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(RESYNC_PAGES, gfp_flags);
 169                 if (!bio)
 170                         goto out_free_bio;
 171                 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
 172                 r1_bio->bios[j] = bio;
 173         }
 174         /*
 175          * Allocate RESYNC_PAGES data pages and attach them to
 176          * the first bio.
 177          * If this is a user-requested check/repair, allocate
 178          * RESYNC_PAGES for each bio.
 179          */
 180         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 181                 need_pages = pi->raid_disks;
 182         else
 183                 need_pages = 1;
 184         for (j = 0; j < pi->raid_disks; j++) {
 185                 struct resync_pages *rp = &rps[j];
 186
 187                 bio = r1_bio->bios[j];
 188
 189                 if (j < need_pages) {
 190                         if (resync_alloc_pages(rp, gfp_flags))
 191                                 goto out_free_pages;
 192                 } else {
 193                         memcpy(rp, &rps[0], sizeof(*rp));
 194                         resync_get_all_pages(rp);
 195                 }
 196
 197                 rp->raid_bio = r1_bio;
 198                 bio->bi_private = rp;
 199         }
 200
 201         r1_bio->master_bio = NULL;
 202
 203         return r1_bio;
 204
 205 out_free_pages:
 206         while (--j >= 0)
 207                 resync_free_pages(&rps[j]);
 208
 209 out_free_bio:
 210         while (++j < pi->raid_disks) {
 211                 bio_uninit(r1_bio->bios[j]);
 212                 kfree(r1_bio->bios[j]);
 213         }
 214         kfree(rps);
 215
 216 out_free_r1bio:
 217         rbio_pool_free(r1_bio, data);
 218         return NULL;
 219 }
 220
 221 static void r1buf_pool_free(void *__r1_bio, void *data)
 222 {
 223         struct pool_info *pi = data;
 224         int i;
 225         struct r1bio *r1bio = __r1_bio;
 226         struct resync_pages *rp = NULL;
 227
 228         for (i = pi->raid_disks; i--; ) {
 229                 rp = get_resync_pages(r1bio->bios[i]);
 230                 resync_free_pages(rp);
 231                 bio_uninit(r1bio->bios[i]);
 232                 kfree(r1bio->bios[i]);
 233         }
 234
 235         /* resync pages array stored in the 1st bio's .bi_private */
 236         kfree(rp);
 237
 238         rbio_pool_free(r1bio, data);
 239 }
 240
 241 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 242 {
 243         int i;
 244
 245         for (i = 0; i < conf->raid_disks * 2; i++) {
 246                 struct bio **bio = r1_bio->bios + i;
 247                 if (!BIO_SPECIAL(*bio))
 248                         bio_put(*bio);
 249                 *bio = NULL;
 250         }
 251 }
 252
 253 static void free_r1bio(struct r1bio *r1_bio)
 254 {
 255         struct r1conf *conf = r1_bio->mddev->private;
 256
 257         put_all_bios(conf, r1_bio);
 258         mempool_free(r1_bio, &conf->r1bio_pool);
 259 }
 260
 261 static void put_buf(struct r1bio *r1_bio)
 262 {
 263         struct r1conf *conf = r1_bio->mddev->private;
 264         sector_t sect = r1_bio->sector;
 265         int i;
 266
 267         for (i = 0; i < conf->raid_disks * 2; i++) {
 268                 struct bio *bio = r1_bio->bios[i];
 269                 if (bio->bi_end_io)
 270                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 271         }
 272
 273         mempool_free(r1_bio, &conf->r1buf_pool);
 274
 275         lower_barrier(conf, sect);
 276 }
 277
 278 static void reschedule_retry(struct r1bio *r1_bio)
 279 {
 280         unsigned long flags;
 281         struct mddev *mddev = r1_bio->mddev;
 282         struct r1conf *conf = mddev->private;
 283         int idx;
 284
 285         idx = sector_to_idx(r1_bio->sector);
 286         spin_lock_irqsave(&conf->device_lock, flags);
 287         list_add(&r1_bio->retry_list, &conf->retry_list);
 288         atomic_inc(&conf->nr_queued[idx]);
 289         spin_unlock_irqrestore(&conf->device_lock, flags);
 290
 291         wake_up(&conf->wait_barrier);
 292         md_wakeup_thread(mddev->thread);
 293 }
 294
 295 /*
 296  * raid_end_bio_io() is called when we have finished servicing a mirrored
 297  * operation and are ready to return a success/failure code to the buffer
 298  * cache layer.
 299  */
 300 static void call_bio_endio(struct r1bio *r1_bio)
 301 {
 302         struct bio *bio = r1_bio->master_bio;
 303
 304         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 305                 bio->bi_status = BLK_STS_IOERR;
 306
 307         bio_endio(bio);
 308 }
 309
 310 static void raid_end_bio_io(struct r1bio *r1_bio)
 311 {
 312         struct bio *bio = r1_bio->master_bio;
 313         struct r1conf *conf = r1_bio->mddev->private;
 314         sector_t sector = r1_bio->sector;
 315
 316         /* if nobody has done the final endio yet, do it now */
 317         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 318                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 319                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 320                          (unsigned long long) bio->bi_iter.bi_sector,
 321                          (unsigned long long) bio_end_sector(bio) - 1);
 322
 323                 call_bio_endio(r1_bio);
 324         }
 325
 326         free_r1bio(r1_bio);
 327         /*
 328          * Wake up any possible resync thread that waits for the device
 329          * to go idle.  All I/Os, even write-behind writes, are done.
 330          */
 331         allow_barrier(conf, sector);
 332 }
 333
 334 /*
 335  * Update disk head position estimator based on IRQ completion info.
 336  */
 337 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 338 {
 339         struct r1conf *conf = r1_bio->mddev->private;
 340
 341         conf->mirrors[disk].head_position =
 342                 r1_bio->sector + (r1_bio->sectors);
 343 }
 344
 345 /*
 346  * Find the disk number which triggered given bio
 347  */
 348 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 349 {
 350         int mirror;
 351         struct r1conf *conf = r1_bio->mddev->private;
 352         int raid_disks = conf->raid_disks;
 353
 354         for (mirror = 0; mirror < raid_disks * 2; mirror++)
 355                 if (r1_bio->bios[mirror] == bio)
 356                         break;
 357
 358         BUG_ON(mirror == raid_disks * 2);
 359         update_head_pos(mirror, r1_bio);
 360
 361         return mirror;
 362 }
 363
 364 static void raid1_end_read_request(struct bio *bio)
 365 {
 366         int uptodate = !bio->bi_status;
 367         struct r1bio *r1_bio = bio->bi_private;
 368         struct r1conf *conf = r1_bio->mddev->private;
 369         struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 370
 371         /*
 372          * this branch is our 'one mirror IO has finished' event handler:
 373          */
 374         update_head_pos(r1_bio->read_disk, r1_bio);
 375
 376         if (uptodate)
 377                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 378         else if (test_bit(FailFast, &rdev->flags) &&
 379                  test_bit(R1BIO_FailFast, &r1_bio->state))
 380                 /* This was a fail-fast read so we definitely
 381                  * want to retry */
 382                 ;
 383         else {
 384                 /* If all other devices have failed, we want to return
 385                  * the error upwards rather than fail the last device.
 386                  * Here we redefine "uptodate" to mean "Don't want to retry"
 387                  */
 388                 unsigned long flags;
 389                 spin_lock_irqsave(&conf->device_lock, flags);
 390                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 391                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 392                      test_bit(In_sync, &rdev->flags)))
 393                         uptodate = 1;
 394                 spin_unlock_irqrestore(&conf->device_lock, flags);
 395         }
 396
 397         if (uptodate) {
 398                 raid_end_bio_io(r1_bio);
 399                 rdev_dec_pending(rdev, conf->mddev);
 400         } else {
 401                 /*
 402                  * oops, read error:
 403                  */
 404                 pr_err_ratelimited("md/raid1:%s: %pg: rescheduling sector %llu\n",
 405                                    mdname(conf->mddev),
 406                                    rdev->bdev,
 407                                    (unsigned long long)r1_bio->sector);
 408                 set_bit(R1BIO_ReadError, &r1_bio->state);
 409                 reschedule_retry(r1_bio);
 410                 /* don't drop the reference on read_disk yet */
 411         }
 412 }
 413
 414 static void close_write(struct r1bio *r1_bio)
 415 {
 416         /* it really is the end of this request */
 417         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 418                 bio_free_pages(r1_bio->behind_master_bio);
 419                 bio_put(r1_bio->behind_master_bio);
 420                 r1_bio->behind_master_bio = NULL;
 421         }
 422         /* clear the bitmap if all writes complete successfully */
 423         md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 424                            r1_bio->sectors,
 425                            !test_bit(R1BIO_Degraded, &r1_bio->state),
 426                            test_bit(R1BIO_BehindIO, &r1_bio->state));
 427         md_write_end(r1_bio->mddev);
 428 }
 429
 430 static void r1_bio_write_done(struct r1bio *r1_bio)
 431 {
 432         if (!atomic_dec_and_test(&r1_bio->remaining))
 433                 return;
 434
 435         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 436                 reschedule_retry(r1_bio);
 437         else {
 438                 close_write(r1_bio);
 439                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 440                         reschedule_retry(r1_bio);
 441                 else
 442                         raid_end_bio_io(r1_bio);
 443         }
 444 }
 445
 446 static void raid1_end_write_request(struct bio *bio)
 447 {
 448         struct r1bio *r1_bio = bio->bi_private;
 449         int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 450         struct r1conf *conf = r1_bio->mddev->private;
 451         struct bio *to_put = NULL;
 452         int mirror = find_bio_disk(r1_bio, bio);
 453         struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 454         bool discard_error;
 455         sector_t lo = r1_bio->sector;
 456         sector_t hi = r1_bio->sector + r1_bio->sectors;
 457
 458         discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 459
 460         /*
 461          * 'one mirror IO has finished' event handler:
 462          */
 463         if (bio->bi_status && !discard_error) {
 464                 set_bit(WriteErrorSeen, &rdev->flags);
 465                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
 466                         set_bit(MD_RECOVERY_NEEDED, &
 467                                 conf->mddev->recovery);
 468
 469                 if (test_bit(FailFast, &rdev->flags) &&
 470                     (bio->bi_opf & MD_FAILFAST) &&
 471                     /* We never try FailFast to WriteMostly devices */
 472                     !test_bit(WriteMostly, &rdev->flags)) {
 473                         md_error(r1_bio->mddev, rdev);
 474                 }
 475
 476                 /*
 477                  * When the device is faulty, it is not necessary to
 478                  * handle write error.
 479                  */
 480                 if (!test_bit(Faulty, &rdev->flags))
 481                         set_bit(R1BIO_WriteError, &r1_bio->state);
 482                 else {
 483                         /* Fail the request */
 484                         set_bit(R1BIO_Degraded, &r1_bio->state);
 485                         /* Finished with this branch */
 486                         r1_bio->bios[mirror] = NULL;
 487                         to_put = bio;
 488                 }
 489         } else {
 490                 /*
 491                  * Set R1BIO_Uptodate in our master bio, so that we
 492                  * will return a good error code for to the higher
 493                  * levels even if IO on some other mirrored buffer
 494                  * fails.
 495                  *
 496                  * The 'master' represents the composite IO operation
 497                  * to user-side. So if something waits for IO, then it
 498                  * will wait for the 'master' bio.
 499                  */
 500                 sector_t first_bad;
 501                 int bad_sectors;
 502
 503                 r1_bio->bios[mirror] = NULL;
 504                 to_put = bio;
 505                 /*
 506                  * Do not set R1BIO_Uptodate if the current device is
 507                  * rebuilding or Faulty. This is because we cannot use
 508                  * such device for properly reading the data back (we could
 509                  * potentially use it, if the current write would have felt
 510                  * before rdev->recovery_offset, but for simplicity we don't
 511                  * check this here.
 512                  */
 513                 if (test_bit(In_sync, &rdev->flags) &&
 514                     !test_bit(Faulty, &rdev->flags))
 515                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 516
 517                 /* Maybe we can clear some bad blocks. */
 518                 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 519                                 &first_bad, &bad_sectors) && !discard_error) {
 520                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 521                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 522                 }
 523         }
 524
 525         if (behind) {
 526                 if (test_bit(CollisionCheck, &rdev->flags))
 527                         remove_serial(rdev, lo, hi);
 528                 if (test_bit(WriteMostly, &rdev->flags))
 529                         atomic_dec(&r1_bio->behind_remaining);
 530
 531                 /*
 532                  * In behind mode, we ACK the master bio once the I/O
 533                  * has safely reached all non-writemostly
 534                  * disks. Setting the Returned bit ensures that this
 535                  * gets done only once -- we don't ever want to return
 536                  * -EIO here, instead we'll wait
 537                  */
 538                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 539                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 540                         /* Maybe we can return now */
 541                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 542                                 struct bio *mbio = r1_bio->master_bio;
 543                                 pr_debug("raid1: behind end write sectors"
 544                                          " %llu-%llu\n",
 545                                          (unsigned long long) mbio->bi_iter.bi_sector,
 546                                          (unsigned long long) bio_end_sector(mbio) - 1);
 547                                 call_bio_endio(r1_bio);
 548                         }
 549                 }
 550         } else if (rdev->mddev->serialize_policy)
 551                 remove_serial(rdev, lo, hi);
 552         if (r1_bio->bios[mirror] == NULL)
 553                 rdev_dec_pending(rdev, conf->mddev);
 554
 555         /*
 556          * Let's see if all mirrored write operations have finished
 557          * already.
 558          */
 559         r1_bio_write_done(r1_bio);
 560
 561         if (to_put)
 562                 bio_put(to_put);
 563 }
 564
 565 static sector_t align_to_barrier_unit_end(sector_t start_sector,
 566                                           sector_t sectors)
 567 {
 568         sector_t len;
 569
 570         WARN_ON(sectors == 0);
 571         /*
 572          * len is the number of sectors from start_sector to end of the
 573          * barrier unit which start_sector belongs to.
