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);
  70 static void lower_barrier(struct r1conf *conf);
  71
  72 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  73 {
  74         struct pool_info *pi = data;
  75         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
  76
  77         /* allocate a r1bio with room for raid_disks entries in the bios array */
  78         return kzalloc(size, gfp_flags);
  79 }
  80
  81 static void r1bio_pool_free(void *r1_bio, void *data)
  82 {
  83         kfree(r1_bio);
  84 }
  85
  86 #define RESYNC_BLOCK_SIZE (64*1024)
  87 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  88 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  89 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  90 #define RESYNC_WINDOW (2048*1024)
  91
  92 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  93 {
  94         struct pool_info *pi = data;
  95         struct page *page;
  96         struct r1bio *r1_bio;
  97         struct bio *bio;
  98         int i, j;
  99
 100         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 101         if (!r1_bio)
 102                 return NULL;
 103
 104         /*
 105          * Allocate bios : 1 for reading, n-1 for writing
 106          */
 107         for (j = pi->raid_disks ; j-- ; ) {
 108                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 109                 if (!bio)
 110                         goto out_free_bio;
 111                 r1_bio->bios[j] = bio;
 112         }
 113         /*
 114          * Allocate RESYNC_PAGES data pages and attach them to
 115          * the first bio.
 116          * If this is a user-requested check/repair, allocate
 117          * RESYNC_PAGES for each bio.
 118          */
 119         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 120                 j = pi->raid_disks;
 121         else
 122                 j = 1;
 123         while(j--) {
 124                 bio = r1_bio->bios[j];
 125                 for (i = 0; i < RESYNC_PAGES; i++) {
 126                         page = alloc_page(gfp_flags);
 127                         if (unlikely(!page))
 128                                 goto out_free_pages;
 129
 130                         bio->bi_io_vec[i].bv_page = page;
 131                         bio->bi_vcnt = i+1;
 132                 }
 133         }
 134         /* If not user-requests, copy the page pointers to all bios */
 135         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 136                 for (i=0; i<RESYNC_PAGES ; i++)
 137                         for (j=1; j<pi->raid_disks; j++)
 138                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 139                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 140         }
 141
 142         r1_bio->master_bio = NULL;
 143
 144         return r1_bio;
 145
 146 out_free_pages:
 147         for (j=0 ; j < pi->raid_disks; j++)
 148                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 149                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 150         j = -1;
 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
 239         if (bio->bi_phys_segments) {
 240                 unsigned long flags;
 241                 spin_lock_irqsave(&conf->device_lock, flags);
 242                 bio->bi_phys_segments--;
 243                 done = (bio->bi_phys_segments == 0);
 244                 spin_unlock_irqrestore(&conf->device_lock, flags);
 245         } else
 246                 done = 1;
 247
 248         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 249                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 250         if (done) {
 251                 bio_endio(bio, 0);
 252                 /*
 253                  * Wake up any possible resync thread that waits for the device
 254                  * to go idle.
 255                  */
 256                 allow_barrier(conf);
 257         }
 258 }
 259
 260 static void raid_end_bio_io(struct r1bio *r1_bio)
 261 {
 262         struct bio *bio = r1_bio->master_bio;
 263
 264         /* if nobody has done the final endio yet, do it now */
 265         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 266                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 267                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 268                          (unsigned long long) bio->bi_sector,
 269                          (unsigned long long) bio->bi_sector +
 270                          (bio->bi_size >> 9) - 1);
 271
 272                 call_bio_endio(r1_bio);
 273         }
 274         free_r1bio(r1_bio);
 275 }
 276
 277 /*
 278  * Update disk head position estimator based on IRQ completion info.
 279  */
 280 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 281 {
 282         struct r1conf *conf = r1_bio->mddev->private;
 283
 284         conf->mirrors[disk].head_position =
 285                 r1_bio->sector + (r1_bio->sectors);
 286 }
 287
 288 /*
 289  * Find the disk number which triggered given bio
 290  */
 291 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 292 {
 293         int mirror;
 294         struct r1conf *conf = r1_bio->mddev->private;
 295         int raid_disks = conf->raid_disks;
 296
 297         for (mirror = 0; mirror < raid_disks * 2; mirror++)
 298                 if (r1_bio->bios[mirror] == bio)
 299                         break;
 300
 301         BUG_ON(mirror == raid_disks * 2);
 302         update_head_pos(mirror, r1_bio);
 303
 304         return mirror;
 305 }
 306
 307 static void raid1_end_read_request(struct bio *bio, int error)
 308 {
 309         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 310         struct r1bio *r1_bio = bio->bi_private;
 311         int mirror;
 312         struct r1conf *conf = r1_bio->mddev->private;
 313
 314         mirror = r1_bio->read_disk;
 315         /*
 316          * this branch is our 'one mirror IO has finished' event handler:
 317          */
 318         update_head_pos(mirror, r1_bio);
 319
 320         if (uptodate)
 321                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 322         else {
 323                 /* If all other devices have failed, we want to return
 324                  * the error upwards rather than fail the last device.
 325                  * Here we redefine "uptodate" to mean "Don't want to retry"
 326                  */
 327                 unsigned long flags;
 328                 spin_lock_irqsave(&conf->device_lock, flags);
 329                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 330                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 331                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 332                         uptodate = 1;
 333                 spin_unlock_irqrestore(&conf->device_lock, flags);
 334         }
 335
 336         if (uptodate)
 337                 raid_end_bio_io(r1_bio);
 338         else {
 339                 /*
 340                  * oops, read error:
 341                  */
 342                 char b[BDEVNAME_SIZE];
 343                 printk_ratelimited(
 344                         KERN_ERR "md/raid1:%s: %s: "
 345                         "rescheduling sector %llu\n",
 346                         mdname(conf->mddev),
 347                         bdevname(conf->mirrors[mirror].rdev->bdev,
 348                                  b),
 349                         (unsigned long long)r1_bio->sector);
 350                 set_bit(R1BIO_ReadError, &r1_bio->state);
 351                 reschedule_retry(r1_bio);
 352         }
 353
 354         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 355 }
 356
 357 static void close_write(struct r1bio *r1_bio)
 358 {
 359         /* it really is the end of this request */
 360         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 361                 /* free extra copy of the data pages */
 362                 int i = r1_bio->behind_page_count;
 363                 while (i--)
 364                         safe_put_page(r1_bio->behind_bvecs[i].bv_page);
 365                 kfree(r1_bio->behind_bvecs);
 366                 r1_bio->behind_bvecs = NULL;
 367         }
 368         /* clear the bitmap if all writes complete successfully */
 369         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 370                         r1_bio->sectors,
 371                         !test_bit(R1BIO_Degraded, &r1_bio->state),
 372                         test_bit(R1BIO_BehindIO, &r1_bio->state));
 373         md_write_end(r1_bio->mddev);
 374 }
 375
 376 static void r1_bio_write_done(struct r1bio *r1_bio)
 377 {
 378         if (!atomic_dec_and_test(&r1_bio->remaining))
 379                 return;
 380
 381         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 382                 reschedule_retry(r1_bio);
 383         else {
 384                 close_write(r1_bio);
 385                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 386                         reschedule_retry(r1_bio);
 387                 else
 388                         raid_end_bio_io(r1_bio);
 389         }
 390 }
 391
 392 static void raid1_end_write_request(struct bio *bio, int error)
 393 {
 394         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 395         struct r1bio *r1_bio = bio->bi_private;
 396         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 397         struct r1conf *conf = r1_bio->mddev->private;
 398         struct bio *to_put = NULL;
 399
 400         mirror = find_bio_disk(r1_bio, bio);
 401
 402         /*
 403          * 'one mirror IO has finished' event handler:
 404          */
 405         if (!uptodate) {
 406                 set_bit(WriteErrorSeen,
 407                         &conf->mirrors[mirror].rdev->flags);
 408                 if (!test_and_set_bit(WantReplacement,
 409                                       &conf->mirrors[mirror].rdev->flags))
 410                         set_bit(MD_RECOVERY_NEEDED, &
 411                                 conf->mddev->recovery);
 412
 413                 set_bit(R1BIO_WriteError, &r1_bio->state);
 414         } else {
 415                 /*
 416                  * Set R1BIO_Uptodate in our master bio, so that we
 417                  * will return a good error code for to the higher
 418                  * levels even if IO on some other mirrored buffer
 419                  * fails.
 420                  *
 421                  * The 'master' represents the composite IO operation
 422                  * to user-side. So if something waits for IO, then it
 423                  * will wait for the 'master' bio.
 424                  */
 425                 sector_t first_bad;
 426                 int bad_sectors;
 427
 428                 r1_bio->bios[mirror] = NULL;
 429                 to_put = bio;
 430                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 431
 432                 /* Maybe we can clear some bad blocks. */
 433                 if (is_badblock(conf->mirrors[mirror].rdev,
 434                                 r1_bio->sector, r1_bio->sectors,
 435                                 &first_bad, &bad_sectors)) {
 436                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 437                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 438                 }
 439         }
 440
 441         if (behind) {
 442                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 443                         atomic_dec(&r1_bio->behind_remaining);
 444
 445                 /*
 446                  * In behind mode, we ACK the master bio once the I/O
 447                  * has safely reached all non-writemostly
 448                  * disks. Setting the Returned bit ensures that this
 449                  * gets done only once -- we don't ever want to return
 450                  * -EIO here, instead we'll wait
 451                  */
 452                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 453                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 454                         /* Maybe we can return now */
 455                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 456                                 struct bio *mbio = r1_bio->master_bio;
 457                                 pr_debug("raid1: behind end write sectors"
 458                                          " %llu-%llu\n",
 459                                          (unsigned long long) mbio->bi_sector,
 460                                          (unsigned long long) mbio->bi_sector +
 461                                          (mbio->bi_size >> 9) - 1);
 462                                 call_bio_endio(r1_bio);
 463                         }
 464                 }
 465         }
 466         if (r1_bio->bios[mirror] == NULL)
 467                 rdev_dec_pending(conf->mirrors[mirror].rdev,
 468                                  conf->mddev);
 469
 470         /*
 471          * Let's see if all mirrored write operations have finished
 472          * already.
 473          */
 474         r1_bio_write_done(r1_bio);
 475
 476         if (to_put)
 477                 bio_put(to_put);
 478 }
 479
 480
 481 /*
 482  * This routine returns the disk from which the requested read should
 483  * be done. There is a per-array 'next expected sequential IO' sector
 484  * number - if this matches on the next IO then we use the last disk.
 485  * There is also a per-disk 'last know head position' sector that is
 486  * maintained from IRQ contexts, both the normal and the resync IO
 487  * completion handlers update this position correctly. If there is no
 488  * perfect sequential match then we pick the disk whose head is closest.
 489  *
 490  * If there are 2 mirrors in the same 2 devices, performance degrades
 491  * because position is mirror, not device based.
 492  *
 493  * The rdev for the device selected will have nr_pending incremented.
 494  */
 495 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 496 {
 497         const sector_t this_sector = r1_bio->sector;
 498         int sectors;
 499         int best_good_sectors;
 500         int best_disk, best_dist_disk, best_pending_disk;
 501         int has_nonrot_disk;
 502         int disk;
 503         sector_t best_dist;
 504         unsigned int min_pending;
 505         struct md_rdev *rdev;
 506         int choose_first;
 507         int choose_next_idle;
 508
 509         rcu_read_lock();
 510         /*
 511          * Check if we can balance. We can balance on the whole
 512          * device if no resync is going on, or below the resync window.
 513          * We take the first readable disk when above the resync window.
 514          */
 515  retry:
 516         sectors = r1_bio->sectors;
 517         best_disk = -1;
 518         best_dist_disk = -1;
 519         best_dist = MaxSector;
 520         best_pending_disk = -1;
 521         min_pending = UINT_MAX;
 522         best_good_sectors = 0;
 523         has_nonrot_disk = 0;
 524         choose_next_idle = 0;
 525
 526         if (conf->mddev->recovery_cp < MaxSector &&
 527             (this_sector + sectors >= conf->next_resync))
 528                 choose_first = 1;
 529         else
 530                 choose_first = 0;
 531
 532         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 533                 sector_t dist;
 534                 sector_t first_bad;
 535                 int bad_sectors;
 536                 unsigned int pending;
 537                 bool nonrot;
 538
 539                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 540                 if (r1_bio->bios[disk] == IO_BLOCKED
 541                     || rdev == NULL
 542                     || test_bit(Unmerged, &rdev->flags)
 543                     || test_bit(Faulty, &rdev->flags))
 544                         continue;
 545                 if (!test_bit(In_sync, &rdev->flags) &&
 546                     rdev->recovery_offset < this_sector + sectors)
 547                         continue;
 548                 if (test_bit(WriteMostly, &rdev->flags)) {
 549                         /* Don't balance among write-mostly, just
 550                          * use the first as a last resort */
 551                         if (best_disk < 0) {
 552                                 if (is_badblock(rdev, this_sector, sectors,
 553                                                 &first_bad, &bad_sectors)) {
 554                                         if (first_bad < this_sector)
 555                                                 /* Cannot use this */
 556                                                 continue;
 557                                         best_good_sectors = first_bad - this_sector;
 558                                 } else
 559                                         best_good_sectors = sectors;
 560                                 best_disk = disk;
 561                         }
 562                         continue;
 563                 }
 564                 /* This is a reasonable device to use.  It might
 565                  * even be best.
