drivers/md/raid1.c

   1 /*
   2  * raid1.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5  *
   6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7  *
   8  * RAID-1 management functions.
   9  *
  10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11  *
  12  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14  *
  15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  16  * bitmapped intelligence in resync:
  17  *
  18  *      - bitmap marked during normal i/o
  19  *      - bitmap used to skip nondirty blocks during sync
  20  *
  21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  22  * - persistent bitmap code
  23  *
  24  * This program is free software; you can redistribute it and/or modify
  25  * it under the terms of the GNU General Public License as published by
  26  * the Free Software Foundation; either version 2, or (at your option)
  27  * any later version.
  28  *
  29  * You should have received a copy of the GNU General Public License
  30  * (for example /usr/src/linux/COPYING); if not, write to the Free
  31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  32  */
  33
  34 #include <linux/slab.h>
  35 #include <linux/delay.h>
  36 #include <linux/blkdev.h>
  37 #include <linux/module.h>
  38 #include <linux/seq_file.h>
  39 #include <linux/ratelimit.h>
  40 #include <linux/sched/signal.h>
  41
  42 #include <trace/events/block.h>
  43
  44 #include "md.h"
  45 #include "raid1.h"
  46 #include "bitmap.h"
  47
  48 #define UNSUPPORTED_MDDEV_FLAGS         \
  49         ((1L << MD_HAS_JOURNAL) |       \
  50          (1L << MD_JOURNAL_CLEAN) |     \
  51          (1L << MD_HAS_PPL))
  52
  53 /*
  54  * Number of guaranteed r1bios in case of extreme VM load:
  55  */
  56 #define NR_RAID1_BIOS 256
  57
  58 /* when we get a read error on a read-only array, we redirect to another
  59  * device without failing the first device, or trying to over-write to
  60  * correct the read error.  To keep track of bad blocks on a per-bio
  61  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
  62  */
  63 #define IO_BLOCKED ((struct bio *)1)
  64 /* When we successfully write to a known bad-block, we need to remove the
  65  * bad-block marking which must be done from process context.  So we record
  66  * the success by setting devs[n].bio to IO_MADE_GOOD
  67  */
  68 #define IO_MADE_GOOD ((struct bio *)2)
  69
  70 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
  71
  72 /* When there are this many requests queue to be written by
  73  * the raid1 thread, we become 'congested' to provide back-pressure
  74  * for writeback.
  75  */
  76 static int max_queued_requests = 1024;
  77
  78 static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
  79 static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
  80
  81 #define raid1_log(md, fmt, args...)                             \
  82         do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
  83
  84 #include "raid1-10.c"
  85
  86 /*
  87  * for resync bio, r1bio pointer can be retrieved from the per-bio
  88  * 'struct resync_pages'.
  89  */
  90 static inline struct r1bio *get_resync_r1bio(struct bio *bio)
  91 {
  92         return get_resync_pages(bio)->raid_bio;
  93 }
  94
  95 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  96 {
  97         struct pool_info *pi = data;
  98         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
  99
 100         /* allocate a r1bio with room for raid_disks entries in the bios array */
 101         return kzalloc(size, gfp_flags);
 102 }
 103
 104 static void r1bio_pool_free(void *r1_bio, void *data)
 105 {
 106         kfree(r1_bio);
 107 }
 108
 109 #define RESYNC_DEPTH 32
 110 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
 111 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
 112 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
 113 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
 114 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
 115
 116 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
 117 {
 118         struct pool_info *pi = data;
 119         struct r1bio *r1_bio;
 120         struct bio *bio;
 121         int need_pages;
 122         int j;
 123         struct resync_pages *rps;
 124
 125         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 126         if (!r1_bio)
 127                 return NULL;
 128
 129         rps = kmalloc(sizeof(struct resync_pages) * pi->raid_disks,
 130                       gfp_flags);
 131         if (!rps)
 132                 goto out_free_r1bio;
 133
 134         /*
 135          * Allocate bios : 1 for reading, n-1 for writing
 136          */
 137         for (j = pi->raid_disks ; j-- ; ) {
 138                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 139                 if (!bio)
 140                         goto out_free_bio;
 141                 r1_bio->bios[j] = bio;
 142         }
 143         /*
 144          * Allocate RESYNC_PAGES data pages and attach them to
 145          * the first bio.
 146          * If this is a user-requested check/repair, allocate
 147          * RESYNC_PAGES for each bio.
 148          */
 149         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 150                 need_pages = pi->raid_disks;
 151         else
 152                 need_pages = 1;
 153         for (j = 0; j < pi->raid_disks; j++) {
 154                 struct resync_pages *rp = &rps[j];
 155
 156                 bio = r1_bio->bios[j];
 157
 158                 if (j < need_pages) {
 159                         if (resync_alloc_pages(rp, gfp_flags))
 160                                 goto out_free_pages;
 161                 } else {
 162                         memcpy(rp, &rps[0], sizeof(*rp));
 163                         resync_get_all_pages(rp);
 164                 }
 165
 166                 rp->raid_bio = r1_bio;
 167                 bio->bi_private = rp;
 168         }
 169
 170         r1_bio->master_bio = NULL;
 171
 172         return r1_bio;
 173
 174 out_free_pages:
 175         while (--j >= 0)
 176                 resync_free_pages(&rps[j]);
 177
 178 out_free_bio:
 179         while (++j < pi->raid_disks)
 180                 bio_put(r1_bio->bios[j]);
 181         kfree(rps);
 182
 183 out_free_r1bio:
 184         r1bio_pool_free(r1_bio, data);
 185         return NULL;
 186 }
 187
 188 static void r1buf_pool_free(void *__r1_bio, void *data)
 189 {
 190         struct pool_info *pi = data;
 191         int i;
 192         struct r1bio *r1bio = __r1_bio;
 193         struct resync_pages *rp = NULL;
 194
 195         for (i = pi->raid_disks; i--; ) {
 196                 rp = get_resync_pages(r1bio->bios[i]);
 197                 resync_free_pages(rp);
 198                 bio_put(r1bio->bios[i]);
 199         }
 200
 201         /* resync pages array stored in the 1st bio's .bi_private */
 202         kfree(rp);
 203
 204         r1bio_pool_free(r1bio, data);
 205 }
 206
 207 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 208 {
 209         int i;
 210
 211         for (i = 0; i < conf->raid_disks * 2; i++) {
 212                 struct bio **bio = r1_bio->bios + i;
 213                 if (!BIO_SPECIAL(*bio))
 214                         bio_put(*bio);
 215                 *bio = NULL;
 216         }
 217 }
 218
 219 static void free_r1bio(struct r1bio *r1_bio)
 220 {
 221         struct r1conf *conf = r1_bio->mddev->private;
 222
 223         put_all_bios(conf, r1_bio);
 224         mempool_free(r1_bio, conf->r1bio_pool);
 225 }
 226
 227 static void put_buf(struct r1bio *r1_bio)
 228 {
 229         struct r1conf *conf = r1_bio->mddev->private;
 230         sector_t sect = r1_bio->sector;
 231         int i;
 232
 233         for (i = 0; i < conf->raid_disks * 2; i++) {
 234                 struct bio *bio = r1_bio->bios[i];
 235                 if (bio->bi_end_io)
 236                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 237         }
 238
 239         mempool_free(r1_bio, conf->r1buf_pool);
 240
 241         lower_barrier(conf, sect);
 242 }
 243
 244 static void reschedule_retry(struct r1bio *r1_bio)
 245 {
 246         unsigned long flags;
 247         struct mddev *mddev = r1_bio->mddev;
 248         struct r1conf *conf = mddev->private;
 249         int idx;
 250
 251         idx = sector_to_idx(r1_bio->sector);
 252         spin_lock_irqsave(&conf->device_lock, flags);
 253         list_add(&r1_bio->retry_list, &conf->retry_list);
 254         atomic_inc(&conf->nr_queued[idx]);
 255         spin_unlock_irqrestore(&conf->device_lock, flags);
 256
 257         wake_up(&conf->wait_barrier);
 258         md_wakeup_thread(mddev->thread);
 259 }
 260
 261 /*
 262  * raid_end_bio_io() is called when we have finished servicing a mirrored
 263  * operation and are ready to return a success/failure code to the buffer
 264  * cache layer.
 265  */
 266 static void call_bio_endio(struct r1bio *r1_bio)
 267 {
 268         struct bio *bio = r1_bio->master_bio;
 269         struct r1conf *conf = r1_bio->mddev->private;
 270
 271         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 272                 bio->bi_status = BLK_STS_IOERR;
 273
 274         bio_endio(bio);
 275         /*
 276          * Wake up any possible resync thread that waits for the device
 277          * to go idle.
 278          */
 279         allow_barrier(conf, r1_bio->sector);
 280 }
 281
 282 static void raid_end_bio_io(struct r1bio *r1_bio)
 283 {
 284         struct bio *bio = r1_bio->master_bio;
 285
 286         /* if nobody has done the final endio yet, do it now */
 287         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 288                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 289                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 290                          (unsigned long long) bio->bi_iter.bi_sector,
 291                          (unsigned long long) bio_end_sector(bio) - 1);
 292
 293                 call_bio_endio(r1_bio);
 294         }
 295         free_r1bio(r1_bio);
 296 }
 297
 298 /*
 299  * Update disk head position estimator based on IRQ completion info.
 300  */
 301 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 302 {
 303         struct r1conf *conf = r1_bio->mddev->private;
 304
 305         conf->mirrors[disk].head_position =
 306                 r1_bio->sector + (r1_bio->sectors);
 307 }
 308
 309 /*
 310  * Find the disk number which triggered given bio
 311  */
 312 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 313 {
 314         int mirror;
 315         struct r1conf *conf = r1_bio->mddev->private;
 316         int raid_disks = conf->raid_disks;
 317
 318         for (mirror = 0; mirror < raid_disks * 2; mirror++)
 319                 if (r1_bio->bios[mirror] == bio)
 320                         break;
 321
 322         BUG_ON(mirror == raid_disks * 2);
 323         update_head_pos(mirror, r1_bio);
 324
 325         return mirror;
 326 }
 327
 328 static void raid1_end_read_request(struct bio *bio)
 329 {
 330         int uptodate = !bio->bi_status;
 331         struct r1bio *r1_bio = bio->bi_private;
 332         struct r1conf *conf = r1_bio->mddev->private;
 333         struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 334
 335         /*
 336          * this branch is our 'one mirror IO has finished' event handler:
 337          */
 338         update_head_pos(r1_bio->read_disk, r1_bio);
 339
 340         if (uptodate)
 341                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 342         else if (test_bit(FailFast, &rdev->flags) &&
 343                  test_bit(R1BIO_FailFast, &r1_bio->state))
 344                 /* This was a fail-fast read so we definitely
 345                  * want to retry */
 346                 ;
 347         else {
 348                 /* If all other devices have failed, we want to return
 349                  * the error upwards rather than fail the last device.
 350                  * Here we redefine "uptodate" to mean "Don't want to retry"
 351                  */
 352                 unsigned long flags;
 353                 spin_lock_irqsave(&conf->device_lock, flags);
 354                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 355                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 356                      test_bit(In_sync, &rdev->flags)))
 357                         uptodate = 1;
 358                 spin_unlock_irqrestore(&conf->device_lock, flags);
 359         }
 360
 361         if (uptodate) {
 362                 raid_end_bio_io(r1_bio);
 363                 rdev_dec_pending(rdev, conf->mddev);
 364         } else {
 365                 /*
 366                  * oops, read error:
 367                  */
 368                 char b[BDEVNAME_SIZE];
 369                 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
 370                                    mdname(conf->mddev),
 371                                    bdevname(rdev->bdev, b),
 372                                    (unsigned long long)r1_bio->sector);
 373                 set_bit(R1BIO_ReadError, &r1_bio->state);
 374                 reschedule_retry(r1_bio);
 375                 /* don't drop the reference on read_disk yet */
 376         }
 377 }
 378
 379 static void close_write(struct r1bio *r1_bio)
 380 {
 381         /* it really is the end of this request */
 382         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 383                 bio_free_pages(r1_bio->behind_master_bio);
 384                 bio_put(r1_bio->behind_master_bio);
 385                 r1_bio->behind_master_bio = NULL;
 386         }
 387         /* clear the bitmap if all writes complete successfully */
 388         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 389                         r1_bio->sectors,
 390                         !test_bit(R1BIO_Degraded, &r1_bio->state),
 391                         test_bit(R1BIO_BehindIO, &r1_bio->state));
 392         md_write_end(r1_bio->mddev);
 393 }
 394
 395 static void r1_bio_write_done(struct r1bio *r1_bio)
 396 {
 397         if (!atomic_dec_and_test(&r1_bio->remaining))
 398                 return;
 399
 400         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 401                 reschedule_retry(r1_bio);
 402         else {
 403                 close_write(r1_bio);
 404                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 405                         reschedule_retry(r1_bio);
 406                 else
 407                         raid_end_bio_io(r1_bio);
 408         }
 409 }
 410
 411 static void raid1_end_write_request(struct bio *bio)
 412 {
 413         struct r1bio *r1_bio = bio->bi_private;
 414         int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 415         struct r1conf *conf = r1_bio->mddev->private;
 416         struct bio *to_put = NULL;
 417         int mirror = find_bio_disk(r1_bio, bio);
 418         struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 419         bool discard_error;
 420
 421         discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 422
 423         /*
 424          * 'one mirror IO has finished' event handler:
 425          */
 426         if (bio->bi_status && !discard_error) {
 427                 set_bit(WriteErrorSeen, &rdev->flags);
 428                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
 429                         set_bit(MD_RECOVERY_NEEDED, &
 430                                 conf->mddev->recovery);
 431
 432                 if (test_bit(FailFast, &rdev->flags) &&
 433                     (bio->bi_opf & MD_FAILFAST) &&
 434                     /* We never try FailFast to WriteMostly devices */
 435                     !test_bit(WriteMostly, &rdev->flags)) {
 436                         md_error(r1_bio->mddev, rdev);
 437                         if (!test_bit(Faulty, &rdev->flags))
 438                                 /* This is the only remaining device,
 439                                  * We need to retry the write without
 440                                  * FailFast
 441                                  */
 442                                 set_bit(R1BIO_WriteError, &r1_bio->state);
 443                         else {
 444                                 /* Finished with this branch */
 445                                 r1_bio->bios[mirror] = NULL;
 446                                 to_put = bio;
 447                         }
 448                 } else
 449                         set_bit(R1BIO_WriteError, &r1_bio->state);
 450         } else {
 451                 /*
 452                  * Set R1BIO_Uptodate in our master bio, so that we
 453                  * will return a good error code for to the higher
 454                  * levels even if IO on some other mirrored buffer
 455                  * fails.
