 694         int i;
 695
 696         degraded = 0;
 697         for (i = 0; i < conf->previous_raid_disks; i++) {
 698                 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
 699
 700                 if (rdev && test_bit(Faulty, &rdev->flags))
 701                         rdev = READ_ONCE(conf->disks[i].replacement);
 702                 if (!rdev || test_bit(Faulty, &rdev->flags))
 703                         degraded++;
 704                 else if (test_bit(In_sync, &rdev->flags))
 705                         ;
 706                 else
 707                         /* not in-sync or faulty.
 708                          * If the reshape increases the number of devices,
 709                          * this is being recovered by the reshape, so
 710                          * this 'previous' section is not in_sync.
 711                          * If the number of devices is being reduced however,
 712                          * the device can only be part of the array if
 713                          * we are reverting a reshape, so this section will
 714                          * be in-sync.
 715                          */
 716                         if (conf->raid_disks >= conf->previous_raid_disks)
 717                                 degraded++;
 718         }
 719         if (conf->raid_disks == conf->previous_raid_disks)
 720                 return degraded;
 721         degraded2 = 0;
 722         for (i = 0; i < conf->raid_disks; i++) {
 723                 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
 724
 725                 if (rdev && test_bit(Faulty, &rdev->flags))
 726                         rdev = READ_ONCE(conf->disks[i].replacement);
 727                 if (!rdev || test_bit(Faulty, &rdev->flags))
 728                         degraded2++;
 729                 else if (test_bit(In_sync, &rdev->flags))
 730                         ;
 731                 else
 732                         /* not in-sync or faulty.
 733                          * If reshape increases the number of devices, this
 734                          * section has already been recovered, else it
 735                          * almost certainly hasn't.
 736                          */
 737                         if (conf->raid_disks <= conf->previous_raid_disks)
 738                                 degraded2++;
 739         }
 740         if (degraded2 > degraded)
 741                 return degraded2;
 742         return degraded;
 743 }
 744
 745 static bool has_failed(struct r5conf *conf)
 746 {
 747         int degraded = conf->mddev->degraded;
 748
 749         if (test_bit(MD_BROKEN, &conf->mddev->flags))
 750                 return true;
 751
 752         if (conf->mddev->reshape_position != MaxSector)
 753                 degraded = raid5_calc_degraded(conf);
 754
 755         return degraded > conf->max_degraded;
 756 }
 757
 758 enum stripe_result {
 759         STRIPE_SUCCESS = 0,
 760         STRIPE_RETRY,
 761         STRIPE_SCHEDULE_AND_RETRY,
 762         STRIPE_FAIL,
 763         STRIPE_WAIT_RESHAPE,
 764 };
 765
 766 struct stripe_request_ctx {
 767         /* a reference to the last stripe_head for batching */
 768         struct stripe_head *batch_last;
 769
 770         /* first sector in the request */
 771         sector_t first_sector;
 772
 773         /* last sector in the request */
 774         sector_t last_sector;
 775
 776         /*
 777          * bitmap to track stripe sectors that have been added to stripes
 778          * add one to account for unaligned requests
 779          */
 780         DECLARE_BITMAP(sectors_to_do, RAID5_MAX_REQ_STRIPES + 1);
 781
 782         /* the request had REQ_PREFLUSH, cleared after the first stripe_head */
 783         bool do_flush;
 784 };
 785
 786 /*
 787  * Block until another thread clears R5_INACTIVE_BLOCKED or
 788  * there are fewer than 3/4 the maximum number of active stripes
 789  * and there is an inactive stripe available.
 790  */
 791 static bool is_inactive_blocked(struct r5conf *conf, int hash)
 792 {
 793         if (list_empty(conf->inactive_list + hash))
 794                 return false;
 795
 796         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
 797                 return true;
 798
 799         return (atomic_read(&conf->active_stripes) <
 800                 (conf->max_nr_stripes * 3 / 4));
 801 }
 802
 803 struct stripe_head *raid5_get_active_stripe(struct r5conf *conf,
 804                 struct stripe_request_ctx *ctx, sector_t sector,
 805                 unsigned int flags)
 806 {
 807         struct stripe_head *sh;
 808         int hash = stripe_hash_locks_hash(conf, sector);
 809         int previous = !!(flags & R5_GAS_PREVIOUS);
 810
 811         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 812
 813         spin_lock_irq(conf->hash_locks + hash);
 814
 815         for (;;) {
 816                 if (!(flags & R5_GAS_NOQUIESCE) && conf->quiesce) {
 817                         /*
 818                          * Must release the reference to batch_last before
 819                          * waiting, on quiesce, otherwise the batch_last will
 820                          * hold a reference to a stripe and raid5_quiesce()
 821                          * will deadlock waiting for active_stripes to go to
 822                          * zero.
 823                          */
 824                         if (ctx && ctx->batch_last) {
 825                                 raid5_release_stripe(ctx->batch_last);
 826                                 ctx->batch_last = NULL;
 827                         }
 828
 829                         wait_event_lock_irq(conf->wait_for_quiescent,
 830                                             !conf->quiesce,
 831                                             *(conf->hash_locks + hash));
 832                 }
 833
 834                 sh = find_get_stripe(conf, sector, conf->generation - previous,
 835                                      hash);
 836                 if (sh)
 837                         break;
 838
 839                 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 840                         sh = get_free_stripe(conf, hash);
 841                         if (sh) {
 842                                 r5c_check_stripe_cache_usage(conf);
 843                                 init_stripe(sh, sector, previous);
 844                                 atomic_inc(&sh->count);
 845                                 break;
 846                         }
 847
 848                         if (!test_bit(R5_DID_ALLOC, &conf->cache_state))
 849                                 set_bit(R5_ALLOC_MORE, &conf->cache_state);
 850                 }
 851
 852                 if (flags & R5_GAS_NOBLOCK)
 853                         break;
 854
 855                 set_bit(R5_INACTIVE_BLOCKED, &conf->cache_state);
 856                 r5l_wake_reclaim(conf->log, 0);
 857
 858                 /* release batch_last before wait to avoid risk of deadlock */
 859                 if (ctx && ctx->batch_last) {
 860                         raid5_release_stripe(ctx->batch_last);
 861                         ctx->batch_last = NULL;
 862                 }
 863
 864                 wait_event_lock_irq(conf->wait_for_stripe,
 865                                     is_inactive_blocked(conf, hash),
 866                                     *(conf->hash_locks + hash));
 867                 clear_bit(R5_INACTIVE_BLOCKED, &conf->cache_state);
 868         }
 869
 870         spin_unlock_irq(conf->hash_locks + hash);
 871         return sh;
 872 }
 873
 874 static bool is_full_stripe_write(struct stripe_head *sh)
 875 {
 876         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 877         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 878 }
 879
 880 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 881                 __acquires(&sh1->stripe_lock)
 882                 __acquires(&sh2->stripe_lock)
 883 {
 884         if (sh1 > sh2) {
 885                 spin_lock_irq(&sh2->stripe_lock);
 886                 spin_lock_nested(&sh1->stripe_lock, 1);
 887         } else {
 888                 spin_lock_irq(&sh1->stripe_lock);
 889                 spin_lock_nested(&sh2->stripe_lock, 1);
 890         }
 891 }
 892
 893 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 894                 __releases(&sh1->stripe_lock)
 895                 __releases(&sh2->stripe_lock)
 896 {
 897         spin_unlock(&sh1->stripe_lock);
 898         spin_unlock_irq(&sh2->stripe_lock);
 899 }
 900
 901 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 902 static bool stripe_can_batch(struct stripe_head *sh)
 903 {
 904         struct r5conf *conf = sh->raid_conf;
 905
 906         if (raid5_has_log(conf) || raid5_has_ppl(conf))
 907                 return false;
 908         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 909                is_full_stripe_write(sh);
 910 }
 911
 912 /* we only do back search */
 913 static void stripe_add_to_batch_list(struct r5conf *conf,
 914                 struct stripe_head *sh, struct stripe_head *last_sh)
 915 {
 916         struct stripe_head *head;
 917         sector_t head_sector, tmp_sec;
 918         int hash;
 919         int dd_idx;
 920
 921         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 922         tmp_sec = sh->sector;
 923         if (!sector_div(tmp_sec, conf->chunk_sectors))
 924                 return;
 925         head_sector = sh->sector - RAID5_STRIPE_SECTORS(conf);
 926
 927         if (last_sh && head_sector == last_sh->sector) {
 928                 head = last_sh;
 929                 atomic_inc(&head->count);
 930         } else {
 931                 hash = stripe_hash_locks_hash(conf, head_sector);
 932                 spin_lock_irq(conf->hash_locks + hash);
 933                 head = find_get_stripe(conf, head_sector, conf->generation,
 934                                        hash);
 935                 spin_unlock_irq(conf->hash_locks + hash);
 936                 if (!head)
 937                         return;
 938                 if (!stripe_can_batch(head))
 939                         goto out;
 940         }
 941
 942         lock_two_stripes(head, sh);
 943         /* clear_batch_ready clear the flag */
 944         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 945                 goto unlock_out;
 946
 947         if (sh->batch_head)
 948                 goto unlock_out;
 949
 950         dd_idx = 0;
 951         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 952                 dd_idx++;
 953         if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
 954             bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 955                 goto unlock_out;
 956
 957         if (head->batch_head) {
 958                 spin_lock(&head->batch_head->batch_lock);
 959                 /* This batch list is already running */
 960                 if (!stripe_can_batch(head)) {
 961                         spin_unlock(&head->batch_head->batch_lock);
 962                         goto unlock_out;
 963                 }
 964                 /*
 965                  * We must assign batch_head of this stripe within the
 966                  * batch_lock, otherwise clear_batch_ready of batch head
 967                  * stripe could clear BATCH_READY bit of this stripe and
 968                  * this stripe->batch_head doesn't get assigned, which
 969                  * could confuse clear_batch_ready for this stripe
 970                  */
 971                 sh->batch_head = head->batch_head;
 972
 973                 /*
 974                  * at this point, head's BATCH_READY could be cleared, but we
 975                  * can still add the stripe to batch list
 976                  */
 977                 list_add(&sh->batch_list, &head->batch_list);
 978                 spin_unlock(&head->batch_head->batch_lock);
 979         } else {
 980                 head->batch_head = head;
 981                 sh->batch_head = head->batch_head;
 982                 spin_lock(&head->batch_lock);
 983                 list_add_tail(&sh->batch_list, &head->batch_list);
 984                 spin_unlock(&head->batch_lock);
 985         }
 986
 987         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 988                 if (atomic_dec_return(&conf->preread_active_stripes)
 989                     < IO_THRESHOLD)
 990                         md_wakeup_thread(conf->mddev->thread);
 991
 992         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 993                 int seq = sh->bm_seq;
 994                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 995                     sh->batch_head->bm_seq > seq)
 996                         seq = sh->batch_head->bm_seq;
 997                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 998                 sh->batch_head->bm_seq = seq;
 999         }
1000
1001         atomic_inc(&sh->count);
1002 unlock_out:
1003         unlock_two_stripes(head, sh);
1004 out:
1005         raid5_release_stripe(head);
1006 }
1007
1008 /* Determine if 'data_offset' or 'new_data_offset' should be used
1009  * in this stripe_head.
1010  */
1011 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
1012 {
1013         sector_t progress = conf->reshape_progress;
1014         /* Need a memory barrier to make sure we see the value
1015          * of conf->generation, or ->data_offset that was set before
1016          * reshape_progress was updated.
1017          */
1018         smp_rmb();
1019         if (progress == MaxSector)
1020                 return 0;
1021         if (sh->generation == conf->generation - 1)
1022                 return 0;
1023         /* We are in a reshape, and this is a new-generation stripe,
1024          * so use new_data_offset.
1025          */
1026         return 1;
1027 }
1028
1029 static void dispatch_bio_list(struct bio_list *tmp)
1030 {
1031         struct bio *bio;
1032
1033         while ((bio = bio_list_pop(tmp)))
1034                 submit_bio_noacct(bio);
1035 }
1036
1037 static int cmp_stripe(void *priv, const struct list_head *a,
1038                       const struct list_head *b)
1039 {
1040         const struct r5pending_data *da = list_entry(a,
1041                                 struct r5pending_data, sibling);
1042         const struct r5pending_data *db = list_entry(b,
1043                                 struct r5pending_data, sibling);
1044         if (da->sector > db->sector)
1045                 return 1;
1046         if (da->sector < db->sector)
1047                 return -1;
1048         return 0;
1049 }
1050
1051 static void dispatch_defer_bios(struct r5conf *conf, int target,
1052                                 struct bio_list *list)
1053 {
1054         struct r5pending_data *data;
1055         struct list_head *first, *next = NULL;
1056         int cnt = 0;
1057
1058         if (conf->pending_data_cnt == 0)
1059                 return;
1060
1061         list_sort(NULL, &conf->pending_list, cmp_stripe);
1062
1063         first = conf->pending_list.next;
1064
1065         /* temporarily move the head */
1066         if (conf->next_pending_data)
1067                 list_move_tail(&conf->pending_list,
1068                                 &conf->next_pending_data->sibling);
1069
1070         while (!list_empty(&conf->pending_list)) {
1071                 data = list_first_entry(&conf->pending_list,
1072                         struct r5pending_data, sibling);
1073                 if (&data->sibling == first)
1074                         first = data->sibling.next;
1075                 next = data->sibling.next;
1076
1077                 bio_list_merge(list, &data->bios);
1078                 list_move(&data->sibling, &conf->free_list);
1079                 cnt++;
1080                 if (cnt >= target)
1081                         break;
1082         }
1083         conf->pending_data_cnt -= cnt;
1084         BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
1085
1086         if (next != &conf->pending_list)
1087                 conf->next_pending_data = list_entry(next,
1088                                 struct r5pending_data, sibling);
1089         else
1090                 conf->next_pending_data = NULL;
1091         /* list isn't empty */
1092         if (first != &conf->pending_list)
1093                 list_move_tail(&conf->pending_list, first);
1094 }
1095
1096 static void flush_deferred_bios(struct r5conf *conf)
1097 {
1098         struct bio_list tmp = BIO_EMPTY_LIST;
1099
1100         if (conf->pending_data_cnt == 0)
1101                 return;
1102
1103         spin_lock(&conf->pending_bios_lock);
1104         dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
1105         BUG_ON(conf->pending_data_cnt != 0);
1106         spin_unlock(&conf->pending_bios_lock);
1107
1108         dispatch_bio_list(&tmp);
1109 }
1110
1111 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
1112                                 struct bio_list *bios)
1113 {
1114         struct bio_list tmp = BIO_EMPTY_LIST;
1115         struct r5pending_data *ent;
1116
1117         spin_lock(&conf->pending_bios_lock);
1118         ent = list_first_entry(&conf->free_list, struct r5pending_data,
1119                                                         sibling);
1120         list_move_tail(&ent->sibling, &conf->pending_list);
1121         ent->sector = sector;
1122         bio_list_init(&ent->bios);
1123         bio_list_merge(&ent->bios, bios);
1124         conf->pending_data_cnt++;
1125         if (conf->pending_data_cnt >= PENDING_IO_MAX)
1126                 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
1127
1128         spin_unlock(&conf->pending_bios_lock);
1129
1130         dispatch_bio_list(&tmp);
1131 }
1132
1133 static void
1134 raid5_end_read_request(struct bio *bi);
1135 static void
1136 raid5_end_write_request(struct bio *bi);
1137
1138 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
1139 {
1140         struct r5conf *conf = sh->raid_conf;
1141         int i, disks = sh->disks;
1142         struct stripe_head *head_sh = sh;
1143         struct bio_list pending_bios = BIO_EMPTY_LIST;
1144         struct r5dev *dev;
1145         bool should_defer;
1146
1147         might_sleep();
1148
1149         if (log_stripe(sh, s) == 0)
1150                 return;
1151
1152         should_defer = conf->batch_bio_dispatch && conf->group_cnt;
1153
1154         for (i = disks; i--; ) {
1155                 enum req_op op;
1156                 blk_opf_t op_flags = 0;
1157                 int replace_only = 0;
1158                 struct bio *bi, *rbi;
1159                 struct md_rdev *rdev, *rrdev = NULL;
1160
1161                 sh = head_sh;
1162                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1163                         op = REQ_OP_WRITE;
1164                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1165                                 op_flags = REQ_FUA;
1166                         if (test_bit(R5_Discard, &sh->dev[i].flags))
1167                                 op = REQ_OP_DISCARD;
1168                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1169                         op = REQ_OP_READ;
1170                 else if (test_and_clear_bit(R5_WantReplace,
1171                                             &sh->dev[i].flags)) {
1172                         op = REQ_OP_WRITE;
1173                         replace_only = 1;
1174                 } else
1175                         continue;
1176                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1177                         op_flags |= REQ_SYNC;
1178
1179 again:
1180                 dev = &sh->dev[i];
1181                 bi = &dev->req;
1182                 rbi = &dev->rreq; /* For writing to replacement */
1183
1184                 rdev = conf->disks[i].rdev;
1185                 rrdev = conf->disks[i].replacement;
1186                 if (op_is_write(op)) {
1187                         if (replace_only)
1188                                 rdev = NULL;
1189                         if (rdev == rrdev)
1190                                 /* We raced and saw duplicates */
1191                                 rrdev = NULL;
1192                 } else {
1193                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1194                                 rdev = rrdev;
1195                         rrdev = NULL;
1196                 }
1197
1198                 if (rdev && test_bit(Faulty, &rdev->flags))
1199                         rdev = NULL;
1200                 if (rdev)
1201                         atomic_inc(&rdev->nr_pending);
1202                 if (rrdev && test_bit(Faulty, &rrdev->flags))
1203                         rrdev = NULL;
1204                 if (rrdev)
1205                         atomic_inc(&rrdev->nr_pending);
1206
1207                 /* We have already checked bad blocks for reads.  Now
1208                  * need to check for writes.  We never accept write errors
1209                  * on the replacement, so we don't to check rrdev.
1210                  */
1211                 while (op_is_write(op) && rdev &&
1212                        test_bit(WriteErrorSeen, &rdev->flags)) {
1213                         int bad = rdev_has_badblock(rdev, sh->sector,
1214                                                     RAID5_STRIPE_SECTORS(conf));
1215                         if (!bad)
1216                                 break;
1217
1218                         if (bad < 0) {
1219                                 set_bit(BlockedBadBlocks, &rdev->flags);
1220                                 if (!conf->mddev->external &&
1221                                     conf->mddev->sb_flags) {
1222                                         /* It is very unlikely, but we might
1223                                          * still need to write out the
1224                                          * bad block log - better give it
1225                                          * a chance*/
1226                                         md_check_recovery(conf->mddev);
1227                                 }
1228                                 /*
1229                                  * Because md_wait_for_blocked_rdev
1230                                  * will dec nr_pending, we must
1231                                  * increment it first.
1232                                  */
1233                                 atomic_inc(&rdev->nr_pending);
1234                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
1235                         } else {
1236                                 /* Acknowledged bad block - skip the write */
1237                                 rdev_dec_pending(rdev, conf->mddev);
1238                                 rdev = NULL;
1239                         }
1240                 }
1241
1242                 if (rdev) {
1243                         set_bit(STRIPE_IO_STARTED, &sh->state);
1244
1245                         bio_init(bi, rdev->bdev, &dev->vec, 1, op | op_flags);
1246                         bi->bi_end_io = op_is_write(op)
1247                                 ? raid5_end_write_request
1248                                 : raid5_end_read_request;
1249                         bi->bi_private = sh;
1250
1251                         pr_debug("%s: for %llu schedule op %d on disc %d\n",
1252                                 __func__, (unsigned long long)sh->sector,
1253                                 bi->bi_opf, i);
1254                         atomic_inc(&sh->count);
1255                         if (sh != head_sh)
1256                                 atomic_inc(&head_sh->count);
1257                         if (use_new_offset(conf, sh))
1258                                 bi->bi_iter.bi_sector = (sh->sector
1259                                                  + rdev->new_data_offset);
1260                         else
1261                                 bi->bi_iter.bi_sector = (sh->sector
1262                                                  + rdev->data_offset);
1263                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1264                                 bi->bi_opf |= REQ_NOMERGE;
1265
1266                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1267                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1268
1269                         if (!op_is_write(op) &&
1270                             test_bit(R5_InJournal, &sh->dev[i].flags))
1271                                 /*
1272                                  * issuing read for a page in journal, this
1273                                  * must be preparing for prexor in rmw; read
1274                                  * the data into orig_page
1275                                  */
1276                                 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1277                         else
1278                                 sh->dev[i].vec.bv_page = sh->dev[i].page;
1279                         bi->bi_vcnt = 1;
1280                         bi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1281                         bi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1282                         bi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1283                         /*
1284                          * If this is discard request, set bi_vcnt 0. We don't
1285                          * want to confuse SCSI because SCSI will replace payload
1286                          */
1287                         if (op == REQ_OP_DISCARD)
1288                                 bi->bi_vcnt = 0;
1289                         if (rrdev)
1290                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1291
1292                         mddev_trace_remap(conf->mddev, bi, sh->dev[i].sector);
1293                         if (should_defer && op_is_write(op))
1294                                 bio_list_add(&pending_bios, bi);
1295                         else
1296                                 submit_bio_noacct(bi);
1297                 }
1298                 if (rrdev) {
1299                         set_bit(STRIPE_IO_STARTED, &sh->state);
1300
1301                         bio_init(rbi, rrdev->bdev, &dev->rvec, 1, op | op_flags);
1302                         BUG_ON(!op_is_write(op));
1303                         rbi->bi_end_io = raid5_end_write_request;
1304                         rbi->bi_private = sh;
1305
1306                         pr_debug("%s: for %llu schedule op %d on "
1307                                  "replacement disc %d\n",
1308                                 __func__, (unsigned long long)sh->sector,
1309                                 rbi->bi_opf, i);
1310                         atomic_inc(&sh->count);
1311                         if (sh != head_sh)
1312                                 atomic_inc(&head_sh->count);
1313                         if (use_new_offset(conf, sh))
1314                                 rbi->bi_iter.bi_sector = (sh->sector
1315                                                   + rrdev->new_data_offset);
1316                         else
1317                                 rbi->bi_iter.bi_sector = (sh->sector
1318                                                   + rrdev->data_offset);
1319                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1320                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1321                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1322                         rbi->bi_vcnt = 1;
1323                         rbi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1324                         rbi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1325                         rbi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1326                         /*
1327                          * If this is discard request, set bi_vcnt 0. We don't
1328                          * want to confuse SCSI because SCSI will replace payload
1329                          */
1330                         if (op == REQ_OP_DISCARD)
1331                                 rbi->bi_vcnt = 0;
1332                         mddev_trace_remap(conf->mddev, rbi, sh->dev[i].sector);
1333                         if (should_defer && op_is_write(op))
1334                                 bio_list_add(&pending_bios, rbi);
1335                         else
1336                                 submit_bio_noacct(rbi);
1337                 }
1338                 if (!rdev && !rrdev) {
1339                         pr_debug("skip op %d on disc %d for sector %llu\n",
1340                                 bi->bi_opf, i, (unsigned long long)sh->sector);
1341                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1342                         set_bit(STRIPE_HANDLE, &sh->state);
1343                 }
1344
1345                 if (!head_sh->batch_head)
1346                         continue;
1347                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1348                                       batch_list);
1349                 if (sh != head_sh)
1350                         goto again;
1351         }
1352
1353         if (should_defer && !bio_list_empty(&pending_bios))
1354                 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1355 }
1356
1357 static struct dma_async_tx_descriptor *
1358 async_copy_data(int frombio, struct bio *bio, struct page **page,
1359         unsigned int poff, sector_t sector, struct dma_async_tx_descriptor *tx,
1360         struct stripe_head *sh, int no_skipcopy)
1361 {
1362         struct bio_vec bvl;
1363         struct bvec_iter iter;
1364         struct page *bio_page;
1365         int page_offset;
1366         struct async_submit_ctl submit;
1367         enum async_tx_flags flags = 0;
1368         struct r5conf *conf = sh->raid_conf;
1369
1370         if (bio->bi_iter.bi_sector >= sector)
1371                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1372         else
1373                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1374
1375         if (frombio)
1376                 flags |= ASYNC_TX_FENCE;
1377         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1378
1379         bio_for_each_segment(bvl, bio, iter) {
1380                 int len = bvl.bv_len;
1381                 int clen;
1382                 int b_offset = 0;
1383
1384                 if (page_offset < 0) {
1385                         b_offset = -page_offset;
1386                         page_offset += b_offset;
1387                         len -= b_offset;
1388                 }
1389
1390                 if (len > 0 && page_offset + len > RAID5_STRIPE_SIZE(conf))
1391                         clen = RAID5_STRIPE_SIZE(conf) - page_offset;
1392                 else
1393                         clen = len;
1394
1395                 if (clen > 0) {
1396                         b_offset += bvl.bv_offset;
1397                         bio_page = bvl.bv_page;
1398                         if (frombio) {
1399                                 if (conf->skip_copy &&
1400                                     b_offset == 0 && page_offset == 0 &&
1401                                     clen == RAID5_STRIPE_SIZE(conf) &&
1402                                     !no_skipcopy)
1403                                         *page = bio_page;
1404                                 else
1405                                         tx = async_memcpy(*page, bio_page, page_offset + poff,
1406                                                   b_offset, clen, &submit);
1407                         } else
1408                                 tx = async_memcpy(bio_page, *page, b_offset,
1409                                                   page_offset + poff, clen, &submit);
1410                 }
1411                 /* chain the operations */
1412                 submit.depend_tx = tx;
1413
1414                 if (clen < len) /* hit end of page */
1415                         break;
1416                 page_offset +=  len;
1417         }
1418
1419         return tx;
1420 }
1421
1422 static void ops_complete_biofill(void *stripe_head_ref)
1423 {
1424         struct stripe_head *sh = stripe_head_ref;
1425         int i;
1426         struct r5conf *conf = sh->raid_conf;
1427
1428         pr_debug("%s: stripe %llu\n", __func__,
1429                 (unsigned long long)sh->sector);
1430
1431         /* clear completed biofills */
1432         for (i = sh->disks; i--; ) {
1433                 struct r5dev *dev = &sh->dev[i];
1434
1435                 /* acknowledge completion of a biofill operation */
1436                 /* and check if we need to reply to a read request,
1437                  * new R5_Wantfill requests are held off until
1438                  * !STRIPE_BIOFILL_RUN
1439                  */
1440                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1441                         struct bio *rbi, *rbi2;
1442
1443                         BUG_ON(!dev->read);
1444                         rbi = dev->read;
1445                         dev->read = NULL;
1446                         while (rbi && rbi->bi_iter.bi_sector <
1447                                 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1448                                 rbi2 = r5_next_bio(conf, rbi, dev->sector);
1449                                 bio_endio(rbi);
1450                                 rbi = rbi2;
1451                         }
1452                 }
1453         }
1454         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1455
1456         set_bit(STRIPE_HANDLE, &sh->state);
1457         raid5_release_stripe(sh);
1458 }
1459
1460 static void ops_run_biofill(struct stripe_head *sh)
1461 {
1462         struct dma_async_tx_descriptor *tx = NULL;
1463         struct async_submit_ctl submit;
1464         int i;
1465         struct r5conf *conf = sh->raid_conf;
1466
1467         BUG_ON(sh->batch_head);
1468         pr_debug("%s: stripe %llu\n", __func__,
1469                 (unsigned long long)sh->sector);
1470
1471         for (i = sh->disks; i--; ) {
1472                 struct r5dev *dev = &sh->dev[i];
1473                 if (test_bit(R5_Wantfill, &dev->flags)) {
1474                         struct bio *rbi;
1475                         spin_lock_irq(&sh->stripe_lock);
1476                         dev->read = rbi = dev->toread;
1477                         dev->toread = NULL;
1478                         spin_unlock_irq(&sh->stripe_lock);
1479                         while (rbi && rbi->bi_iter.bi_sector <
1480                                 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1481                                 tx = async_copy_data(0, rbi, &dev->page,
1482                                                      dev->offset,
1483                                                      dev->sector, tx, sh, 0);
1484                                 rbi = r5_next_bio(conf, rbi, dev->sector);
1485                         }
1486                 }
1487         }
1488
1489         atomic_inc(&sh->count);
1490         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1491         async_trigger_callback(&submit);
1492 }
1493
1494 static void mark_target_uptodate(struct stripe_head *sh, int target)
1495 {
1496         struct r5dev *tgt;
1497
1498         if (target < 0)
1499                 return;
1500
1501         tgt = &sh->dev[target];
1502         set_bit(R5_UPTODATE, &tgt->flags);
1503         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1504         clear_bit(R5_Wantcompute, &tgt->flags);
1505 }
1506
1507 static void ops_complete_compute(void *stripe_head_ref)
1508 {
1509         struct stripe_head *sh = stripe_head_ref;
1510
1511         pr_debug("%s: stripe %llu\n", __func__,
1512                 (unsigned long long)sh->sector);
1513
1514         /* mark the computed target(s) as uptodate */
1515         mark_target_uptodate(sh, sh->ops.target);
1516         mark_target_uptodate(sh, sh->ops.target2);
1517
1518         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1519         if (sh->check_state == check_state_compute_run)
1520                 sh->check_state = check_state_compute_result;
1521         set_bit(STRIPE_HANDLE, &sh->state);
1522         raid5_release_stripe(sh);
1523 }
1524
1525 /* return a pointer to the address conversion region of the scribble buffer */
1526 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1527 {
1528         return percpu->scribble + i * percpu->scribble_obj_size;
1529 }
1530
1531 /* return a pointer to the address conversion region of the scribble buffer */
1532 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1533                                  struct raid5_percpu *percpu, int i)
1534 {
1535         return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1536 }
1537
1538 /*
1539  * Return a pointer to record offset address.
1540  */
1541 static unsigned int *
1542 to_addr_offs(struct stripe_head *sh, struct raid5_percpu *percpu)
1543 {
1544         return (unsigned int *) (to_addr_conv(sh, percpu, 0) + sh->disks + 2);
1545 }
1546
1547 static struct dma_async_tx_descriptor *
1548 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1549 {
1550         int disks = sh->disks;
1551         struct page **xor_srcs = to_addr_page(percpu, 0);
1552         unsigned int *off_srcs = to_addr_offs(sh, percpu);
1553         int target = sh->ops.target;
1554         struct r5dev *tgt = &sh->dev[target];
1555         struct page *xor_dest = tgt->page;
1556         unsigned int off_dest = tgt->offset;
1557         int count = 0;
1558         struct dma_async_tx_descriptor *tx;
1559         struct async_submit_ctl submit;
1560         int i;
1561
1562         BUG_ON(sh->batch_head);
1563
1564         pr_debug("%s: stripe %llu block: %d\n",
1565                 __func__, (unsigned long long)sh->sector, target);
1566         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1567
1568         for (i = disks; i--; ) {
1569                 if (i != target) {
1570                         off_srcs[count] = sh->dev[i].offset;
1571                         xor_srcs[count++] = sh->dev[i].page;
1572                 }
1573         }
1574
1575         atomic_inc(&sh->count);
1576
1577         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1578                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1579         if (unlikely(count == 1))
1580                 tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
1581                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1582         else
1583                 tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1584                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1585
1586         return tx;
1587 }
1588
1589 /* set_syndrome_sources - populate source buffers for gen_syndrome
1590  * @srcs - (struct page *) array of size sh->disks
1591  * @offs - (unsigned int) array of offset for each page
1592  * @sh - stripe_head to parse
1593  *
1594  * Populates srcs in proper layout order for the stripe and returns the
1595  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1596  * destination buffer is recorded in srcs[count] and the Q destination
1597  * is recorded in srcs[count+1]].
1598  */
1599 static int set_syndrome_sources(struct page **srcs,
1600                                 unsigned int *offs,
1601                                 struct stripe_head *sh,
1602                                 int srctype)
1603 {
1604         int disks = sh->disks;
1605         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1606         int d0_idx = raid6_d0(sh);
1607         int count;
1608         int i;
1609
1610         for (i = 0; i < disks; i++)
1611                 srcs[i] = NULL;
1612
1613         count = 0;
1614         i = d0_idx;
1615         do {
1616                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1617                 struct r5dev *dev = &sh->dev[i];
1618
1619                 if (i == sh->qd_idx || i == sh->pd_idx ||
1620                     (srctype == SYNDROME_SRC_ALL) ||
1621                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1622                      (test_bit(R5_Wantdrain, &dev->flags) ||
1623                       test_bit(R5_InJournal, &dev->flags))) ||
1624                     (srctype == SYNDROME_SRC_WRITTEN &&
1625                      (dev->written ||
1626                       test_bit(R5_InJournal, &dev->flags)))) {
1627                         if (test_bit(R5_InJournal, &dev->flags))
1628                                 srcs[slot] = sh->dev[i].orig_page;
1629                         else
1630                                 srcs[slot] = sh->dev[i].page;
1631                         /*
1632                          * For R5_InJournal, PAGE_SIZE must be 4KB and will
1633                          * not shared page. In that case, dev[i].offset
1634                          * is 0.
1635                          */
1636                         offs[slot] = sh->dev[i].offset;
1637                 }
1638                 i = raid6_next_disk(i, disks);
1639         } while (i != d0_idx);
1640
1641         return syndrome_disks;
1642 }
1643
1644 static struct dma_async_tx_descriptor *
1645 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1646 {
1647         int disks = sh->disks;
1648         struct page **blocks = to_addr_page(percpu, 0);
1649         unsigned int *offs = to_addr_offs(sh, percpu);
1650         int target;
1651         int qd_idx = sh->qd_idx;
1652         struct dma_async_tx_descriptor *tx;
1653         struct async_submit_ctl submit;
1654         struct r5dev *tgt;
1655         struct page *dest;
1656         unsigned int dest_off;
1657         int i;
1658         int count;
1659
1660         BUG_ON(sh->batch_head);
1661         if (sh->ops.target < 0)
1662                 target = sh->ops.target2;
1663         else if (sh->ops.target2 < 0)
1664                 target = sh->ops.target;
1665         else
1666                 /* we should only have one valid target */
1667                 BUG();
1668         BUG_ON(target < 0);
1669         pr_debug("%s: stripe %llu block: %d\n",
1670                 __func__, (unsigned long long)sh->sector, target);
1671
1672         tgt = &sh->dev[target];
1673         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1674         dest = tgt->page;
1675         dest_off = tgt->offset;
1676
1677         atomic_inc(&sh->count);
1678
1679         if (target == qd_idx) {
1680                 count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1681                 blocks[count] = NULL; /* regenerating p is not necessary */
1682                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1683                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1684                                   ops_complete_compute, sh,
1685                                   to_addr_conv(sh, percpu, 0));
1686                 tx = async_gen_syndrome(blocks, offs, count+2,
1687                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1688         } else {
1689                 /* Compute any data- or p-drive using XOR */
1690                 count = 0;
1691                 for (i = disks; i-- ; ) {
1692                         if (i == target || i == qd_idx)
1693                                 continue;
1694                         offs[count] = sh->dev[i].offset;
1695                         blocks[count++] = sh->dev[i].page;
1696                 }
1697
1698                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1699                                   NULL, ops_complete_compute, sh,
1700                                   to_addr_conv(sh, percpu, 0));
1701                 tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1702                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1703         }
1704
1705         return tx;
1706 }
1707
1708 static struct dma_async_tx_descriptor *
1709 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1710 {
1711         int i, count, disks = sh->disks;
1712         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1713         int d0_idx = raid6_d0(sh);
1714         int faila = -1, failb = -1;
1715         int target = sh->ops.target;
1716         int target2 = sh->ops.target2;
1717         struct r5dev *tgt = &sh->dev[target];
1718         struct r5dev *tgt2 = &sh->dev[target2];
1719         struct dma_async_tx_descriptor *tx;
1720         struct page **blocks = to_addr_page(percpu, 0);
1721         unsigned int *offs = to_addr_offs(sh, percpu);
1722         struct async_submit_ctl submit;
1723
1724         BUG_ON(sh->batch_head);
1725         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1726                  __func__, (unsigned long long)sh->sector, target, target2);
1727         BUG_ON(target < 0 || target2 < 0);
1728         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1729         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1730
1731         /* we need to open-code set_syndrome_sources to handle the
1732          * slot number conversion for 'faila' and 'failb'
1733          */
1734         for (i = 0; i < disks ; i++) {
1735                 offs[i] = 0;
1736                 blocks[i] = NULL;
1737         }
1738         count = 0;
1739         i = d0_idx;
1740         do {
1741                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1742
1743                 offs[slot] = sh->dev[i].offset;
1744                 blocks[slot] = sh->dev[i].page;
1745
1746                 if (i == target)
1747                         faila = slot;
1748                 if (i == target2)
1749                         failb = slot;
1750                 i = raid6_next_disk(i, disks);
1751         } while (i != d0_idx);
1752
1753         BUG_ON(faila == failb);
1754         if (failb < faila)
1755                 swap(faila, failb);
1756         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1757                  __func__, (unsigned long long)sh->sector, faila, failb);
1758
1759         atomic_inc(&sh->count);
1760
1761         if (failb == syndrome_disks+1) {
1762                 /* Q disk is one of the missing disks */
1763                 if (faila == syndrome_disks) {
1764                         /* Missing P+Q, just recompute */
1765                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1766                                           ops_complete_compute, sh,
1767                                           to_addr_conv(sh, percpu, 0));
1768                         return async_gen_syndrome(blocks, offs, syndrome_disks+2,
1769                                                   RAID5_STRIPE_SIZE(sh->raid_conf),
1770                                                   &submit);
1771                 } else {
1772                         struct page *dest;
1773                         unsigned int dest_off;
1774                         int data_target;
1775                         int qd_idx = sh->qd_idx;
1776
1777                         /* Missing D+Q: recompute D from P, then recompute Q */
1778                         if (target == qd_idx)
1779                                 data_target = target2;
1780                         else
1781                                 data_target = target;
1782
1783                         count = 0;
1784                         for (i = disks; i-- ; ) {
1785                                 if (i == data_target || i == qd_idx)
1786                                         continue;
1787                                 offs[count] = sh->dev[i].offset;
1788                                 blocks[count++] = sh->dev[i].page;
1789                         }
1790                         dest = sh->dev[data_target].page;
1791                         dest_off = sh->dev[data_target].offset;
1792                         init_async_submit(&submit,
1793                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1794                                           NULL, NULL, NULL,
1795                                           to_addr_conv(sh, percpu, 0));
1796                         tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1797                                        RAID5_STRIPE_SIZE(sh->raid_conf),
1798                                        &submit);
1799
1800                         count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1801                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1802                                           ops_complete_compute, sh,
1803                                           to_addr_conv(sh, percpu, 0));
1804                         return async_gen_syndrome(blocks, offs, count+2,
1805                                                   RAID5_STRIPE_SIZE(sh->raid_conf),
1806                                                   &submit);
1807                 }
1808         } else {
1809                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1810                                   ops_complete_compute, sh,
1811                                   to_addr_conv(sh, percpu, 0));
1812                 if (failb == syndrome_disks) {
1813                         /* We're missing D+P. */
1814                         return async_raid6_datap_recov(syndrome_disks+2,
1815                                                 RAID5_STRIPE_SIZE(sh->raid_conf),
1816                                                 faila,
1817                                                 blocks, offs, &submit);
1818                 } else {
1819                         /* We're missing D+D. */
1820                         return async_raid6_2data_recov(syndrome_disks+2,
1821                                                 RAID5_STRIPE_SIZE(sh->raid_conf),
1822                                                 faila, failb,
1823                                                 blocks, offs, &submit);
1824                 }
1825         }
1826 }
1827
1828 static void ops_complete_prexor(void *stripe_head_ref)
1829 {
1830         struct stripe_head *sh = stripe_head_ref;
1831
1832         pr_debug("%s: stripe %llu\n", __func__,
1833                 (unsigned long long)sh->sector);
1834
1835         if (r5c_is_writeback(sh->raid_conf->log))
1836                 /*
1837                  * raid5-cache write back uses orig_page during prexor.
