drivers/md/raid5-cache.c

   1 /*
   2  * Copyright (C) 2015 Shaohua Li <shli@fb.com>
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms and conditions of the GNU General Public License,
   6  * version 2, as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  *
  13  */
  14 #include <linux/kernel.h>
  15 #include <linux/wait.h>
  16 #include <linux/blkdev.h>
  17 #include <linux/slab.h>
  18 #include <linux/raid/md_p.h>
  19 #include <linux/crc32c.h>
  20 #include <linux/random.h>
  21 #include "md.h"
  22 #include "raid5.h"
  23
  24 /*
  25  * metadata/data stored in disk with 4k size unit (a block) regardless
  26  * underneath hardware sector size. only works with PAGE_SIZE == 4096
  27  */
  28 #define BLOCK_SECTORS (8)
  29
  30 /*
  31  * reclaim runs every 1/4 disk size or 10G reclaimable space. This can prevent
  32  * recovery scans a very long log
  33  */
  34 #define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
  35 #define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
  36
  37 struct r5l_log {
  38         struct md_rdev *rdev;
  39
  40         u32 uuid_checksum;
  41
  42         sector_t device_size;           /* log device size, round to
  43                                          * BLOCK_SECTORS */
  44         sector_t max_free_space;        /* reclaim run if free space is at
  45                                          * this size */
  46
  47         sector_t last_checkpoint;       /* log tail. where recovery scan
  48                                          * starts from */
  49         u64 last_cp_seq;                /* log tail sequence */
  50
  51         sector_t log_start;             /* log head. where new data appends */
  52         u64 seq;                        /* log head sequence */
  53
  54         sector_t next_checkpoint;
  55         u64 next_cp_seq;
  56
  57         struct mutex io_mutex;
  58         struct r5l_io_unit *current_io; /* current io_unit accepting new data */
  59
  60         spinlock_t io_list_lock;
  61         struct list_head running_ios;   /* io_units which are still running,
  62                                          * and have not yet been completely
  63                                          * written to the log */
  64         struct list_head io_end_ios;    /* io_units which have been completely
  65                                          * written to the log but not yet written
  66                                          * to the RAID */
  67         struct list_head flushing_ios;  /* io_units which are waiting for log
  68                                          * cache flush */
  69         struct list_head flushed_ios;   /* io_units which settle down in log disk */
  70         struct bio flush_bio;
  71
  72         struct kmem_cache *io_kc;
  73
  74         struct md_thread *reclaim_thread;
  75         unsigned long reclaim_target;   /* number of space that need to be
  76                                          * reclaimed.  if it's 0, reclaim spaces
  77                                          * used by io_units which are in
  78                                          * IO_UNIT_STRIPE_END state (eg, reclaim
  79                                          * dones't wait for specific io_unit
  80                                          * switching to IO_UNIT_STRIPE_END
  81                                          * state) */
  82         wait_queue_head_t iounit_wait;
  83
  84         struct list_head no_space_stripes; /* pending stripes, log has no space */
  85         spinlock_t no_space_stripes_lock;
  86 };
  87
  88 /*
  89  * an IO range starts from a meta data block and end at the next meta data
  90  * block. The io unit's the meta data block tracks data/parity followed it. io
  91  * unit is written to log disk with normal write, as we always flush log disk
  92  * first and then start move data to raid disks, there is no requirement to
  93  * write io unit with FLUSH/FUA
  94  */
  95 struct r5l_io_unit {
  96         struct r5l_log *log;
  97
  98         struct page *meta_page; /* store meta block */
  99         int meta_offset;        /* current offset in meta_page */
 100
 101         struct bio_list bios;
 102         atomic_t pending_io;    /* pending bios not written to log yet */
 103         struct bio *current_bio;/* current_bio accepting new data */
 104
 105         atomic_t pending_stripe;/* how many stripes not flushed to raid */
 106         u64 seq;                /* seq number of the metablock */
 107         sector_t log_start;     /* where the io_unit starts */
 108         sector_t log_end;       /* where the io_unit ends */
 109         struct list_head log_sibling; /* log->running_ios */
 110         struct list_head stripe_list; /* stripes added to the io_unit */
 111
 112         int state;
 113 };
 114
 115 /* r5l_io_unit state */
 116 enum r5l_io_unit_state {
 117         IO_UNIT_RUNNING = 0,    /* accepting new IO */
 118         IO_UNIT_IO_START = 1,   /* io_unit bio start writing to log,
 119                                  * don't accepting new bio */
 120         IO_UNIT_IO_END = 2,     /* io_unit bio finish writing to log */
 121         IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */
 122 };
 123
 124 static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
 125 {
 126         start += inc;
 127         if (start >= log->device_size)
 128                 start = start - log->device_size;
 129         return start;
 130 }
 131
 132 static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
 133                                   sector_t end)
 134 {
 135         if (end >= start)
 136                 return end - start;
 137         else
 138                 return end + log->device_size - start;
 139 }
 140
 141 static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
 142 {
 143         sector_t used_size;
 144
 145         used_size = r5l_ring_distance(log, log->last_checkpoint,
 146                                         log->log_start);
 147
