drivers/md/raid1.c

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