drivers/md/raid1.c

   1 /*
   2  * raid1.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5  *
   6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7  *
   8  * RAID-1 management functions.
   9  *
  10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11  *
  12  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14  *
  15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  16  * bitmapped intelligence in resync:
  17  *
  18  *      - bitmap marked during normal i/o
  19  *      - bitmap used to skip nondirty blocks during sync
  20  *
  21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  22  * - persistent bitmap code
  23  *
  24  * This program is free software; you can redistribute it and/or modify
  25  * it under the terms of the GNU General Public License as published by
  26  * the Free Software Foundation; either version 2, or (at your option)
  27  * any later version.
  28  *
  29  * You should have received a copy of the GNU General Public License
  30  * (for example /usr/src/linux/COPYING); if not, write to the Free
  31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  32  */
  33
  34 #include <linux/slab.h>
  35 #include <linux/delay.h>
  36 #include <linux/blkdev.h>
  37 #include <linux/module.h>
  38 #include <linux/seq_file.h>
  39 #include <linux/ratelimit.h>
  40 #include "md.h"
  41 #include "raid1.h"
  42 #include "bitmap.h"
  43
  44 /*
  45  * Number of guaranteed r1bios in case of extreme VM load:
  46  */
  47 #define NR_RAID1_BIOS 256
  48
  49 /* When there are this many requests queue to be written by
  50  * the raid1 thread, we become 'congested' to provide back-pressure
  51  * for writeback.
  52  */
  53 static int max_queued_requests = 1024;
  54
  55 static void allow_barrier(struct r1conf *conf);
  56 static void lower_barrier(struct r1conf *conf);
  57
  58 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  59 {
  60         struct pool_info *pi = data;
  61         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
  62
  63         /* allocate a r1bio with room for raid_disks entries in the bios array */
  64         return kzalloc(size, gfp_flags);
  65 }
  66
  67 static void r1bio_pool_free(void *r1_bio, void *data)
  68 {
  69         kfree(r1_bio);
  70 }
  71
  72 #define RESYNC_BLOCK_SIZE (64*1024)
  73 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  74 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  75 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  76 #define RESYNC_WINDOW (2048*1024)
  77
  78 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  79 {
  80         struct pool_info *pi = data;
  81         struct page *page;
  82         struct r1bio *r1_bio;
  83         struct bio *bio;
  84         int i, j;
  85
  86         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  87         if (!r1_bio)
  88                 return NULL;
  89
  90         /*
  91          * Allocate bios : 1 for reading, n-1 for writing
  92          */
  93         for (j = pi->raid_disks ; j-- ; ) {
  94                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
  95                 if (!bio)
  96                         goto out_free_bio;
  97                 r1_bio->bios[j] = bio;
  98         }
  99         /*
 100          * Allocate RESYNC_PAGES data pages and attach them to
 101          * the first bio.
 102          * If this is a user-requested check/repair, allocate
 103          * RESYNC_PAGES for each bio.
 104          */
 105         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 106                 j = pi->raid_disks;
 107         else
 108                 j = 1;
 109         while(j--) {
 110                 bio = r1_bio->bios[j];
 111                 for (i = 0; i < RESYNC_PAGES; i++) {
 112                         page = alloc_page(gfp_flags);
 113                         if (unlikely(!page))
 114                                 goto out_free_pages;
 115
 116                         bio->bi_io_vec[i].bv_page = page;
 117                         bio->bi_vcnt = i+1;
 118                 }
 119         }
 120         /* If not user-requests, copy the page pointers to all bios */
 121         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 122                 for (i=0; i<RESYNC_PAGES ; i++)
 123                         for (j=1; j<pi->raid_disks; j++)
 124                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 125                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 126         }
 127
 128         r1_bio->master_bio = NULL;
 129
 130         return r1_bio;
 131
 132 out_free_pages:
 133         for (j=0 ; j < pi->raid_disks; j++)
 134                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 135                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 136         j = -1;
 137 out_free_bio:
 138         while ( ++j < pi->raid_disks )
 139                 bio_put(r1_bio->bios[j]);
 140         r1bio_pool_free(r1_bio, data);
 141         return NULL;
 142 }
 143
 144 static void r1buf_pool_free(void *__r1_bio, void *data)
 145 {
 146         struct pool_info *pi = data;
 147         int i,j;
 148         struct r1bio *r1bio = __r1_bio;
 149
 150         for (i = 0; i < RESYNC_PAGES; i++)
 151                 for (j = pi->raid_disks; j-- ;) {
 152                         if (j == 0 ||
 153                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 154                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 155                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 156                 }
 157         for (i=0 ; i < pi->raid_disks; i++)
 158                 bio_put(r1bio->bios[i]);
 159
 160         r1bio_pool_free(r1bio, data);
 161 }
 162
 163 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 164 {
 165         int i;
 166
 167         for (i = 0; i < conf->raid_disks; i++) {
 168                 struct bio **bio = r1_bio->bios + i;
 169                 if (!BIO_SPECIAL(*bio))
 170                         bio_put(*bio);
 171                 *bio = NULL;
 172         }
 173 }
 174
 175 static void free_r1bio(struct r1bio *r1_bio)
 176 {
 177         struct r1conf *conf = r1_bio->mddev->private;
 178
 179         put_all_bios(conf, r1_bio);
 180         mempool_free(r1_bio, conf->r1bio_pool);
 181 }
 182
 183 static void put_buf(struct r1bio *r1_bio)
 184 {
 185         struct r1conf *conf = r1_bio->mddev->private;
 186         int i;
 187
 188         for (i=0; i<conf->raid_disks; i++) {
 189                 struct bio *bio = r1_bio->bios[i];
 190                 if (bio->bi_end_io)
 191                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 192         }
 193
 194         mempool_free(r1_bio, conf->r1buf_pool);
 195
 196         lower_barrier(conf);
 197 }
 198
 199 static void reschedule_retry(struct r1bio *r1_bio)
 200 {
 201         unsigned long flags;
 202         struct mddev *mddev = r1_bio->mddev;
 203         struct r1conf *conf = mddev->private;
 204
 205         spin_lock_irqsave(&conf->device_lock, flags);
 206         list_add(&r1_bio->retry_list, &conf->retry_list);
 207         conf->nr_queued ++;
 208         spin_unlock_irqrestore(&conf->device_lock, flags);
 209
 210         wake_up(&conf->wait_barrier);
 211         md_wakeup_thread(mddev->thread);
 212 }
 213
 214 /*
 215  * raid_end_bio_io() is called when we have finished servicing a mirrored
 216  * operation and are ready to return a success/failure code to the buffer
 217  * cache layer.
 218  */
 219 static void call_bio_endio(struct r1bio *r1_bio)
 220 {
 221         struct bio *bio = r1_bio->master_bio;
 222         int done;
 223         struct r1conf *conf = r1_bio->mddev->private;
 224
 225         if (bio->bi_phys_segments) {
 226                 unsigned long flags;
 227                 spin_lock_irqsave(&conf->device_lock, flags);
 228                 bio->bi_phys_segments--;
 229                 done = (bio->bi_phys_segments == 0);
 230                 spin_unlock_irqrestore(&conf->device_lock, flags);
 231         } else
 232                 done = 1;
 233
 234         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 235                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 236         if (done) {
 237                 bio_endio(bio, 0);
 238                 /*
 239                  * Wake up any possible resync thread that waits for the device
 240                  * to go idle.
 241                  */
 242                 allow_barrier(conf);
 243         }
 244 }
 245
 246 static void raid_end_bio_io(struct r1bio *r1_bio)
 247 {
 248         struct bio *bio = r1_bio->master_bio;
 249
 250         /* if nobody has done the final endio yet, do it now */
 251         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 252                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 253                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 254                          (unsigned long long) bio->bi_sector,
 255                          (unsigned long long) bio->bi_sector +
 256                          (bio->bi_size >> 9) - 1);
 257
 258                 call_bio_endio(r1_bio);
 259         }
 260         free_r1bio(r1_bio);
 261 }
 262
 263 /*
 264  * Update disk head position estimator based on IRQ completion info.
 265  */
 266 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 267 {
 268         struct r1conf *conf = r1_bio->mddev->private;
 269
 270         conf->mirrors[disk].head_position =
 271                 r1_bio->sector + (r1_bio->sectors);
 272 }
 273
 274 /*
 275  * Find the disk number which triggered given bio
 276  */
 277 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 278 {
 279         int mirror;
 280         int raid_disks = r1_bio->mddev->raid_disks;
 281
 282         for (mirror = 0; mirror < raid_disks; mirror++)
 283                 if (r1_bio->bios[mirror] == bio)
 284                         break;
 285
 286         BUG_ON(mirror == raid_disks);
 287         update_head_pos(mirror, r1_bio);
 288
 289         return mirror;
 290 }
 291
 292 static void raid1_end_read_request(struct bio *bio, int error)
 293 {
 294         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 295         struct r1bio *r1_bio = bio->bi_private;
 296         int mirror;
 297         struct r1conf *conf = r1_bio->mddev->private;
 298
 299         mirror = r1_bio->read_disk;
 300         /*
 301          * this branch is our 'one mirror IO has finished' event handler:
 302          */
 303         update_head_pos(mirror, r1_bio);
 304
 305         if (uptodate)
 306                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 307         else {
 308                 /* If all other devices have failed, we want to return
 309                  * the error upwards rather than fail the last device.
 310                  * Here we redefine "uptodate" to mean "Don't want to retry"
 311                  */
 312                 unsigned long flags;
 313                 spin_lock_irqsave(&conf->device_lock, flags);
 314                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 315                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 316                      test_bit(In_sync, &conf->mirrors[mirror].rdev->flags)))
 317                         uptodate = 1;
 318                 spin_unlock_irqrestore(&conf->device_lock, flags);
 319         }
 320
 321         if (uptodate)
 322                 raid_end_bio_io(r1_bio);
 323         else {
 324                 /*
 325                  * oops, read error:
 326                  */
 327                 char b[BDEVNAME_SIZE];
 328                 printk_ratelimited(
 329                         KERN_ERR "md/raid1:%s: %s: "
 330                         "rescheduling sector %llu\n",
 331                         mdname(conf->mddev),
 332                         bdevname(conf->mirrors[mirror].rdev->bdev,
 333                                  b),
 334                         (unsigned long long)r1_bio->sector);
 335                 set_bit(R1BIO_ReadError, &r1_bio->state);
 336                 reschedule_retry(r1_bio);
 337         }
 338
 339         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 340 }
 341
 342 static void close_write(struct r1bio *r1_bio)
 343 {
 344         /* it really is the end of this request */
 345         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 346                 /* free extra copy of the data pages */
 347                 int i = r1_bio->behind_page_count;
 348                 while (i--)
 349                         safe_put_page(r1_bio->behind_bvecs[i].bv_page);
 350                 kfree(r1_bio->behind_bvecs);
 351                 r1_bio->behind_bvecs = NULL;
 352         }
 353         /* clear the bitmap if all writes complete successfully */
 354         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 355                         r1_bio->sectors,
 356                         !test_bit(R1BIO_Degraded, &r1_bio->state),
 357                         test_bit(R1BIO_BehindIO, &r1_bio->state));
 358         md_write_end(r1_bio->mddev);
 359 }
 360
 361 static void r1_bio_write_done(struct r1bio *r1_bio)
 362 {
 363         if (!atomic_dec_and_test(&r1_bio->remaining))
 364                 return;
 365
 366         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 367                 reschedule_retry(r1_bio);
 368         else {
 369                 close_write(r1_bio);
 370                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 371                         reschedule_retry(r1_bio);
 372                 else
 373                         raid_end_bio_io(r1_bio);
 374         }
 375 }
 376
 377 static void raid1_end_write_request(struct bio *bio, int error)
 378 {
 379         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 380         struct r1bio *r1_bio = bio->bi_private;
 381         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 382         struct r1conf *conf = r1_bio->mddev->private;
 383         struct bio *to_put = NULL;
 384
 385         mirror = find_bio_disk(r1_bio, bio);
 386
 387         /*
 388          * 'one mirror IO has finished' event handler:
 389          */
 390         if (!uptodate) {
 391                 set_bit(WriteErrorSeen,
 392                         &conf->mirrors[mirror].rdev->flags);
 393                 set_bit(R1BIO_WriteError, &r1_bio->state);
 394         } else {
 395                 /*
 396                  * Set R1BIO_Uptodate in our master bio, so that we
 397                  * will return a good error code for to the higher
 398                  * levels even if IO on some other mirrored buffer
 399                  * fails.
