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
 300         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 301                 bio->bi_status = BLK_STS_IOERR;
 302
 303         if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
 304                 bio_end_io_acct(bio, r1_bio->start_time);
 305         bio_endio(bio);
 306 }
 307
 308 static void raid_end_bio_io(struct r1bio *r1_bio)
 309 {
 310         struct bio *bio = r1_bio->master_bio;
 311         struct r1conf *conf = r1_bio->mddev->private;
 312
 313         /* if nobody has done the final endio yet, do it now */
 314         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 315                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 316                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 317                          (unsigned long long) bio->bi_iter.bi_sector,
 318                          (unsigned long long) bio_end_sector(bio) - 1);
 319
 320                 call_bio_endio(r1_bio);
 321         }
 322         /*
 323          * Wake up any possible resync thread that waits for the device
 324          * to go idle.  All I/Os, even write-behind writes, are done.
 325          */
 326         allow_barrier(conf, r1_bio->sector);
 327
 328         free_r1bio(r1_bio);
 329 }
 330
 331 /*
 332  * Update disk head position estimator based on IRQ completion info.
 333  */
 334 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 335 {
 336         struct r1conf *conf = r1_bio->mddev->private;
 337
 338         conf->mirrors[disk].head_position =
 339                 r1_bio->sector + (r1_bio->sectors);
 340 }
 341
 342 /*
 343  * Find the disk number which triggered given bio
 344  */
 345 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 346 {
 347         int mirror;
 348         struct r1conf *conf = r1_bio->mddev->private;
 349         int raid_disks = conf->raid_disks;
 350
 351         for (mirror = 0; mirror < raid_disks * 2; mirror++)
 352                 if (r1_bio->bios[mirror] == bio)
 353                         break;
 354
 355         BUG_ON(mirror == raid_disks * 2);
 356         update_head_pos(mirror, r1_bio);
 357
 358         return mirror;
 359 }
 360
 361 static void raid1_end_read_request(struct bio *bio)
 362 {
 363         int uptodate = !bio->bi_status;
 364         struct r1bio *r1_bio = bio->bi_private;
 365         struct r1conf *conf = r1_bio->mddev->private;
 366         struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 367
 368         /*
 369          * this branch is our 'one mirror IO has finished' event handler:
 370          */
 371         update_head_pos(r1_bio->read_disk, r1_bio);
 372
 373         if (uptodate)
 374                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 375         else if (test_bit(FailFast, &rdev->flags) &&
 376                  test_bit(R1BIO_FailFast, &r1_bio->state))
 377                 /* This was a fail-fast read so we definitely
 378                  * want to retry */
 379                 ;
 380         else {
 381                 /* If all other devices have failed, we want to return
 382                  * the error upwards rather than fail the last device.
 383                  * Here we redefine "uptodate" to mean "Don't want to retry"
 384                  */
 385                 unsigned long flags;
 386                 spin_lock_irqsave(&conf->device_lock, flags);
 387                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 388                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 389                      test_bit(In_sync, &rdev->flags)))
 390                         uptodate = 1;
 391                 spin_unlock_irqrestore(&conf->device_lock, flags);
 392         }
 393
 394         if (uptodate) {
 395                 raid_end_bio_io(r1_bio);
 396                 rdev_dec_pending(rdev, conf->mddev);
 397         } else {
 398                 /*
 399                  * oops, read error:
 400                  */
 401                 char b[BDEVNAME_SIZE];
 402                 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
 403                                    mdname(conf->mddev),
 404                                    bdevname(rdev->bdev, b),
 405                                    (unsigned long long)r1_bio->sector);
 406                 set_bit(R1BIO_ReadError, &r1_bio->state);
 407                 reschedule_retry(r1_bio);
 408                 /* don't drop the reference on read_disk yet */
 409         }
 410 }
 411
 412 static void close_write(struct r1bio *r1_bio)
 413 {
 414         /* it really is the end of this request */
 415         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 416                 bio_free_pages(r1_bio->behind_master_bio);
 417                 bio_put(r1_bio->behind_master_bio);
 418                 r1_bio->behind_master_bio = NULL;
 419         }
 420         /* clear the bitmap if all writes complete successfully */
 421         md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 422                            r1_bio->sectors,
 423                            !test_bit(R1BIO_Degraded, &r1_bio->state),
 424                            test_bit(R1BIO_BehindIO, &r1_bio->state));
 425         md_write_end(r1_bio->mddev);
 426 }
 427
 428 static void r1_bio_write_done(struct r1bio *r1_bio)
 429 {
 430         if (!atomic_dec_and_test(&r1_bio->remaining))
 431                 return;
 432
 433         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 434                 reschedule_retry(r1_bio);
 435         else {
 436                 close_write(r1_bio);
 437                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 438                         reschedule_retry(r1_bio);
 439                 else
 440                         raid_end_bio_io(r1_bio);
 441         }
 442 }
 443
 444 static void raid1_end_write_request(struct bio *bio)
 445 {
 446         struct r1bio *r1_bio = bio->bi_private;
 447         int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 448         struct r1conf *conf = r1_bio->mddev->private;
 449         struct bio *to_put = NULL;
 450         int mirror = find_bio_disk(r1_bio, bio);
 451         struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 452         bool discard_error;
 453         sector_t lo = r1_bio->sector;
 454         sector_t hi = r1_bio->sector + r1_bio->sectors;
 455
 456         discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 457
 458         /*
 459          * 'one mirror IO has finished' event handler:
 460          */
 461         if (bio->bi_status && !discard_error) {
 462                 set_bit(WriteErrorSeen, &rdev->flags);
 463                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
 464                         set_bit(MD_RECOVERY_NEEDED, &
 465                                 conf->mddev->recovery);
 466
 467                 if (test_bit(FailFast, &rdev->flags) &&
 468                     (bio->bi_opf & MD_FAILFAST) &&
 469                     /* We never try FailFast to WriteMostly devices */
 470                     !test_bit(WriteMostly, &rdev->flags)) {
 471                         md_error(r1_bio->mddev, rdev);
 472                 }
 473
 474                 /*
 475                  * When the device is faulty, it is not necessary to
 476                  * handle write error.
 477                  */
 478                 if (!test_bit(Faulty, &rdev->flags))
 479                         set_bit(R1BIO_WriteError, &r1_bio->state);
 480                 else {
 481                         /* Fail the request */
 482                         set_bit(R1BIO_Degraded, &r1_bio->state);
 483                         /* Finished with this branch */
 484                         r1_bio->bios[mirror] = NULL;
 485                         to_put = bio;
 486                 }
 487         } else {
 488                 /*
 489                  * Set R1BIO_Uptodate in our master bio, so that we
 490                  * will return a good error code for to the higher
 491                  * levels even if IO on some other mirrored buffer
 492                  * fails.
 493                  *
 494                  * The 'master' represents the composite IO operation
 495                  * to user-side. So if something waits for IO, then it
 496                  * will wait for the 'master' bio.
 497                  */
 498                 sector_t first_bad;
 499                 int bad_sectors;
 500
 501                 r1_bio->bios[mirror] = NULL;
 502                 to_put = bio;
 503                 /*
 504                  * Do not set R1BIO_Uptodate if the current device is
 505                  * rebuilding or Faulty. This is because we cannot use
 506                  * such device for properly reading the data back (we could
 507                  * potentially use it, if the current write would have felt
 508                  * before rdev->recovery_offset, but for simplicity we don't
 509                  * check this here.
 510                  */
 511                 if (test_bit(In_sync, &rdev->flags) &&
 512                     !test_bit(Faulty, &rdev->flags))
 513                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 514
 515                 /* Maybe we can clear some bad blocks. */
 516                 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 517                                 &first_bad, &bad_sectors) && !discard_error) {
 518                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 519                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 520                 }
 521         }
 522
 523         if (behind) {
 524                 if (test_bit(CollisionCheck, &rdev->flags))
 525                         remove_serial(rdev, lo, hi);
 526                 if (test_bit(WriteMostly, &rdev->flags))
 527                         atomic_dec(&r1_bio->behind_remaining);
 528
 529                 /*
 530                  * In behind mode, we ACK the master bio once the I/O
 531                  * has safely reached all non-writemostly
 532                  * disks. Setting the Returned bit ensures that this
 533                  * gets done only once -- we don't ever want to return
 534                  * -EIO here, instead we'll wait
 535                  */
 536                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 537                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 538                         /* Maybe we can return now */
 539                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 540                                 struct bio *mbio = r1_bio->master_bio;
 541                                 pr_debug("raid1: behind end write sectors"
 542                                          " %llu-%llu\n",
 543                                          (unsigned long long) mbio->bi_iter.bi_sector,
 544                                          (unsigned long long) bio_end_sector(mbio) - 1);
 545                                 call_bio_endio(r1_bio);
 546                         }
 547                 }
 548         } else if (rdev->mddev->serialize_policy)
 549                 remove_serial(rdev, lo, hi);
 550         if (r1_bio->bios[mirror] == NULL)
 551                 rdev_dec_pending(rdev, conf->mddev);
 552
 553         /*
 554          * Let's see if all mirrored write operations have finished
 555          * already.
 556          */
 557         r1_bio_write_done(r1_bio);
 558
 559         if (to_put)
 560                 bio_put(to_put);
 561 }
 562
 563 static sector_t align_to_barrier_unit_end(sector_t start_sector,
 564                                           sector_t sectors)
 565 {
 566         sector_t len;
 567
 568         WARN_ON(sectors == 0);
 569         /*
 570          * len is the number of sectors from start_sector to end of the
 571          * barrier unit which start_sector belongs to.
 572          */
 573         len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 574               start_sector;
 575
 576         if (len > sectors)
 577                 len = sectors;
 578
 579         return len;
 580 }
 581
 582 /*
 583  * This routine returns the disk from which the requested read should
 584  * be done. There is a per-array 'next expected sequential IO' sector
 585  * number - if this matches on the next IO then we use the last disk.
 586  * There is also a per-disk 'last know head position' sector that is
 587  * maintained from IRQ contexts, both the normal and the resync IO
 588  * completion handlers update this position correctly. If there is no
 589  * perfect sequential match then we pick the disk whose head is closest.
 590  *
 591  * If there are 2 mirrors in the same 2 devices, performance degrades
 592  * because position is mirror, not device based.
 593  *
 594  * The rdev for the device selected will have nr_pending incremented.
 595  */
 596 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 597 {
 598         const sector_t this_sector = r1_bio->sector;
 599         int sectors;
 600         int best_good_sectors;
 601         int best_disk, best_dist_disk, best_pending_disk;
 602         int has_nonrot_disk;
 603         int disk;
 604         sector_t best_dist;
 605         unsigned int min_pending;
 606         struct md_rdev *rdev;
 607         int choose_first;
 608         int choose_next_idle;
 609
 610         rcu_read_lock();
 611         /*
 612          * Check if we can balance. We can balance on the whole
 613          * device if no resync is going on, or below the resync window.
 614          * We take the first readable disk when above the resync window.
 615          */
 616  retry:
 617         sectors = r1_bio->sectors;
 618         best_disk = -1;
 619         best_dist_disk = -1;
 620         best_dist = MaxSector;
 621         best_pending_disk = -1;
 622         min_pending = UINT_MAX;
 623         best_good_sectors = 0;
 624         has_nonrot_disk = 0;
 625         choose_next_idle = 0;
 626         clear_bit(R1BIO_FailFast, &r1_bio->state);
 627
 628         if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 629             (mddev_is_clustered(conf->mddev) &&
 630             md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 631                     this_sector + sectors)))
 632                 choose_first = 1;
 633         else
 634                 choose_first = 0;
 635
 636         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 637                 sector_t dist;
 638                 sector_t first_bad;
 639                 int bad_sectors;
 640                 unsigned int pending;
 641                 bool nonrot;
 642
 643                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 644                 if (r1_bio->bios[disk] == IO_BLOCKED
 645                     || rdev == NULL
 646                     || test_bit(Faulty, &rdev->flags))
 647                         continue;
 648                 if (!test_bit(In_sync, &rdev->flags) &&
 649                     rdev->recovery_offset < this_sector + sectors)
 650                         continue;
 651                 if (test_bit(WriteMostly, &rdev->flags)) {
 652                         /* Don't balance among write-mostly, just
 653                          * use the first as a last resort */
 654                         if (best_dist_disk < 0) {
 655                                 if (is_badblock(rdev, this_sector, sectors,
 656                                                 &first_bad, &bad_sectors)) {
 657                                         if (first_bad <= this_sector)
 658                                                 /* Cannot use this */
 659                                                 continue;
 660                                         best_good_sectors = first_bad - this_sector;
 661                                 } else
 662                                         best_good_sectors = sectors;
 663                                 best_dist_disk = disk;
 664                                 best_pending_disk = disk;
 665                         }
 666                         continue;
 667                 }
 668                 /* This is a reasonable device to use.  It might
 669                  * even be best.
 670                  */
 671                 if (is_badblock(rdev, this_sector, sectors,
 672                                 &first_bad, &bad_sectors)) {
 673                         if (best_dist < MaxSector)
 674                                 /* already have a better device */
 675                                 continue;
 676                         if (first_bad <= this_sector) {
 677                                 /* cannot read here. If this is the 'primary'
 678                                  * device, then we must not read beyond
 679                                  * bad_sectors from another device..