1838                  * After prexor, it is time to free orig_page
1839                  */
1840                 r5c_release_extra_page(sh);
1841 }
1842
1843 static struct dma_async_tx_descriptor *
1844 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1845                 struct dma_async_tx_descriptor *tx)
1846 {
1847         int disks = sh->disks;
1848         struct page **xor_srcs = to_addr_page(percpu, 0);
1849         unsigned int *off_srcs = to_addr_offs(sh, percpu);
1850         int count = 0, pd_idx = sh->pd_idx, i;
1851         struct async_submit_ctl submit;
1852
1853         /* existing parity data subtracted */
1854         unsigned int off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
1855         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1856
1857         BUG_ON(sh->batch_head);
1858         pr_debug("%s: stripe %llu\n", __func__,
1859                 (unsigned long long)sh->sector);
1860
1861         for (i = disks; i--; ) {
1862                 struct r5dev *dev = &sh->dev[i];
1863                 /* Only process blocks that are known to be uptodate */
1864                 if (test_bit(R5_InJournal, &dev->flags)) {
1865                         /*
1866                          * For this case, PAGE_SIZE must be equal to 4KB and
1867                          * page offset is zero.
1868                          */
1869                         off_srcs[count] = dev->offset;
1870                         xor_srcs[count++] = dev->orig_page;
1871                 } else if (test_bit(R5_Wantdrain, &dev->flags)) {
1872                         off_srcs[count] = dev->offset;
1873                         xor_srcs[count++] = dev->page;
1874                 }
1875         }
1876
1877         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1878                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1879         tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1880                         RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1881
1882         return tx;
1883 }
1884
1885 static struct dma_async_tx_descriptor *
1886 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1887                 struct dma_async_tx_descriptor *tx)
1888 {
1889         struct page **blocks = to_addr_page(percpu, 0);
1890         unsigned int *offs = to_addr_offs(sh, percpu);
1891         int count;
1892         struct async_submit_ctl submit;
1893
1894         pr_debug("%s: stripe %llu\n", __func__,
1895                 (unsigned long long)sh->sector);
1896
1897         count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_WANT_DRAIN);
1898
1899         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1900                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1901         tx = async_gen_syndrome(blocks, offs, count+2,
1902                         RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1903
1904         return tx;
1905 }
1906
1907 static struct dma_async_tx_descriptor *
1908 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1909 {
1910         struct r5conf *conf = sh->raid_conf;
1911         int disks = sh->disks;
1912         int i;
1913         struct stripe_head *head_sh = sh;
1914
1915         pr_debug("%s: stripe %llu\n", __func__,
1916                 (unsigned long long)sh->sector);
1917
1918         for (i = disks; i--; ) {
1919                 struct r5dev *dev;
1920                 struct bio *chosen;
1921
1922                 sh = head_sh;
1923                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1924                         struct bio *wbi;
1925
1926 again:
1927                         dev = &sh->dev[i];
1928                         /*
1929                          * clear R5_InJournal, so when rewriting a page in
1930                          * journal, it is not skipped by r5l_log_stripe()
1931                          */
1932                         clear_bit(R5_InJournal, &dev->flags);
1933                         spin_lock_irq(&sh->stripe_lock);
1934                         chosen = dev->towrite;
1935                         dev->towrite = NULL;
1936                         sh->overwrite_disks = 0;
1937                         BUG_ON(dev->written);
1938                         wbi = dev->written = chosen;
1939                         spin_unlock_irq(&sh->stripe_lock);
1940                         WARN_ON(dev->page != dev->orig_page);
1941
1942                         while (wbi && wbi->bi_iter.bi_sector <
1943                                 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1944                                 if (wbi->bi_opf & REQ_FUA)
1945                                         set_bit(R5_WantFUA, &dev->flags);
1946                                 if (wbi->bi_opf & REQ_SYNC)
1947                                         set_bit(R5_SyncIO, &dev->flags);
1948                                 if (bio_op(wbi) == REQ_OP_DISCARD)
1949                                         set_bit(R5_Discard, &dev->flags);
1950                                 else {
1951                                         tx = async_copy_data(1, wbi, &dev->page,
1952                                                              dev->offset,
1953                                                              dev->sector, tx, sh,
1954                                                              r5c_is_writeback(conf->log));
1955                                         if (dev->page != dev->orig_page &&
1956                                             !r5c_is_writeback(conf->log)) {
1957                                                 set_bit(R5_SkipCopy, &dev->flags);
1958                                                 clear_bit(R5_UPTODATE, &dev->flags);
1959                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1960                                         }
1961                                 }
1962                                 wbi = r5_next_bio(conf, wbi, dev->sector);
1963                         }
1964
1965                         if (head_sh->batch_head) {
1966                                 sh = list_first_entry(&sh->batch_list,
1967                                                       struct stripe_head,
1968                                                       batch_list);
1969                                 if (sh == head_sh)
1970                                         continue;
1971                                 goto again;
1972                         }
1973                 }
1974         }
1975
1976         return tx;
1977 }
1978
1979 static void ops_complete_reconstruct(void *stripe_head_ref)
1980 {
1981         struct stripe_head *sh = stripe_head_ref;
1982         int disks = sh->disks;
1983         int pd_idx = sh->pd_idx;
1984         int qd_idx = sh->qd_idx;
1985         int i;
1986         bool fua = false, sync = false, discard = false;
1987
1988         pr_debug("%s: stripe %llu\n", __func__,
1989                 (unsigned long long)sh->sector);
1990
1991         for (i = disks; i--; ) {
1992                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1993                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1994                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1995         }
1996
1997         for (i = disks; i--; ) {
1998                 struct r5dev *dev = &sh->dev[i];
1999
2000                 if (dev->written || i == pd_idx || i == qd_idx) {
2001                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
2002                                 set_bit(R5_UPTODATE, &dev->flags);
2003                                 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
2004                                         set_bit(R5_Expanded, &dev->flags);
2005                         }
2006                         if (fua)
2007                                 set_bit(R5_WantFUA, &dev->flags);
2008                         if (sync)
2009                                 set_bit(R5_SyncIO, &dev->flags);
2010                 }
2011         }
2012
2013         if (sh->reconstruct_state == reconstruct_state_drain_run)
2014                 sh->reconstruct_state = reconstruct_state_drain_result;
2015         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
2016                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
2017         else {
2018                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
2019                 sh->reconstruct_state = reconstruct_state_result;
2020         }
2021
2022         set_bit(STRIPE_HANDLE, &sh->state);
2023         raid5_release_stripe(sh);
2024 }
2025
2026 static void
2027 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
2028                      struct dma_async_tx_descriptor *tx)
2029 {
2030         int disks = sh->disks;
2031         struct page **xor_srcs;
2032         unsigned int *off_srcs;
2033         struct async_submit_ctl submit;
2034         int count, pd_idx = sh->pd_idx, i;
2035         struct page *xor_dest;
2036         unsigned int off_dest;
2037         int prexor = 0;
2038         unsigned long flags;
2039         int j = 0;
2040         struct stripe_head *head_sh = sh;
2041         int last_stripe;
2042
2043         pr_debug("%s: stripe %llu\n", __func__,
2044                 (unsigned long long)sh->sector);
2045
2046         for (i = 0; i < sh->disks; i++) {
2047                 if (pd_idx == i)
2048                         continue;
2049                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
2050                         break;
2051         }
2052         if (i >= sh->disks) {
2053                 atomic_inc(&sh->count);
2054                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
2055                 ops_complete_reconstruct(sh);
2056                 return;
2057         }
2058 again:
2059         count = 0;
2060         xor_srcs = to_addr_page(percpu, j);
2061         off_srcs = to_addr_offs(sh, percpu);
2062         /* check if prexor is active which means only process blocks
2063          * that are part of a read-modify-write (written)
2064          */
2065         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2066                 prexor = 1;
2067                 off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
2068                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
2069                 for (i = disks; i--; ) {
2070                         struct r5dev *dev = &sh->dev[i];
2071                         if (head_sh->dev[i].written ||
2072                             test_bit(R5_InJournal, &head_sh->dev[i].flags)) {
2073                                 off_srcs[count] = dev->offset;
2074                                 xor_srcs[count++] = dev->page;
2075                         }
2076                 }
2077         } else {
2078                 xor_dest = sh->dev[pd_idx].page;
2079                 off_dest = sh->dev[pd_idx].offset;
2080                 for (i = disks; i--; ) {
2081                         struct r5dev *dev = &sh->dev[i];
2082                         if (i != pd_idx) {
2083                                 off_srcs[count] = dev->offset;
2084                                 xor_srcs[count++] = dev->page;
2085                         }
2086                 }
2087         }
2088
2089         /* 1/ if we prexor'd then the dest is reused as a source
2090          * 2/ if we did not prexor then we are redoing the parity
2091          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
2092          * for the synchronous xor case
2093          */
2094         last_stripe = !head_sh->batch_head ||
2095                 list_first_entry(&sh->batch_list,
2096                                  struct stripe_head, batch_list) == head_sh;
2097         if (last_stripe) {
2098                 flags = ASYNC_TX_ACK |
2099                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
2100
2101                 atomic_inc(&head_sh->count);
2102                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
2103                                   to_addr_conv(sh, percpu, j));
2104         } else {
2105                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
2106                 init_async_submit(&submit, flags, tx, NULL, NULL,
2107                                   to_addr_conv(sh, percpu, j));
2108         }
2109
2110         if (unlikely(count == 1))
2111                 tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
2112                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2113         else
2114                 tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2115                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2116         if (!last_stripe) {
2117                 j++;
2118                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
2119                                       batch_list);
2120                 goto again;
2121         }
2122 }
2123
2124 static void
2125 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
2126                      struct dma_async_tx_descriptor *tx)
2127 {
2128         struct async_submit_ctl submit;
2129         struct page **blocks;
2130         unsigned int *offs;
2131         int count, i, j = 0;
2132         struct stripe_head *head_sh = sh;
2133         int last_stripe;
2134         int synflags;
2135         unsigned long txflags;
2136
2137         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
2138
2139         for (i = 0; i < sh->disks; i++) {
2140                 if (sh->pd_idx == i || sh->qd_idx == i)
2141                         continue;
2142                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
2143                         break;
2144         }
2145         if (i >= sh->disks) {
2146                 atomic_inc(&sh->count);
2147                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2148                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2149                 ops_complete_reconstruct(sh);
2150                 return;
2151         }
2152
2153 again:
2154         blocks = to_addr_page(percpu, j);
2155         offs = to_addr_offs(sh, percpu);
2156
2157         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2158                 synflags = SYNDROME_SRC_WRITTEN;
2159                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
2160         } else {
2161                 synflags = SYNDROME_SRC_ALL;
2162                 txflags = ASYNC_TX_ACK;
2163         }
2164
2165         count = set_syndrome_sources(blocks, offs, sh, synflags);
2166         last_stripe = !head_sh->batch_head ||
2167                 list_first_entry(&sh->batch_list,
2168                                  struct stripe_head, batch_list) == head_sh;
2169
2170         if (last_stripe) {
2171                 atomic_inc(&head_sh->count);
2172                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
2173                                   head_sh, to_addr_conv(sh, percpu, j));
2174         } else
2175                 init_async_submit(&submit, 0, tx, NULL, NULL,
2176                                   to_addr_conv(sh, percpu, j));
2177         tx = async_gen_syndrome(blocks, offs, count+2,
2178                         RAID5_STRIPE_SIZE(sh->raid_conf),  &submit);
2179         if (!last_stripe) {
2180                 j++;
2181                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
2182                                       batch_list);
2183                 goto again;
2184         }
2185 }
2186
2187 static void ops_complete_check(void *stripe_head_ref)
2188 {
2189         struct stripe_head *sh = stripe_head_ref;
2190
2191         pr_debug("%s: stripe %llu\n", __func__,
2192                 (unsigned long long)sh->sector);
2193
2194         sh->check_state = check_state_check_result;
2195         set_bit(STRIPE_HANDLE, &sh->state);
2196         raid5_release_stripe(sh);
2197 }
2198
2199 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2200 {
2201         int disks = sh->disks;
2202         int pd_idx = sh->pd_idx;
2203         int qd_idx = sh->qd_idx;
2204         struct page *xor_dest;
2205         unsigned int off_dest;
2206         struct page **xor_srcs = to_addr_page(percpu, 0);
2207         unsigned int *off_srcs = to_addr_offs(sh, percpu);
2208         struct dma_async_tx_descriptor *tx;
2209         struct async_submit_ctl submit;
2210         int count;
2211         int i;
2212
2213         pr_debug("%s: stripe %llu\n", __func__,
2214                 (unsigned long long)sh->sector);
2215
2216         BUG_ON(sh->batch_head);
2217         count = 0;
2218         xor_dest = sh->dev[pd_idx].page;
2219         off_dest = sh->dev[pd_idx].offset;
2220         off_srcs[count] = off_dest;
2221         xor_srcs[count++] = xor_dest;
2222         for (i = disks; i--; ) {
2223                 if (i == pd_idx || i == qd_idx)
2224                         continue;
2225                 off_srcs[count] = sh->dev[i].offset;
2226                 xor_srcs[count++] = sh->dev[i].page;
2227         }
2228
2229         init_async_submit(&submit, 0, NULL, NULL, NULL,
2230                           to_addr_conv(sh, percpu, 0));
2231         tx = async_xor_val_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2232                            RAID5_STRIPE_SIZE(sh->raid_conf),
2233                            &sh->ops.zero_sum_result, &submit);
2234
2235         atomic_inc(&sh->count);
2236         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2237         tx = async_trigger_callback(&submit);
2238 }
2239
2240 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2241 {
2242         struct page **srcs = to_addr_page(percpu, 0);
2243         unsigned int *offs = to_addr_offs(sh, percpu);
2244         struct async_submit_ctl submit;
2245         int count;
2246
2247         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2248                 (unsigned long long)sh->sector, checkp);
2249
2250         BUG_ON(sh->batch_head);
2251         count = set_syndrome_sources(srcs, offs, sh, SYNDROME_SRC_ALL);
2252         if (!checkp)
2253                 srcs[count] = NULL;
2254
2255         atomic_inc(&sh->count);
2256         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2257                           sh, to_addr_conv(sh, percpu, 0));
2258         async_syndrome_val(srcs, offs, count+2,
2259                            RAID5_STRIPE_SIZE(sh->raid_conf),
2260                            &sh->ops.zero_sum_result, percpu->spare_page, 0, &submit);
2261 }
2262
2263 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2264 {
2265         int overlap_clear = 0, i, disks = sh->disks;
2266         struct dma_async_tx_descriptor *tx = NULL;
2267         struct r5conf *conf = sh->raid_conf;
2268         int level = conf->level;
2269         struct raid5_percpu *percpu;
2270
2271         local_lock(&conf->percpu->lock);
2272         percpu = this_cpu_ptr(conf->percpu);
2273         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2274                 ops_run_biofill(sh);
2275                 overlap_clear++;
2276         }
2277
2278         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2279                 if (level < 6)
2280                         tx = ops_run_compute5(sh, percpu);
2281                 else {
2282                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
2283                                 tx = ops_run_compute6_1(sh, percpu);
2284                         else
2285                                 tx = ops_run_compute6_2(sh, percpu);
2286                 }
2287                 /* terminate the chain if reconstruct is not set to be run */
2288                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2289                         async_tx_ack(tx);
2290         }
2291
2292         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2293                 if (level < 6)
2294                         tx = ops_run_prexor5(sh, percpu, tx);
2295                 else
2296                         tx = ops_run_prexor6(sh, percpu, tx);
2297         }
2298
2299         if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2300                 tx = ops_run_partial_parity(sh, percpu, tx);
2301
2302         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2303                 tx = ops_run_biodrain(sh, tx);
2304                 overlap_clear++;
2305         }
2306
2307         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2308                 if (level < 6)
2309                         ops_run_reconstruct5(sh, percpu, tx);
2310                 else
2311                         ops_run_reconstruct6(sh, percpu, tx);
2312         }
2313
2314         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2315                 if (sh->check_state == check_state_run)
2316                         ops_run_check_p(sh, percpu);
2317                 else if (sh->check_state == check_state_run_q)
2318                         ops_run_check_pq(sh, percpu, 0);
2319                 else if (sh->check_state == check_state_run_pq)
2320                         ops_run_check_pq(sh, percpu, 1);
2321                 else
2322                         BUG();
2323         }
2324
2325         if (overlap_clear && !sh->batch_head) {
2326                 for (i = disks; i--; ) {
2327                         struct r5dev *dev = &sh->dev[i];
2328                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2329                                 wake_up_bit(&dev->flags, R5_Overlap);
2330                 }
2331         }
2332         local_unlock(&conf->percpu->lock);
2333 }
2334
2335 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2336 {
2337 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2338         kfree(sh->pages);
2339 #endif
2340         if (sh->ppl_page)
2341                 __free_page(sh->ppl_page);
2342         kmem_cache_free(sc, sh);
2343 }
2344
2345 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2346         int disks, struct r5conf *conf)
2347 {
2348         struct stripe_head *sh;
2349
2350         sh = kmem_cache_zalloc(sc, gfp);
2351         if (sh) {
2352                 spin_lock_init(&sh->stripe_lock);
2353                 spin_lock_init(&sh->batch_lock);
2354                 INIT_LIST_HEAD(&sh->batch_list);
2355                 INIT_LIST_HEAD(&sh->lru);
2356                 INIT_LIST_HEAD(&sh->r5c);
2357                 INIT_LIST_HEAD(&sh->log_list);
2358                 atomic_set(&sh->count, 1);
2359                 sh->raid_conf = conf;
2360                 sh->log_start = MaxSector;
2361
2362                 if (raid5_has_ppl(conf)) {
2363                         sh->ppl_page = alloc_page(gfp);
2364                         if (!sh->ppl_page) {
2365                                 free_stripe(sc, sh);
2366                                 return NULL;
2367                         }
2368                 }
2369 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2370                 if (init_stripe_shared_pages(sh, conf, disks)) {
2371                         free_stripe(sc, sh);
2372                         return NULL;
2373                 }
2374 #endif
2375         }
2376         return sh;
2377 }
2378 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2379 {
2380         struct stripe_head *sh;
2381
2382         sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2383         if (!sh)
2384                 return 0;
2385
2386         if (grow_buffers(sh, gfp)) {
2387                 shrink_buffers(sh);
2388                 free_stripe(conf->slab_cache, sh);
2389                 return 0;
2390         }
2391         sh->hash_lock_index =
2392                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2393         /* we just created an active stripe so... */
2394         atomic_inc(&conf->active_stripes);
2395
2396         raid5_release_stripe(sh);
2397         WRITE_ONCE(conf->max_nr_stripes, conf->max_nr_stripes + 1);
2398         return 1;
2399 }
2400
2401 static int grow_stripes(struct r5conf *conf, int num)
2402 {
2403         struct kmem_cache *sc;
2404         size_t namelen = sizeof(conf->cache_name[0]);
2405         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2406
2407         if (mddev_is_dm(conf->mddev))
2408                 snprintf(conf->cache_name[0], namelen,
2409                         "raid%d-%p", conf->level, conf->mddev);
2410         else
2411                 snprintf(conf->cache_name[0], namelen,
2412                         "raid%d-%s", conf->level, mdname(conf->mddev));
2413         snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2414
2415         conf->active_name = 0;
2416         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2417                                struct_size_t(struct stripe_head, dev, devs),
2418                                0, 0, NULL);
2419         if (!sc)
2420                 return 1;
2421         conf->slab_cache = sc;
2422         conf->pool_size = devs;
2423         while (num--)
2424                 if (!grow_one_stripe(conf, GFP_KERNEL))
2425                         return 1;
2426
2427         return 0;
2428 }
2429
2430 /**
2431  * scribble_alloc - allocate percpu scribble buffer for required size
2432  *                  of the scribble region
2433  * @percpu: from for_each_present_cpu() of the caller
2434  * @num: total number of disks in the array
2435  * @cnt: scribble objs count for required size of the scribble region
2436  *
2437  * The scribble buffer size must be enough to contain:
2438  * 1/ a struct page pointer for each device in the array +2
2439  * 2/ room to convert each entry in (1) to its corresponding dma
2440  *    (dma_map_page()) or page (page_address()) address.
2441  *
2442  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2443  * calculate over all devices (not just the data blocks), using zeros in place
2444  * of the P and Q blocks.
2445  */
2446 static int scribble_alloc(struct raid5_percpu *percpu,
2447                           int num, int cnt)
2448 {
2449         size_t obj_size =
2450                 sizeof(struct page *) * (num + 2) +
2451                 sizeof(addr_conv_t) * (num + 2) +
2452                 sizeof(unsigned int) * (num + 2);
2453         void *scribble;
2454
2455         /*
2456          * If here is in raid array suspend context, it is in memalloc noio
2457          * context as well, there is no potential recursive memory reclaim
2458          * I/Os with the GFP_KERNEL flag.
2459          */
2460         scribble = kvmalloc_array(cnt, obj_size, GFP_KERNEL);
2461         if (!scribble)
2462                 return -ENOMEM;
2463
2464         kvfree(percpu->scribble);
2465
2466         percpu->scribble = scribble;
2467         percpu->scribble_obj_size = obj_size;
2468         return 0;
2469 }
2470
2471 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2472 {
2473         unsigned long cpu;
2474         int err = 0;
2475
2476         /* Never shrink. */
2477         if (conf->scribble_disks >= new_disks &&
2478             conf->scribble_sectors >= new_sectors)
2479                 return 0;
2480
2481         raid5_quiesce(conf->mddev, true);
2482         cpus_read_lock();
2483
2484         for_each_present_cpu(cpu) {
2485                 struct raid5_percpu *percpu;
2486
2487                 percpu = per_cpu_ptr(conf->percpu, cpu);
2488                 err = scribble_alloc(percpu, new_disks,
2489                                      new_sectors / RAID5_STRIPE_SECTORS(conf));
2490                 if (err)
2491                         break;
2492         }
2493
2494         cpus_read_unlock();
2495         raid5_quiesce(conf->mddev, false);
2496
2497         if (!err) {
2498                 conf->scribble_disks = new_disks;
2499                 conf->scribble_sectors = new_sectors;
2500         }
2501         return err;
2502 }
2503
2504 static int resize_stripes(struct r5conf *conf, int newsize)
2505 {
2506         /* Make all the stripes able to hold 'newsize' devices.
2507          * New slots in each stripe get 'page' set to a new page.
2508          *
2509          * This happens in stages:
2510          * 1/ create a new kmem_cache and allocate the required number of
2511          *    stripe_heads.
2512          * 2/ gather all the old stripe_heads and transfer the pages across
2513          *    to the new stripe_heads.  This will have the side effect of
2514          *    freezing the array as once all stripe_heads have been collected,
2515          *    no IO will be possible.  Old stripe heads are freed once their
2516          *    pages have been transferred over, and the old kmem_cache is
2517          *    freed when all stripes are done.
2518          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2519          *    we simple return a failure status - no need to clean anything up.
2520          * 4/ allocate new pages for the new slots in the new stripe_heads.
2521          *    If this fails, we don't bother trying the shrink the
2522          *    stripe_heads down again, we just leave them as they are.
2523          *    As each stripe_head is processed the new one is released into
2524          *    active service.
2525          *
2526          * Once step2 is started, we cannot afford to wait for a write,
2527          * so we use GFP_NOIO allocations.
2528          */
2529         struct stripe_head *osh, *nsh;
2530         LIST_HEAD(newstripes);
2531         struct disk_info *ndisks;
2532         int err = 0;
2533         struct kmem_cache *sc;
2534         int i;
2535         int hash, cnt;
2536
2537         md_allow_write(conf->mddev);
2538
2539         /* Step 1 */
2540         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2541                                struct_size_t(struct stripe_head, dev, newsize),
2542                                0, 0, NULL);
2543         if (!sc)
2544                 return -ENOMEM;
2545
2546         /* Need to ensure auto-resizing doesn't interfere */
2547         mutex_lock(&conf->cache_size_mutex);
2548
2549         for (i = conf->max_nr_stripes; i; i--) {
2550                 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2551                 if (!nsh)
2552                         break;
2553
2554                 list_add(&nsh->lru, &newstripes);
2555         }
2556         if (i) {
2557                 /* didn't get enough, give up */
2558                 while (!list_empty(&newstripes)) {
2559                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2560                         list_del(&nsh->lru);
2561                         free_stripe(sc, nsh);
2562                 }
2563                 kmem_cache_destroy(sc);
2564                 mutex_unlock(&conf->cache_size_mutex);
2565                 return -ENOMEM;
2566         }
2567         /* Step 2 - Must use GFP_NOIO now.
2568          * OK, we have enough stripes, start collecting inactive
2569          * stripes and copying them over
2570          */
2571         hash = 0;
2572         cnt = 0;
2573         list_for_each_entry(nsh, &newstripes, lru) {
2574                 lock_device_hash_lock(conf, hash);
2575                 wait_event_cmd(conf->wait_for_stripe,
2576                                     !list_empty(conf->inactive_list + hash),
2577                                     unlock_device_hash_lock(conf, hash),
2578                                     lock_device_hash_lock(conf, hash));
2579                 osh = get_free_stripe(conf, hash);
2580                 unlock_device_hash_lock(conf, hash);
2581
2582 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2583         for (i = 0; i < osh->nr_pages; i++) {
2584                 nsh->pages[i] = osh->pages[i];
2585                 osh->pages[i] = NULL;
2586         }
2587 #endif
2588                 for(i=0; i<conf->pool_size; i++) {
2589                         nsh->dev[i].page = osh->dev[i].page;
2590                         nsh->dev[i].orig_page = osh->dev[i].page;
2591                         nsh->dev[i].offset = osh->dev[i].offset;
2592                 }
2593                 nsh->hash_lock_index = hash;
2594                 free_stripe(conf->slab_cache, osh);
2595                 cnt++;
2596                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2597                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2598                         hash++;
2599                         cnt = 0;
2600                 }
2601         }
2602         kmem_cache_destroy(conf->slab_cache);
2603
2604         /* Step 3.
2605          * At this point, we are holding all the stripes so the array
2606          * is completely stalled, so now is a good time to resize
2607          * conf->disks and the scribble region
2608          */
2609         ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2610         if (ndisks) {
2611                 for (i = 0; i < conf->pool_size; i++)
2612                         ndisks[i] = conf->disks[i];
2613
2614                 for (i = conf->pool_size; i < newsize; i++) {
2615                         ndisks[i].extra_page = alloc_page(GFP_NOIO);
2616                         if (!ndisks[i].extra_page)
2617                                 err = -ENOMEM;
2618                 }
2619
2620                 if (err) {
2621                         for (i = conf->pool_size; i < newsize; i++)
2622                                 if (ndisks[i].extra_page)
2623                                         put_page(ndisks[i].extra_page);
2624                         kfree(ndisks);
2625                 } else {
2626                         kfree(conf->disks);
2627                         conf->disks = ndisks;
2628                 }
2629         } else
2630                 err = -ENOMEM;
2631
2632         conf->slab_cache = sc;
2633         conf->active_name = 1-conf->active_name;
2634
2635         /* Step 4, return new stripes to service */
2636         while(!list_empty(&newstripes)) {
2637                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2638                 list_del_init(&nsh->lru);
2639
2640 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2641                 for (i = 0; i < nsh->nr_pages; i++) {
2642                         if (nsh->pages[i])
2643                                 continue;
2644                         nsh->pages[i] = alloc_page(GFP_NOIO);
2645                         if (!nsh->pages[i])
2646                                 err = -ENOMEM;
2647                 }
2648
2649                 for (i = conf->raid_disks; i < newsize; i++) {
2650                         if (nsh->dev[i].page)
2651                                 continue;
2652                         nsh->dev[i].page = raid5_get_dev_page(nsh, i);
2653                         nsh->dev[i].orig_page = nsh->dev[i].page;
2654                         nsh->dev[i].offset = raid5_get_page_offset(nsh, i);
2655                 }
2656 #else
2657                 for (i=conf->raid_disks; i < newsize; i++)
2658                         if (nsh->dev[i].page == NULL) {
2659                                 struct page *p = alloc_page(GFP_NOIO);
2660                                 nsh->dev[i].page = p;
2661                                 nsh->dev[i].orig_page = p;
2662                                 nsh->dev[i].offset = 0;
2663                                 if (!p)
2664                                         err = -ENOMEM;
2665                         }
2666 #endif
2667                 raid5_release_stripe(nsh);
2668         }
2669         /* critical section pass, GFP_NOIO no longer needed */
2670
2671         if (!err)
2672                 conf->pool_size = newsize;
2673         mutex_unlock(&conf->cache_size_mutex);
2674
2675         return err;
2676 }
2677
2678 static int drop_one_stripe(struct r5conf *conf)
2679 {
2680         struct stripe_head *sh;
2681         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2682
2683         spin_lock_irq(conf->hash_locks + hash);
2684         sh = get_free_stripe(conf, hash);
2685         spin_unlock_irq(conf->hash_locks + hash);
2686         if (!sh)
2687                 return 0;
2688         BUG_ON(atomic_read(&sh->count));
2689         shrink_buffers(sh);
2690         free_stripe(conf->slab_cache, sh);
2691         atomic_dec(&conf->active_stripes);
2692         WRITE_ONCE(conf->max_nr_stripes, conf->max_nr_stripes - 1);
2693         return 1;
2694 }
2695
2696 static void shrink_stripes(struct r5conf *conf)
2697 {
2698         while (conf->max_nr_stripes &&
2699                drop_one_stripe(conf))
2700                 ;
2701
2702         kmem_cache_destroy(conf->slab_cache);
2703         conf->slab_cache = NULL;
2704 }
2705
2706 static void raid5_end_read_request(struct bio * bi)
2707 {
2708         struct stripe_head *sh = bi->bi_private;
2709         struct r5conf *conf = sh->raid_conf;
2710         int disks = sh->disks, i;
2711         struct md_rdev *rdev = NULL;
2712         sector_t s;
2713
2714         for (i=0 ; i<disks; i++)
2715                 if (bi == &sh->dev[i].req)
2716                         break;
2717
2718         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2719                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2720                 bi->bi_status);
2721         if (i == disks) {
2722                 BUG();
2723                 return;
2724         }
2725         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2726                 /* If replacement finished while this request was outstanding,
2727                  * 'replacement' might be NULL already.
2728                  * In that case it moved down to 'rdev'.
2729                  * rdev is not removed until all requests are finished.
2730                  */
2731                 rdev = conf->disks[i].replacement;
2732         if (!rdev)
2733                 rdev = conf->disks[i].rdev;
2734
2735         if (use_new_offset(conf, sh))
2736                 s = sh->sector + rdev->new_data_offset;
2737         else
2738                 s = sh->sector + rdev->data_offset;
2739         if (!bi->bi_status) {
2740                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2741                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2742                         /* Note that this cannot happen on a
2743                          * replacement device.  We just fail those on
2744                          * any error
2745                          */
2746                         pr_info_ratelimited(
2747                                 "md/raid:%s: read error corrected (%lu sectors at %llu on %pg)\n",
2748                                 mdname(conf->mddev), RAID5_STRIPE_SECTORS(conf),
2749                                 (unsigned long long)s,
2750                                 rdev->bdev);
2751                         atomic_add(RAID5_STRIPE_SECTORS(conf), &rdev->corrected_errors);
2752                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2753                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2754                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2755                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2756
2757                 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2758                         /*
2759                          * end read for a page in journal, this
2760                          * must be preparing for prexor in rmw
2761                          */
2762                         set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2763
2764                 if (atomic_read(&rdev->read_errors))
2765                         atomic_set(&rdev->read_errors, 0);
2766         } else {
2767                 int retry = 0;
2768                 int set_bad = 0;
2769
2770                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2771                 if (!(bi->bi_status == BLK_STS_PROTECTION))
2772                         atomic_inc(&rdev->read_errors);
2773                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2774                         pr_warn_ratelimited(
2775                                 "md/raid:%s: read error on replacement device (sector %llu on %pg).\n",
2776                                 mdname(conf->mddev),
2777                                 (unsigned long long)s,
2778                                 rdev->bdev);
2779                 else if (conf->mddev->degraded >= conf->max_degraded) {
2780                         set_bad = 1;
2781                         pr_warn_ratelimited(
2782                                 "md/raid:%s: read error not correctable (sector %llu on %pg).\n",
2783                                 mdname(conf->mddev),
2784                                 (unsigned long long)s,
2785                                 rdev->bdev);
2786                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2787                         /* Oh, no!!! */
2788                         set_bad = 1;
2789                         pr_warn_ratelimited(
2790                                 "md/raid:%s: read error NOT corrected!! (sector %llu on %pg).\n",
2791                                 mdname(conf->mddev),
2792                                 (unsigned long long)s,
2793                                 rdev->bdev);
2794                 } else if (atomic_read(&rdev->read_errors)
2795                          > conf->max_nr_stripes) {
2796                         if (!test_bit(Faulty, &rdev->flags)) {
2797                                 pr_warn("md/raid:%s: %d read_errors > %d stripes\n",
2798                                     mdname(conf->mddev),
2799                                     atomic_read(&rdev->read_errors),
2800                                     conf->max_nr_stripes);
2801                                 pr_warn("md/raid:%s: Too many read errors, failing device %pg.\n",
2802                                     mdname(conf->mddev), rdev->bdev);
2803                         }
2804                 } else
2805                         retry = 1;
2806                 if (set_bad && test_bit(In_sync, &rdev->flags)
2807                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2808                         retry = 1;
2809                 if (retry)
2810                         if (sh->qd_idx >= 0 && sh->pd_idx == i)
2811                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2812                         else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2813                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2814                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2815                         } else
2816                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2817                 else {
2818                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2819                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2820                         if (!(set_bad
2821                               && test_bit(In_sync, &rdev->flags)
2822                               && rdev_set_badblocks(
2823                                       rdev, sh->sector, RAID5_STRIPE_SECTORS(conf), 0)))
2824                                 md_error(conf->mddev, rdev);
2825                 }
2826         }
2827         rdev_dec_pending(rdev, conf->mddev);
2828         bio_uninit(bi);
2829         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2830         set_bit(STRIPE_HANDLE, &sh->state);
2831         raid5_release_stripe(sh);
2832 }
2833
2834 static void raid5_end_write_request(struct bio *bi)
2835 {
2836         struct stripe_head *sh = bi->bi_private;
2837         struct r5conf *conf = sh->raid_conf;
2838         int disks = sh->disks, i;
2839         struct md_rdev *rdev;
2840         int replacement = 0;
2841
2842         for (i = 0 ; i < disks; i++) {
2843                 if (bi == &sh->dev[i].req) {
2844                         rdev = conf->disks[i].rdev;
2845                         break;
2846                 }
2847                 if (bi == &sh->dev[i].rreq) {
2848                         rdev = conf->disks[i].replacement;
2849                         if (rdev)
2850                                 replacement = 1;
2851                         else
2852                                 /* rdev was removed and 'replacement'
2853                                  * replaced it.  rdev is not removed
2854                                  * until all requests are finished.
