drivers/md/raid1.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * raid1.c : Multiple Devices driver for Linux
   4  *
   5  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   6  *
   7  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   8  *
   9  * RAID-1 management functions.
  10  *
  11  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  12  *
  13  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  14  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  15  *
  16  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  17  * bitmapped intelligence in resync:
  18  *
  19  *      - bitmap marked during normal i/o
  20  *      - bitmap used to skip nondirty blocks during sync
  21  *
  22  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  23  * - persistent bitmap code
  24  */
  25
  26 #include <linux/slab.h>
  27 #include <linux/delay.h>
  28 #include <linux/blkdev.h>
  29 #include <linux/module.h>
  30 #include <linux/seq_file.h>
  31 #include <linux/ratelimit.h>
  32 #include <linux/interval_tree_generic.h>
  33
  34 #include <trace/events/block.h>
  35
  36 #include "md.h"
  37 #include "raid1.h"
  38 #include "md-bitmap.h"
  39
  40 #define UNSUPPORTED_MDDEV_FLAGS         \
  41         ((1L << MD_HAS_JOURNAL) |       \
  42          (1L << MD_JOURNAL_CLEAN) |     \
  43          (1L << MD_HAS_PPL) |           \
  44          (1L << MD_HAS_MULTIPLE_PPLS))
  45
  46 static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
  47 static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
  48
  49 #define raid1_log(md, fmt, args...)                             \
  50         do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
  51
  52 #include "raid1-10.c"
  53
  54 #define START(node) ((node)->start)
  55 #define LAST(node) ((node)->last)
  56 INTERVAL_TREE_DEFINE(struct serial_info, node, sector_t, _subtree_last,
  57                      START, LAST, static inline, raid1_rb);
  58
  59 static int check_and_add_serial(struct md_rdev *rdev, struct r1bio *r1_bio,
  60                                 struct serial_info *si, int idx)
  61 {
  62         unsigned long flags;
  63         int ret = 0;
  64         sector_t lo = r1_bio->sector;
  65         sector_t hi = lo + r1_bio->sectors;
  66         struct serial_in_rdev *serial = &rdev->serial[idx];
  67
  68         spin_lock_irqsave(&serial->serial_lock, flags);
  69         /* collision happened */
  70         if (raid1_rb_iter_first(&serial->serial_rb, lo, hi))
  71                 ret = -EBUSY;
  72         else {
  73                 si->start = lo;
  74                 si->last = hi;
  75                 raid1_rb_insert(si, &serial->serial_rb);
  76         }
  77         spin_unlock_irqrestore(&serial->serial_lock, flags);
  78
  79         return ret;
  80 }
  81
  82 static void wait_for_serialization(struct md_rdev *rdev, struct r1bio *r1_bio)
  83 {
  84         struct mddev *mddev = rdev->mddev;
  85         struct serial_info *si;
  86         int idx = sector_to_idx(r1_bio->sector);
  87         struct serial_in_rdev *serial = &rdev->serial[idx];
  88
  89         if (WARN_ON(!mddev->serial_info_pool))
  90                 return;
  91         si = mempool_alloc(mddev->serial_info_pool, GFP_NOIO);
  92         wait_event(serial->serial_io_wait,
  93                    check_and_add_serial(rdev, r1_bio, si, idx) == 0);
  94 }
  95
  96 static void remove_serial(struct md_rdev *rdev, sector_t lo, sector_t hi)
  97 {
  98         struct serial_info *si;
  99         unsigned long flags;
 100         int found = 0;
 101         struct mddev *mddev = rdev->mddev;
 102         int idx = sector_to_idx(lo);
 103         struct serial_in_rdev *serial = &rdev->serial[idx];
 104
 105         spin_lock_irqsave(&serial->serial_lock, flags);
 106         for (si = raid1_rb_iter_first(&serial->serial_rb, lo, hi);
 107              si; si = raid1_rb_iter_next(si, lo, hi)) {
 108                 if (si->start == lo && si->last == hi) {
 109                         raid1_rb_remove(si, &serial->serial_rb);
 110                         mempool_free(si, mddev->serial_info_pool);
 111                         found = 1;
 112                         break;
 113                 }
 114         }
 115         if (!found)
 116                 WARN(1, "The write IO is not recorded for serialization\n");
 117         spin_unlock_irqrestore(&serial->serial_lock, flags);
 118         wake_up(&serial->serial_io_wait);
 119 }
 120
 121 /*
 122  * for resync bio, r1bio pointer can be retrieved from the per-bio
 123  * 'struct resync_pages'.
 124  */
 125 static inline struct r1bio *get_resync_r1bio(struct bio *bio)
 126 {
 127         return get_resync_pages(bio)->raid_bio;
 128 }
 129
 130 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
 131 {
 132         struct pool_info *pi = data;
 133         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
 134
 135         /* allocate a r1bio with room for raid_disks entries in the bios array */
 136         return kzalloc(size, gfp_flags);
 137 }
 138
 139 #define RESYNC_DEPTH 32
 140 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
 141 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
 142 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
 143 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
 144 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
 145
 146 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
 147 {
 148         struct pool_info *pi = data;
 149         struct r1bio *r1_bio;
 150         struct bio *bio;
 151         int need_pages;
 152         int j;
 153         struct resync_pages *rps;
 154
 155         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 156         if (!r1_bio)
 157                 return NULL;
 158
 159         rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
 160                             gfp_flags);
 161         if (!rps)
 162                 goto out_free_r1bio;
 163
 164         /*
 165          * Allocate bios : 1 for reading, n-1 for writing
 166          */
 167         for (j = pi->raid_disks ; j-- ; ) {
 168                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 169                 if (!bio)
 170                         goto out_free_bio;
 171                 r1_bio->bios[j] = bio;
 172         }
 173         /*
 174          * Allocate RESYNC_PAGES data pages and attach them to
 175          * the first bio.
 176          * If this is a user-requested check/repair, allocate
 177          * RESYNC_PAGES for each bio.
 178          */
 179         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 180                 need_pages = pi->raid_disks;
 181         else
 182                 need_pages = 1;
 183         for (j = 0; j < pi->raid_disks; j++) {
 184                 struct resync_pages *rp = &rps[j];
 185
 186                 bio = r1_bio->bios[j];
 187
 188                 if (j < need_pages) {
 189                         if (resync_alloc_pages(rp, gfp_flags))
 190                                 goto out_free_pages;
 191                 } else {
 192                         memcpy(rp, &rps[0], sizeof(*rp));
 193                         resync_get_all_pages(rp);
 194                 }
 195
 196                 rp->raid_bio = r1_bio;
 197                 bio->bi_private = rp;
 198         }
 199
 200         r1_bio->master_bio = NULL;
 201
 202         return r1_bio;
 203
 204 out_free_pages:
 205         while (--j >= 0)
 206                 resync_free_pages(&rps[j]);
 207
 208 out_free_bio:
 209         while (++j < pi->raid_disks)
 210                 bio_put(r1_bio->bios[j]);
 211         kfree(rps);
 212
 213 out_free_r1bio:
 214         rbio_pool_free(r1_bio, data);
 215         return NULL;
 216 }
 217
 218 static void r1buf_pool_free(void *__r1_bio, void *data)
 219 {
 220         struct pool_info *pi = data;
 221         int i;
 222         struct r1bio *r1bio = __r1_bio;
 223         struct resync_pages *rp = NULL;
 224
 225         for (i = pi->raid_disks; i--; ) {
 226                 rp = get_resync_pages(r1bio->bios[i]);
 227                 resync_free_pages(rp);
 228                 bio_put(r1bio->bios[i]);
 229         }
 230
 231         /* resync pages array stored in the 1st bio's .bi_private */
 232         kfree(rp);
 233
 234         rbio_pool_free(r1bio, data);
 235 }
 236
 237 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 238 {
 239         int i;
 240
 241         for (i = 0; i < conf->raid_disks * 2; i++) {
 242                 struct bio **bio = r1_bio->bios + i;
 243                 if (!BIO_SPECIAL(*bio))
 244                         bio_put(*bio);
 245                 *bio = NULL;
 246         }
 247 }
 248
 249 static void free_r1bio(struct r1bio *r1_bio)
 250 {
 251         struct r1conf *conf = r1_bio->mddev->private;
 252
 253         put_all_bios(conf, r1_bio);
 254         mempool_free(r1_bio, &conf->r1bio_pool);
 255 }
 256
 257 static void put_buf(struct r1bio *r1_bio)
 258 {
 259         struct r1conf *conf = r1_bio->mddev->private;
 260         sector_t sect = r1_bio->sector;
 261         int i;
 262
 263         for (i = 0; i < conf->raid_disks * 2; i++) {
 264                 struct bio *bio = r1_bio->bios[i];
 265                 if (bio->bi_end_io)
 266                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 267         }
 268
 269         mempool_free(r1_bio, &conf->r1buf_pool);
 270
 271         lower_barrier(conf, sect);
 272 }
 273
 274 static void reschedule_retry(struct r1bio *r1_bio)
 275 {
 276         unsigned long flags;
 277         struct mddev *mddev = r1_bio->mddev;
 278         struct r1conf *conf = mddev->private;
 279         int idx;
 280
 281         idx = sector_to_idx(r1_bio->sector);
 282         spin_lock_irqsave(&conf->device_lock, flags);
 283         list_add(&r1_bio->retry_list, &conf->retry_list);
 284         atomic_inc(&conf->nr_queued[idx]);
 285         spin_unlock_irqrestore(&conf->device_lock, flags);
 286
 287         wake_up(&conf->wait_barrier);
 288         md_wakeup_thread(mddev->thread);
 289 }
 290
 291 /*
 292  * raid_end_bio_io() is called when we have finished servicing a mirrored
 293  * operation and are ready to return a success/failure code to the buffer
 294  * cache layer.
 295  */
 296 static void call_bio_endio(struct r1bio *r1_bio)
 297 {
 298         struct bio *bio = r1_bio->master_bio;
 299         struct r1conf *conf = r1_bio->mddev->private;
 300
 301         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 302                 bio->bi_status = BLK_STS_IOERR;
 303
 304         bio_endio(bio);
 305         /*
 306          * Wake up any possible resync thread that waits for the device
 307          * to go idle.
 308          */
 309         allow_barrier(conf, r1_bio->sector);
 310 }
 311
 312 static void raid_end_bio_io(struct r1bio *r1_bio)
 313 {
 314         struct bio *bio = r1_bio->master_bio;
 315
 316         /* if nobody has done the final endio yet, do it now */
 317         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 318                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 319                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 320                          (unsigned long long) bio->bi_iter.bi_sector,
 321                          (unsigned long long) bio_end_sector(bio) - 1);
 322
 323                 call_bio_endio(r1_bio);
 324         }
 325         free_r1bio(r1_bio);
 326 }
 327
 328 /*
 329  * Update disk head position estimator based on IRQ completion info.
 330  */
 331 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 332 {
 333         struct r1conf *conf = r1_bio->mddev->private;
 334
 335         conf->mirrors[disk].head_position =
 336                 r1_bio->sector + (r1_bio->sectors);
 337 }
 338
 339 /*
 340  * Find the disk number which triggered given bio
 341  */
 342 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 343 {
 344         int mirror;
 345         struct r1conf *conf = r1_bio->mddev->private;
 346         int raid_disks = conf->raid_disks;
 347
 348         for (mirror = 0; mirror < raid_disks * 2; mirror++)
 349                 if (r1_bio->bios[mirror] == bio)
 350                         break;
 351
 352         BUG_ON(mirror == raid_disks * 2);
 353         update_head_pos(mirror, r1_bio);
 354
 355         return mirror;
 356 }
 357
 358 static void raid1_end_read_request(struct bio *bio)
 359 {
 360         int uptodate = !bio->bi_status;
 361         struct r1bio *r1_bio = bio->bi_private;
 362         struct r1conf *conf = r1_bio->mddev->private;
 363         struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 364
 365         /*
 366          * this branch is our 'one mirror IO has finished' event handler:
 367          */
 368         update_head_pos(r1_bio->read_disk, r1_bio);
 369
 370         if (uptodate)
 371                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 372         else if (test_bit(FailFast, &rdev->flags) &&
 373                  test_bit(R1BIO_FailFast, &r1_bio->state))
 374                 /* This was a fail-fast read so we definitely
 375                  * want to retry */
 376                 ;
 377         else {
 378                 /* If all other devices have failed, we want to return
 379                  * the error upwards rather than fail the last device.
 380                  * Here we redefine "uptodate" to mean "Don't want to retry"
 381                  */
 382                 unsigned long flags;
 383                 spin_lock_irqsave(&conf->device_lock, flags);
 384                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 385                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 386                      test_bit(In_sync, &rdev->flags)))
 387                         uptodate = 1;
 388                 spin_unlock_irqrestore(&conf->device_lock, flags);
 389         }
 390
 391         if (uptodate) {
 392                 raid_end_bio_io(r1_bio);
 393                 rdev_dec_pending(rdev, conf->mddev);
 394         } else {
 395                 /*
 396                  * oops, read error:
 397                  */
 398                 char b[BDEVNAME_SIZE];
 399                 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
 400                                    mdname(conf->mddev),
 401                                    bdevname(rdev->bdev, b),
 402                                    (unsigned long long)r1_bio->sector);
 403                 set_bit(R1BIO_ReadError, &r1_bio->state);
 404                 reschedule_retry(r1_bio);
 405                 /* don't drop the reference on read_disk yet */
 406         }
 407 }
 408
 409 static void close_write(struct r1bio *r1_bio)
 410 {
 411         /* it really is the end of this request */
 412         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 413                 bio_free_pages(r1_bio->behind_master_bio);
 414                 bio_put(r1_bio->behind_master_bio);
 415                 r1_bio->behind_master_bio = NULL;
 416         }
 417         /* clear the bitmap if all writes complete successfully */
 418         md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 419                            r1_bio->sectors,
 420                            !test_bit(R1BIO_Degraded, &r1_bio->state),
 421                            test_bit(R1BIO_BehindIO, &r1_bio->state));
 422         md_write_end(r1_bio->mddev);
 423 }
 424
 425 static void r1_bio_write_done(struct r1bio *r1_bio)
 426 {
 427         if (!atomic_dec_and_test(&r1_bio->remaining))
 428                 return;
 429
 430         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 431                 reschedule_retry(r1_bio);
 432         else {
 433                 close_write(r1_bio);
 434                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 435                         reschedule_retry(r1_bio);
 436                 else
 437                         raid_end_bio_io(r1_bio);
 438         }
 439 }
 440
 441 static void raid1_end_write_request(struct bio *bio)
 442 {
 443         struct r1bio *r1_bio = bio->bi_private;
 444         int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 445         struct r1conf *conf = r1_bio->mddev->private;
 446         struct bio *to_put = NULL;
 447         int mirror = find_bio_disk(r1_bio, bio);
 448         struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 449         bool discard_error;
 450         sector_t lo = r1_bio->sector;
 451         sector_t hi = r1_bio->sector + r1_bio->sectors;
 452
 453         discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 454
 455         /*
 456          * 'one mirror IO has finished' event handler:
 457          */
 458         if (bio->bi_status && !discard_error) {
 459                 set_bit(WriteErrorSeen, &rdev->flags);
 460                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
 461                         set_bit(MD_RECOVERY_NEEDED, &
 462                                 conf->mddev->recovery);
 463
 464                 if (test_bit(FailFast, &rdev->flags) &&
 465                     (bio->bi_opf & MD_FAILFAST) &&
 466                     /* We never try FailFast to WriteMostly devices */
 467                     !test_bit(WriteMostly, &rdev->flags)) {
 468                         md_error(r1_bio->mddev, rdev);
 469                 }
 470
 471                 /*
 472                  * When the device is faulty, it is not necessary to
 473                  * handle write error.