 574          */
 575         len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 576               start_sector;
 577
 578         if (len > sectors)
 579                 len = sectors;
 580
 581         return len;
 582 }
 583
 584 /*
 585  * This routine returns the disk from which the requested read should
 586  * be done. There is a per-array 'next expected sequential IO' sector
 587  * number - if this matches on the next IO then we use the last disk.
 588  * There is also a per-disk 'last know head position' sector that is
 589  * maintained from IRQ contexts, both the normal and the resync IO
 590  * completion handlers update this position correctly. If there is no
 591  * perfect sequential match then we pick the disk whose head is closest.
 592  *
 593  * If there are 2 mirrors in the same 2 devices, performance degrades
 594  * because position is mirror, not device based.
 595  *
 596  * The rdev for the device selected will have nr_pending incremented.
 597  */
 598 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 599 {
 600         const sector_t this_sector = r1_bio->sector;
 601         int sectors;
 602         int best_good_sectors;
 603         int best_disk, best_dist_disk, best_pending_disk;
 604         int has_nonrot_disk;
 605         int disk;
 606         sector_t best_dist;
 607         unsigned int min_pending;
 608         struct md_rdev *rdev;
 609         int choose_first;
 610         int choose_next_idle;
 611
 612         rcu_read_lock();
 613         /*
 614          * Check if we can balance. We can balance on the whole
 615          * device if no resync is going on, or below the resync window.
 616          * We take the first readable disk when above the resync window.
 617          */
 618  retry:
 619         sectors = r1_bio->sectors;
 620         best_disk = -1;
 621         best_dist_disk = -1;
 622         best_dist = MaxSector;
 623         best_pending_disk = -1;
 624         min_pending = UINT_MAX;
 625         best_good_sectors = 0;
 626         has_nonrot_disk = 0;
 627         choose_next_idle = 0;
 628         clear_bit(R1BIO_FailFast, &r1_bio->state);
 629
 630         if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 631             (mddev_is_clustered(conf->mddev) &&
 632             md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 633                     this_sector + sectors)))
 634                 choose_first = 1;
 635         else
 636                 choose_first = 0;
 637
 638         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 639                 sector_t dist;
 640                 sector_t first_bad;
 641                 int bad_sectors;
 642                 unsigned int pending;
 643                 bool nonrot;
 644
 645                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 646                 if (r1_bio->bios[disk] == IO_BLOCKED
 647                     || rdev == NULL
 648                     || test_bit(Faulty, &rdev->flags))
 649                         continue;
 650                 if (!test_bit(In_sync, &rdev->flags) &&
 651                     rdev->recovery_offset < this_sector + sectors)
 652                         continue;
 653                 if (test_bit(WriteMostly, &rdev->flags)) {
 654                         /* Don't balance among write-mostly, just
 655                          * use the first as a last resort */
 656                         if (best_dist_disk < 0) {
 657                                 if (is_badblock(rdev, this_sector, sectors,
 658                                                 &first_bad, &bad_sectors)) {
 659                                         if (first_bad <= this_sector)
 660                                                 /* Cannot use this */
 661                                                 continue;
 662                                         best_good_sectors = first_bad - this_sector;
 663                                 } else
 664                                         best_good_sectors = sectors;
 665                                 best_dist_disk = disk;
 666                                 best_pending_disk = disk;
 667                         }
 668                         continue;
 669                 }
 670                 /* This is a reasonable device to use.  It might
 671                  * even be best.
 672                  */
 673                 if (is_badblock(rdev, this_sector, sectors,
 674                                 &first_bad, &bad_sectors)) {
 675                         if (best_dist < MaxSector)
 676                                 /* already have a better device */
 677                                 continue;
 678                         if (first_bad <= this_sector) {
 679                                 /* cannot read here. If this is the 'primary'
 680                                  * device, then we must not read beyond
 681                                  * bad_sectors from another device..
 682                                  */
 683                                 bad_sectors -= (this_sector - first_bad);
 684                                 if (choose_first && sectors > bad_sectors)
 685                                         sectors = bad_sectors;
 686                                 if (best_good_sectors > sectors)
 687                                         best_good_sectors = sectors;
 688
 689                         } else {
 690                                 sector_t good_sectors = first_bad - this_sector;
 691                                 if (good_sectors > best_good_sectors) {
 692                                         best_good_sectors = good_sectors;
 693                                         best_disk = disk;
 694                                 }
 695                                 if (choose_first)
 696                                         break;
 697                         }
 698                         continue;
 699                 } else {
 700                         if ((sectors > best_good_sectors) && (best_disk >= 0))
 701                                 best_disk = -1;
 702                         best_good_sectors = sectors;
 703                 }
 704
 705                 if (best_disk >= 0)
 706                         /* At least two disks to choose from so failfast is OK */
 707                         set_bit(R1BIO_FailFast, &r1_bio->state);
 708
 709                 nonrot = bdev_nonrot(rdev->bdev);
 710                 has_nonrot_disk |= nonrot;
 711                 pending = atomic_read(&rdev->nr_pending);
 712                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 713                 if (choose_first) {
 714                         best_disk = disk;
 715                         break;
 716                 }
 717                 /* Don't change to another disk for sequential reads */
 718                 if (conf->mirrors[disk].next_seq_sect == this_sector
 719                     || dist == 0) {
 720                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 721                         struct raid1_info *mirror = &conf->mirrors[disk];
 722
 723                         best_disk = disk;
 724                         /*
 725                          * If buffered sequential IO size exceeds optimal
 726                          * iosize, check if there is idle disk. If yes, choose
 727                          * the idle disk. read_balance could already choose an
 728                          * idle disk before noticing it's a sequential IO in
 729                          * this disk. This doesn't matter because this disk
 730                          * will idle, next time it will be utilized after the
 731                          * first disk has IO size exceeds optimal iosize. In
 732                          * this way, iosize of the first disk will be optimal
 733                          * iosize at least. iosize of the second disk might be
 734                          * small, but not a big deal since when the second disk
 735                          * starts IO, the first disk is likely still busy.
 736                          */
 737                         if (nonrot && opt_iosize > 0 &&
 738                             mirror->seq_start != MaxSector &&
 739                             mirror->next_seq_sect > opt_iosize &&
 740                             mirror->next_seq_sect - opt_iosize >=
 741                             mirror->seq_start) {
 742                                 choose_next_idle = 1;
 743                                 continue;
 744                         }
 745                         break;
 746                 }
 747
 748                 if (choose_next_idle)
 749                         continue;
 750
 751                 if (min_pending > pending) {
 752                         min_pending = pending;
 753                         best_pending_disk = disk;
 754                 }
 755
 756                 if (dist < best_dist) {
 757                         best_dist = dist;
 758                         best_dist_disk = disk;
 759                 }
 760         }
 761
 762         /*
 763          * If all disks are rotational, choose the closest disk. If any disk is
 764          * non-rotational, choose the disk with less pending request even the
 765          * disk is rotational, which might/might not be optimal for raids with
 766          * mixed ratation/non-rotational disks depending on workload.
 767          */
 768         if (best_disk == -1) {
 769                 if (has_nonrot_disk || min_pending == 0)
 770                         best_disk = best_pending_disk;
 771                 else
 772                         best_disk = best_dist_disk;
 773         }
 774
 775         if (best_disk >= 0) {
 776                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 777                 if (!rdev)
 778                         goto retry;
 779                 atomic_inc(&rdev->nr_pending);
 780                 sectors = best_good_sectors;
 781
 782                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 783                         conf->mirrors[best_disk].seq_start = this_sector;
 784
 785                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 786         }
 787         rcu_read_unlock();
 788         *max_sectors = sectors;
 789
 790         return best_disk;
 791 }
 792
 793 static void wake_up_barrier(struct r1conf *conf)
 794 {
 795         if (wq_has_sleeper(&conf->wait_barrier))
 796                 wake_up(&conf->wait_barrier);
 797 }
 798
 799 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 800 {
 801         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
 802         raid1_prepare_flush_writes(conf->mddev->bitmap);
 803         wake_up_barrier(conf);
 804
 805         while (bio) { /* submit pending writes */
 806                 struct bio *next = bio->bi_next;
 807
 808                 raid1_submit_write(bio);
 809                 bio = next;
 810                 cond_resched();
 811         }
 812 }
 813
 814 static void flush_pending_writes(struct r1conf *conf)
 815 {
 816         /* Any writes that have been queued but are awaiting
 817          * bitmap updates get flushed here.
 818          */
 819         spin_lock_irq(&conf->device_lock);
 820
 821         if (conf->pending_bio_list.head) {
 822                 struct blk_plug plug;
 823                 struct bio *bio;
 824
 825                 bio = bio_list_get(&conf->pending_bio_list);
 826                 spin_unlock_irq(&conf->device_lock);
 827
 828                 /*
 829                  * As this is called in a wait_event() loop (see freeze_array),
 830                  * current->state might be TASK_UNINTERRUPTIBLE which will
 831                  * cause a warning when we prepare to wait again.  As it is
 832                  * rare that this path is taken, it is perfectly safe to force
 833                  * us to go around the wait_event() loop again, so the warning
 834                  * is a false-positive.  Silence the warning by resetting
 835                  * thread state
 836                  */
 837                 __set_current_state(TASK_RUNNING);
 838                 blk_start_plug(&plug);
 839                 flush_bio_list(conf, bio);
 840                 blk_finish_plug(&plug);
 841         } else
 842                 spin_unlock_irq(&conf->device_lock);
 843 }
 844
 845 /* Barriers....
 846  * Sometimes we need to suspend IO while we do something else,
 847  * either some resync/recovery, or reconfigure the array.
 848  * To do this we raise a 'barrier'.
 849  * The 'barrier' is a counter that can be raised multiple times
 850  * to count how many activities are happening which preclude
 851  * normal IO.
 852  * We can only raise the barrier if there is no pending IO.
 853  * i.e. if nr_pending == 0.
 854  * We choose only to raise the barrier if no-one is waiting for the
 855  * barrier to go down.  This means that as soon as an IO request
 856  * is ready, no other operations which require a barrier will start
 857  * until the IO request has had a chance.
 858  *
 859  * So: regular IO calls 'wait_barrier'.  When that returns there
 860  *    is no backgroup IO happening,  It must arrange to call
 861  *    allow_barrier when it has finished its IO.
 862  * backgroup IO calls must call raise_barrier.  Once that returns
 863  *    there is no normal IO happeing.  It must arrange to call
 864  *    lower_barrier when the particular background IO completes.
 865  *
 866  * If resync/recovery is interrupted, returns -EINTR;
 867  * Otherwise, returns 0.
 868  */
 869 static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
 870 {
 871         int idx = sector_to_idx(sector_nr);
 872
 873         spin_lock_irq(&conf->resync_lock);
 874
 875         /* Wait until no block IO is waiting */
 876         wait_event_lock_irq(conf->wait_barrier,
 877                             !atomic_read(&conf->nr_waiting[idx]),
 878                             conf->resync_lock);
 879
 880         /* block any new IO from starting */
 881         atomic_inc(&conf->barrier[idx]);
 882         /*
 883          * In raise_barrier() we firstly increase conf->barrier[idx] then
 884          * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 885          * increase conf->nr_pending[idx] then check conf->barrier[idx].
 886          * A memory barrier here to make sure conf->nr_pending[idx] won't
 887          * be fetched before conf->barrier[idx] is increased. Otherwise
 888          * there will be a race between raise_barrier() and _wait_barrier().
 889          */
 890         smp_mb__after_atomic();
 891
 892         /* For these conditions we must wait:
 893          * A: while the array is in frozen state
 894          * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 895          *    existing in corresponding I/O barrier bucket.
 896          * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 897          *    max resync count which allowed on current I/O barrier bucket.
 898          */
 899         wait_event_lock_irq(conf->wait_barrier,
 900                             (!conf->array_frozen &&
 901                              !atomic_read(&conf->nr_pending[idx]) &&
 902                              atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
 903                                 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
 904                             conf->resync_lock);
 905
 906         if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
 907                 atomic_dec(&conf->barrier[idx]);
 908                 spin_unlock_irq(&conf->resync_lock);
 909                 wake_up(&conf->wait_barrier);
 910                 return -EINTR;
 911         }
 912
 913         atomic_inc(&conf->nr_sync_pending);
 914         spin_unlock_irq(&conf->resync_lock);
 915
 916         return 0;
 917 }
 918
 919 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 920 {
 921         int idx = sector_to_idx(sector_nr);
 922
 923         BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 924
 925         atomic_dec(&conf->barrier[idx]);
 926         atomic_dec(&conf->nr_sync_pending);
 927         wake_up(&conf->wait_barrier);
 928 }
 929
 930 static bool _wait_barrier(struct r1conf *conf, int idx, bool nowait)
 931 {
 932         bool ret = true;
 933
 934         /*
 935          * We need to increase conf->nr_pending[idx] very early here,
 936          * then raise_barrier() can be blocked when it waits for
 937          * conf->nr_pending[idx] to be 0. Then we can avoid holding
 938          * conf->resync_lock when there is no barrier raised in same
 939          * barrier unit bucket. Also if the array is frozen, I/O
 940          * should be blocked until array is unfrozen.
 941          */
 942         atomic_inc(&conf->nr_pending[idx]);
 943         /*
 944          * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
 945          * check conf->barrier[idx]. In raise_barrier() we firstly increase
 946          * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
 947          * barrier is necessary here to make sure conf->barrier[idx] won't be
 948          * fetched before conf->nr_pending[idx] is increased. Otherwise there
 949          * will be a race between _wait_barrier() and raise_barrier().
 950          */
 951         smp_mb__after_atomic();
 952
 953         /*
 954          * Don't worry about checking two atomic_t variables at same time
 955          * here. If during we check conf->barrier[idx], the array is
 956          * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
 957          * 0, it is safe to return and make the I/O continue. Because the
 958          * array is frozen, all I/O returned here will eventually complete
 959          * or be queued, no race will happen. See code comment in
 960          * frozen_array().