 456                  *
 457                  * The 'master' represents the composite IO operation
 458                  * to user-side. So if something waits for IO, then it
 459                  * will wait for the 'master' bio.
 460                  */
 461                 sector_t first_bad;
 462                 int bad_sectors;
 463
 464                 r1_bio->bios[mirror] = NULL;
 465                 to_put = bio;
 466                 /*
 467                  * Do not set R1BIO_Uptodate if the current device is
 468                  * rebuilding or Faulty. This is because we cannot use
 469                  * such device for properly reading the data back (we could
 470                  * potentially use it, if the current write would have felt
 471                  * before rdev->recovery_offset, but for simplicity we don't
 472                  * check this here.
 473                  */
 474                 if (test_bit(In_sync, &rdev->flags) &&
 475                     !test_bit(Faulty, &rdev->flags))
 476                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 477
 478                 /* Maybe we can clear some bad blocks. */
 479                 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 480                                 &first_bad, &bad_sectors) && !discard_error) {
 481                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 482                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 483                 }
 484         }
 485
 486         if (behind) {
 487                 if (test_bit(WriteMostly, &rdev->flags))
 488                         atomic_dec(&r1_bio->behind_remaining);
 489
 490                 /*
 491                  * In behind mode, we ACK the master bio once the I/O
 492                  * has safely reached all non-writemostly
 493                  * disks. Setting the Returned bit ensures that this
 494                  * gets done only once -- we don't ever want to return
 495                  * -EIO here, instead we'll wait
 496                  */
 497                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 498                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 499                         /* Maybe we can return now */
 500                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 501                                 struct bio *mbio = r1_bio->master_bio;
 502                                 pr_debug("raid1: behind end write sectors"
 503                                          " %llu-%llu\n",
 504                                          (unsigned long long) mbio->bi_iter.bi_sector,
 505                                          (unsigned long long) bio_end_sector(mbio) - 1);
 506                                 call_bio_endio(r1_bio);
 507                         }
 508                 }
 509         }
 510         if (r1_bio->bios[mirror] == NULL)
 511                 rdev_dec_pending(rdev, conf->mddev);
 512
 513         /*
 514          * Let's see if all mirrored write operations have finished
 515          * already.
 516          */
 517         r1_bio_write_done(r1_bio);
 518
 519         if (to_put)
 520                 bio_put(to_put);
 521 }
 522
 523 static sector_t align_to_barrier_unit_end(sector_t start_sector,
 524                                           sector_t sectors)
 525 {
 526         sector_t len;
 527
 528         WARN_ON(sectors == 0);
 529         /*
 530          * len is the number of sectors from start_sector to end of the
 531          * barrier unit which start_sector belongs to.
 532          */
 533         len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 534               start_sector;
 535
 536         if (len > sectors)
 537                 len = sectors;
 538
 539         return len;
 540 }
 541
 542 /*
 543  * This routine returns the disk from which the requested read should
 544  * be done. There is a per-array 'next expected sequential IO' sector
 545  * number - if this matches on the next IO then we use the last disk.
 546  * There is also a per-disk 'last know head position' sector that is
 547  * maintained from IRQ contexts, both the normal and the resync IO
 548  * completion handlers update this position correctly. If there is no
 549  * perfect sequential match then we pick the disk whose head is closest.
 550  *
 551  * If there are 2 mirrors in the same 2 devices, performance degrades
 552  * because position is mirror, not device based.
 553  *
 554  * The rdev for the device selected will have nr_pending incremented.
 555  */
 556 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 557 {
 558         const sector_t this_sector = r1_bio->sector;
 559         int sectors;
 560         int best_good_sectors;
 561         int best_disk, best_dist_disk, best_pending_disk;
 562         int has_nonrot_disk;
 563         int disk;
 564         sector_t best_dist;
 565         unsigned int min_pending;
 566         struct md_rdev *rdev;
 567         int choose_first;
 568         int choose_next_idle;
 569
 570         rcu_read_lock();
 571         /*
 572          * Check if we can balance. We can balance on the whole
 573          * device if no resync is going on, or below the resync window.
 574          * We take the first readable disk when above the resync window.
 575          */
 576  retry:
 577         sectors = r1_bio->sectors;
 578         best_disk = -1;
 579         best_dist_disk = -1;
 580         best_dist = MaxSector;
 581         best_pending_disk = -1;
 582         min_pending = UINT_MAX;
 583         best_good_sectors = 0;
 584         has_nonrot_disk = 0;
 585         choose_next_idle = 0;
 586         clear_bit(R1BIO_FailFast, &r1_bio->state);
 587
 588         if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 589             (mddev_is_clustered(conf->mddev) &&
 590             md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 591                     this_sector + sectors)))
 592                 choose_first = 1;
 593         else
 594                 choose_first = 0;
 595
 596         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 597                 sector_t dist;
 598                 sector_t first_bad;
 599                 int bad_sectors;
 600                 unsigned int pending;
 601                 bool nonrot;
 602
 603                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 604                 if (r1_bio->bios[disk] == IO_BLOCKED
 605                     || rdev == NULL
 606                     || test_bit(Faulty, &rdev->flags))
 607                         continue;
 608                 if (!test_bit(In_sync, &rdev->flags) &&
 609                     rdev->recovery_offset < this_sector + sectors)
 610                         continue;
 611                 if (test_bit(WriteMostly, &rdev->flags)) {
 612                         /* Don't balance among write-mostly, just
 613                          * use the first as a last resort */
 614                         if (best_dist_disk < 0) {
 615                                 if (is_badblock(rdev, this_sector, sectors,
 616                                                 &first_bad, &bad_sectors)) {
 617                                         if (first_bad <= this_sector)
 618                                                 /* Cannot use this */
 619                                                 continue;
 620                                         best_good_sectors = first_bad - this_sector;
 621                                 } else
 622                                         best_good_sectors = sectors;
 623                                 best_dist_disk = disk;
 624                                 best_pending_disk = disk;
 625                         }
 626                         continue;
 627                 }
 628                 /* This is a reasonable device to use.  It might
 629                  * even be best.
 630                  */
 631                 if (is_badblock(rdev, this_sector, sectors,
 632                                 &first_bad, &bad_sectors)) {
 633                         if (best_dist < MaxSector)
 634                                 /* already have a better device */
 635                                 continue;
 636                         if (first_bad <= this_sector) {
 637                                 /* cannot read here. If this is the 'primary'
 638                                  * device, then we must not read beyond
 639                                  * bad_sectors from another device..
 640                                  */
 641                                 bad_sectors -= (this_sector - first_bad);
 642                                 if (choose_first && sectors > bad_sectors)
 643                                         sectors = bad_sectors;
 644                                 if (best_good_sectors > sectors)
 645                                         best_good_sectors = sectors;
 646
 647                         } else {
 648                                 sector_t good_sectors = first_bad - this_sector;
 649                                 if (good_sectors > best_good_sectors) {
 650                                         best_good_sectors = good_sectors;
 651                                         best_disk = disk;
 652                                 }
 653                                 if (choose_first)
 654                                         break;
 655                         }
 656                         continue;
 657                 } else {
 658                         if ((sectors > best_good_sectors) && (best_disk >= 0))
 659                                 best_disk = -1;
 660                         best_good_sectors = sectors;
 661                 }
 662
 663                 if (best_disk >= 0)
 664                         /* At least two disks to choose from so failfast is OK */
 665                         set_bit(R1BIO_FailFast, &r1_bio->state);
 666
 667                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 668                 has_nonrot_disk |= nonrot;
 669                 pending = atomic_read(&rdev->nr_pending);
 670                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 671                 if (choose_first) {
 672                         best_disk = disk;
 673                         break;
 674                 }
 675                 /* Don't change to another disk for sequential reads */
 676                 if (conf->mirrors[disk].next_seq_sect == this_sector
 677                     || dist == 0) {
 678                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 679                         struct raid1_info *mirror = &conf->mirrors[disk];
 680
 681                         best_disk = disk;
 682                         /*
 683                          * If buffered sequential IO size exceeds optimal
 684                          * iosize, check if there is idle disk. If yes, choose
 685                          * the idle disk. read_balance could already choose an
 686                          * idle disk before noticing it's a sequential IO in
 687                          * this disk. This doesn't matter because this disk
 688                          * will idle, next time it will be utilized after the
 689                          * first disk has IO size exceeds optimal iosize. In
 690                          * this way, iosize of the first disk will be optimal
 691                          * iosize at least. iosize of the second disk might be
 692                          * small, but not a big deal since when the second disk
 693                          * starts IO, the first disk is likely still busy.
 694                          */
 695                         if (nonrot && opt_iosize > 0 &&
 696                             mirror->seq_start != MaxSector &&
 697                             mirror->next_seq_sect > opt_iosize &&
 698                             mirror->next_seq_sect - opt_iosize >=
 699                             mirror->seq_start) {
 700                                 choose_next_idle = 1;
 701                                 continue;
 702                         }
 703                         break;
 704                 }
 705
 706                 if (choose_next_idle)
 707                         continue;
 708
 709                 if (min_pending > pending) {
 710                         min_pending = pending;
 711                         best_pending_disk = disk;
 712                 }
 713
 714                 if (dist < best_dist) {
 715                         best_dist = dist;
 716                         best_dist_disk = disk;
 717                 }
 718         }
 719
 720         /*
 721          * If all disks are rotational, choose the closest disk. If any disk is
 722          * non-rotational, choose the disk with less pending request even the
 723          * disk is rotational, which might/might not be optimal for raids with
 724          * mixed ratation/non-rotational disks depending on workload.
 725          */
 726         if (best_disk == -1) {
 727                 if (has_nonrot_disk || min_pending == 0)
 728                         best_disk = best_pending_disk;
 729                 else
 730                         best_disk = best_dist_disk;
 731         }
 732
 733         if (best_disk >= 0) {
 734                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 735                 if (!rdev)
 736                         goto retry;
 737                 atomic_inc(&rdev->nr_pending);
 738                 sectors = best_good_sectors;
 739
 740                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 741                         conf->mirrors[best_disk].seq_start = this_sector;
 742
 743                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 744         }
 745         rcu_read_unlock();
 746         *max_sectors = sectors;
 747
 748         return best_disk;
 749 }
 750
 751 static int raid1_congested(struct mddev *mddev, int bits)
 752 {
 753         struct r1conf *conf = mddev->private;
 754         int i, ret = 0;
 755
 756         if ((bits & (1 << WB_async_congested)) &&
 757             conf->pending_count >= max_queued_requests)
 758                 return 1;
 759
 760         rcu_read_lock();
 761         for (i = 0; i < conf->raid_disks * 2; i++) {
 762                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 763                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 764                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 765
 766                         BUG_ON(!q);
 767
 768                         /* Note the '|| 1' - when read_balance prefers
 769                          * non-congested targets, it can be removed
 770                          */
 771                         if ((bits & (1 << WB_async_congested)) || 1)
 772                                 ret |= bdi_congested(q->backing_dev_info, bits);
 773                         else
 774                                 ret &= bdi_congested(q->backing_dev_info, bits);
 775                 }
 776         }
 777         rcu_read_unlock();
 778         return ret;
 779 }
 780
 781 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 782 {
 783         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
 784         bitmap_unplug(conf->mddev->bitmap);
 785         wake_up(&conf->wait_barrier);
 786
 787         while (bio) { /* submit pending writes */
 788                 struct bio *next = bio->bi_next;
 789                 struct md_rdev *rdev = (void*)bio->bi_bdev;
 790                 bio->bi_next = NULL;
 791                 bio->bi_bdev = rdev->bdev;
 792                 if (test_bit(Faulty, &rdev->flags)) {
 793                         bio->bi_status = BLK_STS_IOERR;
 794                         bio_endio(bio);
 795                 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
 796                                     !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 797                         /* Just ignore it */
 798                         bio_endio(bio);
 799                 else
 800                         generic_make_request(bio);
 801                 bio = next;
 802         }
 803 }
 804
 805 static void flush_pending_writes(struct r1conf *conf)
 806 {
 807         /* Any writes that have been queued but are awaiting
 808          * bitmap updates get flushed here.
 809          */
 810         spin_lock_irq(&conf->device_lock);
 811
 812         if (conf->pending_bio_list.head) {
 813                 struct bio *bio;
 814                 bio = bio_list_get(&conf->pending_bio_list);
 815                 conf->pending_count = 0;
 816                 spin_unlock_irq(&conf->device_lock);
 817                 flush_bio_list(conf, bio);
 818         } else
 819                 spin_unlock_irq(&conf->device_lock);
 820 }
 821
 822 /* Barriers....
 823  * Sometimes we need to suspend IO while we do something else,
 824  * either some resync/recovery, or reconfigure the array.
 825  * To do this we raise a 'barrier'.
 826  * The 'barrier' is a counter that can be raised multiple times
 827  * to count how many activities are happening which preclude
 828  * normal IO.
 829  * We can only raise the barrier if there is no pending IO.
 830  * i.e. if nr_pending == 0.
 831  * We choose only to raise the barrier if no-one is waiting for the
 832  * barrier to go down.  This means that as soon as an IO request
 833  * is ready, no other operations which require a barrier will start
 834  * until the IO request has had a chance.
 835  *
 836  * So: regular IO calls 'wait_barrier'.  When that returns there
 837  *    is no backgroup IO happening,  It must arrange to call
 838  *    allow_barrier when it has finished its IO.
 839  * backgroup IO calls must call raise_barrier.  Once that returns
 840  *    there is no normal IO happeing.  It must arrange to call
 841  *    lower_barrier when the particular background IO completes.