 148         return log->device_size > used_size + size;
 149 }
 150
 151 static struct r5l_io_unit *r5l_alloc_io_unit(struct r5l_log *log)
 152 {
 153         struct r5l_io_unit *io;
 154         /* We can't handle memory allocate failure so far */
 155         gfp_t gfp = GFP_NOIO | __GFP_NOFAIL;
 156
 157         io = kmem_cache_zalloc(log->io_kc, gfp);
 158         io->log = log;
 159         io->meta_page = alloc_page(gfp | __GFP_ZERO);
 160
 161         bio_list_init(&io->bios);
 162         INIT_LIST_HEAD(&io->log_sibling);
 163         INIT_LIST_HEAD(&io->stripe_list);
 164         io->state = IO_UNIT_RUNNING;
 165         return io;
 166 }
 167
 168 static void r5l_free_io_unit(struct r5l_log *log, struct r5l_io_unit *io)
 169 {
 170         __free_page(io->meta_page);
 171         kmem_cache_free(log->io_kc, io);
 172 }
 173
 174 static void r5l_move_io_unit_list(struct list_head *from, struct list_head *to,
 175                                   enum r5l_io_unit_state state)
 176 {
 177         struct r5l_io_unit *io;
 178
 179         while (!list_empty(from)) {
 180                 io = list_first_entry(from, struct r5l_io_unit, log_sibling);
 181                 /* don't change list order */
 182                 if (io->state >= state)
 183                         list_move_tail(&io->log_sibling, to);
 184                 else
 185                         break;
 186         }
 187 }
 188
 189 static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
 190                                     enum r5l_io_unit_state state)
 191 {
 192         if (WARN_ON(io->state >= state))
 193                 return;
 194         io->state = state;
 195 }
 196
 197 /* XXX: totally ignores I/O errors */
 198 static void r5l_log_endio(struct bio *bio)
 199 {
 200         struct r5l_io_unit *io = bio->bi_private;
 201         struct r5l_log *log = io->log;
 202         unsigned long flags;
 203
 204         bio_put(bio);
 205
 206         if (!atomic_dec_and_test(&io->pending_io))
 207                 return;
 208
 209         spin_lock_irqsave(&log->io_list_lock, flags);
 210         __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
 211         r5l_move_io_unit_list(&log->running_ios, &log->io_end_ios,
 212                         IO_UNIT_IO_END);
 213         spin_unlock_irqrestore(&log->io_list_lock, flags);
 214
 215         md_wakeup_thread(log->rdev->mddev->thread);
 216 }
 217
 218 static void r5l_submit_current_io(struct r5l_log *log)
 219 {
 220         struct r5l_io_unit *io = log->current_io;
 221         struct r5l_meta_block *block;
 222         struct bio *bio;
 223         unsigned long flags;
 224         u32 crc;
 225
 226         if (!io)
 227                 return;
 228
 229         block = page_address(io->meta_page);
 230         block->meta_size = cpu_to_le32(io->meta_offset);
 231         crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
 232         block->checksum = cpu_to_le32(crc);
 233
 234         log->current_io = NULL;
 235         spin_lock_irqsave(&log->io_list_lock, flags);
 236         __r5l_set_io_unit_state(io, IO_UNIT_IO_START);
 237         spin_unlock_irqrestore(&log->io_list_lock, flags);
 238
 239         while ((bio = bio_list_pop(&io->bios))) {
 240                 /* all IO must start from rdev->data_offset */
 241                 bio->bi_iter.bi_sector += log->rdev->data_offset;
 242                 submit_bio(WRITE, bio);
 243         }
 244 }
 245
 246 static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
 247 {
 248         struct r5l_io_unit *io;
 249         struct r5l_meta_block *block;
 250         struct bio *bio;
 251
 252         io = r5l_alloc_io_unit(log);
 253
 254         block = page_address(io->meta_page);
 255         block->magic = cpu_to_le32(R5LOG_MAGIC);
 256         block->version = R5LOG_VERSION;
 257         block->seq = cpu_to_le64(log->seq);
 258         block->position = cpu_to_le64(log->log_start);
 259
 260         io->log_start = log->log_start;
 261         io->meta_offset = sizeof(struct r5l_meta_block);
 262         io->seq = log->seq;
 263
 264         bio = bio_kmalloc(GFP_NOIO | __GFP_NOFAIL, BIO_MAX_PAGES);
 265         io->current_bio = bio;
 266         bio->bi_rw = WRITE;
 267         bio->bi_bdev = log->rdev->bdev;
 268         bio->bi_iter.bi_sector = log->log_start;
 269         bio_add_page(bio, io->meta_page, PAGE_SIZE, 0);
 270         bio->bi_end_io = r5l_log_endio;
 271         bio->bi_private = io;
 272
 273         bio_list_add(&io->bios, bio);
 274         atomic_inc(&io->pending_io);
 275
 276         log->seq++;
 277         log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
 278         io->log_end = log->log_start;
 279         /* current bio hit disk end */
 280         if (log->log_start == 0)
 281                 io->current_bio = NULL;
 282
 283         spin_lock_irq(&log->io_list_lock);
 284         list_add_tail(&io->log_sibling, &log->running_ios);
 285         spin_unlock_irq(&log->io_list_lock);
 286
 287         return io;
 288 }
 289
 290 static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
 291 {
 292         struct r5l_io_unit *io;
 293
 294         io = log->current_io;
 295         if (io && io->meta_offset + payload_size > PAGE_SIZE)
 296                 r5l_submit_current_io(log);
 297         io = log->current_io;
 298         if (io)
 299                 return 0;
 300
 301         log->current_io = r5l_new_meta(log);
 302         return 0;
 303 }
 304
 305 static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
 306                                     sector_t location,
 307                                     u32 checksum1, u32 checksum2,
 308                                     bool checksum2_valid)
 309 {
 310         struct r5l_io_unit *io = log->current_io;
 311         struct r5l_payload_data_parity *payload;
 312
 313         payload = page_address(io->meta_page) + io->meta_offset;
 314         payload->header.type = cpu_to_le16(type);
 315         payload->header.flags = cpu_to_le16(0);
