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