drivers/md/raid1.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * raid1.c : Multiple Devices driver for Linux
   4  *
   5  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   6  *
   7  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   8  *
   9  * RAID-1 management functions.
  10  *
  11  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  12  *
  13  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  14  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  15  *
  16  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  17  * bitmapped intelligence in resync:
  18  *
  19  *      - bitmap marked during normal i/o
  20  *      - bitmap used to skip nondirty blocks during sync
  21  *
  22  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  23  * - persistent bitmap code
  24  */
  25
  26 #include <linux/slab.h>
  27 #include <linux/delay.h>
  28 #include <linux/blkdev.h>
  29 #include <linux/module.h>
  30 #include <linux/seq_file.h>
  31 #include <linux/ratelimit.h>
  32 #include <linux/interval_tree_generic.h>
  33
  34 #include <trace/events/block.h>
  35
  36 #include "md.h"
  37 #include "raid1.h"
  38 #include "md-bitmap.h"
  39
  40 #define UNSUPPORTED_MDDEV_FLAGS         \
  41         ((1L << MD_HAS_JOURNAL) |       \
  42          (1L << MD_JOURNAL_CLEAN) |     \
  43          (1L << MD_HAS_PPL) |           \
  44          (1L << MD_HAS_MULTIPLE_PPLS))
  45
  46 static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
  47 static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
  48
  49 #define raid1_log(md, fmt, args...)                             \
  50         do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
  51
  52 #include "raid1-10.c"
  53
  54 #define START(node) ((node)->start)
  55 #define LAST(node) ((node)->last)
  56 INTERVAL_TREE_DEFINE(struct serial_info, node, sector_t, _subtree_last,
  57                      START, LAST, static inline, raid1_rb);
  58
  59 static int check_and_add_serial(struct md_rdev *rdev, struct r1bio *r1_bio,
  60                                 struct serial_info *si, int idx)
  61 {
  62         unsigned long flags;
  63         int ret = 0;
  64         sector_t lo = r1_bio->sector;
  65         sector_t hi = lo + r1_bio->sectors;
  66         struct serial_in_rdev *serial = &rdev->serial[idx];
  67
  68         spin_lock_irqsave(&serial->serial_lock, flags);
  69         /* collision happened */
  70         if (raid1_rb_iter_first(&serial->serial_rb, lo, hi))
  71                 ret = -EBUSY;
  72         else {
  73                 si->start = lo;
  74                 si->last = hi;
  75                 raid1_rb_insert(si, &serial->serial_rb);
  76         }
  77         spin_unlock_irqrestore(&serial->serial_lock, flags);
  78
  79         return ret;
  80 }
  81
  82 static void wait_for_serialization(struct md_rdev *rdev, struct r1bio *r1_bio)
  83 {
  84         struct mddev *mddev = rdev->mddev;
  85         struct serial_info *si;
  86         int idx = sector_to_idx(r1_bio->sector);
  87         struct serial_in_rdev *serial = &rdev->serial[idx];
  88
  89         if (WARN_ON(!mddev->serial_info_pool))
  90                 return;
  91         si = mempool_alloc(mddev->serial_info_pool, GFP_NOIO);
  92         wait_event(serial->serial_io_wait,
  93                    check_and_add_serial(rdev, r1_bio, si, idx) == 0);
  94 }
  95
  96 static void remove_serial(struct md_rdev *rdev, sector_t lo, sector_t hi)
  97 {
  98         struct serial_info *si;
  99         unsigned long flags;
 100         int found = 0;
 101         struct mddev *mddev = rdev->mddev;
 102         int idx = sector_to_idx(lo);
 103         struct serial_in_rdev *serial = &rdev->serial[idx];
 104
 105         spin_lock_irqsave(&serial->serial_lock, flags);
 106         for (si = raid1_rb_iter_first(&serial->serial_rb, lo, hi);
 107              si; si = raid1_rb_iter_next(si, lo, hi)) {
 108                 if (si->start == lo && si->last == hi) {
 109                         raid1_rb_remove(si, &serial->serial_rb);
 110                         mempool_free(si, mddev->serial_info_pool);
 111                         found = 1;
 112                         break;
 113                 }
 114         }
 115         if (!found)
 116                 WARN(1, "The write IO is not recorded for serialization\n");
 117         spin_unlock_irqrestore(&serial->serial_lock, flags);
 118         wake_up(&serial->serial_io_wait);
 119 }
 120
 121 /*
 122  * for resync bio, r1bio pointer can be retrieved from the per-bio
 123  * 'struct resync_pages'.
 124  */
 125 static inline struct r1bio *get_resync_r1bio(struct bio *bio)
 126 {
 127         return get_resync_pages(bio)->raid_bio;
 128 }
 129
 130 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
 131 {
 132         struct pool_info *pi = data;
 133         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
 134
 135         /* allocate a r1bio with room for raid_disks entries in the bios array */
 136         return kzalloc(size, gfp_flags);
 137 }
 138
 139 #define RESYNC_DEPTH 32
 140 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
 141 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
 142 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
 143 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
 144 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
 145
 146 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
 147 {
 148         struct pool_info *pi = data;
 149         struct r1bio *r1_bio;
 150         struct bio *bio;
 151         int need_pages;
 152         int j;
 153         struct resync_pages *rps;
 154
 155         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 156         if (!r1_bio)
 157                 return NULL;
 158
 159         rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
 160                             gfp_flags);
 161         if (!rps)
 162                 goto out_free_r1bio;
 163
 164         /*
 165          * Allocate bios : 1 for reading, n-1 for writing
 166          */
 167         for (j = pi->raid_disks ; j-- ; ) {
 168                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 169                 if (!bio)
 170                         goto out_free_bio;
 171                 r1_bio->bios[j] = bio;
 172         }
 173         /*
 174          * Allocate RESYNC_PAGES data pages and attach them to
 175          * the first bio.
 176          * If this is a user-requested check/repair, allocate
 177          * RESYNC_PAGES for each bio.
 178          */
 179         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 180                 need_pages = pi->raid_disks;
 181         else
 182                 need_pages = 1;
 183         for (j = 0; j < pi->raid_disks; j++) {
 184                 struct resync_pages *rp = &rps[j];
 185
 186                 bio = r1_bio->bios[j];
 187
 188                 if (j < need_pages) {
 189                         if (resync_alloc_pages(rp, gfp_flags))
 190                                 goto out_free_pages;
 191                 } else {
 192                         memcpy(rp, &rps[0], sizeof(*rp));
 193                         resync_get_all_pages(rp);
 194                 }
 195
 196                 rp->raid_bio = r1_bio;
 197                 bio->bi_private = rp;
 198         }
 199
 200         r1_bio->master_bio = NULL;
 201
 202         return r1_bio;
 203
 204 out_free_pages:
 205         while (--j >= 0)
 206                 resync_free_pages(&rps[j]);
 207
 208 out_free_bio:
 209         while (++j < pi->raid_disks)
 210                 bio_put(r1_bio->bios[j]);
 211         kfree(rps);
 212
 213 out_free_r1bio:
 214         rbio_pool_free(r1_bio, data);
 215         return NULL;
 216 }
 217
 218 static void r1buf_pool_free(void *__r1_bio, void *data)
 219 {
 220         struct pool_info *pi = data;
 221         int i;
 222         struct r1bio *r1bio = __r1_bio;
 223         struct resync_pages *rp = NULL;
 224
 225         for (i = pi->raid_disks; i--; ) {
 226                 rp = get_resync_pages(r1bio->bios[i]);
 227                 resync_free_pages(rp);
 228                 bio_put(r1bio->bios[i]);
 229         }
 230
 231         /* resync pages array stored in the 1st bio's .bi_private */
 232         kfree(rp);
 233
 234         rbio_pool_free(r1bio, data);
 235 }
 236
 237 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 238 {
 239         int i;
 240
 241         for (i = 0; i < conf->raid_disks * 2; i++) {
 242                 struct bio **bio = r1_bio->bios + i;
 243                 if (!BIO_SPECIAL(*bio))
 244                         bio_put(*bio);
 245                 *bio = NULL;
 246         }
 247 }
 248
 249 static void free_r1bio(struct r1bio *r1_bio)
 250 {
 251         struct r1conf *conf = r1_bio->mddev->private;
 252
 253         put_all_bios(conf, r1_bio);
 254         mempool_free(r1_bio, &conf->r1bio_pool);
 255 }
 256
 257 static void put_buf(struct r1bio *r1_bio)
 258 {
 259         struct r1conf *conf = r1_bio->mddev->private;
 260         sector_t sect = r1_bio->sector;
 261         int i;
 262
 263         for (i = 0; i < conf->raid_disks * 2; i++) {
 264                 struct bio *bio = r1_bio->bios[i];
 265                 if (bio->bi_end_io)
 266                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 267         }
 268
 269         mempool_free(r1_bio, &conf->r1buf_pool);
 270
 271         lower_barrier(conf, sect);
 272 }
 273
 274 static void reschedule_retry(struct r1bio *r1_bio)
 275 {
 276         unsigned long flags;
 277         struct mddev *mddev = r1_bio->mddev;
 278         struct r1conf *conf = mddev->private;
 279         int idx;
 280
 281         idx = sector_to_idx(r1_bio->sector);
 282         spin_lock_irqsave(&conf->device_lock, flags);
 283         list_add(&r1_bio->retry_list, &conf->retry_list);
 284         atomic_inc(&conf->nr_queued[idx]);
 285         spin_unlock_irqrestore(&conf->device_lock, flags);
 286
 287         wake_up(&conf->wait_barrier);
 288         md_wakeup_thread(mddev->thread);
 289 }
 290
 291 /*
 292  * raid_end_bio_io() is called when we have finished servicing a mirrored
 293  * operation and are ready to return a success/failure code to the buffer
 294  * cache layer.
 295  */
 296 static void call_bio_endio(struct r1bio *r1_bio)
 297 {
 298         struct bio *bio = r1_bio->master_bio;
 299
 300         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 301                 bio->bi_status = BLK_STS_IOERR;
 302
 303         if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
 304                 bio_end_io_acct(bio, r1_bio->start_time);
 305         bio_endio(bio);
 306 }
 307
 308 static void raid_end_bio_io(struct r1bio *r1_bio)
 309 {
 310         struct bio *bio = r1_bio->master_bio;
 311         struct r1conf *conf = r1_bio->mddev->private;
 312
 313         /* if nobody has done the final endio yet, do it now */
 314         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 315                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 316                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 317                          (unsigned long long) bio->bi_iter.bi_sector,
 318                          (unsigned long long) bio_end_sector(bio) - 1);
 319
 320                 call_bio_endio(r1_bio);
 321         }
 322         /*
 323          * Wake up any possible resync thread that waits for the device
 324          * to go idle.  All I/Os, even write-behind writes, are done.
 325          */
 326         allow_barrier(conf, r1_bio->sector);
 327
 328         free_r1bio(r1_bio);
 329 }
 330
 331 /*
 332  * Update disk head position estimator based on IRQ completion info.
 333  */
 334 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 335 {
 336         struct r1conf *conf = r1_bio->mddev->private;
 337
 338         conf->mirrors[disk].head_position =
 339                 r1_bio->sector + (r1_bio->sectors);
 340 }
 341
 342 /*
 343  * Find the disk number which triggered given bio
 344  */
 345 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 346 {
 347         int mirror;
 348         struct r1conf *conf = r1_bio->mddev->private;
 349         int raid_disks = conf->raid_disks;
 350
 351         for (mirror = 0; mirror < raid_disks * 2; mirror++)
 352                 if (r1_bio->bios[mirror] == bio)
 353                         break;
 354
 355         BUG_ON(mirror == raid_disks * 2);
 356         update_head_pos(mirror, r1_bio);
 357
 358         return mirror;
 359 }
 360
 361 static void raid1_end_read_request(struct bio *bio)
 362 {
 363         int uptodate = !bio->bi_status;
 364         struct r1bio *r1_bio = bio->bi_private;
 365         struct r1conf *conf = r1_bio->mddev->private;
 366         struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 367
 368         /*
 369          * this branch is our 'one mirror IO has finished' event handler:
 370          */
 371         update_head_pos(r1_bio->read_disk, r1_bio);
 372
 373         if (uptodate)
 374                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 375         else if (test_bit(FailFast, &rdev->flags) &&
 376                  test_bit(R1BIO_FailFast, &r1_bio->state))
 377                 /* This was a fail-fast read so we definitely
 378                  * want to retry */
 379                 ;
 380         else {
 381                 /* If all other devices have failed, we want to return
 382                  * the error upwards rather than fail the last device.
 383                  * Here we redefine "uptodate" to mean "Don't want to retry"
 384                  */
 385                 unsigned long flags;
 386                 spin_lock_irqsave(&conf->device_lock, flags);
 387                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 388                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 389                      test_bit(In_sync, &rdev->flags)))
 390                         uptodate = 1;
 391                 spin_unlock_irqrestore(&conf->device_lock, flags);
 392         }
 393
 394         if (uptodate) {
 395                 raid_end_bio_io(r1_bio);
 396                 rdev_dec_pending(rdev, conf->mddev);
 397         } else {
 398                 /*
 399                  * oops, read error:
 400                  */
 401                 char b[BDEVNAME_SIZE];
 402                 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
 403                                    mdname(conf->mddev),
 404                                    bdevname(rdev->bdev, b),
 405                                    (unsigned long long)r1_bio->sector);
 406                 set_bit(R1BIO_ReadError, &r1_bio->state);
 407                 reschedule_retry(r1_bio);
 408                 /* don't drop the reference on read_disk yet */
 409         }
 410 }
 411
 412 static void close_write(struct r1bio *r1_bio)
 413 {
 414         /* it really is the end of this request */
 415         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 416                 bio_free_pages(r1_bio->behind_master_bio);
 417                 bio_put(r1_bio->behind_master_bio);
 418                 r1_bio->behind_master_bio = NULL;
 419         }
 420         /* clear the bitmap if all writes complete successfully */
 421         md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 422                            r1_bio->sectors,
 423                            !test_bit(R1BIO_Degraded, &r1_bio->state),
 424                            test_bit(R1BIO_BehindIO, &r1_bio->state));
 425         md_write_end(r1_bio->mddev);
 426 }
 427
 428 static void r1_bio_write_done(struct r1bio *r1_bio)
 429 {
 430         if (!atomic_dec_and_test(&r1_bio->remaining))
 431                 return;
 432
 433         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 434                 reschedule_retry(r1_bio);
 435         else {
 436                 close_write(r1_bio);
 437                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 438                         reschedule_retry(r1_bio);
 439                 else
 440                         raid_end_bio_io(r1_bio);
 441         }
 442 }
 443
 444 static void raid1_end_write_request(struct bio *bio)
 445 {
 446         struct r1bio *r1_bio = bio->bi_private;
 447         int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 448         struct r1conf *conf = r1_bio->mddev->private;
 449         struct bio *to_put = NULL;
 450         int mirror = find_bio_disk(r1_bio, bio);
 451         struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 452         bool discard_error;
 453         sector_t lo = r1_bio->sector;
 454         sector_t hi = r1_bio->sector + r1_bio->sectors;
 455
 456         discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 457
 458         /*
 459          * 'one mirror IO has finished' event handler:
 460          */
 461         if (bio->bi_status && !discard_error) {
 462                 set_bit(WriteErrorSeen, &rdev->flags);
 463                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
 464                         set_bit(MD_RECOVERY_NEEDED, &
 465                                 conf->mddev->recovery);
 466
 467                 if (test_bit(FailFast, &rdev->flags) &&
 468                     (bio->bi_opf & MD_FAILFAST) &&
 469                     /* We never try FailFast to WriteMostly devices */
 470                     !test_bit(WriteMostly, &rdev->flags)) {
 471                         md_error(r1_bio->mddev, rdev);
 472                 }
 473
 474                 /*
 475                  * When the device is faulty, it is not necessary to
 476                  * handle write error.