 316         payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
 317                                     (PAGE_SHIFT - 9));
 318         payload->location = cpu_to_le64(location);
 319         payload->checksum[0] = cpu_to_le32(checksum1);
 320         if (checksum2_valid)
 321                 payload->checksum[1] = cpu_to_le32(checksum2);
 322
 323         io->meta_offset += sizeof(struct r5l_payload_data_parity) +
 324                 sizeof(__le32) * (1 + !!checksum2_valid);
 325 }
 326
 327 static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
 328 {
 329         struct r5l_io_unit *io = log->current_io;
 330
 331 alloc_bio:
 332         if (!io->current_bio) {
 333                 struct bio *bio;
 334
 335                 bio = bio_kmalloc(GFP_NOIO | __GFP_NOFAIL, BIO_MAX_PAGES);
 336                 bio->bi_rw = WRITE;
 337                 bio->bi_bdev = log->rdev->bdev;
 338                 bio->bi_iter.bi_sector = log->log_start;
 339                 bio->bi_end_io = r5l_log_endio;
 340                 bio->bi_private = io;
 341                 bio_list_add(&io->bios, bio);
 342                 atomic_inc(&io->pending_io);
 343                 io->current_bio = bio;
 344         }
 345         if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0)) {
 346                 io->current_bio = NULL;
 347                 goto alloc_bio;
 348         }
 349         log->log_start = r5l_ring_add(log, log->log_start,
 350                                       BLOCK_SECTORS);
 351         /* current bio hit disk end */
 352         if (log->log_start == 0)
 353                 io->current_bio = NULL;
 354
 355         io->log_end = log->log_start;
 356 }
 357
 358 static void r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
 359                            int data_pages, int parity_pages)
 360 {
 361         int i;
 362         int meta_size;
 363         struct r5l_io_unit *io;
 364
 365         meta_size =
 366                 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
 367                  * data_pages) +
 368                 sizeof(struct r5l_payload_data_parity) +
 369                 sizeof(__le32) * parity_pages;
 370
 371         r5l_get_meta(log, meta_size);
 372         io = log->current_io;
 373
 374         for (i = 0; i < sh->disks; i++) {
 375                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
 376                         continue;
 377                 if (i == sh->pd_idx || i == sh->qd_idx)
 378                         continue;
 379                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
 380                                         raid5_compute_blocknr(sh, i, 0),
 381                                         sh->dev[i].log_checksum, 0, false);
 382                 r5l_append_payload_page(log, sh->dev[i].page);
 383         }
 384
 385         if (sh->qd_idx >= 0) {
 386                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 387                                         sh->sector, sh->dev[sh->pd_idx].log_checksum,
 388                                         sh->dev[sh->qd_idx].log_checksum, true);
 389                 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 390                 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
 391         } else {
 392                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 393                                         sh->sector, sh->dev[sh->pd_idx].log_checksum,
 394                                         0, false);
 395                 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 396         }
 397
 398         list_add_tail(&sh->log_list, &io->stripe_list);
 399         atomic_inc(&io->pending_stripe);
 400         sh->log_io = io;
 401 }
 402
 403 static void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
 404 /*
 405  * running in raid5d, where reclaim could wait for raid5d too (when it flushes
 406  * data from log to raid disks), so we shouldn't wait for reclaim here
 407  */
 408 int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
 409 {
 410         int write_disks = 0;
 411         int data_pages, parity_pages;
 412         int meta_size;
 413         int reserve;
 414         int i;
 415
 416         if (!log)
 417                 return -EAGAIN;
 418         /* Don't support stripe batch */
 419         if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
 420             test_bit(STRIPE_SYNCING, &sh->state)) {
 421                 /* the stripe is written to log, we start writing it to raid */
 422                 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
 423                 return -EAGAIN;
 424         }
 425
 426         for (i = 0; i < sh->disks; i++) {
 427                 void *addr;
 428
 429                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
 430                         continue;
 431                 write_disks++;
 432                 /* checksum is already calculated in last run */
 433                 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
 434                         continue;
 435                 addr = kmap_atomic(sh->dev[i].page);
 436                 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
 437                                                     addr, PAGE_SIZE);
 438                 kunmap_atomic(addr);
 439         }
 440         parity_pages = 1 + !!(sh->qd_idx >= 0);
 441         data_pages = write_disks - parity_pages;
 442
 443         meta_size =
 444                 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
 445                  * data_pages) +
 446                 sizeof(struct r5l_payload_data_parity) +
 447                 sizeof(__le32) * parity_pages;
 448         /* Doesn't work with very big raid array */
 449         if (meta_size + sizeof(struct r5l_meta_block) > PAGE_SIZE)
 450                 return -EINVAL;
 451
 452         set_bit(STRIPE_LOG_TRAPPED, &sh->state);
 453         /*
 454          * The stripe must enter state machine again to finish the write, so
 455          * don't delay.
