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         bio_endio(bio);
 304 }
 305
 306 static void raid_end_bio_io(struct r1bio *r1_bio)
 307 {
 308         struct bio *bio = r1_bio->master_bio;
 309         struct r1conf *conf = r1_bio->mddev->private;
 310
 311         /* if nobody has done the final endio yet, do it now */
 312         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 313                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 314                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 315                          (unsigned long long) bio->bi_iter.bi_sector,
 316                          (unsigned long long) bio_end_sector(bio) - 1);
 317
 318                 call_bio_endio(r1_bio);
 319         }
 320         /*
 321          * Wake up any possible resync thread that waits for the device
 322          * to go idle.  All I/Os, even write-behind writes, are done.
 323          */
 324         allow_barrier(conf, r1_bio->sector);
 325
 326         free_r1bio(r1_bio);
 327 }
 328
 329 /*
 330  * Update disk head position estimator based on IRQ completion info.
 331  */
 332 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 333 {
 334         struct r1conf *conf = r1_bio->mddev->private;
 335
 336         conf->mirrors[disk].head_position =
 337                 r1_bio->sector + (r1_bio->sectors);
 338 }
 339
 340 /*
 341  * Find the disk number which triggered given bio
 342  */
 343 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 344 {
 345         int mirror;
 346         struct r1conf *conf = r1_bio->mddev->private;
 347         int raid_disks = conf->raid_disks;
 348
 349         for (mirror = 0; mirror < raid_disks * 2; mirror++)
 350                 if (r1_bio->bios[mirror] == bio)
 351                         break;
 352
 353         BUG_ON(mirror == raid_disks * 2);
 354         update_head_pos(mirror, r1_bio);
 355
 356         return mirror;
 357 }
 358
 359 static void raid1_end_read_request(struct bio *bio)
 360 {
 361         int uptodate = !bio->bi_status;
 362         struct r1bio *r1_bio = bio->bi_private;
 363         struct r1conf *conf = r1_bio->mddev->private;
 364         struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 365
 366         /*
 367          * this branch is our 'one mirror IO has finished' event handler:
 368          */
 369         update_head_pos(r1_bio->read_disk, r1_bio);
 370
 371         if (uptodate)
 372                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 373         else if (test_bit(FailFast, &rdev->flags) &&
 374                  test_bit(R1BIO_FailFast, &r1_bio->state))
 375                 /* This was a fail-fast read so we definitely
 376                  * want to retry */
 377                 ;
 378         else {
 379                 /* If all other devices have failed, we want to return
 380                  * the error upwards rather than fail the last device.
 381                  * Here we redefine "uptodate" to mean "Don't want to retry"
 382                  */
 383                 unsigned long flags;
 384                 spin_lock_irqsave(&conf->device_lock, flags);
 385                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 386                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 387                      test_bit(In_sync, &rdev->flags)))
 388                         uptodate = 1;
 389                 spin_unlock_irqrestore(&conf->device_lock, flags);
 390         }
 391
 392         if (uptodate) {
 393                 raid_end_bio_io(r1_bio);
 394                 rdev_dec_pending(rdev, conf->mddev);
 395         } else {
 396                 /*
 397                  * oops, read error:
 398                  */
 399                 char b[BDEVNAME_SIZE];
 400                 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
 401                                    mdname(conf->mddev),
 402                                    bdevname(rdev->bdev, b),
 403                                    (unsigned long long)r1_bio->sector);
 404                 set_bit(R1BIO_ReadError, &r1_bio->state);
 405                 reschedule_retry(r1_bio);
 406                 /* don't drop the reference on read_disk yet */
 407         }
 408 }
 409
 410 static void close_write(struct r1bio *r1_bio)
 411 {
 412         /* it really is the end of this request */
 413         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 414                 bio_free_pages(r1_bio->behind_master_bio);
 415                 bio_put(r1_bio->behind_master_bio);
 416                 r1_bio->behind_master_bio = NULL;
 417         }
 418         /* clear the bitmap if all writes complete successfully */
 419         md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 420                            r1_bio->sectors,
 421                            !test_bit(R1BIO_Degraded, &r1_bio->state),
 422                            test_bit(R1BIO_BehindIO, &r1_bio->state));
 423         md_write_end(r1_bio->mddev);
 424 }
 425
 426 static void r1_bio_write_done(struct r1bio *r1_bio)
 427 {
 428         if (!atomic_dec_and_test(&r1_bio->remaining))
 429                 return;
 430
 431         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 432                 reschedule_retry(r1_bio);
 433         else {
 434                 close_write(r1_bio);
 435                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 436                         reschedule_retry(r1_bio);
 437                 else
 438                         raid_end_bio_io(r1_bio);
 439         }
 440 }
 441
 442 static void raid1_end_write_request(struct bio *bio)
 443 {
 444         struct r1bio *r1_bio = bio->bi_private;
 445         int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 446         struct r1conf *conf = r1_bio->mddev->private;
 447         struct bio *to_put = NULL;
 448         int mirror = find_bio_disk(r1_bio, bio);
 449         struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 450         bool discard_error;
 451         sector_t lo = r1_bio->sector;
 452         sector_t hi = r1_bio->sector + r1_bio->sectors;
 453
 454         discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 455
 456         /*
 457          * 'one mirror IO has finished' event handler:
 458          */
 459         if (bio->bi_status && !discard_error) {
 460                 set_bit(WriteErrorSeen, &rdev->flags);
 461                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
 462                         set_bit(MD_RECOVERY_NEEDED, &
 463                                 conf->mddev->recovery);
 464
 465                 if (test_bit(FailFast, &rdev->flags) &&
 466                     (bio->bi_opf & MD_FAILFAST) &&
 467                     /* We never try FailFast to WriteMostly devices */
 468                     !test_bit(WriteMostly, &rdev->flags)) {
 469                         md_error(r1_bio->mddev, rdev);
 470                 }
 471
 472                 /*
 473                  * When the device is faulty, it is not necessary to
 474                  * handle write error.
 475                  * For failfast, this is the only remaining device,
 476                  * We need to retry the write without FailFast.
 477                  */
 478                 if (!test_bit(Faulty, &rdev->flags))
 479                         set_bit(R1BIO_WriteError, &r1_bio->state);
 480                 else {
 481                         /* Finished with this branch */
 482                         r1_bio->bios[mirror] = NULL;
 483                         to_put = bio;
 484                 }
 485         } else {
 486                 /*
 487                  * Set R1BIO_Uptodate in our master bio, so that we
 488                  * will return a good error code for to the higher
 489                  * levels even if IO on some other mirrored buffer
 490                  * fails.
 491                  *
 492                  * The 'master' represents the composite IO operation
 493                  * to user-side. So if something waits for IO, then it
 494                  * will wait for the 'master' bio.
 495                  */
 496                 sector_t first_bad;
 497                 int bad_sectors;
 498
 499                 r1_bio->bios[mirror] = NULL;
 500                 to_put = bio;
 501                 /*
 502                  * Do not set R1BIO_Uptodate if the current device is
 503                  * rebuilding or Faulty. This is because we cannot use
 504                  * such device for properly reading the data back (we could
 505                  * potentially use it, if the current write would have felt
 506                  * before rdev->recovery_offset, but for simplicity we don't
 507                  * check this here.
 508                  */
 509                 if (test_bit(In_sync, &rdev->flags) &&
 510                     !test_bit(Faulty, &rdev->flags))
 511                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 512
 513                 /* Maybe we can clear some bad blocks. */
 514                 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 515                                 &first_bad, &bad_sectors) && !discard_error) {
 516                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 517                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 518                 }
 519         }
 520
 521         if (behind) {
 522                 if (test_bit(CollisionCheck, &rdev->flags))
 523                         remove_serial(rdev, lo, hi);
 524                 if (test_bit(WriteMostly, &rdev->flags))
 525                         atomic_dec(&r1_bio->behind_remaining);
 526
 527                 /*
 528                  * In behind mode, we ACK the master bio once the I/O
 529                  * has safely reached all non-writemostly
 530                  * disks. Setting the Returned bit ensures that this
 531                  * gets done only once -- we don't ever want to return
 532                  * -EIO here, instead we'll wait
 533                  */
 534                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 535                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 536                         /* Maybe we can return now */
 537                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 538                                 struct bio *mbio = r1_bio->master_bio;
 539                                 pr_debug("raid1: behind end write sectors"
 540                                          " %llu-%llu\n",
 541                                          (unsigned long long) mbio->bi_iter.bi_sector,
 542                                          (unsigned long long) bio_end_sector(mbio) - 1);
 543                                 call_bio_endio(r1_bio);
 544                         }
 545                 }
 546         } else if (rdev->mddev->serialize_policy)
 547                 remove_serial(rdev, lo, hi);
 548         if (r1_bio->bios[mirror] == NULL)
 549                 rdev_dec_pending(rdev, conf->mddev);
 550
 551         /*
 552          * Let's see if all mirrored write operations have finished
 553          * already.
 554          */
 555         r1_bio_write_done(r1_bio);
 556
 557         if (to_put)
 558                 bio_put(to_put);
 559 }
 560
 561 static sector_t align_to_barrier_unit_end(sector_t start_sector,
 562                                           sector_t sectors)
 563 {
 564         sector_t len;
 565
 566         WARN_ON(sectors == 0);
 567         /*
 568          * len is the number of sectors from start_sector to end of the
 569          * barrier unit which start_sector belongs to.
 570          */
 571         len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 572               start_sector;
 573
 574         if (len > sectors)
 575                 len = sectors;
 576
 577         return len;
 578 }
 579
 580 /*
 581  * This routine returns the disk from which the requested read should
 582  * be done. There is a per-array 'next expected sequential IO' sector
 583  * number - if this matches on the next IO then we use the last disk.
 584  * There is also a per-disk 'last know head position' sector that is
 585  * maintained from IRQ contexts, both the normal and the resync IO
 586  * completion handlers update this position correctly. If there is no
 587  * perfect sequential match then we pick the disk whose head is closest.
 588  *
 589  * If there are 2 mirrors in the same 2 devices, performance degrades
 590  * because position is mirror, not device based.
 591  *
 592  * The rdev for the device selected will have nr_pending incremented.
 593  */
 594 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 595 {
 596         const sector_t this_sector = r1_bio->sector;
 597         int sectors;
 598         int best_good_sectors;
 599         int best_disk, best_dist_disk, best_pending_disk;
 600         int has_nonrot_disk;
 601         int disk;
 602         sector_t best_dist;
 603         unsigned int min_pending;
 604         struct md_rdev *rdev;
 605         int choose_first;
 606         int choose_next_idle;
 607
 608         rcu_read_lock();
 609         /*
 610          * Check if we can balance. We can balance on the whole
 611          * device if no resync is going on, or below the resync window.
 612          * We take the first readable disk when above the resync window.
 613          */
 614  retry:
 615         sectors = r1_bio->sectors;
 616         best_disk = -1;
 617         best_dist_disk = -1;
 618         best_dist = MaxSector;
 619         best_pending_disk = -1;
 620         min_pending = UINT_MAX;
 621         best_good_sectors = 0;
 622         has_nonrot_disk = 0;
 623         choose_next_idle = 0;
 624         clear_bit(R1BIO_FailFast, &r1_bio->state);
 625
 626         if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 627             (mddev_is_clustered(conf->mddev) &&
 628             md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 629                     this_sector + sectors)))
 630                 choose_first = 1;
 631         else
 632                 choose_first = 0;
 633
 634         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 635                 sector_t dist;
 636                 sector_t first_bad;
 637                 int bad_sectors;
 638                 unsigned int pending;
 639                 bool nonrot;
 640
 641                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 642                 if (r1_bio->bios[disk] == IO_BLOCKED
 643                     || rdev == NULL
 644                     || test_bit(Faulty, &rdev->flags))
 645                         continue;
 646                 if (!test_bit(In_sync, &rdev->flags) &&
 647                     rdev->recovery_offset < this_sector + sectors)
 648                         continue;
 649                 if (test_bit(WriteMostly, &rdev->flags)) {
 650                         /* Don't balance among write-mostly, just
 651                          * use the first as a last resort */
 652                         if (best_dist_disk < 0) {
 653                                 if (is_badblock(rdev, this_sector, sectors,
 654                                                 &first_bad, &bad_sectors)) {
 655                                         if (first_bad <= this_sector)
 656                                                 /* Cannot use this */
 657                                                 continue;
 658                                         best_good_sectors = first_bad - this_sector;
 659                                 } else
 660                                         best_good_sectors = sectors;
 661                                 best_dist_disk = disk;
 662                                 best_pending_disk = disk;
 663                         }
 664                         continue;
 665                 }
 666                 /* This is a reasonable device to use.  It might
 667                  * even be best.
 668                  */
 669                 if (is_badblock(rdev, this_sector, sectors,
 670                                 &first_bad, &bad_sectors)) {
 671                         if (best_dist < MaxSector)
 672                                 /* already have a better device */
 673                                 continue;
 674                         if (first_bad <= this_sector) {
 675                                 /* cannot read here. If this is the 'primary'
 676                                  * device, then we must not read beyond
 677                                  * bad_sectors from another device..