 477                  */
 478                 if (!test_bit(Faulty, &rdev->flags))
 479                         set_bit(R1BIO_WriteError, &r1_bio->state);
 480                 else {
 481                         /* Fail the request */
 482                         set_bit(R1BIO_Degraded, &r1_bio->state);
 483                         /* Finished with this branch */
 484                         r1_bio->bios[mirror] = NULL;
 485                         to_put = bio;
 486                 }
 487         } else {
 488                 /*
 489                  * Set R1BIO_Uptodate in our master bio, so that we
 490                  * will return a good error code for to the higher
 491                  * levels even if IO on some other mirrored buffer
 492                  * fails.
 493                  *
 494                  * The 'master' represents the composite IO operation
 495                  * to user-side. So if something waits for IO, then it
 496                  * will wait for the 'master' bio.
 497                  */
 498                 sector_t first_bad;
 499                 int bad_sectors;
 500
 501                 r1_bio->bios[mirror] = NULL;
 502                 to_put = bio;
 503                 /*
 504                  * Do not set R1BIO_Uptodate if the current device is
 505                  * rebuilding or Faulty. This is because we cannot use
 506                  * such device for properly reading the data back (we could
 507                  * potentially use it, if the current write would have felt
 508                  * before rdev->recovery_offset, but for simplicity we don't
 509                  * check this here.
 510                  */
 511                 if (test_bit(In_sync, &rdev->flags) &&
 512                     !test_bit(Faulty, &rdev->flags))
 513                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 514
 515                 /* Maybe we can clear some bad blocks. */
 516                 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 517                                 &first_bad, &bad_sectors) && !discard_error) {
 518                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 519                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 520                 }
 521         }
 522
 523         if (behind) {
 524                 if (test_bit(CollisionCheck, &rdev->flags))
 525                         remove_serial(rdev, lo, hi);
 526                 if (test_bit(WriteMostly, &rdev->flags))
 527                         atomic_dec(&r1_bio->behind_remaining);
 528
 529                 /*
 530                  * In behind mode, we ACK the master bio once the I/O
 531                  * has safely reached all non-writemostly
 532                  * disks. Setting the Returned bit ensures that this
 533                  * gets done only once -- we don't ever want to return
 534                  * -EIO here, instead we'll wait
 535                  */
 536                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 537                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 538                         /* Maybe we can return now */
 539                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 540                                 struct bio *mbio = r1_bio->master_bio;
 541                                 pr_debug("raid1: behind end write sectors"
 542                                          " %llu-%llu\n",
 543                                          (unsigned long long) mbio->bi_iter.bi_sector,
 544                                          (unsigned long long) bio_end_sector(mbio) - 1);
 545                                 call_bio_endio(r1_bio);
 546                         }
 547                 }
 548         } else if (rdev->mddev->serialize_policy)
 549                 remove_serial(rdev, lo, hi);
 550         if (r1_bio->bios[mirror] == NULL)
 551                 rdev_dec_pending(rdev, conf->mddev);
 552
 553         /*
 554          * Let's see if all mirrored write operations have finished
 555          * already.
 556          */
 557         r1_bio_write_done(r1_bio);
 558
 559         if (to_put)
 560                 bio_put(to_put);
 561 }
 562
 563 static sector_t align_to_barrier_unit_end(sector_t start_sector,
 564                                           sector_t sectors)
 565 {
 566         sector_t len;
 567
 568         WARN_ON(sectors == 0);
 569         /*
 570          * len is the number of sectors from start_sector to end of the
 571          * barrier unit which start_sector belongs to.
 572          */
 573         len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 574               start_sector;
 575
 576         if (len > sectors)
 577                 len = sectors;
 578
 579         return len;
 580 }
 581
 582 /*
 583  * This routine returns the disk from which the requested read should
 584  * be done. There is a per-array 'next expected sequential IO' sector
 585  * number - if this matches on the next IO then we use the last disk.
 586  * There is also a per-disk 'last know head position' sector that is
 587  * maintained from IRQ contexts, both the normal and the resync IO
 588  * completion handlers update this position correctly. If there is no
 589  * perfect sequential match then we pick the disk whose head is closest.
 590  *
 591  * If there are 2 mirrors in the same 2 devices, performance degrades
 592  * because position is mirror, not device based.
 593  *
 594  * The rdev for the device selected will have nr_pending incremented.
 595  */
 596 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 597 {
 598         const sector_t this_sector = r1_bio->sector;
 599         int sectors;
 600         int best_good_sectors;
 601         int best_disk, best_dist_disk, best_pending_disk;
 602         int has_nonrot_disk;
 603         int disk;
 604         sector_t best_dist;
 605         unsigned int min_pending;
 606         struct md_rdev *rdev;
 607         int choose_first;
 608         int choose_next_idle;
 609
 610         rcu_read_lock();
 611         /*
 612          * Check if we can balance. We can balance on the whole
 613          * device if no resync is going on, or below the resync window.
 614          * We take the first readable disk when above the resync window.
 615          */
 616  retry:
 617         sectors = r1_bio->sectors;
 618         best_disk = -1;
 619         best_dist_disk = -1;
 620         best_dist = MaxSector;
 621         best_pending_disk = -1;
 622         min_pending = UINT_MAX;
 623         best_good_sectors = 0;
 624         has_nonrot_disk = 0;
 625         choose_next_idle = 0;
 626         clear_bit(R1BIO_FailFast, &r1_bio->state);
 627
 628         if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 629             (mddev_is_clustered(conf->mddev) &&
 630             md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 631                     this_sector + sectors)))
 632                 choose_first = 1;
 633         else
 634                 choose_first = 0;
 635
 636         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 637                 sector_t dist;
 638                 sector_t first_bad;
 639                 int bad_sectors;
 640                 unsigned int pending;
 641                 bool nonrot;
 642
 643                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 644                 if (r1_bio->bios[disk] == IO_BLOCKED
 645                     || rdev == NULL
 646                     || test_bit(Faulty, &rdev->flags))
 647                         continue;
 648                 if (!test_bit(In_sync, &rdev->flags) &&
 649                     rdev->recovery_offset < this_sector + sectors)
 650                         continue;
 651                 if (test_bit(WriteMostly, &rdev->flags)) {
 652                         /* Don't balance among write-mostly, just
 653                          * use the first as a last resort */
 654                         if (best_dist_disk < 0) {
 655                                 if (is_badblock(rdev, this_sector, sectors,
 656                                                 &first_bad, &bad_sectors)) {
 657                                         if (first_bad <= this_sector)
 658                                                 /* Cannot use this */
 659                                                 continue;
 660                                         best_good_sectors = first_bad - this_sector;
 661                                 } else
 662                                         best_good_sectors = sectors;
 663                                 best_dist_disk = disk;
 664                                 best_pending_disk = disk;
 665                         }
 666                         continue;
 667                 }
 668                 /* This is a reasonable device to use.  It might
 669                  * even be best.
 670                  */
 671                 if (is_badblock(rdev, this_sector, sectors,
 672                                 &first_bad, &bad_sectors)) {
 673                         if (best_dist < MaxSector)
 674                                 /* already have a better device */
 675                                 continue;
 676                         if (first_bad <= this_sector) {
 677                                 /* cannot read here. If this is the 'primary'
 678                                  * device, then we must not read beyond
 679                                  * bad_sectors from another device..
 680                                  */
 681                                 bad_sectors -= (this_sector - first_bad);
 682                                 if (choose_first && sectors > bad_sectors)
 683                                         sectors = bad_sectors;
 684                                 if (best_good_sectors > sectors)
 685                                         best_good_sectors = sectors;
 686
 687                         } else {
 688                                 sector_t good_sectors = first_bad - this_sector;
 689                                 if (good_sectors > best_good_sectors) {
 690                                         best_good_sectors = good_sectors;
 691                                         best_disk = disk;
 692                                 }
 693                                 if (choose_first)
 694                                         break;
 695                         }
 696                         continue;
 697                 } else {
 698                         if ((sectors > best_good_sectors) && (best_disk >= 0))
 699                                 best_disk = -1;
 700                         best_good_sectors = sectors;
 701                 }
 702
 703                 if (best_disk >= 0)
 704                         /* At least two disks to choose from so failfast is OK */
 705                         set_bit(R1BIO_FailFast, &r1_bio->state);
 706
 707                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 708                 has_nonrot_disk |= nonrot;
 709                 pending = atomic_read(&rdev->nr_pending);
 710                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 711                 if (choose_first) {
 712                         best_disk = disk;
 713                         break;
 714                 }
 715                 /* Don't change to another disk for sequential reads */
 716                 if (conf->mirrors[disk].next_seq_sect == this_sector
 717                     || dist == 0) {
 718                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 719                         struct raid1_info *mirror = &conf->mirrors[disk];
 720
 721                         best_disk = disk;
 722                         /*
 723                          * If buffered sequential IO size exceeds optimal
 724                          * iosize, check if there is idle disk. If yes, choose
 725                          * the idle disk. read_balance could already choose an
 726                          * idle disk before noticing it's a sequential IO in
 727                          * this disk. This doesn't matter because this disk
 728                          * will idle, next time it will be utilized after the
 729                          * first disk has IO size exceeds optimal iosize. In
 730                          * this way, iosize of the first disk will be optimal
 731                          * iosize at least. iosize of the second disk might be
 732                          * small, but not a big deal since when the second disk
 733                          * starts IO, the first disk is likely still busy.
 734                          */
 735                         if (nonrot && opt_iosize > 0 &&
 736                             mirror->seq_start != MaxSector &&
 737                             mirror->next_seq_sect > opt_iosize &&
 738                             mirror->next_seq_sect - opt_iosize >=
 739                             mirror->seq_start) {
 740                                 choose_next_idle = 1;
 741                                 continue;
 742                         }
 743                         break;
 744                 }
 745
 746                 if (choose_next_idle)
 747                         continue;
 748
 749                 if (min_pending > pending) {
 750                         min_pending = pending;
 751                         best_pending_disk = disk;
 752                 }
 753
 754                 if (dist < best_dist) {
 755                         best_dist = dist;
 756                         best_dist_disk = disk;
 757                 }
 758         }
 759
 760         /*
 761          * If all disks are rotational, choose the closest disk. If any disk is
 762          * non-rotational, choose the disk with less pending request even the
 763          * disk is rotational, which might/might not be optimal for raids with
 764          * mixed ratation/non-rotational disks depending on workload.
 765          */
 766         if (best_disk == -1) {
 767                 if (has_nonrot_disk || min_pending == 0)
 768                         best_disk = best_pending_disk;
 769                 else
 770                         best_disk = best_dist_disk;
 771         }
 772
 773         if (best_disk >= 0) {
 774                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 775                 if (!rdev)
 776                         goto retry;
 777                 atomic_inc(&rdev->nr_pending);
 778                 sectors = best_good_sectors;
 779
 780                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 781                         conf->mirrors[best_disk].seq_start = this_sector;
 782
 783                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 784         }
 785         rcu_read_unlock();
 786         *max_sectors = sectors;
 787
 788         return best_disk;
 789 }
 790
 791 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 792 {
 793         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
 794         md_bitmap_unplug(conf->mddev->bitmap);
 795         wake_up(&conf->wait_barrier);
 796
 797         while (bio) { /* submit pending writes */
 798                 struct bio *next = bio->bi_next;
 799                 struct md_rdev *rdev = (void *)bio->bi_bdev;
 800                 bio->bi_next = NULL;
 801                 bio_set_dev(bio, rdev->bdev);
 802                 if (test_bit(Faulty, &rdev->flags)) {
 803                         bio_io_error(bio);
 804                 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
 805                                     !blk_queue_discard(bio->bi_bdev->bd_disk->queue)))
 806                         /* Just ignore it */
 807                         bio_endio(bio);
 808                 else
 809                         submit_bio_noacct(bio);
 810                 bio = next;
 811                 cond_resched();
 812         }
 813 }
 814
 815 static void flush_pending_writes(struct r1conf *conf)
 816 {
 817         /* Any writes that have been queued but are awaiting
 818          * bitmap updates get flushed here.
 819          */
 820         spin_lock_irq(&conf->device_lock);
 821
 822         if (conf->pending_bio_list.head) {
 823                 struct blk_plug plug;
 824                 struct bio *bio;
 825
 826                 bio = bio_list_get(&conf->pending_bio_list);
 827                 spin_unlock_irq(&conf->device_lock);
 828
 829                 /*
 830                  * As this is called in a wait_event() loop (see freeze_array),
 831                  * current->state might be TASK_UNINTERRUPTIBLE which will
 832                  * cause a warning when we prepare to wait again.  As it is
 833                  * rare that this path is taken, it is perfectly safe to force
 834                  * us to go around the wait_event() loop again, so the warning
 835                  * is a false-positive.  Silence the warning by resetting
 836                  * thread state
 837                  */
 838                 __set_current_state(TASK_RUNNING);
 839                 blk_start_plug(&plug);
 840                 flush_bio_list(conf, bio);
 841                 blk_finish_plug(&plug);
 842         } else
 843                 spin_unlock_irq(&conf->device_lock);
 844 }
 845
 846 /* Barriers....
 847  * Sometimes we need to suspend IO while we do something else,
 848  * either some resync/recovery, or reconfigure the array.
 849  * To do this we raise a 'barrier'.
 850  * The 'barrier' is a counter that can be raised multiple times
 851  * to count how many activities are happening which preclude
 852  * normal IO.
 853  * We can only raise the barrier if there is no pending IO.
 854  * i.e. if nr_pending == 0.
