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/seq_file.h>
  38 #include <linux/ratelimit.h>
  39 #include "md.h"
  40 #include "raid1.h"
  41 #include "bitmap.h"
  42
  43 /*
  44  * Number of guaranteed r1bios in case of extreme VM load:
  45  */
  46 #define NR_RAID1_BIOS 256
  47
  48 /* When there are this many requests queue to be written by
  49  * the raid1 thread, we become 'congested' to provide back-pressure
  50  * for writeback.
  51  */
  52 static int max_queued_requests = 1024;
  53
  54 static void allow_barrier(struct r1conf *conf);
  55 static void lower_barrier(struct r1conf *conf);
  56
  57 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  58 {
  59         struct pool_info *pi = data;
  60         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
  61
  62         /* allocate a r1bio with room for raid_disks entries in the bios array */
  63         return kzalloc(size, gfp_flags);
  64 }
  65
  66 static void r1bio_pool_free(void *r1_bio, void *data)
  67 {
  68         kfree(r1_bio);
  69 }
  70
  71 #define RESYNC_BLOCK_SIZE (64*1024)
  72 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  73 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  74 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  75 #define RESYNC_WINDOW (2048*1024)
  76
  77 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  78 {
  79         struct pool_info *pi = data;
  80         struct page *page;
  81         struct r1bio *r1_bio;
  82         struct bio *bio;
  83         int i, j;
  84
  85         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  86         if (!r1_bio)
  87                 return NULL;
  88
  89         /*
  90          * Allocate bios : 1 for reading, n-1 for writing
  91          */
  92         for (j = pi->raid_disks ; j-- ; ) {
  93                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
  94                 if (!bio)
  95                         goto out_free_bio;
  96                 r1_bio->bios[j] = bio;
  97         }
  98         /*
  99          * Allocate RESYNC_PAGES data pages and attach them to
 100          * the first bio.
 101          * If this is a user-requested check/repair, allocate
 102          * RESYNC_PAGES for each bio.
 103          */
 104         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 105                 j = pi->raid_disks;
 106         else
 107                 j = 1;
 108         while(j--) {
 109                 bio = r1_bio->bios[j];
 110                 for (i = 0; i < RESYNC_PAGES; i++) {
 111                         page = alloc_page(gfp_flags);
 112                         if (unlikely(!page))
 113                                 goto out_free_pages;
 114
 115                         bio->bi_io_vec[i].bv_page = page;
 116                         bio->bi_vcnt = i+1;
 117                 }
 118         }
 119         /* If not user-requests, copy the page pointers to all bios */
 120         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 121                 for (i=0; i<RESYNC_PAGES ; i++)
 122                         for (j=1; j<pi->raid_disks; j++)
 123                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 124                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 125         }
 126
 127         r1_bio->master_bio = NULL;
 128
 129         return r1_bio;
 130
 131 out_free_pages:
 132         for (j=0 ; j < pi->raid_disks; j++)
 133                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 134                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 135         j = -1;
 136 out_free_bio:
 137         while ( ++j < pi->raid_disks )
 138                 bio_put(r1_bio->bios[j]);
 139         r1bio_pool_free(r1_bio, data);
 140         return NULL;
 141 }
 142
 143 static void r1buf_pool_free(void *__r1_bio, void *data)
 144 {
 145         struct pool_info *pi = data;
 146         int i,j;
 147         struct r1bio *r1bio = __r1_bio;
 148
 149         for (i = 0; i < RESYNC_PAGES; i++)
 150                 for (j = pi->raid_disks; j-- ;) {
 151                         if (j == 0 ||
 152                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 153                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 154                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 155                 }
 156         for (i=0 ; i < pi->raid_disks; i++)
 157                 bio_put(r1bio->bios[i]);
 158
 159         r1bio_pool_free(r1bio, data);
 160 }
 161
 162 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 163 {
 164         int i;
 165
 166         for (i = 0; i < conf->raid_disks; i++) {
 167                 struct bio **bio = r1_bio->bios + i;
 168                 if (!BIO_SPECIAL(*bio))
 169                         bio_put(*bio);
 170                 *bio = NULL;
 171         }
 172 }
 173
 174 static void free_r1bio(struct r1bio *r1_bio)
 175 {
 176         struct r1conf *conf = r1_bio->mddev->private;
 177
 178         put_all_bios(conf, r1_bio);
 179         mempool_free(r1_bio, conf->r1bio_pool);
 180 }
 181
 182 static void put_buf(struct r1bio *r1_bio)
 183 {
 184         struct r1conf *conf = r1_bio->mddev->private;
 185         int i;
 186
 187         for (i=0; i<conf->raid_disks; i++) {
 188                 struct bio *bio = r1_bio->bios[i];
 189                 if (bio->bi_end_io)
 190                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 191         }
 192
 193         mempool_free(r1_bio, conf->r1buf_pool);
 194
 195         lower_barrier(conf);
 196 }
 197
 198 static void reschedule_retry(struct r1bio *r1_bio)
 199 {
 200         unsigned long flags;
 201         struct mddev *mddev = r1_bio->mddev;
 202         struct r1conf *conf = mddev->private;
 203
 204         spin_lock_irqsave(&conf->device_lock, flags);
 205         list_add(&r1_bio->retry_list, &conf->retry_list);
 206         conf->nr_queued ++;
 207         spin_unlock_irqrestore(&conf->device_lock, flags);
 208
 209         wake_up(&conf->wait_barrier);
 210         md_wakeup_thread(mddev->thread);
 211 }
 212
 213 /*
 214  * raid_end_bio_io() is called when we have finished servicing a mirrored
 215  * operation and are ready to return a success/failure code to the buffer
 216  * cache layer.
 217  */
 218 static void call_bio_endio(struct r1bio *r1_bio)
 219 {
 220         struct bio *bio = r1_bio->master_bio;
 221         int done;
 222         struct r1conf *conf = r1_bio->mddev->private;
 223
 224         if (bio->bi_phys_segments) {
 225                 unsigned long flags;
 226                 spin_lock_irqsave(&conf->device_lock, flags);
 227                 bio->bi_phys_segments--;
 228                 done = (bio->bi_phys_segments == 0);
 229                 spin_unlock_irqrestore(&conf->device_lock, flags);
 230         } else
 231                 done = 1;
 232
 233         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 234                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 235         if (done) {
 236                 bio_endio(bio, 0);
 237                 /*
 238                  * Wake up any possible resync thread that waits for the device
 239                  * to go idle.
 240                  */
 241                 allow_barrier(conf);
 242         }
 243 }
 244
 245 static void raid_end_bio_io(struct r1bio *r1_bio)
 246 {
 247         struct bio *bio = r1_bio->master_bio;
 248
 249         /* if nobody has done the final endio yet, do it now */
 250         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 251                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 252                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 253                          (unsigned long long) bio->bi_sector,
 254                          (unsigned long long) bio->bi_sector +
 255                          (bio->bi_size >> 9) - 1);
 256
 257                 call_bio_endio(r1_bio);
 258         }
 259         free_r1bio(r1_bio);
 260 }
 261
 262 /*
 263  * Update disk head position estimator based on IRQ completion info.
 264  */
 265 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 266 {
 267         struct r1conf *conf = r1_bio->mddev->private;
 268
 269         conf->mirrors[disk].head_position =
 270                 r1_bio->sector + (r1_bio->sectors);
 271 }
 272
 273 /*
 274  * Find the disk number which triggered given bio
 275  */
 276 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 277 {
 278         int mirror;
 279         int raid_disks = r1_bio->mddev->raid_disks;
 280
 281         for (mirror = 0; mirror < raid_disks; mirror++)
 282                 if (r1_bio->bios[mirror] == bio)
 283                         break;
 284
 285         BUG_ON(mirror == raid_disks);
 286         update_head_pos(mirror, r1_bio);
 287
 288         return mirror;
 289 }
 290
 291 static void raid1_end_read_request(struct bio *bio, int error)
 292 {
 293         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 294         struct r1bio *r1_bio = bio->bi_private;
 295         int mirror;
 296         struct r1conf *conf = r1_bio->mddev->private;
 297
 298         mirror = r1_bio->read_disk;
 299         /*
 300          * this branch is our 'one mirror IO has finished' event handler:
 301          */
 302         update_head_pos(mirror, r1_bio);
 303
 304         if (uptodate)
 305                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 306         else {
 307                 /* If all other devices have failed, we want to return
 308                  * the error upwards rather than fail the last device.
 309                  * Here we redefine "uptodate" to mean "Don't want to retry"
 310                  */
 311                 unsigned long flags;
 312                 spin_lock_irqsave(&conf->device_lock, flags);
 313                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 314                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 315                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 316                         uptodate = 1;
 317                 spin_unlock_irqrestore(&conf->device_lock, flags);
 318         }
 319
 320         if (uptodate)
 321                 raid_end_bio_io(r1_bio);
 322         else {
 323                 /*
 324                  * oops, read error:
 325                  */
 326                 char b[BDEVNAME_SIZE];
 327                 printk_ratelimited(
 328                         KERN_ERR "md/raid1:%s: %s: "
 329                         "rescheduling sector %llu\n",
 330                         mdname(conf->mddev),
 331                         bdevname(conf->mirrors[mirror].rdev->bdev,
 332                                  b),
 333                         (unsigned long long)r1_bio->sector);
 334                 set_bit(R1BIO_ReadError, &r1_bio->state);
 335                 reschedule_retry(r1_bio);
 336         }
 337
 338         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 339 }
 340
 341 static void close_write(struct r1bio *r1_bio)
 342 {
 343         /* it really is the end of this request */
 344         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 345                 /* free extra copy of the data pages */
 346                 int i = r1_bio->behind_page_count;
 347                 while (i--)
 348                         safe_put_page(r1_bio->behind_bvecs[i].bv_page);
 349                 kfree(r1_bio->behind_bvecs);
 350                 r1_bio->behind_bvecs = NULL;
 351         }
 352         /* clear the bitmap if all writes complete successfully */
 353         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 354                         r1_bio->sectors,
 355                         !test_bit(R1BIO_Degraded, &r1_bio->state),
 356                         test_bit(R1BIO_BehindIO, &r1_bio->state));
 357         md_write_end(r1_bio->mddev);
 358 }
 359
 360 static void r1_bio_write_done(struct r1bio *r1_bio)
 361 {
 362         if (!atomic_dec_and_test(&r1_bio->remaining))
 363                 return;
 364
 365         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 366                 reschedule_retry(r1_bio);
 367         else {
 368                 close_write(r1_bio);
 369                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 370                         reschedule_retry(r1_bio);
 371                 else
 372                         raid_end_bio_io(r1_bio);
 373         }
 374 }
 375
 376 static void raid1_end_write_request(struct bio *bio, int error)
 377 {
 378         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 379         struct r1bio *r1_bio = bio->bi_private;
 380         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 381         struct r1conf *conf = r1_bio->mddev->private;
 382         struct bio *to_put = NULL;
 383
 384         mirror = find_bio_disk(r1_bio, bio);
 385
 386         /*
 387          * 'one mirror IO has finished' event handler:
 388          */
 389         if (!uptodate) {
 390                 set_bit(WriteErrorSeen,
 391                         &conf->mirrors[mirror].rdev->flags);
 392                 set_bit(R1BIO_WriteError, &r1_bio->state);
 393         } else {
 394                 /*
 395                  * Set R1BIO_Uptodate in our master bio, so that we
 396                  * will return a good error code for to the higher
 397                  * levels even if IO on some other mirrored buffer
 398                  * fails.
 399                  *
 400                  * The 'master' represents the composite IO operation
 401                  * to user-side. So if something waits for IO, then it
 402                  * will wait for the 'master' bio.