 678                                  */
 679                                 bad_sectors -= (this_sector - first_bad);
 680                                 if (choose_first && sectors > bad_sectors)
 681                                         sectors = bad_sectors;
 682                                 if (best_good_sectors > sectors)
 683                                         best_good_sectors = sectors;
 684
 685                         } else {
 686                                 sector_t good_sectors = first_bad - this_sector;
 687                                 if (good_sectors > best_good_sectors) {
 688                                         best_good_sectors = good_sectors;
 689                                         best_disk = disk;
 690                                 }
 691                                 if (choose_first)
 692                                         break;
 693                         }
 694                         continue;
 695                 } else {
 696                         if ((sectors > best_good_sectors) && (best_disk >= 0))
 697                                 best_disk = -1;
 698                         best_good_sectors = sectors;
 699                 }
 700
 701                 if (best_disk >= 0)
 702                         /* At least two disks to choose from so failfast is OK */
 703                         set_bit(R1BIO_FailFast, &r1_bio->state);
 704
 705                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 706                 has_nonrot_disk |= nonrot;
 707                 pending = atomic_read(&rdev->nr_pending);
 708                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 709                 if (choose_first) {
 710                         best_disk = disk;
 711                         break;
 712                 }
 713                 /* Don't change to another disk for sequential reads */
 714                 if (conf->mirrors[disk].next_seq_sect == this_sector
 715                     || dist == 0) {
 716                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 717                         struct raid1_info *mirror = &conf->mirrors[disk];
 718
 719                         best_disk = disk;
 720                         /*
 721                          * If buffered sequential IO size exceeds optimal
 722                          * iosize, check if there is idle disk. If yes, choose
 723                          * the idle disk. read_balance could already choose an
 724                          * idle disk before noticing it's a sequential IO in
 725                          * this disk. This doesn't matter because this disk
 726                          * will idle, next time it will be utilized after the
 727                          * first disk has IO size exceeds optimal iosize. In
 728                          * this way, iosize of the first disk will be optimal
 729                          * iosize at least. iosize of the second disk might be
 730                          * small, but not a big deal since when the second disk
 731                          * starts IO, the first disk is likely still busy.
 732                          */
 733                         if (nonrot && opt_iosize > 0 &&
 734                             mirror->seq_start != MaxSector &&
 735                             mirror->next_seq_sect > opt_iosize &&
 736                             mirror->next_seq_sect - opt_iosize >=
 737                             mirror->seq_start) {
 738                                 choose_next_idle = 1;
 739                                 continue;
 740                         }
 741                         break;
 742                 }
 743
 744                 if (choose_next_idle)
 745                         continue;
 746
 747                 if (min_pending > pending) {
 748                         min_pending = pending;
 749                         best_pending_disk = disk;
 750                 }
 751
 752                 if (dist < best_dist) {
 753                         best_dist = dist;
 754                         best_dist_disk = disk;
 755                 }
 756         }
 757
 758         /*
 759          * If all disks are rotational, choose the closest disk. If any disk is
 760          * non-rotational, choose the disk with less pending request even the
 761          * disk is rotational, which might/might not be optimal for raids with
 762          * mixed ratation/non-rotational disks depending on workload.
 763          */
 764         if (best_disk == -1) {
 765                 if (has_nonrot_disk || min_pending == 0)
 766                         best_disk = best_pending_disk;
 767                 else
 768                         best_disk = best_dist_disk;
 769         }
 770
 771         if (best_disk >= 0) {
 772                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 773                 if (!rdev)
 774                         goto retry;
 775                 atomic_inc(&rdev->nr_pending);
 776                 sectors = best_good_sectors;
 777
 778                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 779                         conf->mirrors[best_disk].seq_start = this_sector;
 780
 781                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 782         }
 783         rcu_read_unlock();
 784         *max_sectors = sectors;
 785
 786         return best_disk;
 787 }
 788
 789 static int raid1_congested(struct mddev *mddev, int bits)
 790 {
 791         struct r1conf *conf = mddev->private;
 792         int i, ret = 0;
 793
 794         if ((bits & (1 << WB_async_congested)) &&
 795             conf->pending_count >= max_queued_requests)
 796                 return 1;
 797
 798         rcu_read_lock();
 799         for (i = 0; i < conf->raid_disks * 2; i++) {
 800                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 801                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 802                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 803
 804                         BUG_ON(!q);
 805
 806                         /* Note the '|| 1' - when read_balance prefers
 807                          * non-congested targets, it can be removed
 808                          */
 809                         if ((bits & (1 << WB_async_congested)) || 1)
 810                                 ret |= bdi_congested(q->backing_dev_info, bits);
 811                         else
 812                                 ret &= bdi_congested(q->backing_dev_info, bits);
 813                 }
 814         }
 815         rcu_read_unlock();
 816         return ret;
 817 }
 818
 819 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 820 {
 821         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
 822         md_bitmap_unplug(conf->mddev->bitmap);
 823         wake_up(&conf->wait_barrier);
 824
 825         while (bio) { /* submit pending writes */
 826                 struct bio *next = bio->bi_next;
 827                 struct md_rdev *rdev = (void *)bio->bi_disk;
 828                 bio->bi_next = NULL;
 829                 bio_set_dev(bio, rdev->bdev);
 830                 if (test_bit(Faulty, &rdev->flags)) {
 831                         bio_io_error(bio);
 832                 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
 833                                     !blk_queue_discard(bio->bi_disk->queue)))
 834                         /* Just ignore it */
 835                         bio_endio(bio);
 836                 else
 837                         generic_make_request(bio);
 838                 bio = next;
 839                 cond_resched();
 840         }
 841 }
 842
 843 static void flush_pending_writes(struct r1conf *conf)
 844 {
 845         /* Any writes that have been queued but are awaiting
 846          * bitmap updates get flushed here.
 847          */
 848         spin_lock_irq(&conf->device_lock);
 849
 850         if (conf->pending_bio_list.head) {
 851                 struct blk_plug plug;
 852                 struct bio *bio;
 853
 854                 bio = bio_list_get(&conf->pending_bio_list);
 855                 conf->pending_count = 0;
 856                 spin_unlock_irq(&conf->device_lock);
 857
 858                 /*
 859                  * As this is called in a wait_event() loop (see freeze_array),
 860                  * current->state might be TASK_UNINTERRUPTIBLE which will
 861                  * cause a warning when we prepare to wait again.  As it is
 862                  * rare that this path is taken, it is perfectly safe to force
 863                  * us to go around the wait_event() loop again, so the warning
 864                  * is a false-positive.  Silence the warning by resetting
 865                  * thread state
 866                  */
 867                 __set_current_state(TASK_RUNNING);
 868                 blk_start_plug(&plug);
 869                 flush_bio_list(conf, bio);
 870                 blk_finish_plug(&plug);
 871         } else
 872                 spin_unlock_irq(&conf->device_lock);
 873 }
 874
 875 /* Barriers....
 876  * Sometimes we need to suspend IO while we do something else,
 877  * either some resync/recovery, or reconfigure the array.
 878  * To do this we raise a 'barrier'.
 879  * The 'barrier' is a counter that can be raised multiple times
 880  * to count how many activities are happening which preclude
 881  * normal IO.
 882  * We can only raise the barrier if there is no pending IO.
 883  * i.e. if nr_pending == 0.
 884  * We choose only to raise the barrier if no-one is waiting for the
 885  * barrier to go down.  This means that as soon as an IO request
 886  * is ready, no other operations which require a barrier will start
 887  * until the IO request has had a chance.
 888  *
 889  * So: regular IO calls 'wait_barrier'.  When that returns there
 890  *    is no backgroup IO happening,  It must arrange to call
 891  *    allow_barrier when it has finished its IO.
 892  * backgroup IO calls must call raise_barrier.  Once that returns
 893  *    there is no normal IO happeing.  It must arrange to call
 894  *    lower_barrier when the particular background IO completes.
 895  *
 896  * If resync/recovery is interrupted, returns -EINTR;
 897  * Otherwise, returns 0.
 898  */
 899 static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
 900 {
 901         int idx = sector_to_idx(sector_nr);
 902
 903         spin_lock_irq(&conf->resync_lock);
 904
 905         /* Wait until no block IO is waiting */
 906         wait_event_lock_irq(conf->wait_barrier,
 907                             !atomic_read(&conf->nr_waiting[idx]),
 908                             conf->resync_lock);
 909
 910         /* block any new IO from starting */
 911         atomic_inc(&conf->barrier[idx]);
 912         /*
 913          * In raise_barrier() we firstly increase conf->barrier[idx] then
 914          * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 915          * increase conf->nr_pending[idx] then check conf->barrier[idx].
 916          * A memory barrier here to make sure conf->nr_pending[idx] won't
 917          * be fetched before conf->barrier[idx] is increased. Otherwise
 918          * there will be a race between raise_barrier() and _wait_barrier().
 919          */
 920         smp_mb__after_atomic();
 921
 922         /* For these conditions we must wait:
 923          * A: while the array is in frozen state
 924          * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 925          *    existing in corresponding I/O barrier bucket.
 926          * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 927          *    max resync count which allowed on current I/O barrier bucket.
 928          */
 929         wait_event_lock_irq(conf->wait_barrier,
 930                             (!conf->array_frozen &&
 931                              !atomic_read(&conf->nr_pending[idx]) &&
 932                              atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
 933                                 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
 934                             conf->resync_lock);
 935
 936         if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
 937                 atomic_dec(&conf->barrier[idx]);
 938                 spin_unlock_irq(&conf->resync_lock);
 939                 wake_up(&conf->wait_barrier);
 940                 return -EINTR;
 941         }
 942
 943         atomic_inc(&conf->nr_sync_pending);
 944         spin_unlock_irq(&conf->resync_lock);
 945
 946         return 0;
 947 }
 948
 949 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 950 {
 951         int idx = sector_to_idx(sector_nr);
 952
 953         BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 954
 955         atomic_dec(&conf->barrier[idx]);
 956         atomic_dec(&conf->nr_sync_pending);
 957         wake_up(&conf->wait_barrier);
 958 }
 959
 960 static void _wait_barrier(struct r1conf *conf, int idx)
 961 {
 962         /*
 963          * We need to increase conf->nr_pending[idx] very early here,
 964          * then raise_barrier() can be blocked when it waits for
 965          * conf->nr_pending[idx] to be 0. Then we can avoid holding
 966          * conf->resync_lock when there is no barrier raised in same
 967          * barrier unit bucket. Also if the array is frozen, I/O
 968          * should be blocked until array is unfrozen.
 969          */
 970         atomic_inc(&conf->nr_pending[idx]);
 971         /*
 972          * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
 973          * check conf->barrier[idx]. In raise_barrier() we firstly increase
 974          * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
 975          * barrier is necessary here to make sure conf->barrier[idx] won't be
 976          * fetched before conf->nr_pending[idx] is increased. Otherwise there
 977          * will be a race between _wait_barrier() and raise_barrier().
 978          */
 979         smp_mb__after_atomic();
 980
 981         /*
 982          * Don't worry about checking two atomic_t variables at same time
 983          * here. If during we check conf->barrier[idx], the array is
 984          * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
 985          * 0, it is safe to return and make the I/O continue. Because the
 986          * array is frozen, all I/O returned here will eventually complete
 987          * or be queued, no race will happen. See code comment in
 988          * frozen_array().
 989          */
 990         if (!READ_ONCE(conf->array_frozen) &&
 991             !atomic_read(&conf->barrier[idx]))
 992                 return;
 993
 994         /*
 995          * After holding conf->resync_lock, conf->nr_pending[idx]
 996          * should be decreased before waiting for barrier to drop.
 997          * Otherwise, we may encounter a race condition because
 998          * raise_barrer() might be waiting for conf->nr_pending[idx]
 999          * to be 0 at same time.
1000          */
1001         spin_lock_irq(&conf->resync_lock);
1002         atomic_inc(&conf->nr_waiting[idx]);
1003         atomic_dec(&conf->nr_pending[idx]);
1004         /*
1005          * In case freeze_array() is waiting for
1006          * get_unqueued_pending() == extra
1007          */
1008         wake_up(&conf->wait_barrier);
1009         /* Wait for the barrier in same barrier unit bucket to drop. */
1010         wait_event_lock_irq(conf->wait_barrier,
1011                             !conf->array_frozen &&
1012                              !atomic_read(&conf->barrier[idx]),
1013                             conf->resync_lock);
1014         atomic_inc(&conf->nr_pending[idx]);
1015         atomic_dec(&conf->nr_waiting[idx]);
1016         spin_unlock_irq(&conf->resync_lock);
1017 }
1018
1019 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
1020 {
1021         int idx = sector_to_idx(sector_nr);
1022
1023         /*
1024          * Very similar to _wait_barrier(). The difference is, for read
1025          * I/O we don't need wait for sync I/O, but if the whole array
1026          * is frozen, the read I/O still has to wait until the array is
1027          * unfrozen. Since there is no ordering requirement with
1028          * conf->barrier[idx] here, memory barrier is unnecessary as well.