 400                  *
 401                  * The 'master' represents the composite IO operation
 402                  * to user-side. So if something waits for IO, then it
 403                  * will wait for the 'master' bio.
 404                  */
 405                 sector_t first_bad;
 406                 int bad_sectors;
 407
 408                 r1_bio->bios[mirror] = NULL;
 409                 to_put = bio;
 410                 /*
 411                  * Do not set R1BIO_Uptodate if the current device is
 412                  * rebuilding or Faulty. This is because we cannot use
 413                  * such device for properly reading the data back (we could
 414                  * potentially use it, if the current write would have felt
 415                  * before rdev->recovery_offset, but for simplicity we don't
 416                  * check this here.
 417                  */
 418                 if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) &&
 419                     !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))
 420                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 421
 422                 /* Maybe we can clear some bad blocks. */
 423                 if (is_badblock(conf->mirrors[mirror].rdev,
 424                                 r1_bio->sector, r1_bio->sectors,
 425                                 &first_bad, &bad_sectors)) {
 426                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 427                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 428                 }
 429         }
 430
 431         if (behind) {
 432                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 433                         atomic_dec(&r1_bio->behind_remaining);
 434
 435                 /*
 436                  * In behind mode, we ACK the master bio once the I/O
 437                  * has safely reached all non-writemostly
 438                  * disks. Setting the Returned bit ensures that this
 439                  * gets done only once -- we don't ever want to return
 440                  * -EIO here, instead we'll wait
 441                  */
 442                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 443                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 444                         /* Maybe we can return now */
 445                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 446                                 struct bio *mbio = r1_bio->master_bio;
 447                                 pr_debug("raid1: behind end write sectors"
 448                                          " %llu-%llu\n",
 449                                          (unsigned long long) mbio->bi_sector,
 450                                          (unsigned long long) mbio->bi_sector +
 451                                          (mbio->bi_size >> 9) - 1);
 452                                 call_bio_endio(r1_bio);
 453                         }
 454                 }
 455         }
 456         if (r1_bio->bios[mirror] == NULL)
 457                 rdev_dec_pending(conf->mirrors[mirror].rdev,
 458                                  conf->mddev);
 459
 460         /*
 461          * Let's see if all mirrored write operations have finished
 462          * already.
 463          */
 464         r1_bio_write_done(r1_bio);
 465
 466         if (to_put)
 467                 bio_put(to_put);
 468 }
 469
 470
 471 /*
 472  * This routine returns the disk from which the requested read should
 473  * be done. There is a per-array 'next expected sequential IO' sector
 474  * number - if this matches on the next IO then we use the last disk.
 475  * There is also a per-disk 'last know head position' sector that is
 476  * maintained from IRQ contexts, both the normal and the resync IO
 477  * completion handlers update this position correctly. If there is no
 478  * perfect sequential match then we pick the disk whose head is closest.
 479  *
 480  * If there are 2 mirrors in the same 2 devices, performance degrades
 481  * because position is mirror, not device based.
 482  *
 483  * The rdev for the device selected will have nr_pending incremented.
 484  */
 485 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 486 {
 487         const sector_t this_sector = r1_bio->sector;
 488         int sectors;
 489         int best_good_sectors;
 490         int start_disk;
 491         int best_disk;
 492         int i;
 493         sector_t best_dist;
 494         struct md_rdev *rdev;
 495         int choose_first;
 496
 497         rcu_read_lock();
 498         /*
 499          * Check if we can balance. We can balance on the whole
 500          * device if no resync is going on, or below the resync window.
 501          * We take the first readable disk when above the resync window.
 502          */
 503  retry:
 504         sectors = r1_bio->sectors;
 505         best_disk = -1;
 506         best_dist = MaxSector;
 507         best_good_sectors = 0;
 508
 509         if (conf->mddev->recovery_cp < MaxSector &&
 510             (this_sector + sectors >= conf->next_resync)) {
 511                 choose_first = 1;
 512                 start_disk = 0;
 513         } else {
 514                 choose_first = 0;
 515                 start_disk = conf->last_used;
 516         }
 517
 518         for (i = 0 ; i < conf->raid_disks ; i++) {
 519                 sector_t dist;
 520                 sector_t first_bad;
 521                 int bad_sectors;
 522
 523                 int disk = start_disk + i;
 524                 if (disk >= conf->raid_disks)
 525                         disk -= conf->raid_disks;
 526
 527                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 528                 if (r1_bio->bios[disk] == IO_BLOCKED
 529                     || rdev == NULL
 530                     || test_bit(Faulty, &rdev->flags))
 531                         continue;
 532                 if (!test_bit(In_sync, &rdev->flags) &&
 533                     rdev->recovery_offset < this_sector + sectors)
 534                         continue;
 535                 if (test_bit(WriteMostly, &rdev->flags)) {
 536                         /* Don't balance among write-mostly, just
 537                          * use the first as a last resort */
 538                         if (best_disk < 0) {
 539                                 if (is_badblock(rdev, this_sector, sectors,
 540                                                 &first_bad, &bad_sectors)) {
 541                                         if (first_bad < this_sector)
 542                                                 /* Cannot use this */
 543                                                 continue;
 544                                         best_good_sectors = first_bad - this_sector;
 545                                 } else
 546                                         best_good_sectors = sectors;
 547                                 best_disk = disk;
 548                         }
 549                         continue;
 550                 }
 551                 /* This is a reasonable device to use.  It might
 552                  * even be best.
 553                  */
 554                 if (is_badblock(rdev, this_sector, sectors,
 555                                 &first_bad, &bad_sectors)) {
 556                         if (best_dist < MaxSector)
 557                                 /* already have a better device */
 558                                 continue;
 559                         if (first_bad <= this_sector) {
 560                                 /* cannot read here. If this is the 'primary'
 561                                  * device, then we must not read beyond
 562                                  * bad_sectors from another device..
 563                                  */
 564                                 bad_sectors -= (this_sector - first_bad);
 565                                 if (choose_first && sectors > bad_sectors)
 566                                         sectors = bad_sectors;
 567                                 if (best_good_sectors > sectors)
 568                                         best_good_sectors = sectors;
 569
 570                         } else {
 571                                 sector_t good_sectors = first_bad - this_sector;
 572                                 if (good_sectors > best_good_sectors) {
 573                                         best_good_sectors = good_sectors;
 574                                         best_disk = disk;
 575                                 }
 576                                 if (choose_first)
 577                                         break;
 578                         }
 579                         continue;
 580                 } else
 581                         best_good_sectors = sectors;
 582
 583                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 584                 if (choose_first
 585                     /* Don't change to another disk for sequential reads */
 586                     || conf->next_seq_sect == this_sector
 587                     || dist == 0
 588                     /* If device is idle, use it */
 589                     || atomic_read(&rdev->nr_pending) == 0) {
 590                         best_disk = disk;
 591                         break;
 592                 }
 593                 if (dist < best_dist) {
 594                         best_dist = dist;
 595                         best_disk = disk;
 596                 }
 597         }
 598
 599         if (best_disk >= 0) {
 600                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 601                 if (!rdev)
 602                         goto retry;
 603                 atomic_inc(&rdev->nr_pending);
 604                 if (test_bit(Faulty, &rdev->flags)) {
 605                         /* cannot risk returning a device that failed
 606                          * before we inc'ed nr_pending
 607                          */
 608                         rdev_dec_pending(rdev, conf->mddev);
 609                         goto retry;
 610                 }
 611                 sectors = best_good_sectors;
 612                 conf->next_seq_sect = this_sector + sectors;
 613                 conf->last_used = best_disk;
 614         }
 615         rcu_read_unlock();
 616         *max_sectors = sectors;
 617
 618         return best_disk;
 619 }
 620
 621 int md_raid1_congested(struct mddev *mddev, int bits)
 622 {
 623         struct r1conf *conf = mddev->private;
 624         int i, ret = 0;
 625
 626         if ((bits & (1 << BDI_async_congested)) &&
 627             conf->pending_count >= max_queued_requests)
 628                 return 1;
 629
 630         rcu_read_lock();
 631         for (i = 0; i < mddev->raid_disks; i++) {
 632                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 633                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 634                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 635
 636                         BUG_ON(!q);
 637
 638                         /* Note the '|| 1' - when read_balance prefers
 639                          * non-congested targets, it can be removed
 640                          */
 641                         if ((bits & (1<<BDI_async_congested)) || 1)
 642                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 643                         else
 644                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 645                 }
 646         }
 647         rcu_read_unlock();
 648         return ret;
 649 }
 650 EXPORT_SYMBOL_GPL(md_raid1_congested);
 651
 652 static int raid1_congested(void *data, int bits)
 653 {
 654         struct mddev *mddev = data;
 655
 656         return mddev_congested(mddev, bits) ||
 657                 md_raid1_congested(mddev, bits);
 658 }
 659
 660 static void flush_pending_writes(struct r1conf *conf)
 661 {
 662         /* Any writes that have been queued but are awaiting
 663          * bitmap updates get flushed here.
 664          */
 665         spin_lock_irq(&conf->device_lock);
 666
 667         if (conf->pending_bio_list.head) {
 668                 struct bio *bio;
 669                 bio = bio_list_get(&conf->pending_bio_list);
 670                 conf->pending_count = 0;
 671                 spin_unlock_irq(&conf->device_lock);
 672                 /* flush any pending bitmap writes to
 673                  * disk before proceeding w/ I/O */
 674                 bitmap_unplug(conf->mddev->bitmap);
 675                 wake_up(&conf->wait_barrier);
 676
 677                 while (bio) { /* submit pending writes */
 678                         struct bio *next = bio->bi_next;
 679                         bio->bi_next = NULL;
 680                         generic_make_request(bio);
 681                         bio = next;
 682                 }
 683         } else
 684                 spin_unlock_irq(&conf->device_lock);
 685 }
 686
 687 /* Barriers....
 688  * Sometimes we need to suspend IO while we do something else,
 689  * either some resync/recovery, or reconfigure the array.
 690  * To do this we raise a 'barrier'.
 691  * The 'barrier' is a counter that can be raised multiple times
 692  * to count how many activities are happening which preclude
 693  * normal IO.
 694  * We can only raise the barrier if there is no pending IO.
 695  * i.e. if nr_pending == 0.
 696  * We choose only to raise the barrier if no-one is waiting for the
 697  * barrier to go down.  This means that as soon as an IO request
 698  * is ready, no other operations which require a barrier will start
 699  * until the IO request has had a chance.
 700  *
 701  * So: regular IO calls 'wait_barrier'.  When that returns there
 702  *    is no backgroup IO happening,  It must arrange to call
 703  *    allow_barrier when it has finished its IO.
 704  * backgroup IO calls must call raise_barrier.  Once that returns
 705  *    there is no normal IO happeing.  It must arrange to call
 706  *    lower_barrier when the particular background IO completes.
 707  */
 708 #define RESYNC_DEPTH 32
 709
 710 static void raise_barrier(struct r1conf *conf)
 711 {
 712         spin_lock_irq(&conf->resync_lock);
 713
 714         /* Wait until no block IO is waiting */
 715         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 716                             conf->resync_lock, );
 717
 718         /* block any new IO from starting */
 719         conf->barrier++;
 720
 721         /* Now wait for all pending IO to complete */
 722         wait_event_lock_irq(conf->wait_barrier,
 723                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 724                             conf->resync_lock, );
 725
 726         spin_unlock_irq(&conf->resync_lock);
 727 }
 728
 729 static void lower_barrier(struct r1conf *conf)
 730 {
 731         unsigned long flags;
 732         BUG_ON(conf->barrier <= 0);
 733         spin_lock_irqsave(&conf->resync_lock, flags);
 734         conf->barrier--;
 735         spin_unlock_irqrestore(&conf->resync_lock, flags);
 736         wake_up(&conf->wait_barrier);
 737 }
 738
 739 static void wait_barrier(struct r1conf *conf)
 740 {
 741         spin_lock_irq(&conf->resync_lock);
 742         if (conf->barrier) {
 743                 conf->nr_waiting++;
 744                 /* Wait for the barrier to drop.