 680                                  */
 681                                 bad_sectors -= (this_sector - first_bad);
 682                                 if (choose_first && sectors > bad_sectors)
 683                                         sectors = bad_sectors;
 684                                 if (best_good_sectors > sectors)
 685                                         best_good_sectors = sectors;
 686
 687                         } else {
 688                                 sector_t good_sectors = first_bad - this_sector;
 689                                 if (good_sectors > best_good_sectors) {
 690                                         best_good_sectors = good_sectors;
 691                                         best_disk = disk;
 692                                 }
 693                                 if (choose_first)
 694                                         break;
 695                         }
 696                         continue;
 697                 } else {
 698                         if ((sectors > best_good_sectors) && (best_disk >= 0))
 699                                 best_disk = -1;
 700                         best_good_sectors = sectors;
 701                 }
 702
 703                 if (best_disk >= 0)
 704                         /* At least two disks to choose from so failfast is OK */
 705                         set_bit(R1BIO_FailFast, &r1_bio->state);
 706
 707                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 708                 has_nonrot_disk |= nonrot;
 709                 pending = atomic_read(&rdev->nr_pending);
 710                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 711                 if (choose_first) {
 712                         best_disk = disk;
 713                         break;
 714                 }
 715                 /* Don't change to another disk for sequential reads */
 716                 if (conf->mirrors[disk].next_seq_sect == this_sector
 717                     || dist == 0) {
 718                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 719                         struct raid1_info *mirror = &conf->mirrors[disk];
 720
 721                         best_disk = disk;
 722                         /*
 723                          * If buffered sequential IO size exceeds optimal
 724                          * iosize, check if there is idle disk. If yes, choose
 725                          * the idle disk. read_balance could already choose an
 726                          * idle disk before noticing it's a sequential IO in
 727                          * this disk. This doesn't matter because this disk
 728                          * will idle, next time it will be utilized after the
 729                          * first disk has IO size exceeds optimal iosize. In
 730                          * this way, iosize of the first disk will be optimal
 731                          * iosize at least. iosize of the second disk might be
 732                          * small, but not a big deal since when the second disk
 733                          * starts IO, the first disk is likely still busy.
 734                          */
 735                         if (nonrot && opt_iosize > 0 &&
 736                             mirror->seq_start != MaxSector &&
 737                             mirror->next_seq_sect > opt_iosize &&
 738                             mirror->next_seq_sect - opt_iosize >=
 739                             mirror->seq_start) {
 740                                 choose_next_idle = 1;
 741                                 continue;
 742                         }
 743                         break;
 744                 }
 745
 746                 if (choose_next_idle)
 747                         continue;
 748
 749                 if (min_pending > pending) {
 750                         min_pending = pending;
 751                         best_pending_disk = disk;
 752                 }
 753
 754                 if (dist < best_dist) {
 755                         best_dist = dist;
 756                         best_dist_disk = disk;
 757                 }
 758         }
 759
 760         /*
 761          * If all disks are rotational, choose the closest disk. If any disk is
 762          * non-rotational, choose the disk with less pending request even the
 763          * disk is rotational, which might/might not be optimal for raids with
 764          * mixed ratation/non-rotational disks depending on workload.
 765          */
 766         if (best_disk == -1) {
 767                 if (has_nonrot_disk || min_pending == 0)
 768                         best_disk = best_pending_disk;
 769                 else
 770                         best_disk = best_dist_disk;
 771         }
 772
 773         if (best_disk >= 0) {
 774                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 775                 if (!rdev)
 776                         goto retry;
 777                 atomic_inc(&rdev->nr_pending);
 778                 sectors = best_good_sectors;
 779
 780                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 781                         conf->mirrors[best_disk].seq_start = this_sector;
 782
 783                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 784         }
 785         rcu_read_unlock();
 786         *max_sectors = sectors;
 787
 788         return best_disk;
 789 }
 790
 791 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 792 {
 793         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
 794         md_bitmap_unplug(conf->mddev->bitmap);
 795         wake_up(&conf->wait_barrier);
 796
 797         while (bio) { /* submit pending writes */
 798                 struct bio *next = bio->bi_next;
 799                 struct md_rdev *rdev = (void *)bio->bi_bdev;
 800                 bio->bi_next = NULL;
 801                 bio_set_dev(bio, rdev->bdev);
 802                 if (test_bit(Faulty, &rdev->flags)) {
 803                         bio_io_error(bio);
 804                 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
 805                                     !blk_queue_discard(bio->bi_bdev->bd_disk->queue)))
 806                         /* Just ignore it */
 807                         bio_endio(bio);
 808                 else
 809                         submit_bio_noacct(bio);
 810                 bio = next;
 811                 cond_resched();
 812         }
 813 }
 814
 815 static void flush_pending_writes(struct r1conf *conf)
 816 {
 817         /* Any writes that have been queued but are awaiting
 818          * bitmap updates get flushed here.
 819          */
 820         spin_lock_irq(&conf->device_lock);
 821
 822         if (conf->pending_bio_list.head) {
 823                 struct blk_plug plug;
 824                 struct bio *bio;
 825
 826                 bio = bio_list_get(&conf->pending_bio_list);
 827                 conf->pending_count = 0;
 828                 spin_unlock_irq(&conf->device_lock);
 829
 830                 /*
 831                  * As this is called in a wait_event() loop (see freeze_array),
 832                  * current->state might be TASK_UNINTERRUPTIBLE which will
 833                  * cause a warning when we prepare to wait again.  As it is
 834                  * rare that this path is taken, it is perfectly safe to force
 835                  * us to go around the wait_event() loop again, so the warning
 836                  * is a false-positive.  Silence the warning by resetting
 837                  * thread state
 838                  */
 839                 __set_current_state(TASK_RUNNING);
 840                 blk_start_plug(&plug);
 841                 flush_bio_list(conf, bio);
 842                 blk_finish_plug(&plug);
 843         } else
 844                 spin_unlock_irq(&conf->device_lock);
 845 }
 846
 847 /* Barriers....
 848  * Sometimes we need to suspend IO while we do something else,
 849  * either some resync/recovery, or reconfigure the array.
 850  * To do this we raise a 'barrier'.
 851  * The 'barrier' is a counter that can be raised multiple times
 852  * to count how many activities are happening which preclude
 853  * normal IO.
 854  * We can only raise the barrier if there is no pending IO.
 855  * i.e. if nr_pending == 0.
 856  * We choose only to raise the barrier if no-one is waiting for the
 857  * barrier to go down.  This means that as soon as an IO request
 858  * is ready, no other operations which require a barrier will start
 859  * until the IO request has had a chance.
 860  *
 861  * So: regular IO calls 'wait_barrier'.  When that returns there
 862  *    is no backgroup IO happening,  It must arrange to call
 863  *    allow_barrier when it has finished its IO.
 864  * backgroup IO calls must call raise_barrier.  Once that returns
 865  *    there is no normal IO happeing.  It must arrange to call
 866  *    lower_barrier when the particular background IO completes.
 867  *
 868  * If resync/recovery is interrupted, returns -EINTR;
 869  * Otherwise, returns 0.
 870  */
 871 static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
 872 {
 873         int idx = sector_to_idx(sector_nr);
 874
 875         spin_lock_irq(&conf->resync_lock);
 876
 877         /* Wait until no block IO is waiting */
 878         wait_event_lock_irq(conf->wait_barrier,
 879                             !atomic_read(&conf->nr_waiting[idx]),
 880                             conf->resync_lock);
 881
 882         /* block any new IO from starting */
 883         atomic_inc(&conf->barrier[idx]);
 884         /*
 885          * In raise_barrier() we firstly increase conf->barrier[idx] then
 886          * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 887          * increase conf->nr_pending[idx] then check conf->barrier[idx].
 888          * A memory barrier here to make sure conf->nr_pending[idx] won't
 889          * be fetched before conf->barrier[idx] is increased. Otherwise
 890          * there will be a race between raise_barrier() and _wait_barrier().
 891          */
 892         smp_mb__after_atomic();
 893
 894         /* For these conditions we must wait:
 895          * A: while the array is in frozen state
 896          * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 897          *    existing in corresponding I/O barrier bucket.
 898          * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 899          *    max resync count which allowed on current I/O barrier bucket.
 900          */
 901         wait_event_lock_irq(conf->wait_barrier,
 902                             (!conf->array_frozen &&
 903                              !atomic_read(&conf->nr_pending[idx]) &&
 904                              atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
 905                                 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
 906                             conf->resync_lock);
 907
 908         if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
 909                 atomic_dec(&conf->barrier[idx]);
 910                 spin_unlock_irq(&conf->resync_lock);
 911                 wake_up(&conf->wait_barrier);
 912                 return -EINTR;
 913         }
 914
 915         atomic_inc(&conf->nr_sync_pending);
 916         spin_unlock_irq(&conf->resync_lock);
 917
 918         return 0;
 919 }
 920
 921 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 922 {
 923         int idx = sector_to_idx(sector_nr);
 924
 925         BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 926
 927         atomic_dec(&conf->barrier[idx]);
 928         atomic_dec(&conf->nr_sync_pending);
 929         wake_up(&conf->wait_barrier);
 930 }
 931
 932 static void _wait_barrier(struct r1conf *conf, int idx)
 933 {
 934         /*
 935          * We need to increase conf->nr_pending[idx] very early here,
 936          * then raise_barrier() can be blocked when it waits for
 937          * conf->nr_pending[idx] to be 0. Then we can avoid holding
 938          * conf->resync_lock when there is no barrier raised in same
 939          * barrier unit bucket. Also if the array is frozen, I/O
 940          * should be blocked until array is unfrozen.
 941          */
 942         atomic_inc(&conf->nr_pending[idx]);
 943         /*
 944          * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
 945          * check conf->barrier[idx]. In raise_barrier() we firstly increase
 946          * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
 947          * barrier is necessary here to make sure conf->barrier[idx] won't be
 948          * fetched before conf->nr_pending[idx] is increased. Otherwise there
 949          * will be a race between _wait_barrier() and raise_barrier().
 950          */
 951         smp_mb__after_atomic();
 952
 953         /*
 954          * Don't worry about checking two atomic_t variables at same time
 955          * here. If during we check conf->barrier[idx], the array is
 956          * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
 957          * 0, it is safe to return and make the I/O continue. Because the
 958          * array is frozen, all I/O returned here will eventually complete
 959          * or be queued, no race will happen. See code comment in
 960          * frozen_array().
 961          */
 962         if (!READ_ONCE(conf->array_frozen) &&
 963             !atomic_read(&conf->barrier[idx]))
 964                 return;
 965
 966         /*
 967          * After holding conf->resync_lock, conf->nr_pending[idx]
 968          * should be decreased before waiting for barrier to drop.
 969          * Otherwise, we may encounter a race condition because
 970          * raise_barrer() might be waiting for conf->nr_pending[idx]
 971          * to be 0 at same time.
 972          */
 973         spin_lock_irq(&conf->resync_lock);
 974         atomic_inc(&conf->nr_waiting[idx]);
 975         atomic_dec(&conf->nr_pending[idx]);
 976         /*
 977          * In case freeze_array() is waiting for
 978          * get_unqueued_pending() == extra
 979          */
 980         wake_up(&conf->wait_barrier);
 981         /* Wait for the barrier in same barrier unit bucket to drop. */
 982         wait_event_lock_irq(conf->wait_barrier,
 983                             !conf->array_frozen &&
 984                              !atomic_read(&conf->barrier[idx]),
 985                             conf->resync_lock);
 986         atomic_inc(&conf->nr_pending[idx]);
 987         atomic_dec(&conf->nr_waiting[idx]);
 988         spin_unlock_irq(&conf->resync_lock);
 989 }
 990
 991 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
 992 {
 993         int idx = sector_to_idx(sector_nr);
 994
 995         /*
 996          * Very similar to _wait_barrier(). The difference is, for read
 997          * I/O we don't need wait for sync I/O, but if the whole array
 998          * is frozen, the read I/O still has to wait until the array is
 999          * unfrozen. Since there is no ordering requirement with
1000          * conf->barrier[idx] here, memory barrier is unnecessary as well.
1001          */
1002         atomic_inc(&conf->nr_pending[idx]);
1003
1004         if (!READ_ONCE(conf->array_frozen))
1005                 return;
1006
1007         spin_lock_irq(&conf->resync_lock);
1008         atomic_inc(&conf->nr_waiting[idx]);
1009         atomic_dec(&conf->nr_pending[idx]);
1010         /*
1011          * In case freeze_array() is waiting for
1012          * get_unqueued_pending() == extra
1013          */
1014         wake_up(&conf->wait_barrier);
1015         /* Wait for array to be unfrozen */
1016         wait_event_lock_irq(conf->wait_barrier,
1017                             !conf->array_frozen,
1018                             conf->resync_lock);
1019         atomic_inc(&conf->nr_pending[idx]);
1020         atomic_dec(&conf->nr_waiting[idx]);
1021         spin_unlock_irq(&conf->resync_lock);
1022 }
1023
1024 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1025 {
1026         int idx = sector_to_idx(sector_nr);
1027
1028         _wait_barrier(conf, idx);
1029 }
1030
1031 static void _allow_barrier(struct r1conf *conf, int idx)
1032 {
1033         atomic_dec(&conf->nr_pending[idx]);
1034         wake_up(&conf->wait_barrier);
1035 }
1036
1037 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1038 {
1039         int idx = sector_to_idx(sector_nr);
1040
1041         _allow_barrier(conf, idx);
1042 }
1043
1044 /* conf->resync_lock should be held */
1045 static int get_unqueued_pending(struct r1conf *conf)
1046 {
1047         int idx, ret;
1048
1049         ret = atomic_read(&conf->nr_sync_pending);
1050         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1051                 ret += atomic_read(&conf->nr_pending[idx]) -
1052                         atomic_read(&conf->nr_queued[idx]);
1053
1054         return ret;
1055 }
1056
1057 static void freeze_array(struct r1conf *conf, int extra)
1058 {
1059         /* Stop sync I/O and normal I/O and wait for everything to
1060          * go quiet.
