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                 }
 451
 452                 /*
 453                  * When the device is faulty, it is not necessary to
 454                  * handle write error.
 455                  * For failfast, this is the only remaining device,
 456                  * We need to retry the write without FailFast.
 457                  */
 458                 if (!test_bit(Faulty, &rdev->flags))
 459                         set_bit(R1BIO_WriteError, &r1_bio->state);
 460                 else {
 461                         /* Finished with this branch */
 462                         r1_bio->bios[mirror] = NULL;
 463                         to_put = bio;
 464                 }
 465         } else {
 466                 /*
 467                  * Set R1BIO_Uptodate in our master bio, so that we
 468                  * will return a good error code for to the higher
 469                  * levels even if IO on some other mirrored buffer
 470                  * fails.
 471                  *
 472                  * The 'master' represents the composite IO operation
 473                  * to user-side. So if something waits for IO, then it
 474                  * will wait for the 'master' bio.
 475                  */
 476                 sector_t first_bad;
 477                 int bad_sectors;
 478
 479                 r1_bio->bios[mirror] = NULL;
 480                 to_put = bio;
 481                 /*
 482                  * Do not set R1BIO_Uptodate if the current device is
 483                  * rebuilding or Faulty. This is because we cannot use
 484                  * such device for properly reading the data back (we could
 485                  * potentially use it, if the current write would have felt
 486                  * before rdev->recovery_offset, but for simplicity we don't
 487                  * check this here.
 488                  */
 489                 if (test_bit(In_sync, &rdev->flags) &&
 490                     !test_bit(Faulty, &rdev->flags))
 491                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 492
 493                 /* Maybe we can clear some bad blocks. */
 494                 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 495                                 &first_bad, &bad_sectors) && !discard_error) {
 496                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 497                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 498                 }
 499         }
 500
 501         if (behind) {
 502                 if (test_bit(WBCollisionCheck, &rdev->flags)) {
 503                         sector_t lo = r1_bio->sector;
 504                         sector_t hi = r1_bio->sector + r1_bio->sectors;
 505
 506                         remove_wb(rdev, lo, hi);
 507                 }
 508                 if (test_bit(WriteMostly, &rdev->flags))
 509                         atomic_dec(&r1_bio->behind_remaining);
 510
 511                 /*
 512                  * In behind mode, we ACK the master bio once the I/O
 513                  * has safely reached all non-writemostly
 514                  * disks. Setting the Returned bit ensures that this
 515                  * gets done only once -- we don't ever want to return
 516                  * -EIO here, instead we'll wait
 517                  */
 518                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 519                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 520                         /* Maybe we can return now */
 521                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 522                                 struct bio *mbio = r1_bio->master_bio;
 523                                 pr_debug("raid1: behind end write sectors"
 524                                          " %llu-%llu\n",
 525                                          (unsigned long long) mbio->bi_iter.bi_sector,
 526                                          (unsigned long long) bio_end_sector(mbio) - 1);
 527                                 call_bio_endio(r1_bio);
 528                         }
 529                 }
 530         }
 531         if (r1_bio->bios[mirror] == NULL)
 532                 rdev_dec_pending(rdev, conf->mddev);
 533
 534         /*
 535          * Let's see if all mirrored write operations have finished
 536          * already.
 537          */
 538         r1_bio_write_done(r1_bio);
 539
 540         if (to_put)
 541                 bio_put(to_put);
 542 }
 543
 544 static sector_t align_to_barrier_unit_end(sector_t start_sector,
 545                                           sector_t sectors)
 546 {
 547         sector_t len;
 548
 549         WARN_ON(sectors == 0);
 550         /*
 551          * len is the number of sectors from start_sector to end of the
 552          * barrier unit which start_sector belongs to.
 553          */
 554         len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 555               start_sector;
 556
 557         if (len > sectors)
 558                 len = sectors;
 559
 560         return len;
 561 }
 562
 563 /*
 564  * This routine returns the disk from which the requested read should
 565  * be done. There is a per-array 'next expected sequential IO' sector
 566  * number - if this matches on the next IO then we use the last disk.
 567  * There is also a per-disk 'last know head position' sector that is
 568  * maintained from IRQ contexts, both the normal and the resync IO
 569  * completion handlers update this position correctly. If there is no
 570  * perfect sequential match then we pick the disk whose head is closest.
 571  *
 572  * If there are 2 mirrors in the same 2 devices, performance degrades
 573  * because position is mirror, not device based.
 574  *
 575  * The rdev for the device selected will have nr_pending incremented.
 576  */
 577 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 578 {
 579         const sector_t this_sector = r1_bio->sector;
 580         int sectors;
 581         int best_good_sectors;
 582         int best_disk, best_dist_disk, best_pending_disk;
 583         int has_nonrot_disk;
 584         int disk;
 585         sector_t best_dist;
 586         unsigned int min_pending;
 587         struct md_rdev *rdev;
 588         int choose_first;
 589         int choose_next_idle;
 590
 591         rcu_read_lock();
 592         /*
 593          * Check if we can balance. We can balance on the whole
 594          * device if no resync is going on, or below the resync window.
 595          * We take the first readable disk when above the resync window.
 596          */
 597  retry:
 598         sectors = r1_bio->sectors;
 599         best_disk = -1;
 600         best_dist_disk = -1;
 601         best_dist = MaxSector;
 602         best_pending_disk = -1;
 603         min_pending = UINT_MAX;
 604         best_good_sectors = 0;
 605         has_nonrot_disk = 0;
 606         choose_next_idle = 0;
 607         clear_bit(R1BIO_FailFast, &r1_bio->state);
 608
 609         if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 610             (mddev_is_clustered(conf->mddev) &&
 611             md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 612                     this_sector + sectors)))
 613                 choose_first = 1;
 614         else
 615                 choose_first = 0;
 616
 617         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 618                 sector_t dist;
 619                 sector_t first_bad;
 620                 int bad_sectors;
 621                 unsigned int pending;
 622                 bool nonrot;
 623
 624                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 625                 if (r1_bio->bios[disk] == IO_BLOCKED
 626                     || rdev == NULL
 627                     || test_bit(Faulty, &rdev->flags))
 628                         continue;
 629                 if (!test_bit(In_sync, &rdev->flags) &&
 630                     rdev->recovery_offset < this_sector + sectors)
 631                         continue;
 632                 if (test_bit(WriteMostly, &rdev->flags)) {
 633                         /* Don't balance among write-mostly, just
 634                          * use the first as a last resort */
 635                         if (best_dist_disk < 0) {
 636                                 if (is_badblock(rdev, this_sector, sectors,
 637                                                 &first_bad, &bad_sectors)) {
 638                                         if (first_bad <= this_sector)
 639                                                 /* Cannot use this */
 640                                                 continue;
 641                                         best_good_sectors = first_bad - this_sector;
 642                                 } else
 643                                         best_good_sectors = sectors;
 644                                 best_dist_disk = disk;
 645                                 best_pending_disk = disk;
 646                         }
 647                         continue;
 648                 }
 649                 /* This is a reasonable device to use.  It might
 650                  * even be best.
 651                  */
 652                 if (is_badblock(rdev, this_sector, sectors,
 653                                 &first_bad, &bad_sectors)) {
 654                         if (best_dist < MaxSector)
 655                                 /* already have a better device */
 656                                 continue;
 657                         if (first_bad <= this_sector) {
 658                                 /* cannot read here. If this is the 'primary'
 659                                  * device, then we must not read beyond
 660                                  * bad_sectors from another device..
 661                                  */
 662                                 bad_sectors -= (this_sector - first_bad);
 663                                 if (choose_first && sectors > bad_sectors)
 664                                         sectors = bad_sectors;
 665                                 if (best_good_sectors > sectors)
 666                                         best_good_sectors = sectors;
 667
 668                         } else {
 669                                 sector_t good_sectors = first_bad - this_sector;
 670                                 if (good_sectors > best_good_sectors) {
 671                                         best_good_sectors = good_sectors;
 672                                         best_disk = disk;
 673                                 }
 674                                 if (choose_first)
 675                                         break;
 676                         }
 677                         continue;
 678                 } else {
 679                         if ((sectors > best_good_sectors) && (best_disk >= 0))
 680                                 best_disk = -1;
 681                         best_good_sectors = sectors;
 682                 }
 683
 684                 if (best_disk >= 0)
 685                         /* At least two disks to choose from so failfast is OK */
 686                         set_bit(R1BIO_FailFast, &r1_bio->state);
 687
 688                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 689                 has_nonrot_disk |= nonrot;
 690                 pending = atomic_read(&rdev->nr_pending);
 691                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 692                 if (choose_first) {
 693                         best_disk = disk;
 694                         break;
 695                 }
 696                 /* Don't change to another disk for sequential reads */
 697                 if (conf->mirrors[disk].next_seq_sect == this_sector
 698                     || dist == 0) {
 699                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 700                         struct raid1_info *mirror = &conf->mirrors[disk];
 701
 702                         best_disk = disk;
 703                         /*
 704                          * If buffered sequential IO size exceeds optimal
 705                          * iosize, check if there is idle disk. If yes, choose
 706                          * the idle disk. read_balance could already choose an
 707                          * idle disk before noticing it's a sequential IO in
 708                          * this disk. This doesn't matter because this disk
 709                          * will idle, next time it will be utilized after the
 710                          * first disk has IO size exceeds optimal iosize. In
 711                          * this way, iosize of the first disk will be optimal
 712                          * iosize at least. iosize of the second disk might be
 713                          * small, but not a big deal since when the second disk
 714                          * starts IO, the first disk is likely still busy.
 715                          */
 716                         if (nonrot && opt_iosize > 0 &&
 717                             mirror->seq_start != MaxSector &&
 718                             mirror->next_seq_sect > opt_iosize &&
 719                             mirror->next_seq_sect - opt_iosize >=
 720                             mirror->seq_start) {
 721                                 choose_next_idle = 1;
 722                                 continue;
 723                         }
 724                         break;
 725                 }
 726
 727                 if (choose_next_idle)
 728                         continue;
 729
 730                 if (min_pending > pending) {
 731                         min_pending = pending;
 732                         best_pending_disk = disk;
 733                 }
 734
 735                 if (dist < best_dist) {
 736                         best_dist = dist;
 737                         best_dist_disk = disk;
 738                 }
 739         }
 740
 741         /*
 742          * If all disks are rotational, choose the closest disk. If any disk is
 743          * non-rotational, choose the disk with less pending request even the
 744          * disk is rotational, which might/might not be optimal for raids with
 745          * mixed ratation/non-rotational disks depending on workload.
 746          */
 747         if (best_disk == -1) {
 748                 if (has_nonrot_disk || min_pending == 0)
 749                         best_disk = best_pending_disk;
 750                 else
 751                         best_disk = best_dist_disk;
 752         }
 753
 754         if (best_disk >= 0) {
 755                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 756                 if (!rdev)
 757                         goto retry;
 758                 atomic_inc(&rdev->nr_pending);
 759                 sectors = best_good_sectors;
 760
 761                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 762                         conf->mirrors[best_disk].seq_start = this_sector;
 763
 764                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 765         }
 766         rcu_read_unlock();
 767         *max_sectors = sectors;
 768
 769         return best_disk;
 770 }
 771
 772 static int raid1_congested(struct mddev *mddev, int bits)
 773 {
 774         struct r1conf *conf = mddev->private;
 775         int i, ret = 0;
 776
 777         if ((bits & (1 << WB_async_congested)) &&
 778             conf->pending_count >= max_queued_requests)
 779                 return 1;
 780
 781         rcu_read_lock();
 782         for (i = 0; i < conf->raid_disks * 2; i++) {
 783                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 784                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 785                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 786
 787                         BUG_ON(!q);
 788
 789                         /* Note the '|| 1' - when read_balance prefers
 790                          * non-congested targets, it can be removed
 791                          */
 792                         if ((bits & (1 << WB_async_congested)) || 1)
 793                                 ret |= bdi_congested(q->backing_dev_info, bits);
 794                         else
 795                                 ret &= bdi_congested(q->backing_dev_info, bits);
 796                 }
 797         }
 798         rcu_read_unlock();
 799         return ret;
 800 }
 801
 802 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 803 {
 804         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
 805         md_bitmap_unplug(conf->mddev->bitmap);
 806         wake_up(&conf->wait_barrier);
 807
 808         while (bio) { /* submit pending writes */
 809                 struct bio *next = bio->bi_next;
 810                 struct md_rdev *rdev = (void *)bio->bi_disk;
 811                 bio->bi_next = NULL;
 812                 bio_set_dev(bio, rdev->bdev);
 813                 if (test_bit(Faulty, &rdev->flags)) {
 814                         bio_io_error(bio);
 815                 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
 816                                     !blk_queue_discard(bio->bi_disk->queue)))
 817                         /* Just ignore it */
 818                         bio_endio(bio);
 819                 else
 820                         generic_make_request(bio);
 821                 bio = next;
 822         }
 823 }
 824
 825 static void flush_pending_writes(struct r1conf *conf)
 826 {
 827         /* Any writes that have been queued but are awaiting
 828          * bitmap updates get flushed here.