2855                                  */
2856                                 rdev = conf->disks[i].rdev;
2857                         break;
2858                 }
2859         }
2860         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2861                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2862                 bi->bi_status);
2863         if (i == disks) {
2864                 BUG();
2865                 return;
2866         }
2867
2868         if (replacement) {
2869                 if (bi->bi_status)
2870                         md_error(conf->mddev, rdev);
2871                 else if (rdev_has_badblock(rdev, sh->sector,
2872                                            RAID5_STRIPE_SECTORS(conf)))
2873                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2874         } else {
2875                 if (bi->bi_status) {
2876                         set_bit(WriteErrorSeen, &rdev->flags);
2877                         set_bit(R5_WriteError, &sh->dev[i].flags);
2878                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2879                                 set_bit(MD_RECOVERY_NEEDED,
2880                                         &rdev->mddev->recovery);
2881                 } else if (rdev_has_badblock(rdev, sh->sector,
2882                                              RAID5_STRIPE_SECTORS(conf))) {
2883                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2884                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2885                                 /* That was a successful write so make
2886                                  * sure it looks like we already did
2887                                  * a re-write.
2888                                  */
2889                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2890                 }
2891         }
2892         rdev_dec_pending(rdev, conf->mddev);
2893
2894         if (sh->batch_head && bi->bi_status && !replacement)
2895                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2896
2897         bio_uninit(bi);
2898         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2899                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2900         set_bit(STRIPE_HANDLE, &sh->state);
2901
2902         if (sh->batch_head && sh != sh->batch_head)
2903                 raid5_release_stripe(sh->batch_head);
2904         raid5_release_stripe(sh);
2905 }
2906
2907 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2908 {
2909         struct r5conf *conf = mddev->private;
2910         unsigned long flags;
2911         pr_debug("raid456: error called\n");
2912
2913         pr_crit("md/raid:%s: Disk failure on %pg, disabling device.\n",
2914                 mdname(mddev), rdev->bdev);
2915
2916         spin_lock_irqsave(&conf->device_lock, flags);
2917         set_bit(Faulty, &rdev->flags);
2918         clear_bit(In_sync, &rdev->flags);
2919         mddev->degraded = raid5_calc_degraded(conf);
2920
2921         if (has_failed(conf)) {
2922                 set_bit(MD_BROKEN, &conf->mddev->flags);
2923                 conf->recovery_disabled = mddev->recovery_disabled;
2924
2925                 pr_crit("md/raid:%s: Cannot continue operation (%d/%d failed).\n",
2926                         mdname(mddev), mddev->degraded, conf->raid_disks);
2927         } else {
2928                 pr_crit("md/raid:%s: Operation continuing on %d devices.\n",
2929                         mdname(mddev), conf->raid_disks - mddev->degraded);
2930         }
2931
2932         spin_unlock_irqrestore(&conf->device_lock, flags);
2933         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2934
2935         set_bit(Blocked, &rdev->flags);
2936         set_mask_bits(&mddev->sb_flags, 0,
2937                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2938         r5c_update_on_rdev_error(mddev, rdev);
2939 }
2940
2941 /*
2942  * Input: a 'big' sector number,
2943  * Output: index of the data and parity disk, and the sector # in them.
2944  */
2945 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2946                               int previous, int *dd_idx,
2947                               struct stripe_head *sh)
2948 {
2949         sector_t stripe, stripe2;
2950         sector_t chunk_number;
2951         unsigned int chunk_offset;
2952         int pd_idx, qd_idx;
2953         int ddf_layout = 0;
2954         sector_t new_sector;
2955         int algorithm = previous ? conf->prev_algo
2956                                  : conf->algorithm;
2957         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2958                                          : conf->chunk_sectors;
2959         int raid_disks = previous ? conf->previous_raid_disks
2960                                   : conf->raid_disks;
2961         int data_disks = raid_disks - conf->max_degraded;
2962
2963         /* First compute the information on this sector */
2964
2965         /*
2966          * Compute the chunk number and the sector offset inside the chunk
2967          */
2968         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2969         chunk_number = r_sector;
2970
2971         /*
2972          * Compute the stripe number
2973          */
2974         stripe = chunk_number;
2975         *dd_idx = sector_div(stripe, data_disks);
2976         stripe2 = stripe;
2977         /*
2978          * Select the parity disk based on the user selected algorithm.
2979          */
2980         pd_idx = qd_idx = -1;
2981         switch(conf->level) {
2982         case 4:
2983                 pd_idx = data_disks;
2984                 break;
2985         case 5:
2986                 switch (algorithm) {
2987                 case ALGORITHM_LEFT_ASYMMETRIC:
2988                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2989                         if (*dd_idx >= pd_idx)
2990                                 (*dd_idx)++;
2991                         break;
2992                 case ALGORITHM_RIGHT_ASYMMETRIC:
2993                         pd_idx = sector_div(stripe2, raid_disks);
2994                         if (*dd_idx >= pd_idx)
2995                                 (*dd_idx)++;
2996                         break;
2997                 case ALGORITHM_LEFT_SYMMETRIC:
2998                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2999                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3000                         break;
3001                 case ALGORITHM_RIGHT_SYMMETRIC:
3002                         pd_idx = sector_div(stripe2, raid_disks);
3003                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3004                         break;
3005                 case ALGORITHM_PARITY_0:
3006                         pd_idx = 0;
3007                         (*dd_idx)++;
3008                         break;
3009                 case ALGORITHM_PARITY_N:
3010                         pd_idx = data_disks;
3011                         break;
3012                 default:
3013                         BUG();
3014                 }
3015                 break;
3016         case 6:
3017
3018                 switch (algorithm) {
3019                 case ALGORITHM_LEFT_ASYMMETRIC:
3020                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3021                         qd_idx = pd_idx + 1;
3022                         if (pd_idx == raid_disks-1) {
3023                                 (*dd_idx)++;    /* Q D D D P */
3024                                 qd_idx = 0;
3025                         } else if (*dd_idx >= pd_idx)
3026                                 (*dd_idx) += 2; /* D D P Q D */
3027                         break;
3028                 case ALGORITHM_RIGHT_ASYMMETRIC:
3029                         pd_idx = sector_div(stripe2, raid_disks);
3030                         qd_idx = pd_idx + 1;
3031                         if (pd_idx == raid_disks-1) {
3032                                 (*dd_idx)++;    /* Q D D D P */
3033                                 qd_idx = 0;
3034                         } else if (*dd_idx >= pd_idx)
3035                                 (*dd_idx) += 2; /* D D P Q D */
3036                         break;
3037                 case ALGORITHM_LEFT_SYMMETRIC:
3038                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3039                         qd_idx = (pd_idx + 1) % raid_disks;
3040                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3041                         break;
3042                 case ALGORITHM_RIGHT_SYMMETRIC:
3043                         pd_idx = sector_div(stripe2, raid_disks);
3044                         qd_idx = (pd_idx + 1) % raid_disks;
3045                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3046                         break;
3047
3048                 case ALGORITHM_PARITY_0:
3049                         pd_idx = 0;
3050                         qd_idx = 1;
3051                         (*dd_idx) += 2;
3052                         break;
3053                 case ALGORITHM_PARITY_N:
3054                         pd_idx = data_disks;
3055                         qd_idx = data_disks + 1;
3056                         break;
3057
3058                 case ALGORITHM_ROTATING_ZERO_RESTART:
3059                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
3060                          * of blocks for computing Q is different.
3061                          */
3062                         pd_idx = sector_div(stripe2, raid_disks);
3063                         qd_idx = pd_idx + 1;
3064                         if (pd_idx == raid_disks-1) {
3065                                 (*dd_idx)++;    /* Q D D D P */
3066                                 qd_idx = 0;
3067                         } else if (*dd_idx >= pd_idx)
3068                                 (*dd_idx) += 2; /* D D P Q D */
3069                         ddf_layout = 1;
3070                         break;
3071
3072                 case ALGORITHM_ROTATING_N_RESTART:
3073                         /* Same a left_asymmetric, by first stripe is
3074                          * D D D P Q  rather than
3075                          * Q D D D P
3076                          */
3077                         stripe2 += 1;
3078                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3079                         qd_idx = pd_idx + 1;
3080                         if (pd_idx == raid_disks-1) {
3081                                 (*dd_idx)++;    /* Q D D D P */
3082                                 qd_idx = 0;
3083                         } else if (*dd_idx >= pd_idx)
3084                                 (*dd_idx) += 2; /* D D P Q D */
3085                         ddf_layout = 1;
3086                         break;
3087
3088                 case ALGORITHM_ROTATING_N_CONTINUE:
3089                         /* Same as left_symmetric but Q is before P */
3090                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3091                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
3092                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3093                         ddf_layout = 1;
3094                         break;
3095
3096                 case ALGORITHM_LEFT_ASYMMETRIC_6:
3097                         /* RAID5 left_asymmetric, with Q on last device */
3098                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3099                         if (*dd_idx >= pd_idx)
3100                                 (*dd_idx)++;
3101                         qd_idx = raid_disks - 1;
3102                         break;
3103
3104                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3105                         pd_idx = sector_div(stripe2, raid_disks-1);
3106                         if (*dd_idx >= pd_idx)
3107                                 (*dd_idx)++;
3108                         qd_idx = raid_disks - 1;
3109                         break;
3110
3111                 case ALGORITHM_LEFT_SYMMETRIC_6:
3112                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3113                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3114                         qd_idx = raid_disks - 1;
3115                         break;
3116
3117                 case ALGORITHM_RIGHT_SYMMETRIC_6:
3118                         pd_idx = sector_div(stripe2, raid_disks-1);
3119                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3120                         qd_idx = raid_disks - 1;
3121                         break;
3122
3123                 case ALGORITHM_PARITY_0_6:
3124                         pd_idx = 0;
3125                         (*dd_idx)++;
3126                         qd_idx = raid_disks - 1;
3127                         break;
3128
3129                 default:
3130                         BUG();
3131                 }
3132                 break;
3133         }
3134
3135         if (sh) {
3136                 sh->pd_idx = pd_idx;
3137                 sh->qd_idx = qd_idx;
3138                 sh->ddf_layout = ddf_layout;
3139         }
3140         /*
3141          * Finally, compute the new sector number
3142          */
3143         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
3144         return new_sector;
3145 }
3146
3147 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
3148 {
3149         struct r5conf *conf = sh->raid_conf;
3150         int raid_disks = sh->disks;
3151         int data_disks = raid_disks - conf->max_degraded;
3152         sector_t new_sector = sh->sector, check;
3153         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
3154                                          : conf->chunk_sectors;
3155         int algorithm = previous ? conf->prev_algo
3156                                  : conf->algorithm;
3157         sector_t stripe;
3158         int chunk_offset;
3159         sector_t chunk_number;
3160         int dummy1, dd_idx = i;
3161         sector_t r_sector;
3162         struct stripe_head sh2;
3163
3164         chunk_offset = sector_div(new_sector, sectors_per_chunk);
3165         stripe = new_sector;
3166
3167         if (i == sh->pd_idx)
3168                 return 0;
3169         switch(conf->level) {
3170         case 4: break;
3171         case 5:
3172                 switch (algorithm) {
3173                 case ALGORITHM_LEFT_ASYMMETRIC:
3174                 case ALGORITHM_RIGHT_ASYMMETRIC:
3175                         if (i > sh->pd_idx)
3176                                 i--;
3177                         break;
3178                 case ALGORITHM_LEFT_SYMMETRIC:
3179                 case ALGORITHM_RIGHT_SYMMETRIC:
3180                         if (i < sh->pd_idx)
3181                                 i += raid_disks;
3182                         i -= (sh->pd_idx + 1);
3183                         break;
3184                 case ALGORITHM_PARITY_0:
3185                         i -= 1;
3186                         break;
3187                 case ALGORITHM_PARITY_N:
3188                         break;
3189                 default:
3190                         BUG();
3191                 }
3192                 break;
3193         case 6:
3194                 if (i == sh->qd_idx)
3195                         return 0; /* It is the Q disk */
3196                 switch (algorithm) {
3197                 case ALGORITHM_LEFT_ASYMMETRIC:
3198                 case ALGORITHM_RIGHT_ASYMMETRIC:
3199                 case ALGORITHM_ROTATING_ZERO_RESTART:
3200                 case ALGORITHM_ROTATING_N_RESTART:
3201                         if (sh->pd_idx == raid_disks-1)
3202                                 i--;    /* Q D D D P */
3203                         else if (i > sh->pd_idx)
3204                                 i -= 2; /* D D P Q D */
3205                         break;
3206                 case ALGORITHM_LEFT_SYMMETRIC:
3207                 case ALGORITHM_RIGHT_SYMMETRIC:
3208                         if (sh->pd_idx == raid_disks-1)
3209                                 i--; /* Q D D D P */
3210                         else {
3211                                 /* D D P Q D */
3212                                 if (i < sh->pd_idx)
3213                                         i += raid_disks;
3214                                 i -= (sh->pd_idx + 2);
3215                         }
3216                         break;
3217                 case ALGORITHM_PARITY_0:
3218                         i -= 2;
3219                         break;
3220                 case ALGORITHM_PARITY_N:
3221                         break;
3222                 case ALGORITHM_ROTATING_N_CONTINUE:
3223                         /* Like left_symmetric, but P is before Q */
3224                         if (sh->pd_idx == 0)
3225                                 i--;    /* P D D D Q */
3226                         else {
3227                                 /* D D Q P D */
3228                                 if (i < sh->pd_idx)
3229                                         i += raid_disks;
3230                                 i -= (sh->pd_idx + 1);
3231                         }
3232                         break;
3233                 case ALGORITHM_LEFT_ASYMMETRIC_6:
3234                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3235                         if (i > sh->pd_idx)
3236                                 i--;
3237                         break;
3238                 case ALGORITHM_LEFT_SYMMETRIC_6:
3239                 case ALGORITHM_RIGHT_SYMMETRIC_6:
3240                         if (i < sh->pd_idx)
3241                                 i += data_disks + 1;
3242                         i -= (sh->pd_idx + 1);
3243                         break;
3244                 case ALGORITHM_PARITY_0_6:
3245                         i -= 1;
3246                         break;
3247                 default:
3248                         BUG();
3249                 }
3250                 break;
3251         }
3252
3253         chunk_number = stripe * data_disks + i;
3254         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3255
3256         check = raid5_compute_sector(conf, r_sector,
3257                                      previous, &dummy1, &sh2);
3258         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3259                 || sh2.qd_idx != sh->qd_idx) {
3260                 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3261                         mdname(conf->mddev));
3262                 return 0;
3263         }
3264         return r_sector;
3265 }
3266
3267 /*
3268  * There are cases where we want handle_stripe_dirtying() and
3269  * schedule_reconstruction() to delay towrite to some dev of a stripe.
3270  *
3271  * This function checks whether we want to delay the towrite. Specifically,
3272  * we delay the towrite when:
3273  *
3274  *   1. degraded stripe has a non-overwrite to the missing dev, AND this
3275  *      stripe has data in journal (for other devices).
3276  *
3277  *      In this case, when reading data for the non-overwrite dev, it is
3278  *      necessary to handle complex rmw of write back cache (prexor with
3279  *      orig_page, and xor with page). To keep read path simple, we would
3280  *      like to flush data in journal to RAID disks first, so complex rmw
3281  *      is handled in the write patch (handle_stripe_dirtying).
3282  *
3283  *   2. when journal space is critical (R5C_LOG_CRITICAL=1)
3284  *
3285  *      It is important to be able to flush all stripes in raid5-cache.
3286  *      Therefore, we need reserve some space on the journal device for
3287  *      these flushes. If flush operation includes pending writes to the
3288  *      stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3289  *      for the flush out. If we exclude these pending writes from flush
3290  *      operation, we only need (conf->max_degraded + 1) pages per stripe.
3291  *      Therefore, excluding pending writes in these cases enables more
3292  *      efficient use of the journal device.
3293  *
3294  *      Note: To make sure the stripe makes progress, we only delay
3295  *      towrite for stripes with data already in journal (injournal > 0).
3296  *      When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3297  *      no_space_stripes list.
3298  *
3299  *   3. during journal failure
3300  *      In journal failure, we try to flush all cached data to raid disks
3301  *      based on data in stripe cache. The array is read-only to upper
3302  *      layers, so we would skip all pending writes.
3303  *
3304  */
3305 static inline bool delay_towrite(struct r5conf *conf,
3306                                  struct r5dev *dev,
3307                                  struct stripe_head_state *s)
3308 {
3309         /* case 1 above */
3310         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3311             !test_bit(R5_Insync, &dev->flags) && s->injournal)
3312                 return true;
3313         /* case 2 above */
3314         if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3315             s->injournal > 0)
3316                 return true;
3317         /* case 3 above */
3318         if (s->log_failed && s->injournal)
3319                 return true;
3320         return false;
3321 }
3322
3323 static void
3324 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3325                          int rcw, int expand)
3326 {
3327         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3328         struct r5conf *conf = sh->raid_conf;
3329         int level = conf->level;
3330
3331         if (rcw) {
3332                 /*
3333                  * In some cases, handle_stripe_dirtying initially decided to
3334                  * run rmw and allocates extra page for prexor. However, rcw is
3335                  * cheaper later on. We need to free the extra page now,
3336                  * because we won't be able to do that in ops_complete_prexor().
3337                  */
3338                 r5c_release_extra_page(sh);
3339
3340                 for (i = disks; i--; ) {
3341                         struct r5dev *dev = &sh->dev[i];
3342
3343                         if (dev->towrite && !delay_towrite(conf, dev, s)) {
3344                                 set_bit(R5_LOCKED, &dev->flags);
3345                                 set_bit(R5_Wantdrain, &dev->flags);
3346                                 if (!expand)
3347                                         clear_bit(R5_UPTODATE, &dev->flags);
3348                                 s->locked++;
3349                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3350                                 set_bit(R5_LOCKED, &dev->flags);
3351                                 s->locked++;
3352                         }
3353                 }
3354                 /* if we are not expanding this is a proper write request, and
3355                  * there will be bios with new data to be drained into the
3356                  * stripe cache
3357                  */
3358                 if (!expand) {
3359                         if (!s->locked)
3360                                 /* False alarm, nothing to do */
3361                                 return;
3362                         sh->reconstruct_state = reconstruct_state_drain_run;
3363                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3364                 } else
3365                         sh->reconstruct_state = reconstruct_state_run;
3366
3367                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3368
3369                 if (s->locked + conf->max_degraded == disks)
3370                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3371                                 atomic_inc(&conf->pending_full_writes);
3372         } else {
3373                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3374                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3375                 BUG_ON(level == 6 &&
3376                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3377                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3378
3379                 for (i = disks; i--; ) {
3380                         struct r5dev *dev = &sh->dev[i];
3381                         if (i == pd_idx || i == qd_idx)
3382                                 continue;
3383
3384                         if (dev->towrite &&
3385                             (test_bit(R5_UPTODATE, &dev->flags) ||
3386                              test_bit(R5_Wantcompute, &dev->flags))) {
3387                                 set_bit(R5_Wantdrain, &dev->flags);
3388                                 set_bit(R5_LOCKED, &dev->flags);
3389                                 clear_bit(R5_UPTODATE, &dev->flags);
3390                                 s->locked++;
3391                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3392                                 set_bit(R5_LOCKED, &dev->flags);
3393                                 s->locked++;
3394                         }
3395                 }
3396                 if (!s->locked)
3397                         /* False alarm - nothing to do */
3398                         return;
3399                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3400                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3401                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3402                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3403         }
3404
3405         /* keep the parity disk(s) locked while asynchronous operations
3406          * are in flight
3407          */
3408         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3409         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3410         s->locked++;
3411
3412         if (level == 6) {
3413                 int qd_idx = sh->qd_idx;
3414                 struct r5dev *dev = &sh->dev[qd_idx];
3415
3416                 set_bit(R5_LOCKED, &dev->flags);
3417                 clear_bit(R5_UPTODATE, &dev->flags);
3418                 s->locked++;
3419         }
3420
3421         if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3422             test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3423             !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3424             test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3425                 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3426
3427         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3428                 __func__, (unsigned long long)sh->sector,
3429                 s->locked, s->ops_request);
3430 }
3431
3432 static bool stripe_bio_overlaps(struct stripe_head *sh, struct bio *bi,
3433                                 int dd_idx, int forwrite)
3434 {
3435         struct r5conf *conf = sh->raid_conf;
3436         struct bio **bip;
3437
3438         pr_debug("checking bi b#%llu to stripe s#%llu\n",
3439                  bi->bi_iter.bi_sector, sh->sector);
3440
3441         /* Don't allow new IO added to stripes in batch list */
3442         if (sh->batch_head)
3443                 return true;
3444
3445         if (forwrite)
3446                 bip = &sh->dev[dd_idx].towrite;
3447         else
3448                 bip = &sh->dev[dd_idx].toread;
3449
3450         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3451                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3452                         return true;
3453                 bip = &(*bip)->bi_next;
3454         }
3455
3456         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3457                 return true;
3458
3459         if (forwrite && raid5_has_ppl(conf)) {
3460                 /*
3461                  * With PPL only writes to consecutive data chunks within a
3462                  * stripe are allowed because for a single stripe_head we can
3463                  * only have one PPL entry at a time, which describes one data
3464                  * range. Not really an overlap, but R5_Overlap can be
3465                  * used to handle this.
3466                  */
3467                 sector_t sector;
3468                 sector_t first = 0;
3469                 sector_t last = 0;
3470                 int count = 0;
3471                 int i;
3472
3473                 for (i = 0; i < sh->disks; i++) {
3474                         if (i != sh->pd_idx &&
3475                             (i == dd_idx || sh->dev[i].towrite)) {
3476                                 sector = sh->dev[i].sector;
3477                                 if (count == 0 || sector < first)
3478                                         first = sector;
3479                                 if (sector > last)
3480                                         last = sector;
3481                                 count++;
3482                         }
3483                 }
3484
3485                 if (first + conf->chunk_sectors * (count - 1) != last)
3486                         return true;
3487         }
3488
3489         return false;
3490 }
3491
3492 static void __add_stripe_bio(struct stripe_head *sh, struct bio *bi,
3493                              int dd_idx, int forwrite, int previous)
3494 {
3495         struct r5conf *conf = sh->raid_conf;
3496         struct bio **bip;
3497         int firstwrite = 0;
3498
3499         if (forwrite) {
3500                 bip = &sh->dev[dd_idx].towrite;
3501                 if (!*bip)
3502                         firstwrite = 1;
3503         } else {
3504                 bip = &sh->dev[dd_idx].toread;
3505         }
3506
3507         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector)
3508                 bip = &(*bip)->bi_next;
3509
3510         if (!forwrite || previous)
3511                 clear_bit(STRIPE_BATCH_READY, &sh->state);
3512
3513         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3514         if (*bip)
3515                 bi->bi_next = *bip;
3516         *bip = bi;
3517         bio_inc_remaining(bi);
3518         md_write_inc(conf->mddev, bi);
3519
3520         if (forwrite) {
3521                 /* check if page is covered */
3522                 sector_t sector = sh->dev[dd_idx].sector;
3523                 for (bi=sh->dev[dd_idx].towrite;
3524                      sector < sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf) &&
3525                              bi && bi->bi_iter.bi_sector <= sector;
3526                      bi = r5_next_bio(conf, bi, sh->dev[dd_idx].sector)) {
3527                         if (bio_end_sector(bi) >= sector)
3528                                 sector = bio_end_sector(bi);
3529                 }
3530                 if (sector >= sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf))
3531                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3532                                 sh->overwrite_disks++;
3533         }
3534
3535         pr_debug("added bi b#%llu to stripe s#%llu, disk %d, logical %llu\n",
3536                  (*bip)->bi_iter.bi_sector, sh->sector, dd_idx,
3537                  sh->dev[dd_idx].sector);
3538
3539         if (conf->mddev->bitmap && firstwrite && !sh->batch_head) {
3540                 sh->bm_seq = conf->seq_flush+1;
3541                 set_bit(STRIPE_BIT_DELAY, &sh->state);
3542         }
3543 }
3544
3545 /*
3546  * Each stripe/dev can have one or more bios attached.
3547  * toread/towrite point to the first in a chain.
3548  * The bi_next chain must be in order.
3549  */
3550 static bool add_stripe_bio(struct stripe_head *sh, struct bio *bi,
3551                            int dd_idx, int forwrite, int previous)
3552 {
3553         spin_lock_irq(&sh->stripe_lock);
3554
3555         if (stripe_bio_overlaps(sh, bi, dd_idx, forwrite)) {
3556                 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3557                 spin_unlock_irq(&sh->stripe_lock);
3558                 return false;
3559         }
3560
3561         __add_stripe_bio(sh, bi, dd_idx, forwrite, previous);
3562         spin_unlock_irq(&sh->stripe_lock);
3563         return true;
3564 }
3565
3566 static void end_reshape(struct r5conf *conf);
3567
3568 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3569                             struct stripe_head *sh)
3570 {
3571         int sectors_per_chunk =
3572                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3573         int dd_idx;
3574         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3575         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3576
3577         raid5_compute_sector(conf,
3578                              stripe * (disks - conf->max_degraded)
3579                              *sectors_per_chunk + chunk_offset,
3580                              previous,
3581                              &dd_idx, sh);
3582 }
3583
3584 static void
3585 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3586                      struct stripe_head_state *s, int disks)
3587 {
3588         int i;
3589         BUG_ON(sh->batch_head);
3590         for (i = disks; i--; ) {
3591                 struct bio *bi;
3592
3593                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3594                         struct md_rdev *rdev = conf->disks[i].rdev;
3595
3596                         if (rdev && test_bit(In_sync, &rdev->flags) &&
3597                             !test_bit(Faulty, &rdev->flags))
3598                                 atomic_inc(&rdev->nr_pending);
3599                         else
3600                                 rdev = NULL;
3601                         if (rdev) {
3602                                 if (!rdev_set_badblocks(
3603                                             rdev,
3604                                             sh->sector,
3605                                             RAID5_STRIPE_SECTORS(conf), 0))
3606                                         md_error(conf->mddev, rdev);
3607                                 rdev_dec_pending(rdev, conf->mddev);
3608                         }
3609                 }
3610                 spin_lock_irq(&sh->stripe_lock);
3611                 /* fail all writes first */
3612                 bi = sh->dev[i].towrite;
3613                 sh->dev[i].towrite = NULL;
3614                 sh->overwrite_disks = 0;
3615                 spin_unlock_irq(&sh->stripe_lock);
3616
3617                 log_stripe_write_finished(sh);
3618
3619                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3620                         wake_up_bit(&sh->dev[i].flags, R5_Overlap);
3621
3622                 while (bi && bi->bi_iter.bi_sector <
3623                         sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3624                         struct bio *nextbi = r5_next_bio(conf, bi, sh->dev[i].sector);
3625
3626                         md_write_end(conf->mddev);
3627                         bio_io_error(bi);
3628                         bi = nextbi;
3629                 }
3630                 /* and fail all 'written' */
3631                 bi = sh->dev[i].written;
3632                 sh->dev[i].written = NULL;
3633                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3634                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3635                         sh->dev[i].page = sh->dev[i].orig_page;
3636                 }
3637
3638                 while (bi && bi->bi_iter.bi_sector <
3639                        sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3640                         struct bio *bi2 = r5_next_bio(conf, bi, sh->dev[i].sector);
3641
3642                         md_write_end(conf->mddev);
3643                         bio_io_error(bi);
3644                         bi = bi2;
3645                 }
3646
3647                 /* fail any reads if this device is non-operational and
3648                  * the data has not reached the cache yet.
3649                  */
3650                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3651                     s->failed > conf->max_degraded &&
3652                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3653                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3654                         spin_lock_irq(&sh->stripe_lock);
3655                         bi = sh->dev[i].toread;
3656                         sh->dev[i].toread = NULL;
3657                         spin_unlock_irq(&sh->stripe_lock);
3658                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3659                                 wake_up_bit(&sh->dev[i].flags, R5_Overlap);
3660                         if (bi)
3661                                 s->to_read--;
3662                         while (bi && bi->bi_iter.bi_sector <
3663                                sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3664                                 struct bio *nextbi =
3665                                         r5_next_bio(conf, bi, sh->dev[i].sector);
3666
3667                                 bio_io_error(bi);
3668                                 bi = nextbi;
3669                         }
3670                 }
3671                 /* If we were in the middle of a write the parity block might
3672                  * still be locked - so just clear all R5_LOCKED flags
3673                  */
3674                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3675         }
3676         s->to_write = 0;
3677         s->written = 0;
3678
3679         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3680                 if (atomic_dec_and_test(&conf->pending_full_writes))
3681                         md_wakeup_thread(conf->mddev->thread);
3682 }
3683
3684 static void
3685 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3686                    struct stripe_head_state *s)
3687 {
3688         int abort = 0;
3689         int i;
3690
3691         BUG_ON(sh->batch_head);
3692         clear_bit(STRIPE_SYNCING, &sh->state);
3693         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3694                 wake_up_bit(&sh->dev[sh->pd_idx].flags, R5_Overlap);
3695         s->syncing = 0;
3696         s->replacing = 0;
3697         /* There is nothing more to do for sync/check/repair.
3698          * Don't even need to abort as that is handled elsewhere
3699          * if needed, and not always wanted e.g. if there is a known
3700          * bad block here.
3701          * For recover/replace we need to record a bad block on all
3702          * non-sync devices, or abort the recovery
3703          */
3704         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3705                 /* During recovery devices cannot be removed, so
3706                  * locking and refcounting of rdevs is not needed
3707                  */
3708                 for (i = 0; i < conf->raid_disks; i++) {
3709                         struct md_rdev *rdev = conf->disks[i].rdev;
3710
3711                         if (rdev
3712                             && !test_bit(Faulty, &rdev->flags)
3713                             && !test_bit(In_sync, &rdev->flags)
3714                             && !rdev_set_badblocks(rdev, sh->sector,
3715                                                    RAID5_STRIPE_SECTORS(conf), 0))
3716                                 abort = 1;
3717                         rdev = conf->disks[i].replacement;
3718
3719                         if (rdev
3720                             && !test_bit(Faulty, &rdev->flags)
3721                             && !test_bit(In_sync, &rdev->flags)
3722                             && !rdev_set_badblocks(rdev, sh->sector,
3723                                                    RAID5_STRIPE_SECTORS(conf), 0))
3724                                 abort = 1;
3725                 }
3726                 if (abort)
3727                         conf->recovery_disabled =
3728                                 conf->mddev->recovery_disabled;
3729         }
3730         md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), !abort);
3731 }
3732
3733 static int want_replace(struct stripe_head *sh, int disk_idx)
3734 {
3735         struct md_rdev *rdev;
3736         int rv = 0;
3737
3738         rdev = sh->raid_conf->disks[disk_idx].replacement;
3739         if (rdev
3740             && !test_bit(Faulty, &rdev->flags)
3741             && !test_bit(In_sync, &rdev->flags)
3742             && (rdev->recovery_offset <= sh->sector
3743                 || rdev->mddev->recovery_cp <= sh->sector))
3744                 rv = 1;
3745         return rv;
3746 }
3747
3748 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3749                            int disk_idx, int disks)
3750 {
3751         struct r5dev *dev = &sh->dev[disk_idx];
3752         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3753                                   &sh->dev[s->failed_num[1]] };
3754         int i;
3755         bool force_rcw = (sh->raid_conf->rmw_level == PARITY_DISABLE_RMW);
3756
3757
3758         if (test_bit(R5_LOCKED, &dev->flags) ||
3759             test_bit(R5_UPTODATE, &dev->flags))
3760                 /* No point reading this as we already have it or have
3761                  * decided to get it.
3762                  */
3763                 return 0;
3764
3765         if (dev->toread ||
3766             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3767                 /* We need this block to directly satisfy a request */
3768                 return 1;
3769
3770         if (s->syncing || s->expanding ||
3771             (s->replacing && want_replace(sh, disk_idx)))
3772                 /* When syncing, or expanding we read everything.
3773                  * When replacing, we need the replaced block.
3774                  */
3775                 return 1;
3776
3777         if ((s->failed >= 1 && fdev[0]->toread) ||
3778             (s->failed >= 2 && fdev[1]->toread))
3779                 /* If we want to read from a failed device, then
3780                  * we need to actually read every other device.
3781                  */
3782                 return 1;
3783
3784         /* Sometimes neither read-modify-write nor reconstruct-write
3785          * cycles can work.  In those cases we read every block we
3786          * can.  Then the parity-update is certain to have enough to
3787          * work with.
3788          * This can only be a problem when we need to write something,
3789          * and some device has failed.  If either of those tests
3790          * fail we need look no further.
3791          */
3792         if (!s->failed || !s->to_write)
3793                 return 0;
3794
3795         if (test_bit(R5_Insync, &dev->flags) &&
3796             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3797                 /* Pre-reads at not permitted until after short delay
3798                  * to gather multiple requests.  However if this
3799                  * device is no Insync, the block could only be computed
3800                  * and there is no need to delay that.
3801                  */
3802                 return 0;
3803
3804         for (i = 0; i < s->failed && i < 2; i++) {
3805                 if (fdev[i]->towrite &&
3806                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3807                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3808                         /* If we have a partial write to a failed
3809                          * device, then we will need to reconstruct
3810                          * the content of that device, so all other
3811                          * devices must be read.
3812                          */
3813                         return 1;
3814
3815                 if (s->failed >= 2 &&
3816                     (fdev[i]->towrite ||
3817                      s->failed_num[i] == sh->pd_idx ||
3818                      s->failed_num[i] == sh->qd_idx) &&
3819                     !test_bit(R5_UPTODATE, &fdev[i]->flags))
3820                         /* In max degraded raid6, If the failed disk is P, Q,
3821                          * or we want to read the failed disk, we need to do
3822                          * reconstruct-write.
3823                          */
3824                         force_rcw = true;
3825         }
3826
3827         /* If we are forced to do a reconstruct-write, because parity
3828          * cannot be trusted and we are currently recovering it, there
3829          * is extra need to be careful.
3830          * If one of the devices that we would need to read, because
3831          * it is not being overwritten (and maybe not written at all)
3832          * is missing/faulty, then we need to read everything we can.
3833          */
3834         if (!force_rcw &&
3835             sh->sector < sh->raid_conf->mddev->recovery_cp)
3836                 /* reconstruct-write isn't being forced */
3837                 return 0;
3838         for (i = 0; i < s->failed && i < 2; i++) {
3839                 if (s->failed_num[i] != sh->pd_idx &&
3840                     s->failed_num[i] != sh->qd_idx &&
3841                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3842                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3843                         return 1;
3844         }
3845
3846         return 0;
3847 }
3848
3849 /* fetch_block - checks the given member device to see if its data needs
3850  * to be read or computed to satisfy a request.
3851  *
3852  * Returns 1 when no more member devices need to be checked, otherwise returns
3853  * 0 to tell the loop in handle_stripe_fill to continue
3854  */
3855 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3856                        int disk_idx, int disks)
3857 {
3858         struct r5dev *dev = &sh->dev[disk_idx];
3859
3860         /* is the data in this block needed, and can we get it? */
3861         if (need_this_block(sh, s, disk_idx, disks)) {
3862                 /* we would like to get this block, possibly by computing it,
3863                  * otherwise read it if the backing disk is insync
3864                  */
3865                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3866                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3867                 BUG_ON(sh->batch_head);
3868
3869                 /*
3870                  * In the raid6 case if the only non-uptodate disk is P
3871                  * then we already trusted P to compute the other failed
3872                  * drives. It is safe to compute rather than re-read P.
3873                  * In other cases we only compute blocks from failed
3874                  * devices, otherwise check/repair might fail to detect
3875                  * a real inconsistency.
3876                  */
3877
3878                 if ((s->uptodate == disks - 1) &&
3879                     ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3880                     (s->failed && (disk_idx == s->failed_num[0] ||
3881                                    disk_idx == s->failed_num[1])))) {
3882                         /* have disk failed, and we're requested to fetch it;
3883                          * do compute it
3884                          */
3885                         pr_debug("Computing stripe %llu block %d\n",
3886                                (unsigned long long)sh->sector, disk_idx);
3887                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3888                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3889                         set_bit(R5_Wantcompute, &dev->flags);
3890                         sh->ops.target = disk_idx;
3891                         sh->ops.target2 = -1; /* no 2nd target */
3892                         s->req_compute = 1;
3893                         /* Careful: from this point on 'uptodate' is in the eye
3894                          * of raid_run_ops which services 'compute' operations
3895                          * before writes. R5_Wantcompute flags a block that will
3896                          * be R5_UPTODATE by the time it is needed for a
3897                          * subsequent operation.