 855  * We choose only to raise the barrier if no-one is waiting for the
 856  * barrier to go down.  This means that as soon as an IO request
 857  * is ready, no other operations which require a barrier will start
 858  * until the IO request has had a chance.
 859  *
 860  * So: regular IO calls 'wait_barrier'.  When that returns there
 861  *    is no backgroup IO happening,  It must arrange to call
 862  *    allow_barrier when it has finished its IO.
 863  * backgroup IO calls must call raise_barrier.  Once that returns
 864  *    there is no normal IO happeing.  It must arrange to call
 865  *    lower_barrier when the particular background IO completes.
 866  *
 867  * If resync/recovery is interrupted, returns -EINTR;
 868  * Otherwise, returns 0.
 869  */
 870 static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
 871 {
 872         int idx = sector_to_idx(sector_nr);
 873
 874         spin_lock_irq(&conf->resync_lock);
 875
 876         /* Wait until no block IO is waiting */
 877         wait_event_lock_irq(conf->wait_barrier,
 878                             !atomic_read(&conf->nr_waiting[idx]),
 879                             conf->resync_lock);
 880
 881         /* block any new IO from starting */
 882         atomic_inc(&conf->barrier[idx]);
 883         /*
 884          * In raise_barrier() we firstly increase conf->barrier[idx] then
 885          * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 886          * increase conf->nr_pending[idx] then check conf->barrier[idx].
 887          * A memory barrier here to make sure conf->nr_pending[idx] won't
 888          * be fetched before conf->barrier[idx] is increased. Otherwise
 889          * there will be a race between raise_barrier() and _wait_barrier().
 890          */
 891         smp_mb__after_atomic();
 892
 893         /* For these conditions we must wait:
 894          * A: while the array is in frozen state
 895          * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 896          *    existing in corresponding I/O barrier bucket.
 897          * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 898          *    max resync count which allowed on current I/O barrier bucket.
 899          */
 900         wait_event_lock_irq(conf->wait_barrier,
 901                             (!conf->array_frozen &&
 902                              !atomic_read(&conf->nr_pending[idx]) &&
 903                              atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
 904                                 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
 905                             conf->resync_lock);
 906
 907         if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
 908                 atomic_dec(&conf->barrier[idx]);
 909                 spin_unlock_irq(&conf->resync_lock);
 910                 wake_up(&conf->wait_barrier);
 911                 return -EINTR;
 912         }
 913
 914         atomic_inc(&conf->nr_sync_pending);
 915         spin_unlock_irq(&conf->resync_lock);
 916
 917         return 0;
 918 }
 919
 920 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 921 {
 922         int idx = sector_to_idx(sector_nr);
 923
 924         BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 925
 926         atomic_dec(&conf->barrier[idx]);
 927         atomic_dec(&conf->nr_sync_pending);
 928         wake_up(&conf->wait_barrier);
 929 }
 930
 931 static bool _wait_barrier(struct r1conf *conf, int idx, bool nowait)
 932 {
 933         bool ret = true;
 934
 935         /*
 936          * We need to increase conf->nr_pending[idx] very early here,
 937          * then raise_barrier() can be blocked when it waits for
 938          * conf->nr_pending[idx] to be 0. Then we can avoid holding
 939          * conf->resync_lock when there is no barrier raised in same
 940          * barrier unit bucket. Also if the array is frozen, I/O
 941          * should be blocked until array is unfrozen.
 942          */
 943         atomic_inc(&conf->nr_pending[idx]);
 944         /*
 945          * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
 946          * check conf->barrier[idx]. In raise_barrier() we firstly increase
 947          * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
 948          * barrier is necessary here to make sure conf->barrier[idx] won't be
 949          * fetched before conf->nr_pending[idx] is increased. Otherwise there
 950          * will be a race between _wait_barrier() and raise_barrier().
 951          */
 952         smp_mb__after_atomic();
 953
 954         /*
 955          * Don't worry about checking two atomic_t variables at same time
 956          * here. If during we check conf->barrier[idx], the array is
 957          * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
 958          * 0, it is safe to return and make the I/O continue. Because the
 959          * array is frozen, all I/O returned here will eventually complete
 960          * or be queued, no race will happen. See code comment in
 961          * frozen_array().
 962          */
 963         if (!READ_ONCE(conf->array_frozen) &&
 964             !atomic_read(&conf->barrier[idx]))
 965                 return ret;
 966
 967         /*
 968          * After holding conf->resync_lock, conf->nr_pending[idx]
 969          * should be decreased before waiting for barrier to drop.
 970          * Otherwise, we may encounter a race condition because
 971          * raise_barrer() might be waiting for conf->nr_pending[idx]
 972          * to be 0 at same time.
 973          */
 974         spin_lock_irq(&conf->resync_lock);
 975         atomic_inc(&conf->nr_waiting[idx]);
 976         atomic_dec(&conf->nr_pending[idx]);
 977         /*
 978          * In case freeze_array() is waiting for
 979          * get_unqueued_pending() == extra
 980          */
 981         wake_up(&conf->wait_barrier);
 982         /* Wait for the barrier in same barrier unit bucket to drop. */
 983
 984         /* Return false when nowait flag is set */
 985         if (nowait) {
 986                 ret = false;
 987         } else {
 988                 wait_event_lock_irq(conf->wait_barrier,
 989                                 !conf->array_frozen &&
 990                                 !atomic_read(&conf->barrier[idx]),
 991                                 conf->resync_lock);
 992                 atomic_inc(&conf->nr_pending[idx]);
 993         }
 994
 995         atomic_dec(&conf->nr_waiting[idx]);
 996         spin_unlock_irq(&conf->resync_lock);
 997         return ret;
 998 }
 999
1000 static bool wait_read_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1001 {
1002         int idx = sector_to_idx(sector_nr);
1003         bool ret = true;
1004
1005         /*
1006          * Very similar to _wait_barrier(). The difference is, for read
1007          * I/O we don't need wait for sync I/O, but if the whole array
1008          * is frozen, the read I/O still has to wait until the array is
1009          * unfrozen. Since there is no ordering requirement with
1010          * conf->barrier[idx] here, memory barrier is unnecessary as well.
1011          */
1012         atomic_inc(&conf->nr_pending[idx]);
1013
1014         if (!READ_ONCE(conf->array_frozen))
1015                 return ret;
1016
1017         spin_lock_irq(&conf->resync_lock);
1018         atomic_inc(&conf->nr_waiting[idx]);
1019         atomic_dec(&conf->nr_pending[idx]);
1020         /*
1021          * In case freeze_array() is waiting for
1022          * get_unqueued_pending() == extra
1023          */
1024         wake_up(&conf->wait_barrier);
1025         /* Wait for array to be unfrozen */
1026
1027         /* Return false when nowait flag is set */
1028         if (nowait) {
1029                 /* Return false when nowait flag is set */
1030                 ret = false;
1031         } else {
1032                 wait_event_lock_irq(conf->wait_barrier,
1033                                 !conf->array_frozen,
1034                                 conf->resync_lock);
1035                 atomic_inc(&conf->nr_pending[idx]);
1036         }
1037
1038         atomic_dec(&conf->nr_waiting[idx]);
1039         spin_unlock_irq(&conf->resync_lock);
1040         return ret;
1041 }
1042
1043 static bool wait_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1044 {
1045         int idx = sector_to_idx(sector_nr);
1046
1047         return _wait_barrier(conf, idx, nowait);
1048 }
1049
1050 static void _allow_barrier(struct r1conf *conf, int idx)
1051 {
1052         atomic_dec(&conf->nr_pending[idx]);
1053         wake_up(&conf->wait_barrier);
1054 }
1055
1056 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1057 {
1058         int idx = sector_to_idx(sector_nr);
1059
1060         _allow_barrier(conf, idx);
1061 }
1062
1063 /* conf->resync_lock should be held */
1064 static int get_unqueued_pending(struct r1conf *conf)
1065 {
1066         int idx, ret;
1067
1068         ret = atomic_read(&conf->nr_sync_pending);
1069         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1070                 ret += atomic_read(&conf->nr_pending[idx]) -
1071                         atomic_read(&conf->nr_queued[idx]);
1072
1073         return ret;
1074 }
1075
1076 static void freeze_array(struct r1conf *conf, int extra)
1077 {
1078         /* Stop sync I/O and normal I/O and wait for everything to
1079          * go quiet.
1080          * This is called in two situations:
1081          * 1) management command handlers (reshape, remove disk, quiesce).
1082          * 2) one normal I/O request failed.
1083
1084          * After array_frozen is set to 1, new sync IO will be blocked at
1085          * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1086          * or wait_read_barrier(). The flying I/Os will either complete or be
1087          * queued. When everything goes quite, there are only queued I/Os left.
1088
1089          * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1090          * barrier bucket index which this I/O request hits. When all sync and
1091          * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1092          * of all conf->nr_queued[]. But normal I/O failure is an exception,
1093          * in handle_read_error(), we may call freeze_array() before trying to
1094          * fix the read error. In this case, the error read I/O is not queued,
1095          * so get_unqueued_pending() == 1.
1096          *
1097          * Therefore before this function returns, we need to wait until
1098          * get_unqueued_pendings(conf) gets equal to extra. For
1099          * normal I/O context, extra is 1, in rested situations extra is 0.
1100          */
1101         spin_lock_irq(&conf->resync_lock);
1102         conf->array_frozen = 1;
1103         raid1_log(conf->mddev, "wait freeze");
1104         wait_event_lock_irq_cmd(
1105                 conf->wait_barrier,
1106                 get_unqueued_pending(conf) == extra,
1107                 conf->resync_lock,
1108                 flush_pending_writes(conf));
1109         spin_unlock_irq(&conf->resync_lock);
1110 }
1111 static void unfreeze_array(struct r1conf *conf)
1112 {
1113         /* reverse the effect of the freeze */
1114         spin_lock_irq(&conf->resync_lock);
1115         conf->array_frozen = 0;
1116         spin_unlock_irq(&conf->resync_lock);
1117         wake_up(&conf->wait_barrier);
1118 }
1119
1120 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1121                                            struct bio *bio)
1122 {
1123         int size = bio->bi_iter.bi_size;
1124         unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1125         int i = 0;
1126         struct bio *behind_bio = NULL;
1127
1128         behind_bio = bio_alloc_bioset(NULL, vcnt, 0, GFP_NOIO,
1129                                       &r1_bio->mddev->bio_set);
1130         if (!behind_bio)
1131                 return;
1132
1133         /* discard op, we don't support writezero/writesame yet */
1134         if (!bio_has_data(bio)) {
1135                 behind_bio->bi_iter.bi_size = size;
1136                 goto skip_copy;
1137         }
1138
1139         while (i < vcnt && size) {
1140                 struct page *page;
1141                 int len = min_t(int, PAGE_SIZE, size);
1142
1143                 page = alloc_page(GFP_NOIO);
1144                 if (unlikely(!page))
1145                         goto free_pages;
1146
1147                 bio_add_page(behind_bio, page, len, 0);
1148
1149                 size -= len;
1150                 i++;
1151         }
1152
1153         bio_copy_data(behind_bio, bio);
1154 skip_copy:
1155         r1_bio->behind_master_bio = behind_bio;
1156         set_bit(R1BIO_BehindIO, &r1_bio->state);
1157
1158         return;
1159
1160 free_pages:
1161         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1162                  bio->bi_iter.bi_size);
1163         bio_free_pages(behind_bio);
1164         bio_put(behind_bio);
1165 }
1166
1167 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1168 {
1169         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1170                                                   cb);
1171         struct mddev *mddev = plug->cb.data;
1172         struct r1conf *conf = mddev->private;
1173         struct bio *bio;
1174
1175         if (from_schedule || current->bio_list) {
1176                 spin_lock_irq(&conf->device_lock);
1177                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1178                 spin_unlock_irq(&conf->device_lock);
1179                 wake_up(&conf->wait_barrier);
1180                 md_wakeup_thread(mddev->thread);
1181                 kfree(plug);
1182                 return;
1183         }
1184
1185         /* we aren't scheduling, so we can do the write-out directly. */
1186         bio = bio_list_get(&plug->pending);
1187         flush_bio_list(conf, bio);
1188         kfree(plug);
1189 }
1190
1191 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1192 {
1193         r1_bio->master_bio = bio;
1194         r1_bio->sectors = bio_sectors(bio);
1195         r1_bio->state = 0;
1196         r1_bio->mddev = mddev;
1197         r1_bio->sector = bio->bi_iter.bi_sector;
1198 }
1199
1200 static inline struct r1bio *
1201 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1202 {
1203         struct r1conf *conf = mddev->private;
1204         struct r1bio *r1_bio;
1205
1206         r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1207         /* Ensure no bio records IO_BLOCKED */
1208         memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1209         init_r1bio(r1_bio, mddev, bio);
1210         return r1_bio;
1211 }
1212
1213 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1214                                int max_read_sectors, struct r1bio *r1_bio)
1215 {
1216         struct r1conf *conf = mddev->private;
1217         struct raid1_info *mirror;
1218         struct bio *read_bio;
1219         struct bitmap *bitmap = mddev->bitmap;
1220         const int op = bio_op(bio);
1221         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1222         int max_sectors;
1223         int rdisk;
1224         bool r1bio_existed = !!r1_bio;
1225         char b[BDEVNAME_SIZE];
1226
1227         /*
1228          * If r1_bio is set, we are blocking the raid1d thread
1229          * so there is a tiny risk of deadlock.  So ask for
1230          * emergency memory if needed.
1231          */
1232         gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1233
1234         if (r1bio_existed) {
1235                 /* Need to get the block device name carefully */
1236                 struct md_rdev *rdev;
1237                 rcu_read_lock();
1238                 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1239                 if (rdev)
1240                         bdevname(rdev->bdev, b);
1241                 else
1242                         strcpy(b, "???");
1243                 rcu_read_unlock();
1244         }
1245
1246         /*
1247          * Still need barrier for READ in case that whole
1248          * array is frozen.
1249          */
1250         if (!wait_read_barrier(conf, bio->bi_iter.bi_sector,
1251                                 bio->bi_opf & REQ_NOWAIT)) {
1252                 bio_wouldblock_error(bio);
1253                 return;
1254         }
1255
1256         if (!r1_bio)
1257                 r1_bio = alloc_r1bio(mddev, bio);
1258         else
1259                 init_r1bio(r1_bio, mddev, bio);
1260         r1_bio->sectors = max_read_sectors;
1261
1262         /*
1263          * make_request() can abort the operation when read-ahead is being
1264          * used and no empty request is available.