1029          */
1030         atomic_inc(&conf->nr_pending[idx]);
1031
1032         if (!READ_ONCE(conf->array_frozen))
1033                 return;
1034
1035         spin_lock_irq(&conf->resync_lock);
1036         atomic_inc(&conf->nr_waiting[idx]);
1037         atomic_dec(&conf->nr_pending[idx]);
1038         /*
1039          * In case freeze_array() is waiting for
1040          * get_unqueued_pending() == extra
1041          */
1042         wake_up(&conf->wait_barrier);
1043         /* Wait for array to be unfrozen */
1044         wait_event_lock_irq(conf->wait_barrier,
1045                             !conf->array_frozen,
1046                             conf->resync_lock);
1047         atomic_inc(&conf->nr_pending[idx]);
1048         atomic_dec(&conf->nr_waiting[idx]);
1049         spin_unlock_irq(&conf->resync_lock);
1050 }
1051
1052 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1053 {
1054         int idx = sector_to_idx(sector_nr);
1055
1056         _wait_barrier(conf, idx);
1057 }
1058
1059 static void _allow_barrier(struct r1conf *conf, int idx)
1060 {
1061         atomic_dec(&conf->nr_pending[idx]);
1062         wake_up(&conf->wait_barrier);
1063 }
1064
1065 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1066 {
1067         int idx = sector_to_idx(sector_nr);
1068
1069         _allow_barrier(conf, idx);
1070 }
1071
1072 /* conf->resync_lock should be held */
1073 static int get_unqueued_pending(struct r1conf *conf)
1074 {
1075         int idx, ret;
1076
1077         ret = atomic_read(&conf->nr_sync_pending);
1078         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1079                 ret += atomic_read(&conf->nr_pending[idx]) -
1080                         atomic_read(&conf->nr_queued[idx]);
1081
1082         return ret;
1083 }
1084
1085 static void freeze_array(struct r1conf *conf, int extra)
1086 {
1087         /* Stop sync I/O and normal I/O and wait for everything to
1088          * go quiet.
1089          * This is called in two situations:
1090          * 1) management command handlers (reshape, remove disk, quiesce).
1091          * 2) one normal I/O request failed.
1092
1093          * After array_frozen is set to 1, new sync IO will be blocked at
1094          * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1095          * or wait_read_barrier(). The flying I/Os will either complete or be
1096          * queued. When everything goes quite, there are only queued I/Os left.
1097
1098          * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1099          * barrier bucket index which this I/O request hits. When all sync and
1100          * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1101          * of all conf->nr_queued[]. But normal I/O failure is an exception,
1102          * in handle_read_error(), we may call freeze_array() before trying to
1103          * fix the read error. In this case, the error read I/O is not queued,
1104          * so get_unqueued_pending() == 1.
1105          *
1106          * Therefore before this function returns, we need to wait until
1107          * get_unqueued_pendings(conf) gets equal to extra. For
1108          * normal I/O context, extra is 1, in rested situations extra is 0.
1109          */
1110         spin_lock_irq(&conf->resync_lock);
1111         conf->array_frozen = 1;
1112         raid1_log(conf->mddev, "wait freeze");
1113         wait_event_lock_irq_cmd(
1114                 conf->wait_barrier,
1115                 get_unqueued_pending(conf) == extra,
1116                 conf->resync_lock,
1117                 flush_pending_writes(conf));
1118         spin_unlock_irq(&conf->resync_lock);
1119 }
1120 static void unfreeze_array(struct r1conf *conf)
1121 {
1122         /* reverse the effect of the freeze */
1123         spin_lock_irq(&conf->resync_lock);
1124         conf->array_frozen = 0;
1125         spin_unlock_irq(&conf->resync_lock);
1126         wake_up(&conf->wait_barrier);
1127 }
1128
1129 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1130                                            struct bio *bio)
1131 {
1132         int size = bio->bi_iter.bi_size;
1133         unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1134         int i = 0;
1135         struct bio *behind_bio = NULL;
1136
1137         behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1138         if (!behind_bio)
1139                 return;
1140
1141         /* discard op, we don't support writezero/writesame yet */
1142         if (!bio_has_data(bio)) {
1143                 behind_bio->bi_iter.bi_size = size;
1144                 goto skip_copy;
1145         }
1146
1147         behind_bio->bi_write_hint = bio->bi_write_hint;
1148
1149         while (i < vcnt && size) {
1150                 struct page *page;
1151                 int len = min_t(int, PAGE_SIZE, size);
1152
1153                 page = alloc_page(GFP_NOIO);
1154                 if (unlikely(!page))
1155                         goto free_pages;
1156
1157                 bio_add_page(behind_bio, page, len, 0);
1158
1159                 size -= len;
1160                 i++;
1161         }
1162
1163         bio_copy_data(behind_bio, bio);
1164 skip_copy:
1165         r1_bio->behind_master_bio = behind_bio;
1166         set_bit(R1BIO_BehindIO, &r1_bio->state);
1167
1168         return;
1169
1170 free_pages:
1171         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1172                  bio->bi_iter.bi_size);
1173         bio_free_pages(behind_bio);
1174         bio_put(behind_bio);
1175 }
1176
1177 struct raid1_plug_cb {
1178         struct blk_plug_cb      cb;
1179         struct bio_list         pending;
1180         int                     pending_cnt;
1181 };
1182
1183 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1184 {
1185         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1186                                                   cb);
1187         struct mddev *mddev = plug->cb.data;
1188         struct r1conf *conf = mddev->private;
1189         struct bio *bio;
1190
1191         if (from_schedule || current->bio_list) {
1192                 spin_lock_irq(&conf->device_lock);
1193                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1194                 conf->pending_count += plug->pending_cnt;
1195                 spin_unlock_irq(&conf->device_lock);
1196                 wake_up(&conf->wait_barrier);
1197                 md_wakeup_thread(mddev->thread);
1198                 kfree(plug);
1199                 return;
1200         }
1201
1202         /* we aren't scheduling, so we can do the write-out directly. */
1203         bio = bio_list_get(&plug->pending);
1204         flush_bio_list(conf, bio);
1205         kfree(plug);
1206 }
1207
1208 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1209 {
1210         r1_bio->master_bio = bio;
1211         r1_bio->sectors = bio_sectors(bio);
1212         r1_bio->state = 0;
1213         r1_bio->mddev = mddev;
1214         r1_bio->sector = bio->bi_iter.bi_sector;
1215 }
1216
1217 static inline struct r1bio *
1218 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1219 {
1220         struct r1conf *conf = mddev->private;
1221         struct r1bio *r1_bio;
1222
1223         r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1224         /* Ensure no bio records IO_BLOCKED */
1225         memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1226         init_r1bio(r1_bio, mddev, bio);
1227         return r1_bio;
1228 }
1229
1230 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1231                                int max_read_sectors, struct r1bio *r1_bio)
1232 {
1233         struct r1conf *conf = mddev->private;
1234         struct raid1_info *mirror;
1235         struct bio *read_bio;
1236         struct bitmap *bitmap = mddev->bitmap;
1237         const int op = bio_op(bio);
1238         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1239         int max_sectors;
1240         int rdisk;
1241         bool print_msg = !!r1_bio;
1242         char b[BDEVNAME_SIZE];
1243
1244         /*
1245          * If r1_bio is set, we are blocking the raid1d thread
1246          * so there is a tiny risk of deadlock.  So ask for
1247          * emergency memory if needed.
1248          */
1249         gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1250
1251         if (print_msg) {
1252                 /* Need to get the block device name carefully */
1253                 struct md_rdev *rdev;
1254                 rcu_read_lock();
1255                 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1256                 if (rdev)
1257                         bdevname(rdev->bdev, b);
1258                 else
1259                         strcpy(b, "???");
1260                 rcu_read_unlock();
1261         }
1262
1263         /*
1264          * Still need barrier for READ in case that whole
1265          * array is frozen.
1266          */
1267         wait_read_barrier(conf, bio->bi_iter.bi_sector);
1268
1269         if (!r1_bio)
1270                 r1_bio = alloc_r1bio(mddev, bio);
1271         else
1272                 init_r1bio(r1_bio, mddev, bio);
1273         r1_bio->sectors = max_read_sectors;
1274
1275         /*
1276          * make_request() can abort the operation when read-ahead is being
1277          * used and no empty request is available.
1278          */
1279         rdisk = read_balance(conf, r1_bio, &max_sectors);
1280
1281         if (rdisk < 0) {
1282                 /* couldn't find anywhere to read from */
1283                 if (print_msg) {
1284                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1285                                             mdname(mddev),
1286                                             b,
1287                                             (unsigned long long)r1_bio->sector);
1288                 }
1289                 raid_end_bio_io(r1_bio);
1290                 return;
1291         }
1292         mirror = conf->mirrors + rdisk;
1293
1294         if (print_msg)
1295                 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1296                                     mdname(mddev),
1297                                     (unsigned long long)r1_bio->sector,
1298                                     bdevname(mirror->rdev->bdev, b));
1299
1300         if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1301             bitmap) {
1302                 /*
1303                  * Reading from a write-mostly device must take care not to
1304                  * over-take any writes that are 'behind'
1305                  */
1306                 raid1_log(mddev, "wait behind writes");
1307                 wait_event(bitmap->behind_wait,
1308                            atomic_read(&bitmap->behind_writes) == 0);
1309         }
1310
1311         if (max_sectors < bio_sectors(bio)) {
1312                 struct bio *split = bio_split(bio, max_sectors,
1313                                               gfp, &conf->bio_split);
1314                 bio_chain(split, bio);
1315                 generic_make_request(bio);
1316                 bio = split;
1317                 r1_bio->master_bio = bio;
1318                 r1_bio->sectors = max_sectors;
1319         }
1320
1321         r1_bio->read_disk = rdisk;
1322
1323         read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1324
1325         r1_bio->bios[rdisk] = read_bio;
1326
1327         read_bio->bi_iter.bi_sector = r1_bio->sector +
1328                 mirror->rdev->data_offset;
1329         bio_set_dev(read_bio, mirror->rdev->bdev);
1330         read_bio->bi_end_io = raid1_end_read_request;
1331         bio_set_op_attrs(read_bio, op, do_sync);
1332         if (test_bit(FailFast, &mirror->rdev->flags) &&
1333             test_bit(R1BIO_FailFast, &r1_bio->state))
1334                 read_bio->bi_opf |= MD_FAILFAST;
1335         read_bio->bi_private = r1_bio;
1336
1337         if (mddev->gendisk)
1338                 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1339                                 disk_devt(mddev->gendisk), r1_bio->sector);
1340
1341         generic_make_request(read_bio);
1342 }
1343
1344 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1345                                 int max_write_sectors)
1346 {
1347         struct r1conf *conf = mddev->private;
1348         struct r1bio *r1_bio;
1349         int i, disks;
1350         struct bitmap *bitmap = mddev->bitmap;
1351         unsigned long flags;
1352         struct md_rdev *blocked_rdev;
1353         struct blk_plug_cb *cb;
1354         struct raid1_plug_cb *plug = NULL;
1355         int first_clone;
1356         int max_sectors;
1357
1358         if (mddev_is_clustered(mddev) &&
1359              md_cluster_ops->area_resyncing(mddev, WRITE,
1360                      bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1361
1362                 DEFINE_WAIT(w);
1363                 for (;;) {
1364                         prepare_to_wait(&conf->wait_barrier,
1365                                         &w, TASK_IDLE);
1366                         if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1367                                                         bio->bi_iter.bi_sector,
1368                                                         bio_end_sector(bio)))
1369                                 break;
1370                         schedule();
1371                 }
1372                 finish_wait(&conf->wait_barrier, &w);
1373         }
1374
1375         /*
1376          * Register the new request and wait if the reconstruction
1377          * thread has put up a bar for new requests.
1378          * Continue immediately if no resync is active currently.
1379          */
1380         wait_barrier(conf, bio->bi_iter.bi_sector);
1381
1382         r1_bio = alloc_r1bio(mddev, bio);
1383         r1_bio->sectors = max_write_sectors;
1384
1385         if (conf->pending_count >= max_queued_requests) {
1386                 md_wakeup_thread(mddev->thread);
1387                 raid1_log(mddev, "wait queued");
1388                 wait_event(conf->wait_barrier,
1389                            conf->pending_count < max_queued_requests);
1390         }
1391         /* first select target devices under rcu_lock and
1392          * inc refcount on their rdev.  Record them by setting
1393          * bios[x] to bio
1394          * If there are known/acknowledged bad blocks on any device on
1395          * which we have seen a write error, we want to avoid writing those
1396          * blocks.
1397          * This potentially requires several writes to write around
1398          * the bad blocks.  Each set of writes gets it's own r1bio
1399          * with a set of bios attached.
1400          */
1401
1402         disks = conf->raid_disks * 2;
1403  retry_write:
1404         blocked_rdev = NULL;
1405         rcu_read_lock();
1406         max_sectors = r1_bio->sectors;
1407         for (i = 0;  i < disks; i++) {
1408                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1409                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1410                         atomic_inc(&rdev->nr_pending);
1411                         blocked_rdev = rdev;
1412                         break;
1413                 }
1414                 r1_bio->bios[i] = NULL;
1415                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1416                         if (i < conf->raid_disks)
1417                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1418                         continue;
1419                 }
1420
1421                 atomic_inc(&rdev->nr_pending);
1422                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1423                         sector_t first_bad;
1424                         int bad_sectors;
1425                         int is_bad;
1426
1427                         is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1428                                              &first_bad, &bad_sectors);
1429                         if (is_bad < 0) {
1430                                 /* mustn't write here until the bad block is
1431                                  * acknowledged*/
1432                                 set_bit(BlockedBadBlocks, &rdev->flags);
1433                                 blocked_rdev = rdev;
1434                                 break;
1435                         }
1436                         if (is_bad && first_bad <= r1_bio->sector) {
1437                                 /* Cannot write here at all */
1438                                 bad_sectors -= (r1_bio->sector - first_bad);
1439                                 if (bad_sectors < max_sectors)
1440                                         /* mustn't write more than bad_sectors
1441                                          * to other devices yet
1442                                          */
1443                                         max_sectors = bad_sectors;
1444                                 rdev_dec_pending(rdev, mddev);
1445                                 /* We don't set R1BIO_Degraded as that
1446                                  * only applies if the disk is
1447                                  * missing, so it might be re-added,
1448                                  * and we want to know to recover this
1449                                  * chunk.