 566                  */
 567                 if (is_badblock(rdev, this_sector, sectors,
 568                                 &first_bad, &bad_sectors)) {
 569                         if (best_dist < MaxSector)
 570                                 /* already have a better device */
 571                                 continue;
 572                         if (first_bad <= this_sector) {
 573                                 /* cannot read here. If this is the 'primary'
 574                                  * device, then we must not read beyond
 575                                  * bad_sectors from another device..
 576                                  */
 577                                 bad_sectors -= (this_sector - first_bad);
 578                                 if (choose_first && sectors > bad_sectors)
 579                                         sectors = bad_sectors;
 580                                 if (best_good_sectors > sectors)
 581                                         best_good_sectors = sectors;
 582
 583                         } else {
 584                                 sector_t good_sectors = first_bad - this_sector;
 585                                 if (good_sectors > best_good_sectors) {
 586                                         best_good_sectors = good_sectors;
 587                                         best_disk = disk;
 588                                 }
 589                                 if (choose_first)
 590                                         break;
 591                         }
 592                         continue;
 593                 } else
 594                         best_good_sectors = sectors;
 595
 596                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 597                 has_nonrot_disk |= nonrot;
 598                 pending = atomic_read(&rdev->nr_pending);
 599                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 600                 if (choose_first) {
 601                         best_disk = disk;
 602                         break;
 603                 }
 604                 /* Don't change to another disk for sequential reads */
 605                 if (conf->mirrors[disk].next_seq_sect == this_sector
 606                     || dist == 0) {
 607                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 608                         struct raid1_info *mirror = &conf->mirrors[disk];
 609
 610                         best_disk = disk;
 611                         /*
 612                          * If buffered sequential IO size exceeds optimal
 613                          * iosize, check if there is idle disk. If yes, choose
 614                          * the idle disk. read_balance could already choose an
 615                          * idle disk before noticing it's a sequential IO in
 616                          * this disk. This doesn't matter because this disk
 617                          * will idle, next time it will be utilized after the
 618                          * first disk has IO size exceeds optimal iosize. In
 619                          * this way, iosize of the first disk will be optimal
 620                          * iosize at least. iosize of the second disk might be
 621                          * small, but not a big deal since when the second disk
 622                          * starts IO, the first disk is likely still busy.
 623                          */
 624                         if (nonrot && opt_iosize > 0 &&
 625                             mirror->seq_start != MaxSector &&
 626                             mirror->next_seq_sect > opt_iosize &&
 627                             mirror->next_seq_sect - opt_iosize >=
 628                             mirror->seq_start) {
 629                                 choose_next_idle = 1;
 630                                 continue;
 631                         }
 632                         break;
 633                 }
 634                 /* If device is idle, use it */
 635                 if (pending == 0) {
 636                         best_disk = disk;
 637                         break;
 638                 }
 639
 640                 if (choose_next_idle)
 641                         continue;
 642
 643                 if (min_pending > pending) {
 644                         min_pending = pending;
 645                         best_pending_disk = disk;
 646                 }
 647
 648                 if (dist < best_dist) {
 649                         best_dist = dist;
 650                         best_dist_disk = disk;
 651                 }
 652         }
 653
 654         /*
 655          * If all disks are rotational, choose the closest disk. If any disk is
 656          * non-rotational, choose the disk with less pending request even the
 657          * disk is rotational, which might/might not be optimal for raids with
 658          * mixed ratation/non-rotational disks depending on workload.
 659          */
 660         if (best_disk == -1) {
 661                 if (has_nonrot_disk)
 662                         best_disk = best_pending_disk;
 663                 else
 664                         best_disk = best_dist_disk;
 665         }
 666
 667         if (best_disk >= 0) {
 668                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 669                 if (!rdev)
 670                         goto retry;
 671                 atomic_inc(&rdev->nr_pending);
 672                 if (test_bit(Faulty, &rdev->flags)) {
 673                         /* cannot risk returning a device that failed
 674                          * before we inc'ed nr_pending
 675                          */
 676                         rdev_dec_pending(rdev, conf->mddev);
 677                         goto retry;
 678                 }
 679                 sectors = best_good_sectors;
 680
 681                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 682                         conf->mirrors[best_disk].seq_start = this_sector;
 683
 684                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 685         }
 686         rcu_read_unlock();
 687         *max_sectors = sectors;
 688
 689         return best_disk;
 690 }
 691
 692 static int raid1_mergeable_bvec(struct request_queue *q,
 693                                 struct bvec_merge_data *bvm,
 694                                 struct bio_vec *biovec)
 695 {
 696         struct mddev *mddev = q->queuedata;
 697         struct r1conf *conf = mddev->private;
 698         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 699         int max = biovec->bv_len;
 700
 701         if (mddev->merge_check_needed) {
 702                 int disk;
 703                 rcu_read_lock();
 704                 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
 705                         struct md_rdev *rdev = rcu_dereference(
 706                                 conf->mirrors[disk].rdev);
 707                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
 708                                 struct request_queue *q =
 709                                         bdev_get_queue(rdev->bdev);
 710                                 if (q->merge_bvec_fn) {
 711                                         bvm->bi_sector = sector +
 712                                                 rdev->data_offset;
 713                                         bvm->bi_bdev = rdev->bdev;
 714                                         max = min(max, q->merge_bvec_fn(
 715                                                           q, bvm, biovec));
 716                                 }
 717                         }
 718                 }
 719                 rcu_read_unlock();
 720         }
 721         return max;
 722
 723 }
 724
 725 int md_raid1_congested(struct mddev *mddev, int bits)
 726 {
 727         struct r1conf *conf = mddev->private;
 728         int i, ret = 0;
 729
 730         if ((bits & (1 << BDI_async_congested)) &&
 731             conf->pending_count >= max_queued_requests)
 732                 return 1;
 733
 734         rcu_read_lock();
 735         for (i = 0; i < conf->raid_disks * 2; i++) {
 736                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 737                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 738                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 739
 740                         BUG_ON(!q);
 741
 742                         /* Note the '|| 1' - when read_balance prefers
 743                          * non-congested targets, it can be removed
 744                          */
 745                         if ((bits & (1<<BDI_async_congested)) || 1)
 746                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 747                         else
 748                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 749                 }
 750         }
 751         rcu_read_unlock();
 752         return ret;
 753 }
 754 EXPORT_SYMBOL_GPL(md_raid1_congested);
 755
 756 static int raid1_congested(void *data, int bits)
 757 {
 758         struct mddev *mddev = data;
 759
 760         return mddev_congested(mddev, bits) ||
 761                 md_raid1_congested(mddev, bits);
 762 }
 763
 764 static void flush_pending_writes(struct r1conf *conf)
 765 {
 766         /* Any writes that have been queued but are awaiting
 767          * bitmap updates get flushed here.
 768          */
 769         spin_lock_irq(&conf->device_lock);
 770
 771         if (conf->pending_bio_list.head) {
 772                 struct bio *bio;
 773                 bio = bio_list_get(&conf->pending_bio_list);
 774                 conf->pending_count = 0;
 775                 spin_unlock_irq(&conf->device_lock);
 776                 /* flush any pending bitmap writes to
 777                  * disk before proceeding w/ I/O */
 778                 bitmap_unplug(conf->mddev->bitmap);
 779                 wake_up(&conf->wait_barrier);
 780
 781                 while (bio) { /* submit pending writes */
 782                         struct bio *next = bio->bi_next;
 783                         bio->bi_next = NULL;
 784                         generic_make_request(bio);
 785                         bio = next;
 786                 }
 787         } else
 788                 spin_unlock_irq(&conf->device_lock);
 789 }
 790
 791 /* Barriers....
 792  * Sometimes we need to suspend IO while we do something else,
 793  * either some resync/recovery, or reconfigure the array.
 794  * To do this we raise a 'barrier'.
 795  * The 'barrier' is a counter that can be raised multiple times
 796  * to count how many activities are happening which preclude
 797  * normal IO.
 798  * We can only raise the barrier if there is no pending IO.
 799  * i.e. if nr_pending == 0.
 800  * We choose only to raise the barrier if no-one is waiting for the
 801  * barrier to go down.  This means that as soon as an IO request
 802  * is ready, no other operations which require a barrier will start
 803  * until the IO request has had a chance.
 804  *
 805  * So: regular IO calls 'wait_barrier'.  When that returns there
 806  *    is no backgroup IO happening,  It must arrange to call
 807  *    allow_barrier when it has finished its IO.
 808  * backgroup IO calls must call raise_barrier.  Once that returns
 809  *    there is no normal IO happeing.  It must arrange to call
 810  *    lower_barrier when the particular background IO completes.
 811  */
 812 #define RESYNC_DEPTH 32
 813
 814 static void raise_barrier(struct r1conf *conf)
 815 {
 816         spin_lock_irq(&conf->resync_lock);
 817
 818         /* Wait until no block IO is waiting */
 819         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 820                             conf->resync_lock, );
 821
 822         /* block any new IO from starting */
 823         conf->barrier++;
 824
 825         /* Now wait for all pending IO to complete */
 826         wait_event_lock_irq(conf->wait_barrier,
 827                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 828                             conf->resync_lock, );
 829
 830         spin_unlock_irq(&conf->resync_lock);
 831 }
 832
 833 static void lower_barrier(struct r1conf *conf)
 834 {
 835         unsigned long flags;
 836         BUG_ON(conf->barrier <= 0);
 837         spin_lock_irqsave(&conf->resync_lock, flags);
 838         conf->barrier--;
 839         spin_unlock_irqrestore(&conf->resync_lock, flags);
 840         wake_up(&conf->wait_barrier);
 841 }
 842
 843 static void wait_barrier(struct r1conf *conf)
 844 {
 845         spin_lock_irq(&conf->resync_lock);
 846         if (conf->barrier) {
 847                 conf->nr_waiting++;
 848                 /* Wait for the barrier to drop.
 849                  * However if there are already pending
 850                  * requests (preventing the barrier from
 851                  * rising completely), and the
 852                  * pre-process bio queue isn't empty,
 853                  * then don't wait, as we need to empty
 854                  * that queue to get the nr_pending
 855                  * count down.
 856                  */
 857                 wait_event_lock_irq(conf->wait_barrier,
 858                                     !conf->barrier ||
 859                                     (conf->nr_pending &&
 860                                      current->bio_list &&
 861                                      !bio_list_empty(current->bio_list)),
 862                                     conf->resync_lock,
 863                         );
 864                 conf->nr_waiting--;
 865         }
 866         conf->nr_pending++;
 867         spin_unlock_irq(&conf->resync_lock);
 868 }
 869
 870 static void allow_barrier(struct r1conf *conf)
 871 {
 872         unsigned long flags;
 873         spin_lock_irqsave(&conf->resync_lock, flags);
 874         conf->nr_pending--;
 875         spin_unlock_irqrestore(&conf->resync_lock, flags);
 876         wake_up(&conf->wait_barrier);
 877 }
 878
 879 static void freeze_array(struct r1conf *conf)
 880 {
 881         /* stop syncio and normal IO and wait for everything to
 882          * go quite.
 883          * We increment barrier and nr_waiting, and then
 884          * wait until nr_pending match nr_queued+1
 885          * This is called in the context of one normal IO request
 886          * that has failed. Thus any sync request that might be pending
 887          * will be blocked by nr_pending, and we need to wait for
 888          * pending IO requests to complete or be queued for re-try.
 889          * Thus the number queued (nr_queued) plus this request (1)
 890          * must match the number of pending IOs (nr_pending) before
 891          * we continue.
