drivers/md/raid1.c

   1 /*
   2  * raid1.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5  *
   6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7  *
   8  * RAID-1 management functions.
   9  *
  10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11  *
  12  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14  *
  15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  16  * bitmapped intelligence in resync:
  17  *
  18  *      - bitmap marked during normal i/o
  19  *      - bitmap used to skip nondirty blocks during sync
  20  *
  21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  22  * - persistent bitmap code
  23  *
  24  * This program is free software; you can redistribute it and/or modify
  25  * it under the terms of the GNU General Public License as published by
  26  * the Free Software Foundation; either version 2, or (at your option)
  27  * any later version.
  28  *
  29  * You should have received a copy of the GNU General Public License
  30  * (for example /usr/src/linux/COPYING); if not, write to the Free
  31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  32  */
  33
  34 #include <linux/slab.h>
  35 #include <linux/delay.h>
  36 #include <linux/blkdev.h>
  37 #include <linux/seq_file.h>
  38 #include <linux/ratelimit.h>
  39 #include "md.h"
  40 #include "raid1.h"
  41 #include "bitmap.h"
  42
  43 #define DEBUG 0
  44 #define PRINTK(x...) do { if (DEBUG) printk(x); } while (0)
  45
  46 /*
  47  * Number of guaranteed r1bios in case of extreme VM load:
  48  */
  49 #define NR_RAID1_BIOS 256
  50
  51
  52 static void allow_barrier(conf_t *conf);
  53 static void lower_barrier(conf_t *conf);
  54
  55 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  56 {
  57         struct pool_info *pi = data;
  58         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
  59
  60         /* allocate a r1bio with room for raid_disks entries in the bios array */
  61         return kzalloc(size, gfp_flags);
  62 }
  63
  64 static void r1bio_pool_free(void *r1_bio, void *data)
  65 {
  66         kfree(r1_bio);
  67 }
  68
  69 #define RESYNC_BLOCK_SIZE (64*1024)
  70 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  71 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  72 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  73 #define RESYNC_WINDOW (2048*1024)
  74
  75 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  76 {
  77         struct pool_info *pi = data;
  78         struct page *page;
  79         r1bio_t *r1_bio;
  80         struct bio *bio;
  81         int i, j;
  82
  83         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  84         if (!r1_bio)
  85                 return NULL;
  86
  87         /*
  88          * Allocate bios : 1 for reading, n-1 for writing
  89          */
  90         for (j = pi->raid_disks ; j-- ; ) {
  91                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
  92                 if (!bio)
  93                         goto out_free_bio;
  94                 r1_bio->bios[j] = bio;
  95         }
  96         /*
  97          * Allocate RESYNC_PAGES data pages and attach them to
  98          * the first bio.
  99          * If this is a user-requested check/repair, allocate
 100          * RESYNC_PAGES for each bio.
 101          */
 102         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 103                 j = pi->raid_disks;
 104         else
 105                 j = 1;
 106         while(j--) {
 107                 bio = r1_bio->bios[j];
 108                 for (i = 0; i < RESYNC_PAGES; i++) {
 109                         page = alloc_page(gfp_flags);
 110                         if (unlikely(!page))
 111                                 goto out_free_pages;
 112
 113                         bio->bi_io_vec[i].bv_page = page;
 114                         bio->bi_vcnt = i+1;
 115                 }
 116         }
 117         /* If not user-requests, copy the page pointers to all bios */
 118         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 119                 for (i=0; i<RESYNC_PAGES ; i++)
 120                         for (j=1; j<pi->raid_disks; j++)
 121                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 122                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 123         }
 124
 125         r1_bio->master_bio = NULL;
 126
 127         return r1_bio;
 128
 129 out_free_pages:
 130         for (j=0 ; j < pi->raid_disks; j++)
 131                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 132                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 133         j = -1;
 134 out_free_bio:
 135         while ( ++j < pi->raid_disks )
 136                 bio_put(r1_bio->bios[j]);
 137         r1bio_pool_free(r1_bio, data);
 138         return NULL;
 139 }
 140
 141 static void r1buf_pool_free(void *__r1_bio, void *data)
 142 {
 143         struct pool_info *pi = data;
 144         int i,j;
 145         r1bio_t *r1bio = __r1_bio;
 146
 147         for (i = 0; i < RESYNC_PAGES; i++)
 148                 for (j = pi->raid_disks; j-- ;) {
 149                         if (j == 0 ||
 150                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 151                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 152                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 153                 }
 154         for (i=0 ; i < pi->raid_disks; i++)
 155                 bio_put(r1bio->bios[i]);
 156
 157         r1bio_pool_free(r1bio, data);
 158 }
 159
 160 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 161 {
 162         int i;
 163
