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