1265          */
1266         rdisk = read_balance(conf, r1_bio, &max_sectors);
1267
1268         if (rdisk < 0) {
1269                 /* couldn't find anywhere to read from */
1270                 if (r1bio_existed) {
1271                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1272                                             mdname(mddev),
1273                                             b,
1274                                             (unsigned long long)r1_bio->sector);
1275                 }
1276                 raid_end_bio_io(r1_bio);
1277                 return;
1278         }
1279         mirror = conf->mirrors + rdisk;
1280
1281         if (r1bio_existed)
1282                 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1283                                     mdname(mddev),
1284                                     (unsigned long long)r1_bio->sector,
1285                                     bdevname(mirror->rdev->bdev, b));
1286
1287         if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1288             bitmap) {
1289                 /*
1290                  * Reading from a write-mostly device must take care not to
1291                  * over-take any writes that are 'behind'
1292                  */
1293                 raid1_log(mddev, "wait behind writes");
1294                 wait_event(bitmap->behind_wait,
1295                            atomic_read(&bitmap->behind_writes) == 0);
1296         }
1297
1298         if (max_sectors < bio_sectors(bio)) {
1299                 struct bio *split = bio_split(bio, max_sectors,
1300                                               gfp, &conf->bio_split);
1301                 bio_chain(split, bio);
1302                 submit_bio_noacct(bio);
1303                 bio = split;
1304                 r1_bio->master_bio = bio;
1305                 r1_bio->sectors = max_sectors;
1306         }
1307
1308         r1_bio->read_disk = rdisk;
1309
1310         if (!r1bio_existed && blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1311                 r1_bio->start_time = bio_start_io_acct(bio);
1312
1313         read_bio = bio_alloc_clone(mirror->rdev->bdev, bio, gfp,
1314                                    &mddev->bio_set);
1315
1316         r1_bio->bios[rdisk] = read_bio;
1317
1318         read_bio->bi_iter.bi_sector = r1_bio->sector +
1319                 mirror->rdev->data_offset;
1320         read_bio->bi_end_io = raid1_end_read_request;
1321         bio_set_op_attrs(read_bio, op, do_sync);
1322         if (test_bit(FailFast, &mirror->rdev->flags) &&
1323             test_bit(R1BIO_FailFast, &r1_bio->state))
1324                 read_bio->bi_opf |= MD_FAILFAST;
1325         read_bio->bi_private = r1_bio;
1326
1327         if (mddev->gendisk)
1328                 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1329                                       r1_bio->sector);
1330
1331         submit_bio_noacct(read_bio);
1332 }
1333
1334 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1335                                 int max_write_sectors)
1336 {
1337         struct r1conf *conf = mddev->private;
1338         struct r1bio *r1_bio;
1339         int i, disks;
1340         struct bitmap *bitmap = mddev->bitmap;
1341         unsigned long flags;
1342         struct md_rdev *blocked_rdev;
1343         struct blk_plug_cb *cb;
1344         struct raid1_plug_cb *plug = NULL;
1345         int first_clone;
1346         int max_sectors;
1347         bool write_behind = false;
1348
1349         if (mddev_is_clustered(mddev) &&
1350              md_cluster_ops->area_resyncing(mddev, WRITE,
1351                      bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1352
1353                 DEFINE_WAIT(w);
1354                 if (bio->bi_opf & REQ_NOWAIT) {
1355                         bio_wouldblock_error(bio);
1356                         return;
1357                 }
1358                 for (;;) {
1359                         prepare_to_wait(&conf->wait_barrier,
1360                                         &w, TASK_IDLE);
1361                         if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1362                                                         bio->bi_iter.bi_sector,
1363                                                         bio_end_sector(bio)))
1364                                 break;
1365                         schedule();
1366                 }
1367                 finish_wait(&conf->wait_barrier, &w);
1368         }
1369
1370         /*
1371          * Register the new request and wait if the reconstruction
1372          * thread has put up a bar for new requests.
1373          * Continue immediately if no resync is active currently.
1374          */
1375         if (!wait_barrier(conf, bio->bi_iter.bi_sector,
1376                                 bio->bi_opf & REQ_NOWAIT)) {
1377                 bio_wouldblock_error(bio);
1378                 return;
1379         }
1380
1381         r1_bio = alloc_r1bio(mddev, bio);
1382         r1_bio->sectors = max_write_sectors;
1383
1384         /* first select target devices under rcu_lock and
1385          * inc refcount on their rdev.  Record them by setting
1386          * bios[x] to bio
1387          * If there are known/acknowledged bad blocks on any device on
1388          * which we have seen a write error, we want to avoid writing those
1389          * blocks.
1390          * This potentially requires several writes to write around
1391          * the bad blocks.  Each set of writes gets it's own r1bio
1392          * with a set of bios attached.
1393          */
1394
1395         disks = conf->raid_disks * 2;
1396  retry_write:
1397         blocked_rdev = NULL;
1398         rcu_read_lock();
1399         max_sectors = r1_bio->sectors;
1400         for (i = 0;  i < disks; i++) {
1401                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1402
1403                 /*
1404                  * The write-behind io is only attempted on drives marked as
1405                  * write-mostly, which means we could allocate write behind
1406                  * bio later.
1407                  */
1408                 if (rdev && test_bit(WriteMostly, &rdev->flags))
1409                         write_behind = true;
1410
1411                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1412                         atomic_inc(&rdev->nr_pending);
1413                         blocked_rdev = rdev;
1414                         break;
1415                 }
1416                 r1_bio->bios[i] = NULL;
1417                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1418                         if (i < conf->raid_disks)
1419                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1420                         continue;
1421                 }
1422
1423                 atomic_inc(&rdev->nr_pending);
1424                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1425                         sector_t first_bad;
1426                         int bad_sectors;
1427                         int is_bad;
1428
1429                         is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1430                                              &first_bad, &bad_sectors);
1431                         if (is_bad < 0) {
1432                                 /* mustn't write here until the bad block is
1433                                  * acknowledged*/
1434                                 set_bit(BlockedBadBlocks, &rdev->flags);
1435                                 blocked_rdev = rdev;
1436                                 break;
1437                         }
1438                         if (is_bad && first_bad <= r1_bio->sector) {
1439                                 /* Cannot write here at all */
1440                                 bad_sectors -= (r1_bio->sector - first_bad);
1441                                 if (bad_sectors < max_sectors)
1442                                         /* mustn't write more than bad_sectors
1443                                          * to other devices yet
1444                                          */
1445                                         max_sectors = bad_sectors;
1446                                 rdev_dec_pending(rdev, mddev);
1447                                 /* We don't set R1BIO_Degraded as that
1448                                  * only applies if the disk is
1449                                  * missing, so it might be re-added,
1450                                  * and we want to know to recover this
1451                                  * chunk.
1452                                  * In this case the device is here,
1453                                  * and the fact that this chunk is not
1454                                  * in-sync is recorded in the bad
1455                                  * block log
1456                                  */
1457                                 continue;
1458                         }
1459                         if (is_bad) {
1460                                 int good_sectors = first_bad - r1_bio->sector;
1461                                 if (good_sectors < max_sectors)
1462                                         max_sectors = good_sectors;
1463                         }
1464                 }
1465                 r1_bio->bios[i] = bio;
1466         }
1467         rcu_read_unlock();
1468
1469         if (unlikely(blocked_rdev)) {
1470                 /* Wait for this device to become unblocked */
1471                 int j;
1472
1473                 for (j = 0; j < i; j++)
1474                         if (r1_bio->bios[j])
1475                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1476                 r1_bio->state = 0;
1477                 allow_barrier(conf, bio->bi_iter.bi_sector);
1478
1479                 if (bio->bi_opf & REQ_NOWAIT) {
1480                         bio_wouldblock_error(bio);
1481                         return;
1482                 }
1483                 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1484                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1485                 wait_barrier(conf, bio->bi_iter.bi_sector, false);
1486                 goto retry_write;
1487         }
1488
1489         /*
1490          * When using a bitmap, we may call alloc_behind_master_bio below.
1491          * alloc_behind_master_bio allocates a copy of the data payload a page
1492          * at a time and thus needs a new bio that can fit the whole payload
1493          * this bio in page sized chunks.
1494          */
1495         if (write_behind && bitmap)
1496                 max_sectors = min_t(int, max_sectors,
1497                                     BIO_MAX_VECS * (PAGE_SIZE >> 9));
1498         if (max_sectors < bio_sectors(bio)) {
1499                 struct bio *split = bio_split(bio, max_sectors,
1500                                               GFP_NOIO, &conf->bio_split);
1501                 bio_chain(split, bio);
1502                 submit_bio_noacct(bio);
1503                 bio = split;
1504                 r1_bio->master_bio = bio;
1505                 r1_bio->sectors = max_sectors;
1506         }
1507
1508         if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1509                 r1_bio->start_time = bio_start_io_acct(bio);
1510         atomic_set(&r1_bio->remaining, 1);
1511         atomic_set(&r1_bio->behind_remaining, 0);
1512
1513         first_clone = 1;
1514
1515         for (i = 0; i < disks; i++) {
1516                 struct bio *mbio = NULL;
1517                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1518                 if (!r1_bio->bios[i])
1519                         continue;
1520
1521                 if (first_clone) {
1522                         /* do behind I/O ?
1523                          * Not if there are too many, or cannot
1524                          * allocate memory, or a reader on WriteMostly
1525                          * is waiting for behind writes to flush */
1526                         if (bitmap &&
1527                             test_bit(WriteMostly, &rdev->flags) &&
1528                             (atomic_read(&bitmap->behind_writes)
1529                              < mddev->bitmap_info.max_write_behind) &&
1530                             !waitqueue_active(&bitmap->behind_wait)) {
1531                                 alloc_behind_master_bio(r1_bio, bio);
1532                         }
1533
1534                         md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1535                                              test_bit(R1BIO_BehindIO, &r1_bio->state));
1536                         first_clone = 0;
1537                 }
1538
1539                 if (r1_bio->behind_master_bio) {
1540                         mbio = bio_alloc_clone(rdev->bdev,
1541                                                r1_bio->behind_master_bio,
1542                                                GFP_NOIO, &mddev->bio_set);
1543                         if (test_bit(CollisionCheck, &rdev->flags))
1544                                 wait_for_serialization(rdev, r1_bio);
1545                         if (test_bit(WriteMostly, &rdev->flags))
1546                                 atomic_inc(&r1_bio->behind_remaining);
1547                 } else {
1548                         mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
1549                                                &mddev->bio_set);
1550
1551                         if (mddev->serialize_policy)
1552                                 wait_for_serialization(rdev, r1_bio);
1553                 }
1554
1555                 r1_bio->bios[i] = mbio;
1556
1557                 mbio->bi_iter.bi_sector = (r1_bio->sector + rdev->data_offset);
1558                 mbio->bi_end_io = raid1_end_write_request;
1559                 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1560                 if (test_bit(FailFast, &rdev->flags) &&
1561                     !test_bit(WriteMostly, &rdev->flags) &&
1562                     conf->raid_disks - mddev->degraded > 1)
1563                         mbio->bi_opf |= MD_FAILFAST;
1564                 mbio->bi_private = r1_bio;
1565
1566                 atomic_inc(&r1_bio->remaining);
1567
1568                 if (mddev->gendisk)
1569                         trace_block_bio_remap(mbio, disk_devt(mddev->gendisk),
1570                                               r1_bio->sector);
1571                 /* flush_pending_writes() needs access to the rdev so...*/
1572                 mbio->bi_bdev = (void *)rdev;
1573
1574                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1575                 if (cb)
1576                         plug = container_of(cb, struct raid1_plug_cb, cb);
1577                 else
1578                         plug = NULL;
1579                 if (plug) {
1580                         bio_list_add(&plug->pending, mbio);
1581                 } else {
1582                         spin_lock_irqsave(&conf->device_lock, flags);
1583                         bio_list_add(&conf->pending_bio_list, mbio);
1584                         spin_unlock_irqrestore(&conf->device_lock, flags);
1585                         md_wakeup_thread(mddev->thread);
1586                 }
1587         }
1588
1589         r1_bio_write_done(r1_bio);
1590
1591         /* In case raid1d snuck in to freeze_array */
1592         wake_up(&conf->wait_barrier);
1593 }
1594
1595 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1596 {
1597         sector_t sectors;
1598
1599         if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1600             && md_flush_request(mddev, bio))
1601                 return true;
1602
1603         /*
1604          * There is a limit to the maximum size, but
1605          * the read/write handler might find a lower limit
1606          * due to bad blocks.  To avoid multiple splits,
1607          * we pass the maximum number of sectors down
1608          * and let the lower level perform the split.
1609          */
1610         sectors = align_to_barrier_unit_end(
1611                 bio->bi_iter.bi_sector, bio_sectors(bio));
1612
1613         if (bio_data_dir(bio) == READ)
1614                 raid1_read_request(mddev, bio, sectors, NULL);
1615         else {
1616                 if (!md_write_start(mddev,bio))
1617                         return false;
1618                 raid1_write_request(mddev, bio, sectors);
1619         }
1620         return true;
1621 }
1622
1623 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1624 {
1625         struct r1conf *conf = mddev->private;
1626         int i;
1627
1628         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1629                    conf->raid_disks - mddev->degraded);
1630         rcu_read_lock();
1631         for (i = 0; i < conf->raid_disks; i++) {
1632                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1633                 seq_printf(seq, "%s",
1634                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1635         }
1636         rcu_read_unlock();
1637         seq_printf(seq, "]");
1638 }
1639
1640 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1641 {
1642         char b[BDEVNAME_SIZE];
1643         struct r1conf *conf = mddev->private;
1644         unsigned long flags;
1645
1646         /*
1647          * If it is not operational, then we have already marked it as dead
1648          * else if it is the last working disks with "fail_last_dev == false",
1649          * ignore the error, let the next level up know.
1650          * else mark the drive as failed
1651          */
1652         spin_lock_irqsave(&conf->device_lock, flags);
1653         if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1654             && (conf->raid_disks - mddev->degraded) == 1) {
1655                 /*
1656                  * Don't fail the drive, act as though we were just a
1657                  * normal single drive.
1658                  * However don't try a recovery from this drive as
1659                  * it is very likely to fail.
1660                  */
1661                 conf->recovery_disabled = mddev->recovery_disabled;
1662                 spin_unlock_irqrestore(&conf->device_lock, flags);
1663                 return;
1664         }
1665         set_bit(Blocked, &rdev->flags);
1666         if (test_and_clear_bit(In_sync, &rdev->flags))
1667                 mddev->degraded++;
1668         set_bit(Faulty, &rdev->flags);
1669         spin_unlock_irqrestore(&conf->device_lock, flags);
1670         /*
1671          * if recovery is running, make sure it aborts.