 829          */
 830         spin_lock_irq(&conf->device_lock);
 831
 832         if (conf->pending_bio_list.head) {
 833                 struct blk_plug plug;
 834                 struct bio *bio;
 835
 836                 bio = bio_list_get(&conf->pending_bio_list);
 837                 conf->pending_count = 0;
 838                 spin_unlock_irq(&conf->device_lock);
 839
 840                 /*
 841                  * As this is called in a wait_event() loop (see freeze_array),
 842                  * current->state might be TASK_UNINTERRUPTIBLE which will
 843                  * cause a warning when we prepare to wait again.  As it is
 844                  * rare that this path is taken, it is perfectly safe to force
 845                  * us to go around the wait_event() loop again, so the warning
 846                  * is a false-positive.  Silence the warning by resetting
 847                  * thread state
 848                  */
 849                 __set_current_state(TASK_RUNNING);
 850                 blk_start_plug(&plug);
 851                 flush_bio_list(conf, bio);
 852                 blk_finish_plug(&plug);
 853         } else
 854                 spin_unlock_irq(&conf->device_lock);
 855 }
 856
 857 /* Barriers....
 858  * Sometimes we need to suspend IO while we do something else,
 859  * either some resync/recovery, or reconfigure the array.
 860  * To do this we raise a 'barrier'.
 861  * The 'barrier' is a counter that can be raised multiple times
 862  * to count how many activities are happening which preclude
 863  * normal IO.
 864  * We can only raise the barrier if there is no pending IO.
 865  * i.e. if nr_pending == 0.
 866  * We choose only to raise the barrier if no-one is waiting for the
 867  * barrier to go down.  This means that as soon as an IO request
 868  * is ready, no other operations which require a barrier will start
 869  * until the IO request has had a chance.
 870  *
 871  * So: regular IO calls 'wait_barrier'.  When that returns there
 872  *    is no backgroup IO happening,  It must arrange to call
 873  *    allow_barrier when it has finished its IO.
 874  * backgroup IO calls must call raise_barrier.  Once that returns
 875  *    there is no normal IO happeing.  It must arrange to call
 876  *    lower_barrier when the particular background IO completes.
 877  */
 878 static sector_t raise_barrier(struct r1conf *conf, sector_t sector_nr)
 879 {
 880         int idx = sector_to_idx(sector_nr);
 881
 882         spin_lock_irq(&conf->resync_lock);
 883
 884         /* Wait until no block IO is waiting */
 885         wait_event_lock_irq(conf->wait_barrier,
 886                             !atomic_read(&conf->nr_waiting[idx]),
 887                             conf->resync_lock);
 888
 889         /* block any new IO from starting */
 890         atomic_inc(&conf->barrier[idx]);
 891         /*
 892          * In raise_barrier() we firstly increase conf->barrier[idx] then
 893          * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 894          * increase conf->nr_pending[idx] then check conf->barrier[idx].
 895          * A memory barrier here to make sure conf->nr_pending[idx] won't
 896          * be fetched before conf->barrier[idx] is increased. Otherwise
 897          * there will be a race between raise_barrier() and _wait_barrier().
 898          */
 899         smp_mb__after_atomic();
 900
 901         /* For these conditions we must wait:
 902          * A: while the array is in frozen state
 903          * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 904          *    existing in corresponding I/O barrier bucket.
 905          * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 906          *    max resync count which allowed on current I/O barrier bucket.
 907          */
 908         wait_event_lock_irq(conf->wait_barrier,
 909                             (!conf->array_frozen &&
 910                              !atomic_read(&conf->nr_pending[idx]) &&
 911                              atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
 912                                 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
 913                             conf->resync_lock);
 914
 915         if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
 916                 atomic_dec(&conf->barrier[idx]);
 917                 spin_unlock_irq(&conf->resync_lock);
 918                 wake_up(&conf->wait_barrier);
 919                 return -EINTR;
 920         }
 921
 922         atomic_inc(&conf->nr_sync_pending);
 923         spin_unlock_irq(&conf->resync_lock);
 924
 925         return 0;
 926 }
 927
 928 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 929 {
 930         int idx = sector_to_idx(sector_nr);
 931
 932         BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 933
 934         atomic_dec(&conf->barrier[idx]);
 935         atomic_dec(&conf->nr_sync_pending);
 936         wake_up(&conf->wait_barrier);
 937 }
 938
 939 static void _wait_barrier(struct r1conf *conf, int idx)
 940 {
 941         /*
 942          * We need to increase conf->nr_pending[idx] very early here,
 943          * then raise_barrier() can be blocked when it waits for
 944          * conf->nr_pending[idx] to be 0. Then we can avoid holding
 945          * conf->resync_lock when there is no barrier raised in same
 946          * barrier unit bucket. Also if the array is frozen, I/O
 947          * should be blocked until array is unfrozen.
 948          */
 949         atomic_inc(&conf->nr_pending[idx]);
 950         /*
 951          * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
 952          * check conf->barrier[idx]. In raise_barrier() we firstly increase
 953          * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
 954          * barrier is necessary here to make sure conf->barrier[idx] won't be
 955          * fetched before conf->nr_pending[idx] is increased. Otherwise there
 956          * will be a race between _wait_barrier() and raise_barrier().
 957          */
 958         smp_mb__after_atomic();
 959
 960         /*
 961          * Don't worry about checking two atomic_t variables at same time
 962          * here. If during we check conf->barrier[idx], the array is
 963          * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
 964          * 0, it is safe to return and make the I/O continue. Because the
 965          * array is frozen, all I/O returned here will eventually complete
 966          * or be queued, no race will happen. See code comment in
 967          * frozen_array().
 968          */
 969         if (!READ_ONCE(conf->array_frozen) &&
 970             !atomic_read(&conf->barrier[idx]))
 971                 return;
 972
 973         /*
 974          * After holding conf->resync_lock, conf->nr_pending[idx]
 975          * should be decreased before waiting for barrier to drop.
 976          * Otherwise, we may encounter a race condition because
 977          * raise_barrer() might be waiting for conf->nr_pending[idx]
 978          * to be 0 at same time.
 979          */
 980         spin_lock_irq(&conf->resync_lock);
 981         atomic_inc(&conf->nr_waiting[idx]);
 982         atomic_dec(&conf->nr_pending[idx]);
 983         /*
 984          * In case freeze_array() is waiting for
 985          * get_unqueued_pending() == extra
 986          */
 987         wake_up(&conf->wait_barrier);
 988         /* Wait for the barrier in same barrier unit bucket to drop. */
 989         wait_event_lock_irq(conf->wait_barrier,
 990                             !conf->array_frozen &&
 991                              !atomic_read(&conf->barrier[idx]),
 992                             conf->resync_lock);
 993         atomic_inc(&conf->nr_pending[idx]);
 994         atomic_dec(&conf->nr_waiting[idx]);
 995         spin_unlock_irq(&conf->resync_lock);
 996 }
 997
 998 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
 999 {
1000         int idx = sector_to_idx(sector_nr);
1001
1002         /*
1003          * Very similar to _wait_barrier(). The difference is, for read
1004          * I/O we don't need wait for sync I/O, but if the whole array
1005          * is frozen, the read I/O still has to wait until the array is
1006          * unfrozen. Since there is no ordering requirement with
1007          * conf->barrier[idx] here, memory barrier is unnecessary as well.
1008          */
1009         atomic_inc(&conf->nr_pending[idx]);
1010
1011         if (!READ_ONCE(conf->array_frozen))
1012                 return;
1013
1014         spin_lock_irq(&conf->resync_lock);
1015         atomic_inc(&conf->nr_waiting[idx]);
1016         atomic_dec(&conf->nr_pending[idx]);
1017         /*
1018          * In case freeze_array() is waiting for
1019          * get_unqueued_pending() == extra
1020          */
1021         wake_up(&conf->wait_barrier);
1022         /* Wait for array to be unfrozen */
1023         wait_event_lock_irq(conf->wait_barrier,
1024                             !conf->array_frozen,
1025                             conf->resync_lock);
1026         atomic_inc(&conf->nr_pending[idx]);
1027         atomic_dec(&conf->nr_waiting[idx]);
1028         spin_unlock_irq(&conf->resync_lock);
1029 }
1030
1031 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1032 {
1033         int idx = sector_to_idx(sector_nr);
1034
1035         _wait_barrier(conf, idx);
1036 }
1037
1038 static void _allow_barrier(struct r1conf *conf, int idx)
1039 {
1040         atomic_dec(&conf->nr_pending[idx]);
1041         wake_up(&conf->wait_barrier);
1042 }
1043
1044 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1045 {
1046         int idx = sector_to_idx(sector_nr);
1047
1048         _allow_barrier(conf, idx);
1049 }
1050
1051 /* conf->resync_lock should be held */
1052 static int get_unqueued_pending(struct r1conf *conf)
1053 {
1054         int idx, ret;
1055
1056         ret = atomic_read(&conf->nr_sync_pending);
1057         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1058                 ret += atomic_read(&conf->nr_pending[idx]) -
1059                         atomic_read(&conf->nr_queued[idx]);
1060
1061         return ret;
1062 }
1063
1064 static void freeze_array(struct r1conf *conf, int extra)
1065 {
1066         /* Stop sync I/O and normal I/O and wait for everything to
1067          * go quiet.
1068          * This is called in two situations:
1069          * 1) management command handlers (reshape, remove disk, quiesce).
1070          * 2) one normal I/O request failed.
1071
1072          * After array_frozen is set to 1, new sync IO will be blocked at
1073          * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1074          * or wait_read_barrier(). The flying I/Os will either complete or be
1075          * queued. When everything goes quite, there are only queued I/Os left.
1076
1077          * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1078          * barrier bucket index which this I/O request hits. When all sync and
1079          * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1080          * of all conf->nr_queued[]. But normal I/O failure is an exception,
1081          * in handle_read_error(), we may call freeze_array() before trying to
1082          * fix the read error. In this case, the error read I/O is not queued,
1083          * so get_unqueued_pending() == 1.
1084          *
1085          * Therefore before this function returns, we need to wait until
1086          * get_unqueued_pendings(conf) gets equal to extra. For
1087          * normal I/O context, extra is 1, in rested situations extra is 0.