1061          * This is called in two situations:
1062          * 1) management command handlers (reshape, remove disk, quiesce).
1063          * 2) one normal I/O request failed.
1064
1065          * After array_frozen is set to 1, new sync IO will be blocked at
1066          * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1067          * or wait_read_barrier(). The flying I/Os will either complete or be
1068          * queued. When everything goes quite, there are only queued I/Os left.
1069
1070          * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1071          * barrier bucket index which this I/O request hits. When all sync and
1072          * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1073          * of all conf->nr_queued[]. But normal I/O failure is an exception,
1074          * in handle_read_error(), we may call freeze_array() before trying to
1075          * fix the read error. In this case, the error read I/O is not queued,
1076          * so get_unqueued_pending() == 1.
1077          *
1078          * Therefore before this function returns, we need to wait until
1079          * get_unqueued_pendings(conf) gets equal to extra. For
1080          * normal I/O context, extra is 1, in rested situations extra is 0.
1081          */
1082         spin_lock_irq(&conf->resync_lock);
1083         conf->array_frozen = 1;
1084         raid1_log(conf->mddev, "wait freeze");
1085         wait_event_lock_irq_cmd(
1086                 conf->wait_barrier,
1087                 get_unqueued_pending(conf) == extra,
1088                 conf->resync_lock,
1089                 flush_pending_writes(conf));
1090         spin_unlock_irq(&conf->resync_lock);
1091 }
1092 static void unfreeze_array(struct r1conf *conf)
1093 {
1094         /* reverse the effect of the freeze */
1095         spin_lock_irq(&conf->resync_lock);
1096         conf->array_frozen = 0;
1097         spin_unlock_irq(&conf->resync_lock);
1098         wake_up(&conf->wait_barrier);
1099 }
1100
1101 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1102                                            struct bio *bio)
1103 {
1104         int size = bio->bi_iter.bi_size;
1105         unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1106         int i = 0;
1107         struct bio *behind_bio = NULL;
1108
1109         behind_bio = bio_alloc_bioset(GFP_NOIO, vcnt, &r1_bio->mddev->bio_set);
1110         if (!behind_bio)
1111                 return;
1112
1113         /* discard op, we don't support writezero/writesame yet */
1114         if (!bio_has_data(bio)) {
1115                 behind_bio->bi_iter.bi_size = size;
1116                 goto skip_copy;
1117         }
1118
1119         behind_bio->bi_write_hint = bio->bi_write_hint;
1120
1121         while (i < vcnt && size) {
1122                 struct page *page;
1123                 int len = min_t(int, PAGE_SIZE, size);
1124
1125                 page = alloc_page(GFP_NOIO);
1126                 if (unlikely(!page))
1127                         goto free_pages;
1128
1129                 bio_add_page(behind_bio, page, len, 0);
1130
1131                 size -= len;
1132                 i++;
1133         }
1134
1135         bio_copy_data(behind_bio, bio);
1136 skip_copy:
1137         r1_bio->behind_master_bio = behind_bio;
1138         set_bit(R1BIO_BehindIO, &r1_bio->state);
1139
1140         return;
1141
1142 free_pages:
1143         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1144                  bio->bi_iter.bi_size);
1145         bio_free_pages(behind_bio);
1146         bio_put(behind_bio);
1147 }
1148
1149 struct raid1_plug_cb {
1150         struct blk_plug_cb      cb;
1151         struct bio_list         pending;
1152         int                     pending_cnt;
1153 };
1154
1155 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1156 {
1157         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1158                                                   cb);
1159         struct mddev *mddev = plug->cb.data;
1160         struct r1conf *conf = mddev->private;
1161         struct bio *bio;
1162
1163         if (from_schedule || current->bio_list) {
1164                 spin_lock_irq(&conf->device_lock);
1165                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1166                 conf->pending_count += plug->pending_cnt;
1167                 spin_unlock_irq(&conf->device_lock);
1168                 wake_up(&conf->wait_barrier);
1169                 md_wakeup_thread(mddev->thread);
1170                 kfree(plug);
1171                 return;
1172         }
1173
1174         /* we aren't scheduling, so we can do the write-out directly. */
1175         bio = bio_list_get(&plug->pending);
1176         flush_bio_list(conf, bio);
1177         kfree(plug);
1178 }
1179
1180 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1181 {
1182         r1_bio->master_bio = bio;
1183         r1_bio->sectors = bio_sectors(bio);
1184         r1_bio->state = 0;
1185         r1_bio->mddev = mddev;
1186         r1_bio->sector = bio->bi_iter.bi_sector;
1187 }
1188
1189 static inline struct r1bio *
1190 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1191 {
1192         struct r1conf *conf = mddev->private;
1193         struct r1bio *r1_bio;
1194
1195         r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1196         /* Ensure no bio records IO_BLOCKED */
1197         memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1198         init_r1bio(r1_bio, mddev, bio);
1199         return r1_bio;
1200 }
1201
1202 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1203                                int max_read_sectors, struct r1bio *r1_bio)
1204 {
1205         struct r1conf *conf = mddev->private;
1206         struct raid1_info *mirror;
1207         struct bio *read_bio;
1208         struct bitmap *bitmap = mddev->bitmap;
1209         const int op = bio_op(bio);
1210         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1211         int max_sectors;
1212         int rdisk;
1213         bool r1bio_existed = !!r1_bio;
1214         char b[BDEVNAME_SIZE];
1215
1216         /*
1217          * If r1_bio is set, we are blocking the raid1d thread
1218          * so there is a tiny risk of deadlock.  So ask for
1219          * emergency memory if needed.
1220          */
1221         gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1222
1223         if (r1bio_existed) {
1224                 /* Need to get the block device name carefully */
1225                 struct md_rdev *rdev;
1226                 rcu_read_lock();
1227                 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1228                 if (rdev)
1229                         bdevname(rdev->bdev, b);
1230                 else
1231                         strcpy(b, "???");
1232                 rcu_read_unlock();
1233         }
1234
1235         /*
1236          * Still need barrier for READ in case that whole
1237          * array is frozen.
1238          */
1239         wait_read_barrier(conf, bio->bi_iter.bi_sector);
1240
1241         if (!r1_bio)
1242                 r1_bio = alloc_r1bio(mddev, bio);
1243         else
1244                 init_r1bio(r1_bio, mddev, bio);
1245         r1_bio->sectors = max_read_sectors;
1246
1247         /*
1248          * make_request() can abort the operation when read-ahead is being
1249          * used and no empty request is available.
1250          */
1251         rdisk = read_balance(conf, r1_bio, &max_sectors);
1252
1253         if (rdisk < 0) {
1254                 /* couldn't find anywhere to read from */
1255                 if (r1bio_existed) {
1256                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1257                                             mdname(mddev),
1258                                             b,
1259                                             (unsigned long long)r1_bio->sector);
1260                 }
1261                 raid_end_bio_io(r1_bio);
1262                 return;
1263         }
1264         mirror = conf->mirrors + rdisk;
1265
1266         if (r1bio_existed)
1267                 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1268                                     mdname(mddev),
1269                                     (unsigned long long)r1_bio->sector,
1270                                     bdevname(mirror->rdev->bdev, b));
1271
1272         if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1273             bitmap) {
1274                 /*
1275                  * Reading from a write-mostly device must take care not to
1276                  * over-take any writes that are 'behind'
1277                  */
1278                 raid1_log(mddev, "wait behind writes");
1279                 wait_event(bitmap->behind_wait,
1280                            atomic_read(&bitmap->behind_writes) == 0);
1281         }
1282
1283         if (max_sectors < bio_sectors(bio)) {
1284                 struct bio *split = bio_split(bio, max_sectors,
1285                                               gfp, &conf->bio_split);
1286                 bio_chain(split, bio);
1287                 submit_bio_noacct(bio);
1288                 bio = split;
1289                 r1_bio->master_bio = bio;
1290                 r1_bio->sectors = max_sectors;
1291         }
1292
1293         r1_bio->read_disk = rdisk;
1294
1295         if (!r1bio_existed && blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1296                 r1_bio->start_time = bio_start_io_acct(bio);
1297
1298         read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1299
1300         r1_bio->bios[rdisk] = read_bio;
1301
1302         read_bio->bi_iter.bi_sector = r1_bio->sector +
1303                 mirror->rdev->data_offset;
1304         bio_set_dev(read_bio, mirror->rdev->bdev);
1305         read_bio->bi_end_io = raid1_end_read_request;
1306         bio_set_op_attrs(read_bio, op, do_sync);
1307         if (test_bit(FailFast, &mirror->rdev->flags) &&
1308             test_bit(R1BIO_FailFast, &r1_bio->state))
1309                 read_bio->bi_opf |= MD_FAILFAST;
1310         read_bio->bi_private = r1_bio;
1311
1312         if (mddev->gendisk)
1313                 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1314                                       r1_bio->sector);
1315
1316         submit_bio_noacct(read_bio);
1317 }
1318
1319 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1320                                 int max_write_sectors)
1321 {
1322         struct r1conf *conf = mddev->private;
1323         struct r1bio *r1_bio;
1324         int i, disks;
1325         struct bitmap *bitmap = mddev->bitmap;
1326         unsigned long flags;
1327         struct md_rdev *blocked_rdev;
1328         struct blk_plug_cb *cb;
1329         struct raid1_plug_cb *plug = NULL;
1330         int first_clone;
1331         int max_sectors;
1332         bool write_behind = false;
1333
1334         if (mddev_is_clustered(mddev) &&
1335              md_cluster_ops->area_resyncing(mddev, WRITE,
1336                      bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1337
1338                 DEFINE_WAIT(w);
1339                 for (;;) {
1340                         prepare_to_wait(&conf->wait_barrier,
1341                                         &w, TASK_IDLE);
1342                         if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1343                                                         bio->bi_iter.bi_sector,
1344                                                         bio_end_sector(bio)))
1345                                 break;
1346                         schedule();
1347                 }
1348                 finish_wait(&conf->wait_barrier, &w);
1349         }
1350
1351         /*
1352          * Register the new request and wait if the reconstruction
1353          * thread has put up a bar for new requests.
1354          * Continue immediately if no resync is active currently.
1355          */
1356         wait_barrier(conf, bio->bi_iter.bi_sector);
1357
1358         r1_bio = alloc_r1bio(mddev, bio);
1359         r1_bio->sectors = max_write_sectors;
1360
1361         if (conf->pending_count >= max_queued_requests) {
1362                 md_wakeup_thread(mddev->thread);
1363                 raid1_log(mddev, "wait queued");
1364                 wait_event(conf->wait_barrier,
1365                            conf->pending_count < max_queued_requests);
1366         }
1367         /* first select target devices under rcu_lock and
1368          * inc refcount on their rdev.  Record them by setting
1369          * bios[x] to bio
1370          * If there are known/acknowledged bad blocks on any device on
1371          * which we have seen a write error, we want to avoid writing those
1372          * blocks.
1373          * This potentially requires several writes to write around
1374          * the bad blocks.  Each set of writes gets it's own r1bio
1375          * with a set of bios attached.
1376          */
1377
1378         disks = conf->raid_disks * 2;
1379  retry_write:
1380         blocked_rdev = NULL;
1381         rcu_read_lock();
1382         max_sectors = r1_bio->sectors;
1383         for (i = 0;  i < disks; i++) {
1384                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1385
1386                 /*
1387                  * The write-behind io is only attempted on drives marked as
1388                  * write-mostly, which means we could allocate write behind
1389                  * bio later.
1390                  */
1391                 if (rdev && test_bit(WriteMostly, &rdev->flags))
1392                         write_behind = true;
1393
1394                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1395                         atomic_inc(&rdev->nr_pending);
1396                         blocked_rdev = rdev;
1397                         break;
1398                 }
1399                 r1_bio->bios[i] = NULL;
1400                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1401                         if (i < conf->raid_disks)
1402                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1403                         continue;
1404                 }
1405
1406                 atomic_inc(&rdev->nr_pending);
1407                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1408                         sector_t first_bad;
1409                         int bad_sectors;
1410                         int is_bad;
1411
1412                         is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1413                                              &first_bad, &bad_sectors);
1414                         if (is_bad < 0) {
1415                                 /* mustn't write here until the bad block is
1416                                  * acknowledged*/
1417                                 set_bit(BlockedBadBlocks, &rdev->flags);
1418                                 blocked_rdev = rdev;
1419                                 break;
1420                         }
1421                         if (is_bad && first_bad <= r1_bio->sector) {
1422                                 /* Cannot write here at all */
1423                                 bad_sectors -= (r1_bio->sector - first_bad);
1424                                 if (bad_sectors < max_sectors)
1425                                         /* mustn't write more than bad_sectors
1426                                          * to other devices yet
1427                                          */
1428                                         max_sectors = bad_sectors;
1429                                 rdev_dec_pending(rdev, mddev);
1430                                 /* We don't set R1BIO_Degraded as that
1431                                  * only applies if the disk is
1432                                  * missing, so it might be re-added,
1433                                  * and we want to know to recover this
1434                                  * chunk.