 456          */
 457         clear_bit(STRIPE_DELAYED, &sh->state);
 458         atomic_inc(&sh->count);
 459
 460         mutex_lock(&log->io_mutex);
 461         /* meta + data */
 462         reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
 463         if (r5l_has_free_space(log, reserve))
 464                 r5l_log_stripe(log, sh, data_pages, parity_pages);
 465         else {
 466                 spin_lock(&log->no_space_stripes_lock);
 467                 list_add_tail(&sh->log_list, &log->no_space_stripes);
 468                 spin_unlock(&log->no_space_stripes_lock);
 469
 470                 r5l_wake_reclaim(log, reserve);
 471         }
 472         mutex_unlock(&log->io_mutex);
 473
 474         return 0;
 475 }
 476
 477 void r5l_write_stripe_run(struct r5l_log *log)
 478 {
 479         if (!log)
 480                 return;
 481         mutex_lock(&log->io_mutex);
 482         r5l_submit_current_io(log);
 483         mutex_unlock(&log->io_mutex);
 484 }
 485
 486 int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
 487 {
 488         if (!log)
 489                 return -ENODEV;
 490         /*
 491          * we flush log disk cache first, then write stripe data to raid disks.
 492          * So if bio is finished, the log disk cache is flushed already. The
 493          * recovery guarantees we can recovery the bio from log disk, so we
 494          * don't need to flush again
 495          */
 496         if (bio->bi_iter.bi_size == 0) {
 497                 bio_endio(bio);
 498                 return 0;
 499         }
 500         bio->bi_rw &= ~REQ_FLUSH;
 501         return -EAGAIN;
 502 }
 503
 504 /* This will run after log space is reclaimed */
 505 static void r5l_run_no_space_stripes(struct r5l_log *log)
 506 {
 507         struct stripe_head *sh;
 508
 509         spin_lock(&log->no_space_stripes_lock);
 510         while (!list_empty(&log->no_space_stripes)) {
 511                 sh = list_first_entry(&log->no_space_stripes,
 512                                       struct stripe_head, log_list);
 513                 list_del_init(&sh->log_list);
 514                 set_bit(STRIPE_HANDLE, &sh->state);
 515                 raid5_release_stripe(sh);
 516         }
 517         spin_unlock(&log->no_space_stripes_lock);
 518 }
 519
 520 static sector_t r5l_reclaimable_space(struct r5l_log *log)
 521 {
 522         return r5l_ring_distance(log, log->last_checkpoint,
 523                                  log->next_checkpoint);
 524 }
 525
 526 static bool r5l_complete_flushed_ios(struct r5l_log *log)
 527 {
 528         struct r5l_io_unit *io, *next;
 529         bool found = false;
 530
 531         assert_spin_locked(&log->io_list_lock);
 532
 533         list_for_each_entry_safe(io, next, &log->flushed_ios, log_sibling) {
 534                 /* don't change list order */
 535                 if (io->state < IO_UNIT_STRIPE_END)
 536                         break;
 537
 538                 log->next_checkpoint = io->log_start;
 539                 log->next_cp_seq = io->seq;
 540
 541                 list_del(&io->log_sibling);
 542                 r5l_free_io_unit(log, io);
 543
 544                 found = true;
 545         }
 546
 547         return found;
 548 }
 549
 550 static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
 551 {
 552         struct r5l_log *log = io->log;
 553         unsigned long flags;
 554
 555         spin_lock_irqsave(&log->io_list_lock, flags);
 556         __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
 557
 558         if (!r5l_complete_flushed_ios(log)) {
 559                 spin_unlock_irqrestore(&log->io_list_lock, flags);