 403                  */
 404                 sector_t first_bad;
 405                 int bad_sectors;
 406
 407                 r1_bio->bios[mirror] = NULL;
 408                 to_put = bio;
 409                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 410
 411                 /* Maybe we can clear some bad blocks. */
 412                 if (is_badblock(conf->mirrors[mirror].rdev,
 413                                 r1_bio->sector, r1_bio->sectors,
 414                                 &first_bad, &bad_sectors)) {
 415                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 416                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 417                 }
 418         }
 419
 420         if (behind) {
 421                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 422                         atomic_dec(&r1_bio->behind_remaining);
 423
 424                 /*
 425                  * In behind mode, we ACK the master bio once the I/O
 426                  * has safely reached all non-writemostly
 427                  * disks. Setting the Returned bit ensures that this
 428                  * gets done only once -- we don't ever want to return
 429                  * -EIO here, instead we'll wait
 430                  */
 431                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 432                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 433                         /* Maybe we can return now */
 434                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 435                                 struct bio *mbio = r1_bio->master_bio;
 436                                 pr_debug("raid1: behind end write sectors"
 437                                          " %llu-%llu\n",
 438                                          (unsigned long long) mbio->bi_sector,
 439                                          (unsigned long long) mbio->bi_sector +
 440                                          (mbio->bi_size >> 9) - 1);
 441                                 call_bio_endio(r1_bio);
 442                         }
 443                 }
 444         }
 445         if (r1_bio->bios[mirror] == NULL)
 446                 rdev_dec_pending(conf->mirrors[mirror].rdev,
 447                                  conf->mddev);
 448
 449         /*
 450          * Let's see if all mirrored write operations have finished
 451          * already.
 452          */
 453         r1_bio_write_done(r1_bio);
 454
 455         if (to_put)
 456                 bio_put(to_put);
 457 }
 458
 459
 460 /*
 461  * This routine returns the disk from which the requested read should
 462  * be done. There is a per-array 'next expected sequential IO' sector
 463  * number - if this matches on the next IO then we use the last disk.
 464  * There is also a per-disk 'last know head position' sector that is
 465  * maintained from IRQ contexts, both the normal and the resync IO
 466  * completion handlers update this position correctly. If there is no
 467  * perfect sequential match then we pick the disk whose head is closest.
 468  *
 469  * If there are 2 mirrors in the same 2 devices, performance degrades
 470  * because position is mirror, not device based.
 471  *
 472  * The rdev for the device selected will have nr_pending incremented.
 473  */
 474 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 475 {
 476         const sector_t this_sector = r1_bio->sector;
 477         int sectors;
 478         int best_good_sectors;
 479         int start_disk;
 480         int best_disk;
 481         int i;
 482         sector_t best_dist;
 483         struct md_rdev *rdev;
 484         int choose_first;
 485
 486         rcu_read_lock();
 487         /*
 488          * Check if we can balance. We can balance on the whole
 489          * device if no resync is going on, or below the resync window.
 490          * We take the first readable disk when above the resync window.
 491          */
 492  retry:
 493         sectors = r1_bio->sectors;
 494         best_disk = -1;
 495         best_dist = MaxSector;
 496         best_good_sectors = 0;
 497
 498         if (conf->mddev->recovery_cp < MaxSector &&
 499             (this_sector + sectors >= conf->next_resync)) {
 500                 choose_first = 1;
 501                 start_disk = 0;
 502         } else {
 503                 choose_first = 0;
 504                 start_disk = conf->last_used;
 505         }
 506
 507         for (i = 0 ; i < conf->raid_disks ; i++) {
 508                 sector_t dist;
 509                 sector_t first_bad;
 510                 int bad_sectors;
 511
 512                 int disk = start_disk + i;
 513                 if (disk >= conf->raid_disks)
 514                         disk -= conf->raid_disks;
 515
 516                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 517                 if (r1_bio->bios[disk] == IO_BLOCKED
 518                     || rdev == NULL
 519                     || test_bit(Faulty, &rdev->flags))
 520                         continue;
 521                 if (!test_bit(In_sync, &rdev->flags) &&
 522                     rdev->recovery_offset < this_sector + sectors)
 523                         continue;
 524                 if (test_bit(WriteMostly, &rdev->flags)) {
 525                         /* Don't balance among write-mostly, just
 526                          * use the first as a last resort */
 527                         if (best_disk < 0)
 528                                 best_disk = disk;
 529                         continue;
 530                 }
 531                 /* This is a reasonable device to use.  It might
 532                  * even be best.
 533                  */
 534                 if (is_badblock(rdev, this_sector, sectors,
 535                                 &first_bad, &bad_sectors)) {
 536                         if (best_dist < MaxSector)
 537                                 /* already have a better device */
 538                                 continue;
 539                         if (first_bad <= this_sector) {
 540                                 /* cannot read here. If this is the 'primary'
 541                                  * device, then we must not read beyond
 542                                  * bad_sectors from another device..
 543                                  */
 544                                 bad_sectors -= (this_sector - first_bad);
 545                                 if (choose_first && sectors > bad_sectors)
 546                                         sectors = bad_sectors;
 547                                 if (best_good_sectors > sectors)
 548                                         best_good_sectors = sectors;
 549
 550                         } else {
 551                                 sector_t good_sectors = first_bad - this_sector;
 552                                 if (good_sectors > best_good_sectors) {
 553                                         best_good_sectors = good_sectors;
 554                                         best_disk = disk;
 555                                 }
 556                                 if (choose_first)
 557                                         break;
 558                         }
 559                         continue;
 560                 } else
 561                         best_good_sectors = sectors;
 562
 563                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 564                 if (choose_first
 565                     /* Don't change to another disk for sequential reads */
 566                     || conf->next_seq_sect == this_sector
 567                     || dist == 0
 568                     /* If device is idle, use it */
 569                     || atomic_read(&rdev->nr_pending) == 0) {
 570                         best_disk = disk;
 571                         break;
 572                 }
 573                 if (dist < best_dist) {
 574                         best_dist = dist;
 575                         best_disk = disk;
 576                 }
 577         }
 578
 579         if (best_disk >= 0) {
 580                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 581                 if (!rdev)
 582                         goto retry;
 583                 atomic_inc(&rdev->nr_pending);
 584                 if (test_bit(Faulty, &rdev->flags)) {
 585                         /* cannot risk returning a device that failed
 586                          * before we inc'ed nr_pending
 587                          */
 588                         rdev_dec_pending(rdev, conf->mddev);
 589                         goto retry;
 590                 }
 591                 sectors = best_good_sectors;
 592                 conf->next_seq_sect = this_sector + sectors;
 593                 conf->last_used = best_disk;
 594         }
 595         rcu_read_unlock();
 596         *max_sectors = sectors;
 597
 598         return best_disk;
 599 }
 600
 601 int md_raid1_congested(struct mddev *mddev, int bits)
 602 {
 603         struct r1conf *conf = mddev->private;
 604         int i, ret = 0;
 605
 606         if ((bits & (1 << BDI_async_congested)) &&
 607             conf->pending_count >= max_queued_requests)
 608                 return 1;
 609
 610         rcu_read_lock();
 611         for (i = 0; i < mddev->raid_disks; i++) {
 612                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 613                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 614                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 615
 616                         BUG_ON(!q);
 617
 618                         /* Note the '|| 1' - when read_balance prefers
 619                          * non-congested targets, it can be removed
 620                          */
 621                         if ((bits & (1<<BDI_async_congested)) || 1)
 622                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 623                         else
 624                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 625                 }
 626         }
 627         rcu_read_unlock();
 628         return ret;
 629 }
 630 EXPORT_SYMBOL_GPL(md_raid1_congested);
 631
 632 static int raid1_congested(void *data, int bits)
 633 {
 634         struct mddev *mddev = data;
 635
 636         return mddev_congested(mddev, bits) ||
 637                 md_raid1_congested(mddev, bits);
 638 }
 639
 640 static void flush_pending_writes(struct r1conf *conf)
 641 {
 642         /* Any writes that have been queued but are awaiting
 643          * bitmap updates get flushed here.
 644          */
 645         spin_lock_irq(&conf->device_lock);
 646
 647         if (conf->pending_bio_list.head) {
 648                 struct bio *bio;
 649                 bio = bio_list_get(&conf->pending_bio_list);
 650                 conf->pending_count = 0;
 651                 spin_unlock_irq(&conf->device_lock);
 652                 /* flush any pending bitmap writes to
 653                  * disk before proceeding w/ I/O */
 654                 bitmap_unplug(conf->mddev->bitmap);
 655                 wake_up(&conf->wait_barrier);
 656
 657                 while (bio) { /* submit pending writes */
 658                         struct bio *next = bio->bi_next;
 659                         bio->bi_next = NULL;
 660                         generic_make_request(bio);
 661                         bio = next;
 662                 }
 663         } else
 664                 spin_unlock_irq(&conf->device_lock);
 665 }
 666
 667 /* Barriers....
 668  * Sometimes we need to suspend IO while we do something else,
 669  * either some resync/recovery, or reconfigure the array.
 670  * To do this we raise a 'barrier'.
 671  * The 'barrier' is a counter that can be raised multiple times
 672  * to count how many activities are happening which preclude
 673  * normal IO.
 674  * We can only raise the barrier if there is no pending IO.
 675  * i.e. if nr_pending == 0.
 676  * We choose only to raise the barrier if no-one is waiting for the
 677  * barrier to go down.  This means that as soon as an IO request
 678  * is ready, no other operations which require a barrier will start
 679  * until the IO request has had a chance.
 680  *
 681  * So: regular IO calls 'wait_barrier'.  When that returns there
 682  *    is no backgroup IO happening,  It must arrange to call
 683  *    allow_barrier when it has finished its IO.
 684  * backgroup IO calls must call raise_barrier.  Once that returns
 685  *    there is no normal IO happeing.  It must arrange to call
 686  *    lower_barrier when the particular background IO completes.
 687  */
 688 #define RESYNC_DEPTH 32
 689
 690 static void raise_barrier(struct r1conf *conf)
 691 {
 692         spin_lock_irq(&conf->resync_lock);
 693
 694         /* Wait until no block IO is waiting */
 695         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 696                             conf->resync_lock, );
 697
 698         /* block any new IO from starting */
 699         conf->barrier++;
 700
 701         /* Now wait for all pending IO to complete */
 702         wait_event_lock_irq(conf->wait_barrier,
 703                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 704                             conf->resync_lock, );
 705
 706         spin_unlock_irq(&conf->resync_lock);
 707 }
 708
 709 static void lower_barrier(struct r1conf *conf)
 710 {
 711         unsigned long flags;
 712         BUG_ON(conf->barrier <= 0);
 713         spin_lock_irqsave(&conf->resync_lock, flags);
 714         conf->barrier--;
 715         spin_unlock_irqrestore(&conf->resync_lock, flags);
 716         wake_up(&conf->wait_barrier);
 717 }
 718
 719 static void wait_barrier(struct r1conf *conf)
 720 {
 721         spin_lock_irq(&conf->resync_lock);
 722         if (conf->barrier) {
 723                 conf->nr_waiting++;
 724                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 725                                     conf->resync_lock,
 726                                     );
 727                 conf->nr_waiting--;
 728         }
 729         conf->nr_pending++;
 730         spin_unlock_irq(&conf->resync_lock);
 731 }
 732
 733 static void allow_barrier(struct r1conf *conf)
 734 {
 735         unsigned long flags;
 736         spin_lock_irqsave(&conf->resync_lock, flags);
 737         conf->nr_pending--;
 738         spin_unlock_irqrestore(&conf->resync_lock, flags);
 739         wake_up(&conf->wait_barrier);
 740 }
 741
 742 static void freeze_array(struct r1conf *conf)
 743 {
 744         /* stop syncio and normal IO and wait for everything to
 745          * go quite.