1435                                  * In this case the device is here,
1436                                  * and the fact that this chunk is not
1437                                  * in-sync is recorded in the bad
1438                                  * block log
1439                                  */
1440                                 continue;
1441                         }
1442                         if (is_bad) {
1443                                 int good_sectors = first_bad - r1_bio->sector;
1444                                 if (good_sectors < max_sectors)
1445                                         max_sectors = good_sectors;
1446                         }
1447                 }
1448                 r1_bio->bios[i] = bio;
1449         }
1450         rcu_read_unlock();
1451
1452         if (unlikely(blocked_rdev)) {
1453                 /* Wait for this device to become unblocked */
1454                 int j;
1455
1456                 for (j = 0; j < i; j++)
1457                         if (r1_bio->bios[j])
1458                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1459                 r1_bio->state = 0;
1460                 allow_barrier(conf, bio->bi_iter.bi_sector);
1461                 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1462                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1463                 wait_barrier(conf, bio->bi_iter.bi_sector);
1464                 goto retry_write;
1465         }
1466
1467         /*
1468          * When using a bitmap, we may call alloc_behind_master_bio below.
1469          * alloc_behind_master_bio allocates a copy of the data payload a page
1470          * at a time and thus needs a new bio that can fit the whole payload
1471          * this bio in page sized chunks.
1472          */
1473         if (write_behind && bitmap)
1474                 max_sectors = min_t(int, max_sectors,
1475                                     BIO_MAX_VECS * (PAGE_SIZE >> 9));
1476         if (max_sectors < bio_sectors(bio)) {
1477                 struct bio *split = bio_split(bio, max_sectors,
1478                                               GFP_NOIO, &conf->bio_split);
1479                 bio_chain(split, bio);
1480                 submit_bio_noacct(bio);
1481                 bio = split;
1482                 r1_bio->master_bio = bio;
1483                 r1_bio->sectors = max_sectors;
1484         }
1485
1486         if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1487                 r1_bio->start_time = bio_start_io_acct(bio);
1488         atomic_set(&r1_bio->remaining, 1);
1489         atomic_set(&r1_bio->behind_remaining, 0);
1490
1491         first_clone = 1;
1492
1493         for (i = 0; i < disks; i++) {
1494                 struct bio *mbio = NULL;
1495                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1496                 if (!r1_bio->bios[i])
1497                         continue;
1498
1499                 if (first_clone) {
1500                         /* do behind I/O ?
1501                          * Not if there are too many, or cannot
1502                          * allocate memory, or a reader on WriteMostly
1503                          * is waiting for behind writes to flush */
1504                         if (bitmap &&
1505                             test_bit(WriteMostly, &rdev->flags) &&
1506                             (atomic_read(&bitmap->behind_writes)
1507                              < mddev->bitmap_info.max_write_behind) &&
1508                             !waitqueue_active(&bitmap->behind_wait)) {
1509                                 alloc_behind_master_bio(r1_bio, bio);
1510                         }
1511
1512                         md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1513                                              test_bit(R1BIO_BehindIO, &r1_bio->state));
1514                         first_clone = 0;
1515                 }
1516
1517                 if (r1_bio->behind_master_bio)
1518                         mbio = bio_clone_fast(r1_bio->behind_master_bio,
1519                                               GFP_NOIO, &mddev->bio_set);
1520                 else
1521                         mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1522
1523                 if (r1_bio->behind_master_bio) {
1524                         if (test_bit(CollisionCheck, &rdev->flags))
1525                                 wait_for_serialization(rdev, r1_bio);
1526                         if (test_bit(WriteMostly, &rdev->flags))
1527                                 atomic_inc(&r1_bio->behind_remaining);
1528                 } else if (mddev->serialize_policy)
1529                         wait_for_serialization(rdev, r1_bio);
1530
1531                 r1_bio->bios[i] = mbio;
1532
1533                 mbio->bi_iter.bi_sector = (r1_bio->sector +
1534                                    conf->mirrors[i].rdev->data_offset);
1535                 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1536                 mbio->bi_end_io = raid1_end_write_request;
1537                 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1538                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1539                     !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1540                     conf->raid_disks - mddev->degraded > 1)
1541                         mbio->bi_opf |= MD_FAILFAST;
1542                 mbio->bi_private = r1_bio;
1543
1544                 atomic_inc(&r1_bio->remaining);
1545
1546                 if (mddev->gendisk)
1547                         trace_block_bio_remap(mbio, disk_devt(mddev->gendisk),
1548                                               r1_bio->sector);
1549                 /* flush_pending_writes() needs access to the rdev so...*/
1550                 mbio->bi_bdev = (void *)conf->mirrors[i].rdev;
1551
1552                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1553                 if (cb)
1554                         plug = container_of(cb, struct raid1_plug_cb, cb);
1555                 else
1556                         plug = NULL;
1557                 if (plug) {
1558                         bio_list_add(&plug->pending, mbio);
1559                         plug->pending_cnt++;
1560                 } else {
1561                         spin_lock_irqsave(&conf->device_lock, flags);
1562                         bio_list_add(&conf->pending_bio_list, mbio);
1563                         conf->pending_count++;
1564                         spin_unlock_irqrestore(&conf->device_lock, flags);
1565                         md_wakeup_thread(mddev->thread);
1566                 }
1567         }
1568
1569         r1_bio_write_done(r1_bio);
1570
1571         /* In case raid1d snuck in to freeze_array */
1572         wake_up(&conf->wait_barrier);
1573 }
1574
1575 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1576 {
1577         sector_t sectors;
1578
1579         if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1580             && md_flush_request(mddev, bio))
1581                 return true;
1582
1583         /*
1584          * There is a limit to the maximum size, but
1585          * the read/write handler might find a lower limit
1586          * due to bad blocks.  To avoid multiple splits,
1587          * we pass the maximum number of sectors down
1588          * and let the lower level perform the split.
1589          */
1590         sectors = align_to_barrier_unit_end(
1591                 bio->bi_iter.bi_sector, bio_sectors(bio));
1592
1593         if (bio_data_dir(bio) == READ)
1594                 raid1_read_request(mddev, bio, sectors, NULL);
1595         else {
1596                 if (!md_write_start(mddev,bio))
1597                         return false;
1598                 raid1_write_request(mddev, bio, sectors);
1599         }
1600         return true;
1601 }
1602
1603 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1604 {
1605         struct r1conf *conf = mddev->private;
1606         int i;
1607
1608         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1609                    conf->raid_disks - mddev->degraded);
1610         rcu_read_lock();
1611         for (i = 0; i < conf->raid_disks; i++) {
1612                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1613                 seq_printf(seq, "%s",
1614                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1615         }
1616         rcu_read_unlock();
1617         seq_printf(seq, "]");
1618 }
1619
1620 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1621 {
1622         char b[BDEVNAME_SIZE];
1623         struct r1conf *conf = mddev->private;
1624         unsigned long flags;
1625
1626         /*
1627          * If it is not operational, then we have already marked it as dead
1628          * else if it is the last working disks with "fail_last_dev == false",
1629          * ignore the error, let the next level up know.
1630          * else mark the drive as failed
1631          */
1632         spin_lock_irqsave(&conf->device_lock, flags);
1633         if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1634             && (conf->raid_disks - mddev->degraded) == 1) {
1635                 /*
1636                  * Don't fail the drive, act as though we were just a
1637                  * normal single drive.
1638                  * However don't try a recovery from this drive as
1639                  * it is very likely to fail.
1640                  */
1641                 conf->recovery_disabled = mddev->recovery_disabled;
1642                 spin_unlock_irqrestore(&conf->device_lock, flags);
1643                 return;
1644         }
1645         set_bit(Blocked, &rdev->flags);
1646         if (test_and_clear_bit(In_sync, &rdev->flags))
1647                 mddev->degraded++;
1648         set_bit(Faulty, &rdev->flags);
1649         spin_unlock_irqrestore(&conf->device_lock, flags);
1650         /*
1651          * if recovery is running, make sure it aborts.
1652          */
1653         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1654         set_mask_bits(&mddev->sb_flags, 0,
1655                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1656         pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1657                 "md/raid1:%s: Operation continuing on %d devices.\n",
1658                 mdname(mddev), bdevname(rdev->bdev, b),
1659                 mdname(mddev), conf->raid_disks - mddev->degraded);
1660 }
1661
1662 static void print_conf(struct r1conf *conf)
1663 {
1664         int i;
1665
1666         pr_debug("RAID1 conf printout:\n");
1667         if (!conf) {
1668                 pr_debug("(!conf)\n");
1669                 return;
1670         }
1671         pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1672                  conf->raid_disks);
1673
1674         rcu_read_lock();
1675         for (i = 0; i < conf->raid_disks; i++) {
1676                 char b[BDEVNAME_SIZE];
1677                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1678                 if (rdev)
1679                         pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1680                                  i, !test_bit(In_sync, &rdev->flags),
1681                                  !test_bit(Faulty, &rdev->flags),
1682                                  bdevname(rdev->bdev,b));
1683         }
1684         rcu_read_unlock();
1685 }
1686
1687 static void close_sync(struct r1conf *conf)
1688 {
1689         int idx;
1690
1691         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1692                 _wait_barrier(conf, idx);
1693                 _allow_barrier(conf, idx);
1694         }
1695
1696         mempool_exit(&conf->r1buf_pool);
1697 }
1698
1699 static int raid1_spare_active(struct mddev *mddev)
1700 {
1701         int i;
1702         struct r1conf *conf = mddev->private;
1703         int count = 0;
1704         unsigned long flags;
1705
1706         /*
1707          * Find all failed disks within the RAID1 configuration
1708          * and mark them readable.
1709          * Called under mddev lock, so rcu protection not needed.
1710          * device_lock used to avoid races with raid1_end_read_request
1711          * which expects 'In_sync' flags and ->degraded to be consistent.
1712          */
1713         spin_lock_irqsave(&conf->device_lock, flags);
1714         for (i = 0; i < conf->raid_disks; i++) {
1715                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1716                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1717                 if (repl
1718                     && !test_bit(Candidate, &repl->flags)
1719                     && repl->recovery_offset == MaxSector
1720                     && !test_bit(Faulty, &repl->flags)
1721                     && !test_and_set_bit(In_sync, &repl->flags)) {
1722                         /* replacement has just become active */
1723                         if (!rdev ||
1724                             !test_and_clear_bit(In_sync, &rdev->flags))
1725                                 count++;
1726                         if (rdev) {
1727                                 /* Replaced device not technically
1728                                  * faulty, but we need to be sure
1729                                  * it gets removed and never re-added
1730                                  */
1731                                 set_bit(Faulty, &rdev->flags);
1732                                 sysfs_notify_dirent_safe(
1733                                         rdev->sysfs_state);
1734                         }
1735                 }
1736                 if (rdev
1737                     && rdev->recovery_offset == MaxSector
1738                     && !test_bit(Faulty, &rdev->flags)
1739                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1740                         count++;
1741                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1742                 }
1743         }
1744         mddev->degraded -= count;
1745         spin_unlock_irqrestore(&conf->device_lock, flags);
1746
1747         print_conf(conf);
1748         return count;
1749 }
1750
1751 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1752 {
1753         struct r1conf *conf = mddev->private;
1754         int err = -EEXIST;
1755         int mirror = 0;
1756         struct raid1_info *p;
1757         int first = 0;
1758         int last = conf->raid_disks - 1;
1759
1760         if (mddev->recovery_disabled == conf->recovery_disabled)
1761                 return -EBUSY;
1762
1763         if (md_integrity_add_rdev(rdev, mddev))
1764                 return -ENXIO;
1765
1766         if (rdev->raid_disk >= 0)
1767                 first = last = rdev->raid_disk;
1768
1769         /*
1770          * find the disk ... but prefer rdev->saved_raid_disk
1771          * if possible.
1772          */
1773         if (rdev->saved_raid_disk >= 0 &&
1774             rdev->saved_raid_disk >= first &&
1775             rdev->saved_raid_disk < conf->raid_disks &&
1776             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1777                 first = last = rdev->saved_raid_disk;
1778
1779         for (mirror = first; mirror <= last; mirror++) {
1780                 p = conf->mirrors + mirror;
1781                 if (!p->rdev) {
1782                         if (mddev->gendisk)
1783                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1784                                                   rdev->data_offset << 9);
1785
1786                         p->head_position = 0;
1787                         rdev->raid_disk = mirror;
1788                         err = 0;
1789                         /* As all devices are equivalent, we don't need a full recovery
1790                          * if this was recently any drive of the array
1791                          */
1792                         if (rdev->saved_raid_disk < 0)
1793                                 conf->fullsync = 1;
1794                         rcu_assign_pointer(p->rdev, rdev);
1795                         break;
1796                 }
1797                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1798                     p[conf->raid_disks].rdev == NULL) {
1799                         /* Add this device as a replacement */
1800                         clear_bit(In_sync, &rdev->flags);
1801                         set_bit(Replacement, &rdev->flags);
1802                         rdev->raid_disk = mirror;
1803                         err = 0;
1804                         conf->fullsync = 1;
1805                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1806                         break;
1807                 }
1808         }
1809         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1810                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1811         print_conf(conf);
1812         return err;
1813 }
1814
1815 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1816 {
1817         struct r1conf *conf = mddev->private;
1818         int err = 0;
1819         int number = rdev->raid_disk;
1820         struct raid1_info *p = conf->mirrors + number;
1821
1822         if (rdev != p->rdev)
1823                 p = conf->mirrors + conf->raid_disks + number;
1824
1825         print_conf(conf);
1826         if (rdev == p->rdev) {
1827                 if (test_bit(In_sync, &rdev->flags) ||
1828                     atomic_read(&rdev->nr_pending)) {
1829                         err = -EBUSY;
1830                         goto abort;
1831                 }
1832                 /* Only remove non-faulty devices if recovery
1833                  * is not possible.
