drivers/md/raid1.c

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