 961          */
 962         if (!READ_ONCE(conf->array_frozen) &&
 963             !atomic_read(&conf->barrier[idx]))
 964                 return ret;
 965
 966         /*
 967          * After holding conf->resync_lock, conf->nr_pending[idx]
 968          * should be decreased before waiting for barrier to drop.
 969          * Otherwise, we may encounter a race condition because
 970          * raise_barrer() might be waiting for conf->nr_pending[idx]
 971          * to be 0 at same time.
 972          */
 973         spin_lock_irq(&conf->resync_lock);
 974         atomic_inc(&conf->nr_waiting[idx]);
 975         atomic_dec(&conf->nr_pending[idx]);
 976         /*
 977          * In case freeze_array() is waiting for
 978          * get_unqueued_pending() == extra
 979          */
 980         wake_up_barrier(conf);
 981         /* Wait for the barrier in same barrier unit bucket to drop. */
 982
 983         /* Return false when nowait flag is set */
 984         if (nowait) {
 985                 ret = false;
 986         } else {
 987                 wait_event_lock_irq(conf->wait_barrier,
 988                                 !conf->array_frozen &&
 989                                 !atomic_read(&conf->barrier[idx]),
 990                                 conf->resync_lock);
 991                 atomic_inc(&conf->nr_pending[idx]);
 992         }
 993
 994         atomic_dec(&conf->nr_waiting[idx]);
 995         spin_unlock_irq(&conf->resync_lock);
 996         return ret;
 997 }
 998
 999 static bool wait_read_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1000 {
1001         int idx = sector_to_idx(sector_nr);
1002         bool ret = true;
1003
1004         /*
1005          * Very similar to _wait_barrier(). The difference is, for read
1006          * I/O we don't need wait for sync I/O, but if the whole array
1007          * is frozen, the read I/O still has to wait until the array is
1008          * unfrozen. Since there is no ordering requirement with
1009          * conf->barrier[idx] here, memory barrier is unnecessary as well.
1010          */
1011         atomic_inc(&conf->nr_pending[idx]);
1012
1013         if (!READ_ONCE(conf->array_frozen))
1014                 return ret;
1015
1016         spin_lock_irq(&conf->resync_lock);
1017         atomic_inc(&conf->nr_waiting[idx]);
1018         atomic_dec(&conf->nr_pending[idx]);
1019         /*
1020          * In case freeze_array() is waiting for
1021          * get_unqueued_pending() == extra
1022          */
1023         wake_up_barrier(conf);
1024         /* Wait for array to be unfrozen */
1025
1026         /* Return false when nowait flag is set */
1027         if (nowait) {
1028                 /* Return false when nowait flag is set */
1029                 ret = false;
1030         } else {
1031                 wait_event_lock_irq(conf->wait_barrier,
1032                                 !conf->array_frozen,
1033                                 conf->resync_lock);
1034                 atomic_inc(&conf->nr_pending[idx]);
1035         }
1036
1037         atomic_dec(&conf->nr_waiting[idx]);
1038         spin_unlock_irq(&conf->resync_lock);
1039         return ret;
1040 }
1041
1042 static bool wait_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1043 {
1044         int idx = sector_to_idx(sector_nr);
1045
1046         return _wait_barrier(conf, idx, nowait);
1047 }
1048
1049 static void _allow_barrier(struct r1conf *conf, int idx)
1050 {
1051         atomic_dec(&conf->nr_pending[idx]);
1052         wake_up_barrier(conf);
1053 }
1054
1055 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1056 {
1057         int idx = sector_to_idx(sector_nr);
1058
1059         _allow_barrier(conf, idx);
1060 }
1061
1062 /* conf->resync_lock should be held */
1063 static int get_unqueued_pending(struct r1conf *conf)
1064 {
1065         int idx, ret;
1066
1067         ret = atomic_read(&conf->nr_sync_pending);
1068         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1069                 ret += atomic_read(&conf->nr_pending[idx]) -
1070                         atomic_read(&conf->nr_queued[idx]);
1071
1072         return ret;
1073 }
1074
1075 static void freeze_array(struct r1conf *conf, int extra)
1076 {
1077         /* Stop sync I/O and normal I/O and wait for everything to
1078          * go quiet.
1079          * This is called in two situations:
1080          * 1) management command handlers (reshape, remove disk, quiesce).
1081          * 2) one normal I/O request failed.
1082
1083          * After array_frozen is set to 1, new sync IO will be blocked at
1084          * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1085          * or wait_read_barrier(). The flying I/Os will either complete or be
1086          * queued. When everything goes quite, there are only queued I/Os left.
1087
1088          * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1089          * barrier bucket index which this I/O request hits. When all sync and
1090          * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1091          * of all conf->nr_queued[]. But normal I/O failure is an exception,
1092          * in handle_read_error(), we may call freeze_array() before trying to
1093          * fix the read error. In this case, the error read I/O is not queued,
1094          * so get_unqueued_pending() == 1.
1095          *
1096          * Therefore before this function returns, we need to wait until
1097          * get_unqueued_pendings(conf) gets equal to extra. For
1098          * normal I/O context, extra is 1, in rested situations extra is 0.
1099          */
1100         spin_lock_irq(&conf->resync_lock);
1101         conf->array_frozen = 1;
1102         raid1_log(conf->mddev, "wait freeze");
1103         wait_event_lock_irq_cmd(
1104                 conf->wait_barrier,
1105                 get_unqueued_pending(conf) == extra,
1106                 conf->resync_lock,
1107                 flush_pending_writes(conf));
1108         spin_unlock_irq(&conf->resync_lock);
1109 }
1110 static void unfreeze_array(struct r1conf *conf)
1111 {
1112         /* reverse the effect of the freeze */
1113         spin_lock_irq(&conf->resync_lock);
1114         conf->array_frozen = 0;
1115         spin_unlock_irq(&conf->resync_lock);
1116         wake_up(&conf->wait_barrier);
1117 }
1118
1119 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1120                                            struct bio *bio)
1121 {
1122         int size = bio->bi_iter.bi_size;
1123         unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1124         int i = 0;
1125         struct bio *behind_bio = NULL;
1126
1127         behind_bio = bio_alloc_bioset(NULL, vcnt, 0, GFP_NOIO,
1128                                       &r1_bio->mddev->bio_set);
1129         if (!behind_bio)
1130                 return;
1131
1132         /* discard op, we don't support writezero/writesame yet */
1133         if (!bio_has_data(bio)) {
1134                 behind_bio->bi_iter.bi_size = size;
1135                 goto skip_copy;
1136         }
1137
1138         while (i < vcnt && size) {
1139                 struct page *page;
1140                 int len = min_t(int, PAGE_SIZE, size);
1141
1142                 page = alloc_page(GFP_NOIO);
1143                 if (unlikely(!page))
1144                         goto free_pages;
1145
1146                 if (!bio_add_page(behind_bio, page, len, 0)) {
1147                         put_page(page);
1148                         goto free_pages;
1149                 }
1150
1151                 size -= len;
1152                 i++;
1153         }
1154
1155         bio_copy_data(behind_bio, bio);
1156 skip_copy:
1157         r1_bio->behind_master_bio = behind_bio;
1158         set_bit(R1BIO_BehindIO, &r1_bio->state);
1159
1160         return;
1161
1162 free_pages:
1163         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1164                  bio->bi_iter.bi_size);
1165         bio_free_pages(behind_bio);
1166         bio_put(behind_bio);
1167 }
1168
1169 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1170 {
1171         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1172                                                   cb);
1173         struct mddev *mddev = plug->cb.data;
1174         struct r1conf *conf = mddev->private;
1175         struct bio *bio;
1176
1177         if (from_schedule) {
1178                 spin_lock_irq(&conf->device_lock);
1179                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1180                 spin_unlock_irq(&conf->device_lock);
1181                 wake_up_barrier(conf);
1182                 md_wakeup_thread(mddev->thread);
1183                 kfree(plug);
1184                 return;
1185         }
1186
1187         /* we aren't scheduling, so we can do the write-out directly. */
1188         bio = bio_list_get(&plug->pending);
1189         flush_bio_list(conf, bio);
1190         kfree(plug);
1191 }
1192
1193 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1194 {
1195         r1_bio->master_bio = bio;
1196         r1_bio->sectors = bio_sectors(bio);
1197         r1_bio->state = 0;
1198         r1_bio->mddev = mddev;
1199         r1_bio->sector = bio->bi_iter.bi_sector;
1200 }
1201
1202 static inline struct r1bio *
1203 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1204 {
1205         struct r1conf *conf = mddev->private;
1206         struct r1bio *r1_bio;
1207
1208         r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1209         /* Ensure no bio records IO_BLOCKED */
1210         memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1211         init_r1bio(r1_bio, mddev, bio);
1212         return r1_bio;
1213 }
1214
1215 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1216                                int max_read_sectors, struct r1bio *r1_bio)
1217 {
1218         struct r1conf *conf = mddev->private;
1219         struct raid1_info *mirror;
1220         struct bio *read_bio;
1221         struct bitmap *bitmap = mddev->bitmap;
1222         const enum req_op op = bio_op(bio);
1223         const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1224         int max_sectors;
1225         int rdisk;
1226         bool r1bio_existed = !!r1_bio;
1227         char b[BDEVNAME_SIZE];
1228
1229         /*
1230          * If r1_bio is set, we are blocking the raid1d thread
1231          * so there is a tiny risk of deadlock.  So ask for
1232          * emergency memory if needed.
1233          */
1234         gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1235
1236         if (r1bio_existed) {
1237                 /* Need to get the block device name carefully */
1238                 struct md_rdev *rdev;
1239                 rcu_read_lock();
1240                 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1241                 if (rdev)
1242                         snprintf(b, sizeof(b), "%pg", rdev->bdev);
1243                 else
1244                         strcpy(b, "???");
1245                 rcu_read_unlock();
1246         }
1247
1248         /*
1249          * Still need barrier for READ in case that whole
1250          * array is frozen.
1251          */
1252         if (!wait_read_barrier(conf, bio->bi_iter.bi_sector,
1253                                 bio->bi_opf & REQ_NOWAIT)) {
1254                 bio_wouldblock_error(bio);
1255                 return;
1256         }
1257
1258         if (!r1_bio)
1259                 r1_bio = alloc_r1bio(mddev, bio);
1260         else
1261                 init_r1bio(r1_bio, mddev, bio);
1262         r1_bio->sectors = max_read_sectors;
1263
1264         /*
1265          * make_request() can abort the operation when read-ahead is being
1266          * used and no empty request is available.
1267          */
1268         rdisk = read_balance(conf, r1_bio, &max_sectors);
1269
1270         if (rdisk < 0) {
1271                 /* couldn't find anywhere to read from */
1272                 if (r1bio_existed) {
1273                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1274                                             mdname(mddev),
1275                                             b,
1276                                             (unsigned long long)r1_bio->sector);
1277                 }
1278                 raid_end_bio_io(r1_bio);
1279                 return;
1280         }
1281         mirror = conf->mirrors + rdisk;
1282
1283         if (r1bio_existed)
1284                 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %pg\n",
1285                                     mdname(mddev),
1286                                     (unsigned long long)r1_bio->sector,
1287                                     mirror->rdev->bdev);
1288
1289         if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1290             bitmap) {
1291                 /*
1292                  * Reading from a write-mostly device must take care not to
1293                  * over-take any writes that are 'behind'
1294                  */
1295                 raid1_log(mddev, "wait behind writes");
1296                 wait_event(bitmap->behind_wait,
1297                            atomic_read(&bitmap->behind_writes) == 0);
1298         }
1299
1300         if (max_sectors < bio_sectors(bio)) {
1301                 struct bio *split = bio_split(bio, max_sectors,
1302                                               gfp, &conf->bio_split);
1303                 bio_chain(split, bio);
1304                 submit_bio_noacct(bio);
1305                 bio = split;
1306                 r1_bio->master_bio = bio;
1307                 r1_bio->sectors = max_sectors;
1308         }
1309
1310         r1_bio->read_disk = rdisk;
1311         if (!r1bio_existed) {
1312                 md_account_bio(mddev, &bio);
1313                 r1_bio->master_bio = bio;
1314         }
1315         read_bio = bio_alloc_clone(mirror->rdev->bdev, bio, gfp,
1316                                    &mddev->bio_set);
1317
1318         r1_bio->bios[rdisk] = read_bio;
1319
1320         read_bio->bi_iter.bi_sector = r1_bio->sector +
1321                 mirror->rdev->data_offset;
1322         read_bio->bi_end_io = raid1_end_read_request;
1323         read_bio->bi_opf = op | do_sync;
1324         if (test_bit(FailFast, &mirror->rdev->flags) &&
1325             test_bit(R1BIO_FailFast, &r1_bio->state))
1326                 read_bio->bi_opf |= MD_FAILFAST;
1327         read_bio->bi_private = r1_bio;
1328
1329         if (mddev->gendisk)
1330                 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1331                                       r1_bio->sector);
1332
1333         submit_bio_noacct(read_bio);
1334 }
1335
1336 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1337                                 int max_write_sectors)
1338 {
1339         struct r1conf *conf = mddev->private;
1340         struct r1bio *r1_bio;
1341         int i, disks;
1342         struct bitmap *bitmap = mddev->bitmap;
1343         unsigned long flags;
1344         struct md_rdev *blocked_rdev;
1345         int first_clone;
1346         int max_sectors;
1347         bool write_behind = false;
1348
1349         if (mddev_is_clustered(mddev) &&
1350              md_cluster_ops->area_resyncing(mddev, WRITE,
1351                      bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1352
1353                 DEFINE_WAIT(w);
1354                 if (bio->bi_opf & REQ_NOWAIT) {
1355                         bio_wouldblock_error(bio);
1356                         return;
1357                 }
1358                 for (;;) {
1359                         prepare_to_wait(&conf->wait_barrier,
1360                                         &w, TASK_IDLE);
1361                         if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1362                                                         bio->bi_iter.bi_sector,
1363                                                         bio_end_sector(bio)))
1364                                 break;
1365                         schedule();
1366                 }
1367                 finish_wait(&conf->wait_barrier, &w);
1368         }
1369
1370         /*
1371          * Register the new request and wait if the reconstruction
1372          * thread has put up a bar for new requests.