1834                  */
1835                 if (!test_bit(Faulty, &rdev->flags) &&
1836                     mddev->recovery_disabled != conf->recovery_disabled &&
1837                     mddev->degraded < conf->raid_disks) {
1838                         err = -EBUSY;
1839                         goto abort;
1840                 }
1841                 p->rdev = NULL;
1842                 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1843                         synchronize_rcu();
1844                         if (atomic_read(&rdev->nr_pending)) {
1845                                 /* lost the race, try later */
1846                                 err = -EBUSY;
1847                                 p->rdev = rdev;
1848                                 goto abort;
1849                         }
1850                 }
1851                 if (conf->mirrors[conf->raid_disks + number].rdev) {
1852                         /* We just removed a device that is being replaced.
1853                          * Move down the replacement.  We drain all IO before
1854                          * doing this to avoid confusion.
1855                          */
1856                         struct md_rdev *repl =
1857                                 conf->mirrors[conf->raid_disks + number].rdev;
1858                         freeze_array(conf, 0);
1859                         if (atomic_read(&repl->nr_pending)) {
1860                                 /* It means that some queued IO of retry_list
1861                                  * hold repl. Thus, we cannot set replacement
1862                                  * as NULL, avoiding rdev NULL pointer
1863                                  * dereference in sync_request_write and
1864                                  * handle_write_finished.
1865                                  */
1866                                 err = -EBUSY;
1867                                 unfreeze_array(conf);
1868                                 goto abort;
1869                         }
1870                         clear_bit(Replacement, &repl->flags);
1871                         p->rdev = repl;
1872                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1873                         unfreeze_array(conf);
1874                 }
1875
1876                 clear_bit(WantReplacement, &rdev->flags);
1877                 err = md_integrity_register(mddev);
1878         }
1879 abort:
1880
1881         print_conf(conf);
1882         return err;
1883 }
1884
1885 static void end_sync_read(struct bio *bio)
1886 {
1887         struct r1bio *r1_bio = get_resync_r1bio(bio);
1888
1889         update_head_pos(r1_bio->read_disk, r1_bio);
1890
1891         /*
1892          * we have read a block, now it needs to be re-written,
1893          * or re-read if the read failed.
1894          * We don't do much here, just schedule handling by raid1d
1895          */
1896         if (!bio->bi_status)
1897                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1898
1899         if (atomic_dec_and_test(&r1_bio->remaining))
1900                 reschedule_retry(r1_bio);
1901 }
1902
1903 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1904 {
1905         sector_t sync_blocks = 0;
1906         sector_t s = r1_bio->sector;
1907         long sectors_to_go = r1_bio->sectors;
1908
1909         /* make sure these bits don't get cleared. */
1910         do {
1911                 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1912                 s += sync_blocks;
1913                 sectors_to_go -= sync_blocks;
1914         } while (sectors_to_go > 0);
1915 }
1916
1917 static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1918 {
1919         if (atomic_dec_and_test(&r1_bio->remaining)) {
1920                 struct mddev *mddev = r1_bio->mddev;
1921                 int s = r1_bio->sectors;
1922
1923                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1924                     test_bit(R1BIO_WriteError, &r1_bio->state))
1925                         reschedule_retry(r1_bio);
1926                 else {
1927                         put_buf(r1_bio);
1928                         md_done_sync(mddev, s, uptodate);
1929                 }
1930         }
1931 }
1932
1933 static void end_sync_write(struct bio *bio)
1934 {
1935         int uptodate = !bio->bi_status;
1936         struct r1bio *r1_bio = get_resync_r1bio(bio);
1937         struct mddev *mddev = r1_bio->mddev;
1938         struct r1conf *conf = mddev->private;
1939         sector_t first_bad;
1940         int bad_sectors;
1941         struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1942
1943         if (!uptodate) {
1944                 abort_sync_write(mddev, r1_bio);
1945                 set_bit(WriteErrorSeen, &rdev->flags);
1946                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1947                         set_bit(MD_RECOVERY_NEEDED, &
1948                                 mddev->recovery);
1949                 set_bit(R1BIO_WriteError, &r1_bio->state);
1950         } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1951                                &first_bad, &bad_sectors) &&
1952                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1953                                 r1_bio->sector,
1954                                 r1_bio->sectors,
1955                                 &first_bad, &bad_sectors)
1956                 )
1957                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1958
1959         put_sync_write_buf(r1_bio, uptodate);
1960 }
1961
1962 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1963                             int sectors, struct page *page, int rw)
1964 {
1965         if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1966                 /* success */
1967                 return 1;
1968         if (rw == WRITE) {
1969                 set_bit(WriteErrorSeen, &rdev->flags);
1970                 if (!test_and_set_bit(WantReplacement,
1971                                       &rdev->flags))
1972                         set_bit(MD_RECOVERY_NEEDED, &
1973                                 rdev->mddev->recovery);
1974         }
1975         /* need to record an error - either for the block or the device */
1976         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1977                 md_error(rdev->mddev, rdev);
1978         return 0;
1979 }
1980
1981 static int fix_sync_read_error(struct r1bio *r1_bio)
1982 {
1983         /* Try some synchronous reads of other devices to get
1984          * good data, much like with normal read errors.  Only
1985          * read into the pages we already have so we don't
1986          * need to re-issue the read request.
1987          * We don't need to freeze the array, because being in an
1988          * active sync request, there is no normal IO, and
1989          * no overlapping syncs.
1990          * We don't need to check is_badblock() again as we
1991          * made sure that anything with a bad block in range
1992          * will have bi_end_io clear.
1993          */
1994         struct mddev *mddev = r1_bio->mddev;
1995         struct r1conf *conf = mddev->private;
1996         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1997         struct page **pages = get_resync_pages(bio)->pages;
1998         sector_t sect = r1_bio->sector;
1999         int sectors = r1_bio->sectors;
2000         int idx = 0;
2001         struct md_rdev *rdev;
2002
2003         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2004         if (test_bit(FailFast, &rdev->flags)) {
2005                 /* Don't try recovering from here - just fail it
2006                  * ... unless it is the last working device of course */
2007                 md_error(mddev, rdev);
2008                 if (test_bit(Faulty, &rdev->flags))
2009                         /* Don't try to read from here, but make sure
2010                          * put_buf does it's thing
2011                          */
2012                         bio->bi_end_io = end_sync_write;
2013         }
2014
2015         while(sectors) {
2016                 int s = sectors;
2017                 int d = r1_bio->read_disk;
2018                 int success = 0;
2019                 int start;
2020
2021                 if (s > (PAGE_SIZE>>9))
2022                         s = PAGE_SIZE >> 9;
2023                 do {
2024                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2025                                 /* No rcu protection needed here devices
2026                                  * can only be removed when no resync is
2027                                  * active, and resync is currently active
2028                                  */
2029                                 rdev = conf->mirrors[d].rdev;
2030                                 if (sync_page_io(rdev, sect, s<<9,
2031                                                  pages[idx],
2032                                                  REQ_OP_READ, 0, false)) {
2033                                         success = 1;
2034                                         break;
2035                                 }
2036                         }
2037                         d++;
2038                         if (d == conf->raid_disks * 2)
2039                                 d = 0;
2040                 } while (!success && d != r1_bio->read_disk);
2041
2042                 if (!success) {
2043                         char b[BDEVNAME_SIZE];
2044                         int abort = 0;
2045                         /* Cannot read from anywhere, this block is lost.
2046                          * Record a bad block on each device.  If that doesn't
2047                          * work just disable and interrupt the recovery.
2048                          * Don't fail devices as that won't really help.
2049                          */
2050                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2051                                             mdname(mddev), bio_devname(bio, b),
2052                                             (unsigned long long)r1_bio->sector);
2053                         for (d = 0; d < conf->raid_disks * 2; d++) {
2054                                 rdev = conf->mirrors[d].rdev;
2055                                 if (!rdev || test_bit(Faulty, &rdev->flags))
2056                                         continue;
2057                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
2058                                         abort = 1;
2059                         }
2060                         if (abort) {
2061                                 conf->recovery_disabled =
2062                                         mddev->recovery_disabled;
2063                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2064                                 md_done_sync(mddev, r1_bio->sectors, 0);
2065                                 put_buf(r1_bio);
2066                                 return 0;
2067                         }
2068                         /* Try next page */
2069                         sectors -= s;
2070                         sect += s;
2071                         idx++;
2072                         continue;
2073                 }
2074
2075                 start = d;
2076                 /* write it back and re-read */
2077                 while (d != r1_bio->read_disk) {
2078                         if (d == 0)
2079                                 d = conf->raid_disks * 2;
2080                         d--;
2081                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2082                                 continue;
2083                         rdev = conf->mirrors[d].rdev;
2084                         if (r1_sync_page_io(rdev, sect, s,
2085                                             pages[idx],
2086                                             WRITE) == 0) {
2087                                 r1_bio->bios[d]->bi_end_io = NULL;
2088                                 rdev_dec_pending(rdev, mddev);
2089                         }
2090                 }
2091                 d = start;
2092                 while (d != r1_bio->read_disk) {
2093                         if (d == 0)
2094                                 d = conf->raid_disks * 2;
2095                         d--;
2096                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2097                                 continue;
2098                         rdev = conf->mirrors[d].rdev;
2099                         if (r1_sync_page_io(rdev, sect, s,
2100                                             pages[idx],
2101                                             READ) != 0)
2102                                 atomic_add(s, &rdev->corrected_errors);
2103                 }
2104                 sectors -= s;
2105                 sect += s;
2106                 idx ++;
2107         }
2108         set_bit(R1BIO_Uptodate, &r1_bio->state);
2109         bio->bi_status = 0;
2110         return 1;
2111 }
2112
2113 static void process_checks(struct r1bio *r1_bio)
2114 {
2115         /* We have read all readable devices.  If we haven't
2116          * got the block, then there is no hope left.
2117          * If we have, then we want to do a comparison
2118          * and skip the write if everything is the same.
2119          * If any blocks failed to read, then we need to
2120          * attempt an over-write
2121          */
2122         struct mddev *mddev = r1_bio->mddev;
2123         struct r1conf *conf = mddev->private;
2124         int primary;
2125         int i;
2126         int vcnt;
2127
2128         /* Fix variable parts of all bios */
2129         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2130         for (i = 0; i < conf->raid_disks * 2; i++) {
2131                 blk_status_t status;
2132                 struct bio *b = r1_bio->bios[i];
2133                 struct resync_pages *rp = get_resync_pages(b);
2134                 if (b->bi_end_io != end_sync_read)
2135                         continue;
2136                 /* fixup the bio for reuse, but preserve errno */
2137                 status = b->bi_status;
2138                 bio_reset(b);
2139                 b->bi_status = status;
2140                 b->bi_iter.bi_sector = r1_bio->sector +
2141                         conf->mirrors[i].rdev->data_offset;
2142                 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2143                 b->bi_end_io = end_sync_read;
2144                 rp->raid_bio = r1_bio;
2145                 b->bi_private = rp;
2146
2147                 /* initialize bvec table again */
2148                 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2149         }
2150         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2151                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2152                     !r1_bio->bios[primary]->bi_status) {
2153                         r1_bio->bios[primary]->bi_end_io = NULL;
2154                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2155                         break;
2156                 }
2157         r1_bio->read_disk = primary;
2158         for (i = 0; i < conf->raid_disks * 2; i++) {
2159                 int j = 0;
2160                 struct bio *pbio = r1_bio->bios[primary];
2161                 struct bio *sbio = r1_bio->bios[i];
2162                 blk_status_t status = sbio->bi_status;
2163                 struct page **ppages = get_resync_pages(pbio)->pages;
2164                 struct page **spages = get_resync_pages(sbio)->pages;
2165                 struct bio_vec *bi;
2166                 int page_len[RESYNC_PAGES] = { 0 };
2167                 struct bvec_iter_all iter_all;
2168
2169                 if (sbio->bi_end_io != end_sync_read)
2170                         continue;
2171                 /* Now we can 'fixup' the error value */
2172                 sbio->bi_status = 0;
2173
2174                 bio_for_each_segment_all(bi, sbio, iter_all)
2175                         page_len[j++] = bi->bv_len;
2176
2177                 if (!status) {
2178                         for (j = vcnt; j-- ; ) {
2179                                 if (memcmp(page_address(ppages[j]),
2180                                            page_address(spages[j]),
2181                                            page_len[j]))
2182                                         break;
2183                         }
2184                 } else
2185                         j = 0;
2186                 if (j >= 0)
2187                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2188                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2189                               && !status)) {
2190                         /* No need to write to this device. */
2191                         sbio->bi_end_io = NULL;
2192                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2193                         continue;
2194                 }
2195
2196                 bio_copy_data(sbio, pbio);
2197         }
2198 }
2199
2200 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2201 {
2202         struct r1conf *conf = mddev->private;
2203         int i;
2204         int disks = conf->raid_disks * 2;
2205         struct bio *wbio;
2206
2207         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2208                 /* ouch - failed to read all of that. */
2209                 if (!fix_sync_read_error(r1_bio))
2210                         return;
2211
2212         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2213                 process_checks(r1_bio);
2214
2215         /*
2216          * schedule writes
2217          */
2218         atomic_set(&r1_bio->remaining, 1);
2219         for (i = 0; i < disks ; i++) {
2220                 wbio = r1_bio->bios[i];
2221                 if (wbio->bi_end_io == NULL ||
2222                     (wbio->bi_end_io == end_sync_read &&
2223                      (i == r1_bio->read_disk ||
2224                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2225                         continue;
2226                 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2227                         abort_sync_write(mddev, r1_bio);
2228                         continue;
2229                 }
2230
2231                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2232                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2233                         wbio->bi_opf |= MD_FAILFAST;
2234
2235                 wbio->bi_end_io = end_sync_write;
2236                 atomic_inc(&r1_bio->remaining);
2237                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2238
2239                 submit_bio_noacct(wbio);
2240         }
2241
2242         put_sync_write_buf(r1_bio, 1);
2243 }
2244
2245 /*
2246  * This is a kernel thread which:
2247  *
2248  *      1.      Retries failed read operations on working mirrors.