 746          * We increment barrier and nr_waiting, and then
 747          * wait until nr_pending match nr_queued+1
 748          * This is called in the context of one normal IO request
 749          * that has failed. Thus any sync request that might be pending
 750          * will be blocked by nr_pending, and we need to wait for
 751          * pending IO requests to complete or be queued for re-try.
 752          * Thus the number queued (nr_queued) plus this request (1)
 753          * must match the number of pending IOs (nr_pending) before
 754          * we continue.
 755          */
 756         spin_lock_irq(&conf->resync_lock);
 757         conf->barrier++;
 758         conf->nr_waiting++;
 759         wait_event_lock_irq(conf->wait_barrier,
 760                             conf->nr_pending == conf->nr_queued+1,
 761                             conf->resync_lock,
 762                             flush_pending_writes(conf));
 763         spin_unlock_irq(&conf->resync_lock);
 764 }
 765 static void unfreeze_array(struct r1conf *conf)
 766 {
 767         /* reverse the effect of the freeze */
 768         spin_lock_irq(&conf->resync_lock);
 769         conf->barrier--;
 770         conf->nr_waiting--;
 771         wake_up(&conf->wait_barrier);
 772         spin_unlock_irq(&conf->resync_lock);
 773 }
 774
 775
 776 /* duplicate the data pages for behind I/O
 777  */
 778 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
 779 {
 780         int i;
 781         struct bio_vec *bvec;
 782         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
 783                                         GFP_NOIO);
 784         if (unlikely(!bvecs))
 785                 return;
 786
 787         bio_for_each_segment(bvec, bio, i) {
 788                 bvecs[i] = *bvec;
 789                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
 790                 if (unlikely(!bvecs[i].bv_page))
 791                         goto do_sync_io;
 792                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
 793                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 794                 kunmap(bvecs[i].bv_page);
 795                 kunmap(bvec->bv_page);
 796         }
 797         r1_bio->behind_bvecs = bvecs;
 798         r1_bio->behind_page_count = bio->bi_vcnt;
 799         set_bit(R1BIO_BehindIO, &r1_bio->state);
 800         return;
 801
 802 do_sync_io:
 803         for (i = 0; i < bio->bi_vcnt; i++)
 804                 if (bvecs[i].bv_page)
 805                         put_page(bvecs[i].bv_page);
 806         kfree(bvecs);
 807         pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 808 }
 809
 810 static int make_request(struct mddev *mddev, struct bio * bio)
 811 {
 812         struct r1conf *conf = mddev->private;
 813         struct mirror_info *mirror;
 814         struct r1bio *r1_bio;
 815         struct bio *read_bio;
 816         int i, disks;
 817         struct bitmap *bitmap;
 818         unsigned long flags;
 819         const int rw = bio_data_dir(bio);
 820         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 821         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
 822         struct md_rdev *blocked_rdev;
 823         int plugged;
 824         int first_clone;
 825         int sectors_handled;
 826         int max_sectors;
 827
 828         /*
 829          * Register the new request and wait if the reconstruction
 830          * thread has put up a bar for new requests.
 831          * Continue immediately if no resync is active currently.
 832          */
 833
 834         md_write_start(mddev, bio); /* wait on superblock update early */
 835
 836         if (bio_data_dir(bio) == WRITE &&
 837             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
 838             bio->bi_sector < mddev->suspend_hi) {
 839                 /* As the suspend_* range is controlled by
 840                  * userspace, we want an interruptible
 841                  * wait.
 842                  */
 843                 DEFINE_WAIT(w);
 844                 for (;;) {
 845                         flush_signals(current);
 846                         prepare_to_wait(&conf->wait_barrier,
 847                                         &w, TASK_INTERRUPTIBLE);
 848                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
 849                             bio->bi_sector >= mddev->suspend_hi)
 850                                 break;
 851                         schedule();
 852                 }
 853                 finish_wait(&conf->wait_barrier, &w);
 854         }
 855
 856         wait_barrier(conf);
 857
 858         bitmap = mddev->bitmap;
 859
 860         /*
 861          * make_request() can abort the operation when READA is being
 862          * used and no empty request is available.
 863          *
 864          */
 865         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 866
 867         r1_bio->master_bio = bio;
 868         r1_bio->sectors = bio->bi_size >> 9;
 869         r1_bio->state = 0;
 870         r1_bio->mddev = mddev;
 871         r1_bio->sector = bio->bi_sector;
 872
 873         /* We might need to issue multiple reads to different
 874          * devices if there are bad blocks around, so we keep
 875          * track of the number of reads in bio->bi_phys_segments.
 876          * If this is 0, there is only one r1_bio and no locking
 877          * will be needed when requests complete.  If it is
 878          * non-zero, then it is the number of not-completed requests.
 879          */
 880         bio->bi_phys_segments = 0;
 881         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
 882
 883         if (rw == READ) {
 884                 /*
 885                  * read balancing logic:
 886                  */
 887                 int rdisk;
 888
 889 read_again:
 890                 rdisk = read_balance(conf, r1_bio, &max_sectors);
 891
 892                 if (rdisk < 0) {
 893                         /* couldn't find anywhere to read from */
 894                         raid_end_bio_io(r1_bio);
 895                         return 0;
 896                 }
 897                 mirror = conf->mirrors + rdisk;
 898
 899                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
 900                     bitmap) {
 901                         /* Reading from a write-mostly device must
 902                          * take care not to over-take any writes
 903                          * that are 'behind'
 904                          */
 905                         wait_event(bitmap->behind_wait,
 906                                    atomic_read(&bitmap->behind_writes) == 0);
 907                 }
 908                 r1_bio->read_disk = rdisk;
 909
 910                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 911                 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
 912                             max_sectors);
 913
 914                 r1_bio->bios[rdisk] = read_bio;
 915
 916                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 917                 read_bio->bi_bdev = mirror->rdev->bdev;
 918                 read_bio->bi_end_io = raid1_end_read_request;
 919                 read_bio->bi_rw = READ | do_sync;
 920                 read_bio->bi_private = r1_bio;
 921
 922                 if (max_sectors < r1_bio->sectors) {
 923                         /* could not read all from this device, so we will
 924                          * need another r1_bio.
 925                          */
 926
 927                         sectors_handled = (r1_bio->sector + max_sectors
 928                                            - bio->bi_sector);
 929                         r1_bio->sectors = max_sectors;
 930                         spin_lock_irq(&conf->device_lock);
 931                         if (bio->bi_phys_segments == 0)
 932                                 bio->bi_phys_segments = 2;
 933                         else
 934                                 bio->bi_phys_segments++;
 935                         spin_unlock_irq(&conf->device_lock);
 936                         /* Cannot call generic_make_request directly
 937                          * as that will be queued in __make_request
 938                          * and subsequent mempool_alloc might block waiting
 939                          * for it.  So hand bio over to raid1d.
 940                          */
 941                         reschedule_retry(r1_bio);
 942
 943                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 944
 945                         r1_bio->master_bio = bio;
 946                         r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
 947                         r1_bio->state = 0;
 948                         r1_bio->mddev = mddev;
 949                         r1_bio->sector = bio->bi_sector + sectors_handled;
 950                         goto read_again;
 951                 } else
 952                         generic_make_request(read_bio);
 953                 return 0;
 954         }
 955
 956         /*
 957          * WRITE:
 958          */
 959         if (conf->pending_count >= max_queued_requests) {
 960                 md_wakeup_thread(mddev->thread);
 961                 wait_event(conf->wait_barrier,
 962                            conf->pending_count < max_queued_requests);
 963         }
 964         /* first select target devices under rcu_lock and
 965          * inc refcount on their rdev.  Record them by setting
 966          * bios[x] to bio
 967          * If there are known/acknowledged bad blocks on any device on
 968          * which we have seen a write error, we want to avoid writing those
 969          * blocks.
 970          * This potentially requires several writes to write around
 971          * the bad blocks.  Each set of writes gets it's own r1bio
 972          * with a set of bios attached.
 973          */
 974         plugged = mddev_check_plugged(mddev);
 975
 976         disks = conf->raid_disks;
 977  retry_write:
 978         blocked_rdev = NULL;
 979         rcu_read_lock();
 980         max_sectors = r1_bio->sectors;
 981         for (i = 0;  i < disks; i++) {
 982                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 983                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 984                         atomic_inc(&rdev->nr_pending);
 985                         blocked_rdev = rdev;
 986                         break;
 987                 }
 988                 r1_bio->bios[i] = NULL;
 989                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
 990                         set_bit(R1BIO_Degraded, &r1_bio->state);
 991                         continue;
 992                 }
 993
 994                 atomic_inc(&rdev->nr_pending);
 995                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
 996                         sector_t first_bad;
 997                         int bad_sectors;
 998                         int is_bad;
 999
1000                         is_bad = is_badblock(rdev, r1_bio->sector,
1001                                              max_sectors,
1002                                              &first_bad, &bad_sectors);
1003                         if (is_bad < 0) {
1004                                 /* mustn't write here until the bad block is
1005                                  * acknowledged*/
1006                                 set_bit(BlockedBadBlocks, &rdev->flags);
1007                                 blocked_rdev = rdev;
1008                                 break;
1009                         }
1010                         if (is_bad && first_bad <= r1_bio->sector) {
1011                                 /* Cannot write here at all */
1012                                 bad_sectors -= (r1_bio->sector - first_bad);
1013                                 if (bad_sectors < max_sectors)
1014                                         /* mustn't write more than bad_sectors
1015                                          * to other devices yet
1016                                          */
1017                                         max_sectors = bad_sectors;
1018                                 rdev_dec_pending(rdev, mddev);
1019                                 /* We don't set R1BIO_Degraded as that
1020                                  * only applies if the disk is
1021                                  * missing, so it might be re-added,
1022                                  * and we want to know to recover this
1023                                  * chunk.
1024                                  * In this case the device is here,
1025                                  * and the fact that this chunk is not
1026                                  * in-sync is recorded in the bad
1027                                  * block log
1028                                  */
1029                                 continue;
1030                         }
1031                         if (is_bad) {
1032                                 int good_sectors = first_bad - r1_bio->sector;
1033                                 if (good_sectors < max_sectors)
1034                                         max_sectors = good_sectors;
1035                         }
1036                 }
1037                 r1_bio->bios[i] = bio;
1038         }
1039         rcu_read_unlock();
1040
1041         if (unlikely(blocked_rdev)) {
1042                 /* Wait for this device to become unblocked */
1043                 int j;
1044
1045                 for (j = 0; j < i; j++)
1046                         if (r1_bio->bios[j])
1047                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1048                 r1_bio->state = 0;
1049                 allow_barrier(conf);
1050                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1051                 wait_barrier(conf);
1052                 goto retry_write;
1053         }
1054
1055         if (max_sectors < r1_bio->sectors) {
1056                 /* We are splitting this write into multiple parts, so
1057                  * we need to prepare for allocating another r1_bio.
1058                  */
1059                 r1_bio->sectors = max_sectors;
1060                 spin_lock_irq(&conf->device_lock);
1061                 if (bio->bi_phys_segments == 0)
1062                         bio->bi_phys_segments = 2;
1063                 else
1064                         bio->bi_phys_segments++;
1065                 spin_unlock_irq(&conf->device_lock);
1066         }
1067         sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1068
1069         atomic_set(&r1_bio->remaining, 1);
1070         atomic_set(&r1_bio->behind_remaining, 0);
1071
1072         first_clone = 1;
1073         for (i = 0; i < disks; i++) {
1074                 struct bio *mbio;
1075                 if (!r1_bio->bios[i])
1076                         continue;
1077
1078                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1079                 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1080
1081                 if (first_clone) {
1082                         /* do behind I/O ?