 745                  * However if there are already pending
 746                  * requests (preventing the barrier from
 747                  * rising completely), and the
 748                  * pre-process bio queue isn't empty,
 749                  * then don't wait, as we need to empty
 750                  * that queue to get the nr_pending
 751                  * count down.
 752                  */
 753                 wait_event_lock_irq(conf->wait_barrier,
 754                                     !conf->barrier ||
 755                                     (conf->nr_pending &&
 756                                      current->bio_list &&
 757                                      !bio_list_empty(current->bio_list)),
 758                                     conf->resync_lock,
 759                         );
 760                 conf->nr_waiting--;
 761         }
 762         conf->nr_pending++;
 763         spin_unlock_irq(&conf->resync_lock);
 764 }
 765
 766 static void allow_barrier(struct r1conf *conf)
 767 {
 768         unsigned long flags;
 769         spin_lock_irqsave(&conf->resync_lock, flags);
 770         conf->nr_pending--;
 771         spin_unlock_irqrestore(&conf->resync_lock, flags);
 772         wake_up(&conf->wait_barrier);
 773 }
 774
 775 static void freeze_array(struct r1conf *conf)
 776 {
 777         /* stop syncio and normal IO and wait for everything to
 778          * go quite.
 779          * We increment barrier and nr_waiting, and then
 780          * wait until nr_pending match nr_queued+1
 781          * This is called in the context of one normal IO request
 782          * that has failed. Thus any sync request that might be pending
 783          * will be blocked by nr_pending, and we need to wait for
 784          * pending IO requests to complete or be queued for re-try.
 785          * Thus the number queued (nr_queued) plus this request (1)
 786          * must match the number of pending IOs (nr_pending) before
 787          * we continue.
 788          */
 789         spin_lock_irq(&conf->resync_lock);
 790         conf->barrier++;
 791         conf->nr_waiting++;
 792         wait_event_lock_irq(conf->wait_barrier,
 793                             conf->nr_pending == conf->nr_queued+1,
 794                             conf->resync_lock,
 795                             flush_pending_writes(conf));
 796         spin_unlock_irq(&conf->resync_lock);
 797 }
 798 static void unfreeze_array(struct r1conf *conf)
 799 {
 800         /* reverse the effect of the freeze */
 801         spin_lock_irq(&conf->resync_lock);
 802         conf->barrier--;
 803         conf->nr_waiting--;
 804         wake_up(&conf->wait_barrier);
 805         spin_unlock_irq(&conf->resync_lock);
 806 }
 807
 808
 809 /* duplicate the data pages for behind I/O
 810  */
 811 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
 812 {
 813         int i;
 814         struct bio_vec *bvec;
 815         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
 816                                         GFP_NOIO);
 817         if (unlikely(!bvecs))
 818                 return;
 819
 820         bio_for_each_segment(bvec, bio, i) {
 821                 bvecs[i] = *bvec;
 822                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
 823                 if (unlikely(!bvecs[i].bv_page))
 824                         goto do_sync_io;
 825                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
 826                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 827                 kunmap(bvecs[i].bv_page);
 828                 kunmap(bvec->bv_page);
 829         }
 830         r1_bio->behind_bvecs = bvecs;
 831         r1_bio->behind_page_count = bio->bi_vcnt;
 832         set_bit(R1BIO_BehindIO, &r1_bio->state);
 833         return;
 834
 835 do_sync_io:
 836         for (i = 0; i < bio->bi_vcnt; i++)
 837                 if (bvecs[i].bv_page)
 838                         put_page(bvecs[i].bv_page);
 839         kfree(bvecs);
 840         pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 841 }
 842
 843 static void make_request(struct mddev *mddev, struct bio * bio)
 844 {
 845         struct r1conf *conf = mddev->private;
 846         struct mirror_info *mirror;
 847         struct r1bio *r1_bio;
 848         struct bio *read_bio;
 849         int i, disks;
 850         struct bitmap *bitmap;
 851         unsigned long flags;
 852         const int rw = bio_data_dir(bio);
 853         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 854         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
 855         struct md_rdev *blocked_rdev;
 856         int plugged;
 857         int first_clone;
 858         int sectors_handled;
 859         int max_sectors;
 860
 861         /*
 862          * Register the new request and wait if the reconstruction
 863          * thread has put up a bar for new requests.
 864          * Continue immediately if no resync is active currently.
 865          */
 866
 867         md_write_start(mddev, bio); /* wait on superblock update early */
 868
 869         if (bio_data_dir(bio) == WRITE &&
 870             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
 871             bio->bi_sector < mddev->suspend_hi) {
 872                 /* As the suspend_* range is controlled by
 873                  * userspace, we want an interruptible
 874                  * wait.
 875                  */
 876                 DEFINE_WAIT(w);
 877                 for (;;) {
 878                         sigset_t full, old;
 879                         prepare_to_wait(&conf->wait_barrier,
 880                                         &w, TASK_INTERRUPTIBLE);
 881                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
 882                             bio->bi_sector >= mddev->suspend_hi)
 883                                 break;
 884                         sigfillset(&full);
 885                         sigprocmask(SIG_BLOCK, &full, &old);
 886                         schedule();
 887                         sigprocmask(SIG_SETMASK, &old, NULL);
 888                 }
 889                 finish_wait(&conf->wait_barrier, &w);
 890         }
 891
 892         wait_barrier(conf);
 893
 894         bitmap = mddev->bitmap;
 895
 896         /*
 897          * make_request() can abort the operation when READA is being
 898          * used and no empty request is available.
 899          *
 900          */
 901         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 902
 903         r1_bio->master_bio = bio;
 904         r1_bio->sectors = bio->bi_size >> 9;
 905         r1_bio->state = 0;
 906         r1_bio->mddev = mddev;
 907         r1_bio->sector = bio->bi_sector;
 908
 909         /* We might need to issue multiple reads to different
 910          * devices if there are bad blocks around, so we keep
 911          * track of the number of reads in bio->bi_phys_segments.
 912          * If this is 0, there is only one r1_bio and no locking
 913          * will be needed when requests complete.  If it is
 914          * non-zero, then it is the number of not-completed requests.
 915          */
 916         bio->bi_phys_segments = 0;
 917         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
 918
 919         if (rw == READ) {
 920                 /*
 921                  * read balancing logic:
 922                  */
 923                 int rdisk;
 924
 925 read_again:
 926                 rdisk = read_balance(conf, r1_bio, &max_sectors);
 927
 928                 if (rdisk < 0) {
 929                         /* couldn't find anywhere to read from */
 930                         raid_end_bio_io(r1_bio);
 931                         return;
 932                 }
 933                 mirror = conf->mirrors + rdisk;
 934
 935                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
 936                     bitmap) {
 937                         /* Reading from a write-mostly device must
 938                          * take care not to over-take any writes
 939                          * that are 'behind'
 940                          */
 941                         wait_event(bitmap->behind_wait,
 942                                    atomic_read(&bitmap->behind_writes) == 0);
 943                 }
 944                 r1_bio->read_disk = rdisk;
 945
 946                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 947                 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
 948                             max_sectors);
 949
 950                 r1_bio->bios[rdisk] = read_bio;
 951
 952                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 953                 read_bio->bi_bdev = mirror->rdev->bdev;
 954                 read_bio->bi_end_io = raid1_end_read_request;
 955                 read_bio->bi_rw = READ | do_sync;
 956                 read_bio->bi_private = r1_bio;
 957
 958                 if (max_sectors < r1_bio->sectors) {
 959                         /* could not read all from this device, so we will
 960                          * need another r1_bio.
 961                          */
 962
 963                         sectors_handled = (r1_bio->sector + max_sectors
 964                                            - bio->bi_sector);
 965                         r1_bio->sectors = max_sectors;
 966                         spin_lock_irq(&conf->device_lock);
 967                         if (bio->bi_phys_segments == 0)
 968                                 bio->bi_phys_segments = 2;
 969                         else
 970                                 bio->bi_phys_segments++;
 971                         spin_unlock_irq(&conf->device_lock);
 972                         /* Cannot call generic_make_request directly
 973                          * as that will be queued in __make_request
 974                          * and subsequent mempool_alloc might block waiting
 975                          * for it.  So hand bio over to raid1d.
 976                          */
 977                         reschedule_retry(r1_bio);
 978
 979                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 980
 981                         r1_bio->master_bio = bio;
 982                         r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
 983                         r1_bio->state = 0;
 984                         r1_bio->mddev = mddev;
 985                         r1_bio->sector = bio->bi_sector + sectors_handled;
 986                         goto read_again;
 987                 } else
 988                         generic_make_request(read_bio);
 989                 return;
 990         }
 991
 992         /*
 993          * WRITE:
 994          */
 995         if (conf->pending_count >= max_queued_requests) {
 996                 md_wakeup_thread(mddev->thread);
 997                 wait_event(conf->wait_barrier,
 998                            conf->pending_count < max_queued_requests);
 999         }
1000         /* first select target devices under rcu_lock and
1001          * inc refcount on their rdev.  Record them by setting
1002          * bios[x] to bio
1003          * If there are known/acknowledged bad blocks on any device on
1004          * which we have seen a write error, we want to avoid writing those
1005          * blocks.
1006          * This potentially requires several writes to write around
1007          * the bad blocks.  Each set of writes gets it's own r1bio
1008          * with a set of bios attached.
1009          */
1010         plugged = mddev_check_plugged(mddev);
1011
1012         disks = conf->raid_disks;
1013  retry_write:
1014         blocked_rdev = NULL;
1015         rcu_read_lock();
1016         max_sectors = r1_bio->sectors;
1017         for (i = 0;  i < disks; i++) {
1018                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1019                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1020                         atomic_inc(&rdev->nr_pending);
1021                         blocked_rdev = rdev;
1022                         break;
1023                 }
1024                 r1_bio->bios[i] = NULL;
1025                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1026                         set_bit(R1BIO_Degraded, &r1_bio->state);
1027                         continue;
1028                 }
1029
1030                 atomic_inc(&rdev->nr_pending);
1031                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1032                         sector_t first_bad;
1033                         int bad_sectors;
1034                         int is_bad;
1035
1036                         is_bad = is_badblock(rdev, r1_bio->sector,
1037                                              max_sectors,
1038                                              &first_bad, &bad_sectors);
1039                         if (is_bad < 0) {
1040                                 /* mustn't write here until the bad block is
1041                                  * acknowledged*/
1042                                 set_bit(BlockedBadBlocks, &rdev->flags);
1043                                 blocked_rdev = rdev;
1044                                 break;
1045                         }
1046                         if (is_bad && first_bad <= r1_bio->sector) {
1047                                 /* Cannot write here at all */
1048                                 bad_sectors -= (r1_bio->sector - first_bad);
1049                                 if (bad_sectors < max_sectors)
1050                                         /* mustn't write more than bad_sectors
1051                                          * to other devices yet
1052                                          */
1053                                         max_sectors = bad_sectors;
1054                                 rdev_dec_pending(rdev, mddev);
1055                                 /* We don't set R1BIO_Degraded as that
1056                                  * only applies if the disk is
1057                                  * missing, so it might be re-added,
1058                                  * and we want to know to recover this
1059                                  * chunk.
1060                                  * In this case the device is here,
1061                                  * and the fact that this chunk is not
1062                                  * in-sync is recorded in the bad
1063                                  * block log
1064                                  */
1065                                 continue;
1066                         }
1067                         if (is_bad) {
1068                                 int good_sectors = first_bad - r1_bio->sector;
1069                                 if (good_sectors < max_sectors)
1070                                         max_sectors = good_sectors;
1071                         }
1072                 }
1073                 r1_bio->bios[i] = bio;
1074         }
1075         rcu_read_unlock();
1076
1077         if (unlikely(blocked_rdev)) {
1078                 /* Wait for this device to become unblocked */
1079                 int j;
1080
1081                 for (j = 0; j < i; j++)
1082                         if (r1_bio->bios[j])
1083                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1084                 r1_bio->state = 0;
1085                 allow_barrier(conf);
1086                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1087                 wait_barrier(conf);
1088                 goto retry_write;
1089         }
1090
1091         if (max_sectors < r1_bio->sectors) {
1092                 /* We are splitting this write into multiple parts, so
1093                  * we need to prepare for allocating another r1_bio.