1088          */
1089         spin_lock_irq(&conf->resync_lock);
1090         conf->array_frozen = 1;
1091         raid1_log(conf->mddev, "wait freeze");
1092         wait_event_lock_irq_cmd(
1093                 conf->wait_barrier,
1094                 get_unqueued_pending(conf) == extra,
1095                 conf->resync_lock,
1096                 flush_pending_writes(conf));
1097         spin_unlock_irq(&conf->resync_lock);
1098 }
1099 static void unfreeze_array(struct r1conf *conf)
1100 {
1101         /* reverse the effect of the freeze */
1102         spin_lock_irq(&conf->resync_lock);
1103         conf->array_frozen = 0;
1104         spin_unlock_irq(&conf->resync_lock);
1105         wake_up(&conf->wait_barrier);
1106 }
1107
1108 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1109                                            struct bio *bio)
1110 {
1111         int size = bio->bi_iter.bi_size;
1112         unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1113         int i = 0;
1114         struct bio *behind_bio = NULL;
1115
1116         behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1117         if (!behind_bio)
1118                 return;
1119
1120         /* discard op, we don't support writezero/writesame yet */
1121         if (!bio_has_data(bio)) {
1122                 behind_bio->bi_iter.bi_size = size;
1123                 goto skip_copy;
1124         }
1125
1126         behind_bio->bi_write_hint = bio->bi_write_hint;
1127
1128         while (i < vcnt && size) {
1129                 struct page *page;
1130                 int len = min_t(int, PAGE_SIZE, size);
1131
1132                 page = alloc_page(GFP_NOIO);
1133                 if (unlikely(!page))
1134                         goto free_pages;
1135
1136                 bio_add_page(behind_bio, page, len, 0);
1137
1138                 size -= len;
1139                 i++;
1140         }
1141
1142         bio_copy_data(behind_bio, bio);
1143 skip_copy:
1144         r1_bio->behind_master_bio = behind_bio;
1145         set_bit(R1BIO_BehindIO, &r1_bio->state);
1146
1147         return;
1148
1149 free_pages:
1150         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1151                  bio->bi_iter.bi_size);
1152         bio_free_pages(behind_bio);
1153         bio_put(behind_bio);
1154 }
1155
1156 struct raid1_plug_cb {
1157         struct blk_plug_cb      cb;
1158         struct bio_list         pending;
1159         int                     pending_cnt;
1160 };
1161
1162 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1163 {
1164         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1165                                                   cb);
1166         struct mddev *mddev = plug->cb.data;
1167         struct r1conf *conf = mddev->private;
1168         struct bio *bio;
1169
1170         if (from_schedule || current->bio_list) {
1171                 spin_lock_irq(&conf->device_lock);
1172                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1173                 conf->pending_count += plug->pending_cnt;
1174                 spin_unlock_irq(&conf->device_lock);
1175                 wake_up(&conf->wait_barrier);
1176                 md_wakeup_thread(mddev->thread);
1177                 kfree(plug);
1178                 return;
1179         }
1180
1181         /* we aren't scheduling, so we can do the write-out directly. */
1182         bio = bio_list_get(&plug->pending);
1183         flush_bio_list(conf, bio);
1184         kfree(plug);
1185 }
1186
1187 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1188 {
1189         r1_bio->master_bio = bio;
1190         r1_bio->sectors = bio_sectors(bio);
1191         r1_bio->state = 0;
1192         r1_bio->mddev = mddev;
1193         r1_bio->sector = bio->bi_iter.bi_sector;
1194 }
1195
1196 static inline struct r1bio *
1197 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1198 {
1199         struct r1conf *conf = mddev->private;
1200         struct r1bio *r1_bio;
1201
1202         r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1203         /* Ensure no bio records IO_BLOCKED */
1204         memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1205         init_r1bio(r1_bio, mddev, bio);
1206         return r1_bio;
1207 }
1208
1209 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1210                                int max_read_sectors, struct r1bio *r1_bio)
1211 {
1212         struct r1conf *conf = mddev->private;
1213         struct raid1_info *mirror;
1214         struct bio *read_bio;
1215         struct bitmap *bitmap = mddev->bitmap;
1216         const int op = bio_op(bio);
1217         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1218         int max_sectors;
1219         int rdisk;
1220         bool print_msg = !!r1_bio;
1221         char b[BDEVNAME_SIZE];
1222
1223         /*
1224          * If r1_bio is set, we are blocking the raid1d thread
1225          * so there is a tiny risk of deadlock.  So ask for
1226          * emergency memory if needed.
1227          */
1228         gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1229
1230         if (print_msg) {
1231                 /* Need to get the block device name carefully */
1232                 struct md_rdev *rdev;
1233                 rcu_read_lock();
1234                 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1235                 if (rdev)
1236                         bdevname(rdev->bdev, b);
1237                 else
1238                         strcpy(b, "???");
1239                 rcu_read_unlock();
1240         }
1241
1242         /*
1243          * Still need barrier for READ in case that whole
1244          * array is frozen.
1245          */
1246         wait_read_barrier(conf, bio->bi_iter.bi_sector);
1247
1248         if (!r1_bio)
1249                 r1_bio = alloc_r1bio(mddev, bio);
1250         else
1251                 init_r1bio(r1_bio, mddev, bio);
1252         r1_bio->sectors = max_read_sectors;
1253
1254         /*
1255          * make_request() can abort the operation when read-ahead is being
1256          * used and no empty request is available.
1257          */
1258         rdisk = read_balance(conf, r1_bio, &max_sectors);
1259
1260         if (rdisk < 0) {
1261                 /* couldn't find anywhere to read from */
1262                 if (print_msg) {
1263                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1264                                             mdname(mddev),
1265                                             b,
1266                                             (unsigned long long)r1_bio->sector);
1267                 }
1268                 raid_end_bio_io(r1_bio);
1269                 return;
1270         }
1271         mirror = conf->mirrors + rdisk;
1272
1273         if (print_msg)
1274                 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1275                                     mdname(mddev),
1276                                     (unsigned long long)r1_bio->sector,
1277                                     bdevname(mirror->rdev->bdev, b));
1278
1279         if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1280             bitmap) {
1281                 /*
1282                  * Reading from a write-mostly device must take care not to
1283                  * over-take any writes that are 'behind'
1284                  */
1285                 raid1_log(mddev, "wait behind writes");
1286                 wait_event(bitmap->behind_wait,
1287                            atomic_read(&bitmap->behind_writes) == 0);
1288         }
1289
1290         if (max_sectors < bio_sectors(bio)) {
1291                 struct bio *split = bio_split(bio, max_sectors,
1292                                               gfp, &conf->bio_split);
1293                 bio_chain(split, bio);
1294                 generic_make_request(bio);
1295                 bio = split;
1296                 r1_bio->master_bio = bio;
1297                 r1_bio->sectors = max_sectors;
1298         }
1299
1300         r1_bio->read_disk = rdisk;
1301
1302         read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1303
1304         r1_bio->bios[rdisk] = read_bio;
1305
1306         read_bio->bi_iter.bi_sector = r1_bio->sector +
1307                 mirror->rdev->data_offset;
1308         bio_set_dev(read_bio, mirror->rdev->bdev);
1309         read_bio->bi_end_io = raid1_end_read_request;
1310         bio_set_op_attrs(read_bio, op, do_sync);
1311         if (test_bit(FailFast, &mirror->rdev->flags) &&
1312             test_bit(R1BIO_FailFast, &r1_bio->state))
1313                 read_bio->bi_opf |= MD_FAILFAST;
1314         read_bio->bi_private = r1_bio;
1315
1316         if (mddev->gendisk)
1317                 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1318                                 disk_devt(mddev->gendisk), r1_bio->sector);
1319
1320         generic_make_request(read_bio);
1321 }
1322
1323 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1324                                 int max_write_sectors)
1325 {
1326         struct r1conf *conf = mddev->private;
1327         struct r1bio *r1_bio;
1328         int i, disks;
1329         struct bitmap *bitmap = mddev->bitmap;
1330         unsigned long flags;
1331         struct md_rdev *blocked_rdev;
1332         struct blk_plug_cb *cb;
1333         struct raid1_plug_cb *plug = NULL;
1334         int first_clone;
1335         int max_sectors;
1336
1337         if (mddev_is_clustered(mddev) &&
1338              md_cluster_ops->area_resyncing(mddev, WRITE,
1339                      bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1340
1341                 DEFINE_WAIT(w);
1342                 for (;;) {
1343                         prepare_to_wait(&conf->wait_barrier,
1344                                         &w, TASK_IDLE);
1345                         if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1346                                                         bio->bi_iter.bi_sector,
1347                                                         bio_end_sector(bio)))
1348                                 break;
1349                         schedule();
1350                 }
1351                 finish_wait(&conf->wait_barrier, &w);
1352         }
1353
1354         /*
1355          * Register the new request and wait if the reconstruction
1356          * thread has put up a bar for new requests.
1357          * Continue immediately if no resync is active currently.
1358          */
1359         wait_barrier(conf, bio->bi_iter.bi_sector);
1360
1361         r1_bio = alloc_r1bio(mddev, bio);
1362         r1_bio->sectors = max_write_sectors;
1363
1364         if (conf->pending_count >= max_queued_requests) {
1365                 md_wakeup_thread(mddev->thread);
1366                 raid1_log(mddev, "wait queued");
1367                 wait_event(conf->wait_barrier,
1368                            conf->pending_count < max_queued_requests);
1369         }
1370         /* first select target devices under rcu_lock and
1371          * inc refcount on their rdev.  Record them by setting
1372          * bios[x] to bio
1373          * If there are known/acknowledged bad blocks on any device on
1374          * which we have seen a write error, we want to avoid writing those
1375          * blocks.
1376          * This potentially requires several writes to write around
1377          * the bad blocks.  Each set of writes gets it's own r1bio
1378          * with a set of bios attached.
1379          */
1380
1381         disks = conf->raid_disks * 2;
1382  retry_write:
1383         blocked_rdev = NULL;
1384         rcu_read_lock();
1385         max_sectors = r1_bio->sectors;
1386         for (i = 0;  i < disks; i++) {
1387                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1388                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1389                         atomic_inc(&rdev->nr_pending);
1390                         blocked_rdev = rdev;
1391                         break;
1392                 }
1393                 r1_bio->bios[i] = NULL;
1394                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1395                         if (i < conf->raid_disks)
1396                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1397                         continue;
1398                 }
1399
1400                 atomic_inc(&rdev->nr_pending);
1401                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1402                         sector_t first_bad;
1403                         int bad_sectors;
1404                         int is_bad;
1405
1406                         is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1407                                              &first_bad, &bad_sectors);
1408                         if (is_bad < 0) {
1409                                 /* mustn't write here until the bad block is
1410                                  * acknowledged*/
1411                                 set_bit(BlockedBadBlocks, &rdev->flags);
1412                                 blocked_rdev = rdev;
1413                                 break;
1414                         }
1415                         if (is_bad && first_bad <= r1_bio->sector) {
1416                                 /* Cannot write here at all */
1417                                 bad_sectors -= (r1_bio->sector - first_bad);
1418                                 if (bad_sectors < max_sectors)
1419                                         /* mustn't write more than bad_sectors
1420                                          * to other devices yet
1421                                          */
1422                                         max_sectors = bad_sectors;
1423                                 rdev_dec_pending(rdev, mddev);
1424                                 /* We don't set R1BIO_Degraded as that
1425                                  * only applies if the disk is
1426                                  * missing, so it might be re-added,
1427                                  * and we want to know to recover this
1428                                  * chunk.
1429                                  * In this case the device is here,
1430                                  * and the fact that this chunk is not
1431                                  * in-sync is recorded in the bad
1432                                  * block log
1433                                  */
1434                                 continue;
1435                         }
1436                         if (is_bad) {
1437                                 int good_sectors = first_bad - r1_bio->sector;
1438                                 if (good_sectors < max_sectors)
1439                                         max_sectors = good_sectors;
1440                         }
1441                 }
1442                 r1_bio->bios[i] = bio;
1443         }
1444         rcu_read_unlock();
1445
1446         if (unlikely(blocked_rdev)) {
1447                 /* Wait for this device to become unblocked */
1448                 int j;
1449
1450                 for (j = 0; j < i; j++)
1451                         if (r1_bio->bios[j])
1452                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1453                 r1_bio->state = 0;
1454                 allow_barrier(conf, bio->bi_iter.bi_sector);
1455                 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1456                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1457                 wait_barrier(conf, bio->bi_iter.bi_sector);
1458                 goto retry_write;
1459         }
1460
1461         if (max_sectors < bio_sectors(bio)) {
1462                 struct bio *split = bio_split(bio, max_sectors,
1463                                               GFP_NOIO, &conf->bio_split);
1464                 bio_chain(split, bio);
1465                 generic_make_request(bio);
1466                 bio = split;
1467                 r1_bio->master_bio = bio;
1468                 r1_bio->sectors = max_sectors;
1469         }
1470
1471         atomic_set(&r1_bio->remaining, 1);
1472         atomic_set(&r1_bio->behind_remaining, 0);
1473
1474         first_clone = 1;
1475
1476         for (i = 0; i < disks; i++) {
1477                 struct bio *mbio = NULL;
1478                 if (!r1_bio->bios[i])
1479                         continue;
1480
1481                 if (first_clone) {
1482                         /* do behind I/O ?