1450                                  * In this case the device is here,
1451                                  * and the fact that this chunk is not
1452                                  * in-sync is recorded in the bad
1453                                  * block log
1454                                  */
1455                                 continue;
1456                         }
1457                         if (is_bad) {
1458                                 int good_sectors = first_bad - r1_bio->sector;
1459                                 if (good_sectors < max_sectors)
1460                                         max_sectors = good_sectors;
1461                         }
1462                 }
1463                 r1_bio->bios[i] = bio;
1464         }
1465         rcu_read_unlock();
1466
1467         if (unlikely(blocked_rdev)) {
1468                 /* Wait for this device to become unblocked */
1469                 int j;
1470
1471                 for (j = 0; j < i; j++)
1472                         if (r1_bio->bios[j])
1473                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1474                 r1_bio->state = 0;
1475                 allow_barrier(conf, bio->bi_iter.bi_sector);
1476                 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1477                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1478                 wait_barrier(conf, bio->bi_iter.bi_sector);
1479                 goto retry_write;
1480         }
1481
1482         if (max_sectors < bio_sectors(bio)) {
1483                 struct bio *split = bio_split(bio, max_sectors,
1484                                               GFP_NOIO, &conf->bio_split);
1485                 bio_chain(split, bio);
1486                 generic_make_request(bio);
1487                 bio = split;
1488                 r1_bio->master_bio = bio;
1489                 r1_bio->sectors = max_sectors;
1490         }
1491
1492         atomic_set(&r1_bio->remaining, 1);
1493         atomic_set(&r1_bio->behind_remaining, 0);
1494
1495         first_clone = 1;
1496
1497         for (i = 0; i < disks; i++) {
1498                 struct bio *mbio = NULL;
1499                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1500                 if (!r1_bio->bios[i])
1501                         continue;
1502
1503                 if (first_clone) {
1504                         /* do behind I/O ?
1505                          * Not if there are too many, or cannot
1506                          * allocate memory, or a reader on WriteMostly
1507                          * is waiting for behind writes to flush */
1508                         if (bitmap &&
1509                             (atomic_read(&bitmap->behind_writes)
1510                              < mddev->bitmap_info.max_write_behind) &&
1511                             !waitqueue_active(&bitmap->behind_wait)) {
1512                                 alloc_behind_master_bio(r1_bio, bio);
1513                         }
1514
1515                         md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1516                                              test_bit(R1BIO_BehindIO, &r1_bio->state));
1517                         first_clone = 0;
1518                 }
1519
1520                 if (r1_bio->behind_master_bio)
1521                         mbio = bio_clone_fast(r1_bio->behind_master_bio,
1522                                               GFP_NOIO, &mddev->bio_set);
1523                 else
1524                         mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1525
1526                 if (r1_bio->behind_master_bio) {
1527                         if (test_bit(CollisionCheck, &rdev->flags))
1528                                 wait_for_serialization(rdev, r1_bio);
1529                         if (test_bit(WriteMostly, &rdev->flags))
1530                                 atomic_inc(&r1_bio->behind_remaining);
1531                 } else if (mddev->serialize_policy)
1532                         wait_for_serialization(rdev, r1_bio);
1533
1534                 r1_bio->bios[i] = mbio;
1535
1536                 mbio->bi_iter.bi_sector = (r1_bio->sector +
1537                                    conf->mirrors[i].rdev->data_offset);
1538                 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1539                 mbio->bi_end_io = raid1_end_write_request;
1540                 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1541                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1542                     !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1543                     conf->raid_disks - mddev->degraded > 1)
1544                         mbio->bi_opf |= MD_FAILFAST;
1545                 mbio->bi_private = r1_bio;
1546
1547                 atomic_inc(&r1_bio->remaining);
1548
1549                 if (mddev->gendisk)
1550                         trace_block_bio_remap(mbio->bi_disk->queue,
1551                                               mbio, disk_devt(mddev->gendisk),
1552                                               r1_bio->sector);
1553                 /* flush_pending_writes() needs access to the rdev so...*/
1554                 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1555
1556                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1557                 if (cb)
1558                         plug = container_of(cb, struct raid1_plug_cb, cb);
1559                 else
1560                         plug = NULL;
1561                 if (plug) {
1562                         bio_list_add(&plug->pending, mbio);
1563                         plug->pending_cnt++;
1564                 } else {
1565                         spin_lock_irqsave(&conf->device_lock, flags);
1566                         bio_list_add(&conf->pending_bio_list, mbio);
1567                         conf->pending_count++;
1568                         spin_unlock_irqrestore(&conf->device_lock, flags);
1569                         md_wakeup_thread(mddev->thread);
1570                 }
1571         }
1572
1573         r1_bio_write_done(r1_bio);
1574
1575         /* In case raid1d snuck in to freeze_array */
1576         wake_up(&conf->wait_barrier);
1577 }
1578
1579 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1580 {
1581         sector_t sectors;
1582
1583         if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1584             && md_flush_request(mddev, bio))
1585                 return true;
1586
1587         /*
1588          * There is a limit to the maximum size, but
1589          * the read/write handler might find a lower limit
1590          * due to bad blocks.  To avoid multiple splits,
1591          * we pass the maximum number of sectors down
1592          * and let the lower level perform the split.
1593          */
1594         sectors = align_to_barrier_unit_end(
1595                 bio->bi_iter.bi_sector, bio_sectors(bio));
1596
1597         if (bio_data_dir(bio) == READ)
1598                 raid1_read_request(mddev, bio, sectors, NULL);
1599         else {
1600                 if (!md_write_start(mddev,bio))
1601                         return false;
1602                 raid1_write_request(mddev, bio, sectors);
1603         }
1604         return true;
1605 }
1606
1607 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1608 {
1609         struct r1conf *conf = mddev->private;
1610         int i;
1611
1612         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1613                    conf->raid_disks - mddev->degraded);
1614         rcu_read_lock();
1615         for (i = 0; i < conf->raid_disks; i++) {
1616                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1617                 seq_printf(seq, "%s",
1618                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1619         }
1620         rcu_read_unlock();
1621         seq_printf(seq, "]");
1622 }
1623
1624 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1625 {
1626         char b[BDEVNAME_SIZE];
1627         struct r1conf *conf = mddev->private;
1628         unsigned long flags;
1629
1630         /*
1631          * If it is not operational, then we have already marked it as dead
1632          * else if it is the last working disks with "fail_last_dev == false",
1633          * ignore the error, let the next level up know.
1634          * else mark the drive as failed
1635          */
1636         spin_lock_irqsave(&conf->device_lock, flags);
1637         if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1638             && (conf->raid_disks - mddev->degraded) == 1) {
1639                 /*
1640                  * Don't fail the drive, act as though we were just a
1641                  * normal single drive.
1642                  * However don't try a recovery from this drive as
1643                  * it is very likely to fail.
1644                  */
1645                 conf->recovery_disabled = mddev->recovery_disabled;
1646                 spin_unlock_irqrestore(&conf->device_lock, flags);
1647                 return;
1648         }
1649         set_bit(Blocked, &rdev->flags);
1650         if (test_and_clear_bit(In_sync, &rdev->flags))
1651                 mddev->degraded++;
1652         set_bit(Faulty, &rdev->flags);
1653         spin_unlock_irqrestore(&conf->device_lock, flags);
1654         /*
1655          * if recovery is running, make sure it aborts.
1656          */
1657         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1658         set_mask_bits(&mddev->sb_flags, 0,
1659                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1660         pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1661                 "md/raid1:%s: Operation continuing on %d devices.\n",
1662                 mdname(mddev), bdevname(rdev->bdev, b),
1663                 mdname(mddev), conf->raid_disks - mddev->degraded);
1664 }
1665
1666 static void print_conf(struct r1conf *conf)
1667 {
1668         int i;
1669
1670         pr_debug("RAID1 conf printout:\n");
1671         if (!conf) {
1672                 pr_debug("(!conf)\n");
1673                 return;
1674         }
1675         pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1676                  conf->raid_disks);
1677
1678         rcu_read_lock();
1679         for (i = 0; i < conf->raid_disks; i++) {
1680                 char b[BDEVNAME_SIZE];
1681                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1682                 if (rdev)
1683                         pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1684                                  i, !test_bit(In_sync, &rdev->flags),
1685                                  !test_bit(Faulty, &rdev->flags),
1686                                  bdevname(rdev->bdev,b));
1687         }
1688         rcu_read_unlock();
1689 }
1690
1691 static void close_sync(struct r1conf *conf)
1692 {
1693         int idx;
1694
1695         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1696                 _wait_barrier(conf, idx);
1697                 _allow_barrier(conf, idx);
1698         }
1699
1700         mempool_exit(&conf->r1buf_pool);
1701 }
1702
1703 static int raid1_spare_active(struct mddev *mddev)
1704 {
1705         int i;
1706         struct r1conf *conf = mddev->private;
1707         int count = 0;
1708         unsigned long flags;
1709
1710         /*
1711          * Find all failed disks within the RAID1 configuration
1712          * and mark them readable.
1713          * Called under mddev lock, so rcu protection not needed.
1714          * device_lock used to avoid races with raid1_end_read_request
1715          * which expects 'In_sync' flags and ->degraded to be consistent.
1716          */
1717         spin_lock_irqsave(&conf->device_lock, flags);
1718         for (i = 0; i < conf->raid_disks; i++) {
1719                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1720                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1721                 if (repl
1722                     && !test_bit(Candidate, &repl->flags)
1723                     && repl->recovery_offset == MaxSector
1724                     && !test_bit(Faulty, &repl->flags)
1725                     && !test_and_set_bit(In_sync, &repl->flags)) {
1726                         /* replacement has just become active */
1727                         if (!rdev ||
1728                             !test_and_clear_bit(In_sync, &rdev->flags))
1729                                 count++;
1730                         if (rdev) {
1731                                 /* Replaced device not technically
1732                                  * faulty, but we need to be sure
1733                                  * it gets removed and never re-added
1734                                  */
1735                                 set_bit(Faulty, &rdev->flags);
1736                                 sysfs_notify_dirent_safe(
1737                                         rdev->sysfs_state);
1738                         }
1739                 }
1740                 if (rdev
1741                     && rdev->recovery_offset == MaxSector
1742                     && !test_bit(Faulty, &rdev->flags)
1743                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1744                         count++;
1745                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1746                 }
1747         }
1748         mddev->degraded -= count;
1749         spin_unlock_irqrestore(&conf->device_lock, flags);
1750
1751         print_conf(conf);
1752         return count;
1753 }
1754
1755 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1756 {
1757         struct r1conf *conf = mddev->private;
1758         int err = -EEXIST;
1759         int mirror = 0;
1760         struct raid1_info *p;
1761         int first = 0;
1762         int last = conf->raid_disks - 1;
1763
1764         if (mddev->recovery_disabled == conf->recovery_disabled)
1765                 return -EBUSY;
1766
1767         if (md_integrity_add_rdev(rdev, mddev))
1768                 return -ENXIO;
1769
1770         if (rdev->raid_disk >= 0)
1771                 first = last = rdev->raid_disk;
1772
1773         /*
1774          * find the disk ... but prefer rdev->saved_raid_disk
1775          * if possible.
1776          */
1777         if (rdev->saved_raid_disk >= 0 &&
1778             rdev->saved_raid_disk >= first &&
1779             rdev->saved_raid_disk < conf->raid_disks &&
1780             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1781                 first = last = rdev->saved_raid_disk;
1782
1783         for (mirror = first; mirror <= last; mirror++) {
1784                 p = conf->mirrors + mirror;
1785                 if (!p->rdev) {
1786                         if (mddev->gendisk)
1787                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1788                                                   rdev->data_offset << 9);
1789
1790                         p->head_position = 0;
1791                         rdev->raid_disk = mirror;
1792                         err = 0;
1793                         /* As all devices are equivalent, we don't need a full recovery
1794                          * if this was recently any drive of the array
1795                          */
1796                         if (rdev->saved_raid_disk < 0)
1797                                 conf->fullsync = 1;
1798                         rcu_assign_pointer(p->rdev, rdev);
1799                         break;
1800                 }
1801                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1802                     p[conf->raid_disks].rdev == NULL) {
1803                         /* Add this device as a replacement */
1804                         clear_bit(In_sync, &rdev->flags);
1805                         set_bit(Replacement, &rdev->flags);
1806                         rdev->raid_disk = mirror;
1807                         err = 0;
1808                         conf->fullsync = 1;
1809                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1810                         break;
1811                 }
1812         }
1813         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1814                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1815         print_conf(conf);
1816         return err;
1817 }
1818
1819 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1820 {
1821         struct r1conf *conf = mddev->private;
1822         int err = 0;
1823         int number = rdev->raid_disk;
1824         struct raid1_info *p = conf->mirrors + number;
1825
1826         if (rdev != p->rdev)
1827                 p = conf->mirrors + conf->raid_disks + number;
1828
1829         print_conf(conf);
1830         if (rdev == p->rdev) {
1831                 if (test_bit(In_sync, &rdev->flags) ||
1832                     atomic_read(&rdev->nr_pending)) {
1833                         err = -EBUSY;
1834                         goto abort;
1835                 }
1836                 /* Only remove non-faulty devices if recovery
1837                  * is not possible.