 560                 return;
 561         }
 562
 563         if (r5l_reclaimable_space(log) > log->max_free_space)
 564                 r5l_wake_reclaim(log, 0);
 565
 566         spin_unlock_irqrestore(&log->io_list_lock, flags);
 567         wake_up(&log->iounit_wait);
 568 }
 569
 570 void r5l_stripe_write_finished(struct stripe_head *sh)
 571 {
 572         struct r5l_io_unit *io;
 573
 574         io = sh->log_io;
 575         sh->log_io = NULL;
 576
 577         if (io && atomic_dec_and_test(&io->pending_stripe))
 578                 __r5l_stripe_write_finished(io);
 579 }
 580
 581 static void r5l_log_flush_endio(struct bio *bio)
 582 {
 583         struct r5l_log *log = container_of(bio, struct r5l_log,
 584                 flush_bio);
 585         unsigned long flags;
 586         struct r5l_io_unit *io;
 587         struct stripe_head *sh;
 588
 589         spin_lock_irqsave(&log->io_list_lock, flags);
 590         list_for_each_entry(io, &log->flushing_ios, log_sibling) {
 591                 while (!list_empty(&io->stripe_list)) {
 592                         sh = list_first_entry(&io->stripe_list,
 593                                 struct stripe_head, log_list);
 594                         list_del_init(&sh->log_list);
 595                         set_bit(STRIPE_HANDLE, &sh->state);
 596                         raid5_release_stripe(sh);
 597                 }
 598         }
 599         list_splice_tail_init(&log->flushing_ios, &log->flushed_ios);
 600         spin_unlock_irqrestore(&log->io_list_lock, flags);
 601 }
 602
 603 /*
 604  * Starting dispatch IO to raid.
 605  * io_unit(meta) consists of a log. There is one situation we want to avoid. A
 606  * broken meta in the middle of a log causes recovery can't find meta at the
 607  * head of log. If operations require meta at the head persistent in log, we
 608  * must make sure meta before it persistent in log too. A case is:
 609  *
 610  * stripe data/parity is in log, we start write stripe to raid disks. stripe
 611  * data/parity must be persistent in log before we do the write to raid disks.
 612  *
 613  * The solution is we restrictly maintain io_unit list order. In this case, we
 614  * only write stripes of an io_unit to raid disks till the io_unit is the first
 615  * one whose data/parity is in log.
 616  */
 617 void r5l_flush_stripe_to_raid(struct r5l_log *log)
 618 {
 619         bool do_flush;
 620         if (!log)
 621                 return;
 622
 623         spin_lock_irq(&log->io_list_lock);
 624         /* flush bio is running */
 625         if (!list_empty(&log->flushing_ios)) {
 626                 spin_unlock_irq(&log->io_list_lock);
 627                 return;
 628         }
 629         list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
 630         do_flush = !list_empty(&log->flushing_ios);
 631         spin_unlock_irq(&log->io_list_lock);
 632
 633         if (!do_flush)
 634                 return;
 635         bio_reset(&log->flush_bio);
 636         log->flush_bio.bi_bdev = log->rdev->bdev;
 637         log->flush_bio.bi_end_io = r5l_log_flush_endio;
 638         submit_bio(WRITE_FLUSH, &log->flush_bio);
 639 }
 640
 641 static void r5l_write_super(struct r5l_log *log, sector_t cp);
 642 static void r5l_do_reclaim(struct r5l_log *log)
 643 {
 644         sector_t reclaim_target = xchg(&log->reclaim_target, 0);
 645         sector_t reclaimable;
 646         sector_t next_checkpoint;
 647         u64 next_cp_seq;
 648
 649         spin_lock_irq(&log->io_list_lock);
 650         /*
 651          * move proper io_unit to reclaim list. We should not change the order.