1094                  */
1095                 r1_bio->sectors = max_sectors;
1096                 spin_lock_irq(&conf->device_lock);
1097                 if (bio->bi_phys_segments == 0)
1098                         bio->bi_phys_segments = 2;
1099                 else
1100                         bio->bi_phys_segments++;
1101                 spin_unlock_irq(&conf->device_lock);
1102         }
1103         sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1104
1105         atomic_set(&r1_bio->remaining, 1);
1106         atomic_set(&r1_bio->behind_remaining, 0);
1107
1108         first_clone = 1;
1109         for (i = 0; i < disks; i++) {
1110                 struct bio *mbio;
1111                 if (!r1_bio->bios[i])
1112                         continue;
1113
1114                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1115                 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1116
1117                 if (first_clone) {
1118                         /* do behind I/O ?
1119                          * Not if there are too many, or cannot
1120                          * allocate memory, or a reader on WriteMostly
1121                          * is waiting for behind writes to flush */
1122                         if (bitmap &&
1123                             (atomic_read(&bitmap->behind_writes)
1124                              < mddev->bitmap_info.max_write_behind) &&
1125                             !waitqueue_active(&bitmap->behind_wait))
1126                                 alloc_behind_pages(mbio, r1_bio);
1127
1128                         bitmap_startwrite(bitmap, r1_bio->sector,
1129                                           r1_bio->sectors,
1130                                           test_bit(R1BIO_BehindIO,
1131                                                    &r1_bio->state));
1132                         first_clone = 0;
1133                 }
1134                 if (r1_bio->behind_bvecs) {
1135                         struct bio_vec *bvec;
1136                         int j;
1137
1138                         /* Yes, I really want the '__' version so that
1139                          * we clear any unused pointer in the io_vec, rather
1140                          * than leave them unchanged.  This is important
1141                          * because when we come to free the pages, we won't
1142                          * know the original bi_idx, so we just free
1143                          * them all
1144                          */
1145                         bio_for_each_segment_all(bvec, mbio, j)
1146                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1147                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1148                                 atomic_inc(&r1_bio->behind_remaining);
1149                 }
1150
1151                 r1_bio->bios[i] = mbio;
1152
1153                 mbio->bi_sector = (r1_bio->sector +
1154                                    conf->mirrors[i].rdev->data_offset);
1155                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1156                 mbio->bi_end_io = raid1_end_write_request;
1157                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
1158                 mbio->bi_private = r1_bio;
1159
1160                 atomic_inc(&r1_bio->remaining);
1161                 spin_lock_irqsave(&conf->device_lock, flags);
1162                 bio_list_add(&conf->pending_bio_list, mbio);
1163                 conf->pending_count++;
1164                 spin_unlock_irqrestore(&conf->device_lock, flags);
1165         }
1166         /* Mustn't call r1_bio_write_done before this next test,
1167          * as it could result in the bio being freed.
1168          */
1169         if (sectors_handled < (bio->bi_size >> 9)) {
1170                 r1_bio_write_done(r1_bio);
1171                 /* We need another r1_bio.  It has already been counted
1172                  * in bio->bi_phys_segments
1173                  */
1174                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1175                 r1_bio->master_bio = bio;
1176                 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1177                 r1_bio->state = 0;
1178                 r1_bio->mddev = mddev;
1179                 r1_bio->sector = bio->bi_sector + sectors_handled;
1180                 goto retry_write;
1181         }
1182
1183         r1_bio_write_done(r1_bio);
1184
1185         /* In case raid1d snuck in to freeze_array */
1186         wake_up(&conf->wait_barrier);
1187
1188         if (do_sync || !bitmap || !plugged)
1189                 md_wakeup_thread(mddev->thread);
1190 }
1191
1192 static void status(struct seq_file *seq, struct mddev *mddev)
1193 {
1194         struct r1conf *conf = mddev->private;
1195         int i;
1196
1197         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1198                    conf->raid_disks - mddev->degraded);
1199         rcu_read_lock();
1200         for (i = 0; i < conf->raid_disks; i++) {
1201                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1202                 seq_printf(seq, "%s",
1203                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1204         }
1205         rcu_read_unlock();
1206         seq_printf(seq, "]");
1207 }
1208
1209
1210 static void error(struct mddev *mddev, struct md_rdev *rdev)
1211 {
1212         char b[BDEVNAME_SIZE];
1213         struct r1conf *conf = mddev->private;
1214         unsigned long flags;
1215
1216         /*
1217          * If it is not operational, then we have already marked it as dead
1218          * else if it is the last working disks, ignore the error, let the
1219          * next level up know.
1220          * else mark the drive as failed
1221          */
1222         if (test_bit(In_sync, &rdev->flags)
1223             && (conf->raid_disks - mddev->degraded) == 1) {
1224                 /*
1225                  * Don't fail the drive, act as though we were just a
1226                  * normal single drive.
1227                  * However don't try a recovery from this drive as
1228                  * it is very likely to fail.
1229                  */
1230                 conf->recovery_disabled = mddev->recovery_disabled;
1231                 return;
1232         }
1233         set_bit(Blocked, &rdev->flags);
1234         spin_lock_irqsave(&conf->device_lock, flags);
1235         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1236                 mddev->degraded++;
1237                 set_bit(Faulty, &rdev->flags);
1238         } else
1239                 set_bit(Faulty, &rdev->flags);
1240         spin_unlock_irqrestore(&conf->device_lock, flags);
1241         /*
1242          * if recovery is running, make sure it aborts.
1243          */
1244         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1245         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1246         set_bit(MD_CHANGE_PENDING, &mddev->flags);
1247         printk(KERN_ALERT
1248                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1249                "md/raid1:%s: Operation continuing on %d devices.\n",
1250                mdname(mddev), bdevname(rdev->bdev, b),
1251                mdname(mddev), conf->raid_disks - mddev->degraded);
1252 }
1253
1254 static void print_conf(struct r1conf *conf)
1255 {
1256         int i;
1257
1258         printk(KERN_DEBUG "RAID1 conf printout:\n");
1259         if (!conf) {
1260                 printk(KERN_DEBUG "(!conf)\n");
1261                 return;
1262         }
1263         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1264                 conf->raid_disks);
1265
1266         rcu_read_lock();
1267         for (i = 0; i < conf->raid_disks; i++) {
1268                 char b[BDEVNAME_SIZE];
1269                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1270                 if (rdev)
1271                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1272                                i, !test_bit(In_sync, &rdev->flags),
1273                                !test_bit(Faulty, &rdev->flags),
1274                                bdevname(rdev->bdev,b));
1275         }
1276         rcu_read_unlock();
1277 }
1278
1279 static void close_sync(struct r1conf *conf)
1280 {
1281         wait_barrier(conf);
1282         allow_barrier(conf);
1283
1284         mempool_destroy(conf->r1buf_pool);
1285         conf->r1buf_pool = NULL;
1286 }
1287
1288 static int raid1_spare_active(struct mddev *mddev)
1289 {
1290         int i;
1291         struct r1conf *conf = mddev->private;
1292         int count = 0;
1293         unsigned long flags;
1294
1295         /*
1296          * Find all failed disks within the RAID1 configuration
1297          * and mark them readable.
1298          * Called under mddev lock, so rcu protection not needed.
1299          * device_lock used to avoid races with raid1_end_read_request
1300          * which expects 'In_sync' flags and ->degraded to be consistent.
1301          */
1302         spin_lock_irqsave(&conf->device_lock, flags);
1303         for (i = 0; i < conf->raid_disks; i++) {
1304                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1305                 if (rdev
1306                     && !test_bit(Faulty, &rdev->flags)
1307                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1308                         count++;
1309                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1310                 }
1311         }
1312         mddev->degraded -= count;
1313         spin_unlock_irqrestore(&conf->device_lock, flags);
1314
1315         print_conf(conf);
1316         return count;
1317 }
1318
1319
1320 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1321 {
1322         struct r1conf *conf = mddev->private;
1323         int err = -EEXIST;
1324         int mirror = 0;
1325         struct mirror_info *p;
1326         int first = 0;
1327         int last = mddev->raid_disks - 1;
1328
1329         if (mddev->recovery_disabled == conf->recovery_disabled)
1330                 return -EBUSY;
1331
1332         if (rdev->raid_disk >= 0)
1333                 first = last = rdev->raid_disk;
1334
1335         for (mirror = first; mirror <= last; mirror++)
1336                 if ( !(p=conf->mirrors+mirror)->rdev) {
1337
1338                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1339                                           rdev->data_offset << 9);
1340                         /* as we don't honour merge_bvec_fn, we must
1341                          * never risk violating it, so limit
1342                          * ->max_segments to one lying with a single
1343                          * page, as a one page request is never in
1344                          * violation.
1345                          */
1346                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1347                                 blk_queue_max_segments(mddev->queue, 1);
1348                                 blk_queue_segment_boundary(mddev->queue,
1349                                                            PAGE_CACHE_SIZE - 1);
1350                         }
1351
1352                         p->head_position = 0;
1353                         rdev->raid_disk = mirror;
1354                         err = 0;
1355                         /* As all devices are equivalent, we don't need a full recovery
1356                          * if this was recently any drive of the array
1357                          */
1358                         if (rdev->saved_raid_disk < 0)
1359                                 conf->fullsync = 1;
1360                         rcu_assign_pointer(p->rdev, rdev);
1361                         break;
1362                 }
1363         md_integrity_add_rdev(rdev, mddev);
1364         print_conf(conf);
1365         return err;
1366 }
1367
1368 static int raid1_remove_disk(struct mddev *mddev, int number)
1369 {
1370         struct r1conf *conf = mddev->private;
1371         int err = 0;
1372         struct md_rdev *rdev;
1373         struct mirror_info *p = conf->mirrors+ number;
1374
1375         print_conf(conf);
1376         rdev = p->rdev;
1377         if (rdev) {
1378                 if (test_bit(In_sync, &rdev->flags) ||
1379                     atomic_read(&rdev->nr_pending)) {
1380                         err = -EBUSY;
1381                         goto abort;
1382                 }
1383                 /* Only remove non-faulty devices if recovery
1384                  * is not possible.
1385                  */
1386                 if (!test_bit(Faulty, &rdev->flags) &&
1387                     mddev->recovery_disabled != conf->recovery_disabled &&
1388                     mddev->degraded < conf->raid_disks) {
1389                         err = -EBUSY;
1390                         goto abort;
1391                 }
1392                 p->rdev = NULL;
1393                 synchronize_rcu();
1394                 if (atomic_read(&rdev->nr_pending)) {
1395                         /* lost the race, try later */
1396                         err = -EBUSY;
1397                         p->rdev = rdev;
1398                         goto abort;
1399                 }
1400                 err = md_integrity_register(mddev);
1401         }
1402 abort:
1403
1404         print_conf(conf);
1405         return err;
1406 }
1407
1408
1409 static void end_sync_read(struct bio *bio, int error)
1410 {
1411         struct r1bio *r1_bio = bio->bi_private;
1412
1413         update_head_pos(r1_bio->read_disk, r1_bio);
1414
1415         /*
1416          * we have read a block, now it needs to be re-written,
1417          * or re-read if the read failed.