1838                  */
1839                 if (!test_bit(Faulty, &rdev->flags) &&
1840                     mddev->recovery_disabled != conf->recovery_disabled &&
1841                     mddev->degraded < conf->raid_disks) {
1842                         err = -EBUSY;
1843                         goto abort;
1844                 }
1845                 p->rdev = NULL;
1846                 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1847                         synchronize_rcu();
1848                         if (atomic_read(&rdev->nr_pending)) {
1849                                 /* lost the race, try later */
1850                                 err = -EBUSY;
1851                                 p->rdev = rdev;
1852                                 goto abort;
1853                         }
1854                 }
1855                 if (conf->mirrors[conf->raid_disks + number].rdev) {
1856                         /* We just removed a device that is being replaced.
1857                          * Move down the replacement.  We drain all IO before
1858                          * doing this to avoid confusion.
1859                          */
1860                         struct md_rdev *repl =
1861                                 conf->mirrors[conf->raid_disks + number].rdev;
1862                         freeze_array(conf, 0);
1863                         if (atomic_read(&repl->nr_pending)) {
1864                                 /* It means that some queued IO of retry_list
1865                                  * hold repl. Thus, we cannot set replacement
1866                                  * as NULL, avoiding rdev NULL pointer
1867                                  * dereference in sync_request_write and
1868                                  * handle_write_finished.
1869                                  */
1870                                 err = -EBUSY;
1871                                 unfreeze_array(conf);
1872                                 goto abort;
1873                         }
1874                         clear_bit(Replacement, &repl->flags);
1875                         p->rdev = repl;
1876                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1877                         unfreeze_array(conf);
1878                 }
1879
1880                 clear_bit(WantReplacement, &rdev->flags);
1881                 err = md_integrity_register(mddev);
1882         }
1883 abort:
1884
1885         print_conf(conf);
1886         return err;
1887 }
1888
1889 static void end_sync_read(struct bio *bio)
1890 {
1891         struct r1bio *r1_bio = get_resync_r1bio(bio);
1892
1893         update_head_pos(r1_bio->read_disk, r1_bio);
1894
1895         /*
1896          * we have read a block, now it needs to be re-written,
1897          * or re-read if the read failed.
1898          * We don't do much here, just schedule handling by raid1d
1899          */
1900         if (!bio->bi_status)
1901                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1902
1903         if (atomic_dec_and_test(&r1_bio->remaining))
1904                 reschedule_retry(r1_bio);
1905 }
1906
1907 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1908 {
1909         sector_t sync_blocks = 0;
1910         sector_t s = r1_bio->sector;
1911         long sectors_to_go = r1_bio->sectors;
1912
1913         /* make sure these bits don't get cleared. */
1914         do {
1915                 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1916                 s += sync_blocks;
1917                 sectors_to_go -= sync_blocks;
1918         } while (sectors_to_go > 0);
1919 }
1920
1921 static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1922 {
1923         if (atomic_dec_and_test(&r1_bio->remaining)) {
1924                 struct mddev *mddev = r1_bio->mddev;
1925                 int s = r1_bio->sectors;
1926
1927                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1928                     test_bit(R1BIO_WriteError, &r1_bio->state))
1929                         reschedule_retry(r1_bio);
1930                 else {
1931                         put_buf(r1_bio);
1932                         md_done_sync(mddev, s, uptodate);
1933                 }
1934         }
1935 }
1936
1937 static void end_sync_write(struct bio *bio)
1938 {
1939         int uptodate = !bio->bi_status;
1940         struct r1bio *r1_bio = get_resync_r1bio(bio);
1941         struct mddev *mddev = r1_bio->mddev;
1942         struct r1conf *conf = mddev->private;
1943         sector_t first_bad;
1944         int bad_sectors;
1945         struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1946
1947         if (!uptodate) {
1948                 abort_sync_write(mddev, r1_bio);
1949                 set_bit(WriteErrorSeen, &rdev->flags);
1950                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1951                         set_bit(MD_RECOVERY_NEEDED, &
1952                                 mddev->recovery);
1953                 set_bit(R1BIO_WriteError, &r1_bio->state);
1954         } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1955                                &first_bad, &bad_sectors) &&
1956                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1957                                 r1_bio->sector,
1958                                 r1_bio->sectors,
1959                                 &first_bad, &bad_sectors)
1960                 )
1961                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1962
1963         put_sync_write_buf(r1_bio, uptodate);
1964 }
1965
1966 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1967                             int sectors, struct page *page, int rw)
1968 {
1969         if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1970                 /* success */
1971                 return 1;
1972         if (rw == WRITE) {
1973                 set_bit(WriteErrorSeen, &rdev->flags);
1974                 if (!test_and_set_bit(WantReplacement,
1975                                       &rdev->flags))
1976                         set_bit(MD_RECOVERY_NEEDED, &
1977                                 rdev->mddev->recovery);
1978         }
1979         /* need to record an error - either for the block or the device */
1980         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1981                 md_error(rdev->mddev, rdev);
1982         return 0;
1983 }
1984
1985 static int fix_sync_read_error(struct r1bio *r1_bio)
1986 {
1987         /* Try some synchronous reads of other devices to get
1988          * good data, much like with normal read errors.  Only
1989          * read into the pages we already have so we don't
1990          * need to re-issue the read request.
1991          * We don't need to freeze the array, because being in an
1992          * active sync request, there is no normal IO, and
1993          * no overlapping syncs.
1994          * We don't need to check is_badblock() again as we
1995          * made sure that anything with a bad block in range
1996          * will have bi_end_io clear.
1997          */
1998         struct mddev *mddev = r1_bio->mddev;
1999         struct r1conf *conf = mddev->private;
2000         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
2001         struct page **pages = get_resync_pages(bio)->pages;
2002         sector_t sect = r1_bio->sector;
2003         int sectors = r1_bio->sectors;
2004         int idx = 0;
2005         struct md_rdev *rdev;
2006
2007         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2008         if (test_bit(FailFast, &rdev->flags)) {
2009                 /* Don't try recovering from here - just fail it
2010                  * ... unless it is the last working device of course */
2011                 md_error(mddev, rdev);
2012                 if (test_bit(Faulty, &rdev->flags))
2013                         /* Don't try to read from here, but make sure
2014                          * put_buf does it's thing
2015                          */
2016                         bio->bi_end_io = end_sync_write;
2017         }
2018
2019         while(sectors) {
2020                 int s = sectors;
2021                 int d = r1_bio->read_disk;
2022                 int success = 0;
2023                 int start;
2024
2025                 if (s > (PAGE_SIZE>>9))
2026                         s = PAGE_SIZE >> 9;
2027                 do {
2028                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2029                                 /* No rcu protection needed here devices
2030                                  * can only be removed when no resync is
2031                                  * active, and resync is currently active
2032                                  */
2033                                 rdev = conf->mirrors[d].rdev;
2034                                 if (sync_page_io(rdev, sect, s<<9,
2035                                                  pages[idx],
2036                                                  REQ_OP_READ, 0, false)) {
2037                                         success = 1;
2038                                         break;
2039                                 }
2040                         }
2041                         d++;
2042                         if (d == conf->raid_disks * 2)
2043                                 d = 0;
2044                 } while (!success && d != r1_bio->read_disk);
2045
2046                 if (!success) {
2047                         char b[BDEVNAME_SIZE];
2048                         int abort = 0;
2049                         /* Cannot read from anywhere, this block is lost.
2050                          * Record a bad block on each device.  If that doesn't
2051                          * work just disable and interrupt the recovery.
2052                          * Don't fail devices as that won't really help.
2053                          */
2054                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2055                                             mdname(mddev), bio_devname(bio, b),
2056                                             (unsigned long long)r1_bio->sector);
2057                         for (d = 0; d < conf->raid_disks * 2; d++) {
2058                                 rdev = conf->mirrors[d].rdev;
2059                                 if (!rdev || test_bit(Faulty, &rdev->flags))
2060                                         continue;
2061                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
2062                                         abort = 1;
2063                         }
2064                         if (abort) {
2065                                 conf->recovery_disabled =
2066                                         mddev->recovery_disabled;
2067                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2068                                 md_done_sync(mddev, r1_bio->sectors, 0);
2069                                 put_buf(r1_bio);
2070                                 return 0;
2071                         }
2072                         /* Try next page */
2073                         sectors -= s;
2074                         sect += s;
2075                         idx++;
2076                         continue;
2077                 }
2078
2079                 start = d;
2080                 /* write it back and re-read */
2081                 while (d != r1_bio->read_disk) {
2082                         if (d == 0)
2083                                 d = conf->raid_disks * 2;
2084                         d--;
2085                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2086                                 continue;
2087                         rdev = conf->mirrors[d].rdev;
2088                         if (r1_sync_page_io(rdev, sect, s,
2089                                             pages[idx],
2090                                             WRITE) == 0) {
2091                                 r1_bio->bios[d]->bi_end_io = NULL;
2092                                 rdev_dec_pending(rdev, mddev);
2093                         }
2094                 }
2095                 d = start;
2096                 while (d != r1_bio->read_disk) {
2097                         if (d == 0)
2098                                 d = conf->raid_disks * 2;
2099                         d--;
2100                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2101                                 continue;
2102                         rdev = conf->mirrors[d].rdev;
2103                         if (r1_sync_page_io(rdev, sect, s,
2104                                             pages[idx],
2105                                             READ) != 0)
2106                                 atomic_add(s, &rdev->corrected_errors);
2107                 }
2108                 sectors -= s;
2109                 sect += s;
2110                 idx ++;
2111         }
2112         set_bit(R1BIO_Uptodate, &r1_bio->state);
2113         bio->bi_status = 0;
2114         return 1;
2115 }
2116
2117 static void process_checks(struct r1bio *r1_bio)
2118 {
2119         /* We have read all readable devices.  If we haven't
2120          * got the block, then there is no hope left.
2121          * If we have, then we want to do a comparison
2122          * and skip the write if everything is the same.
2123          * If any blocks failed to read, then we need to
2124          * attempt an over-write
2125          */
2126         struct mddev *mddev = r1_bio->mddev;
2127         struct r1conf *conf = mddev->private;
2128         int primary;
2129         int i;
2130         int vcnt;
2131
2132         /* Fix variable parts of all bios */
2133         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2134         for (i = 0; i < conf->raid_disks * 2; i++) {
2135                 blk_status_t status;
2136                 struct bio *b = r1_bio->bios[i];
2137                 struct resync_pages *rp = get_resync_pages(b);
2138                 if (b->bi_end_io != end_sync_read)
2139                         continue;
2140                 /* fixup the bio for reuse, but preserve errno */
2141                 status = b->bi_status;
2142                 bio_reset(b);
2143                 b->bi_status = status;
2144                 b->bi_iter.bi_sector = r1_bio->sector +
2145                         conf->mirrors[i].rdev->data_offset;
2146                 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2147                 b->bi_end_io = end_sync_read;
2148                 rp->raid_bio = r1_bio;
2149                 b->bi_private = rp;
2150
2151                 /* initialize bvec table again */
2152                 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2153         }
2154         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2155                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2156                     !r1_bio->bios[primary]->bi_status) {
2157                         r1_bio->bios[primary]->bi_end_io = NULL;
2158                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2159                         break;
2160                 }
2161         r1_bio->read_disk = primary;
2162         for (i = 0; i < conf->raid_disks * 2; i++) {
2163                 int j = 0;
2164                 struct bio *pbio = r1_bio->bios[primary];
2165                 struct bio *sbio = r1_bio->bios[i];
2166                 blk_status_t status = sbio->bi_status;
2167                 struct page **ppages = get_resync_pages(pbio)->pages;
2168                 struct page **spages = get_resync_pages(sbio)->pages;
2169                 struct bio_vec *bi;
2170                 int page_len[RESYNC_PAGES] = { 0 };
2171                 struct bvec_iter_all iter_all;
2172
2173                 if (sbio->bi_end_io != end_sync_read)
2174                         continue;
2175                 /* Now we can 'fixup' the error value */
2176                 sbio->bi_status = 0;
2177
2178                 bio_for_each_segment_all(bi, sbio, iter_all)
2179                         page_len[j++] = bi->bv_len;
2180
2181                 if (!status) {
2182                         for (j = vcnt; j-- ; ) {
2183                                 if (memcmp(page_address(ppages[j]),
2184                                            page_address(spages[j]),
2185                                            page_len[j]))
2186                                         break;
2187                         }
2188                 } else
2189                         j = 0;
2190                 if (j >= 0)
2191                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2192                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2193                               && !status)) {
2194                         /* No need to write to this device. */
2195                         sbio->bi_end_io = NULL;
2196                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2197                         continue;
2198                 }
2199
2200                 bio_copy_data(sbio, pbio);
2201         }
2202 }
2203
2204 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2205 {
2206         struct r1conf *conf = mddev->private;
2207         int i;
2208         int disks = conf->raid_disks * 2;
2209         struct bio *wbio;
2210
2211         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2212                 /* ouch - failed to read all of that. */
2213                 if (!fix_sync_read_error(r1_bio))
2214                         return;
2215
2216         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2217                 process_checks(r1_bio);
2218
2219         /*
2220          * schedule writes
2221          */
2222         atomic_set(&r1_bio->remaining, 1);
2223         for (i = 0; i < disks ; i++) {
2224                 wbio = r1_bio->bios[i];
2225                 if (wbio->bi_end_io == NULL ||
2226                     (wbio->bi_end_io == end_sync_read &&
2227                      (i == r1_bio->read_disk ||
2228                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2229                         continue;
2230                 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2231                         abort_sync_write(mddev, r1_bio);
2232                         continue;
2233                 }
2234
2235                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2236                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2237                         wbio->bi_opf |= MD_FAILFAST;
2238
2239                 wbio->bi_end_io = end_sync_write;
2240                 atomic_inc(&r1_bio->remaining);
2241                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2242
2243                 generic_make_request(wbio);
2244         }
2245
2246         put_sync_write_buf(r1_bio, 1);
2247 }
2248
2249 /*
2250  * This is a kernel thread which:
2251  *
2252  *      1.      Retries failed read operations on working mirrors.