1083                          * Not if there are too many, or cannot
1084                          * allocate memory, or a reader on WriteMostly
1085                          * is waiting for behind writes to flush */
1086                         if (bitmap &&
1087                             (atomic_read(&bitmap->behind_writes)
1088                              < mddev->bitmap_info.max_write_behind) &&
1089                             !waitqueue_active(&bitmap->behind_wait))
1090                                 alloc_behind_pages(mbio, r1_bio);
1091
1092                         bitmap_startwrite(bitmap, r1_bio->sector,
1093                                           r1_bio->sectors,
1094                                           test_bit(R1BIO_BehindIO,
1095                                                    &r1_bio->state));
1096                         first_clone = 0;
1097                 }
1098                 if (r1_bio->behind_bvecs) {
1099                         struct bio_vec *bvec;
1100                         int j;
1101
1102                         /* Yes, I really want the '__' version so that
1103                          * we clear any unused pointer in the io_vec, rather
1104                          * than leave them unchanged.  This is important
1105                          * because when we come to free the pages, we won't
1106                          * know the original bi_idx, so we just free
1107                          * them all
1108                          */
1109                         __bio_for_each_segment(bvec, mbio, j, 0)
1110                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1111                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1112                                 atomic_inc(&r1_bio->behind_remaining);
1113                 }
1114
1115                 r1_bio->bios[i] = mbio;
1116
1117                 mbio->bi_sector = (r1_bio->sector +
1118                                    conf->mirrors[i].rdev->data_offset);
1119                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1120                 mbio->bi_end_io = raid1_end_write_request;
1121                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
1122                 mbio->bi_private = r1_bio;
1123
1124                 atomic_inc(&r1_bio->remaining);
1125                 spin_lock_irqsave(&conf->device_lock, flags);
1126                 bio_list_add(&conf->pending_bio_list, mbio);
1127                 conf->pending_count++;
1128                 spin_unlock_irqrestore(&conf->device_lock, flags);
1129         }
1130         /* Mustn't call r1_bio_write_done before this next test,
1131          * as it could result in the bio being freed.
1132          */
1133         if (sectors_handled < (bio->bi_size >> 9)) {
1134                 r1_bio_write_done(r1_bio);
1135                 /* We need another r1_bio.  It has already been counted
1136                  * in bio->bi_phys_segments
1137                  */
1138                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1139                 r1_bio->master_bio = bio;
1140                 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1141                 r1_bio->state = 0;
1142                 r1_bio->mddev = mddev;
1143                 r1_bio->sector = bio->bi_sector + sectors_handled;
1144                 goto retry_write;
1145         }
1146
1147         r1_bio_write_done(r1_bio);
1148
1149         /* In case raid1d snuck in to freeze_array */
1150         wake_up(&conf->wait_barrier);
1151
1152         if (do_sync || !bitmap || !plugged)
1153                 md_wakeup_thread(mddev->thread);
1154
1155         return 0;
1156 }
1157
1158 static void status(struct seq_file *seq, struct mddev *mddev)
1159 {
1160         struct r1conf *conf = mddev->private;
1161         int i;
1162
1163         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1164                    conf->raid_disks - mddev->degraded);
1165         rcu_read_lock();
1166         for (i = 0; i < conf->raid_disks; i++) {
1167                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1168                 seq_printf(seq, "%s",
1169                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1170         }
1171         rcu_read_unlock();
1172         seq_printf(seq, "]");
1173 }
1174
1175
1176 static void error(struct mddev *mddev, struct md_rdev *rdev)
1177 {
1178         char b[BDEVNAME_SIZE];
1179         struct r1conf *conf = mddev->private;
1180
1181         /*
1182          * If it is not operational, then we have already marked it as dead
1183          * else if it is the last working disks, ignore the error, let the
1184          * next level up know.
1185          * else mark the drive as failed
1186          */
1187         if (test_bit(In_sync, &rdev->flags)
1188             && (conf->raid_disks - mddev->degraded) == 1) {
1189                 /*
1190                  * Don't fail the drive, act as though we were just a
1191                  * normal single drive.
1192                  * However don't try a recovery from this drive as
1193                  * it is very likely to fail.
1194                  */
1195                 conf->recovery_disabled = mddev->recovery_disabled;
1196                 return;
1197         }
1198         set_bit(Blocked, &rdev->flags);
1199         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1200                 unsigned long flags;
1201                 spin_lock_irqsave(&conf->device_lock, flags);
1202                 mddev->degraded++;
1203                 set_bit(Faulty, &rdev->flags);
1204                 spin_unlock_irqrestore(&conf->device_lock, flags);
1205                 /*
1206                  * if recovery is running, make sure it aborts.
1207                  */
1208                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1209         } else
1210                 set_bit(Faulty, &rdev->flags);
1211         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1212         printk(KERN_ALERT
1213                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1214                "md/raid1:%s: Operation continuing on %d devices.\n",
1215                mdname(mddev), bdevname(rdev->bdev, b),
1216                mdname(mddev), conf->raid_disks - mddev->degraded);
1217 }
1218
1219 static void print_conf(struct r1conf *conf)
1220 {
1221         int i;
1222
1223         printk(KERN_DEBUG "RAID1 conf printout:\n");
1224         if (!conf) {
1225                 printk(KERN_DEBUG "(!conf)\n");
1226                 return;
1227         }
1228         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1229                 conf->raid_disks);
1230
1231         rcu_read_lock();
1232         for (i = 0; i < conf->raid_disks; i++) {
1233                 char b[BDEVNAME_SIZE];
1234                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1235                 if (rdev)
1236                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1237                                i, !test_bit(In_sync, &rdev->flags),
1238                                !test_bit(Faulty, &rdev->flags),
1239                                bdevname(rdev->bdev,b));
1240         }
1241         rcu_read_unlock();
1242 }
1243
1244 static void close_sync(struct r1conf *conf)
1245 {
1246         wait_barrier(conf);
1247         allow_barrier(conf);
1248
1249         mempool_destroy(conf->r1buf_pool);
1250         conf->r1buf_pool = NULL;
1251 }
1252
1253 static int raid1_spare_active(struct mddev *mddev)
1254 {
1255         int i;
1256         struct r1conf *conf = mddev->private;
1257         int count = 0;
1258         unsigned long flags;
1259
1260         /*
1261          * Find all failed disks within the RAID1 configuration
1262          * and mark them readable.
1263          * Called under mddev lock, so rcu protection not needed.
1264          */
1265         for (i = 0; i < conf->raid_disks; i++) {
1266                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1267                 if (rdev
1268                     && !test_bit(Faulty, &rdev->flags)
1269                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1270                         count++;
1271                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1272                 }
1273         }
1274         spin_lock_irqsave(&conf->device_lock, flags);
1275         mddev->degraded -= count;
1276         spin_unlock_irqrestore(&conf->device_lock, flags);
1277
1278         print_conf(conf);
1279         return count;
1280 }
1281
1282
1283 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1284 {
1285         struct r1conf *conf = mddev->private;
1286         int err = -EEXIST;
1287         int mirror = 0;
1288         struct mirror_info *p;
1289         int first = 0;
1290         int last = mddev->raid_disks - 1;
1291
1292         if (mddev->recovery_disabled == conf->recovery_disabled)
1293                 return -EBUSY;
1294
1295         if (rdev->raid_disk >= 0)
1296                 first = last = rdev->raid_disk;
1297
1298         for (mirror = first; mirror <= last; mirror++)
1299                 if ( !(p=conf->mirrors+mirror)->rdev) {
1300
1301                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1302                                           rdev->data_offset << 9);
1303                         /* as we don't honour merge_bvec_fn, we must
1304                          * never risk violating it, so limit
1305                          * ->max_segments to one lying with a single
1306                          * page, as a one page request is never in
1307                          * violation.
1308                          */
1309                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1310                                 blk_queue_max_segments(mddev->queue, 1);
1311                                 blk_queue_segment_boundary(mddev->queue,
1312                                                            PAGE_CACHE_SIZE - 1);
1313                         }
1314
1315                         p->head_position = 0;
1316                         rdev->raid_disk = mirror;
1317                         err = 0;
1318                         /* As all devices are equivalent, we don't need a full recovery
1319                          * if this was recently any drive of the array
1320                          */
1321                         if (rdev->saved_raid_disk < 0)
1322                                 conf->fullsync = 1;
1323                         rcu_assign_pointer(p->rdev, rdev);
1324                         break;
1325                 }
1326         md_integrity_add_rdev(rdev, mddev);
1327         print_conf(conf);
1328         return err;
1329 }
1330
1331 static int raid1_remove_disk(struct mddev *mddev, int number)
1332 {
1333         struct r1conf *conf = mddev->private;
1334         int err = 0;
1335         struct md_rdev *rdev;
1336         struct mirror_info *p = conf->mirrors+ number;
1337
1338         print_conf(conf);
1339         rdev = p->rdev;
1340         if (rdev) {
1341                 if (test_bit(In_sync, &rdev->flags) ||
1342                     atomic_read(&rdev->nr_pending)) {
1343                         err = -EBUSY;
1344                         goto abort;
1345                 }
1346                 /* Only remove non-faulty devices if recovery
1347                  * is not possible.
1348                  */
1349                 if (!test_bit(Faulty, &rdev->flags) &&
1350                     mddev->recovery_disabled != conf->recovery_disabled &&
1351                     mddev->degraded < conf->raid_disks) {
1352                         err = -EBUSY;
1353                         goto abort;
1354                 }
1355                 p->rdev = NULL;
1356                 synchronize_rcu();
1357                 if (atomic_read(&rdev->nr_pending)) {
1358                         /* lost the race, try later */
1359                         err = -EBUSY;
1360                         p->rdev = rdev;
1361                         goto abort;
1362                 }
1363                 err = md_integrity_register(mddev);
1364         }
1365 abort:
1366
1367         print_conf(conf);
1368         return err;
1369 }
1370
1371
1372 static void end_sync_read(struct bio *bio, int error)
1373 {
1374         struct r1bio *r1_bio = bio->bi_private;
1375
1376         update_head_pos(r1_bio->read_disk, r1_bio);
1377
1378         /*
1379          * we have read a block, now it needs to be re-written,
1380          * or re-read if the read failed.
1381          * We don't do much here, just schedule handling by raid1d
1382          */
1383         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1384                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1385
1386         if (atomic_dec_and_test(&r1_bio->remaining))
1387                 reschedule_retry(r1_bio);
1388 }
1389
1390 static void end_sync_write(struct bio *bio, int error)
1391 {
1392         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1393         struct r1bio *r1_bio = bio->bi_private;
1394         struct mddev *mddev = r1_bio->mddev;
1395         struct r1conf *conf = mddev->private;
1396         int mirror=0;
1397         sector_t first_bad;
1398         int bad_sectors;
1399
1400         mirror = find_bio_disk(r1_bio, bio);
1401
1402         if (!uptodate) {
1403                 sector_t sync_blocks = 0;
1404                 sector_t s = r1_bio->sector;
1405                 long sectors_to_go = r1_bio->sectors;
1406                 /* make sure these bits doesn't get cleared. */
1407                 do {
1408                         bitmap_end_sync(mddev->bitmap, s,
1409                                         &sync_blocks, 1);
1410                         s += sync_blocks;
1411                         sectors_to_go -= sync_blocks;
1412                 } while (sectors_to_go > 0);
1413                 set_bit(WriteErrorSeen,
1414                         &conf->mirrors[mirror].rdev->flags);
1415                 set_bit(R1BIO_WriteError, &r1_bio->state);
1416         } else if (is_badblock(conf->mirrors[mirror].rdev,
1417                                r1_bio->sector,
1418                                r1_bio->sectors,
1419                                &first_bad, &bad_sectors) &&
1420                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1421                                 r1_bio->sector,
1422                                 r1_bio->sectors,
1423                                 &first_bad, &bad_sectors)
1424                 )
1425                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1426
1427         if (atomic_dec_and_test(&r1_bio->remaining)) {
1428                 int s = r1_bio->sectors;
1429                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1430                     test_bit(R1BIO_WriteError, &r1_bio->state))
1431                         reschedule_retry(r1_bio);
1432                 else {
1433                         put_buf(r1_bio);
1434                         md_done_sync(mddev, s, uptodate);
1435                 }
1436         }
1437 }
1438
1439 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1440                             int sectors, struct page *page, int rw)
1441 {
1442         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1443                 /* success */
1444                 return 1;
1445         if (rw == WRITE)
1446                 set_bit(WriteErrorSeen, &rdev->flags);
1447         /* need to record an error - either for the block or the device */
1448         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1449                 md_error(rdev->mddev, rdev);
1450         return 0;
1451 }
1452
1453 static int fix_sync_read_error(struct r1bio *r1_bio)
1454 {
1455         /* Try some synchronous reads of other devices to get
1456          * good data, much like with normal read errors.  Only
1457          * read into the pages we already have so we don't
1458          * need to re-issue the read request.