2249  *      2.      Updates the raid superblock when problems encounter.
2250  *      3.      Performs writes following reads for array synchronising.
2251  */
2252
2253 static void fix_read_error(struct r1conf *conf, int read_disk,
2254                            sector_t sect, int sectors)
2255 {
2256         struct mddev *mddev = conf->mddev;
2257         while(sectors) {
2258                 int s = sectors;
2259                 int d = read_disk;
2260                 int success = 0;
2261                 int start;
2262                 struct md_rdev *rdev;
2263
2264                 if (s > (PAGE_SIZE>>9))
2265                         s = PAGE_SIZE >> 9;
2266
2267                 do {
2268                         sector_t first_bad;
2269                         int bad_sectors;
2270
2271                         rcu_read_lock();
2272                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2273                         if (rdev &&
2274                             (test_bit(In_sync, &rdev->flags) ||
2275                              (!test_bit(Faulty, &rdev->flags) &&
2276                               rdev->recovery_offset >= sect + s)) &&
2277                             is_badblock(rdev, sect, s,
2278                                         &first_bad, &bad_sectors) == 0) {
2279                                 atomic_inc(&rdev->nr_pending);
2280                                 rcu_read_unlock();
2281                                 if (sync_page_io(rdev, sect, s<<9,
2282                                          conf->tmppage, REQ_OP_READ, 0, false))
2283                                         success = 1;
2284                                 rdev_dec_pending(rdev, mddev);
2285                                 if (success)
2286                                         break;
2287                         } else
2288                                 rcu_read_unlock();
2289                         d++;
2290                         if (d == conf->raid_disks * 2)
2291                                 d = 0;
2292                 } while (!success && d != read_disk);
2293
2294                 if (!success) {
2295                         /* Cannot read from anywhere - mark it bad */
2296                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2297                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2298                                 md_error(mddev, rdev);
2299                         break;
2300                 }
2301                 /* write it back and re-read */
2302                 start = d;
2303                 while (d != read_disk) {
2304                         if (d==0)
2305                                 d = conf->raid_disks * 2;
2306                         d--;
2307                         rcu_read_lock();
2308                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2309                         if (rdev &&
2310                             !test_bit(Faulty, &rdev->flags)) {
2311                                 atomic_inc(&rdev->nr_pending);
2312                                 rcu_read_unlock();
2313                                 r1_sync_page_io(rdev, sect, s,
2314                                                 conf->tmppage, WRITE);
2315                                 rdev_dec_pending(rdev, mddev);
2316                         } else
2317                                 rcu_read_unlock();
2318                 }
2319                 d = start;
2320                 while (d != read_disk) {
2321                         char b[BDEVNAME_SIZE];
2322                         if (d==0)
2323                                 d = conf->raid_disks * 2;
2324                         d--;
2325                         rcu_read_lock();
2326                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2327                         if (rdev &&
2328                             !test_bit(Faulty, &rdev->flags)) {
2329                                 atomic_inc(&rdev->nr_pending);
2330                                 rcu_read_unlock();
2331                                 if (r1_sync_page_io(rdev, sect, s,
2332                                                     conf->tmppage, READ)) {
2333                                         atomic_add(s, &rdev->corrected_errors);
2334                                         pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2335                                                 mdname(mddev), s,
2336                                                 (unsigned long long)(sect +
2337                                                                      rdev->data_offset),
2338                                                 bdevname(rdev->bdev, b));
2339                                 }
2340                                 rdev_dec_pending(rdev, mddev);
2341                         } else
2342                                 rcu_read_unlock();
2343                 }
2344                 sectors -= s;
2345                 sect += s;
2346         }
2347 }
2348
2349 static int narrow_write_error(struct r1bio *r1_bio, int i)
2350 {
2351         struct mddev *mddev = r1_bio->mddev;
2352         struct r1conf *conf = mddev->private;
2353         struct md_rdev *rdev = conf->mirrors[i].rdev;
2354
2355         /* bio has the data to be written to device 'i' where
2356          * we just recently had a write error.
2357          * We repeatedly clone the bio and trim down to one block,
2358          * then try the write.  Where the write fails we record
2359          * a bad block.
2360          * It is conceivable that the bio doesn't exactly align with
2361          * blocks.  We must handle this somehow.
2362          *
2363          * We currently own a reference on the rdev.
2364          */
2365
2366         int block_sectors;
2367         sector_t sector;
2368         int sectors;
2369         int sect_to_write = r1_bio->sectors;
2370         int ok = 1;
2371
2372         if (rdev->badblocks.shift < 0)
2373                 return 0;
2374
2375         block_sectors = roundup(1 << rdev->badblocks.shift,
2376                                 bdev_logical_block_size(rdev->bdev) >> 9);
2377         sector = r1_bio->sector;
2378         sectors = ((sector + block_sectors)
2379                    & ~(sector_t)(block_sectors - 1))
2380                 - sector;
2381
2382         while (sect_to_write) {
2383                 struct bio *wbio;
2384                 if (sectors > sect_to_write)
2385                         sectors = sect_to_write;
2386                 /* Write at 'sector' for 'sectors'*/
2387
2388                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2389                         wbio = bio_clone_fast(r1_bio->behind_master_bio,
2390                                               GFP_NOIO,
2391                                               &mddev->bio_set);
2392                 } else {
2393                         wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2394                                               &mddev->bio_set);
2395                 }
2396
2397                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2398                 wbio->bi_iter.bi_sector = r1_bio->sector;
2399                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2400
2401                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2402                 wbio->bi_iter.bi_sector += rdev->data_offset;
2403                 bio_set_dev(wbio, rdev->bdev);
2404
2405                 if (submit_bio_wait(wbio) < 0)
2406                         /* failure! */
2407                         ok = rdev_set_badblocks(rdev, sector,
2408                                                 sectors, 0)
2409                                 && ok;
2410
2411                 bio_put(wbio);
2412                 sect_to_write -= sectors;
2413                 sector += sectors;
2414                 sectors = block_sectors;
2415         }
2416         return ok;
2417 }
2418
2419 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2420 {
2421         int m;
2422         int s = r1_bio->sectors;
2423         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2424                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2425                 struct bio *bio = r1_bio->bios[m];
2426                 if (bio->bi_end_io == NULL)
2427                         continue;
2428                 if (!bio->bi_status &&
2429                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2430                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2431                 }
2432                 if (bio->bi_status &&
2433                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2434                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2435                                 md_error(conf->mddev, rdev);
2436                 }
2437         }
2438         put_buf(r1_bio);
2439         md_done_sync(conf->mddev, s, 1);
2440 }
2441
2442 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2443 {
2444         int m, idx;
2445         bool fail = false;
2446
2447         for (m = 0; m < conf->raid_disks * 2 ; m++)
2448                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2449                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2450                         rdev_clear_badblocks(rdev,
2451                                              r1_bio->sector,
2452                                              r1_bio->sectors, 0);
2453                         rdev_dec_pending(rdev, conf->mddev);
2454                 } else if (r1_bio->bios[m] != NULL) {
2455                         /* This drive got a write error.  We need to
2456                          * narrow down and record precise write
2457                          * errors.
2458                          */
2459                         fail = true;
2460                         if (!narrow_write_error(r1_bio, m)) {
2461                                 md_error(conf->mddev,
2462                                          conf->mirrors[m].rdev);
2463                                 /* an I/O failed, we can't clear the bitmap */
2464                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2465                         }
2466                         rdev_dec_pending(conf->mirrors[m].rdev,
2467                                          conf->mddev);
2468                 }
2469         if (fail) {
2470                 spin_lock_irq(&conf->device_lock);
2471                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2472                 idx = sector_to_idx(r1_bio->sector);
2473                 atomic_inc(&conf->nr_queued[idx]);
2474                 spin_unlock_irq(&conf->device_lock);
2475                 /*
2476                  * In case freeze_array() is waiting for condition
2477                  * get_unqueued_pending() == extra to be true.
2478                  */
2479                 wake_up(&conf->wait_barrier);
2480                 md_wakeup_thread(conf->mddev->thread);
2481         } else {
2482                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2483                         close_write(r1_bio);
2484                 raid_end_bio_io(r1_bio);
2485         }
2486 }
2487
2488 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2489 {
2490         struct mddev *mddev = conf->mddev;
2491         struct bio *bio;
2492         struct md_rdev *rdev;
2493
2494         clear_bit(R1BIO_ReadError, &r1_bio->state);
2495         /* we got a read error. Maybe the drive is bad.  Maybe just
2496          * the block and we can fix it.
2497          * We freeze all other IO, and try reading the block from
2498          * other devices.  When we find one, we re-write
2499          * and check it that fixes the read error.
2500          * This is all done synchronously while the array is
2501          * frozen
2502          */
2503
2504         bio = r1_bio->bios[r1_bio->read_disk];
2505         bio_put(bio);
2506         r1_bio->bios[r1_bio->read_disk] = NULL;
2507
2508         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2509         if (mddev->ro == 0
2510             && !test_bit(FailFast, &rdev->flags)) {
2511                 freeze_array(conf, 1);
2512                 fix_read_error(conf, r1_bio->read_disk,
2513                                r1_bio->sector, r1_bio->sectors);
2514                 unfreeze_array(conf);
2515         } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2516                 md_error(mddev, rdev);
2517         } else {
2518                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2519         }
2520
2521         rdev_dec_pending(rdev, conf->mddev);
2522         allow_barrier(conf, r1_bio->sector);
2523         bio = r1_bio->master_bio;
2524
2525         /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2526         r1_bio->state = 0;
2527         raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2528 }
2529
2530 static void raid1d(struct md_thread *thread)
2531 {
2532         struct mddev *mddev = thread->mddev;
2533         struct r1bio *r1_bio;
2534         unsigned long flags;
2535         struct r1conf *conf = mddev->private;
2536         struct list_head *head = &conf->retry_list;
2537         struct blk_plug plug;
2538         int idx;
2539
2540         md_check_recovery(mddev);
2541
2542         if (!list_empty_careful(&conf->bio_end_io_list) &&
2543             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2544                 LIST_HEAD(tmp);
2545                 spin_lock_irqsave(&conf->device_lock, flags);
2546                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2547                         list_splice_init(&conf->bio_end_io_list, &tmp);
2548                 spin_unlock_irqrestore(&conf->device_lock, flags);
2549                 while (!list_empty(&tmp)) {
2550                         r1_bio = list_first_entry(&tmp, struct r1bio,
2551                                                   retry_list);
2552                         list_del(&r1_bio->retry_list);
2553                         idx = sector_to_idx(r1_bio->sector);
2554                         atomic_dec(&conf->nr_queued[idx]);
2555                         if (mddev->degraded)
2556                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2557                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2558                                 close_write(r1_bio);
2559                         raid_end_bio_io(r1_bio);
2560                 }
2561         }
2562
2563         blk_start_plug(&plug);
2564         for (;;) {
2565
2566                 flush_pending_writes(conf);
2567
2568                 spin_lock_irqsave(&conf->device_lock, flags);
2569                 if (list_empty(head)) {
2570                         spin_unlock_irqrestore(&conf->device_lock, flags);
2571                         break;
2572                 }
2573                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2574                 list_del(head->prev);
2575                 idx = sector_to_idx(r1_bio->sector);
2576                 atomic_dec(&conf->nr_queued[idx]);
2577                 spin_unlock_irqrestore(&conf->device_lock, flags);
2578
2579                 mddev = r1_bio->mddev;
2580                 conf = mddev->private;
2581                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2582                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2583                             test_bit(R1BIO_WriteError, &r1_bio->state))
2584                                 handle_sync_write_finished(conf, r1_bio);
2585                         else
2586                                 sync_request_write(mddev, r1_bio);
2587                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2588                            test_bit(R1BIO_WriteError, &r1_bio->state))
2589                         handle_write_finished(conf, r1_bio);
2590                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2591                         handle_read_error(conf, r1_bio);
2592                 else
2593                         WARN_ON_ONCE(1);
2594
2595                 cond_resched();
2596                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2597                         md_check_recovery(mddev);
2598         }
2599         blk_finish_plug(&plug);
2600 }
2601
2602 static int init_resync(struct r1conf *conf)
2603 {
2604         int buffs;
2605
2606         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2607         BUG_ON(mempool_initialized(&conf->r1buf_pool));
2608
2609         return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2610                             r1buf_pool_free, conf->poolinfo);
2611 }
2612
2613 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2614 {
2615         struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2616         struct resync_pages *rps;
2617         struct bio *bio;
2618         int i;
2619
2620         for (i = conf->poolinfo->raid_disks; i--; ) {
2621                 bio = r1bio->bios[i];
2622                 rps = bio->bi_private;
2623                 bio_reset(bio);
2624                 bio->bi_private = rps;
2625         }
2626         r1bio->master_bio = NULL;
2627         return r1bio;
2628 }
2629
2630 /*
2631  * perform a "sync" on one "block"
2632  *
2633  * We need to make sure that no normal I/O request - particularly write
2634  * requests - conflict with active sync requests.