 474                  * For failfast, this is the only remaining device,
 475                  * We need to retry the write without FailFast.
 476                  */
 477                 if (!test_bit(Faulty, &rdev->flags))
 478                         set_bit(R1BIO_WriteError, &r1_bio->state);
 479                 else {
 480                         /* Finished with this branch */
 481                         r1_bio->bios[mirror] = NULL;
 482                         to_put = bio;
 483                 }
 484         } else {
 485                 /*
 486                  * Set R1BIO_Uptodate in our master bio, so that we
 487                  * will return a good error code for to the higher
 488                  * levels even if IO on some other mirrored buffer
 489                  * fails.
 490                  *
 491                  * The 'master' represents the composite IO operation
 492                  * to user-side. So if something waits for IO, then it
 493                  * will wait for the 'master' bio.
 494                  */
 495                 sector_t first_bad;
 496                 int bad_sectors;
 497
 498                 r1_bio->bios[mirror] = NULL;
 499                 to_put = bio;
 500                 /*
 501                  * Do not set R1BIO_Uptodate if the current device is
 502                  * rebuilding or Faulty. This is because we cannot use
 503                  * such device for properly reading the data back (we could
 504                  * potentially use it, if the current write would have felt
 505                  * before rdev->recovery_offset, but for simplicity we don't
 506                  * check this here.
 507                  */
 508                 if (test_bit(In_sync, &rdev->flags) &&
 509                     !test_bit(Faulty, &rdev->flags))
 510                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 511
 512                 /* Maybe we can clear some bad blocks. */
 513                 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 514                                 &first_bad, &bad_sectors) && !discard_error) {
 515                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 516                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 517                 }
 518         }
 519
 520         if (behind) {
 521                 if (test_bit(CollisionCheck, &rdev->flags))
 522                         remove_serial(rdev, lo, hi);
 523                 if (test_bit(WriteMostly, &rdev->flags))
 524                         atomic_dec(&r1_bio->behind_remaining);
 525
 526                 /*
 527                  * In behind mode, we ACK the master bio once the I/O
 528                  * has safely reached all non-writemostly
 529                  * disks. Setting the Returned bit ensures that this
 530                  * gets done only once -- we don't ever want to return
 531                  * -EIO here, instead we'll wait
 532                  */
 533                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 534                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 535                         /* Maybe we can return now */
 536                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 537                                 struct bio *mbio = r1_bio->master_bio;
 538                                 pr_debug("raid1: behind end write sectors"
 539                                          " %llu-%llu\n",
 540                                          (unsigned long long) mbio->bi_iter.bi_sector,
 541                                          (unsigned long long) bio_end_sector(mbio) - 1);
 542                                 call_bio_endio(r1_bio);
 543                         }
 544                 }
 545         } else if (rdev->mddev->serialize_policy)
 546                 remove_serial(rdev, lo, hi);
 547         if (r1_bio->bios[mirror] == NULL)
 548                 rdev_dec_pending(rdev, conf->mddev);
 549
 550         /*
 551          * Let's see if all mirrored write operations have finished
 552          * already.
 553          */
 554         r1_bio_write_done(r1_bio);
 555
 556         if (to_put)
 557                 bio_put(to_put);
 558 }
 559
 560 static sector_t align_to_barrier_unit_end(sector_t start_sector,
 561                                           sector_t sectors)
 562 {
 563         sector_t len;
 564
 565         WARN_ON(sectors == 0);
 566         /*
 567          * len is the number of sectors from start_sector to end of the
 568          * barrier unit which start_sector belongs to.
 569          */
 570         len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 571               start_sector;
 572
 573         if (len > sectors)
 574                 len = sectors;
 575
 576         return len;
 577 }
 578
 579 /*
 580  * This routine returns the disk from which the requested read should
 581  * be done. There is a per-array 'next expected sequential IO' sector
 582  * number - if this matches on the next IO then we use the last disk.
 583  * There is also a per-disk 'last know head position' sector that is
 584  * maintained from IRQ contexts, both the normal and the resync IO
 585  * completion handlers update this position correctly. If there is no
 586  * perfect sequential match then we pick the disk whose head is closest.
 587  *
 588  * If there are 2 mirrors in the same 2 devices, performance degrades
 589  * because position is mirror, not device based.
 590  *
 591  * The rdev for the device selected will have nr_pending incremented.
 592  */
 593 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 594 {
 595         const sector_t this_sector = r1_bio->sector;
 596         int sectors;
 597         int best_good_sectors;
 598         int best_disk, best_dist_disk, best_pending_disk;
 599         int has_nonrot_disk;
 600         int disk;
 601         sector_t best_dist;
 602         unsigned int min_pending;
 603         struct md_rdev *rdev;
 604         int choose_first;
 605         int choose_next_idle;
 606
 607         rcu_read_lock();
 608         /*
 609          * Check if we can balance. We can balance on the whole
 610          * device if no resync is going on, or below the resync window.
 611          * We take the first readable disk when above the resync window.
 612          */
 613  retry:
 614         sectors = r1_bio->sectors;
 615         best_disk = -1;
 616         best_dist_disk = -1;
 617         best_dist = MaxSector;
 618         best_pending_disk = -1;
 619         min_pending = UINT_MAX;
 620         best_good_sectors = 0;
 621         has_nonrot_disk = 0;
 622         choose_next_idle = 0;
 623         clear_bit(R1BIO_FailFast, &r1_bio->state);
 624
 625         if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 626             (mddev_is_clustered(conf->mddev) &&
 627             md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 628                     this_sector + sectors)))
 629                 choose_first = 1;
 630         else
 631                 choose_first = 0;
 632
 633         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 634                 sector_t dist;
 635                 sector_t first_bad;
 636                 int bad_sectors;
 637                 unsigned int pending;
 638                 bool nonrot;
 639
 640                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 641                 if (r1_bio->bios[disk] == IO_BLOCKED
 642                     || rdev == NULL
 643                     || test_bit(Faulty, &rdev->flags))
 644                         continue;
 645                 if (!test_bit(In_sync, &rdev->flags) &&
 646                     rdev->recovery_offset < this_sector + sectors)
 647                         continue;
 648                 if (test_bit(WriteMostly, &rdev->flags)) {
 649                         /* Don't balance among write-mostly, just
 650                          * use the first as a last resort */
 651                         if (best_dist_disk < 0) {
 652                                 if (is_badblock(rdev, this_sector, sectors,
 653                                                 &first_bad, &bad_sectors)) {
 654                                         if (first_bad <= this_sector)
 655                                                 /* Cannot use this */
 656                                                 continue;
 657                                         best_good_sectors = first_bad - this_sector;
 658                                 } else
 659                                         best_good_sectors = sectors;
 660                                 best_dist_disk = disk;
 661                                 best_pending_disk = disk;
 662                         }
 663                         continue;
 664                 }
 665                 /* This is a reasonable device to use.  It might
 666                  * even be best.
 667                  */
 668                 if (is_badblock(rdev, this_sector, sectors,
 669                                 &first_bad, &bad_sectors)) {
 670                         if (best_dist < MaxSector)
 671                                 /* already have a better device */
 672                                 continue;
 673                         if (first_bad <= this_sector) {
 674                                 /* cannot read here. If this is the 'primary'
 675                                  * device, then we must not read beyond
 676                                  * bad_sectors from another device..
 677                                  */
 678                                 bad_sectors -= (this_sector - first_bad);
 679                                 if (choose_first && sectors > bad_sectors)
 680                                         sectors = bad_sectors;
 681                                 if (best_good_sectors > sectors)
 682                                         best_good_sectors = sectors;
 683
 684                         } else {
 685                                 sector_t good_sectors = first_bad - this_sector;
 686                                 if (good_sectors > best_good_sectors) {
 687                                         best_good_sectors = good_sectors;
 688                                         best_disk = disk;
 689                                 }
 690                                 if (choose_first)
 691                                         break;
 692                         }
 693                         continue;
 694                 } else {
 695                         if ((sectors > best_good_sectors) && (best_disk >= 0))
 696                                 best_disk = -1;
 697                         best_good_sectors = sectors;
 698                 }
 699
 700                 if (best_disk >= 0)
 701                         /* At least two disks to choose from so failfast is OK */
 702                         set_bit(R1BIO_FailFast, &r1_bio->state);
 703
 704                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 705                 has_nonrot_disk |= nonrot;
 706                 pending = atomic_read(&rdev->nr_pending);
 707                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 708                 if (choose_first) {
 709                         best_disk = disk;
 710                         break;
 711                 }
 712                 /* Don't change to another disk for sequential reads */
 713                 if (conf->mirrors[disk].next_seq_sect == this_sector
 714                     || dist == 0) {
 715                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 716                         struct raid1_info *mirror = &conf->mirrors[disk];
 717
 718                         best_disk = disk;
 719                         /*
 720                          * If buffered sequential IO size exceeds optimal
 721                          * iosize, check if there is idle disk. If yes, choose
 722                          * the idle disk. read_balance could already choose an
 723                          * idle disk before noticing it's a sequential IO in
 724                          * this disk. This doesn't matter because this disk
 725                          * will idle, next time it will be utilized after the
 726                          * first disk has IO size exceeds optimal iosize. In
 727                          * this way, iosize of the first disk will be optimal
 728                          * iosize at least. iosize of the second disk might be
 729                          * small, but not a big deal since when the second disk
 730                          * starts IO, the first disk is likely still busy.
 731                          */
 732                         if (nonrot && opt_iosize > 0 &&
 733                             mirror->seq_start != MaxSector &&
 734                             mirror->next_seq_sect > opt_iosize &&
 735                             mirror->next_seq_sect - opt_iosize >=
 736                             mirror->seq_start) {
 737                                 choose_next_idle = 1;
 738                                 continue;
 739                         }
 740                         break;
 741                 }
 742
 743                 if (choose_next_idle)
 744                         continue;
 745
 746                 if (min_pending > pending) {
 747                         min_pending = pending;
 748                         best_pending_disk = disk;
 749                 }
 750
 751                 if (dist < best_dist) {
 752                         best_dist = dist;
 753                         best_dist_disk = disk;
 754                 }
 755         }
 756
 757         /*
 758          * If all disks are rotational, choose the closest disk. If any disk is
 759          * non-rotational, choose the disk with less pending request even the
 760          * disk is rotational, which might/might not be optimal for raids with
 761          * mixed ratation/non-rotational disks depending on workload.
 762          */
 763         if (best_disk == -1) {
 764                 if (has_nonrot_disk || min_pending == 0)
 765                         best_disk = best_pending_disk;
 766                 else
 767                         best_disk = best_dist_disk;
 768         }
 769
 770         if (best_disk >= 0) {
 771                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 772                 if (!rdev)
 773                         goto retry;
 774                 atomic_inc(&rdev->nr_pending);
 775                 sectors = best_good_sectors;
 776
 777                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 778                         conf->mirrors[best_disk].seq_start = this_sector;
 779
 780                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 781         }
 782         rcu_read_unlock();
 783         *max_sectors = sectors;
 784
 785         return best_disk;
 786 }
 787
 788 static int raid1_congested(struct mddev *mddev, int bits)
 789 {
 790         struct r1conf *conf = mddev->private;
 791         int i, ret = 0;
 792
 793         if ((bits & (1 << WB_async_congested)) &&
 794             conf->pending_count >= max_queued_requests)
 795                 return 1;
 796
 797         rcu_read_lock();
 798         for (i = 0; i < conf->raid_disks * 2; i++) {
 799                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 800                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 801                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 802
 803                         BUG_ON(!q);
 804
 805                         /* Note the '|| 1' - when read_balance prefers
 806                          * non-congested targets, it can be removed
 807                          */
 808                         if ((bits & (1 << WB_async_congested)) || 1)
 809                                 ret |= bdi_congested(q->backing_dev_info, bits);
 810                         else
 811                                 ret &= bdi_congested(q->backing_dev_info, bits);
 812                 }
 813         }
 814         rcu_read_unlock();
 815         return ret;
 816 }
 817
 818 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 819 {
 820         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
 821         md_bitmap_unplug(conf->mddev->bitmap);
 822         wake_up(&conf->wait_barrier);
 823
 824         while (bio) { /* submit pending writes */
 825                 struct bio *next = bio->bi_next;
 826                 struct md_rdev *rdev = (void *)bio->bi_disk;
 827                 bio->bi_next = NULL;
 828                 bio_set_dev(bio, rdev->bdev);
 829                 if (test_bit(Faulty, &rdev->flags)) {
 830                         bio_io_error(bio);
 831                 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
 832                                     !blk_queue_discard(bio->bi_disk->queue)))
 833                         /* Just ignore it */
 834                         bio_endio(bio);
 835                 else
 836                         generic_make_request(bio);
 837                 bio = next;
 838                 cond_resched();
 839         }
 840 }
 841
 842 static void flush_pending_writes(struct r1conf *conf)
 843 {
 844         /* Any writes that have been queued but are awaiting
 845          * bitmap updates get flushed here.