 652          * reclaimable/unreclaimable io_unit can be mixed in the list, we
 653          * shouldn't reuse space of an unreclaimable io_unit
 654          */
 655         while (1) {
 656                 reclaimable = r5l_reclaimable_space(log);
 657                 if (reclaimable >= reclaim_target ||
 658                     (list_empty(&log->running_ios) &&
 659                      list_empty(&log->io_end_ios) &&
 660                      list_empty(&log->flushing_ios) &&
 661                      list_empty(&log->flushed_ios)))
 662                         break;
 663
 664                 md_wakeup_thread(log->rdev->mddev->thread);
 665                 wait_event_lock_irq(log->iounit_wait,
 666                                     r5l_reclaimable_space(log) > reclaimable,
 667                                     log->io_list_lock);
 668         }
 669
 670         next_checkpoint = log->next_checkpoint;
 671         next_cp_seq = log->next_cp_seq;
 672         spin_unlock_irq(&log->io_list_lock);
 673
 674         BUG_ON(reclaimable < 0);
 675         if (reclaimable == 0)
 676                 return;
 677
 678         /*
 679          * write_super will flush cache of each raid disk. We must write super
 680          * here, because the log area might be reused soon and we don't want to
 681          * confuse recovery
 682          */
 683         r5l_write_super(log, next_checkpoint);
 684
 685         mutex_lock(&log->io_mutex);
 686         log->last_checkpoint = next_checkpoint;
 687         log->last_cp_seq = next_cp_seq;
 688         mutex_unlock(&log->io_mutex);
 689
 690         r5l_run_no_space_stripes(log);
 691 }
 692
 693 static void r5l_reclaim_thread(struct md_thread *thread)
 694 {
 695         struct mddev *mddev = thread->mddev;
 696         struct r5conf *conf = mddev->private;
 697         struct r5l_log *log = conf->log;
 698
 699         if (!log)
 700                 return;
 701         r5l_do_reclaim(log);
 702 }
 703
 704 static void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
 705 {
 706         unsigned long target;
 707         unsigned long new = (unsigned long)space; /* overflow in theory */
 708
 709         do {
 710                 target = log->reclaim_target;
 711                 if (new < target)
 712                         return;
 713         } while (cmpxchg(&log->reclaim_target, target, new) != target);
 714         md_wakeup_thread(log->reclaim_thread);
 715 }
 716
 717 void r5l_quiesce(struct r5l_log *log, int state)
 718 {
 719         if (!log || state == 2)
 720                 return;
 721         if (state == 0) {
 722                 log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
 723                                         log->rdev->mddev, "reclaim");
 724         } else if (state == 1) {
 725                 /*
 726                  * at this point all stripes are finished, so io_unit is at
 727                  * least in STRIPE_END state
 728                  */
 729                 r5l_wake_reclaim(log, -1L);
 730                 md_unregister_thread(&log->reclaim_thread);
 731                 r5l_do_reclaim(log);
 732         }
 733 }
 734
 735 struct r5l_recovery_ctx {
 736         struct page *meta_page;         /* current meta */
 737         sector_t meta_total_blocks;     /* total size of current meta and data */
 738         sector_t pos;                   /* recovery position */
 739         u64 seq;                        /* recovery position seq */
 740 };
 741
 742 static int r5l_read_meta_block(struct r5l_log *log,
 743                                struct r5l_recovery_ctx *ctx)
 744 {
 745         struct page *page = ctx->meta_page;
 746         struct r5l_meta_block *mb;
 747         u32 crc, stored_crc;
 748
 749         if (!sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, READ, false))
 750                 return -EIO;
 751
 752         mb = page_address(page);
 753         stored_crc = le32_to_cpu(mb->checksum);
 754         mb->checksum = 0;
 755
 756         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
 757             le64_to_cpu(mb->seq) != ctx->seq ||
 758             mb->version != R5LOG_VERSION ||
 759             le64_to_cpu(mb->position) != ctx->pos)
 760                 return -EINVAL;
 761
 762         crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
 763         if (stored_crc != crc)
 764                 return -EINVAL;
 765
 766         if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
 767                 return -EINVAL;
 768
 769         ctx->meta_total_blocks = BLOCK_SECTORS;
 770
 771         return 0;
 772 }
 773
 774 static int r5l_recovery_flush_one_stripe(struct r5l_log *log,
 775                                          struct r5l_recovery_ctx *ctx,
 776                                          sector_t stripe_sect,
 777                                          int *offset, sector_t *log_offset)
 778 {
 779         struct r5conf *conf = log->rdev->mddev->private;
 780         struct stripe_head *sh;
 781         struct r5l_payload_data_parity *payload;
 782         int disk_index;
 783
 784         sh = raid5_get_active_stripe(conf, stripe_sect, 0, 0, 0);
 785         while (1) {
 786                 payload = page_address(ctx->meta_page) + *offset;
 787
 788                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
 789                         raid5_compute_sector(conf,
 790                                              le64_to_cpu(payload->location), 0,
 791                                              &disk_index, sh);
 792
 793                         sync_page_io(log->rdev, *log_offset, PAGE_SIZE,
 794                                      sh->dev[disk_index].page, READ, false);
 795                         sh->dev[disk_index].log_checksum =
 796                                 le32_to_cpu(payload->checksum[0]);
 797                         set_bit(R5_Wantwrite, &sh->dev[disk_index].flags);
 798                         ctx->meta_total_blocks += BLOCK_SECTORS;
 799                 } else {
 800                         disk_index = sh->pd_idx;
 801                         sync_page_io(log->rdev, *log_offset, PAGE_SIZE,
 802                                      sh->dev[disk_index].page, READ, false);
 803                         sh->dev[disk_index].log_checksum =
 804                                 le32_to_cpu(payload->checksum[0]);
 805                         set_bit(R5_Wantwrite, &sh->dev[disk_index].flags);
 806
 807                         if (sh->qd_idx >= 0) {
 808                                 disk_index = sh->qd_idx;