1483                          * Not if there are too many, or cannot
1484                          * allocate memory, or a reader on WriteMostly
1485                          * is waiting for behind writes to flush */
1486                         if (bitmap &&
1487                             (atomic_read(&bitmap->behind_writes)
1488                              < mddev->bitmap_info.max_write_behind) &&
1489                             !waitqueue_active(&bitmap->behind_wait)) {
1490                                 alloc_behind_master_bio(r1_bio, bio);
1491                         }
1492
1493                         md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1494                                              test_bit(R1BIO_BehindIO, &r1_bio->state));
1495                         first_clone = 0;
1496                 }
1497
1498                 if (r1_bio->behind_master_bio)
1499                         mbio = bio_clone_fast(r1_bio->behind_master_bio,
1500                                               GFP_NOIO, &mddev->bio_set);
1501                 else
1502                         mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1503
1504                 if (r1_bio->behind_master_bio) {
1505                         struct md_rdev *rdev = conf->mirrors[i].rdev;
1506
1507                         if (test_bit(WBCollisionCheck, &rdev->flags)) {
1508                                 sector_t lo = r1_bio->sector;
1509                                 sector_t hi = r1_bio->sector + r1_bio->sectors;
1510
1511                                 wait_event(rdev->wb_io_wait,
1512                                            check_and_add_wb(rdev, lo, hi) == 0);
1513                         }
1514                         if (test_bit(WriteMostly, &rdev->flags))
1515                                 atomic_inc(&r1_bio->behind_remaining);
1516                 }
1517
1518                 r1_bio->bios[i] = mbio;
1519
1520                 mbio->bi_iter.bi_sector = (r1_bio->sector +
1521                                    conf->mirrors[i].rdev->data_offset);
1522                 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1523                 mbio->bi_end_io = raid1_end_write_request;
1524                 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1525                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1526                     !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1527                     conf->raid_disks - mddev->degraded > 1)
1528                         mbio->bi_opf |= MD_FAILFAST;
1529                 mbio->bi_private = r1_bio;
1530
1531                 atomic_inc(&r1_bio->remaining);
1532
1533                 if (mddev->gendisk)
1534                         trace_block_bio_remap(mbio->bi_disk->queue,
1535                                               mbio, disk_devt(mddev->gendisk),
1536                                               r1_bio->sector);
1537                 /* flush_pending_writes() needs access to the rdev so...*/
1538                 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1539
1540                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1541                 if (cb)
1542                         plug = container_of(cb, struct raid1_plug_cb, cb);
1543                 else
1544                         plug = NULL;
1545                 if (plug) {
1546                         bio_list_add(&plug->pending, mbio);
1547                         plug->pending_cnt++;
1548                 } else {
1549                         spin_lock_irqsave(&conf->device_lock, flags);
1550                         bio_list_add(&conf->pending_bio_list, mbio);
1551                         conf->pending_count++;
1552                         spin_unlock_irqrestore(&conf->device_lock, flags);
1553                         md_wakeup_thread(mddev->thread);
1554                 }
1555         }
1556
1557         r1_bio_write_done(r1_bio);
1558
1559         /* In case raid1d snuck in to freeze_array */
1560         wake_up(&conf->wait_barrier);
1561 }
1562
1563 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1564 {
1565         sector_t sectors;
1566
1567         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1568                 md_flush_request(mddev, bio);
1569                 return true;
1570         }
1571
1572         /*
1573          * There is a limit to the maximum size, but
1574          * the read/write handler might find a lower limit
1575          * due to bad blocks.  To avoid multiple splits,
1576          * we pass the maximum number of sectors down
1577          * and let the lower level perform the split.
1578          */
1579         sectors = align_to_barrier_unit_end(
1580                 bio->bi_iter.bi_sector, bio_sectors(bio));
1581
1582         if (bio_data_dir(bio) == READ)
1583                 raid1_read_request(mddev, bio, sectors, NULL);
1584         else {
1585                 if (!md_write_start(mddev,bio))
1586                         return false;
1587                 raid1_write_request(mddev, bio, sectors);
1588         }
1589         return true;
1590 }
1591
1592 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1593 {
1594         struct r1conf *conf = mddev->private;
1595         int i;
1596
1597         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1598                    conf->raid_disks - mddev->degraded);
1599         rcu_read_lock();
1600         for (i = 0; i < conf->raid_disks; i++) {
1601                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1602                 seq_printf(seq, "%s",
1603                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1604         }
1605         rcu_read_unlock();
1606         seq_printf(seq, "]");
1607 }
1608
1609 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1610 {
1611         char b[BDEVNAME_SIZE];
1612         struct r1conf *conf = mddev->private;
1613         unsigned long flags;
1614
1615         /*
1616          * If it is not operational, then we have already marked it as dead
1617          * else if it is the last working disks, ignore the error, let the
1618          * next level up know.
1619          * else mark the drive as failed
1620          */
1621         spin_lock_irqsave(&conf->device_lock, flags);
1622         if (test_bit(In_sync, &rdev->flags)
1623             && (conf->raid_disks - mddev->degraded) == 1) {
1624                 /*
1625                  * Don't fail the drive, act as though we were just a
1626                  * normal single drive.
1627                  * However don't try a recovery from this drive as
1628                  * it is very likely to fail.
1629                  */
1630                 conf->recovery_disabled = mddev->recovery_disabled;
1631                 spin_unlock_irqrestore(&conf->device_lock, flags);
1632                 return;
1633         }
1634         set_bit(Blocked, &rdev->flags);
1635         if (test_and_clear_bit(In_sync, &rdev->flags))
1636                 mddev->degraded++;
1637         set_bit(Faulty, &rdev->flags);
1638         spin_unlock_irqrestore(&conf->device_lock, flags);
1639         /*
1640          * if recovery is running, make sure it aborts.
1641          */
1642         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1643         set_mask_bits(&mddev->sb_flags, 0,
1644                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1645         pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1646                 "md/raid1:%s: Operation continuing on %d devices.\n",
1647                 mdname(mddev), bdevname(rdev->bdev, b),
1648                 mdname(mddev), conf->raid_disks - mddev->degraded);
1649 }
1650
1651 static void print_conf(struct r1conf *conf)
1652 {
1653         int i;
1654
1655         pr_debug("RAID1 conf printout:\n");
1656         if (!conf) {
1657                 pr_debug("(!conf)\n");
1658                 return;
1659         }
1660         pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1661                  conf->raid_disks);
1662
1663         rcu_read_lock();
1664         for (i = 0; i < conf->raid_disks; i++) {
1665                 char b[BDEVNAME_SIZE];
1666                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1667                 if (rdev)
1668                         pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1669                                  i, !test_bit(In_sync, &rdev->flags),
1670                                  !test_bit(Faulty, &rdev->flags),
1671                                  bdevname(rdev->bdev,b));
1672         }
1673         rcu_read_unlock();
1674 }
1675
1676 static void close_sync(struct r1conf *conf)
1677 {
1678         int idx;
1679
1680         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1681                 _wait_barrier(conf, idx);
1682                 _allow_barrier(conf, idx);
1683         }
1684
1685         mempool_exit(&conf->r1buf_pool);
1686 }
1687
1688 static int raid1_spare_active(struct mddev *mddev)
1689 {
1690         int i;
1691         struct r1conf *conf = mddev->private;
1692         int count = 0;
1693         unsigned long flags;
1694
1695         /*
1696          * Find all failed disks within the RAID1 configuration
1697          * and mark them readable.
1698          * Called under mddev lock, so rcu protection not needed.
1699          * device_lock used to avoid races with raid1_end_read_request
1700          * which expects 'In_sync' flags and ->degraded to be consistent.
1701          */
1702         spin_lock_irqsave(&conf->device_lock, flags);
1703         for (i = 0; i < conf->raid_disks; i++) {
1704                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1705                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1706                 if (repl
1707                     && !test_bit(Candidate, &repl->flags)
1708                     && repl->recovery_offset == MaxSector
1709                     && !test_bit(Faulty, &repl->flags)
1710                     && !test_and_set_bit(In_sync, &repl->flags)) {
1711                         /* replacement has just become active */
1712                         if (!rdev ||
1713                             !test_and_clear_bit(In_sync, &rdev->flags))
1714                                 count++;
1715                         if (rdev) {
1716                                 /* Replaced device not technically
1717                                  * faulty, but we need to be sure
1718                                  * it gets removed and never re-added
1719                                  */
1720                                 set_bit(Faulty, &rdev->flags);
1721                                 sysfs_notify_dirent_safe(
1722                                         rdev->sysfs_state);
1723                         }
1724                 }
1725                 if (rdev
1726                     && rdev->recovery_offset == MaxSector
1727                     && !test_bit(Faulty, &rdev->flags)
1728                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1729                         count++;
1730                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1731                 }
1732         }
1733         mddev->degraded -= count;
1734         spin_unlock_irqrestore(&conf->device_lock, flags);
1735
1736         print_conf(conf);
1737         return count;
1738 }
1739
1740 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1741 {
1742         struct r1conf *conf = mddev->private;
1743         int err = -EEXIST;
1744         int mirror = 0;
1745         struct raid1_info *p;
1746         int first = 0;
1747         int last = conf->raid_disks - 1;
1748
1749         if (mddev->recovery_disabled == conf->recovery_disabled)
1750                 return -EBUSY;
1751
1752         if (md_integrity_add_rdev(rdev, mddev))
1753                 return -ENXIO;
1754
1755         if (rdev->raid_disk >= 0)
1756                 first = last = rdev->raid_disk;
1757
1758         /*
1759          * find the disk ... but prefer rdev->saved_raid_disk
1760          * if possible.
1761          */
1762         if (rdev->saved_raid_disk >= 0 &&
1763             rdev->saved_raid_disk >= first &&
1764             rdev->saved_raid_disk < conf->raid_disks &&
1765             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1766                 first = last = rdev->saved_raid_disk;
1767
1768         for (mirror = first; mirror <= last; mirror++) {
1769                 p = conf->mirrors + mirror;
1770                 if (!p->rdev) {
1771                         if (mddev->gendisk)
1772                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1773                                                   rdev->data_offset << 9);
1774
1775                         p->head_position = 0;
1776                         rdev->raid_disk = mirror;
1777                         err = 0;
1778                         /* As all devices are equivalent, we don't need a full recovery
1779                          * if this was recently any drive of the array
1780                          */
1781                         if (rdev->saved_raid_disk < 0)
1782                                 conf->fullsync = 1;
1783                         rcu_assign_pointer(p->rdev, rdev);
1784                         break;
1785                 }
1786                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1787                     p[conf->raid_disks].rdev == NULL) {
1788                         /* Add this device as a replacement */
1789                         clear_bit(In_sync, &rdev->flags);
1790                         set_bit(Replacement, &rdev->flags);
1791                         rdev->raid_disk = mirror;
1792                         err = 0;
1793                         conf->fullsync = 1;
1794                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1795                         break;
1796                 }
1797         }
1798         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1799                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1800         print_conf(conf);
1801         return err;
1802 }
1803
1804 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1805 {
1806         struct r1conf *conf = mddev->private;
1807         int err = 0;
1808         int number = rdev->raid_disk;
1809         struct raid1_info *p = conf->mirrors + number;
1810
1811         if (rdev != p->rdev)
1812                 p = conf->mirrors + conf->raid_disks + number;
1813
1814         print_conf(conf);
1815         if (rdev == p->rdev) {
1816                 if (test_bit(In_sync, &rdev->flags) ||
1817                     atomic_read(&rdev->nr_pending)) {
1818                         err = -EBUSY;
1819                         goto abort;
1820                 }
1821                 /* Only remove non-faulty devices if recovery
1822                  * is not possible.
1823                  */
1824                 if (!test_bit(Faulty, &rdev->flags) &&
1825                     mddev->recovery_disabled != conf->recovery_disabled &&
1826                     mddev->degraded < conf->raid_disks) {
1827                         err = -EBUSY;
1828                         goto abort;
1829                 }
1830                 p->rdev = NULL;
1831                 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1832                         synchronize_rcu();
1833                         if (atomic_read(&rdev->nr_pending)) {
1834                                 /* lost the race, try later */
1835                                 err = -EBUSY;
1836                                 p->rdev = rdev;
1837                                 goto abort;
1838                         }
1839                 }
1840                 if (conf->mirrors[conf->raid_disks + number].rdev) {
1841                         /* We just removed a device that is being replaced.
1842                          * Move down the replacement.  We drain all IO before
1843                          * doing this to avoid confusion.
1844                          */
1845                         struct md_rdev *repl =
1846                                 conf->mirrors[conf->raid_disks + number].rdev;
1847                         freeze_array(conf, 0);
1848                         if (atomic_read(&repl->nr_pending)) {
1849                                 /* It means that some queued IO of retry_list
1850                                  * hold repl. Thus, we cannot set replacement
1851                                  * as NULL, avoiding rdev NULL pointer
1852                                  * dereference in sync_request_write and
1853                                  * handle_write_finished.
1854                                  */
1855                                 err = -EBUSY;
1856                                 unfreeze_array(conf);
1857                                 goto abort;
1858                         }
1859                         clear_bit(Replacement, &repl->flags);
1860                         p->rdev = repl;
1861                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1862                         unfreeze_array(conf);
1863                 }
1864
1865                 clear_bit(WantReplacement, &rdev->flags);
1866                 err = md_integrity_register(mddev);
1867         }
1868 abort:
1869
1870         print_conf(conf);
1871         return err;
1872 }
1873
1874 static void end_sync_read(struct bio *bio)
1875 {
1876         struct r1bio *r1_bio = get_resync_r1bio(bio);
1877
1878         update_head_pos(r1_bio->read_disk, r1_bio);
1879
1880         /*
1881          * we have read a block, now it needs to be re-written,
1882          * or re-read if the read failed.