1459          * We don't need to freeze the array, because being in an
1460          * active sync request, there is no normal IO, and
1461          * no overlapping syncs.
1462          * We don't need to check is_badblock() again as we
1463          * made sure that anything with a bad block in range
1464          * will have bi_end_io clear.
1465          */
1466         struct mddev *mddev = r1_bio->mddev;
1467         struct r1conf *conf = mddev->private;
1468         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1469         sector_t sect = r1_bio->sector;
1470         int sectors = r1_bio->sectors;
1471         int idx = 0;
1472
1473         while(sectors) {
1474                 int s = sectors;
1475                 int d = r1_bio->read_disk;
1476                 int success = 0;
1477                 struct md_rdev *rdev;
1478                 int start;
1479
1480                 if (s > (PAGE_SIZE>>9))
1481                         s = PAGE_SIZE >> 9;
1482                 do {
1483                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1484                                 /* No rcu protection needed here devices
1485                                  * can only be removed when no resync is
1486                                  * active, and resync is currently active
1487                                  */
1488                                 rdev = conf->mirrors[d].rdev;
1489                                 if (sync_page_io(rdev, sect, s<<9,
1490                                                  bio->bi_io_vec[idx].bv_page,
1491                                                  READ, false)) {
1492                                         success = 1;
1493                                         break;
1494                                 }
1495                         }
1496                         d++;
1497                         if (d == conf->raid_disks)
1498                                 d = 0;
1499                 } while (!success && d != r1_bio->read_disk);
1500
1501                 if (!success) {
1502                         char b[BDEVNAME_SIZE];
1503                         int abort = 0;
1504                         /* Cannot read from anywhere, this block is lost.
1505                          * Record a bad block on each device.  If that doesn't
1506                          * work just disable and interrupt the recovery.
1507                          * Don't fail devices as that won't really help.
1508                          */
1509                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1510                                " for block %llu\n",
1511                                mdname(mddev),
1512                                bdevname(bio->bi_bdev, b),
1513                                (unsigned long long)r1_bio->sector);
1514                         for (d = 0; d < conf->raid_disks; d++) {
1515                                 rdev = conf->mirrors[d].rdev;
1516                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1517                                         continue;
1518                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1519                                         abort = 1;
1520                         }
1521                         if (abort) {
1522                                 conf->recovery_disabled =
1523                                         mddev->recovery_disabled;
1524                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1525                                 md_done_sync(mddev, r1_bio->sectors, 0);
1526                                 put_buf(r1_bio);
1527                                 return 0;
1528                         }
1529                         /* Try next page */
1530                         sectors -= s;
1531                         sect += s;
1532                         idx++;
1533                         continue;
1534                 }
1535
1536                 start = d;
1537                 /* write it back and re-read */
1538                 while (d != r1_bio->read_disk) {
1539                         if (d == 0)
1540                                 d = conf->raid_disks;
1541                         d--;
1542                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1543                                 continue;
1544                         rdev = conf->mirrors[d].rdev;
1545                         if (r1_sync_page_io(rdev, sect, s,
1546                                             bio->bi_io_vec[idx].bv_page,
1547                                             WRITE) == 0) {
1548                                 r1_bio->bios[d]->bi_end_io = NULL;
1549                                 rdev_dec_pending(rdev, mddev);
1550                         }
1551                 }
1552                 d = start;
1553                 while (d != r1_bio->read_disk) {
1554                         if (d == 0)
1555                                 d = conf->raid_disks;
1556                         d--;
1557                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1558                                 continue;
1559                         rdev = conf->mirrors[d].rdev;
1560                         if (r1_sync_page_io(rdev, sect, s,
1561                                             bio->bi_io_vec[idx].bv_page,
1562                                             READ) != 0)
1563                                 atomic_add(s, &rdev->corrected_errors);
1564                 }
1565                 sectors -= s;
1566                 sect += s;
1567                 idx ++;
1568         }
1569         set_bit(R1BIO_Uptodate, &r1_bio->state);
1570         set_bit(BIO_UPTODATE, &bio->bi_flags);
1571         return 1;
1572 }
1573
1574 static int process_checks(struct r1bio *r1_bio)
1575 {
1576         /* We have read all readable devices.  If we haven't
1577          * got the block, then there is no hope left.
1578          * If we have, then we want to do a comparison
1579          * and skip the write if everything is the same.
1580          * If any blocks failed to read, then we need to
1581          * attempt an over-write
1582          */
1583         struct mddev *mddev = r1_bio->mddev;
1584         struct r1conf *conf = mddev->private;
1585         int primary;
1586         int i;
1587
1588         for (primary = 0; primary < conf->raid_disks; primary++)
1589                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1590                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1591                         r1_bio->bios[primary]->bi_end_io = NULL;
1592                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1593                         break;
1594                 }
1595         r1_bio->read_disk = primary;
1596         for (i = 0; i < conf->raid_disks; i++) {
1597                 int j;
1598                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1599                 struct bio *pbio = r1_bio->bios[primary];
1600                 struct bio *sbio = r1_bio->bios[i];
1601                 int size;
1602
1603                 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1604                         continue;
1605
1606                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1607                         for (j = vcnt; j-- ; ) {
1608                                 struct page *p, *s;
1609                                 p = pbio->bi_io_vec[j].bv_page;
1610                                 s = sbio->bi_io_vec[j].bv_page;
1611                                 if (memcmp(page_address(p),
1612                                            page_address(s),
1613                                            PAGE_SIZE))
1614                                         break;
1615                         }
1616                 } else
1617                         j = 0;
1618                 if (j >= 0)
1619                         mddev->resync_mismatches += r1_bio->sectors;
1620                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1621                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1622                         /* No need to write to this device. */
1623                         sbio->bi_end_io = NULL;
1624                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1625                         continue;
1626                 }
1627                 /* fixup the bio for reuse */
1628                 sbio->bi_vcnt = vcnt;
1629                 sbio->bi_size = r1_bio->sectors << 9;
1630                 sbio->bi_idx = 0;
1631                 sbio->bi_phys_segments = 0;
1632                 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1633                 sbio->bi_flags |= 1 << BIO_UPTODATE;
1634                 sbio->bi_next = NULL;
1635                 sbio->bi_sector = r1_bio->sector +
1636                         conf->mirrors[i].rdev->data_offset;
1637                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1638                 size = sbio->bi_size;
1639                 for (j = 0; j < vcnt ; j++) {
1640                         struct bio_vec *bi;
1641                         bi = &sbio->bi_io_vec[j];
1642                         bi->bv_offset = 0;
1643                         if (size > PAGE_SIZE)
1644                                 bi->bv_len = PAGE_SIZE;
1645                         else
1646                                 bi->bv_len = size;
1647                         size -= PAGE_SIZE;
1648                         memcpy(page_address(bi->bv_page),
1649                                page_address(pbio->bi_io_vec[j].bv_page),
1650                                PAGE_SIZE);
1651                 }
1652         }
1653         return 0;
1654 }
1655
1656 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1657 {
1658         struct r1conf *conf = mddev->private;
1659         int i;
1660         int disks = conf->raid_disks;
1661         struct bio *bio, *wbio;
1662
1663         bio = r1_bio->bios[r1_bio->read_disk];
1664
1665         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1666                 /* ouch - failed to read all of that. */
1667                 if (!fix_sync_read_error(r1_bio))
1668                         return;
1669
1670         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1671                 if (process_checks(r1_bio) < 0)
1672                         return;
1673         /*
1674          * schedule writes
1675          */
1676         atomic_set(&r1_bio->remaining, 1);
1677         for (i = 0; i < disks ; i++) {
1678                 wbio = r1_bio->bios[i];
1679                 if (wbio->bi_end_io == NULL ||
1680                     (wbio->bi_end_io == end_sync_read &&
1681                      (i == r1_bio->read_disk ||
1682                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1683                         continue;
1684
1685                 wbio->bi_rw = WRITE;
1686                 wbio->bi_end_io = end_sync_write;
1687                 atomic_inc(&r1_bio->remaining);
1688                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1689
1690                 generic_make_request(wbio);
1691         }
1692
1693         if (atomic_dec_and_test(&r1_bio->remaining)) {
1694                 /* if we're here, all write(s) have completed, so clean up */
1695                 md_done_sync(mddev, r1_bio->sectors, 1);
1696                 put_buf(r1_bio);
1697         }
1698 }
1699
1700 /*
1701  * This is a kernel thread which:
1702  *
1703  *      1.      Retries failed read operations on working mirrors.
1704  *      2.      Updates the raid superblock when problems encounter.
1705  *      3.      Performs writes following reads for array synchronising.
1706  */
1707
1708 static void fix_read_error(struct r1conf *conf, int read_disk,
1709                            sector_t sect, int sectors)
1710 {
1711         struct mddev *mddev = conf->mddev;
1712         while(sectors) {
1713                 int s = sectors;
1714                 int d = read_disk;
1715                 int success = 0;
1716                 int start;
1717                 struct md_rdev *rdev;
1718
1719                 if (s > (PAGE_SIZE>>9))
1720                         s = PAGE_SIZE >> 9;
1721
1722                 do {
1723                         /* Note: no rcu protection needed here
1724                          * as this is synchronous in the raid1d thread
1725                          * which is the thread that might remove
1726                          * a device.  If raid1d ever becomes multi-threaded....