1672          */
1673         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1674         set_mask_bits(&mddev->sb_flags, 0,
1675                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1676         pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1677                 "md/raid1:%s: Operation continuing on %d devices.\n",
1678                 mdname(mddev), bdevname(rdev->bdev, b),
1679                 mdname(mddev), conf->raid_disks - mddev->degraded);
1680 }
1681
1682 static void print_conf(struct r1conf *conf)
1683 {
1684         int i;
1685
1686         pr_debug("RAID1 conf printout:\n");
1687         if (!conf) {
1688                 pr_debug("(!conf)\n");
1689                 return;
1690         }
1691         pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1692                  conf->raid_disks);
1693
1694         rcu_read_lock();
1695         for (i = 0; i < conf->raid_disks; i++) {
1696                 char b[BDEVNAME_SIZE];
1697                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1698                 if (rdev)
1699                         pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1700                                  i, !test_bit(In_sync, &rdev->flags),
1701                                  !test_bit(Faulty, &rdev->flags),
1702                                  bdevname(rdev->bdev,b));
1703         }
1704         rcu_read_unlock();
1705 }
1706
1707 static void close_sync(struct r1conf *conf)
1708 {
1709         int idx;
1710
1711         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1712                 _wait_barrier(conf, idx, false);
1713                 _allow_barrier(conf, idx);
1714         }
1715
1716         mempool_exit(&conf->r1buf_pool);
1717 }
1718
1719 static int raid1_spare_active(struct mddev *mddev)
1720 {
1721         int i;
1722         struct r1conf *conf = mddev->private;
1723         int count = 0;
1724         unsigned long flags;
1725
1726         /*
1727          * Find all failed disks within the RAID1 configuration
1728          * and mark them readable.
1729          * Called under mddev lock, so rcu protection not needed.
1730          * device_lock used to avoid races with raid1_end_read_request
1731          * which expects 'In_sync' flags and ->degraded to be consistent.
1732          */
1733         spin_lock_irqsave(&conf->device_lock, flags);
1734         for (i = 0; i < conf->raid_disks; i++) {
1735                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1736                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1737                 if (repl
1738                     && !test_bit(Candidate, &repl->flags)
1739                     && repl->recovery_offset == MaxSector
1740                     && !test_bit(Faulty, &repl->flags)
1741                     && !test_and_set_bit(In_sync, &repl->flags)) {
1742                         /* replacement has just become active */
1743                         if (!rdev ||
1744                             !test_and_clear_bit(In_sync, &rdev->flags))
1745                                 count++;
1746                         if (rdev) {
1747                                 /* Replaced device not technically
1748                                  * faulty, but we need to be sure
1749                                  * it gets removed and never re-added
1750                                  */
1751                                 set_bit(Faulty, &rdev->flags);
1752                                 sysfs_notify_dirent_safe(
1753                                         rdev->sysfs_state);
1754                         }
1755                 }
1756                 if (rdev
1757                     && rdev->recovery_offset == MaxSector
1758                     && !test_bit(Faulty, &rdev->flags)
1759                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1760                         count++;
1761                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1762                 }
1763         }
1764         mddev->degraded -= count;
1765         spin_unlock_irqrestore(&conf->device_lock, flags);
1766
1767         print_conf(conf);
1768         return count;
1769 }
1770
1771 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1772 {
1773         struct r1conf *conf = mddev->private;
1774         int err = -EEXIST;
1775         int mirror = 0;
1776         struct raid1_info *p;
1777         int first = 0;
1778         int last = conf->raid_disks - 1;
1779
1780         if (mddev->recovery_disabled == conf->recovery_disabled)
1781                 return -EBUSY;
1782
1783         if (md_integrity_add_rdev(rdev, mddev))
1784                 return -ENXIO;
1785
1786         if (rdev->raid_disk >= 0)
1787                 first = last = rdev->raid_disk;
1788
1789         /*
1790          * find the disk ... but prefer rdev->saved_raid_disk
1791          * if possible.
1792          */
1793         if (rdev->saved_raid_disk >= 0 &&
1794             rdev->saved_raid_disk >= first &&
1795             rdev->saved_raid_disk < conf->raid_disks &&
1796             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1797                 first = last = rdev->saved_raid_disk;
1798
1799         for (mirror = first; mirror <= last; mirror++) {
1800                 p = conf->mirrors + mirror;
1801                 if (!p->rdev) {
1802                         if (mddev->gendisk)
1803                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1804                                                   rdev->data_offset << 9);
1805
1806                         p->head_position = 0;
1807                         rdev->raid_disk = mirror;
1808                         err = 0;
1809                         /* As all devices are equivalent, we don't need a full recovery
1810                          * if this was recently any drive of the array
1811                          */
1812                         if (rdev->saved_raid_disk < 0)
1813                                 conf->fullsync = 1;
1814                         rcu_assign_pointer(p->rdev, rdev);
1815                         break;
1816                 }
1817                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1818                     p[conf->raid_disks].rdev == NULL) {
1819                         /* Add this device as a replacement */
1820                         clear_bit(In_sync, &rdev->flags);
1821                         set_bit(Replacement, &rdev->flags);
1822                         rdev->raid_disk = mirror;
1823                         err = 0;
1824                         conf->fullsync = 1;
1825                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1826                         break;
1827                 }
1828         }
1829         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1830                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1831         print_conf(conf);
1832         return err;
1833 }
1834
1835 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1836 {
1837         struct r1conf *conf = mddev->private;
1838         int err = 0;
1839         int number = rdev->raid_disk;
1840         struct raid1_info *p = conf->mirrors + number;
1841
1842         if (rdev != p->rdev)
1843                 p = conf->mirrors + conf->raid_disks + number;
1844
1845         print_conf(conf);
1846         if (rdev == p->rdev) {
1847                 if (test_bit(In_sync, &rdev->flags) ||
1848                     atomic_read(&rdev->nr_pending)) {
1849                         err = -EBUSY;
1850                         goto abort;
1851                 }
1852                 /* Only remove non-faulty devices if recovery
1853                  * is not possible.
1854                  */
1855                 if (!test_bit(Faulty, &rdev->flags) &&
1856                     mddev->recovery_disabled != conf->recovery_disabled &&
1857                     mddev->degraded < conf->raid_disks) {
1858                         err = -EBUSY;
1859                         goto abort;
1860                 }
1861                 p->rdev = NULL;
1862                 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1863                         synchronize_rcu();
1864                         if (atomic_read(&rdev->nr_pending)) {
1865                                 /* lost the race, try later */
1866                                 err = -EBUSY;
1867                                 p->rdev = rdev;
1868                                 goto abort;
1869                         }
1870                 }
1871                 if (conf->mirrors[conf->raid_disks + number].rdev) {
1872                         /* We just removed a device that is being replaced.
1873                          * Move down the replacement.  We drain all IO before
1874                          * doing this to avoid confusion.
1875                          */
1876                         struct md_rdev *repl =
1877                                 conf->mirrors[conf->raid_disks + number].rdev;
1878                         freeze_array(conf, 0);
1879                         if (atomic_read(&repl->nr_pending)) {
1880                                 /* It means that some queued IO of retry_list
1881                                  * hold repl. Thus, we cannot set replacement
1882                                  * as NULL, avoiding rdev NULL pointer
1883                                  * dereference in sync_request_write and
1884                                  * handle_write_finished.
1885                                  */
1886                                 err = -EBUSY;
1887                                 unfreeze_array(conf);
1888                                 goto abort;
1889                         }
1890                         clear_bit(Replacement, &repl->flags);
1891                         p->rdev = repl;
1892                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1893                         unfreeze_array(conf);
1894                 }
1895
1896                 clear_bit(WantReplacement, &rdev->flags);
1897                 err = md_integrity_register(mddev);
1898         }
1899 abort:
1900
1901         print_conf(conf);
1902         return err;
1903 }
1904
1905 static void end_sync_read(struct bio *bio)
1906 {
1907         struct r1bio *r1_bio = get_resync_r1bio(bio);
1908
1909         update_head_pos(r1_bio->read_disk, r1_bio);
1910
1911         /*
1912          * we have read a block, now it needs to be re-written,
1913          * or re-read if the read failed.
1914          * We don't do much here, just schedule handling by raid1d
1915          */
1916         if (!bio->bi_status)
1917                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1918
1919         if (atomic_dec_and_test(&r1_bio->remaining))
1920                 reschedule_retry(r1_bio);
1921 }
1922
1923 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1924 {
1925         sector_t sync_blocks = 0;
1926         sector_t s = r1_bio->sector;
1927         long sectors_to_go = r1_bio->sectors;
1928
1929         /* make sure these bits don't get cleared. */
1930         do {
1931                 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1932                 s += sync_blocks;
1933                 sectors_to_go -= sync_blocks;
1934         } while (sectors_to_go > 0);
1935 }
1936
1937 static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1938 {
1939         if (atomic_dec_and_test(&r1_bio->remaining)) {
1940                 struct mddev *mddev = r1_bio->mddev;
1941                 int s = r1_bio->sectors;
1942
1943                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1944                     test_bit(R1BIO_WriteError, &r1_bio->state))
1945                         reschedule_retry(r1_bio);
1946                 else {
1947                         put_buf(r1_bio);
1948                         md_done_sync(mddev, s, uptodate);
1949                 }
1950         }
1951 }
1952
1953 static void end_sync_write(struct bio *bio)
1954 {
1955         int uptodate = !bio->bi_status;
1956         struct r1bio *r1_bio = get_resync_r1bio(bio);
1957         struct mddev *mddev = r1_bio->mddev;
1958         struct r1conf *conf = mddev->private;
1959         sector_t first_bad;
1960         int bad_sectors;
1961         struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1962
1963         if (!uptodate) {
1964                 abort_sync_write(mddev, r1_bio);
1965                 set_bit(WriteErrorSeen, &rdev->flags);
1966                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1967                         set_bit(MD_RECOVERY_NEEDED, &
1968                                 mddev->recovery);
1969                 set_bit(R1BIO_WriteError, &r1_bio->state);
1970         } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1971                                &first_bad, &bad_sectors) &&
1972                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1973                                 r1_bio->sector,
1974                                 r1_bio->sectors,
1975                                 &first_bad, &bad_sectors)
1976                 )
1977                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1978
1979         put_sync_write_buf(r1_bio, uptodate);
1980 }
1981
1982 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1983                             int sectors, struct page *page, int rw)
1984 {
1985         if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1986                 /* success */
1987                 return 1;
1988         if (rw == WRITE) {
1989                 set_bit(WriteErrorSeen, &rdev->flags);
1990                 if (!test_and_set_bit(WantReplacement,
1991                                       &rdev->flags))
1992                         set_bit(MD_RECOVERY_NEEDED, &
1993                                 rdev->mddev->recovery);
1994         }
1995         /* need to record an error - either for the block or the device */
1996         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1997                 md_error(rdev->mddev, rdev);
1998         return 0;
1999 }
2000
2001 static int fix_sync_read_error(struct r1bio *r1_bio)
2002 {
2003         /* Try some synchronous reads of other devices to get
2004          * good data, much like with normal read errors.  Only
2005          * read into the pages we already have so we don't
2006          * need to re-issue the read request.
2007          * We don't need to freeze the array, because being in an
2008          * active sync request, there is no normal IO, and
2009          * no overlapping syncs.
2010          * We don't need to check is_badblock() again as we
2011          * made sure that anything with a bad block in range
2012          * will have bi_end_io clear.
2013          */
2014         struct mddev *mddev = r1_bio->mddev;
2015         struct r1conf *conf = mddev->private;
2016         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
2017         struct page **pages = get_resync_pages(bio)->pages;
2018         sector_t sect = r1_bio->sector;
2019         int sectors = r1_bio->sectors;
2020         int idx = 0;
2021         struct md_rdev *rdev;
2022
2023         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2024         if (test_bit(FailFast, &rdev->flags)) {
2025                 /* Don't try recovering from here - just fail it
2026                  * ... unless it is the last working device of course */
2027                 md_error(mddev, rdev);
2028                 if (test_bit(Faulty, &rdev->flags))
2029                         /* Don't try to read from here, but make sure
2030                          * put_buf does it's thing
2031                          */
2032                         bio->bi_end_io = end_sync_write;
2033         }
2034
2035         while(sectors) {
2036                 int s = sectors;
2037                 int d = r1_bio->read_disk;
2038                 int success = 0;
2039                 int start;
2040
2041                 if (s > (PAGE_SIZE>>9))
2042                         s = PAGE_SIZE >> 9;
2043                 do {
2044                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2045                                 /* No rcu protection needed here devices
2046                                  * can only be removed when no resync is
2047                                  * active, and resync is currently active
2048                                  */
2049                                 rdev = conf->mirrors[d].rdev;
2050                                 if (sync_page_io(rdev, sect, s<<9,
2051                                                  pages[idx],
2052                                                  REQ_OP_READ, 0, false)) {
2053                                         success = 1;
2054                                         break;
2055                                 }
2056                         }
2057                         d++;
2058                         if (d == conf->raid_disks * 2)
2059                                 d = 0;
2060                 } while (!success && d != r1_bio->read_disk);
2061
2062                 if (!success) {
2063                         int abort = 0;
2064                         /* Cannot read from anywhere, this block is lost.
2065                          * Record a bad block on each device.  If that doesn't
2066                          * work just disable and interrupt the recovery.
2067                          * Don't fail devices as that won't really help.
2068                          */
2069                         pr_crit_ratelimited("md/raid1:%s: %pg: unrecoverable I/O read error for block %llu\n",
2070                                             mdname(mddev), bio->bi_bdev,
2071                                             (unsigned long long)r1_bio->sector);
2072                         for (d = 0; d < conf->raid_disks * 2; d++) {
2073                                 rdev = conf->mirrors[d].rdev;
2074                                 if (!rdev || test_bit(Faulty, &rdev->flags))
2075                                         continue;
2076                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
2077                                         abort = 1;
2078                         }
2079                         if (abort) {
2080                                 conf->recovery_disabled =
2081                                         mddev->recovery_disabled;
2082                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2083                                 md_done_sync(mddev, r1_bio->sectors, 0);
2084                                 put_buf(r1_bio);
2085                                 return 0;
2086                         }
2087                         /* Try next page */
2088                         sectors -= s;
2089                         sect += s;
2090                         idx++;
2091                         continue;
2092                 }
2093
2094                 start = d;
2095                 /* write it back and re-read */
2096                 while (d != r1_bio->read_disk) {
2097                         if (d == 0)
2098                                 d = conf->raid_disks * 2;
2099                         d--;
2100                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2101                                 continue;
2102                         rdev = conf->mirrors[d].rdev;
2103                         if (r1_sync_page_io(rdev, sect, s,
2104                                             pages[idx],
2105                                             WRITE) == 0) {
2106                                 r1_bio->bios[d]->bi_end_io = NULL;
2107                                 rdev_dec_pending(rdev, mddev);
2108                         }
2109                 }
2110                 d = start;
2111                 while (d != r1_bio->read_disk) {
2112                         if (d == 0)
2113                                 d = conf->raid_disks * 2;
2114                         d--;
2115                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2116                                 continue;
2117                         rdev = conf->mirrors[d].rdev;
2118                         if (r1_sync_page_io(rdev, sect, s,
2119                                             pages[idx],
2120                                             READ) != 0)
2121                                 atomic_add(s, &rdev->corrected_errors);
2122                 }
2123                 sectors -= s;
2124                 sect += s;
2125                 idx ++;
2126         }
2127         set_bit(R1BIO_Uptodate, &r1_bio->state);
2128         bio->bi_status = 0;
2129         return 1;
2130 }
2131
2132 static void process_checks(struct r1bio *r1_bio)
2133 {
2134         /* We have read all readable devices.  If we haven't
2135          * got the block, then there is no hope left.