1883          * We don't do much here, just schedule handling by raid1d
1884          */
1885         if (!bio->bi_status)
1886                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1887
1888         if (atomic_dec_and_test(&r1_bio->remaining))
1889                 reschedule_retry(r1_bio);
1890 }
1891
1892 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1893 {
1894         sector_t sync_blocks = 0;
1895         sector_t s = r1_bio->sector;
1896         long sectors_to_go = r1_bio->sectors;
1897
1898         /* make sure these bits don't get cleared. */
1899         do {
1900                 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1901                 s += sync_blocks;
1902                 sectors_to_go -= sync_blocks;
1903         } while (sectors_to_go > 0);
1904 }
1905
1906 static void end_sync_write(struct bio *bio)
1907 {
1908         int uptodate = !bio->bi_status;
1909         struct r1bio *r1_bio = get_resync_r1bio(bio);
1910         struct mddev *mddev = r1_bio->mddev;
1911         struct r1conf *conf = mddev->private;
1912         sector_t first_bad;
1913         int bad_sectors;
1914         struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1915
1916         if (!uptodate) {
1917                 abort_sync_write(mddev, r1_bio);
1918                 set_bit(WriteErrorSeen, &rdev->flags);
1919                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1920                         set_bit(MD_RECOVERY_NEEDED, &
1921                                 mddev->recovery);
1922                 set_bit(R1BIO_WriteError, &r1_bio->state);
1923         } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1924                                &first_bad, &bad_sectors) &&
1925                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1926                                 r1_bio->sector,
1927                                 r1_bio->sectors,
1928                                 &first_bad, &bad_sectors)
1929                 )
1930                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1931
1932         if (atomic_dec_and_test(&r1_bio->remaining)) {
1933                 int s = r1_bio->sectors;
1934                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1935                     test_bit(R1BIO_WriteError, &r1_bio->state))
1936                         reschedule_retry(r1_bio);
1937                 else {
1938                         put_buf(r1_bio);
1939                         md_done_sync(mddev, s, uptodate);
1940                 }
1941         }
1942 }
1943
1944 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1945                             int sectors, struct page *page, int rw)
1946 {
1947         if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1948                 /* success */
1949                 return 1;
1950         if (rw == WRITE) {
1951                 set_bit(WriteErrorSeen, &rdev->flags);
1952                 if (!test_and_set_bit(WantReplacement,
1953                                       &rdev->flags))
1954                         set_bit(MD_RECOVERY_NEEDED, &
1955                                 rdev->mddev->recovery);
1956         }
1957         /* need to record an error - either for the block or the device */
1958         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1959                 md_error(rdev->mddev, rdev);
1960         return 0;
1961 }
1962
1963 static int fix_sync_read_error(struct r1bio *r1_bio)
1964 {
1965         /* Try some synchronous reads of other devices to get
1966          * good data, much like with normal read errors.  Only
1967          * read into the pages we already have so we don't
1968          * need to re-issue the read request.
1969          * We don't need to freeze the array, because being in an
1970          * active sync request, there is no normal IO, and
1971          * no overlapping syncs.
1972          * We don't need to check is_badblock() again as we
1973          * made sure that anything with a bad block in range
1974          * will have bi_end_io clear.
1975          */
1976         struct mddev *mddev = r1_bio->mddev;
1977         struct r1conf *conf = mddev->private;
1978         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1979         struct page **pages = get_resync_pages(bio)->pages;
1980         sector_t sect = r1_bio->sector;
1981         int sectors = r1_bio->sectors;
1982         int idx = 0;
1983         struct md_rdev *rdev;
1984
1985         rdev = conf->mirrors[r1_bio->read_disk].rdev;
1986         if (test_bit(FailFast, &rdev->flags)) {
1987                 /* Don't try recovering from here - just fail it
1988                  * ... unless it is the last working device of course */
1989                 md_error(mddev, rdev);
1990                 if (test_bit(Faulty, &rdev->flags))
1991                         /* Don't try to read from here, but make sure
1992                          * put_buf does it's thing
1993                          */
1994                         bio->bi_end_io = end_sync_write;
1995         }
1996
1997         while(sectors) {
1998                 int s = sectors;
1999                 int d = r1_bio->read_disk;
2000                 int success = 0;
2001                 int start;
2002
2003                 if (s > (PAGE_SIZE>>9))
2004                         s = PAGE_SIZE >> 9;
2005                 do {
2006                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2007                                 /* No rcu protection needed here devices
2008                                  * can only be removed when no resync is
2009                                  * active, and resync is currently active
2010                                  */
2011                                 rdev = conf->mirrors[d].rdev;
2012                                 if (sync_page_io(rdev, sect, s<<9,
2013                                                  pages[idx],
2014                                                  REQ_OP_READ, 0, false)) {
2015                                         success = 1;
2016                                         break;
2017                                 }
2018                         }
2019                         d++;
2020                         if (d == conf->raid_disks * 2)
2021                                 d = 0;
2022                 } while (!success && d != r1_bio->read_disk);
2023
2024                 if (!success) {
2025                         char b[BDEVNAME_SIZE];
2026                         int abort = 0;
2027                         /* Cannot read from anywhere, this block is lost.
2028                          * Record a bad block on each device.  If that doesn't
2029                          * work just disable and interrupt the recovery.
2030                          * Don't fail devices as that won't really help.
2031                          */
2032                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2033                                             mdname(mddev), bio_devname(bio, b),
2034                                             (unsigned long long)r1_bio->sector);
2035                         for (d = 0; d < conf->raid_disks * 2; d++) {
2036                                 rdev = conf->mirrors[d].rdev;
2037                                 if (!rdev || test_bit(Faulty, &rdev->flags))
2038                                         continue;
2039                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
2040                                         abort = 1;
2041                         }
2042                         if (abort) {
2043                                 conf->recovery_disabled =
2044                                         mddev->recovery_disabled;
2045                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2046                                 md_done_sync(mddev, r1_bio->sectors, 0);
2047                                 put_buf(r1_bio);
2048                                 return 0;
2049                         }
2050                         /* Try next page */
2051                         sectors -= s;
2052                         sect += s;
2053                         idx++;
2054                         continue;
2055                 }
2056
2057                 start = d;
2058                 /* write it back and re-read */
2059                 while (d != r1_bio->read_disk) {
2060                         if (d == 0)
2061                                 d = conf->raid_disks * 2;
2062                         d--;
2063                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2064                                 continue;
2065                         rdev = conf->mirrors[d].rdev;
2066                         if (r1_sync_page_io(rdev, sect, s,
2067                                             pages[idx],
2068                                             WRITE) == 0) {
2069                                 r1_bio->bios[d]->bi_end_io = NULL;
2070                                 rdev_dec_pending(rdev, mddev);
2071                         }
2072                 }
2073                 d = start;
2074                 while (d != r1_bio->read_disk) {
2075                         if (d == 0)
2076                                 d = conf->raid_disks * 2;
2077                         d--;
2078                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2079                                 continue;
2080                         rdev = conf->mirrors[d].rdev;
2081                         if (r1_sync_page_io(rdev, sect, s,
2082                                             pages[idx],
2083                                             READ) != 0)
2084                                 atomic_add(s, &rdev->corrected_errors);
2085                 }
2086                 sectors -= s;
2087                 sect += s;
2088                 idx ++;
2089         }
2090         set_bit(R1BIO_Uptodate, &r1_bio->state);
2091         bio->bi_status = 0;
2092         return 1;
2093 }
2094
2095 static void process_checks(struct r1bio *r1_bio)
2096 {
2097         /* We have read all readable devices.  If we haven't
2098          * got the block, then there is no hope left.
2099          * If we have, then we want to do a comparison
2100          * and skip the write if everything is the same.
2101          * If any blocks failed to read, then we need to
2102          * attempt an over-write
2103          */
2104         struct mddev *mddev = r1_bio->mddev;
2105         struct r1conf *conf = mddev->private;
2106         int primary;
2107         int i;
2108         int vcnt;
2109
2110         /* Fix variable parts of all bios */
2111         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2112         for (i = 0; i < conf->raid_disks * 2; i++) {
2113                 blk_status_t status;
2114                 struct bio *b = r1_bio->bios[i];
2115                 struct resync_pages *rp = get_resync_pages(b);
2116                 if (b->bi_end_io != end_sync_read)
2117                         continue;
2118                 /* fixup the bio for reuse, but preserve errno */
2119                 status = b->bi_status;
2120                 bio_reset(b);
2121                 b->bi_status = status;
2122                 b->bi_iter.bi_sector = r1_bio->sector +
2123                         conf->mirrors[i].rdev->data_offset;
2124                 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2125                 b->bi_end_io = end_sync_read;
2126                 rp->raid_bio = r1_bio;
2127                 b->bi_private = rp;
2128
2129                 /* initialize bvec table again */
2130                 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2131         }
2132         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2133                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2134                     !r1_bio->bios[primary]->bi_status) {
2135                         r1_bio->bios[primary]->bi_end_io = NULL;
2136                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2137                         break;
2138                 }
2139         r1_bio->read_disk = primary;
2140         for (i = 0; i < conf->raid_disks * 2; i++) {
2141                 int j = 0;
2142                 struct bio *pbio = r1_bio->bios[primary];
2143                 struct bio *sbio = r1_bio->bios[i];
2144                 blk_status_t status = sbio->bi_status;
2145                 struct page **ppages = get_resync_pages(pbio)->pages;
2146                 struct page **spages = get_resync_pages(sbio)->pages;
2147                 struct bio_vec *bi;
2148                 int page_len[RESYNC_PAGES] = { 0 };
2149                 struct bvec_iter_all iter_all;
2150
2151                 if (sbio->bi_end_io != end_sync_read)
2152                         continue;
2153                 /* Now we can 'fixup' the error value */
2154                 sbio->bi_status = 0;
2155
2156                 bio_for_each_segment_all(bi, sbio, iter_all)
2157                         page_len[j++] = bi->bv_len;
2158
2159                 if (!status) {
2160                         for (j = vcnt; j-- ; ) {
2161                                 if (memcmp(page_address(ppages[j]),
2162                                            page_address(spages[j]),
2163                                            page_len[j]))
2164                                         break;
2165                         }
2166                 } else
2167                         j = 0;
2168                 if (j >= 0)
2169                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2170                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2171                               && !status)) {
2172                         /* No need to write to this device. */
2173                         sbio->bi_end_io = NULL;
2174                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2175                         continue;
2176                 }
2177
2178                 bio_copy_data(sbio, pbio);
2179         }
2180 }
2181
2182 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2183 {
2184         struct r1conf *conf = mddev->private;
2185         int i;
2186         int disks = conf->raid_disks * 2;
2187         struct bio *wbio;
2188
2189         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2190                 /* ouch - failed to read all of that. */
2191                 if (!fix_sync_read_error(r1_bio))
2192                         return;
2193
2194         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2195                 process_checks(r1_bio);
2196
2197         /*
2198          * schedule writes
2199          */
2200         atomic_set(&r1_bio->remaining, 1);
2201         for (i = 0; i < disks ; i++) {
2202                 wbio = r1_bio->bios[i];
2203                 if (wbio->bi_end_io == NULL ||
2204                     (wbio->bi_end_io == end_sync_read &&
2205                      (i == r1_bio->read_disk ||
2206                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2207                         continue;
2208                 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2209                         abort_sync_write(mddev, r1_bio);
2210                         continue;
2211                 }
2212
2213                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2214                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2215                         wbio->bi_opf |= MD_FAILFAST;
2216
2217                 wbio->bi_end_io = end_sync_write;
2218                 atomic_inc(&r1_bio->remaining);
2219                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2220
2221                 generic_make_request(wbio);
2222         }
2223
2224         if (atomic_dec_and_test(&r1_bio->remaining)) {
2225                 /* if we're here, all write(s) have completed, so clean up */
2226                 int s = r1_bio->sectors;
2227                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2228                     test_bit(R1BIO_WriteError, &r1_bio->state))
2229                         reschedule_retry(r1_bio);
2230                 else {
2231                         put_buf(r1_bio);
2232                         md_done_sync(mddev, s, 1);
2233                 }
2234         }
2235 }
2236
2237 /*
2238  * This is a kernel thread which:
2239  *
2240  *      1.      Retries failed read operations on working mirrors.
2241  *      2.      Updates the raid superblock when problems encounter.
2242  *      3.      Performs writes following reads for array synchronising.