 842  */
 843 static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
 844 {
 845         int idx = sector_to_idx(sector_nr);
 846
 847         spin_lock_irq(&conf->resync_lock);
 848
 849         /* Wait until no block IO is waiting */
 850         wait_event_lock_irq(conf->wait_barrier,
 851                             !atomic_read(&conf->nr_waiting[idx]),
 852                             conf->resync_lock);
 853
 854         /* block any new IO from starting */
 855         atomic_inc(&conf->barrier[idx]);
 856         /*
 857          * In raise_barrier() we firstly increase conf->barrier[idx] then
 858          * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 859          * increase conf->nr_pending[idx] then check conf->barrier[idx].
 860          * A memory barrier here to make sure conf->nr_pending[idx] won't
 861          * be fetched before conf->barrier[idx] is increased. Otherwise
 862          * there will be a race between raise_barrier() and _wait_barrier().
 863          */
 864         smp_mb__after_atomic();
 865
 866         /* For these conditions we must wait:
 867          * A: while the array is in frozen state
 868          * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 869          *    existing in corresponding I/O barrier bucket.
 870          * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 871          *    max resync count which allowed on current I/O barrier bucket.
 872          */
 873         wait_event_lock_irq(conf->wait_barrier,
 874                             !conf->array_frozen &&
 875                              !atomic_read(&conf->nr_pending[idx]) &&
 876                              atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH,
 877                             conf->resync_lock);
 878
 879         atomic_inc(&conf->nr_sync_pending);
 880         spin_unlock_irq(&conf->resync_lock);
 881 }
 882
 883 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 884 {
 885         int idx = sector_to_idx(sector_nr);
 886
 887         BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 888
 889         atomic_dec(&conf->barrier[idx]);
 890         atomic_dec(&conf->nr_sync_pending);
 891         wake_up(&conf->wait_barrier);
 892 }
 893
 894 static void _wait_barrier(struct r1conf *conf, int idx)
 895 {
 896         /*
 897          * We need to increase conf->nr_pending[idx] very early here,
 898          * then raise_barrier() can be blocked when it waits for
 899          * conf->nr_pending[idx] to be 0. Then we can avoid holding
 900          * conf->resync_lock when there is no barrier raised in same
 901          * barrier unit bucket. Also if the array is frozen, I/O
 902          * should be blocked until array is unfrozen.
 903          */
 904         atomic_inc(&conf->nr_pending[idx]);
 905         /*
 906          * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
 907          * check conf->barrier[idx]. In raise_barrier() we firstly increase
 908          * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
 909          * barrier is necessary here to make sure conf->barrier[idx] won't be
 910          * fetched before conf->nr_pending[idx] is increased. Otherwise there
 911          * will be a race between _wait_barrier() and raise_barrier().
 912          */
 913         smp_mb__after_atomic();
 914
 915         /*
 916          * Don't worry about checking two atomic_t variables at same time
 917          * here. If during we check conf->barrier[idx], the array is
 918          * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
 919          * 0, it is safe to return and make the I/O continue. Because the
 920          * array is frozen, all I/O returned here will eventually complete
 921          * or be queued, no race will happen. See code comment in
 922          * frozen_array().
 923          */
 924         if (!READ_ONCE(conf->array_frozen) &&
 925             !atomic_read(&conf->barrier[idx]))
 926                 return;
 927
 928         /*
 929          * After holding conf->resync_lock, conf->nr_pending[idx]
 930          * should be decreased before waiting for barrier to drop.
 931          * Otherwise, we may encounter a race condition because
 932          * raise_barrer() might be waiting for conf->nr_pending[idx]
 933          * to be 0 at same time.
 934          */
 935         spin_lock_irq(&conf->resync_lock);
 936         atomic_inc(&conf->nr_waiting[idx]);
 937         atomic_dec(&conf->nr_pending[idx]);
 938         /*
 939          * In case freeze_array() is waiting for
 940          * get_unqueued_pending() == extra
 941          */
 942         wake_up(&conf->wait_barrier);
 943         /* Wait for the barrier in same barrier unit bucket to drop. */
 944         wait_event_lock_irq(conf->wait_barrier,
 945                             !conf->array_frozen &&
 946                              !atomic_read(&conf->barrier[idx]),
 947                             conf->resync_lock);
 948         atomic_inc(&conf->nr_pending[idx]);
 949         atomic_dec(&conf->nr_waiting[idx]);
 950         spin_unlock_irq(&conf->resync_lock);
 951 }
 952
 953 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
 954 {
 955         int idx = sector_to_idx(sector_nr);
 956
 957         /*
 958          * Very similar to _wait_barrier(). The difference is, for read
 959          * I/O we don't need wait for sync I/O, but if the whole array
 960          * is frozen, the read I/O still has to wait until the array is
 961          * unfrozen. Since there is no ordering requirement with
 962          * conf->barrier[idx] here, memory barrier is unnecessary as well.
 963          */
 964         atomic_inc(&conf->nr_pending[idx]);
 965
 966         if (!READ_ONCE(conf->array_frozen))
 967                 return;
 968
 969         spin_lock_irq(&conf->resync_lock);
 970         atomic_inc(&conf->nr_waiting[idx]);
 971         atomic_dec(&conf->nr_pending[idx]);
 972         /*
 973          * In case freeze_array() is waiting for
 974          * get_unqueued_pending() == extra
 975          */
 976         wake_up(&conf->wait_barrier);
 977         /* Wait for array to be unfrozen */
 978         wait_event_lock_irq(conf->wait_barrier,
 979                             !conf->array_frozen,
 980                             conf->resync_lock);
 981         atomic_inc(&conf->nr_pending[idx]);
 982         atomic_dec(&conf->nr_waiting[idx]);
 983         spin_unlock_irq(&conf->resync_lock);
 984 }
 985
 986 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
 987 {
 988         int idx = sector_to_idx(sector_nr);
 989
 990         _wait_barrier(conf, idx);
 991 }
 992
 993 static void wait_all_barriers(struct r1conf *conf)
 994 {
 995         int idx;
 996
 997         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
 998                 _wait_barrier(conf, idx);
 999 }
1000
1001 static void _allow_barrier(struct r1conf *conf, int idx)
1002 {
1003         atomic_dec(&conf->nr_pending[idx]);
1004         wake_up(&conf->wait_barrier);
1005 }
1006
1007 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1008 {
1009         int idx = sector_to_idx(sector_nr);
1010
1011         _allow_barrier(conf, idx);
1012 }
1013
1014 static void allow_all_barriers(struct r1conf *conf)
1015 {
1016         int idx;
1017
1018         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1019                 _allow_barrier(conf, idx);
1020 }
1021
1022 /* conf->resync_lock should be held */
1023 static int get_unqueued_pending(struct r1conf *conf)
1024 {
1025         int idx, ret;
1026
1027         ret = atomic_read(&conf->nr_sync_pending);
1028         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1029                 ret += atomic_read(&conf->nr_pending[idx]) -
1030                         atomic_read(&conf->nr_queued[idx]);
1031
1032         return ret;
1033 }
1034
1035 static void freeze_array(struct r1conf *conf, int extra)
1036 {
1037         /* Stop sync I/O and normal I/O and wait for everything to
1038          * go quiet.
1039          * This is called in two situations:
1040          * 1) management command handlers (reshape, remove disk, quiesce).
1041          * 2) one normal I/O request failed.
1042
1043          * After array_frozen is set to 1, new sync IO will be blocked at
1044          * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1045          * or wait_read_barrier(). The flying I/Os will either complete or be
1046          * queued. When everything goes quite, there are only queued I/Os left.
1047
1048          * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1049          * barrier bucket index which this I/O request hits. When all sync and
1050          * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1051          * of all conf->nr_queued[]. But normal I/O failure is an exception,
1052          * in handle_read_error(), we may call freeze_array() before trying to
1053          * fix the read error. In this case, the error read I/O is not queued,
1054          * so get_unqueued_pending() == 1.
1055          *
1056          * Therefore before this function returns, we need to wait until
1057          * get_unqueued_pendings(conf) gets equal to extra. For
1058          * normal I/O context, extra is 1, in rested situations extra is 0.
1059          */
1060         spin_lock_irq(&conf->resync_lock);
1061         conf->array_frozen = 1;
1062         raid1_log(conf->mddev, "wait freeze");
1063         wait_event_lock_irq_cmd(
1064                 conf->wait_barrier,
1065                 get_unqueued_pending(conf) == extra,
1066                 conf->resync_lock,
1067                 flush_pending_writes(conf));
1068         spin_unlock_irq(&conf->resync_lock);
1069 }
1070 static void unfreeze_array(struct r1conf *conf)
1071 {
1072         /* reverse the effect of the freeze */
1073         spin_lock_irq(&conf->resync_lock);
1074         conf->array_frozen = 0;
1075         spin_unlock_irq(&conf->resync_lock);
1076         wake_up(&conf->wait_barrier);
1077 }
1078
1079 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1080                                            struct bio *bio)
1081 {
1082         int size = bio->bi_iter.bi_size;
1083         unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1084         int i = 0;
1085         struct bio *behind_bio = NULL;
1086
1087         behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1088         if (!behind_bio)
1089                 return;
1090
1091         /* discard op, we don't support writezero/writesame yet */
1092         if (!bio_has_data(bio)) {
1093                 behind_bio->bi_iter.bi_size = size;
1094                 goto skip_copy;
1095         }
1096
1097         while (i < vcnt && size) {
1098                 struct page *page;
1099                 int len = min_t(int, PAGE_SIZE, size);
1100
1101                 page = alloc_page(GFP_NOIO);
1102                 if (unlikely(!page))
1103                         goto free_pages;
1104
1105                 bio_add_page(behind_bio, page, len, 0);
1106
1107                 size -= len;
1108                 i++;
1109         }
1110
1111         bio_copy_data(behind_bio, bio);
1112 skip_copy:
1113         r1_bio->behind_master_bio = behind_bio;;
1114         set_bit(R1BIO_BehindIO, &r1_bio->state);
1115
1116         return;
1117
1118 free_pages:
1119         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1120                  bio->bi_iter.bi_size);
1121         bio_free_pages(behind_bio);
1122         bio_put(behind_bio);
1123 }
1124
1125 struct raid1_plug_cb {
1126         struct blk_plug_cb      cb;
1127         struct bio_list         pending;
1128         int                     pending_cnt;
1129 };
1130
1131 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1132 {
1133         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1134                                                   cb);
1135         struct mddev *mddev = plug->cb.data;
1136         struct r1conf *conf = mddev->private;
1137         struct bio *bio;
1138
1139         if (from_schedule || current->bio_list) {
1140                 spin_lock_irq(&conf->device_lock);
1141                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1142                 conf->pending_count += plug->pending_cnt;
1143                 spin_unlock_irq(&conf->device_lock);
1144                 wake_up(&conf->wait_barrier);
1145                 md_wakeup_thread(mddev->thread);
1146                 kfree(plug);
1147                 return;
1148         }
1149
1150         /* we aren't scheduling, so we can do the write-out directly. */
1151         bio = bio_list_get(&plug->pending);
1152         flush_bio_list(conf, bio);
1153         kfree(plug);
1154 }
1155
1156 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1157 {
1158         r1_bio->master_bio = bio;
1159         r1_bio->sectors = bio_sectors(bio);
1160         r1_bio->state = 0;
1161         r1_bio->mddev = mddev;
1162         r1_bio->sector = bio->bi_iter.bi_sector;
1163 }
1164
1165 static inline struct r1bio *
1166 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1167 {
1168         struct r1conf *conf = mddev->private;
1169         struct r1bio *r1_bio;
1170
1171         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1172         /* Ensure no bio records IO_BLOCKED */
1173         memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1174         init_r1bio(r1_bio, mddev, bio);
1175         return r1_bio;
1176 }
1177
1178 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1179                                int max_read_sectors, struct r1bio *r1_bio)
1180 {
1181         struct r1conf *conf = mddev->private;
1182         struct raid1_info *mirror;
1183         struct bio *read_bio;
1184         struct bitmap *bitmap = mddev->bitmap;
1185         const int op = bio_op(bio);
1186         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1187         int max_sectors;
1188         int rdisk;
1189         bool print_msg = !!r1_bio;
1190         char b[BDEVNAME_SIZE];
1191
1192         /*
1193          * If r1_bio is set, we are blocking the raid1d thread
1194          * so there is a tiny risk of deadlock.  So ask for
1195          * emergency memory if needed.
1196          */
1197         gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1198
1199         if (print_msg) {
1200                 /* Need to get the block device name carefully */
1201                 struct md_rdev *rdev;
1202                 rcu_read_lock();
1203                 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1204                 if (rdev)
1205                         bdevname(rdev->bdev, b);
1206                 else
1207                         strcpy(b, "???");
1208                 rcu_read_unlock();
1209         }
1210
1211         /*
1212          * Still need barrier for READ in case that whole
1213          * array is frozen.
1214          */
1215         wait_read_barrier(conf, bio->bi_iter.bi_sector);
1216
1217         if (!r1_bio)
1218                 r1_bio = alloc_r1bio(mddev, bio);
1219         else
1220                 init_r1bio(r1_bio, mddev, bio);
1221         r1_bio->sectors = max_read_sectors;
1222
1223         /*
1224          * make_request() can abort the operation when read-ahead is being
1225          * used and no empty request is available.