 892          */
 893         spin_lock_irq(&conf->resync_lock);
 894         conf->barrier++;
 895         conf->nr_waiting++;
 896         wait_event_lock_irq(conf->wait_barrier,
 897                             conf->nr_pending == conf->nr_queued+1,
 898                             conf->resync_lock,
 899                             flush_pending_writes(conf));
 900         spin_unlock_irq(&conf->resync_lock);
 901 }
 902 static void unfreeze_array(struct r1conf *conf)
 903 {
 904         /* reverse the effect of the freeze */
 905         spin_lock_irq(&conf->resync_lock);
 906         conf->barrier--;
 907         conf->nr_waiting--;
 908         wake_up(&conf->wait_barrier);
 909         spin_unlock_irq(&conf->resync_lock);
 910 }
 911
 912
 913 /* duplicate the data pages for behind I/O
 914  */
 915 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
 916 {
 917         int i;
 918         struct bio_vec *bvec;
 919         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
 920                                         GFP_NOIO);
 921         if (unlikely(!bvecs))
 922                 return;
 923
 924         bio_for_each_segment(bvec, bio, i) {
 925                 bvecs[i] = *bvec;
 926                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
 927                 if (unlikely(!bvecs[i].bv_page))
 928                         goto do_sync_io;
 929                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
 930                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 931                 kunmap(bvecs[i].bv_page);
 932                 kunmap(bvec->bv_page);
 933         }
 934         r1_bio->behind_bvecs = bvecs;
 935         r1_bio->behind_page_count = bio->bi_vcnt;
 936         set_bit(R1BIO_BehindIO, &r1_bio->state);
 937         return;
 938
 939 do_sync_io:
 940         for (i = 0; i < bio->bi_vcnt; i++)
 941                 if (bvecs[i].bv_page)
 942                         put_page(bvecs[i].bv_page);
 943         kfree(bvecs);
 944         pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 945 }
 946
 947 struct raid1_plug_cb {
 948         struct blk_plug_cb      cb;
 949         struct bio_list         pending;
 950         int                     pending_cnt;
 951 };
 952
 953 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
 954 {
 955         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
 956                                                   cb);
 957         struct mddev *mddev = plug->cb.data;
 958         struct r1conf *conf = mddev->private;
 959         struct bio *bio;
 960
 961         if (from_schedule) {
 962                 spin_lock_irq(&conf->device_lock);
 963                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
 964                 conf->pending_count += plug->pending_cnt;
 965                 spin_unlock_irq(&conf->device_lock);
 966                 md_wakeup_thread(mddev->thread);
 967                 kfree(plug);
 968                 return;
 969         }
 970
 971         /* we aren't scheduling, so we can do the write-out directly. */
 972         bio = bio_list_get(&plug->pending);
 973         bitmap_unplug(mddev->bitmap);
 974         wake_up(&conf->wait_barrier);
 975
 976         while (bio) { /* submit pending writes */
 977                 struct bio *next = bio->bi_next;
 978                 bio->bi_next = NULL;
 979                 generic_make_request(bio);
 980                 bio = next;
 981         }
 982         kfree(plug);
 983 }
 984
 985 static void make_request(struct mddev *mddev, struct bio * bio)
 986 {
 987         struct r1conf *conf = mddev->private;
 988         struct raid1_info *mirror;
 989         struct r1bio *r1_bio;
 990         struct bio *read_bio;
 991         int i, disks;
 992         struct bitmap *bitmap;
 993         unsigned long flags;
 994         const int rw = bio_data_dir(bio);
 995         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 996         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
 997         struct md_rdev *blocked_rdev;
 998         struct blk_plug_cb *cb;
 999         struct raid1_plug_cb *plug = NULL;
1000         int first_clone;
1001         int sectors_handled;
1002         int max_sectors;
1003
1004         /*
1005          * Register the new request and wait if the reconstruction
1006          * thread has put up a bar for new requests.
1007          * Continue immediately if no resync is active currently.
1008          */
1009
1010         md_write_start(mddev, bio); /* wait on superblock update early */
1011
1012         if (bio_data_dir(bio) == WRITE &&
1013             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
1014             bio->bi_sector < mddev->suspend_hi) {
1015                 /* As the suspend_* range is controlled by
1016                  * userspace, we want an interruptible
1017                  * wait.
1018                  */
1019                 DEFINE_WAIT(w);
1020                 for (;;) {
1021                         flush_signals(current);
1022                         prepare_to_wait(&conf->wait_barrier,
1023                                         &w, TASK_INTERRUPTIBLE);
1024                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
1025                             bio->bi_sector >= mddev->suspend_hi)
1026                                 break;
1027                         schedule();
1028                 }
1029                 finish_wait(&conf->wait_barrier, &w);
1030         }
1031
1032         wait_barrier(conf);
1033
1034         bitmap = mddev->bitmap;
1035
1036         /*
1037          * make_request() can abort the operation when READA is being
1038          * used and no empty request is available.
1039          *
1040          */
1041         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1042
1043         r1_bio->master_bio = bio;
1044         r1_bio->sectors = bio->bi_size >> 9;
1045         r1_bio->state = 0;
1046         r1_bio->mddev = mddev;
1047         r1_bio->sector = bio->bi_sector;
1048
1049         /* We might need to issue multiple reads to different
1050          * devices if there are bad blocks around, so we keep
1051          * track of the number of reads in bio->bi_phys_segments.
1052          * If this is 0, there is only one r1_bio and no locking
1053          * will be needed when requests complete.  If it is
1054          * non-zero, then it is the number of not-completed requests.
1055          */
1056         bio->bi_phys_segments = 0;
1057         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1058
1059         if (rw == READ) {
1060                 /*
1061                  * read balancing logic:
1062                  */
1063                 int rdisk;
1064
1065 read_again:
1066                 rdisk = read_balance(conf, r1_bio, &max_sectors);
1067
1068                 if (rdisk < 0) {
1069                         /* couldn't find anywhere to read from */
1070                         raid_end_bio_io(r1_bio);
1071                         return;
1072                 }
1073                 mirror = conf->mirrors + rdisk;
1074
1075                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1076                     bitmap) {
1077                         /* Reading from a write-mostly device must
1078                          * take care not to over-take any writes
1079                          * that are 'behind'
1080                          */
1081                         wait_event(bitmap->behind_wait,
1082                                    atomic_read(&bitmap->behind_writes) == 0);
1083                 }
1084                 r1_bio->read_disk = rdisk;
1085
1086                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1087                 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
1088                             max_sectors);
1089
1090                 r1_bio->bios[rdisk] = read_bio;
1091
1092                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
1093                 read_bio->bi_bdev = mirror->rdev->bdev;
1094                 read_bio->bi_end_io = raid1_end_read_request;
1095                 read_bio->bi_rw = READ | do_sync;
1096                 read_bio->bi_private = r1_bio;
1097
1098                 if (max_sectors < r1_bio->sectors) {
1099                         /* could not read all from this device, so we will
1100                          * need another r1_bio.
1101                          */
1102
1103                         sectors_handled = (r1_bio->sector + max_sectors
1104                                            - bio->bi_sector);
1105                         r1_bio->sectors = max_sectors;
1106                         spin_lock_irq(&conf->device_lock);
1107                         if (bio->bi_phys_segments == 0)
1108                                 bio->bi_phys_segments = 2;
1109                         else
1110                                 bio->bi_phys_segments++;
1111                         spin_unlock_irq(&conf->device_lock);
1112                         /* Cannot call generic_make_request directly
1113                          * as that will be queued in __make_request
1114                          * and subsequent mempool_alloc might block waiting
1115                          * for it.  So hand bio over to raid1d.
1116                          */
1117                         reschedule_retry(r1_bio);
1118
1119                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1120
1121                         r1_bio->master_bio = bio;
1122                         r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1123                         r1_bio->state = 0;
1124                         r1_bio->mddev = mddev;
1125                         r1_bio->sector = bio->bi_sector + sectors_handled;
1126                         goto read_again;
1127                 } else
1128                         generic_make_request(read_bio);
1129                 return;
1130         }
1131
1132         /*
1133          * WRITE:
1134          */
1135         if (conf->pending_count >= max_queued_requests) {
1136                 md_wakeup_thread(mddev->thread);
1137                 wait_event(conf->wait_barrier,
1138                            conf->pending_count < max_queued_requests);
1139         }
1140         /* first select target devices under rcu_lock and
1141          * inc refcount on their rdev.  Record them by setting
1142          * bios[x] to bio
1143          * If there are known/acknowledged bad blocks on any device on
1144          * which we have seen a write error, we want to avoid writing those
1145          * blocks.
1146          * This potentially requires several writes to write around
1147          * the bad blocks.  Each set of writes gets it's own r1bio
1148          * with a set of bios attached.
1149          */
1150
1151         disks = conf->raid_disks * 2;
1152  retry_write:
1153         blocked_rdev = NULL;
1154         rcu_read_lock();
1155         max_sectors = r1_bio->sectors;
1156         for (i = 0;  i < disks; i++) {
1157                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1158                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1159                         atomic_inc(&rdev->nr_pending);
1160                         blocked_rdev = rdev;
1161                         break;
1162                 }
1163                 r1_bio->bios[i] = NULL;
1164                 if (!rdev || test_bit(Faulty, &rdev->flags)
1165                     || test_bit(Unmerged, &rdev->flags)) {
1166                         if (i < conf->raid_disks)
1167                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1168                         continue;
1169                 }
1170
1171                 atomic_inc(&rdev->nr_pending);
1172                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1173                         sector_t first_bad;
1174                         int bad_sectors;
1175                         int is_bad;
1176
1177                         is_bad = is_badblock(rdev, r1_bio->sector,
1178                                              max_sectors,
1179                                              &first_bad, &bad_sectors);
1180                         if (is_bad < 0) {
1181                                 /* mustn't write here until the bad block is
1182                                  * acknowledged*/
1183                                 set_bit(BlockedBadBlocks, &rdev->flags);
1184                                 blocked_rdev = rdev;
1185                                 break;
1186                         }
1187                         if (is_bad && first_bad <= r1_bio->sector) {
1188                                 /* Cannot write here at all */
1189                                 bad_sectors -= (r1_bio->sector - first_bad);
1190                                 if (bad_sectors < max_sectors)
1191                                         /* mustn't write more than bad_sectors
1192                                          * to other devices yet
1193                                          */
1194                                         max_sectors = bad_sectors;
1195                                 rdev_dec_pending(rdev, mddev);
1196                                 /* We don't set R1BIO_Degraded as that
1197                                  * only applies if the disk is
1198                                  * missing, so it might be re-added,
1199                                  * and we want to know to recover this
1200                                  * chunk.
1201                                  * In this case the device is here,
1202                                  * and the fact that this chunk is not
1203                                  * in-sync is recorded in the bad
1204                                  * block log
1205                                  */
1206                                 continue;
1207                         }
1208                         if (is_bad) {
1209                                 int good_sectors = first_bad - r1_bio->sector;
1210                                 if (good_sectors < max_sectors)
1211                                         max_sectors = good_sectors;
1212                         }
1213                 }
1214                 r1_bio->bios[i] = bio;
1215         }
1216         rcu_read_unlock();
1217
1218         if (unlikely(blocked_rdev)) {
1219                 /* Wait for this device to become unblocked */
1220                 int j;
1221
1222                 for (j = 0; j < i; j++)
1223                         if (r1_bio->bios[j])
1224                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1225                 r1_bio->state = 0;
1226                 allow_barrier(conf);
1227                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1228                 wait_barrier(conf);
1229                 goto retry_write;
1230         }
1231
1232         if (max_sectors < r1_bio->sectors) {
1233                 /* We are splitting this write into multiple parts, so
1234                  * we need to prepare for allocating another r1_bio.
1235                  */
1236                 r1_bio->sectors = max_sectors;
1237                 spin_lock_irq(&conf->device_lock);
1238                 if (bio->bi_phys_segments == 0)
1239                         bio->bi_phys_segments = 2;
1240                 else
1241                         bio->bi_phys_segments++;
1242                 spin_unlock_irq(&conf->device_lock);
1243         }
1244         sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1245
1246         atomic_set(&r1_bio->remaining, 1);
1247         atomic_set(&r1_bio->behind_remaining, 0);
1248
1249         first_clone = 1;
1250         for (i = 0; i < disks; i++) {
1251                 struct bio *mbio;
1252                 if (!r1_bio->bios[i])
1253                         continue;
1254
1255                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1256                 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1257
1258                 if (first_clone) {
1259                         /* do behind I/O ?