3898                          */
3899                         s->uptodate++;
3900                         return 1;
3901                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3902                         /* Computing 2-failure is *very* expensive; only
3903                          * do it if failed >= 2
3904                          */
3905                         int other;
3906                         for (other = disks; other--; ) {
3907                                 if (other == disk_idx)
3908                                         continue;
3909                                 if (!test_bit(R5_UPTODATE,
3910                                       &sh->dev[other].flags))
3911                                         break;
3912                         }
3913                         BUG_ON(other < 0);
3914                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3915                                (unsigned long long)sh->sector,
3916                                disk_idx, other);
3917                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3918                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3919                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3920                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3921                         sh->ops.target = disk_idx;
3922                         sh->ops.target2 = other;
3923                         s->uptodate += 2;
3924                         s->req_compute = 1;
3925                         return 1;
3926                 } else if (test_bit(R5_Insync, &dev->flags)) {
3927                         set_bit(R5_LOCKED, &dev->flags);
3928                         set_bit(R5_Wantread, &dev->flags);
3929                         s->locked++;
3930                         pr_debug("Reading block %d (sync=%d)\n",
3931                                 disk_idx, s->syncing);
3932                 }
3933         }
3934
3935         return 0;
3936 }
3937
3938 /*
3939  * handle_stripe_fill - read or compute data to satisfy pending requests.
3940  */
3941 static void handle_stripe_fill(struct stripe_head *sh,
3942                                struct stripe_head_state *s,
3943                                int disks)
3944 {
3945         int i;
3946
3947         /* look for blocks to read/compute, skip this if a compute
3948          * is already in flight, or if the stripe contents are in the
3949          * midst of changing due to a write
3950          */
3951         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3952             !sh->reconstruct_state) {
3953
3954                 /*
3955                  * For degraded stripe with data in journal, do not handle
3956                  * read requests yet, instead, flush the stripe to raid
3957                  * disks first, this avoids handling complex rmw of write
3958                  * back cache (prexor with orig_page, and then xor with
3959                  * page) in the read path
3960                  */
3961                 if (s->to_read && s->injournal && s->failed) {
3962                         if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3963                                 r5c_make_stripe_write_out(sh);
3964                         goto out;
3965                 }
3966
3967                 for (i = disks; i--; )
3968                         if (fetch_block(sh, s, i, disks))
3969                                 break;
3970         }
3971 out:
3972         set_bit(STRIPE_HANDLE, &sh->state);
3973 }
3974
3975 static void break_stripe_batch_list(struct stripe_head *head_sh,
3976                                     unsigned long handle_flags);
3977 /* handle_stripe_clean_event
3978  * any written block on an uptodate or failed drive can be returned.
3979  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3980  * never LOCKED, so we don't need to test 'failed' directly.
3981  */
3982 static void handle_stripe_clean_event(struct r5conf *conf,
3983         struct stripe_head *sh, int disks)
3984 {
3985         int i;
3986         struct r5dev *dev;
3987         int discard_pending = 0;
3988         struct stripe_head *head_sh = sh;
3989         bool do_endio = false;
3990
3991         for (i = disks; i--; )
3992                 if (sh->dev[i].written) {
3993                         dev = &sh->dev[i];
3994                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3995                             (test_bit(R5_UPTODATE, &dev->flags) ||
3996                              test_bit(R5_Discard, &dev->flags) ||
3997                              test_bit(R5_SkipCopy, &dev->flags))) {
3998                                 /* We can return any write requests */
3999                                 struct bio *wbi, *wbi2;
4000                                 pr_debug("Return write for disc %d\n", i);
4001                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
4002                                         clear_bit(R5_UPTODATE, &dev->flags);
4003                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
4004                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
4005                                 }
4006                                 do_endio = true;
4007
4008 returnbi:
4009                                 dev->page = dev->orig_page;
4010                                 wbi = dev->written;
4011                                 dev->written = NULL;
4012                                 while (wbi && wbi->bi_iter.bi_sector <
4013                                         dev->sector + RAID5_STRIPE_SECTORS(conf)) {
4014                                         wbi2 = r5_next_bio(conf, wbi, dev->sector);
4015                                         md_write_end(conf->mddev);
4016                                         bio_endio(wbi);
4017                                         wbi = wbi2;
4018                                 }
4019
4020                                 if (head_sh->batch_head) {
4021                                         sh = list_first_entry(&sh->batch_list,
4022                                                               struct stripe_head,
4023                                                               batch_list);
4024                                         if (sh != head_sh) {
4025                                                 dev = &sh->dev[i];
4026                                                 goto returnbi;
4027                                         }
4028                                 }
4029                                 sh = head_sh;
4030                                 dev = &sh->dev[i];
4031                         } else if (test_bit(R5_Discard, &dev->flags))
4032                                 discard_pending = 1;
4033                 }
4034
4035         log_stripe_write_finished(sh);
4036
4037         if (!discard_pending &&
4038             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
4039                 int hash;
4040                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
4041                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4042                 if (sh->qd_idx >= 0) {
4043                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
4044                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
4045                 }
4046                 /* now that discard is done we can proceed with any sync */
4047                 clear_bit(STRIPE_DISCARD, &sh->state);
4048                 /*
4049                  * SCSI discard will change some bio fields and the stripe has
4050                  * no updated data, so remove it from hash list and the stripe
4051                  * will be reinitialized
4052                  */
4053 unhash:
4054                 hash = sh->hash_lock_index;
4055                 spin_lock_irq(conf->hash_locks + hash);
4056                 remove_hash(sh);
4057                 spin_unlock_irq(conf->hash_locks + hash);
4058                 if (head_sh->batch_head) {
4059                         sh = list_first_entry(&sh->batch_list,
4060                                               struct stripe_head, batch_list);
4061                         if (sh != head_sh)
4062                                         goto unhash;
4063                 }
4064                 sh = head_sh;
4065
4066                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
4067                         set_bit(STRIPE_HANDLE, &sh->state);
4068
4069         }
4070
4071         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
4072                 if (atomic_dec_and_test(&conf->pending_full_writes))
4073                         md_wakeup_thread(conf->mddev->thread);
4074
4075         if (head_sh->batch_head && do_endio)
4076                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
4077 }
4078
4079 /*
4080  * For RMW in write back cache, we need extra page in prexor to store the
4081  * old data. This page is stored in dev->orig_page.
4082  *
4083  * This function checks whether we have data for prexor. The exact logic
4084  * is:
4085  *       R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
4086  */
4087 static inline bool uptodate_for_rmw(struct r5dev *dev)
4088 {
4089         return (test_bit(R5_UPTODATE, &dev->flags)) &&
4090                 (!test_bit(R5_InJournal, &dev->flags) ||
4091                  test_bit(R5_OrigPageUPTDODATE, &dev->flags));
4092 }
4093
4094 static int handle_stripe_dirtying(struct r5conf *conf,
4095                                   struct stripe_head *sh,
4096                                   struct stripe_head_state *s,
4097                                   int disks)
4098 {
4099         int rmw = 0, rcw = 0, i;
4100         sector_t recovery_cp = conf->mddev->recovery_cp;
4101
4102         /* Check whether resync is now happening or should start.
4103          * If yes, then the array is dirty (after unclean shutdown or
4104          * initial creation), so parity in some stripes might be inconsistent.
4105          * In this case, we need to always do reconstruct-write, to ensure
4106          * that in case of drive failure or read-error correction, we
4107          * generate correct data from the parity.
4108          */
4109         if (conf->rmw_level == PARITY_DISABLE_RMW ||
4110             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
4111              s->failed == 0)) {
4112                 /* Calculate the real rcw later - for now make it
4113                  * look like rcw is cheaper
4114                  */
4115                 rcw = 1; rmw = 2;
4116                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
4117                          conf->rmw_level, (unsigned long long)recovery_cp,
4118                          (unsigned long long)sh->sector);
4119         } else for (i = disks; i--; ) {
4120                 /* would I have to read this buffer for read_modify_write */
4121                 struct r5dev *dev = &sh->dev[i];
4122                 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4123                      i == sh->pd_idx || i == sh->qd_idx ||
4124                      test_bit(R5_InJournal, &dev->flags)) &&
4125                     !test_bit(R5_LOCKED, &dev->flags) &&
4126                     !(uptodate_for_rmw(dev) ||
4127                       test_bit(R5_Wantcompute, &dev->flags))) {
4128                         if (test_bit(R5_Insync, &dev->flags))
4129                                 rmw++;
4130                         else
4131                                 rmw += 2*disks;  /* cannot read it */
4132                 }
4133                 /* Would I have to read this buffer for reconstruct_write */
4134                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4135                     i != sh->pd_idx && i != sh->qd_idx &&
4136                     !test_bit(R5_LOCKED, &dev->flags) &&
4137                     !(test_bit(R5_UPTODATE, &dev->flags) ||
4138                       test_bit(R5_Wantcompute, &dev->flags))) {
4139                         if (test_bit(R5_Insync, &dev->flags))
4140                                 rcw++;
4141                         else
4142                                 rcw += 2*disks;
4143                 }
4144         }
4145
4146         pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
4147                  (unsigned long long)sh->sector, sh->state, rmw, rcw);
4148         set_bit(STRIPE_HANDLE, &sh->state);
4149         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
4150                 /* prefer read-modify-write, but need to get some data */
4151                 mddev_add_trace_msg(conf->mddev, "raid5 rmw %llu %d",
4152                                 sh->sector, rmw);
4153
4154                 for (i = disks; i--; ) {
4155                         struct r5dev *dev = &sh->dev[i];
4156                         if (test_bit(R5_InJournal, &dev->flags) &&
4157                             dev->page == dev->orig_page &&
4158                             !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
4159                                 /* alloc page for prexor */
4160                                 struct page *p = alloc_page(GFP_NOIO);
4161
4162                                 if (p) {
4163                                         dev->orig_page = p;
4164                                         continue;
4165                                 }
4166
4167                                 /*
4168                                  * alloc_page() failed, try use
4169                                  * disk_info->extra_page
4170                                  */
4171                                 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
4172                                                       &conf->cache_state)) {
4173                                         r5c_use_extra_page(sh);
4174                                         break;
4175                                 }
4176
4177                                 /* extra_page in use, add to delayed_list */
4178                                 set_bit(STRIPE_DELAYED, &sh->state);
4179                                 s->waiting_extra_page = 1;
4180                                 return -EAGAIN;
4181                         }
4182                 }
4183
4184                 for (i = disks; i--; ) {
4185                         struct r5dev *dev = &sh->dev[i];
4186                         if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4187                              i == sh->pd_idx || i == sh->qd_idx ||
4188                              test_bit(R5_InJournal, &dev->flags)) &&
4189                             !test_bit(R5_LOCKED, &dev->flags) &&
4190                             !(uptodate_for_rmw(dev) ||
4191                               test_bit(R5_Wantcompute, &dev->flags)) &&
4192                             test_bit(R5_Insync, &dev->flags)) {
4193                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
4194                                              &sh->state)) {
4195                                         pr_debug("Read_old block %d for r-m-w\n",
4196                                                  i);
4197                                         set_bit(R5_LOCKED, &dev->flags);
4198                                         set_bit(R5_Wantread, &dev->flags);
4199                                         s->locked++;
4200                                 } else
4201                                         set_bit(STRIPE_DELAYED, &sh->state);
4202                         }
4203                 }
4204         }
4205         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
4206                 /* want reconstruct write, but need to get some data */
4207                 int qread =0;
4208                 rcw = 0;
4209                 for (i = disks; i--; ) {
4210                         struct r5dev *dev = &sh->dev[i];
4211                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4212                             i != sh->pd_idx && i != sh->qd_idx &&
4213                             !test_bit(R5_LOCKED, &dev->flags) &&
4214                             !(test_bit(R5_UPTODATE, &dev->flags) ||
4215                               test_bit(R5_Wantcompute, &dev->flags))) {
4216                                 rcw++;
4217                                 if (test_bit(R5_Insync, &dev->flags) &&
4218                                     test_bit(STRIPE_PREREAD_ACTIVE,
4219                                              &sh->state)) {
4220                                         pr_debug("Read_old block "
4221                                                 "%d for Reconstruct\n", i);
4222                                         set_bit(R5_LOCKED, &dev->flags);
4223                                         set_bit(R5_Wantread, &dev->flags);
4224                                         s->locked++;
4225                                         qread++;
4226                                 } else
4227                                         set_bit(STRIPE_DELAYED, &sh->state);
4228                         }
4229                 }
4230                 if (rcw && !mddev_is_dm(conf->mddev))
4231                         blk_add_trace_msg(conf->mddev->gendisk->queue,
4232                                 "raid5 rcw %llu %d %d %d",
4233                                 (unsigned long long)sh->sector, rcw, qread,
4234                                 test_bit(STRIPE_DELAYED, &sh->state));
4235         }
4236
4237         if (rcw > disks && rmw > disks &&
4238             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4239                 set_bit(STRIPE_DELAYED, &sh->state);
4240
4241         /* now if nothing is locked, and if we have enough data,
4242          * we can start a write request
4243          */
4244         /* since handle_stripe can be called at any time we need to handle the
4245          * case where a compute block operation has been submitted and then a
4246          * subsequent call wants to start a write request.  raid_run_ops only
4247          * handles the case where compute block and reconstruct are requested
4248          * simultaneously.  If this is not the case then new writes need to be
4249          * held off until the compute completes.
4250          */
4251         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4252             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4253              !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4254                 schedule_reconstruction(sh, s, rcw == 0, 0);
4255         return 0;
4256 }
4257
4258 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4259                                 struct stripe_head_state *s, int disks)
4260 {
4261         struct r5dev *dev = NULL;
4262
4263         BUG_ON(sh->batch_head);
4264         set_bit(STRIPE_HANDLE, &sh->state);
4265
4266         switch (sh->check_state) {
4267         case check_state_idle:
4268                 /* start a new check operation if there are no failures */
4269                 if (s->failed == 0) {
4270                         BUG_ON(s->uptodate != disks);
4271                         sh->check_state = check_state_run;
4272                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4273                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4274                         s->uptodate--;
4275                         break;
4276                 }
4277                 dev = &sh->dev[s->failed_num[0]];
4278                 fallthrough;
4279         case check_state_compute_result:
4280                 sh->check_state = check_state_idle;
4281                 if (!dev)
4282                         dev = &sh->dev[sh->pd_idx];
4283
4284                 /* check that a write has not made the stripe insync */
4285                 if (test_bit(STRIPE_INSYNC, &sh->state))
4286                         break;
4287
4288                 /* either failed parity check, or recovery is happening */
4289                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4290                 BUG_ON(s->uptodate != disks);
4291
4292                 set_bit(R5_LOCKED, &dev->flags);
4293                 s->locked++;
4294                 set_bit(R5_Wantwrite, &dev->flags);
4295
4296                 set_bit(STRIPE_INSYNC, &sh->state);
4297                 break;
4298         case check_state_run:
4299                 break; /* we will be called again upon completion */
4300         case check_state_check_result:
4301                 sh->check_state = check_state_idle;
4302
4303                 /* if a failure occurred during the check operation, leave
4304                  * STRIPE_INSYNC not set and let the stripe be handled again
4305                  */
4306                 if (s->failed)
4307                         break;
4308
4309                 /* handle a successful check operation, if parity is correct
4310                  * we are done.  Otherwise update the mismatch count and repair
4311                  * parity if !MD_RECOVERY_CHECK
4312                  */
4313                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4314                         /* parity is correct (on disc,
4315                          * not in buffer any more)
4316                          */
4317                         set_bit(STRIPE_INSYNC, &sh->state);
4318                 else {
4319                         atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4320                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4321                                 /* don't try to repair!! */
4322                                 set_bit(STRIPE_INSYNC, &sh->state);
4323                                 pr_warn_ratelimited("%s: mismatch sector in range "
4324                                                     "%llu-%llu\n", mdname(conf->mddev),
4325                                                     (unsigned long long) sh->sector,
4326                                                     (unsigned long long) sh->sector +
4327                                                     RAID5_STRIPE_SECTORS(conf));
4328                         } else {
4329                                 sh->check_state = check_state_compute_run;
4330                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4331                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4332                                 set_bit(R5_Wantcompute,
4333                                         &sh->dev[sh->pd_idx].flags);
4334                                 sh->ops.target = sh->pd_idx;
4335                                 sh->ops.target2 = -1;
4336                                 s->uptodate++;
4337                         }
4338                 }
4339                 break;
4340         case check_state_compute_run:
4341                 break;
4342         default:
4343                 pr_err("%s: unknown check_state: %d sector: %llu\n",
4344                        __func__, sh->check_state,
4345                        (unsigned long long) sh->sector);
4346                 BUG();
4347         }
4348 }
4349
4350 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4351                                   struct stripe_head_state *s,
4352                                   int disks)
4353 {
4354         int pd_idx = sh->pd_idx;
4355         int qd_idx = sh->qd_idx;
4356         struct r5dev *dev;
4357
4358         BUG_ON(sh->batch_head);
4359         set_bit(STRIPE_HANDLE, &sh->state);
4360
4361         BUG_ON(s->failed > 2);
4362
4363         /* Want to check and possibly repair P and Q.
4364          * However there could be one 'failed' device, in which
4365          * case we can only check one of them, possibly using the
4366          * other to generate missing data
4367          */
4368
4369         switch (sh->check_state) {
4370         case check_state_idle:
4371                 /* start a new check operation if there are < 2 failures */
4372                 if (s->failed == s->q_failed) {
4373                         /* The only possible failed device holds Q, so it
4374                          * makes sense to check P (If anything else were failed,
4375                          * we would have used P to recreate it).
4376                          */
4377                         sh->check_state = check_state_run;
4378                 }
4379                 if (!s->q_failed && s->failed < 2) {
4380                         /* Q is not failed, and we didn't use it to generate
4381                          * anything, so it makes sense to check it
4382                          */
4383                         if (sh->check_state == check_state_run)
4384                                 sh->check_state = check_state_run_pq;
4385                         else
4386                                 sh->check_state = check_state_run_q;
4387                 }
4388
4389                 /* discard potentially stale zero_sum_result */
4390                 sh->ops.zero_sum_result = 0;
4391
4392                 if (sh->check_state == check_state_run) {
4393                         /* async_xor_zero_sum destroys the contents of P */
4394                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4395                         s->uptodate--;
4396                 }
4397                 if (sh->check_state >= check_state_run &&
4398                     sh->check_state <= check_state_run_pq) {
4399                         /* async_syndrome_zero_sum preserves P and Q, so
4400                          * no need to mark them !uptodate here
4401                          */
4402                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4403                         break;
4404                 }
4405
4406                 /* we have 2-disk failure */
4407                 BUG_ON(s->failed != 2);
4408                 fallthrough;
4409         case check_state_compute_result:
4410                 sh->check_state = check_state_idle;
4411
4412                 /* check that a write has not made the stripe insync */
4413                 if (test_bit(STRIPE_INSYNC, &sh->state))
4414                         break;
4415
4416                 /* now write out any block on a failed drive,
4417                  * or P or Q if they were recomputed
4418                  */
4419                 dev = NULL;
4420                 if (s->failed == 2) {
4421                         dev = &sh->dev[s->failed_num[1]];
4422                         s->locked++;
4423                         set_bit(R5_LOCKED, &dev->flags);
4424                         set_bit(R5_Wantwrite, &dev->flags);
4425                 }
4426                 if (s->failed >= 1) {
4427                         dev = &sh->dev[s->failed_num[0]];
4428                         s->locked++;
4429                         set_bit(R5_LOCKED, &dev->flags);
4430                         set_bit(R5_Wantwrite, &dev->flags);
4431                 }
4432                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4433                         dev = &sh->dev[pd_idx];
4434                         s->locked++;
4435                         set_bit(R5_LOCKED, &dev->flags);
4436                         set_bit(R5_Wantwrite, &dev->flags);
4437                 }
4438                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4439                         dev = &sh->dev[qd_idx];
4440                         s->locked++;
4441                         set_bit(R5_LOCKED, &dev->flags);
4442                         set_bit(R5_Wantwrite, &dev->flags);
4443                 }
4444                 if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4445                               "%s: disk%td not up to date\n",
4446                               mdname(conf->mddev),
4447                               dev - (struct r5dev *) &sh->dev)) {
4448                         clear_bit(R5_LOCKED, &dev->flags);
4449                         clear_bit(R5_Wantwrite, &dev->flags);
4450                         s->locked--;
4451                 }
4452
4453                 set_bit(STRIPE_INSYNC, &sh->state);
4454                 break;
4455         case check_state_run:
4456         case check_state_run_q:
4457         case check_state_run_pq:
4458                 break; /* we will be called again upon completion */
4459         case check_state_check_result:
4460                 sh->check_state = check_state_idle;
4461
4462                 /* handle a successful check operation, if parity is correct
4463                  * we are done.  Otherwise update the mismatch count and repair
4464                  * parity if !MD_RECOVERY_CHECK
4465                  */
4466                 if (sh->ops.zero_sum_result == 0) {
4467                         /* both parities are correct */
4468                         if (!s->failed)
4469                                 set_bit(STRIPE_INSYNC, &sh->state);
4470                         else {
4471                                 /* in contrast to the raid5 case we can validate
4472                                  * parity, but still have a failure to write
4473                                  * back
4474                                  */
4475                                 sh->check_state = check_state_compute_result;
4476                                 /* Returning at this point means that we may go
4477                                  * off and bring p and/or q uptodate again so
4478                                  * we make sure to check zero_sum_result again
4479                                  * to verify if p or q need writeback
4480                                  */
4481                         }
4482                 } else {
4483                         atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4484                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4485                                 /* don't try to repair!! */
4486                                 set_bit(STRIPE_INSYNC, &sh->state);
4487                                 pr_warn_ratelimited("%s: mismatch sector in range "
4488                                                     "%llu-%llu\n", mdname(conf->mddev),
4489                                                     (unsigned long long) sh->sector,
4490                                                     (unsigned long long) sh->sector +
4491                                                     RAID5_STRIPE_SECTORS(conf));
4492                         } else {
4493                                 int *target = &sh->ops.target;
4494
4495                                 sh->ops.target = -1;
4496                                 sh->ops.target2 = -1;
4497                                 sh->check_state = check_state_compute_run;
4498                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4499                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4500                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4501                                         set_bit(R5_Wantcompute,
4502                                                 &sh->dev[pd_idx].flags);
4503                                         *target = pd_idx;
4504                                         target = &sh->ops.target2;
4505                                         s->uptodate++;
4506                                 }
4507                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4508                                         set_bit(R5_Wantcompute,
4509                                                 &sh->dev[qd_idx].flags);
4510                                         *target = qd_idx;
4511                                         s->uptodate++;
4512                                 }
4513                         }
4514                 }
4515                 break;
4516         case check_state_compute_run:
4517                 break;
4518         default:
4519                 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4520                         __func__, sh->check_state,
4521                         (unsigned long long) sh->sector);
4522                 BUG();
4523         }
4524 }
4525
4526 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4527 {
4528         int i;
4529
4530         /* We have read all the blocks in this stripe and now we need to
4531          * copy some of them into a target stripe for expand.
4532          */
4533         struct dma_async_tx_descriptor *tx = NULL;
4534         BUG_ON(sh->batch_head);
4535         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4536         for (i = 0; i < sh->disks; i++)
4537                 if (i != sh->pd_idx && i != sh->qd_idx) {
4538                         int dd_idx, j;
4539                         struct stripe_head *sh2;
4540                         struct async_submit_ctl submit;
4541
4542                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
4543                         sector_t s = raid5_compute_sector(conf, bn, 0,
4544                                                           &dd_idx, NULL);
4545                         sh2 = raid5_get_active_stripe(conf, NULL, s,
4546                                 R5_GAS_NOBLOCK | R5_GAS_NOQUIESCE);
4547                         if (sh2 == NULL)
4548                                 /* so far only the early blocks of this stripe
4549                                  * have been requested.  When later blocks
4550                                  * get requested, we will try again
4551                                  */
4552                                 continue;
4553                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4554                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4555                                 /* must have already done this block */
4556                                 raid5_release_stripe(sh2);
4557                                 continue;
4558                         }
4559
4560                         /* place all the copies on one channel */
4561                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4562                         tx = async_memcpy(sh2->dev[dd_idx].page,
4563                                           sh->dev[i].page, sh2->dev[dd_idx].offset,
4564                                           sh->dev[i].offset, RAID5_STRIPE_SIZE(conf),
4565                                           &submit);
4566
4567                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4568                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4569                         for (j = 0; j < conf->raid_disks; j++)
4570                                 if (j != sh2->pd_idx &&
4571                                     j != sh2->qd_idx &&
4572                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
4573                                         break;
4574                         if (j == conf->raid_disks) {
4575                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4576                                 set_bit(STRIPE_HANDLE, &sh2->state);
4577                         }
4578                         raid5_release_stripe(sh2);
4579
4580                 }
4581         /* done submitting copies, wait for them to complete */
4582         async_tx_quiesce(&tx);
4583 }
4584
4585 /*
4586  * handle_stripe - do things to a stripe.
4587  *
4588  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4589  * state of various bits to see what needs to be done.
4590  * Possible results:
4591  *    return some read requests which now have data
4592  *    return some write requests which are safely on storage
4593  *    schedule a read on some buffers
4594  *    schedule a write of some buffers
4595  *    return confirmation of parity correctness
4596  *
4597  */
4598
4599 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4600 {
4601         struct r5conf *conf = sh->raid_conf;
4602         int disks = sh->disks;
4603         struct r5dev *dev;
4604         int i;
4605         int do_recovery = 0;
4606
4607         memset(s, 0, sizeof(*s));
4608
4609         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4610         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4611         s->failed_num[0] = -1;
4612         s->failed_num[1] = -1;
4613         s->log_failed = r5l_log_disk_error(conf);
4614
4615         /* Now to look around and see what can be done */
4616         for (i=disks; i--; ) {
4617                 struct md_rdev *rdev;
4618                 int is_bad = 0;
4619
4620                 dev = &sh->dev[i];
4621
4622                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4623                          i, dev->flags,
4624                          dev->toread, dev->towrite, dev->written);
4625                 /* maybe we can reply to a read
4626                  *
4627                  * new wantfill requests are only permitted while
4628                  * ops_complete_biofill is guaranteed to be inactive
4629                  */
4630                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4631                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4632                         set_bit(R5_Wantfill, &dev->flags);
4633
4634                 /* now count some things */
4635                 if (test_bit(R5_LOCKED, &dev->flags))
4636                         s->locked++;
4637                 if (test_bit(R5_UPTODATE, &dev->flags))
4638                         s->uptodate++;
4639                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4640                         s->compute++;
4641                         BUG_ON(s->compute > 2);
4642                 }
4643
4644                 if (test_bit(R5_Wantfill, &dev->flags))
4645                         s->to_fill++;
4646                 else if (dev->toread)
4647                         s->to_read++;
4648                 if (dev->towrite) {
4649                         s->to_write++;
4650                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4651                                 s->non_overwrite++;
4652                 }
4653                 if (dev->written)
4654                         s->written++;
4655                 /* Prefer to use the replacement for reads, but only
4656                  * if it is recovered enough and has no bad blocks.
4657                  */
4658                 rdev = conf->disks[i].replacement;
4659                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4660                     rdev->recovery_offset >= sh->sector + RAID5_STRIPE_SECTORS(conf) &&
4661                     !rdev_has_badblock(rdev, sh->sector,
4662                                        RAID5_STRIPE_SECTORS(conf)))
4663                         set_bit(R5_ReadRepl, &dev->flags);
4664                 else {
4665                         if (rdev && !test_bit(Faulty, &rdev->flags))
4666                                 set_bit(R5_NeedReplace, &dev->flags);
4667                         else
4668                                 clear_bit(R5_NeedReplace, &dev->flags);
4669                         rdev = conf->disks[i].rdev;
4670                         clear_bit(R5_ReadRepl, &dev->flags);
4671                 }
4672                 if (rdev && test_bit(Faulty, &rdev->flags))
4673                         rdev = NULL;
4674                 if (rdev) {
4675                         is_bad = rdev_has_badblock(rdev, sh->sector,
4676                                                    RAID5_STRIPE_SECTORS(conf));
4677                         if (s->blocked_rdev == NULL) {
4678                                 if (is_bad < 0)
4679                                         set_bit(BlockedBadBlocks, &rdev->flags);
4680                                 if (rdev_blocked(rdev)) {
4681                                         s->blocked_rdev = rdev;
4682                                         atomic_inc(&rdev->nr_pending);
4683                                 }
4684                         }
4685                 }
4686                 clear_bit(R5_Insync, &dev->flags);
4687                 if (!rdev)
4688                         /* Not in-sync */;
4689                 else if (is_bad) {
4690                         /* also not in-sync */
4691                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4692                             test_bit(R5_UPTODATE, &dev->flags)) {
4693                                 /* treat as in-sync, but with a read error
4694                                  * which we can now try to correct
4695                                  */
4696                                 set_bit(R5_Insync, &dev->flags);
4697                                 set_bit(R5_ReadError, &dev->flags);
4698                         }
4699                 } else if (test_bit(In_sync, &rdev->flags))
4700                         set_bit(R5_Insync, &dev->flags);
4701                 else if (sh->sector + RAID5_STRIPE_SECTORS(conf) <= rdev->recovery_offset)
4702                         /* in sync if before recovery_offset */
4703                         set_bit(R5_Insync, &dev->flags);
4704                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4705                          test_bit(R5_Expanded, &dev->flags))
4706                         /* If we've reshaped into here, we assume it is Insync.
4707                          * We will shortly update recovery_offset to make
4708                          * it official.
4709                          */
4710                         set_bit(R5_Insync, &dev->flags);
4711
4712                 if (test_bit(R5_WriteError, &dev->flags)) {
4713                         /* This flag does not apply to '.replacement'
4714                          * only to .rdev, so make sure to check that*/
4715                         struct md_rdev *rdev2 = conf->disks[i].rdev;
4716
4717                         if (rdev2 == rdev)
4718                                 clear_bit(R5_Insync, &dev->flags);
4719                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4720                                 s->handle_bad_blocks = 1;
4721                                 atomic_inc(&rdev2->nr_pending);
4722                         } else
4723                                 clear_bit(R5_WriteError, &dev->flags);
4724                 }
4725                 if (test_bit(R5_MadeGood, &dev->flags)) {
4726                         /* This flag does not apply to '.replacement'
4727                          * only to .rdev, so make sure to check that*/
4728                         struct md_rdev *rdev2 = conf->disks[i].rdev;
4729
4730                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4731                                 s->handle_bad_blocks = 1;
4732                                 atomic_inc(&rdev2->nr_pending);
4733                         } else
4734                                 clear_bit(R5_MadeGood, &dev->flags);
4735                 }
4736                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4737                         struct md_rdev *rdev2 = conf->disks[i].replacement;
4738
4739                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4740                                 s->handle_bad_blocks = 1;
4741                                 atomic_inc(&rdev2->nr_pending);
4742                         } else
4743                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4744                 }
4745                 if (!test_bit(R5_Insync, &dev->flags)) {
4746                         /* The ReadError flag will just be confusing now */
4747                         clear_bit(R5_ReadError, &dev->flags);
4748                         clear_bit(R5_ReWrite, &dev->flags);
4749                 }
4750                 if (test_bit(R5_ReadError, &dev->flags))
4751                         clear_bit(R5_Insync, &dev->flags);
4752                 if (!test_bit(R5_Insync, &dev->flags)) {
4753                         if (s->failed < 2)
4754                                 s->failed_num[s->failed] = i;
4755                         s->failed++;
4756                         if (rdev && !test_bit(Faulty, &rdev->flags))
4757                                 do_recovery = 1;
4758                         else if (!rdev) {
4759                                 rdev = conf->disks[i].replacement;
4760                                 if (rdev && !test_bit(Faulty, &rdev->flags))
4761                                         do_recovery = 1;
4762                         }
4763                 }
4764
4765                 if (test_bit(R5_InJournal, &dev->flags))
4766                         s->injournal++;
4767                 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4768                         s->just_cached++;
4769         }
4770         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4771                 /* If there is a failed device being replaced,
4772                  *     we must be recovering.
4773                  * else if we are after recovery_cp, we must be syncing
4774                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4775                  * else we can only be replacing
4776                  * sync and recovery both need to read all devices, and so
4777                  * use the same flag.
4778                  */
4779                 if (do_recovery ||
4780                     sh->sector >= conf->mddev->recovery_cp ||
4781                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4782                         s->syncing = 1;
4783                 else
4784                         s->replacing = 1;
4785         }
4786 }
4787
4788 /*
4789  * Return '1' if this is a member of batch, or '0' if it is a lone stripe or
4790  * a head which can now be handled.
4791  */
4792 static int clear_batch_ready(struct stripe_head *sh)
4793 {
4794         struct stripe_head *tmp;
4795         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4796                 return (sh->batch_head && sh->batch_head != sh);
4797         spin_lock(&sh->stripe_lock);
4798         if (!sh->batch_head) {
4799                 spin_unlock(&sh->stripe_lock);
4800                 return 0;
4801         }
4802
4803         /*
4804          * this stripe could be added to a batch list before we check
4805          * BATCH_READY, skips it
4806          */
4807         if (sh->batch_head != sh) {
4808                 spin_unlock(&sh->stripe_lock);
4809                 return 1;
4810         }
4811         spin_lock(&sh->batch_lock);
4812         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4813                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4814         spin_unlock(&sh->batch_lock);
4815         spin_unlock(&sh->stripe_lock);
4816
4817         /*
4818          * BATCH_READY is cleared, no new stripes can be added.
4819          * batch_list can be accessed without lock
4820          */
4821         return 0;
4822 }
4823
4824 static void break_stripe_batch_list(struct stripe_head *head_sh,
4825                                     unsigned long handle_flags)
4826 {
4827         struct stripe_head *sh, *next;
4828         int i;
4829
4830         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4831
4832                 list_del_init(&sh->batch_list);
4833
4834                 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4835                                           (1 << STRIPE_SYNCING) |
4836                                           (1 << STRIPE_REPLACED) |
4837                                           (1 << STRIPE_DELAYED) |
4838                                           (1 << STRIPE_BIT_DELAY) |
4839                                           (1 << STRIPE_FULL_WRITE) |
4840                                           (1 << STRIPE_BIOFILL_RUN) |
4841                                           (1 << STRIPE_COMPUTE_RUN)  |
4842                                           (1 << STRIPE_DISCARD) |
4843                                           (1 << STRIPE_BATCH_READY) |
4844                                           (1 << STRIPE_BATCH_ERR)),
4845                         "stripe state: %lx\n", sh->state);
4846                 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4847                                               (1 << STRIPE_REPLACED)),
4848                         "head stripe state: %lx\n", head_sh->state);
4849
4850                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4851                                             (1 << STRIPE_PREREAD_ACTIVE) |
4852                                             (1 << STRIPE_ON_UNPLUG_LIST)),
4853                               head_sh->state & (1 << STRIPE_INSYNC));
4854
4855                 sh->check_state = head_sh->check_state;
4856                 sh->reconstruct_state = head_sh->reconstruct_state;
4857                 spin_lock_irq(&sh->stripe_lock);
4858                 sh->batch_head = NULL;
4859                 spin_unlock_irq(&sh->stripe_lock);
4860                 for (i = 0; i < sh->disks; i++) {
4861                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4862                                 wake_up_bit(&sh->dev[i].flags, R5_Overlap);
4863                         sh->dev[i].flags = head_sh->dev[i].flags &
4864                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4865                 }
4866                 if (handle_flags == 0 ||
4867                     sh->state & handle_flags)
4868                         set_bit(STRIPE_HANDLE, &sh->state);
4869                 raid5_release_stripe(sh);
4870         }
4871         spin_lock_irq(&head_sh->stripe_lock);
4872         head_sh->batch_head = NULL;
4873         spin_unlock_irq(&head_sh->stripe_lock);
4874         for (i = 0; i < head_sh->disks; i++)
4875                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4876                         wake_up_bit(&head_sh->dev[i].flags, R5_Overlap);
4877         if (head_sh->state & handle_flags)
4878                 set_bit(STRIPE_HANDLE, &head_sh->state);
4879 }
4880
4881 static void handle_stripe(struct stripe_head *sh)
4882 {
4883         struct stripe_head_state s;
4884         struct r5conf *conf = sh->raid_conf;
4885         int i;
4886         int prexor;
4887         int disks = sh->disks;
4888         struct r5dev *pdev, *qdev;
4889
4890         clear_bit(STRIPE_HANDLE, &sh->state);
4891
4892         /*
4893          * handle_stripe should not continue handle the batched stripe, only
4894          * the head of batch list or lone stripe can continue. Otherwise we
4895          * could see break_stripe_batch_list warns about the STRIPE_ACTIVE
4896          * is set for the batched stripe.
4897          */
4898         if (clear_batch_ready(sh))
4899                 return;
4900
4901         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4902                 /* already being handled, ensure it gets handled
4903                  * again when current action finishes */
4904                 set_bit(STRIPE_HANDLE, &sh->state);
4905                 return;
4906         }
4907
4908         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4909                 break_stripe_batch_list(sh, 0);
4910
4911         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4912                 spin_lock(&sh->stripe_lock);
4913                 /*
4914                  * Cannot process 'sync' concurrently with 'discard'.
4915                  * Flush data in r5cache before 'sync'.