1226          */
1227         rdisk = read_balance(conf, r1_bio, &max_sectors);
1228
1229         if (rdisk < 0) {
1230                 /* couldn't find anywhere to read from */
1231                 if (print_msg) {
1232                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1233                                             mdname(mddev),
1234                                             b,
1235                                             (unsigned long long)r1_bio->sector);
1236                 }
1237                 raid_end_bio_io(r1_bio);
1238                 return;
1239         }
1240         mirror = conf->mirrors + rdisk;
1241
1242         if (print_msg)
1243                 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1244                                     mdname(mddev),
1245                                     (unsigned long long)r1_bio->sector,
1246                                     bdevname(mirror->rdev->bdev, b));
1247
1248         if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1249             bitmap) {
1250                 /*
1251                  * Reading from a write-mostly device must take care not to
1252                  * over-take any writes that are 'behind'
1253                  */
1254                 raid1_log(mddev, "wait behind writes");
1255                 wait_event(bitmap->behind_wait,
1256                            atomic_read(&bitmap->behind_writes) == 0);
1257         }
1258
1259         if (max_sectors < bio_sectors(bio)) {
1260                 struct bio *split = bio_split(bio, max_sectors,
1261                                               gfp, conf->bio_split);
1262                 bio_chain(split, bio);
1263                 generic_make_request(bio);
1264                 bio = split;
1265                 r1_bio->master_bio = bio;
1266                 r1_bio->sectors = max_sectors;
1267         }
1268
1269         r1_bio->read_disk = rdisk;
1270
1271         read_bio = bio_clone_fast(bio, gfp, mddev->bio_set);
1272
1273         r1_bio->bios[rdisk] = read_bio;
1274
1275         read_bio->bi_iter.bi_sector = r1_bio->sector +
1276                 mirror->rdev->data_offset;
1277         read_bio->bi_bdev = mirror->rdev->bdev;
1278         read_bio->bi_end_io = raid1_end_read_request;
1279         bio_set_op_attrs(read_bio, op, do_sync);
1280         if (test_bit(FailFast, &mirror->rdev->flags) &&
1281             test_bit(R1BIO_FailFast, &r1_bio->state))
1282                 read_bio->bi_opf |= MD_FAILFAST;
1283         read_bio->bi_private = r1_bio;
1284
1285         if (mddev->gendisk)
1286                 trace_block_bio_remap(bdev_get_queue(read_bio->bi_bdev),
1287                                       read_bio, disk_devt(mddev->gendisk),
1288                                       r1_bio->sector);
1289
1290         generic_make_request(read_bio);
1291 }
1292
1293 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1294                                 int max_write_sectors)
1295 {
1296         struct r1conf *conf = mddev->private;
1297         struct r1bio *r1_bio;
1298         int i, disks;
1299         struct bitmap *bitmap = mddev->bitmap;
1300         unsigned long flags;
1301         struct md_rdev *blocked_rdev;
1302         struct blk_plug_cb *cb;
1303         struct raid1_plug_cb *plug = NULL;
1304         int first_clone;
1305         int max_sectors;
1306
1307         /*
1308          * Register the new request and wait if the reconstruction
1309          * thread has put up a bar for new requests.
1310          * Continue immediately if no resync is active currently.
1311          */
1312
1313
1314         if ((bio_end_sector(bio) > mddev->suspend_lo &&
1315             bio->bi_iter.bi_sector < mddev->suspend_hi) ||
1316             (mddev_is_clustered(mddev) &&
1317              md_cluster_ops->area_resyncing(mddev, WRITE,
1318                      bio->bi_iter.bi_sector, bio_end_sector(bio)))) {
1319
1320                 /*
1321                  * As the suspend_* range is controlled by userspace, we want
1322                  * an interruptible wait.
1323                  */
1324                 DEFINE_WAIT(w);
1325                 for (;;) {
1326                         sigset_t full, old;
1327                         prepare_to_wait(&conf->wait_barrier,
1328                                         &w, TASK_INTERRUPTIBLE);
1329                         if (bio_end_sector(bio) <= mddev->suspend_lo ||
1330                             bio->bi_iter.bi_sector >= mddev->suspend_hi ||
1331                             (mddev_is_clustered(mddev) &&
1332                              !md_cluster_ops->area_resyncing(mddev, WRITE,
1333                                      bio->bi_iter.bi_sector,
1334                                      bio_end_sector(bio))))
1335                                 break;
1336                         sigfillset(&full);
1337                         sigprocmask(SIG_BLOCK, &full, &old);
1338                         schedule();
1339                         sigprocmask(SIG_SETMASK, &old, NULL);
1340                 }
1341                 finish_wait(&conf->wait_barrier, &w);
1342         }
1343         wait_barrier(conf, bio->bi_iter.bi_sector);
1344
1345         r1_bio = alloc_r1bio(mddev, bio);
1346         r1_bio->sectors = max_write_sectors;
1347
1348         if (conf->pending_count >= max_queued_requests) {
1349                 md_wakeup_thread(mddev->thread);
1350                 raid1_log(mddev, "wait queued");
1351                 wait_event(conf->wait_barrier,
1352                            conf->pending_count < max_queued_requests);
1353         }
1354         /* first select target devices under rcu_lock and
1355          * inc refcount on their rdev.  Record them by setting
1356          * bios[x] to bio
1357          * If there are known/acknowledged bad blocks on any device on
1358          * which we have seen a write error, we want to avoid writing those
1359          * blocks.
1360          * This potentially requires several writes to write around
1361          * the bad blocks.  Each set of writes gets it's own r1bio
1362          * with a set of bios attached.
1363          */
1364
1365         disks = conf->raid_disks * 2;
1366  retry_write:
1367         blocked_rdev = NULL;
1368         rcu_read_lock();
1369         max_sectors = r1_bio->sectors;
1370         for (i = 0;  i < disks; i++) {
1371                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1372                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1373                         atomic_inc(&rdev->nr_pending);
1374                         blocked_rdev = rdev;
1375                         break;
1376                 }
1377                 r1_bio->bios[i] = NULL;
1378                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1379                         if (i < conf->raid_disks)
1380                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1381                         continue;
1382                 }
1383
1384                 atomic_inc(&rdev->nr_pending);
1385                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1386                         sector_t first_bad;
1387                         int bad_sectors;
1388                         int is_bad;
1389
1390                         is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1391                                              &first_bad, &bad_sectors);
1392                         if (is_bad < 0) {
1393                                 /* mustn't write here until the bad block is
1394                                  * acknowledged*/
1395                                 set_bit(BlockedBadBlocks, &rdev->flags);
1396                                 blocked_rdev = rdev;
1397                                 break;
1398                         }
1399                         if (is_bad && first_bad <= r1_bio->sector) {
1400                                 /* Cannot write here at all */
1401                                 bad_sectors -= (r1_bio->sector - first_bad);
1402                                 if (bad_sectors < max_sectors)
1403                                         /* mustn't write more than bad_sectors
1404                                          * to other devices yet
1405                                          */
1406                                         max_sectors = bad_sectors;
1407                                 rdev_dec_pending(rdev, mddev);
1408                                 /* We don't set R1BIO_Degraded as that
1409                                  * only applies if the disk is
1410                                  * missing, so it might be re-added,
1411                                  * and we want to know to recover this
1412                                  * chunk.
1413                                  * In this case the device is here,
1414                                  * and the fact that this chunk is not
1415                                  * in-sync is recorded in the bad
1416                                  * block log
1417                                  */
1418                                 continue;
1419                         }
1420                         if (is_bad) {
1421                                 int good_sectors = first_bad - r1_bio->sector;
1422                                 if (good_sectors < max_sectors)
1423                                         max_sectors = good_sectors;
1424                         }
1425                 }
1426                 r1_bio->bios[i] = bio;
1427         }
1428         rcu_read_unlock();
1429
1430         if (unlikely(blocked_rdev)) {
1431                 /* Wait for this device to become unblocked */
1432                 int j;
1433
1434                 for (j = 0; j < i; j++)
1435                         if (r1_bio->bios[j])
1436                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1437                 r1_bio->state = 0;
1438                 allow_barrier(conf, bio->bi_iter.bi_sector);
1439                 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1440                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1441                 wait_barrier(conf, bio->bi_iter.bi_sector);
1442                 goto retry_write;
1443         }
1444
1445         if (max_sectors < bio_sectors(bio)) {
1446                 struct bio *split = bio_split(bio, max_sectors,
1447                                               GFP_NOIO, conf->bio_split);
1448                 bio_chain(split, bio);
1449                 generic_make_request(bio);
1450                 bio = split;
1451                 r1_bio->master_bio = bio;
1452                 r1_bio->sectors = max_sectors;
1453         }
1454
1455         atomic_set(&r1_bio->remaining, 1);
1456         atomic_set(&r1_bio->behind_remaining, 0);
1457
1458         first_clone = 1;
1459
1460         for (i = 0; i < disks; i++) {
1461                 struct bio *mbio = NULL;
1462                 if (!r1_bio->bios[i])
1463                         continue;
1464
1465
1466                 if (first_clone) {
1467                         /* do behind I/O ?
1468                          * Not if there are too many, or cannot
1469                          * allocate memory, or a reader on WriteMostly
1470                          * is waiting for behind writes to flush */
1471                         if (bitmap &&
1472                             (atomic_read(&bitmap->behind_writes)
1473                              < mddev->bitmap_info.max_write_behind) &&
1474                             !waitqueue_active(&bitmap->behind_wait)) {
1475                                 alloc_behind_master_bio(r1_bio, bio);
1476                         }
1477
1478                         bitmap_startwrite(bitmap, r1_bio->sector,
1479                                           r1_bio->sectors,
1480                                           test_bit(R1BIO_BehindIO,
1481                                                    &r1_bio->state));
1482                         first_clone = 0;
1483                 }
1484
1485                 if (r1_bio->behind_master_bio)
1486                         mbio = bio_clone_fast(r1_bio->behind_master_bio,
1487                                               GFP_NOIO, mddev->bio_set);
1488                 else
1489                         mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1490
1491                 if (r1_bio->behind_master_bio) {
1492                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1493                                 atomic_inc(&r1_bio->behind_remaining);
1494                 }
1495
1496                 r1_bio->bios[i] = mbio;
1497
1498                 mbio->bi_iter.bi_sector = (r1_bio->sector +
1499                                    conf->mirrors[i].rdev->data_offset);
1500                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1501                 mbio->bi_end_io = raid1_end_write_request;
1502                 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1503                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1504                     !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1505                     conf->raid_disks - mddev->degraded > 1)
1506                         mbio->bi_opf |= MD_FAILFAST;
1507                 mbio->bi_private = r1_bio;
1508
1509                 atomic_inc(&r1_bio->remaining);
1510
1511                 if (mddev->gendisk)
1512                         trace_block_bio_remap(bdev_get_queue(mbio->bi_bdev),
1513                                               mbio, disk_devt(mddev->gendisk),
1514                                               r1_bio->sector);
1515                 /* flush_pending_writes() needs access to the rdev so...*/
1516                 mbio->bi_bdev = (void*)conf->mirrors[i].rdev;
1517
1518                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1519                 if (cb)
1520                         plug = container_of(cb, struct raid1_plug_cb, cb);
1521                 else
1522                         plug = NULL;
1523                 if (plug) {
1524                         bio_list_add(&plug->pending, mbio);
1525                         plug->pending_cnt++;
1526                 } else {
1527                         spin_lock_irqsave(&conf->device_lock, flags);
1528                         bio_list_add(&conf->pending_bio_list, mbio);
1529                         conf->pending_count++;
1530                         spin_unlock_irqrestore(&conf->device_lock, flags);
1531                         md_wakeup_thread(mddev->thread);
1532                 }
1533         }
1534
1535         r1_bio_write_done(r1_bio);
1536
1537         /* In case raid1d snuck in to freeze_array */
1538         wake_up(&conf->wait_barrier);
1539 }
1540
1541 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1542 {
1543         sector_t sectors;
1544
1545         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1546                 md_flush_request(mddev, bio);
1547                 return true;
1548         }
1549
1550         /*
1551          * There is a limit to the maximum size, but
1552          * the read/write handler might find a lower limit
1553          * due to bad blocks.  To avoid multiple splits,
1554          * we pass the maximum number of sectors down
1555          * and let the lower level perform the split.
1556          */
1557         sectors = align_to_barrier_unit_end(
1558                 bio->bi_iter.bi_sector, bio_sectors(bio));
1559
1560         if (bio_data_dir(bio) == READ)
1561                 raid1_read_request(mddev, bio, sectors, NULL);
1562         else {
1563                 if (!md_write_start(mddev,bio))
1564                         return false;
1565                 raid1_write_request(mddev, bio, sectors);
1566         }
1567         return true;
1568 }
1569
1570 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1571 {
1572         struct r1conf *conf = mddev->private;
1573         int i;
1574
1575         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1576                    conf->raid_disks - mddev->degraded);
1577         rcu_read_lock();
1578         for (i = 0; i < conf->raid_disks; i++) {
1579                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1580                 seq_printf(seq, "%s",
1581                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1582         }
1583         rcu_read_unlock();
1584         seq_printf(seq, "]");
1585 }
1586
1587 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1588 {
1589         char b[BDEVNAME_SIZE];
1590         struct r1conf *conf = mddev->private;
1591         unsigned long flags;
1592
1593         /*
1594          * If it is not operational, then we have already marked it as dead
1595          * else if it is the last working disks, ignore the error, let the
1596          * next level up know.
1597          * else mark the drive as failed
1598          */
1599         spin_lock_irqsave(&conf->device_lock, flags);
1600         if (test_bit(In_sync, &rdev->flags)
1601             && (conf->raid_disks - mddev->degraded) == 1) {
1602                 /*
1603                  * Don't fail the drive, act as though we were just a
1604                  * normal single drive.
1605                  * However don't try a recovery from this drive as
1606                  * it is very likely to fail.
1607                  */
1608                 conf->recovery_disabled = mddev->recovery_disabled;
1609                 spin_unlock_irqrestore(&conf->device_lock, flags);
1610                 return;
1611         }
1612         set_bit(Blocked, &rdev->flags);
1613         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1614                 mddev->degraded++;
1615                 set_bit(Faulty, &rdev->flags);
1616         } else
1617                 set_bit(Faulty, &rdev->flags);
1618         spin_unlock_irqrestore(&conf->device_lock, flags);
1619         /*
1620          * if recovery is running, make sure it aborts.
1621          */
1622         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1623         set_mask_bits(&mddev->sb_flags, 0,
1624                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1625         pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1626                 "md/raid1:%s: Operation continuing on %d devices.\n",
1627                 mdname(mddev), bdevname(rdev->bdev, b),
1628                 mdname(mddev), conf->raid_disks - mddev->degraded);
1629 }
1630
1631 static void print_conf(struct r1conf *conf)
1632 {
1633         int i;
1634
1635         pr_debug("RAID1 conf printout:\n");
1636         if (!conf) {
1637                 pr_debug("(!conf)\n");
1638                 return;
1639         }
1640         pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1641                  conf->raid_disks);
1642
1643         rcu_read_lock();
1644         for (i = 0; i < conf->raid_disks; i++) {
1645                 char b[BDEVNAME_SIZE];
1646                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1647                 if (rdev)
1648                         pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1649                                  i, !test_bit(In_sync, &rdev->flags),
1650                                  !test_bit(Faulty, &rdev->flags),
1651                                  bdevname(rdev->bdev,b));
1652         }
1653         rcu_read_unlock();
1654 }
1655
1656 static void close_sync(struct r1conf *conf)
1657 {
1658         wait_all_barriers(conf);
1659         allow_all_barriers(conf);
1660
1661         mempool_destroy(conf->r1buf_pool);
1662         conf->r1buf_pool = NULL;
1663 }
1664
1665 static int raid1_spare_active(struct mddev *mddev)
1666 {
1667         int i;
1668         struct r1conf *conf = mddev->private;
1669         int count = 0;
1670         unsigned long flags;
1671
1672         /*
1673          * Find all failed disks within the RAID1 configuration
1674          * and mark them readable.