 846          */
 847         spin_lock_irq(&conf->device_lock);
 848
 849         if (conf->pending_bio_list.head) {
 850                 struct blk_plug plug;
 851                 struct bio *bio;
 852
 853                 bio = bio_list_get(&conf->pending_bio_list);
 854                 conf->pending_count = 0;
 855                 spin_unlock_irq(&conf->device_lock);
 856
 857                 /*
 858                  * As this is called in a wait_event() loop (see freeze_array),
 859                  * current->state might be TASK_UNINTERRUPTIBLE which will
 860                  * cause a warning when we prepare to wait again.  As it is
 861                  * rare that this path is taken, it is perfectly safe to force
 862                  * us to go around the wait_event() loop again, so the warning
 863                  * is a false-positive.  Silence the warning by resetting
 864                  * thread state
 865                  */
 866                 __set_current_state(TASK_RUNNING);
 867                 blk_start_plug(&plug);
 868                 flush_bio_list(conf, bio);
 869                 blk_finish_plug(&plug);
 870         } else
 871                 spin_unlock_irq(&conf->device_lock);
 872 }
 873
 874 /* Barriers....
 875  * Sometimes we need to suspend IO while we do something else,
 876  * either some resync/recovery, or reconfigure the array.
 877  * To do this we raise a 'barrier'.
 878  * The 'barrier' is a counter that can be raised multiple times
 879  * to count how many activities are happening which preclude
 880  * normal IO.
 881  * We can only raise the barrier if there is no pending IO.
 882  * i.e. if nr_pending == 0.
 883  * We choose only to raise the barrier if no-one is waiting for the
 884  * barrier to go down.  This means that as soon as an IO request
 885  * is ready, no other operations which require a barrier will start
 886  * until the IO request has had a chance.
 887  *
 888  * So: regular IO calls 'wait_barrier'.  When that returns there
 889  *    is no backgroup IO happening,  It must arrange to call
 890  *    allow_barrier when it has finished its IO.
 891  * backgroup IO calls must call raise_barrier.  Once that returns
 892  *    there is no normal IO happeing.  It must arrange to call
 893  *    lower_barrier when the particular background IO completes.
 894  *
 895  * If resync/recovery is interrupted, returns -EINTR;
 896  * Otherwise, returns 0.
 897  */
 898 static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
 899 {
 900         int idx = sector_to_idx(sector_nr);
 901
 902         spin_lock_irq(&conf->resync_lock);
 903
 904         /* Wait until no block IO is waiting */
 905         wait_event_lock_irq(conf->wait_barrier,
 906                             !atomic_read(&conf->nr_waiting[idx]),
 907                             conf->resync_lock);
 908
 909         /* block any new IO from starting */
 910         atomic_inc(&conf->barrier[idx]);
 911         /*
 912          * In raise_barrier() we firstly increase conf->barrier[idx] then
 913          * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 914          * increase conf->nr_pending[idx] then check conf->barrier[idx].
 915          * A memory barrier here to make sure conf->nr_pending[idx] won't
 916          * be fetched before conf->barrier[idx] is increased. Otherwise
 917          * there will be a race between raise_barrier() and _wait_barrier().
 918          */
 919         smp_mb__after_atomic();
 920
 921         /* For these conditions we must wait:
 922          * A: while the array is in frozen state
 923          * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 924          *    existing in corresponding I/O barrier bucket.
 925          * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 926          *    max resync count which allowed on current I/O barrier bucket.
 927          */
 928         wait_event_lock_irq(conf->wait_barrier,
 929                             (!conf->array_frozen &&
 930                              !atomic_read(&conf->nr_pending[idx]) &&
 931                              atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
 932                                 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
 933                             conf->resync_lock);
 934
 935         if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
 936                 atomic_dec(&conf->barrier[idx]);
 937                 spin_unlock_irq(&conf->resync_lock);
 938                 wake_up(&conf->wait_barrier);
 939                 return -EINTR;
 940         }
 941
 942         atomic_inc(&conf->nr_sync_pending);
 943         spin_unlock_irq(&conf->resync_lock);
 944
 945         return 0;
 946 }
 947
 948 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 949 {
 950         int idx = sector_to_idx(sector_nr);
 951
 952         BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 953
 954         atomic_dec(&conf->barrier[idx]);
 955         atomic_dec(&conf->nr_sync_pending);
 956         wake_up(&conf->wait_barrier);
 957 }
 958
 959 static void _wait_barrier(struct r1conf *conf, int idx)
 960 {
 961         /*
 962          * We need to increase conf->nr_pending[idx] very early here,
 963          * then raise_barrier() can be blocked when it waits for
 964          * conf->nr_pending[idx] to be 0. Then we can avoid holding
 965          * conf->resync_lock when there is no barrier raised in same
 966          * barrier unit bucket. Also if the array is frozen, I/O
 967          * should be blocked until array is unfrozen.
 968          */
 969         atomic_inc(&conf->nr_pending[idx]);
 970         /*
 971          * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
 972          * check conf->barrier[idx]. In raise_barrier() we firstly increase
 973          * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
 974          * barrier is necessary here to make sure conf->barrier[idx] won't be
 975          * fetched before conf->nr_pending[idx] is increased. Otherwise there
 976          * will be a race between _wait_barrier() and raise_barrier().
 977          */
 978         smp_mb__after_atomic();
 979
 980         /*
 981          * Don't worry about checking two atomic_t variables at same time
 982          * here. If during we check conf->barrier[idx], the array is
 983          * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
 984          * 0, it is safe to return and make the I/O continue. Because the
 985          * array is frozen, all I/O returned here will eventually complete
 986          * or be queued, no race will happen. See code comment in
 987          * frozen_array().
 988          */
 989         if (!READ_ONCE(conf->array_frozen) &&
 990             !atomic_read(&conf->barrier[idx]))
 991                 return;
 992
 993         /*
 994          * After holding conf->resync_lock, conf->nr_pending[idx]
 995          * should be decreased before waiting for barrier to drop.
 996          * Otherwise, we may encounter a race condition because
 997          * raise_barrer() might be waiting for conf->nr_pending[idx]
 998          * to be 0 at same time.
 999          */
1000         spin_lock_irq(&conf->resync_lock);
1001         atomic_inc(&conf->nr_waiting[idx]);
1002         atomic_dec(&conf->nr_pending[idx]);
1003         /*
1004          * In case freeze_array() is waiting for
1005          * get_unqueued_pending() == extra
1006          */
1007         wake_up(&conf->wait_barrier);
1008         /* Wait for the barrier in same barrier unit bucket to drop. */
1009         wait_event_lock_irq(conf->wait_barrier,
1010                             !conf->array_frozen &&
1011                              !atomic_read(&conf->barrier[idx]),
1012                             conf->resync_lock);
1013         atomic_inc(&conf->nr_pending[idx]);
1014         atomic_dec(&conf->nr_waiting[idx]);
1015         spin_unlock_irq(&conf->resync_lock);
1016 }
1017
1018 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
1019 {
1020         int idx = sector_to_idx(sector_nr);
1021
1022         /*
1023          * Very similar to _wait_barrier(). The difference is, for read
1024          * I/O we don't need wait for sync I/O, but if the whole array
1025          * is frozen, the read I/O still has to wait until the array is
1026          * unfrozen. Since there is no ordering requirement with
1027          * conf->barrier[idx] here, memory barrier is unnecessary as well.
1028          */
1029         atomic_inc(&conf->nr_pending[idx]);
1030
1031         if (!READ_ONCE(conf->array_frozen))
1032                 return;
1033
1034         spin_lock_irq(&conf->resync_lock);
1035         atomic_inc(&conf->nr_waiting[idx]);
1036         atomic_dec(&conf->nr_pending[idx]);
1037         /*
1038          * In case freeze_array() is waiting for
1039          * get_unqueued_pending() == extra
1040          */
1041         wake_up(&conf->wait_barrier);
1042         /* Wait for array to be unfrozen */
1043         wait_event_lock_irq(conf->wait_barrier,
1044                             !conf->array_frozen,
1045                             conf->resync_lock);
1046         atomic_inc(&conf->nr_pending[idx]);
1047         atomic_dec(&conf->nr_waiting[idx]);
1048         spin_unlock_irq(&conf->resync_lock);
1049 }
1050
1051 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1052 {
1053         int idx = sector_to_idx(sector_nr);
1054
1055         _wait_barrier(conf, idx);
1056 }
1057
1058 static void _allow_barrier(struct r1conf *conf, int idx)
1059 {
1060         atomic_dec(&conf->nr_pending[idx]);
1061         wake_up(&conf->wait_barrier);
1062 }
1063
1064 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1065 {
1066         int idx = sector_to_idx(sector_nr);
1067
1068         _allow_barrier(conf, idx);
1069 }
1070
1071 /* conf->resync_lock should be held */
1072 static int get_unqueued_pending(struct r1conf *conf)
1073 {
1074         int idx, ret;
1075
1076         ret = atomic_read(&conf->nr_sync_pending);
1077         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1078                 ret += atomic_read(&conf->nr_pending[idx]) -
1079                         atomic_read(&conf->nr_queued[idx]);
1080
1081         return ret;
1082 }
1083
1084 static void freeze_array(struct r1conf *conf, int extra)
1085 {
1086         /* Stop sync I/O and normal I/O and wait for everything to
1087          * go quiet.
1088          * This is called in two situations:
1089          * 1) management command handlers (reshape, remove disk, quiesce).
1090          * 2) one normal I/O request failed.
1091
1092          * After array_frozen is set to 1, new sync IO will be blocked at
1093          * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1094          * or wait_read_barrier(). The flying I/Os will either complete or be
1095          * queued. When everything goes quite, there are only queued I/Os left.
1096
1097          * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1098          * barrier bucket index which this I/O request hits. When all sync and
1099          * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1100          * of all conf->nr_queued[]. But normal I/O failure is an exception,
1101          * in handle_read_error(), we may call freeze_array() before trying to
1102          * fix the read error. In this case, the error read I/O is not queued,
1103          * so get_unqueued_pending() == 1.
1104          *
1105          * Therefore before this function returns, we need to wait until
1106          * get_unqueued_pendings(conf) gets equal to extra. For
1107          * normal I/O context, extra is 1, in rested situations extra is 0.
1108          */
1109         spin_lock_irq(&conf->resync_lock);
1110         conf->array_frozen = 1;
1111         raid1_log(conf->mddev, "wait freeze");
1112         wait_event_lock_irq_cmd(
1113                 conf->wait_barrier,
1114                 get_unqueued_pending(conf) == extra,
1115                 conf->resync_lock,
1116                 flush_pending_writes(conf));
1117         spin_unlock_irq(&conf->resync_lock);
1118 }
1119 static void unfreeze_array(struct r1conf *conf)
1120 {
1121         /* reverse the effect of the freeze */
1122         spin_lock_irq(&conf->resync_lock);
1123         conf->array_frozen = 0;
1124         spin_unlock_irq(&conf->resync_lock);
1125         wake_up(&conf->wait_barrier);
1126 }
1127
1128 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1129                                            struct bio *bio)
1130 {
1131         int size = bio->bi_iter.bi_size;
1132         unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1133         int i = 0;
1134         struct bio *behind_bio = NULL;
1135
1136         behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1137         if (!behind_bio)
1138                 return;
1139
1140         /* discard op, we don't support writezero/writesame yet */
1141         if (!bio_has_data(bio)) {
1142                 behind_bio->bi_iter.bi_size = size;
1143                 goto skip_copy;
1144         }
1145
1146         behind_bio->bi_write_hint = bio->bi_write_hint;
1147
1148         while (i < vcnt && size) {
1149                 struct page *page;
1150                 int len = min_t(int, PAGE_SIZE, size);
1151
1152                 page = alloc_page(GFP_NOIO);
1153                 if (unlikely(!page))
1154                         goto free_pages;
1155
1156                 bio_add_page(behind_bio, page, len, 0);
1157
1158                 size -= len;
1159                 i++;
1160         }
1161
1162         bio_copy_data(behind_bio, bio);
1163 skip_copy:
1164         r1_bio->behind_master_bio = behind_bio;
1165         set_bit(R1BIO_BehindIO, &r1_bio->state);
1166
1167         return;
1168
1169 free_pages:
1170         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1171                  bio->bi_iter.bi_size);
1172         bio_free_pages(behind_bio);
1173         bio_put(behind_bio);
1174 }
1175
1176 struct raid1_plug_cb {
1177         struct blk_plug_cb      cb;
1178         struct bio_list         pending;
1179         int                     pending_cnt;
1180 };
1181
1182 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1183 {
1184         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1185                                                   cb);
1186         struct mddev *mddev = plug->cb.data;
1187         struct r1conf *conf = mddev->private;
1188         struct bio *bio;
1189
1190         if (from_schedule || current->bio_list) {
1191                 spin_lock_irq(&conf->device_lock);
1192                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1193                 conf->pending_count += plug->pending_cnt;
1194                 spin_unlock_irq(&conf->device_lock);
1195                 wake_up(&conf->wait_barrier);
1196                 md_wakeup_thread(mddev->thread);
1197                 kfree(plug);
1198                 return;
1199         }
1200
1201         /* we aren't scheduling, so we can do the write-out directly. */
1202         bio = bio_list_get(&plug->pending);
1203         flush_bio_list(conf, bio);
1204         kfree(plug);
1205 }
1206
1207 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1208 {
1209         r1_bio->master_bio = bio;
1210         r1_bio->sectors = bio_sectors(bio);
1211         r1_bio->state = 0;
1212         r1_bio->mddev = mddev;
1213         r1_bio->sector = bio->bi_iter.bi_sector;
1214 }
1215
1216 static inline struct r1bio *
1217 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1218 {
1219         struct r1conf *conf = mddev->private;
1220         struct r1bio *r1_bio;
1221
1222         r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1223         /* Ensure no bio records IO_BLOCKED */
1224         memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1225         init_r1bio(r1_bio, mddev, bio);
1226         return r1_bio;
1227 }
1228
1229 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1230                                int max_read_sectors, struct r1bio *r1_bio)
1231 {
1232         struct r1conf *conf = mddev->private;
1233         struct raid1_info *mirror;
1234         struct bio *read_bio;
1235         struct bitmap *bitmap = mddev->bitmap;
1236         const int op = bio_op(bio);
1237         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1238         int max_sectors;
1239         int rdisk;
1240         bool print_msg = !!r1_bio;
1241         char b[BDEVNAME_SIZE];
1242
1243         /*
1244          * If r1_bio is set, we are blocking the raid1d thread
1245          * so there is a tiny risk of deadlock.  So ask for
1246          * emergency memory if needed.
1247          */
1248         gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1249
1250         if (print_msg) {
1251                 /* Need to get the block device name carefully */
1252                 struct md_rdev *rdev;
1253                 rcu_read_lock();
1254                 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1255                 if (rdev)
1256                         bdevname(rdev->bdev, b);
1257                 else
1258                         strcpy(b, "???");
1259                 rcu_read_unlock();
1260         }
1261
1262         /*
1263          * Still need barrier for READ in case that whole
1264          * array is frozen.