1260                          * Not if there are too many, or cannot
1261                          * allocate memory, or a reader on WriteMostly
1262                          * is waiting for behind writes to flush */
1263                         if (bitmap &&
1264                             (atomic_read(&bitmap->behind_writes)
1265                              < mddev->bitmap_info.max_write_behind) &&
1266                             !waitqueue_active(&bitmap->behind_wait))
1267                                 alloc_behind_pages(mbio, r1_bio);
1268
1269                         bitmap_startwrite(bitmap, r1_bio->sector,
1270                                           r1_bio->sectors,
1271                                           test_bit(R1BIO_BehindIO,
1272                                                    &r1_bio->state));
1273                         first_clone = 0;
1274                 }
1275                 if (r1_bio->behind_bvecs) {
1276                         struct bio_vec *bvec;
1277                         int j;
1278
1279                         /* Yes, I really want the '__' version so that
1280                          * we clear any unused pointer in the io_vec, rather
1281                          * than leave them unchanged.  This is important
1282                          * because when we come to free the pages, we won't
1283                          * know the original bi_idx, so we just free
1284                          * them all
1285                          */
1286                         __bio_for_each_segment(bvec, mbio, j, 0)
1287                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1288                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1289                                 atomic_inc(&r1_bio->behind_remaining);
1290                 }
1291
1292                 r1_bio->bios[i] = mbio;
1293
1294                 mbio->bi_sector = (r1_bio->sector +
1295                                    conf->mirrors[i].rdev->data_offset);
1296                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1297                 mbio->bi_end_io = raid1_end_write_request;
1298                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
1299                 mbio->bi_private = r1_bio;
1300
1301                 atomic_inc(&r1_bio->remaining);
1302
1303                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1304                 if (cb)
1305                         plug = container_of(cb, struct raid1_plug_cb, cb);
1306                 else
1307                         plug = NULL;
1308                 spin_lock_irqsave(&conf->device_lock, flags);
1309                 if (plug) {
1310                         bio_list_add(&plug->pending, mbio);
1311                         plug->pending_cnt++;
1312                 } else {
1313                         bio_list_add(&conf->pending_bio_list, mbio);
1314                         conf->pending_count++;
1315                 }
1316                 spin_unlock_irqrestore(&conf->device_lock, flags);
1317                 if (!plug)
1318                         md_wakeup_thread(mddev->thread);
1319         }
1320         /* Mustn't call r1_bio_write_done before this next test,
1321          * as it could result in the bio being freed.
1322          */
1323         if (sectors_handled < (bio->bi_size >> 9)) {
1324                 r1_bio_write_done(r1_bio);
1325                 /* We need another r1_bio.  It has already been counted
1326                  * in bio->bi_phys_segments
1327                  */
1328                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1329                 r1_bio->master_bio = bio;
1330                 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1331                 r1_bio->state = 0;
1332                 r1_bio->mddev = mddev;
1333                 r1_bio->sector = bio->bi_sector + sectors_handled;
1334                 goto retry_write;
1335         }
1336
1337         r1_bio_write_done(r1_bio);
1338
1339         /* In case raid1d snuck in to freeze_array */
1340         wake_up(&conf->wait_barrier);
1341 }
1342
1343 static void status(struct seq_file *seq, struct mddev *mddev)
1344 {
1345         struct r1conf *conf = mddev->private;
1346         int i;
1347
1348         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1349                    conf->raid_disks - mddev->degraded);
1350         rcu_read_lock();
1351         for (i = 0; i < conf->raid_disks; i++) {
1352                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1353                 seq_printf(seq, "%s",
1354                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1355         }
1356         rcu_read_unlock();
1357         seq_printf(seq, "]");
1358 }
1359
1360
1361 static void error(struct mddev *mddev, struct md_rdev *rdev)
1362 {
1363         char b[BDEVNAME_SIZE];
1364         struct r1conf *conf = mddev->private;
1365
1366         /*
1367          * If it is not operational, then we have already marked it as dead
1368          * else if it is the last working disks, ignore the error, let the
1369          * next level up know.
1370          * else mark the drive as failed
1371          */
1372         if (test_bit(In_sync, &rdev->flags)
1373             && (conf->raid_disks - mddev->degraded) == 1) {
1374                 /*
1375                  * Don't fail the drive, act as though we were just a
1376                  * normal single drive.
1377                  * However don't try a recovery from this drive as
1378                  * it is very likely to fail.
1379                  */
1380                 conf->recovery_disabled = mddev->recovery_disabled;
1381                 return;
1382         }
1383         set_bit(Blocked, &rdev->flags);
1384         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1385                 unsigned long flags;
1386                 spin_lock_irqsave(&conf->device_lock, flags);
1387                 mddev->degraded++;
1388                 set_bit(Faulty, &rdev->flags);
1389                 spin_unlock_irqrestore(&conf->device_lock, flags);
1390                 /*
1391                  * if recovery is running, make sure it aborts.
1392                  */
1393                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1394         } else
1395                 set_bit(Faulty, &rdev->flags);
1396         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1397         printk(KERN_ALERT
1398                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1399                "md/raid1:%s: Operation continuing on %d devices.\n",
1400                mdname(mddev), bdevname(rdev->bdev, b),
1401                mdname(mddev), conf->raid_disks - mddev->degraded);
1402 }
1403
1404 static void print_conf(struct r1conf *conf)
1405 {
1406         int i;
1407
1408         printk(KERN_DEBUG "RAID1 conf printout:\n");
1409         if (!conf) {
1410                 printk(KERN_DEBUG "(!conf)\n");
1411                 return;
1412         }
1413         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1414                 conf->raid_disks);
1415
1416         rcu_read_lock();
1417         for (i = 0; i < conf->raid_disks; i++) {
1418                 char b[BDEVNAME_SIZE];
1419                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1420                 if (rdev)
1421                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1422                                i, !test_bit(In_sync, &rdev->flags),
1423                                !test_bit(Faulty, &rdev->flags),
1424                                bdevname(rdev->bdev,b));
1425         }
1426         rcu_read_unlock();
1427 }
1428
1429 static void close_sync(struct r1conf *conf)
1430 {
1431         wait_barrier(conf);
1432         allow_barrier(conf);
1433
1434         mempool_destroy(conf->r1buf_pool);
1435         conf->r1buf_pool = NULL;
1436 }
1437
1438 static int raid1_spare_active(struct mddev *mddev)
1439 {
1440         int i;
1441         struct r1conf *conf = mddev->private;
1442         int count = 0;
1443         unsigned long flags;
1444
1445         /*
1446          * Find all failed disks within the RAID1 configuration
1447          * and mark them readable.
1448          * Called under mddev lock, so rcu protection not needed.
1449          */
1450         for (i = 0; i < conf->raid_disks; i++) {
1451                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1452                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1453                 if (repl
1454                     && repl->recovery_offset == MaxSector
1455                     && !test_bit(Faulty, &repl->flags)
1456                     && !test_and_set_bit(In_sync, &repl->flags)) {
1457                         /* replacement has just become active */
1458                         if (!rdev ||
1459                             !test_and_clear_bit(In_sync, &rdev->flags))
1460                                 count++;
1461                         if (rdev) {
1462                                 /* Replaced device not technically
1463                                  * faulty, but we need to be sure
1464                                  * it gets removed and never re-added
1465                                  */
1466                                 set_bit(Faulty, &rdev->flags);
1467                                 sysfs_notify_dirent_safe(
1468                                         rdev->sysfs_state);
1469                         }
1470                 }
1471                 if (rdev
1472                     && !test_bit(Faulty, &rdev->flags)
1473                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1474                         count++;
1475                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1476                 }
1477         }
1478         spin_lock_irqsave(&conf->device_lock, flags);
1479         mddev->degraded -= count;
1480         spin_unlock_irqrestore(&conf->device_lock, flags);
1481
1482         print_conf(conf);
1483         return count;
1484 }
1485
1486
1487 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1488 {
1489         struct r1conf *conf = mddev->private;
1490         int err = -EEXIST;
1491         int mirror = 0;
1492         struct raid1_info *p;
1493         int first = 0;
1494         int last = conf->raid_disks - 1;
1495         struct request_queue *q = bdev_get_queue(rdev->bdev);
1496
1497         if (mddev->recovery_disabled == conf->recovery_disabled)
1498                 return -EBUSY;
1499
1500         if (rdev->raid_disk >= 0)
1501                 first = last = rdev->raid_disk;
1502
1503         if (q->merge_bvec_fn) {
1504                 set_bit(Unmerged, &rdev->flags);
1505                 mddev->merge_check_needed = 1;
1506         }
1507
1508         for (mirror = first; mirror <= last; mirror++) {
1509                 p = conf->mirrors+mirror;
1510                 if (!p->rdev) {
1511
1512                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1513                                           rdev->data_offset << 9);
1514
1515                         p->head_position = 0;
1516                         rdev->raid_disk = mirror;
1517                         err = 0;
1518                         /* As all devices are equivalent, we don't need a full recovery
1519                          * if this was recently any drive of the array
1520                          */
1521                         if (rdev->saved_raid_disk < 0)
1522                                 conf->fullsync = 1;
1523                         rcu_assign_pointer(p->rdev, rdev);
1524                         break;
1525                 }
1526                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1527                     p[conf->raid_disks].rdev == NULL) {
1528                         /* Add this device as a replacement */
1529                         clear_bit(In_sync, &rdev->flags);
1530                         set_bit(Replacement, &rdev->flags);
1531                         rdev->raid_disk = mirror;
1532                         err = 0;
1533                         conf->fullsync = 1;
1534                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1535                         break;
1536                 }
1537         }
1538         if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1539                 /* Some requests might not have seen this new
1540                  * merge_bvec_fn.  We must wait for them to complete
1541                  * before merging the device fully.
1542                  * First we make sure any code which has tested
1543                  * our function has submitted the request, then
1544                  * we wait for all outstanding requests to complete.
1545                  */
1546                 synchronize_sched();
1547                 raise_barrier(conf);
1548                 lower_barrier(conf);
1549                 clear_bit(Unmerged, &rdev->flags);
1550         }
1551         md_integrity_add_rdev(rdev, mddev);
1552         print_conf(conf);
1553         return err;
1554 }
1555
1556 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1557 {
1558         struct r1conf *conf = mddev->private;
1559         int err = 0;
1560         int number = rdev->raid_disk;
1561         struct raid1_info *p = conf->mirrors + number;
1562
1563         if (rdev != p->rdev)
1564                 p = conf->mirrors + conf->raid_disks + number;
1565
1566         print_conf(conf);
1567         if (rdev == p->rdev) {
1568                 if (test_bit(In_sync, &rdev->flags) ||
1569                     atomic_read(&rdev->nr_pending)) {
1570                         err = -EBUSY;
1571                         goto abort;
1572                 }
1573                 /* Only remove non-faulty devices if recovery
1574                  * is not possible.
1575                  */
1576                 if (!test_bit(Faulty, &rdev->flags) &&
1577                     mddev->recovery_disabled != conf->recovery_disabled &&
1578                     mddev->degraded < conf->raid_disks) {
1579                         err = -EBUSY;
1580                         goto abort;
1581                 }
1582                 p->rdev = NULL;
1583                 synchronize_rcu();
1584                 if (atomic_read(&rdev->nr_pending)) {
1585                         /* lost the race, try later */
1586                         err = -EBUSY;
1587                         p->rdev = rdev;
1588                         goto abort;
1589                 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1590                         /* We just removed a device that is being replaced.
1591                          * Move down the replacement.  We drain all IO before
1592                          * doing this to avoid confusion.
1593                          */
1594                         struct md_rdev *repl =
1595                                 conf->mirrors[conf->raid_disks + number].rdev;
1596                         raise_barrier(conf);
1597                         clear_bit(Replacement, &repl->flags);
1598                         p->rdev = repl;
1599                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1600                         lower_barrier(conf);
1601                         clear_bit(WantReplacement, &rdev->flags);
1602                 } else
1603                         clear_bit(WantReplacement, &rdev->flags);
1604                 err = md_integrity_register(mddev);
1605         }
1606 abort:
1607
1608         print_conf(conf);
1609         return err;
1610 }
1611
1612
1613 static void end_sync_read(struct bio *bio, int error)
1614 {
1615         struct r1bio *r1_bio = bio->bi_private;
1616
1617         update_head_pos(r1_bio->read_disk, r1_bio);
1618
1619         /*
1620          * we have read a block, now it needs to be re-written,
1621          * or re-read if the read failed.