1675          * Called under mddev lock, so rcu protection not needed.
1676          * device_lock used to avoid races with raid1_end_read_request
1677          * which expects 'In_sync' flags and ->degraded to be consistent.
1678          */
1679         spin_lock_irqsave(&conf->device_lock, flags);
1680         for (i = 0; i < conf->raid_disks; i++) {
1681                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1682                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1683                 if (repl
1684                     && !test_bit(Candidate, &repl->flags)
1685                     && repl->recovery_offset == MaxSector
1686                     && !test_bit(Faulty, &repl->flags)
1687                     && !test_and_set_bit(In_sync, &repl->flags)) {
1688                         /* replacement has just become active */
1689                         if (!rdev ||
1690                             !test_and_clear_bit(In_sync, &rdev->flags))
1691                                 count++;
1692                         if (rdev) {
1693                                 /* Replaced device not technically
1694                                  * faulty, but we need to be sure
1695                                  * it gets removed and never re-added
1696                                  */
1697                                 set_bit(Faulty, &rdev->flags);
1698                                 sysfs_notify_dirent_safe(
1699                                         rdev->sysfs_state);
1700                         }
1701                 }
1702                 if (rdev
1703                     && rdev->recovery_offset == MaxSector
1704                     && !test_bit(Faulty, &rdev->flags)
1705                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1706                         count++;
1707                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1708                 }
1709         }
1710         mddev->degraded -= count;
1711         spin_unlock_irqrestore(&conf->device_lock, flags);
1712
1713         print_conf(conf);
1714         return count;
1715 }
1716
1717 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1718 {
1719         struct r1conf *conf = mddev->private;
1720         int err = -EEXIST;
1721         int mirror = 0;
1722         struct raid1_info *p;
1723         int first = 0;
1724         int last = conf->raid_disks - 1;
1725
1726         if (mddev->recovery_disabled == conf->recovery_disabled)
1727                 return -EBUSY;
1728
1729         if (md_integrity_add_rdev(rdev, mddev))
1730                 return -ENXIO;
1731
1732         if (rdev->raid_disk >= 0)
1733                 first = last = rdev->raid_disk;
1734
1735         /*
1736          * find the disk ... but prefer rdev->saved_raid_disk
1737          * if possible.
1738          */
1739         if (rdev->saved_raid_disk >= 0 &&
1740             rdev->saved_raid_disk >= first &&
1741             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1742                 first = last = rdev->saved_raid_disk;
1743
1744         for (mirror = first; mirror <= last; mirror++) {
1745                 p = conf->mirrors+mirror;
1746                 if (!p->rdev) {
1747
1748                         if (mddev->gendisk)
1749                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1750                                                   rdev->data_offset << 9);
1751
1752                         p->head_position = 0;
1753                         rdev->raid_disk = mirror;
1754                         err = 0;
1755                         /* As all devices are equivalent, we don't need a full recovery
1756                          * if this was recently any drive of the array
1757                          */
1758                         if (rdev->saved_raid_disk < 0)
1759                                 conf->fullsync = 1;
1760                         rcu_assign_pointer(p->rdev, rdev);
1761                         break;
1762                 }
1763                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1764                     p[conf->raid_disks].rdev == NULL) {
1765                         /* Add this device as a replacement */
1766                         clear_bit(In_sync, &rdev->flags);
1767                         set_bit(Replacement, &rdev->flags);
1768                         rdev->raid_disk = mirror;
1769                         err = 0;
1770                         conf->fullsync = 1;
1771                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1772                         break;
1773                 }
1774         }
1775         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1776                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1777         print_conf(conf);
1778         return err;
1779 }
1780
1781 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1782 {
1783         struct r1conf *conf = mddev->private;
1784         int err = 0;
1785         int number = rdev->raid_disk;
1786         struct raid1_info *p = conf->mirrors + number;
1787
1788         if (rdev != p->rdev)
1789                 p = conf->mirrors + conf->raid_disks + number;
1790
1791         print_conf(conf);
1792         if (rdev == p->rdev) {
1793                 if (test_bit(In_sync, &rdev->flags) ||
1794                     atomic_read(&rdev->nr_pending)) {
1795                         err = -EBUSY;
1796                         goto abort;
1797                 }
1798                 /* Only remove non-faulty devices if recovery
1799                  * is not possible.
1800                  */
1801                 if (!test_bit(Faulty, &rdev->flags) &&
1802                     mddev->recovery_disabled != conf->recovery_disabled &&
1803                     mddev->degraded < conf->raid_disks) {
1804                         err = -EBUSY;
1805                         goto abort;
1806                 }
1807                 p->rdev = NULL;
1808                 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1809                         synchronize_rcu();
1810                         if (atomic_read(&rdev->nr_pending)) {
1811                                 /* lost the race, try later */
1812                                 err = -EBUSY;
1813                                 p->rdev = rdev;
1814                                 goto abort;
1815                         }
1816                 }
1817                 if (conf->mirrors[conf->raid_disks + number].rdev) {
1818                         /* We just removed a device that is being replaced.
1819                          * Move down the replacement.  We drain all IO before
1820                          * doing this to avoid confusion.
1821                          */
1822                         struct md_rdev *repl =
1823                                 conf->mirrors[conf->raid_disks + number].rdev;
1824                         freeze_array(conf, 0);
1825                         clear_bit(Replacement, &repl->flags);
1826                         p->rdev = repl;
1827                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1828                         unfreeze_array(conf);
1829                 }
1830
1831                 clear_bit(WantReplacement, &rdev->flags);
1832                 err = md_integrity_register(mddev);
1833         }
1834 abort:
1835
1836         print_conf(conf);
1837         return err;
1838 }
1839
1840 static void end_sync_read(struct bio *bio)
1841 {
1842         struct r1bio *r1_bio = get_resync_r1bio(bio);
1843
1844         update_head_pos(r1_bio->read_disk, r1_bio);
1845
1846         /*
1847          * we have read a block, now it needs to be re-written,
1848          * or re-read if the read failed.
1849          * We don't do much here, just schedule handling by raid1d
1850          */
1851         if (!bio->bi_status)
1852                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1853
1854         if (atomic_dec_and_test(&r1_bio->remaining))
1855                 reschedule_retry(r1_bio);
1856 }
1857
1858 static void end_sync_write(struct bio *bio)
1859 {
1860         int uptodate = !bio->bi_status;
1861         struct r1bio *r1_bio = get_resync_r1bio(bio);
1862         struct mddev *mddev = r1_bio->mddev;
1863         struct r1conf *conf = mddev->private;
1864         sector_t first_bad;
1865         int bad_sectors;
1866         struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1867
1868         if (!uptodate) {
1869                 sector_t sync_blocks = 0;
1870                 sector_t s = r1_bio->sector;
1871                 long sectors_to_go = r1_bio->sectors;
1872                 /* make sure these bits doesn't get cleared. */
1873                 do {
1874                         bitmap_end_sync(mddev->bitmap, s,
1875                                         &sync_blocks, 1);
1876                         s += sync_blocks;
1877                         sectors_to_go -= sync_blocks;
1878                 } while (sectors_to_go > 0);
1879                 set_bit(WriteErrorSeen, &rdev->flags);
1880                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1881                         set_bit(MD_RECOVERY_NEEDED, &
1882                                 mddev->recovery);
1883                 set_bit(R1BIO_WriteError, &r1_bio->state);
1884         } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1885                                &first_bad, &bad_sectors) &&
1886                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1887                                 r1_bio->sector,
1888                                 r1_bio->sectors,
1889                                 &first_bad, &bad_sectors)
1890                 )
1891                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1892
1893         if (atomic_dec_and_test(&r1_bio->remaining)) {
1894                 int s = r1_bio->sectors;
1895                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1896                     test_bit(R1BIO_WriteError, &r1_bio->state))
1897                         reschedule_retry(r1_bio);
1898                 else {
1899                         put_buf(r1_bio);
1900                         md_done_sync(mddev, s, uptodate);
1901                 }
1902         }
1903 }
1904
1905 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1906                             int sectors, struct page *page, int rw)
1907 {
1908         if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1909                 /* success */
1910                 return 1;
1911         if (rw == WRITE) {
1912                 set_bit(WriteErrorSeen, &rdev->flags);
1913                 if (!test_and_set_bit(WantReplacement,
1914                                       &rdev->flags))
1915                         set_bit(MD_RECOVERY_NEEDED, &
1916                                 rdev->mddev->recovery);
1917         }
1918         /* need to record an error - either for the block or the device */
1919         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1920                 md_error(rdev->mddev, rdev);
1921         return 0;
1922 }
1923
1924 static int fix_sync_read_error(struct r1bio *r1_bio)
1925 {
1926         /* Try some synchronous reads of other devices to get
1927          * good data, much like with normal read errors.  Only
1928          * read into the pages we already have so we don't
1929          * need to re-issue the read request.
1930          * We don't need to freeze the array, because being in an
1931          * active sync request, there is no normal IO, and
1932          * no overlapping syncs.
1933          * We don't need to check is_badblock() again as we
1934          * made sure that anything with a bad block in range
1935          * will have bi_end_io clear.
1936          */
1937         struct mddev *mddev = r1_bio->mddev;
1938         struct r1conf *conf = mddev->private;
1939         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1940         struct page **pages = get_resync_pages(bio)->pages;
1941         sector_t sect = r1_bio->sector;
1942         int sectors = r1_bio->sectors;
1943         int idx = 0;
1944         struct md_rdev *rdev;
1945
1946         rdev = conf->mirrors[r1_bio->read_disk].rdev;
1947         if (test_bit(FailFast, &rdev->flags)) {
1948                 /* Don't try recovering from here - just fail it
1949                  * ... unless it is the last working device of course */
1950                 md_error(mddev, rdev);
1951                 if (test_bit(Faulty, &rdev->flags))
1952                         /* Don't try to read from here, but make sure
1953                          * put_buf does it's thing
1954                          */
1955                         bio->bi_end_io = end_sync_write;
1956         }
1957
1958         while(sectors) {
1959                 int s = sectors;
1960                 int d = r1_bio->read_disk;
1961                 int success = 0;
1962                 int start;
1963
1964                 if (s > (PAGE_SIZE>>9))
1965                         s = PAGE_SIZE >> 9;
1966                 do {
1967                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1968                                 /* No rcu protection needed here devices
1969                                  * can only be removed when no resync is
1970                                  * active, and resync is currently active
1971                                  */
1972                                 rdev = conf->mirrors[d].rdev;
1973                                 if (sync_page_io(rdev, sect, s<<9,
1974                                                  pages[idx],
1975                                                  REQ_OP_READ, 0, false)) {
1976                                         success = 1;
1977                                         break;
1978                                 }
1979                         }
1980                         d++;
1981                         if (d == conf->raid_disks * 2)
1982                                 d = 0;
1983                 } while (!success && d != r1_bio->read_disk);
1984
1985                 if (!success) {
1986                         char b[BDEVNAME_SIZE];
1987                         int abort = 0;
1988                         /* Cannot read from anywhere, this block is lost.
1989                          * Record a bad block on each device.  If that doesn't
1990                          * work just disable and interrupt the recovery.
1991                          * Don't fail devices as that won't really help.
1992                          */
1993                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1994                                             mdname(mddev),
1995                                             bdevname(bio->bi_bdev, b),
1996                                             (unsigned long long)r1_bio->sector);
1997                         for (d = 0; d < conf->raid_disks * 2; d++) {
1998                                 rdev = conf->mirrors[d].rdev;
1999                                 if (!rdev || test_bit(Faulty, &rdev->flags))
2000                                         continue;
2001                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
2002                                         abort = 1;
2003                         }
2004                         if (abort) {
2005                                 conf->recovery_disabled =
2006                                         mddev->recovery_disabled;
2007                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2008                                 md_done_sync(mddev, r1_bio->sectors, 0);
2009                                 put_buf(r1_bio);
2010                                 return 0;
2011                         }
2012                         /* Try next page */
2013                         sectors -= s;
2014                         sect += s;
2015                         idx++;
2016                         continue;
2017                 }
2018
2019                 start = d;
2020                 /* write it back and re-read */
2021                 while (d != r1_bio->read_disk) {
2022                         if (d == 0)
2023                                 d = conf->raid_disks * 2;
2024                         d--;
2025                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2026                                 continue;
2027                         rdev = conf->mirrors[d].rdev;
2028                         if (r1_sync_page_io(rdev, sect, s,
2029                                             pages[idx],
2030                                             WRITE) == 0) {
2031                                 r1_bio->bios[d]->bi_end_io = NULL;
2032                                 rdev_dec_pending(rdev, mddev);
2033                         }
2034                 }
2035                 d = start;
2036                 while (d != r1_bio->read_disk) {
2037                         if (d == 0)
2038                                 d = conf->raid_disks * 2;
2039                         d--;
2040                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2041                                 continue;
2042                         rdev = conf->mirrors[d].rdev;
2043                         if (r1_sync_page_io(rdev, sect, s,
2044                                             pages[idx],
2045                                             READ) != 0)
2046                                 atomic_add(s, &rdev->corrected_errors);
2047                 }
2048                 sectors -= s;
2049                 sect += s;
2050                 idx ++;
2051         }
2052         set_bit(R1BIO_Uptodate, &r1_bio->state);
2053         bio->bi_status = 0;
2054         return 1;
2055 }
2056
2057 static void process_checks(struct r1bio *r1_bio)
2058 {
2059         /* We have read all readable devices.  If we haven't
2060          * got the block, then there is no hope left.
2061          * If we have, then we want to do a comparison
2062          * and skip the write if everything is the same.
