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
  33 #include <trace/events/block.h>
  34
  35 #include "md.h"
  36 #include "raid1.h"
  37 #include "md-bitmap.h"
  38
  39 #define UNSUPPORTED_MDDEV_FLAGS         \
  40         ((1L << MD_HAS_JOURNAL) |       \
  41          (1L << MD_JOURNAL_CLEAN) |     \
  42          (1L << MD_HAS_PPL) |           \
  43          (1L << MD_HAS_MULTIPLE_PPLS))
  44
  45 static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
  46 static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
  47
  48 #define raid1_log(md, fmt, args...)                             \
  49         do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
  50
  51 #include "raid1-10.c"
  52
  53 static int check_and_add_wb(struct md_rdev *rdev, sector_t lo, sector_t hi)
  54 {
  55         struct wb_info *wi, *temp_wi;
  56         unsigned long flags;
  57         int ret = 0;
  58         struct mddev *mddev = rdev->mddev;
  59
  60         wi = mempool_alloc(mddev->wb_info_pool, GFP_NOIO);
  61
  62         spin_lock_irqsave(&rdev->wb_list_lock, flags);
  63         list_for_each_entry(temp_wi, &rdev->wb_list, list) {
  64                 /* collision happened */
  65                 if (hi > temp_wi->lo && lo < temp_wi->hi) {
  66                         ret = -EBUSY;
  67                         break;
  68                 }
  69         }
  70
  71         if (!ret) {
  72                 wi->lo = lo;
  73                 wi->hi = hi;
  74                 list_add(&wi->list, &rdev->wb_list);
  75         } else
  76                 mempool_free(wi, mddev->wb_info_pool);
  77         spin_unlock_irqrestore(&rdev->wb_list_lock, flags);
  78
  79         return ret;
  80 }
  81
  82 static void remove_wb(struct md_rdev *rdev, sector_t lo, sector_t hi)
  83 {
  84         struct wb_info *wi;
  85         unsigned long flags;
  86         int found = 0;
  87         struct mddev *mddev = rdev->mddev;
  88
  89         spin_lock_irqsave(&rdev->wb_list_lock, flags);
  90         list_for_each_entry(wi, &rdev->wb_list, list)
  91                 if (hi == wi->hi && lo == wi->lo) {
  92                         list_del(&wi->list);
  93                         mempool_free(wi, mddev->wb_info_pool);
  94                         found = 1;
  95                         break;
  96                 }
  97
  98         if (!found)
  99                 WARN(1, "The write behind IO is not recorded\n");
 100         spin_unlock_irqrestore(&rdev->wb_list_lock, flags);
 101         wake_up(&rdev->wb_io_wait);
 102 }
 103
 104 /*
 105  * for resync bio, r1bio pointer can be retrieved from the per-bio
 106  * 'struct resync_pages'.
 107  */
 108 static inline struct r1bio *get_resync_r1bio(struct bio *bio)
 109 {
 110         return get_resync_pages(bio)->raid_bio;
 111 }
 112
 113 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
 114 {
 115         struct pool_info *pi = data;
 116         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
 117
 118         /* allocate a r1bio with room for raid_disks entries in the bios array */
 119         return kzalloc(size, gfp_flags);
 120 }
 121
 122 #define RESYNC_DEPTH 32
 123 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
 124 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
 125 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
 126 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
 127 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
 128
 129 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
 130 {
 131         struct pool_info *pi = data;
 132         struct r1bio *r1_bio;
 133         struct bio *bio;
 134         int need_pages;
 135         int j;
 136         struct resync_pages *rps;
 137
 138         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 139         if (!r1_bio)
 140                 return NULL;
 141
 142         rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
 143                             gfp_flags);
 144         if (!rps)
 145                 goto out_free_r1bio;
 146
 147         /*
 148          * Allocate bios : 1 for reading, n-1 for writing
 149          */
 150         for (j = pi->raid_disks ; j-- ; ) {
 151                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 152                 if (!bio)
 153                         goto out_free_bio;
 154                 r1_bio->bios[j] = bio;
 155         }
 156         /*
 157          * Allocate RESYNC_PAGES data pages and attach them to
 158          * the first bio.
 159          * If this is a user-requested check/repair, allocate
 160          * RESYNC_PAGES for each bio.
 161          */
 162         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 163                 need_pages = pi->raid_disks;
 164         else
 165                 need_pages = 1;
 166         for (j = 0; j < pi->raid_disks; j++) {
 167                 struct resync_pages *rp = &rps[j];
 168
 169                 bio = r1_bio->bios[j];
 170
 171                 if (j < need_pages) {
 172                         if (resync_alloc_pages(rp, gfp_flags))
 173                                 goto out_free_pages;
 174                 } else {
 175                         memcpy(rp, &rps[0], sizeof(*rp));
 176                         resync_get_all_pages(rp);
 177                 }
 178
 179                 rp->raid_bio = r1_bio;
 180                 bio->bi_private = rp;
 181         }
 182
 183         r1_bio->master_bio = NULL;
 184
 185         return r1_bio;
 186
 187 out_free_pages:
 188         while (--j >= 0)
 189                 resync_free_pages(&rps[j]);
 190
 191 out_free_bio:
 192         while (++j < pi->raid_disks)
 193                 bio_put(r1_bio->bios[j]);
 194         kfree(rps);
 195
 196 out_free_r1bio:
 197         rbio_pool_free(r1_bio, data);
 198         return NULL;
 199 }
 200
 201 static void r1buf_pool_free(void *__r1_bio, void *data)
 202 {
 203         struct pool_info *pi = data;
 204         int i;
 205         struct r1bio *r1bio = __r1_bio;
 206         struct resync_pages *rp = NULL;
 207
 208         for (i = pi->raid_disks; i--; ) {
 209                 rp = get_resync_pages(r1bio->bios[i]);
 210                 resync_free_pages(rp);
 211                 bio_put(r1bio->bios[i]);
 212         }
 213
 214         /* resync pages array stored in the 1st bio's .bi_private */
 215         kfree(rp);
 216
 217         rbio_pool_free(r1bio, data);
 218 }
 219
 220 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 221 {
 222         int i;
 223
 224         for (i = 0; i < conf->raid_disks * 2; i++) {
 225                 struct bio **bio = r1_bio->bios + i;
 226                 if (!BIO_SPECIAL(*bio))
 227                         bio_put(*bio);
 228                 *bio = NULL;
 229         }
 230 }
 231
 232 static void free_r1bio(struct r1bio *r1_bio)
 233 {
 234         struct r1conf *conf = r1_bio->mddev->private;
 235
 236         put_all_bios(conf, r1_bio);
 237         mempool_free(r1_bio, &conf->r1bio_pool);
 238 }
 239
 240 static void put_buf(struct r1bio *r1_bio)
 241 {
 242         struct r1conf *conf = r1_bio->mddev->private;
 243         sector_t sect = r1_bio->sector;
 244         int i;
 245
 246         for (i = 0; i < conf->raid_disks * 2; i++) {
 247                 struct bio *bio = r1_bio->bios[i];
 248                 if (bio->bi_end_io)
 249                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 250         }
 251
 252         mempool_free(r1_bio, &conf->r1buf_pool);
 253
 254         lower_barrier(conf, sect);
 255 }
 256
 257 static void reschedule_retry(struct r1bio *r1_bio)
 258 {
 259         unsigned long flags;
 260         struct mddev *mddev = r1_bio->mddev;
 261         struct r1conf *conf = mddev->private;
 262         int idx;
 263
 264         idx = sector_to_idx(r1_bio->sector);
 265         spin_lock_irqsave(&conf->device_lock, flags);
 266         list_add(&r1_bio->retry_list, &conf->retry_list);
 267         atomic_inc(&conf->nr_queued[idx]);
 268         spin_unlock_irqrestore(&conf->device_lock, flags);
 269
 270         wake_up(&conf->wait_barrier);
 271         md_wakeup_thread(mddev->thread);
 272 }
 273
 274 /*
 275  * raid_end_bio_io() is called when we have finished servicing a mirrored
 276  * operation and are ready to return a success/failure code to the buffer
 277  * cache layer.
 278  */
 279 static void call_bio_endio(struct r1bio *r1_bio)
 280 {
 281         struct bio *bio = r1_bio->master_bio;
 282         struct r1conf *conf = r1_bio->mddev->private;
 283
 284         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 285                 bio->bi_status = BLK_STS_IOERR;
 286
 287         bio_endio(bio);
 288         /*
 289          * Wake up any possible resync thread that waits for the device
 290          * to go idle.
 291          */
 292         allow_barrier(conf, r1_bio->sector);
 293 }
 294
 295 static void raid_end_bio_io(struct r1bio *r1_bio)
 296 {
 297         struct bio *bio = r1_bio->master_bio;
 298
 299         /* if nobody has done the final endio yet, do it now */
 300         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 301                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 302                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 303                          (unsigned long long) bio->bi_iter.bi_sector,
 304                          (unsigned long long) bio_end_sector(bio) - 1);
 305
 306                 call_bio_endio(r1_bio);
 307         }
 308         free_r1bio(r1_bio);
 309 }
 310
 311 /*
 312  * Update disk head position estimator based on IRQ completion info.
 313  */
 314 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 315 {
 316         struct r1conf *conf = r1_bio->mddev->private;
 317
 318         conf->mirrors[disk].head_position =
 319                 r1_bio->sector + (r1_bio->sectors);
 320 }
 321
 322 /*
 323  * Find the disk number which triggered given bio
 324  */
 325 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 326 {
 327         int mirror;
 328         struct r1conf *conf = r1_bio->mddev->private;
 329         int raid_disks = conf->raid_disks;
 330
 331         for (mirror = 0; mirror < raid_disks * 2; mirror++)
 332                 if (r1_bio->bios[mirror] == bio)
 333                         break;
 334
 335         BUG_ON(mirror == raid_disks * 2);
 336         update_head_pos(mirror, r1_bio);
 337
 338         return mirror;
 339 }
 340
 341 static void raid1_end_read_request(struct bio *bio)
 342 {
 343         int uptodate = !bio->bi_status;
 344         struct r1bio *r1_bio = bio->bi_private;
 345         struct r1conf *conf = r1_bio->mddev->private;
 346         struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 347
 348         /*
 349          * this branch is our 'one mirror IO has finished' event handler:
 350          */
 351         update_head_pos(r1_bio->read_disk, r1_bio);
 352
 353         if (uptodate)
 354                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 355         else if (test_bit(FailFast, &rdev->flags) &&
 356                  test_bit(R1BIO_FailFast, &r1_bio->state))
 357                 /* This was a fail-fast read so we definitely
 358                  * want to retry */
 359                 ;
 360         else {
 361                 /* If all other devices have failed, we want to return
 362                  * the error upwards rather than fail the last device.
 363                  * Here we redefine "uptodate" to mean "Don't want to retry"
 364                  */
 365                 unsigned long flags;
 366                 spin_lock_irqsave(&conf->device_lock, flags);
 367                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 368                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 369                      test_bit(In_sync, &rdev->flags)))
 370                         uptodate = 1;
 371                 spin_unlock_irqrestore(&conf->device_lock, flags);
 372         }
 373
 374         if (uptodate) {
 375                 raid_end_bio_io(r1_bio);
 376                 rdev_dec_pending(rdev, conf->mddev);
 377         } else {
 378                 /*
 379                  * oops, read error:
 380                  */
 381                 char b[BDEVNAME_SIZE];
 382                 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
 383                                    mdname(conf->mddev),
 384                                    bdevname(rdev->bdev, b),
 385                                    (unsigned long long)r1_bio->sector);
 386                 set_bit(R1BIO_ReadError, &r1_bio->state);
 387                 reschedule_retry(r1_bio);
 388                 /* don't drop the reference on read_disk yet */
 389         }
 390 }
 391
 392 static void close_write(struct r1bio *r1_bio)
 393 {
 394         /* it really is the end of this request */
 395         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 396                 bio_free_pages(r1_bio->behind_master_bio);
 397                 bio_put(r1_bio->behind_master_bio);
 398                 r1_bio->behind_master_bio = NULL;
 399         }
 400         /* clear the bitmap if all writes complete successfully */
 401         md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 402                            r1_bio->sectors,
 403                            !test_bit(R1BIO_Degraded, &r1_bio->state),
 404                            test_bit(R1BIO_BehindIO, &r1_bio->state));
 405         md_write_end(r1_bio->mddev);
 406 }
 407
 408 static void r1_bio_write_done(struct r1bio *r1_bio)
 409 {
 410         if (!atomic_dec_and_test(&r1_bio->remaining))
 411                 return;
 412
 413         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 414                 reschedule_retry(r1_bio);
 415         else {
 416                 close_write(r1_bio);
 417                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 418                         reschedule_retry(r1_bio);
 419                 else
 420                         raid_end_bio_io(r1_bio);
 421         }
 422 }
 423
 424 static void raid1_end_write_request(struct bio *bio)
 425 {
 426         struct r1bio *r1_bio = bio->bi_private;
 427         int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 428         struct r1conf *conf = r1_bio->mddev->private;
 429         struct bio *to_put = NULL;
 430         int mirror = find_bio_disk(r1_bio, bio);
 431         struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 432         bool discard_error;
 433
 434         discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 435
 436         /*
 437          * 'one mirror IO has finished' event handler:
 438          */
 439         if (bio->bi_status && !discard_error) {
 440                 set_bit(WriteErrorSeen, &rdev->flags);
 441                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
 442                         set_bit(MD_RECOVERY_NEEDED, &
 443                                 conf->mddev->recovery);
 444
 445                 if (test_bit(FailFast, &rdev->flags) &&
 446                     (bio->bi_opf & MD_FAILFAST) &&
 447                     /* We never try FailFast to WriteMostly devices */
 448                     !test_bit(WriteMostly, &rdev->flags)) {
 449                         md_error(r1_bio->mddev, rdev);
 450                         if (!test_bit(Faulty, &rdev->flags))
 451                                 /* This is the only remaining device,
 452                                  * We need to retry the write without
 453                                  * FailFast
 454                                  */
 455                                 set_bit(R1BIO_WriteError, &r1_bio->state);
 456                         else {
 457                                 /* Finished with this branch */
 458                                 r1_bio->bios[mirror] = NULL;
 459                                 to_put = bio;
 460                         }
 461                 } else
 462                         set_bit(R1BIO_WriteError, &r1_bio->state);
 463         } else {
 464                 /*
 465                  * Set R1BIO_Uptodate in our master bio, so that we
 466                  * will return a good error code for to the higher
 467                  * levels even if IO on some other mirrored buffer
 468                  * fails.