1622          * We don't do much here, just schedule handling by raid1d
1623          */
1624         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1625                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1626
1627         if (atomic_dec_and_test(&r1_bio->remaining))
1628                 reschedule_retry(r1_bio);
1629 }
1630
1631 static void end_sync_write(struct bio *bio, int error)
1632 {
1633         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1634         struct r1bio *r1_bio = bio->bi_private;
1635         struct mddev *mddev = r1_bio->mddev;
1636         struct r1conf *conf = mddev->private;
1637         int mirror=0;
1638         sector_t first_bad;
1639         int bad_sectors;
1640
1641         mirror = find_bio_disk(r1_bio, bio);
1642
1643         if (!uptodate) {
1644                 sector_t sync_blocks = 0;
1645                 sector_t s = r1_bio->sector;
1646                 long sectors_to_go = r1_bio->sectors;
1647                 /* make sure these bits doesn't get cleared. */
1648                 do {
1649                         bitmap_end_sync(mddev->bitmap, s,
1650                                         &sync_blocks, 1);
1651                         s += sync_blocks;
1652                         sectors_to_go -= sync_blocks;
1653                 } while (sectors_to_go > 0);
1654                 set_bit(WriteErrorSeen,
1655                         &conf->mirrors[mirror].rdev->flags);
1656                 if (!test_and_set_bit(WantReplacement,
1657                                       &conf->mirrors[mirror].rdev->flags))
1658                         set_bit(MD_RECOVERY_NEEDED, &
1659                                 mddev->recovery);
1660                 set_bit(R1BIO_WriteError, &r1_bio->state);
1661         } else if (is_badblock(conf->mirrors[mirror].rdev,
1662                                r1_bio->sector,
1663                                r1_bio->sectors,
1664                                &first_bad, &bad_sectors) &&
1665                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1666                                 r1_bio->sector,
1667                                 r1_bio->sectors,
1668                                 &first_bad, &bad_sectors)
1669                 )
1670                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1671
1672         if (atomic_dec_and_test(&r1_bio->remaining)) {
1673                 int s = r1_bio->sectors;
1674                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1675                     test_bit(R1BIO_WriteError, &r1_bio->state))
1676                         reschedule_retry(r1_bio);
1677                 else {
1678                         put_buf(r1_bio);
1679                         md_done_sync(mddev, s, uptodate);
1680                 }
1681         }
1682 }
1683
1684 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1685                             int sectors, struct page *page, int rw)
1686 {
1687         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1688                 /* success */
1689                 return 1;
1690         if (rw == WRITE) {
1691                 set_bit(WriteErrorSeen, &rdev->flags);
1692                 if (!test_and_set_bit(WantReplacement,
1693                                       &rdev->flags))
1694                         set_bit(MD_RECOVERY_NEEDED, &
1695                                 rdev->mddev->recovery);
1696         }
1697         /* need to record an error - either for the block or the device */
1698         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1699                 md_error(rdev->mddev, rdev);
1700         return 0;
1701 }
1702
1703 static int fix_sync_read_error(struct r1bio *r1_bio)
1704 {
1705         /* Try some synchronous reads of other devices to get
1706          * good data, much like with normal read errors.  Only
1707          * read into the pages we already have so we don't
1708          * need to re-issue the read request.
1709          * We don't need to freeze the array, because being in an
1710          * active sync request, there is no normal IO, and
1711          * no overlapping syncs.
1712          * We don't need to check is_badblock() again as we
1713          * made sure that anything with a bad block in range
1714          * will have bi_end_io clear.
1715          */
1716         struct mddev *mddev = r1_bio->mddev;
1717         struct r1conf *conf = mddev->private;
1718         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1719         sector_t sect = r1_bio->sector;
1720         int sectors = r1_bio->sectors;
1721         int idx = 0;
1722
1723         while(sectors) {
1724                 int s = sectors;
1725                 int d = r1_bio->read_disk;
1726                 int success = 0;
1727                 struct md_rdev *rdev;
1728                 int start;
1729
1730                 if (s > (PAGE_SIZE>>9))
1731                         s = PAGE_SIZE >> 9;
1732                 do {
1733                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1734                                 /* No rcu protection needed here devices
1735                                  * can only be removed when no resync is
1736                                  * active, and resync is currently active
1737                                  */
1738                                 rdev = conf->mirrors[d].rdev;
1739                                 if (sync_page_io(rdev, sect, s<<9,
1740                                                  bio->bi_io_vec[idx].bv_page,
1741                                                  READ, false)) {
1742                                         success = 1;
1743                                         break;
1744                                 }
1745                         }
1746                         d++;
1747                         if (d == conf->raid_disks * 2)
1748                                 d = 0;
1749                 } while (!success && d != r1_bio->read_disk);
1750
1751                 if (!success) {
1752                         char b[BDEVNAME_SIZE];
1753                         int abort = 0;
1754                         /* Cannot read from anywhere, this block is lost.
1755                          * Record a bad block on each device.  If that doesn't
1756                          * work just disable and interrupt the recovery.
1757                          * Don't fail devices as that won't really help.
1758                          */
1759                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1760                                " for block %llu\n",
1761                                mdname(mddev),
1762                                bdevname(bio->bi_bdev, b),
1763                                (unsigned long long)r1_bio->sector);
1764                         for (d = 0; d < conf->raid_disks * 2; d++) {
1765                                 rdev = conf->mirrors[d].rdev;
1766                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1767                                         continue;
1768                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1769                                         abort = 1;
1770                         }
1771                         if (abort) {
1772                                 conf->recovery_disabled =
1773                                         mddev->recovery_disabled;
1774                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1775                                 md_done_sync(mddev, r1_bio->sectors, 0);
1776                                 put_buf(r1_bio);
1777                                 return 0;
1778                         }
1779                         /* Try next page */
1780                         sectors -= s;
1781                         sect += s;
1782                         idx++;
1783                         continue;
1784                 }
1785
1786                 start = d;
1787                 /* write it back and re-read */
1788                 while (d != r1_bio->read_disk) {
1789                         if (d == 0)
1790                                 d = conf->raid_disks * 2;
1791                         d--;
1792                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1793                                 continue;
1794                         rdev = conf->mirrors[d].rdev;
1795                         if (r1_sync_page_io(rdev, sect, s,
1796                                             bio->bi_io_vec[idx].bv_page,
1797                                             WRITE) == 0) {
1798                                 r1_bio->bios[d]->bi_end_io = NULL;
1799                                 rdev_dec_pending(rdev, mddev);
1800                         }
1801                 }
1802                 d = start;
1803                 while (d != r1_bio->read_disk) {
1804                         if (d == 0)
1805                                 d = conf->raid_disks * 2;
1806                         d--;
1807                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1808                                 continue;
1809                         rdev = conf->mirrors[d].rdev;
1810                         if (r1_sync_page_io(rdev, sect, s,
1811                                             bio->bi_io_vec[idx].bv_page,
1812                                             READ) != 0)
1813                                 atomic_add(s, &rdev->corrected_errors);
1814                 }
1815                 sectors -= s;
1816                 sect += s;
1817                 idx ++;
1818         }
1819         set_bit(R1BIO_Uptodate, &r1_bio->state);
1820         set_bit(BIO_UPTODATE, &bio->bi_flags);
1821         return 1;
1822 }
1823
1824 static int process_checks(struct r1bio *r1_bio)
1825 {
1826         /* We have read all readable devices.  If we haven't
1827          * got the block, then there is no hope left.
1828          * If we have, then we want to do a comparison
1829          * and skip the write if everything is the same.
1830          * If any blocks failed to read, then we need to
1831          * attempt an over-write
1832          */
1833         struct mddev *mddev = r1_bio->mddev;
1834         struct r1conf *conf = mddev->private;
1835         int primary;
1836         int i;
1837         int vcnt;
1838
1839         for (primary = 0; primary < conf->raid_disks * 2; primary++)
1840                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1841                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1842                         r1_bio->bios[primary]->bi_end_io = NULL;
1843                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1844                         break;
1845                 }
1846         r1_bio->read_disk = primary;
1847         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1848         for (i = 0; i < conf->raid_disks * 2; i++) {
1849                 int j;
1850                 struct bio *pbio = r1_bio->bios[primary];
1851                 struct bio *sbio = r1_bio->bios[i];
1852                 int size;
1853
1854                 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1855                         continue;
1856
1857                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1858                         for (j = vcnt; j-- ; ) {
1859                                 struct page *p, *s;
1860                                 p = pbio->bi_io_vec[j].bv_page;
1861                                 s = sbio->bi_io_vec[j].bv_page;
1862                                 if (memcmp(page_address(p),
1863                                            page_address(s),
1864                                            sbio->bi_io_vec[j].bv_len))
1865                                         break;
1866                         }
1867                 } else
1868                         j = 0;
1869                 if (j >= 0)
1870                         mddev->resync_mismatches += r1_bio->sectors;
1871                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1872                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1873                         /* No need to write to this device. */
1874                         sbio->bi_end_io = NULL;
1875                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1876                         continue;
1877                 }
1878                 /* fixup the bio for reuse */
1879                 sbio->bi_vcnt = vcnt;
1880                 sbio->bi_size = r1_bio->sectors << 9;
1881                 sbio->bi_idx = 0;
1882                 sbio->bi_phys_segments = 0;
1883                 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1884                 sbio->bi_flags |= 1 << BIO_UPTODATE;
1885                 sbio->bi_next = NULL;
1886                 sbio->bi_sector = r1_bio->sector +
1887                         conf->mirrors[i].rdev->data_offset;
1888                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1889                 size = sbio->bi_size;
1890                 for (j = 0; j < vcnt ; j++) {
1891                         struct bio_vec *bi;
1892                         bi = &sbio->bi_io_vec[j];
1893                         bi->bv_offset = 0;
1894                         if (size > PAGE_SIZE)
1895                                 bi->bv_len = PAGE_SIZE;
1896                         else
1897                                 bi->bv_len = size;
1898                         size -= PAGE_SIZE;
1899                         memcpy(page_address(bi->bv_page),
1900                                page_address(pbio->bi_io_vec[j].bv_page),
1901                                PAGE_SIZE);
1902                 }
1903         }
1904         return 0;
1905 }
1906
1907 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1908 {
1909         struct r1conf *conf = mddev->private;
1910         int i;
1911         int disks = conf->raid_disks * 2;
1912         struct bio *bio, *wbio;
1913
1914         bio = r1_bio->bios[r1_bio->read_disk];
1915
1916         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1917                 /* ouch - failed to read all of that. */
1918                 if (!fix_sync_read_error(r1_bio))
1919                         return;
1920
1921         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1922                 if (process_checks(r1_bio) < 0)
1923                         return;
1924         /*
1925          * schedule writes
1926          */
1927         atomic_set(&r1_bio->remaining, 1);
1928         for (i = 0; i < disks ; i++) {
1929                 wbio = r1_bio->bios[i];
1930                 if (wbio->bi_end_io == NULL ||
1931                     (wbio->bi_end_io == end_sync_read &&
1932                      (i == r1_bio->read_disk ||
1933                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1934                         continue;
1935
1936                 wbio->bi_rw = WRITE;
1937                 wbio->bi_end_io = end_sync_write;
1938                 atomic_inc(&r1_bio->remaining);
1939                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1940
1941                 generic_make_request(wbio);
1942         }
1943
1944         if (atomic_dec_and_test(&r1_bio->remaining)) {
1945                 /* if we're here, all write(s) have completed, so clean up */
1946                 int s = r1_bio->sectors;
1947                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1948                     test_bit(R1BIO_WriteError, &r1_bio->state))
1949                         reschedule_retry(r1_bio);
1950                 else {
1951                         put_buf(r1_bio);
1952                         md_done_sync(mddev, s, 1);
1953                 }
1954         }
1955 }
1956
1957 /*
1958  * This is a kernel thread which:
1959  *
1960  *      1.      Retries failed read operations on working mirrors.
1961  *      2.      Updates the raid superblock when problems encounter.
1962  *      3.      Performs writes following reads for array synchronising.
1963  */
1964
1965 static void fix_read_error(struct r1conf *conf, int read_disk,
1966                            sector_t sect, int sectors)
1967 {
1968         struct mddev *mddev = conf->mddev;
1969         while(sectors) {
1970                 int s = sectors;
1971                 int d = read_disk;
1972                 int success = 0;
1973                 int start;
1974                 struct md_rdev *rdev;
1975
1976                 if (s > (PAGE_SIZE>>9))
1977                         s = PAGE_SIZE >> 9;
1978
1979                 do {
1980                         /* Note: no rcu protection needed here
1981                          * as this is synchronous in the raid1d thread
1982                          * which is the thread that might remove
1983                          * a device.  If raid1d ever becomes multi-threaded....