2063          * If any blocks failed to read, then we need to
2064          * attempt an over-write
2065          */
2066         struct mddev *mddev = r1_bio->mddev;
2067         struct r1conf *conf = mddev->private;
2068         int primary;
2069         int i;
2070         int vcnt;
2071
2072         /* Fix variable parts of all bios */
2073         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2074         for (i = 0; i < conf->raid_disks * 2; i++) {
2075                 blk_status_t status;
2076                 struct bio *b = r1_bio->bios[i];
2077                 struct resync_pages *rp = get_resync_pages(b);
2078                 if (b->bi_end_io != end_sync_read)
2079                         continue;
2080                 /* fixup the bio for reuse, but preserve errno */
2081                 status = b->bi_status;
2082                 bio_reset(b);
2083                 b->bi_status = status;
2084                 b->bi_iter.bi_sector = r1_bio->sector +
2085                         conf->mirrors[i].rdev->data_offset;
2086                 b->bi_bdev = conf->mirrors[i].rdev->bdev;
2087                 b->bi_end_io = end_sync_read;
2088                 rp->raid_bio = r1_bio;
2089                 b->bi_private = rp;
2090
2091                 /* initialize bvec table again */
2092                 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2093         }
2094         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2095                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2096                     !r1_bio->bios[primary]->bi_status) {
2097                         r1_bio->bios[primary]->bi_end_io = NULL;
2098                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2099                         break;
2100                 }
2101         r1_bio->read_disk = primary;
2102         for (i = 0; i < conf->raid_disks * 2; i++) {
2103                 int j;
2104                 struct bio *pbio = r1_bio->bios[primary];
2105                 struct bio *sbio = r1_bio->bios[i];
2106                 blk_status_t status = sbio->bi_status;
2107                 struct page **ppages = get_resync_pages(pbio)->pages;
2108                 struct page **spages = get_resync_pages(sbio)->pages;
2109                 struct bio_vec *bi;
2110                 int page_len[RESYNC_PAGES] = { 0 };
2111
2112                 if (sbio->bi_end_io != end_sync_read)
2113                         continue;
2114                 /* Now we can 'fixup' the error value */
2115                 sbio->bi_status = 0;
2116
2117                 bio_for_each_segment_all(bi, sbio, j)
2118                         page_len[j] = bi->bv_len;
2119
2120                 if (!status) {
2121                         for (j = vcnt; j-- ; ) {
2122                                 if (memcmp(page_address(ppages[j]),
2123                                            page_address(spages[j]),
2124                                            page_len[j]))
2125                                         break;
2126                         }
2127                 } else
2128                         j = 0;
2129                 if (j >= 0)
2130                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2131                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2132                               && !status)) {
2133                         /* No need to write to this device. */
2134                         sbio->bi_end_io = NULL;
2135                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2136                         continue;
2137                 }
2138
2139                 bio_copy_data(sbio, pbio);
2140         }
2141 }
2142
2143 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2144 {
2145         struct r1conf *conf = mddev->private;
2146         int i;
2147         int disks = conf->raid_disks * 2;
2148         struct bio *wbio;
2149
2150         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2151                 /* ouch - failed to read all of that. */
2152                 if (!fix_sync_read_error(r1_bio))
2153                         return;
2154
2155         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2156                 process_checks(r1_bio);
2157
2158         /*
2159          * schedule writes
2160          */
2161         atomic_set(&r1_bio->remaining, 1);
2162         for (i = 0; i < disks ; i++) {
2163                 wbio = r1_bio->bios[i];
2164                 if (wbio->bi_end_io == NULL ||
2165                     (wbio->bi_end_io == end_sync_read &&
2166                      (i == r1_bio->read_disk ||
2167                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2168                         continue;
2169                 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2170                         continue;
2171
2172                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2173                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2174                         wbio->bi_opf |= MD_FAILFAST;
2175
2176                 wbio->bi_end_io = end_sync_write;
2177                 atomic_inc(&r1_bio->remaining);
2178                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2179
2180                 generic_make_request(wbio);
2181         }
2182
2183         if (atomic_dec_and_test(&r1_bio->remaining)) {
2184                 /* if we're here, all write(s) have completed, so clean up */
2185                 int s = r1_bio->sectors;
2186                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2187                     test_bit(R1BIO_WriteError, &r1_bio->state))
2188                         reschedule_retry(r1_bio);
2189                 else {
2190                         put_buf(r1_bio);
2191                         md_done_sync(mddev, s, 1);
2192                 }
2193         }
2194 }
2195
2196 /*
2197  * This is a kernel thread which:
2198  *
2199  *      1.      Retries failed read operations on working mirrors.
2200  *      2.      Updates the raid superblock when problems encounter.
2201  *      3.      Performs writes following reads for array synchronising.
2202  */
2203
2204 static void fix_read_error(struct r1conf *conf, int read_disk,
2205                            sector_t sect, int sectors)
2206 {
2207         struct mddev *mddev = conf->mddev;
2208         while(sectors) {
2209                 int s = sectors;
2210                 int d = read_disk;
2211                 int success = 0;
2212                 int start;
2213                 struct md_rdev *rdev;
2214
2215                 if (s > (PAGE_SIZE>>9))
2216                         s = PAGE_SIZE >> 9;
2217
2218                 do {
2219                         sector_t first_bad;
2220                         int bad_sectors;
2221
2222                         rcu_read_lock();
2223                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2224                         if (rdev &&
2225                             (test_bit(In_sync, &rdev->flags) ||
2226                              (!test_bit(Faulty, &rdev->flags) &&
2227                               rdev->recovery_offset >= sect + s)) &&
2228                             is_badblock(rdev, sect, s,
2229                                         &first_bad, &bad_sectors) == 0) {
2230                                 atomic_inc(&rdev->nr_pending);
2231                                 rcu_read_unlock();
2232                                 if (sync_page_io(rdev, sect, s<<9,
2233                                          conf->tmppage, REQ_OP_READ, 0, false))
2234                                         success = 1;
2235                                 rdev_dec_pending(rdev, mddev);
2236                                 if (success)
2237                                         break;
2238                         } else
2239                                 rcu_read_unlock();
2240                         d++;
2241                         if (d == conf->raid_disks * 2)
2242                                 d = 0;
2243                 } while (!success && d != read_disk);
2244
2245                 if (!success) {
2246                         /* Cannot read from anywhere - mark it bad */
2247                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2248                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2249                                 md_error(mddev, rdev);
2250                         break;
2251                 }
2252                 /* write it back and re-read */
2253                 start = d;
2254                 while (d != read_disk) {
2255                         if (d==0)
2256                                 d = conf->raid_disks * 2;
2257                         d--;
2258                         rcu_read_lock();
2259                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2260                         if (rdev &&
2261                             !test_bit(Faulty, &rdev->flags)) {
2262                                 atomic_inc(&rdev->nr_pending);
2263                                 rcu_read_unlock();
2264                                 r1_sync_page_io(rdev, sect, s,
2265                                                 conf->tmppage, WRITE);
2266                                 rdev_dec_pending(rdev, mddev);
2267                         } else
2268                                 rcu_read_unlock();
2269                 }
2270                 d = start;
2271                 while (d != read_disk) {
2272                         char b[BDEVNAME_SIZE];
2273                         if (d==0)
2274                                 d = conf->raid_disks * 2;
2275                         d--;
2276                         rcu_read_lock();
2277                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2278                         if (rdev &&
2279                             !test_bit(Faulty, &rdev->flags)) {
2280                                 atomic_inc(&rdev->nr_pending);
2281                                 rcu_read_unlock();
2282                                 if (r1_sync_page_io(rdev, sect, s,
2283                                                     conf->tmppage, READ)) {
2284                                         atomic_add(s, &rdev->corrected_errors);
2285                                         pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2286                                                 mdname(mddev), s,
2287                                                 (unsigned long long)(sect +
2288                                                                      rdev->data_offset),
2289                                                 bdevname(rdev->bdev, b));
2290                                 }
2291                                 rdev_dec_pending(rdev, mddev);
2292                         } else
2293                                 rcu_read_unlock();
2294                 }
2295                 sectors -= s;
2296                 sect += s;
2297         }
2298 }
2299
2300 static int narrow_write_error(struct r1bio *r1_bio, int i)
2301 {
2302         struct mddev *mddev = r1_bio->mddev;
2303         struct r1conf *conf = mddev->private;
2304         struct md_rdev *rdev = conf->mirrors[i].rdev;
2305
2306         /* bio has the data to be written to device 'i' where
2307          * we just recently had a write error.
2308          * We repeatedly clone the bio and trim down to one block,
2309          * then try the write.  Where the write fails we record
2310          * a bad block.
2311          * It is conceivable that the bio doesn't exactly align with
2312          * blocks.  We must handle this somehow.
2313          *
2314          * We currently own a reference on the rdev.
2315          */
2316
2317         int block_sectors;
2318         sector_t sector;
2319         int sectors;
2320         int sect_to_write = r1_bio->sectors;
2321         int ok = 1;
2322
2323         if (rdev->badblocks.shift < 0)
2324                 return 0;
2325
2326         block_sectors = roundup(1 << rdev->badblocks.shift,
2327                                 bdev_logical_block_size(rdev->bdev) >> 9);
2328         sector = r1_bio->sector;
2329         sectors = ((sector + block_sectors)
2330                    & ~(sector_t)(block_sectors - 1))
2331                 - sector;
2332
2333         while (sect_to_write) {
2334                 struct bio *wbio;
2335                 if (sectors > sect_to_write)
2336                         sectors = sect_to_write;
2337                 /* Write at 'sector' for 'sectors'*/
2338
2339                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2340                         wbio = bio_clone_fast(r1_bio->behind_master_bio,
2341                                               GFP_NOIO,
2342                                               mddev->bio_set);
2343                 } else {
2344                         wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2345                                               mddev->bio_set);
2346                 }
2347
2348                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2349                 wbio->bi_iter.bi_sector = r1_bio->sector;
2350                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2351
2352                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2353                 wbio->bi_iter.bi_sector += rdev->data_offset;
2354                 wbio->bi_bdev = rdev->bdev;
2355
2356                 if (submit_bio_wait(wbio) < 0)
2357                         /* failure! */
2358                         ok = rdev_set_badblocks(rdev, sector,
2359                                                 sectors, 0)
2360                                 && ok;
2361
2362                 bio_put(wbio);
2363                 sect_to_write -= sectors;
2364                 sector += sectors;
2365                 sectors = block_sectors;
2366         }
2367         return ok;
2368 }
2369
2370 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2371 {
2372         int m;
2373         int s = r1_bio->sectors;
2374         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2375                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2376                 struct bio *bio = r1_bio->bios[m];
2377                 if (bio->bi_end_io == NULL)
2378                         continue;
2379                 if (!bio->bi_status &&
2380                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2381                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2382                 }
2383                 if (bio->bi_status &&
2384                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2385                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2386                                 md_error(conf->mddev, rdev);
2387                 }
2388         }
2389         put_buf(r1_bio);
2390         md_done_sync(conf->mddev, s, 1);
2391 }
2392
2393 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2394 {
2395         int m, idx;
2396         bool fail = false;
2397
2398         for (m = 0; m < conf->raid_disks * 2 ; m++)
2399                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2400                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2401                         rdev_clear_badblocks(rdev,
2402                                              r1_bio->sector,
2403                                              r1_bio->sectors, 0);
2404                         rdev_dec_pending(rdev, conf->mddev);
2405                 } else if (r1_bio->bios[m] != NULL) {
2406                         /* This drive got a write error.  We need to
2407                          * narrow down and record precise write
2408                          * errors.
2409                          */
2410                         fail = true;
2411                         if (!narrow_write_error(r1_bio, m)) {
2412                                 md_error(conf->mddev,
2413                                          conf->mirrors[m].rdev);
2414                                 /* an I/O failed, we can't clear the bitmap */
2415                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2416                         }
2417                         rdev_dec_pending(conf->mirrors[m].rdev,
2418                                          conf->mddev);
2419                 }
2420         if (fail) {
2421                 spin_lock_irq(&conf->device_lock);
2422                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2423                 idx = sector_to_idx(r1_bio->sector);
2424                 atomic_inc(&conf->nr_queued[idx]);
2425                 spin_unlock_irq(&conf->device_lock);
2426                 /*
2427                  * In case freeze_array() is waiting for condition
2428                  * get_unqueued_pending() == extra to be true.
2429                  */
2430                 wake_up(&conf->wait_barrier);
2431                 md_wakeup_thread(conf->mddev->thread);
2432         } else {
2433                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2434                         close_write(r1_bio);
2435                 raid_end_bio_io(r1_bio);
2436         }
2437 }
2438
2439 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2440 {
2441         struct mddev *mddev = conf->mddev;
2442         struct bio *bio;
2443         struct md_rdev *rdev;
2444         dev_t bio_dev;
2445         sector_t bio_sector;
2446
2447         clear_bit(R1BIO_ReadError, &r1_bio->state);
2448         /* we got a read error. Maybe the drive is bad.  Maybe just
2449          * the block and we can fix it.
2450          * We freeze all other IO, and try reading the block from
2451          * other devices.  When we find one, we re-write
2452          * and check it that fixes the read error.