4916                  */
4917                 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4918                     !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4919                     !test_bit(STRIPE_DISCARD, &sh->state) &&
4920                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4921                         set_bit(STRIPE_SYNCING, &sh->state);
4922                         clear_bit(STRIPE_INSYNC, &sh->state);
4923                         clear_bit(STRIPE_REPLACED, &sh->state);
4924                 }
4925                 spin_unlock(&sh->stripe_lock);
4926         }
4927         clear_bit(STRIPE_DELAYED, &sh->state);
4928
4929         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4930                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4931                (unsigned long long)sh->sector, sh->state,
4932                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4933                sh->check_state, sh->reconstruct_state);
4934
4935         analyse_stripe(sh, &s);
4936
4937         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4938                 goto finish;
4939
4940         if (s.handle_bad_blocks ||
4941             test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4942                 set_bit(STRIPE_HANDLE, &sh->state);
4943                 goto finish;
4944         }
4945
4946         if (unlikely(s.blocked_rdev)) {
4947                 if (s.syncing || s.expanding || s.expanded ||
4948                     s.replacing || s.to_write || s.written) {
4949                         set_bit(STRIPE_HANDLE, &sh->state);
4950                         goto finish;
4951                 }
4952                 /* There is nothing for the blocked_rdev to block */
4953                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4954                 s.blocked_rdev = NULL;
4955         }
4956
4957         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4958                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4959                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4960         }
4961
4962         pr_debug("locked=%d uptodate=%d to_read=%d"
4963                " to_write=%d failed=%d failed_num=%d,%d\n",
4964                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4965                s.failed_num[0], s.failed_num[1]);
4966         /*
4967          * check if the array has lost more than max_degraded devices and,
4968          * if so, some requests might need to be failed.
4969          *
4970          * When journal device failed (log_failed), we will only process
4971          * the stripe if there is data need write to raid disks
4972          */
4973         if (s.failed > conf->max_degraded ||
4974             (s.log_failed && s.injournal == 0)) {
4975                 sh->check_state = 0;
4976                 sh->reconstruct_state = 0;
4977                 break_stripe_batch_list(sh, 0);
4978                 if (s.to_read+s.to_write+s.written)
4979                         handle_failed_stripe(conf, sh, &s, disks);
4980                 if (s.syncing + s.replacing)
4981                         handle_failed_sync(conf, sh, &s);
4982         }
4983
4984         /* Now we check to see if any write operations have recently
4985          * completed
4986          */
4987         prexor = 0;
4988         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4989                 prexor = 1;
4990         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4991             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4992                 sh->reconstruct_state = reconstruct_state_idle;
4993
4994                 /* All the 'written' buffers and the parity block are ready to
4995                  * be written back to disk
4996                  */
4997                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4998                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4999                 BUG_ON(sh->qd_idx >= 0 &&
5000                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
5001                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
5002                 for (i = disks; i--; ) {
5003                         struct r5dev *dev = &sh->dev[i];
5004                         if (test_bit(R5_LOCKED, &dev->flags) &&
5005                                 (i == sh->pd_idx || i == sh->qd_idx ||
5006                                  dev->written || test_bit(R5_InJournal,
5007                                                           &dev->flags))) {
5008                                 pr_debug("Writing block %d\n", i);
5009                                 set_bit(R5_Wantwrite, &dev->flags);
5010                                 if (prexor)
5011                                         continue;
5012                                 if (s.failed > 1)
5013                                         continue;
5014                                 if (!test_bit(R5_Insync, &dev->flags) ||
5015                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
5016                                      s.failed == 0))
5017                                         set_bit(STRIPE_INSYNC, &sh->state);
5018                         }
5019                 }
5020                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5021                         s.dec_preread_active = 1;
5022         }
5023
5024         /*
5025          * might be able to return some write requests if the parity blocks
5026          * are safe, or on a failed drive
5027          */
5028         pdev = &sh->dev[sh->pd_idx];
5029         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
5030                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
5031         qdev = &sh->dev[sh->qd_idx];
5032         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
5033                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
5034                 || conf->level < 6;
5035
5036         if (s.written &&
5037             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
5038                              && !test_bit(R5_LOCKED, &pdev->flags)
5039                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
5040                                  test_bit(R5_Discard, &pdev->flags))))) &&
5041             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
5042                              && !test_bit(R5_LOCKED, &qdev->flags)
5043                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
5044                                  test_bit(R5_Discard, &qdev->flags))))))
5045                 handle_stripe_clean_event(conf, sh, disks);
5046
5047         if (s.just_cached)
5048                 r5c_handle_cached_data_endio(conf, sh, disks);
5049         log_stripe_write_finished(sh);
5050
5051         /* Now we might consider reading some blocks, either to check/generate
5052          * parity, or to satisfy requests
5053          * or to load a block that is being partially written.
5054          */
5055         if (s.to_read || s.non_overwrite
5056             || (s.to_write && s.failed)
5057             || (s.syncing && (s.uptodate + s.compute < disks))
5058             || s.replacing
5059             || s.expanding)
5060                 handle_stripe_fill(sh, &s, disks);
5061
5062         /*
5063          * When the stripe finishes full journal write cycle (write to journal
5064          * and raid disk), this is the clean up procedure so it is ready for
5065          * next operation.
5066          */
5067         r5c_finish_stripe_write_out(conf, sh, &s);
5068
5069         /*
5070          * Now to consider new write requests, cache write back and what else,
5071          * if anything should be read.  We do not handle new writes when:
5072          * 1/ A 'write' operation (copy+xor) is already in flight.
5073          * 2/ A 'check' operation is in flight, as it may clobber the parity
5074          *    block.
5075          * 3/ A r5c cache log write is in flight.
5076          */
5077
5078         if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
5079                 if (!r5c_is_writeback(conf->log)) {
5080                         if (s.to_write)
5081                                 handle_stripe_dirtying(conf, sh, &s, disks);
5082                 } else { /* write back cache */
5083                         int ret = 0;
5084
5085                         /* First, try handle writes in caching phase */
5086                         if (s.to_write)
5087                                 ret = r5c_try_caching_write(conf, sh, &s,
5088                                                             disks);
5089                         /*
5090                          * If caching phase failed: ret == -EAGAIN
5091                          *    OR
5092                          * stripe under reclaim: !caching && injournal
5093                          *
5094                          * fall back to handle_stripe_dirtying()
5095                          */
5096                         if (ret == -EAGAIN ||
5097                             /* stripe under reclaim: !caching && injournal */
5098                             (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
5099                              s.injournal > 0)) {
5100                                 ret = handle_stripe_dirtying(conf, sh, &s,
5101                                                              disks);
5102                                 if (ret == -EAGAIN)
5103                                         goto finish;
5104                         }
5105                 }
5106         }
5107
5108         /* maybe we need to check and possibly fix the parity for this stripe
5109          * Any reads will already have been scheduled, so we just see if enough
5110          * data is available.  The parity check is held off while parity
5111          * dependent operations are in flight.
5112          */
5113         if (sh->check_state ||
5114             (s.syncing && s.locked == 0 &&
5115              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5116              !test_bit(STRIPE_INSYNC, &sh->state))) {
5117                 if (conf->level == 6)
5118                         handle_parity_checks6(conf, sh, &s, disks);
5119                 else
5120                         handle_parity_checks5(conf, sh, &s, disks);
5121         }
5122
5123         if ((s.replacing || s.syncing) && s.locked == 0
5124             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
5125             && !test_bit(STRIPE_REPLACED, &sh->state)) {
5126                 /* Write out to replacement devices where possible */
5127                 for (i = 0; i < conf->raid_disks; i++)
5128                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
5129                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
5130                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
5131                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
5132                                 s.locked++;
5133                         }
5134                 if (s.replacing)
5135                         set_bit(STRIPE_INSYNC, &sh->state);
5136                 set_bit(STRIPE_REPLACED, &sh->state);
5137         }
5138         if ((s.syncing || s.replacing) && s.locked == 0 &&
5139             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5140             test_bit(STRIPE_INSYNC, &sh->state)) {
5141                 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5142                 clear_bit(STRIPE_SYNCING, &sh->state);
5143                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
5144                         wake_up_bit(&sh->dev[sh->pd_idx].flags, R5_Overlap);
5145         }
5146
5147         /* If the failed drives are just a ReadError, then we might need
5148          * to progress the repair/check process
5149          */
5150         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
5151                 for (i = 0; i < s.failed; i++) {
5152                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
5153                         if (test_bit(R5_ReadError, &dev->flags)
5154                             && !test_bit(R5_LOCKED, &dev->flags)
5155                             && test_bit(R5_UPTODATE, &dev->flags)
5156                                 ) {
5157                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
5158                                         set_bit(R5_Wantwrite, &dev->flags);
5159                                         set_bit(R5_ReWrite, &dev->flags);
5160                                 } else
5161                                         /* let's read it back */
5162                                         set_bit(R5_Wantread, &dev->flags);
5163                                 set_bit(R5_LOCKED, &dev->flags);
5164                                 s.locked++;
5165                         }
5166                 }
5167
5168         /* Finish reconstruct operations initiated by the expansion process */
5169         if (sh->reconstruct_state == reconstruct_state_result) {
5170                 struct stripe_head *sh_src
5171                         = raid5_get_active_stripe(conf, NULL, sh->sector,
5172                                         R5_GAS_PREVIOUS | R5_GAS_NOBLOCK |
5173                                         R5_GAS_NOQUIESCE);
5174                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
5175                         /* sh cannot be written until sh_src has been read.
5176                          * so arrange for sh to be delayed a little
5177                          */
5178                         set_bit(STRIPE_DELAYED, &sh->state);
5179                         set_bit(STRIPE_HANDLE, &sh->state);
5180                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
5181                                               &sh_src->state))
5182                                 atomic_inc(&conf->preread_active_stripes);
5183                         raid5_release_stripe(sh_src);
5184                         goto finish;
5185                 }
5186                 if (sh_src)
5187                         raid5_release_stripe(sh_src);
5188
5189                 sh->reconstruct_state = reconstruct_state_idle;
5190                 clear_bit(STRIPE_EXPANDING, &sh->state);
5191                 for (i = conf->raid_disks; i--; ) {
5192                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
5193                         set_bit(R5_LOCKED, &sh->dev[i].flags);
5194                         s.locked++;
5195                 }
5196         }
5197
5198         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
5199             !sh->reconstruct_state) {
5200                 /* Need to write out all blocks after computing parity */
5201                 sh->disks = conf->raid_disks;
5202                 stripe_set_idx(sh->sector, conf, 0, sh);
5203                 schedule_reconstruction(sh, &s, 1, 1);
5204         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
5205                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
5206                 atomic_dec(&conf->reshape_stripes);
5207                 wake_up(&conf->wait_for_reshape);
5208                 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5209         }
5210
5211         if (s.expanding && s.locked == 0 &&
5212             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
5213                 handle_stripe_expansion(conf, sh);
5214
5215 finish:
5216         /* wait for this device to become unblocked */
5217         if (unlikely(s.blocked_rdev)) {
5218                 if (conf->mddev->external)
5219                         md_wait_for_blocked_rdev(s.blocked_rdev,
5220                                                  conf->mddev);
5221                 else
5222                         /* Internal metadata will immediately
5223                          * be written by raid5d, so we don't
5224                          * need to wait here.
5225                          */
5226                         rdev_dec_pending(s.blocked_rdev,
5227                                          conf->mddev);
5228         }
5229
5230         if (s.handle_bad_blocks)
5231                 for (i = disks; i--; ) {
5232                         struct md_rdev *rdev;
5233                         struct r5dev *dev = &sh->dev[i];
5234                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5235                                 /* We own a safe reference to the rdev */
5236                                 rdev = conf->disks[i].rdev;
5237                                 if (!rdev_set_badblocks(rdev, sh->sector,
5238                                                         RAID5_STRIPE_SECTORS(conf), 0))
5239                                         md_error(conf->mddev, rdev);
5240                                 rdev_dec_pending(rdev, conf->mddev);
5241                         }
5242                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5243                                 rdev = conf->disks[i].rdev;
5244                                 rdev_clear_badblocks(rdev, sh->sector,
5245                                                      RAID5_STRIPE_SECTORS(conf), 0);
5246                                 rdev_dec_pending(rdev, conf->mddev);
5247                         }
5248                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5249                                 rdev = conf->disks[i].replacement;
5250                                 if (!rdev)
5251                                         /* rdev have been moved down */
5252                                         rdev = conf->disks[i].rdev;
5253                                 rdev_clear_badblocks(rdev, sh->sector,
5254                                                      RAID5_STRIPE_SECTORS(conf), 0);
5255                                 rdev_dec_pending(rdev, conf->mddev);
5256                         }
5257                 }
5258
5259         if (s.ops_request)
5260                 raid_run_ops(sh, s.ops_request);
5261
5262         ops_run_io(sh, &s);
5263
5264         if (s.dec_preread_active) {
5265                 /* We delay this until after ops_run_io so that if make_request
5266                  * is waiting on a flush, it won't continue until the writes
5267                  * have actually been submitted.
5268                  */
5269                 atomic_dec(&conf->preread_active_stripes);
5270                 if (atomic_read(&conf->preread_active_stripes) <
5271                     IO_THRESHOLD)
5272                         md_wakeup_thread(conf->mddev->thread);
5273         }
5274
5275         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5276 }
5277
5278 static void raid5_activate_delayed(struct r5conf *conf)
5279         __must_hold(&conf->device_lock)
5280 {
5281         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5282                 while (!list_empty(&conf->delayed_list)) {
5283                         struct list_head *l = conf->delayed_list.next;
5284                         struct stripe_head *sh;
5285                         sh = list_entry(l, struct stripe_head, lru);
5286                         list_del_init(l);
5287                         clear_bit(STRIPE_DELAYED, &sh->state);
5288                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5289                                 atomic_inc(&conf->preread_active_stripes);
5290                         list_add_tail(&sh->lru, &conf->hold_list);
5291                         raid5_wakeup_stripe_thread(sh);
5292                 }
5293         }
5294 }
5295
5296 static void activate_bit_delay(struct r5conf *conf,
5297                 struct list_head *temp_inactive_list)
5298         __must_hold(&conf->device_lock)
5299 {
5300         struct list_head head;
5301         list_add(&head, &conf->bitmap_list);
5302         list_del_init(&conf->bitmap_list);
5303         while (!list_empty(&head)) {
5304                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5305                 int hash;
5306                 list_del_init(&sh->lru);
5307                 atomic_inc(&sh->count);
5308                 hash = sh->hash_lock_index;
5309                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5310         }
5311 }
5312
5313 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5314 {
5315         struct r5conf *conf = mddev->private;
5316         sector_t sector = bio->bi_iter.bi_sector;
5317         unsigned int chunk_sectors;
5318         unsigned int bio_sectors = bio_sectors(bio);
5319
5320         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5321         return  chunk_sectors >=
5322                 ((sector & (chunk_sectors - 1)) + bio_sectors);
5323 }
5324
5325 /*
5326  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
5327  *  later sampled by raid5d.
5328  */
5329 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5330 {
5331         unsigned long flags;
5332
5333         spin_lock_irqsave(&conf->device_lock, flags);
5334
5335         bi->bi_next = conf->retry_read_aligned_list;
5336         conf->retry_read_aligned_list = bi;
5337
5338         spin_unlock_irqrestore(&conf->device_lock, flags);
5339         md_wakeup_thread(conf->mddev->thread);
5340 }
5341
5342 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5343                                          unsigned int *offset)
5344 {
5345         struct bio *bi;
5346
5347         bi = conf->retry_read_aligned;
5348         if (bi) {
5349                 *offset = conf->retry_read_offset;
5350                 conf->retry_read_aligned = NULL;
5351                 return bi;
5352         }
5353         bi = conf->retry_read_aligned_list;
5354         if(bi) {
5355                 conf->retry_read_aligned_list = bi->bi_next;
5356                 bi->bi_next = NULL;
5357                 *offset = 0;
5358         }
5359
5360         return bi;
5361 }
5362
5363 /*
5364  *  The "raid5_align_endio" should check if the read succeeded and if it
5365  *  did, call bio_endio on the original bio (having bio_put the new bio
5366  *  first).
5367  *  If the read failed..
5368  */
5369 static void raid5_align_endio(struct bio *bi)
5370 {
5371         struct bio *raid_bi = bi->bi_private;
5372         struct md_rdev *rdev = (void *)raid_bi->bi_next;
5373         struct mddev *mddev = rdev->mddev;
5374         struct r5conf *conf = mddev->private;
5375         blk_status_t error = bi->bi_status;
5376
5377         bio_put(bi);
5378         raid_bi->bi_next = NULL;
5379         rdev_dec_pending(rdev, conf->mddev);
5380
5381         if (!error) {
5382                 bio_endio(raid_bi);
5383                 if (atomic_dec_and_test(&conf->active_aligned_reads))
5384                         wake_up(&conf->wait_for_quiescent);
5385                 return;
5386         }
5387
5388         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5389
5390         add_bio_to_retry(raid_bi, conf);
5391 }
5392
5393 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5394 {
5395         struct r5conf *conf = mddev->private;
5396         struct bio *align_bio;
5397         struct md_rdev *rdev;
5398         sector_t sector, end_sector;
5399         int dd_idx;
5400         bool did_inc;
5401
5402         if (!in_chunk_boundary(mddev, raid_bio)) {
5403                 pr_debug("%s: non aligned\n", __func__);
5404                 return 0;
5405         }
5406
5407         sector = raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector, 0,
5408                                       &dd_idx, NULL);
5409         end_sector = sector + bio_sectors(raid_bio);
5410
5411         if (r5c_big_stripe_cached(conf, sector))
5412                 return 0;
5413
5414         rdev = conf->disks[dd_idx].replacement;
5415         if (!rdev || test_bit(Faulty, &rdev->flags) ||
5416             rdev->recovery_offset < end_sector) {
5417                 rdev = conf->disks[dd_idx].rdev;
5418                 if (!rdev)
5419                         return 0;
5420                 if (test_bit(Faulty, &rdev->flags) ||
5421                     !(test_bit(In_sync, &rdev->flags) ||
5422                       rdev->recovery_offset >= end_sector))
5423                         return 0;
5424         }
5425
5426         atomic_inc(&rdev->nr_pending);
5427
5428         if (rdev_has_badblock(rdev, sector, bio_sectors(raid_bio))) {
5429                 rdev_dec_pending(rdev, mddev);
5430                 return 0;
5431         }
5432
5433         md_account_bio(mddev, &raid_bio);
5434         raid_bio->bi_next = (void *)rdev;
5435
5436         align_bio = bio_alloc_clone(rdev->bdev, raid_bio, GFP_NOIO,
5437                                     &mddev->bio_set);
5438         align_bio->bi_end_io = raid5_align_endio;
5439         align_bio->bi_private = raid_bio;
5440         align_bio->bi_iter.bi_sector = sector;
5441
5442         /* No reshape active, so we can trust rdev->data_offset */
5443         align_bio->bi_iter.bi_sector += rdev->data_offset;
5444
5445         did_inc = false;
5446         if (conf->quiesce == 0) {
5447                 atomic_inc(&conf->active_aligned_reads);
5448                 did_inc = true;
5449         }
5450         /* need a memory barrier to detect the race with raid5_quiesce() */
5451         if (!did_inc || smp_load_acquire(&conf->quiesce) != 0) {
5452                 /* quiesce is in progress, so we need to undo io activation and wait
5453                  * for it to finish
5454                  */
5455                 if (did_inc && atomic_dec_and_test(&conf->active_aligned_reads))
5456                         wake_up(&conf->wait_for_quiescent);
5457                 spin_lock_irq(&conf->device_lock);
5458                 wait_event_lock_irq(conf->wait_for_quiescent, conf->quiesce == 0,
5459                                     conf->device_lock);
5460                 atomic_inc(&conf->active_aligned_reads);
5461                 spin_unlock_irq(&conf->device_lock);
5462         }
5463
5464         mddev_trace_remap(mddev, align_bio, raid_bio->bi_iter.bi_sector);
5465         submit_bio_noacct(align_bio);
5466         return 1;
5467 }
5468
5469 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5470 {
5471         struct bio *split;
5472         sector_t sector = raid_bio->bi_iter.bi_sector;
5473         unsigned chunk_sects = mddev->chunk_sectors;
5474         unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5475
5476         if (sectors < bio_sectors(raid_bio)) {
5477                 struct r5conf *conf = mddev->private;
5478                 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5479                 bio_chain(split, raid_bio);
5480                 submit_bio_noacct(raid_bio);
5481                 raid_bio = split;
5482         }
5483
5484         if (!raid5_read_one_chunk(mddev, raid_bio))
5485                 return raid_bio;
5486
5487         return NULL;
5488 }
5489
5490 /* __get_priority_stripe - get the next stripe to process
5491  *
5492  * Full stripe writes are allowed to pass preread active stripes up until
5493  * the bypass_threshold is exceeded.  In general the bypass_count
5494  * increments when the handle_list is handled before the hold_list; however, it
5495  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5496  * stripe with in flight i/o.  The bypass_count will be reset when the
5497  * head of the hold_list has changed, i.e. the head was promoted to the
5498  * handle_list.
5499  */
5500 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5501         __must_hold(&conf->device_lock)
5502 {
5503         struct stripe_head *sh, *tmp;
5504         struct list_head *handle_list = NULL;
5505         struct r5worker_group *wg;
5506         bool second_try = !r5c_is_writeback(conf->log) &&
5507                 !r5l_log_disk_error(conf);
5508         bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5509                 r5l_log_disk_error(conf);
5510
5511 again:
5512         wg = NULL;
5513         sh = NULL;
5514         if (conf->worker_cnt_per_group == 0) {
5515                 handle_list = try_loprio ? &conf->loprio_list :
5516                                         &conf->handle_list;
5517         } else if (group != ANY_GROUP) {
5518                 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5519                                 &conf->worker_groups[group].handle_list;
5520                 wg = &conf->worker_groups[group];
5521         } else {
5522                 int i;
5523                 for (i = 0; i < conf->group_cnt; i++) {
5524                         handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5525                                 &conf->worker_groups[i].handle_list;
5526                         wg = &conf->worker_groups[i];
5527                         if (!list_empty(handle_list))
5528                                 break;
5529                 }
5530         }
5531
5532         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5533                   __func__,
5534                   list_empty(handle_list) ? "empty" : "busy",
5535                   list_empty(&conf->hold_list) ? "empty" : "busy",
5536                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
5537
5538         if (!list_empty(handle_list)) {
5539                 sh = list_entry(handle_list->next, typeof(*sh), lru);
5540
5541                 if (list_empty(&conf->hold_list))
5542                         conf->bypass_count = 0;
5543                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5544                         if (conf->hold_list.next == conf->last_hold)
5545                                 conf->bypass_count++;
5546                         else {
5547                                 conf->last_hold = conf->hold_list.next;
5548                                 conf->bypass_count -= conf->bypass_threshold;
5549                                 if (conf->bypass_count < 0)
5550                                         conf->bypass_count = 0;
5551                         }
5552                 }
5553         } else if (!list_empty(&conf->hold_list) &&
5554                    ((conf->bypass_threshold &&
5555                      conf->bypass_count > conf->bypass_threshold) ||
5556                     atomic_read(&conf->pending_full_writes) == 0)) {
5557
5558                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
5559                         if (conf->worker_cnt_per_group == 0 ||
5560                             group == ANY_GROUP ||
5561                             !cpu_online(tmp->cpu) ||
5562                             cpu_to_group(tmp->cpu) == group) {
5563                                 sh = tmp;
5564                                 break;
5565                         }
5566                 }
5567
5568                 if (sh) {
5569                         conf->bypass_count -= conf->bypass_threshold;
5570                         if (conf->bypass_count < 0)
5571                                 conf->bypass_count = 0;
5572                 }
5573                 wg = NULL;
5574         }
5575
5576         if (!sh) {
5577                 if (second_try)
5578                         return NULL;
5579                 second_try = true;
5580                 try_loprio = !try_loprio;
5581                 goto again;
5582         }
5583
5584         if (wg) {
5585                 wg->stripes_cnt--;
5586                 sh->group = NULL;
5587         }
5588         list_del_init(&sh->lru);
5589         BUG_ON(atomic_inc_return(&sh->count) != 1);
5590         return sh;
5591 }
5592
5593 struct raid5_plug_cb {
5594         struct blk_plug_cb      cb;
5595         struct list_head        list;
5596         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5597 };
5598
5599 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5600 {
5601         struct raid5_plug_cb *cb = container_of(
5602                 blk_cb, struct raid5_plug_cb, cb);
5603         struct stripe_head *sh;
5604         struct mddev *mddev = cb->cb.data;
5605         struct r5conf *conf = mddev->private;
5606         int cnt = 0;
5607         int hash;
5608
5609         if (cb->list.next && !list_empty(&cb->list)) {
5610                 spin_lock_irq(&conf->device_lock);
5611                 while (!list_empty(&cb->list)) {
5612                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
5613                         list_del_init(&sh->lru);
5614                         /*
5615                          * avoid race release_stripe_plug() sees
5616                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5617                          * is still in our list
5618                          */
5619                         smp_mb__before_atomic();
5620                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5621                         /*
5622                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5623                          * case, the count is always > 1 here
5624                          */
5625                         hash = sh->hash_lock_index;
5626                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5627                         cnt++;
5628                 }
5629                 spin_unlock_irq(&conf->device_lock);
5630         }
5631         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5632                                      NR_STRIPE_HASH_LOCKS);
5633         if (!mddev_is_dm(mddev))
5634                 trace_block_unplug(mddev->gendisk->queue, cnt, !from_schedule);
5635         kfree(cb);
5636 }
5637
5638 static void release_stripe_plug(struct mddev *mddev,
5639                                 struct stripe_head *sh)
5640 {
5641         struct blk_plug_cb *blk_cb = blk_check_plugged(
5642                 raid5_unplug, mddev,
5643                 sizeof(struct raid5_plug_cb));
5644         struct raid5_plug_cb *cb;
5645
5646         if (!blk_cb) {
5647                 raid5_release_stripe(sh);
5648                 return;
5649         }
5650
5651         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5652
5653         if (cb->list.next == NULL) {
5654                 int i;
5655                 INIT_LIST_HEAD(&cb->list);
5656                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5657                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5658         }
5659
5660         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5661                 list_add_tail(&sh->lru, &cb->list);
5662         else
5663                 raid5_release_stripe(sh);
5664 }
5665
5666 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5667 {
5668         struct r5conf *conf = mddev->private;
5669         sector_t logical_sector, last_sector;
5670         struct stripe_head *sh;
5671         int stripe_sectors;
5672
5673         /* We need to handle this when io_uring supports discard/trim */
5674         if (WARN_ON_ONCE(bi->bi_opf & REQ_NOWAIT))
5675                 return;
5676
5677         if (mddev->reshape_position != MaxSector)
5678                 /* Skip discard while reshape is happening */
5679                 return;
5680
5681         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
5682         last_sector = bio_end_sector(bi);
5683
5684         bi->bi_next = NULL;
5685
5686         stripe_sectors = conf->chunk_sectors *
5687                 (conf->raid_disks - conf->max_degraded);
5688         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5689                                                stripe_sectors);
5690         sector_div(last_sector, stripe_sectors);
5691
5692         logical_sector *= conf->chunk_sectors;
5693         last_sector *= conf->chunk_sectors;
5694
5695         for (; logical_sector < last_sector;
5696              logical_sector += RAID5_STRIPE_SECTORS(conf)) {
5697                 DEFINE_WAIT(w);
5698                 int d;
5699         again:
5700                 sh = raid5_get_active_stripe(conf, NULL, logical_sector, 0);
5701                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5702                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5703                         raid5_release_stripe(sh);
5704                         wait_on_bit(&sh->dev[sh->pd_idx].flags, R5_Overlap,
5705                                     TASK_UNINTERRUPTIBLE);
5706                         goto again;
5707                 }
5708                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5709                 spin_lock_irq(&sh->stripe_lock);
5710                 for (d = 0; d < conf->raid_disks; d++) {
5711                         if (d == sh->pd_idx || d == sh->qd_idx)
5712                                 continue;
5713                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5714                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5715                                 spin_unlock_irq(&sh->stripe_lock);
5716                                 raid5_release_stripe(sh);
5717                                 wait_on_bit(&sh->dev[d].flags, R5_Overlap,
5718                                             TASK_UNINTERRUPTIBLE);
5719                                 goto again;
5720                         }
5721                 }
5722                 set_bit(STRIPE_DISCARD, &sh->state);
5723                 sh->overwrite_disks = 0;
5724                 for (d = 0; d < conf->raid_disks; d++) {
5725                         if (d == sh->pd_idx || d == sh->qd_idx)
5726                                 continue;
5727                         sh->dev[d].towrite = bi;
5728                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5729                         bio_inc_remaining(bi);
5730                         md_write_inc(mddev, bi);
5731                         sh->overwrite_disks++;
5732                 }
5733                 spin_unlock_irq(&sh->stripe_lock);
5734                 if (conf->mddev->bitmap) {
5735                         sh->bm_seq = conf->seq_flush + 1;
5736                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5737                 }
5738
5739                 set_bit(STRIPE_HANDLE, &sh->state);
5740                 clear_bit(STRIPE_DELAYED, &sh->state);
5741                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5742                         atomic_inc(&conf->preread_active_stripes);
5743                 release_stripe_plug(mddev, sh);
5744         }
5745
5746         bio_endio(bi);
5747 }
5748
5749 static bool ahead_of_reshape(struct mddev *mddev, sector_t sector,
5750                              sector_t reshape_sector)
5751 {
5752         return mddev->reshape_backwards ? sector < reshape_sector :
5753                                           sector >= reshape_sector;
5754 }
5755
5756 static bool range_ahead_of_reshape(struct mddev *mddev, sector_t min,
5757                                    sector_t max, sector_t reshape_sector)
5758 {
5759         return mddev->reshape_backwards ? max < reshape_sector :
5760                                           min >= reshape_sector;
5761 }
5762
5763 static bool stripe_ahead_of_reshape(struct mddev *mddev, struct r5conf *conf,
5764                                     struct stripe_head *sh)
5765 {
5766         sector_t max_sector = 0, min_sector = MaxSector;
5767         bool ret = false;
5768         int dd_idx;
5769
5770         for (dd_idx = 0; dd_idx < sh->disks; dd_idx++) {
5771                 if (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
5772                         continue;
5773
5774                 min_sector = min(min_sector, sh->dev[dd_idx].sector);
5775                 max_sector = max(max_sector, sh->dev[dd_idx].sector);
5776         }
5777
5778         spin_lock_irq(&conf->device_lock);
5779
5780         if (!range_ahead_of_reshape(mddev, min_sector, max_sector,
5781                                      conf->reshape_progress))
5782                 /* mismatch, need to try again */
5783                 ret = true;
5784
5785         spin_unlock_irq(&conf->device_lock);
5786
5787         return ret;
5788 }
5789
5790 static int add_all_stripe_bios(struct r5conf *conf,
5791                 struct stripe_request_ctx *ctx, struct stripe_head *sh,
5792                 struct bio *bi, int forwrite, int previous)
5793 {
5794         int dd_idx;
5795
5796         spin_lock_irq(&sh->stripe_lock);
5797
5798         for (dd_idx = 0; dd_idx < sh->disks; dd_idx++) {
5799                 struct r5dev *dev = &sh->dev[dd_idx];
5800
5801                 if (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
5802                         continue;
5803
5804                 if (dev->sector < ctx->first_sector ||
5805                     dev->sector >= ctx->last_sector)
5806                         continue;
5807
5808                 if (stripe_bio_overlaps(sh, bi, dd_idx, forwrite)) {
5809                         set_bit(R5_Overlap, &dev->flags);
5810                         spin_unlock_irq(&sh->stripe_lock);
5811                         raid5_release_stripe(sh);
5812                         /* release batch_last before wait to avoid risk of deadlock */
5813                         if (ctx->batch_last) {
5814                                 raid5_release_stripe(ctx->batch_last);
5815                                 ctx->batch_last = NULL;
5816                         }
5817                         md_wakeup_thread(conf->mddev->thread);
5818                         wait_on_bit(&dev->flags, R5_Overlap, TASK_UNINTERRUPTIBLE);
5819                         return 0;
5820                 }
5821         }
5822
5823         for (dd_idx = 0; dd_idx < sh->disks; dd_idx++) {
5824                 struct r5dev *dev = &sh->dev[dd_idx];
5825
5826                 if (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
5827                         continue;
5828
5829                 if (dev->sector < ctx->first_sector ||
5830                     dev->sector >= ctx->last_sector)
5831                         continue;
5832
5833                 __add_stripe_bio(sh, bi, dd_idx, forwrite, previous);
5834                 clear_bit((dev->sector - ctx->first_sector) >>
5835                           RAID5_STRIPE_SHIFT(conf), ctx->sectors_to_do);
5836         }
5837
5838         spin_unlock_irq(&sh->stripe_lock);
5839         return 1;
5840 }
5841
5842 enum reshape_loc {
5843         LOC_NO_RESHAPE,
5844         LOC_AHEAD_OF_RESHAPE,
5845         LOC_INSIDE_RESHAPE,
5846         LOC_BEHIND_RESHAPE,
5847 };
5848
5849 static enum reshape_loc get_reshape_loc(struct mddev *mddev,
5850                 struct r5conf *conf, sector_t logical_sector)
5851 {
5852         sector_t reshape_progress, reshape_safe;
5853
5854         if (likely(conf->reshape_progress == MaxSector))
5855                 return LOC_NO_RESHAPE;
5856         /*
5857          * Spinlock is needed as reshape_progress may be
5858          * 64bit on a 32bit platform, and so it might be
5859          * possible to see a half-updated value
5860          * Of course reshape_progress could change after
5861          * the lock is dropped, so once we get a reference
5862          * to the stripe that we think it is, we will have
5863          * to check again.
5864          */
5865         spin_lock_irq(&conf->device_lock);
5866         reshape_progress = conf->reshape_progress;
5867         reshape_safe = conf->reshape_safe;
5868         spin_unlock_irq(&conf->device_lock);
5869         if (reshape_progress == MaxSector)
5870                 return LOC_NO_RESHAPE;
5871         if (ahead_of_reshape(mddev, logical_sector, reshape_progress))
5872                 return LOC_AHEAD_OF_RESHAPE;
5873         if (ahead_of_reshape(mddev, logical_sector, reshape_safe))
5874                 return LOC_INSIDE_RESHAPE;
5875         return LOC_BEHIND_RESHAPE;
5876 }
5877
5878 static void raid5_bitmap_sector(struct mddev *mddev, sector_t *offset,
5879                                 unsigned long *sectors)
5880 {
5881         struct r5conf *conf = mddev->private;
5882         sector_t start = *offset;
5883         sector_t end = start + *sectors;
5884         sector_t prev_start = start;
5885         sector_t prev_end = end;
5886         int sectors_per_chunk;
5887         enum reshape_loc loc;
5888         int dd_idx;
5889
5890         sectors_per_chunk = conf->chunk_sectors *
5891                 (conf->raid_disks - conf->max_degraded);
5892         start = round_down(start, sectors_per_chunk);
5893         end = round_up(end, sectors_per_chunk);
5894
5895         start = raid5_compute_sector(conf, start, 0, &dd_idx, NULL);
5896         end = raid5_compute_sector(conf, end, 0, &dd_idx, NULL);
5897
5898         /*
5899          * For LOC_INSIDE_RESHAPE, this IO will wait for reshape to make
5900          * progress, hence it's the same as LOC_BEHIND_RESHAPE.
5901          */
5902         loc = get_reshape_loc(mddev, conf, prev_start);
5903         if (likely(loc != LOC_AHEAD_OF_RESHAPE)) {
5904                 *offset = start;
5905                 *sectors = end - start;
5906                 return;
5907         }
5908
5909         sectors_per_chunk = conf->prev_chunk_sectors *
5910                 (conf->previous_raid_disks - conf->max_degraded);
5911         prev_start = round_down(prev_start, sectors_per_chunk);
5912         prev_end = round_down(prev_end, sectors_per_chunk);
5913
5914         prev_start = raid5_compute_sector(conf, prev_start, 1, &dd_idx, NULL);
5915         prev_end = raid5_compute_sector(conf, prev_end, 1, &dd_idx, NULL);
5916
5917         /*
5918          * for LOC_AHEAD_OF_RESHAPE, reshape can make progress before this IO
5919          * is handled in make_stripe_request(), we can't know this here hence
5920          * we set bits for both.
5921          */
5922         *offset = min(start, prev_start);
5923         *sectors = max(end, prev_end) - *offset;
5924 }
5925
5926 static enum stripe_result make_stripe_request(struct mddev *mddev,
5927                 struct r5conf *conf, struct stripe_request_ctx *ctx,
5928                 sector_t logical_sector, struct bio *bi)
5929 {
5930         const int rw = bio_data_dir(bi);
5931         enum stripe_result ret;
5932         struct stripe_head *sh;
5933         enum reshape_loc loc;
5934         sector_t new_sector;
5935         int previous = 0, flags = 0;
5936         int seq, dd_idx;
5937
5938         seq = read_seqcount_begin(&conf->gen_lock);
5939         loc = get_reshape_loc(mddev, conf, logical_sector);
5940         if (loc == LOC_INSIDE_RESHAPE) {
5941                 ret = STRIPE_SCHEDULE_AND_RETRY;
5942                 goto out;
5943         }
5944         if (loc == LOC_AHEAD_OF_RESHAPE)
5945                 previous = 1;
5946
5947         new_sector = raid5_compute_sector(conf, logical_sector, previous,
5948                                           &dd_idx, NULL);
5949         pr_debug("raid456: %s, sector %llu logical %llu\n", __func__,
5950                  new_sector, logical_sector);
5951
5952         if (previous)
5953                 flags |= R5_GAS_PREVIOUS;
5954         if (bi->bi_opf & REQ_RAHEAD)
5955                 flags |= R5_GAS_NOBLOCK;
5956         sh = raid5_get_active_stripe(conf, ctx, new_sector, flags);
5957         if (unlikely(!sh)) {
5958                 /* cannot get stripe, just give-up */
5959                 bi->bi_status = BLK_STS_IOERR;
5960                 return STRIPE_FAIL;
5961         }
5962
5963         if (unlikely(previous) &&
5964             stripe_ahead_of_reshape(mddev, conf, sh)) {
5965                 /*
5966                  * Expansion moved on while waiting for a stripe.