1727                          */
1728                         sector_t first_bad;
1729                         int bad_sectors;
1730
1731                         rdev = conf->mirrors[d].rdev;
1732                         if (rdev &&
1733                             test_bit(In_sync, &rdev->flags) &&
1734                             is_badblock(rdev, sect, s,
1735                                         &first_bad, &bad_sectors) == 0 &&
1736                             sync_page_io(rdev, sect, s<<9,
1737                                          conf->tmppage, READ, false))
1738                                 success = 1;
1739                         else {
1740                                 d++;
1741                                 if (d == conf->raid_disks)
1742                                         d = 0;
1743                         }
1744                 } while (!success && d != read_disk);
1745
1746                 if (!success) {
1747                         /* Cannot read from anywhere - mark it bad */
1748                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
1749                         if (!rdev_set_badblocks(rdev, sect, s, 0))
1750                                 md_error(mddev, rdev);
1751                         break;
1752                 }
1753                 /* write it back and re-read */
1754                 start = d;
1755                 while (d != read_disk) {
1756                         if (d==0)
1757                                 d = conf->raid_disks;
1758                         d--;
1759                         rdev = conf->mirrors[d].rdev;
1760                         if (rdev &&
1761                             test_bit(In_sync, &rdev->flags))
1762                                 r1_sync_page_io(rdev, sect, s,
1763                                                 conf->tmppage, WRITE);
1764                 }
1765                 d = start;
1766                 while (d != read_disk) {
1767                         char b[BDEVNAME_SIZE];
1768                         if (d==0)
1769                                 d = conf->raid_disks;
1770                         d--;
1771                         rdev = conf->mirrors[d].rdev;
1772                         if (rdev &&
1773                             test_bit(In_sync, &rdev->flags)) {
1774                                 if (r1_sync_page_io(rdev, sect, s,
1775                                                     conf->tmppage, READ)) {
1776                                         atomic_add(s, &rdev->corrected_errors);
1777                                         printk(KERN_INFO
1778                                                "md/raid1:%s: read error corrected "
1779                                                "(%d sectors at %llu on %s)\n",
1780                                                mdname(mddev), s,
1781                                                (unsigned long long)(sect +
1782                                                    rdev->data_offset),
1783                                                bdevname(rdev->bdev, b));
1784                                 }
1785                         }
1786                 }
1787                 sectors -= s;
1788                 sect += s;
1789         }
1790 }
1791
1792 static void bi_complete(struct bio *bio, int error)
1793 {
1794         complete((struct completion *)bio->bi_private);
1795 }
1796
1797 static int submit_bio_wait(int rw, struct bio *bio)
1798 {
1799         struct completion event;
1800         rw |= REQ_SYNC;
1801
1802         init_completion(&event);
1803         bio->bi_private = &event;
1804         bio->bi_end_io = bi_complete;
1805         submit_bio(rw, bio);
1806         wait_for_completion(&event);
1807
1808         return test_bit(BIO_UPTODATE, &bio->bi_flags);
1809 }
1810
1811 static int narrow_write_error(struct r1bio *r1_bio, int i)
1812 {
1813         struct mddev *mddev = r1_bio->mddev;
1814         struct r1conf *conf = mddev->private;
1815         struct md_rdev *rdev = conf->mirrors[i].rdev;
1816         int vcnt, idx;
1817         struct bio_vec *vec;
1818
1819         /* bio has the data to be written to device 'i' where
1820          * we just recently had a write error.
1821          * We repeatedly clone the bio and trim down to one block,
1822          * then try the write.  Where the write fails we record
1823          * a bad block.
1824          * It is conceivable that the bio doesn't exactly align with
1825          * blocks.  We must handle this somehow.
1826          *
1827          * We currently own a reference on the rdev.
1828          */
1829
1830         int block_sectors;
1831         sector_t sector;
1832         int sectors;
1833         int sect_to_write = r1_bio->sectors;
1834         int ok = 1;
1835
1836         if (rdev->badblocks.shift < 0)
1837                 return 0;
1838
1839         block_sectors = 1 << rdev->badblocks.shift;
1840         sector = r1_bio->sector;
1841         sectors = ((sector + block_sectors)
1842                    & ~(sector_t)(block_sectors - 1))
1843                 - sector;
1844
1845         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
1846                 vcnt = r1_bio->behind_page_count;
1847                 vec = r1_bio->behind_bvecs;
1848                 idx = 0;
1849                 while (vec[idx].bv_page == NULL)
1850                         idx++;
1851         } else {
1852                 vcnt = r1_bio->master_bio->bi_vcnt;
1853                 vec = r1_bio->master_bio->bi_io_vec;
1854                 idx = r1_bio->master_bio->bi_idx;
1855         }
1856         while (sect_to_write) {
1857                 struct bio *wbio;
1858                 if (sectors > sect_to_write)
1859                         sectors = sect_to_write;
1860                 /* Write at 'sector' for 'sectors'*/
1861
1862                 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
1863                 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
1864                 wbio->bi_sector = r1_bio->sector;
1865                 wbio->bi_rw = WRITE;
1866                 wbio->bi_vcnt = vcnt;
1867                 wbio->bi_size = r1_bio->sectors << 9;
1868                 wbio->bi_idx = idx;
1869
1870                 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
1871                 wbio->bi_sector += rdev->data_offset;
1872                 wbio->bi_bdev = rdev->bdev;
1873                 if (submit_bio_wait(WRITE, wbio) == 0)
1874                         /* failure! */
1875                         ok = rdev_set_badblocks(rdev, sector,
1876                                                 sectors, 0)
1877                                 && ok;
1878
1879                 bio_put(wbio);
1880                 sect_to_write -= sectors;
1881                 sector += sectors;
1882                 sectors = block_sectors;
1883         }
1884         return ok;
1885 }
1886
1887 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
1888 {
1889         int m;
1890         int s = r1_bio->sectors;
1891         for (m = 0; m < conf->raid_disks ; m++) {
1892                 struct md_rdev *rdev = conf->mirrors[m].rdev;
1893                 struct bio *bio = r1_bio->bios[m];
1894                 if (bio->bi_end_io == NULL)
1895                         continue;
1896                 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
1897                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
1898                         rdev_clear_badblocks(rdev, r1_bio->sector, s);
1899                 }
1900                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
1901                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
1902                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
1903                                 md_error(conf->mddev, rdev);
1904                 }
1905         }
1906         put_buf(r1_bio);
1907         md_done_sync(conf->mddev, s, 1);
1908 }
1909
1910 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
1911 {
1912         int m;
1913         for (m = 0; m < conf->raid_disks ; m++)
1914                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
1915                         struct md_rdev *rdev = conf->mirrors[m].rdev;
1916                         rdev_clear_badblocks(rdev,
1917                                              r1_bio->sector,
1918                                              r1_bio->sectors);
1919                         rdev_dec_pending(rdev, conf->mddev);
1920                 } else if (r1_bio->bios[m] != NULL) {
1921                         /* This drive got a write error.  We need to
1922                          * narrow down and record precise write
1923                          * errors.
1924                          */
1925                         if (!narrow_write_error(r1_bio, m)) {
1926                                 md_error(conf->mddev,
1927                                          conf->mirrors[m].rdev);
1928                                 /* an I/O failed, we can't clear the bitmap */
1929                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1930                         }
1931                         rdev_dec_pending(conf->mirrors[m].rdev,
1932                                          conf->mddev);
1933                 }
1934         if (test_bit(R1BIO_WriteError, &r1_bio->state))
1935                 close_write(r1_bio);
1936         raid_end_bio_io(r1_bio);
1937 }
1938
1939 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
1940 {
1941         int disk;
1942         int max_sectors;
1943         struct mddev *mddev = conf->mddev;
1944         struct bio *bio;
1945         char b[BDEVNAME_SIZE];
1946         struct md_rdev *rdev;
1947
1948         clear_bit(R1BIO_ReadError, &r1_bio->state);
1949         /* we got a read error. Maybe the drive is bad.  Maybe just
1950          * the block and we can fix it.
1951          * We freeze all other IO, and try reading the block from
1952          * other devices.  When we find one, we re-write
1953          * and check it that fixes the read error.
1954          * This is all done synchronously while the array is
1955          * frozen
1956          */
1957         if (mddev->ro == 0) {
1958                 freeze_array(conf);
1959                 fix_read_error(conf, r1_bio->read_disk,
1960                                r1_bio->sector, r1_bio->sectors);
1961                 unfreeze_array(conf);
1962         } else
1963                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1964
1965         bio = r1_bio->bios[r1_bio->read_disk];
1966         bdevname(bio->bi_bdev, b);
1967 read_more:
1968         disk = read_balance(conf, r1_bio, &max_sectors);
1969         if (disk == -1) {
1970                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1971                        " read error for block %llu\n",
1972                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
1973                 raid_end_bio_io(r1_bio);
1974         } else {
1975                 const unsigned long do_sync
1976                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
1977                 if (bio) {
1978                         r1_bio->bios[r1_bio->read_disk] =
1979                                 mddev->ro ? IO_BLOCKED : NULL;
1980                         bio_put(bio);
1981                 }
1982                 r1_bio->read_disk = disk;
1983                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
1984                 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
1985                 r1_bio->bios[r1_bio->read_disk] = bio;
1986                 rdev = conf->mirrors[disk].rdev;
1987                 printk_ratelimited(KERN_ERR
1988                                    "md/raid1:%s: redirecting sector %llu"
1989                                    " to other mirror: %s\n",
1990                                    mdname(mddev),
1991                                    (unsigned long long)r1_bio->sector,
1992                                    bdevname(rdev->bdev, b));
1993                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1994                 bio->bi_bdev = rdev->bdev;
1995                 bio->bi_end_io = raid1_end_read_request;
1996                 bio->bi_rw = READ | do_sync;
1997                 bio->bi_private = r1_bio;
1998                 if (max_sectors < r1_bio->sectors) {
1999                         /* Drat - have to split this up more */
2000                         struct bio *mbio = r1_bio->master_bio;
2001                         int sectors_handled = (r1_bio->sector + max_sectors
2002                                                - mbio->bi_sector);
2003                         r1_bio->sectors = max_sectors;
2004                         spin_lock_irq(&conf->device_lock);
2005                         if (mbio->bi_phys_segments == 0)
2006                                 mbio->bi_phys_segments = 2;
2007                         else
2008                                 mbio->bi_phys_segments++;
2009                         spin_unlock_irq(&conf->device_lock);
2010                         generic_make_request(bio);
2011                         bio = NULL;
2012
2013                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2014
2015                         r1_bio->master_bio = mbio;
2016                         r1_bio->sectors = (mbio->bi_size >> 9)
2017                                           - sectors_handled;
2018                         r1_bio->state = 0;
2019                         set_bit(R1BIO_ReadError, &r1_bio->state);
2020                         r1_bio->mddev = mddev;
2021                         r1_bio->sector = mbio->bi_sector + sectors_handled;
2022
2023                         goto read_more;
2024                 } else
2025                         generic_make_request(bio);
2026         }
2027 }
2028
2029 static void raid1d(struct mddev *mddev)
2030 {
2031         struct r1bio *r1_bio;
2032         unsigned long flags;
2033         struct r1conf *conf = mddev->private;
2034         struct list_head *head = &conf->retry_list;
2035         struct blk_plug plug;
2036
2037         md_check_recovery(mddev);
2038
2039         blk_start_plug(&plug);
2040         for (;;) {
2041
2042                 if (atomic_read(&mddev->plug_cnt) == 0)
2043                         flush_pending_writes(conf);
2044
2045                 spin_lock_irqsave(&conf->device_lock, flags);
2046                 if (list_empty(head)) {
2047                         spin_unlock_irqrestore(&conf->device_lock, flags);
2048                         break;
2049                 }
2050                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2051                 list_del(head->prev);
2052                 conf->nr_queued--;
2053                 spin_unlock_irqrestore(&conf->device_lock, flags);
2054
2055                 mddev = r1_bio->mddev;
2056                 conf = mddev->private;
2057                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2058                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2059                             test_bit(R1BIO_WriteError, &r1_bio->state))
2060                                 handle_sync_write_finished(conf, r1_bio);
2061                         else
2062                                 sync_request_write(mddev, r1_bio);
2063                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2064                            test_bit(R1BIO_WriteError, &r1_bio->state))
2065                         handle_write_finished(conf, r1_bio);
2066                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2067                         handle_read_error(conf, r1_bio);
2068                 else
2069                         /* just a partial read to be scheduled from separate
2070                          * context
2071                          */
2072                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2073
2074                 cond_resched();
2075                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2076                         md_check_recovery(mddev);
2077         }
2078         blk_finish_plug(&plug);
2079 }
2080
2081
2082 static int init_resync(struct r1conf *conf)
2083 {
2084         int buffs;
2085
2086         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2087         BUG_ON(conf->r1buf_pool);
2088         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2089                                           conf->poolinfo);
2090         if (!conf->r1buf_pool)
2091                 return -ENOMEM;
2092         conf->next_resync = 0;
2093         return 0;
2094 }
2095
2096 /*
2097  * perform a "sync" on one "block"
2098  *
2099  * We need to make sure that no normal I/O request - particularly write
2100  * requests - conflict with active sync requests.