1418          * We don't do much here, just schedule handling by raid1d
1419          */
1420         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1421                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1422
1423         if (atomic_dec_and_test(&r1_bio->remaining))
1424                 reschedule_retry(r1_bio);
1425 }
1426
1427 static void end_sync_write(struct bio *bio, int error)
1428 {
1429         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1430         struct r1bio *r1_bio = bio->bi_private;
1431         struct mddev *mddev = r1_bio->mddev;
1432         struct r1conf *conf = mddev->private;
1433         int mirror=0;
1434         sector_t first_bad;
1435         int bad_sectors;
1436
1437         mirror = find_bio_disk(r1_bio, bio);
1438
1439         if (!uptodate) {
1440                 sector_t sync_blocks = 0;
1441                 sector_t s = r1_bio->sector;
1442                 long sectors_to_go = r1_bio->sectors;
1443                 /* make sure these bits doesn't get cleared. */
1444                 do {
1445                         bitmap_end_sync(mddev->bitmap, s,
1446                                         &sync_blocks, 1);
1447                         s += sync_blocks;
1448                         sectors_to_go -= sync_blocks;
1449                 } while (sectors_to_go > 0);
1450                 set_bit(WriteErrorSeen,
1451                         &conf->mirrors[mirror].rdev->flags);
1452                 set_bit(R1BIO_WriteError, &r1_bio->state);
1453         } else if (is_badblock(conf->mirrors[mirror].rdev,
1454                                r1_bio->sector,
1455                                r1_bio->sectors,
1456                                &first_bad, &bad_sectors) &&
1457                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1458                                 r1_bio->sector,
1459                                 r1_bio->sectors,
1460                                 &first_bad, &bad_sectors)
1461                 )
1462                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1463
1464         if (atomic_dec_and_test(&r1_bio->remaining)) {
1465                 int s = r1_bio->sectors;
1466                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1467                     test_bit(R1BIO_WriteError, &r1_bio->state))
1468                         reschedule_retry(r1_bio);
1469                 else {
1470                         put_buf(r1_bio);
1471                         md_done_sync(mddev, s, uptodate);
1472                 }
1473         }
1474 }
1475
1476 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1477                             int sectors, struct page *page, int rw)
1478 {
1479         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1480                 /* success */
1481                 return 1;
1482         if (rw == WRITE)
1483                 set_bit(WriteErrorSeen, &rdev->flags);
1484         /* need to record an error - either for the block or the device */
1485         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1486                 md_error(rdev->mddev, rdev);
1487         return 0;
1488 }
1489
1490 static int fix_sync_read_error(struct r1bio *r1_bio)
1491 {
1492         /* Try some synchronous reads of other devices to get
1493          * good data, much like with normal read errors.  Only
1494          * read into the pages we already have so we don't
1495          * need to re-issue the read request.
1496          * We don't need to freeze the array, because being in an
1497          * active sync request, there is no normal IO, and
1498          * no overlapping syncs.
1499          * We don't need to check is_badblock() again as we
1500          * made sure that anything with a bad block in range
1501          * will have bi_end_io clear.
1502          */
1503         struct mddev *mddev = r1_bio->mddev;
1504         struct r1conf *conf = mddev->private;
1505         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1506         sector_t sect = r1_bio->sector;
1507         int sectors = r1_bio->sectors;
1508         int idx = 0;
1509
1510         while(sectors) {
1511                 int s = sectors;
1512                 int d = r1_bio->read_disk;
1513                 int success = 0;
1514                 struct md_rdev *rdev;
1515                 int start;
1516
1517                 if (s > (PAGE_SIZE>>9))
1518                         s = PAGE_SIZE >> 9;
1519                 do {
1520                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1521                                 /* No rcu protection needed here devices
1522                                  * can only be removed when no resync is
1523                                  * active, and resync is currently active
1524                                  */
1525                                 rdev = conf->mirrors[d].rdev;
1526                                 if (sync_page_io(rdev, sect, s<<9,
1527                                                  bio->bi_io_vec[idx].bv_page,
1528                                                  READ, false)) {
1529                                         success = 1;
1530                                         break;
1531                                 }
1532                         }
1533                         d++;
1534                         if (d == conf->raid_disks)
1535                                 d = 0;
1536                 } while (!success && d != r1_bio->read_disk);
1537
1538                 if (!success) {
1539                         char b[BDEVNAME_SIZE];
1540                         int abort = 0;
1541                         /* Cannot read from anywhere, this block is lost.
1542                          * Record a bad block on each device.  If that doesn't
1543                          * work just disable and interrupt the recovery.
1544                          * Don't fail devices as that won't really help.
1545                          */
1546                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1547                                " for block %llu\n",
1548                                mdname(mddev),
1549                                bdevname(bio->bi_bdev, b),
1550                                (unsigned long long)r1_bio->sector);
1551                         for (d = 0; d < conf->raid_disks; d++) {
1552                                 rdev = conf->mirrors[d].rdev;
1553                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1554                                         continue;
1555                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1556                                         abort = 1;
1557                         }
1558                         if (abort) {
1559                                 conf->recovery_disabled =
1560                                         mddev->recovery_disabled;
1561                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1562                                 md_done_sync(mddev, r1_bio->sectors, 0);
1563                                 put_buf(r1_bio);
1564                                 return 0;
1565                         }
1566                         /* Try next page */
1567                         sectors -= s;
1568                         sect += s;
1569                         idx++;
1570                         continue;
1571                 }
1572
1573                 start = d;
1574                 /* write it back and re-read */
1575                 while (d != r1_bio->read_disk) {
1576                         if (d == 0)
1577                                 d = conf->raid_disks;
1578                         d--;
1579                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1580                                 continue;
1581                         rdev = conf->mirrors[d].rdev;
1582                         if (r1_sync_page_io(rdev, sect, s,
1583                                             bio->bi_io_vec[idx].bv_page,
1584                                             WRITE) == 0) {
1585                                 r1_bio->bios[d]->bi_end_io = NULL;
1586                                 rdev_dec_pending(rdev, mddev);
1587                         }
1588                 }
1589                 d = start;
1590                 while (d != r1_bio->read_disk) {
1591                         if (d == 0)
1592                                 d = conf->raid_disks;
1593                         d--;
1594                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1595                                 continue;
1596                         rdev = conf->mirrors[d].rdev;
1597                         if (r1_sync_page_io(rdev, sect, s,
1598                                             bio->bi_io_vec[idx].bv_page,
1599                                             READ) != 0)
1600                                 atomic_add(s, &rdev->corrected_errors);
1601                 }
1602                 sectors -= s;
1603                 sect += s;
1604                 idx ++;
1605         }
1606         set_bit(R1BIO_Uptodate, &r1_bio->state);
1607         set_bit(BIO_UPTODATE, &bio->bi_flags);
1608         return 1;
1609 }
1610
1611 static int process_checks(struct r1bio *r1_bio)
1612 {
1613         /* We have read all readable devices.  If we haven't
1614          * got the block, then there is no hope left.
1615          * If we have, then we want to do a comparison
1616          * and skip the write if everything is the same.
1617          * If any blocks failed to read, then we need to
1618          * attempt an over-write
1619          */
1620         struct mddev *mddev = r1_bio->mddev;
1621         struct r1conf *conf = mddev->private;
1622         int primary;
1623         int i;
1624
1625         for (primary = 0; primary < conf->raid_disks; primary++)
1626                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1627                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1628                         r1_bio->bios[primary]->bi_end_io = NULL;
1629                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1630                         break;
1631                 }
1632         r1_bio->read_disk = primary;
1633         for (i = 0; i < conf->raid_disks; i++) {
1634                 int j;
1635                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1636                 struct bio *pbio = r1_bio->bios[primary];
1637                 struct bio *sbio = r1_bio->bios[i];
1638                 int size;
1639
1640                 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1641                         continue;
1642
1643                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1644                         for (j = vcnt; j-- ; ) {
1645                                 struct page *p, *s;
1646                                 p = pbio->bi_io_vec[j].bv_page;
1647                                 s = sbio->bi_io_vec[j].bv_page;
1648                                 if (memcmp(page_address(p),
1649                                            page_address(s),
1650                                            PAGE_SIZE))
1651                                         break;
1652                         }
1653                 } else
1654                         j = 0;
1655                 if (j >= 0)
1656                         mddev->resync_mismatches += r1_bio->sectors;
1657                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1658                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1659                         /* No need to write to this device. */
1660                         sbio->bi_end_io = NULL;
1661                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1662                         continue;
1663                 }
1664                 /* fixup the bio for reuse */
1665                 sbio->bi_vcnt = vcnt;
1666                 sbio->bi_size = r1_bio->sectors << 9;
1667                 sbio->bi_idx = 0;
1668                 sbio->bi_phys_segments = 0;
1669                 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1670                 sbio->bi_flags |= 1 << BIO_UPTODATE;
1671                 sbio->bi_next = NULL;
1672                 sbio->bi_sector = r1_bio->sector +
1673                         conf->mirrors[i].rdev->data_offset;
1674                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1675                 size = sbio->bi_size;
1676                 for (j = 0; j < vcnt ; j++) {
1677                         struct bio_vec *bi;
1678                         bi = &sbio->bi_io_vec[j];
1679                         bi->bv_offset = 0;
1680                         if (size > PAGE_SIZE)
1681                                 bi->bv_len = PAGE_SIZE;
1682                         else
1683                                 bi->bv_len = size;
1684                         size -= PAGE_SIZE;
1685                         memcpy(page_address(bi->bv_page),
1686                                page_address(pbio->bi_io_vec[j].bv_page),
1687                                PAGE_SIZE);
1688                 }
1689         }
1690         return 0;
1691 }
1692
1693 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1694 {
1695         struct r1conf *conf = mddev->private;
1696         int i;
1697         int disks = conf->raid_disks;
1698         struct bio *bio, *wbio;
1699
1700         bio = r1_bio->bios[r1_bio->read_disk];
1701
1702         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1703                 /* ouch - failed to read all of that. */
1704                 if (!fix_sync_read_error(r1_bio))
1705                         return;
1706
1707         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1708                 if (process_checks(r1_bio) < 0)
1709                         return;
1710         /*
1711          * schedule writes
1712          */
1713         atomic_set(&r1_bio->remaining, 1);
1714         for (i = 0; i < disks ; i++) {
1715                 wbio = r1_bio->bios[i];
1716                 if (wbio->bi_end_io == NULL ||
1717                     (wbio->bi_end_io == end_sync_read &&
1718                      (i == r1_bio->read_disk ||
1719                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1720                         continue;
1721
1722                 wbio->bi_rw = WRITE;
1723                 wbio->bi_end_io = end_sync_write;
1724                 atomic_inc(&r1_bio->remaining);
1725                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1726
1727                 generic_make_request(wbio);
1728         }
1729
1730         if (atomic_dec_and_test(&r1_bio->remaining)) {
1731                 /* if we're here, all write(s) have completed, so clean up */
1732                 int s = r1_bio->sectors;
1733                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1734                     test_bit(R1BIO_WriteError, &r1_bio->state))
1735                         reschedule_retry(r1_bio);
1736                 else {
1737                         put_buf(r1_bio);
1738                         md_done_sync(mddev, s, 1);
1739                 }
1740         }
1741 }
1742
1743 /*
1744  * This is a kernel thread which:
1745  *
1746  *      1.      Retries failed read operations on working mirrors.
1747  *      2.      Updates the raid superblock when problems encounter.
1748  *      3.      Performs writes following reads for array synchronising.
1749  */
1750
1751 static void fix_read_error(struct r1conf *conf, int read_disk,
1752                            sector_t sect, int sectors)
1753 {
1754         struct mddev *mddev = conf->mddev;
1755         while(sectors) {
1756                 int s = sectors;
1757                 int d = read_disk;
1758                 int success = 0;
1759                 int start;
1760                 struct md_rdev *rdev;
1761
1762                 if (s > (PAGE_SIZE>>9))
1763                         s = PAGE_SIZE >> 9;
1764
1765                 do {
1766                         /* Note: no rcu protection needed here
1767                          * as this is synchronous in the raid1d thread
1768                          * which is the thread that might remove
1769                          * a device.  If raid1d ever becomes multi-threaded....
