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