1373          * Continue immediately if no resync is active currently.
1374          */
1375         if (!wait_barrier(conf, bio->bi_iter.bi_sector,
1376                                 bio->bi_opf & REQ_NOWAIT)) {
1377                 bio_wouldblock_error(bio);
1378                 return;
1379         }
1380
1381  retry_write:
1382         r1_bio = alloc_r1bio(mddev, bio);
1383         r1_bio->sectors = max_write_sectors;
1384
1385         /* first select target devices under rcu_lock and
1386          * inc refcount on their rdev.  Record them by setting
1387          * bios[x] to bio
1388          * If there are known/acknowledged bad blocks on any device on
1389          * which we have seen a write error, we want to avoid writing those
1390          * blocks.
1391          * This potentially requires several writes to write around
1392          * the bad blocks.  Each set of writes gets it's own r1bio
1393          * with a set of bios attached.
1394          */
1395
1396         disks = conf->raid_disks * 2;
1397         blocked_rdev = NULL;
1398         rcu_read_lock();
1399         max_sectors = r1_bio->sectors;
1400         for (i = 0;  i < disks; i++) {
1401                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1402
1403                 /*
1404                  * The write-behind io is only attempted on drives marked as
1405                  * write-mostly, which means we could allocate write behind
1406                  * bio later.
1407                  */
1408                 if (rdev && test_bit(WriteMostly, &rdev->flags))
1409                         write_behind = true;
1410
1411                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1412                         atomic_inc(&rdev->nr_pending);
1413                         blocked_rdev = rdev;
1414                         break;
1415                 }
1416                 r1_bio->bios[i] = NULL;
1417                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1418                         if (i < conf->raid_disks)
1419                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1420                         continue;
1421                 }
1422
1423                 atomic_inc(&rdev->nr_pending);
1424                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1425                         sector_t first_bad;
1426                         int bad_sectors;
1427                         int is_bad;
1428
1429                         is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1430                                              &first_bad, &bad_sectors);
1431                         if (is_bad < 0) {
1432                                 /* mustn't write here until the bad block is
1433                                  * acknowledged*/
1434                                 set_bit(BlockedBadBlocks, &rdev->flags);
1435                                 blocked_rdev = rdev;
1436                                 break;
1437                         }
1438                         if (is_bad && first_bad <= r1_bio->sector) {
1439                                 /* Cannot write here at all */
1440                                 bad_sectors -= (r1_bio->sector - first_bad);
1441                                 if (bad_sectors < max_sectors)
1442                                         /* mustn't write more than bad_sectors
1443                                          * to other devices yet
1444                                          */
1445                                         max_sectors = bad_sectors;
1446                                 rdev_dec_pending(rdev, mddev);
1447                                 /* We don't set R1BIO_Degraded as that
1448                                  * only applies if the disk is
1449                                  * missing, so it might be re-added,
1450                                  * and we want to know to recover this
1451                                  * chunk.
1452                                  * In this case the device is here,
1453                                  * and the fact that this chunk is not
1454                                  * in-sync is recorded in the bad
1455                                  * block log
1456                                  */
1457                                 continue;
1458                         }
1459                         if (is_bad) {
1460                                 int good_sectors = first_bad - r1_bio->sector;
1461                                 if (good_sectors < max_sectors)
1462                                         max_sectors = good_sectors;
1463                         }
1464                 }
1465                 r1_bio->bios[i] = bio;
1466         }
1467         rcu_read_unlock();
1468
1469         if (unlikely(blocked_rdev)) {
1470                 /* Wait for this device to become unblocked */
1471                 int j;
1472
1473                 for (j = 0; j < i; j++)
1474                         if (r1_bio->bios[j])
1475                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1476                 free_r1bio(r1_bio);
1477                 allow_barrier(conf, bio->bi_iter.bi_sector);
1478
1479                 if (bio->bi_opf & REQ_NOWAIT) {
1480                         bio_wouldblock_error(bio);
1481                         return;
1482                 }
1483                 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1484                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1485                 wait_barrier(conf, bio->bi_iter.bi_sector, false);
1486                 goto retry_write;
1487         }
1488
1489         /*
1490          * When using a bitmap, we may call alloc_behind_master_bio below.
1491          * alloc_behind_master_bio allocates a copy of the data payload a page
1492          * at a time and thus needs a new bio that can fit the whole payload
1493          * this bio in page sized chunks.
1494          */
1495         if (write_behind && bitmap)
1496                 max_sectors = min_t(int, max_sectors,
1497                                     BIO_MAX_VECS * (PAGE_SIZE >> 9));
1498         if (max_sectors < bio_sectors(bio)) {
1499                 struct bio *split = bio_split(bio, max_sectors,
1500                                               GFP_NOIO, &conf->bio_split);
1501                 bio_chain(split, bio);
1502                 submit_bio_noacct(bio);
1503                 bio = split;
1504                 r1_bio->master_bio = bio;
1505                 r1_bio->sectors = max_sectors;
1506         }
1507
1508         md_account_bio(mddev, &bio);
1509         r1_bio->master_bio = bio;
1510         atomic_set(&r1_bio->remaining, 1);
1511         atomic_set(&r1_bio->behind_remaining, 0);
1512
1513         first_clone = 1;
1514
1515         for (i = 0; i < disks; i++) {
1516                 struct bio *mbio = NULL;
1517                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1518                 if (!r1_bio->bios[i])
1519                         continue;
1520
1521                 if (first_clone) {
1522                         /* do behind I/O ?
1523                          * Not if there are too many, or cannot
1524                          * allocate memory, or a reader on WriteMostly
1525                          * is waiting for behind writes to flush */
1526                         if (bitmap && write_behind &&
1527                             (atomic_read(&bitmap->behind_writes)
1528                              < mddev->bitmap_info.max_write_behind) &&
1529                             !waitqueue_active(&bitmap->behind_wait)) {
1530                                 alloc_behind_master_bio(r1_bio, bio);
1531                         }
1532
1533                         md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1534                                              test_bit(R1BIO_BehindIO, &r1_bio->state));
1535                         first_clone = 0;
1536                 }
1537
1538                 if (r1_bio->behind_master_bio) {
1539                         mbio = bio_alloc_clone(rdev->bdev,
1540                                                r1_bio->behind_master_bio,
1541                                                GFP_NOIO, &mddev->bio_set);
1542                         if (test_bit(CollisionCheck, &rdev->flags))
1543                                 wait_for_serialization(rdev, r1_bio);
1544                         if (test_bit(WriteMostly, &rdev->flags))
1545                                 atomic_inc(&r1_bio->behind_remaining);
1546                 } else {
1547                         mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
1548                                                &mddev->bio_set);
1549
1550                         if (mddev->serialize_policy)
1551                                 wait_for_serialization(rdev, r1_bio);
1552                 }
1553
1554                 r1_bio->bios[i] = mbio;
1555
1556                 mbio->bi_iter.bi_sector = (r1_bio->sector + rdev->data_offset);
1557                 mbio->bi_end_io = raid1_end_write_request;
1558                 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1559                 if (test_bit(FailFast, &rdev->flags) &&
1560                     !test_bit(WriteMostly, &rdev->flags) &&
1561                     conf->raid_disks - mddev->degraded > 1)
1562                         mbio->bi_opf |= MD_FAILFAST;
1563                 mbio->bi_private = r1_bio;
1564
1565                 atomic_inc(&r1_bio->remaining);
1566
1567                 if (mddev->gendisk)
1568                         trace_block_bio_remap(mbio, disk_devt(mddev->gendisk),
1569                                               r1_bio->sector);
1570                 /* flush_pending_writes() needs access to the rdev so...*/
1571                 mbio->bi_bdev = (void *)rdev;
1572                 if (!raid1_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) {
1573                         spin_lock_irqsave(&conf->device_lock, flags);
1574                         bio_list_add(&conf->pending_bio_list, mbio);
1575                         spin_unlock_irqrestore(&conf->device_lock, flags);
1576                         md_wakeup_thread(mddev->thread);
1577                 }
1578         }
1579
1580         r1_bio_write_done(r1_bio);
1581
1582         /* In case raid1d snuck in to freeze_array */
1583         wake_up_barrier(conf);
1584 }
1585
1586 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1587 {
1588         sector_t sectors;
1589
1590         if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1591             && md_flush_request(mddev, bio))
1592                 return true;
1593
1594         /*
1595          * There is a limit to the maximum size, but
1596          * the read/write handler might find a lower limit
1597          * due to bad blocks.  To avoid multiple splits,
1598          * we pass the maximum number of sectors down
1599          * and let the lower level perform the split.
1600          */
1601         sectors = align_to_barrier_unit_end(
1602                 bio->bi_iter.bi_sector, bio_sectors(bio));
1603
1604         if (bio_data_dir(bio) == READ)
1605                 raid1_read_request(mddev, bio, sectors, NULL);
1606         else {
1607                 if (!md_write_start(mddev,bio))
1608                         return false;
1609                 raid1_write_request(mddev, bio, sectors);
1610         }
1611         return true;
1612 }
1613
1614 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1615 {
1616         struct r1conf *conf = mddev->private;
1617         int i;
1618
1619         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1620                    conf->raid_disks - mddev->degraded);
1621         rcu_read_lock();
1622         for (i = 0; i < conf->raid_disks; i++) {
1623                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1624                 seq_printf(seq, "%s",
1625                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1626         }
1627         rcu_read_unlock();
1628         seq_printf(seq, "]");
1629 }
1630
1631 /**
1632  * raid1_error() - RAID1 error handler.
1633  * @mddev: affected md device.
1634  * @rdev: member device to fail.
1635  *
1636  * The routine acknowledges &rdev failure and determines new @mddev state.
1637  * If it failed, then:
1638  *      - &MD_BROKEN flag is set in &mddev->flags.
1639  *      - recovery is disabled.
1640  * Otherwise, it must be degraded:
1641  *      - recovery is interrupted.
1642  *      - &mddev->degraded is bumped.
1643  *
1644  * @rdev is marked as &Faulty excluding case when array is failed and
1645  * &mddev->fail_last_dev is off.
1646  */
1647 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1648 {
1649         struct r1conf *conf = mddev->private;
1650         unsigned long flags;
1651
1652         spin_lock_irqsave(&conf->device_lock, flags);
1653
1654         if (test_bit(In_sync, &rdev->flags) &&
1655             (conf->raid_disks - mddev->degraded) == 1) {
1656                 set_bit(MD_BROKEN, &mddev->flags);
1657
1658                 if (!mddev->fail_last_dev) {
1659                         conf->recovery_disabled = mddev->recovery_disabled;
1660                         spin_unlock_irqrestore(&conf->device_lock, flags);
1661                         return;
1662                 }
1663         }
1664         set_bit(Blocked, &rdev->flags);
1665         if (test_and_clear_bit(In_sync, &rdev->flags))
1666                 mddev->degraded++;
1667         set_bit(Faulty, &rdev->flags);
1668         spin_unlock_irqrestore(&conf->device_lock, flags);
1669         /*
1670          * if recovery is running, make sure it aborts.
1671          */
1672         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1673         set_mask_bits(&mddev->sb_flags, 0,
1674                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1675         pr_crit("md/raid1:%s: Disk failure on %pg, disabling device.\n"
1676                 "md/raid1:%s: Operation continuing on %d devices.\n",
1677                 mdname(mddev), rdev->bdev,
1678                 mdname(mddev), conf->raid_disks - mddev->degraded);
1679 }
1680
1681 static void print_conf(struct r1conf *conf)
1682 {
1683         int i;
1684
1685         pr_debug("RAID1 conf printout:\n");
1686         if (!conf) {
1687                 pr_debug("(!conf)\n");
1688                 return;
1689         }
1690         pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1691                  conf->raid_disks);
1692
1693         rcu_read_lock();
1694         for (i = 0; i < conf->raid_disks; i++) {
1695                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1696                 if (rdev)
1697                         pr_debug(" disk %d, wo:%d, o:%d, dev:%pg\n",
1698                                  i, !test_bit(In_sync, &rdev->flags),
1699                                  !test_bit(Faulty, &rdev->flags),
1700                                  rdev->bdev);
1701         }
1702         rcu_read_unlock();
1703 }
1704
1705 static void close_sync(struct r1conf *conf)
1706 {
1707         int idx;
1708
1709         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1710                 _wait_barrier(conf, idx, false);
1711                 _allow_barrier(conf, idx);
1712         }
1713
1714         mempool_exit(&conf->r1buf_pool);
1715 }
1716
1717 static int raid1_spare_active(struct mddev *mddev)
1718 {
1719         int i;
1720         struct r1conf *conf = mddev->private;
1721         int count = 0;
1722         unsigned long flags;
1723
1724         /*
1725          * Find all failed disks within the RAID1 configuration
1726          * and mark them readable.
1727          * Called under mddev lock, so rcu protection not needed.
1728          * device_lock used to avoid races with raid1_end_read_request
1729          * which expects 'In_sync' flags and ->degraded to be consistent.