 469                  *
 470                  * The 'master' represents the composite IO operation
 471                  * to user-side. So if something waits for IO, then it
 472                  * will wait for the 'master' bio.
 473                  */
 474                 sector_t first_bad;
 475                 int bad_sectors;
 476
 477                 r1_bio->bios[mirror] = NULL;
 478                 to_put = bio;
 479                 /*
 480                  * Do not set R1BIO_Uptodate if the current device is
 481                  * rebuilding or Faulty. This is because we cannot use
 482                  * such device for properly reading the data back (we could
 483                  * potentially use it, if the current write would have felt
 484                  * before rdev->recovery_offset, but for simplicity we don't
 485                  * check this here.
 486                  */
 487                 if (test_bit(In_sync, &rdev->flags) &&
 488                     !test_bit(Faulty, &rdev->flags))
 489                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 490
 491                 /* Maybe we can clear some bad blocks. */
 492                 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 493                                 &first_bad, &bad_sectors) && !discard_error) {
 494                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 495                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 496                 }
 497         }
 498
 499         if (behind) {
 500                 if (test_bit(WBCollisionCheck, &rdev->flags)) {
 501                         sector_t lo = r1_bio->sector;
 502                         sector_t hi = r1_bio->sector + r1_bio->sectors;
 503
 504                         remove_wb(rdev, lo, hi);
 505                 }
 506                 if (test_bit(WriteMostly, &rdev->flags))
 507                         atomic_dec(&r1_bio->behind_remaining);
 508
 509                 /*
 510                  * In behind mode, we ACK the master bio once the I/O
 511                  * has safely reached all non-writemostly
 512                  * disks. Setting the Returned bit ensures that this
 513                  * gets done only once -- we don't ever want to return
 514                  * -EIO here, instead we'll wait
 515                  */
 516                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 517                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 518                         /* Maybe we can return now */
 519                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 520                                 struct bio *mbio = r1_bio->master_bio;
 521                                 pr_debug("raid1: behind end write sectors"
 522                                          " %llu-%llu\n",
 523                                          (unsigned long long) mbio->bi_iter.bi_sector,
 524                                          (unsigned long long) bio_end_sector(mbio) - 1);
 525                                 call_bio_endio(r1_bio);
 526                         }
 527                 }
 528         }
 529         if (r1_bio->bios[mirror] == NULL)
 530                 rdev_dec_pending(rdev, conf->mddev);
 531
 532         /*
 533          * Let's see if all mirrored write operations have finished
 534          * already.
 535          */
 536         r1_bio_write_done(r1_bio);
 537
 538         if (to_put)
 539                 bio_put(to_put);
 540 }
 541
 542 static sector_t align_to_barrier_unit_end(sector_t start_sector,
 543                                           sector_t sectors)
 544 {
 545         sector_t len;
 546
 547         WARN_ON(sectors == 0);
 548         /*
 549          * len is the number of sectors from start_sector to end of the
 550          * barrier unit which start_sector belongs to.
 551          */
 552         len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 553               start_sector;
 554
 555         if (len > sectors)
 556                 len = sectors;
 557
 558         return len;
 559 }
 560
 561 /*
 562  * This routine returns the disk from which the requested read should
 563  * be done. There is a per-array 'next expected sequential IO' sector
 564  * number - if this matches on the next IO then we use the last disk.
 565  * There is also a per-disk 'last know head position' sector that is
 566  * maintained from IRQ contexts, both the normal and the resync IO
 567  * completion handlers update this position correctly. If there is no
 568  * perfect sequential match then we pick the disk whose head is closest.
 569  *
 570  * If there are 2 mirrors in the same 2 devices, performance degrades
 571  * because position is mirror, not device based.
 572  *
 573  * The rdev for the device selected will have nr_pending incremented.
 574  */
 575 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 576 {
 577         const sector_t this_sector = r1_bio->sector;
 578         int sectors;
 579         int best_good_sectors;
 580         int best_disk, best_dist_disk, best_pending_disk;
 581         int has_nonrot_disk;
 582         int disk;
 583         sector_t best_dist;
 584         unsigned int min_pending;
 585         struct md_rdev *rdev;
 586         int choose_first;
 587         int choose_next_idle;
 588
 589         rcu_read_lock();
 590         /*
 591          * Check if we can balance. We can balance on the whole
 592          * device if no resync is going on, or below the resync window.
 593          * We take the first readable disk when above the resync window.
 594          */
 595  retry:
 596         sectors = r1_bio->sectors;
 597         best_disk = -1;
 598         best_dist_disk = -1;
 599         best_dist = MaxSector;
 600         best_pending_disk = -1;
 601         min_pending = UINT_MAX;
 602         best_good_sectors = 0;
 603         has_nonrot_disk = 0;
 604         choose_next_idle = 0;
 605         clear_bit(R1BIO_FailFast, &r1_bio->state);
 606
 607         if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 608             (mddev_is_clustered(conf->mddev) &&
 609             md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 610                     this_sector + sectors)))
 611                 choose_first = 1;
 612         else
 613                 choose_first = 0;
 614
 615         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 616                 sector_t dist;
 617                 sector_t first_bad;
 618                 int bad_sectors;
 619                 unsigned int pending;
 620                 bool nonrot;
 621
 622                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 623                 if (r1_bio->bios[disk] == IO_BLOCKED
 624                     || rdev == NULL
 625                     || test_bit(Faulty, &rdev->flags))
 626                         continue;
 627                 if (!test_bit(In_sync, &rdev->flags) &&
 628                     rdev->recovery_offset < this_sector + sectors)
 629                         continue;
 630                 if (test_bit(WriteMostly, &rdev->flags)) {
 631                         /* Don't balance among write-mostly, just
 632                          * use the first as a last resort */
 633                         if (best_dist_disk < 0) {
 634                                 if (is_badblock(rdev, this_sector, sectors,
 635                                                 &first_bad, &bad_sectors)) {
 636                                         if (first_bad <= this_sector)
 637                                                 /* Cannot use this */
 638                                                 continue;
 639                                         best_good_sectors = first_bad - this_sector;
 640                                 } else
 641                                         best_good_sectors = sectors;
 642                                 best_dist_disk = disk;
 643                                 best_pending_disk = disk;
 644                         }
 645                         continue;
 646                 }
 647                 /* This is a reasonable device to use.  It might
 648                  * even be best.
 649                  */
 650                 if (is_badblock(rdev, this_sector, sectors,
 651                                 &first_bad, &bad_sectors)) {
 652                         if (best_dist < MaxSector)
 653                                 /* already have a better device */
 654                                 continue;
 655                         if (first_bad <= this_sector) {
 656                                 /* cannot read here. If this is the 'primary'
 657                                  * device, then we must not read beyond
 658                                  * bad_sectors from another device..
 659                                  */
 660                                 bad_sectors -= (this_sector - first_bad);
 661                                 if (choose_first && sectors > bad_sectors)
 662                                         sectors = bad_sectors;
 663                                 if (best_good_sectors > sectors)
 664                                         best_good_sectors = sectors;
 665
 666                         } else {
 667                                 sector_t good_sectors = first_bad - this_sector;
 668                                 if (good_sectors > best_good_sectors) {
 669                                         best_good_sectors = good_sectors;
 670                                         best_disk = disk;
 671                                 }
 672                                 if (choose_first)
 673                                         break;
 674                         }
 675                         continue;
 676                 } else {
 677                         if ((sectors > best_good_sectors) && (best_disk >= 0))
 678                                 best_disk = -1;
 679                         best_good_sectors = sectors;
 680                 }
 681
 682                 if (best_disk >= 0)
 683                         /* At least two disks to choose from so failfast is OK */
 684                         set_bit(R1BIO_FailFast, &r1_bio->state);
 685
 686                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 687                 has_nonrot_disk |= nonrot;
 688                 pending = atomic_read(&rdev->nr_pending);
 689                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 690                 if (choose_first) {
 691                         best_disk = disk;
 692                         break;
 693                 }
 694                 /* Don't change to another disk for sequential reads */
 695                 if (conf->mirrors[disk].next_seq_sect == this_sector
 696                     || dist == 0) {
 697                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 698                         struct raid1_info *mirror = &conf->mirrors[disk];
 699
 700                         best_disk = disk;
 701                         /*
 702                          * If buffered sequential IO size exceeds optimal
 703                          * iosize, check if there is idle disk. If yes, choose
 704                          * the idle disk. read_balance could already choose an
 705                          * idle disk before noticing it's a sequential IO in
 706                          * this disk. This doesn't matter because this disk
 707                          * will idle, next time it will be utilized after the
 708                          * first disk has IO size exceeds optimal iosize. In
 709                          * this way, iosize of the first disk will be optimal
 710                          * iosize at least. iosize of the second disk might be
 711                          * small, but not a big deal since when the second disk
 712                          * starts IO, the first disk is likely still busy.
 713                          */
 714                         if (nonrot && opt_iosize > 0 &&
 715                             mirror->seq_start != MaxSector &&
 716                             mirror->next_seq_sect > opt_iosize &&
 717                             mirror->next_seq_sect - opt_iosize >=
 718                             mirror->seq_start) {
 719                                 choose_next_idle = 1;
 720                                 continue;
 721                         }
 722                         break;
 723                 }
 724
 725                 if (choose_next_idle)
 726                         continue;
 727
 728                 if (min_pending > pending) {
 729                         min_pending = pending;
 730                         best_pending_disk = disk;
 731                 }
 732
 733                 if (dist < best_dist) {
 734                         best_dist = dist;
 735                         best_dist_disk = disk;
 736                 }
 737         }
 738
 739         /*
 740          * If all disks are rotational, choose the closest disk. If any disk is
 741          * non-rotational, choose the disk with less pending request even the
 742          * disk is rotational, which might/might not be optimal for raids with
 743          * mixed ratation/non-rotational disks depending on workload.
 744          */
 745         if (best_disk == -1) {
 746                 if (has_nonrot_disk || min_pending == 0)
 747                         best_disk = best_pending_disk;
 748                 else
 749                         best_disk = best_dist_disk;
 750         }
 751
 752         if (best_disk >= 0) {
 753                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 754                 if (!rdev)
 755                         goto retry;
 756                 atomic_inc(&rdev->nr_pending);
 757                 sectors = best_good_sectors;
 758
 759                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 760                         conf->mirrors[best_disk].seq_start = this_sector;
 761
 762                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 763         }
 764         rcu_read_unlock();
 765         *max_sectors = sectors;
 766
 767         return best_disk;
 768 }
 769
 770 static int raid1_congested(struct mddev *mddev, int bits)
 771 {
 772         struct r1conf *conf = mddev->private;
 773         int i, ret = 0;
 774
 775         if ((bits & (1 << WB_async_congested)) &&
 776             conf->pending_count >= max_queued_requests)
 777                 return 1;
 778
 779         rcu_read_lock();
 780         for (i = 0; i < conf->raid_disks * 2; i++) {
 781                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 782                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 783                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 784
 785                         BUG_ON(!q);
 786
 787                         /* Note the '|| 1' - when read_balance prefers
 788                          * non-congested targets, it can be removed
 789                          */
 790                         if ((bits & (1 << WB_async_congested)) || 1)
 791                                 ret |= bdi_congested(q->backing_dev_info, bits);
 792                         else
 793                                 ret &= bdi_congested(q->backing_dev_info, bits);
 794                 }
 795         }
 796         rcu_read_unlock();
 797         return ret;
 798 }
 799
 800 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 801 {
 802         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
 803         md_bitmap_unplug(conf->mddev->bitmap);
 804         wake_up(&conf->wait_barrier);
 805
 806         while (bio) { /* submit pending writes */
 807                 struct bio *next = bio->bi_next;
 808                 struct md_rdev *rdev = (void *)bio->bi_disk;
 809                 bio->bi_next = NULL;
 810                 bio_set_dev(bio, rdev->bdev);
 811                 if (test_bit(Faulty, &rdev->flags)) {
 812                         bio_io_error(bio);
 813                 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
 814                                     !blk_queue_discard(bio->bi_disk->queue)))
 815                         /* Just ignore it */
 816                         bio_endio(bio);
 817                 else
 818                         generic_make_request(bio);
 819                 bio = next;
 820         }
 821 }
 822
 823 static void flush_pending_writes(struct r1conf *conf)
 824 {
 825         /* Any writes that have been queued but are awaiting
 826          * bitmap updates get flushed here.
 827          */
 828         spin_lock_irq(&conf->device_lock);
 829
 830         if (conf->pending_bio_list.head) {
 831                 struct blk_plug plug;
 832                 struct bio *bio;
 833
 834                 bio = bio_list_get(&conf->pending_bio_list);
 835                 conf->pending_count = 0;
 836                 spin_unlock_irq(&conf->device_lock);
 837
 838                 /*
 839                  * As this is called in a wait_event() loop (see freeze_array),
 840                  * current->state might be TASK_UNINTERRUPTIBLE which will
 841                  * cause a warning when we prepare to wait again.  As it is
 842                  * rare that this path is taken, it is perfectly safe to force
 843                  * us to go around the wait_event() loop again, so the warning
 844                  * is a false-positive.  Silence the warning by resetting
 845                  * thread state
 846                  */
 847                 __set_current_state(TASK_RUNNING);
 848                 blk_start_plug(&plug);
 849                 flush_bio_list(conf, bio);
 850                 blk_finish_plug(&plug);
 851         } else
 852                 spin_unlock_irq(&conf->device_lock);
 853 }
 854
 855 /* Barriers....
 856  * Sometimes we need to suspend IO while we do something else,
 857  * either some resync/recovery, or reconfigure the array.
 858  * To do this we raise a 'barrier'.
 859  * The 'barrier' is a counter that can be raised multiple times
 860  * to count how many activities are happening which preclude
 861  * normal IO.
 862  * We can only raise the barrier if there is no pending IO.
 863  * i.e. if nr_pending == 0.
 864  * We choose only to raise the barrier if no-one is waiting for the
 865  * barrier to go down.  This means that as soon as an IO request
 866  * is ready, no other operations which require a barrier will start
 867  * until the IO request has had a chance.
 868  *
 869  * So: regular IO calls 'wait_barrier'.  When that returns there
 870  *    is no backgroup IO happening,  It must arrange to call
 871  *    allow_barrier when it has finished its IO.