1984                          */
1985                         sector_t first_bad;
1986                         int bad_sectors;
1987
1988                         rdev = conf->mirrors[d].rdev;
1989                         if (rdev &&
1990                             (test_bit(In_sync, &rdev->flags) ||
1991                              (!test_bit(Faulty, &rdev->flags) &&
1992                               rdev->recovery_offset >= sect + s)) &&
1993                             is_badblock(rdev, sect, s,
1994                                         &first_bad, &bad_sectors) == 0 &&
1995                             sync_page_io(rdev, sect, s<<9,
1996                                          conf->tmppage, READ, false))
1997                                 success = 1;
1998                         else {
1999                                 d++;
2000                                 if (d == conf->raid_disks * 2)
2001                                         d = 0;
2002                         }
2003                 } while (!success && d != read_disk);
2004
2005                 if (!success) {
2006                         /* Cannot read from anywhere - mark it bad */
2007                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2008                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2009                                 md_error(mddev, rdev);
2010                         break;
2011                 }
2012                 /* write it back and re-read */
2013                 start = d;
2014                 while (d != read_disk) {
2015                         if (d==0)
2016                                 d = conf->raid_disks * 2;
2017                         d--;
2018                         rdev = conf->mirrors[d].rdev;
2019                         if (rdev &&
2020                             test_bit(In_sync, &rdev->flags))
2021                                 r1_sync_page_io(rdev, sect, s,
2022                                                 conf->tmppage, WRITE);
2023                 }
2024                 d = start;
2025                 while (d != read_disk) {
2026                         char b[BDEVNAME_SIZE];
2027                         if (d==0)
2028                                 d = conf->raid_disks * 2;
2029                         d--;
2030                         rdev = conf->mirrors[d].rdev;
2031                         if (rdev &&
2032                             test_bit(In_sync, &rdev->flags)) {
2033                                 if (r1_sync_page_io(rdev, sect, s,
2034                                                     conf->tmppage, READ)) {
2035                                         atomic_add(s, &rdev->corrected_errors);
2036                                         printk(KERN_INFO
2037                                                "md/raid1:%s: read error corrected "
2038                                                "(%d sectors at %llu on %s)\n",
2039                                                mdname(mddev), s,
2040                                                (unsigned long long)(sect +
2041                                                    rdev->data_offset),
2042                                                bdevname(rdev->bdev, b));
2043                                 }
2044                         }
2045                 }
2046                 sectors -= s;
2047                 sect += s;
2048         }
2049 }
2050
2051 static void bi_complete(struct bio *bio, int error)
2052 {
2053         complete((struct completion *)bio->bi_private);
2054 }
2055
2056 static int submit_bio_wait(int rw, struct bio *bio)
2057 {
2058         struct completion event;
2059         rw |= REQ_SYNC;
2060
2061         init_completion(&event);
2062         bio->bi_private = &event;
2063         bio->bi_end_io = bi_complete;
2064         submit_bio(rw, bio);
2065         wait_for_completion(&event);
2066
2067         return test_bit(BIO_UPTODATE, &bio->bi_flags);
2068 }
2069
2070 static int narrow_write_error(struct r1bio *r1_bio, int i)
2071 {
2072         struct mddev *mddev = r1_bio->mddev;
2073         struct r1conf *conf = mddev->private;
2074         struct md_rdev *rdev = conf->mirrors[i].rdev;
2075         int vcnt, idx;
2076         struct bio_vec *vec;
2077
2078         /* bio has the data to be written to device 'i' where
2079          * we just recently had a write error.
2080          * We repeatedly clone the bio and trim down to one block,
2081          * then try the write.  Where the write fails we record
2082          * a bad block.
2083          * It is conceivable that the bio doesn't exactly align with
2084          * blocks.  We must handle this somehow.
2085          *
2086          * We currently own a reference on the rdev.
2087          */
2088
2089         int block_sectors;
2090         sector_t sector;
2091         int sectors;
2092         int sect_to_write = r1_bio->sectors;
2093         int ok = 1;
2094
2095         if (rdev->badblocks.shift < 0)
2096                 return 0;
2097
2098         block_sectors = 1 << rdev->badblocks.shift;
2099         sector = r1_bio->sector;
2100         sectors = ((sector + block_sectors)
2101                    & ~(sector_t)(block_sectors - 1))
2102                 - sector;
2103
2104         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2105                 vcnt = r1_bio->behind_page_count;
2106                 vec = r1_bio->behind_bvecs;
2107                 idx = 0;
2108                 while (vec[idx].bv_page == NULL)
2109                         idx++;
2110         } else {
2111                 vcnt = r1_bio->master_bio->bi_vcnt;
2112                 vec = r1_bio->master_bio->bi_io_vec;
2113                 idx = r1_bio->master_bio->bi_idx;
2114         }
2115         while (sect_to_write) {
2116                 struct bio *wbio;
2117                 if (sectors > sect_to_write)
2118                         sectors = sect_to_write;
2119                 /* Write at 'sector' for 'sectors'*/
2120
2121                 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2122                 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2123                 wbio->bi_sector = r1_bio->sector;
2124                 wbio->bi_rw = WRITE;
2125                 wbio->bi_vcnt = vcnt;
2126                 wbio->bi_size = r1_bio->sectors << 9;
2127                 wbio->bi_idx = idx;
2128
2129                 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2130                 wbio->bi_sector += rdev->data_offset;
2131                 wbio->bi_bdev = rdev->bdev;
2132                 if (submit_bio_wait(WRITE, wbio) == 0)
2133                         /* failure! */
2134                         ok = rdev_set_badblocks(rdev, sector,
2135                                                 sectors, 0)
2136                                 && ok;
2137
2138                 bio_put(wbio);
2139                 sect_to_write -= sectors;
2140                 sector += sectors;
2141                 sectors = block_sectors;
2142         }
2143         return ok;
2144 }
2145
2146 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2147 {
2148         int m;
2149         int s = r1_bio->sectors;
2150         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2151                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2152                 struct bio *bio = r1_bio->bios[m];
2153                 if (bio->bi_end_io == NULL)
2154                         continue;
2155                 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2156                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2157                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2158                 }
2159                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2160                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2161                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2162                                 md_error(conf->mddev, rdev);
2163                 }
2164         }
2165         put_buf(r1_bio);
2166         md_done_sync(conf->mddev, s, 1);
2167 }
2168
2169 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2170 {
2171         int m;
2172         for (m = 0; m < conf->raid_disks * 2 ; m++)
2173                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2174                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2175                         rdev_clear_badblocks(rdev,
2176                                              r1_bio->sector,
2177                                              r1_bio->sectors, 0);
2178                         rdev_dec_pending(rdev, conf->mddev);
2179                 } else if (r1_bio->bios[m] != NULL) {
2180                         /* This drive got a write error.  We need to
2181                          * narrow down and record precise write
2182                          * errors.
2183                          */
2184                         if (!narrow_write_error(r1_bio, m)) {
2185                                 md_error(conf->mddev,
2186                                          conf->mirrors[m].rdev);
2187                                 /* an I/O failed, we can't clear the bitmap */
2188                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2189                         }
2190                         rdev_dec_pending(conf->mirrors[m].rdev,
2191                                          conf->mddev);
2192                 }
2193         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2194                 close_write(r1_bio);
2195         raid_end_bio_io(r1_bio);
2196 }
2197
2198 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2199 {
2200         int disk;
2201         int max_sectors;
2202         struct mddev *mddev = conf->mddev;
2203         struct bio *bio;
2204         char b[BDEVNAME_SIZE];
2205         struct md_rdev *rdev;
2206
2207         clear_bit(R1BIO_ReadError, &r1_bio->state);
2208         /* we got a read error. Maybe the drive is bad.  Maybe just
2209          * the block and we can fix it.
2210          * We freeze all other IO, and try reading the block from
2211          * other devices.  When we find one, we re-write
2212          * and check it that fixes the read error.
2213          * This is all done synchronously while the array is
2214          * frozen
2215          */
2216         if (mddev->ro == 0) {
2217                 freeze_array(conf);
2218                 fix_read_error(conf, r1_bio->read_disk,
2219                                r1_bio->sector, r1_bio->sectors);
2220                 unfreeze_array(conf);
2221         } else
2222                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2223
2224         bio = r1_bio->bios[r1_bio->read_disk];
2225         bdevname(bio->bi_bdev, b);
2226 read_more:
2227         disk = read_balance(conf, r1_bio, &max_sectors);
2228         if (disk == -1) {
2229                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2230                        " read error for block %llu\n",
2231                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
2232                 raid_end_bio_io(r1_bio);
2233         } else {
2234                 const unsigned long do_sync
2235                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
2236                 if (bio) {
2237                         r1_bio->bios[r1_bio->read_disk] =
2238                                 mddev->ro ? IO_BLOCKED : NULL;
2239                         bio_put(bio);
2240                 }
2241                 r1_bio->read_disk = disk;
2242                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2243                 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2244                 r1_bio->bios[r1_bio->read_disk] = bio;
2245                 rdev = conf->mirrors[disk].rdev;
2246                 printk_ratelimited(KERN_ERR
2247                                    "md/raid1:%s: redirecting sector %llu"
2248                                    " to other mirror: %s\n",
2249                                    mdname(mddev),
2250                                    (unsigned long long)r1_bio->sector,
2251                                    bdevname(rdev->bdev, b));
2252                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
2253                 bio->bi_bdev = rdev->bdev;
2254                 bio->bi_end_io = raid1_end_read_request;
2255                 bio->bi_rw = READ | do_sync;
2256                 bio->bi_private = r1_bio;
2257                 if (max_sectors < r1_bio->sectors) {
2258                         /* Drat - have to split this up more */
2259                         struct bio *mbio = r1_bio->master_bio;
2260                         int sectors_handled = (r1_bio->sector + max_sectors
2261                                                - mbio->bi_sector);
2262                         r1_bio->sectors = max_sectors;
2263                         spin_lock_irq(&conf->device_lock);
2264                         if (mbio->bi_phys_segments == 0)
2265                                 mbio->bi_phys_segments = 2;
2266                         else
2267                                 mbio->bi_phys_segments++;
2268                         spin_unlock_irq(&conf->device_lock);
2269                         generic_make_request(bio);
2270                         bio = NULL;
2271
2272                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2273
2274                         r1_bio->master_bio = mbio;
2275                         r1_bio->sectors = (mbio->bi_size >> 9)
2276                                           - sectors_handled;
2277                         r1_bio->state = 0;
2278                         set_bit(R1BIO_ReadError, &r1_bio->state);
2279                         r1_bio->mddev = mddev;
2280                         r1_bio->sector = mbio->bi_sector + sectors_handled;
2281
2282                         goto read_more;
2283                 } else
2284                         generic_make_request(bio);
2285         }
2286 }
2287
2288 static void raid1d(struct mddev *mddev)
2289 {
2290         struct r1bio *r1_bio;
2291         unsigned long flags;
2292         struct r1conf *conf = mddev->private;
2293         struct list_head *head = &conf->retry_list;
2294         struct blk_plug plug;
2295
2296         md_check_recovery(mddev);
2297
2298         blk_start_plug(&plug);
2299         for (;;) {
2300
2301                 flush_pending_writes(conf);
2302
2303                 spin_lock_irqsave(&conf->device_lock, flags);
2304                 if (list_empty(head)) {
2305                         spin_unlock_irqrestore(&conf->device_lock, flags);
2306                         break;
2307                 }
2308                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2309                 list_del(head->prev);
2310                 conf->nr_queued--;
2311                 spin_unlock_irqrestore(&conf->device_lock, flags);
2312
2313                 mddev = r1_bio->mddev;
2314                 conf = mddev->private;
2315                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2316                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2317                             test_bit(R1BIO_WriteError, &r1_bio->state))
2318                                 handle_sync_write_finished(conf, r1_bio);
2319                         else
2320                                 sync_request_write(mddev, r1_bio);
2321                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2322                            test_bit(R1BIO_WriteError, &r1_bio->state))
2323                         handle_write_finished(conf, r1_bio);
2324                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2325                         handle_read_error(conf, r1_bio);
2326                 else
2327                         /* just a partial read to be scheduled from separate
2328                          * context
2329                          */
2330                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2331
2332                 cond_resched();
2333                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2334                         md_check_recovery(mddev);
2335         }
2336         blk_finish_plug(&plug);
2337 }
2338
2339
2340 static int init_resync(struct r1conf *conf)
2341 {
2342         int buffs;
2343
2344         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2345         BUG_ON(conf->r1buf_pool);
2346         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2347                                           conf->poolinfo);
2348         if (!conf->r1buf_pool)
2349                 return -ENOMEM;
2350         conf->next_resync = 0;
2351         return 0;
2352 }
2353
2354 /*
2355  * perform a "sync" on one "block"
2356  *
2357  * We need to make sure that no normal I/O request - particularly write
2358  * requests - conflict with active sync requests.