1265          */
1266         wait_read_barrier(conf, bio->bi_iter.bi_sector);
1267
1268         if (!r1_bio)
1269                 r1_bio = alloc_r1bio(mddev, bio);
1270         else
1271                 init_r1bio(r1_bio, mddev, bio);
1272         r1_bio->sectors = max_read_sectors;
1273
1274         /*
1275          * make_request() can abort the operation when read-ahead is being
1276          * used and no empty request is available.
1277          */
1278         rdisk = read_balance(conf, r1_bio, &max_sectors);
1279
1280         if (rdisk < 0) {
1281                 /* couldn't find anywhere to read from */
1282                 if (print_msg) {
1283                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1284                                             mdname(mddev),
1285                                             b,
1286                                             (unsigned long long)r1_bio->sector);
1287                 }
1288                 raid_end_bio_io(r1_bio);
1289                 return;
1290         }
1291         mirror = conf->mirrors + rdisk;
1292
1293         if (print_msg)
1294                 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1295                                     mdname(mddev),
1296                                     (unsigned long long)r1_bio->sector,
1297                                     bdevname(mirror->rdev->bdev, b));
1298
1299         if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1300             bitmap) {
1301                 /*
1302                  * Reading from a write-mostly device must take care not to
1303                  * over-take any writes that are 'behind'
1304                  */
1305                 raid1_log(mddev, "wait behind writes");
1306                 wait_event(bitmap->behind_wait,
1307                            atomic_read(&bitmap->behind_writes) == 0);
1308         }
1309
1310         if (max_sectors < bio_sectors(bio)) {
1311                 struct bio *split = bio_split(bio, max_sectors,
1312                                               gfp, &conf->bio_split);
1313                 bio_chain(split, bio);
1314                 generic_make_request(bio);
1315                 bio = split;
1316                 r1_bio->master_bio = bio;
1317                 r1_bio->sectors = max_sectors;
1318         }
1319
1320         r1_bio->read_disk = rdisk;
1321
1322         read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1323
1324         r1_bio->bios[rdisk] = read_bio;
1325
1326         read_bio->bi_iter.bi_sector = r1_bio->sector +
1327                 mirror->rdev->data_offset;
1328         bio_set_dev(read_bio, mirror->rdev->bdev);
1329         read_bio->bi_end_io = raid1_end_read_request;
1330         bio_set_op_attrs(read_bio, op, do_sync);
1331         if (test_bit(FailFast, &mirror->rdev->flags) &&
1332             test_bit(R1BIO_FailFast, &r1_bio->state))
1333                 read_bio->bi_opf |= MD_FAILFAST;
1334         read_bio->bi_private = r1_bio;
1335
1336         if (mddev->gendisk)
1337                 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1338                                 disk_devt(mddev->gendisk), r1_bio->sector);
1339
1340         generic_make_request(read_bio);
1341 }
1342
1343 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1344                                 int max_write_sectors)
1345 {
1346         struct r1conf *conf = mddev->private;
1347         struct r1bio *r1_bio;
1348         int i, disks;
1349         struct bitmap *bitmap = mddev->bitmap;
1350         unsigned long flags;
1351         struct md_rdev *blocked_rdev;
1352         struct blk_plug_cb *cb;
1353         struct raid1_plug_cb *plug = NULL;
1354         int first_clone;
1355         int max_sectors;
1356
1357         if (mddev_is_clustered(mddev) &&
1358              md_cluster_ops->area_resyncing(mddev, WRITE,
1359                      bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1360
1361                 DEFINE_WAIT(w);
1362                 for (;;) {
1363                         prepare_to_wait(&conf->wait_barrier,
1364                                         &w, TASK_IDLE);
1365                         if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1366                                                         bio->bi_iter.bi_sector,
1367                                                         bio_end_sector(bio)))
1368                                 break;
1369                         schedule();
1370                 }
1371                 finish_wait(&conf->wait_barrier, &w);
1372         }
1373
1374         /*
1375          * Register the new request and wait if the reconstruction
1376          * thread has put up a bar for new requests.
1377          * Continue immediately if no resync is active currently.
1378          */
1379         wait_barrier(conf, bio->bi_iter.bi_sector);
1380
1381         r1_bio = alloc_r1bio(mddev, bio);
1382         r1_bio->sectors = max_write_sectors;
1383
1384         if (conf->pending_count >= max_queued_requests) {
1385                 md_wakeup_thread(mddev->thread);
1386                 raid1_log(mddev, "wait queued");
1387                 wait_event(conf->wait_barrier,
1388                            conf->pending_count < max_queued_requests);
1389         }
1390         /* first select target devices under rcu_lock and
1391          * inc refcount on their rdev.  Record them by setting
1392          * bios[x] to bio
1393          * If there are known/acknowledged bad blocks on any device on
1394          * which we have seen a write error, we want to avoid writing those
1395          * blocks.
1396          * This potentially requires several writes to write around
1397          * the bad blocks.  Each set of writes gets it's own r1bio
1398          * with a set of bios attached.
1399          */
1400
1401         disks = conf->raid_disks * 2;
1402  retry_write:
1403         blocked_rdev = NULL;
1404         rcu_read_lock();
1405         max_sectors = r1_bio->sectors;
1406         for (i = 0;  i < disks; i++) {
1407                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1408                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1409                         atomic_inc(&rdev->nr_pending);
1410                         blocked_rdev = rdev;
1411                         break;
1412                 }
1413                 r1_bio->bios[i] = NULL;
1414                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1415                         if (i < conf->raid_disks)
1416                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1417                         continue;
1418                 }
1419
1420                 atomic_inc(&rdev->nr_pending);
1421                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1422                         sector_t first_bad;
1423                         int bad_sectors;
1424                         int is_bad;
1425
1426                         is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1427                                              &first_bad, &bad_sectors);
1428                         if (is_bad < 0) {
1429                                 /* mustn't write here until the bad block is
1430                                  * acknowledged*/
1431                                 set_bit(BlockedBadBlocks, &rdev->flags);
1432                                 blocked_rdev = rdev;
1433                                 break;
1434                         }
1435                         if (is_bad && first_bad <= r1_bio->sector) {
1436                                 /* Cannot write here at all */
1437                                 bad_sectors -= (r1_bio->sector - first_bad);
1438                                 if (bad_sectors < max_sectors)
1439                                         /* mustn't write more than bad_sectors
1440                                          * to other devices yet
1441                                          */
1442                                         max_sectors = bad_sectors;
1443                                 rdev_dec_pending(rdev, mddev);
1444                                 /* We don't set R1BIO_Degraded as that
1445                                  * only applies if the disk is
1446                                  * missing, so it might be re-added,
1447                                  * and we want to know to recover this
1448                                  * chunk.
1449                                  * In this case the device is here,
1450                                  * and the fact that this chunk is not
1451                                  * in-sync is recorded in the bad
1452                                  * block log
1453                                  */
1454                                 continue;
1455                         }
1456                         if (is_bad) {
1457                                 int good_sectors = first_bad - r1_bio->sector;
1458                                 if (good_sectors < max_sectors)
1459                                         max_sectors = good_sectors;
1460                         }
1461                 }
1462                 r1_bio->bios[i] = bio;
1463         }
1464         rcu_read_unlock();
1465
1466         if (unlikely(blocked_rdev)) {
1467                 /* Wait for this device to become unblocked */
1468                 int j;
1469
1470                 for (j = 0; j < i; j++)
1471                         if (r1_bio->bios[j])
1472                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1473                 r1_bio->state = 0;
1474                 allow_barrier(conf, bio->bi_iter.bi_sector);
1475                 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1476                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1477                 wait_barrier(conf, bio->bi_iter.bi_sector);
1478                 goto retry_write;
1479         }
1480
1481         if (max_sectors < bio_sectors(bio)) {
1482                 struct bio *split = bio_split(bio, max_sectors,
1483                                               GFP_NOIO, &conf->bio_split);
1484                 bio_chain(split, bio);
1485                 generic_make_request(bio);
1486                 bio = split;
1487                 r1_bio->master_bio = bio;
1488                 r1_bio->sectors = max_sectors;
1489         }
1490
1491         atomic_set(&r1_bio->remaining, 1);
1492         atomic_set(&r1_bio->behind_remaining, 0);
1493
1494         first_clone = 1;
1495
1496         for (i = 0; i < disks; i++) {
1497                 struct bio *mbio = NULL;
1498                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1499                 if (!r1_bio->bios[i])
1500                         continue;
1501
1502                 if (first_clone) {
1503                         /* do behind I/O ?
1504                          * Not if there are too many, or cannot
1505                          * allocate memory, or a reader on WriteMostly
1506                          * is waiting for behind writes to flush */
1507                         if (bitmap &&
1508                             (atomic_read(&bitmap->behind_writes)
1509                              < mddev->bitmap_info.max_write_behind) &&
1510                             !waitqueue_active(&bitmap->behind_wait)) {
1511                                 alloc_behind_master_bio(r1_bio, bio);
1512                         }
1513
1514                         md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1515                                              test_bit(R1BIO_BehindIO, &r1_bio->state));
1516                         first_clone = 0;
1517                 }
1518
1519                 if (r1_bio->behind_master_bio)
1520                         mbio = bio_clone_fast(r1_bio->behind_master_bio,
1521                                               GFP_NOIO, &mddev->bio_set);
1522                 else
1523                         mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1524
1525                 if (r1_bio->behind_master_bio) {
1526                         if (test_bit(CollisionCheck, &rdev->flags))
1527                                 wait_for_serialization(rdev, r1_bio);
1528                         if (test_bit(WriteMostly, &rdev->flags))
1529                                 atomic_inc(&r1_bio->behind_remaining);
1530                 } else if (mddev->serialize_policy)
1531                         wait_for_serialization(rdev, r1_bio);
1532
1533                 r1_bio->bios[i] = mbio;
1534
1535                 mbio->bi_iter.bi_sector = (r1_bio->sector +
1536                                    conf->mirrors[i].rdev->data_offset);
1537                 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1538                 mbio->bi_end_io = raid1_end_write_request;
1539                 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1540                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1541                     !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1542                     conf->raid_disks - mddev->degraded > 1)
1543                         mbio->bi_opf |= MD_FAILFAST;
1544                 mbio->bi_private = r1_bio;
1545
1546                 atomic_inc(&r1_bio->remaining);
1547
1548                 if (mddev->gendisk)
1549                         trace_block_bio_remap(mbio->bi_disk->queue,
1550                                               mbio, disk_devt(mddev->gendisk),
1551                                               r1_bio->sector);
1552                 /* flush_pending_writes() needs access to the rdev so...*/
1553                 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1554
1555                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1556                 if (cb)
1557                         plug = container_of(cb, struct raid1_plug_cb, cb);
1558                 else
1559                         plug = NULL;
1560                 if (plug) {
1561                         bio_list_add(&plug->pending, mbio);
1562                         plug->pending_cnt++;
1563                 } else {
1564                         spin_lock_irqsave(&conf->device_lock, flags);
1565                         bio_list_add(&conf->pending_bio_list, mbio);
1566                         conf->pending_count++;
1567                         spin_unlock_irqrestore(&conf->device_lock, flags);
1568                         md_wakeup_thread(mddev->thread);
1569                 }
1570         }
1571
1572         r1_bio_write_done(r1_bio);
1573
1574         /* In case raid1d snuck in to freeze_array */
1575         wake_up(&conf->wait_barrier);
1576 }
1577
1578 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1579 {
1580         sector_t sectors;
1581
1582         if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1583             && md_flush_request(mddev, bio))
1584                 return true;
1585
1586         /*
1587          * There is a limit to the maximum size, but
1588          * the read/write handler might find a lower limit
1589          * due to bad blocks.  To avoid multiple splits,
1590          * we pass the maximum number of sectors down
1591          * and let the lower level perform the split.
1592          */
1593         sectors = align_to_barrier_unit_end(
1594                 bio->bi_iter.bi_sector, bio_sectors(bio));
1595
1596         if (bio_data_dir(bio) == READ)
1597                 raid1_read_request(mddev, bio, sectors, NULL);
1598         else {
1599                 if (!md_write_start(mddev,bio))
1600                         return false;
1601                 raid1_write_request(mddev, bio, sectors);
1602         }
1603         return true;
1604 }
1605
1606 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1607 {
1608         struct r1conf *conf = mddev->private;
1609         int i;
1610
1611         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1612                    conf->raid_disks - mddev->degraded);
1613         rcu_read_lock();
1614         for (i = 0; i < conf->raid_disks; i++) {
1615                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1616                 seq_printf(seq, "%s",
1617                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1618         }
1619         rcu_read_unlock();
1620         seq_printf(seq, "]");
1621 }
1622
1623 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1624 {
1625         char b[BDEVNAME_SIZE];
1626         struct r1conf *conf = mddev->private;
1627         unsigned long flags;
1628
1629         /*
1630          * If it is not operational, then we have already marked it as dead
1631          * else if it is the last working disks with "fail_last_dev == false",
1632          * ignore the error, let the next level up know.
1633          * else mark the drive as failed
1634          */
1635         spin_lock_irqsave(&conf->device_lock, flags);
1636         if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1637             && (conf->raid_disks - mddev->degraded) == 1) {
1638                 /*
1639                  * Don't fail the drive, act as though we were just a
1640                  * normal single drive.
1641                  * However don't try a recovery from this drive as
1642                  * it is very likely to fail.
1643                  */
1644                 conf->recovery_disabled = mddev->recovery_disabled;
1645                 spin_unlock_irqrestore(&conf->device_lock, flags);
1646                 return;
1647         }
1648         set_bit(Blocked, &rdev->flags);
1649         if (test_and_clear_bit(In_sync, &rdev->flags))
1650                 mddev->degraded++;
1651         set_bit(Faulty, &rdev->flags);
1652         spin_unlock_irqrestore(&conf->device_lock, flags);
1653         /*
1654          * if recovery is running, make sure it aborts.