2635  *
2636  * This is achieved by tracking pending requests and a 'barrier' concept
2637  * that can be installed to exclude normal IO requests.
2638  */
2639
2640 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2641                                    int *skipped)
2642 {
2643         struct r1conf *conf = mddev->private;
2644         struct r1bio *r1_bio;
2645         struct bio *bio;
2646         sector_t max_sector, nr_sectors;
2647         int disk = -1;
2648         int i;
2649         int wonly = -1;
2650         int write_targets = 0, read_targets = 0;
2651         sector_t sync_blocks;
2652         int still_degraded = 0;
2653         int good_sectors = RESYNC_SECTORS;
2654         int min_bad = 0; /* number of sectors that are bad in all devices */
2655         int idx = sector_to_idx(sector_nr);
2656         int page_idx = 0;
2657
2658         if (!mempool_initialized(&conf->r1buf_pool))
2659                 if (init_resync(conf))
2660                         return 0;
2661
2662         max_sector = mddev->dev_sectors;
2663         if (sector_nr >= max_sector) {
2664                 /* If we aborted, we need to abort the
2665                  * sync on the 'current' bitmap chunk (there will
2666                  * only be one in raid1 resync.
2667                  * We can find the current addess in mddev->curr_resync
2668                  */
2669                 if (mddev->curr_resync < max_sector) /* aborted */
2670                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2671                                            &sync_blocks, 1);
2672                 else /* completed sync */
2673                         conf->fullsync = 0;
2674
2675                 md_bitmap_close_sync(mddev->bitmap);
2676                 close_sync(conf);
2677
2678                 if (mddev_is_clustered(mddev)) {
2679                         conf->cluster_sync_low = 0;
2680                         conf->cluster_sync_high = 0;
2681                 }
2682                 return 0;
2683         }
2684
2685         if (mddev->bitmap == NULL &&
2686             mddev->recovery_cp == MaxSector &&
2687             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2688             conf->fullsync == 0) {
2689                 *skipped = 1;
2690                 return max_sector - sector_nr;
2691         }
2692         /* before building a request, check if we can skip these blocks..
2693          * This call the bitmap_start_sync doesn't actually record anything
2694          */
2695         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2696             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2697                 /* We can skip this block, and probably several more */
2698                 *skipped = 1;
2699                 return sync_blocks;
2700         }
2701
2702         /*
2703          * If there is non-resync activity waiting for a turn, then let it
2704          * though before starting on this new sync request.
2705          */
2706         if (atomic_read(&conf->nr_waiting[idx]))
2707                 schedule_timeout_uninterruptible(1);
2708
2709         /* we are incrementing sector_nr below. To be safe, we check against
2710          * sector_nr + two times RESYNC_SECTORS
2711          */
2712
2713         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2714                 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2715
2716
2717         if (raise_barrier(conf, sector_nr))
2718                 return 0;
2719
2720         r1_bio = raid1_alloc_init_r1buf(conf);
2721
2722         rcu_read_lock();
2723         /*
2724          * If we get a correctably read error during resync or recovery,
2725          * we might want to read from a different device.  So we
2726          * flag all drives that could conceivably be read from for READ,
2727          * and any others (which will be non-In_sync devices) for WRITE.
2728          * If a read fails, we try reading from something else for which READ
2729          * is OK.
2730          */
2731
2732         r1_bio->mddev = mddev;
2733         r1_bio->sector = sector_nr;
2734         r1_bio->state = 0;
2735         set_bit(R1BIO_IsSync, &r1_bio->state);
2736         /* make sure good_sectors won't go across barrier unit boundary */
2737         good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2738
2739         for (i = 0; i < conf->raid_disks * 2; i++) {
2740                 struct md_rdev *rdev;
2741                 bio = r1_bio->bios[i];
2742
2743                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2744                 if (rdev == NULL ||
2745                     test_bit(Faulty, &rdev->flags)) {
2746                         if (i < conf->raid_disks)
2747                                 still_degraded = 1;
2748                 } else if (!test_bit(In_sync, &rdev->flags)) {
2749                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2750                         bio->bi_end_io = end_sync_write;
2751                         write_targets ++;
2752                 } else {
2753                         /* may need to read from here */
2754                         sector_t first_bad = MaxSector;
2755                         int bad_sectors;
2756
2757                         if (is_badblock(rdev, sector_nr, good_sectors,
2758                                         &first_bad, &bad_sectors)) {
2759                                 if (first_bad > sector_nr)
2760                                         good_sectors = first_bad - sector_nr;
2761                                 else {
2762                                         bad_sectors -= (sector_nr - first_bad);
2763                                         if (min_bad == 0 ||
2764                                             min_bad > bad_sectors)
2765                                                 min_bad = bad_sectors;
2766                                 }
2767                         }
2768                         if (sector_nr < first_bad) {
2769                                 if (test_bit(WriteMostly, &rdev->flags)) {
2770                                         if (wonly < 0)
2771                                                 wonly = i;
2772                                 } else {
2773                                         if (disk < 0)
2774                                                 disk = i;
2775                                 }
2776                                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2777                                 bio->bi_end_io = end_sync_read;
2778                                 read_targets++;
2779                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2780                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2781                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2782                                 /*
2783                                  * The device is suitable for reading (InSync),
2784                                  * but has bad block(s) here. Let's try to correct them,
2785                                  * if we are doing resync or repair. Otherwise, leave
2786                                  * this device alone for this sync request.
2787                                  */
2788                                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2789                                 bio->bi_end_io = end_sync_write;
2790                                 write_targets++;
2791                         }
2792                 }
2793                 if (rdev && bio->bi_end_io) {
2794                         atomic_inc(&rdev->nr_pending);
2795                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2796                         bio_set_dev(bio, rdev->bdev);
2797                         if (test_bit(FailFast, &rdev->flags))
2798                                 bio->bi_opf |= MD_FAILFAST;
2799                 }
2800         }
2801         rcu_read_unlock();
2802         if (disk < 0)
2803                 disk = wonly;
2804         r1_bio->read_disk = disk;
2805
2806         if (read_targets == 0 && min_bad > 0) {
2807                 /* These sectors are bad on all InSync devices, so we
2808                  * need to mark them bad on all write targets
2809                  */
2810                 int ok = 1;
2811                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2812                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2813                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2814                                 ok = rdev_set_badblocks(rdev, sector_nr,
2815                                                         min_bad, 0
2816                                         ) && ok;
2817                         }
2818                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2819                 *skipped = 1;
2820                 put_buf(r1_bio);
2821
2822                 if (!ok) {
2823                         /* Cannot record the badblocks, so need to
2824                          * abort the resync.
2825                          * If there are multiple read targets, could just
2826                          * fail the really bad ones ???
2827                          */
2828                         conf->recovery_disabled = mddev->recovery_disabled;
2829                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2830                         return 0;
2831                 } else
2832                         return min_bad;
2833
2834         }
2835         if (min_bad > 0 && min_bad < good_sectors) {
2836                 /* only resync enough to reach the next bad->good
2837                  * transition */
2838                 good_sectors = min_bad;
2839         }
2840
2841         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2842                 /* extra read targets are also write targets */
2843                 write_targets += read_targets-1;
2844
2845         if (write_targets == 0 || read_targets == 0) {
2846                 /* There is nowhere to write, so all non-sync
2847                  * drives must be failed - so we are finished
2848                  */
2849                 sector_t rv;
2850                 if (min_bad > 0)
2851                         max_sector = sector_nr + min_bad;
2852                 rv = max_sector - sector_nr;
2853                 *skipped = 1;
2854                 put_buf(r1_bio);
2855                 return rv;
2856         }
2857
2858         if (max_sector > mddev->resync_max)
2859                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2860         if (max_sector > sector_nr + good_sectors)
2861                 max_sector = sector_nr + good_sectors;
2862         nr_sectors = 0;
2863         sync_blocks = 0;
2864         do {
2865                 struct page *page;
2866                 int len = PAGE_SIZE;
2867                 if (sector_nr + (len>>9) > max_sector)
2868                         len = (max_sector - sector_nr) << 9;
2869                 if (len == 0)
2870                         break;
2871                 if (sync_blocks == 0) {
2872                         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2873                                                   &sync_blocks, still_degraded) &&
2874                             !conf->fullsync &&
2875                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2876                                 break;
2877                         if ((len >> 9) > sync_blocks)
2878                                 len = sync_blocks<<9;
2879                 }
2880
2881                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2882                         struct resync_pages *rp;
2883
2884                         bio = r1_bio->bios[i];
2885                         rp = get_resync_pages(bio);
2886                         if (bio->bi_end_io) {
2887                                 page = resync_fetch_page(rp, page_idx);
2888
2889                                 /*
2890                                  * won't fail because the vec table is big
2891                                  * enough to hold all these pages
2892                                  */
2893                                 bio_add_page(bio, page, len, 0);
2894                         }
2895                 }
2896                 nr_sectors += len>>9;
2897                 sector_nr += len>>9;
2898                 sync_blocks -= (len>>9);
2899         } while (++page_idx < RESYNC_PAGES);
2900
2901         r1_bio->sectors = nr_sectors;
2902
2903         if (mddev_is_clustered(mddev) &&
2904                         conf->cluster_sync_high < sector_nr + nr_sectors) {
2905                 conf->cluster_sync_low = mddev->curr_resync_completed;
2906                 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2907                 /* Send resync message */
2908                 md_cluster_ops->resync_info_update(mddev,
2909                                 conf->cluster_sync_low,
2910                                 conf->cluster_sync_high);
2911         }
2912
2913         /* For a user-requested sync, we read all readable devices and do a
2914          * compare
2915          */
2916         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2917                 atomic_set(&r1_bio->remaining, read_targets);
2918                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2919                         bio = r1_bio->bios[i];
2920                         if (bio->bi_end_io == end_sync_read) {
2921                                 read_targets--;
2922                                 md_sync_acct_bio(bio, nr_sectors);
2923                                 if (read_targets == 1)
2924                                         bio->bi_opf &= ~MD_FAILFAST;
2925                                 submit_bio_noacct(bio);
2926                         }
2927                 }
2928         } else {
2929                 atomic_set(&r1_bio->remaining, 1);
2930                 bio = r1_bio->bios[r1_bio->read_disk];
2931                 md_sync_acct_bio(bio, nr_sectors);
2932                 if (read_targets == 1)
2933                         bio->bi_opf &= ~MD_FAILFAST;
2934                 submit_bio_noacct(bio);
2935         }
2936         return nr_sectors;
2937 }
2938
2939 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2940 {
2941         if (sectors)
2942                 return sectors;
2943
2944         return mddev->dev_sectors;
2945 }
2946
2947 static struct r1conf *setup_conf(struct mddev *mddev)
2948 {
2949         struct r1conf *conf;
2950         int i;
2951         struct raid1_info *disk;
2952         struct md_rdev *rdev;
2953         int err = -ENOMEM;
2954
2955         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2956         if (!conf)
2957                 goto abort;
2958
2959         conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2960                                    sizeof(atomic_t), GFP_KERNEL);
2961         if (!conf->nr_pending)
2962                 goto abort;
2963
2964         conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2965                                    sizeof(atomic_t), GFP_KERNEL);
2966         if (!conf->nr_waiting)
2967                 goto abort;
2968
2969         conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2970                                   sizeof(atomic_t), GFP_KERNEL);
2971         if (!conf->nr_queued)
2972                 goto abort;
2973
2974         conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2975                                 sizeof(atomic_t), GFP_KERNEL);
2976         if (!conf->barrier)
2977                 goto abort;
2978
2979         conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2980                                             mddev->raid_disks, 2),
2981                                 GFP_KERNEL);
2982         if (!conf->mirrors)
2983                 goto abort;
2984
2985         conf->tmppage = alloc_page(GFP_KERNEL);
2986         if (!conf->tmppage)
2987                 goto abort;
2988
2989         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2990         if (!conf->poolinfo)
2991                 goto abort;