 872  * backgroup IO calls must call raise_barrier.  Once that returns
 873  *    there is no normal IO happeing.  It must arrange to call
 874  *    lower_barrier when the particular background IO completes.
 875  */
 876 static sector_t raise_barrier(struct r1conf *conf, sector_t sector_nr)
 877 {
 878         int idx = sector_to_idx(sector_nr);
 879
 880         spin_lock_irq(&conf->resync_lock);
 881
 882         /* Wait until no block IO is waiting */
 883         wait_event_lock_irq(conf->wait_barrier,
 884                             !atomic_read(&conf->nr_waiting[idx]),
 885                             conf->resync_lock);
 886
 887         /* block any new IO from starting */
 888         atomic_inc(&conf->barrier[idx]);
 889         /*
 890          * In raise_barrier() we firstly increase conf->barrier[idx] then
 891          * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 892          * increase conf->nr_pending[idx] then check conf->barrier[idx].
 893          * A memory barrier here to make sure conf->nr_pending[idx] won't
 894          * be fetched before conf->barrier[idx] is increased. Otherwise
 895          * there will be a race between raise_barrier() and _wait_barrier().
 896          */
 897         smp_mb__after_atomic();
 898
 899         /* For these conditions we must wait:
 900          * A: while the array is in frozen state
 901          * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 902          *    existing in corresponding I/O barrier bucket.
 903          * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 904          *    max resync count which allowed on current I/O barrier bucket.
 905          */
 906         wait_event_lock_irq(conf->wait_barrier,
 907                             (!conf->array_frozen &&
 908                              !atomic_read(&conf->nr_pending[idx]) &&
 909                              atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
 910                                 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
 911                             conf->resync_lock);
 912
 913         if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
 914                 atomic_dec(&conf->barrier[idx]);
 915                 spin_unlock_irq(&conf->resync_lock);
 916                 wake_up(&conf->wait_barrier);
 917                 return -EINTR;
 918         }
 919
 920         atomic_inc(&conf->nr_sync_pending);
 921         spin_unlock_irq(&conf->resync_lock);
 922
 923         return 0;
 924 }
 925
 926 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 927 {
 928         int idx = sector_to_idx(sector_nr);
 929
 930         BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 931
 932         atomic_dec(&conf->barrier[idx]);
 933         atomic_dec(&conf->nr_sync_pending);
 934         wake_up(&conf->wait_barrier);
 935 }
 936
 937 static void _wait_barrier(struct r1conf *conf, int idx)
 938 {
 939         /*
 940          * We need to increase conf->nr_pending[idx] very early here,
 941          * then raise_barrier() can be blocked when it waits for
 942          * conf->nr_pending[idx] to be 0. Then we can avoid holding
 943          * conf->resync_lock when there is no barrier raised in same
 944          * barrier unit bucket. Also if the array is frozen, I/O
 945          * should be blocked until array is unfrozen.
 946          */
 947         atomic_inc(&conf->nr_pending[idx]);
 948         /*
 949          * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
 950          * check conf->barrier[idx]. In raise_barrier() we firstly increase
 951          * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
 952          * barrier is necessary here to make sure conf->barrier[idx] won't be
 953          * fetched before conf->nr_pending[idx] is increased. Otherwise there
 954          * will be a race between _wait_barrier() and raise_barrier().
 955          */
 956         smp_mb__after_atomic();
 957
 958         /*
 959          * Don't worry about checking two atomic_t variables at same time
 960          * here. If during we check conf->barrier[idx], the array is
 961          * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
 962          * 0, it is safe to return and make the I/O continue. Because the
 963          * array is frozen, all I/O returned here will eventually complete
 964          * or be queued, no race will happen. See code comment in
 965          * frozen_array().
 966          */
 967         if (!READ_ONCE(conf->array_frozen) &&
 968             !atomic_read(&conf->barrier[idx]))
 969                 return;
 970
 971         /*
 972          * After holding conf->resync_lock, conf->nr_pending[idx]
 973          * should be decreased before waiting for barrier to drop.
 974          * Otherwise, we may encounter a race condition because
 975          * raise_barrer() might be waiting for conf->nr_pending[idx]
 976          * to be 0 at same time.
 977          */
 978         spin_lock_irq(&conf->resync_lock);
 979         atomic_inc(&conf->nr_waiting[idx]);
 980         atomic_dec(&conf->nr_pending[idx]);
 981         /*
 982          * In case freeze_array() is waiting for
 983          * get_unqueued_pending() == extra
 984          */
 985         wake_up(&conf->wait_barrier);
 986         /* Wait for the barrier in same barrier unit bucket to drop. */
 987         wait_event_lock_irq(conf->wait_barrier,
 988                             !conf->array_frozen &&
 989                              !atomic_read(&conf->barrier[idx]),
 990                             conf->resync_lock);
 991         atomic_inc(&conf->nr_pending[idx]);
 992         atomic_dec(&conf->nr_waiting[idx]);
 993         spin_unlock_irq(&conf->resync_lock);
 994 }
 995
 996 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
 997 {
 998         int idx = sector_to_idx(sector_nr);
 999
1000         /*
1001          * Very similar to _wait_barrier(). The difference is, for read
1002          * I/O we don't need wait for sync I/O, but if the whole array
1003          * is frozen, the read I/O still has to wait until the array is
1004          * unfrozen. Since there is no ordering requirement with
1005          * conf->barrier[idx] here, memory barrier is unnecessary as well.
1006          */
1007         atomic_inc(&conf->nr_pending[idx]);
1008
1009         if (!READ_ONCE(conf->array_frozen))
1010                 return;
1011
1012         spin_lock_irq(&conf->resync_lock);
1013         atomic_inc(&conf->nr_waiting[idx]);
1014         atomic_dec(&conf->nr_pending[idx]);
1015         /*
1016          * In case freeze_array() is waiting for
1017          * get_unqueued_pending() == extra
1018          */
1019         wake_up(&conf->wait_barrier);
1020         /* Wait for array to be unfrozen */
1021         wait_event_lock_irq(conf->wait_barrier,
1022                             !conf->array_frozen,
1023                             conf->resync_lock);
1024         atomic_inc(&conf->nr_pending[idx]);
1025         atomic_dec(&conf->nr_waiting[idx]);
1026         spin_unlock_irq(&conf->resync_lock);
1027 }
1028
1029 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1030 {
1031         int idx = sector_to_idx(sector_nr);
1032
1033         _wait_barrier(conf, idx);
1034 }
1035
1036 static void _allow_barrier(struct r1conf *conf, int idx)
1037 {
1038         atomic_dec(&conf->nr_pending[idx]);
1039         wake_up(&conf->wait_barrier);
1040 }
1041
1042 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1043 {
1044         int idx = sector_to_idx(sector_nr);
1045
1046         _allow_barrier(conf, idx);
1047 }
1048
1049 /* conf->resync_lock should be held */
1050 static int get_unqueued_pending(struct r1conf *conf)
1051 {
1052         int idx, ret;
1053
1054         ret = atomic_read(&conf->nr_sync_pending);
1055         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1056                 ret += atomic_read(&conf->nr_pending[idx]) -
1057                         atomic_read(&conf->nr_queued[idx]);
1058
1059         return ret;
1060 }
1061
1062 static void freeze_array(struct r1conf *conf, int extra)
1063 {
1064         /* Stop sync I/O and normal I/O and wait for everything to
1065          * go quiet.
1066          * This is called in two situations:
1067          * 1) management command handlers (reshape, remove disk, quiesce).
1068          * 2) one normal I/O request failed.
1069
1070          * After array_frozen is set to 1, new sync IO will be blocked at
1071          * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1072          * or wait_read_barrier(). The flying I/Os will either complete or be
1073          * queued. When everything goes quite, there are only queued I/Os left.
1074
1075          * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1076          * barrier bucket index which this I/O request hits. When all sync and
1077          * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1078          * of all conf->nr_queued[]. But normal I/O failure is an exception,
1079          * in handle_read_error(), we may call freeze_array() before trying to
1080          * fix the read error. In this case, the error read I/O is not queued,
1081          * so get_unqueued_pending() == 1.
1082          *
1083          * Therefore before this function returns, we need to wait until
1084          * get_unqueued_pendings(conf) gets equal to extra. For
1085          * normal I/O context, extra is 1, in rested situations extra is 0.
1086          */
1087         spin_lock_irq(&conf->resync_lock);
1088         conf->array_frozen = 1;
1089         raid1_log(conf->mddev, "wait freeze");
1090         wait_event_lock_irq_cmd(
1091                 conf->wait_barrier,
1092                 get_unqueued_pending(conf) == extra,
1093                 conf->resync_lock,
1094                 flush_pending_writes(conf));
1095         spin_unlock_irq(&conf->resync_lock);
1096 }
1097 static void unfreeze_array(struct r1conf *conf)
1098 {
1099         /* reverse the effect of the freeze */
1100         spin_lock_irq(&conf->resync_lock);
1101         conf->array_frozen = 0;
1102         spin_unlock_irq(&conf->resync_lock);
1103         wake_up(&conf->wait_barrier);
1104 }
1105
1106 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1107                                            struct bio *bio)
1108 {
1109         int size = bio->bi_iter.bi_size;
1110         unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1111         int i = 0;
1112         struct bio *behind_bio = NULL;
1113
1114         behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1115         if (!behind_bio)
1116                 return;
1117
1118         /* discard op, we don't support writezero/writesame yet */
1119         if (!bio_has_data(bio)) {
1120                 behind_bio->bi_iter.bi_size = size;
1121                 goto skip_copy;
1122         }
1123
1124         behind_bio->bi_write_hint = bio->bi_write_hint;
1125
1126         while (i < vcnt && size) {
1127                 struct page *page;
1128                 int len = min_t(int, PAGE_SIZE, size);
1129
1130                 page = alloc_page(GFP_NOIO);
1131                 if (unlikely(!page))
1132                         goto free_pages;
1133
1134                 bio_add_page(behind_bio, page, len, 0);
1135
1136                 size -= len;
1137                 i++;
1138         }
1139
1140         bio_copy_data(behind_bio, bio);
1141 skip_copy:
1142         r1_bio->behind_master_bio = behind_bio;
1143         set_bit(R1BIO_BehindIO, &r1_bio->state);
1144
1145         return;
1146
1147 free_pages:
1148         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1149                  bio->bi_iter.bi_size);
1150         bio_free_pages(behind_bio);
1151         bio_put(behind_bio);
1152 }
1153
1154 struct raid1_plug_cb {
1155         struct blk_plug_cb      cb;
1156         struct bio_list         pending;
1157         int                     pending_cnt;
1158 };
1159
1160 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1161 {
1162         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1163                                                   cb);
1164         struct mddev *mddev = plug->cb.data;
1165         struct r1conf *conf = mddev->private;
1166         struct bio *bio;
1167
1168         if (from_schedule || current->bio_list) {
1169                 spin_lock_irq(&conf->device_lock);
1170                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1171                 conf->pending_count += plug->pending_cnt;
1172                 spin_unlock_irq(&conf->device_lock);
1173                 wake_up(&conf->wait_barrier);
1174                 md_wakeup_thread(mddev->thread);
1175                 kfree(plug);
1176                 return;
1177         }
1178
1179         /* we aren't scheduling, so we can do the write-out directly. */
1180         bio = bio_list_get(&plug->pending);
1181         flush_bio_list(conf, bio);
1182         kfree(plug);
1183 }
1184
1185 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1186 {
1187         r1_bio->master_bio = bio;
1188         r1_bio->sectors = bio_sectors(bio);
1189         r1_bio->state = 0;
1190         r1_bio->mddev = mddev;
1191         r1_bio->sector = bio->bi_iter.bi_sector;
1192 }
1193
1194 static inline struct r1bio *
1195 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1196 {
1197         struct r1conf *conf = mddev->private;
1198         struct r1bio *r1_bio;
1199
1200         r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1201         /* Ensure no bio records IO_BLOCKED */
1202         memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1203         init_r1bio(r1_bio, mddev, bio);
1204         return r1_bio;
1205 }
1206
1207 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1208                                int max_read_sectors, struct r1bio *r1_bio)
1209 {
1210         struct r1conf *conf = mddev->private;
1211         struct raid1_info *mirror;
1212         struct bio *read_bio;
1213         struct bitmap *bitmap = mddev->bitmap;
1214         const int op = bio_op(bio);
1215         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1216         int max_sectors;
1217         int rdisk;
1218         bool print_msg = !!r1_bio;
1219         char b[BDEVNAME_SIZE];
1220
1221         /*
1222          * If r1_bio is set, we are blocking the raid1d thread
1223          * so there is a tiny risk of deadlock.  So ask for
1224          * emergency memory if needed.
1225          */
1226         gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1227
1228         if (print_msg) {
1229                 /* Need to get the block device name carefully */
1230                 struct md_rdev *rdev;
1231                 rcu_read_lock();
1232                 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1233                 if (rdev)
1234                         bdevname(rdev->bdev, b);
1235                 else
1236                         strcpy(b, "???");
1237                 rcu_read_unlock();
1238         }
1239
1240         /*
1241          * Still need barrier for READ in case that whole
1242          * array is frozen.
1243          */
1244         wait_read_barrier(conf, bio->bi_iter.bi_sector);
1245
1246         if (!r1_bio)
1247                 r1_bio = alloc_r1bio(mddev, bio);
1248         else
1249                 init_r1bio(r1_bio, mddev, bio);
1250         r1_bio->sectors = max_read_sectors;
1251
1252         /*
1253          * make_request() can abort the operation when read-ahead is being
1254          * used and no empty request is available.