1730          */
1731         spin_lock_irqsave(&conf->device_lock, flags);
1732         for (i = 0; i < conf->raid_disks; i++) {
1733                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1734                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1735                 if (repl
1736                     && !test_bit(Candidate, &repl->flags)
1737                     && repl->recovery_offset == MaxSector
1738                     && !test_bit(Faulty, &repl->flags)
1739                     && !test_and_set_bit(In_sync, &repl->flags)) {
1740                         /* replacement has just become active */
1741                         if (!rdev ||
1742                             !test_and_clear_bit(In_sync, &rdev->flags))
1743                                 count++;
1744                         if (rdev) {
1745                                 /* Replaced device not technically
1746                                  * faulty, but we need to be sure
1747                                  * it gets removed and never re-added
1748                                  */
1749                                 set_bit(Faulty, &rdev->flags);
1750                                 sysfs_notify_dirent_safe(
1751                                         rdev->sysfs_state);
1752                         }
1753                 }
1754                 if (rdev
1755                     && rdev->recovery_offset == MaxSector
1756                     && !test_bit(Faulty, &rdev->flags)
1757                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1758                         count++;
1759                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1760                 }
1761         }
1762         mddev->degraded -= count;
1763         spin_unlock_irqrestore(&conf->device_lock, flags);
1764
1765         print_conf(conf);
1766         return count;
1767 }
1768
1769 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1770 {
1771         struct r1conf *conf = mddev->private;
1772         int err = -EEXIST;
1773         int mirror = 0, repl_slot = -1;
1774         struct raid1_info *p;
1775         int first = 0;
1776         int last = conf->raid_disks - 1;
1777
1778         if (mddev->recovery_disabled == conf->recovery_disabled)
1779                 return -EBUSY;
1780
1781         if (md_integrity_add_rdev(rdev, mddev))
1782                 return -ENXIO;
1783
1784         if (rdev->raid_disk >= 0)
1785                 first = last = rdev->raid_disk;
1786
1787         /*
1788          * find the disk ... but prefer rdev->saved_raid_disk
1789          * if possible.
1790          */
1791         if (rdev->saved_raid_disk >= 0 &&
1792             rdev->saved_raid_disk >= first &&
1793             rdev->saved_raid_disk < conf->raid_disks &&
1794             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1795                 first = last = rdev->saved_raid_disk;
1796
1797         for (mirror = first; mirror <= last; mirror++) {
1798                 p = conf->mirrors + mirror;
1799                 if (!p->rdev) {
1800                         if (mddev->gendisk)
1801                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1802                                                   rdev->data_offset << 9);
1803
1804                         p->head_position = 0;
1805                         rdev->raid_disk = mirror;
1806                         err = 0;
1807                         /* As all devices are equivalent, we don't need a full recovery
1808                          * if this was recently any drive of the array
1809                          */
1810                         if (rdev->saved_raid_disk < 0)
1811                                 conf->fullsync = 1;
1812                         rcu_assign_pointer(p->rdev, rdev);
1813                         break;
1814                 }
1815                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1816                     p[conf->raid_disks].rdev == NULL && repl_slot < 0)
1817                         repl_slot = mirror;
1818         }
1819
1820         if (err && repl_slot >= 0) {
1821                 /* Add this device as a replacement */
1822                 p = conf->mirrors + repl_slot;
1823                 clear_bit(In_sync, &rdev->flags);
1824                 set_bit(Replacement, &rdev->flags);
1825                 rdev->raid_disk = repl_slot;
1826                 err = 0;
1827                 conf->fullsync = 1;
1828                 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1829         }
1830
1831         print_conf(conf);
1832         return err;
1833 }
1834
1835 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1836 {
1837         struct r1conf *conf = mddev->private;
1838         int err = 0;
1839         int number = rdev->raid_disk;
1840
1841         if (unlikely(number >= conf->raid_disks))
1842                 goto abort;
1843
1844         struct raid1_info *p = conf->mirrors + number;
1845
1846         if (rdev != p->rdev)
1847                 p = conf->mirrors + conf->raid_disks + number;
1848
1849         print_conf(conf);
1850         if (rdev == p->rdev) {
1851                 if (test_bit(In_sync, &rdev->flags) ||
1852                     atomic_read(&rdev->nr_pending)) {
1853                         err = -EBUSY;
1854                         goto abort;
1855                 }
1856                 /* Only remove non-faulty devices if recovery
1857                  * is not possible.
1858                  */
1859                 if (!test_bit(Faulty, &rdev->flags) &&
1860                     mddev->recovery_disabled != conf->recovery_disabled &&
1861                     mddev->degraded < conf->raid_disks) {
1862                         err = -EBUSY;
1863                         goto abort;
1864                 }
1865                 p->rdev = NULL;
1866                 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1867                         synchronize_rcu();
1868                         if (atomic_read(&rdev->nr_pending)) {
1869                                 /* lost the race, try later */
1870                                 err = -EBUSY;
1871                                 p->rdev = rdev;
1872                                 goto abort;
1873                         }
1874                 }
1875                 if (conf->mirrors[conf->raid_disks + number].rdev) {
1876                         /* We just removed a device that is being replaced.
1877                          * Move down the replacement.  We drain all IO before
1878                          * doing this to avoid confusion.
1879                          */
1880                         struct md_rdev *repl =
1881                                 conf->mirrors[conf->raid_disks + number].rdev;
1882                         freeze_array(conf, 0);
1883                         if (atomic_read(&repl->nr_pending)) {
1884                                 /* It means that some queued IO of retry_list
1885                                  * hold repl. Thus, we cannot set replacement
1886                                  * as NULL, avoiding rdev NULL pointer
1887                                  * dereference in sync_request_write and
1888                                  * handle_write_finished.
1889                                  */
1890                                 err = -EBUSY;
1891                                 unfreeze_array(conf);
1892                                 goto abort;
1893                         }
1894                         clear_bit(Replacement, &repl->flags);
1895                         p->rdev = repl;
1896                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1897                         unfreeze_array(conf);
1898                 }
1899
1900                 clear_bit(WantReplacement, &rdev->flags);
1901                 err = md_integrity_register(mddev);
1902         }
1903 abort:
1904
1905         print_conf(conf);
1906         return err;
1907 }
1908
1909 static void end_sync_read(struct bio *bio)
1910 {
1911         struct r1bio *r1_bio = get_resync_r1bio(bio);
1912
1913         update_head_pos(r1_bio->read_disk, r1_bio);
1914
1915         /*
1916          * we have read a block, now it needs to be re-written,
1917          * or re-read if the read failed.
1918          * We don't do much here, just schedule handling by raid1d
1919          */
1920         if (!bio->bi_status)
1921                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1922
1923         if (atomic_dec_and_test(&r1_bio->remaining))
1924                 reschedule_retry(r1_bio);
1925 }
1926
1927 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1928 {
1929         sector_t sync_blocks = 0;
1930         sector_t s = r1_bio->sector;
1931         long sectors_to_go = r1_bio->sectors;
1932
1933         /* make sure these bits don't get cleared. */
1934         do {
1935                 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1936                 s += sync_blocks;
1937                 sectors_to_go -= sync_blocks;
1938         } while (sectors_to_go > 0);
1939 }
1940
1941 static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1942 {
1943         if (atomic_dec_and_test(&r1_bio->remaining)) {
1944                 struct mddev *mddev = r1_bio->mddev;
1945                 int s = r1_bio->sectors;
1946
1947                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1948                     test_bit(R1BIO_WriteError, &r1_bio->state))
1949                         reschedule_retry(r1_bio);
1950                 else {
1951                         put_buf(r1_bio);
1952                         md_done_sync(mddev, s, uptodate);
1953                 }
1954         }
1955 }
1956
1957 static void end_sync_write(struct bio *bio)
1958 {
1959         int uptodate = !bio->bi_status;
1960         struct r1bio *r1_bio = get_resync_r1bio(bio);
1961         struct mddev *mddev = r1_bio->mddev;
1962         struct r1conf *conf = mddev->private;
1963         sector_t first_bad;
1964         int bad_sectors;
1965         struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1966
1967         if (!uptodate) {
1968                 abort_sync_write(mddev, r1_bio);
1969                 set_bit(WriteErrorSeen, &rdev->flags);
1970                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1971                         set_bit(MD_RECOVERY_NEEDED, &
1972                                 mddev->recovery);
1973                 set_bit(R1BIO_WriteError, &r1_bio->state);
1974         } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1975                                &first_bad, &bad_sectors) &&
1976                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1977                                 r1_bio->sector,
1978                                 r1_bio->sectors,
1979                                 &first_bad, &bad_sectors)
1980                 )
1981                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1982
1983         put_sync_write_buf(r1_bio, uptodate);
1984 }
1985
1986 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1987                            int sectors, struct page *page, int rw)
1988 {
1989         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1990                 /* success */
1991                 return 1;
1992         if (rw == WRITE) {
1993                 set_bit(WriteErrorSeen, &rdev->flags);
1994                 if (!test_and_set_bit(WantReplacement,
1995                                       &rdev->flags))
1996                         set_bit(MD_RECOVERY_NEEDED, &
1997                                 rdev->mddev->recovery);
1998         }
1999         /* need to record an error - either for the block or the device */
2000         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2001                 md_error(rdev->mddev, rdev);
2002         return 0;
2003 }
2004
2005 static int fix_sync_read_error(struct r1bio *r1_bio)
2006 {
2007         /* Try some synchronous reads of other devices to get
2008          * good data, much like with normal read errors.  Only
2009          * read into the pages we already have so we don't
2010          * need to re-issue the read request.
2011          * We don't need to freeze the array, because being in an
2012          * active sync request, there is no normal IO, and
2013          * no overlapping syncs.
2014          * We don't need to check is_badblock() again as we
2015          * made sure that anything with a bad block in range
2016          * will have bi_end_io clear.
2017          */
2018         struct mddev *mddev = r1_bio->mddev;
2019         struct r1conf *conf = mddev->private;
2020         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
2021         struct page **pages = get_resync_pages(bio)->pages;
2022         sector_t sect = r1_bio->sector;
2023         int sectors = r1_bio->sectors;
2024         int idx = 0;
2025         struct md_rdev *rdev;
2026
2027         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2028         if (test_bit(FailFast, &rdev->flags)) {
2029                 /* Don't try recovering from here - just fail it
2030                  * ... unless it is the last working device of course */
2031                 md_error(mddev, rdev);
2032                 if (test_bit(Faulty, &rdev->flags))
2033                         /* Don't try to read from here, but make sure
2034                          * put_buf does it's thing
2035                          */
2036                         bio->bi_end_io = end_sync_write;
2037         }
2038
2039         while(sectors) {
2040                 int s = sectors;
2041                 int d = r1_bio->read_disk;
2042                 int success = 0;
2043                 int start;
2044
2045                 if (s > (PAGE_SIZE>>9))
2046                         s = PAGE_SIZE >> 9;
2047                 do {
2048                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2049                                 /* No rcu protection needed here devices
2050                                  * can only be removed when no resync is
2051                                  * active, and resync is currently active
2052                                  */
2053                                 rdev = conf->mirrors[d].rdev;
2054                                 if (sync_page_io(rdev, sect, s<<9,
2055                                                  pages[idx],
2056                                                  REQ_OP_READ, false)) {
2057                                         success = 1;
2058                                         break;
2059                                 }
2060                         }
2061                         d++;
2062                         if (d == conf->raid_disks * 2)
2063                                 d = 0;
2064                 } while (!success && d != r1_bio->read_disk);
2065
2066                 if (!success) {
2067                         int abort = 0;
2068                         /* Cannot read from anywhere, this block is lost.
2069                          * Record a bad block on each device.  If that doesn't
2070                          * work just disable and interrupt the recovery.
2071                          * Don't fail devices as that won't really help.
2072                          */
2073                         pr_crit_ratelimited("md/raid1:%s: %pg: unrecoverable I/O read error for block %llu\n",
2074                                             mdname(mddev), bio->bi_bdev,
2075                                             (unsigned long long)r1_bio->sector);
2076                         for (d = 0; d < conf->raid_disks * 2; d++) {
2077                                 rdev = conf->mirrors[d].rdev;
2078                                 if (!rdev || test_bit(Faulty, &rdev->flags))
2079                                         continue;
2080                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
2081                                         abort = 1;
2082                         }
2083                         if (abort) {
2084                                 conf->recovery_disabled =
2085                                         mddev->recovery_disabled;
2086                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2087                                 md_done_sync(mddev, r1_bio->sectors, 0);
2088                                 put_buf(r1_bio);
2089                                 return 0;
2090                         }
2091                         /* Try next page */
2092                         sectors -= s;
2093                         sect += s;
2094                         idx++;
2095                         continue;
2096                 }
2097
2098                 start = d;
2099                 /* write it back and re-read */
2100                 while (d != r1_bio->read_disk) {
2101                         if (d == 0)
2102                                 d = conf->raid_disks * 2;
2103                         d--;
2104                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2105                                 continue;
2106                         rdev = conf->mirrors[d].rdev;
2107                         if (r1_sync_page_io(rdev, sect, s,
2108                                             pages[idx],
2109                                             WRITE) == 0) {
2110                                 r1_bio->bios[d]->bi_end_io = NULL;
2111                                 rdev_dec_pending(rdev, mddev);
2112                         }
2113                 }
2114                 d = start;
2115                 while (d != r1_bio->read_disk) {
2116                         if (d == 0)
2117                                 d = conf->raid_disks * 2;
2118                         d--;
2119                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2120                                 continue;
2121                         rdev = conf->mirrors[d].rdev;
2122                         if (r1_sync_page_io(rdev, sect, s,
2123                                             pages[idx],
2124                                             READ) != 0)
2125                                 atomic_add(s, &rdev->corrected_errors);
2126                 }
2127                 sectors -= s;
2128                 sect += s;
2129                 idx ++;
2130         }
2131         set_bit(R1BIO_Uptodate, &r1_bio->state);
2132         bio->bi_status = 0;
2133         return 1;
2134 }
2135
2136 static void process_checks(struct r1bio *r1_bio)
2137 {
2138         /* We have read all readable devices.  If we haven't
2139          * got the block, then there is no hope left.
2140          * If we have, then we want to do a comparison
2141          * and skip the write if everything is the same.