5967                  * Expansion could still move past after this
5968                  * test, but as we are holding a reference to
5969                  * 'sh', we know that if that happens,
5970                  *  STRIPE_EXPANDING will get set and the expansion
5971                  * won't proceed until we finish with the stripe.
5972                  */
5973                 ret = STRIPE_SCHEDULE_AND_RETRY;
5974                 goto out_release;
5975         }
5976
5977         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5978                 /* Might have got the wrong stripe_head by accident */
5979                 ret = STRIPE_RETRY;
5980                 goto out_release;
5981         }
5982
5983         if (test_bit(STRIPE_EXPANDING, &sh->state)) {
5984                 md_wakeup_thread(mddev->thread);
5985                 ret = STRIPE_SCHEDULE_AND_RETRY;
5986                 goto out_release;
5987         }
5988
5989         if (!add_all_stripe_bios(conf, ctx, sh, bi, rw, previous)) {
5990                 ret = STRIPE_RETRY;
5991                 goto out;
5992         }
5993
5994         if (stripe_can_batch(sh)) {
5995                 stripe_add_to_batch_list(conf, sh, ctx->batch_last);
5996                 if (ctx->batch_last)
5997                         raid5_release_stripe(ctx->batch_last);
5998                 atomic_inc(&sh->count);
5999                 ctx->batch_last = sh;
6000         }
6001
6002         if (ctx->do_flush) {
6003                 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
6004                 /* we only need flush for one stripe */
6005                 ctx->do_flush = false;
6006         }
6007
6008         set_bit(STRIPE_HANDLE, &sh->state);
6009         clear_bit(STRIPE_DELAYED, &sh->state);
6010         if ((!sh->batch_head || sh == sh->batch_head) &&
6011             (bi->bi_opf & REQ_SYNC) &&
6012             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
6013                 atomic_inc(&conf->preread_active_stripes);
6014
6015         release_stripe_plug(mddev, sh);
6016         return STRIPE_SUCCESS;
6017
6018 out_release:
6019         raid5_release_stripe(sh);
6020 out:
6021         if (ret == STRIPE_SCHEDULE_AND_RETRY && reshape_interrupted(mddev)) {
6022                 bi->bi_status = BLK_STS_RESOURCE;
6023                 ret = STRIPE_WAIT_RESHAPE;
6024                 pr_err_ratelimited("dm-raid456: io across reshape position while reshape can't make progress");
6025         }
6026         return ret;
6027 }
6028
6029 /*
6030  * If the bio covers multiple data disks, find sector within the bio that has
6031  * the lowest chunk offset in the first chunk.
6032  */
6033 static sector_t raid5_bio_lowest_chunk_sector(struct r5conf *conf,
6034                                               struct bio *bi)
6035 {
6036         int sectors_per_chunk = conf->chunk_sectors;
6037         int raid_disks = conf->raid_disks;
6038         int dd_idx;
6039         struct stripe_head sh;
6040         unsigned int chunk_offset;
6041         sector_t r_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6042         sector_t sector;
6043
6044         /* We pass in fake stripe_head to get back parity disk numbers */
6045         sector = raid5_compute_sector(conf, r_sector, 0, &dd_idx, &sh);
6046         chunk_offset = sector_div(sector, sectors_per_chunk);
6047         if (sectors_per_chunk - chunk_offset >= bio_sectors(bi))
6048                 return r_sector;
6049         /*
6050          * Bio crosses to the next data disk. Check whether it's in the same
6051          * chunk.
6052          */
6053         dd_idx++;
6054         while (dd_idx == sh.pd_idx || dd_idx == sh.qd_idx)
6055                 dd_idx++;
6056         if (dd_idx >= raid_disks)
6057                 return r_sector;
6058         return r_sector + sectors_per_chunk - chunk_offset;
6059 }
6060
6061 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
6062 {
6063         DEFINE_WAIT_FUNC(wait, woken_wake_function);
6064         bool on_wq;
6065         struct r5conf *conf = mddev->private;
6066         sector_t logical_sector;
6067         struct stripe_request_ctx ctx = {};
6068         const int rw = bio_data_dir(bi);
6069         enum stripe_result res;
6070         int s, stripe_cnt;
6071
6072         if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
6073                 int ret = log_handle_flush_request(conf, bi);
6074
6075                 if (ret == 0)
6076                         return true;
6077                 if (ret == -ENODEV) {
6078                         if (md_flush_request(mddev, bi))
6079                                 return true;
6080                 }
6081                 /* ret == -EAGAIN, fallback */
6082                 /*
6083                  * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
6084                  * we need to flush journal device
6085                  */
6086                 ctx.do_flush = bi->bi_opf & REQ_PREFLUSH;
6087         }
6088
6089         md_write_start(mddev, bi);
6090         /*
6091          * If array is degraded, better not do chunk aligned read because
6092          * later we might have to read it again in order to reconstruct
6093          * data on failed drives.
6094          */
6095         if (rw == READ && mddev->degraded == 0 &&
6096             mddev->reshape_position == MaxSector) {
6097                 bi = chunk_aligned_read(mddev, bi);
6098                 if (!bi)
6099                         return true;
6100         }
6101
6102         if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
6103                 make_discard_request(mddev, bi);
6104                 md_write_end(mddev);
6105                 return true;
6106         }
6107
6108         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6109         ctx.first_sector = logical_sector;
6110         ctx.last_sector = bio_end_sector(bi);
6111         bi->bi_next = NULL;
6112
6113         stripe_cnt = DIV_ROUND_UP_SECTOR_T(ctx.last_sector - logical_sector,
6114                                            RAID5_STRIPE_SECTORS(conf));
6115         bitmap_set(ctx.sectors_to_do, 0, stripe_cnt);
6116
6117         pr_debug("raid456: %s, logical %llu to %llu\n", __func__,
6118                  bi->bi_iter.bi_sector, ctx.last_sector);
6119
6120         /* Bail out if conflicts with reshape and REQ_NOWAIT is set */
6121         if ((bi->bi_opf & REQ_NOWAIT) &&
6122             get_reshape_loc(mddev, conf, logical_sector) == LOC_INSIDE_RESHAPE) {
6123                 bio_wouldblock_error(bi);
6124                 if (rw == WRITE)
6125                         md_write_end(mddev);
6126                 return true;
6127         }
6128         md_account_bio(mddev, &bi);
6129
6130         /*
6131          * Lets start with the stripe with the lowest chunk offset in the first
6132          * chunk. That has the best chances of creating IOs adjacent to
6133          * previous IOs in case of sequential IO and thus creates the most
6134          * sequential IO pattern. We don't bother with the optimization when
6135          * reshaping as the performance benefit is not worth the complexity.
6136          */
6137         if (likely(conf->reshape_progress == MaxSector)) {
6138                 logical_sector = raid5_bio_lowest_chunk_sector(conf, bi);
6139                 on_wq = false;
6140         } else {
6141                 add_wait_queue(&conf->wait_for_reshape, &wait);
6142                 on_wq = true;
6143         }
6144         s = (logical_sector - ctx.first_sector) >> RAID5_STRIPE_SHIFT(conf);
6145
6146         while (1) {
6147                 res = make_stripe_request(mddev, conf, &ctx, logical_sector,
6148                                           bi);
6149                 if (res == STRIPE_FAIL || res == STRIPE_WAIT_RESHAPE)
6150                         break;
6151
6152                 if (res == STRIPE_RETRY)
6153                         continue;
6154
6155                 if (res == STRIPE_SCHEDULE_AND_RETRY) {
6156                         WARN_ON_ONCE(!on_wq);
6157                         /*
6158                          * Must release the reference to batch_last before
6159                          * scheduling and waiting for work to be done,
6160                          * otherwise the batch_last stripe head could prevent
6161                          * raid5_activate_delayed() from making progress
6162                          * and thus deadlocking.
6163                          */
6164                         if (ctx.batch_last) {
6165                                 raid5_release_stripe(ctx.batch_last);
6166                                 ctx.batch_last = NULL;
6167                         }
6168
6169                         wait_woken(&wait, TASK_UNINTERRUPTIBLE,
6170                                    MAX_SCHEDULE_TIMEOUT);
6171                         continue;
6172                 }
6173
6174                 s = find_next_bit_wrap(ctx.sectors_to_do, stripe_cnt, s);
6175                 if (s == stripe_cnt)
6176                         break;
6177
6178                 logical_sector = ctx.first_sector +
6179                         (s << RAID5_STRIPE_SHIFT(conf));
6180         }
6181         if (unlikely(on_wq))
6182                 remove_wait_queue(&conf->wait_for_reshape, &wait);
6183
6184         if (ctx.batch_last)
6185                 raid5_release_stripe(ctx.batch_last);
6186
6187         if (rw == WRITE)
6188                 md_write_end(mddev);
6189         if (res == STRIPE_WAIT_RESHAPE) {
6190                 md_free_cloned_bio(bi);
6191                 return false;
6192         }
6193
6194         bio_endio(bi);
6195         return true;
6196 }
6197
6198 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
6199
6200 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
6201 {
6202         /* reshaping is quite different to recovery/resync so it is
6203          * handled quite separately ... here.
6204          *
6205          * On each call to sync_request, we gather one chunk worth of
6206          * destination stripes and flag them as expanding.
6207          * Then we find all the source stripes and request reads.
6208          * As the reads complete, handle_stripe will copy the data
6209          * into the destination stripe and release that stripe.
6210          */
6211         struct r5conf *conf = mddev->private;
6212         struct stripe_head *sh;
6213         struct md_rdev *rdev;
6214         sector_t first_sector, last_sector;
6215         int raid_disks = conf->previous_raid_disks;
6216         int data_disks = raid_disks - conf->max_degraded;
6217         int new_data_disks = conf->raid_disks - conf->max_degraded;
6218         int i;
6219         int dd_idx;
6220         sector_t writepos, readpos, safepos;
6221         sector_t stripe_addr;
6222         int reshape_sectors;
6223         struct list_head stripes;
6224         sector_t retn;
6225
6226         if (sector_nr == 0) {
6227                 /* If restarting in the middle, skip the initial sectors */
6228                 if (mddev->reshape_backwards &&
6229                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
6230                         sector_nr = raid5_size(mddev, 0, 0)
6231                                 - conf->reshape_progress;
6232                 } else if (mddev->reshape_backwards &&
6233                            conf->reshape_progress == MaxSector) {
6234                         /* shouldn't happen, but just in case, finish up.*/
6235                         sector_nr = MaxSector;
6236                 } else if (!mddev->reshape_backwards &&
6237                            conf->reshape_progress > 0)
6238                         sector_nr = conf->reshape_progress;
6239                 sector_div(sector_nr, new_data_disks);
6240                 if (sector_nr) {
6241                         mddev->curr_resync_completed = sector_nr;
6242                         sysfs_notify_dirent_safe(mddev->sysfs_completed);
6243                         *skipped = 1;
6244                         retn = sector_nr;
6245                         goto finish;
6246                 }
6247         }
6248
6249         /* We need to process a full chunk at a time.
6250          * If old and new chunk sizes differ, we need to process the
6251          * largest of these
6252          */
6253
6254         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
6255
6256         /* We update the metadata at least every 10 seconds, or when
6257          * the data about to be copied would over-write the source of
6258          * the data at the front of the range.  i.e. one new_stripe
6259          * along from reshape_progress new_maps to after where
6260          * reshape_safe old_maps to
6261          */
6262         writepos = conf->reshape_progress;
6263         sector_div(writepos, new_data_disks);
6264         readpos = conf->reshape_progress;
6265         sector_div(readpos, data_disks);
6266         safepos = conf->reshape_safe;
6267         sector_div(safepos, data_disks);
6268         if (mddev->reshape_backwards) {
6269                 if (WARN_ON(writepos < reshape_sectors))
6270                         return MaxSector;
6271
6272                 writepos -= reshape_sectors;
6273                 readpos += reshape_sectors;
6274                 safepos += reshape_sectors;
6275         } else {
6276                 writepos += reshape_sectors;
6277                 /* readpos and safepos are worst-case calculations.
6278                  * A negative number is overly pessimistic, and causes
6279                  * obvious problems for unsigned storage.  So clip to 0.
6280                  */
6281                 readpos -= min_t(sector_t, reshape_sectors, readpos);
6282                 safepos -= min_t(sector_t, reshape_sectors, safepos);
6283         }
6284
6285         /* Having calculated the 'writepos' possibly use it
6286          * to set 'stripe_addr' which is where we will write to.
6287          */
6288         if (mddev->reshape_backwards) {
6289                 if (WARN_ON(conf->reshape_progress == 0))
6290                         return MaxSector;
6291
6292                 stripe_addr = writepos;
6293                 if (WARN_ON((mddev->dev_sectors &
6294                     ~((sector_t)reshape_sectors - 1)) -
6295                     reshape_sectors - stripe_addr != sector_nr))
6296                         return MaxSector;
6297         } else {
6298                 if (WARN_ON(writepos != sector_nr + reshape_sectors))
6299                         return MaxSector;
6300
6301                 stripe_addr = sector_nr;
6302         }
6303
6304         /* 'writepos' is the most advanced device address we might write.
6305          * 'readpos' is the least advanced device address we might read.
6306          * 'safepos' is the least address recorded in the metadata as having
6307          *     been reshaped.
6308          * If there is a min_offset_diff, these are adjusted either by
6309          * increasing the safepos/readpos if diff is negative, or
6310          * increasing writepos if diff is positive.
6311          * If 'readpos' is then behind 'writepos', there is no way that we can
6312          * ensure safety in the face of a crash - that must be done by userspace
6313          * making a backup of the data.  So in that case there is no particular
6314          * rush to update metadata.
6315          * Otherwise if 'safepos' is behind 'writepos', then we really need to
6316          * update the metadata to advance 'safepos' to match 'readpos' so that
6317          * we can be safe in the event of a crash.
6318          * So we insist on updating metadata if safepos is behind writepos and
6319          * readpos is beyond writepos.
6320          * In any case, update the metadata every 10 seconds.
6321          * Maybe that number should be configurable, but I'm not sure it is
6322          * worth it.... maybe it could be a multiple of safemode_delay???
6323          */
6324         if (conf->min_offset_diff < 0) {
6325                 safepos += -conf->min_offset_diff;
6326                 readpos += -conf->min_offset_diff;
6327         } else
6328                 writepos += conf->min_offset_diff;
6329
6330         if ((mddev->reshape_backwards
6331              ? (safepos > writepos && readpos < writepos)
6332              : (safepos < writepos && readpos > writepos)) ||
6333             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
6334                 /* Cannot proceed until we've updated the superblock... */
6335                 wait_event(conf->wait_for_reshape,
6336                            atomic_read(&conf->reshape_stripes)==0
6337                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6338                 if (atomic_read(&conf->reshape_stripes) != 0)
6339                         return 0;
6340                 mddev->reshape_position = conf->reshape_progress;
6341                 mddev->curr_resync_completed = sector_nr;
6342                 if (!mddev->reshape_backwards)
6343                         /* Can update recovery_offset */
6344                         rdev_for_each(rdev, mddev)
6345                                 if (rdev->raid_disk >= 0 &&
6346                                     !test_bit(Journal, &rdev->flags) &&
6347                                     !test_bit(In_sync, &rdev->flags) &&
6348                                     rdev->recovery_offset < sector_nr)
6349                                         rdev->recovery_offset = sector_nr;
6350
6351                 conf->reshape_checkpoint = jiffies;
6352                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6353                 md_wakeup_thread(mddev->thread);
6354                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
6355                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6356                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6357                         return 0;
6358                 spin_lock_irq(&conf->device_lock);
6359                 conf->reshape_safe = mddev->reshape_position;
6360                 spin_unlock_irq(&conf->device_lock);
6361                 wake_up(&conf->wait_for_reshape);
6362                 sysfs_notify_dirent_safe(mddev->sysfs_completed);
6363         }
6364
6365         INIT_LIST_HEAD(&stripes);
6366         for (i = 0; i < reshape_sectors; i += RAID5_STRIPE_SECTORS(conf)) {
6367                 int j;
6368                 int skipped_disk = 0;
6369                 sh = raid5_get_active_stripe(conf, NULL, stripe_addr+i,
6370                                              R5_GAS_NOQUIESCE);
6371                 set_bit(STRIPE_EXPANDING, &sh->state);
6372                 atomic_inc(&conf->reshape_stripes);
6373                 /* If any of this stripe is beyond the end of the old
6374                  * array, then we need to zero those blocks
6375                  */
6376                 for (j=sh->disks; j--;) {
6377                         sector_t s;
6378                         if (j == sh->pd_idx)
6379                                 continue;
6380                         if (conf->level == 6 &&
6381                             j == sh->qd_idx)
6382                                 continue;
6383                         s = raid5_compute_blocknr(sh, j, 0);
6384                         if (s < raid5_size(mddev, 0, 0)) {
6385                                 skipped_disk = 1;
6386                                 continue;
6387                         }
6388                         memset(page_address(sh->dev[j].page), 0, RAID5_STRIPE_SIZE(conf));
6389                         set_bit(R5_Expanded, &sh->dev[j].flags);
6390                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
6391                 }
6392                 if (!skipped_disk) {
6393                         set_bit(STRIPE_EXPAND_READY, &sh->state);
6394                         set_bit(STRIPE_HANDLE, &sh->state);
6395                 }
6396                 list_add(&sh->lru, &stripes);
6397         }
6398         spin_lock_irq(&conf->device_lock);
6399         if (mddev->reshape_backwards)
6400                 conf->reshape_progress -= reshape_sectors * new_data_disks;
6401         else
6402                 conf->reshape_progress += reshape_sectors * new_data_disks;
6403         spin_unlock_irq(&conf->device_lock);
6404         /* Ok, those stripe are ready. We can start scheduling
6405          * reads on the source stripes.
6406          * The source stripes are determined by mapping the first and last
6407          * block on the destination stripes.
6408          */
6409         first_sector =
6410                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
6411                                      1, &dd_idx, NULL);
6412         last_sector =
6413                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
6414                                             * new_data_disks - 1),
6415                                      1, &dd_idx, NULL);
6416         if (last_sector >= mddev->dev_sectors)
6417                 last_sector = mddev->dev_sectors - 1;
6418         while (first_sector <= last_sector) {
6419                 sh = raid5_get_active_stripe(conf, NULL, first_sector,
6420                                 R5_GAS_PREVIOUS | R5_GAS_NOQUIESCE);
6421                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
6422                 set_bit(STRIPE_HANDLE, &sh->state);
6423                 raid5_release_stripe(sh);
6424                 first_sector += RAID5_STRIPE_SECTORS(conf);
6425         }
6426         /* Now that the sources are clearly marked, we can release
6427          * the destination stripes
6428          */
6429         while (!list_empty(&stripes)) {
6430                 sh = list_entry(stripes.next, struct stripe_head, lru);
6431                 list_del_init(&sh->lru);
6432                 raid5_release_stripe(sh);
6433         }
6434         /* If this takes us to the resync_max point where we have to pause,
6435          * then we need to write out the superblock.
6436          */
6437         sector_nr += reshape_sectors;
6438         retn = reshape_sectors;
6439 finish:
6440         if (mddev->curr_resync_completed > mddev->resync_max ||
6441             (sector_nr - mddev->curr_resync_completed) * 2
6442             >= mddev->resync_max - mddev->curr_resync_completed) {
6443                 /* Cannot proceed until we've updated the superblock... */
6444                 wait_event(conf->wait_for_reshape,
6445                            atomic_read(&conf->reshape_stripes) == 0
6446                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6447                 if (atomic_read(&conf->reshape_stripes) != 0)
6448                         goto ret;
6449                 mddev->reshape_position = conf->reshape_progress;
6450                 mddev->curr_resync_completed = sector_nr;
6451                 if (!mddev->reshape_backwards)
6452                         /* Can update recovery_offset */
6453                         rdev_for_each(rdev, mddev)
6454                                 if (rdev->raid_disk >= 0 &&
6455                                     !test_bit(Journal, &rdev->flags) &&
6456                                     !test_bit(In_sync, &rdev->flags) &&
6457                                     rdev->recovery_offset < sector_nr)
6458                                         rdev->recovery_offset = sector_nr;
6459                 conf->reshape_checkpoint = jiffies;
6460                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6461                 md_wakeup_thread(mddev->thread);
6462                 wait_event(mddev->sb_wait,
6463                            !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6464                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6465                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6466                         goto ret;
6467                 spin_lock_irq(&conf->device_lock);
6468                 conf->reshape_safe = mddev->reshape_position;
6469                 spin_unlock_irq(&conf->device_lock);
6470                 wake_up(&conf->wait_for_reshape);
6471                 sysfs_notify_dirent_safe(mddev->sysfs_completed);
6472         }
6473 ret:
6474         return retn;
6475 }
6476
6477 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6478                                           sector_t max_sector, int *skipped)
6479 {
6480         struct r5conf *conf = mddev->private;
6481         struct stripe_head *sh;
6482         sector_t sync_blocks;
6483         bool still_degraded = false;
6484         int i;
6485
6486         if (sector_nr >= max_sector) {
6487                 /* just being told to finish up .. nothing much to do */
6488
6489                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6490                         end_reshape(conf);
6491                         return 0;
6492                 }
6493
6494                 if (mddev->curr_resync < max_sector) /* aborted */
6495                         mddev->bitmap_ops->end_sync(mddev, mddev->curr_resync,
6496                                                     &sync_blocks);
6497                 else /* completed sync */
6498                         conf->fullsync = 0;
6499                 mddev->bitmap_ops->close_sync(mddev);
6500
6501                 return 0;
6502         }
6503
6504         /* Allow raid5_quiesce to complete */
6505         wait_event(conf->wait_for_reshape, conf->quiesce != 2);
6506
6507         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6508                 return reshape_request(mddev, sector_nr, skipped);
6509
6510         /* No need to check resync_max as we never do more than one
6511          * stripe, and as resync_max will always be on a chunk boundary,
6512          * if the check in md_do_sync didn't fire, there is no chance
6513          * of overstepping resync_max here
6514          */
6515
6516         /* if there is too many failed drives and we are trying
6517          * to resync, then assert that we are finished, because there is
6518          * nothing we can do.
6519          */
6520         if (mddev->degraded >= conf->max_degraded &&
6521             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6522                 sector_t rv = mddev->dev_sectors - sector_nr;
6523                 *skipped = 1;
6524                 return rv;
6525         }
6526         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6527             !conf->fullsync &&
6528             !mddev->bitmap_ops->start_sync(mddev, sector_nr, &sync_blocks,
6529                                            true) &&
6530             sync_blocks >= RAID5_STRIPE_SECTORS(conf)) {
6531                 /* we can skip this block, and probably more */
6532                 do_div(sync_blocks, RAID5_STRIPE_SECTORS(conf));
6533                 *skipped = 1;
6534                 /* keep things rounded to whole stripes */
6535                 return sync_blocks * RAID5_STRIPE_SECTORS(conf);
6536         }
6537
6538         mddev->bitmap_ops->cond_end_sync(mddev, sector_nr, false);
6539
6540         sh = raid5_get_active_stripe(conf, NULL, sector_nr,
6541                                      R5_GAS_NOBLOCK);
6542         if (sh == NULL) {
6543                 sh = raid5_get_active_stripe(conf, NULL, sector_nr, 0);
6544                 /* make sure we don't swamp the stripe cache if someone else
6545                  * is trying to get access
6546                  */
6547                 schedule_timeout_uninterruptible(1);
6548         }
6549         /* Need to check if array will still be degraded after recovery/resync
6550          * Note in case of > 1 drive failures it's possible we're rebuilding
6551          * one drive while leaving another faulty drive in array.
6552          */
6553         for (i = 0; i < conf->raid_disks; i++) {
6554                 struct md_rdev *rdev = conf->disks[i].rdev;
6555
6556                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6557                         still_degraded = true;
6558         }
6559
6560         mddev->bitmap_ops->start_sync(mddev, sector_nr, &sync_blocks,
6561                                       still_degraded);
6562
6563         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6564         set_bit(STRIPE_HANDLE, &sh->state);
6565
6566         raid5_release_stripe(sh);
6567
6568         return RAID5_STRIPE_SECTORS(conf);
6569 }
6570
6571 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6572                                unsigned int offset)
6573 {
6574         /* We may not be able to submit a whole bio at once as there
6575          * may not be enough stripe_heads available.
6576          * We cannot pre-allocate enough stripe_heads as we may need
6577          * more than exist in the cache (if we allow ever large chunks).
6578          * So we do one stripe head at a time and record in
6579          * ->bi_hw_segments how many have been done.
6580          *
6581          * We *know* that this entire raid_bio is in one chunk, so
6582          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6583          */
6584         struct stripe_head *sh;
6585         int dd_idx;
6586         sector_t sector, logical_sector, last_sector;
6587         int scnt = 0;
6588         int handled = 0;
6589
6590         logical_sector = raid_bio->bi_iter.bi_sector &
6591                 ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6592         sector = raid5_compute_sector(conf, logical_sector,
6593                                       0, &dd_idx, NULL);
6594         last_sector = bio_end_sector(raid_bio);
6595
6596         for (; logical_sector < last_sector;
6597              logical_sector += RAID5_STRIPE_SECTORS(conf),
6598                      sector += RAID5_STRIPE_SECTORS(conf),
6599                      scnt++) {
6600
6601                 if (scnt < offset)
6602                         /* already done this stripe */
6603                         continue;
6604
6605                 sh = raid5_get_active_stripe(conf, NULL, sector,
6606                                 R5_GAS_NOBLOCK | R5_GAS_NOQUIESCE);
6607                 if (!sh) {
6608                         /* failed to get a stripe - must wait */
6609                         conf->retry_read_aligned = raid_bio;
6610                         conf->retry_read_offset = scnt;
6611                         return handled;
6612                 }
6613
6614                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6615                         raid5_release_stripe(sh);
6616                         conf->retry_read_aligned = raid_bio;
6617                         conf->retry_read_offset = scnt;
6618                         return handled;
6619                 }
6620
6621                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6622                 handle_stripe(sh);
6623                 raid5_release_stripe(sh);
6624                 handled++;
6625         }
6626
6627         bio_endio(raid_bio);
6628
6629         if (atomic_dec_and_test(&conf->active_aligned_reads))
6630                 wake_up(&conf->wait_for_quiescent);
6631         return handled;
6632 }
6633
6634 static int handle_active_stripes(struct r5conf *conf, int group,
6635                                  struct r5worker *worker,
6636                                  struct list_head *temp_inactive_list)
6637                 __must_hold(&conf->device_lock)
6638 {
6639         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6640         int i, batch_size = 0, hash;
6641         bool release_inactive = false;
6642
6643         while (batch_size < MAX_STRIPE_BATCH &&
6644                         (sh = __get_priority_stripe(conf, group)) != NULL)
6645                 batch[batch_size++] = sh;
6646
6647         if (batch_size == 0) {
6648                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6649                         if (!list_empty(temp_inactive_list + i))
6650                                 break;
6651                 if (i == NR_STRIPE_HASH_LOCKS) {
6652                         spin_unlock_irq(&conf->device_lock);
6653                         log_flush_stripe_to_raid(conf);
6654                         spin_lock_irq(&conf->device_lock);
6655                         return batch_size;
6656                 }
6657                 release_inactive = true;
6658         }
6659         spin_unlock_irq(&conf->device_lock);
6660
6661         release_inactive_stripe_list(conf, temp_inactive_list,
6662                                      NR_STRIPE_HASH_LOCKS);
6663
6664         r5l_flush_stripe_to_raid(conf->log);
6665         if (release_inactive) {
6666                 spin_lock_irq(&conf->device_lock);
6667                 return 0;
6668         }
6669
6670         for (i = 0; i < batch_size; i++)
6671                 handle_stripe(batch[i]);
6672         log_write_stripe_run(conf);
6673
6674         cond_resched();
6675
6676         spin_lock_irq(&conf->device_lock);
6677         for (i = 0; i < batch_size; i++) {
6678                 hash = batch[i]->hash_lock_index;
6679                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6680         }
6681         return batch_size;
6682 }
6683
6684 static void raid5_do_work(struct work_struct *work)
6685 {
6686         struct r5worker *worker = container_of(work, struct r5worker, work);
6687         struct r5worker_group *group = worker->group;
6688         struct r5conf *conf = group->conf;
6689         struct mddev *mddev = conf->mddev;
6690         int group_id = group - conf->worker_groups;
6691         int handled;
6692         struct blk_plug plug;
6693
6694         pr_debug("+++ raid5worker active\n");
6695
6696         blk_start_plug(&plug);
6697         handled = 0;
6698         spin_lock_irq(&conf->device_lock);
6699         while (1) {
6700                 int batch_size, released;
6701
6702                 released = release_stripe_list(conf, worker->temp_inactive_list);
6703
6704                 batch_size = handle_active_stripes(conf, group_id, worker,
6705                                                    worker->temp_inactive_list);
6706                 worker->working = false;
6707                 if (!batch_size && !released)
6708                         break;
6709                 handled += batch_size;
6710                 wait_event_lock_irq(mddev->sb_wait,
6711                         !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6712                         conf->device_lock);
6713         }
6714         pr_debug("%d stripes handled\n", handled);
6715
6716         spin_unlock_irq(&conf->device_lock);
6717
6718         flush_deferred_bios(conf);
6719
6720         r5l_flush_stripe_to_raid(conf->log);
6721
6722         async_tx_issue_pending_all();
6723         blk_finish_plug(&plug);
6724
6725         pr_debug("--- raid5worker inactive\n");
6726 }
6727
6728 /*
6729  * This is our raid5 kernel thread.
6730  *
6731  * We scan the hash table for stripes which can be handled now.
6732  * During the scan, completed stripes are saved for us by the interrupt
6733  * handler, so that they will not have to wait for our next wakeup.
6734  */
6735 static void raid5d(struct md_thread *thread)
6736 {
6737         struct mddev *mddev = thread->mddev;
6738         struct r5conf *conf = mddev->private;
6739         int handled;
6740         struct blk_plug plug;
6741
6742         pr_debug("+++ raid5d active\n");
6743
6744         md_check_recovery(mddev);
6745
6746         blk_start_plug(&plug);
6747         handled = 0;
6748         spin_lock_irq(&conf->device_lock);
6749         while (1) {
6750                 struct bio *bio;
6751                 int batch_size, released;
6752                 unsigned int offset;
6753
6754                 if (test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
6755                         break;
6756
6757                 released = release_stripe_list(conf, conf->temp_inactive_list);
6758                 if (released)
6759                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
6760
6761                 if (
6762                     !list_empty(&conf->bitmap_list)) {
6763                         /* Now is a good time to flush some bitmap updates */
6764                         conf->seq_flush++;
6765                         spin_unlock_irq(&conf->device_lock);
6766                         mddev->bitmap_ops->unplug(mddev, true);
6767                         spin_lock_irq(&conf->device_lock);
6768                         conf->seq_write = conf->seq_flush;
6769                         activate_bit_delay(conf, conf->temp_inactive_list);
6770                 }
6771                 raid5_activate_delayed(conf);
6772
6773                 while ((bio = remove_bio_from_retry(conf, &offset))) {
6774                         int ok;
6775                         spin_unlock_irq(&conf->device_lock);
6776                         ok = retry_aligned_read(conf, bio, offset);
6777                         spin_lock_irq(&conf->device_lock);
6778                         if (!ok)
6779                                 break;
6780                         handled++;
6781                 }
6782
6783                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6784                                                    conf->temp_inactive_list);
6785                 if (!batch_size && !released)
6786                         break;
6787                 handled += batch_size;
6788
6789                 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6790                         spin_unlock_irq(&conf->device_lock);
6791                         md_check_recovery(mddev);
6792                         spin_lock_irq(&conf->device_lock);
6793                 }
6794         }
6795         pr_debug("%d stripes handled\n", handled);
6796
6797         spin_unlock_irq(&conf->device_lock);
6798         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6799             mutex_trylock(&conf->cache_size_mutex)) {
6800                 grow_one_stripe(conf, __GFP_NOWARN);
6801                 /* Set flag even if allocation failed.  This helps
6802                  * slow down allocation requests when mem is short
6803                  */
6804                 set_bit(R5_DID_ALLOC, &conf->cache_state);
6805                 mutex_unlock(&conf->cache_size_mutex);
6806         }
6807
6808         flush_deferred_bios(conf);
6809
6810         r5l_flush_stripe_to_raid(conf->log);
6811
6812         async_tx_issue_pending_all();
6813         blk_finish_plug(&plug);
6814
6815         pr_debug("--- raid5d inactive\n");
6816 }
6817
6818 static ssize_t
6819 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6820 {
6821         struct r5conf *conf;
6822         int ret = 0;
6823         spin_lock(&mddev->lock);
6824         conf = mddev->private;
6825         if (conf)
6826                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6827         spin_unlock(&mddev->lock);
6828         return ret;
6829 }
6830
6831 int
6832 raid5_set_cache_size(struct mddev *mddev, int size)
6833 {
6834         int result = 0;
6835         struct r5conf *conf = mddev->private;
6836
6837         if (size <= 16 || size > 32768)
6838                 return -EINVAL;
6839
6840         WRITE_ONCE(conf->min_nr_stripes, size);
6841         mutex_lock(&conf->cache_size_mutex);
6842         while (size < conf->max_nr_stripes &&
6843                drop_one_stripe(conf))
6844                 ;
6845         mutex_unlock(&conf->cache_size_mutex);
6846
6847         md_allow_write(mddev);
6848
6849         mutex_lock(&conf->cache_size_mutex);
6850         while (size > conf->max_nr_stripes)
6851                 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6852                         WRITE_ONCE(conf->min_nr_stripes, conf->max_nr_stripes);
6853                         result = -ENOMEM;
6854                         break;
6855                 }
6856         mutex_unlock(&conf->cache_size_mutex);
6857
6858         return result;
6859 }
6860 EXPORT_SYMBOL(raid5_set_cache_size);
6861
6862 static ssize_t
6863 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6864 {
6865         struct r5conf *conf;
6866         unsigned long new;
6867         int err;
6868
6869         if (len >= PAGE_SIZE)
6870                 return -EINVAL;
6871         if (kstrtoul(page, 10, &new))
6872                 return -EINVAL;
6873         err = mddev_lock(mddev);
6874         if (err)
6875                 return err;
6876         conf = mddev->private;
6877         if (!conf)
6878                 err = -ENODEV;
6879         else
6880                 err = raid5_set_cache_size(mddev, new);
6881         mddev_unlock(mddev);
6882
6883         return err ?: len;
6884 }
6885
6886 static struct md_sysfs_entry
6887 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6888                                 raid5_show_stripe_cache_size,
6889                                 raid5_store_stripe_cache_size);
6890
6891 static ssize_t
6892 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6893 {
6894         struct r5conf *conf = mddev->private;
6895         if (conf)
6896                 return sprintf(page, "%d\n", conf->rmw_level);
6897         else
6898                 return 0;
6899 }
6900
6901 static ssize_t
6902 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6903 {
6904         struct r5conf *conf = mddev->private;
6905         unsigned long new;
6906
6907         if (!conf)
6908                 return -ENODEV;
6909
6910         if (len >= PAGE_SIZE)
6911                 return -EINVAL;
6912
6913         if (kstrtoul(page, 10, &new))
6914                 return -EINVAL;
6915
6916         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6917                 return -EINVAL;
6918
6919         if (new != PARITY_DISABLE_RMW &&
6920             new != PARITY_ENABLE_RMW &&
6921             new != PARITY_PREFER_RMW)
6922                 return -EINVAL;
6923
6924         conf->rmw_level = new;
6925         return len;
6926 }
6927
6928 static struct md_sysfs_entry
6929 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6930                          raid5_show_rmw_level,
6931                          raid5_store_rmw_level);
6932
6933 static ssize_t
6934 raid5_show_stripe_size(struct mddev  *mddev, char *page)
6935 {
6936         struct r5conf *conf;
6937         int ret = 0;
6938
6939         spin_lock(&mddev->lock);
6940         conf = mddev->private;
6941         if (conf)
6942                 ret = sprintf(page, "%lu\n", RAID5_STRIPE_SIZE(conf));
6943         spin_unlock(&mddev->lock);
6944         return ret;
6945 }
6946
6947 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
6948 static ssize_t
6949 raid5_store_stripe_size(struct mddev  *mddev, const char *page, size_t len)
6950 {
6951         struct r5conf *conf;
6952         unsigned long new;
6953         int err;
6954         int size;
6955
6956         if (len >= PAGE_SIZE)
6957                 return -EINVAL;
6958         if (kstrtoul(page, 10, &new))
6959                 return -EINVAL;
6960
6961         /*
6962          * The value should not be bigger than PAGE_SIZE. It requires to
6963          * be multiple of DEFAULT_STRIPE_SIZE and the value should be power
6964          * of two.