2243  */
2244
2245 static void fix_read_error(struct r1conf *conf, int read_disk,
2246                            sector_t sect, int sectors)
2247 {
2248         struct mddev *mddev = conf->mddev;
2249         while(sectors) {
2250                 int s = sectors;
2251                 int d = read_disk;
2252                 int success = 0;
2253                 int start;
2254                 struct md_rdev *rdev;
2255
2256                 if (s > (PAGE_SIZE>>9))
2257                         s = PAGE_SIZE >> 9;
2258
2259                 do {
2260                         sector_t first_bad;
2261                         int bad_sectors;
2262
2263                         rcu_read_lock();
2264                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2265                         if (rdev &&
2266                             (test_bit(In_sync, &rdev->flags) ||
2267                              (!test_bit(Faulty, &rdev->flags) &&
2268                               rdev->recovery_offset >= sect + s)) &&
2269                             is_badblock(rdev, sect, s,
2270                                         &first_bad, &bad_sectors) == 0) {
2271                                 atomic_inc(&rdev->nr_pending);
2272                                 rcu_read_unlock();
2273                                 if (sync_page_io(rdev, sect, s<<9,
2274                                          conf->tmppage, REQ_OP_READ, 0, false))
2275                                         success = 1;
2276                                 rdev_dec_pending(rdev, mddev);
2277                                 if (success)
2278                                         break;
2279                         } else
2280                                 rcu_read_unlock();
2281                         d++;
2282                         if (d == conf->raid_disks * 2)
2283                                 d = 0;
2284                 } while (!success && d != read_disk);
2285
2286                 if (!success) {
2287                         /* Cannot read from anywhere - mark it bad */
2288                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2289                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2290                                 md_error(mddev, rdev);
2291                         break;
2292                 }
2293                 /* write it back and re-read */
2294                 start = d;
2295                 while (d != read_disk) {
2296                         if (d==0)
2297                                 d = conf->raid_disks * 2;
2298                         d--;
2299                         rcu_read_lock();
2300                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2301                         if (rdev &&
2302                             !test_bit(Faulty, &rdev->flags)) {
2303                                 atomic_inc(&rdev->nr_pending);
2304                                 rcu_read_unlock();
2305                                 r1_sync_page_io(rdev, sect, s,
2306                                                 conf->tmppage, WRITE);
2307                                 rdev_dec_pending(rdev, mddev);
2308                         } else
2309                                 rcu_read_unlock();
2310                 }
2311                 d = start;
2312                 while (d != read_disk) {
2313                         char b[BDEVNAME_SIZE];
2314                         if (d==0)
2315                                 d = conf->raid_disks * 2;
2316                         d--;
2317                         rcu_read_lock();
2318                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2319                         if (rdev &&
2320                             !test_bit(Faulty, &rdev->flags)) {
2321                                 atomic_inc(&rdev->nr_pending);
2322                                 rcu_read_unlock();
2323                                 if (r1_sync_page_io(rdev, sect, s,
2324                                                     conf->tmppage, READ)) {
2325                                         atomic_add(s, &rdev->corrected_errors);
2326                                         pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2327                                                 mdname(mddev), s,
2328                                                 (unsigned long long)(sect +
2329                                                                      rdev->data_offset),
2330                                                 bdevname(rdev->bdev, b));
2331                                 }
2332                                 rdev_dec_pending(rdev, mddev);
2333                         } else
2334                                 rcu_read_unlock();
2335                 }
2336                 sectors -= s;
2337                 sect += s;
2338         }
2339 }
2340
2341 static int narrow_write_error(struct r1bio *r1_bio, int i)
2342 {
2343         struct mddev *mddev = r1_bio->mddev;
2344         struct r1conf *conf = mddev->private;
2345         struct md_rdev *rdev = conf->mirrors[i].rdev;
2346
2347         /* bio has the data to be written to device 'i' where
2348          * we just recently had a write error.
2349          * We repeatedly clone the bio and trim down to one block,
2350          * then try the write.  Where the write fails we record
2351          * a bad block.
2352          * It is conceivable that the bio doesn't exactly align with
2353          * blocks.  We must handle this somehow.
2354          *
2355          * We currently own a reference on the rdev.
2356          */
2357
2358         int block_sectors;
2359         sector_t sector;
2360         int sectors;
2361         int sect_to_write = r1_bio->sectors;
2362         int ok = 1;
2363
2364         if (rdev->badblocks.shift < 0)
2365                 return 0;
2366
2367         block_sectors = roundup(1 << rdev->badblocks.shift,
2368                                 bdev_logical_block_size(rdev->bdev) >> 9);
2369         sector = r1_bio->sector;
2370         sectors = ((sector + block_sectors)
2371                    & ~(sector_t)(block_sectors - 1))
2372                 - sector;
2373
2374         while (sect_to_write) {
2375                 struct bio *wbio;
2376                 if (sectors > sect_to_write)
2377                         sectors = sect_to_write;
2378                 /* Write at 'sector' for 'sectors'*/
2379
2380                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2381                         wbio = bio_clone_fast(r1_bio->behind_master_bio,
2382                                               GFP_NOIO,
2383                                               &mddev->bio_set);
2384                 } else {
2385                         wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2386                                               &mddev->bio_set);
2387                 }
2388
2389                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2390                 wbio->bi_iter.bi_sector = r1_bio->sector;
2391                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2392
2393                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2394                 wbio->bi_iter.bi_sector += rdev->data_offset;
2395                 bio_set_dev(wbio, rdev->bdev);
2396
2397                 if (submit_bio_wait(wbio) < 0)
2398                         /* failure! */
2399                         ok = rdev_set_badblocks(rdev, sector,
2400                                                 sectors, 0)
2401                                 && ok;
2402
2403                 bio_put(wbio);
2404                 sect_to_write -= sectors;
2405                 sector += sectors;
2406                 sectors = block_sectors;
2407         }
2408         return ok;
2409 }
2410
2411 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2412 {
2413         int m;
2414         int s = r1_bio->sectors;
2415         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2416                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2417                 struct bio *bio = r1_bio->bios[m];
2418                 if (bio->bi_end_io == NULL)
2419                         continue;
2420                 if (!bio->bi_status &&
2421                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2422                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2423                 }
2424                 if (bio->bi_status &&
2425                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2426                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2427                                 md_error(conf->mddev, rdev);
2428                 }
2429         }
2430         put_buf(r1_bio);
2431         md_done_sync(conf->mddev, s, 1);
2432 }
2433
2434 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2435 {
2436         int m, idx;
2437         bool fail = false;
2438
2439         for (m = 0; m < conf->raid_disks * 2 ; m++)
2440                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2441                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2442                         rdev_clear_badblocks(rdev,
2443                                              r1_bio->sector,
2444                                              r1_bio->sectors, 0);
2445                         rdev_dec_pending(rdev, conf->mddev);
2446                 } else if (r1_bio->bios[m] != NULL) {
2447                         /* This drive got a write error.  We need to
2448                          * narrow down and record precise write
2449                          * errors.
2450                          */
2451                         fail = true;
2452                         if (!narrow_write_error(r1_bio, m)) {
2453                                 md_error(conf->mddev,
2454                                          conf->mirrors[m].rdev);
2455                                 /* an I/O failed, we can't clear the bitmap */
2456                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2457                         }
2458                         rdev_dec_pending(conf->mirrors[m].rdev,
2459                                          conf->mddev);
2460                 }
2461         if (fail) {
2462                 spin_lock_irq(&conf->device_lock);
2463                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2464                 idx = sector_to_idx(r1_bio->sector);
2465                 atomic_inc(&conf->nr_queued[idx]);
2466                 spin_unlock_irq(&conf->device_lock);
2467                 /*
2468                  * In case freeze_array() is waiting for condition
2469                  * get_unqueued_pending() == extra to be true.
2470                  */
2471                 wake_up(&conf->wait_barrier);
2472                 md_wakeup_thread(conf->mddev->thread);
2473         } else {
2474                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2475                         close_write(r1_bio);
2476                 raid_end_bio_io(r1_bio);
2477         }
2478 }
2479
2480 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2481 {
2482         struct mddev *mddev = conf->mddev;
2483         struct bio *bio;
2484         struct md_rdev *rdev;
2485
2486         clear_bit(R1BIO_ReadError, &r1_bio->state);
2487         /* we got a read error. Maybe the drive is bad.  Maybe just
2488          * the block and we can fix it.
2489          * We freeze all other IO, and try reading the block from
2490          * other devices.  When we find one, we re-write
2491          * and check it that fixes the read error.
2492          * This is all done synchronously while the array is
2493          * frozen
2494          */
2495
2496         bio = r1_bio->bios[r1_bio->read_disk];
2497         bio_put(bio);
2498         r1_bio->bios[r1_bio->read_disk] = NULL;
2499
2500         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2501         if (mddev->ro == 0
2502             && !test_bit(FailFast, &rdev->flags)) {
2503                 freeze_array(conf, 1);
2504                 fix_read_error(conf, r1_bio->read_disk,
2505                                r1_bio->sector, r1_bio->sectors);
2506                 unfreeze_array(conf);
2507         } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2508                 md_error(mddev, rdev);
2509         } else {
2510                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2511         }
2512
2513         rdev_dec_pending(rdev, conf->mddev);
2514         allow_barrier(conf, r1_bio->sector);
2515         bio = r1_bio->master_bio;
2516
2517         /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2518         r1_bio->state = 0;
2519         raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2520 }
2521
2522 static void raid1d(struct md_thread *thread)
2523 {
2524         struct mddev *mddev = thread->mddev;
2525         struct r1bio *r1_bio;
2526         unsigned long flags;
2527         struct r1conf *conf = mddev->private;
2528         struct list_head *head = &conf->retry_list;
2529         struct blk_plug plug;
2530         int idx;
2531
2532         md_check_recovery(mddev);
2533
2534         if (!list_empty_careful(&conf->bio_end_io_list) &&
2535             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2536                 LIST_HEAD(tmp);
2537                 spin_lock_irqsave(&conf->device_lock, flags);
2538                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2539                         list_splice_init(&conf->bio_end_io_list, &tmp);
2540                 spin_unlock_irqrestore(&conf->device_lock, flags);
2541                 while (!list_empty(&tmp)) {
2542                         r1_bio = list_first_entry(&tmp, struct r1bio,
2543                                                   retry_list);
2544                         list_del(&r1_bio->retry_list);
2545                         idx = sector_to_idx(r1_bio->sector);
2546                         atomic_dec(&conf->nr_queued[idx]);
2547                         if (mddev->degraded)
2548                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2549                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2550                                 close_write(r1_bio);
2551                         raid_end_bio_io(r1_bio);
2552                 }
2553         }
2554
2555         blk_start_plug(&plug);
2556         for (;;) {
2557
2558                 flush_pending_writes(conf);
2559
2560                 spin_lock_irqsave(&conf->device_lock, flags);
2561                 if (list_empty(head)) {
2562                         spin_unlock_irqrestore(&conf->device_lock, flags);
2563                         break;
2564                 }
2565                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2566                 list_del(head->prev);
2567                 idx = sector_to_idx(r1_bio->sector);
2568                 atomic_dec(&conf->nr_queued[idx]);
2569                 spin_unlock_irqrestore(&conf->device_lock, flags);
2570
2571                 mddev = r1_bio->mddev;
2572                 conf = mddev->private;
2573                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2574                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2575                             test_bit(R1BIO_WriteError, &r1_bio->state))
2576                                 handle_sync_write_finished(conf, r1_bio);
2577                         else
2578                                 sync_request_write(mddev, r1_bio);
2579                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2580                            test_bit(R1BIO_WriteError, &r1_bio->state))
2581                         handle_write_finished(conf, r1_bio);
2582                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2583                         handle_read_error(conf, r1_bio);
2584                 else
2585                         WARN_ON_ONCE(1);
2586
2587                 cond_resched();
2588                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2589                         md_check_recovery(mddev);
2590         }
2591         blk_finish_plug(&plug);
2592 }
2593
2594 static int init_resync(struct r1conf *conf)
2595 {
2596         int buffs;
2597
2598         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2599         BUG_ON(mempool_initialized(&conf->r1buf_pool));
2600
2601         return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2602                             r1buf_pool_free, conf->poolinfo);
2603 }
2604
2605 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2606 {
2607         struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2608         struct resync_pages *rps;
2609         struct bio *bio;
2610         int i;
2611
2612         for (i = conf->poolinfo->raid_disks; i--; ) {
2613                 bio = r1bio->bios[i];
2614                 rps = bio->bi_private;
2615                 bio_reset(bio);
2616                 bio->bi_private = rps;
2617         }
2618         r1bio->master_bio = NULL;
2619         return r1bio;
2620 }
2621
2622 /*
2623  * perform a "sync" on one "block"
2624  *
2625  * We need to make sure that no normal I/O request - particularly write
2626  * requests - conflict with active sync requests.
2627  *
2628  * This is achieved by tracking pending requests and a 'barrier' concept
2629  * that can be installed to exclude normal IO requests.