1770                          */
1771                         sector_t first_bad;
1772                         int bad_sectors;
1773
1774                         rdev = conf->mirrors[d].rdev;
1775                         if (rdev &&
1776                             test_bit(In_sync, &rdev->flags) &&
1777                             is_badblock(rdev, sect, s,
1778                                         &first_bad, &bad_sectors) == 0 &&
1779                             sync_page_io(rdev, sect, s<<9,
1780                                          conf->tmppage, READ, false))
1781                                 success = 1;
1782                         else {
1783                                 d++;
1784                                 if (d == conf->raid_disks)
1785                                         d = 0;
1786                         }
1787                 } while (!success && d != read_disk);
1788
1789                 if (!success) {
1790                         /* Cannot read from anywhere - mark it bad */
1791                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
1792                         if (!rdev_set_badblocks(rdev, sect, s, 0))
1793                                 md_error(mddev, rdev);
1794                         break;
1795                 }
1796                 /* write it back and re-read */
1797                 start = d;
1798                 while (d != read_disk) {
1799                         if (d==0)
1800                                 d = conf->raid_disks;
1801                         d--;
1802                         rdev = conf->mirrors[d].rdev;
1803                         if (rdev &&
1804                             test_bit(In_sync, &rdev->flags))
1805                                 r1_sync_page_io(rdev, sect, s,
1806                                                 conf->tmppage, WRITE);
1807                 }
1808                 d = start;
1809                 while (d != read_disk) {
1810                         char b[BDEVNAME_SIZE];
1811                         if (d==0)
1812                                 d = conf->raid_disks;
1813                         d--;
1814                         rdev = conf->mirrors[d].rdev;
1815                         if (rdev &&
1816                             test_bit(In_sync, &rdev->flags)) {
1817                                 if (r1_sync_page_io(rdev, sect, s,
1818                                                     conf->tmppage, READ)) {
1819                                         atomic_add(s, &rdev->corrected_errors);
1820                                         printk(KERN_INFO
1821                                                "md/raid1:%s: read error corrected "
1822                                                "(%d sectors at %llu on %s)\n",
1823                                                mdname(mddev), s,
1824                                                (unsigned long long)(sect +
1825                                                    rdev->data_offset),
1826                                                bdevname(rdev->bdev, b));
1827                                 }
1828                         }
1829                 }
1830                 sectors -= s;
1831                 sect += s;
1832         }
1833 }
1834
1835 static void bi_complete(struct bio *bio, int error)
1836 {
1837         complete((struct completion *)bio->bi_private);
1838 }
1839
1840 static int submit_bio_wait(int rw, struct bio *bio)
1841 {
1842         struct completion event;
1843         rw |= REQ_SYNC;
1844
1845         init_completion(&event);
1846         bio->bi_private = &event;
1847         bio->bi_end_io = bi_complete;
1848         submit_bio(rw, bio);
1849         wait_for_completion(&event);
1850
1851         return test_bit(BIO_UPTODATE, &bio->bi_flags);
1852 }
1853
1854 static int narrow_write_error(struct r1bio *r1_bio, int i)
1855 {
1856         struct mddev *mddev = r1_bio->mddev;
1857         struct r1conf *conf = mddev->private;
1858         struct md_rdev *rdev = conf->mirrors[i].rdev;
1859         int vcnt, idx;
1860         struct bio_vec *vec;
1861
1862         /* bio has the data to be written to device 'i' where
1863          * we just recently had a write error.
1864          * We repeatedly clone the bio and trim down to one block,
1865          * then try the write.  Where the write fails we record
1866          * a bad block.
1867          * It is conceivable that the bio doesn't exactly align with
1868          * blocks.  We must handle this somehow.
1869          *
1870          * We currently own a reference on the rdev.
1871          */
1872
1873         int block_sectors;
1874         sector_t sector;
1875         int sectors;
1876         int sect_to_write = r1_bio->sectors;
1877         int ok = 1;
1878
1879         if (rdev->badblocks.shift < 0)
1880                 return 0;
1881
1882         block_sectors = 1 << rdev->badblocks.shift;
1883         sector = r1_bio->sector;
1884         sectors = ((sector + block_sectors)
1885                    & ~(sector_t)(block_sectors - 1))
1886                 - sector;
1887
1888         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
1889                 vcnt = r1_bio->behind_page_count;
1890                 vec = r1_bio->behind_bvecs;
1891                 idx = 0;
1892                 while (vec[idx].bv_page == NULL)
1893                         idx++;
1894         } else {
1895                 vcnt = r1_bio->master_bio->bi_vcnt;
1896                 vec = r1_bio->master_bio->bi_io_vec;
1897                 idx = r1_bio->master_bio->bi_idx;
1898         }
1899         while (sect_to_write) {
1900                 struct bio *wbio;
1901                 if (sectors > sect_to_write)
1902                         sectors = sect_to_write;
1903                 /* Write at 'sector' for 'sectors'*/
1904
1905                 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
1906                 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
1907                 wbio->bi_sector = r1_bio->sector;
1908                 wbio->bi_rw = WRITE;
1909                 wbio->bi_vcnt = vcnt;
1910                 wbio->bi_size = r1_bio->sectors << 9;
1911                 wbio->bi_idx = idx;
1912
1913                 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
1914                 wbio->bi_sector += rdev->data_offset;
1915                 wbio->bi_bdev = rdev->bdev;
1916                 if (submit_bio_wait(WRITE, wbio) == 0)
1917                         /* failure! */
1918                         ok = rdev_set_badblocks(rdev, sector,
1919                                                 sectors, 0)
1920                                 && ok;
1921
1922                 bio_put(wbio);
1923                 sect_to_write -= sectors;
1924                 sector += sectors;
1925                 sectors = block_sectors;
1926         }
1927         return ok;
1928 }
1929
1930 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
1931 {
1932         int m;
1933         int s = r1_bio->sectors;
1934         for (m = 0; m < conf->raid_disks ; m++) {
1935                 struct md_rdev *rdev = conf->mirrors[m].rdev;
1936                 struct bio *bio = r1_bio->bios[m];
1937                 if (bio->bi_end_io == NULL)
1938                         continue;
1939                 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
1940                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
1941                         rdev_clear_badblocks(rdev, r1_bio->sector, s);
1942                 }
1943                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
1944                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
1945                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
1946                                 md_error(conf->mddev, rdev);
1947                 }
1948         }
1949         put_buf(r1_bio);
1950         md_done_sync(conf->mddev, s, 1);
1951 }
1952
1953 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
1954 {
1955         int m;
1956         bool fail = false;
1957         for (m = 0; m < conf->raid_disks ; m++)
1958                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
1959                         struct md_rdev *rdev = conf->mirrors[m].rdev;
1960                         rdev_clear_badblocks(rdev,
1961                                              r1_bio->sector,
1962                                              r1_bio->sectors);
1963                         rdev_dec_pending(rdev, conf->mddev);
1964                 } else if (r1_bio->bios[m] != NULL) {
1965                         /* This drive got a write error.  We need to
1966                          * narrow down and record precise write
1967                          * errors.
1968                          */
1969                         fail = true;
1970                         if (!narrow_write_error(r1_bio, m)) {
1971                                 md_error(conf->mddev,
1972                                          conf->mirrors[m].rdev);
1973                                 /* an I/O failed, we can't clear the bitmap */
1974                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1975                         }
1976                         rdev_dec_pending(conf->mirrors[m].rdev,
1977                                          conf->mddev);
1978                 }
1979         if (fail) {
1980                 spin_lock_irq(&conf->device_lock);
1981                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
1982                 conf->nr_queued++;
1983                 spin_unlock_irq(&conf->device_lock);
1984                 md_wakeup_thread(conf->mddev->thread);
1985         } else {
1986                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
1987                         close_write(r1_bio);
1988                 raid_end_bio_io(r1_bio);
1989         }
1990 }
1991
1992 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
1993 {
1994         int disk;
1995         int max_sectors;
1996         struct mddev *mddev = conf->mddev;
1997         struct bio *bio;
1998         char b[BDEVNAME_SIZE];
1999         struct md_rdev *rdev;
2000
2001         clear_bit(R1BIO_ReadError, &r1_bio->state);
2002         /* we got a read error. Maybe the drive is bad.  Maybe just
2003          * the block and we can fix it.
2004          * We freeze all other IO, and try reading the block from
2005          * other devices.  When we find one, we re-write
2006          * and check it that fixes the read error.
2007          * This is all done synchronously while the array is
2008          * frozen
2009          */
2010         if (mddev->ro == 0) {
2011                 freeze_array(conf);
2012                 fix_read_error(conf, r1_bio->read_disk,
2013                                r1_bio->sector, r1_bio->sectors);
2014                 unfreeze_array(conf);
2015         } else
2016                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2017
2018         bio = r1_bio->bios[r1_bio->read_disk];
2019         bdevname(bio->bi_bdev, b);
2020 read_more:
2021         disk = read_balance(conf, r1_bio, &max_sectors);
2022         if (disk == -1) {
2023                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2024                        " read error for block %llu\n",
2025                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
2026                 raid_end_bio_io(r1_bio);
2027         } else {
2028                 const unsigned long do_sync
2029                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
2030                 if (bio) {
2031                         r1_bio->bios[r1_bio->read_disk] =
2032                                 mddev->ro ? IO_BLOCKED : NULL;
2033                         bio_put(bio);
2034                 }
2035                 r1_bio->read_disk = disk;
2036                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2037                 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2038                 r1_bio->bios[r1_bio->read_disk] = bio;
2039                 rdev = conf->mirrors[disk].rdev;
2040                 printk_ratelimited(KERN_ERR
2041                                    "md/raid1:%s: redirecting sector %llu"
2042                                    " to other mirror: %s\n",
2043                                    mdname(mddev),
2044                                    (unsigned long long)r1_bio->sector,
2045                                    bdevname(rdev->bdev, b));
2046                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
2047                 bio->bi_bdev = rdev->bdev;
2048                 bio->bi_end_io = raid1_end_read_request;
2049                 bio->bi_rw = READ | do_sync;
2050                 bio->bi_private = r1_bio;
2051                 if (max_sectors < r1_bio->sectors) {
2052                         /* Drat - have to split this up more */
2053                         struct bio *mbio = r1_bio->master_bio;
2054                         int sectors_handled = (r1_bio->sector + max_sectors
2055                                                - mbio->bi_sector);
2056                         r1_bio->sectors = max_sectors;
2057                         spin_lock_irq(&conf->device_lock);
2058                         if (mbio->bi_phys_segments == 0)
2059                                 mbio->bi_phys_segments = 2;
2060                         else
2061                                 mbio->bi_phys_segments++;
2062                         spin_unlock_irq(&conf->device_lock);
2063                         generic_make_request(bio);
2064                         bio = NULL;
2065
2066                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2067
2068                         r1_bio->master_bio = mbio;
2069                         r1_bio->sectors = (mbio->bi_size >> 9)
2070                                           - sectors_handled;
2071                         r1_bio->state = 0;
2072                         set_bit(R1BIO_ReadError, &r1_bio->state);
2073                         r1_bio->mddev = mddev;
2074                         r1_bio->sector = mbio->bi_sector + sectors_handled;
2075
2076                         goto read_more;
2077                 } else
2078                         generic_make_request(bio);
2079         }
2080 }
2081
2082 static void raid1d(struct mddev *mddev)
2083 {
2084         struct r1bio *r1_bio;
2085         unsigned long flags;
2086         struct r1conf *conf = mddev->private;
2087         struct list_head *head = &conf->retry_list;
2088         struct blk_plug plug;
2089
2090         md_check_recovery(mddev);
2091
2092         if (!list_empty_careful(&conf->bio_end_io_list) &&
2093             !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2094                 LIST_HEAD(tmp);
2095                 spin_lock_irqsave(&conf->device_lock, flags);
2096                 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2097                         while (!list_empty(&conf->bio_end_io_list)) {
2098                                 list_move(conf->bio_end_io_list.prev, &tmp);
2099                                 conf->nr_queued--;
2100                         }
2101                 }
2102                 spin_unlock_irqrestore(&conf->device_lock, flags);
2103                 while (!list_empty(&tmp)) {
2104                         r1_bio = list_first_entry(&conf->bio_end_io_list,
2105                                                   struct r1bio, retry_list);
2106                         list_del(&r1_bio->retry_list);
2107                         if (mddev->degraded)
2108                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2109                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2110                                 close_write(r1_bio);
2111                         raid_end_bio_io(r1_bio);
2112                 }
2113         }
2114
2115         blk_start_plug(&plug);
2116         for (;;) {
2117
2118                 if (atomic_read(&mddev->plug_cnt) == 0)
2119                         flush_pending_writes(conf);
2120
2121                 spin_lock_irqsave(&conf->device_lock, flags);
2122                 if (list_empty(head)) {
2123                         spin_unlock_irqrestore(&conf->device_lock, flags);
2124                         break;
2125                 }
2126                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2127                 list_del(head->prev);
2128                 conf->nr_queued--;
2129                 spin_unlock_irqrestore(&conf->device_lock, flags);
2130
2131                 mddev = r1_bio->mddev;
2132                 conf = mddev->private;
2133                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2134                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2135                             test_bit(R1BIO_WriteError, &r1_bio->state))
2136                                 handle_sync_write_finished(conf, r1_bio);
2137                         else
2138                                 sync_request_write(mddev, r1_bio);
2139                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2140                            test_bit(R1BIO_WriteError, &r1_bio->state))
2141                         handle_write_finished(conf, r1_bio);
2142                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2143                         handle_read_error(conf, r1_bio);
2144                 else
2145                         /* just a partial read to be scheduled from separate
2146                          * context
2147                          */
2148                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2149
2150                 cond_resched();
2151                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2152                         md_check_recovery(mddev);
2153         }
2154         blk_finish_plug(&plug);
2155 }
2156
2157
2158 static int init_resync(struct r1conf *conf)
2159 {
2160         int buffs;
2161
2162         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2163         BUG_ON(conf->r1buf_pool);
2164         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2165                                           conf->poolinfo);
2166         if (!conf->r1buf_pool)
2167                 return -ENOMEM;
2168         conf->next_resync = 0;
2169         return 0;
2170 }
2171
2172 /*
2173  * perform a "sync" on one "block"
2174  *
2175  * We need to make sure that no normal I/O request - particularly write
2176  * requests - conflict with active sync requests.
