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