 809                                 sync_page_io(log->rdev,
 810                                              r5l_ring_add(log, *log_offset, BLOCK_SECTORS),
 811                                              PAGE_SIZE, sh->dev[disk_index].page,
 812                                              READ, false);
 813                                 sh->dev[disk_index].log_checksum =
 814                                         le32_to_cpu(payload->checksum[1]);
 815                                 set_bit(R5_Wantwrite,
 816                                         &sh->dev[disk_index].flags);
 817                         }
 818                         ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
 819                 }
 820
 821                 *log_offset = r5l_ring_add(log, *log_offset,
 822                                            le32_to_cpu(payload->size));
 823                 *offset += sizeof(struct r5l_payload_data_parity) +
 824                         sizeof(__le32) *
 825                         (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
 826                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
 827                         break;
 828         }
 829
 830         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
 831                 void *addr;
 832                 u32 checksum;
 833
 834                 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
 835                         continue;
 836                 addr = kmap_atomic(sh->dev[disk_index].page);
 837                 checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
 838                 kunmap_atomic(addr);
 839                 if (checksum != sh->dev[disk_index].log_checksum)
 840                         goto error;
 841         }
 842
 843         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
 844                 struct md_rdev *rdev, *rrdev;
 845
 846                 if (!test_and_clear_bit(R5_Wantwrite,
 847                                         &sh->dev[disk_index].flags))
 848                         continue;
 849
 850                 /* in case device is broken */
 851                 rdev = rcu_dereference(conf->disks[disk_index].rdev);
 852                 if (rdev)
 853                         sync_page_io(rdev, stripe_sect, PAGE_SIZE,
 854                                      sh->dev[disk_index].page, WRITE, false);
 855                 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
 856                 if (rrdev)
 857                         sync_page_io(rrdev, stripe_sect, PAGE_SIZE,
 858                                      sh->dev[disk_index].page, WRITE, false);
 859         }
 860         raid5_release_stripe(sh);
 861         return 0;
 862
 863 error:
 864         for (disk_index = 0; disk_index < sh->disks; disk_index++)
 865                 sh->dev[disk_index].flags = 0;
 866         raid5_release_stripe(sh);
 867         return -EINVAL;
 868 }
 869
 870 static int r5l_recovery_flush_one_meta(struct r5l_log *log,
 871                                        struct r5l_recovery_ctx *ctx)
 872 {
 873         struct r5conf *conf = log->rdev->mddev->private;
 874         struct r5l_payload_data_parity *payload;
 875         struct r5l_meta_block *mb;
 876         int offset;
 877         sector_t log_offset;
 878         sector_t stripe_sector;
 879
 880         mb = page_address(ctx->meta_page);
 881         offset = sizeof(struct r5l_meta_block);
 882         log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
 883
 884         while (offset < le32_to_cpu(mb->meta_size)) {
 885                 int dd;
 886
 887                 payload = (void *)mb + offset;
 888                 stripe_sector = raid5_compute_sector(conf,
 889                                                      le64_to_cpu(payload->location), 0, &dd, NULL);
 890                 if (r5l_recovery_flush_one_stripe(log, ctx, stripe_sector,
 891                                                   &offset, &log_offset))
 892                         return -EINVAL;
 893         }
 894         return 0;
 895 }
 896
 897 /* copy data/parity from log to raid disks */
 898 static void r5l_recovery_flush_log(struct r5l_log *log,
 899                                    struct r5l_recovery_ctx *ctx)
 900 {
 901         while (1) {
 902                 if (r5l_read_meta_block(log, ctx))
 903                         return;
 904                 if (r5l_recovery_flush_one_meta(log, ctx))
 905                         return;
 906                 ctx->seq++;
 907                 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
 908         }
 909 }
 910
 911 static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
 912                                           u64 seq)
 913 {
 914         struct page *page;
 915         struct r5l_meta_block *mb;
 916         u32 crc;
 917
 918         page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 919         if (!page)
 920                 return -ENOMEM;
 921         mb = page_address(page);
 922         mb->magic = cpu_to_le32(R5LOG_MAGIC);
 923         mb->version = R5LOG_VERSION;
 924         mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
 925         mb->seq = cpu_to_le64(seq);
 926         mb->position = cpu_to_le64(pos);
 927         crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
 928         mb->checksum = cpu_to_le32(crc);
 929
 930         if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, WRITE_FUA, false)) {
 931                 __free_page(page);
 932                 return -EIO;
 933         }
 934         __free_page(page);
 935         return 0;
 936 }
 937
 938 static int r5l_recovery_log(struct r5l_log *log)
 939 {
 940         struct r5l_recovery_ctx ctx;
 941
 942         ctx.pos = log->last_checkpoint;
 943         ctx.seq = log->last_cp_seq;
 944         ctx.meta_page = alloc_page(GFP_KERNEL);
 945         if (!ctx.meta_page)
 946                 return -ENOMEM;
 947
 948         r5l_recovery_flush_log(log, &ctx);
 949         __free_page(ctx.meta_page);
 950
 951         /*
 952          * we did a recovery. Now ctx.pos points to an invalid meta block. New
 953          * log will start here. but we can't let superblock point to last valid
 954          * meta block. The log might looks like:
 955          * | meta 1| meta 2| meta 3|
 956          * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
 957          * superblock points to meta 1, we write a new valid meta 2n.  if crash
 958          * happens again, new recovery will start from meta 1. Since meta 2n is
 959          * valid now, recovery will think meta 3 is valid, which is wrong.