1255          */
1256         rdisk = read_balance(conf, r1_bio, &max_sectors);
1257
1258         if (rdisk < 0) {
1259                 /* couldn't find anywhere to read from */
1260                 if (print_msg) {
1261                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1262                                             mdname(mddev),
1263                                             b,
1264                                             (unsigned long long)r1_bio->sector);
1265                 }
1266                 raid_end_bio_io(r1_bio);
1267                 return;
1268         }
1269         mirror = conf->mirrors + rdisk;
1270
1271         if (print_msg)
1272                 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1273                                     mdname(mddev),
1274                                     (unsigned long long)r1_bio->sector,
1275                                     bdevname(mirror->rdev->bdev, b));
1276
1277         if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1278             bitmap) {
1279                 /*
1280                  * Reading from a write-mostly device must take care not to
1281                  * over-take any writes that are 'behind'
1282                  */
1283                 raid1_log(mddev, "wait behind writes");
1284                 wait_event(bitmap->behind_wait,
1285                            atomic_read(&bitmap->behind_writes) == 0);
1286         }
1287
1288         if (max_sectors < bio_sectors(bio)) {
1289                 struct bio *split = bio_split(bio, max_sectors,
1290                                               gfp, &conf->bio_split);
1291                 bio_chain(split, bio);
1292                 generic_make_request(bio);
1293                 bio = split;
1294                 r1_bio->master_bio = bio;
1295                 r1_bio->sectors = max_sectors;
1296         }
1297
1298         r1_bio->read_disk = rdisk;
1299
1300         read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1301
1302         r1_bio->bios[rdisk] = read_bio;
1303
1304         read_bio->bi_iter.bi_sector = r1_bio->sector +
1305                 mirror->rdev->data_offset;
1306         bio_set_dev(read_bio, mirror->rdev->bdev);
1307         read_bio->bi_end_io = raid1_end_read_request;
1308         bio_set_op_attrs(read_bio, op, do_sync);
1309         if (test_bit(FailFast, &mirror->rdev->flags) &&
1310             test_bit(R1BIO_FailFast, &r1_bio->state))
1311                 read_bio->bi_opf |= MD_FAILFAST;
1312         read_bio->bi_private = r1_bio;
1313
1314         if (mddev->gendisk)
1315                 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1316                                 disk_devt(mddev->gendisk), r1_bio->sector);
1317
1318         generic_make_request(read_bio);
1319 }
1320
1321 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1322                                 int max_write_sectors)
1323 {
1324         struct r1conf *conf = mddev->private;
1325         struct r1bio *r1_bio;
1326         int i, disks;
1327         struct bitmap *bitmap = mddev->bitmap;
1328         unsigned long flags;
1329         struct md_rdev *blocked_rdev;
1330         struct blk_plug_cb *cb;
1331         struct raid1_plug_cb *plug = NULL;
1332         int first_clone;
1333         int max_sectors;
1334
1335         if (mddev_is_clustered(mddev) &&
1336              md_cluster_ops->area_resyncing(mddev, WRITE,
1337                      bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1338
1339                 DEFINE_WAIT(w);
1340                 for (;;) {
1341                         prepare_to_wait(&conf->wait_barrier,
1342                                         &w, TASK_IDLE);
1343                         if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1344                                                         bio->bi_iter.bi_sector,
1345                                                         bio_end_sector(bio)))
1346                                 break;
1347                         schedule();
1348                 }
1349                 finish_wait(&conf->wait_barrier, &w);
1350         }
1351
1352         /*
1353          * Register the new request and wait if the reconstruction
1354          * thread has put up a bar for new requests.
1355          * Continue immediately if no resync is active currently.
1356          */
1357         wait_barrier(conf, bio->bi_iter.bi_sector);
1358
1359         r1_bio = alloc_r1bio(mddev, bio);
1360         r1_bio->sectors = max_write_sectors;
1361
1362         if (conf->pending_count >= max_queued_requests) {
1363                 md_wakeup_thread(mddev->thread);
1364                 raid1_log(mddev, "wait queued");
1365                 wait_event(conf->wait_barrier,
1366                            conf->pending_count < max_queued_requests);
1367         }
1368         /* first select target devices under rcu_lock and
1369          * inc refcount on their rdev.  Record them by setting
1370          * bios[x] to bio
1371          * If there are known/acknowledged bad blocks on any device on
1372          * which we have seen a write error, we want to avoid writing those
1373          * blocks.
1374          * This potentially requires several writes to write around
1375          * the bad blocks.  Each set of writes gets it's own r1bio
1376          * with a set of bios attached.
1377          */
1378
1379         disks = conf->raid_disks * 2;
1380  retry_write:
1381         blocked_rdev = NULL;
1382         rcu_read_lock();
1383         max_sectors = r1_bio->sectors;
1384         for (i = 0;  i < disks; i++) {
1385                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1386                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1387                         atomic_inc(&rdev->nr_pending);
1388                         blocked_rdev = rdev;
1389                         break;
1390                 }
1391                 r1_bio->bios[i] = NULL;
1392                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1393                         if (i < conf->raid_disks)
1394                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1395                         continue;
1396                 }
1397
1398                 atomic_inc(&rdev->nr_pending);
1399                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1400                         sector_t first_bad;
1401                         int bad_sectors;
1402                         int is_bad;
1403
1404                         is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1405                                              &first_bad, &bad_sectors);
1406                         if (is_bad < 0) {
1407                                 /* mustn't write here until the bad block is
1408                                  * acknowledged*/
1409                                 set_bit(BlockedBadBlocks, &rdev->flags);
1410                                 blocked_rdev = rdev;
1411                                 break;
1412                         }
1413                         if (is_bad && first_bad <= r1_bio->sector) {
1414                                 /* Cannot write here at all */
1415                                 bad_sectors -= (r1_bio->sector - first_bad);
1416                                 if (bad_sectors < max_sectors)
1417                                         /* mustn't write more than bad_sectors
1418                                          * to other devices yet
1419                                          */
1420                                         max_sectors = bad_sectors;
1421                                 rdev_dec_pending(rdev, mddev);
1422                                 /* We don't set R1BIO_Degraded as that
1423                                  * only applies if the disk is
1424                                  * missing, so it might be re-added,
1425                                  * and we want to know to recover this
1426                                  * chunk.
1427                                  * In this case the device is here,
1428                                  * and the fact that this chunk is not
1429                                  * in-sync is recorded in the bad
1430                                  * block log
1431                                  */
1432                                 continue;
1433                         }
1434                         if (is_bad) {
1435                                 int good_sectors = first_bad - r1_bio->sector;
1436                                 if (good_sectors < max_sectors)
1437                                         max_sectors = good_sectors;
1438                         }
1439                 }
1440                 r1_bio->bios[i] = bio;
1441         }
1442         rcu_read_unlock();
1443
1444         if (unlikely(blocked_rdev)) {
1445                 /* Wait for this device to become unblocked */
1446                 int j;
1447
1448                 for (j = 0; j < i; j++)
1449                         if (r1_bio->bios[j])
1450                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1451                 r1_bio->state = 0;
1452                 allow_barrier(conf, bio->bi_iter.bi_sector);
1453                 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1454                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1455                 wait_barrier(conf, bio->bi_iter.bi_sector);
1456                 goto retry_write;
1457         }
1458
1459         if (max_sectors < bio_sectors(bio)) {
1460                 struct bio *split = bio_split(bio, max_sectors,
1461                                               GFP_NOIO, &conf->bio_split);
1462                 bio_chain(split, bio);
1463                 generic_make_request(bio);
1464                 bio = split;
1465                 r1_bio->master_bio = bio;
1466                 r1_bio->sectors = max_sectors;
1467         }
1468
1469         atomic_set(&r1_bio->remaining, 1);
1470         atomic_set(&r1_bio->behind_remaining, 0);
1471
1472         first_clone = 1;
1473
1474         for (i = 0; i < disks; i++) {
1475                 struct bio *mbio = NULL;
1476                 if (!r1_bio->bios[i])
1477                         continue;
1478
1479                 if (first_clone) {
1480                         /* do behind I/O ?
1481                          * Not if there are too many, or cannot
1482                          * allocate memory, or a reader on WriteMostly
1483                          * is waiting for behind writes to flush */
1484                         if (bitmap &&
1485                             (atomic_read(&bitmap->behind_writes)
1486                              < mddev->bitmap_info.max_write_behind) &&
1487                             !waitqueue_active(&bitmap->behind_wait)) {
1488                                 alloc_behind_master_bio(r1_bio, bio);
1489                         }
1490
1491                         md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1492                                              test_bit(R1BIO_BehindIO, &r1_bio->state));
1493                         first_clone = 0;
1494                 }
1495
1496                 if (r1_bio->behind_master_bio)
1497                         mbio = bio_clone_fast(r1_bio->behind_master_bio,
1498                                               GFP_NOIO, &mddev->bio_set);
1499                 else
1500                         mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1501
1502                 if (r1_bio->behind_master_bio) {
1503                         struct md_rdev *rdev = conf->mirrors[i].rdev;
1504
1505                         if (test_bit(WBCollisionCheck, &rdev->flags)) {
1506                                 sector_t lo = r1_bio->sector;
1507                                 sector_t hi = r1_bio->sector + r1_bio->sectors;
1508
1509                                 wait_event(rdev->wb_io_wait,
1510                                            check_and_add_wb(rdev, lo, hi) == 0);
1511                         }
1512                         if (test_bit(WriteMostly, &rdev->flags))
1513                                 atomic_inc(&r1_bio->behind_remaining);
1514                 }
1515
1516                 r1_bio->bios[i] = mbio;
1517
1518                 mbio->bi_iter.bi_sector = (r1_bio->sector +
1519                                    conf->mirrors[i].rdev->data_offset);
1520                 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1521                 mbio->bi_end_io = raid1_end_write_request;
1522                 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1523                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1524                     !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1525                     conf->raid_disks - mddev->degraded > 1)
1526                         mbio->bi_opf |= MD_FAILFAST;
1527                 mbio->bi_private = r1_bio;
1528
1529                 atomic_inc(&r1_bio->remaining);
1530
1531                 if (mddev->gendisk)
1532                         trace_block_bio_remap(mbio->bi_disk->queue,
1533                                               mbio, disk_devt(mddev->gendisk),
1534                                               r1_bio->sector);
1535                 /* flush_pending_writes() needs access to the rdev so...*/
1536                 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1537
1538                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1539                 if (cb)
1540                         plug = container_of(cb, struct raid1_plug_cb, cb);
1541                 else
1542                         plug = NULL;
1543                 if (plug) {
1544                         bio_list_add(&plug->pending, mbio);
1545                         plug->pending_cnt++;
1546                 } else {
1547                         spin_lock_irqsave(&conf->device_lock, flags);
1548                         bio_list_add(&conf->pending_bio_list, mbio);
1549                         conf->pending_count++;
1550                         spin_unlock_irqrestore(&conf->device_lock, flags);
1551                         md_wakeup_thread(mddev->thread);
1552                 }
1553         }
1554
1555         r1_bio_write_done(r1_bio);
1556
1557         /* In case raid1d snuck in to freeze_array */
1558         wake_up(&conf->wait_barrier);
1559 }
1560
1561 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1562 {
1563         sector_t sectors;
1564
1565         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1566                 md_flush_request(mddev, bio);
1567                 return true;
1568         }
1569
1570         /*
1571          * There is a limit to the maximum size, but
1572          * the read/write handler might find a lower limit
1573          * due to bad blocks.  To avoid multiple splits,
1574          * we pass the maximum number of sectors down
1575          * and let the lower level perform the split.
1576          */
1577         sectors = align_to_barrier_unit_end(
1578                 bio->bi_iter.bi_sector, bio_sectors(bio));
1579
1580         if (bio_data_dir(bio) == READ)
1581                 raid1_read_request(mddev, bio, sectors, NULL);
1582         else {
1583                 if (!md_write_start(mddev,bio))
1584                         return false;
1585                 raid1_write_request(mddev, bio, sectors);
1586         }
1587         return true;
1588 }
1589
1590 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1591 {
1592         struct r1conf *conf = mddev->private;
1593         int i;
1594
1595         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1596                    conf->raid_disks - mddev->degraded);
1597         rcu_read_lock();
1598         for (i = 0; i < conf->raid_disks; i++) {
1599                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1600                 seq_printf(seq, "%s",
1601                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1602         }
1603         rcu_read_unlock();
1604         seq_printf(seq, "]");
1605 }
1606
1607 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1608 {
1609         char b[BDEVNAME_SIZE];
1610         struct r1conf *conf = mddev->private;
1611         unsigned long flags;
1612
1613         /*
1614          * If it is not operational, then we have already marked it as dead
1615          * else if it is the last working disks, ignore the error, let the
1616          * next level up know.
1617          * else mark the drive as failed
1618          */
1619         spin_lock_irqsave(&conf->device_lock, flags);
1620         if (test_bit(In_sync, &rdev->flags)
1621             && (conf->raid_disks - mddev->degraded) == 1) {
1622                 /*
1623                  * Don't fail the drive, act as though we were just a
1624                  * normal single drive.
1625                  * However don't try a recovery from this drive as
1626                  * it is very likely to fail.
1627                  */
1628                 conf->recovery_disabled = mddev->recovery_disabled;
1629                 spin_unlock_irqrestore(&conf->device_lock, flags);
1630                 return;
1631         }
1632         set_bit(Blocked, &rdev->flags);
1633         if (test_and_clear_bit(In_sync, &rdev->flags))
1634                 mddev->degraded++;
1635         set_bit(Faulty, &rdev->flags);
1636         spin_unlock_irqrestore(&conf->device_lock, flags);
1637         /*
1638          * if recovery is running, make sure it aborts.
1639          */
1640         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1641         set_mask_bits(&mddev->sb_flags, 0,
1642                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1643         pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1644                 "md/raid1:%s: Operation continuing on %d devices.\n",
1645                 mdname(mddev), bdevname(rdev->bdev, b),
1646                 mdname(mddev), conf->raid_disks - mddev->degraded);
1647 }
1648
1649 static void print_conf(struct r1conf *conf)
1650 {
1651         int i;
1652
1653         pr_debug("RAID1 conf printout:\n");
1654         if (!conf) {
1655                 pr_debug("(!conf)\n");
1656                 return;
1657         }
1658         pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1659                  conf->raid_disks);
1660
1661         rcu_read_lock();
1662         for (i = 0; i < conf->raid_disks; i++) {
1663                 char b[BDEVNAME_SIZE];
1664                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1665                 if (rdev)
1666                         pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1667                                  i, !test_bit(In_sync, &rdev->flags),
1668                                  !test_bit(Faulty, &rdev->flags),
1669                                  bdevname(rdev->bdev,b));
1670         }
1671         rcu_read_unlock();
1672 }
1673
1674 static void close_sync(struct r1conf *conf)
1675 {
1676         int idx;
1677
1678         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1679                 _wait_barrier(conf, idx);
1680                 _allow_barrier(conf, idx);
1681         }
1682
1683         mempool_exit(&conf->r1buf_pool);
1684 }
1685
1686 static int raid1_spare_active(struct mddev *mddev)
1687 {
1688         int i;
1689         struct r1conf *conf = mddev->private;
1690         int count = 0;
1691         unsigned long flags;
1692
1693         /*
1694          * Find all failed disks within the RAID1 configuration
1695          * and mark them readable.