1655          */
1656         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1657         set_mask_bits(&mddev->sb_flags, 0,
1658                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1659         pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1660                 "md/raid1:%s: Operation continuing on %d devices.\n",
1661                 mdname(mddev), bdevname(rdev->bdev, b),
1662                 mdname(mddev), conf->raid_disks - mddev->degraded);
1663 }
1664
1665 static void print_conf(struct r1conf *conf)
1666 {
1667         int i;
1668
1669         pr_debug("RAID1 conf printout:\n");
1670         if (!conf) {
1671                 pr_debug("(!conf)\n");
1672                 return;
1673         }
1674         pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1675                  conf->raid_disks);
1676
1677         rcu_read_lock();
1678         for (i = 0; i < conf->raid_disks; i++) {
1679                 char b[BDEVNAME_SIZE];
1680                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1681                 if (rdev)
1682                         pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1683                                  i, !test_bit(In_sync, &rdev->flags),
1684                                  !test_bit(Faulty, &rdev->flags),
1685                                  bdevname(rdev->bdev,b));
1686         }
1687         rcu_read_unlock();
1688 }
1689
1690 static void close_sync(struct r1conf *conf)
1691 {
1692         int idx;
1693
1694         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1695                 _wait_barrier(conf, idx);
1696                 _allow_barrier(conf, idx);
1697         }
1698
1699         mempool_exit(&conf->r1buf_pool);
1700 }
1701
1702 static int raid1_spare_active(struct mddev *mddev)
1703 {
1704         int i;
1705         struct r1conf *conf = mddev->private;
1706         int count = 0;
1707         unsigned long flags;
1708
1709         /*
1710          * Find all failed disks within the RAID1 configuration
1711          * and mark them readable.
1712          * Called under mddev lock, so rcu protection not needed.
1713          * device_lock used to avoid races with raid1_end_read_request
1714          * which expects 'In_sync' flags and ->degraded to be consistent.
1715          */
1716         spin_lock_irqsave(&conf->device_lock, flags);
1717         for (i = 0; i < conf->raid_disks; i++) {
1718                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1719                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1720                 if (repl
1721                     && !test_bit(Candidate, &repl->flags)
1722                     && repl->recovery_offset == MaxSector
1723                     && !test_bit(Faulty, &repl->flags)
1724                     && !test_and_set_bit(In_sync, &repl->flags)) {
1725                         /* replacement has just become active */
1726                         if (!rdev ||
1727                             !test_and_clear_bit(In_sync, &rdev->flags))
1728                                 count++;
1729                         if (rdev) {
1730                                 /* Replaced device not technically
1731                                  * faulty, but we need to be sure
1732                                  * it gets removed and never re-added
1733                                  */
1734                                 set_bit(Faulty, &rdev->flags);
1735                                 sysfs_notify_dirent_safe(
1736                                         rdev->sysfs_state);
1737                         }
1738                 }
1739                 if (rdev
1740                     && rdev->recovery_offset == MaxSector
1741                     && !test_bit(Faulty, &rdev->flags)
1742                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1743                         count++;
1744                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1745                 }
1746         }
1747         mddev->degraded -= count;
1748         spin_unlock_irqrestore(&conf->device_lock, flags);
1749
1750         print_conf(conf);
1751         return count;
1752 }
1753
1754 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1755 {
1756         struct r1conf *conf = mddev->private;
1757         int err = -EEXIST;
1758         int mirror = 0;
1759         struct raid1_info *p;
1760         int first = 0;
1761         int last = conf->raid_disks - 1;
1762
1763         if (mddev->recovery_disabled == conf->recovery_disabled)
1764                 return -EBUSY;
1765
1766         if (md_integrity_add_rdev(rdev, mddev))
1767                 return -ENXIO;
1768
1769         if (rdev->raid_disk >= 0)
1770                 first = last = rdev->raid_disk;
1771
1772         /*
1773          * find the disk ... but prefer rdev->saved_raid_disk
1774          * if possible.
1775          */
1776         if (rdev->saved_raid_disk >= 0 &&
1777             rdev->saved_raid_disk >= first &&
1778             rdev->saved_raid_disk < conf->raid_disks &&
1779             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1780                 first = last = rdev->saved_raid_disk;
1781
1782         for (mirror = first; mirror <= last; mirror++) {
1783                 p = conf->mirrors + mirror;
1784                 if (!p->rdev) {
1785                         if (mddev->gendisk)
1786                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1787                                                   rdev->data_offset << 9);
1788
1789                         p->head_position = 0;
1790                         rdev->raid_disk = mirror;
1791                         err = 0;
1792                         /* As all devices are equivalent, we don't need a full recovery
1793                          * if this was recently any drive of the array
1794                          */
1795                         if (rdev->saved_raid_disk < 0)
1796                                 conf->fullsync = 1;
1797                         rcu_assign_pointer(p->rdev, rdev);
1798                         break;
1799                 }
1800                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1801                     p[conf->raid_disks].rdev == NULL) {
1802                         /* Add this device as a replacement */
1803                         clear_bit(In_sync, &rdev->flags);
1804                         set_bit(Replacement, &rdev->flags);
1805                         rdev->raid_disk = mirror;
1806                         err = 0;
1807                         conf->fullsync = 1;
1808                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1809                         break;
1810                 }
1811         }
1812         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1813                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1814         print_conf(conf);
1815         return err;
1816 }
1817
1818 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1819 {
1820         struct r1conf *conf = mddev->private;
1821         int err = 0;
1822         int number = rdev->raid_disk;
1823         struct raid1_info *p = conf->mirrors + number;
1824
1825         if (rdev != p->rdev)
1826                 p = conf->mirrors + conf->raid_disks + number;
1827
1828         print_conf(conf);
1829         if (rdev == p->rdev) {
1830                 if (test_bit(In_sync, &rdev->flags) ||
1831                     atomic_read(&rdev->nr_pending)) {
1832                         err = -EBUSY;
1833                         goto abort;
1834                 }
1835                 /* Only remove non-faulty devices if recovery
1836                  * is not possible.
1837                  */
1838                 if (!test_bit(Faulty, &rdev->flags) &&
1839                     mddev->recovery_disabled != conf->recovery_disabled &&
1840                     mddev->degraded < conf->raid_disks) {
1841                         err = -EBUSY;
1842                         goto abort;
1843                 }
1844                 p->rdev = NULL;
1845                 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1846                         synchronize_rcu();
1847                         if (atomic_read(&rdev->nr_pending)) {
1848                                 /* lost the race, try later */
1849                                 err = -EBUSY;
1850                                 p->rdev = rdev;
1851                                 goto abort;
1852                         }
1853                 }
1854                 if (conf->mirrors[conf->raid_disks + number].rdev) {
1855                         /* We just removed a device that is being replaced.
1856                          * Move down the replacement.  We drain all IO before
1857                          * doing this to avoid confusion.
1858                          */
1859                         struct md_rdev *repl =
1860                                 conf->mirrors[conf->raid_disks + number].rdev;
1861                         freeze_array(conf, 0);
1862                         if (atomic_read(&repl->nr_pending)) {
1863                                 /* It means that some queued IO of retry_list
1864                                  * hold repl. Thus, we cannot set replacement
1865                                  * as NULL, avoiding rdev NULL pointer
1866                                  * dereference in sync_request_write and
1867                                  * handle_write_finished.
1868                                  */
1869                                 err = -EBUSY;
1870                                 unfreeze_array(conf);
1871                                 goto abort;
1872                         }
1873                         clear_bit(Replacement, &repl->flags);
1874                         p->rdev = repl;
1875                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1876                         unfreeze_array(conf);
1877                 }
1878
1879                 clear_bit(WantReplacement, &rdev->flags);
1880                 err = md_integrity_register(mddev);
1881         }
1882 abort:
1883
1884         print_conf(conf);
1885         return err;
1886 }
1887
1888 static void end_sync_read(struct bio *bio)
1889 {
1890         struct r1bio *r1_bio = get_resync_r1bio(bio);
1891
1892         update_head_pos(r1_bio->read_disk, r1_bio);
1893
1894         /*
1895          * we have read a block, now it needs to be re-written,
1896          * or re-read if the read failed.
1897          * We don't do much here, just schedule handling by raid1d
1898          */
1899         if (!bio->bi_status)
1900                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1901
1902         if (atomic_dec_and_test(&r1_bio->remaining))
1903                 reschedule_retry(r1_bio);
1904 }
1905
1906 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1907 {
1908         sector_t sync_blocks = 0;
1909         sector_t s = r1_bio->sector;
1910         long sectors_to_go = r1_bio->sectors;
1911
1912         /* make sure these bits don't get cleared. */
1913         do {
1914                 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1915                 s += sync_blocks;
1916                 sectors_to_go -= sync_blocks;
1917         } while (sectors_to_go > 0);
1918 }
1919
1920 static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1921 {
1922         if (atomic_dec_and_test(&r1_bio->remaining)) {
1923                 struct mddev *mddev = r1_bio->mddev;
1924                 int s = r1_bio->sectors;
1925
1926                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1927                     test_bit(R1BIO_WriteError, &r1_bio->state))
1928                         reschedule_retry(r1_bio);
1929                 else {
1930                         put_buf(r1_bio);
1931                         md_done_sync(mddev, s, uptodate);
1932                 }
1933         }
1934 }
1935
1936 static void end_sync_write(struct bio *bio)
1937 {
1938         int uptodate = !bio->bi_status;
1939         struct r1bio *r1_bio = get_resync_r1bio(bio);
1940         struct mddev *mddev = r1_bio->mddev;
1941         struct r1conf *conf = mddev->private;
1942         sector_t first_bad;
1943         int bad_sectors;
1944         struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1945
1946         if (!uptodate) {
1947                 abort_sync_write(mddev, r1_bio);
1948                 set_bit(WriteErrorSeen, &rdev->flags);
1949                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1950                         set_bit(MD_RECOVERY_NEEDED, &
1951                                 mddev->recovery);
1952                 set_bit(R1BIO_WriteError, &r1_bio->state);
1953         } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1954                                &first_bad, &bad_sectors) &&
1955                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1956                                 r1_bio->sector,
1957                                 r1_bio->sectors,
1958                                 &first_bad, &bad_sectors)
1959                 )
1960                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1961
1962         put_sync_write_buf(r1_bio, uptodate);
1963 }
1964
1965 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1966                             int sectors, struct page *page, int rw)
1967 {
1968         if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1969                 /* success */
1970                 return 1;
1971         if (rw == WRITE) {
1972                 set_bit(WriteErrorSeen, &rdev->flags);
1973                 if (!test_and_set_bit(WantReplacement,
1974                                       &rdev->flags))
1975                         set_bit(MD_RECOVERY_NEEDED, &
1976                                 rdev->mddev->recovery);
1977         }
1978         /* need to record an error - either for the block or the device */
1979         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1980                 md_error(rdev->mddev, rdev);
1981         return 0;
1982 }
1983
1984 static int fix_sync_read_error(struct r1bio *r1_bio)
1985 {
1986         /* Try some synchronous reads of other devices to get
1987          * good data, much like with normal read errors.  Only
1988          * read into the pages we already have so we don't
1989          * need to re-issue the read request.
1990          * We don't need to freeze the array, because being in an
1991          * active sync request, there is no normal IO, and
1992          * no overlapping syncs.
1993          * We don't need to check is_badblock() again as we
1994          * made sure that anything with a bad block in range
1995          * will have bi_end_io clear.
1996          */
1997         struct mddev *mddev = r1_bio->mddev;
1998         struct r1conf *conf = mddev->private;
1999         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
2000         struct page **pages = get_resync_pages(bio)->pages;
2001         sector_t sect = r1_bio->sector;
2002         int sectors = r1_bio->sectors;
2003         int idx = 0;
2004         struct md_rdev *rdev;
2005
2006         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2007         if (test_bit(FailFast, &rdev->flags)) {
2008                 /* Don't try recovering from here - just fail it
2009                  * ... unless it is the last working device of course */
2010                 md_error(mddev, rdev);
2011                 if (test_bit(Faulty, &rdev->flags))
2012                         /* Don't try to read from here, but make sure
2013                          * put_buf does it's thing
2014                          */
2015                         bio->bi_end_io = end_sync_write;
2016         }
2017
2018         while(sectors) {
2019                 int s = sectors;
2020                 int d = r1_bio->read_disk;
2021                 int success = 0;
2022                 int start;
2023
2024                 if (s > (PAGE_SIZE>>9))
2025                         s = PAGE_SIZE >> 9;
2026                 do {
2027                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2028                                 /* No rcu protection needed here devices
2029                                  * can only be removed when no resync is
2030                                  * active, and resync is currently active
2031                                  */
2032                                 rdev = conf->mirrors[d].rdev;
2033                                 if (sync_page_io(rdev, sect, s<<9,
2034                                                  pages[idx],
2035                                                  REQ_OP_READ, 0, false)) {
2036                                         success = 1;
2037                                         break;
2038                                 }
2039                         }
2040                         d++;
2041                         if (d == conf->raid_disks * 2)
2042                                 d = 0;
2043                 } while (!success && d != r1_bio->read_disk);
2044
2045                 if (!success) {
2046                         char b[BDEVNAME_SIZE];
2047                         int abort = 0;
2048                         /* Cannot read from anywhere, this block is lost.
2049                          * Record a bad block on each device.  If that doesn't
2050                          * work just disable and interrupt the recovery.
2051                          * Don't fail devices as that won't really help.
2052                          */
2053                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2054                                             mdname(mddev), bio_devname(bio, b),
2055                                             (unsigned long long)r1_bio->sector);
2056                         for (d = 0; d < conf->raid_disks * 2; d++) {
2057                                 rdev = conf->mirrors[d].rdev;
2058                                 if (!rdev || test_bit(Faulty, &rdev->flags))
2059                                         continue;
2060                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
2061                                         abort = 1;
2062                         }
2063                         if (abort) {
2064                                 conf->recovery_disabled =
2065                                         mddev->recovery_disabled;
2066                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2067                                 md_done_sync(mddev, r1_bio->sectors, 0);
2068                                 put_buf(r1_bio);
2069                                 return 0;
2070                         }
2071                         /* Try next page */
2072                         sectors -= s;
2073                         sect += s;
2074                         idx++;
2075                         continue;
2076                 }
2077
2078                 start = d;
2079                 /* write it back and re-read */
2080                 while (d != r1_bio->read_disk) {
2081                         if (d == 0)
2082                                 d = conf->raid_disks * 2;
2083                         d--;
2084                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2085                                 continue;
2086                         rdev = conf->mirrors[d].rdev;
2087                         if (r1_sync_page_io(rdev, sect, s,
2088                                             pages[idx],
2089                                             WRITE) == 0) {
2090                                 r1_bio->bios[d]->bi_end_io = NULL;
2091                                 rdev_dec_pending(rdev, mddev);
2092                         }
2093                 }
2094                 d = start;
2095                 while (d != r1_bio->read_disk) {
2096                         if (d == 0)
2097                                 d = conf->raid_disks * 2;
2098                         d--;
2099                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2100                                 continue;
2101                         rdev = conf->mirrors[d].rdev;
2102                         if (r1_sync_page_io(rdev, sect, s,
2103                                             pages[idx],
2104                                             READ) != 0)
2105                                 atomic_add(s, &rdev->corrected_errors);
2106                 }
2107                 sectors -= s;
2108                 sect += s;
2109                 idx ++;
2110         }
2111         set_bit(R1BIO_Uptodate, &r1_bio->state);
2112         bio->bi_status = 0;
2113         return 1;
2114 }
2115
2116 static void process_checks(struct r1bio *r1_bio)
2117 {
2118         /* We have read all readable devices.  If we haven't
2119          * got the block, then there is no hope left.