2142          * If any blocks failed to read, then we need to
2143          * attempt an over-write
2144          */
2145         struct mddev *mddev = r1_bio->mddev;
2146         struct r1conf *conf = mddev->private;
2147         int primary;
2148         int i;
2149         int vcnt;
2150
2151         /* Fix variable parts of all bios */
2152         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2153         for (i = 0; i < conf->raid_disks * 2; i++) {
2154                 blk_status_t status;
2155                 struct bio *b = r1_bio->bios[i];
2156                 struct resync_pages *rp = get_resync_pages(b);
2157                 if (b->bi_end_io != end_sync_read)
2158                         continue;
2159                 /* fixup the bio for reuse, but preserve errno */
2160                 status = b->bi_status;
2161                 bio_reset(b, conf->mirrors[i].rdev->bdev, REQ_OP_READ);
2162                 b->bi_status = status;
2163                 b->bi_iter.bi_sector = r1_bio->sector +
2164                         conf->mirrors[i].rdev->data_offset;
2165                 b->bi_end_io = end_sync_read;
2166                 rp->raid_bio = r1_bio;
2167                 b->bi_private = rp;
2168
2169                 /* initialize bvec table again */
2170                 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2171         }
2172         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2173                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2174                     !r1_bio->bios[primary]->bi_status) {
2175                         r1_bio->bios[primary]->bi_end_io = NULL;
2176                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2177                         break;
2178                 }
2179         r1_bio->read_disk = primary;
2180         for (i = 0; i < conf->raid_disks * 2; i++) {
2181                 int j = 0;
2182                 struct bio *pbio = r1_bio->bios[primary];
2183                 struct bio *sbio = r1_bio->bios[i];
2184                 blk_status_t status = sbio->bi_status;
2185                 struct page **ppages = get_resync_pages(pbio)->pages;
2186                 struct page **spages = get_resync_pages(sbio)->pages;
2187                 struct bio_vec *bi;
2188                 int page_len[RESYNC_PAGES] = { 0 };
2189                 struct bvec_iter_all iter_all;
2190
2191                 if (sbio->bi_end_io != end_sync_read)
2192                         continue;
2193                 /* Now we can 'fixup' the error value */
2194                 sbio->bi_status = 0;
2195
2196                 bio_for_each_segment_all(bi, sbio, iter_all)
2197                         page_len[j++] = bi->bv_len;
2198
2199                 if (!status) {
2200                         for (j = vcnt; j-- ; ) {
2201                                 if (memcmp(page_address(ppages[j]),
2202                                            page_address(spages[j]),
2203                                            page_len[j]))
2204                                         break;
2205                         }
2206                 } else
2207                         j = 0;
2208                 if (j >= 0)
2209                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2210                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2211                               && !status)) {
2212                         /* No need to write to this device. */
2213                         sbio->bi_end_io = NULL;
2214                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2215                         continue;
2216                 }
2217
2218                 bio_copy_data(sbio, pbio);
2219         }
2220 }
2221
2222 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2223 {
2224         struct r1conf *conf = mddev->private;
2225         int i;
2226         int disks = conf->raid_disks * 2;
2227         struct bio *wbio;
2228
2229         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2230                 /* ouch - failed to read all of that. */
2231                 if (!fix_sync_read_error(r1_bio))
2232                         return;
2233
2234         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2235                 process_checks(r1_bio);
2236
2237         /*
2238          * schedule writes
2239          */
2240         atomic_set(&r1_bio->remaining, 1);
2241         for (i = 0; i < disks ; i++) {
2242                 wbio = r1_bio->bios[i];
2243                 if (wbio->bi_end_io == NULL ||
2244                     (wbio->bi_end_io == end_sync_read &&
2245                      (i == r1_bio->read_disk ||
2246                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2247                         continue;
2248                 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2249                         abort_sync_write(mddev, r1_bio);
2250                         continue;
2251                 }
2252
2253                 wbio->bi_opf = REQ_OP_WRITE;
2254                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2255                         wbio->bi_opf |= MD_FAILFAST;
2256
2257                 wbio->bi_end_io = end_sync_write;
2258                 atomic_inc(&r1_bio->remaining);
2259                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2260
2261                 submit_bio_noacct(wbio);
2262         }
2263
2264         put_sync_write_buf(r1_bio, 1);
2265 }
2266
2267 /*
2268  * This is a kernel thread which:
2269  *
2270  *      1.      Retries failed read operations on working mirrors.
2271  *      2.      Updates the raid superblock when problems encounter.
2272  *      3.      Performs writes following reads for array synchronising.
2273  */
2274
2275 static void fix_read_error(struct r1conf *conf, int read_disk,
2276                            sector_t sect, int sectors)
2277 {
2278         struct mddev *mddev = conf->mddev;
2279         while(sectors) {
2280                 int s = sectors;
2281                 int d = read_disk;
2282                 int success = 0;
2283                 int start;
2284                 struct md_rdev *rdev;
2285
2286                 if (s > (PAGE_SIZE>>9))
2287                         s = PAGE_SIZE >> 9;
2288
2289                 do {
2290                         sector_t first_bad;
2291                         int bad_sectors;
2292
2293                         rcu_read_lock();
2294                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2295                         if (rdev &&
2296                             (test_bit(In_sync, &rdev->flags) ||
2297                              (!test_bit(Faulty, &rdev->flags) &&
2298                               rdev->recovery_offset >= sect + s)) &&
2299                             is_badblock(rdev, sect, s,
2300                                         &first_bad, &bad_sectors) == 0) {
2301                                 atomic_inc(&rdev->nr_pending);
2302                                 rcu_read_unlock();
2303                                 if (sync_page_io(rdev, sect, s<<9,
2304                                          conf->tmppage, REQ_OP_READ, false))
2305                                         success = 1;
2306                                 rdev_dec_pending(rdev, mddev);
2307                                 if (success)
2308                                         break;
2309                         } else
2310                                 rcu_read_unlock();
2311                         d++;
2312                         if (d == conf->raid_disks * 2)
2313                                 d = 0;
2314                 } while (d != read_disk);
2315
2316                 if (!success) {
2317                         /* Cannot read from anywhere - mark it bad */
2318                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2319                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2320                                 md_error(mddev, rdev);
2321                         break;
2322                 }
2323                 /* write it back and re-read */
2324                 start = d;
2325                 while (d != read_disk) {
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                                 r1_sync_page_io(rdev, sect, s,
2336                                                 conf->tmppage, WRITE);
2337                                 rdev_dec_pending(rdev, mddev);
2338                         } else
2339                                 rcu_read_unlock();
2340                 }
2341                 d = start;
2342                 while (d != read_disk) {
2343                         if (d==0)
2344                                 d = conf->raid_disks * 2;
2345                         d--;
2346                         rcu_read_lock();
2347                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2348                         if (rdev &&
2349                             !test_bit(Faulty, &rdev->flags)) {
2350                                 atomic_inc(&rdev->nr_pending);
2351                                 rcu_read_unlock();
2352                                 if (r1_sync_page_io(rdev, sect, s,
2353                                                     conf->tmppage, READ)) {
2354                                         atomic_add(s, &rdev->corrected_errors);
2355                                         pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %pg)\n",
2356                                                 mdname(mddev), s,
2357                                                 (unsigned long long)(sect +
2358                                                                      rdev->data_offset),
2359                                                 rdev->bdev);
2360                                 }
2361                                 rdev_dec_pending(rdev, mddev);
2362                         } else
2363                                 rcu_read_unlock();
2364                 }
2365                 sectors -= s;
2366                 sect += s;
2367         }
2368 }
2369
2370 static int narrow_write_error(struct r1bio *r1_bio, int i)
2371 {
2372         struct mddev *mddev = r1_bio->mddev;
2373         struct r1conf *conf = mddev->private;
2374         struct md_rdev *rdev = conf->mirrors[i].rdev;
2375
2376         /* bio has the data to be written to device 'i' where
2377          * we just recently had a write error.
2378          * We repeatedly clone the bio and trim down to one block,
2379          * then try the write.  Where the write fails we record
2380          * a bad block.
2381          * It is conceivable that the bio doesn't exactly align with
2382          * blocks.  We must handle this somehow.
2383          *
2384          * We currently own a reference on the rdev.
2385          */
2386
2387         int block_sectors;
2388         sector_t sector;
2389         int sectors;
2390         int sect_to_write = r1_bio->sectors;
2391         int ok = 1;
2392
2393         if (rdev->badblocks.shift < 0)
2394                 return 0;
2395
2396         block_sectors = roundup(1 << rdev->badblocks.shift,
2397                                 bdev_logical_block_size(rdev->bdev) >> 9);
2398         sector = r1_bio->sector;
2399         sectors = ((sector + block_sectors)
2400                    & ~(sector_t)(block_sectors - 1))
2401                 - sector;
2402
2403         while (sect_to_write) {
2404                 struct bio *wbio;
2405                 if (sectors > sect_to_write)
2406                         sectors = sect_to_write;
2407                 /* Write at 'sector' for 'sectors'*/
2408
2409                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2410                         wbio = bio_alloc_clone(rdev->bdev,
2411                                                r1_bio->behind_master_bio,
2412                                                GFP_NOIO, &mddev->bio_set);
2413                 } else {
2414                         wbio = bio_alloc_clone(rdev->bdev, r1_bio->master_bio,
2415                                                GFP_NOIO, &mddev->bio_set);
2416                 }
2417
2418                 wbio->bi_opf = REQ_OP_WRITE;
2419                 wbio->bi_iter.bi_sector = r1_bio->sector;
2420                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2421
2422                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2423                 wbio->bi_iter.bi_sector += rdev->data_offset;
2424
2425                 if (submit_bio_wait(wbio) < 0)
2426                         /* failure! */
2427                         ok = rdev_set_badblocks(rdev, sector,
2428                                                 sectors, 0)
2429                                 && ok;
2430
2431                 bio_put(wbio);
2432                 sect_to_write -= sectors;
2433                 sector += sectors;
2434                 sectors = block_sectors;
2435         }
2436         return ok;
2437 }
2438
2439 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2440 {
2441         int m;
2442         int s = r1_bio->sectors;
2443         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2444                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2445                 struct bio *bio = r1_bio->bios[m];
2446                 if (bio->bi_end_io == NULL)
2447                         continue;
2448                 if (!bio->bi_status &&
2449                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2450                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2451                 }
2452                 if (bio->bi_status &&
2453                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2454                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2455                                 md_error(conf->mddev, rdev);
2456                 }
2457         }
2458         put_buf(r1_bio);
2459         md_done_sync(conf->mddev, s, 1);
2460 }
2461
2462 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2463 {
2464         int m, idx;
2465         bool fail = false;
2466
2467         for (m = 0; m < conf->raid_disks * 2 ; m++)
2468                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2469                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2470                         rdev_clear_badblocks(rdev,
2471                                              r1_bio->sector,
2472                                              r1_bio->sectors, 0);
2473                         rdev_dec_pending(rdev, conf->mddev);
2474                 } else if (r1_bio->bios[m] != NULL) {
2475                         /* This drive got a write error.  We need to
2476                          * narrow down and record precise write
2477                          * errors.
2478                          */
2479                         fail = true;
2480                         if (!narrow_write_error(r1_bio, m)) {
2481                                 md_error(conf->mddev,
2482                                          conf->mirrors[m].rdev);
2483                                 /* an I/O failed, we can't clear the bitmap */
2484                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2485                         }
2486                         rdev_dec_pending(conf->mirrors[m].rdev,
2487                                          conf->mddev);
2488                 }
2489         if (fail) {
2490                 spin_lock_irq(&conf->device_lock);
2491                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2492                 idx = sector_to_idx(r1_bio->sector);
2493                 atomic_inc(&conf->nr_queued[idx]);
2494                 spin_unlock_irq(&conf->device_lock);
2495                 /*
2496                  * In case freeze_array() is waiting for condition
2497                  * get_unqueued_pending() == extra to be true.
2498                  */
2499                 wake_up(&conf->wait_barrier);
2500                 md_wakeup_thread(conf->mddev->thread);
2501         } else {
2502                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2503                         close_write(r1_bio);
2504                 raid_end_bio_io(r1_bio);
2505         }
2506 }
2507
2508 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2509 {
2510         struct mddev *mddev = conf->mddev;
2511         struct bio *bio;
2512         struct md_rdev *rdev;
2513         sector_t sector;
2514
2515         clear_bit(R1BIO_ReadError, &r1_bio->state);
2516         /* we got a read error. Maybe the drive is bad.  Maybe just
2517          * the block and we can fix it.
2518          * We freeze all other IO, and try reading the block from
2519          * other devices.  When we find one, we re-write
2520          * and check it that fixes the read error.