1696          * Called under mddev lock, so rcu protection not needed.
1697          * device_lock used to avoid races with raid1_end_read_request
1698          * which expects 'In_sync' flags and ->degraded to be consistent.
1699          */
1700         spin_lock_irqsave(&conf->device_lock, flags);
1701         for (i = 0; i < conf->raid_disks; i++) {
1702                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1703                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1704                 if (repl
1705                     && !test_bit(Candidate, &repl->flags)
1706                     && repl->recovery_offset == MaxSector
1707                     && !test_bit(Faulty, &repl->flags)
1708                     && !test_and_set_bit(In_sync, &repl->flags)) {
1709                         /* replacement has just become active */
1710                         if (!rdev ||
1711                             !test_and_clear_bit(In_sync, &rdev->flags))
1712                                 count++;
1713                         if (rdev) {
1714                                 /* Replaced device not technically
1715                                  * faulty, but we need to be sure
1716                                  * it gets removed and never re-added
1717                                  */
1718                                 set_bit(Faulty, &rdev->flags);
1719                                 sysfs_notify_dirent_safe(
1720                                         rdev->sysfs_state);
1721                         }
1722                 }
1723                 if (rdev
1724                     && rdev->recovery_offset == MaxSector
1725                     && !test_bit(Faulty, &rdev->flags)
1726                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1727                         count++;
1728                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1729                 }
1730         }
1731         mddev->degraded -= count;
1732         spin_unlock_irqrestore(&conf->device_lock, flags);
1733
1734         print_conf(conf);
1735         return count;
1736 }
1737
1738 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1739 {
1740         struct r1conf *conf = mddev->private;
1741         int err = -EEXIST;
1742         int mirror = 0;
1743         struct raid1_info *p;
1744         int first = 0;
1745         int last = conf->raid_disks - 1;
1746
1747         if (mddev->recovery_disabled == conf->recovery_disabled)
1748                 return -EBUSY;
1749
1750         if (md_integrity_add_rdev(rdev, mddev))
1751                 return -ENXIO;
1752
1753         if (rdev->raid_disk >= 0)
1754                 first = last = rdev->raid_disk;
1755
1756         /*
1757          * find the disk ... but prefer rdev->saved_raid_disk
1758          * if possible.
1759          */
1760         if (rdev->saved_raid_disk >= 0 &&
1761             rdev->saved_raid_disk >= first &&
1762             rdev->saved_raid_disk < conf->raid_disks &&
1763             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1764                 first = last = rdev->saved_raid_disk;
1765
1766         for (mirror = first; mirror <= last; mirror++) {
1767                 p = conf->mirrors + mirror;
1768                 if (!p->rdev) {
1769                         if (mddev->gendisk)
1770                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1771                                                   rdev->data_offset << 9);
1772
1773                         p->head_position = 0;
1774                         rdev->raid_disk = mirror;
1775                         err = 0;
1776                         /* As all devices are equivalent, we don't need a full recovery
1777                          * if this was recently any drive of the array
1778                          */
1779                         if (rdev->saved_raid_disk < 0)
1780                                 conf->fullsync = 1;
1781                         rcu_assign_pointer(p->rdev, rdev);
1782                         break;
1783                 }
1784                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1785                     p[conf->raid_disks].rdev == NULL) {
1786                         /* Add this device as a replacement */
1787                         clear_bit(In_sync, &rdev->flags);
1788                         set_bit(Replacement, &rdev->flags);
1789                         rdev->raid_disk = mirror;
1790                         err = 0;
1791                         conf->fullsync = 1;
1792                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1793                         break;
1794                 }
1795         }
1796         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1797                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1798         print_conf(conf);
1799         return err;
1800 }
1801
1802 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1803 {
1804         struct r1conf *conf = mddev->private;
1805         int err = 0;
1806         int number = rdev->raid_disk;
1807         struct raid1_info *p = conf->mirrors + number;
1808
1809         if (rdev != p->rdev)
1810                 p = conf->mirrors + conf->raid_disks + number;
1811
1812         print_conf(conf);
1813         if (rdev == p->rdev) {
1814                 if (test_bit(In_sync, &rdev->flags) ||
1815                     atomic_read(&rdev->nr_pending)) {
1816                         err = -EBUSY;
1817                         goto abort;
1818                 }
1819                 /* Only remove non-faulty devices if recovery
1820                  * is not possible.
1821                  */
1822                 if (!test_bit(Faulty, &rdev->flags) &&
1823                     mddev->recovery_disabled != conf->recovery_disabled &&
1824                     mddev->degraded < conf->raid_disks) {
1825                         err = -EBUSY;
1826                         goto abort;
1827                 }
1828                 p->rdev = NULL;
1829                 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1830                         synchronize_rcu();
1831                         if (atomic_read(&rdev->nr_pending)) {
1832                                 /* lost the race, try later */
1833                                 err = -EBUSY;
1834                                 p->rdev = rdev;
1835                                 goto abort;
1836                         }
1837                 }
1838                 if (conf->mirrors[conf->raid_disks + number].rdev) {
1839                         /* We just removed a device that is being replaced.
1840                          * Move down the replacement.  We drain all IO before
1841                          * doing this to avoid confusion.
1842                          */
1843                         struct md_rdev *repl =
1844                                 conf->mirrors[conf->raid_disks + number].rdev;
1845                         freeze_array(conf, 0);
1846                         if (atomic_read(&repl->nr_pending)) {
1847                                 /* It means that some queued IO of retry_list
1848                                  * hold repl. Thus, we cannot set replacement
1849                                  * as NULL, avoiding rdev NULL pointer
1850                                  * dereference in sync_request_write and
1851                                  * handle_write_finished.
1852                                  */
1853                                 err = -EBUSY;
1854                                 unfreeze_array(conf);
1855                                 goto abort;
1856                         }
1857                         clear_bit(Replacement, &repl->flags);
1858                         p->rdev = repl;
1859                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1860                         unfreeze_array(conf);
1861                 }
1862
1863                 clear_bit(WantReplacement, &rdev->flags);
1864                 err = md_integrity_register(mddev);
1865         }
1866 abort:
1867
1868         print_conf(conf);
1869         return err;
1870 }
1871
1872 static void end_sync_read(struct bio *bio)
1873 {
1874         struct r1bio *r1_bio = get_resync_r1bio(bio);
1875
1876         update_head_pos(r1_bio->read_disk, r1_bio);
1877
1878         /*
1879          * we have read a block, now it needs to be re-written,
1880          * or re-read if the read failed.
1881          * We don't do much here, just schedule handling by raid1d
1882          */
1883         if (!bio->bi_status)
1884                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1885
1886         if (atomic_dec_and_test(&r1_bio->remaining))
1887                 reschedule_retry(r1_bio);
1888 }
1889
1890 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1891 {
1892         sector_t sync_blocks = 0;
1893         sector_t s = r1_bio->sector;
1894         long sectors_to_go = r1_bio->sectors;
1895
1896         /* make sure these bits don't get cleared. */
1897         do {
1898                 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1899                 s += sync_blocks;
1900                 sectors_to_go -= sync_blocks;
1901         } while (sectors_to_go > 0);
1902 }
1903
1904 static void end_sync_write(struct bio *bio)
1905 {
1906         int uptodate = !bio->bi_status;
1907         struct r1bio *r1_bio = get_resync_r1bio(bio);
1908         struct mddev *mddev = r1_bio->mddev;
1909         struct r1conf *conf = mddev->private;
1910         sector_t first_bad;
1911         int bad_sectors;
1912         struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1913
1914         if (!uptodate) {
1915                 abort_sync_write(mddev, r1_bio);
1916                 set_bit(WriteErrorSeen, &rdev->flags);
1917                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1918                         set_bit(MD_RECOVERY_NEEDED, &
1919                                 mddev->recovery);
1920                 set_bit(R1BIO_WriteError, &r1_bio->state);
1921         } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1922                                &first_bad, &bad_sectors) &&
1923                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1924                                 r1_bio->sector,
1925                                 r1_bio->sectors,
1926                                 &first_bad, &bad_sectors)
1927                 )
1928                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1929
1930         if (atomic_dec_and_test(&r1_bio->remaining)) {
1931                 int s = r1_bio->sectors;
1932                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1933                     test_bit(R1BIO_WriteError, &r1_bio->state))
1934                         reschedule_retry(r1_bio);
1935                 else {
1936                         put_buf(r1_bio);
1937                         md_done_sync(mddev, s, uptodate);
1938                 }
1939         }
1940 }
1941
1942 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1943                             int sectors, struct page *page, int rw)
1944 {
1945         if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1946                 /* success */
1947                 return 1;
1948         if (rw == WRITE) {
1949                 set_bit(WriteErrorSeen, &rdev->flags);
1950                 if (!test_and_set_bit(WantReplacement,
1951                                       &rdev->flags))
1952                         set_bit(MD_RECOVERY_NEEDED, &
1953                                 rdev->mddev->recovery);
1954         }
1955         /* need to record an error - either for the block or the device */
1956         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1957                 md_error(rdev->mddev, rdev);
1958         return 0;
1959 }
1960
1961 static int fix_sync_read_error(struct r1bio *r1_bio)
1962 {
1963         /* Try some synchronous reads of other devices to get
1964          * good data, much like with normal read errors.  Only
1965          * read into the pages we already have so we don't
1966          * need to re-issue the read request.
1967          * We don't need to freeze the array, because being in an
1968          * active sync request, there is no normal IO, and
1969          * no overlapping syncs.
1970          * We don't need to check is_badblock() again as we
1971          * made sure that anything with a bad block in range
1972          * will have bi_end_io clear.
1973          */
1974         struct mddev *mddev = r1_bio->mddev;
1975         struct r1conf *conf = mddev->private;
1976         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1977         struct page **pages = get_resync_pages(bio)->pages;
1978         sector_t sect = r1_bio->sector;
1979         int sectors = r1_bio->sectors;
1980         int idx = 0;
1981         struct md_rdev *rdev;
1982
1983         rdev = conf->mirrors[r1_bio->read_disk].rdev;
1984         if (test_bit(FailFast, &rdev->flags)) {
1985                 /* Don't try recovering from here - just fail it
1986                  * ... unless it is the last working device of course */
1987                 md_error(mddev, rdev);
1988                 if (test_bit(Faulty, &rdev->flags))
1989                         /* Don't try to read from here, but make sure
1990                          * put_buf does it's thing
1991                          */
1992                         bio->bi_end_io = end_sync_write;
1993         }
1994
1995         while(sectors) {
1996                 int s = sectors;
1997                 int d = r1_bio->read_disk;
1998                 int success = 0;
1999                 int start;
2000
2001                 if (s > (PAGE_SIZE>>9))
2002                         s = PAGE_SIZE >> 9;
2003                 do {
2004                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2005                                 /* No rcu protection needed here devices
2006                                  * can only be removed when no resync is
2007                                  * active, and resync is currently active
2008                                  */
2009                                 rdev = conf->mirrors[d].rdev;
2010                                 if (sync_page_io(rdev, sect, s<<9,
2011                                                  pages[idx],
2012                                                  REQ_OP_READ, 0, false)) {
2013                                         success = 1;
2014                                         break;
2015                                 }
2016                         }
2017                         d++;
2018                         if (d == conf->raid_disks * 2)
2019                                 d = 0;
2020                 } while (!success && d != r1_bio->read_disk);
2021
2022                 if (!success) {
2023                         char b[BDEVNAME_SIZE];
2024                         int abort = 0;
2025                         /* Cannot read from anywhere, this block is lost.
2026                          * Record a bad block on each device.  If that doesn't
2027                          * work just disable and interrupt the recovery.
2028                          * Don't fail devices as that won't really help.
2029                          */
2030                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2031                                             mdname(mddev), bio_devname(bio, b),
2032                                             (unsigned long long)r1_bio->sector);
2033                         for (d = 0; d < conf->raid_disks * 2; d++) {
2034                                 rdev = conf->mirrors[d].rdev;
2035                                 if (!rdev || test_bit(Faulty, &rdev->flags))
2036                                         continue;
2037                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
2038                                         abort = 1;
2039                         }
2040                         if (abort) {
2041                                 conf->recovery_disabled =
2042                                         mddev->recovery_disabled;
2043                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2044                                 md_done_sync(mddev, r1_bio->sectors, 0);
2045                                 put_buf(r1_bio);
2046                                 return 0;
2047                         }
2048                         /* Try next page */
2049                         sectors -= s;
2050                         sect += s;
2051                         idx++;
2052                         continue;
2053                 }
2054
2055                 start = d;
2056                 /* write it back and re-read */
2057                 while (d != r1_bio->read_disk) {
2058                         if (d == 0)
2059                                 d = conf->raid_disks * 2;
2060                         d--;
2061                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2062                                 continue;
2063                         rdev = conf->mirrors[d].rdev;
2064                         if (r1_sync_page_io(rdev, sect, s,
2065                                             pages[idx],
2066                                             WRITE) == 0) {
2067                                 r1_bio->bios[d]->bi_end_io = NULL;
2068                                 rdev_dec_pending(rdev, mddev);
2069                         }
2070                 }
2071                 d = start;
2072                 while (d != r1_bio->read_disk) {
2073                         if (d == 0)
2074                                 d = conf->raid_disks * 2;
2075                         d--;
2076                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2077                                 continue;
2078                         rdev = conf->mirrors[d].rdev;
2079                         if (r1_sync_page_io(rdev, sect, s,
2080                                             pages[idx],
2081                                             READ) != 0)
2082                                 atomic_add(s, &rdev->corrected_errors);
2083                 }
2084                 sectors -= s;
2085                 sect += s;
2086                 idx ++;
2087         }
2088         set_bit(R1BIO_Uptodate, &r1_bio->state);
2089         bio->bi_status = 0;
2090         return 1;
2091 }
2092
2093 static void process_checks(struct r1bio *r1_bio)
2094 {
2095         /* We have read all readable devices.  If we haven't
2096          * got the block, then there is no hope left.
2097          * If we have, then we want to do a comparison
2098          * and skip the write if everything is the same.