2136          * If we have, then we want to do a comparison
2137          * and skip the write if everything is the same.
2138          * If any blocks failed to read, then we need to
2139          * attempt an over-write
2140          */
2141         struct mddev *mddev = r1_bio->mddev;
2142         struct r1conf *conf = mddev->private;
2143         int primary;
2144         int i;
2145         int vcnt;
2146
2147         /* Fix variable parts of all bios */
2148         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2149         for (i = 0; i < conf->raid_disks * 2; i++) {
2150                 blk_status_t status;
2151                 struct bio *b = r1_bio->bios[i];
2152                 struct resync_pages *rp = get_resync_pages(b);
2153                 if (b->bi_end_io != end_sync_read)
2154                         continue;
2155                 /* fixup the bio for reuse, but preserve errno */
2156                 status = b->bi_status;
2157                 bio_reset(b, conf->mirrors[i].rdev->bdev, REQ_OP_READ);
2158                 b->bi_status = status;
2159                 b->bi_iter.bi_sector = r1_bio->sector +
2160                         conf->mirrors[i].rdev->data_offset;
2161                 b->bi_end_io = end_sync_read;
2162                 rp->raid_bio = r1_bio;
2163                 b->bi_private = rp;
2164
2165                 /* initialize bvec table again */
2166                 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2167         }
2168         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2169                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2170                     !r1_bio->bios[primary]->bi_status) {
2171                         r1_bio->bios[primary]->bi_end_io = NULL;
2172                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2173                         break;
2174                 }
2175         r1_bio->read_disk = primary;
2176         for (i = 0; i < conf->raid_disks * 2; i++) {
2177                 int j = 0;
2178                 struct bio *pbio = r1_bio->bios[primary];
2179                 struct bio *sbio = r1_bio->bios[i];
2180                 blk_status_t status = sbio->bi_status;
2181                 struct page **ppages = get_resync_pages(pbio)->pages;
2182                 struct page **spages = get_resync_pages(sbio)->pages;
2183                 struct bio_vec *bi;
2184                 int page_len[RESYNC_PAGES] = { 0 };
2185                 struct bvec_iter_all iter_all;
2186
2187                 if (sbio->bi_end_io != end_sync_read)
2188                         continue;
2189                 /* Now we can 'fixup' the error value */
2190                 sbio->bi_status = 0;
2191
2192                 bio_for_each_segment_all(bi, sbio, iter_all)
2193                         page_len[j++] = bi->bv_len;
2194
2195                 if (!status) {
2196                         for (j = vcnt; j-- ; ) {
2197                                 if (memcmp(page_address(ppages[j]),
2198                                            page_address(spages[j]),
2199                                            page_len[j]))
2200                                         break;
2201                         }
2202                 } else
2203                         j = 0;
2204                 if (j >= 0)
2205                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2206                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2207                               && !status)) {
2208                         /* No need to write to this device. */
2209                         sbio->bi_end_io = NULL;
2210                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2211                         continue;
2212                 }
2213
2214                 bio_copy_data(sbio, pbio);
2215         }
2216 }
2217
2218 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2219 {
2220         struct r1conf *conf = mddev->private;
2221         int i;
2222         int disks = conf->raid_disks * 2;
2223         struct bio *wbio;
2224
2225         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2226                 /* ouch - failed to read all of that. */
2227                 if (!fix_sync_read_error(r1_bio))
2228                         return;
2229
2230         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2231                 process_checks(r1_bio);
2232
2233         /*
2234          * schedule writes
2235          */
2236         atomic_set(&r1_bio->remaining, 1);
2237         for (i = 0; i < disks ; i++) {
2238                 wbio = r1_bio->bios[i];
2239                 if (wbio->bi_end_io == NULL ||
2240                     (wbio->bi_end_io == end_sync_read &&
2241                      (i == r1_bio->read_disk ||
2242                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2243                         continue;
2244                 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2245                         abort_sync_write(mddev, r1_bio);
2246                         continue;
2247                 }
2248
2249                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2250                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2251                         wbio->bi_opf |= MD_FAILFAST;
2252
2253                 wbio->bi_end_io = end_sync_write;
2254                 atomic_inc(&r1_bio->remaining);
2255                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2256
2257                 submit_bio_noacct(wbio);
2258         }
2259
2260         put_sync_write_buf(r1_bio, 1);
2261 }
2262
2263 /*
2264  * This is a kernel thread which:
2265  *
2266  *      1.      Retries failed read operations on working mirrors.
2267  *      2.      Updates the raid superblock when problems encounter.
2268  *      3.      Performs writes following reads for array synchronising.
2269  */
2270
2271 static void fix_read_error(struct r1conf *conf, int read_disk,
2272                            sector_t sect, int sectors)
2273 {
2274         struct mddev *mddev = conf->mddev;
2275         while(sectors) {
2276                 int s = sectors;
2277                 int d = read_disk;
2278                 int success = 0;
2279                 int start;
2280                 struct md_rdev *rdev;
2281
2282                 if (s > (PAGE_SIZE>>9))
2283                         s = PAGE_SIZE >> 9;
2284
2285                 do {
2286                         sector_t first_bad;
2287                         int bad_sectors;
2288
2289                         rcu_read_lock();
2290                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2291                         if (rdev &&
2292                             (test_bit(In_sync, &rdev->flags) ||
2293                              (!test_bit(Faulty, &rdev->flags) &&
2294                               rdev->recovery_offset >= sect + s)) &&
2295                             is_badblock(rdev, sect, s,
2296                                         &first_bad, &bad_sectors) == 0) {
2297                                 atomic_inc(&rdev->nr_pending);
2298                                 rcu_read_unlock();
2299                                 if (sync_page_io(rdev, sect, s<<9,
2300                                          conf->tmppage, REQ_OP_READ, 0, false))
2301                                         success = 1;
2302                                 rdev_dec_pending(rdev, mddev);
2303                                 if (success)
2304                                         break;
2305                         } else
2306                                 rcu_read_unlock();
2307                         d++;
2308                         if (d == conf->raid_disks * 2)
2309                                 d = 0;
2310                 } while (!success && d != read_disk);
2311
2312                 if (!success) {
2313                         /* Cannot read from anywhere - mark it bad */
2314                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2315                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2316                                 md_error(mddev, rdev);
2317                         break;
2318                 }
2319                 /* write it back and re-read */
2320                 start = d;
2321                 while (d != read_disk) {
2322                         if (d==0)
2323                                 d = conf->raid_disks * 2;
2324                         d--;
2325                         rcu_read_lock();
2326                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2327                         if (rdev &&
2328                             !test_bit(Faulty, &rdev->flags)) {
2329                                 atomic_inc(&rdev->nr_pending);
2330                                 rcu_read_unlock();
2331                                 r1_sync_page_io(rdev, sect, s,
2332                                                 conf->tmppage, WRITE);
2333                                 rdev_dec_pending(rdev, mddev);
2334                         } else
2335                                 rcu_read_unlock();
2336                 }
2337                 d = start;
2338                 while (d != read_disk) {
2339                         char b[BDEVNAME_SIZE];
2340                         if (d==0)
2341                                 d = conf->raid_disks * 2;
2342                         d--;
2343                         rcu_read_lock();
2344                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2345                         if (rdev &&
2346                             !test_bit(Faulty, &rdev->flags)) {
2347                                 atomic_inc(&rdev->nr_pending);
2348                                 rcu_read_unlock();
2349                                 if (r1_sync_page_io(rdev, sect, s,
2350                                                     conf->tmppage, READ)) {
2351                                         atomic_add(s, &rdev->corrected_errors);
2352                                         pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2353                                                 mdname(mddev), s,
2354                                                 (unsigned long long)(sect +
2355                                                                      rdev->data_offset),
2356                                                 bdevname(rdev->bdev, b));
2357                                 }
2358                                 rdev_dec_pending(rdev, mddev);
2359                         } else
2360                                 rcu_read_unlock();
2361                 }
2362                 sectors -= s;
2363                 sect += s;
2364         }
2365 }
2366
2367 static int narrow_write_error(struct r1bio *r1_bio, int i)
2368 {
2369         struct mddev *mddev = r1_bio->mddev;
2370         struct r1conf *conf = mddev->private;
2371         struct md_rdev *rdev = conf->mirrors[i].rdev;
2372
2373         /* bio has the data to be written to device 'i' where
2374          * we just recently had a write error.
2375          * We repeatedly clone the bio and trim down to one block,
2376          * then try the write.  Where the write fails we record
2377          * a bad block.
2378          * It is conceivable that the bio doesn't exactly align with
2379          * blocks.  We must handle this somehow.
2380          *
2381          * We currently own a reference on the rdev.
2382          */
2383
2384         int block_sectors;
2385         sector_t sector;
2386         int sectors;
2387         int sect_to_write = r1_bio->sectors;
2388         int ok = 1;
2389
2390         if (rdev->badblocks.shift < 0)
2391                 return 0;
2392
2393         block_sectors = roundup(1 << rdev->badblocks.shift,
2394                                 bdev_logical_block_size(rdev->bdev) >> 9);
2395         sector = r1_bio->sector;
2396         sectors = ((sector + block_sectors)
2397                    & ~(sector_t)(block_sectors - 1))
2398                 - sector;
2399
2400         while (sect_to_write) {
2401                 struct bio *wbio;
2402                 if (sectors > sect_to_write)
2403                         sectors = sect_to_write;
2404                 /* Write at 'sector' for 'sectors'*/
2405
2406                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2407                         wbio = bio_alloc_clone(rdev->bdev,
2408                                                r1_bio->behind_master_bio,
2409                                                GFP_NOIO, &mddev->bio_set);
2410                 } else {
2411                         wbio = bio_alloc_clone(rdev->bdev, r1_bio->master_bio,
2412                                                GFP_NOIO, &mddev->bio_set);
2413                 }
2414
2415                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2416                 wbio->bi_iter.bi_sector = r1_bio->sector;
2417                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2418
2419                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2420                 wbio->bi_iter.bi_sector += rdev->data_offset;
2421
2422                 if (submit_bio_wait(wbio) < 0)
2423                         /* failure! */
2424                         ok = rdev_set_badblocks(rdev, sector,
2425                                                 sectors, 0)
2426                                 && ok;
2427
2428                 bio_put(wbio);
2429                 sect_to_write -= sectors;
2430                 sector += sectors;
2431                 sectors = block_sectors;
2432         }
2433         return ok;
2434 }
2435
2436 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2437 {
2438         int m;
2439         int s = r1_bio->sectors;
2440         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2441                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2442                 struct bio *bio = r1_bio->bios[m];
2443                 if (bio->bi_end_io == NULL)
2444                         continue;
2445                 if (!bio->bi_status &&
2446                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2447                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2448                 }
2449                 if (bio->bi_status &&
2450                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2451                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2452                                 md_error(conf->mddev, rdev);
2453                 }
2454         }
2455         put_buf(r1_bio);
2456         md_done_sync(conf->mddev, s, 1);
2457 }
2458
2459 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2460 {
2461         int m, idx;
2462         bool fail = false;
2463
2464         for (m = 0; m < conf->raid_disks * 2 ; m++)
2465                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2466                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2467                         rdev_clear_badblocks(rdev,
2468                                              r1_bio->sector,
2469                                              r1_bio->sectors, 0);
2470                         rdev_dec_pending(rdev, conf->mddev);
2471                 } else if (r1_bio->bios[m] != NULL) {
2472                         /* This drive got a write error.  We need to
2473                          * narrow down and record precise write
2474                          * errors.
2475                          */
2476                         fail = true;
2477                         if (!narrow_write_error(r1_bio, m)) {
2478                                 md_error(conf->mddev,
2479                                          conf->mirrors[m].rdev);
2480                                 /* an I/O failed, we can't clear the bitmap */
2481                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2482                         }
2483                         rdev_dec_pending(conf->mirrors[m].rdev,
2484                                          conf->mddev);
2485                 }
2486         if (fail) {
2487                 spin_lock_irq(&conf->device_lock);
2488                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2489                 idx = sector_to_idx(r1_bio->sector);
2490                 atomic_inc(&conf->nr_queued[idx]);
2491                 spin_unlock_irq(&conf->device_lock);
2492                 /*
2493                  * In case freeze_array() is waiting for condition
2494                  * get_unqueued_pending() == extra to be true.
2495                  */
2496                 wake_up(&conf->wait_barrier);
2497                 md_wakeup_thread(conf->mddev->thread);
2498         } else {
2499                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2500                         close_write(r1_bio);
2501                 raid_end_bio_io(r1_bio);
2502         }
2503 }
2504
2505 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2506 {
2507         struct mddev *mddev = conf->mddev;
2508         struct bio *bio;
2509         struct md_rdev *rdev;
2510
2511         clear_bit(R1BIO_ReadError, &r1_bio->state);
2512         /* we got a read error. Maybe the drive is bad.  Maybe just
2513          * the block and we can fix it.
2514          * We freeze all other IO, and try reading the block from
2515          * other devices.  When we find one, we re-write
2516          * and check it that fixes the read error.