6965          */
6966         if (new % DEFAULT_STRIPE_SIZE != 0 ||
6967                         new > PAGE_SIZE || new == 0 ||
6968                         new != roundup_pow_of_two(new))
6969                 return -EINVAL;
6970
6971         err = mddev_suspend_and_lock(mddev);
6972         if (err)
6973                 return err;
6974
6975         conf = mddev->private;
6976         if (!conf) {
6977                 err = -ENODEV;
6978                 goto out_unlock;
6979         }
6980
6981         if (new == conf->stripe_size)
6982                 goto out_unlock;
6983
6984         pr_debug("md/raid: change stripe_size from %lu to %lu\n",
6985                         conf->stripe_size, new);
6986
6987         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
6988             mddev->reshape_position != MaxSector || mddev->sysfs_active) {
6989                 err = -EBUSY;
6990                 goto out_unlock;
6991         }
6992
6993         mutex_lock(&conf->cache_size_mutex);
6994         size = conf->max_nr_stripes;
6995
6996         shrink_stripes(conf);
6997
6998         conf->stripe_size = new;
6999         conf->stripe_shift = ilog2(new) - 9;
7000         conf->stripe_sectors = new >> 9;
7001         if (grow_stripes(conf, size)) {
7002                 pr_warn("md/raid:%s: couldn't allocate buffers\n",
7003                                 mdname(mddev));
7004                 err = -ENOMEM;
7005         }
7006         mutex_unlock(&conf->cache_size_mutex);
7007
7008 out_unlock:
7009         mddev_unlock_and_resume(mddev);
7010         return err ?: len;
7011 }
7012
7013 static struct md_sysfs_entry
7014 raid5_stripe_size = __ATTR(stripe_size, 0644,
7015                          raid5_show_stripe_size,
7016                          raid5_store_stripe_size);
7017 #else
7018 static struct md_sysfs_entry
7019 raid5_stripe_size = __ATTR(stripe_size, 0444,
7020                          raid5_show_stripe_size,
7021                          NULL);
7022 #endif
7023
7024 static ssize_t
7025 raid5_show_preread_threshold(struct mddev *mddev, char *page)
7026 {
7027         struct r5conf *conf;
7028         int ret = 0;
7029         spin_lock(&mddev->lock);
7030         conf = mddev->private;
7031         if (conf)
7032                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
7033         spin_unlock(&mddev->lock);
7034         return ret;
7035 }
7036
7037 static ssize_t
7038 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
7039 {
7040         struct r5conf *conf;
7041         unsigned long new;
7042         int err;
7043
7044         if (len >= PAGE_SIZE)
7045                 return -EINVAL;
7046         if (kstrtoul(page, 10, &new))
7047                 return -EINVAL;
7048
7049         err = mddev_lock(mddev);
7050         if (err)
7051                 return err;
7052         conf = mddev->private;
7053         if (!conf)
7054                 err = -ENODEV;
7055         else if (new > conf->min_nr_stripes)
7056                 err = -EINVAL;
7057         else
7058                 conf->bypass_threshold = new;
7059         mddev_unlock(mddev);
7060         return err ?: len;
7061 }
7062
7063 static struct md_sysfs_entry
7064 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
7065                                         S_IRUGO | S_IWUSR,
7066                                         raid5_show_preread_threshold,
7067                                         raid5_store_preread_threshold);
7068
7069 static ssize_t
7070 raid5_show_skip_copy(struct mddev *mddev, char *page)
7071 {
7072         struct r5conf *conf;
7073         int ret = 0;
7074         spin_lock(&mddev->lock);
7075         conf = mddev->private;
7076         if (conf)
7077                 ret = sprintf(page, "%d\n", conf->skip_copy);
7078         spin_unlock(&mddev->lock);
7079         return ret;
7080 }
7081
7082 static ssize_t
7083 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
7084 {
7085         struct r5conf *conf;
7086         unsigned long new;
7087         int err;
7088
7089         if (len >= PAGE_SIZE)
7090                 return -EINVAL;
7091         if (kstrtoul(page, 10, &new))
7092                 return -EINVAL;
7093         new = !!new;
7094
7095         err = mddev_suspend_and_lock(mddev);
7096         if (err)
7097                 return err;
7098         conf = mddev->private;
7099         if (!conf)
7100                 err = -ENODEV;
7101         else if (new != conf->skip_copy) {
7102                 struct request_queue *q = mddev->gendisk->queue;
7103                 struct queue_limits lim = queue_limits_start_update(q);
7104
7105                 conf->skip_copy = new;
7106                 if (new)
7107                         lim.features |= BLK_FEAT_STABLE_WRITES;
7108                 else
7109                         lim.features &= ~BLK_FEAT_STABLE_WRITES;
7110                 err = queue_limits_commit_update(q, &lim);
7111         }
7112         mddev_unlock_and_resume(mddev);
7113         return err ?: len;
7114 }
7115
7116 static struct md_sysfs_entry
7117 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
7118                                         raid5_show_skip_copy,
7119                                         raid5_store_skip_copy);
7120
7121 static ssize_t
7122 stripe_cache_active_show(struct mddev *mddev, char *page)
7123 {
7124         struct r5conf *conf = mddev->private;
7125         if (conf)
7126                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
7127         else
7128                 return 0;
7129 }
7130
7131 static struct md_sysfs_entry
7132 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
7133
7134 static ssize_t
7135 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
7136 {
7137         struct r5conf *conf;
7138         int ret = 0;
7139         spin_lock(&mddev->lock);
7140         conf = mddev->private;
7141         if (conf)
7142                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
7143         spin_unlock(&mddev->lock);
7144         return ret;
7145 }
7146
7147 static int alloc_thread_groups(struct r5conf *conf, int cnt,
7148                                int *group_cnt,
7149                                struct r5worker_group **worker_groups);
7150 static ssize_t
7151 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
7152 {
7153         struct r5conf *conf;
7154         unsigned int new;
7155         int err;
7156         struct r5worker_group *new_groups, *old_groups;
7157         int group_cnt;
7158
7159         if (len >= PAGE_SIZE)
7160                 return -EINVAL;
7161         if (kstrtouint(page, 10, &new))
7162                 return -EINVAL;
7163         /* 8192 should be big enough */
7164         if (new > 8192)
7165                 return -EINVAL;
7166
7167         err = mddev_suspend_and_lock(mddev);
7168         if (err)
7169                 return err;
7170         raid5_quiesce(mddev, true);
7171
7172         conf = mddev->private;
7173         if (!conf)
7174                 err = -ENODEV;
7175         else if (new != conf->worker_cnt_per_group) {
7176                 old_groups = conf->worker_groups;
7177                 if (old_groups)
7178                         flush_workqueue(raid5_wq);
7179
7180                 err = alloc_thread_groups(conf, new, &group_cnt, &new_groups);
7181                 if (!err) {
7182                         spin_lock_irq(&conf->device_lock);
7183                         conf->group_cnt = group_cnt;
7184                         conf->worker_cnt_per_group = new;
7185                         conf->worker_groups = new_groups;
7186                         spin_unlock_irq(&conf->device_lock);
7187
7188                         if (old_groups)
7189                                 kfree(old_groups[0].workers);
7190                         kfree(old_groups);
7191                 }
7192         }
7193
7194         raid5_quiesce(mddev, false);
7195         mddev_unlock_and_resume(mddev);
7196
7197         return err ?: len;
7198 }
7199
7200 static struct md_sysfs_entry
7201 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
7202                                 raid5_show_group_thread_cnt,
7203                                 raid5_store_group_thread_cnt);
7204
7205 static struct attribute *raid5_attrs[] =  {
7206         &raid5_stripecache_size.attr,
7207         &raid5_stripecache_active.attr,
7208         &raid5_preread_bypass_threshold.attr,
7209         &raid5_group_thread_cnt.attr,
7210         &raid5_skip_copy.attr,
7211         &raid5_rmw_level.attr,
7212         &raid5_stripe_size.attr,
7213         &r5c_journal_mode.attr,
7214         &ppl_write_hint.attr,
7215         NULL,
7216 };
7217 static const struct attribute_group raid5_attrs_group = {
7218         .name = NULL,
7219         .attrs = raid5_attrs,
7220 };
7221
7222 static int alloc_thread_groups(struct r5conf *conf, int cnt, int *group_cnt,
7223                                struct r5worker_group **worker_groups)
7224 {
7225         int i, j, k;
7226         ssize_t size;
7227         struct r5worker *workers;
7228
7229         if (cnt == 0) {
7230                 *group_cnt = 0;
7231                 *worker_groups = NULL;
7232                 return 0;
7233         }
7234         *group_cnt = num_possible_nodes();
7235         size = sizeof(struct r5worker) * cnt;
7236         workers = kcalloc(size, *group_cnt, GFP_NOIO);
7237         *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
7238                                  GFP_NOIO);
7239         if (!*worker_groups || !workers) {
7240                 kfree(workers);
7241                 kfree(*worker_groups);
7242                 return -ENOMEM;
7243         }
7244
7245         for (i = 0; i < *group_cnt; i++) {
7246                 struct r5worker_group *group;
7247
7248                 group = &(*worker_groups)[i];
7249                 INIT_LIST_HEAD(&group->handle_list);
7250                 INIT_LIST_HEAD(&group->loprio_list);
7251                 group->conf = conf;
7252                 group->workers = workers + i * cnt;
7253
7254                 for (j = 0; j < cnt; j++) {
7255                         struct r5worker *worker = group->workers + j;
7256                         worker->group = group;
7257                         INIT_WORK(&worker->work, raid5_do_work);
7258
7259                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
7260                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
7261                 }
7262         }
7263
7264         return 0;
7265 }
7266
7267 static void free_thread_groups(struct r5conf *conf)
7268 {
7269         if (conf->worker_groups)
7270                 kfree(conf->worker_groups[0].workers);
7271         kfree(conf->worker_groups);
7272         conf->worker_groups = NULL;
7273 }
7274
7275 static sector_t
7276 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
7277 {
7278         struct r5conf *conf = mddev->private;
7279
7280         if (!sectors)
7281                 sectors = mddev->dev_sectors;
7282         if (!raid_disks)
7283                 /* size is defined by the smallest of previous and new size */
7284                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
7285
7286         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7287         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
7288         return sectors * (raid_disks - conf->max_degraded);
7289 }
7290
7291 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7292 {
7293         safe_put_page(percpu->spare_page);
7294         percpu->spare_page = NULL;
7295         kvfree(percpu->scribble);
7296         percpu->scribble = NULL;
7297 }
7298
7299 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7300 {
7301         if (conf->level == 6 && !percpu->spare_page) {
7302                 percpu->spare_page = alloc_page(GFP_KERNEL);
7303                 if (!percpu->spare_page)
7304                         return -ENOMEM;
7305         }
7306
7307         if (scribble_alloc(percpu,
7308                            max(conf->raid_disks,
7309                                conf->previous_raid_disks),
7310                            max(conf->chunk_sectors,
7311                                conf->prev_chunk_sectors)
7312                            / RAID5_STRIPE_SECTORS(conf))) {
7313                 free_scratch_buffer(conf, percpu);
7314                 return -ENOMEM;
7315         }
7316
7317         local_lock_init(&percpu->lock);
7318         return 0;
7319 }
7320
7321 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
7322 {
7323         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7324
7325         free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
7326         return 0;
7327 }
7328
7329 static void raid5_free_percpu(struct r5conf *conf)
7330 {
7331         if (!conf->percpu)
7332                 return;
7333
7334         cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7335         free_percpu(conf->percpu);
7336 }
7337
7338 static void free_conf(struct r5conf *conf)
7339 {
7340         int i;
7341
7342         log_exit(conf);
7343
7344         shrinker_free(conf->shrinker);
7345         free_thread_groups(conf);
7346         shrink_stripes(conf);
7347         raid5_free_percpu(conf);
7348         for (i = 0; i < conf->pool_size; i++)
7349                 if (conf->disks[i].extra_page)
7350                         put_page(conf->disks[i].extra_page);
7351         kfree(conf->disks);
7352         bioset_exit(&conf->bio_split);
7353         kfree(conf->stripe_hashtbl);
7354         kfree(conf->pending_data);
7355         kfree(conf);
7356 }
7357
7358 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
7359 {
7360         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7361         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
7362
7363         if (alloc_scratch_buffer(conf, percpu)) {
7364                 pr_warn("%s: failed memory allocation for cpu%u\n",
7365                         __func__, cpu);
7366                 return -ENOMEM;
7367         }
7368         return 0;
7369 }
7370
7371 static int raid5_alloc_percpu(struct r5conf *conf)
7372 {
7373         int err = 0;
7374
7375         conf->percpu = alloc_percpu(struct raid5_percpu);
7376         if (!conf->percpu)
7377                 return -ENOMEM;
7378
7379         err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7380         if (!err) {
7381                 conf->scribble_disks = max(conf->raid_disks,
7382                         conf->previous_raid_disks);
7383                 conf->scribble_sectors = max(conf->chunk_sectors,
7384                         conf->prev_chunk_sectors);
7385         }
7386         return err;
7387 }
7388
7389 static unsigned long raid5_cache_scan(struct shrinker *shrink,
7390                                       struct shrink_control *sc)
7391 {
7392         struct r5conf *conf = shrink->private_data;
7393         unsigned long ret = SHRINK_STOP;
7394
7395         if (mutex_trylock(&conf->cache_size_mutex)) {
7396                 ret= 0;
7397                 while (ret < sc->nr_to_scan &&
7398                        conf->max_nr_stripes > conf->min_nr_stripes) {
7399                         if (drop_one_stripe(conf) == 0) {
7400                                 ret = SHRINK_STOP;
7401                                 break;
7402                         }
7403                         ret++;
7404                 }
7405                 mutex_unlock(&conf->cache_size_mutex);
7406         }
7407         return ret;
7408 }
7409
7410 static unsigned long raid5_cache_count(struct shrinker *shrink,
7411                                        struct shrink_control *sc)
7412 {
7413         struct r5conf *conf = shrink->private_data;
7414         int max_stripes = READ_ONCE(conf->max_nr_stripes);
7415         int min_stripes = READ_ONCE(conf->min_nr_stripes);
7416
7417         if (max_stripes < min_stripes)
7418                 /* unlikely, but not impossible */
7419                 return 0;
7420         return max_stripes - min_stripes;
7421 }
7422
7423 static struct r5conf *setup_conf(struct mddev *mddev)
7424 {
7425         struct r5conf *conf;
7426         int raid_disk, memory, max_disks;
7427         struct md_rdev *rdev;
7428         struct disk_info *disk;
7429         char pers_name[6];
7430         int i;
7431         int group_cnt;
7432         struct r5worker_group *new_group;
7433         int ret = -ENOMEM;
7434
7435         if (mddev->new_level != 5
7436             && mddev->new_level != 4
7437             && mddev->new_level != 6) {
7438                 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
7439                         mdname(mddev), mddev->new_level);
7440                 return ERR_PTR(-EIO);
7441         }
7442         if ((mddev->new_level == 5
7443              && !algorithm_valid_raid5(mddev->new_layout)) ||
7444             (mddev->new_level == 6
7445              && !algorithm_valid_raid6(mddev->new_layout))) {
7446                 pr_warn("md/raid:%s: layout %d not supported\n",
7447                         mdname(mddev), mddev->new_layout);
7448                 return ERR_PTR(-EIO);
7449         }
7450         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
7451                 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
7452                         mdname(mddev), mddev->raid_disks);
7453                 return ERR_PTR(-EINVAL);
7454         }
7455
7456         if (!mddev->new_chunk_sectors ||
7457             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
7458             !is_power_of_2(mddev->new_chunk_sectors)) {
7459                 pr_warn("md/raid:%s: invalid chunk size %d\n",
7460                         mdname(mddev), mddev->new_chunk_sectors << 9);
7461                 return ERR_PTR(-EINVAL);
7462         }
7463
7464         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
7465         if (conf == NULL)
7466                 goto abort;
7467
7468 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
7469         conf->stripe_size = DEFAULT_STRIPE_SIZE;
7470         conf->stripe_shift = ilog2(DEFAULT_STRIPE_SIZE) - 9;
7471         conf->stripe_sectors = DEFAULT_STRIPE_SIZE >> 9;
7472 #endif
7473         INIT_LIST_HEAD(&conf->free_list);
7474         INIT_LIST_HEAD(&conf->pending_list);
7475         conf->pending_data = kcalloc(PENDING_IO_MAX,
7476                                      sizeof(struct r5pending_data),
7477                                      GFP_KERNEL);
7478         if (!conf->pending_data)
7479                 goto abort;
7480         for (i = 0; i < PENDING_IO_MAX; i++)
7481                 list_add(&conf->pending_data[i].sibling, &conf->free_list);
7482         /* Don't enable multi-threading by default*/
7483         if (!alloc_thread_groups(conf, 0, &group_cnt, &new_group)) {
7484                 conf->group_cnt = group_cnt;
7485                 conf->worker_cnt_per_group = 0;
7486                 conf->worker_groups = new_group;
7487         } else
7488                 goto abort;
7489         spin_lock_init(&conf->device_lock);
7490         seqcount_spinlock_init(&conf->gen_lock, &conf->device_lock);
7491         mutex_init(&conf->cache_size_mutex);
7492
7493         init_waitqueue_head(&conf->wait_for_quiescent);
7494         init_waitqueue_head(&conf->wait_for_stripe);
7495         init_waitqueue_head(&conf->wait_for_reshape);
7496         INIT_LIST_HEAD(&conf->handle_list);
7497         INIT_LIST_HEAD(&conf->loprio_list);
7498         INIT_LIST_HEAD(&conf->hold_list);
7499         INIT_LIST_HEAD(&conf->delayed_list);
7500         INIT_LIST_HEAD(&conf->bitmap_list);
7501         init_llist_head(&conf->released_stripes);
7502         atomic_set(&conf->active_stripes, 0);
7503         atomic_set(&conf->preread_active_stripes, 0);
7504         atomic_set(&conf->active_aligned_reads, 0);
7505         spin_lock_init(&conf->pending_bios_lock);
7506         conf->batch_bio_dispatch = true;
7507         rdev_for_each(rdev, mddev) {
7508                 if (test_bit(Journal, &rdev->flags))
7509                         continue;
7510                 if (bdev_nonrot(rdev->bdev)) {
7511                         conf->batch_bio_dispatch = false;
7512                         break;
7513                 }
7514         }
7515
7516         conf->bypass_threshold = BYPASS_THRESHOLD;
7517         conf->recovery_disabled = mddev->recovery_disabled - 1;
7518
7519         conf->raid_disks = mddev->raid_disks;
7520         if (mddev->reshape_position == MaxSector)
7521                 conf->previous_raid_disks = mddev->raid_disks;
7522         else
7523                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
7524         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
7525
7526         conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
7527                               GFP_KERNEL);
7528
7529         if (!conf->disks)
7530                 goto abort;
7531
7532         for (i = 0; i < max_disks; i++) {
7533                 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
7534                 if (!conf->disks[i].extra_page)
7535                         goto abort;
7536         }
7537
7538         ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
7539         if (ret)
7540                 goto abort;
7541         conf->mddev = mddev;
7542
7543         ret = -ENOMEM;
7544         conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL);
7545         if (!conf->stripe_hashtbl)
7546                 goto abort;
7547
7548         /* We init hash_locks[0] separately to that it can be used
7549          * as the reference lock in the spin_lock_nest_lock() call
7550          * in lock_all_device_hash_locks_irq in order to convince
7551          * lockdep that we know what we are doing.
7552          */
7553         spin_lock_init(conf->hash_locks);
7554         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
7555                 spin_lock_init(conf->hash_locks + i);
7556
7557         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7558                 INIT_LIST_HEAD(conf->inactive_list + i);
7559
7560         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7561                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
7562
7563         atomic_set(&conf->r5c_cached_full_stripes, 0);
7564         INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7565         atomic_set(&conf->r5c_cached_partial_stripes, 0);
7566         INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7567         atomic_set(&conf->r5c_flushing_full_stripes, 0);
7568         atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7569
7570         conf->level = mddev->new_level;
7571         conf->chunk_sectors = mddev->new_chunk_sectors;
7572         ret = raid5_alloc_percpu(conf);
7573         if (ret)
7574                 goto abort;
7575
7576         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7577
7578         ret = -EIO;
7579         rdev_for_each(rdev, mddev) {
7580                 raid_disk = rdev->raid_disk;
7581                 if (raid_disk >= max_disks
7582                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7583                         continue;
7584                 disk = conf->disks + raid_disk;
7585
7586                 if (test_bit(Replacement, &rdev->flags)) {
7587                         if (disk->replacement)
7588                                 goto abort;
7589                         disk->replacement = rdev;
7590                 } else {
7591                         if (disk->rdev)
7592                                 goto abort;
7593                         disk->rdev = rdev;
7594                 }
7595
7596                 if (test_bit(In_sync, &rdev->flags)) {
7597                         pr_info("md/raid:%s: device %pg operational as raid disk %d\n",
7598                                 mdname(mddev), rdev->bdev, raid_disk);
7599                 } else if (rdev->saved_raid_disk != raid_disk)
7600                         /* Cannot rely on bitmap to complete recovery */
7601                         conf->fullsync = 1;
7602         }
7603
7604         conf->level = mddev->new_level;
7605         if (conf->level == 6) {
7606                 conf->max_degraded = 2;
7607                 if (raid6_call.xor_syndrome)
7608                         conf->rmw_level = PARITY_ENABLE_RMW;
7609                 else
7610                         conf->rmw_level = PARITY_DISABLE_RMW;
7611         } else {
7612                 conf->max_degraded = 1;
7613                 conf->rmw_level = PARITY_ENABLE_RMW;
7614         }
7615         conf->algorithm = mddev->new_layout;
7616         conf->reshape_progress = mddev->reshape_position;
7617         if (conf->reshape_progress != MaxSector) {
7618                 conf->prev_chunk_sectors = mddev->chunk_sectors;
7619                 conf->prev_algo = mddev->layout;
7620         } else {
7621                 conf->prev_chunk_sectors = conf->chunk_sectors;
7622                 conf->prev_algo = conf->algorithm;
7623         }
7624
7625         conf->min_nr_stripes = NR_STRIPES;
7626         if (mddev->reshape_position != MaxSector) {
7627                 int stripes = max_t(int,
7628                         ((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4,
7629                         ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4);
7630                 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7631                 if (conf->min_nr_stripes != NR_STRIPES)
7632                         pr_info("md/raid:%s: force stripe size %d for reshape\n",
7633                                 mdname(mddev), conf->min_nr_stripes);
7634         }
7635         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7636                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7637         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7638         if (grow_stripes(conf, conf->min_nr_stripes)) {
7639                 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7640                         mdname(mddev), memory);
7641                 ret = -ENOMEM;
7642                 goto abort;
7643         } else
7644                 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7645         /*
7646          * Losing a stripe head costs more than the time to refill it,
7647          * it reduces the queue depth and so can hurt throughput.
7648          * So set it rather large, scaled by number of devices.
7649          */
7650         conf->shrinker = shrinker_alloc(0, "md-raid5:%s", mdname(mddev));
7651         if (!conf->shrinker) {
7652                 ret = -ENOMEM;
7653                 pr_warn("md/raid:%s: couldn't allocate shrinker.\n",
7654                         mdname(mddev));
7655                 goto abort;
7656         }
7657
7658         conf->shrinker->seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7659         conf->shrinker->scan_objects = raid5_cache_scan;
7660         conf->shrinker->count_objects = raid5_cache_count;
7661         conf->shrinker->batch = 128;
7662         conf->shrinker->private_data = conf;
7663
7664         shrinker_register(conf->shrinker);
7665
7666         sprintf(pers_name, "raid%d", mddev->new_level);
7667         rcu_assign_pointer(conf->thread,
7668                            md_register_thread(raid5d, mddev, pers_name));
7669         if (!conf->thread) {
7670                 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7671                         mdname(mddev));
7672                 ret = -ENOMEM;
7673                 goto abort;
7674         }
7675
7676         return conf;
7677
7678  abort:
7679         if (conf)
7680                 free_conf(conf);
7681         return ERR_PTR(ret);
7682 }
7683
7684 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7685 {
7686         switch (algo) {
7687         case ALGORITHM_PARITY_0:
7688                 if (raid_disk < max_degraded)
7689                         return 1;
7690                 break;
7691         case ALGORITHM_PARITY_N:
7692                 if (raid_disk >= raid_disks - max_degraded)
7693                         return 1;
7694                 break;
7695         case ALGORITHM_PARITY_0_6:
7696                 if (raid_disk == 0 ||
7697                     raid_disk == raid_disks - 1)
7698                         return 1;
7699                 break;
7700         case ALGORITHM_LEFT_ASYMMETRIC_6:
7701         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7702         case ALGORITHM_LEFT_SYMMETRIC_6:
7703         case ALGORITHM_RIGHT_SYMMETRIC_6:
7704                 if (raid_disk == raid_disks - 1)
7705                         return 1;
7706         }
7707         return 0;
7708 }
7709
7710 static int raid5_set_limits(struct mddev *mddev)
7711 {
7712         struct r5conf *conf = mddev->private;
7713         struct queue_limits lim;
7714         int data_disks, stripe;
7715         struct md_rdev *rdev;
7716
7717         /*
7718          * The read-ahead size must cover two whole stripes, which is
7719          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices.
7720          */
7721         data_disks = conf->previous_raid_disks - conf->max_degraded;
7722
7723         /*
7724          * We can only discard a whole stripe. It doesn't make sense to
7725          * discard data disk but write parity disk
7726          */
7727         stripe = roundup_pow_of_two(data_disks * (mddev->chunk_sectors << 9));
7728
7729         md_init_stacking_limits(&lim);
7730         lim.io_min = mddev->chunk_sectors << 9;
7731         lim.io_opt = lim.io_min * (conf->raid_disks - conf->max_degraded);
7732         lim.features |= BLK_FEAT_RAID_PARTIAL_STRIPES_EXPENSIVE;
7733         lim.discard_granularity = stripe;
7734         lim.max_write_zeroes_sectors = 0;
7735         mddev_stack_rdev_limits(mddev, &lim, 0);
7736         rdev_for_each(rdev, mddev)
7737                 queue_limits_stack_bdev(&lim, rdev->bdev, rdev->new_data_offset,
7738                                 mddev->gendisk->disk_name);
7739
7740         /*
7741          * Zeroing is required for discard, otherwise data could be lost.
7742          *
7743          * Consider a scenario: discard a stripe (the stripe could be
7744          * inconsistent if discard_zeroes_data is 0); write one disk of the
7745          * stripe (the stripe could be inconsistent again depending on which
7746          * disks are used to calculate parity); the disk is broken; The stripe
7747          * data of this disk is lost.
7748          *
7749          * We only allow DISCARD if the sysadmin has confirmed that only safe
7750          * devices are in use by setting a module parameter.  A better idea
7751          * might be to turn DISCARD into WRITE_ZEROES requests, as that is
7752          * required to be safe.
7753          */
7754         if (!devices_handle_discard_safely ||
7755             lim.max_discard_sectors < (stripe >> 9) ||
7756             lim.discard_granularity < stripe)
7757                 lim.max_hw_discard_sectors = 0;
7758
7759         /*
7760          * Requests require having a bitmap for each stripe.
7761          * Limit the max sectors based on this.
7762          */
7763         lim.max_hw_sectors = RAID5_MAX_REQ_STRIPES << RAID5_STRIPE_SHIFT(conf);
7764
7765         /* No restrictions on the number of segments in the request */
7766         lim.max_segments = USHRT_MAX;
7767
7768         return queue_limits_set(mddev->gendisk->queue, &lim);
7769 }
7770
7771 static int raid5_run(struct mddev *mddev)
7772 {
7773         struct r5conf *conf;
7774         int dirty_parity_disks = 0;
7775         struct md_rdev *rdev;
7776         struct md_rdev *journal_dev = NULL;
7777         sector_t reshape_offset = 0;
7778         int i;
7779         long long min_offset_diff = 0;
7780         int first = 1;
7781         int ret = -EIO;
7782
7783         if (mddev->recovery_cp != MaxSector)
7784                 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7785                           mdname(mddev));
7786
7787         rdev_for_each(rdev, mddev) {
7788                 long long diff;
7789
7790                 if (test_bit(Journal, &rdev->flags)) {
7791                         journal_dev = rdev;
7792                         continue;
7793                 }
7794                 if (rdev->raid_disk < 0)
7795                         continue;
7796                 diff = (rdev->new_data_offset - rdev->data_offset);
7797                 if (first) {
7798                         min_offset_diff = diff;
7799                         first = 0;
7800                 } else if (mddev->reshape_backwards &&
7801                          diff < min_offset_diff)
7802                         min_offset_diff = diff;
7803                 else if (!mddev->reshape_backwards &&
7804                          diff > min_offset_diff)
7805                         min_offset_diff = diff;
7806         }
7807
7808         if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7809             (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7810                 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7811                           mdname(mddev));
7812                 return -EINVAL;
7813         }
7814
7815         if (mddev->reshape_position != MaxSector) {
7816                 /* Check that we can continue the reshape.
7817                  * Difficulties arise if the stripe we would write to
7818                  * next is at or after the stripe we would read from next.
7819                  * For a reshape that changes the number of devices, this
7820                  * is only possible for a very short time, and mdadm makes
7821                  * sure that time appears to have past before assembling
7822                  * the array.  So we fail if that time hasn't passed.
7823                  * For a reshape that keeps the number of devices the same
7824                  * mdadm must be monitoring the reshape can keeping the
7825                  * critical areas read-only and backed up.  It will start
7826                  * the array in read-only mode, so we check for that.
7827                  */
7828                 sector_t here_new, here_old;
7829                 int old_disks;
7830                 int max_degraded = (mddev->level == 6 ? 2 : 1);
7831                 int chunk_sectors;
7832                 int new_data_disks;
7833
7834                 if (journal_dev) {
7835                         pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7836                                 mdname(mddev));
7837                         return -EINVAL;
7838                 }
7839
7840                 if (mddev->new_level != mddev->level) {
7841                         pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7842                                 mdname(mddev));
7843                         return -EINVAL;
7844                 }
7845                 old_disks = mddev->raid_disks - mddev->delta_disks;
7846                 /* reshape_position must be on a new-stripe boundary, and one
7847                  * further up in new geometry must map after here in old
7848                  * geometry.
7849                  * If the chunk sizes are different, then as we perform reshape
7850                  * in units of the largest of the two, reshape_position needs
7851                  * be a multiple of the largest chunk size times new data disks.
7852                  */
7853                 here_new = mddev->reshape_position;
7854                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7855                 new_data_disks = mddev->raid_disks - max_degraded;
7856                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7857                         pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7858                                 mdname(mddev));
7859                         return -EINVAL;
7860                 }
7861                 reshape_offset = here_new * chunk_sectors;
7862                 /* here_new is the stripe we will write to */
7863                 here_old = mddev->reshape_position;
7864                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7865                 /* here_old is the first stripe that we might need to read
7866                  * from */
7867                 if (mddev->delta_disks == 0) {
7868                         /* We cannot be sure it is safe to start an in-place
7869                          * reshape.  It is only safe if user-space is monitoring
7870                          * and taking constant backups.
7871                          * mdadm always starts a situation like this in
7872                          * readonly mode so it can take control before
7873                          * allowing any writes.  So just check for that.
7874                          */
7875                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7876                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
7877                                 /* not really in-place - so OK */;
7878                         else if (mddev->ro == 0) {
7879                                 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7880                                         mdname(mddev));
7881                                 return -EINVAL;
7882                         }
7883                 } else if (mddev->reshape_backwards
7884                     ? (here_new * chunk_sectors + min_offset_diff <=
7885                        here_old * chunk_sectors)
7886                     : (here_new * chunk_sectors >=
7887                        here_old * chunk_sectors + (-min_offset_diff))) {
7888                         /* Reading from the same stripe as writing to - bad */
7889                         pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7890                                 mdname(mddev));
7891                         return -EINVAL;
7892                 }
7893                 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7894                 /* OK, we should be able to continue; */
7895         } else {
7896                 BUG_ON(mddev->level != mddev->new_level);
7897                 BUG_ON(mddev->layout != mddev->new_layout);
7898                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7899                 BUG_ON(mddev->delta_disks != 0);
7900         }
7901
7902         if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7903             test_bit(MD_HAS_PPL, &mddev->flags)) {
7904                 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7905                         mdname(mddev));
7906                 clear_bit(MD_HAS_PPL, &mddev->flags);
7907                 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7908         }
7909
7910         if (mddev->private == NULL)
7911                 conf = setup_conf(mddev);
7912         else
7913                 conf = mddev->private;
7914
7915         if (IS_ERR(conf))
7916                 return PTR_ERR(conf);
7917
7918         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7919                 if (!journal_dev) {
7920                         pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7921                                 mdname(mddev));
7922                         mddev->ro = 1;
7923                         set_disk_ro(mddev->gendisk, 1);
7924                 } else if (mddev->recovery_cp == MaxSector)
7925                         set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7926         }
7927
7928         conf->min_offset_diff = min_offset_diff;
7929         rcu_assign_pointer(mddev->thread, conf->thread);
7930         rcu_assign_pointer(conf->thread, NULL);
7931         mddev->private = conf;
7932
7933         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7934              i++) {
7935                 rdev = conf->disks[i].rdev;
7936                 if (!rdev)
7937                         continue;
7938                 if (conf->disks[i].replacement &&
7939                     conf->reshape_progress != MaxSector) {
7940                         /* replacements and reshape simply do not mix. */
7941                         pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7942                         goto abort;
7943                 }
7944                 if (test_bit(In_sync, &rdev->flags))
7945                         continue;
7946                 /* This disc is not fully in-sync.  However if it
7947                  * just stored parity (beyond the recovery_offset),
7948                  * when we don't need to be concerned about the
7949                  * array being dirty.
7950                  * When reshape goes 'backwards', we never have
7951                  * partially completed devices, so we only need
7952                  * to worry about reshape going forwards.
7953                  */
7954                 /* Hack because v0.91 doesn't store recovery_offset properly. */
7955                 if (mddev->major_version == 0 &&
7956                     mddev->minor_version > 90)
7957                         rdev->recovery_offset = reshape_offset;
7958
7959                 if (rdev->recovery_offset < reshape_offset) {
7960                         /* We need to check old and new layout */
7961                         if (!only_parity(rdev->raid_disk,
7962                                          conf->algorithm,
7963                                          conf->raid_disks,
7964                                          conf->max_degraded))
7965                                 continue;
7966                 }
7967                 if (!only_parity(rdev->raid_disk,
7968                                  conf->prev_algo,
7969                                  conf->previous_raid_disks,
7970                                  conf->max_degraded))
7971                         continue;
7972                 dirty_parity_disks++;
7973         }
7974
7975         /*
7976          * 0 for a fully functional array, 1 or 2 for a degraded array.
7977          */
7978         mddev->degraded = raid5_calc_degraded(conf);
7979
7980         if (has_failed(conf)) {
7981                 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7982                         mdname(mddev), mddev->degraded, conf->raid_disks);
7983                 goto abort;
7984         }
7985
7986         /* device size must be a multiple of chunk size */
7987         mddev->dev_sectors &= ~((sector_t)mddev->chunk_sectors - 1);
7988         mddev->resync_max_sectors = mddev->dev_sectors;
7989
7990         if (mddev->degraded > dirty_parity_disks &&
7991             mddev->recovery_cp != MaxSector) {
7992                 if (test_bit(MD_HAS_PPL, &mddev->flags))
7993                         pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7994                                 mdname(mddev));
7995                 else if (mddev->ok_start_degraded)
7996                         pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7997                                 mdname(mddev));
7998                 else {
7999                         pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
8000                                 mdname(mddev));
8001                         goto abort;
8002                 }
8003         }
8004
8005         pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
8006                 mdname(mddev), conf->level,
8007                 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
8008                 mddev->new_layout);
8009
8010         print_raid5_conf(conf);
8011
8012         if (conf->reshape_progress != MaxSector) {
8013                 conf->reshape_safe = conf->reshape_progress;
8014                 atomic_set(&conf->reshape_stripes, 0);
8015                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8016                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
8017                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
8018                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8019         }
8020
8021         /* Ok, everything is just fine now */
8022         if (mddev->to_remove == &raid5_attrs_group)
8023                 mddev->to_remove = NULL;
8024         else if (mddev->kobj.sd &&
8025             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
8026                 pr_warn("raid5: failed to create sysfs attributes for %s\n",
8027                         mdname(mddev));
8028         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
8029
8030         if (!mddev_is_dm(mddev)) {
8031                 ret = raid5_set_limits(mddev);
8032                 if (ret)
8033                         goto abort;
8034         }
8035
8036         if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
8037                 goto abort;
8038
8039         return 0;
8040 abort:
8041         md_unregister_thread(mddev, &mddev->thread);
8042         print_raid5_conf(conf);
8043         free_conf(conf);
8044         mddev->private = NULL;
8045         pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
8046         return ret;
8047 }
8048
8049 static void raid5_free(struct mddev *mddev, void *priv)
8050 {
8051         struct r5conf *conf = priv;
8052
8053         free_conf(conf);
8054         mddev->to_remove = &raid5_attrs_group;
8055 }
8056
8057 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
8058 {
8059         struct r5conf *conf = mddev->private;
8060         int i;
8061
8062         lockdep_assert_held(&mddev->lock);
8063
8064         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
8065                 conf->chunk_sectors / 2, mddev->layout);
8066         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
8067         for (i = 0; i < conf->raid_disks; i++) {
8068                 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
8069
8070                 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
8071         }
8072         seq_printf (seq, "]");
8073 }
8074
8075 static void print_raid5_conf(struct r5conf *conf)
8076 {
8077         struct md_rdev *rdev;
8078         int i;
8079
8080         pr_debug("RAID conf printout:\n");
8081         if (!conf) {
8082                 pr_debug("(conf==NULL)\n");
8083                 return;
8084         }
8085         pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
8086                conf->raid_disks,
8087                conf->raid_disks - conf->mddev->degraded);
8088
8089         for (i = 0; i < conf->raid_disks; i++) {
8090                 rdev = conf->disks[i].rdev;
8091                 if (rdev)
8092                         pr_debug(" disk %d, o:%d, dev:%pg\n",
8093                                i, !test_bit(Faulty, &rdev->flags),
8094                                rdev->bdev);
8095         }
8096 }
8097
8098 static int raid5_spare_active(struct mddev *mddev)
8099 {
8100         int i;
8101         struct r5conf *conf = mddev->private;
8102         struct md_rdev *rdev, *replacement;
8103         int count = 0;
8104         unsigned long flags;
8105
8106         for (i = 0; i < conf->raid_disks; i++) {
8107                 rdev = conf->disks[i].rdev;
8108                 replacement = conf->disks[i].replacement;
8109                 if (replacement
8110                     && replacement->recovery_offset == MaxSector
8111                     && !test_bit(Faulty, &replacement->flags)
8112                     && !test_and_set_bit(In_sync, &replacement->flags)) {
8113                         /* Replacement has just become active. */
8114                         if (!rdev
8115                             || !test_and_clear_bit(In_sync, &rdev->flags))
8116                                 count++;
8117                         if (rdev) {
8118                                 /* Replaced device not technically faulty,
8119                                  * but we need to be sure it gets removed
8120                                  * and never re-added.