2630  */
2631
2632 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2633                                    int *skipped)
2634 {
2635         struct r1conf *conf = mddev->private;
2636         struct r1bio *r1_bio;
2637         struct bio *bio;
2638         sector_t max_sector, nr_sectors;
2639         int disk = -1;
2640         int i;
2641         int wonly = -1;
2642         int write_targets = 0, read_targets = 0;
2643         sector_t sync_blocks;
2644         int still_degraded = 0;
2645         int good_sectors = RESYNC_SECTORS;
2646         int min_bad = 0; /* number of sectors that are bad in all devices */
2647         int idx = sector_to_idx(sector_nr);
2648         int page_idx = 0;
2649
2650         if (!mempool_initialized(&conf->r1buf_pool))
2651                 if (init_resync(conf))
2652                         return 0;
2653
2654         max_sector = mddev->dev_sectors;
2655         if (sector_nr >= max_sector) {
2656                 /* If we aborted, we need to abort the
2657                  * sync on the 'current' bitmap chunk (there will
2658                  * only be one in raid1 resync.
2659                  * We can find the current addess in mddev->curr_resync
2660                  */
2661                 if (mddev->curr_resync < max_sector) /* aborted */
2662                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2663                                            &sync_blocks, 1);
2664                 else /* completed sync */
2665                         conf->fullsync = 0;
2666
2667                 md_bitmap_close_sync(mddev->bitmap);
2668                 close_sync(conf);
2669
2670                 if (mddev_is_clustered(mddev)) {
2671                         conf->cluster_sync_low = 0;
2672                         conf->cluster_sync_high = 0;
2673                 }
2674                 return 0;
2675         }
2676
2677         if (mddev->bitmap == NULL &&
2678             mddev->recovery_cp == MaxSector &&
2679             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2680             conf->fullsync == 0) {
2681                 *skipped = 1;
2682                 return max_sector - sector_nr;
2683         }
2684         /* before building a request, check if we can skip these blocks..
2685          * This call the bitmap_start_sync doesn't actually record anything
2686          */
2687         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2688             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2689                 /* We can skip this block, and probably several more */
2690                 *skipped = 1;
2691                 return sync_blocks;
2692         }
2693
2694         /*
2695          * If there is non-resync activity waiting for a turn, then let it
2696          * though before starting on this new sync request.
2697          */
2698         if (atomic_read(&conf->nr_waiting[idx]))
2699                 schedule_timeout_uninterruptible(1);
2700
2701         /* we are incrementing sector_nr below. To be safe, we check against
2702          * sector_nr + two times RESYNC_SECTORS
2703          */
2704
2705         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2706                 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2707
2708
2709         if (raise_barrier(conf, sector_nr))
2710                 return 0;
2711
2712         r1_bio = raid1_alloc_init_r1buf(conf);
2713
2714         rcu_read_lock();
2715         /*
2716          * If we get a correctably read error during resync or recovery,
2717          * we might want to read from a different device.  So we
2718          * flag all drives that could conceivably be read from for READ,
2719          * and any others (which will be non-In_sync devices) for WRITE.
2720          * If a read fails, we try reading from something else for which READ
2721          * is OK.
2722          */
2723
2724         r1_bio->mddev = mddev;
2725         r1_bio->sector = sector_nr;
2726         r1_bio->state = 0;
2727         set_bit(R1BIO_IsSync, &r1_bio->state);
2728         /* make sure good_sectors won't go across barrier unit boundary */
2729         good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2730
2731         for (i = 0; i < conf->raid_disks * 2; i++) {
2732                 struct md_rdev *rdev;
2733                 bio = r1_bio->bios[i];
2734
2735                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2736                 if (rdev == NULL ||
2737                     test_bit(Faulty, &rdev->flags)) {
2738                         if (i < conf->raid_disks)
2739                                 still_degraded = 1;
2740                 } else if (!test_bit(In_sync, &rdev->flags)) {
2741                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2742                         bio->bi_end_io = end_sync_write;
2743                         write_targets ++;
2744                 } else {
2745                         /* may need to read from here */
2746                         sector_t first_bad = MaxSector;
2747                         int bad_sectors;
2748
2749                         if (is_badblock(rdev, sector_nr, good_sectors,
2750                                         &first_bad, &bad_sectors)) {
2751                                 if (first_bad > sector_nr)
2752                                         good_sectors = first_bad - sector_nr;
2753                                 else {
2754                                         bad_sectors -= (sector_nr - first_bad);
2755                                         if (min_bad == 0 ||
2756                                             min_bad > bad_sectors)
2757                                                 min_bad = bad_sectors;
2758                                 }
2759                         }
2760                         if (sector_nr < first_bad) {
2761                                 if (test_bit(WriteMostly, &rdev->flags)) {
2762                                         if (wonly < 0)
2763                                                 wonly = i;
2764                                 } else {
2765                                         if (disk < 0)
2766                                                 disk = i;
2767                                 }
2768                                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2769                                 bio->bi_end_io = end_sync_read;
2770                                 read_targets++;
2771                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2772                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2773                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2774                                 /*
2775                                  * The device is suitable for reading (InSync),
2776                                  * but has bad block(s) here. Let's try to correct them,
2777                                  * if we are doing resync or repair. Otherwise, leave
2778                                  * this device alone for this sync request.
2779                                  */
2780                                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2781                                 bio->bi_end_io = end_sync_write;
2782                                 write_targets++;
2783                         }
2784                 }
2785                 if (bio->bi_end_io) {
2786                         atomic_inc(&rdev->nr_pending);
2787                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2788                         bio_set_dev(bio, rdev->bdev);
2789                         if (test_bit(FailFast, &rdev->flags))
2790                                 bio->bi_opf |= MD_FAILFAST;
2791                 }
2792         }
2793         rcu_read_unlock();
2794         if (disk < 0)
2795                 disk = wonly;
2796         r1_bio->read_disk = disk;
2797
2798         if (read_targets == 0 && min_bad > 0) {
2799                 /* These sectors are bad on all InSync devices, so we
2800                  * need to mark them bad on all write targets
2801                  */
2802                 int ok = 1;
2803                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2804                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2805                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2806                                 ok = rdev_set_badblocks(rdev, sector_nr,
2807                                                         min_bad, 0
2808                                         ) && ok;
2809                         }
2810                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2811                 *skipped = 1;
2812                 put_buf(r1_bio);
2813
2814                 if (!ok) {
2815                         /* Cannot record the badblocks, so need to
2816                          * abort the resync.
2817                          * If there are multiple read targets, could just
2818                          * fail the really bad ones ???
2819                          */
2820                         conf->recovery_disabled = mddev->recovery_disabled;
2821                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2822                         return 0;
2823                 } else
2824                         return min_bad;
2825
2826         }
2827         if (min_bad > 0 && min_bad < good_sectors) {
2828                 /* only resync enough to reach the next bad->good
2829                  * transition */
2830                 good_sectors = min_bad;
2831         }
2832
2833         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2834                 /* extra read targets are also write targets */
2835                 write_targets += read_targets-1;
2836
2837         if (write_targets == 0 || read_targets == 0) {
2838                 /* There is nowhere to write, so all non-sync
2839                  * drives must be failed - so we are finished
2840                  */
2841                 sector_t rv;
2842                 if (min_bad > 0)
2843                         max_sector = sector_nr + min_bad;
2844                 rv = max_sector - sector_nr;
2845                 *skipped = 1;
2846                 put_buf(r1_bio);
2847                 return rv;
2848         }
2849
2850         if (max_sector > mddev->resync_max)
2851                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2852         if (max_sector > sector_nr + good_sectors)
2853                 max_sector = sector_nr + good_sectors;
2854         nr_sectors = 0;
2855         sync_blocks = 0;
2856         do {
2857                 struct page *page;
2858                 int len = PAGE_SIZE;
2859                 if (sector_nr + (len>>9) > max_sector)
2860                         len = (max_sector - sector_nr) << 9;
2861                 if (len == 0)
2862                         break;
2863                 if (sync_blocks == 0) {
2864                         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2865                                                   &sync_blocks, still_degraded) &&
2866                             !conf->fullsync &&
2867                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2868                                 break;
2869                         if ((len >> 9) > sync_blocks)
2870                                 len = sync_blocks<<9;
2871                 }
2872
2873                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2874                         struct resync_pages *rp;
2875
2876                         bio = r1_bio->bios[i];
2877                         rp = get_resync_pages(bio);
2878                         if (bio->bi_end_io) {
2879                                 page = resync_fetch_page(rp, page_idx);
2880
2881                                 /*
2882                                  * won't fail because the vec table is big
2883                                  * enough to hold all these pages
2884                                  */
2885                                 bio_add_page(bio, page, len, 0);
2886                         }
2887                 }
2888                 nr_sectors += len>>9;
2889                 sector_nr += len>>9;
2890                 sync_blocks -= (len>>9);
2891         } while (++page_idx < RESYNC_PAGES);
2892
2893         r1_bio->sectors = nr_sectors;
2894
2895         if (mddev_is_clustered(mddev) &&
2896                         conf->cluster_sync_high < sector_nr + nr_sectors) {
2897                 conf->cluster_sync_low = mddev->curr_resync_completed;
2898                 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2899                 /* Send resync message */
2900                 md_cluster_ops->resync_info_update(mddev,
2901                                 conf->cluster_sync_low,
2902                                 conf->cluster_sync_high);
2903         }
2904
2905         /* For a user-requested sync, we read all readable devices and do a
2906          * compare
2907          */
2908         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2909                 atomic_set(&r1_bio->remaining, read_targets);
2910                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2911                         bio = r1_bio->bios[i];
2912                         if (bio->bi_end_io == end_sync_read) {
2913                                 read_targets--;
2914                                 md_sync_acct_bio(bio, nr_sectors);
2915                                 if (read_targets == 1)
2916                                         bio->bi_opf &= ~MD_FAILFAST;
2917                                 generic_make_request(bio);
2918                         }
2919                 }
2920         } else {
2921                 atomic_set(&r1_bio->remaining, 1);
2922                 bio = r1_bio->bios[r1_bio->read_disk];
2923                 md_sync_acct_bio(bio, nr_sectors);
2924                 if (read_targets == 1)
2925                         bio->bi_opf &= ~MD_FAILFAST;
2926                 generic_make_request(bio);
2927         }
2928         return nr_sectors;
2929 }
2930
2931 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2932 {
2933         if (sectors)
2934                 return sectors;
2935
2936         return mddev->dev_sectors;
2937 }
2938
2939 static struct r1conf *setup_conf(struct mddev *mddev)
2940 {
2941         struct r1conf *conf;
2942         int i;
2943         struct raid1_info *disk;
2944         struct md_rdev *rdev;
2945         int err = -ENOMEM;
2946
2947         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2948         if (!conf)
2949                 goto abort;
2950
2951         conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2952                                    sizeof(atomic_t), GFP_KERNEL);
2953         if (!conf->nr_pending)
2954                 goto abort;
2955
2956         conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2957                                    sizeof(atomic_t), GFP_KERNEL);
2958         if (!conf->nr_waiting)
2959                 goto abort;
2960
2961         conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2962                                   sizeof(atomic_t), GFP_KERNEL);
2963         if (!conf->nr_queued)
2964                 goto abort;
2965
2966         conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2967                                 sizeof(atomic_t), GFP_KERNEL);
2968         if (!conf->barrier)
2969                 goto abort;
2970
2971         conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2972                                             mddev->raid_disks, 2),
2973                                 GFP_KERNEL);
2974         if (!conf->mirrors)
2975                 goto abort;
2976
2977         conf->tmppage = alloc_page(GFP_KERNEL);
2978         if (!conf->tmppage)
2979                 goto abort;
2980
2981         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2982         if (!conf->poolinfo)
2983                 goto abort;
2984         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2985         err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2986                            rbio_pool_free, conf->poolinfo);