2521          * This is all done synchronously while the array is
2522          * frozen
2523          */
2524
2525         bio = r1_bio->bios[r1_bio->read_disk];
2526         bio_put(bio);
2527         r1_bio->bios[r1_bio->read_disk] = NULL;
2528
2529         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2530         if (mddev->ro == 0
2531             && !test_bit(FailFast, &rdev->flags)) {
2532                 freeze_array(conf, 1);
2533                 fix_read_error(conf, r1_bio->read_disk,
2534                                r1_bio->sector, r1_bio->sectors);
2535                 unfreeze_array(conf);
2536         } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2537                 md_error(mddev, rdev);
2538         } else {
2539                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2540         }
2541
2542         rdev_dec_pending(rdev, conf->mddev);
2543         sector = r1_bio->sector;
2544         bio = r1_bio->master_bio;
2545
2546         /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2547         r1_bio->state = 0;
2548         raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2549         allow_barrier(conf, sector);
2550 }
2551
2552 static void raid1d(struct md_thread *thread)
2553 {
2554         struct mddev *mddev = thread->mddev;
2555         struct r1bio *r1_bio;
2556         unsigned long flags;
2557         struct r1conf *conf = mddev->private;
2558         struct list_head *head = &conf->retry_list;
2559         struct blk_plug plug;
2560         int idx;
2561
2562         md_check_recovery(mddev);
2563
2564         if (!list_empty_careful(&conf->bio_end_io_list) &&
2565             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2566                 LIST_HEAD(tmp);
2567                 spin_lock_irqsave(&conf->device_lock, flags);
2568                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2569                         list_splice_init(&conf->bio_end_io_list, &tmp);
2570                 spin_unlock_irqrestore(&conf->device_lock, flags);
2571                 while (!list_empty(&tmp)) {
2572                         r1_bio = list_first_entry(&tmp, struct r1bio,
2573                                                   retry_list);
2574                         list_del(&r1_bio->retry_list);
2575                         idx = sector_to_idx(r1_bio->sector);
2576                         atomic_dec(&conf->nr_queued[idx]);
2577                         if (mddev->degraded)
2578                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2579                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2580                                 close_write(r1_bio);
2581                         raid_end_bio_io(r1_bio);
2582                 }
2583         }
2584
2585         blk_start_plug(&plug);
2586         for (;;) {
2587
2588                 flush_pending_writes(conf);
2589
2590                 spin_lock_irqsave(&conf->device_lock, flags);
2591                 if (list_empty(head)) {
2592                         spin_unlock_irqrestore(&conf->device_lock, flags);
2593                         break;
2594                 }
2595                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2596                 list_del(head->prev);
2597                 idx = sector_to_idx(r1_bio->sector);
2598                 atomic_dec(&conf->nr_queued[idx]);
2599                 spin_unlock_irqrestore(&conf->device_lock, flags);
2600
2601                 mddev = r1_bio->mddev;
2602                 conf = mddev->private;
2603                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2604                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2605                             test_bit(R1BIO_WriteError, &r1_bio->state))
2606                                 handle_sync_write_finished(conf, r1_bio);
2607                         else
2608                                 sync_request_write(mddev, r1_bio);
2609                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2610                            test_bit(R1BIO_WriteError, &r1_bio->state))
2611                         handle_write_finished(conf, r1_bio);
2612                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2613                         handle_read_error(conf, r1_bio);
2614                 else
2615                         WARN_ON_ONCE(1);
2616
2617                 cond_resched();
2618                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2619                         md_check_recovery(mddev);
2620         }
2621         blk_finish_plug(&plug);
2622 }
2623
2624 static int init_resync(struct r1conf *conf)
2625 {
2626         int buffs;
2627
2628         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2629         BUG_ON(mempool_initialized(&conf->r1buf_pool));
2630
2631         return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2632                             r1buf_pool_free, conf->poolinfo);
2633 }
2634
2635 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2636 {
2637         struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2638         struct resync_pages *rps;
2639         struct bio *bio;
2640         int i;
2641
2642         for (i = conf->poolinfo->raid_disks; i--; ) {
2643                 bio = r1bio->bios[i];
2644                 rps = bio->bi_private;
2645                 bio_reset(bio, NULL, 0);
2646                 bio->bi_private = rps;
2647         }
2648         r1bio->master_bio = NULL;
2649         return r1bio;
2650 }
2651
2652 /*
2653  * perform a "sync" on one "block"
2654  *
2655  * We need to make sure that no normal I/O request - particularly write
2656  * requests - conflict with active sync requests.
2657  *
2658  * This is achieved by tracking pending requests and a 'barrier' concept
2659  * that can be installed to exclude normal IO requests.
2660  */
2661
2662 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2663                                    int *skipped)
2664 {
2665         struct r1conf *conf = mddev->private;
2666         struct r1bio *r1_bio;
2667         struct bio *bio;
2668         sector_t max_sector, nr_sectors;
2669         int disk = -1;
2670         int i;
2671         int wonly = -1;
2672         int write_targets = 0, read_targets = 0;
2673         sector_t sync_blocks;
2674         int still_degraded = 0;
2675         int good_sectors = RESYNC_SECTORS;
2676         int min_bad = 0; /* number of sectors that are bad in all devices */
2677         int idx = sector_to_idx(sector_nr);
2678         int page_idx = 0;
2679
2680         if (!mempool_initialized(&conf->r1buf_pool))
2681                 if (init_resync(conf))
2682                         return 0;
2683
2684         max_sector = mddev->dev_sectors;
2685         if (sector_nr >= max_sector) {
2686                 /* If we aborted, we need to abort the
2687                  * sync on the 'current' bitmap chunk (there will
2688                  * only be one in raid1 resync.
2689                  * We can find the current addess in mddev->curr_resync
2690                  */
2691                 if (mddev->curr_resync < max_sector) /* aborted */
2692                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2693                                            &sync_blocks, 1);
2694                 else /* completed sync */
2695                         conf->fullsync = 0;
2696
2697                 md_bitmap_close_sync(mddev->bitmap);
2698                 close_sync(conf);
2699
2700                 if (mddev_is_clustered(mddev)) {
2701                         conf->cluster_sync_low = 0;
2702                         conf->cluster_sync_high = 0;
2703                 }
2704                 return 0;
2705         }
2706
2707         if (mddev->bitmap == NULL &&
2708             mddev->recovery_cp == MaxSector &&
2709             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2710             conf->fullsync == 0) {
2711                 *skipped = 1;
2712                 return max_sector - sector_nr;
2713         }
2714         /* before building a request, check if we can skip these blocks..
2715          * This call the bitmap_start_sync doesn't actually record anything
2716          */
2717         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2718             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2719                 /* We can skip this block, and probably several more */
2720                 *skipped = 1;
2721                 return sync_blocks;
2722         }
2723
2724         /*
2725          * If there is non-resync activity waiting for a turn, then let it
2726          * though before starting on this new sync request.
2727          */
2728         if (atomic_read(&conf->nr_waiting[idx]))
2729                 schedule_timeout_uninterruptible(1);
2730
2731         /* we are incrementing sector_nr below. To be safe, we check against
2732          * sector_nr + two times RESYNC_SECTORS
2733          */
2734
2735         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2736                 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2737
2738
2739         if (raise_barrier(conf, sector_nr))
2740                 return 0;
2741
2742         r1_bio = raid1_alloc_init_r1buf(conf);
2743
2744         rcu_read_lock();
2745         /*
2746          * If we get a correctably read error during resync or recovery,
2747          * we might want to read from a different device.  So we
2748          * flag all drives that could conceivably be read from for READ,
2749          * and any others (which will be non-In_sync devices) for WRITE.
2750          * If a read fails, we try reading from something else for which READ
2751          * is OK.
2752          */
2753
2754         r1_bio->mddev = mddev;
2755         r1_bio->sector = sector_nr;
2756         r1_bio->state = 0;
2757         set_bit(R1BIO_IsSync, &r1_bio->state);
2758         /* make sure good_sectors won't go across barrier unit boundary */
2759         good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2760
2761         for (i = 0; i < conf->raid_disks * 2; i++) {
2762                 struct md_rdev *rdev;
2763                 bio = r1_bio->bios[i];
2764
2765                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2766                 if (rdev == NULL ||
2767                     test_bit(Faulty, &rdev->flags)) {
2768                         if (i < conf->raid_disks)
2769                                 still_degraded = 1;
2770                 } else if (!test_bit(In_sync, &rdev->flags)) {
2771                         bio->bi_opf = REQ_OP_WRITE;
2772                         bio->bi_end_io = end_sync_write;
2773                         write_targets ++;
2774                 } else {
2775                         /* may need to read from here */
2776                         sector_t first_bad = MaxSector;
2777                         int bad_sectors;
2778
2779                         if (is_badblock(rdev, sector_nr, good_sectors,
2780                                         &first_bad, &bad_sectors)) {
2781                                 if (first_bad > sector_nr)
2782                                         good_sectors = first_bad - sector_nr;
2783                                 else {
2784                                         bad_sectors -= (sector_nr - first_bad);
2785                                         if (min_bad == 0 ||
2786                                             min_bad > bad_sectors)
2787                                                 min_bad = bad_sectors;
2788                                 }
2789                         }
2790                         if (sector_nr < first_bad) {
2791                                 if (test_bit(WriteMostly, &rdev->flags)) {
2792                                         if (wonly < 0)
2793                                                 wonly = i;
2794                                 } else {
2795                                         if (disk < 0)
2796                                                 disk = i;
2797                                 }
2798                                 bio->bi_opf = REQ_OP_READ;
2799                                 bio->bi_end_io = end_sync_read;
2800                                 read_targets++;
2801                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2802                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2803                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2804                                 /*
2805                                  * The device is suitable for reading (InSync),
2806                                  * but has bad block(s) here. Let's try to correct them,
2807                                  * if we are doing resync or repair. Otherwise, leave
2808                                  * this device alone for this sync request.
2809                                  */
2810                                 bio->bi_opf = REQ_OP_WRITE;
2811                                 bio->bi_end_io = end_sync_write;
2812                                 write_targets++;
2813                         }
2814                 }
2815                 if (rdev && bio->bi_end_io) {
2816                         atomic_inc(&rdev->nr_pending);
2817                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2818                         bio_set_dev(bio, rdev->bdev);
2819                         if (test_bit(FailFast, &rdev->flags))
2820                                 bio->bi_opf |= MD_FAILFAST;
2821                 }
2822         }
2823         rcu_read_unlock();
2824         if (disk < 0)
2825                 disk = wonly;
2826         r1_bio->read_disk = disk;
2827
2828         if (read_targets == 0 && min_bad > 0) {
2829                 /* These sectors are bad on all InSync devices, so we
2830                  * need to mark them bad on all write targets
2831                  */
2832                 int ok = 1;
2833                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2834                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2835                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2836                                 ok = rdev_set_badblocks(rdev, sector_nr,
2837                                                         min_bad, 0
2838                                         ) && ok;
2839                         }
2840                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2841                 *skipped = 1;
2842                 put_buf(r1_bio);
2843
2844                 if (!ok) {
2845                         /* Cannot record the badblocks, so need to
2846                          * abort the resync.
2847                          * If there are multiple read targets, could just
2848                          * fail the really bad ones ???
2849                          */
2850                         conf->recovery_disabled = mddev->recovery_disabled;
2851                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2852                         return 0;
2853                 } else
2854                         return min_bad;
2855
2856         }
2857         if (min_bad > 0 && min_bad < good_sectors) {
2858                 /* only resync enough to reach the next bad->good
2859                  * transition */
2860                 good_sectors = min_bad;
2861         }
2862
2863         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2864                 /* extra read targets are also write targets */
2865                 write_targets += read_targets-1;
2866
2867         if (write_targets == 0 || read_targets == 0) {
2868                 /* There is nowhere to write, so all non-sync
2869                  * drives must be failed - so we are finished
2870                  */
2871                 sector_t rv;
2872                 if (min_bad > 0)
2873                         max_sector = sector_nr + min_bad;
2874                 rv = max_sector - sector_nr;
2875                 *skipped = 1;
2876                 put_buf(r1_bio);
2877                 return rv;
2878         }
2879
2880         if (max_sector > mddev->resync_max)
2881                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2882         if (max_sector > sector_nr + good_sectors)
2883                 max_sector = sector_nr + good_sectors;
2884         nr_sectors = 0;
2885         sync_blocks = 0;
2886         do {
2887                 struct page *page;
2888                 int len = PAGE_SIZE;
2889                 if (sector_nr + (len>>9) > max_sector)
2890                         len = (max_sector - sector_nr) << 9;
2891                 if (len == 0)
2892                         break;
2893                 if (sync_blocks == 0) {
2894                         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2895                                                   &sync_blocks, still_degraded) &&
2896                             !conf->fullsync &&
2897                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2898                                 break;
2899                         if ((len >> 9) > sync_blocks)
2900                                 len = sync_blocks<<9;
2901                 }
2902
2903                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2904                         struct resync_pages *rp;
2905
2906                         bio = r1_bio->bios[i];
2907                         rp = get_resync_pages(bio);
2908                         if (bio->bi_end_io) {
2909                                 page = resync_fetch_page(rp, page_idx);
2910
2911                                 /*
2912                                  * won't fail because the vec table is big
2913                                  * enough to hold all these pages
2914                                  */
2915                                 __bio_add_page(bio, page, len, 0);
2916                         }
2917                 }
2918                 nr_sectors += len>>9;
2919                 sector_nr += len>>9;
2920                 sync_blocks -= (len>>9);
2921         } while (++page_idx < RESYNC_PAGES);
2922
2923         r1_bio->sectors = nr_sectors;
2924
2925         if (mddev_is_clustered(mddev) &&
2926                         conf->cluster_sync_high < sector_nr + nr_sectors) {
2927                 conf->cluster_sync_low = mddev->curr_resync_completed;
2928                 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2929                 /* Send resync message */
2930                 md_cluster_ops->resync_info_update(mddev,
2931                                 conf->cluster_sync_low,
2932                                 conf->cluster_sync_high);
2933         }
2934
2935         /* For a user-requested sync, we read all readable devices and do a