2253  *      2.      Updates the raid superblock when problems encounter.
2254  *      3.      Performs writes following reads for array synchronising.
2255  */
2256
2257 static void fix_read_error(struct r1conf *conf, int read_disk,
2258                            sector_t sect, int sectors)
2259 {
2260         struct mddev *mddev = conf->mddev;
2261         while(sectors) {
2262                 int s = sectors;
2263                 int d = read_disk;
2264                 int success = 0;
2265                 int start;
2266                 struct md_rdev *rdev;
2267
2268                 if (s > (PAGE_SIZE>>9))
2269                         s = PAGE_SIZE >> 9;
2270
2271                 do {
2272                         sector_t first_bad;
2273                         int bad_sectors;
2274
2275                         rcu_read_lock();
2276                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2277                         if (rdev &&
2278                             (test_bit(In_sync, &rdev->flags) ||
2279                              (!test_bit(Faulty, &rdev->flags) &&
2280                               rdev->recovery_offset >= sect + s)) &&
2281                             is_badblock(rdev, sect, s,
2282                                         &first_bad, &bad_sectors) == 0) {
2283                                 atomic_inc(&rdev->nr_pending);
2284                                 rcu_read_unlock();
2285                                 if (sync_page_io(rdev, sect, s<<9,
2286                                          conf->tmppage, REQ_OP_READ, 0, false))
2287                                         success = 1;
2288                                 rdev_dec_pending(rdev, mddev);
2289                                 if (success)
2290                                         break;
2291                         } else
2292                                 rcu_read_unlock();
2293                         d++;
2294                         if (d == conf->raid_disks * 2)
2295                                 d = 0;
2296                 } while (!success && d != read_disk);
2297
2298                 if (!success) {
2299                         /* Cannot read from anywhere - mark it bad */
2300                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2301                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2302                                 md_error(mddev, rdev);
2303                         break;
2304                 }
2305                 /* write it back and re-read */
2306                 start = d;
2307                 while (d != read_disk) {
2308                         if (d==0)
2309                                 d = conf->raid_disks * 2;
2310                         d--;
2311                         rcu_read_lock();
2312                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2313                         if (rdev &&
2314                             !test_bit(Faulty, &rdev->flags)) {
2315                                 atomic_inc(&rdev->nr_pending);
2316                                 rcu_read_unlock();
2317                                 r1_sync_page_io(rdev, sect, s,
2318                                                 conf->tmppage, WRITE);
2319                                 rdev_dec_pending(rdev, mddev);
2320                         } else
2321                                 rcu_read_unlock();
2322                 }
2323                 d = start;
2324                 while (d != read_disk) {
2325                         char b[BDEVNAME_SIZE];
2326                         if (d==0)
2327                                 d = conf->raid_disks * 2;
2328                         d--;
2329                         rcu_read_lock();
2330                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2331                         if (rdev &&
2332                             !test_bit(Faulty, &rdev->flags)) {
2333                                 atomic_inc(&rdev->nr_pending);
2334                                 rcu_read_unlock();
2335                                 if (r1_sync_page_io(rdev, sect, s,
2336                                                     conf->tmppage, READ)) {
2337                                         atomic_add(s, &rdev->corrected_errors);
2338                                         pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2339                                                 mdname(mddev), s,
2340                                                 (unsigned long long)(sect +
2341                                                                      rdev->data_offset),
2342                                                 bdevname(rdev->bdev, b));
2343                                 }
2344                                 rdev_dec_pending(rdev, mddev);
2345                         } else
2346                                 rcu_read_unlock();
2347                 }
2348                 sectors -= s;
2349                 sect += s;
2350         }
2351 }
2352
2353 static int narrow_write_error(struct r1bio *r1_bio, int i)
2354 {
2355         struct mddev *mddev = r1_bio->mddev;
2356         struct r1conf *conf = mddev->private;
2357         struct md_rdev *rdev = conf->mirrors[i].rdev;
2358
2359         /* bio has the data to be written to device 'i' where
2360          * we just recently had a write error.
2361          * We repeatedly clone the bio and trim down to one block,
2362          * then try the write.  Where the write fails we record
2363          * a bad block.
2364          * It is conceivable that the bio doesn't exactly align with
2365          * blocks.  We must handle this somehow.
2366          *
2367          * We currently own a reference on the rdev.
2368          */
2369
2370         int block_sectors;
2371         sector_t sector;
2372         int sectors;
2373         int sect_to_write = r1_bio->sectors;
2374         int ok = 1;
2375
2376         if (rdev->badblocks.shift < 0)
2377                 return 0;
2378
2379         block_sectors = roundup(1 << rdev->badblocks.shift,
2380                                 bdev_logical_block_size(rdev->bdev) >> 9);
2381         sector = r1_bio->sector;
2382         sectors = ((sector + block_sectors)
2383                    & ~(sector_t)(block_sectors - 1))
2384                 - sector;
2385
2386         while (sect_to_write) {
2387                 struct bio *wbio;
2388                 if (sectors > sect_to_write)
2389                         sectors = sect_to_write;
2390                 /* Write at 'sector' for 'sectors'*/
2391
2392                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2393                         wbio = bio_clone_fast(r1_bio->behind_master_bio,
2394                                               GFP_NOIO,
2395                                               &mddev->bio_set);
2396                 } else {
2397                         wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2398                                               &mddev->bio_set);
2399                 }
2400
2401                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2402                 wbio->bi_iter.bi_sector = r1_bio->sector;
2403                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2404
2405                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2406                 wbio->bi_iter.bi_sector += rdev->data_offset;
2407                 bio_set_dev(wbio, rdev->bdev);
2408
2409                 if (submit_bio_wait(wbio) < 0)
2410                         /* failure! */
2411                         ok = rdev_set_badblocks(rdev, sector,
2412                                                 sectors, 0)
2413                                 && ok;
2414
2415                 bio_put(wbio);
2416                 sect_to_write -= sectors;
2417                 sector += sectors;
2418                 sectors = block_sectors;
2419         }
2420         return ok;
2421 }
2422
2423 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2424 {
2425         int m;
2426         int s = r1_bio->sectors;
2427         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2428                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2429                 struct bio *bio = r1_bio->bios[m];
2430                 if (bio->bi_end_io == NULL)
2431                         continue;
2432                 if (!bio->bi_status &&
2433                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2434                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2435                 }
2436                 if (bio->bi_status &&
2437                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2438                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2439                                 md_error(conf->mddev, rdev);
2440                 }
2441         }
2442         put_buf(r1_bio);
2443         md_done_sync(conf->mddev, s, 1);
2444 }
2445
2446 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2447 {
2448         int m, idx;
2449         bool fail = false;
2450
2451         for (m = 0; m < conf->raid_disks * 2 ; m++)
2452                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2453                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2454                         rdev_clear_badblocks(rdev,
2455                                              r1_bio->sector,
2456                                              r1_bio->sectors, 0);
2457                         rdev_dec_pending(rdev, conf->mddev);
2458                 } else if (r1_bio->bios[m] != NULL) {
2459                         /* This drive got a write error.  We need to
2460                          * narrow down and record precise write
2461                          * errors.
2462                          */
2463                         fail = true;
2464                         if (!narrow_write_error(r1_bio, m)) {
2465                                 md_error(conf->mddev,
2466                                          conf->mirrors[m].rdev);
2467                                 /* an I/O failed, we can't clear the bitmap */
2468                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2469                         }
2470                         rdev_dec_pending(conf->mirrors[m].rdev,
2471                                          conf->mddev);
2472                 }
2473         if (fail) {
2474                 spin_lock_irq(&conf->device_lock);
2475                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2476                 idx = sector_to_idx(r1_bio->sector);
2477                 atomic_inc(&conf->nr_queued[idx]);
2478                 spin_unlock_irq(&conf->device_lock);
2479                 /*
2480                  * In case freeze_array() is waiting for condition
2481                  * get_unqueued_pending() == extra to be true.
2482                  */
2483                 wake_up(&conf->wait_barrier);
2484                 md_wakeup_thread(conf->mddev->thread);
2485         } else {
2486                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2487                         close_write(r1_bio);
2488                 raid_end_bio_io(r1_bio);
2489         }
2490 }
2491
2492 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2493 {
2494         struct mddev *mddev = conf->mddev;
2495         struct bio *bio;
2496         struct md_rdev *rdev;
2497
2498         clear_bit(R1BIO_ReadError, &r1_bio->state);
2499         /* we got a read error. Maybe the drive is bad.  Maybe just
2500          * the block and we can fix it.
2501          * We freeze all other IO, and try reading the block from
2502          * other devices.  When we find one, we re-write
2503          * and check it that fixes the read error.
2504          * This is all done synchronously while the array is
2505          * frozen
2506          */
2507
2508         bio = r1_bio->bios[r1_bio->read_disk];
2509         bio_put(bio);
2510         r1_bio->bios[r1_bio->read_disk] = NULL;
2511
2512         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2513         if (mddev->ro == 0
2514             && !test_bit(FailFast, &rdev->flags)) {
2515                 freeze_array(conf, 1);
2516                 fix_read_error(conf, r1_bio->read_disk,
2517                                r1_bio->sector, r1_bio->sectors);
2518                 unfreeze_array(conf);
2519         } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2520                 md_error(mddev, rdev);
2521         } else {
2522                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2523         }
2524
2525         rdev_dec_pending(rdev, conf->mddev);
2526         allow_barrier(conf, r1_bio->sector);
2527         bio = r1_bio->master_bio;
2528
2529         /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2530         r1_bio->state = 0;
2531         raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2532 }
2533
2534 static void raid1d(struct md_thread *thread)
2535 {
2536         struct mddev *mddev = thread->mddev;
2537         struct r1bio *r1_bio;
2538         unsigned long flags;
2539         struct r1conf *conf = mddev->private;
2540         struct list_head *head = &conf->retry_list;
2541         struct blk_plug plug;
2542         int idx;
2543
2544         md_check_recovery(mddev);
2545
2546         if (!list_empty_careful(&conf->bio_end_io_list) &&
2547             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2548                 LIST_HEAD(tmp);
2549                 spin_lock_irqsave(&conf->device_lock, flags);
2550                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2551                         list_splice_init(&conf->bio_end_io_list, &tmp);
2552                 spin_unlock_irqrestore(&conf->device_lock, flags);
2553                 while (!list_empty(&tmp)) {
2554                         r1_bio = list_first_entry(&tmp, struct r1bio,
2555                                                   retry_list);
2556                         list_del(&r1_bio->retry_list);
2557                         idx = sector_to_idx(r1_bio->sector);
2558                         atomic_dec(&conf->nr_queued[idx]);
2559                         if (mddev->degraded)
2560                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2561                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2562                                 close_write(r1_bio);
2563                         raid_end_bio_io(r1_bio);
2564                 }
2565         }
2566
2567         blk_start_plug(&plug);
2568         for (;;) {
2569
2570                 flush_pending_writes(conf);
2571
2572                 spin_lock_irqsave(&conf->device_lock, flags);
2573                 if (list_empty(head)) {
2574                         spin_unlock_irqrestore(&conf->device_lock, flags);
2575                         break;
2576                 }
2577                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2578                 list_del(head->prev);
2579                 idx = sector_to_idx(r1_bio->sector);
2580                 atomic_dec(&conf->nr_queued[idx]);
2581                 spin_unlock_irqrestore(&conf->device_lock, flags);
2582
2583                 mddev = r1_bio->mddev;
2584                 conf = mddev->private;
2585                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2586                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2587                             test_bit(R1BIO_WriteError, &r1_bio->state))
2588                                 handle_sync_write_finished(conf, r1_bio);
2589                         else
2590                                 sync_request_write(mddev, r1_bio);
2591                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2592                            test_bit(R1BIO_WriteError, &r1_bio->state))
2593                         handle_write_finished(conf, r1_bio);
2594                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2595                         handle_read_error(conf, r1_bio);
2596                 else
2597                         WARN_ON_ONCE(1);
2598
2599                 cond_resched();
2600                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2601                         md_check_recovery(mddev);
2602         }
2603         blk_finish_plug(&plug);
2604 }
2605
2606 static int init_resync(struct r1conf *conf)
2607 {
2608         int buffs;
2609
2610         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2611         BUG_ON(mempool_initialized(&conf->r1buf_pool));
2612
2613         return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2614                             r1buf_pool_free, conf->poolinfo);
2615 }
2616
2617 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2618 {
2619         struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2620         struct resync_pages *rps;
2621         struct bio *bio;
2622         int i;
2623
2624         for (i = conf->poolinfo->raid_disks; i--; ) {
2625                 bio = r1bio->bios[i];
2626                 rps = bio->bi_private;
2627                 bio_reset(bio);
2628                 bio->bi_private = rps;
2629         }
2630         r1bio->master_bio = NULL;
2631         return r1bio;
2632 }
2633
2634 /*
2635  * perform a "sync" on one "block"
2636  *
2637  * We need to make sure that no normal I/O request - particularly write
2638  * requests - conflict with active sync requests.