2099          * If any blocks failed to read, then we need to
2100          * attempt an over-write
2101          */
2102         struct mddev *mddev = r1_bio->mddev;
2103         struct r1conf *conf = mddev->private;
2104         int primary;
2105         int i;
2106         int vcnt;
2107
2108         /* Fix variable parts of all bios */
2109         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2110         for (i = 0; i < conf->raid_disks * 2; i++) {
2111                 blk_status_t status;
2112                 struct bio *b = r1_bio->bios[i];
2113                 struct resync_pages *rp = get_resync_pages(b);
2114                 if (b->bi_end_io != end_sync_read)
2115                         continue;
2116                 /* fixup the bio for reuse, but preserve errno */
2117                 status = b->bi_status;
2118                 bio_reset(b);
2119                 b->bi_status = status;
2120                 b->bi_iter.bi_sector = r1_bio->sector +
2121                         conf->mirrors[i].rdev->data_offset;
2122                 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2123                 b->bi_end_io = end_sync_read;
2124                 rp->raid_bio = r1_bio;
2125                 b->bi_private = rp;
2126
2127                 /* initialize bvec table again */
2128                 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2129         }
2130         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2131                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2132                     !r1_bio->bios[primary]->bi_status) {
2133                         r1_bio->bios[primary]->bi_end_io = NULL;
2134                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2135                         break;
2136                 }
2137         r1_bio->read_disk = primary;
2138         for (i = 0; i < conf->raid_disks * 2; i++) {
2139                 int j = 0;
2140                 struct bio *pbio = r1_bio->bios[primary];
2141                 struct bio *sbio = r1_bio->bios[i];
2142                 blk_status_t status = sbio->bi_status;
2143                 struct page **ppages = get_resync_pages(pbio)->pages;
2144                 struct page **spages = get_resync_pages(sbio)->pages;
2145                 struct bio_vec *bi;
2146                 int page_len[RESYNC_PAGES] = { 0 };
2147                 struct bvec_iter_all iter_all;
2148
2149                 if (sbio->bi_end_io != end_sync_read)
2150                         continue;
2151                 /* Now we can 'fixup' the error value */
2152                 sbio->bi_status = 0;
2153
2154                 bio_for_each_segment_all(bi, sbio, iter_all)
2155                         page_len[j++] = bi->bv_len;
2156
2157                 if (!status) {
2158                         for (j = vcnt; j-- ; ) {
2159                                 if (memcmp(page_address(ppages[j]),
2160                                            page_address(spages[j]),
2161                                            page_len[j]))
2162                                         break;
2163                         }
2164                 } else
2165                         j = 0;
2166                 if (j >= 0)
2167                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2168                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2169                               && !status)) {
2170                         /* No need to write to this device. */
2171                         sbio->bi_end_io = NULL;
2172                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2173                         continue;
2174                 }
2175
2176                 bio_copy_data(sbio, pbio);
2177         }
2178 }
2179
2180 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2181 {
2182         struct r1conf *conf = mddev->private;
2183         int i;
2184         int disks = conf->raid_disks * 2;
2185         struct bio *wbio;
2186
2187         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2188                 /* ouch - failed to read all of that. */
2189                 if (!fix_sync_read_error(r1_bio))
2190                         return;
2191
2192         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2193                 process_checks(r1_bio);
2194
2195         /*
2196          * schedule writes
2197          */
2198         atomic_set(&r1_bio->remaining, 1);
2199         for (i = 0; i < disks ; i++) {
2200                 wbio = r1_bio->bios[i];
2201                 if (wbio->bi_end_io == NULL ||
2202                     (wbio->bi_end_io == end_sync_read &&
2203                      (i == r1_bio->read_disk ||
2204                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2205                         continue;
2206                 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2207                         abort_sync_write(mddev, r1_bio);
2208                         continue;
2209                 }
2210
2211                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2212                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2213                         wbio->bi_opf |= MD_FAILFAST;
2214
2215                 wbio->bi_end_io = end_sync_write;
2216                 atomic_inc(&r1_bio->remaining);
2217                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2218
2219                 generic_make_request(wbio);
2220         }
2221
2222         if (atomic_dec_and_test(&r1_bio->remaining)) {
2223                 /* if we're here, all write(s) have completed, so clean up */
2224                 int s = r1_bio->sectors;
2225                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2226                     test_bit(R1BIO_WriteError, &r1_bio->state))
2227                         reschedule_retry(r1_bio);
2228                 else {
2229                         put_buf(r1_bio);
2230                         md_done_sync(mddev, s, 1);
2231                 }
2232         }
2233 }
2234
2235 /*
2236  * This is a kernel thread which:
2237  *
2238  *      1.      Retries failed read operations on working mirrors.
2239  *      2.      Updates the raid superblock when problems encounter.
2240  *      3.      Performs writes following reads for array synchronising.
2241  */
2242
2243 static void fix_read_error(struct r1conf *conf, int read_disk,
2244                            sector_t sect, int sectors)
2245 {
2246         struct mddev *mddev = conf->mddev;
2247         while(sectors) {
2248                 int s = sectors;
2249                 int d = read_disk;
2250                 int success = 0;
2251                 int start;
2252                 struct md_rdev *rdev;
2253
2254                 if (s > (PAGE_SIZE>>9))
2255                         s = PAGE_SIZE >> 9;
2256
2257                 do {
2258                         sector_t first_bad;
2259                         int bad_sectors;
2260
2261                         rcu_read_lock();
2262                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2263                         if (rdev &&
2264                             (test_bit(In_sync, &rdev->flags) ||
2265                              (!test_bit(Faulty, &rdev->flags) &&
2266                               rdev->recovery_offset >= sect + s)) &&
2267                             is_badblock(rdev, sect, s,
2268                                         &first_bad, &bad_sectors) == 0) {
2269                                 atomic_inc(&rdev->nr_pending);
2270                                 rcu_read_unlock();
2271                                 if (sync_page_io(rdev, sect, s<<9,
2272                                          conf->tmppage, REQ_OP_READ, 0, false))
2273                                         success = 1;
2274                                 rdev_dec_pending(rdev, mddev);
2275                                 if (success)
2276                                         break;
2277                         } else
2278                                 rcu_read_unlock();
2279                         d++;
2280                         if (d == conf->raid_disks * 2)
2281                                 d = 0;
2282                 } while (!success && d != read_disk);
2283
2284                 if (!success) {
2285                         /* Cannot read from anywhere - mark it bad */
2286                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2287                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2288                                 md_error(mddev, rdev);
2289                         break;
2290                 }
2291                 /* write it back and re-read */
2292                 start = d;
2293                 while (d != read_disk) {
2294                         if (d==0)
2295                                 d = conf->raid_disks * 2;
2296                         d--;
2297                         rcu_read_lock();
2298                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2299                         if (rdev &&
2300                             !test_bit(Faulty, &rdev->flags)) {
2301                                 atomic_inc(&rdev->nr_pending);
2302                                 rcu_read_unlock();
2303                                 r1_sync_page_io(rdev, sect, s,
2304                                                 conf->tmppage, WRITE);
2305                                 rdev_dec_pending(rdev, mddev);
2306                         } else
2307                                 rcu_read_unlock();
2308                 }
2309                 d = start;
2310                 while (d != read_disk) {
2311                         char b[BDEVNAME_SIZE];
2312                         if (d==0)
2313                                 d = conf->raid_disks * 2;
2314                         d--;
2315                         rcu_read_lock();
2316                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2317                         if (rdev &&
2318                             !test_bit(Faulty, &rdev->flags)) {
2319                                 atomic_inc(&rdev->nr_pending);
2320                                 rcu_read_unlock();
2321                                 if (r1_sync_page_io(rdev, sect, s,
2322                                                     conf->tmppage, READ)) {
2323                                         atomic_add(s, &rdev->corrected_errors);
2324                                         pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2325                                                 mdname(mddev), s,
2326                                                 (unsigned long long)(sect +
2327                                                                      rdev->data_offset),
2328                                                 bdevname(rdev->bdev, b));
2329                                 }
2330                                 rdev_dec_pending(rdev, mddev);
2331                         } else
2332                                 rcu_read_unlock();
2333                 }
2334                 sectors -= s;
2335                 sect += s;
2336         }
2337 }
2338
2339 static int narrow_write_error(struct r1bio *r1_bio, int i)
2340 {
2341         struct mddev *mddev = r1_bio->mddev;
2342         struct r1conf *conf = mddev->private;
2343         struct md_rdev *rdev = conf->mirrors[i].rdev;
2344
2345         /* bio has the data to be written to device 'i' where
2346          * we just recently had a write error.
2347          * We repeatedly clone the bio and trim down to one block,
2348          * then try the write.  Where the write fails we record
2349          * a bad block.
2350          * It is conceivable that the bio doesn't exactly align with
2351          * blocks.  We must handle this somehow.
2352          *
2353          * We currently own a reference on the rdev.
2354          */
2355
2356         int block_sectors;
2357         sector_t sector;
2358         int sectors;
2359         int sect_to_write = r1_bio->sectors;
2360         int ok = 1;
2361
2362         if (rdev->badblocks.shift < 0)
2363                 return 0;
2364
2365         block_sectors = roundup(1 << rdev->badblocks.shift,
2366                                 bdev_logical_block_size(rdev->bdev) >> 9);
2367         sector = r1_bio->sector;
2368         sectors = ((sector + block_sectors)
2369                    & ~(sector_t)(block_sectors - 1))
2370                 - sector;
2371
2372         while (sect_to_write) {
2373                 struct bio *wbio;
2374                 if (sectors > sect_to_write)
2375                         sectors = sect_to_write;
2376                 /* Write at 'sector' for 'sectors'*/
2377
2378                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2379                         wbio = bio_clone_fast(r1_bio->behind_master_bio,
2380                                               GFP_NOIO,
2381                                               &mddev->bio_set);
2382                 } else {
2383                         wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2384                                               &mddev->bio_set);
2385                 }
2386
2387                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2388                 wbio->bi_iter.bi_sector = r1_bio->sector;
2389                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2390
2391                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2392                 wbio->bi_iter.bi_sector += rdev->data_offset;
2393                 bio_set_dev(wbio, rdev->bdev);
2394
2395                 if (submit_bio_wait(wbio) < 0)
2396                         /* failure! */
2397                         ok = rdev_set_badblocks(rdev, sector,
2398                                                 sectors, 0)
2399                                 && ok;
2400
2401                 bio_put(wbio);
2402                 sect_to_write -= sectors;
2403                 sector += sectors;
2404                 sectors = block_sectors;
2405         }
2406         return ok;
2407 }
2408
2409 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2410 {
2411         int m;
2412         int s = r1_bio->sectors;
2413         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2414                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2415                 struct bio *bio = r1_bio->bios[m];
2416                 if (bio->bi_end_io == NULL)
2417                         continue;
2418                 if (!bio->bi_status &&
2419                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2420                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2421                 }
2422                 if (bio->bi_status &&
2423                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2424                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2425                                 md_error(conf->mddev, rdev);
2426                 }
2427         }
2428         put_buf(r1_bio);
2429         md_done_sync(conf->mddev, s, 1);
2430 }
2431
2432 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2433 {
2434         int m, idx;
2435         bool fail = false;
2436
2437         for (m = 0; m < conf->raid_disks * 2 ; m++)
2438                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2439                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2440                         rdev_clear_badblocks(rdev,
2441                                              r1_bio->sector,
2442                                              r1_bio->sectors, 0);
2443                         rdev_dec_pending(rdev, conf->mddev);
2444                 } else if (r1_bio->bios[m] != NULL) {
2445                         /* This drive got a write error.  We need to
2446                          * narrow down and record precise write
2447                          * errors.
2448                          */
2449                         fail = true;
2450                         if (!narrow_write_error(r1_bio, m)) {
2451                                 md_error(conf->mddev,
2452                                          conf->mirrors[m].rdev);
2453                                 /* an I/O failed, we can't clear the bitmap */
2454                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2455                         }
2456                         rdev_dec_pending(conf->mirrors[m].rdev,
2457                                          conf->mddev);
2458                 }
2459         if (fail) {
2460                 spin_lock_irq(&conf->device_lock);
2461                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2462                 idx = sector_to_idx(r1_bio->sector);
2463                 atomic_inc(&conf->nr_queued[idx]);
2464                 spin_unlock_irq(&conf->device_lock);
2465                 /*
2466                  * In case freeze_array() is waiting for condition
2467                  * get_unqueued_pending() == extra to be true.
2468                  */
2469                 wake_up(&conf->wait_barrier);
2470                 md_wakeup_thread(conf->mddev->thread);
2471         } else {
2472                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2473                         close_write(r1_bio);
2474                 raid_end_bio_io(r1_bio);
2475         }
2476 }
2477
2478 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2479 {
2480         struct mddev *mddev = conf->mddev;
2481         struct bio *bio;
2482         struct md_rdev *rdev;
2483
2484         clear_bit(R1BIO_ReadError, &r1_bio->state);
2485         /* we got a read error. Maybe the drive is bad.  Maybe just
2486          * the block and we can fix it.
2487          * We freeze all other IO, and try reading the block from
2488          * other devices.  When we find one, we re-write
2489          * and check it that fixes the read error.