2936          * compare
2937          */
2938         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2939                 atomic_set(&r1_bio->remaining, read_targets);
2940                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2941                         bio = r1_bio->bios[i];
2942                         if (bio->bi_end_io == end_sync_read) {
2943                                 read_targets--;
2944                                 md_sync_acct_bio(bio, nr_sectors);
2945                                 if (read_targets == 1)
2946                                         bio->bi_opf &= ~MD_FAILFAST;
2947                                 submit_bio_noacct(bio);
2948                         }
2949                 }
2950         } else {
2951                 atomic_set(&r1_bio->remaining, 1);
2952                 bio = r1_bio->bios[r1_bio->read_disk];
2953                 md_sync_acct_bio(bio, nr_sectors);
2954                 if (read_targets == 1)
2955                         bio->bi_opf &= ~MD_FAILFAST;
2956                 submit_bio_noacct(bio);
2957         }
2958         return nr_sectors;
2959 }
2960
2961 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2962 {
2963         if (sectors)
2964                 return sectors;
2965
2966         return mddev->dev_sectors;
2967 }
2968
2969 static struct r1conf *setup_conf(struct mddev *mddev)
2970 {
2971         struct r1conf *conf;
2972         int i;
2973         struct raid1_info *disk;
2974         struct md_rdev *rdev;
2975         int err = -ENOMEM;
2976
2977         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2978         if (!conf)
2979                 goto abort;
2980
2981         conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2982                                    sizeof(atomic_t), GFP_KERNEL);
2983         if (!conf->nr_pending)
2984                 goto abort;
2985
2986         conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2987                                    sizeof(atomic_t), GFP_KERNEL);
2988         if (!conf->nr_waiting)
2989                 goto abort;
2990
2991         conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2992                                   sizeof(atomic_t), GFP_KERNEL);
2993         if (!conf->nr_queued)
2994                 goto abort;
2995
2996         conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2997                                 sizeof(atomic_t), GFP_KERNEL);
2998         if (!conf->barrier)
2999                 goto abort;
3000
3001         conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3002                                             mddev->raid_disks, 2),
3003                                 GFP_KERNEL);
3004         if (!conf->mirrors)
3005                 goto abort;
3006
3007         conf->tmppage = alloc_page(GFP_KERNEL);
3008         if (!conf->tmppage)
3009                 goto abort;
3010
3011         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
3012         if (!conf->poolinfo)
3013                 goto abort;
3014         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
3015         err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
3016                            rbio_pool_free, conf->poolinfo);
3017         if (err)
3018                 goto abort;
3019
3020         err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3021         if (err)
3022                 goto abort;
3023
3024         conf->poolinfo->mddev = mddev;
3025
3026         err = -EINVAL;
3027         spin_lock_init(&conf->device_lock);
3028         rdev_for_each(rdev, mddev) {
3029                 int disk_idx = rdev->raid_disk;
3030                 if (disk_idx >= mddev->raid_disks
3031                     || disk_idx < 0)
3032                         continue;
3033                 if (test_bit(Replacement, &rdev->flags))
3034                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
3035                 else
3036                         disk = conf->mirrors + disk_idx;
3037
3038                 if (disk->rdev)
3039                         goto abort;
3040                 disk->rdev = rdev;
3041                 disk->head_position = 0;
3042                 disk->seq_start = MaxSector;
3043         }
3044         conf->raid_disks = mddev->raid_disks;
3045         conf->mddev = mddev;
3046         INIT_LIST_HEAD(&conf->retry_list);
3047         INIT_LIST_HEAD(&conf->bio_end_io_list);
3048
3049         spin_lock_init(&conf->resync_lock);
3050         init_waitqueue_head(&conf->wait_barrier);
3051
3052         bio_list_init(&conf->pending_bio_list);
3053         conf->recovery_disabled = mddev->recovery_disabled - 1;
3054
3055         err = -EIO;
3056         for (i = 0; i < conf->raid_disks * 2; i++) {
3057
3058                 disk = conf->mirrors + i;
3059
3060                 if (i < conf->raid_disks &&
3061                     disk[conf->raid_disks].rdev) {
3062                         /* This slot has a replacement. */
3063                         if (!disk->rdev) {
3064                                 /* No original, just make the replacement
3065                                  * a recovering spare
3066                                  */
3067                                 disk->rdev =
3068                                         disk[conf->raid_disks].rdev;
3069                                 disk[conf->raid_disks].rdev = NULL;
3070                         } else if (!test_bit(In_sync, &disk->rdev->flags))
3071                                 /* Original is not in_sync - bad */
3072                                 goto abort;
3073                 }
3074
3075                 if (!disk->rdev ||
3076                     !test_bit(In_sync, &disk->rdev->flags)) {
3077                         disk->head_position = 0;
3078                         if (disk->rdev &&
3079                             (disk->rdev->saved_raid_disk < 0))
3080                                 conf->fullsync = 1;
3081                 }
3082         }
3083
3084         err = -ENOMEM;
3085         rcu_assign_pointer(conf->thread,
3086                            md_register_thread(raid1d, mddev, "raid1"));
3087         if (!conf->thread)
3088                 goto abort;
3089
3090         return conf;
3091
3092  abort:
3093         if (conf) {
3094                 mempool_exit(&conf->r1bio_pool);
3095                 kfree(conf->mirrors);
3096                 safe_put_page(conf->tmppage);
3097                 kfree(conf->poolinfo);
3098                 kfree(conf->nr_pending);
3099                 kfree(conf->nr_waiting);
3100                 kfree(conf->nr_queued);
3101                 kfree(conf->barrier);
3102                 bioset_exit(&conf->bio_split);
3103                 kfree(conf);
3104         }
3105         return ERR_PTR(err);
3106 }
3107
3108 static void raid1_free(struct mddev *mddev, void *priv);
3109 static int raid1_run(struct mddev *mddev)
3110 {
3111         struct r1conf *conf;
3112         int i;
3113         struct md_rdev *rdev;
3114         int ret;
3115
3116         if (mddev->level != 1) {
3117                 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3118                         mdname(mddev), mddev->level);
3119                 return -EIO;
3120         }
3121         if (mddev->reshape_position != MaxSector) {
3122                 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3123                         mdname(mddev));
3124                 return -EIO;
3125         }
3126         if (mddev_init_writes_pending(mddev) < 0)
3127                 return -ENOMEM;
3128         /*
3129          * copy the already verified devices into our private RAID1
3130          * bookkeeping area. [whatever we allocate in run(),
3131          * should be freed in raid1_free()]
3132          */
3133         if (mddev->private == NULL)
3134                 conf = setup_conf(mddev);
3135         else
3136                 conf = mddev->private;
3137
3138         if (IS_ERR(conf))
3139                 return PTR_ERR(conf);
3140
3141         if (mddev->queue)
3142                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3143
3144         rdev_for_each(rdev, mddev) {
3145                 if (!mddev->gendisk)
3146                         continue;
3147                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3148                                   rdev->data_offset << 9);
3149         }
3150
3151         mddev->degraded = 0;
3152         for (i = 0; i < conf->raid_disks; i++)
3153                 if (conf->mirrors[i].rdev == NULL ||
3154                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3155                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3156                         mddev->degraded++;
3157         /*
3158          * RAID1 needs at least one disk in active
3159          */
3160         if (conf->raid_disks - mddev->degraded < 1) {
3161                 md_unregister_thread(mddev, &conf->thread);
3162                 ret = -EINVAL;
3163                 goto abort;
3164         }
3165
3166         if (conf->raid_disks - mddev->degraded == 1)
3167                 mddev->recovery_cp = MaxSector;
3168
3169         if (mddev->recovery_cp != MaxSector)
3170                 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3171                         mdname(mddev));
3172         pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3173                 mdname(mddev), mddev->raid_disks - mddev->degraded,
3174                 mddev->raid_disks);
3175
3176         /*
3177          * Ok, everything is just fine now
3178          */
3179         rcu_assign_pointer(mddev->thread, conf->thread);
3180         rcu_assign_pointer(conf->thread, NULL);
3181         mddev->private = conf;
3182         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3183
3184         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3185
3186         ret = md_integrity_register(mddev);
3187         if (ret) {
3188                 md_unregister_thread(mddev, &mddev->thread);
3189                 goto abort;
3190         }
3191         return 0;
3192
3193 abort:
3194         raid1_free(mddev, conf);
3195         return ret;
3196 }
3197
3198 static void raid1_free(struct mddev *mddev, void *priv)
3199 {
3200         struct r1conf *conf = priv;
3201
3202         mempool_exit(&conf->r1bio_pool);
3203         kfree(conf->mirrors);
3204         safe_put_page(conf->tmppage);
3205         kfree(conf->poolinfo);
3206         kfree(conf->nr_pending);
3207         kfree(conf->nr_waiting);
3208         kfree(conf->nr_queued);
3209         kfree(conf->barrier);
3210         bioset_exit(&conf->bio_split);
3211         kfree(conf);
3212 }
3213
3214 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3215 {
3216         /* no resync is happening, and there is enough space
3217          * on all devices, so we can resize.
3218          * We need to make sure resync covers any new space.
3219          * If the array is shrinking we should possibly wait until
3220          * any io in the removed space completes, but it hardly seems
3221          * worth it.
3222          */
3223         sector_t newsize = raid1_size(mddev, sectors, 0);
3224         if (mddev->external_size &&
3225             mddev->array_sectors > newsize)
3226                 return -EINVAL;
3227         if (mddev->bitmap) {
3228                 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3229                 if (ret)
3230                         return ret;
3231         }
3232         md_set_array_sectors(mddev, newsize);
3233         if (sectors > mddev->dev_sectors &&
3234             mddev->recovery_cp > mddev->dev_sectors) {
3235                 mddev->recovery_cp = mddev->dev_sectors;
3236                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3237         }
3238         mddev->dev_sectors = sectors;
3239         mddev->resync_max_sectors = sectors;
3240         return 0;
3241 }
3242
3243 static int raid1_reshape(struct mddev *mddev)
3244 {
3245         /* We need to:
3246          * 1/ resize the r1bio_pool
3247          * 2/ resize conf->mirrors
3248          *
3249          * We allocate a new r1bio_pool if we can.
3250          * Then raise a device barrier and wait until all IO stops.
3251          * Then resize conf->mirrors and swap in the new r1bio pool.
3252          *
3253          * At the same time, we "pack" the devices so that all the missing
3254          * devices have the higher raid_disk numbers.
3255          */
3256         mempool_t newpool, oldpool;
3257         struct pool_info *newpoolinfo;
3258         struct raid1_info *newmirrors;
3259         struct r1conf *conf = mddev->private;
3260         int cnt, raid_disks;
3261         unsigned long flags;
3262         int d, d2;
3263         int ret;
3264
3265         memset(&newpool, 0, sizeof(newpool));
3266         memset(&oldpool, 0, sizeof(oldpool));
3267
3268         /* Cannot change chunk_size, layout, or level */
3269         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3270             mddev->layout != mddev->new_layout ||
3271             mddev->level != mddev->new_level) {
3272                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3273                 mddev->new_layout = mddev->layout;
3274                 mddev->new_level = mddev->level;
3275                 return -EINVAL;
3276         }
3277
3278         if (!mddev_is_clustered(mddev))
3279                 md_allow_write(mddev);
3280
3281         raid_disks = mddev->raid_disks + mddev->delta_disks;
3282
3283         if (raid_disks < conf->raid_disks) {
3284                 cnt=0;
3285                 for (d= 0; d < conf->raid_disks; d++)
3286                         if (conf->mirrors[d].rdev)
3287                                 cnt++;
3288                 if (cnt > raid_disks)
3289                         return -EBUSY;
3290         }
3291
3292         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3293         if (!newpoolinfo)
3294                 return -ENOMEM;
3295         newpoolinfo->mddev = mddev;
3296         newpoolinfo->raid_disks = raid_disks * 2;
3297
3298         ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3299                            rbio_pool_free, newpoolinfo);
3300         if (ret) {
3301                 kfree(newpoolinfo);
3302                 return ret;
3303         }
3304         newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3305                                          raid_disks, 2),
3306                              GFP_KERNEL);
3307         if (!newmirrors) {
3308                 kfree(newpoolinfo);
3309                 mempool_exit(&newpool);
3310                 return -ENOMEM;
3311         }
3312
3313         freeze_array(conf, 0);
3314
3315         /* ok, everything is stopped */
3316         oldpool = conf->r1bio_pool;
3317         conf->r1bio_pool = newpool;
3318
3319         for (d = d2 = 0; d < conf->raid_disks; d++) {
3320                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3321                 if (rdev && rdev->raid_disk != d2) {
3322                         sysfs_unlink_rdev(mddev, rdev);
3323                         rdev->raid_disk = d2;
3324                         sysfs_unlink_rdev(mddev, rdev);
3325                         if (sysfs_link_rdev(mddev, rdev))
3326                                 pr_warn("md/raid1:%s: cannot register rd%d\n",
3327                                         mdname(mddev), rdev->raid_disk);
3328                 }
3329                 if (rdev)
3330                         newmirrors[d2++].rdev = rdev;
3331         }
3332         kfree(conf->mirrors);
3333         conf->mirrors = newmirrors;
3334         kfree(conf->poolinfo);
3335         conf->poolinfo = newpoolinfo;
3336
3337         spin_lock_irqsave(&conf->device_lock, flags);
3338         mddev->degraded += (raid_disks - conf->raid_disks);
3339         spin_unlock_irqrestore(&conf->device_lock, flags);
3340         conf->raid_disks = mddev->raid_disks = raid_disks;
3341         mddev->delta_disks = 0;
3342
3343         unfreeze_array(conf);
3344
3345         set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3346         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3347         md_wakeup_thread(mddev->thread);
3348
3349         mempool_exit(&oldpool);
3350         return 0;
3351 }
3352
3353 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3354 {
3355         struct r1conf *conf = mddev->private;
3356
3357         if (quiesce)
3358                 freeze_array(conf, 0);
3359         else
3360                 unfreeze_array(conf);
3361 }
3362
3363 static void *raid1_takeover(struct mddev *mddev)
3364 {
3365         /* raid1 can take over:
3366          *  raid5 with 2 devices, any layout or chunk size
3367          */
3368         if (mddev->level == 5 && mddev->raid_disks == 2) {
3369                 struct r1conf *conf;
3370                 mddev->new_level = 1;
3371                 mddev->new_layout = 0;
3372                 mddev->new_chunk_sectors = 0;
3373                 conf = setup_conf(mddev);
3374                 if (!IS_ERR(conf)) {
3375                         /* Array must appear to be quiesced */
3376                         conf->array_frozen = 1;
3377                         mddev_clear_unsupported_flags(mddev,
3378                                 UNSUPPORTED_MDDEV_FLAGS);
3379                 }
3380                 return conf;
3381         }
3382         return ERR_PTR(-EINVAL);
3383 }
3384
3385 static struct md_personality raid1_personality =
3386 {
3387         .name           = "raid1",
3388         .level          = 1,
3389         .owner          = THIS_MODULE,
3390         .make_request   = raid1_make_request,
3391         .run            = raid1_run,
3392         .free           = raid1_free,
3393         .status         = raid1_status,
3394         .error_handler  = raid1_error,
3395         .hot_add_disk   = raid1_add_disk,
3396         .hot_remove_disk= raid1_remove_disk,
3397         .spare_active   = raid1_spare_active,
3398         .sync_request   = raid1_sync_request,
3399         .resize         = raid1_resize,
3400         .size           = raid1_size,
3401         .check_reshape  = raid1_reshape,
3402         .quiesce        = raid1_quiesce,
3403         .takeover       = raid1_takeover,
3404 };
3405
3406 static int __init raid_init(void)
3407 {
3408         return register_md_personality(&raid1_personality);
3409 }
3410
3411 static void raid_exit(void)
3412 {
3413         unregister_md_personality(&raid1_personality);
3414 }
3415
3416 module_init(raid_init);
3417 module_exit(raid_exit);
3418 MODULE_LICENSE("GPL");
3419 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3420 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3421 MODULE_ALIAS("md-raid1");
3422 MODULE_ALIAS("md-level-1");