2987         if (err)
2988                 goto abort;
2989
2990         err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2991         if (err)
2992                 goto abort;
2993
2994         conf->poolinfo->mddev = mddev;
2995
2996         err = -EINVAL;
2997         spin_lock_init(&conf->device_lock);
2998         rdev_for_each(rdev, mddev) {
2999                 int disk_idx = rdev->raid_disk;
3000                 if (disk_idx >= mddev->raid_disks
3001                     || disk_idx < 0)
3002                         continue;
3003                 if (test_bit(Replacement, &rdev->flags))
3004                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
3005                 else
3006                         disk = conf->mirrors + disk_idx;
3007
3008                 if (disk->rdev)
3009                         goto abort;
3010                 disk->rdev = rdev;
3011                 disk->head_position = 0;
3012                 disk->seq_start = MaxSector;
3013         }
3014         conf->raid_disks = mddev->raid_disks;
3015         conf->mddev = mddev;
3016         INIT_LIST_HEAD(&conf->retry_list);
3017         INIT_LIST_HEAD(&conf->bio_end_io_list);
3018
3019         spin_lock_init(&conf->resync_lock);
3020         init_waitqueue_head(&conf->wait_barrier);
3021
3022         bio_list_init(&conf->pending_bio_list);
3023         conf->pending_count = 0;
3024         conf->recovery_disabled = mddev->recovery_disabled - 1;
3025
3026         err = -EIO;
3027         for (i = 0; i < conf->raid_disks * 2; i++) {
3028
3029                 disk = conf->mirrors + i;
3030
3031                 if (i < conf->raid_disks &&
3032                     disk[conf->raid_disks].rdev) {
3033                         /* This slot has a replacement. */
3034                         if (!disk->rdev) {
3035                                 /* No original, just make the replacement
3036                                  * a recovering spare
3037                                  */
3038                                 disk->rdev =
3039                                         disk[conf->raid_disks].rdev;
3040                                 disk[conf->raid_disks].rdev = NULL;
3041                         } else if (!test_bit(In_sync, &disk->rdev->flags))
3042                                 /* Original is not in_sync - bad */
3043                                 goto abort;
3044                 }
3045
3046                 if (!disk->rdev ||
3047                     !test_bit(In_sync, &disk->rdev->flags)) {
3048                         disk->head_position = 0;
3049                         if (disk->rdev &&
3050                             (disk->rdev->saved_raid_disk < 0))
3051                                 conf->fullsync = 1;
3052                 }
3053         }
3054
3055         err = -ENOMEM;
3056         conf->thread = md_register_thread(raid1d, mddev, "raid1");
3057         if (!conf->thread)
3058                 goto abort;
3059
3060         return conf;
3061
3062  abort:
3063         if (conf) {
3064                 mempool_exit(&conf->r1bio_pool);
3065                 kfree(conf->mirrors);
3066                 safe_put_page(conf->tmppage);
3067                 kfree(conf->poolinfo);
3068                 kfree(conf->nr_pending);
3069                 kfree(conf->nr_waiting);
3070                 kfree(conf->nr_queued);
3071                 kfree(conf->barrier);
3072                 bioset_exit(&conf->bio_split);
3073                 kfree(conf);
3074         }
3075         return ERR_PTR(err);
3076 }
3077
3078 static void raid1_free(struct mddev *mddev, void *priv);
3079 static int raid1_run(struct mddev *mddev)
3080 {
3081         struct r1conf *conf;
3082         int i;
3083         struct md_rdev *rdev;
3084         int ret;
3085         bool discard_supported = false;
3086
3087         if (mddev->level != 1) {
3088                 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3089                         mdname(mddev), mddev->level);
3090                 return -EIO;
3091         }
3092         if (mddev->reshape_position != MaxSector) {
3093                 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3094                         mdname(mddev));
3095                 return -EIO;
3096         }
3097         if (mddev_init_writes_pending(mddev) < 0)
3098                 return -ENOMEM;
3099         /*
3100          * copy the already verified devices into our private RAID1
3101          * bookkeeping area. [whatever we allocate in run(),
3102          * should be freed in raid1_free()]
3103          */
3104         if (mddev->private == NULL)
3105                 conf = setup_conf(mddev);
3106         else
3107                 conf = mddev->private;
3108
3109         if (IS_ERR(conf))
3110                 return PTR_ERR(conf);
3111
3112         if (mddev->queue) {
3113                 blk_queue_max_write_same_sectors(mddev->queue, 0);
3114                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3115         }
3116
3117         rdev_for_each(rdev, mddev) {
3118                 if (!mddev->gendisk)
3119                         continue;
3120                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3121                                   rdev->data_offset << 9);
3122                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3123                         discard_supported = true;
3124         }
3125
3126         mddev->degraded = 0;
3127         for (i = 0; i < conf->raid_disks; i++)
3128                 if (conf->mirrors[i].rdev == NULL ||
3129                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3130                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3131                         mddev->degraded++;
3132         /*
3133          * RAID1 needs at least one disk in active
3134          */
3135         if (conf->raid_disks - mddev->degraded < 1) {
3136                 ret = -EINVAL;
3137                 goto abort;
3138         }
3139
3140         if (conf->raid_disks - mddev->degraded == 1)
3141                 mddev->recovery_cp = MaxSector;
3142
3143         if (mddev->recovery_cp != MaxSector)
3144                 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3145                         mdname(mddev));
3146         pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3147                 mdname(mddev), mddev->raid_disks - mddev->degraded,
3148                 mddev->raid_disks);
3149
3150         /*
3151          * Ok, everything is just fine now
3152          */
3153         mddev->thread = conf->thread;
3154         conf->thread = NULL;
3155         mddev->private = conf;
3156         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3157
3158         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3159
3160         if (mddev->queue) {
3161                 if (discard_supported)
3162                         blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3163                                                 mddev->queue);
3164                 else
3165                         blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3166                                                   mddev->queue);
3167         }
3168
3169         ret = md_integrity_register(mddev);
3170         if (ret) {
3171                 md_unregister_thread(&mddev->thread);
3172                 goto abort;
3173         }
3174         return 0;
3175
3176 abort:
3177         raid1_free(mddev, conf);
3178         return ret;
3179 }
3180
3181 static void raid1_free(struct mddev *mddev, void *priv)
3182 {
3183         struct r1conf *conf = priv;
3184
3185         mempool_exit(&conf->r1bio_pool);
3186         kfree(conf->mirrors);
3187         safe_put_page(conf->tmppage);
3188         kfree(conf->poolinfo);
3189         kfree(conf->nr_pending);
3190         kfree(conf->nr_waiting);
3191         kfree(conf->nr_queued);
3192         kfree(conf->barrier);
3193         bioset_exit(&conf->bio_split);
3194         kfree(conf);
3195 }
3196
3197 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3198 {
3199         /* no resync is happening, and there is enough space
3200          * on all devices, so we can resize.
3201          * We need to make sure resync covers any new space.
3202          * If the array is shrinking we should possibly wait until
3203          * any io in the removed space completes, but it hardly seems
3204          * worth it.
3205          */
3206         sector_t newsize = raid1_size(mddev, sectors, 0);
3207         if (mddev->external_size &&
3208             mddev->array_sectors > newsize)
3209                 return -EINVAL;
3210         if (mddev->bitmap) {
3211                 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3212                 if (ret)
3213                         return ret;
3214         }
3215         md_set_array_sectors(mddev, newsize);
3216         if (sectors > mddev->dev_sectors &&
3217             mddev->recovery_cp > mddev->dev_sectors) {
3218                 mddev->recovery_cp = mddev->dev_sectors;
3219                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3220         }
3221         mddev->dev_sectors = sectors;
3222         mddev->resync_max_sectors = sectors;
3223         return 0;
3224 }
3225
3226 static int raid1_reshape(struct mddev *mddev)
3227 {
3228         /* We need to:
3229          * 1/ resize the r1bio_pool
3230          * 2/ resize conf->mirrors
3231          *
3232          * We allocate a new r1bio_pool if we can.
3233          * Then raise a device barrier and wait until all IO stops.
3234          * Then resize conf->mirrors and swap in the new r1bio pool.
3235          *
3236          * At the same time, we "pack" the devices so that all the missing
3237          * devices have the higher raid_disk numbers.
3238          */
3239         mempool_t newpool, oldpool;
3240         struct pool_info *newpoolinfo;
3241         struct raid1_info *newmirrors;
3242         struct r1conf *conf = mddev->private;
3243         int cnt, raid_disks;
3244         unsigned long flags;
3245         int d, d2;
3246         int ret;
3247
3248         memset(&newpool, 0, sizeof(newpool));
3249         memset(&oldpool, 0, sizeof(oldpool));
3250
3251         /* Cannot change chunk_size, layout, or level */
3252         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3253             mddev->layout != mddev->new_layout ||
3254             mddev->level != mddev->new_level) {
3255                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3256                 mddev->new_layout = mddev->layout;
3257                 mddev->new_level = mddev->level;
3258                 return -EINVAL;
3259         }
3260
3261         if (!mddev_is_clustered(mddev))
3262                 md_allow_write(mddev);
3263
3264         raid_disks = mddev->raid_disks + mddev->delta_disks;
3265
3266         if (raid_disks < conf->raid_disks) {
3267                 cnt=0;
3268                 for (d= 0; d < conf->raid_disks; d++)
3269                         if (conf->mirrors[d].rdev)
3270                                 cnt++;
3271                 if (cnt > raid_disks)
3272                         return -EBUSY;
3273         }
3274
3275         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3276         if (!newpoolinfo)
3277                 return -ENOMEM;
3278         newpoolinfo->mddev = mddev;
3279         newpoolinfo->raid_disks = raid_disks * 2;
3280
3281         ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3282                            rbio_pool_free, newpoolinfo);
3283         if (ret) {
3284                 kfree(newpoolinfo);
3285                 return ret;
3286         }
3287         newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3288                                          raid_disks, 2),
3289                              GFP_KERNEL);
3290         if (!newmirrors) {
3291                 kfree(newpoolinfo);
3292                 mempool_exit(&newpool);
3293                 return -ENOMEM;
3294         }
3295
3296         freeze_array(conf, 0);
3297
3298         /* ok, everything is stopped */
3299         oldpool = conf->r1bio_pool;
3300         conf->r1bio_pool = newpool;
3301
3302         for (d = d2 = 0; d < conf->raid_disks; d++) {
3303                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3304                 if (rdev && rdev->raid_disk != d2) {
3305                         sysfs_unlink_rdev(mddev, rdev);
3306                         rdev->raid_disk = d2;
3307                         sysfs_unlink_rdev(mddev, rdev);
3308                         if (sysfs_link_rdev(mddev, rdev))
3309                                 pr_warn("md/raid1:%s: cannot register rd%d\n",
3310                                         mdname(mddev), rdev->raid_disk);
3311                 }
3312                 if (rdev)
3313                         newmirrors[d2++].rdev = rdev;
3314         }
3315         kfree(conf->mirrors);
3316         conf->mirrors = newmirrors;
3317         kfree(conf->poolinfo);
3318         conf->poolinfo = newpoolinfo;
3319
3320         spin_lock_irqsave(&conf->device_lock, flags);
3321         mddev->degraded += (raid_disks - conf->raid_disks);
3322         spin_unlock_irqrestore(&conf->device_lock, flags);
3323         conf->raid_disks = mddev->raid_disks = raid_disks;
3324         mddev->delta_disks = 0;
3325
3326         unfreeze_array(conf);
3327
3328         set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3329         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3330         md_wakeup_thread(mddev->thread);
3331
3332         mempool_exit(&oldpool);
3333         return 0;
3334 }
3335
3336 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3337 {
3338         struct r1conf *conf = mddev->private;
3339
3340         if (quiesce)
3341                 freeze_array(conf, 0);
3342         else
3343                 unfreeze_array(conf);
3344 }
3345
3346 static void *raid1_takeover(struct mddev *mddev)
3347 {
3348         /* raid1 can take over:
3349          *  raid5 with 2 devices, any layout or chunk size
3350          */
3351         if (mddev->level == 5 && mddev->raid_disks == 2) {
3352                 struct r1conf *conf;
3353                 mddev->new_level = 1;
3354                 mddev->new_layout = 0;
3355                 mddev->new_chunk_sectors = 0;
3356                 conf = setup_conf(mddev);
3357                 if (!IS_ERR(conf)) {
3358                         /* Array must appear to be quiesced */
3359                         conf->array_frozen = 1;
3360                         mddev_clear_unsupported_flags(mddev,
3361                                 UNSUPPORTED_MDDEV_FLAGS);
3362                 }
3363                 return conf;
3364         }
3365         return ERR_PTR(-EINVAL);
3366 }
3367
3368 static struct md_personality raid1_personality =
3369 {
3370         .name           = "raid1",
3371         .level          = 1,
3372         .owner          = THIS_MODULE,
3373         .make_request   = raid1_make_request,
3374         .run            = raid1_run,
3375         .free           = raid1_free,
3376         .status         = raid1_status,
3377         .error_handler  = raid1_error,
3378         .hot_add_disk   = raid1_add_disk,
3379         .hot_remove_disk= raid1_remove_disk,
3380         .spare_active   = raid1_spare_active,
3381         .sync_request   = raid1_sync_request,
3382         .resize         = raid1_resize,
3383         .size           = raid1_size,
3384         .check_reshape  = raid1_reshape,
3385         .quiesce        = raid1_quiesce,
3386         .takeover       = raid1_takeover,
3387         .congested      = raid1_congested,
3388 };
3389
3390 static int __init raid_init(void)
3391 {
3392         return register_md_personality(&raid1_personality);
3393 }
3394
3395 static void raid_exit(void)
3396 {
3397         unregister_md_personality(&raid1_personality);
3398 }
3399
3400 module_init(raid_init);
3401 module_exit(raid_exit);
3402 MODULE_LICENSE("GPL");
3403 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3404 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3405 MODULE_ALIAS("md-raid1");
3406 MODULE_ALIAS("md-level-1");
3407
3408 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);