2359  *
2360  * This is achieved by tracking pending requests and a 'barrier' concept
2361  * that can be installed to exclude normal IO requests.
2362  */
2363
2364 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2365 {
2366         struct r1conf *conf = mddev->private;
2367         struct r1bio *r1_bio;
2368         struct bio *bio;
2369         sector_t max_sector, nr_sectors;
2370         int disk = -1;
2371         int i;
2372         int wonly = -1;
2373         int write_targets = 0, read_targets = 0;
2374         sector_t sync_blocks;
2375         int still_degraded = 0;
2376         int good_sectors = RESYNC_SECTORS;
2377         int min_bad = 0; /* number of sectors that are bad in all devices */
2378
2379         if (!conf->r1buf_pool)
2380                 if (init_resync(conf))
2381                         return 0;
2382
2383         max_sector = mddev->dev_sectors;
2384         if (sector_nr >= max_sector) {
2385                 /* If we aborted, we need to abort the
2386                  * sync on the 'current' bitmap chunk (there will
2387                  * only be one in raid1 resync.
2388                  * We can find the current addess in mddev->curr_resync
2389                  */
2390                 if (mddev->curr_resync < max_sector) /* aborted */
2391                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2392                                                 &sync_blocks, 1);
2393                 else /* completed sync */
2394                         conf->fullsync = 0;
2395
2396                 bitmap_close_sync(mddev->bitmap);
2397                 close_sync(conf);
2398                 return 0;
2399         }
2400
2401         if (mddev->bitmap == NULL &&
2402             mddev->recovery_cp == MaxSector &&
2403             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2404             conf->fullsync == 0) {
2405                 *skipped = 1;
2406                 return max_sector - sector_nr;
2407         }
2408         /* before building a request, check if we can skip these blocks..
2409          * This call the bitmap_start_sync doesn't actually record anything
2410          */
2411         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2412             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2413                 /* We can skip this block, and probably several more */
2414                 *skipped = 1;
2415                 return sync_blocks;
2416         }
2417         /*
2418          * If there is non-resync activity waiting for a turn,
2419          * and resync is going fast enough,
2420          * then let it though before starting on this new sync request.
2421          */
2422         if (!go_faster && conf->nr_waiting)
2423                 msleep_interruptible(1000);
2424
2425         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2426         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2427         raise_barrier(conf);
2428
2429         conf->next_resync = sector_nr;
2430
2431         rcu_read_lock();
2432         /*
2433          * If we get a correctably read error during resync or recovery,
2434          * we might want to read from a different device.  So we
2435          * flag all drives that could conceivably be read from for READ,
2436          * and any others (which will be non-In_sync devices) for WRITE.
2437          * If a read fails, we try reading from something else for which READ
2438          * is OK.
2439          */
2440
2441         r1_bio->mddev = mddev;
2442         r1_bio->sector = sector_nr;
2443         r1_bio->state = 0;
2444         set_bit(R1BIO_IsSync, &r1_bio->state);
2445
2446         for (i = 0; i < conf->raid_disks * 2; i++) {
2447                 struct md_rdev *rdev;
2448                 bio = r1_bio->bios[i];
2449
2450                 /* take from bio_init */
2451                 bio->bi_next = NULL;
2452                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
2453                 bio->bi_flags |= 1 << BIO_UPTODATE;
2454                 bio->bi_rw = READ;
2455                 bio->bi_vcnt = 0;
2456                 bio->bi_idx = 0;
2457                 bio->bi_phys_segments = 0;
2458                 bio->bi_size = 0;
2459                 bio->bi_end_io = NULL;
2460                 bio->bi_private = NULL;
2461
2462                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2463                 if (rdev == NULL ||
2464                     test_bit(Faulty, &rdev->flags)) {
2465                         if (i < conf->raid_disks)
2466                                 still_degraded = 1;
2467                 } else if (!test_bit(In_sync, &rdev->flags)) {
2468                         bio->bi_rw = WRITE;
2469                         bio->bi_end_io = end_sync_write;
2470                         write_targets ++;
2471                 } else {
2472                         /* may need to read from here */
2473                         sector_t first_bad = MaxSector;
2474                         int bad_sectors;
2475
2476                         if (is_badblock(rdev, sector_nr, good_sectors,
2477                                         &first_bad, &bad_sectors)) {
2478                                 if (first_bad > sector_nr)
2479                                         good_sectors = first_bad - sector_nr;
2480                                 else {
2481                                         bad_sectors -= (sector_nr - first_bad);
2482                                         if (min_bad == 0 ||
2483                                             min_bad > bad_sectors)
2484                                                 min_bad = bad_sectors;
2485                                 }
2486                         }
2487                         if (sector_nr < first_bad) {
2488                                 if (test_bit(WriteMostly, &rdev->flags)) {
2489                                         if (wonly < 0)
2490                                                 wonly = i;
2491                                 } else {
2492                                         if (disk < 0)
2493                                                 disk = i;
2494                                 }
2495                                 bio->bi_rw = READ;
2496                                 bio->bi_end_io = end_sync_read;
2497                                 read_targets++;
2498                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2499                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2500                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2501                                 /*
2502                                  * The device is suitable for reading (InSync),
2503                                  * but has bad block(s) here. Let's try to correct them,
2504                                  * if we are doing resync or repair. Otherwise, leave
2505                                  * this device alone for this sync request.
2506                                  */
2507                                 bio->bi_rw = WRITE;
2508                                 bio->bi_end_io = end_sync_write;
2509                                 write_targets++;
2510                         }
2511                 }
2512                 if (bio->bi_end_io) {
2513                         atomic_inc(&rdev->nr_pending);
2514                         bio->bi_sector = sector_nr + rdev->data_offset;
2515                         bio->bi_bdev = rdev->bdev;
2516                         bio->bi_private = r1_bio;
2517                 }
2518         }
2519         rcu_read_unlock();
2520         if (disk < 0)
2521                 disk = wonly;
2522         r1_bio->read_disk = disk;
2523
2524         if (read_targets == 0 && min_bad > 0) {
2525                 /* These sectors are bad on all InSync devices, so we
2526                  * need to mark them bad on all write targets
2527                  */
2528                 int ok = 1;
2529                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2530                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2531                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2532                                 ok = rdev_set_badblocks(rdev, sector_nr,
2533                                                         min_bad, 0
2534                                         ) && ok;
2535                         }
2536                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2537                 *skipped = 1;
2538                 put_buf(r1_bio);
2539
2540                 if (!ok) {
2541                         /* Cannot record the badblocks, so need to
2542                          * abort the resync.
2543                          * If there are multiple read targets, could just
2544                          * fail the really bad ones ???
2545                          */
2546                         conf->recovery_disabled = mddev->recovery_disabled;
2547                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2548                         return 0;
2549                 } else
2550                         return min_bad;
2551
2552         }
2553         if (min_bad > 0 && min_bad < good_sectors) {
2554                 /* only resync enough to reach the next bad->good
2555                  * transition */
2556                 good_sectors = min_bad;
2557         }
2558
2559         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2560                 /* extra read targets are also write targets */
2561                 write_targets += read_targets-1;
2562
2563         if (write_targets == 0 || read_targets == 0) {
2564                 /* There is nowhere to write, so all non-sync
2565                  * drives must be failed - so we are finished
2566                  */
2567                 sector_t rv;
2568                 if (min_bad > 0)
2569                         max_sector = sector_nr + min_bad;
2570                 rv = max_sector - sector_nr;
2571                 *skipped = 1;
2572                 put_buf(r1_bio);
2573                 return rv;
2574         }
2575
2576         if (max_sector > mddev->resync_max)
2577                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2578         if (max_sector > sector_nr + good_sectors)
2579                 max_sector = sector_nr + good_sectors;
2580         nr_sectors = 0;
2581         sync_blocks = 0;
2582         do {
2583                 struct page *page;
2584                 int len = PAGE_SIZE;
2585                 if (sector_nr + (len>>9) > max_sector)
2586                         len = (max_sector - sector_nr) << 9;
2587                 if (len == 0)
2588                         break;
2589                 if (sync_blocks == 0) {
2590                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2591                                                &sync_blocks, still_degraded) &&
2592                             !conf->fullsync &&
2593                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2594                                 break;
2595                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2596                         if ((len >> 9) > sync_blocks)
2597                                 len = sync_blocks<<9;
2598                 }
2599
2600                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2601                         bio = r1_bio->bios[i];
2602                         if (bio->bi_end_io) {
2603                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2604                                 if (bio_add_page(bio, page, len, 0) == 0) {
2605                                         /* stop here */
2606                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2607                                         while (i > 0) {
2608                                                 i--;
2609                                                 bio = r1_bio->bios[i];
2610                                                 if (bio->bi_end_io==NULL)
2611                                                         continue;
2612                                                 /* remove last page from this bio */
2613                                                 bio->bi_vcnt--;
2614                                                 bio->bi_size -= len;
2615                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2616                                         }
2617                                         goto bio_full;
2618                                 }
2619                         }
2620                 }
2621                 nr_sectors += len>>9;
2622                 sector_nr += len>>9;
2623                 sync_blocks -= (len>>9);
2624         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2625  bio_full:
2626         r1_bio->sectors = nr_sectors;
2627
2628         /* For a user-requested sync, we read all readable devices and do a
2629          * compare
2630          */
2631         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2632                 atomic_set(&r1_bio->remaining, read_targets);
2633                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2634                         bio = r1_bio->bios[i];
2635                         if (bio->bi_end_io == end_sync_read) {
2636                                 read_targets--;
2637                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2638                                 generic_make_request(bio);
2639                         }
2640                 }
2641         } else {
2642                 atomic_set(&r1_bio->remaining, 1);
2643                 bio = r1_bio->bios[r1_bio->read_disk];
2644                 md_sync_acct(bio->bi_bdev, nr_sectors);
2645                 generic_make_request(bio);
2646
2647         }
2648         return nr_sectors;
2649 }
2650
2651 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2652 {
2653         if (sectors)
2654                 return sectors;
2655
2656         return mddev->dev_sectors;
2657 }
2658
2659 static struct r1conf *setup_conf(struct mddev *mddev)
2660 {
2661         struct r1conf *conf;
2662         int i;
2663         struct raid1_info *disk;
2664         struct md_rdev *rdev;
2665         int err = -ENOMEM;
2666
2667         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2668         if (!conf)
2669                 goto abort;
2670
2671         conf->mirrors = kzalloc(sizeof(struct raid1_info)
2672                                 * mddev->raid_disks * 2,
2673                                  GFP_KERNEL);
2674         if (!conf->mirrors)
2675                 goto abort;
2676
2677         conf->tmppage = alloc_page(GFP_KERNEL);
2678         if (!conf->tmppage)
2679                 goto abort;
2680
2681         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2682         if (!conf->poolinfo)
2683                 goto abort;
2684         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2685         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2686                                           r1bio_pool_free,
2687                                           conf->poolinfo);
2688         if (!conf->r1bio_pool)
2689                 goto abort;
2690
2691         conf->poolinfo->mddev = mddev;
2692
2693         err = -EINVAL;
2694         spin_lock_init(&conf->device_lock);
2695         rdev_for_each(rdev, mddev) {
