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