2490          * This is all done synchronously while the array is
2491          * frozen
2492          */
2493
2494         bio = r1_bio->bios[r1_bio->read_disk];
2495         bio_put(bio);
2496         r1_bio->bios[r1_bio->read_disk] = NULL;
2497
2498         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2499         if (mddev->ro == 0
2500             && !test_bit(FailFast, &rdev->flags)) {
2501                 freeze_array(conf, 1);
2502                 fix_read_error(conf, r1_bio->read_disk,
2503                                r1_bio->sector, r1_bio->sectors);
2504                 unfreeze_array(conf);
2505         } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2506                 md_error(mddev, rdev);
2507         } else {
2508                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2509         }
2510
2511         rdev_dec_pending(rdev, conf->mddev);
2512         allow_barrier(conf, r1_bio->sector);
2513         bio = r1_bio->master_bio;
2514
2515         /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2516         r1_bio->state = 0;
2517         raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2518 }
2519
2520 static void raid1d(struct md_thread *thread)
2521 {
2522         struct mddev *mddev = thread->mddev;
2523         struct r1bio *r1_bio;
2524         unsigned long flags;
2525         struct r1conf *conf = mddev->private;
2526         struct list_head *head = &conf->retry_list;
2527         struct blk_plug plug;
2528         int idx;
2529
2530         md_check_recovery(mddev);
2531
2532         if (!list_empty_careful(&conf->bio_end_io_list) &&
2533             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2534                 LIST_HEAD(tmp);
2535                 spin_lock_irqsave(&conf->device_lock, flags);
2536                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2537                         list_splice_init(&conf->bio_end_io_list, &tmp);
2538                 spin_unlock_irqrestore(&conf->device_lock, flags);
2539                 while (!list_empty(&tmp)) {
2540                         r1_bio = list_first_entry(&tmp, struct r1bio,
2541                                                   retry_list);
2542                         list_del(&r1_bio->retry_list);
2543                         idx = sector_to_idx(r1_bio->sector);
2544                         atomic_dec(&conf->nr_queued[idx]);
2545                         if (mddev->degraded)
2546                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2547                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2548                                 close_write(r1_bio);
2549                         raid_end_bio_io(r1_bio);
2550                 }
2551         }
2552
2553         blk_start_plug(&plug);
2554         for (;;) {
2555
2556                 flush_pending_writes(conf);
2557
2558                 spin_lock_irqsave(&conf->device_lock, flags);
2559                 if (list_empty(head)) {
2560                         spin_unlock_irqrestore(&conf->device_lock, flags);
2561                         break;
2562                 }
2563                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2564                 list_del(head->prev);
2565                 idx = sector_to_idx(r1_bio->sector);
2566                 atomic_dec(&conf->nr_queued[idx]);
2567                 spin_unlock_irqrestore(&conf->device_lock, flags);
2568
2569                 mddev = r1_bio->mddev;
2570                 conf = mddev->private;
2571                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2572                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2573                             test_bit(R1BIO_WriteError, &r1_bio->state))
2574                                 handle_sync_write_finished(conf, r1_bio);
2575                         else
2576                                 sync_request_write(mddev, r1_bio);
2577                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2578                            test_bit(R1BIO_WriteError, &r1_bio->state))
2579                         handle_write_finished(conf, r1_bio);
2580                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2581                         handle_read_error(conf, r1_bio);
2582                 else
2583                         WARN_ON_ONCE(1);
2584
2585                 cond_resched();
2586                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2587                         md_check_recovery(mddev);
2588         }
2589         blk_finish_plug(&plug);
2590 }
2591
2592 static int init_resync(struct r1conf *conf)
2593 {
2594         int buffs;
2595
2596         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2597         BUG_ON(mempool_initialized(&conf->r1buf_pool));
2598
2599         return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2600                             r1buf_pool_free, conf->poolinfo);
2601 }
2602
2603 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2604 {
2605         struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2606         struct resync_pages *rps;
2607         struct bio *bio;
2608         int i;
2609
2610         for (i = conf->poolinfo->raid_disks; i--; ) {
2611                 bio = r1bio->bios[i];
2612                 rps = bio->bi_private;
2613                 bio_reset(bio);
2614                 bio->bi_private = rps;
2615         }
2616         r1bio->master_bio = NULL;
2617         return r1bio;
2618 }
2619
2620 /*
2621  * perform a "sync" on one "block"
2622  *
2623  * We need to make sure that no normal I/O request - particularly write
2624  * requests - conflict with active sync requests.
2625  *
2626  * This is achieved by tracking pending requests and a 'barrier' concept
2627  * that can be installed to exclude normal IO requests.
2628  */
2629
2630 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2631                                    int *skipped)
2632 {
2633         struct r1conf *conf = mddev->private;
2634         struct r1bio *r1_bio;
2635         struct bio *bio;
2636         sector_t max_sector, nr_sectors;
2637         int disk = -1;
2638         int i;
2639         int wonly = -1;
2640         int write_targets = 0, read_targets = 0;
2641         sector_t sync_blocks;
2642         int still_degraded = 0;
2643         int good_sectors = RESYNC_SECTORS;
2644         int min_bad = 0; /* number of sectors that are bad in all devices */
2645         int idx = sector_to_idx(sector_nr);
2646         int page_idx = 0;
2647
2648         if (!mempool_initialized(&conf->r1buf_pool))
2649                 if (init_resync(conf))
2650                         return 0;
2651
2652         max_sector = mddev->dev_sectors;
2653         if (sector_nr >= max_sector) {
2654                 /* If we aborted, we need to abort the
2655                  * sync on the 'current' bitmap chunk (there will
2656                  * only be one in raid1 resync.
2657                  * We can find the current addess in mddev->curr_resync
2658                  */
2659                 if (mddev->curr_resync < max_sector) /* aborted */
2660                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2661                                            &sync_blocks, 1);
2662                 else /* completed sync */
2663                         conf->fullsync = 0;
2664
2665                 md_bitmap_close_sync(mddev->bitmap);
2666                 close_sync(conf);
2667
2668                 if (mddev_is_clustered(mddev)) {
2669                         conf->cluster_sync_low = 0;
2670                         conf->cluster_sync_high = 0;
2671                 }
2672                 return 0;
2673         }
2674
2675         if (mddev->bitmap == NULL &&
2676             mddev->recovery_cp == MaxSector &&
2677             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2678             conf->fullsync == 0) {
2679                 *skipped = 1;
2680                 return max_sector - sector_nr;
2681         }
2682         /* before building a request, check if we can skip these blocks..
2683          * This call the bitmap_start_sync doesn't actually record anything
2684          */
2685         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2686             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2687                 /* We can skip this block, and probably several more */
2688                 *skipped = 1;
2689                 return sync_blocks;
2690         }
2691
2692         /*
2693          * If there is non-resync activity waiting for a turn, then let it
2694          * though before starting on this new sync request.
2695          */
2696         if (atomic_read(&conf->nr_waiting[idx]))
2697                 schedule_timeout_uninterruptible(1);
2698
2699         /* we are incrementing sector_nr below. To be safe, we check against
2700          * sector_nr + two times RESYNC_SECTORS
2701          */
2702
2703         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2704                 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2705
2706
2707         if (raise_barrier(conf, sector_nr))
2708                 return 0;
2709
2710         r1_bio = raid1_alloc_init_r1buf(conf);
2711
2712         rcu_read_lock();
2713         /*
2714          * If we get a correctably read error during resync or recovery,
2715          * we might want to read from a different device.  So we
2716          * flag all drives that could conceivably be read from for READ,
2717          * and any others (which will be non-In_sync devices) for WRITE.
2718          * If a read fails, we try reading from something else for which READ
2719          * is OK.
2720          */
2721
2722         r1_bio->mddev = mddev;
2723         r1_bio->sector = sector_nr;
2724         r1_bio->state = 0;
2725         set_bit(R1BIO_IsSync, &r1_bio->state);
2726         /* make sure good_sectors won't go across barrier unit boundary */
2727         good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2728
2729         for (i = 0; i < conf->raid_disks * 2; i++) {
2730                 struct md_rdev *rdev;
2731                 bio = r1_bio->bios[i];
2732
2733                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2734                 if (rdev == NULL ||
2735                     test_bit(Faulty, &rdev->flags)) {
2736                         if (i < conf->raid_disks)
2737                                 still_degraded = 1;
2738                 } else if (!test_bit(In_sync, &rdev->flags)) {
2739                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2740                         bio->bi_end_io = end_sync_write;
2741                         write_targets ++;
2742                 } else {
2743                         /* may need to read from here */
2744                         sector_t first_bad = MaxSector;
2745                         int bad_sectors;
2746
2747                         if (is_badblock(rdev, sector_nr, good_sectors,
2748                                         &first_bad, &bad_sectors)) {
2749                                 if (first_bad > sector_nr)
2750                                         good_sectors = first_bad - sector_nr;
2751                                 else {
2752                                         bad_sectors -= (sector_nr - first_bad);
2753                                         if (min_bad == 0 ||
2754                                             min_bad > bad_sectors)
2755                                                 min_bad = bad_sectors;
2756                                 }
2757                         }
2758                         if (sector_nr < first_bad) {
2759                                 if (test_bit(WriteMostly, &rdev->flags)) {
2760                                         if (wonly < 0)
2761                                                 wonly = i;
2762                                 } else {
2763                                         if (disk < 0)
2764                                                 disk = i;
2765                                 }
2766                                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2767                                 bio->bi_end_io = end_sync_read;
2768                                 read_targets++;
2769                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2770                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2771                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2772                                 /*
2773                                  * The device is suitable for reading (InSync),
2774                                  * but has bad block(s) here. Let's try to correct them,
2775                                  * if we are doing resync or repair. Otherwise, leave
2776                                  * this device alone for this sync request.
2777                                  */
2778                                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2779                                 bio->bi_end_io = end_sync_write;
2780                                 write_targets++;
2781                         }
2782                 }
2783                 if (bio->bi_end_io) {
2784                         atomic_inc(&rdev->nr_pending);
2785                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2786                         bio_set_dev(bio, rdev->bdev);
2787                         if (test_bit(FailFast, &rdev->flags))
2788                                 bio->bi_opf |= MD_FAILFAST;
2789                 }
2790         }
2791         rcu_read_unlock();
2792         if (disk < 0)
2793                 disk = wonly;
2794         r1_bio->read_disk = disk;
2795
2796         if (read_targets == 0 && min_bad > 0) {
2797                 /* These sectors are bad on all InSync devices, so we
2798                  * need to mark them bad on all write targets
2799                  */
2800                 int ok = 1;
2801                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2802                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2803                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2804                                 ok = rdev_set_badblocks(rdev, sector_nr,
2805                                                         min_bad, 0
2806                                         ) && ok;
2807                         }
2808                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2809                 *skipped = 1;
2810                 put_buf(r1_bio);
2811
2812                 if (!ok) {
2813                         /* Cannot record the badblocks, so need to
2814                          * abort the resync.
2815                          * If there are multiple read targets, could just
2816                          * fail the really bad ones ???
2817                          */
2818                         conf->recovery_disabled = mddev->recovery_disabled;
2819                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2820                         return 0;
2821                 } else
2822                         return min_bad;
2823
2824         }
2825         if (min_bad > 0 && min_bad < good_sectors) {
2826                 /* only resync enough to reach the next bad->good
2827                  * transition */
2828                 good_sectors = min_bad;
2829         }
2830
2831         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2832                 /* extra read targets are also write targets */
2833                 write_targets += read_targets-1;
2834
2835         if (write_targets == 0 || read_targets == 0) {
2836                 /* There is nowhere to write, so all non-sync
2837                  * drives must be failed - so we are finished
2838                  */
2839                 sector_t rv;
2840                 if (min_bad > 0)
2841                         max_sector = sector_nr + min_bad;
2842                 rv = max_sector - sector_nr;
2843                 *skipped = 1;
2844                 put_buf(r1_bio);
2845                 return rv;
2846         }
2847
2848         if (max_sector > mddev->resync_max)
2849                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2850         if (max_sector > sector_nr + good_sectors)
2851                 max_sector = sector_nr + good_sectors;
2852         nr_sectors = 0;
2853         sync_blocks = 0;
2854         do {
2855                 struct page *page;
2856                 int len = PAGE_SIZE;
2857                 if (sector_nr + (len>>9) > max_sector)
2858                         len = (max_sector - sector_nr) << 9;
2859                 if (len == 0)
2860                         break;
2861                 if (sync_blocks == 0) {
2862                         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2863                                                   &sync_blocks, still_degraded) &&
2864                             !conf->fullsync &&
2865                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2866                                 break;
2867                         if ((len >> 9) > sync_blocks)
2868                                 len = sync_blocks<<9;
2869                 }
2870
2871                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2872                         struct resync_pages *rp;
2873
2874                         bio = r1_bio->bios[i];
2875                         rp = get_resync_pages(bio);
2876                         if (bio->bi_end_io) {
2877                                 page = resync_fetch_page(rp, page_idx);
2878
2879                                 /*
2880                                  * won't fail because the vec table is big
2881                                  * enough to hold all these pages
2882                                  */
2883                                 bio_add_page(bio, page, len, 0);
2884                         }
2885                 }
2886                 nr_sectors += len>>9;
2887                 sector_nr += len>>9;
2888                 sync_blocks -= (len>>9);
2889         } while (++page_idx < RESYNC_PAGES);
2890
2891         r1_bio->sectors = nr_sectors;
2892
2893         if (mddev_is_clustered(mddev) &&
2894                         conf->cluster_sync_high < sector_nr + nr_sectors) {
2895                 conf->cluster_sync_low = mddev->curr_resync_completed;
2896                 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2897                 /* Send resync message */
2898                 md_cluster_ops->resync_info_update(mddev,
2899                                 conf->cluster_sync_low,
2900                                 conf->cluster_sync_high);
2901         }
2902
2903         /* For a user-requested sync, we read all readable devices and do a
2904          * compare
2905          */
2906         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2907                 atomic_set(&r1_bio->remaining, read_targets);