2101  *
2102  * This is achieved by tracking pending requests and a 'barrier' concept
2103  * that can be installed to exclude normal IO requests.
2104  */
2105
2106 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2107 {
2108         struct r1conf *conf = mddev->private;
2109         struct r1bio *r1_bio;
2110         struct bio *bio;
2111         sector_t max_sector, nr_sectors;
2112         int disk = -1;
2113         int i;
2114         int wonly = -1;
2115         int write_targets = 0, read_targets = 0;
2116         sector_t sync_blocks;
2117         int still_degraded = 0;
2118         int good_sectors = RESYNC_SECTORS;
2119         int min_bad = 0; /* number of sectors that are bad in all devices */
2120
2121         if (!conf->r1buf_pool)
2122                 if (init_resync(conf))
2123                         return 0;
2124
2125         max_sector = mddev->dev_sectors;
2126         if (sector_nr >= max_sector) {
2127                 /* If we aborted, we need to abort the
2128                  * sync on the 'current' bitmap chunk (there will
2129                  * only be one in raid1 resync.
2130                  * We can find the current addess in mddev->curr_resync
2131                  */
2132                 if (mddev->curr_resync < max_sector) /* aborted */
2133                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2134                                                 &sync_blocks, 1);
2135                 else /* completed sync */
2136                         conf->fullsync = 0;
2137
2138                 bitmap_close_sync(mddev->bitmap);
2139                 close_sync(conf);
2140                 return 0;
2141         }
2142
2143         if (mddev->bitmap == NULL &&
2144             mddev->recovery_cp == MaxSector &&
2145             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2146             conf->fullsync == 0) {
2147                 *skipped = 1;
2148                 return max_sector - sector_nr;
2149         }
2150         /* before building a request, check if we can skip these blocks..
2151          * This call the bitmap_start_sync doesn't actually record anything
2152          */
2153         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2154             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2155                 /* We can skip this block, and probably several more */
2156                 *skipped = 1;
2157                 return sync_blocks;
2158         }
2159         /*
2160          * If there is non-resync activity waiting for a turn,
2161          * and resync is going fast enough,
2162          * then let it though before starting on this new sync request.
2163          */
2164         if (!go_faster && conf->nr_waiting)
2165                 msleep_interruptible(1000);
2166
2167         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2168         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2169         raise_barrier(conf);
2170
2171         conf->next_resync = sector_nr;
2172
2173         rcu_read_lock();
2174         /*
2175          * If we get a correctably read error during resync or recovery,
2176          * we might want to read from a different device.  So we
2177          * flag all drives that could conceivably be read from for READ,
2178          * and any others (which will be non-In_sync devices) for WRITE.
2179          * If a read fails, we try reading from something else for which READ
2180          * is OK.
2181          */
2182
2183         r1_bio->mddev = mddev;
2184         r1_bio->sector = sector_nr;
2185         r1_bio->state = 0;
2186         set_bit(R1BIO_IsSync, &r1_bio->state);
2187
2188         for (i=0; i < conf->raid_disks; i++) {
2189                 struct md_rdev *rdev;
2190                 bio = r1_bio->bios[i];
2191
2192                 /* take from bio_init */
2193                 bio->bi_next = NULL;
2194                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
2195                 bio->bi_flags |= 1 << BIO_UPTODATE;
2196                 bio->bi_comp_cpu = -1;
2197                 bio->bi_rw = READ;
2198                 bio->bi_vcnt = 0;
2199                 bio->bi_idx = 0;
2200                 bio->bi_phys_segments = 0;
2201                 bio->bi_size = 0;
2202                 bio->bi_end_io = NULL;
2203                 bio->bi_private = NULL;
2204
2205                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2206                 if (rdev == NULL ||
2207                     test_bit(Faulty, &rdev->flags)) {
2208                         still_degraded = 1;
2209                 } else if (!test_bit(In_sync, &rdev->flags)) {
2210                         bio->bi_rw = WRITE;
2211                         bio->bi_end_io = end_sync_write;
2212                         write_targets ++;
2213                 } else {
2214                         /* may need to read from here */
2215                         sector_t first_bad = MaxSector;
2216                         int bad_sectors;
2217
2218                         if (is_badblock(rdev, sector_nr, good_sectors,
2219                                         &first_bad, &bad_sectors)) {
2220                                 if (first_bad > sector_nr)
2221                                         good_sectors = first_bad - sector_nr;
2222                                 else {
2223                                         bad_sectors -= (sector_nr - first_bad);
2224                                         if (min_bad == 0 ||
2225                                             min_bad > bad_sectors)
2226                                                 min_bad = bad_sectors;
2227                                 }
2228                         }
2229                         if (sector_nr < first_bad) {
2230                                 if (test_bit(WriteMostly, &rdev->flags)) {
2231                                         if (wonly < 0)
2232                                                 wonly = i;
2233                                 } else {
2234                                         if (disk < 0)
2235                                                 disk = i;
2236                                 }
2237                                 bio->bi_rw = READ;
2238                                 bio->bi_end_io = end_sync_read;
2239                                 read_targets++;
2240                         }
2241                 }
2242                 if (bio->bi_end_io) {
2243                         atomic_inc(&rdev->nr_pending);
2244                         bio->bi_sector = sector_nr + rdev->data_offset;
2245                         bio->bi_bdev = rdev->bdev;
2246                         bio->bi_private = r1_bio;
2247                 }
2248         }
2249         rcu_read_unlock();
2250         if (disk < 0)
2251                 disk = wonly;
2252         r1_bio->read_disk = disk;
2253
2254         if (read_targets == 0 && min_bad > 0) {
2255                 /* These sectors are bad on all InSync devices, so we
2256                  * need to mark them bad on all write targets
2257                  */
2258                 int ok = 1;
2259                 for (i = 0 ; i < conf->raid_disks ; i++)
2260                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2261                                 struct md_rdev *rdev =
2262                                         rcu_dereference(conf->mirrors[i].rdev);
2263                                 ok = rdev_set_badblocks(rdev, sector_nr,
2264                                                         min_bad, 0
2265                                         ) && ok;
2266                         }
2267                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2268                 *skipped = 1;
2269                 put_buf(r1_bio);
2270
2271                 if (!ok) {
2272                         /* Cannot record the badblocks, so need to
2273                          * abort the resync.
2274                          * If there are multiple read targets, could just
2275                          * fail the really bad ones ???
2276                          */
2277                         conf->recovery_disabled = mddev->recovery_disabled;
2278                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2279                         return 0;
2280                 } else
2281                         return min_bad;
2282
2283         }
2284         if (min_bad > 0 && min_bad < good_sectors) {
2285                 /* only resync enough to reach the next bad->good
2286                  * transition */
2287                 good_sectors = min_bad;
2288         }
2289
2290         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2291                 /* extra read targets are also write targets */
2292                 write_targets += read_targets-1;
2293
2294         if (write_targets == 0 || read_targets == 0) {
2295                 /* There is nowhere to write, so all non-sync
2296                  * drives must be failed - so we are finished
2297                  */
2298                 sector_t rv = max_sector - sector_nr;
2299                 *skipped = 1;
2300                 put_buf(r1_bio);
2301                 return rv;
2302         }
2303
2304         if (max_sector > mddev->resync_max)
2305                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2306         if (max_sector > sector_nr + good_sectors)
2307                 max_sector = sector_nr + good_sectors;
2308         nr_sectors = 0;
2309         sync_blocks = 0;
2310         do {
2311                 struct page *page;
2312                 int len = PAGE_SIZE;
2313                 if (sector_nr + (len>>9) > max_sector)
2314                         len = (max_sector - sector_nr) << 9;
2315                 if (len == 0)
2316                         break;
2317                 if (sync_blocks == 0) {
2318                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2319                                                &sync_blocks, still_degraded) &&
2320                             !conf->fullsync &&
2321                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2322                                 break;
2323                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2324                         if ((len >> 9) > sync_blocks)
2325                                 len = sync_blocks<<9;
2326                 }
2327
2328                 for (i=0 ; i < conf->raid_disks; i++) {
2329                         bio = r1_bio->bios[i];
2330                         if (bio->bi_end_io) {
2331                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2332                                 if (bio_add_page(bio, page, len, 0) == 0) {
2333                                         /* stop here */
2334                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2335                                         while (i > 0) {
2336                                                 i--;
2337                                                 bio = r1_bio->bios[i];
2338                                                 if (bio->bi_end_io==NULL)
2339                                                         continue;
2340                                                 /* remove last page from this bio */
2341                                                 bio->bi_vcnt--;
2342                                                 bio->bi_size -= len;
2343                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2344                                         }
2345                                         goto bio_full;
2346                                 }
2347                         }
2348                 }
2349                 nr_sectors += len>>9;
2350                 sector_nr += len>>9;
2351                 sync_blocks -= (len>>9);
2352         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2353  bio_full:
2354         r1_bio->sectors = nr_sectors;
2355
2356         /* For a user-requested sync, we read all readable devices and do a
2357          * compare
2358          */
2359         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2360                 atomic_set(&r1_bio->remaining, read_targets);
2361                 for (i=0; i<conf->raid_disks; i++) {
2362                         bio = r1_bio->bios[i];
2363                         if (bio->bi_end_io == end_sync_read) {
2364                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2365                                 generic_make_request(bio);
2366                         }
2367                 }
2368         } else {
2369                 atomic_set(&r1_bio->remaining, 1);
2370                 bio = r1_bio->bios[r1_bio->read_disk];
2371                 md_sync_acct(bio->bi_bdev, nr_sectors);
2372                 generic_make_request(bio);
2373
2374         }
2375         return nr_sectors;
2376 }
2377
2378 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2379 {
2380         if (sectors)
2381                 return sectors;
2382
2383         return mddev->dev_sectors;
2384 }
2385
2386 static struct r1conf *setup_conf(struct mddev *mddev)
2387 {
2388         struct r1conf *conf;
2389         int i;
2390         struct mirror_info *disk;
2391         struct md_rdev *rdev;
2392         int err = -ENOMEM;
2393
2394         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2395         if (!conf)
2396                 goto abort;
2397
2398         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2399                                  GFP_KERNEL);
2400         if (!conf->mirrors)
2401                 goto abort;
2402
2403         conf->tmppage = alloc_page(GFP_KERNEL);
2404         if (!conf->tmppage)
2405                 goto abort;
2406
2407         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2408         if (!conf->poolinfo)
2409                 goto abort;
2410         conf->poolinfo->raid_disks = mddev->raid_disks;
2411         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2412                                           r1bio_pool_free,
2413                                           conf->poolinfo);
2414         if (!conf->r1bio_pool)
2415                 goto abort;
2416
2417         conf->poolinfo->mddev = mddev;
2418
2419         spin_lock_init(&conf->device_lock);
2420         list_for_each_entry(rdev, &mddev->disks, same_set) {
2421                 int disk_idx = rdev->raid_disk;
2422                 if (disk_idx >= mddev->raid_disks
2423                     || disk_idx < 0)
2424                         continue;
2425                 disk = conf->mirrors + disk_idx;
2426
2427                 disk->rdev = rdev;
2428
2429                 disk->head_position = 0;
2430         }
2431         conf->raid_disks = mddev->raid_disks;
2432         conf->mddev = mddev;
2433         INIT_LIST_HEAD(&conf->retry_list);
2434
2435         spin_lock_init(&conf->resync_lock);
2436         init_waitqueue_head(&conf->wait_barrier);
2437
2438         bio_list_init(&conf->pending_bio_list);
2439         conf->pending_count = 0;
2440         conf->recovery_disabled = mddev->recovery_disabled - 1;
2441
2442         conf->last_used = -1;
2443         for (i = 0; i < conf->raid_disks; i++) {
2444
2445                 disk = conf->mirrors + i;
2446
2447                 if (!disk->rdev ||
2448                     !test_bit(In_sync, &disk->rdev->flags)) {
2449                         disk->head_position = 0;
2450                         if (disk->rdev)
2451                                 conf->fullsync = 1;
2452                 } else if (conf->last_used < 0)
2453                         /*
2454                          * The first working device is used as a
2455                          * starting point to read balancing.