8121                                  */
8122                                 set_bit(Faulty, &rdev->flags);
8123                                 sysfs_notify_dirent_safe(
8124                                         rdev->sysfs_state);
8125                         }
8126                         sysfs_notify_dirent_safe(replacement->sysfs_state);
8127                 } else if (rdev
8128                     && rdev->recovery_offset == MaxSector
8129                     && !test_bit(Faulty, &rdev->flags)
8130                     && !test_and_set_bit(In_sync, &rdev->flags)) {
8131                         count++;
8132                         sysfs_notify_dirent_safe(rdev->sysfs_state);
8133                 }
8134         }
8135         spin_lock_irqsave(&conf->device_lock, flags);
8136         mddev->degraded = raid5_calc_degraded(conf);
8137         spin_unlock_irqrestore(&conf->device_lock, flags);
8138         print_raid5_conf(conf);
8139         return count;
8140 }
8141
8142 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
8143 {
8144         struct r5conf *conf = mddev->private;
8145         int err = 0;
8146         int number = rdev->raid_disk;
8147         struct md_rdev **rdevp;
8148         struct disk_info *p;
8149         struct md_rdev *tmp;
8150
8151         print_raid5_conf(conf);
8152         if (test_bit(Journal, &rdev->flags) && conf->log) {
8153                 /*
8154                  * we can't wait pending write here, as this is called in
8155                  * raid5d, wait will deadlock.
8156                  * neilb: there is no locking about new writes here,
8157                  * so this cannot be safe.
8158                  */
8159                 if (atomic_read(&conf->active_stripes) ||
8160                     atomic_read(&conf->r5c_cached_full_stripes) ||
8161                     atomic_read(&conf->r5c_cached_partial_stripes)) {
8162                         return -EBUSY;
8163                 }
8164                 log_exit(conf);
8165                 return 0;
8166         }
8167         if (unlikely(number >= conf->pool_size))
8168                 return 0;
8169         p = conf->disks + number;
8170         if (rdev == p->rdev)
8171                 rdevp = &p->rdev;
8172         else if (rdev == p->replacement)
8173                 rdevp = &p->replacement;
8174         else
8175                 return 0;
8176
8177         if (number >= conf->raid_disks &&
8178             conf->reshape_progress == MaxSector)
8179                 clear_bit(In_sync, &rdev->flags);
8180
8181         if (test_bit(In_sync, &rdev->flags) ||
8182             atomic_read(&rdev->nr_pending)) {
8183                 err = -EBUSY;
8184                 goto abort;
8185         }
8186         /* Only remove non-faulty devices if recovery
8187          * isn't possible.
8188          */
8189         if (!test_bit(Faulty, &rdev->flags) &&
8190             mddev->recovery_disabled != conf->recovery_disabled &&
8191             !has_failed(conf) &&
8192             (!p->replacement || p->replacement == rdev) &&
8193             number < conf->raid_disks) {
8194                 err = -EBUSY;
8195                 goto abort;
8196         }
8197         WRITE_ONCE(*rdevp, NULL);
8198         if (!err) {
8199                 err = log_modify(conf, rdev, false);
8200                 if (err)
8201                         goto abort;
8202         }
8203
8204         tmp = p->replacement;
8205         if (tmp) {
8206                 /* We must have just cleared 'rdev' */
8207                 WRITE_ONCE(p->rdev, tmp);
8208                 clear_bit(Replacement, &tmp->flags);
8209                 WRITE_ONCE(p->replacement, NULL);
8210
8211                 if (!err)
8212                         err = log_modify(conf, tmp, true);
8213         }
8214
8215         clear_bit(WantReplacement, &rdev->flags);
8216 abort:
8217
8218         print_raid5_conf(conf);
8219         return err;
8220 }
8221
8222 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
8223 {
8224         struct r5conf *conf = mddev->private;
8225         int ret, err = -EEXIST;
8226         int disk;
8227         struct disk_info *p;
8228         struct md_rdev *tmp;
8229         int first = 0;
8230         int last = conf->raid_disks - 1;
8231
8232         if (test_bit(Journal, &rdev->flags)) {
8233                 if (conf->log)
8234                         return -EBUSY;
8235
8236                 rdev->raid_disk = 0;
8237                 /*
8238                  * The array is in readonly mode if journal is missing, so no
8239                  * write requests running. We should be safe
8240                  */
8241                 ret = log_init(conf, rdev, false);
8242                 if (ret)
8243                         return ret;
8244
8245                 ret = r5l_start(conf->log);
8246                 if (ret)
8247                         return ret;
8248
8249                 return 0;
8250         }
8251         if (mddev->recovery_disabled == conf->recovery_disabled)
8252                 return -EBUSY;
8253
8254         if (rdev->saved_raid_disk < 0 && has_failed(conf))
8255                 /* no point adding a device */
8256                 return -EINVAL;
8257
8258         if (rdev->raid_disk >= 0)
8259                 first = last = rdev->raid_disk;
8260
8261         /*
8262          * find the disk ... but prefer rdev->saved_raid_disk
8263          * if possible.
8264          */
8265         if (rdev->saved_raid_disk >= first &&
8266             rdev->saved_raid_disk <= last &&
8267             conf->disks[rdev->saved_raid_disk].rdev == NULL)
8268                 first = rdev->saved_raid_disk;
8269
8270         for (disk = first; disk <= last; disk++) {
8271                 p = conf->disks + disk;
8272                 if (p->rdev == NULL) {
8273                         clear_bit(In_sync, &rdev->flags);
8274                         rdev->raid_disk = disk;
8275                         if (rdev->saved_raid_disk != disk)
8276                                 conf->fullsync = 1;
8277                         WRITE_ONCE(p->rdev, rdev);
8278
8279                         err = log_modify(conf, rdev, true);
8280
8281                         goto out;
8282                 }
8283         }
8284         for (disk = first; disk <= last; disk++) {
8285                 p = conf->disks + disk;
8286                 tmp = p->rdev;
8287                 if (test_bit(WantReplacement, &tmp->flags) &&
8288                     mddev->reshape_position == MaxSector &&
8289                     p->replacement == NULL) {
8290                         clear_bit(In_sync, &rdev->flags);
8291                         set_bit(Replacement, &rdev->flags);
8292                         rdev->raid_disk = disk;
8293                         err = 0;
8294                         conf->fullsync = 1;
8295                         WRITE_ONCE(p->replacement, rdev);
8296                         break;
8297                 }
8298         }
8299 out:
8300         print_raid5_conf(conf);
8301         return err;
8302 }
8303
8304 static int raid5_resize(struct mddev *mddev, sector_t sectors)
8305 {
8306         /* no resync is happening, and there is enough space
8307          * on all devices, so we can resize.
8308          * We need to make sure resync covers any new space.
8309          * If the array is shrinking we should possibly wait until
8310          * any io in the removed space completes, but it hardly seems
8311          * worth it.
8312          */
8313         sector_t newsize;
8314         struct r5conf *conf = mddev->private;
8315         int ret;
8316
8317         if (raid5_has_log(conf) || raid5_has_ppl(conf))
8318                 return -EINVAL;
8319         sectors &= ~((sector_t)conf->chunk_sectors - 1);
8320         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
8321         if (mddev->external_size &&
8322             mddev->array_sectors > newsize)
8323                 return -EINVAL;
8324
8325         ret = mddev->bitmap_ops->resize(mddev, sectors, 0, false);
8326         if (ret)
8327                 return ret;
8328
8329         md_set_array_sectors(mddev, newsize);
8330         if (sectors > mddev->dev_sectors &&
8331             mddev->recovery_cp > mddev->dev_sectors) {
8332                 mddev->recovery_cp = mddev->dev_sectors;
8333                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8334         }
8335         mddev->dev_sectors = sectors;
8336         mddev->resync_max_sectors = sectors;
8337         return 0;
8338 }
8339
8340 static int check_stripe_cache(struct mddev *mddev)
8341 {
8342         /* Can only proceed if there are plenty of stripe_heads.
8343          * We need a minimum of one full stripe,, and for sensible progress
8344          * it is best to have about 4 times that.
8345          * If we require 4 times, then the default 256 4K stripe_heads will
8346          * allow for chunk sizes up to 256K, which is probably OK.
8347          * If the chunk size is greater, user-space should request more
8348          * stripe_heads first.
8349          */
8350         struct r5conf *conf = mddev->private;
8351         if (((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8352             > conf->min_nr_stripes ||
8353             ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8354             > conf->min_nr_stripes) {
8355                 pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
8356                         mdname(mddev),
8357                         ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
8358                          / RAID5_STRIPE_SIZE(conf))*4);
8359                 return 0;
8360         }
8361         return 1;
8362 }
8363
8364 static int check_reshape(struct mddev *mddev)
8365 {
8366         struct r5conf *conf = mddev->private;
8367
8368         if (raid5_has_log(conf) || raid5_has_ppl(conf))
8369                 return -EINVAL;
8370         if (mddev->delta_disks == 0 &&
8371             mddev->new_layout == mddev->layout &&
8372             mddev->new_chunk_sectors == mddev->chunk_sectors)
8373                 return 0; /* nothing to do */
8374         if (has_failed(conf))
8375                 return -EINVAL;
8376         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
8377                 /* We might be able to shrink, but the devices must
8378                  * be made bigger first.
8379                  * For raid6, 4 is the minimum size.
8380                  * Otherwise 2 is the minimum
8381                  */
8382                 int min = 2;
8383                 if (mddev->level == 6)
8384                         min = 4;
8385                 if (mddev->raid_disks + mddev->delta_disks < min)
8386                         return -EINVAL;
8387         }
8388
8389         if (!check_stripe_cache(mddev))
8390                 return -ENOSPC;
8391
8392         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
8393             mddev->delta_disks > 0)
8394                 if (resize_chunks(conf,
8395                                   conf->previous_raid_disks
8396                                   + max(0, mddev->delta_disks),
8397                                   max(mddev->new_chunk_sectors,
8398                                       mddev->chunk_sectors)
8399                             ) < 0)
8400                         return -ENOMEM;
8401
8402         if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
8403                 return 0; /* never bother to shrink */
8404         return resize_stripes(conf, (conf->previous_raid_disks
8405                                      + mddev->delta_disks));
8406 }
8407
8408 static int raid5_start_reshape(struct mddev *mddev)
8409 {
8410         struct r5conf *conf = mddev->private;
8411         struct md_rdev *rdev;
8412         int spares = 0;
8413         int i;
8414         unsigned long flags;
8415
8416         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
8417                 return -EBUSY;
8418
8419         if (!check_stripe_cache(mddev))
8420                 return -ENOSPC;
8421
8422         if (has_failed(conf))
8423                 return -EINVAL;
8424
8425         /* raid5 can't handle concurrent reshape and recovery */
8426         if (mddev->recovery_cp < MaxSector)
8427                 return -EBUSY;
8428         for (i = 0; i < conf->raid_disks; i++)
8429                 if (conf->disks[i].replacement)
8430                         return -EBUSY;
8431
8432         rdev_for_each(rdev, mddev) {
8433                 if (!test_bit(In_sync, &rdev->flags)
8434                     && !test_bit(Faulty, &rdev->flags))
8435                         spares++;
8436         }
8437
8438         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
8439                 /* Not enough devices even to make a degraded array
8440                  * of that size
8441                  */
8442                 return -EINVAL;
8443
8444         /* Refuse to reduce size of the array.  Any reductions in
8445          * array size must be through explicit setting of array_size
8446          * attribute.
8447          */
8448         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
8449             < mddev->array_sectors) {
8450                 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
8451                         mdname(mddev));
8452                 return -EINVAL;
8453         }
8454
8455         atomic_set(&conf->reshape_stripes, 0);
8456         spin_lock_irq(&conf->device_lock);
8457         write_seqcount_begin(&conf->gen_lock);
8458         conf->previous_raid_disks = conf->raid_disks;
8459         conf->raid_disks += mddev->delta_disks;
8460         conf->prev_chunk_sectors = conf->chunk_sectors;
8461         conf->chunk_sectors = mddev->new_chunk_sectors;
8462         conf->prev_algo = conf->algorithm;
8463         conf->algorithm = mddev->new_layout;
8464         conf->generation++;
8465         /* Code that selects data_offset needs to see the generation update
8466          * if reshape_progress has been set - so a memory barrier needed.
8467          */
8468         smp_mb();
8469         if (mddev->reshape_backwards)
8470                 conf->reshape_progress = raid5_size(mddev, 0, 0);
8471         else
8472                 conf->reshape_progress = 0;
8473         conf->reshape_safe = conf->reshape_progress;
8474         write_seqcount_end(&conf->gen_lock);
8475         spin_unlock_irq(&conf->device_lock);
8476
8477         /* Now make sure any requests that proceeded on the assumption
8478          * the reshape wasn't running - like Discard or Read - have
8479          * completed.
8480          */
8481         raid5_quiesce(mddev, true);
8482         raid5_quiesce(mddev, false);
8483
8484         /* Add some new drives, as many as will fit.
8485          * We know there are enough to make the newly sized array work.
8486          * Don't add devices if we are reducing the number of
8487          * devices in the array.  This is because it is not possible
8488          * to correctly record the "partially reconstructed" state of
8489          * such devices during the reshape and confusion could result.
8490          */
8491         if (mddev->delta_disks >= 0) {
8492                 rdev_for_each(rdev, mddev)
8493                         if (rdev->raid_disk < 0 &&
8494                             !test_bit(Faulty, &rdev->flags)) {
8495                                 if (raid5_add_disk(mddev, rdev) == 0) {
8496                                         if (rdev->raid_disk
8497                                             >= conf->previous_raid_disks)
8498                                                 set_bit(In_sync, &rdev->flags);
8499                                         else
8500                                                 rdev->recovery_offset = 0;
8501
8502                                         /* Failure here is OK */
8503                                         sysfs_link_rdev(mddev, rdev);
8504                                 }
8505                         } else if (rdev->raid_disk >= conf->previous_raid_disks
8506                                    && !test_bit(Faulty, &rdev->flags)) {
8507                                 /* This is a spare that was manually added */
8508                                 set_bit(In_sync, &rdev->flags);
8509                         }
8510
8511                 /* When a reshape changes the number of devices,
8512                  * ->degraded is measured against the larger of the
8513                  * pre and post number of devices.
8514                  */
8515                 spin_lock_irqsave(&conf->device_lock, flags);
8516                 mddev->degraded = raid5_calc_degraded(conf);
8517                 spin_unlock_irqrestore(&conf->device_lock, flags);
8518         }
8519         mddev->raid_disks = conf->raid_disks;
8520         mddev->reshape_position = conf->reshape_progress;
8521         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8522
8523         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8524         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
8525         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
8526         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
8527         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8528         conf->reshape_checkpoint = jiffies;
8529         md_new_event();
8530         return 0;
8531 }
8532
8533 /* This is called from the reshape thread and should make any
8534  * changes needed in 'conf'
8535  */
8536 static void end_reshape(struct r5conf *conf)
8537 {
8538
8539         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
8540                 struct md_rdev *rdev;
8541
8542                 spin_lock_irq(&conf->device_lock);
8543                 conf->previous_raid_disks = conf->raid_disks;
8544                 md_finish_reshape(conf->mddev);
8545                 smp_wmb();
8546                 conf->reshape_progress = MaxSector;
8547                 conf->mddev->reshape_position = MaxSector;
8548                 rdev_for_each(rdev, conf->mddev)
8549                         if (rdev->raid_disk >= 0 &&
8550                             !test_bit(Journal, &rdev->flags) &&
8551                             !test_bit(In_sync, &rdev->flags))
8552                                 rdev->recovery_offset = MaxSector;
8553                 spin_unlock_irq(&conf->device_lock);
8554                 wake_up(&conf->wait_for_reshape);
8555
8556                 mddev_update_io_opt(conf->mddev,
8557                         conf->raid_disks - conf->max_degraded);
8558         }
8559 }
8560
8561 /* This is called from the raid5d thread with mddev_lock held.
8562  * It makes config changes to the device.
8563  */
8564 static void raid5_finish_reshape(struct mddev *mddev)
8565 {
8566         struct r5conf *conf = mddev->private;
8567         struct md_rdev *rdev;
8568
8569         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8570
8571                 if (mddev->delta_disks <= 0) {
8572                         int d;
8573                         spin_lock_irq(&conf->device_lock);
8574                         mddev->degraded = raid5_calc_degraded(conf);
8575                         spin_unlock_irq(&conf->device_lock);
8576                         for (d = conf->raid_disks ;
8577                              d < conf->raid_disks - mddev->delta_disks;
8578                              d++) {
8579                                 rdev = conf->disks[d].rdev;
8580                                 if (rdev)
8581                                         clear_bit(In_sync, &rdev->flags);
8582                                 rdev = conf->disks[d].replacement;
8583                                 if (rdev)
8584                                         clear_bit(In_sync, &rdev->flags);
8585                         }
8586                 }
8587                 mddev->layout = conf->algorithm;
8588                 mddev->chunk_sectors = conf->chunk_sectors;
8589                 mddev->reshape_position = MaxSector;
8590                 mddev->delta_disks = 0;
8591                 mddev->reshape_backwards = 0;
8592         }
8593 }
8594
8595 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8596 {
8597         struct r5conf *conf = mddev->private;
8598
8599         if (quiesce) {
8600                 /* stop all writes */
8601                 lock_all_device_hash_locks_irq(conf);
8602                 /* '2' tells resync/reshape to pause so that all
8603                  * active stripes can drain
8604                  */
8605                 r5c_flush_cache(conf, INT_MAX);
8606                 /* need a memory barrier to make sure read_one_chunk() sees
8607                  * quiesce started and reverts to slow (locked) path.
8608                  */
8609                 smp_store_release(&conf->quiesce, 2);
8610                 wait_event_cmd(conf->wait_for_quiescent,
8611                                     atomic_read(&conf->active_stripes) == 0 &&
8612                                     atomic_read(&conf->active_aligned_reads) == 0,
8613                                     unlock_all_device_hash_locks_irq(conf),
8614                                     lock_all_device_hash_locks_irq(conf));
8615                 conf->quiesce = 1;
8616                 unlock_all_device_hash_locks_irq(conf);
8617                 /* allow reshape to continue */
8618                 wake_up(&conf->wait_for_reshape);
8619         } else {
8620                 /* re-enable writes */
8621                 lock_all_device_hash_locks_irq(conf);
8622                 conf->quiesce = 0;
8623                 wake_up(&conf->wait_for_quiescent);
8624                 wake_up(&conf->wait_for_reshape);
8625                 unlock_all_device_hash_locks_irq(conf);
8626         }
8627         log_quiesce(conf, quiesce);
8628 }
8629
8630 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8631 {
8632         struct r0conf *raid0_conf = mddev->private;
8633         sector_t sectors;
8634
8635         /* for raid0 takeover only one zone is supported */
8636         if (raid0_conf->nr_strip_zones > 1) {
8637                 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8638                         mdname(mddev));
8639                 return ERR_PTR(-EINVAL);
8640         }
8641
8642         sectors = raid0_conf->strip_zone[0].zone_end;
8643         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8644         mddev->dev_sectors = sectors;
8645         mddev->new_level = level;
8646         mddev->new_layout = ALGORITHM_PARITY_N;
8647         mddev->new_chunk_sectors = mddev->chunk_sectors;
8648         mddev->raid_disks += 1;
8649         mddev->delta_disks = 1;
8650         /* make sure it will be not marked as dirty */
8651         mddev->recovery_cp = MaxSector;
8652
8653         return setup_conf(mddev);
8654 }
8655
8656 static void *raid5_takeover_raid1(struct mddev *mddev)
8657 {
8658         int chunksect;
8659         void *ret;
8660
8661         if (mddev->raid_disks != 2 ||
8662             mddev->degraded > 1)
8663                 return ERR_PTR(-EINVAL);
8664
8665         /* Should check if there are write-behind devices? */
8666
8667         chunksect = 64*2; /* 64K by default */
8668
8669         /* The array must be an exact multiple of chunksize */
8670         while (chunksect && (mddev->array_sectors & (chunksect-1)))
8671                 chunksect >>= 1;
8672
8673         if ((chunksect<<9) < RAID5_STRIPE_SIZE((struct r5conf *)mddev->private))
8674                 /* array size does not allow a suitable chunk size */
8675                 return ERR_PTR(-EINVAL);
8676
8677         mddev->new_level = 5;
8678         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8679         mddev->new_chunk_sectors = chunksect;
8680
8681         ret = setup_conf(mddev);
8682         if (!IS_ERR(ret))
8683                 mddev_clear_unsupported_flags(mddev,
8684                         UNSUPPORTED_MDDEV_FLAGS);
8685         return ret;
8686 }
8687
8688 static void *raid5_takeover_raid6(struct mddev *mddev)
8689 {
8690         int new_layout;
8691
8692         switch (mddev->layout) {
8693         case ALGORITHM_LEFT_ASYMMETRIC_6:
8694                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8695                 break;
8696         case ALGORITHM_RIGHT_ASYMMETRIC_6:
8697                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8698                 break;
8699         case ALGORITHM_LEFT_SYMMETRIC_6:
8700                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8701                 break;
8702         case ALGORITHM_RIGHT_SYMMETRIC_6:
8703                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8704                 break;
8705         case ALGORITHM_PARITY_0_6:
8706                 new_layout = ALGORITHM_PARITY_0;
8707                 break;
8708         case ALGORITHM_PARITY_N:
8709                 new_layout = ALGORITHM_PARITY_N;
8710                 break;
8711         default:
8712                 return ERR_PTR(-EINVAL);
8713         }
8714         mddev->new_level = 5;
8715         mddev->new_layout = new_layout;
8716         mddev->delta_disks = -1;
8717         mddev->raid_disks -= 1;
8718         return setup_conf(mddev);
8719 }
8720
8721 static int raid5_check_reshape(struct mddev *mddev)
8722 {
8723         /* For a 2-drive array, the layout and chunk size can be changed
8724          * immediately as not restriping is needed.
8725          * For larger arrays we record the new value - after validation
8726          * to be used by a reshape pass.
8727          */
8728         struct r5conf *conf = mddev->private;
8729         int new_chunk = mddev->new_chunk_sectors;
8730
8731         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8732                 return -EINVAL;
8733         if (new_chunk > 0) {
8734                 if (!is_power_of_2(new_chunk))
8735                         return -EINVAL;
8736                 if (new_chunk < (PAGE_SIZE>>9))
8737                         return -EINVAL;
8738                 if (mddev->array_sectors & (new_chunk-1))
8739                         /* not factor of array size */
8740                         return -EINVAL;
8741         }
8742
8743         /* They look valid */
8744
8745         if (mddev->raid_disks == 2) {
8746                 /* can make the change immediately */
8747                 if (mddev->new_layout >= 0) {
8748                         conf->algorithm = mddev->new_layout;
8749                         mddev->layout = mddev->new_layout;
8750                 }
8751                 if (new_chunk > 0) {
8752                         conf->chunk_sectors = new_chunk ;
8753                         mddev->chunk_sectors = new_chunk;
8754                 }
8755                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8756                 md_wakeup_thread(mddev->thread);
8757         }
8758         return check_reshape(mddev);
8759 }
8760
8761 static int raid6_check_reshape(struct mddev *mddev)
8762 {
8763         int new_chunk = mddev->new_chunk_sectors;
8764
8765         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8766                 return -EINVAL;
8767         if (new_chunk > 0) {
8768                 if (!is_power_of_2(new_chunk))
8769                         return -EINVAL;
8770                 if (new_chunk < (PAGE_SIZE >> 9))
8771                         return -EINVAL;
8772                 if (mddev->array_sectors & (new_chunk-1))
8773                         /* not factor of array size */
8774                         return -EINVAL;
8775         }
8776
8777         /* They look valid */
8778         return check_reshape(mddev);
8779 }
8780
8781 static void *raid5_takeover(struct mddev *mddev)
8782 {
8783         /* raid5 can take over:
8784          *  raid0 - if there is only one strip zone - make it a raid4 layout
8785          *  raid1 - if there are two drives.  We need to know the chunk size
8786          *  raid4 - trivial - just use a raid4 layout.
8787          *  raid6 - Providing it is a *_6 layout
8788          */
8789         if (mddev->level == 0)
8790                 return raid45_takeover_raid0(mddev, 5);
8791         if (mddev->level == 1)
8792                 return raid5_takeover_raid1(mddev);
8793         if (mddev->level == 4) {
8794                 mddev->new_layout = ALGORITHM_PARITY_N;
8795                 mddev->new_level = 5;
8796                 return setup_conf(mddev);
8797         }
8798         if (mddev->level == 6)
8799                 return raid5_takeover_raid6(mddev);
8800
8801         return ERR_PTR(-EINVAL);
8802 }
8803
8804 static void *raid4_takeover(struct mddev *mddev)
8805 {
8806         /* raid4 can take over:
8807          *  raid0 - if there is only one strip zone
8808          *  raid5 - if layout is right
8809          */
8810         if (mddev->level == 0)
8811                 return raid45_takeover_raid0(mddev, 4);
8812         if (mddev->level == 5 &&
8813             mddev->layout == ALGORITHM_PARITY_N) {
8814                 mddev->new_layout = 0;
8815                 mddev->new_level = 4;
8816                 return setup_conf(mddev);
8817         }
8818         return ERR_PTR(-EINVAL);
8819 }
8820
8821 static struct md_personality raid5_personality;
8822
8823 static void *raid6_takeover(struct mddev *mddev)
8824 {
8825         /* Currently can only take over a raid5.  We map the
8826          * personality to an equivalent raid6 personality
8827          * with the Q block at the end.
8828          */
8829         int new_layout;
8830
8831         if (mddev->pers != &raid5_personality)
8832                 return ERR_PTR(-EINVAL);
8833         if (mddev->degraded > 1)
8834                 return ERR_PTR(-EINVAL);
8835         if (mddev->raid_disks > 253)
8836                 return ERR_PTR(-EINVAL);
8837         if (mddev->raid_disks < 3)
8838                 return ERR_PTR(-EINVAL);
8839
8840         switch (mddev->layout) {
8841         case ALGORITHM_LEFT_ASYMMETRIC:
8842                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8843                 break;
8844         case ALGORITHM_RIGHT_ASYMMETRIC:
8845                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8846                 break;
8847         case ALGORITHM_LEFT_SYMMETRIC:
8848                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8849                 break;
8850         case ALGORITHM_RIGHT_SYMMETRIC:
8851                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8852                 break;
8853         case ALGORITHM_PARITY_0:
8854                 new_layout = ALGORITHM_PARITY_0_6;
8855                 break;
8856         case ALGORITHM_PARITY_N:
8857                 new_layout = ALGORITHM_PARITY_N;
8858                 break;
8859         default:
8860                 return ERR_PTR(-EINVAL);
8861         }
8862         mddev->new_level = 6;
8863         mddev->new_layout = new_layout;
8864         mddev->delta_disks = 1;
8865         mddev->raid_disks += 1;
8866         return setup_conf(mddev);
8867 }
8868
8869 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8870 {
8871         struct r5conf *conf;
8872         int err;
8873
8874         err = mddev_suspend_and_lock(mddev);
8875         if (err)
8876                 return err;
8877         conf = mddev->private;
8878         if (!conf) {
8879                 mddev_unlock_and_resume(mddev);
8880                 return -ENODEV;
8881         }
8882
8883         if (strncmp(buf, "ppl", 3) == 0) {
8884                 /* ppl only works with RAID 5 */
8885                 if (!raid5_has_ppl(conf) && conf->level == 5) {
8886                         err = log_init(conf, NULL, true);
8887                         if (!err) {
8888                                 err = resize_stripes(conf, conf->pool_size);
8889                                 if (err)
8890                                         log_exit(conf);
8891                         }
8892                 } else
8893                         err = -EINVAL;
8894         } else if (strncmp(buf, "resync", 6) == 0) {
8895                 if (raid5_has_ppl(conf)) {
8896                         log_exit(conf);
8897                         err = resize_stripes(conf, conf->pool_size);
8898                 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8899                            r5l_log_disk_error(conf)) {
8900                         bool journal_dev_exists = false;
8901                         struct md_rdev *rdev;
8902
8903                         rdev_for_each(rdev, mddev)
8904                                 if (test_bit(Journal, &rdev->flags)) {
8905                                         journal_dev_exists = true;
8906                                         break;
8907                                 }
8908
8909                         if (!journal_dev_exists)
8910                                 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8911                         else  /* need remove journal device first */
8912                                 err = -EBUSY;
8913                 } else
8914                         err = -EINVAL;
8915         } else {
8916                 err = -EINVAL;
8917         }
8918
8919         if (!err)
8920                 md_update_sb(mddev, 1);
8921
8922         mddev_unlock_and_resume(mddev);
8923
8924         return err;
8925 }
8926
8927 static int raid5_start(struct mddev *mddev)
8928 {
8929         struct r5conf *conf = mddev->private;
8930
8931         return r5l_start(conf->log);
8932 }
8933
8934 /*
8935  * This is only used for dm-raid456, caller already frozen sync_thread, hence
8936  * if rehsape is still in progress, io that is waiting for reshape can never be
8937  * done now, hence wake up and handle those IO.
8938  */
8939 static void raid5_prepare_suspend(struct mddev *mddev)
8940 {
8941         struct r5conf *conf = mddev->private;
8942
8943         wake_up(&conf->wait_for_reshape);
8944 }
8945
8946 static struct md_personality raid6_personality =
8947 {
8948         .head = {
8949                 .type   = MD_PERSONALITY,
8950                 .id     = ID_RAID6,
8951                 .name   = "raid6",
8952                 .owner  = THIS_MODULE,
8953         },
8954
8955         .make_request   = raid5_make_request,
8956         .run            = raid5_run,
8957         .start          = raid5_start,
8958         .free           = raid5_free,
8959         .status         = raid5_status,
8960         .error_handler  = raid5_error,
8961         .hot_add_disk   = raid5_add_disk,
8962         .hot_remove_disk= raid5_remove_disk,
8963         .spare_active   = raid5_spare_active,
8964         .sync_request   = raid5_sync_request,
8965         .resize         = raid5_resize,
8966         .size           = raid5_size,
8967         .check_reshape  = raid6_check_reshape,
8968         .start_reshape  = raid5_start_reshape,
8969         .finish_reshape = raid5_finish_reshape,
8970         .quiesce        = raid5_quiesce,
8971         .takeover       = raid6_takeover,
8972         .change_consistency_policy = raid5_change_consistency_policy,
8973         .prepare_suspend = raid5_prepare_suspend,
8974         .bitmap_sector  = raid5_bitmap_sector,
8975 };
8976 static struct md_personality raid5_personality =
8977 {
8978         .head = {
8979                 .type   = MD_PERSONALITY,
8980                 .id     = ID_RAID5,
8981                 .name   = "raid5",
8982                 .owner  = THIS_MODULE,
8983         },
8984
8985         .make_request   = raid5_make_request,
8986         .run            = raid5_run,
8987         .start          = raid5_start,
8988         .free           = raid5_free,
8989         .status         = raid5_status,
8990         .error_handler  = raid5_error,
8991         .hot_add_disk   = raid5_add_disk,
8992         .hot_remove_disk= raid5_remove_disk,
8993         .spare_active   = raid5_spare_active,
8994         .sync_request   = raid5_sync_request,
8995         .resize         = raid5_resize,
8996         .size           = raid5_size,
8997         .check_reshape  = raid5_check_reshape,
8998         .start_reshape  = raid5_start_reshape,
8999         .finish_reshape = raid5_finish_reshape,
9000         .quiesce        = raid5_quiesce,
9001         .takeover       = raid5_takeover,
9002         .change_consistency_policy = raid5_change_consistency_policy,
9003         .prepare_suspend = raid5_prepare_suspend,
9004         .bitmap_sector  = raid5_bitmap_sector,
9005 };
9006
9007 static struct md_personality raid4_personality =
9008 {
9009         .head = {
9010                 .type   = MD_PERSONALITY,
9011                 .id     = ID_RAID4,
9012                 .name   = "raid4",
9013                 .owner  = THIS_MODULE,
9014         },
9015
9016         .make_request   = raid5_make_request,
9017         .run            = raid5_run,
9018         .start          = raid5_start,
9019         .free           = raid5_free,
9020         .status         = raid5_status,
9021         .error_handler  = raid5_error,
9022         .hot_add_disk   = raid5_add_disk,
9023         .hot_remove_disk= raid5_remove_disk,
9024         .spare_active   = raid5_spare_active,
9025         .sync_request   = raid5_sync_request,
9026         .resize         = raid5_resize,
9027         .size           = raid5_size,
9028         .check_reshape  = raid5_check_reshape,
9029         .start_reshape  = raid5_start_reshape,
9030         .finish_reshape = raid5_finish_reshape,
9031         .quiesce        = raid5_quiesce,
9032         .takeover       = raid4_takeover,
9033         .change_consistency_policy = raid5_change_consistency_policy,
9034         .prepare_suspend = raid5_prepare_suspend,
9035         .bitmap_sector  = raid5_bitmap_sector,
9036 };
9037
9038 static int __init raid5_init(void)
9039 {
9040         int ret;
9041
9042         raid5_wq = alloc_workqueue("raid5wq",
9043                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
9044         if (!raid5_wq)
9045                 return -ENOMEM;
9046
9047         ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
9048                                       "md/raid5:prepare",
9049                                       raid456_cpu_up_prepare,
9050                                       raid456_cpu_dead);
9051         if (ret)
9052                 goto err_destroy_wq;
9053
9054         ret = register_md_submodule(&raid6_personality.head);
9055         if (ret)
9056                 goto err_cpuhp_remove;
9057
9058         ret = register_md_submodule(&raid5_personality.head);
9059         if (ret)
9060                 goto err_unregister_raid6;
9061
9062         ret = register_md_submodule(&raid4_personality.head);
9063         if (ret)
9064                 goto err_unregister_raid5;
9065
9066         return 0;
9067
9068 err_unregister_raid5:
9069         unregister_md_submodule(&raid5_personality.head);
9070 err_unregister_raid6:
9071         unregister_md_submodule(&raid6_personality.head);
9072 err_cpuhp_remove:
9073         cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
9074 err_destroy_wq:
9075         destroy_workqueue(raid5_wq);
9076         return ret;
9077 }
9078
9079 static void __exit raid5_exit(void)
9080 {
9081         unregister_md_submodule(&raid6_personality.head);
9082         unregister_md_submodule(&raid5_personality.head);
9083         unregister_md_submodule(&raid4_personality.head);
9084         cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
9085         destroy_workqueue(raid5_wq);
9086 }
9087
9088 module_init(raid5_init);
9089 module_exit(raid5_exit);
9090 MODULE_LICENSE("GPL");
9091 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
9092 MODULE_ALIAS("md-personality-4"); /* RAID5 */
9093 MODULE_ALIAS("md-raid5");
9094 MODULE_ALIAS("md-raid4");
9095 MODULE_ALIAS("md-level-5");
9096 MODULE_ALIAS("md-level-4");
9097 MODULE_ALIAS("md-personality-8"); /* RAID6 */
9098 MODULE_ALIAS("md-raid6");
9099 MODULE_ALIAS("md-level-6");
9100
9101 /* This used to be two separate modules, they were: */
9102 MODULE_ALIAS("raid5");
9103 MODULE_ALIAS("raid6");