2120          * If we have, then we want to do a comparison
2121          * and skip the write if everything is the same.
2122          * If any blocks failed to read, then we need to
2123          * attempt an over-write
2124          */
2125         struct mddev *mddev = r1_bio->mddev;
2126         struct r1conf *conf = mddev->private;
2127         int primary;
2128         int i;
2129         int vcnt;
2130
2131         /* Fix variable parts of all bios */
2132         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2133         for (i = 0; i < conf->raid_disks * 2; i++) {
2134                 blk_status_t status;
2135                 struct bio *b = r1_bio->bios[i];
2136                 struct resync_pages *rp = get_resync_pages(b);
2137                 if (b->bi_end_io != end_sync_read)
2138                         continue;
2139                 /* fixup the bio for reuse, but preserve errno */
2140                 status = b->bi_status;
2141                 bio_reset(b);
2142                 b->bi_status = status;
2143                 b->bi_iter.bi_sector = r1_bio->sector +
2144                         conf->mirrors[i].rdev->data_offset;
2145                 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2146                 b->bi_end_io = end_sync_read;
2147                 rp->raid_bio = r1_bio;
2148                 b->bi_private = rp;
2149
2150                 /* initialize bvec table again */
2151                 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2152         }
2153         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2154                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2155                     !r1_bio->bios[primary]->bi_status) {
2156                         r1_bio->bios[primary]->bi_end_io = NULL;
2157                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2158                         break;
2159                 }
2160         r1_bio->read_disk = primary;
2161         for (i = 0; i < conf->raid_disks * 2; i++) {
2162                 int j = 0;
2163                 struct bio *pbio = r1_bio->bios[primary];
2164                 struct bio *sbio = r1_bio->bios[i];
2165                 blk_status_t status = sbio->bi_status;
2166                 struct page **ppages = get_resync_pages(pbio)->pages;
2167                 struct page **spages = get_resync_pages(sbio)->pages;
2168                 struct bio_vec *bi;
2169                 int page_len[RESYNC_PAGES] = { 0 };
2170                 struct bvec_iter_all iter_all;
2171
2172                 if (sbio->bi_end_io != end_sync_read)
2173                         continue;
2174                 /* Now we can 'fixup' the error value */
2175                 sbio->bi_status = 0;
2176
2177                 bio_for_each_segment_all(bi, sbio, iter_all)
2178                         page_len[j++] = bi->bv_len;
2179
2180                 if (!status) {
2181                         for (j = vcnt; j-- ; ) {
2182                                 if (memcmp(page_address(ppages[j]),
2183                                            page_address(spages[j]),
2184                                            page_len[j]))
2185                                         break;
2186                         }
2187                 } else
2188                         j = 0;
2189                 if (j >= 0)
2190                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2191                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2192                               && !status)) {
2193                         /* No need to write to this device. */
2194                         sbio->bi_end_io = NULL;
2195                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2196                         continue;
2197                 }
2198
2199                 bio_copy_data(sbio, pbio);
2200         }
2201 }
2202
2203 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2204 {
2205         struct r1conf *conf = mddev->private;
2206         int i;
2207         int disks = conf->raid_disks * 2;
2208         struct bio *wbio;
2209
2210         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2211                 /* ouch - failed to read all of that. */
2212                 if (!fix_sync_read_error(r1_bio))
2213                         return;
2214
2215         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2216                 process_checks(r1_bio);
2217
2218         /*
2219          * schedule writes
2220          */
2221         atomic_set(&r1_bio->remaining, 1);
2222         for (i = 0; i < disks ; i++) {
2223                 wbio = r1_bio->bios[i];
2224                 if (wbio->bi_end_io == NULL ||
2225                     (wbio->bi_end_io == end_sync_read &&
2226                      (i == r1_bio->read_disk ||
2227                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2228                         continue;
2229                 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2230                         abort_sync_write(mddev, r1_bio);
2231                         continue;
2232                 }
2233
2234                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2235                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2236                         wbio->bi_opf |= MD_FAILFAST;
2237
2238                 wbio->bi_end_io = end_sync_write;
2239                 atomic_inc(&r1_bio->remaining);
2240                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2241
2242                 generic_make_request(wbio);
2243         }
2244
2245         put_sync_write_buf(r1_bio, 1);
2246 }
2247
2248 /*
2249  * This is a kernel thread which:
2250  *
2251  *      1.      Retries failed read operations on working mirrors.
2252  *      2.      Updates the raid superblock when problems encounter.
2253  *      3.      Performs writes following reads for array synchronising.
2254  */
2255
2256 static void fix_read_error(struct r1conf *conf, int read_disk,
2257                            sector_t sect, int sectors)
2258 {
2259         struct mddev *mddev = conf->mddev;
2260         while(sectors) {
2261                 int s = sectors;
2262                 int d = read_disk;
2263                 int success = 0;
2264                 int start;
2265                 struct md_rdev *rdev;
2266
2267                 if (s > (PAGE_SIZE>>9))
2268                         s = PAGE_SIZE >> 9;
2269
2270                 do {
2271                         sector_t first_bad;
2272                         int bad_sectors;
2273
2274                         rcu_read_lock();
2275                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2276                         if (rdev &&
2277                             (test_bit(In_sync, &rdev->flags) ||
2278                              (!test_bit(Faulty, &rdev->flags) &&
2279                               rdev->recovery_offset >= sect + s)) &&
2280                             is_badblock(rdev, sect, s,
2281                                         &first_bad, &bad_sectors) == 0) {
2282                                 atomic_inc(&rdev->nr_pending);
2283                                 rcu_read_unlock();
2284                                 if (sync_page_io(rdev, sect, s<<9,
2285                                          conf->tmppage, REQ_OP_READ, 0, false))
2286                                         success = 1;
2287                                 rdev_dec_pending(rdev, mddev);
2288                                 if (success)
2289                                         break;
2290                         } else
2291                                 rcu_read_unlock();
2292                         d++;
2293                         if (d == conf->raid_disks * 2)
2294                                 d = 0;
2295                 } while (!success && d != read_disk);
2296
2297                 if (!success) {
2298                         /* Cannot read from anywhere - mark it bad */
2299                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2300                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2301                                 md_error(mddev, rdev);
2302                         break;
2303                 }
2304                 /* write it back and re-read */
2305                 start = d;
2306                 while (d != read_disk) {
2307                         if (d==0)
2308                                 d = conf->raid_disks * 2;
2309                         d--;
2310                         rcu_read_lock();
2311                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2312                         if (rdev &&
2313                             !test_bit(Faulty, &rdev->flags)) {
2314                                 atomic_inc(&rdev->nr_pending);
2315                                 rcu_read_unlock();
2316                                 r1_sync_page_io(rdev, sect, s,
2317                                                 conf->tmppage, WRITE);
2318                                 rdev_dec_pending(rdev, mddev);
2319                         } else
2320                                 rcu_read_unlock();
2321                 }
2322                 d = start;
2323                 while (d != read_disk) {
2324                         char b[BDEVNAME_SIZE];
2325                         if (d==0)
2326                                 d = conf->raid_disks * 2;
2327                         d--;
2328                         rcu_read_lock();
2329                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2330                         if (rdev &&
2331                             !test_bit(Faulty, &rdev->flags)) {
2332                                 atomic_inc(&rdev->nr_pending);
2333                                 rcu_read_unlock();
2334                                 if (r1_sync_page_io(rdev, sect, s,
2335                                                     conf->tmppage, READ)) {
2336                                         atomic_add(s, &rdev->corrected_errors);
2337                                         pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2338                                                 mdname(mddev), s,
2339                                                 (unsigned long long)(sect +
2340                                                                      rdev->data_offset),
2341                                                 bdevname(rdev->bdev, b));
2342                                 }
2343                                 rdev_dec_pending(rdev, mddev);
2344                         } else
2345                                 rcu_read_unlock();
2346                 }
2347                 sectors -= s;
2348                 sect += s;
2349         }
2350 }
2351
2352 static int narrow_write_error(struct r1bio *r1_bio, int i)
2353 {
2354         struct mddev *mddev = r1_bio->mddev;
2355         struct r1conf *conf = mddev->private;
2356         struct md_rdev *rdev = conf->mirrors[i].rdev;
2357
2358         /* bio has the data to be written to device 'i' where
2359          * we just recently had a write error.
2360          * We repeatedly clone the bio and trim down to one block,
2361          * then try the write.  Where the write fails we record
2362          * a bad block.
2363          * It is conceivable that the bio doesn't exactly align with
2364          * blocks.  We must handle this somehow.
2365          *
2366          * We currently own a reference on the rdev.
2367          */
2368
2369         int block_sectors;
2370         sector_t sector;
2371         int sectors;
2372         int sect_to_write = r1_bio->sectors;
2373         int ok = 1;
2374
2375         if (rdev->badblocks.shift < 0)
2376                 return 0;
2377
2378         block_sectors = roundup(1 << rdev->badblocks.shift,
2379                                 bdev_logical_block_size(rdev->bdev) >> 9);
2380         sector = r1_bio->sector;
2381         sectors = ((sector + block_sectors)
2382                    & ~(sector_t)(block_sectors - 1))
2383                 - sector;
2384
2385         while (sect_to_write) {
2386                 struct bio *wbio;
2387                 if (sectors > sect_to_write)
2388                         sectors = sect_to_write;
2389                 /* Write at 'sector' for 'sectors'*/
2390
2391                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2392                         wbio = bio_clone_fast(r1_bio->behind_master_bio,
2393                                               GFP_NOIO,
2394                                               &mddev->bio_set);
2395                 } else {
2396                         wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2397                                               &mddev->bio_set);
2398                 }
2399
2400                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2401                 wbio->bi_iter.bi_sector = r1_bio->sector;
2402                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2403
2404                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2405                 wbio->bi_iter.bi_sector += rdev->data_offset;
2406                 bio_set_dev(wbio, rdev->bdev);
2407
2408                 if (submit_bio_wait(wbio) < 0)
2409                         /* failure! */
2410                         ok = rdev_set_badblocks(rdev, sector,
2411                                                 sectors, 0)
2412                                 && ok;
2413
2414                 bio_put(wbio);
2415                 sect_to_write -= sectors;
2416                 sector += sectors;
2417                 sectors = block_sectors;
2418         }
2419         return ok;
2420 }
2421
2422 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2423 {
2424         int m;
2425         int s = r1_bio->sectors;
2426         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2427                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2428                 struct bio *bio = r1_bio->bios[m];
2429                 if (bio->bi_end_io == NULL)
2430                         continue;
2431                 if (!bio->bi_status &&
2432                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2433                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2434                 }
2435                 if (bio->bi_status &&
2436                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2437                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2438                                 md_error(conf->mddev, rdev);
2439                 }
2440         }
2441         put_buf(r1_bio);
2442         md_done_sync(conf->mddev, s, 1);
2443 }
2444
2445 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2446 {
2447         int m, idx;
2448         bool fail = false;
2449
2450         for (m = 0; m < conf->raid_disks * 2 ; m++)
2451                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2452                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2453                         rdev_clear_badblocks(rdev,
2454                                              r1_bio->sector,
2455                                              r1_bio->sectors, 0);
2456                         rdev_dec_pending(rdev, conf->mddev);
2457                 } else if (r1_bio->bios[m] != NULL) {
2458                         /* This drive got a write error.  We need to
2459                          * narrow down and record precise write
2460                          * errors.
2461                          */
2462                         fail = true;
2463                         if (!narrow_write_error(r1_bio, m)) {
2464                                 md_error(conf->mddev,
2465                                          conf->mirrors[m].rdev);
2466                                 /* an I/O failed, we can't clear the bitmap */
2467                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2468                         }
2469                         rdev_dec_pending(conf->mirrors[m].rdev,
2470                                          conf->mddev);
2471                 }
2472         if (fail) {
2473                 spin_lock_irq(&conf->device_lock);
2474                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2475                 idx = sector_to_idx(r1_bio->sector);
2476                 atomic_inc(&conf->nr_queued[idx]);
2477                 spin_unlock_irq(&conf->device_lock);
2478                 /*
2479                  * In case freeze_array() is waiting for condition
2480                  * get_unqueued_pending() == extra to be true.
2481                  */
2482                 wake_up(&conf->wait_barrier);
2483                 md_wakeup_thread(conf->mddev->thread);
2484         } else {
2485                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2486                         close_write(r1_bio);
2487                 raid_end_bio_io(r1_bio);
2488         }
2489 }
2490
2491 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2492 {
2493         struct mddev *mddev = conf->mddev;
2494         struct bio *bio;
2495         struct md_rdev *rdev;
2496
2497         clear_bit(R1BIO_ReadError, &r1_bio->state);
2498         /* we got a read error. Maybe the drive is bad.  Maybe just
2499          * the block and we can fix it.
2500          * We freeze all other IO, and try reading the block from
2501          * other devices.  When we find one, we re-write
2502          * and check it that fixes the read error.