2453          * This is all done synchronously while the array is
2454          * frozen
2455          */
2456
2457         bio = r1_bio->bios[r1_bio->read_disk];
2458         bio_dev = bio->bi_bdev->bd_dev;
2459         bio_sector = conf->mirrors[r1_bio->read_disk].rdev->data_offset + r1_bio->sector;
2460         bio_put(bio);
2461         r1_bio->bios[r1_bio->read_disk] = NULL;
2462
2463         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2464         if (mddev->ro == 0
2465             && !test_bit(FailFast, &rdev->flags)) {
2466                 freeze_array(conf, 1);
2467                 fix_read_error(conf, r1_bio->read_disk,
2468                                r1_bio->sector, r1_bio->sectors);
2469                 unfreeze_array(conf);
2470         } else {
2471                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2472         }
2473
2474         rdev_dec_pending(rdev, conf->mddev);
2475         allow_barrier(conf, r1_bio->sector);
2476         bio = r1_bio->master_bio;
2477
2478         /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2479         r1_bio->state = 0;
2480         raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2481 }
2482
2483 static void raid1d(struct md_thread *thread)
2484 {
2485         struct mddev *mddev = thread->mddev;
2486         struct r1bio *r1_bio;
2487         unsigned long flags;
2488         struct r1conf *conf = mddev->private;
2489         struct list_head *head = &conf->retry_list;
2490         struct blk_plug plug;
2491         int idx;
2492
2493         md_check_recovery(mddev);
2494
2495         if (!list_empty_careful(&conf->bio_end_io_list) &&
2496             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2497                 LIST_HEAD(tmp);
2498                 spin_lock_irqsave(&conf->device_lock, flags);
2499                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2500                         list_splice_init(&conf->bio_end_io_list, &tmp);
2501                 spin_unlock_irqrestore(&conf->device_lock, flags);
2502                 while (!list_empty(&tmp)) {
2503                         r1_bio = list_first_entry(&tmp, struct r1bio,
2504                                                   retry_list);
2505                         list_del(&r1_bio->retry_list);
2506                         idx = sector_to_idx(r1_bio->sector);
2507                         atomic_dec(&conf->nr_queued[idx]);
2508                         if (mddev->degraded)
2509                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2510                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2511                                 close_write(r1_bio);
2512                         raid_end_bio_io(r1_bio);
2513                 }
2514         }
2515
2516         blk_start_plug(&plug);
2517         for (;;) {
2518
2519                 flush_pending_writes(conf);
2520
2521                 spin_lock_irqsave(&conf->device_lock, flags);
2522                 if (list_empty(head)) {
2523                         spin_unlock_irqrestore(&conf->device_lock, flags);
2524                         break;
2525                 }
2526                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2527                 list_del(head->prev);
2528                 idx = sector_to_idx(r1_bio->sector);
2529                 atomic_dec(&conf->nr_queued[idx]);
2530                 spin_unlock_irqrestore(&conf->device_lock, flags);
2531
2532                 mddev = r1_bio->mddev;
2533                 conf = mddev->private;
2534                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2535                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2536                             test_bit(R1BIO_WriteError, &r1_bio->state))
2537                                 handle_sync_write_finished(conf, r1_bio);
2538                         else
2539                                 sync_request_write(mddev, r1_bio);
2540                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2541                            test_bit(R1BIO_WriteError, &r1_bio->state))
2542                         handle_write_finished(conf, r1_bio);
2543                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2544                         handle_read_error(conf, r1_bio);
2545                 else
2546                         WARN_ON_ONCE(1);
2547
2548                 cond_resched();
2549                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2550                         md_check_recovery(mddev);
2551         }
2552         blk_finish_plug(&plug);
2553 }
2554
2555 static int init_resync(struct r1conf *conf)
2556 {
2557         int buffs;
2558
2559         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2560         BUG_ON(conf->r1buf_pool);
2561         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2562                                           conf->poolinfo);
2563         if (!conf->r1buf_pool)
2564                 return -ENOMEM;
2565         return 0;
2566 }
2567
2568 /*
2569  * perform a "sync" on one "block"
2570  *
2571  * We need to make sure that no normal I/O request - particularly write
2572  * requests - conflict with active sync requests.
2573  *
2574  * This is achieved by tracking pending requests and a 'barrier' concept
2575  * that can be installed to exclude normal IO requests.
2576  */
2577
2578 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2579                                    int *skipped)
2580 {
2581         struct r1conf *conf = mddev->private;
2582         struct r1bio *r1_bio;
2583         struct bio *bio;
2584         sector_t max_sector, nr_sectors;
2585         int disk = -1;
2586         int i;
2587         int wonly = -1;
2588         int write_targets = 0, read_targets = 0;
2589         sector_t sync_blocks;
2590         int still_degraded = 0;
2591         int good_sectors = RESYNC_SECTORS;
2592         int min_bad = 0; /* number of sectors that are bad in all devices */
2593         int idx = sector_to_idx(sector_nr);
2594         int page_idx = 0;
2595
2596         if (!conf->r1buf_pool)
2597                 if (init_resync(conf))
2598                         return 0;
2599
2600         max_sector = mddev->dev_sectors;
2601         if (sector_nr >= max_sector) {
2602                 /* If we aborted, we need to abort the
2603                  * sync on the 'current' bitmap chunk (there will
2604                  * only be one in raid1 resync.
2605                  * We can find the current addess in mddev->curr_resync
2606                  */
2607                 if (mddev->curr_resync < max_sector) /* aborted */
2608                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2609                                                 &sync_blocks, 1);
2610                 else /* completed sync */
2611                         conf->fullsync = 0;
2612
2613                 bitmap_close_sync(mddev->bitmap);
2614                 close_sync(conf);
2615
2616                 if (mddev_is_clustered(mddev)) {
2617                         conf->cluster_sync_low = 0;
2618                         conf->cluster_sync_high = 0;
2619                 }
2620                 return 0;
2621         }
2622
2623         if (mddev->bitmap == NULL &&
2624             mddev->recovery_cp == MaxSector &&
2625             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2626             conf->fullsync == 0) {
2627                 *skipped = 1;
2628                 return max_sector - sector_nr;
2629         }
2630         /* before building a request, check if we can skip these blocks..
2631          * This call the bitmap_start_sync doesn't actually record anything
2632          */
2633         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2634             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2635                 /* We can skip this block, and probably several more */
2636                 *skipped = 1;
2637                 return sync_blocks;
2638         }
2639
2640         /*
2641          * If there is non-resync activity waiting for a turn, then let it
2642          * though before starting on this new sync request.
2643          */
2644         if (atomic_read(&conf->nr_waiting[idx]))
2645                 schedule_timeout_uninterruptible(1);
2646
2647         /* we are incrementing sector_nr below. To be safe, we check against
2648          * sector_nr + two times RESYNC_SECTORS
2649          */
2650
2651         bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2652                 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2653         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2654
2655         raise_barrier(conf, sector_nr);
2656
2657         rcu_read_lock();
2658         /*
2659          * If we get a correctably read error during resync or recovery,
2660          * we might want to read from a different device.  So we
2661          * flag all drives that could conceivably be read from for READ,
2662          * and any others (which will be non-In_sync devices) for WRITE.
2663          * If a read fails, we try reading from something else for which READ
2664          * is OK.
2665          */
2666
2667         r1_bio->mddev = mddev;
2668         r1_bio->sector = sector_nr;
2669         r1_bio->state = 0;
2670         set_bit(R1BIO_IsSync, &r1_bio->state);
2671         /* make sure good_sectors won't go across barrier unit boundary */
2672         good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2673
2674         for (i = 0; i < conf->raid_disks * 2; i++) {
2675                 struct md_rdev *rdev;
2676                 bio = r1_bio->bios[i];
2677
2678                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2679                 if (rdev == NULL ||
2680                     test_bit(Faulty, &rdev->flags)) {
2681                         if (i < conf->raid_disks)
2682                                 still_degraded = 1;
2683                 } else if (!test_bit(In_sync, &rdev->flags)) {
2684                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2685                         bio->bi_end_io = end_sync_write;
2686                         write_targets ++;
2687                 } else {
2688                         /* may need to read from here */
2689                         sector_t first_bad = MaxSector;
2690                         int bad_sectors;
2691
2692                         if (is_badblock(rdev, sector_nr, good_sectors,
2693                                         &first_bad, &bad_sectors)) {
2694                                 if (first_bad > sector_nr)
2695                                         good_sectors = first_bad - sector_nr;
2696                                 else {
2697                                         bad_sectors -= (sector_nr - first_bad);
2698                                         if (min_bad == 0 ||
2699                                             min_bad > bad_sectors)
2700                                                 min_bad = bad_sectors;
2701                                 }
2702                         }
2703                         if (sector_nr < first_bad) {
2704                                 if (test_bit(WriteMostly, &rdev->flags)) {
2705                                         if (wonly < 0)
2706                                                 wonly = i;
2707                                 } else {
2708                                         if (disk < 0)
2709                                                 disk = i;
2710                                 }
2711                                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2712                                 bio->bi_end_io = end_sync_read;
2713                                 read_targets++;
2714                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2715                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2716                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2717                                 /*
2718                                  * The device is suitable for reading (InSync),
2719                                  * but has bad block(s) here. Let's try to correct them,
2720                                  * if we are doing resync or repair. Otherwise, leave
2721                                  * this device alone for this sync request.
2722                                  */
2723                                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2724                                 bio->bi_end_io = end_sync_write;
2725                                 write_targets++;
2726                         }
2727                 }
2728                 if (bio->bi_end_io) {
2729                         atomic_inc(&rdev->nr_pending);
2730                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2731                         bio->bi_bdev = rdev->bdev;
2732                         if (test_bit(FailFast, &rdev->flags))
2733                                 bio->bi_opf |= MD_FAILFAST;
2734                 }
2735         }
2736         rcu_read_unlock();
2737         if (disk < 0)
2738                 disk = wonly;
2739         r1_bio->read_disk = disk;
2740
2741         if (read_targets == 0 && min_bad > 0) {
2742                 /* These sectors are bad on all InSync devices, so we
2743                  * need to mark them bad on all write targets
2744                  */
2745                 int ok = 1;
2746                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2747                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2748                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2749                                 ok = rdev_set_badblocks(rdev, sector_nr,
2750                                                         min_bad, 0
2751                                         ) && ok;
2752                         }
2753                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2754                 *skipped = 1;
2755                 put_buf(r1_bio);
2756
2757                 if (!ok) {
2758                         /* Cannot record the badblocks, so need to
2759                          * abort the resync.
2760                          * If there are multiple read targets, could just
2761                          * fail the really bad ones ???
2762                          */
2763                         conf->recovery_disabled = mddev->recovery_disabled;
2764                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2765                         return 0;
2766                 } else
2767                         return min_bad;
2768
2769         }
2770         if (min_bad > 0 && min_bad < good_sectors) {
2771                 /* only resync enough to reach the next bad->good
2772                  * transition */
2773                 good_sectors = min_bad;
2774         }
2775
2776         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2777                 /* extra read targets are also write targets */
2778                 write_targets += read_targets-1;
2779
2780         if (write_targets == 0 || read_targets == 0) {
2781                 /* There is nowhere to write, so all non-sync
2782                  * drives must be failed - so we are finished
2783                  */
2784                 sector_t rv;
2785                 if (min_bad > 0)
2786                         max_sector = sector_nr + min_bad;
2787                 rv = max_sector - sector_nr;
2788                 *skipped = 1;
2789                 put_buf(r1_bio);
2790                 return rv;
2791         }
2792
2793         if (max_sector > mddev->resync_max)
2794                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2795         if (max_sector > sector_nr + good_sectors)
2796                 max_sector = sector_nr + good_sectors;
2797         nr_sectors = 0;
2798         sync_blocks = 0;
2799         do {
2800                 struct page *page;
2801                 int len = PAGE_SIZE;
2802                 if (sector_nr + (len>>9) > max_sector)
2803                         len = (max_sector - sector_nr) << 9;
2804                 if (len == 0)
2805                         break;
2806                 if (sync_blocks == 0) {
2807                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2808                                                &sync_blocks, still_degraded) &&
2809                             !conf->fullsync &&
2810                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2811                                 break;
2812                         if ((len >> 9) > sync_blocks)
2813                                 len = sync_blocks<<9;
2814                 }
2815
2816                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2817                         struct resync_pages *rp;
2818
2819                         bio = r1_bio->bios[i];
2820                         rp = get_resync_pages(bio);
2821                         if (bio->bi_end_io) {
2822                                 page = resync_fetch_page(rp, page_idx);
2823
2824                                 /*
2825                                  * won't fail because the vec table is big
2826                                  * enough to hold all these pages
2827                                  */
2828                                 bio_add_page(bio, page, len, 0);
2829                         }
2830                 }
2831                 nr_sectors += len>>9;
2832                 sector_nr += len>>9;
2833                 sync_blocks -= (len>>9);
2834         } while (++page_idx < RESYNC_PAGES);
2835
2836         r1_bio->sectors = nr_sectors;
2837
2838         if (mddev_is_clustered(mddev) &&
2839                         conf->cluster_sync_high < sector_nr + nr_sectors) {
2840                 conf->cluster_sync_low = mddev->curr_resync_completed;
2841                 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2842                 /* Send resync message */
2843                 md_cluster_ops->resync_info_update(mddev,
2844                                 conf->cluster_sync_low,
2845                                 conf->cluster_sync_high);
2846         }
2847
2848         /* For a user-requested sync, we read all readable devices and do a
2849          * compare
2850          */
2851         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2852                 atomic_set(&r1_bio->remaining, read_targets);
2853                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2854                         bio = r1_bio->bios[i];
2855                         if (bio->bi_end_io == end_sync_read) {
2856                                 read_targets--;
2857                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2858                                 if (read_targets == 1)