2517          * This is all done synchronously while the array is
2518          * frozen
2519          */
2520
2521         bio = r1_bio->bios[r1_bio->read_disk];
2522         bio_put(bio);
2523         r1_bio->bios[r1_bio->read_disk] = NULL;
2524
2525         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2526         if (mddev->ro == 0
2527             && !test_bit(FailFast, &rdev->flags)) {
2528                 freeze_array(conf, 1);
2529                 fix_read_error(conf, r1_bio->read_disk,
2530                                r1_bio->sector, r1_bio->sectors);
2531                 unfreeze_array(conf);
2532         } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2533                 md_error(mddev, rdev);
2534         } else {
2535                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2536         }
2537
2538         rdev_dec_pending(rdev, conf->mddev);
2539         allow_barrier(conf, r1_bio->sector);
2540         bio = r1_bio->master_bio;
2541
2542         /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2543         r1_bio->state = 0;
2544         raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2545 }
2546
2547 static void raid1d(struct md_thread *thread)
2548 {
2549         struct mddev *mddev = thread->mddev;
2550         struct r1bio *r1_bio;
2551         unsigned long flags;
2552         struct r1conf *conf = mddev->private;
2553         struct list_head *head = &conf->retry_list;
2554         struct blk_plug plug;
2555         int idx;
2556
2557         md_check_recovery(mddev);
2558
2559         if (!list_empty_careful(&conf->bio_end_io_list) &&
2560             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2561                 LIST_HEAD(tmp);
2562                 spin_lock_irqsave(&conf->device_lock, flags);
2563                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2564                         list_splice_init(&conf->bio_end_io_list, &tmp);
2565                 spin_unlock_irqrestore(&conf->device_lock, flags);
2566                 while (!list_empty(&tmp)) {
2567                         r1_bio = list_first_entry(&tmp, struct r1bio,
2568                                                   retry_list);
2569                         list_del(&r1_bio->retry_list);
2570                         idx = sector_to_idx(r1_bio->sector);
2571                         atomic_dec(&conf->nr_queued[idx]);
2572                         if (mddev->degraded)
2573                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2574                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2575                                 close_write(r1_bio);
2576                         raid_end_bio_io(r1_bio);
2577                 }
2578         }
2579
2580         blk_start_plug(&plug);
2581         for (;;) {
2582
2583                 flush_pending_writes(conf);
2584
2585                 spin_lock_irqsave(&conf->device_lock, flags);
2586                 if (list_empty(head)) {
2587                         spin_unlock_irqrestore(&conf->device_lock, flags);
2588                         break;
2589                 }
2590                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2591                 list_del(head->prev);
2592                 idx = sector_to_idx(r1_bio->sector);
2593                 atomic_dec(&conf->nr_queued[idx]);
2594                 spin_unlock_irqrestore(&conf->device_lock, flags);
2595
2596                 mddev = r1_bio->mddev;
2597                 conf = mddev->private;
2598                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2599                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2600                             test_bit(R1BIO_WriteError, &r1_bio->state))
2601                                 handle_sync_write_finished(conf, r1_bio);
2602                         else
2603                                 sync_request_write(mddev, r1_bio);
2604                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2605                            test_bit(R1BIO_WriteError, &r1_bio->state))
2606                         handle_write_finished(conf, r1_bio);
2607                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2608                         handle_read_error(conf, r1_bio);
2609                 else
2610                         WARN_ON_ONCE(1);
2611
2612                 cond_resched();
2613                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2614                         md_check_recovery(mddev);
2615         }
2616         blk_finish_plug(&plug);
2617 }
2618
2619 static int init_resync(struct r1conf *conf)
2620 {
2621         int buffs;
2622
2623         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2624         BUG_ON(mempool_initialized(&conf->r1buf_pool));
2625
2626         return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2627                             r1buf_pool_free, conf->poolinfo);
2628 }
2629
2630 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2631 {
2632         struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2633         struct resync_pages *rps;
2634         struct bio *bio;
2635         int i;
2636
2637         for (i = conf->poolinfo->raid_disks; i--; ) {
2638                 bio = r1bio->bios[i];
2639                 rps = bio->bi_private;
2640                 bio_reset(bio, NULL, 0);
2641                 bio->bi_private = rps;
2642         }
2643         r1bio->master_bio = NULL;
2644         return r1bio;
2645 }
2646
2647 /*
2648  * perform a "sync" on one "block"
2649  *
2650  * We need to make sure that no normal I/O request - particularly write
2651  * requests - conflict with active sync requests.
2652  *
2653  * This is achieved by tracking pending requests and a 'barrier' concept
2654  * that can be installed to exclude normal IO requests.
2655  */
2656
2657 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2658                                    int *skipped)
2659 {
2660         struct r1conf *conf = mddev->private;
2661         struct r1bio *r1_bio;
2662         struct bio *bio;
2663         sector_t max_sector, nr_sectors;
2664         int disk = -1;
2665         int i;
2666         int wonly = -1;
2667         int write_targets = 0, read_targets = 0;
2668         sector_t sync_blocks;
2669         int still_degraded = 0;
2670         int good_sectors = RESYNC_SECTORS;
2671         int min_bad = 0; /* number of sectors that are bad in all devices */
2672         int idx = sector_to_idx(sector_nr);
2673         int page_idx = 0;
2674
2675         if (!mempool_initialized(&conf->r1buf_pool))
2676                 if (init_resync(conf))
2677                         return 0;
2678
2679         max_sector = mddev->dev_sectors;
2680         if (sector_nr >= max_sector) {
2681                 /* If we aborted, we need to abort the
2682                  * sync on the 'current' bitmap chunk (there will
2683                  * only be one in raid1 resync.
2684                  * We can find the current addess in mddev->curr_resync
2685                  */
2686                 if (mddev->curr_resync < max_sector) /* aborted */
2687                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2688                                            &sync_blocks, 1);
2689                 else /* completed sync */
2690                         conf->fullsync = 0;
2691
2692                 md_bitmap_close_sync(mddev->bitmap);
2693                 close_sync(conf);
2694
2695                 if (mddev_is_clustered(mddev)) {
2696                         conf->cluster_sync_low = 0;
2697                         conf->cluster_sync_high = 0;
2698                 }
2699                 return 0;
2700         }
2701
2702         if (mddev->bitmap == NULL &&
2703             mddev->recovery_cp == MaxSector &&
2704             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2705             conf->fullsync == 0) {
2706                 *skipped = 1;
2707                 return max_sector - sector_nr;
2708         }
2709         /* before building a request, check if we can skip these blocks..
2710          * This call the bitmap_start_sync doesn't actually record anything
2711          */
2712         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2713             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2714                 /* We can skip this block, and probably several more */
2715                 *skipped = 1;
2716                 return sync_blocks;
2717         }
2718
2719         /*
2720          * If there is non-resync activity waiting for a turn, then let it
2721          * though before starting on this new sync request.
2722          */
2723         if (atomic_read(&conf->nr_waiting[idx]))
2724                 schedule_timeout_uninterruptible(1);
2725
2726         /* we are incrementing sector_nr below. To be safe, we check against
2727          * sector_nr + two times RESYNC_SECTORS
2728          */
2729
2730         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2731                 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2732
2733
2734         if (raise_barrier(conf, sector_nr))
2735                 return 0;
2736
2737         r1_bio = raid1_alloc_init_r1buf(conf);
2738
2739         rcu_read_lock();
2740         /*
2741          * If we get a correctably read error during resync or recovery,
2742          * we might want to read from a different device.  So we
2743          * flag all drives that could conceivably be read from for READ,
2744          * and any others (which will be non-In_sync devices) for WRITE.
2745          * If a read fails, we try reading from something else for which READ
2746          * is OK.
2747          */
2748
2749         r1_bio->mddev = mddev;
2750         r1_bio->sector = sector_nr;
2751         r1_bio->state = 0;
2752         set_bit(R1BIO_IsSync, &r1_bio->state);
2753         /* make sure good_sectors won't go across barrier unit boundary */
2754         good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2755
2756         for (i = 0; i < conf->raid_disks * 2; i++) {
2757                 struct md_rdev *rdev;
2758                 bio = r1_bio->bios[i];
2759
2760                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2761                 if (rdev == NULL ||
2762                     test_bit(Faulty, &rdev->flags)) {
2763                         if (i < conf->raid_disks)
2764                                 still_degraded = 1;
2765                 } else if (!test_bit(In_sync, &rdev->flags)) {
2766                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2767                         bio->bi_end_io = end_sync_write;
2768                         write_targets ++;
2769                 } else {
2770                         /* may need to read from here */
2771                         sector_t first_bad = MaxSector;
2772                         int bad_sectors;
2773
2774                         if (is_badblock(rdev, sector_nr, good_sectors,
2775                                         &first_bad, &bad_sectors)) {
2776                                 if (first_bad > sector_nr)
2777                                         good_sectors = first_bad - sector_nr;
2778                                 else {
2779                                         bad_sectors -= (sector_nr - first_bad);
2780                                         if (min_bad == 0 ||
2781                                             min_bad > bad_sectors)
2782                                                 min_bad = bad_sectors;
2783                                 }
2784                         }
2785                         if (sector_nr < first_bad) {
2786                                 if (test_bit(WriteMostly, &rdev->flags)) {
2787                                         if (wonly < 0)
2788                                                 wonly = i;
2789                                 } else {
2790                                         if (disk < 0)
2791                                                 disk = i;
2792                                 }
2793                                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2794                                 bio->bi_end_io = end_sync_read;
2795                                 read_targets++;
2796                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2797                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2798                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2799                                 /*
2800                                  * The device is suitable for reading (InSync),
2801                                  * but has bad block(s) here. Let's try to correct them,
2802                                  * if we are doing resync or repair. Otherwise, leave
2803                                  * this device alone for this sync request.
2804                                  */
2805                                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2806                                 bio->bi_end_io = end_sync_write;
2807                                 write_targets++;
2808                         }
2809                 }
2810                 if (rdev && bio->bi_end_io) {
2811                         atomic_inc(&rdev->nr_pending);
2812                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2813                         bio_set_dev(bio, rdev->bdev);
2814                         if (test_bit(FailFast, &rdev->flags))
2815                                 bio->bi_opf |= MD_FAILFAST;
2816                 }
2817         }
2818         rcu_read_unlock();
2819         if (disk < 0)
2820                 disk = wonly;
2821         r1_bio->read_disk = disk;
2822
2823         if (read_targets == 0 && min_bad > 0) {
2824                 /* These sectors are bad on all InSync devices, so we
2825                  * need to mark them bad on all write targets
2826                  */
2827                 int ok = 1;
2828                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2829                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2830                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2831                                 ok = rdev_set_badblocks(rdev, sector_nr,
2832                                                         min_bad, 0
2833                                         ) && ok;
2834                         }
2835                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2836                 *skipped = 1;
2837                 put_buf(r1_bio);
2838
2839                 if (!ok) {
2840                         /* Cannot record the badblocks, so need to
2841                          * abort the resync.
2842                          * If there are multiple read targets, could just
2843                          * fail the really bad ones ???
2844                          */
2845                         conf->recovery_disabled = mddev->recovery_disabled;
2846                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2847                         return 0;
2848                 } else
2849                         return min_bad;
2850
2851         }
2852         if (min_bad > 0 && min_bad < good_sectors) {
2853                 /* only resync enough to reach the next bad->good
2854                  * transition */
2855                 good_sectors = min_bad;
2856         }
2857
2858         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2859                 /* extra read targets are also write targets */
2860                 write_targets += read_targets-1;
2861
2862         if (write_targets == 0 || read_targets == 0) {
2863                 /* There is nowhere to write, so all non-sync
2864                  * drives must be failed - so we are finished
2865                  */
2866                 sector_t rv;
2867                 if (min_bad > 0)
2868                         max_sector = sector_nr + min_bad;
2869                 rv = max_sector - sector_nr;
2870                 *skipped = 1;
2871                 put_buf(r1_bio);
2872                 return rv;
2873         }
2874
2875         if (max_sector > mddev->resync_max)
2876                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2877         if (max_sector > sector_nr + good_sectors)
2878                 max_sector = sector_nr + good_sectors;
2879         nr_sectors = 0;
2880         sync_blocks = 0;
2881         do {
2882                 struct page *page;
2883                 int len = PAGE_SIZE;
2884                 if (sector_nr + (len>>9) > max_sector)
2885                         len = (max_sector - sector_nr) << 9;
2886                 if (len == 0)
2887                         break;
2888                 if (sync_blocks == 0) {
2889                         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2890                                                   &sync_blocks, still_degraded) &&
2891                             !conf->fullsync &&
2892                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2893                                 break;
2894                         if ((len >> 9) > sync_blocks)
2895                                 len = sync_blocks<<9;
2896                 }
2897
2898                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2899                         struct resync_pages *rp;
2900
2901                         bio = r1_bio->bios[i];
2902                         rp = get_resync_pages(bio);
2903                         if (bio->bi_end_io) {
2904                                 page = resync_fetch_page(rp, page_idx);
2905
2906                                 /*
2907                                  * won't fail because the vec table is big
2908                                  * enough to hold all these pages
2909                                  */
2910                                 bio_add_page(bio, page, len, 0);
2911                         }
2912                 }
2913                 nr_sectors += len>>9;
2914                 sector_nr += len>>9;
2915                 sync_blocks -= (len>>9);
2916         } while (++page_idx < RESYNC_PAGES);
2917
2918         r1_bio->sectors = nr_sectors;
2919
2920         if (mddev_is_clustered(mddev) &&
2921                         conf->cluster_sync_high < sector_nr + nr_sectors) {
2922                 conf->cluster_sync_low = mddev->curr_resync_completed;
2923                 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2924                 /* Send resync message */
2925                 md_cluster_ops->resync_info_update(mddev,
2926                                 conf->cluster_sync_low,
2927                                 conf->cluster_sync_high);
2928         }
2929
2930         /* For a user-requested sync, we read all readable devices and do a
2931          * compare
2932          */
2933         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2934                 atomic_set(&r1_bio->remaining, read_targets);