2177  *
2178  * This is achieved by tracking pending requests and a 'barrier' concept
2179  * that can be installed to exclude normal IO requests.
2180  */
2181
2182 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2183 {
2184         struct r1conf *conf = mddev->private;
2185         struct r1bio *r1_bio;
2186         struct bio *bio;
2187         sector_t max_sector, nr_sectors;
2188         int disk = -1;
2189         int i;
2190         int wonly = -1;
2191         int write_targets = 0, read_targets = 0;
2192         sector_t sync_blocks;
2193         int still_degraded = 0;
2194         int good_sectors = RESYNC_SECTORS;
2195         int min_bad = 0; /* number of sectors that are bad in all devices */
2196
2197         if (!conf->r1buf_pool)
2198                 if (init_resync(conf))
2199                         return 0;
2200
2201         max_sector = mddev->dev_sectors;
2202         if (sector_nr >= max_sector) {
2203                 /* If we aborted, we need to abort the
2204                  * sync on the 'current' bitmap chunk (there will
2205                  * only be one in raid1 resync.
2206                  * We can find the current addess in mddev->curr_resync
2207                  */
2208                 if (mddev->curr_resync < max_sector) /* aborted */
2209                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2210                                                 &sync_blocks, 1);
2211                 else /* completed sync */
2212                         conf->fullsync = 0;
2213
2214                 bitmap_close_sync(mddev->bitmap);
2215                 close_sync(conf);
2216                 return 0;
2217         }
2218
2219         if (mddev->bitmap == NULL &&
2220             mddev->recovery_cp == MaxSector &&
2221             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2222             conf->fullsync == 0) {
2223                 *skipped = 1;
2224                 return max_sector - sector_nr;
2225         }
2226         /* before building a request, check if we can skip these blocks..
2227          * This call the bitmap_start_sync doesn't actually record anything
2228          */
2229         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2230             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2231                 /* We can skip this block, and probably several more */
2232                 *skipped = 1;
2233                 return sync_blocks;
2234         }
2235         /*
2236          * If there is non-resync activity waiting for a turn,
2237          * and resync is going fast enough,
2238          * then let it though before starting on this new sync request.
2239          */
2240         if (!go_faster && conf->nr_waiting)
2241                 msleep_interruptible(1000);
2242
2243         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2244         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2245         raise_barrier(conf);
2246
2247         conf->next_resync = sector_nr;
2248
2249         rcu_read_lock();
2250         /*
2251          * If we get a correctably read error during resync or recovery,
2252          * we might want to read from a different device.  So we
2253          * flag all drives that could conceivably be read from for READ,
2254          * and any others (which will be non-In_sync devices) for WRITE.
2255          * If a read fails, we try reading from something else for which READ
2256          * is OK.
2257          */
2258
2259         r1_bio->mddev = mddev;
2260         r1_bio->sector = sector_nr;
2261         r1_bio->state = 0;
2262         set_bit(R1BIO_IsSync, &r1_bio->state);
2263
2264         for (i=0; i < conf->raid_disks; i++) {
2265                 struct md_rdev *rdev;
2266                 bio = r1_bio->bios[i];
2267
2268                 /* take from bio_init */
2269                 bio->bi_next = NULL;
2270                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
2271                 bio->bi_flags |= 1 << BIO_UPTODATE;
2272                 bio->bi_rw = READ;
2273                 bio->bi_vcnt = 0;
2274                 bio->bi_idx = 0;
2275                 bio->bi_phys_segments = 0;
2276                 bio->bi_size = 0;
2277                 bio->bi_end_io = NULL;
2278                 bio->bi_private = NULL;
2279
2280                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2281                 if (rdev == NULL ||
2282                     test_bit(Faulty, &rdev->flags)) {
2283                         still_degraded = 1;
2284                 } else if (!test_bit(In_sync, &rdev->flags)) {
2285                         bio->bi_rw = WRITE;
2286                         bio->bi_end_io = end_sync_write;
2287                         write_targets ++;
2288                 } else {
2289                         /* may need to read from here */
2290                         sector_t first_bad = MaxSector;
2291                         int bad_sectors;
2292
2293                         if (is_badblock(rdev, sector_nr, good_sectors,
2294                                         &first_bad, &bad_sectors)) {
2295                                 if (first_bad > sector_nr)
2296                                         good_sectors = first_bad - sector_nr;
2297                                 else {
2298                                         bad_sectors -= (sector_nr - first_bad);
2299                                         if (min_bad == 0 ||
2300                                             min_bad > bad_sectors)
2301                                                 min_bad = bad_sectors;
2302                                 }
2303                         }
2304                         if (sector_nr < first_bad) {
2305                                 if (test_bit(WriteMostly, &rdev->flags)) {
2306                                         if (wonly < 0)
2307                                                 wonly = i;
2308                                 } else {
2309                                         if (disk < 0)
2310                                                 disk = i;
2311                                 }
2312                                 bio->bi_rw = READ;
2313                                 bio->bi_end_io = end_sync_read;
2314                                 read_targets++;
2315                         }
2316                 }
2317                 if (bio->bi_end_io) {
2318                         atomic_inc(&rdev->nr_pending);
2319                         bio->bi_sector = sector_nr + rdev->data_offset;
2320                         bio->bi_bdev = rdev->bdev;
2321                         bio->bi_private = r1_bio;
2322                 }
2323         }
2324         rcu_read_unlock();
2325         if (disk < 0)
2326                 disk = wonly;
2327         r1_bio->read_disk = disk;
2328
2329         if (read_targets == 0 && min_bad > 0) {
2330                 /* These sectors are bad on all InSync devices, so we
2331                  * need to mark them bad on all write targets
2332                  */
2333                 int ok = 1;
2334                 for (i = 0 ; i < conf->raid_disks ; i++)
2335                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2336                                 struct md_rdev *rdev =
2337                                         rcu_dereference(conf->mirrors[i].rdev);
2338                                 ok = rdev_set_badblocks(rdev, sector_nr,
2339                                                         min_bad, 0
2340                                         ) && ok;
2341                         }
2342                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2343                 *skipped = 1;
2344                 put_buf(r1_bio);
2345
2346                 if (!ok) {
2347                         /* Cannot record the badblocks, so need to
2348                          * abort the resync.
2349                          * If there are multiple read targets, could just
2350                          * fail the really bad ones ???
2351                          */
2352                         conf->recovery_disabled = mddev->recovery_disabled;
2353                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2354                         return 0;
2355                 } else
2356                         return min_bad;
2357
2358         }
2359         if (min_bad > 0 && min_bad < good_sectors) {
2360                 /* only resync enough to reach the next bad->good
2361                  * transition */
2362                 good_sectors = min_bad;
2363         }
2364
2365         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2366                 /* extra read targets are also write targets */
2367                 write_targets += read_targets-1;
2368
2369         if (write_targets == 0 || read_targets == 0) {
2370                 /* There is nowhere to write, so all non-sync
2371                  * drives must be failed - so we are finished
2372                  */
2373                 sector_t rv;
2374                 if (min_bad > 0)
2375                         max_sector = sector_nr + min_bad;
2376                 rv = max_sector - sector_nr;
2377                 *skipped = 1;
2378                 put_buf(r1_bio);
2379                 return rv;
2380         }
2381
2382         if (max_sector > mddev->resync_max)
2383                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2384         if (max_sector > sector_nr + good_sectors)
2385                 max_sector = sector_nr + good_sectors;
2386         nr_sectors = 0;
2387         sync_blocks = 0;
2388         do {
2389                 struct page *page;
2390                 int len = PAGE_SIZE;
2391                 if (sector_nr + (len>>9) > max_sector)
2392                         len = (max_sector - sector_nr) << 9;
2393                 if (len == 0)
2394                         break;
2395                 if (sync_blocks == 0) {
2396                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2397                                                &sync_blocks, still_degraded) &&
2398                             !conf->fullsync &&
2399                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2400                                 break;
2401                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2402                         if ((len >> 9) > sync_blocks)
2403                                 len = sync_blocks<<9;
2404                 }
2405
2406                 for (i=0 ; i < conf->raid_disks; i++) {
2407                         bio = r1_bio->bios[i];
2408                         if (bio->bi_end_io) {
2409                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2410                                 if (bio_add_page(bio, page, len, 0) == 0) {
2411                                         /* stop here */
2412                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2413                                         while (i > 0) {
2414                                                 i--;
2415                                                 bio = r1_bio->bios[i];
2416                                                 if (bio->bi_end_io==NULL)
2417                                                         continue;
2418                                                 /* remove last page from this bio */
2419                                                 bio->bi_vcnt--;
2420                                                 bio->bi_size -= len;
2421                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2422                                         }
2423                                         goto bio_full;
2424                                 }
2425                         }
2426                 }
2427                 nr_sectors += len>>9;
2428                 sector_nr += len>>9;
2429                 sync_blocks -= (len>>9);
2430         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2431  bio_full:
2432         r1_bio->sectors = nr_sectors;
2433
2434         /* For a user-requested sync, we read all readable devices and do a
2435          * compare
2436          */
2437         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2438                 atomic_set(&r1_bio->remaining, read_targets);
2439                 for (i = 0; i < conf->raid_disks && read_targets; i++) {
2440                         bio = r1_bio->bios[i];
2441                         if (bio->bi_end_io == end_sync_read) {
2442                                 read_targets--;
2443                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2444                                 generic_make_request(bio);
2445                         }
2446                 }
2447         } else {
2448                 atomic_set(&r1_bio->remaining, 1);
2449                 bio = r1_bio->bios[r1_bio->read_disk];
2450                 md_sync_acct(bio->bi_bdev, nr_sectors);
2451                 generic_make_request(bio);
2452
2453         }
2454         return nr_sectors;
2455 }
2456
2457 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2458 {
2459         if (sectors)
2460                 return sectors;
2461
2462         return mddev->dev_sectors;
2463 }
2464
2465 static struct r1conf *setup_conf(struct mddev *mddev)
2466 {
2467         struct r1conf *conf;
2468         int i;
2469         struct mirror_info *disk;
2470         struct md_rdev *rdev;
2471         int err = -ENOMEM;
2472
2473         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2474         if (!conf)
2475                 goto abort;
2476
2477         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2478                                  GFP_KERNEL);
2479         if (!conf->mirrors)
2480                 goto abort;
2481
2482         conf->tmppage = alloc_page(GFP_KERNEL);
2483         if (!conf->tmppage)
2484                 goto abort;
2485
2486         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2487         if (!conf->poolinfo)
2488                 goto abort;
2489         conf->poolinfo->raid_disks = mddev->raid_disks;
2490         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2491                                           r1bio_pool_free,
2492                                           conf->poolinfo);
2493         if (!conf->r1bio_pool)
2494                 goto abort;
2495
2496         conf->poolinfo->mddev = mddev;
2497
2498         spin_lock_init(&conf->device_lock);
2499         list_for_each_entry(rdev, &mddev->disks, same_set) {
2500                 int disk_idx = rdev->raid_disk;
2501                 if (disk_idx >= mddev->raid_disks
2502                     || disk_idx < 0)
2503                         continue;
2504                 disk = conf->mirrors + disk_idx;
2505
2506                 disk->rdev = rdev;
2507
2508                 disk->head_position = 0;
2509         }
2510         conf->raid_disks = mddev->raid_disks;
2511         conf->mddev = mddev;
2512         INIT_LIST_HEAD(&conf->retry_list);
2513         INIT_LIST_HEAD(&conf->bio_end_io_list);
2514
2515         spin_lock_init(&conf->resync_lock);
2516         init_waitqueue_head(&conf->wait_barrier);
2517
2518         bio_list_init(&conf->pending_bio_list);
2519         conf->pending_count = 0;
2520         conf->recovery_disabled = mddev->recovery_disabled - 1;
2521
2522         conf->last_used = -1;
2523         for (i = 0; i < conf->raid_disks; i++) {
2524
2525                 disk = conf->mirrors + i;
2526
2527                 if (!disk->rdev ||
2528                     !test_bit(In_sync, &disk->rdev->flags)) {
2529                         disk->head_position = 0;
2530                         if (disk->rdev)
2531                                 conf->fullsync = 1;
2532                 } else if (conf->last_used < 0)
2533                         /*
2534                          * The first working device is used as a
2535                          * starting point to read balancing.