2859                                         bio->bi_opf &= ~MD_FAILFAST;
2860                                 generic_make_request(bio);
2861                         }
2862                 }
2863         } else {
2864                 atomic_set(&r1_bio->remaining, 1);
2865                 bio = r1_bio->bios[r1_bio->read_disk];
2866                 md_sync_acct(bio->bi_bdev, nr_sectors);
2867                 if (read_targets == 1)
2868                         bio->bi_opf &= ~MD_FAILFAST;
2869                 generic_make_request(bio);
2870
2871         }
2872         return nr_sectors;
2873 }
2874
2875 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2876 {
2877         if (sectors)
2878                 return sectors;
2879
2880         return mddev->dev_sectors;
2881 }
2882
2883 static struct r1conf *setup_conf(struct mddev *mddev)
2884 {
2885         struct r1conf *conf;
2886         int i;
2887         struct raid1_info *disk;
2888         struct md_rdev *rdev;
2889         int err = -ENOMEM;
2890
2891         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2892         if (!conf)
2893                 goto abort;
2894
2895         conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2896                                    sizeof(atomic_t), GFP_KERNEL);
2897         if (!conf->nr_pending)
2898                 goto abort;
2899
2900         conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2901                                    sizeof(atomic_t), GFP_KERNEL);
2902         if (!conf->nr_waiting)
2903                 goto abort;
2904
2905         conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2906                                   sizeof(atomic_t), GFP_KERNEL);
2907         if (!conf->nr_queued)
2908                 goto abort;
2909
2910         conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2911                                 sizeof(atomic_t), GFP_KERNEL);
2912         if (!conf->barrier)
2913                 goto abort;
2914
2915         conf->mirrors = kzalloc(sizeof(struct raid1_info)
2916                                 * mddev->raid_disks * 2,
2917                                  GFP_KERNEL);
2918         if (!conf->mirrors)
2919                 goto abort;
2920
2921         conf->tmppage = alloc_page(GFP_KERNEL);
2922         if (!conf->tmppage)
2923                 goto abort;
2924
2925         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2926         if (!conf->poolinfo)
2927                 goto abort;
2928         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2929         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2930                                           r1bio_pool_free,
2931                                           conf->poolinfo);
2932         if (!conf->r1bio_pool)
2933                 goto abort;
2934
2935         conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0);
2936         if (!conf->bio_split)
2937                 goto abort;
2938
2939         conf->poolinfo->mddev = mddev;
2940
2941         err = -EINVAL;
2942         spin_lock_init(&conf->device_lock);
2943         rdev_for_each(rdev, mddev) {
2944                 int disk_idx = rdev->raid_disk;
2945                 if (disk_idx >= mddev->raid_disks
2946                     || disk_idx < 0)
2947                         continue;
2948                 if (test_bit(Replacement, &rdev->flags))
2949                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
2950                 else
2951                         disk = conf->mirrors + disk_idx;
2952
2953                 if (disk->rdev)
2954                         goto abort;
2955                 disk->rdev = rdev;
2956                 disk->head_position = 0;
2957                 disk->seq_start = MaxSector;
2958         }
2959         conf->raid_disks = mddev->raid_disks;
2960         conf->mddev = mddev;
2961         INIT_LIST_HEAD(&conf->retry_list);
2962         INIT_LIST_HEAD(&conf->bio_end_io_list);
2963
2964         spin_lock_init(&conf->resync_lock);
2965         init_waitqueue_head(&conf->wait_barrier);
2966
2967         bio_list_init(&conf->pending_bio_list);
2968         conf->pending_count = 0;
2969         conf->recovery_disabled = mddev->recovery_disabled - 1;
2970
2971         err = -EIO;
2972         for (i = 0; i < conf->raid_disks * 2; i++) {
2973
2974                 disk = conf->mirrors + i;
2975
2976                 if (i < conf->raid_disks &&
2977                     disk[conf->raid_disks].rdev) {
2978                         /* This slot has a replacement. */
2979                         if (!disk->rdev) {
2980                                 /* No original, just make the replacement
2981                                  * a recovering spare
2982                                  */
2983                                 disk->rdev =
2984                                         disk[conf->raid_disks].rdev;
2985                                 disk[conf->raid_disks].rdev = NULL;
2986                         } else if (!test_bit(In_sync, &disk->rdev->flags))
2987                                 /* Original is not in_sync - bad */
2988                                 goto abort;
2989                 }
2990
2991                 if (!disk->rdev ||
2992                     !test_bit(In_sync, &disk->rdev->flags)) {
2993                         disk->head_position = 0;
2994                         if (disk->rdev &&
2995                             (disk->rdev->saved_raid_disk < 0))
2996                                 conf->fullsync = 1;
2997                 }
2998         }
2999
3000         err = -ENOMEM;
3001         conf->thread = md_register_thread(raid1d, mddev, "raid1");
3002         if (!conf->thread)
3003                 goto abort;
3004
3005         return conf;
3006
3007  abort:
3008         if (conf) {
3009                 mempool_destroy(conf->r1bio_pool);
3010                 kfree(conf->mirrors);
3011                 safe_put_page(conf->tmppage);
3012                 kfree(conf->poolinfo);
3013                 kfree(conf->nr_pending);
3014                 kfree(conf->nr_waiting);
3015                 kfree(conf->nr_queued);
3016                 kfree(conf->barrier);
3017                 if (conf->bio_split)
3018                         bioset_free(conf->bio_split);
3019                 kfree(conf);
3020         }
3021         return ERR_PTR(err);
3022 }
3023
3024 static void raid1_free(struct mddev *mddev, void *priv);
3025 static int raid1_run(struct mddev *mddev)
3026 {
3027         struct r1conf *conf;
3028         int i;
3029         struct md_rdev *rdev;
3030         int ret;
3031         bool discard_supported = false;
3032
3033         if (mddev->level != 1) {
3034                 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3035                         mdname(mddev), mddev->level);
3036                 return -EIO;
3037         }
3038         if (mddev->reshape_position != MaxSector) {
3039                 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3040                         mdname(mddev));
3041                 return -EIO;
3042         }
3043         if (mddev_init_writes_pending(mddev) < 0)
3044                 return -ENOMEM;
3045         /*
3046          * copy the already verified devices into our private RAID1
3047          * bookkeeping area. [whatever we allocate in run(),
3048          * should be freed in raid1_free()]
3049          */
3050         if (mddev->private == NULL)
3051                 conf = setup_conf(mddev);
3052         else
3053                 conf = mddev->private;
3054
3055         if (IS_ERR(conf))
3056                 return PTR_ERR(conf);
3057
3058         if (mddev->queue) {
3059                 blk_queue_max_write_same_sectors(mddev->queue, 0);
3060                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3061         }
3062
3063         rdev_for_each(rdev, mddev) {
3064                 if (!mddev->gendisk)
3065                         continue;
3066                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3067                                   rdev->data_offset << 9);
3068                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3069                         discard_supported = true;
3070         }
3071
3072         mddev->degraded = 0;
3073         for (i=0; i < conf->raid_disks; i++)
3074                 if (conf->mirrors[i].rdev == NULL ||
3075                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3076                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3077                         mddev->degraded++;
3078
3079         if (conf->raid_disks - mddev->degraded == 1)
3080                 mddev->recovery_cp = MaxSector;
3081
3082         if (mddev->recovery_cp != MaxSector)
3083                 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3084                         mdname(mddev));
3085         pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3086                 mdname(mddev), mddev->raid_disks - mddev->degraded,
3087                 mddev->raid_disks);
3088
3089         /*
3090          * Ok, everything is just fine now
3091          */
3092         mddev->thread = conf->thread;
3093         conf->thread = NULL;
3094         mddev->private = conf;
3095         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3096
3097         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3098
3099         if (mddev->queue) {
3100                 if (discard_supported)
3101                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3102                                                 mddev->queue);
3103                 else
3104                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3105                                                   mddev->queue);
3106         }
3107
3108         ret =  md_integrity_register(mddev);
3109         if (ret) {
3110                 md_unregister_thread(&mddev->thread);
3111                 raid1_free(mddev, conf);
3112         }
3113         return ret;
3114 }
3115
3116 static void raid1_free(struct mddev *mddev, void *priv)
3117 {
3118         struct r1conf *conf = priv;
3119
3120         mempool_destroy(conf->r1bio_pool);
3121         kfree(conf->mirrors);
3122         safe_put_page(conf->tmppage);
3123         kfree(conf->poolinfo);
3124         kfree(conf->nr_pending);
3125         kfree(conf->nr_waiting);
3126         kfree(conf->nr_queued);
3127         kfree(conf->barrier);
3128         if (conf->bio_split)
3129                 bioset_free(conf->bio_split);
3130         kfree(conf);
3131 }
3132
3133 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3134 {
3135         /* no resync is happening, and there is enough space
3136          * on all devices, so we can resize.
3137          * We need to make sure resync covers any new space.
3138          * If the array is shrinking we should possibly wait until
3139          * any io in the removed space completes, but it hardly seems
3140          * worth it.
3141          */
3142         sector_t newsize = raid1_size(mddev, sectors, 0);
3143         if (mddev->external_size &&
3144             mddev->array_sectors > newsize)
3145                 return -EINVAL;
3146         if (mddev->bitmap) {
3147                 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3148                 if (ret)
3149                         return ret;
3150         }
3151         md_set_array_sectors(mddev, newsize);
3152         if (sectors > mddev->dev_sectors &&
3153             mddev->recovery_cp > mddev->dev_sectors) {
3154                 mddev->recovery_cp = mddev->dev_sectors;
3155                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3156         }
3157         mddev->dev_sectors = sectors;
3158         mddev->resync_max_sectors = sectors;
3159         return 0;
3160 }
3161
3162 static int raid1_reshape(struct mddev *mddev)
3163 {
3164         /* We need to:
3165          * 1/ resize the r1bio_pool
3166          * 2/ resize conf->mirrors
3167          *
3168          * We allocate a new r1bio_pool if we can.
3169          * Then raise a device barrier and wait until all IO stops.
3170          * Then resize conf->mirrors and swap in the new r1bio pool.
3171          *
3172          * At the same time, we "pack" the devices so that all the missing
3173          * devices have the higher raid_disk numbers.
3174          */
3175         mempool_t *newpool, *oldpool;
3176         struct pool_info *newpoolinfo;
3177         struct raid1_info *newmirrors;
3178         struct r1conf *conf = mddev->private;
3179         int cnt, raid_disks;
3180         unsigned long flags;
3181         int d, d2;
3182
3183         /* Cannot change chunk_size, layout, or level */
3184         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3185             mddev->layout != mddev->new_layout ||
3186             mddev->level != mddev->new_level) {
3187                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3188                 mddev->new_layout = mddev->layout;
3189                 mddev->new_level = mddev->level;
3190                 return -EINVAL;
3191         }
3192
3193         if (!mddev_is_clustered(mddev))
3194                 md_allow_write(mddev);
3195
3196         raid_disks = mddev->raid_disks + mddev->delta_disks;
3197
3198         if (raid_disks < conf->raid_disks) {
3199                 cnt=0;
3200                 for (d= 0; d < conf->raid_disks; d++)
3201                         if (conf->mirrors[d].rdev)
3202                                 cnt++;
3203                 if (cnt > raid_disks)
3204                         return -EBUSY;
3205         }
3206
3207         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3208         if (!newpoolinfo)
3209                 return -ENOMEM;
3210         newpoolinfo->mddev = mddev;
3211         newpoolinfo->raid_disks = raid_disks * 2;
3212
3213         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3214                                  r1bio_pool_free, newpoolinfo);
3215         if (!newpool) {
3216                 kfree(newpoolinfo);
3217                 return -ENOMEM;
3218         }
3219         newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3220                              GFP_KERNEL);
3221         if (!newmirrors) {
3222                 kfree(newpoolinfo);
3223                 mempool_destroy(newpool);
3224                 return -ENOMEM;
3225         }
3226
3227         freeze_array(conf, 0);
3228
3229         /* ok, everything is stopped */
3230         oldpool = conf->r1bio_pool;
3231         conf->r1bio_pool = newpool;
3232
3233         for (d = d2 = 0; d < conf->raid_disks; d++) {
3234                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3235                 if (rdev && rdev->raid_disk != d2) {
3236                         sysfs_unlink_rdev(mddev, rdev);
3237                         rdev->raid_disk = d2;
3238                         sysfs_unlink_rdev(mddev, rdev);
3239                         if (sysfs_link_rdev(mddev, rdev))
3240                                 pr_warn("md/raid1:%s: cannot register rd%d\n",
3241                                         mdname(mddev), rdev->raid_disk);
3242                 }
3243                 if (rdev)
3244                         newmirrors[d2++].rdev = rdev;
3245         }
3246         kfree(conf->mirrors);
3247         conf->mirrors = newmirrors;
3248         kfree(conf->poolinfo);
3249         conf->poolinfo = newpoolinfo;
3250
3251         spin_lock_irqsave(&conf->device_lock, flags);
3252         mddev->degraded += (raid_disks - conf->raid_disks);
3253         spin_unlock_irqrestore(&conf->device_lock, flags);
3254         conf->raid_disks = mddev->raid_disks = raid_disks;
3255         mddev->delta_disks = 0;
3256
3257         unfreeze_array(conf);
3258
3259         set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3260         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3261         md_wakeup_thread(mddev->thread);
3262
3263         mempool_destroy(oldpool);
3264         return 0;
3265 }
3266
3267 static void raid1_quiesce(struct mddev *mddev, int state)
3268 {
3269         struct r1conf *conf = mddev->private;
3270
3271         switch(state) {
3272         case 2: /* wake for suspend */
3273                 wake_up(&conf->wait_barrier);
3274                 break;
3275         case 1:
3276                 freeze_array(conf, 0);
3277                 break;
3278         case 0:
3279                 unfreeze_array(conf);
3280                 break;
3281         }
3282 }
3283
3284 static void *raid1_takeover(struct mddev *mddev)
3285 {
3286         /* raid1 can take over:
3287          *  raid5 with 2 devices, any layout or chunk size
3288          */
3289         if (mddev->level == 5 && mddev->raid_disks == 2) {
3290                 struct r1conf *conf;
3291                 mddev->new_level = 1;
3292                 mddev->new_layout = 0;
3293                 mddev->new_chunk_sectors = 0;
3294                 conf = setup_conf(mddev);
3295                 if (!IS_ERR(conf)) {
3296                         /* Array must appear to be quiesced */
3297                         conf->array_frozen = 1;
3298                         mddev_clear_unsupported_flags(mddev,
3299                                 UNSUPPORTED_MDDEV_FLAGS);
3300                 }
3301                 return conf;
3302         }
3303         return ERR_PTR(-EINVAL);
3304 }
3305
3306 static struct md_personality raid1_personality =
3307 {
3308         .name           = "raid1",
3309         .level          = 1,
3310         .owner          = THIS_MODULE,
3311         .make_request   = raid1_make_request,
3312         .run            = raid1_run,
3313         .free           = raid1_free,
3314         .status         = raid1_status,
3315         .error_handler  = raid1_error,
3316         .hot_add_disk   = raid1_add_disk,
3317         .hot_remove_disk= raid1_remove_disk,
3318         .spare_active   = raid1_spare_active,
3319         .sync_request   = raid1_sync_request,
3320         .resize         = raid1_resize,
3321         .size           = raid1_size,
3322         .check_reshape  = raid1_reshape,
3323         .quiesce        = raid1_quiesce,
3324         .takeover       = raid1_takeover,
3325         .congested      = raid1_congested,
3326 };
3327
3328 static int __init raid_init(void)
3329 {
3330         return register_md_personality(&raid1_personality);
3331 }
3332
3333 static void raid_exit(void)
3334 {
3335         unregister_md_personality(&raid1_personality);
3336 }
3337
3338 module_init(raid_init);
3339 module_exit(raid_exit);
3340 MODULE_LICENSE("GPL");
3341 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3342 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3343 MODULE_ALIAS("md-raid1");
3344 MODULE_ALIAS("md-level-1");
3345
3346 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);