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