2639  *
2640  * This is achieved by tracking pending requests and a 'barrier' concept
2641  * that can be installed to exclude normal IO requests.
2642  */
2643
2644 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2645                                    int *skipped)
2646 {
2647         struct r1conf *conf = mddev->private;
2648         struct r1bio *r1_bio;
2649         struct bio *bio;
2650         sector_t max_sector, nr_sectors;
2651         int disk = -1;
2652         int i;
2653         int wonly = -1;
2654         int write_targets = 0, read_targets = 0;
2655         sector_t sync_blocks;
2656         int still_degraded = 0;
2657         int good_sectors = RESYNC_SECTORS;
2658         int min_bad = 0; /* number of sectors that are bad in all devices */
2659         int idx = sector_to_idx(sector_nr);
2660         int page_idx = 0;
2661
2662         if (!mempool_initialized(&conf->r1buf_pool))
2663                 if (init_resync(conf))
2664                         return 0;
2665
2666         max_sector = mddev->dev_sectors;
2667         if (sector_nr >= max_sector) {
2668                 /* If we aborted, we need to abort the
2669                  * sync on the 'current' bitmap chunk (there will
2670                  * only be one in raid1 resync.
2671                  * We can find the current addess in mddev->curr_resync
2672                  */
2673                 if (mddev->curr_resync < max_sector) /* aborted */
2674                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2675                                            &sync_blocks, 1);
2676                 else /* completed sync */
2677                         conf->fullsync = 0;
2678
2679                 md_bitmap_close_sync(mddev->bitmap);
2680                 close_sync(conf);
2681
2682                 if (mddev_is_clustered(mddev)) {
2683                         conf->cluster_sync_low = 0;
2684                         conf->cluster_sync_high = 0;
2685                 }
2686                 return 0;
2687         }
2688
2689         if (mddev->bitmap == NULL &&
2690             mddev->recovery_cp == MaxSector &&
2691             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2692             conf->fullsync == 0) {
2693                 *skipped = 1;
2694                 return max_sector - sector_nr;
2695         }
2696         /* before building a request, check if we can skip these blocks..
2697          * This call the bitmap_start_sync doesn't actually record anything
2698          */
2699         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2700             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2701                 /* We can skip this block, and probably several more */
2702                 *skipped = 1;
2703                 return sync_blocks;
2704         }
2705
2706         /*
2707          * If there is non-resync activity waiting for a turn, then let it
2708          * though before starting on this new sync request.
2709          */
2710         if (atomic_read(&conf->nr_waiting[idx]))
2711                 schedule_timeout_uninterruptible(1);
2712
2713         /* we are incrementing sector_nr below. To be safe, we check against
2714          * sector_nr + two times RESYNC_SECTORS
2715          */
2716
2717         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2718                 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2719
2720
2721         if (raise_barrier(conf, sector_nr))
2722                 return 0;
2723
2724         r1_bio = raid1_alloc_init_r1buf(conf);
2725
2726         rcu_read_lock();
2727         /*
2728          * If we get a correctably read error during resync or recovery,
2729          * we might want to read from a different device.  So we
2730          * flag all drives that could conceivably be read from for READ,
2731          * and any others (which will be non-In_sync devices) for WRITE.
2732          * If a read fails, we try reading from something else for which READ
2733          * is OK.
2734          */
2735
2736         r1_bio->mddev = mddev;
2737         r1_bio->sector = sector_nr;
2738         r1_bio->state = 0;
2739         set_bit(R1BIO_IsSync, &r1_bio->state);
2740         /* make sure good_sectors won't go across barrier unit boundary */
2741         good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2742
2743         for (i = 0; i < conf->raid_disks * 2; i++) {
2744                 struct md_rdev *rdev;
2745                 bio = r1_bio->bios[i];
2746
2747                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2748                 if (rdev == NULL ||
2749                     test_bit(Faulty, &rdev->flags)) {
2750                         if (i < conf->raid_disks)
2751                                 still_degraded = 1;
2752                 } else if (!test_bit(In_sync, &rdev->flags)) {
2753                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2754                         bio->bi_end_io = end_sync_write;
2755                         write_targets ++;
2756                 } else {
2757                         /* may need to read from here */
2758                         sector_t first_bad = MaxSector;
2759                         int bad_sectors;
2760
2761                         if (is_badblock(rdev, sector_nr, good_sectors,
2762                                         &first_bad, &bad_sectors)) {
2763                                 if (first_bad > sector_nr)
2764                                         good_sectors = first_bad - sector_nr;
2765                                 else {
2766                                         bad_sectors -= (sector_nr - first_bad);
2767                                         if (min_bad == 0 ||
2768                                             min_bad > bad_sectors)
2769                                                 min_bad = bad_sectors;
2770                                 }
2771                         }
2772                         if (sector_nr < first_bad) {
2773                                 if (test_bit(WriteMostly, &rdev->flags)) {
2774                                         if (wonly < 0)
2775                                                 wonly = i;
2776                                 } else {
2777                                         if (disk < 0)
2778                                                 disk = i;
2779                                 }
2780                                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2781                                 bio->bi_end_io = end_sync_read;
2782                                 read_targets++;
2783                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2784                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2785                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2786                                 /*
2787                                  * The device is suitable for reading (InSync),
2788                                  * but has bad block(s) here. Let's try to correct them,
2789                                  * if we are doing resync or repair. Otherwise, leave
2790                                  * this device alone for this sync request.
2791                                  */
2792                                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2793                                 bio->bi_end_io = end_sync_write;
2794                                 write_targets++;
2795                         }
2796                 }
2797                 if (rdev && bio->bi_end_io) {
2798                         atomic_inc(&rdev->nr_pending);
2799                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2800                         bio_set_dev(bio, rdev->bdev);
2801                         if (test_bit(FailFast, &rdev->flags))
2802                                 bio->bi_opf |= MD_FAILFAST;
2803                 }
2804         }
2805         rcu_read_unlock();
2806         if (disk < 0)
2807                 disk = wonly;
2808         r1_bio->read_disk = disk;
2809
2810         if (read_targets == 0 && min_bad > 0) {
2811                 /* These sectors are bad on all InSync devices, so we
2812                  * need to mark them bad on all write targets
2813                  */
2814                 int ok = 1;
2815                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2816                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2817                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2818                                 ok = rdev_set_badblocks(rdev, sector_nr,
2819                                                         min_bad, 0
2820                                         ) && ok;
2821                         }
2822                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2823                 *skipped = 1;
2824                 put_buf(r1_bio);
2825
2826                 if (!ok) {
2827                         /* Cannot record the badblocks, so need to
2828                          * abort the resync.
2829                          * If there are multiple read targets, could just
2830                          * fail the really bad ones ???
2831                          */
2832                         conf->recovery_disabled = mddev->recovery_disabled;
2833                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2834                         return 0;
2835                 } else
2836                         return min_bad;
2837
2838         }
2839         if (min_bad > 0 && min_bad < good_sectors) {
2840                 /* only resync enough to reach the next bad->good
2841                  * transition */
2842                 good_sectors = min_bad;
2843         }
2844
2845         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2846                 /* extra read targets are also write targets */
2847                 write_targets += read_targets-1;
2848
2849         if (write_targets == 0 || read_targets == 0) {
2850                 /* There is nowhere to write, so all non-sync
2851                  * drives must be failed - so we are finished
2852                  */
2853                 sector_t rv;
2854                 if (min_bad > 0)
2855                         max_sector = sector_nr + min_bad;
2856                 rv = max_sector - sector_nr;
2857                 *skipped = 1;
2858                 put_buf(r1_bio);
2859                 return rv;
2860         }
2861
2862         if (max_sector > mddev->resync_max)
2863                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2864         if (max_sector > sector_nr + good_sectors)
2865                 max_sector = sector_nr + good_sectors;
2866         nr_sectors = 0;
2867         sync_blocks = 0;
2868         do {
2869                 struct page *page;
2870                 int len = PAGE_SIZE;
2871                 if (sector_nr + (len>>9) > max_sector)
2872                         len = (max_sector - sector_nr) << 9;
2873                 if (len == 0)
2874                         break;
2875                 if (sync_blocks == 0) {
2876                         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2877                                                   &sync_blocks, still_degraded) &&
2878                             !conf->fullsync &&
2879                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2880                                 break;
2881                         if ((len >> 9) > sync_blocks)
2882                                 len = sync_blocks<<9;
2883                 }
2884
2885                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2886                         struct resync_pages *rp;
2887
2888                         bio = r1_bio->bios[i];
2889                         rp = get_resync_pages(bio);
2890                         if (bio->bi_end_io) {
2891                                 page = resync_fetch_page(rp, page_idx);
2892
2893                                 /*
2894                                  * won't fail because the vec table is big
2895                                  * enough to hold all these pages
2896                                  */
2897                                 bio_add_page(bio, page, len, 0);
2898                         }
2899                 }
2900                 nr_sectors += len>>9;
2901                 sector_nr += len>>9;
2902                 sync_blocks -= (len>>9);
2903         } while (++page_idx < RESYNC_PAGES);
2904
2905         r1_bio->sectors = nr_sectors;
2906
2907         if (mddev_is_clustered(mddev) &&
2908                         conf->cluster_sync_high < sector_nr + nr_sectors) {
2909                 conf->cluster_sync_low = mddev->curr_resync_completed;
2910                 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2911                 /* Send resync message */
2912                 md_cluster_ops->resync_info_update(mddev,
2913                                 conf->cluster_sync_low,
2914                                 conf->cluster_sync_high);
2915         }
2916
2917         /* For a user-requested sync, we read all readable devices and do a
2918          * compare
2919          */
2920         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2921                 atomic_set(&r1_bio->remaining, read_targets);
2922                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2923                         bio = r1_bio->bios[i];
2924                         if (bio->bi_end_io == end_sync_read) {
2925                                 read_targets--;
2926                                 md_sync_acct_bio(bio, nr_sectors);
2927                                 if (read_targets == 1)
2928                                         bio->bi_opf &= ~MD_FAILFAST;
2929                                 generic_make_request(bio);
2930                         }
2931                 }
2932         } else {
2933                 atomic_set(&r1_bio->remaining, 1);
2934                 bio = r1_bio->bios[r1_bio->read_disk];
2935                 md_sync_acct_bio(bio, nr_sectors);
2936                 if (read_targets == 1)
2937                         bio->bi_opf &= ~MD_FAILFAST;
2938                 generic_make_request(bio);
2939         }
2940         return nr_sectors;
2941 }
2942
2943 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2944 {
2945         if (sectors)
2946                 return sectors;
2947
2948         return mddev->dev_sectors;
2949 }
2950
2951 static struct r1conf *setup_conf(struct mddev *mddev)
2952 {
2953         struct r1conf *conf;
2954         int i;
2955         struct raid1_info *disk;
2956         struct md_rdev *rdev;
2957         int err = -ENOMEM;
2958
2959         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2960         if (!conf)
2961                 goto abort;
2962
2963         conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2964                                    sizeof(atomic_t), GFP_KERNEL);
2965         if (!conf->nr_pending)
2966                 goto abort;
2967
2968         conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2969                                    sizeof(atomic_t), GFP_KERNEL);
2970         if (!conf->nr_waiting)
2971                 goto abort;
2972
2973         conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2974                                   sizeof(atomic_t), GFP_KERNEL);
2975         if (!conf->nr_queued)
2976                 goto abort;
2977
2978         conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2979                                 sizeof(atomic_t), GFP_KERNEL);
2980         if (!conf->barrier)
2981                 goto abort;
2982
2983         conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2984                                             mddev->raid_disks, 2),
2985                                 GFP_KERNEL);
2986         if (!conf->mirrors)
2987                 goto abort;
2988
2989         conf->tmppage = alloc_page(GFP_KERNEL);
2990         if (!conf->tmppage)
2991                 goto abort;
2992
2993         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2994         if (!conf->poolinfo)
2995                 goto abort;
2996         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2997         err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2998                            rbio_pool_free, conf->poolinfo);