2503          * This is all done synchronously while the array is
2504          * frozen
2505          */
2506
2507         bio = r1_bio->bios[r1_bio->read_disk];
2508         bio_put(bio);
2509         r1_bio->bios[r1_bio->read_disk] = NULL;
2510
2511         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2512         if (mddev->ro == 0
2513             && !test_bit(FailFast, &rdev->flags)) {
2514                 freeze_array(conf, 1);
2515                 fix_read_error(conf, r1_bio->read_disk,
2516                                r1_bio->sector, r1_bio->sectors);
2517                 unfreeze_array(conf);
2518         } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2519                 md_error(mddev, rdev);
2520         } else {
2521                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2522         }
2523
2524         rdev_dec_pending(rdev, conf->mddev);
2525         allow_barrier(conf, r1_bio->sector);
2526         bio = r1_bio->master_bio;
2527
2528         /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2529         r1_bio->state = 0;
2530         raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2531 }
2532
2533 static void raid1d(struct md_thread *thread)
2534 {
2535         struct mddev *mddev = thread->mddev;
2536         struct r1bio *r1_bio;
2537         unsigned long flags;
2538         struct r1conf *conf = mddev->private;
2539         struct list_head *head = &conf->retry_list;
2540         struct blk_plug plug;
2541         int idx;
2542
2543         md_check_recovery(mddev);
2544
2545         if (!list_empty_careful(&conf->bio_end_io_list) &&
2546             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2547                 LIST_HEAD(tmp);
2548                 spin_lock_irqsave(&conf->device_lock, flags);
2549                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2550                         list_splice_init(&conf->bio_end_io_list, &tmp);
2551                 spin_unlock_irqrestore(&conf->device_lock, flags);
2552                 while (!list_empty(&tmp)) {
2553                         r1_bio = list_first_entry(&tmp, struct r1bio,
2554                                                   retry_list);
2555                         list_del(&r1_bio->retry_list);
2556                         idx = sector_to_idx(r1_bio->sector);
2557                         atomic_dec(&conf->nr_queued[idx]);
2558                         if (mddev->degraded)
2559                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2560                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2561                                 close_write(r1_bio);
2562                         raid_end_bio_io(r1_bio);
2563                 }
2564         }
2565
2566         blk_start_plug(&plug);
2567         for (;;) {
2568
2569                 flush_pending_writes(conf);
2570
2571                 spin_lock_irqsave(&conf->device_lock, flags);
2572                 if (list_empty(head)) {
2573                         spin_unlock_irqrestore(&conf->device_lock, flags);
2574                         break;
2575                 }
2576                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2577                 list_del(head->prev);
2578                 idx = sector_to_idx(r1_bio->sector);
2579                 atomic_dec(&conf->nr_queued[idx]);
2580                 spin_unlock_irqrestore(&conf->device_lock, flags);
2581
2582                 mddev = r1_bio->mddev;
2583                 conf = mddev->private;
2584                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2585                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2586                             test_bit(R1BIO_WriteError, &r1_bio->state))
2587                                 handle_sync_write_finished(conf, r1_bio);
2588                         else
2589                                 sync_request_write(mddev, r1_bio);
2590                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2591                            test_bit(R1BIO_WriteError, &r1_bio->state))
2592                         handle_write_finished(conf, r1_bio);
2593                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2594                         handle_read_error(conf, r1_bio);
2595                 else
2596                         WARN_ON_ONCE(1);
2597
2598                 cond_resched();
2599                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2600                         md_check_recovery(mddev);
2601         }
2602         blk_finish_plug(&plug);
2603 }
2604
2605 static int init_resync(struct r1conf *conf)
2606 {
2607         int buffs;
2608
2609         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2610         BUG_ON(mempool_initialized(&conf->r1buf_pool));
2611
2612         return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2613                             r1buf_pool_free, conf->poolinfo);
2614 }
2615
2616 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2617 {
2618         struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2619         struct resync_pages *rps;
2620         struct bio *bio;
2621         int i;
2622
2623         for (i = conf->poolinfo->raid_disks; i--; ) {
2624                 bio = r1bio->bios[i];
2625                 rps = bio->bi_private;
2626                 bio_reset(bio);
2627                 bio->bi_private = rps;
2628         }
2629         r1bio->master_bio = NULL;
2630         return r1bio;
2631 }
2632
2633 /*
2634  * perform a "sync" on one "block"
2635  *
2636  * We need to make sure that no normal I/O request - particularly write
2637  * requests - conflict with active sync requests.
2638  *
2639  * This is achieved by tracking pending requests and a 'barrier' concept
2640  * that can be installed to exclude normal IO requests.
2641  */
2642
2643 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2644                                    int *skipped)
2645 {
2646         struct r1conf *conf = mddev->private;
2647         struct r1bio *r1_bio;
2648         struct bio *bio;
2649         sector_t max_sector, nr_sectors;
2650         int disk = -1;
2651         int i;
2652         int wonly = -1;
2653         int write_targets = 0, read_targets = 0;
2654         sector_t sync_blocks;
2655         int still_degraded = 0;
2656         int good_sectors = RESYNC_SECTORS;
2657         int min_bad = 0; /* number of sectors that are bad in all devices */
2658         int idx = sector_to_idx(sector_nr);
2659         int page_idx = 0;
2660
2661         if (!mempool_initialized(&conf->r1buf_pool))
2662                 if (init_resync(conf))
2663                         return 0;
2664
2665         max_sector = mddev->dev_sectors;
2666         if (sector_nr >= max_sector) {
2667                 /* If we aborted, we need to abort the
2668                  * sync on the 'current' bitmap chunk (there will
2669                  * only be one in raid1 resync.
2670                  * We can find the current addess in mddev->curr_resync
2671                  */
2672                 if (mddev->curr_resync < max_sector) /* aborted */
2673                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2674                                            &sync_blocks, 1);
2675                 else /* completed sync */
2676                         conf->fullsync = 0;
2677
2678                 md_bitmap_close_sync(mddev->bitmap);
2679                 close_sync(conf);
2680
2681                 if (mddev_is_clustered(mddev)) {
2682                         conf->cluster_sync_low = 0;
2683                         conf->cluster_sync_high = 0;
2684                 }
2685                 return 0;
2686         }
2687
2688         if (mddev->bitmap == NULL &&
2689             mddev->recovery_cp == MaxSector &&
2690             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2691             conf->fullsync == 0) {
2692                 *skipped = 1;
2693                 return max_sector - sector_nr;
2694         }
2695         /* before building a request, check if we can skip these blocks..
2696          * This call the bitmap_start_sync doesn't actually record anything
2697          */
2698         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2699             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2700                 /* We can skip this block, and probably several more */
2701                 *skipped = 1;
2702                 return sync_blocks;
2703         }
2704
2705         /*
2706          * If there is non-resync activity waiting for a turn, then let it
2707          * though before starting on this new sync request.
2708          */
2709         if (atomic_read(&conf->nr_waiting[idx]))
2710                 schedule_timeout_uninterruptible(1);
2711
2712         /* we are incrementing sector_nr below. To be safe, we check against
2713          * sector_nr + two times RESYNC_SECTORS
2714          */
2715
2716         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2717                 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2718
2719
2720         if (raise_barrier(conf, sector_nr))
2721                 return 0;
2722
2723         r1_bio = raid1_alloc_init_r1buf(conf);
2724
2725         rcu_read_lock();
2726         /*
2727          * If we get a correctably read error during resync or recovery,
2728          * we might want to read from a different device.  So we
2729          * flag all drives that could conceivably be read from for READ,
2730          * and any others (which will be non-In_sync devices) for WRITE.
2731          * If a read fails, we try reading from something else for which READ
2732          * is OK.
2733          */
2734
2735         r1_bio->mddev = mddev;
2736         r1_bio->sector = sector_nr;
2737         r1_bio->state = 0;
2738         set_bit(R1BIO_IsSync, &r1_bio->state);
2739         /* make sure good_sectors won't go across barrier unit boundary */
2740         good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2741
2742         for (i = 0; i < conf->raid_disks * 2; i++) {
2743                 struct md_rdev *rdev;
2744                 bio = r1_bio->bios[i];
2745
2746                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2747                 if (rdev == NULL ||
2748                     test_bit(Faulty, &rdev->flags)) {
2749                         if (i < conf->raid_disks)
2750                                 still_degraded = 1;
2751                 } else if (!test_bit(In_sync, &rdev->flags)) {
2752                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2753                         bio->bi_end_io = end_sync_write;
2754                         write_targets ++;
2755                 } else {
2756                         /* may need to read from here */
2757                         sector_t first_bad = MaxSector;
2758                         int bad_sectors;
2759
2760                         if (is_badblock(rdev, sector_nr, good_sectors,
2761                                         &first_bad, &bad_sectors)) {
2762                                 if (first_bad > sector_nr)
2763                                         good_sectors = first_bad - sector_nr;
2764                                 else {
2765                                         bad_sectors -= (sector_nr - first_bad);
2766                                         if (min_bad == 0 ||
2767                                             min_bad > bad_sectors)
2768                                                 min_bad = bad_sectors;
2769                                 }
2770                         }
2771                         if (sector_nr < first_bad) {
2772                                 if (test_bit(WriteMostly, &rdev->flags)) {
2773                                         if (wonly < 0)
2774                                                 wonly = i;
2775                                 } else {
2776                                         if (disk < 0)
2777                                                 disk = i;
2778                                 }
2779                                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2780                                 bio->bi_end_io = end_sync_read;
2781                                 read_targets++;
2782                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2783                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2784                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2785                                 /*
2786                                  * The device is suitable for reading (InSync),
2787                                  * but has bad block(s) here. Let's try to correct them,
2788                                  * if we are doing resync or repair. Otherwise, leave
2789                                  * this device alone for this sync request.
2790                                  */
2791                                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2792                                 bio->bi_end_io = end_sync_write;
2793                                 write_targets++;
2794                         }
2795                 }
2796                 if (rdev && bio->bi_end_io) {
2797                         atomic_inc(&rdev->nr_pending);
2798                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2799                         bio_set_dev(bio, rdev->bdev);
2800                         if (test_bit(FailFast, &rdev->flags))
2801                                 bio->bi_opf |= MD_FAILFAST;
2802                 }
2803         }
2804         rcu_read_unlock();
2805         if (disk < 0)
2806                 disk = wonly;
2807         r1_bio->read_disk = disk;
2808
2809         if (read_targets == 0 && min_bad > 0) {
2810                 /* These sectors are bad on all InSync devices, so we
2811                  * need to mark them bad on all write targets
2812                  */
2813                 int ok = 1;
2814                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2815                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2816                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2817                                 ok = rdev_set_badblocks(rdev, sector_nr,
2818                                                         min_bad, 0
2819                                         ) && ok;
2820                         }
2821                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2822                 *skipped = 1;
2823                 put_buf(r1_bio);
2824
2825                 if (!ok) {
2826                         /* Cannot record the badblocks, so need to
2827                          * abort the resync.
2828                          * If there are multiple read targets, could just
2829                          * fail the really bad ones ???
2830                          */
2831                         conf->recovery_disabled = mddev->recovery_disabled;
2832                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2833                         return 0;
2834                 } else
2835                         return min_bad;
2836
2837         }
2838         if (min_bad > 0 && min_bad < good_sectors) {
2839                 /* only resync enough to reach the next bad->good
2840                  * transition */
2841                 good_sectors = min_bad;
2842         }
2843
2844         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2845                 /* extra read targets are also write targets */
2846                 write_targets += read_targets-1;
2847
2848         if (write_targets == 0 || read_targets == 0) {
2849                 /* There is nowhere to write, so all non-sync
2850                  * drives must be failed - so we are finished
2851                  */
2852                 sector_t rv;
2853                 if (min_bad > 0)
2854                         max_sector = sector_nr + min_bad;
2855                 rv = max_sector - sector_nr;
2856                 *skipped = 1;
2857                 put_buf(r1_bio);
2858                 return rv;
2859         }
2860
2861         if (max_sector > mddev->resync_max)
2862                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2863         if (max_sector > sector_nr + good_sectors)
2864                 max_sector = sector_nr + good_sectors;
2865         nr_sectors = 0;
2866         sync_blocks = 0;
2867         do {
2868                 struct page *page;
2869                 int len = PAGE_SIZE;
2870                 if (sector_nr + (len>>9) > max_sector)
2871                         len = (max_sector - sector_nr) << 9;
2872                 if (len == 0)
2873                         break;
2874                 if (sync_blocks == 0) {
2875                         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2876                                                   &sync_blocks, still_degraded) &&
2877                             !conf->fullsync &&
2878                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2879                                 break;
2880                         if ((len >> 9) > sync_blocks)
2881                                 len = sync_blocks<<9;
2882                 }
2883
2884                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2885                         struct resync_pages *rp;
2886
2887                         bio = r1_bio->bios[i];
2888                         rp = get_resync_pages(bio);
2889                         if (bio->bi_end_io) {
2890                                 page = resync_fetch_page(rp, page_idx);
2891
2892                                 /*
2893                                  * won't fail because the vec table is big
2894                                  * enough to hold all these pages
2895                                  */
2896                                 bio_add_page(bio, page, len, 0);
2897                         }
2898                 }
2899                 nr_sectors += len>>9;
2900                 sector_nr += len>>9;
2901                 sync_blocks -= (len>>9);
2902         } while (++page_idx < RESYNC_PAGES);
2903
2904         r1_bio->sectors = nr_sectors;
2905
2906         if (mddev_is_clustered(mddev) &&
2907                         conf->cluster_sync_high < sector_nr + nr_sectors) {
2908                 conf->cluster_sync_low = mddev->curr_resync_completed;
2909                 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2910                 /* Send resync message */
2911                 md_cluster_ops->resync_info_update(mddev,
2912                                 conf->cluster_sync_low,
2913                                 conf->cluster_sync_high);
2914         }
2915
2916         /* For a user-requested sync, we read all readable devices and do a
2917          * compare
2918          */
2919         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2920                 atomic_set(&r1_bio->remaining, read_targets);
2921                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2922                         bio = r1_bio->bios[i];
