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