2999         if (err)
3000                 goto abort;
3001
3002         err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3003         if (err)
3004                 goto abort;
3005
3006         conf->poolinfo->mddev = mddev;
3007
3008         err = -EINVAL;
3009         spin_lock_init(&conf->device_lock);
3010         rdev_for_each(rdev, mddev) {
3011                 int disk_idx = rdev->raid_disk;
3012                 if (disk_idx >= mddev->raid_disks
3013                     || disk_idx < 0)
3014                         continue;
3015                 if (test_bit(Replacement, &rdev->flags))
3016                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
3017                 else
3018                         disk = conf->mirrors + disk_idx;
3019
3020                 if (disk->rdev)
3021                         goto abort;
3022                 disk->rdev = rdev;
3023                 disk->head_position = 0;
3024                 disk->seq_start = MaxSector;
3025         }
3026         conf->raid_disks = mddev->raid_disks;
3027         conf->mddev = mddev;
3028         INIT_LIST_HEAD(&conf->retry_list);
3029         INIT_LIST_HEAD(&conf->bio_end_io_list);
3030
3031         spin_lock_init(&conf->resync_lock);
3032         init_waitqueue_head(&conf->wait_barrier);
3033
3034         bio_list_init(&conf->pending_bio_list);
3035         conf->pending_count = 0;
3036         conf->recovery_disabled = mddev->recovery_disabled - 1;
3037
3038         err = -EIO;
3039         for (i = 0; i < conf->raid_disks * 2; i++) {
3040
3041                 disk = conf->mirrors + i;
3042
3043                 if (i < conf->raid_disks &&
3044                     disk[conf->raid_disks].rdev) {
3045                         /* This slot has a replacement. */
3046                         if (!disk->rdev) {
3047                                 /* No original, just make the replacement
3048                                  * a recovering spare
3049                                  */
3050                                 disk->rdev =
3051                                         disk[conf->raid_disks].rdev;
3052                                 disk[conf->raid_disks].rdev = NULL;
3053                         } else if (!test_bit(In_sync, &disk->rdev->flags))
3054                                 /* Original is not in_sync - bad */
3055                                 goto abort;
3056                 }
3057
3058                 if (!disk->rdev ||
3059                     !test_bit(In_sync, &disk->rdev->flags)) {
3060                         disk->head_position = 0;
3061                         if (disk->rdev &&
3062                             (disk->rdev->saved_raid_disk < 0))
3063                                 conf->fullsync = 1;
3064                 }
3065         }
3066
3067         err = -ENOMEM;
3068         conf->thread = md_register_thread(raid1d, mddev, "raid1");
3069         if (!conf->thread)
3070                 goto abort;
3071
3072         return conf;
3073
3074  abort:
3075         if (conf) {
3076                 mempool_exit(&conf->r1bio_pool);
3077                 kfree(conf->mirrors);
3078                 safe_put_page(conf->tmppage);
3079                 kfree(conf->poolinfo);
3080                 kfree(conf->nr_pending);
3081                 kfree(conf->nr_waiting);
3082                 kfree(conf->nr_queued);
3083                 kfree(conf->barrier);
3084                 bioset_exit(&conf->bio_split);
3085                 kfree(conf);
3086         }
3087         return ERR_PTR(err);
3088 }
3089
3090 static void raid1_free(struct mddev *mddev, void *priv);
3091 static int raid1_run(struct mddev *mddev)
3092 {
3093         struct r1conf *conf;
3094         int i;
3095         struct md_rdev *rdev;
3096         int ret;
3097         bool discard_supported = false;
3098
3099         if (mddev->level != 1) {
3100                 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3101                         mdname(mddev), mddev->level);
3102                 return -EIO;
3103         }
3104         if (mddev->reshape_position != MaxSector) {
3105                 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3106                         mdname(mddev));
3107                 return -EIO;
3108         }
3109         if (mddev_init_writes_pending(mddev) < 0)
3110                 return -ENOMEM;
3111         /*
3112          * copy the already verified devices into our private RAID1
3113          * bookkeeping area. [whatever we allocate in run(),
3114          * should be freed in raid1_free()]
3115          */
3116         if (mddev->private == NULL)
3117                 conf = setup_conf(mddev);
3118         else
3119                 conf = mddev->private;
3120
3121         if (IS_ERR(conf))
3122                 return PTR_ERR(conf);
3123
3124         if (mddev->queue) {
3125                 blk_queue_max_write_same_sectors(mddev->queue, 0);
3126                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3127         }
3128
3129         rdev_for_each(rdev, mddev) {
3130                 if (!mddev->gendisk)
3131                         continue;
3132                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3133                                   rdev->data_offset << 9);
3134                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3135                         discard_supported = true;
3136         }
3137
3138         mddev->degraded = 0;
3139         for (i = 0; i < conf->raid_disks; i++)
3140                 if (conf->mirrors[i].rdev == NULL ||
3141                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3142                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3143                         mddev->degraded++;
3144         /*
3145          * RAID1 needs at least one disk in active
3146          */
3147         if (conf->raid_disks - mddev->degraded < 1) {
3148                 ret = -EINVAL;
3149                 goto abort;
3150         }
3151
3152         if (conf->raid_disks - mddev->degraded == 1)
3153                 mddev->recovery_cp = MaxSector;
3154
3155         if (mddev->recovery_cp != MaxSector)
3156                 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3157                         mdname(mddev));
3158         pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3159                 mdname(mddev), mddev->raid_disks - mddev->degraded,
3160                 mddev->raid_disks);
3161
3162         /*
3163          * Ok, everything is just fine now
3164          */
3165         mddev->thread = conf->thread;
3166         conf->thread = NULL;
3167         mddev->private = conf;
3168         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3169
3170         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3171
3172         if (mddev->queue) {
3173                 if (discard_supported)
3174                         blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3175                                                 mddev->queue);
3176                 else
3177                         blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3178                                                   mddev->queue);
3179         }
3180
3181         ret = md_integrity_register(mddev);
3182         if (ret) {
3183                 md_unregister_thread(&mddev->thread);
3184                 goto abort;
3185         }
3186         return 0;
3187
3188 abort:
3189         raid1_free(mddev, conf);
3190         return ret;
3191 }
3192
3193 static void raid1_free(struct mddev *mddev, void *priv)
3194 {
3195         struct r1conf *conf = priv;
3196
3197         mempool_exit(&conf->r1bio_pool);
3198         kfree(conf->mirrors);
3199         safe_put_page(conf->tmppage);
3200         kfree(conf->poolinfo);
3201         kfree(conf->nr_pending);
3202         kfree(conf->nr_waiting);
3203         kfree(conf->nr_queued);
3204         kfree(conf->barrier);
3205         bioset_exit(&conf->bio_split);
3206         kfree(conf);
3207 }
3208
3209 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3210 {
3211         /* no resync is happening, and there is enough space
3212          * on all devices, so we can resize.
3213          * We need to make sure resync covers any new space.
3214          * If the array is shrinking we should possibly wait until
3215          * any io in the removed space completes, but it hardly seems
3216          * worth it.
3217          */
3218         sector_t newsize = raid1_size(mddev, sectors, 0);
3219         if (mddev->external_size &&
3220             mddev->array_sectors > newsize)
3221                 return -EINVAL;
3222         if (mddev->bitmap) {
3223                 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3224                 if (ret)
3225                         return ret;
3226         }
3227         md_set_array_sectors(mddev, newsize);
3228         if (sectors > mddev->dev_sectors &&
3229             mddev->recovery_cp > mddev->dev_sectors) {
3230                 mddev->recovery_cp = mddev->dev_sectors;
3231                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3232         }
3233         mddev->dev_sectors = sectors;
3234         mddev->resync_max_sectors = sectors;
3235         return 0;
3236 }
3237
3238 static int raid1_reshape(struct mddev *mddev)
3239 {
3240         /* We need to:
3241          * 1/ resize the r1bio_pool
3242          * 2/ resize conf->mirrors
3243          *
3244          * We allocate a new r1bio_pool if we can.
3245          * Then raise a device barrier and wait until all IO stops.
3246          * Then resize conf->mirrors and swap in the new r1bio pool.
3247          *
3248          * At the same time, we "pack" the devices so that all the missing
3249          * devices have the higher raid_disk numbers.
3250          */
3251         mempool_t newpool, oldpool;
3252         struct pool_info *newpoolinfo;
3253         struct raid1_info *newmirrors;
3254         struct r1conf *conf = mddev->private;
3255         int cnt, raid_disks;
3256         unsigned long flags;
3257         int d, d2;
3258         int ret;
3259
3260         memset(&newpool, 0, sizeof(newpool));
3261         memset(&oldpool, 0, sizeof(oldpool));
3262
3263         /* Cannot change chunk_size, layout, or level */
3264         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3265             mddev->layout != mddev->new_layout ||
3266             mddev->level != mddev->new_level) {
3267                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3268                 mddev->new_layout = mddev->layout;
3269                 mddev->new_level = mddev->level;
3270                 return -EINVAL;
3271         }
3272
3273         if (!mddev_is_clustered(mddev))
3274                 md_allow_write(mddev);
3275
3276         raid_disks = mddev->raid_disks + mddev->delta_disks;
3277
3278         if (raid_disks < conf->raid_disks) {
3279                 cnt=0;
3280                 for (d= 0; d < conf->raid_disks; d++)
3281                         if (conf->mirrors[d].rdev)
3282                                 cnt++;
3283                 if (cnt > raid_disks)
3284                         return -EBUSY;
3285         }
3286
3287         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3288         if (!newpoolinfo)
3289                 return -ENOMEM;
3290         newpoolinfo->mddev = mddev;
3291         newpoolinfo->raid_disks = raid_disks * 2;
3292
3293         ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3294                            rbio_pool_free, newpoolinfo);
3295         if (ret) {
3296                 kfree(newpoolinfo);
3297                 return ret;
3298         }
3299         newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3300                                          raid_disks, 2),
3301                              GFP_KERNEL);
3302         if (!newmirrors) {
3303                 kfree(newpoolinfo);
3304                 mempool_exit(&newpool);
3305                 return -ENOMEM;
3306         }
3307
3308         freeze_array(conf, 0);
3309
3310         /* ok, everything is stopped */
3311         oldpool = conf->r1bio_pool;
3312         conf->r1bio_pool = newpool;
3313
3314         for (d = d2 = 0; d < conf->raid_disks; d++) {
3315                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3316                 if (rdev && rdev->raid_disk != d2) {
3317                         sysfs_unlink_rdev(mddev, rdev);
3318                         rdev->raid_disk = d2;
3319                         sysfs_unlink_rdev(mddev, rdev);
3320                         if (sysfs_link_rdev(mddev, rdev))
3321                                 pr_warn("md/raid1:%s: cannot register rd%d\n",
3322                                         mdname(mddev), rdev->raid_disk);
3323                 }
3324                 if (rdev)
3325                         newmirrors[d2++].rdev = rdev;
3326         }
3327         kfree(conf->mirrors);
3328         conf->mirrors = newmirrors;
3329         kfree(conf->poolinfo);
3330         conf->poolinfo = newpoolinfo;
3331
3332         spin_lock_irqsave(&conf->device_lock, flags);
3333         mddev->degraded += (raid_disks - conf->raid_disks);
3334         spin_unlock_irqrestore(&conf->device_lock, flags);
3335         conf->raid_disks = mddev->raid_disks = raid_disks;
3336         mddev->delta_disks = 0;
3337
3338         unfreeze_array(conf);
3339
3340         set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3341         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3342         md_wakeup_thread(mddev->thread);
3343
3344         mempool_exit(&oldpool);
3345         return 0;
3346 }
3347
3348 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3349 {
3350         struct r1conf *conf = mddev->private;
3351
3352         if (quiesce)
3353                 freeze_array(conf, 0);
3354         else
3355                 unfreeze_array(conf);
3356 }
3357
3358 static void *raid1_takeover(struct mddev *mddev)
3359 {
3360         /* raid1 can take over:
3361          *  raid5 with 2 devices, any layout or chunk size
3362          */
3363         if (mddev->level == 5 && mddev->raid_disks == 2) {
3364                 struct r1conf *conf;
3365                 mddev->new_level = 1;
3366                 mddev->new_layout = 0;
3367                 mddev->new_chunk_sectors = 0;
3368                 conf = setup_conf(mddev);
3369                 if (!IS_ERR(conf)) {
3370                         /* Array must appear to be quiesced */
3371                         conf->array_frozen = 1;
3372                         mddev_clear_unsupported_flags(mddev,
3373                                 UNSUPPORTED_MDDEV_FLAGS);
3374                 }
3375                 return conf;
3376         }
3377         return ERR_PTR(-EINVAL);
3378 }
3379
3380 static struct md_personality raid1_personality =
3381 {
3382         .name           = "raid1",
3383         .level          = 1,
3384         .owner          = THIS_MODULE,
3385         .make_request   = raid1_make_request,
3386         .run            = raid1_run,
3387         .free           = raid1_free,
3388         .status         = raid1_status,
3389         .error_handler  = raid1_error,
3390         .hot_add_disk   = raid1_add_disk,
3391         .hot_remove_disk= raid1_remove_disk,
3392         .spare_active   = raid1_spare_active,
3393         .sync_request   = raid1_sync_request,
3394         .resize         = raid1_resize,
3395         .size           = raid1_size,
3396         .check_reshape  = raid1_reshape,
3397         .quiesce        = raid1_quiesce,
3398         .takeover       = raid1_takeover,
3399         .congested      = raid1_congested,
3400 };
3401
3402 static int __init raid_init(void)
3403 {
3404         return register_md_personality(&raid1_personality);
3405 }
3406
3407 static void raid_exit(void)
3408 {
3409         unregister_md_personality(&raid1_personality);
3410 }
3411
3412 module_init(raid_init);
3413 module_exit(raid_exit);
3414 MODULE_LICENSE("GPL");
3415 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3416 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3417 MODULE_ALIAS("md-raid1");
3418 MODULE_ALIAS("md-level-1");
3419
3420 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);