2696                 struct request_queue *q;
2697                 int disk_idx = rdev->raid_disk;
2698                 if (disk_idx >= mddev->raid_disks
2699                     || disk_idx < 0)
2700                         continue;
2701                 if (test_bit(Replacement, &rdev->flags))
2702                         disk = conf->mirrors + conf->raid_disks + disk_idx;
2703                 else
2704                         disk = conf->mirrors + disk_idx;
2705
2706                 if (disk->rdev)
2707                         goto abort;
2708                 disk->rdev = rdev;
2709                 q = bdev_get_queue(rdev->bdev);
2710                 if (q->merge_bvec_fn)
2711                         mddev->merge_check_needed = 1;
2712
2713                 disk->head_position = 0;
2714                 disk->seq_start = MaxSector;
2715         }
2716         conf->raid_disks = mddev->raid_disks;
2717         conf->mddev = mddev;
2718         INIT_LIST_HEAD(&conf->retry_list);
2719
2720         spin_lock_init(&conf->resync_lock);
2721         init_waitqueue_head(&conf->wait_barrier);
2722
2723         bio_list_init(&conf->pending_bio_list);
2724         conf->pending_count = 0;
2725         conf->recovery_disabled = mddev->recovery_disabled - 1;
2726
2727         err = -EIO;
2728         for (i = 0; i < conf->raid_disks * 2; i++) {
2729
2730                 disk = conf->mirrors + i;
2731
2732                 if (i < conf->raid_disks &&
2733                     disk[conf->raid_disks].rdev) {
2734                         /* This slot has a replacement. */
2735                         if (!disk->rdev) {
2736                                 /* No original, just make the replacement
2737                                  * a recovering spare
2738                                  */
2739                                 disk->rdev =
2740                                         disk[conf->raid_disks].rdev;
2741                                 disk[conf->raid_disks].rdev = NULL;
2742                         } else if (!test_bit(In_sync, &disk->rdev->flags))
2743                                 /* Original is not in_sync - bad */
2744                                 goto abort;
2745                 }
2746
2747                 if (!disk->rdev ||
2748                     !test_bit(In_sync, &disk->rdev->flags)) {
2749                         disk->head_position = 0;
2750                         if (disk->rdev &&
2751                             (disk->rdev->saved_raid_disk < 0))
2752                                 conf->fullsync = 1;
2753                 }
2754         }
2755
2756         err = -ENOMEM;
2757         conf->thread = md_register_thread(raid1d, mddev, "raid1");
2758         if (!conf->thread) {
2759                 printk(KERN_ERR
2760                        "md/raid1:%s: couldn't allocate thread\n",
2761                        mdname(mddev));
2762                 goto abort;
2763         }
2764
2765         return conf;
2766
2767  abort:
2768         if (conf) {
2769                 if (conf->r1bio_pool)
2770                         mempool_destroy(conf->r1bio_pool);
2771                 kfree(conf->mirrors);
2772                 safe_put_page(conf->tmppage);
2773                 kfree(conf->poolinfo);
2774                 kfree(conf);
2775         }
2776         return ERR_PTR(err);
2777 }
2778
2779 static int stop(struct mddev *mddev);
2780 static int run(struct mddev *mddev)
2781 {
2782         struct r1conf *conf;
2783         int i;
2784         struct md_rdev *rdev;
2785         int ret;
2786
2787         if (mddev->level != 1) {
2788                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2789                        mdname(mddev), mddev->level);
2790                 return -EIO;
2791         }
2792         if (mddev->reshape_position != MaxSector) {
2793                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2794                        mdname(mddev));
2795                 return -EIO;
2796         }
2797         /*
2798          * copy the already verified devices into our private RAID1
2799          * bookkeeping area. [whatever we allocate in run(),
2800          * should be freed in stop()]
2801          */
2802         if (mddev->private == NULL)
2803                 conf = setup_conf(mddev);
2804         else
2805                 conf = mddev->private;
2806
2807         if (IS_ERR(conf))
2808                 return PTR_ERR(conf);
2809
2810         rdev_for_each(rdev, mddev) {
2811                 if (!mddev->gendisk)
2812                         continue;
2813                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2814                                   rdev->data_offset << 9);
2815         }
2816
2817         mddev->degraded = 0;
2818         for (i=0; i < conf->raid_disks; i++)
2819                 if (conf->mirrors[i].rdev == NULL ||
2820                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2821                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2822                         mddev->degraded++;
2823
2824         if (conf->raid_disks - mddev->degraded == 1)
2825                 mddev->recovery_cp = MaxSector;
2826
2827         if (mddev->recovery_cp != MaxSector)
2828                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2829                        " -- starting background reconstruction\n",
2830                        mdname(mddev));
2831         printk(KERN_INFO
2832                 "md/raid1:%s: active with %d out of %d mirrors\n",
2833                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2834                 mddev->raid_disks);
2835
2836         /*
2837          * Ok, everything is just fine now
2838          */
2839         mddev->thread = conf->thread;
2840         conf->thread = NULL;
2841         mddev->private = conf;
2842
2843         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2844
2845         if (mddev->queue) {
2846                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2847                 mddev->queue->backing_dev_info.congested_data = mddev;
2848                 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2849         }
2850
2851         ret =  md_integrity_register(mddev);
2852         if (ret)
2853                 stop(mddev);
2854         return ret;
2855 }
2856
2857 static int stop(struct mddev *mddev)
2858 {
2859         struct r1conf *conf = mddev->private;
2860         struct bitmap *bitmap = mddev->bitmap;
2861
2862         /* wait for behind writes to complete */
2863         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2864                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2865                        mdname(mddev));
2866                 /* need to kick something here to make sure I/O goes? */
2867                 wait_event(bitmap->behind_wait,
2868                            atomic_read(&bitmap->behind_writes) == 0);
2869         }
2870
2871         raise_barrier(conf);
2872         lower_barrier(conf);
2873
2874         md_unregister_thread(&mddev->thread);
2875         if (conf->r1bio_pool)
2876                 mempool_destroy(conf->r1bio_pool);
2877         kfree(conf->mirrors);
2878         kfree(conf->poolinfo);
2879         kfree(conf);
2880         mddev->private = NULL;
2881         return 0;
2882 }
2883
2884 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2885 {
2886         /* no resync is happening, and there is enough space
2887          * on all devices, so we can resize.
2888          * We need to make sure resync covers any new space.
2889          * If the array is shrinking we should possibly wait until
2890          * any io in the removed space completes, but it hardly seems
2891          * worth it.
2892          */
2893         sector_t newsize = raid1_size(mddev, sectors, 0);
2894         if (mddev->external_size &&
2895             mddev->array_sectors > newsize)
2896                 return -EINVAL;
2897         if (mddev->bitmap) {
2898                 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2899                 if (ret)
2900                         return ret;
2901         }
2902         md_set_array_sectors(mddev, newsize);
2903         set_capacity(mddev->gendisk, mddev->array_sectors);
2904         revalidate_disk(mddev->gendisk);
2905         if (sectors > mddev->dev_sectors &&
2906             mddev->recovery_cp > mddev->dev_sectors) {
2907                 mddev->recovery_cp = mddev->dev_sectors;
2908                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2909         }
2910         mddev->dev_sectors = sectors;
2911         mddev->resync_max_sectors = sectors;
2912         return 0;
2913 }
2914
2915 static int raid1_reshape(struct mddev *mddev)
2916 {
2917         /* We need to:
2918          * 1/ resize the r1bio_pool
2919          * 2/ resize conf->mirrors
2920          *
2921          * We allocate a new r1bio_pool if we can.
2922          * Then raise a device barrier and wait until all IO stops.
2923          * Then resize conf->mirrors and swap in the new r1bio pool.
2924          *
2925          * At the same time, we "pack" the devices so that all the missing
2926          * devices have the higher raid_disk numbers.
2927          */
2928         mempool_t *newpool, *oldpool;
2929         struct pool_info *newpoolinfo;
2930         struct raid1_info *newmirrors;
2931         struct r1conf *conf = mddev->private;
2932         int cnt, raid_disks;
2933         unsigned long flags;
2934         int d, d2, err;
2935
2936         /* Cannot change chunk_size, layout, or level */
2937         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2938             mddev->layout != mddev->new_layout ||
2939             mddev->level != mddev->new_level) {
2940                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2941                 mddev->new_layout = mddev->layout;
2942                 mddev->new_level = mddev->level;
2943                 return -EINVAL;
2944         }
2945
2946         err = md_allow_write(mddev);
2947         if (err)
2948                 return err;
2949
2950         raid_disks = mddev->raid_disks + mddev->delta_disks;
2951
2952         if (raid_disks < conf->raid_disks) {
2953                 cnt=0;
2954                 for (d= 0; d < conf->raid_disks; d++)
2955                         if (conf->mirrors[d].rdev)
2956                                 cnt++;
2957                 if (cnt > raid_disks)
2958                         return -EBUSY;
2959         }
2960
2961         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2962         if (!newpoolinfo)
2963                 return -ENOMEM;
2964         newpoolinfo->mddev = mddev;
2965         newpoolinfo->raid_disks = raid_disks * 2;
2966
2967         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2968                                  r1bio_pool_free, newpoolinfo);
2969         if (!newpool) {
2970                 kfree(newpoolinfo);
2971                 return -ENOMEM;
2972         }
2973         newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
2974                              GFP_KERNEL);
2975         if (!newmirrors) {
2976                 kfree(newpoolinfo);
2977                 mempool_destroy(newpool);
2978                 return -ENOMEM;
2979         }
2980
2981         raise_barrier(conf);
2982
2983         /* ok, everything is stopped */
2984         oldpool = conf->r1bio_pool;
2985         conf->r1bio_pool = newpool;
2986
2987         for (d = d2 = 0; d < conf->raid_disks; d++) {
2988                 struct md_rdev *rdev = conf->mirrors[d].rdev;
2989                 if (rdev && rdev->raid_disk != d2) {
2990                         sysfs_unlink_rdev(mddev, rdev);
2991                         rdev->raid_disk = d2;
2992                         sysfs_unlink_rdev(mddev, rdev);
2993                         if (sysfs_link_rdev(mddev, rdev))
2994                                 printk(KERN_WARNING
2995                                        "md/raid1:%s: cannot register rd%d\n",
2996                                        mdname(mddev), rdev->raid_disk);
2997                 }
2998                 if (rdev)
2999                         newmirrors[d2++].rdev = rdev;
3000         }
3001         kfree(conf->mirrors);
3002         conf->mirrors = newmirrors;
3003         kfree(conf->poolinfo);
3004         conf->poolinfo = newpoolinfo;
3005
3006         spin_lock_irqsave(&conf->device_lock, flags);
3007         mddev->degraded += (raid_disks - conf->raid_disks);
3008         spin_unlock_irqrestore(&conf->device_lock, flags);
3009         conf->raid_disks = mddev->raid_disks = raid_disks;
3010         mddev->delta_disks = 0;
3011
3012         lower_barrier(conf);
3013
3014         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3015         md_wakeup_thread(mddev->thread);
3016
3017         mempool_destroy(oldpool);
3018         return 0;
3019 }
3020
3021 static void raid1_quiesce(struct mddev *mddev, int state)
3022 {
3023         struct r1conf *conf = mddev->private;
3024
3025         switch(state) {
3026         case 2: /* wake for suspend */
3027                 wake_up(&conf->wait_barrier);
3028                 break;
3029         case 1:
3030                 raise_barrier(conf);
3031                 break;
3032         case 0:
3033                 lower_barrier(conf);
3034                 break;
3035         }
3036 }
3037
3038 static void *raid1_takeover(struct mddev *mddev)
3039 {
3040         /* raid1 can take over:
3041          *  raid5 with 2 devices, any layout or chunk size
3042          */
3043         if (mddev->level == 5 && mddev->raid_disks == 2) {
3044                 struct r1conf *conf;
3045                 mddev->new_level = 1;
3046                 mddev->new_layout = 0;
3047                 mddev->new_chunk_sectors = 0;
3048                 conf = setup_conf(mddev);
3049                 if (!IS_ERR(conf))
3050                         conf->barrier = 1;
3051                 return conf;
3052         }
3053         return ERR_PTR(-EINVAL);
3054 }
3055
3056 static struct md_personality raid1_personality =
3057 {
3058         .name           = "raid1",
3059         .level          = 1,
3060         .owner          = THIS_MODULE,
3061         .make_request   = make_request,
3062         .run            = run,
3063         .stop           = stop,
3064         .status         = status,
3065         .error_handler  = error,
3066         .hot_add_disk   = raid1_add_disk,
3067         .hot_remove_disk= raid1_remove_disk,
3068         .spare_active   = raid1_spare_active,
3069         .sync_request   = sync_request,
3070         .resize         = raid1_resize,
3071         .size           = raid1_size,
3072         .check_reshape  = raid1_reshape,
3073         .quiesce        = raid1_quiesce,
3074         .takeover       = raid1_takeover,
3075 };
3076
3077 static int __init raid_init(void)
3078 {
3079         return register_md_personality(&raid1_personality);
3080 }
3081
3082 static void raid_exit(void)
3083 {
3084         unregister_md_personality(&raid1_personality);
3085 }
3086
3087 module_init(raid_init);
3088 module_exit(raid_exit);
3089 MODULE_LICENSE("GPL");
3090 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3091 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3092 MODULE_ALIAS("md-raid1");
3093 MODULE_ALIAS("md-level-1");
3094
3095 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);