 960          * The solution is we create a new meta in meta2 with its seq == meta
 961          * 1's seq + 10 and let superblock points to meta2. The same recovery will
 962          * not think meta 3 is a valid meta, because its seq doesn't match
 963          */
 964         if (ctx.seq > log->last_cp_seq + 1) {
 965                 int ret;
 966
 967                 ret = r5l_log_write_empty_meta_block(log, ctx.pos, ctx.seq + 10);
 968                 if (ret)
 969                         return ret;
 970                 log->seq = ctx.seq + 11;
 971                 log->log_start = r5l_ring_add(log, ctx.pos, BLOCK_SECTORS);
 972                 r5l_write_super(log, ctx.pos);
 973         } else {
 974                 log->log_start = ctx.pos;
 975                 log->seq = ctx.seq;
 976         }
 977         return 0;
 978 }
 979
 980 static void r5l_write_super(struct r5l_log *log, sector_t cp)
 981 {
 982         struct mddev *mddev = log->rdev->mddev;
 983
 984         log->rdev->journal_tail = cp;
 985         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 986 }
 987
 988 static int r5l_load_log(struct r5l_log *log)
 989 {
 990         struct md_rdev *rdev = log->rdev;
 991         struct page *page;
 992         struct r5l_meta_block *mb;
 993         sector_t cp = log->rdev->journal_tail;
 994         u32 stored_crc, expected_crc;
 995         bool create_super = false;
 996         int ret;
 997
 998         /* Make sure it's valid */
 999         if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
1000                 cp = 0;
1001         page = alloc_page(GFP_KERNEL);
1002         if (!page)
1003                 return -ENOMEM;
1004
1005         if (!sync_page_io(rdev, cp, PAGE_SIZE, page, READ, false)) {
1006                 ret = -EIO;
1007                 goto ioerr;
1008         }
1009         mb = page_address(page);
1010
1011         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1012             mb->version != R5LOG_VERSION) {
1013                 create_super = true;
1014                 goto create;
1015         }
1016         stored_crc = le32_to_cpu(mb->checksum);
1017         mb->checksum = 0;
1018         expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1019         if (stored_crc != expected_crc) {
1020                 create_super = true;
1021                 goto create;
1022         }
1023         if (le64_to_cpu(mb->position) != cp) {
1024                 create_super = true;
1025                 goto create;
1026         }
1027 create:
1028         if (create_super) {
1029                 log->last_cp_seq = prandom_u32();
1030                 cp = 0;
1031                 /*
1032                  * Make sure super points to correct address. Log might have
1033                  * data very soon. If super hasn't correct log tail address,
1034                  * recovery can't find the log
1035                  */
1036                 r5l_write_super(log, cp);
1037         } else
1038                 log->last_cp_seq = le64_to_cpu(mb->seq);
1039
1040         log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
1041         log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
1042         if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
1043                 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
1044         log->last_checkpoint = cp;
1045
1046         __free_page(page);
1047
1048         return r5l_recovery_log(log);
1049 ioerr:
1050         __free_page(page);
1051         return ret;
1052 }
1053
1054 int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
1055 {
1056         struct r5l_log *log;
1057
1058         if (PAGE_SIZE != 4096)
1059                 return -EINVAL;
1060         log = kzalloc(sizeof(*log), GFP_KERNEL);
1061         if (!log)
1062                 return -ENOMEM;
1063         log->rdev = rdev;
1064
1065         log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
1066                                        sizeof(rdev->mddev->uuid));
1067
1068         mutex_init(&log->io_mutex);
1069
1070         spin_lock_init(&log->io_list_lock);
1071         INIT_LIST_HEAD(&log->running_ios);
1072         INIT_LIST_HEAD(&log->io_end_ios);
1073         INIT_LIST_HEAD(&log->flushing_ios);
1074         INIT_LIST_HEAD(&log->flushed_ios);
1075         bio_init(&log->flush_bio);
1076
1077         log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
1078         if (!log->io_kc)
1079                 goto io_kc;
1080
1081         log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
1082                                                  log->rdev->mddev, "reclaim");
1083         if (!log->reclaim_thread)
1084                 goto reclaim_thread;
1085         init_waitqueue_head(&log->iounit_wait);
1086
1087         INIT_LIST_HEAD(&log->no_space_stripes);
1088         spin_lock_init(&log->no_space_stripes_lock);
1089
1090         if (r5l_load_log(log))
1091                 goto error;
1092
1093         conf->log = log;
1094         return 0;
1095 error:
1096         md_unregister_thread(&log->reclaim_thread);
1097 reclaim_thread:
1098         kmem_cache_destroy(log->io_kc);
1099 io_kc:
1100         kfree(log);
1101         return -EINVAL;
1102 }
1103
1104 void r5l_exit_log(struct r5l_log *log)
1105 {
1106         md_unregister_thread(&log->reclaim_thread);
1107         kmem_cache_destroy(log->io_kc);
1108         kfree(log);
1109 }