2536                          */
2537                         conf->last_used = i;
2538         }
2539
2540         err = -EIO;
2541         if (conf->last_used < 0) {
2542                 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
2543                        mdname(mddev));
2544                 goto abort;
2545         }
2546         err = -ENOMEM;
2547         conf->thread = md_register_thread(raid1d, mddev, NULL);
2548         if (!conf->thread) {
2549                 printk(KERN_ERR
2550                        "md/raid1:%s: couldn't allocate thread\n",
2551                        mdname(mddev));
2552                 goto abort;
2553         }
2554
2555         return conf;
2556
2557  abort:
2558         if (conf) {
2559                 if (conf->r1bio_pool)
2560                         mempool_destroy(conf->r1bio_pool);
2561                 kfree(conf->mirrors);
2562                 safe_put_page(conf->tmppage);
2563                 kfree(conf->poolinfo);
2564                 kfree(conf);
2565         }
2566         return ERR_PTR(err);
2567 }
2568
2569 static int run(struct mddev *mddev)
2570 {
2571         struct r1conf *conf;
2572         int i;
2573         struct md_rdev *rdev;
2574
2575         if (mddev->level != 1) {
2576                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2577                        mdname(mddev), mddev->level);
2578                 return -EIO;
2579         }
2580         if (mddev->reshape_position != MaxSector) {
2581                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2582                        mdname(mddev));
2583                 return -EIO;
2584         }
2585         /*
2586          * copy the already verified devices into our private RAID1
2587          * bookkeeping area. [whatever we allocate in run(),
2588          * should be freed in stop()]
2589          */
2590         if (mddev->private == NULL)
2591                 conf = setup_conf(mddev);
2592         else
2593                 conf = mddev->private;
2594
2595         if (IS_ERR(conf))
2596                 return PTR_ERR(conf);
2597
2598         list_for_each_entry(rdev, &mddev->disks, same_set) {
2599                 if (!mddev->gendisk)
2600                         continue;
2601                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2602                                   rdev->data_offset << 9);
2603                 /* as we don't honour merge_bvec_fn, we must never risk
2604                  * violating it, so limit ->max_segments to 1 lying within
2605                  * a single page, as a one page request is never in violation.
2606                  */
2607                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2608                         blk_queue_max_segments(mddev->queue, 1);
2609                         blk_queue_segment_boundary(mddev->queue,
2610                                                    PAGE_CACHE_SIZE - 1);
2611                 }
2612         }
2613
2614         mddev->degraded = 0;
2615         for (i=0; i < conf->raid_disks; i++)
2616                 if (conf->mirrors[i].rdev == NULL ||
2617                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2618                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2619                         mddev->degraded++;
2620
2621         if (conf->raid_disks - mddev->degraded == 1)
2622                 mddev->recovery_cp = MaxSector;
2623
2624         if (mddev->recovery_cp != MaxSector)
2625                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2626                        " -- starting background reconstruction\n",
2627                        mdname(mddev));
2628         printk(KERN_INFO
2629                 "md/raid1:%s: active with %d out of %d mirrors\n",
2630                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2631                 mddev->raid_disks);
2632
2633         /*
2634          * Ok, everything is just fine now
2635          */
2636         mddev->thread = conf->thread;
2637         conf->thread = NULL;
2638         mddev->private = conf;
2639
2640         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2641
2642         if (mddev->queue) {
2643                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2644                 mddev->queue->backing_dev_info.congested_data = mddev;
2645         }
2646         return md_integrity_register(mddev);
2647 }
2648
2649 static int stop(struct mddev *mddev)
2650 {
2651         struct r1conf *conf = mddev->private;
2652         struct bitmap *bitmap = mddev->bitmap;
2653
2654         /* wait for behind writes to complete */
2655         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2656                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2657                        mdname(mddev));
2658                 /* need to kick something here to make sure I/O goes? */
2659                 wait_event(bitmap->behind_wait,
2660                            atomic_read(&bitmap->behind_writes) == 0);
2661         }
2662
2663         raise_barrier(conf);
2664         lower_barrier(conf);
2665
2666         md_unregister_thread(&mddev->thread);
2667         if (conf->r1bio_pool)
2668                 mempool_destroy(conf->r1bio_pool);
2669         kfree(conf->mirrors);
2670         kfree(conf->poolinfo);
2671         kfree(conf);
2672         mddev->private = NULL;
2673         return 0;
2674 }
2675
2676 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2677 {
2678         /* no resync is happening, and there is enough space
2679          * on all devices, so we can resize.
2680          * We need to make sure resync covers any new space.
2681          * If the array is shrinking we should possibly wait until
2682          * any io in the removed space completes, but it hardly seems
2683          * worth it.
2684          */
2685         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2686         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2687                 return -EINVAL;
2688         set_capacity(mddev->gendisk, mddev->array_sectors);
2689         revalidate_disk(mddev->gendisk);
2690         if (sectors > mddev->dev_sectors &&
2691             mddev->recovery_cp > mddev->dev_sectors) {
2692                 mddev->recovery_cp = mddev->dev_sectors;
2693                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2694         }
2695         mddev->dev_sectors = sectors;
2696         mddev->resync_max_sectors = sectors;
2697         return 0;
2698 }
2699
2700 static int raid1_reshape(struct mddev *mddev)
2701 {
2702         /* We need to:
2703          * 1/ resize the r1bio_pool
2704          * 2/ resize conf->mirrors
2705          *
2706          * We allocate a new r1bio_pool if we can.
2707          * Then raise a device barrier and wait until all IO stops.
2708          * Then resize conf->mirrors and swap in the new r1bio pool.
2709          *
2710          * At the same time, we "pack" the devices so that all the missing
2711          * devices have the higher raid_disk numbers.
2712          */
2713         mempool_t *newpool, *oldpool;
2714         struct pool_info *newpoolinfo;
2715         struct mirror_info *newmirrors;
2716         struct r1conf *conf = mddev->private;
2717         int cnt, raid_disks;
2718         unsigned long flags;
2719         int d, d2, err;
2720
2721         /* Cannot change chunk_size, layout, or level */
2722         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2723             mddev->layout != mddev->new_layout ||
2724             mddev->level != mddev->new_level) {
2725                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2726                 mddev->new_layout = mddev->layout;
2727                 mddev->new_level = mddev->level;
2728                 return -EINVAL;
2729         }
2730
2731         err = md_allow_write(mddev);
2732         if (err)
2733                 return err;
2734
2735         raid_disks = mddev->raid_disks + mddev->delta_disks;
2736
2737         if (raid_disks < conf->raid_disks) {
2738                 cnt=0;
2739                 for (d= 0; d < conf->raid_disks; d++)
2740                         if (conf->mirrors[d].rdev)
2741                                 cnt++;
2742                 if (cnt > raid_disks)
2743                         return -EBUSY;
2744         }
2745
2746         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2747         if (!newpoolinfo)
2748                 return -ENOMEM;
2749         newpoolinfo->mddev = mddev;
2750         newpoolinfo->raid_disks = raid_disks;
2751
2752         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2753                                  r1bio_pool_free, newpoolinfo);
2754         if (!newpool) {
2755                 kfree(newpoolinfo);
2756                 return -ENOMEM;
2757         }
2758         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2759         if (!newmirrors) {
2760                 kfree(newpoolinfo);
2761                 mempool_destroy(newpool);
2762                 return -ENOMEM;
2763         }
2764
2765         raise_barrier(conf);
2766
2767         /* ok, everything is stopped */
2768         oldpool = conf->r1bio_pool;
2769         conf->r1bio_pool = newpool;
2770
2771         for (d = d2 = 0; d < conf->raid_disks; d++) {
2772                 struct md_rdev *rdev = conf->mirrors[d].rdev;
2773                 if (rdev && rdev->raid_disk != d2) {
2774                         sysfs_unlink_rdev(mddev, rdev);
2775                         rdev->raid_disk = d2;
2776                         sysfs_unlink_rdev(mddev, rdev);
2777                         if (sysfs_link_rdev(mddev, rdev))
2778                                 printk(KERN_WARNING
2779                                        "md/raid1:%s: cannot register rd%d\n",
2780                                        mdname(mddev), rdev->raid_disk);
2781                 }
2782                 if (rdev)
2783                         newmirrors[d2++].rdev = rdev;
2784         }
2785         kfree(conf->mirrors);
2786         conf->mirrors = newmirrors;
2787         kfree(conf->poolinfo);
2788         conf->poolinfo = newpoolinfo;
2789
2790         spin_lock_irqsave(&conf->device_lock, flags);
2791         mddev->degraded += (raid_disks - conf->raid_disks);
2792         spin_unlock_irqrestore(&conf->device_lock, flags);
2793         conf->raid_disks = mddev->raid_disks = raid_disks;
2794         mddev->delta_disks = 0;
2795
2796         conf->last_used = 0; /* just make sure it is in-range */
2797         lower_barrier(conf);
2798
2799         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2800         md_wakeup_thread(mddev->thread);
2801
2802         mempool_destroy(oldpool);
2803         return 0;
2804 }
2805
2806 static void raid1_quiesce(struct mddev *mddev, int state)
2807 {
2808         struct r1conf *conf = mddev->private;
2809
2810         switch(state) {
2811         case 2: /* wake for suspend */
2812                 wake_up(&conf->wait_barrier);
2813                 break;
2814         case 1:
2815                 raise_barrier(conf);
2816                 break;
2817         case 0:
2818                 lower_barrier(conf);
2819                 break;
2820         }
2821 }
2822
2823 static void *raid1_takeover(struct mddev *mddev)
2824 {
2825         /* raid1 can take over:
2826          *  raid5 with 2 devices, any layout or chunk size
2827          */
2828         if (mddev->level == 5 && mddev->raid_disks == 2) {
2829                 struct r1conf *conf;
2830                 mddev->new_level = 1;
2831                 mddev->new_layout = 0;
2832                 mddev->new_chunk_sectors = 0;
2833                 conf = setup_conf(mddev);
2834                 if (!IS_ERR(conf))
2835                         conf->barrier = 1;
2836                 return conf;
2837         }
2838         return ERR_PTR(-EINVAL);
2839 }
2840
2841 static struct md_personality raid1_personality =
2842 {
2843         .name           = "raid1",
2844         .level          = 1,
2845         .owner          = THIS_MODULE,
2846         .make_request   = make_request,
2847         .run            = run,
2848         .stop           = stop,
2849         .status         = status,
2850         .error_handler  = error,
2851         .hot_add_disk   = raid1_add_disk,
2852         .hot_remove_disk= raid1_remove_disk,
2853         .spare_active   = raid1_spare_active,
2854         .sync_request   = sync_request,
2855         .resize         = raid1_resize,
2856         .size           = raid1_size,
2857         .check_reshape  = raid1_reshape,
2858         .quiesce        = raid1_quiesce,
2859         .takeover       = raid1_takeover,
2860 };
2861
2862 static int __init raid_init(void)
2863 {
2864         return register_md_personality(&raid1_personality);
2865 }
2866
2867 static void raid_exit(void)
2868 {
2869         unregister_md_personality(&raid1_personality);
2870 }
2871
2872 module_init(raid_init);
2873 module_exit(raid_exit);
2874 MODULE_LICENSE("GPL");
2875 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2876 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2877 MODULE_ALIAS("md-raid1");
2878 MODULE_ALIAS("md-level-1");
2879
2880 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);