2456                          */
2457                         conf->last_used = i;
2458         }
2459
2460         err = -EIO;
2461         if (conf->last_used < 0) {
2462                 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
2463                        mdname(mddev));
2464                 goto abort;
2465         }
2466         err = -ENOMEM;
2467         conf->thread = md_register_thread(raid1d, mddev, NULL);
2468         if (!conf->thread) {
2469                 printk(KERN_ERR
2470                        "md/raid1:%s: couldn't allocate thread\n",
2471                        mdname(mddev));
2472                 goto abort;
2473         }
2474
2475         return conf;
2476
2477  abort:
2478         if (conf) {
2479                 if (conf->r1bio_pool)
2480                         mempool_destroy(conf->r1bio_pool);
2481                 kfree(conf->mirrors);
2482                 safe_put_page(conf->tmppage);
2483                 kfree(conf->poolinfo);
2484                 kfree(conf);
2485         }
2486         return ERR_PTR(err);
2487 }
2488
2489 static int run(struct mddev *mddev)
2490 {
2491         struct r1conf *conf;
2492         int i;
2493         struct md_rdev *rdev;
2494
2495         if (mddev->level != 1) {
2496                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2497                        mdname(mddev), mddev->level);
2498                 return -EIO;
2499         }
2500         if (mddev->reshape_position != MaxSector) {
2501                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2502                        mdname(mddev));
2503                 return -EIO;
2504         }
2505         /*
2506          * copy the already verified devices into our private RAID1
2507          * bookkeeping area. [whatever we allocate in run(),
2508          * should be freed in stop()]
2509          */
2510         if (mddev->private == NULL)
2511                 conf = setup_conf(mddev);
2512         else
2513                 conf = mddev->private;
2514
2515         if (IS_ERR(conf))
2516                 return PTR_ERR(conf);
2517
2518         list_for_each_entry(rdev, &mddev->disks, same_set) {
2519                 if (!mddev->gendisk)
2520                         continue;
2521                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2522                                   rdev->data_offset << 9);
2523                 /* as we don't honour merge_bvec_fn, we must never risk
2524                  * violating it, so limit ->max_segments to 1 lying within
2525                  * a single page, as a one page request is never in violation.
2526                  */
2527                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2528                         blk_queue_max_segments(mddev->queue, 1);
2529                         blk_queue_segment_boundary(mddev->queue,
2530                                                    PAGE_CACHE_SIZE - 1);
2531                 }
2532         }
2533
2534         mddev->degraded = 0;
2535         for (i=0; i < conf->raid_disks; i++)
2536                 if (conf->mirrors[i].rdev == NULL ||
2537                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2538                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2539                         mddev->degraded++;
2540
2541         if (conf->raid_disks - mddev->degraded == 1)
2542                 mddev->recovery_cp = MaxSector;
2543
2544         if (mddev->recovery_cp != MaxSector)
2545                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2546                        " -- starting background reconstruction\n",
2547                        mdname(mddev));
2548         printk(KERN_INFO
2549                 "md/raid1:%s: active with %d out of %d mirrors\n",
2550                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2551                 mddev->raid_disks);
2552
2553         /*
2554          * Ok, everything is just fine now
2555          */
2556         mddev->thread = conf->thread;
2557         conf->thread = NULL;
2558         mddev->private = conf;
2559
2560         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2561
2562         if (mddev->queue) {
2563                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2564                 mddev->queue->backing_dev_info.congested_data = mddev;
2565         }
2566         return md_integrity_register(mddev);
2567 }
2568
2569 static int stop(struct mddev *mddev)
2570 {
2571         struct r1conf *conf = mddev->private;
2572         struct bitmap *bitmap = mddev->bitmap;
2573
2574         /* wait for behind writes to complete */
2575         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2576                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2577                        mdname(mddev));
2578                 /* need to kick something here to make sure I/O goes? */
2579                 wait_event(bitmap->behind_wait,
2580                            atomic_read(&bitmap->behind_writes) == 0);
2581         }
2582
2583         raise_barrier(conf);
2584         lower_barrier(conf);
2585
2586         md_unregister_thread(&mddev->thread);
2587         if (conf->r1bio_pool)
2588                 mempool_destroy(conf->r1bio_pool);
2589         kfree(conf->mirrors);
2590         kfree(conf->poolinfo);
2591         kfree(conf);
2592         mddev->private = NULL;
2593         return 0;
2594 }
2595
2596 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2597 {
2598         /* no resync is happening, and there is enough space
2599          * on all devices, so we can resize.
2600          * We need to make sure resync covers any new space.
2601          * If the array is shrinking we should possibly wait until
2602          * any io in the removed space completes, but it hardly seems
2603          * worth it.
2604          */
2605         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2606         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2607                 return -EINVAL;
2608         set_capacity(mddev->gendisk, mddev->array_sectors);
2609         revalidate_disk(mddev->gendisk);
2610         if (sectors > mddev->dev_sectors &&
2611             mddev->recovery_cp > mddev->dev_sectors) {
2612                 mddev->recovery_cp = mddev->dev_sectors;
2613                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2614         }
2615         mddev->dev_sectors = sectors;
2616         mddev->resync_max_sectors = sectors;
2617         return 0;
2618 }
2619
2620 static int raid1_reshape(struct mddev *mddev)
2621 {
2622         /* We need to:
2623          * 1/ resize the r1bio_pool
2624          * 2/ resize conf->mirrors
2625          *
2626          * We allocate a new r1bio_pool if we can.
2627          * Then raise a device barrier and wait until all IO stops.
2628          * Then resize conf->mirrors and swap in the new r1bio pool.
2629          *
2630          * At the same time, we "pack" the devices so that all the missing
2631          * devices have the higher raid_disk numbers.
2632          */
2633         mempool_t *newpool, *oldpool;
2634         struct pool_info *newpoolinfo;
2635         struct mirror_info *newmirrors;
2636         struct r1conf *conf = mddev->private;
2637         int cnt, raid_disks;
2638         unsigned long flags;
2639         int d, d2, err;
2640
2641         /* Cannot change chunk_size, layout, or level */
2642         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2643             mddev->layout != mddev->new_layout ||
2644             mddev->level != mddev->new_level) {
2645                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2646                 mddev->new_layout = mddev->layout;
2647                 mddev->new_level = mddev->level;
2648                 return -EINVAL;
2649         }
2650
2651         err = md_allow_write(mddev);
2652         if (err)
2653                 return err;
2654
2655         raid_disks = mddev->raid_disks + mddev->delta_disks;
2656
2657         if (raid_disks < conf->raid_disks) {
2658                 cnt=0;
2659                 for (d= 0; d < conf->raid_disks; d++)
2660                         if (conf->mirrors[d].rdev)
2661                                 cnt++;
2662                 if (cnt > raid_disks)
2663                         return -EBUSY;
2664         }
2665
2666         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2667         if (!newpoolinfo)
2668                 return -ENOMEM;
2669         newpoolinfo->mddev = mddev;
2670         newpoolinfo->raid_disks = raid_disks;
2671
2672         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2673                                  r1bio_pool_free, newpoolinfo);
2674         if (!newpool) {
2675                 kfree(newpoolinfo);
2676                 return -ENOMEM;
2677         }
2678         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2679         if (!newmirrors) {
2680                 kfree(newpoolinfo);
2681                 mempool_destroy(newpool);
2682                 return -ENOMEM;
2683         }
2684
2685         raise_barrier(conf);
2686
2687         /* ok, everything is stopped */
2688         oldpool = conf->r1bio_pool;
2689         conf->r1bio_pool = newpool;
2690
2691         for (d = d2 = 0; d < conf->raid_disks; d++) {
2692                 struct md_rdev *rdev = conf->mirrors[d].rdev;
2693                 if (rdev && rdev->raid_disk != d2) {
2694                         sysfs_unlink_rdev(mddev, rdev);
2695                         rdev->raid_disk = d2;
2696                         sysfs_unlink_rdev(mddev, rdev);
2697                         if (sysfs_link_rdev(mddev, rdev))
2698                                 printk(KERN_WARNING
2699                                        "md/raid1:%s: cannot register rd%d\n",
2700                                        mdname(mddev), rdev->raid_disk);
2701                 }
2702                 if (rdev)
2703                         newmirrors[d2++].rdev = rdev;
2704         }
2705         kfree(conf->mirrors);
2706         conf->mirrors = newmirrors;
2707         kfree(conf->poolinfo);
2708         conf->poolinfo = newpoolinfo;
2709
2710         spin_lock_irqsave(&conf->device_lock, flags);
2711         mddev->degraded += (raid_disks - conf->raid_disks);
2712         spin_unlock_irqrestore(&conf->device_lock, flags);
2713         conf->raid_disks = mddev->raid_disks = raid_disks;
2714         mddev->delta_disks = 0;
2715
2716         conf->last_used = 0; /* just make sure it is in-range */
2717         lower_barrier(conf);
2718
2719         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2720         md_wakeup_thread(mddev->thread);
2721
2722         mempool_destroy(oldpool);
2723         return 0;
2724 }
2725
2726 static void raid1_quiesce(struct mddev *mddev, int state)
2727 {
2728         struct r1conf *conf = mddev->private;
2729
2730         switch(state) {
2731         case 2: /* wake for suspend */
2732                 wake_up(&conf->wait_barrier);
2733                 break;
2734         case 1:
2735                 raise_barrier(conf);
2736                 break;
2737         case 0:
2738                 lower_barrier(conf);
2739                 break;
2740         }
2741 }
2742
2743 static void *raid1_takeover(struct mddev *mddev)
2744 {
2745         /* raid1 can take over:
2746          *  raid5 with 2 devices, any layout or chunk size
2747          */
2748         if (mddev->level == 5 && mddev->raid_disks == 2) {
2749                 struct r1conf *conf;
2750                 mddev->new_level = 1;
2751                 mddev->new_layout = 0;
2752                 mddev->new_chunk_sectors = 0;
2753                 conf = setup_conf(mddev);
2754                 if (!IS_ERR(conf))
2755                         conf->barrier = 1;
2756                 return conf;
2757         }
2758         return ERR_PTR(-EINVAL);
2759 }
2760
2761 static struct md_personality raid1_personality =
2762 {
2763         .name           = "raid1",
2764         .level          = 1,
2765         .owner          = THIS_MODULE,
2766         .make_request   = make_request,
2767         .run            = run,
2768         .stop           = stop,
2769         .status         = status,
2770         .error_handler  = error,
2771         .hot_add_disk   = raid1_add_disk,
2772         .hot_remove_disk= raid1_remove_disk,
2773         .spare_active   = raid1_spare_active,
2774         .sync_request   = sync_request,
2775         .resize         = raid1_resize,
2776         .size           = raid1_size,
2777         .check_reshape  = raid1_reshape,
2778         .quiesce        = raid1_quiesce,
2779         .takeover       = raid1_takeover,
2780 };
2781
2782 static int __init raid_init(void)
2783 {
2784         return register_md_personality(&raid1_personality);
2785 }
2786
2787 static void raid_exit(void)
2788 {
2789         unregister_md_personality(&raid1_personality);
2790 }
2791
2792 module_init(raid_init);
2793 module_exit(raid_exit);
2794 MODULE_LICENSE("GPL");
2795 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2796 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2797 MODULE_ALIAS("md-raid1");
2798 MODULE_ALIAS("md-level-1");
2799
2800 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);