2935                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2936                         bio = r1_bio->bios[i];
2937                         if (bio->bi_end_io == end_sync_read) {
2938                                 read_targets--;
2939                                 md_sync_acct_bio(bio, nr_sectors);
2940                                 if (read_targets == 1)
2941                                         bio->bi_opf &= ~MD_FAILFAST;
2942                                 submit_bio_noacct(bio);
2943                         }
2944                 }
2945         } else {
2946                 atomic_set(&r1_bio->remaining, 1);
2947                 bio = r1_bio->bios[r1_bio->read_disk];
2948                 md_sync_acct_bio(bio, nr_sectors);
2949                 if (read_targets == 1)
2950                         bio->bi_opf &= ~MD_FAILFAST;
2951                 submit_bio_noacct(bio);
2952         }
2953         return nr_sectors;
2954 }
2955
2956 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2957 {
2958         if (sectors)
2959                 return sectors;
2960
2961         return mddev->dev_sectors;
2962 }
2963
2964 static struct r1conf *setup_conf(struct mddev *mddev)
2965 {
2966         struct r1conf *conf;
2967         int i;
2968         struct raid1_info *disk;
2969         struct md_rdev *rdev;
2970         int err = -ENOMEM;
2971
2972         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2973         if (!conf)
2974                 goto abort;
2975
2976         conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2977                                    sizeof(atomic_t), GFP_KERNEL);
2978         if (!conf->nr_pending)
2979                 goto abort;
2980
2981         conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2982                                    sizeof(atomic_t), GFP_KERNEL);
2983         if (!conf->nr_waiting)
2984                 goto abort;
2985
2986         conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2987                                   sizeof(atomic_t), GFP_KERNEL);
2988         if (!conf->nr_queued)
2989                 goto abort;
2990
2991         conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2992                                 sizeof(atomic_t), GFP_KERNEL);
2993         if (!conf->barrier)
2994                 goto abort;
2995
2996         conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2997                                             mddev->raid_disks, 2),
2998                                 GFP_KERNEL);
2999         if (!conf->mirrors)
3000                 goto abort;
3001
3002         conf->tmppage = alloc_page(GFP_KERNEL);
3003         if (!conf->tmppage)
3004                 goto abort;
3005
3006         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
3007         if (!conf->poolinfo)
3008                 goto abort;
3009         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
3010         err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
3011                            rbio_pool_free, conf->poolinfo);
3012         if (err)
3013                 goto abort;
3014
3015         err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3016         if (err)
3017                 goto abort;
3018
3019         conf->poolinfo->mddev = mddev;
3020
3021         err = -EINVAL;
3022         spin_lock_init(&conf->device_lock);
3023         rdev_for_each(rdev, mddev) {
3024                 int disk_idx = rdev->raid_disk;
3025                 if (disk_idx >= mddev->raid_disks
3026                     || disk_idx < 0)
3027                         continue;
3028                 if (test_bit(Replacement, &rdev->flags))
3029                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
3030                 else
3031                         disk = conf->mirrors + disk_idx;
3032
3033                 if (disk->rdev)
3034                         goto abort;
3035                 disk->rdev = rdev;
3036                 disk->head_position = 0;
3037                 disk->seq_start = MaxSector;
3038         }
3039         conf->raid_disks = mddev->raid_disks;
3040         conf->mddev = mddev;
3041         INIT_LIST_HEAD(&conf->retry_list);
3042         INIT_LIST_HEAD(&conf->bio_end_io_list);
3043
3044         spin_lock_init(&conf->resync_lock);
3045         init_waitqueue_head(&conf->wait_barrier);
3046
3047         bio_list_init(&conf->pending_bio_list);
3048         conf->recovery_disabled = mddev->recovery_disabled - 1;
3049
3050         err = -EIO;
3051         for (i = 0; i < conf->raid_disks * 2; i++) {
3052
3053                 disk = conf->mirrors + i;
3054
3055                 if (i < conf->raid_disks &&
3056                     disk[conf->raid_disks].rdev) {
3057                         /* This slot has a replacement. */
3058                         if (!disk->rdev) {
3059                                 /* No original, just make the replacement
3060                                  * a recovering spare
3061                                  */
3062                                 disk->rdev =
3063                                         disk[conf->raid_disks].rdev;
3064                                 disk[conf->raid_disks].rdev = NULL;
3065                         } else if (!test_bit(In_sync, &disk->rdev->flags))
3066                                 /* Original is not in_sync - bad */
3067                                 goto abort;
3068                 }
3069
3070                 if (!disk->rdev ||
3071                     !test_bit(In_sync, &disk->rdev->flags)) {
3072                         disk->head_position = 0;
3073                         if (disk->rdev &&
3074                             (disk->rdev->saved_raid_disk < 0))
3075                                 conf->fullsync = 1;
3076                 }
3077         }
3078
3079         err = -ENOMEM;
3080         conf->thread = md_register_thread(raid1d, mddev, "raid1");
3081         if (!conf->thread)
3082                 goto abort;
3083
3084         return conf;
3085
3086  abort:
3087         if (conf) {
3088                 mempool_exit(&conf->r1bio_pool);
3089                 kfree(conf->mirrors);
3090                 safe_put_page(conf->tmppage);
3091                 kfree(conf->poolinfo);
3092                 kfree(conf->nr_pending);
3093                 kfree(conf->nr_waiting);
3094                 kfree(conf->nr_queued);
3095                 kfree(conf->barrier);
3096                 bioset_exit(&conf->bio_split);
3097                 kfree(conf);
3098         }
3099         return ERR_PTR(err);
3100 }
3101
3102 static void raid1_free(struct mddev *mddev, void *priv);
3103 static int raid1_run(struct mddev *mddev)
3104 {
3105         struct r1conf *conf;
3106         int i;
3107         struct md_rdev *rdev;
3108         int ret;
3109         bool discard_supported = false;
3110
3111         if (mddev->level != 1) {
3112                 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3113                         mdname(mddev), mddev->level);
3114                 return -EIO;
3115         }
3116         if (mddev->reshape_position != MaxSector) {
3117                 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3118                         mdname(mddev));
3119                 return -EIO;
3120         }
3121         if (mddev_init_writes_pending(mddev) < 0)
3122                 return -ENOMEM;
3123         /*
3124          * copy the already verified devices into our private RAID1
3125          * bookkeeping area. [whatever we allocate in run(),
3126          * should be freed in raid1_free()]
3127          */
3128         if (mddev->private == NULL)
3129                 conf = setup_conf(mddev);
3130         else
3131                 conf = mddev->private;
3132
3133         if (IS_ERR(conf))
3134                 return PTR_ERR(conf);
3135
3136         if (mddev->queue)
3137                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3138
3139         rdev_for_each(rdev, mddev) {
3140                 if (!mddev->gendisk)
3141                         continue;
3142                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3143                                   rdev->data_offset << 9);
3144                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3145                         discard_supported = true;
3146         }
3147
3148         mddev->degraded = 0;
3149         for (i = 0; i < conf->raid_disks; i++)
3150                 if (conf->mirrors[i].rdev == NULL ||
3151                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3152                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3153                         mddev->degraded++;
3154         /*
3155          * RAID1 needs at least one disk in active
3156          */
3157         if (conf->raid_disks - mddev->degraded < 1) {
3158                 ret = -EINVAL;
3159                 goto abort;
3160         }
3161
3162         if (conf->raid_disks - mddev->degraded == 1)
3163                 mddev->recovery_cp = MaxSector;
3164
3165         if (mddev->recovery_cp != MaxSector)
3166                 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3167                         mdname(mddev));
3168         pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3169                 mdname(mddev), mddev->raid_disks - mddev->degraded,
3170                 mddev->raid_disks);
3171
3172         /*
3173          * Ok, everything is just fine now
3174          */
3175         mddev->thread = conf->thread;
3176         conf->thread = NULL;
3177         mddev->private = conf;
3178         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3179
3180         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3181
3182         if (mddev->queue) {
3183                 if (discard_supported)
3184                         blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3185                                                 mddev->queue);
3186                 else
3187                         blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3188                                                   mddev->queue);
3189         }
3190
3191         ret = md_integrity_register(mddev);
3192         if (ret) {
3193                 md_unregister_thread(&mddev->thread);
3194                 goto abort;
3195         }
3196         return 0;
3197
3198 abort:
3199         raid1_free(mddev, conf);
3200         return ret;
3201 }
3202
3203 static void raid1_free(struct mddev *mddev, void *priv)
3204 {
3205         struct r1conf *conf = priv;
3206
3207         mempool_exit(&conf->r1bio_pool);
3208         kfree(conf->mirrors);
3209         safe_put_page(conf->tmppage);
3210         kfree(conf->poolinfo);
3211         kfree(conf->nr_pending);
3212         kfree(conf->nr_waiting);
3213         kfree(conf->nr_queued);
3214         kfree(conf->barrier);
3215         bioset_exit(&conf->bio_split);
3216         kfree(conf);
3217 }
3218
3219 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3220 {
3221         /* no resync is happening, and there is enough space
3222          * on all devices, so we can resize.
3223          * We need to make sure resync covers any new space.
3224          * If the array is shrinking we should possibly wait until
3225          * any io in the removed space completes, but it hardly seems
3226          * worth it.
3227          */
3228         sector_t newsize = raid1_size(mddev, sectors, 0);
3229         if (mddev->external_size &&
3230             mddev->array_sectors > newsize)
3231                 return -EINVAL;
3232         if (mddev->bitmap) {
3233                 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3234                 if (ret)
3235                         return ret;
3236         }
3237         md_set_array_sectors(mddev, newsize);
3238         if (sectors > mddev->dev_sectors &&
3239             mddev->recovery_cp > mddev->dev_sectors) {
3240                 mddev->recovery_cp = mddev->dev_sectors;
3241                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3242         }
3243         mddev->dev_sectors = sectors;
3244         mddev->resync_max_sectors = sectors;
3245         return 0;
3246 }
3247
3248 static int raid1_reshape(struct mddev *mddev)
3249 {
3250         /* We need to:
3251          * 1/ resize the r1bio_pool
3252          * 2/ resize conf->mirrors
3253          *
3254          * We allocate a new r1bio_pool if we can.
3255          * Then raise a device barrier and wait until all IO stops.
3256          * Then resize conf->mirrors and swap in the new r1bio pool.
3257          *
3258          * At the same time, we "pack" the devices so that all the missing
3259          * devices have the higher raid_disk numbers.
3260          */
3261         mempool_t newpool, oldpool;
3262         struct pool_info *newpoolinfo;
3263         struct raid1_info *newmirrors;
3264         struct r1conf *conf = mddev->private;
3265         int cnt, raid_disks;
3266         unsigned long flags;
3267         int d, d2;
3268         int ret;
3269
3270         memset(&newpool, 0, sizeof(newpool));
3271         memset(&oldpool, 0, sizeof(oldpool));
3272
3273         /* Cannot change chunk_size, layout, or level */
3274         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3275             mddev->layout != mddev->new_layout ||
3276             mddev->level != mddev->new_level) {
3277                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3278                 mddev->new_layout = mddev->layout;
3279                 mddev->new_level = mddev->level;
3280                 return -EINVAL;
3281         }
3282
3283         if (!mddev_is_clustered(mddev))
3284                 md_allow_write(mddev);
3285
3286         raid_disks = mddev->raid_disks + mddev->delta_disks;
3287
3288         if (raid_disks < conf->raid_disks) {
3289                 cnt=0;
3290                 for (d= 0; d < conf->raid_disks; d++)
3291                         if (conf->mirrors[d].rdev)
3292                                 cnt++;
3293                 if (cnt > raid_disks)
3294                         return -EBUSY;
3295         }
3296
3297         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3298         if (!newpoolinfo)
3299                 return -ENOMEM;
3300         newpoolinfo->mddev = mddev;
3301         newpoolinfo->raid_disks = raid_disks * 2;
3302
3303         ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3304                            rbio_pool_free, newpoolinfo);
3305         if (ret) {
3306                 kfree(newpoolinfo);
3307                 return ret;
3308         }
3309         newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3310                                          raid_disks, 2),
3311                              GFP_KERNEL);
3312         if (!newmirrors) {
3313                 kfree(newpoolinfo);
3314                 mempool_exit(&newpool);
3315                 return -ENOMEM;
3316         }
3317
3318         freeze_array(conf, 0);
3319
3320         /* ok, everything is stopped */
3321         oldpool = conf->r1bio_pool;
3322         conf->r1bio_pool = newpool;
3323
3324         for (d = d2 = 0; d < conf->raid_disks; d++) {
3325                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3326                 if (rdev && rdev->raid_disk != d2) {
3327                         sysfs_unlink_rdev(mddev, rdev);
3328                         rdev->raid_disk = d2;
3329                         sysfs_unlink_rdev(mddev, rdev);
3330                         if (sysfs_link_rdev(mddev, rdev))
3331                                 pr_warn("md/raid1:%s: cannot register rd%d\n",
3332                                         mdname(mddev), rdev->raid_disk);
3333                 }
3334                 if (rdev)
3335                         newmirrors[d2++].rdev = rdev;
3336         }
3337         kfree(conf->mirrors);
3338         conf->mirrors = newmirrors;
3339         kfree(conf->poolinfo);
3340         conf->poolinfo = newpoolinfo;
3341
3342         spin_lock_irqsave(&conf->device_lock, flags);
3343         mddev->degraded += (raid_disks - conf->raid_disks);
3344         spin_unlock_irqrestore(&conf->device_lock, flags);
3345         conf->raid_disks = mddev->raid_disks = raid_disks;
3346         mddev->delta_disks = 0;
3347
3348         unfreeze_array(conf);
3349
3350         set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3351         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3352         md_wakeup_thread(mddev->thread);
3353
3354         mempool_exit(&oldpool);
3355         return 0;
3356 }
3357
3358 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3359 {
3360         struct r1conf *conf = mddev->private;
3361
3362         if (quiesce)
3363                 freeze_array(conf, 0);
3364         else
3365                 unfreeze_array(conf);
3366 }
3367
3368 static void *raid1_takeover(struct mddev *mddev)
3369 {
3370         /* raid1 can take over:
3371          *  raid5 with 2 devices, any layout or chunk size
3372          */
3373         if (mddev->level == 5 && mddev->raid_disks == 2) {
3374                 struct r1conf *conf;
3375                 mddev->new_level = 1;
3376                 mddev->new_layout = 0;
3377                 mddev->new_chunk_sectors = 0;
3378                 conf = setup_conf(mddev);
3379                 if (!IS_ERR(conf)) {
3380                         /* Array must appear to be quiesced */
3381                         conf->array_frozen = 1;
3382                         mddev_clear_unsupported_flags(mddev,
3383                                 UNSUPPORTED_MDDEV_FLAGS);
3384                 }
3385                 return conf;
3386         }
3387         return ERR_PTR(-EINVAL);
3388 }
3389
3390 static struct md_personality raid1_personality =
3391 {
3392         .name           = "raid1",
3393         .level          = 1,
3394         .owner          = THIS_MODULE,
3395         .make_request   = raid1_make_request,
3396         .run            = raid1_run,
3397         .free           = raid1_free,
3398         .status         = raid1_status,
3399         .error_handler  = raid1_error,
3400         .hot_add_disk   = raid1_add_disk,
3401         .hot_remove_disk= raid1_remove_disk,
3402         .spare_active   = raid1_spare_active,
3403         .sync_request   = raid1_sync_request,
3404         .resize         = raid1_resize,
3405         .size           = raid1_size,
3406         .check_reshape  = raid1_reshape,
3407         .quiesce        = raid1_quiesce,
3408         .takeover       = raid1_takeover,
3409 };
3410
3411 static int __init raid_init(void)
3412 {
3413         return register_md_personality(&raid1_personality);
3414 }
3415
3416 static void raid_exit(void)
3417 {
3418         unregister_md_personality(&raid1_personality);
3419 }
3420
3421 module_init(raid_init);
3422 module_exit(raid_exit);
3423 MODULE_LICENSE("GPL");
3424 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3425 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3426 MODULE_ALIAS("md-raid1");
3427 MODULE_ALIAS("md-level-1");