2923                         if (bio->bi_end_io == end_sync_read) {
2924                                 read_targets--;
2925                                 md_sync_acct_bio(bio, nr_sectors);
2926                                 if (read_targets == 1)
2927                                         bio->bi_opf &= ~MD_FAILFAST;
2928                                 generic_make_request(bio);
2929                         }
2930                 }
2931         } else {
2932                 atomic_set(&r1_bio->remaining, 1);
2933                 bio = r1_bio->bios[r1_bio->read_disk];
2934                 md_sync_acct_bio(bio, nr_sectors);
2935                 if (read_targets == 1)
2936                         bio->bi_opf &= ~MD_FAILFAST;
2937                 generic_make_request(bio);
2938         }
2939         return nr_sectors;
2940 }
2941
2942 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2943 {
2944         if (sectors)
2945                 return sectors;
2946
2947         return mddev->dev_sectors;
2948 }
2949
2950 static struct r1conf *setup_conf(struct mddev *mddev)
2951 {
2952         struct r1conf *conf;
2953         int i;
2954         struct raid1_info *disk;
2955         struct md_rdev *rdev;
2956         int err = -ENOMEM;
2957
2958         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2959         if (!conf)
2960                 goto abort;
2961
2962         conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2963                                    sizeof(atomic_t), GFP_KERNEL);
2964         if (!conf->nr_pending)
2965                 goto abort;
2966
2967         conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2968                                    sizeof(atomic_t), GFP_KERNEL);
2969         if (!conf->nr_waiting)
2970                 goto abort;
2971
2972         conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2973                                   sizeof(atomic_t), GFP_KERNEL);
2974         if (!conf->nr_queued)
2975                 goto abort;
2976
2977         conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2978                                 sizeof(atomic_t), GFP_KERNEL);
2979         if (!conf->barrier)
2980                 goto abort;
2981
2982         conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2983                                             mddev->raid_disks, 2),
2984                                 GFP_KERNEL);
2985         if (!conf->mirrors)
2986                 goto abort;
2987
2988         conf->tmppage = alloc_page(GFP_KERNEL);
2989         if (!conf->tmppage)
2990                 goto abort;
2991
2992         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2993         if (!conf->poolinfo)
2994                 goto abort;
2995         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2996         err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2997                            rbio_pool_free, conf->poolinfo);
2998         if (err)
2999                 goto abort;
3000
3001         err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3002         if (err)
3003                 goto abort;
3004
3005         conf->poolinfo->mddev = mddev;
3006
3007         err = -EINVAL;
3008         spin_lock_init(&conf->device_lock);
3009         rdev_for_each(rdev, mddev) {
3010                 int disk_idx = rdev->raid_disk;
3011                 if (disk_idx >= mddev->raid_disks
3012                     || disk_idx < 0)
3013                         continue;
3014                 if (test_bit(Replacement, &rdev->flags))
3015                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
3016                 else
3017                         disk = conf->mirrors + disk_idx;
3018
3019                 if (disk->rdev)
3020                         goto abort;
3021                 disk->rdev = rdev;
3022                 disk->head_position = 0;
3023                 disk->seq_start = MaxSector;
3024         }
3025         conf->raid_disks = mddev->raid_disks;
3026         conf->mddev = mddev;
3027         INIT_LIST_HEAD(&conf->retry_list);
3028         INIT_LIST_HEAD(&conf->bio_end_io_list);
3029
3030         spin_lock_init(&conf->resync_lock);
3031         init_waitqueue_head(&conf->wait_barrier);
3032
3033         bio_list_init(&conf->pending_bio_list);
3034         conf->pending_count = 0;
3035         conf->recovery_disabled = mddev->recovery_disabled - 1;
3036
3037         err = -EIO;
3038         for (i = 0; i < conf->raid_disks * 2; i++) {
3039
3040                 disk = conf->mirrors + i;
3041
3042                 if (i < conf->raid_disks &&
3043                     disk[conf->raid_disks].rdev) {
3044                         /* This slot has a replacement. */
3045                         if (!disk->rdev) {
3046                                 /* No original, just make the replacement
3047                                  * a recovering spare
3048                                  */
3049                                 disk->rdev =
3050                                         disk[conf->raid_disks].rdev;
3051                                 disk[conf->raid_disks].rdev = NULL;
3052                         } else if (!test_bit(In_sync, &disk->rdev->flags))
3053                                 /* Original is not in_sync - bad */
3054                                 goto abort;
3055                 }
3056
3057                 if (!disk->rdev ||
3058                     !test_bit(In_sync, &disk->rdev->flags)) {
3059                         disk->head_position = 0;
3060                         if (disk->rdev &&
3061                             (disk->rdev->saved_raid_disk < 0))
3062                                 conf->fullsync = 1;
3063                 }
3064         }
3065
3066         err = -ENOMEM;
3067         conf->thread = md_register_thread(raid1d, mddev, "raid1");
3068         if (!conf->thread)
3069                 goto abort;
3070
3071         return conf;
3072
3073  abort:
3074         if (conf) {
3075                 mempool_exit(&conf->r1bio_pool);
3076                 kfree(conf->mirrors);
3077                 safe_put_page(conf->tmppage);
3078                 kfree(conf->poolinfo);
3079                 kfree(conf->nr_pending);
3080                 kfree(conf->nr_waiting);
3081                 kfree(conf->nr_queued);
3082                 kfree(conf->barrier);
3083                 bioset_exit(&conf->bio_split);
3084                 kfree(conf);
3085         }
3086         return ERR_PTR(err);
3087 }
3088
3089 static void raid1_free(struct mddev *mddev, void *priv);
3090 static int raid1_run(struct mddev *mddev)
3091 {
3092         struct r1conf *conf;
3093         int i;
3094         struct md_rdev *rdev;
3095         int ret;
3096         bool discard_supported = false;
3097
3098         if (mddev->level != 1) {
3099                 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3100                         mdname(mddev), mddev->level);
3101                 return -EIO;
3102         }
3103         if (mddev->reshape_position != MaxSector) {
3104                 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3105                         mdname(mddev));
3106                 return -EIO;
3107         }
3108         if (mddev_init_writes_pending(mddev) < 0)
3109                 return -ENOMEM;
3110         /*
3111          * copy the already verified devices into our private RAID1
3112          * bookkeeping area. [whatever we allocate in run(),
3113          * should be freed in raid1_free()]
3114          */
3115         if (mddev->private == NULL)
3116                 conf = setup_conf(mddev);
3117         else
3118                 conf = mddev->private;
3119
3120         if (IS_ERR(conf))
3121                 return PTR_ERR(conf);
3122
3123         if (mddev->queue) {
3124                 blk_queue_max_write_same_sectors(mddev->queue, 0);
3125                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3126         }
3127
3128         rdev_for_each(rdev, mddev) {
3129                 if (!mddev->gendisk)
3130                         continue;
3131                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3132                                   rdev->data_offset << 9);
3133                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3134                         discard_supported = true;
3135         }
3136
3137         mddev->degraded = 0;
3138         for (i = 0; i < conf->raid_disks; i++)
3139                 if (conf->mirrors[i].rdev == NULL ||
3140                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3141                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3142                         mddev->degraded++;
3143         /*
3144          * RAID1 needs at least one disk in active
3145          */
3146         if (conf->raid_disks - mddev->degraded < 1) {
3147                 ret = -EINVAL;
3148                 goto abort;
3149         }
3150
3151         if (conf->raid_disks - mddev->degraded == 1)
3152                 mddev->recovery_cp = MaxSector;
3153
3154         if (mddev->recovery_cp != MaxSector)
3155                 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3156                         mdname(mddev));
3157         pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3158                 mdname(mddev), mddev->raid_disks - mddev->degraded,
3159                 mddev->raid_disks);
3160
3161         /*
3162          * Ok, everything is just fine now
3163          */
3164         mddev->thread = conf->thread;
3165         conf->thread = NULL;
3166         mddev->private = conf;
3167         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3168
3169         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3170
3171         if (mddev->queue) {
3172                 if (discard_supported)
3173                         blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3174                                                 mddev->queue);
3175                 else
3176                         blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3177                                                   mddev->queue);
3178         }
3179
3180         ret = md_integrity_register(mddev);
3181         if (ret) {
3182                 md_unregister_thread(&mddev->thread);
3183                 goto abort;
3184         }
3185         return 0;
3186
3187 abort:
3188         raid1_free(mddev, conf);
3189         return ret;
3190 }
3191
3192 static void raid1_free(struct mddev *mddev, void *priv)
3193 {
3194         struct r1conf *conf = priv;
3195
3196         mempool_exit(&conf->r1bio_pool);
3197         kfree(conf->mirrors);
3198         safe_put_page(conf->tmppage);
3199         kfree(conf->poolinfo);
3200         kfree(conf->nr_pending);
3201         kfree(conf->nr_waiting);
3202         kfree(conf->nr_queued);
3203         kfree(conf->barrier);
3204         bioset_exit(&conf->bio_split);
3205         kfree(conf);
3206 }
3207
3208 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3209 {
3210         /* no resync is happening, and there is enough space
3211          * on all devices, so we can resize.
3212          * We need to make sure resync covers any new space.
3213          * If the array is shrinking we should possibly wait until
3214          * any io in the removed space completes, but it hardly seems
3215          * worth it.
3216          */
3217         sector_t newsize = raid1_size(mddev, sectors, 0);
3218         if (mddev->external_size &&
3219             mddev->array_sectors > newsize)
3220                 return -EINVAL;
3221         if (mddev->bitmap) {
3222                 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3223                 if (ret)
3224                         return ret;
3225         }
3226         md_set_array_sectors(mddev, newsize);
3227         if (sectors > mddev->dev_sectors &&
3228             mddev->recovery_cp > mddev->dev_sectors) {
3229                 mddev->recovery_cp = mddev->dev_sectors;
3230                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3231         }
3232         mddev->dev_sectors = sectors;
3233         mddev->resync_max_sectors = sectors;
3234         return 0;
3235 }
3236
3237 static int raid1_reshape(struct mddev *mddev)
3238 {
3239         /* We need to:
3240          * 1/ resize the r1bio_pool
3241          * 2/ resize conf->mirrors
3242          *
3243          * We allocate a new r1bio_pool if we can.
3244          * Then raise a device barrier and wait until all IO stops.
3245          * Then resize conf->mirrors and swap in the new r1bio pool.
3246          *
3247          * At the same time, we "pack" the devices so that all the missing
3248          * devices have the higher raid_disk numbers.
3249          */
3250         mempool_t newpool, oldpool;
3251         struct pool_info *newpoolinfo;
3252         struct raid1_info *newmirrors;
3253         struct r1conf *conf = mddev->private;
3254         int cnt, raid_disks;
3255         unsigned long flags;
3256         int d, d2;
3257         int ret;
3258
3259         memset(&newpool, 0, sizeof(newpool));
3260         memset(&oldpool, 0, sizeof(oldpool));
3261
3262         /* Cannot change chunk_size, layout, or level */
3263         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3264             mddev->layout != mddev->new_layout ||
3265             mddev->level != mddev->new_level) {
3266                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3267                 mddev->new_layout = mddev->layout;
3268                 mddev->new_level = mddev->level;
3269                 return -EINVAL;
3270         }
3271
3272         if (!mddev_is_clustered(mddev))
3273                 md_allow_write(mddev);
3274
3275         raid_disks = mddev->raid_disks + mddev->delta_disks;
3276
3277         if (raid_disks < conf->raid_disks) {
3278                 cnt=0;
3279                 for (d= 0; d < conf->raid_disks; d++)
3280                         if (conf->mirrors[d].rdev)
3281                                 cnt++;
3282                 if (cnt > raid_disks)
3283                         return -EBUSY;
3284         }
3285
3286         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3287         if (!newpoolinfo)
3288                 return -ENOMEM;
3289         newpoolinfo->mddev = mddev;
3290         newpoolinfo->raid_disks = raid_disks * 2;
3291
3292         ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3293                            rbio_pool_free, newpoolinfo);
3294         if (ret) {
3295                 kfree(newpoolinfo);
3296                 return ret;
3297         }
3298         newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3299                                          raid_disks, 2),
3300                              GFP_KERNEL);
3301         if (!newmirrors) {
3302                 kfree(newpoolinfo);
3303                 mempool_exit(&newpool);
3304                 return -ENOMEM;
3305         }
3306
3307         freeze_array(conf, 0);
3308
3309         /* ok, everything is stopped */
3310         oldpool = conf->r1bio_pool;
3311         conf->r1bio_pool = newpool;
3312
3313         for (d = d2 = 0; d < conf->raid_disks; d++) {
3314                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3315                 if (rdev && rdev->raid_disk != d2) {
3316                         sysfs_unlink_rdev(mddev, rdev);
3317                         rdev->raid_disk = d2;
3318                         sysfs_unlink_rdev(mddev, rdev);
3319                         if (sysfs_link_rdev(mddev, rdev))
3320                                 pr_warn("md/raid1:%s: cannot register rd%d\n",
3321                                         mdname(mddev), rdev->raid_disk);
3322                 }
3323                 if (rdev)
3324                         newmirrors[d2++].rdev = rdev;
3325         }
3326         kfree(conf->mirrors);
3327         conf->mirrors = newmirrors;
3328         kfree(conf->poolinfo);
3329         conf->poolinfo = newpoolinfo;
3330
3331         spin_lock_irqsave(&conf->device_lock, flags);
3332         mddev->degraded += (raid_disks - conf->raid_disks);
3333         spin_unlock_irqrestore(&conf->device_lock, flags);
3334         conf->raid_disks = mddev->raid_disks = raid_disks;
3335         mddev->delta_disks = 0;
3336
3337         unfreeze_array(conf);
3338
3339         set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3340         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3341         md_wakeup_thread(mddev->thread);
3342
3343         mempool_exit(&oldpool);
3344         return 0;
3345 }
3346
3347 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3348 {
3349         struct r1conf *conf = mddev->private;
3350
3351         if (quiesce)
3352                 freeze_array(conf, 0);
3353         else
3354                 unfreeze_array(conf);
3355 }
3356
3357 static void *raid1_takeover(struct mddev *mddev)
3358 {
3359         /* raid1 can take over:
3360          *  raid5 with 2 devices, any layout or chunk size
3361          */
3362         if (mddev->level == 5 && mddev->raid_disks == 2) {
3363                 struct r1conf *conf;
3364                 mddev->new_level = 1;
3365                 mddev->new_layout = 0;
3366                 mddev->new_chunk_sectors = 0;
3367                 conf = setup_conf(mddev);
3368                 if (!IS_ERR(conf)) {
3369                         /* Array must appear to be quiesced */
3370                         conf->array_frozen = 1;
3371                         mddev_clear_unsupported_flags(mddev,
3372                                 UNSUPPORTED_MDDEV_FLAGS);
3373                 }
3374                 return conf;
3375         }
3376         return ERR_PTR(-EINVAL);
3377 }
3378
3379 static struct md_personality raid1_personality =
3380 {
3381         .name           = "raid1",
3382         .level          = 1,
3383         .owner          = THIS_MODULE,
3384         .make_request   = raid1_make_request,
3385         .run            = raid1_run,
3386         .free           = raid1_free,
3387         .status         = raid1_status,
3388         .error_handler  = raid1_error,
3389         .hot_add_disk   = raid1_add_disk,
3390         .hot_remove_disk= raid1_remove_disk,
3391         .spare_active   = raid1_spare_active,
3392         .sync_request   = raid1_sync_request,
3393         .resize         = raid1_resize,
3394         .size           = raid1_size,
3395         .check_reshape  = raid1_reshape,
3396         .quiesce        = raid1_quiesce,
3397         .takeover       = raid1_takeover,
3398         .congested      = raid1_congested,
3399 };
3400
3401 static int __init raid_init(void)
3402 {
3403         return register_md_personality(&raid1_personality);
3404 }
3405
3406 static void raid_exit(void)
3407 {
3408         unregister_md_personality(&raid1_personality);
3409 }
3410
3411 module_init(raid_init);
3412 module_exit(raid_exit);
3413 MODULE_LICENSE("GPL");
3414 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3415 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3416 MODULE_ALIAS("md-raid1");
3417 MODULE_ALIAS("md-level-1");
3418
3419 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);