2992         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2993         err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2994                            rbio_pool_free, conf->poolinfo);
2995         if (err)
2996                 goto abort;
2997
2998         err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2999         if (err)
3000                 goto abort;
3001
3002         conf->poolinfo->mddev = mddev;
3003
3004         err = -EINVAL;
3005         spin_lock_init(&conf->device_lock);
3006         rdev_for_each(rdev, mddev) {
3007                 int disk_idx = rdev->raid_disk;
3008                 if (disk_idx >= mddev->raid_disks
3009                     || disk_idx < 0)
3010                         continue;
3011                 if (test_bit(Replacement, &rdev->flags))
3012                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
3013                 else
3014                         disk = conf->mirrors + disk_idx;
3015
3016                 if (disk->rdev)
3017                         goto abort;
3018                 disk->rdev = rdev;
3019                 disk->head_position = 0;
3020                 disk->seq_start = MaxSector;
3021         }
3022         conf->raid_disks = mddev->raid_disks;
3023         conf->mddev = mddev;
3024         INIT_LIST_HEAD(&conf->retry_list);
3025         INIT_LIST_HEAD(&conf->bio_end_io_list);
3026
3027         spin_lock_init(&conf->resync_lock);
3028         init_waitqueue_head(&conf->wait_barrier);
3029
3030         bio_list_init(&conf->pending_bio_list);
3031         conf->pending_count = 0;
3032         conf->recovery_disabled = mddev->recovery_disabled - 1;
3033
3034         err = -EIO;
3035         for (i = 0; i < conf->raid_disks * 2; i++) {
3036
3037                 disk = conf->mirrors + i;
3038
3039                 if (i < conf->raid_disks &&
3040                     disk[conf->raid_disks].rdev) {
3041                         /* This slot has a replacement. */
3042                         if (!disk->rdev) {
3043                                 /* No original, just make the replacement
3044                                  * a recovering spare
3045                                  */
3046                                 disk->rdev =
3047                                         disk[conf->raid_disks].rdev;
3048                                 disk[conf->raid_disks].rdev = NULL;
3049                         } else if (!test_bit(In_sync, &disk->rdev->flags))
3050                                 /* Original is not in_sync - bad */
3051                                 goto abort;
3052                 }
3053
3054                 if (!disk->rdev ||
3055                     !test_bit(In_sync, &disk->rdev->flags)) {
3056                         disk->head_position = 0;
3057                         if (disk->rdev &&
3058                             (disk->rdev->saved_raid_disk < 0))
3059                                 conf->fullsync = 1;
3060                 }
3061         }
3062
3063         err = -ENOMEM;
3064         conf->thread = md_register_thread(raid1d, mddev, "raid1");
3065         if (!conf->thread)
3066                 goto abort;
3067
3068         return conf;
3069
3070  abort:
3071         if (conf) {
3072                 mempool_exit(&conf->r1bio_pool);
3073                 kfree(conf->mirrors);
3074                 safe_put_page(conf->tmppage);
3075                 kfree(conf->poolinfo);
3076                 kfree(conf->nr_pending);
3077                 kfree(conf->nr_waiting);
3078                 kfree(conf->nr_queued);
3079                 kfree(conf->barrier);
3080                 bioset_exit(&conf->bio_split);
3081                 kfree(conf);
3082         }
3083         return ERR_PTR(err);
3084 }
3085
3086 static void raid1_free(struct mddev *mddev, void *priv);
3087 static int raid1_run(struct mddev *mddev)
3088 {
3089         struct r1conf *conf;
3090         int i;
3091         struct md_rdev *rdev;
3092         int ret;
3093         bool discard_supported = false;
3094
3095         if (mddev->level != 1) {
3096                 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3097                         mdname(mddev), mddev->level);
3098                 return -EIO;
3099         }
3100         if (mddev->reshape_position != MaxSector) {
3101                 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3102                         mdname(mddev));
3103                 return -EIO;
3104         }
3105         if (mddev_init_writes_pending(mddev) < 0)
3106                 return -ENOMEM;
3107         /*
3108          * copy the already verified devices into our private RAID1
3109          * bookkeeping area. [whatever we allocate in run(),
3110          * should be freed in raid1_free()]
3111          */
3112         if (mddev->private == NULL)
3113                 conf = setup_conf(mddev);
3114         else
3115                 conf = mddev->private;
3116
3117         if (IS_ERR(conf))
3118                 return PTR_ERR(conf);
3119
3120         if (mddev->queue) {
3121                 blk_queue_max_write_same_sectors(mddev->queue, 0);
3122                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3123         }
3124
3125         rdev_for_each(rdev, mddev) {
3126                 if (!mddev->gendisk)
3127                         continue;
3128                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3129                                   rdev->data_offset << 9);
3130                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3131                         discard_supported = true;
3132         }
3133
3134         mddev->degraded = 0;
3135         for (i = 0; i < conf->raid_disks; i++)
3136                 if (conf->mirrors[i].rdev == NULL ||
3137                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3138                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3139                         mddev->degraded++;
3140         /*
3141          * RAID1 needs at least one disk in active
3142          */
3143         if (conf->raid_disks - mddev->degraded < 1) {
3144                 ret = -EINVAL;
3145                 goto abort;
3146         }
3147
3148         if (conf->raid_disks - mddev->degraded == 1)
3149                 mddev->recovery_cp = MaxSector;
3150
3151         if (mddev->recovery_cp != MaxSector)
3152                 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3153                         mdname(mddev));
3154         pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3155                 mdname(mddev), mddev->raid_disks - mddev->degraded,
3156                 mddev->raid_disks);
3157
3158         /*
3159          * Ok, everything is just fine now
3160          */
3161         mddev->thread = conf->thread;
3162         conf->thread = NULL;
3163         mddev->private = conf;
3164         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3165
3166         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3167
3168         if (mddev->queue) {
3169                 if (discard_supported)
3170                         blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3171                                                 mddev->queue);
3172                 else
3173                         blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3174                                                   mddev->queue);
3175         }
3176
3177         ret = md_integrity_register(mddev);
3178         if (ret) {
3179                 md_unregister_thread(&mddev->thread);
3180                 goto abort;
3181         }
3182         return 0;
3183
3184 abort:
3185         raid1_free(mddev, conf);
3186         return ret;
3187 }
3188
3189 static void raid1_free(struct mddev *mddev, void *priv)
3190 {
3191         struct r1conf *conf = priv;
3192
3193         mempool_exit(&conf->r1bio_pool);
3194         kfree(conf->mirrors);
3195         safe_put_page(conf->tmppage);
3196         kfree(conf->poolinfo);
3197         kfree(conf->nr_pending);
3198         kfree(conf->nr_waiting);
3199         kfree(conf->nr_queued);
3200         kfree(conf->barrier);
3201         bioset_exit(&conf->bio_split);
3202         kfree(conf);
3203 }
3204
3205 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3206 {
3207         /* no resync is happening, and there is enough space
3208          * on all devices, so we can resize.
3209          * We need to make sure resync covers any new space.
3210          * If the array is shrinking we should possibly wait until
3211          * any io in the removed space completes, but it hardly seems
3212          * worth it.
3213          */
3214         sector_t newsize = raid1_size(mddev, sectors, 0);
3215         if (mddev->external_size &&
3216             mddev->array_sectors > newsize)
3217                 return -EINVAL;
3218         if (mddev->bitmap) {
3219                 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3220                 if (ret)
3221                         return ret;
3222         }
3223         md_set_array_sectors(mddev, newsize);
3224         if (sectors > mddev->dev_sectors &&
3225             mddev->recovery_cp > mddev->dev_sectors) {
3226                 mddev->recovery_cp = mddev->dev_sectors;
3227                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3228         }
3229         mddev->dev_sectors = sectors;
3230         mddev->resync_max_sectors = sectors;
3231         return 0;
3232 }
3233
3234 static int raid1_reshape(struct mddev *mddev)
3235 {
3236         /* We need to:
3237          * 1/ resize the r1bio_pool
3238          * 2/ resize conf->mirrors
3239          *
3240          * We allocate a new r1bio_pool if we can.
3241          * Then raise a device barrier and wait until all IO stops.
3242          * Then resize conf->mirrors and swap in the new r1bio pool.
3243          *
3244          * At the same time, we "pack" the devices so that all the missing
3245          * devices have the higher raid_disk numbers.
3246          */
3247         mempool_t newpool, oldpool;
3248         struct pool_info *newpoolinfo;
3249         struct raid1_info *newmirrors;
3250         struct r1conf *conf = mddev->private;
3251         int cnt, raid_disks;
3252         unsigned long flags;
3253         int d, d2;
3254         int ret;
3255
3256         memset(&newpool, 0, sizeof(newpool));
3257         memset(&oldpool, 0, sizeof(oldpool));
3258
3259         /* Cannot change chunk_size, layout, or level */
3260         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3261             mddev->layout != mddev->new_layout ||
3262             mddev->level != mddev->new_level) {
3263                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3264                 mddev->new_layout = mddev->layout;
3265                 mddev->new_level = mddev->level;
3266                 return -EINVAL;
3267         }
3268
3269         if (!mddev_is_clustered(mddev))
3270                 md_allow_write(mddev);
3271
3272         raid_disks = mddev->raid_disks + mddev->delta_disks;
3273
3274         if (raid_disks < conf->raid_disks) {
3275                 cnt=0;
3276                 for (d= 0; d < conf->raid_disks; d++)
3277                         if (conf->mirrors[d].rdev)
3278                                 cnt++;
3279                 if (cnt > raid_disks)
3280                         return -EBUSY;
3281         }
3282
3283         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3284         if (!newpoolinfo)
3285                 return -ENOMEM;
3286         newpoolinfo->mddev = mddev;
3287         newpoolinfo->raid_disks = raid_disks * 2;
3288
3289         ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3290                            rbio_pool_free, newpoolinfo);
3291         if (ret) {
3292                 kfree(newpoolinfo);
3293                 return ret;
3294         }
3295         newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3296                                          raid_disks, 2),
3297                              GFP_KERNEL);
3298         if (!newmirrors) {
3299                 kfree(newpoolinfo);
3300                 mempool_exit(&newpool);
3301                 return -ENOMEM;
3302         }
3303
3304         freeze_array(conf, 0);
3305
3306         /* ok, everything is stopped */
3307         oldpool = conf->r1bio_pool;
3308         conf->r1bio_pool = newpool;
3309
3310         for (d = d2 = 0; d < conf->raid_disks; d++) {
3311                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3312                 if (rdev && rdev->raid_disk != d2) {
3313                         sysfs_unlink_rdev(mddev, rdev);
3314                         rdev->raid_disk = d2;
3315                         sysfs_unlink_rdev(mddev, rdev);
3316                         if (sysfs_link_rdev(mddev, rdev))
3317                                 pr_warn("md/raid1:%s: cannot register rd%d\n",
3318                                         mdname(mddev), rdev->raid_disk);
3319                 }
3320                 if (rdev)
3321                         newmirrors[d2++].rdev = rdev;
3322         }
3323         kfree(conf->mirrors);
3324         conf->mirrors = newmirrors;
3325         kfree(conf->poolinfo);
3326         conf->poolinfo = newpoolinfo;
3327
3328         spin_lock_irqsave(&conf->device_lock, flags);
3329         mddev->degraded += (raid_disks - conf->raid_disks);
3330         spin_unlock_irqrestore(&conf->device_lock, flags);
3331         conf->raid_disks = mddev->raid_disks = raid_disks;
3332         mddev->delta_disks = 0;
3333
3334         unfreeze_array(conf);
3335
3336         set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3337         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3338         md_wakeup_thread(mddev->thread);
3339
3340         mempool_exit(&oldpool);
3341         return 0;
3342 }
3343
3344 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3345 {
3346         struct r1conf *conf = mddev->private;
3347
3348         if (quiesce)
3349                 freeze_array(conf, 0);
3350         else
3351                 unfreeze_array(conf);
3352 }
3353
3354 static void *raid1_takeover(struct mddev *mddev)
3355 {
3356         /* raid1 can take over:
3357          *  raid5 with 2 devices, any layout or chunk size
3358          */
3359         if (mddev->level == 5 && mddev->raid_disks == 2) {
3360                 struct r1conf *conf;
3361                 mddev->new_level = 1;
3362                 mddev->new_layout = 0;
3363                 mddev->new_chunk_sectors = 0;
3364                 conf = setup_conf(mddev);
3365                 if (!IS_ERR(conf)) {
3366                         /* Array must appear to be quiesced */
3367                         conf->array_frozen = 1;
3368                         mddev_clear_unsupported_flags(mddev,
3369                                 UNSUPPORTED_MDDEV_FLAGS);
3370                 }
3371                 return conf;
3372         }
3373         return ERR_PTR(-EINVAL);
3374 }
3375
3376 static struct md_personality raid1_personality =
3377 {
3378         .name           = "raid1",
3379         .level          = 1,
3380         .owner          = THIS_MODULE,
3381         .make_request   = raid1_make_request,
3382         .run            = raid1_run,
3383         .free           = raid1_free,
3384         .status         = raid1_status,
3385         .error_handler  = raid1_error,
3386         .hot_add_disk   = raid1_add_disk,
3387         .hot_remove_disk= raid1_remove_disk,
3388         .spare_active   = raid1_spare_active,
3389         .sync_request   = raid1_sync_request,
3390         .resize         = raid1_resize,
3391         .size           = raid1_size,
3392         .check_reshape  = raid1_reshape,
3393         .quiesce        = raid1_quiesce,
3394         .takeover       = raid1_takeover,
3395 };
3396
3397 static int __init raid_init(void)
3398 {
3399         return register_md_personality(&raid1_personality);
3400 }
3401
3402 static void raid_exit(void)
3403 {
3404         unregister_md_personality(&raid1_personality);
3405 }
3406
3407 module_init(raid_init);
3408 module_exit(raid_exit);
3409 MODULE_LICENSE("GPL");
3410 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3411 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3412 MODULE_ALIAS("md-raid1");
3413 MODULE_ALIAS("md-level-1");
3414
3415 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);