2908                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2909                         bio = r1_bio->bios[i];
2910                         if (bio->bi_end_io == end_sync_read) {
2911                                 read_targets--;
2912                                 md_sync_acct_bio(bio, nr_sectors);
2913                                 if (read_targets == 1)
2914                                         bio->bi_opf &= ~MD_FAILFAST;
2915                                 generic_make_request(bio);
2916                         }
2917                 }
2918         } else {
2919                 atomic_set(&r1_bio->remaining, 1);
2920                 bio = r1_bio->bios[r1_bio->read_disk];
2921                 md_sync_acct_bio(bio, nr_sectors);
2922                 if (read_targets == 1)
2923                         bio->bi_opf &= ~MD_FAILFAST;
2924                 generic_make_request(bio);
2925         }
2926         return nr_sectors;
2927 }
2928
2929 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2930 {
2931         if (sectors)
2932                 return sectors;
2933
2934         return mddev->dev_sectors;
2935 }
2936
2937 static struct r1conf *setup_conf(struct mddev *mddev)
2938 {
2939         struct r1conf *conf;
2940         int i;
2941         struct raid1_info *disk;
2942         struct md_rdev *rdev;
2943         int err = -ENOMEM;
2944
2945         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2946         if (!conf)
2947                 goto abort;
2948
2949         conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2950                                    sizeof(atomic_t), GFP_KERNEL);
2951         if (!conf->nr_pending)
2952                 goto abort;
2953
2954         conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2955                                    sizeof(atomic_t), GFP_KERNEL);
2956         if (!conf->nr_waiting)
2957                 goto abort;
2958
2959         conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2960                                   sizeof(atomic_t), GFP_KERNEL);
2961         if (!conf->nr_queued)
2962                 goto abort;
2963
2964         conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2965                                 sizeof(atomic_t), GFP_KERNEL);
2966         if (!conf->barrier)
2967                 goto abort;
2968
2969         conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2970                                             mddev->raid_disks, 2),
2971                                 GFP_KERNEL);
2972         if (!conf->mirrors)
2973                 goto abort;
2974
2975         conf->tmppage = alloc_page(GFP_KERNEL);
2976         if (!conf->tmppage)
2977                 goto abort;
2978
2979         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2980         if (!conf->poolinfo)
2981                 goto abort;
2982         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2983         err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2984                            rbio_pool_free, conf->poolinfo);
2985         if (err)
2986                 goto abort;
2987
2988         err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2989         if (err)
2990                 goto abort;
2991
2992         conf->poolinfo->mddev = mddev;
2993
2994         err = -EINVAL;
2995         spin_lock_init(&conf->device_lock);
2996         rdev_for_each(rdev, mddev) {
2997                 int disk_idx = rdev->raid_disk;
2998                 if (disk_idx >= mddev->raid_disks
2999                     || disk_idx < 0)
3000                         continue;
3001                 if (test_bit(Replacement, &rdev->flags))
3002                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
3003                 else
3004                         disk = conf->mirrors + disk_idx;
3005
3006                 if (disk->rdev)
3007                         goto abort;
3008                 disk->rdev = rdev;
3009                 disk->head_position = 0;
3010                 disk->seq_start = MaxSector;
3011         }
3012         conf->raid_disks = mddev->raid_disks;
3013         conf->mddev = mddev;
3014         INIT_LIST_HEAD(&conf->retry_list);
3015         INIT_LIST_HEAD(&conf->bio_end_io_list);
3016
3017         spin_lock_init(&conf->resync_lock);
3018         init_waitqueue_head(&conf->wait_barrier);
3019
3020         bio_list_init(&conf->pending_bio_list);
3021         conf->pending_count = 0;
3022         conf->recovery_disabled = mddev->recovery_disabled - 1;
3023
3024         err = -EIO;
3025         for (i = 0; i < conf->raid_disks * 2; i++) {
3026
3027                 disk = conf->mirrors + i;
3028
3029                 if (i < conf->raid_disks &&
3030                     disk[conf->raid_disks].rdev) {
3031                         /* This slot has a replacement. */
3032                         if (!disk->rdev) {
3033                                 /* No original, just make the replacement
3034                                  * a recovering spare
3035                                  */
3036                                 disk->rdev =
3037                                         disk[conf->raid_disks].rdev;
3038                                 disk[conf->raid_disks].rdev = NULL;
3039                         } else if (!test_bit(In_sync, &disk->rdev->flags))
3040                                 /* Original is not in_sync - bad */
3041                                 goto abort;
3042                 }
3043
3044                 if (!disk->rdev ||
3045                     !test_bit(In_sync, &disk->rdev->flags)) {
3046                         disk->head_position = 0;
3047                         if (disk->rdev &&
3048                             (disk->rdev->saved_raid_disk < 0))
3049                                 conf->fullsync = 1;
3050                 }
3051         }
3052
3053         err = -ENOMEM;
3054         conf->thread = md_register_thread(raid1d, mddev, "raid1");
3055         if (!conf->thread)
3056                 goto abort;
3057
3058         return conf;
3059
3060  abort:
3061         if (conf) {
3062                 mempool_exit(&conf->r1bio_pool);
3063                 kfree(conf->mirrors);
3064                 safe_put_page(conf->tmppage);
3065                 kfree(conf->poolinfo);
3066                 kfree(conf->nr_pending);
3067                 kfree(conf->nr_waiting);
3068                 kfree(conf->nr_queued);
3069                 kfree(conf->barrier);
3070                 bioset_exit(&conf->bio_split);
3071                 kfree(conf);
3072         }
3073         return ERR_PTR(err);
3074 }
3075
3076 static void raid1_free(struct mddev *mddev, void *priv);
3077 static int raid1_run(struct mddev *mddev)
3078 {
3079         struct r1conf *conf;
3080         int i;
3081         struct md_rdev *rdev;
3082         int ret;
3083         bool discard_supported = false;
3084
3085         if (mddev->level != 1) {
3086                 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3087                         mdname(mddev), mddev->level);
3088                 return -EIO;
3089         }
3090         if (mddev->reshape_position != MaxSector) {
3091                 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3092                         mdname(mddev));
3093                 return -EIO;
3094         }
3095         if (mddev_init_writes_pending(mddev) < 0)
3096                 return -ENOMEM;
3097         /*
3098          * copy the already verified devices into our private RAID1
3099          * bookkeeping area. [whatever we allocate in run(),
3100          * should be freed in raid1_free()]
3101          */
3102         if (mddev->private == NULL)
3103                 conf = setup_conf(mddev);
3104         else
3105                 conf = mddev->private;
3106
3107         if (IS_ERR(conf))
3108                 return PTR_ERR(conf);
3109
3110         if (mddev->queue) {
3111                 blk_queue_max_write_same_sectors(mddev->queue, 0);
3112                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3113         }
3114
3115         rdev_for_each(rdev, mddev) {
3116                 if (!mddev->gendisk)
3117                         continue;
3118                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3119                                   rdev->data_offset << 9);
3120                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3121                         discard_supported = true;
3122         }
3123
3124         mddev->degraded = 0;
3125         for (i = 0; i < conf->raid_disks; i++)
3126                 if (conf->mirrors[i].rdev == NULL ||
3127                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3128                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3129                         mddev->degraded++;
3130
3131         if (conf->raid_disks - mddev->degraded == 1)
3132                 mddev->recovery_cp = MaxSector;
3133
3134         if (mddev->recovery_cp != MaxSector)
3135                 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3136                         mdname(mddev));
3137         pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3138                 mdname(mddev), mddev->raid_disks - mddev->degraded,
3139                 mddev->raid_disks);
3140
3141         /*
3142          * Ok, everything is just fine now
3143          */
3144         mddev->thread = conf->thread;
3145         conf->thread = NULL;
3146         mddev->private = conf;
3147         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3148
3149         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3150
3151         if (mddev->queue) {
3152                 if (discard_supported)
3153                         blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3154                                                 mddev->queue);
3155                 else
3156                         blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3157                                                   mddev->queue);
3158         }
3159
3160         ret = md_integrity_register(mddev);
3161         if (ret) {
3162                 md_unregister_thread(&mddev->thread);
3163                 raid1_free(mddev, conf);
3164         }
3165         return ret;
3166 }
3167
3168 static void raid1_free(struct mddev *mddev, void *priv)
3169 {
3170         struct r1conf *conf = priv;
3171
3172         mempool_exit(&conf->r1bio_pool);
3173         kfree(conf->mirrors);
3174         safe_put_page(conf->tmppage);
3175         kfree(conf->poolinfo);
3176         kfree(conf->nr_pending);
3177         kfree(conf->nr_waiting);
3178         kfree(conf->nr_queued);
3179         kfree(conf->barrier);
3180         bioset_exit(&conf->bio_split);
3181         kfree(conf);
3182 }
3183
3184 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3185 {
3186         /* no resync is happening, and there is enough space
3187          * on all devices, so we can resize.
3188          * We need to make sure resync covers any new space.
3189          * If the array is shrinking we should possibly wait until
3190          * any io in the removed space completes, but it hardly seems
3191          * worth it.
3192          */
3193         sector_t newsize = raid1_size(mddev, sectors, 0);
3194         if (mddev->external_size &&
3195             mddev->array_sectors > newsize)
3196                 return -EINVAL;
3197         if (mddev->bitmap) {
3198                 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3199                 if (ret)
3200                         return ret;
3201         }
3202         md_set_array_sectors(mddev, newsize);
3203         if (sectors > mddev->dev_sectors &&
3204             mddev->recovery_cp > mddev->dev_sectors) {
3205                 mddev->recovery_cp = mddev->dev_sectors;
3206                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3207         }
3208         mddev->dev_sectors = sectors;
3209         mddev->resync_max_sectors = sectors;
3210         return 0;
3211 }
3212
3213 static int raid1_reshape(struct mddev *mddev)
3214 {
3215         /* We need to:
3216          * 1/ resize the r1bio_pool
3217          * 2/ resize conf->mirrors
3218          *
3219          * We allocate a new r1bio_pool if we can.
3220          * Then raise a device barrier and wait until all IO stops.
3221          * Then resize conf->mirrors and swap in the new r1bio pool.
3222          *
3223          * At the same time, we "pack" the devices so that all the missing
3224          * devices have the higher raid_disk numbers.
3225          */
3226         mempool_t newpool, oldpool;
3227         struct pool_info *newpoolinfo;
3228         struct raid1_info *newmirrors;
3229         struct r1conf *conf = mddev->private;
3230         int cnt, raid_disks;
3231         unsigned long flags;
3232         int d, d2;
3233         int ret;
3234
3235         memset(&newpool, 0, sizeof(newpool));
3236         memset(&oldpool, 0, sizeof(oldpool));
3237
3238         /* Cannot change chunk_size, layout, or level */
3239         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3240             mddev->layout != mddev->new_layout ||
3241             mddev->level != mddev->new_level) {
3242                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3243                 mddev->new_layout = mddev->layout;
3244                 mddev->new_level = mddev->level;
3245                 return -EINVAL;
3246         }
3247
3248         if (!mddev_is_clustered(mddev))
3249                 md_allow_write(mddev);
3250
3251         raid_disks = mddev->raid_disks + mddev->delta_disks;
3252
3253         if (raid_disks < conf->raid_disks) {
3254                 cnt=0;
3255                 for (d= 0; d < conf->raid_disks; d++)
3256                         if (conf->mirrors[d].rdev)
3257                                 cnt++;
3258                 if (cnt > raid_disks)
3259                         return -EBUSY;
3260         }
3261
3262         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3263         if (!newpoolinfo)
3264                 return -ENOMEM;
3265         newpoolinfo->mddev = mddev;
3266         newpoolinfo->raid_disks = raid_disks * 2;
3267
3268         ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3269                            rbio_pool_free, newpoolinfo);
3270         if (ret) {
3271                 kfree(newpoolinfo);
3272                 return ret;
3273         }
3274         newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3275                                          raid_disks, 2),
3276                              GFP_KERNEL);
3277         if (!newmirrors) {
3278                 kfree(newpoolinfo);
3279                 mempool_exit(&newpool);
3280                 return -ENOMEM;
3281         }
3282
3283         freeze_array(conf, 0);
3284
3285         /* ok, everything is stopped */
3286         oldpool = conf->r1bio_pool;
3287         conf->r1bio_pool = newpool;
3288
3289         for (d = d2 = 0; d < conf->raid_disks; d++) {
3290                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3291                 if (rdev && rdev->raid_disk != d2) {
3292                         sysfs_unlink_rdev(mddev, rdev);
3293                         rdev->raid_disk = d2;
3294                         sysfs_unlink_rdev(mddev, rdev);
3295                         if (sysfs_link_rdev(mddev, rdev))
3296                                 pr_warn("md/raid1:%s: cannot register rd%d\n",
3297                                         mdname(mddev), rdev->raid_disk);
3298                 }
3299                 if (rdev)
3300                         newmirrors[d2++].rdev = rdev;
3301         }
3302         kfree(conf->mirrors);
3303         conf->mirrors = newmirrors;
3304         kfree(conf->poolinfo);
3305         conf->poolinfo = newpoolinfo;
3306
3307         spin_lock_irqsave(&conf->device_lock, flags);
3308         mddev->degraded += (raid_disks - conf->raid_disks);
3309         spin_unlock_irqrestore(&conf->device_lock, flags);
3310         conf->raid_disks = mddev->raid_disks = raid_disks;
3311         mddev->delta_disks = 0;
3312
3313         unfreeze_array(conf);
3314
3315         set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3316         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3317         md_wakeup_thread(mddev->thread);
3318
3319         mempool_exit(&oldpool);
3320         return 0;
3321 }
3322
3323 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3324 {
3325         struct r1conf *conf = mddev->private;
3326
3327         if (quiesce)
3328                 freeze_array(conf, 0);
3329         else
3330                 unfreeze_array(conf);
3331 }
3332
3333 static void *raid1_takeover(struct mddev *mddev)
3334 {
3335         /* raid1 can take over:
3336          *  raid5 with 2 devices, any layout or chunk size
3337          */
3338         if (mddev->level == 5 && mddev->raid_disks == 2) {
3339                 struct r1conf *conf;
3340                 mddev->new_level = 1;
3341                 mddev->new_layout = 0;
3342                 mddev->new_chunk_sectors = 0;
3343                 conf = setup_conf(mddev);
3344                 if (!IS_ERR(conf)) {
3345                         /* Array must appear to be quiesced */
3346                         conf->array_frozen = 1;
3347                         mddev_clear_unsupported_flags(mddev,
3348                                 UNSUPPORTED_MDDEV_FLAGS);
3349                 }
3350                 return conf;
3351         }
3352         return ERR_PTR(-EINVAL);
3353 }
3354
3355 static struct md_personality raid1_personality =
3356 {
3357         .name           = "raid1",
3358         .level          = 1,
3359         .owner          = THIS_MODULE,
3360         .make_request   = raid1_make_request,
3361         .run            = raid1_run,
3362         .free           = raid1_free,
3363         .status         = raid1_status,
3364         .error_handler  = raid1_error,
3365         .hot_add_disk   = raid1_add_disk,
3366         .hot_remove_disk= raid1_remove_disk,
3367         .spare_active   = raid1_spare_active,
3368         .sync_request   = raid1_sync_request,
3369         .resize         = raid1_resize,
3370         .size           = raid1_size,
3371         .check_reshape  = raid1_reshape,
3372         .quiesce        = raid1_quiesce,
3373         .takeover       = raid1_takeover,
3374         .congested      = raid1_congested,
3375 };
3376
3377 static int __init raid_init(void)
3378 {
3379         return register_md_personality(&raid1_personality);
3380 }
3381
3382 static void raid_exit(void)
3383 {
3384         unregister_md_personality(&raid1_personality);
3385 }
3386
3387 module_init(raid_init);
3388 module_exit(raid_exit);
3389 MODULE_LICENSE("GPL");
3390 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3391 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3392 MODULE_ALIAS("md-raid1");
3393 MODULE_ALIAS("md-level-1");
3394
3395 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);