 164         for (i = 0; i < conf->raid_disks; i++) {
 165                 struct bio **bio = r1_bio->bios + i;
 166                 if (!BIO_SPECIAL(*bio))
 167                         bio_put(*bio);
 168                 *bio = NULL;
 169         }
 170 }
 171
 172 static void free_r1bio(r1bio_t *r1_bio)
 173 {
 174         conf_t *conf = r1_bio->mddev->private;
 175
 176         put_all_bios(conf, r1_bio);
 177         mempool_free(r1_bio, conf->r1bio_pool);
 178 }
 179
 180 static void put_buf(r1bio_t *r1_bio)
 181 {
 182         conf_t *conf = r1_bio->mddev->private;
 183         int i;
 184
 185         for (i=0; i<conf->raid_disks; i++) {
 186                 struct bio *bio = r1_bio->bios[i];
 187                 if (bio->bi_end_io)
 188                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 189         }
 190
 191         mempool_free(r1_bio, conf->r1buf_pool);
 192
 193         lower_barrier(conf);
 194 }
 195
 196 static void reschedule_retry(r1bio_t *r1_bio)
 197 {
 198         unsigned long flags;
 199         mddev_t *mddev = r1_bio->mddev;
 200         conf_t *conf = mddev->private;
 201
 202         spin_lock_irqsave(&conf->device_lock, flags);
 203         list_add(&r1_bio->retry_list, &conf->retry_list);
 204         conf->nr_queued ++;
 205         spin_unlock_irqrestore(&conf->device_lock, flags);
 206
 207         wake_up(&conf->wait_barrier);
 208         md_wakeup_thread(mddev->thread);
 209 }
 210
 211 /*
 212  * raid_end_bio_io() is called when we have finished servicing a mirrored
 213  * operation and are ready to return a success/failure code to the buffer
 214  * cache layer.
 215  */
 216 static void call_bio_endio(r1bio_t *r1_bio)
 217 {
 218         struct bio *bio = r1_bio->master_bio;
 219         int done;
 220         conf_t *conf = r1_bio->mddev->private;
 221
 222         if (bio->bi_phys_segments) {
 223                 unsigned long flags;
 224                 spin_lock_irqsave(&conf->device_lock, flags);
 225                 bio->bi_phys_segments--;
 226                 done = (bio->bi_phys_segments == 0);
 227                 spin_unlock_irqrestore(&conf->device_lock, flags);
 228         } else
 229                 done = 1;
 230
 231         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 232                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 233         if (done) {
 234                 bio_endio(bio, 0);
 235                 /*
 236                  * Wake up any possible resync thread that waits for the device
 237                  * to go idle.
 238                  */
 239                 allow_barrier(conf);
 240         }
 241 }
 242
 243 static void raid_end_bio_io(r1bio_t *r1_bio)
 244 {
 245         struct bio *bio = r1_bio->master_bio;
 246
 247         /* if nobody has done the final endio yet, do it now */
 248         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 249                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
 250                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 251                         (unsigned long long) bio->bi_sector,
 252                         (unsigned long long) bio->bi_sector +
 253                                 (bio->bi_size >> 9) - 1);
 254
 255                 call_bio_endio(r1_bio);
 256         }
 257         free_r1bio(r1_bio);
 258 }
 259
 260 /*
 261  * Update disk head position estimator based on IRQ completion info.
 262  */
 263 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 264 {
 265         conf_t *conf = r1_bio->mddev->private;
 266
 267         conf->mirrors[disk].head_position =
 268                 r1_bio->sector + (r1_bio->sectors);
 269 }
 270
 271 static void raid1_end_read_request(struct bio *bio, int error)
 272 {
 273         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 274         r1bio_t *r1_bio = bio->bi_private;
 275         int mirror;
 276         conf_t *conf = r1_bio->mddev->private;
 277
 278         mirror = r1_bio->read_disk;
 279         /*
 280          * this branch is our 'one mirror IO has finished' event handler:
 281          */
 282         update_head_pos(mirror, r1_bio);
 283
 284         if (uptodate)
 285                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 286         else {
 287                 /* If all other devices have failed, we want to return
 288                  * the error upwards rather than fail the last device.
 289                  * Here we redefine "uptodate" to mean "Don't want to retry"
 290                  */
 291                 unsigned long flags;
 292                 spin_lock_irqsave(&conf->device_lock, flags);
 293                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 294                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 295                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 296                         uptodate = 1;
 297                 spin_unlock_irqrestore(&conf->device_lock, flags);
 298         }
 299
 300         if (uptodate)
 301                 raid_end_bio_io(r1_bio);
 302         else {
 303                 /*
 304                  * oops, read error:
 305                  */
 306                 char b[BDEVNAME_SIZE];
 307                 printk_ratelimited(
 308                         KERN_ERR "md/raid1:%s: %s: "
 309                         "rescheduling sector %llu\n",
 310                         mdname(conf->mddev),
 311                         bdevname(conf->mirrors[mirror].rdev->bdev,
 312                                  b),
 313                         (unsigned long long)r1_bio->sector);
 314                 set_bit(R1BIO_ReadError, &r1_bio->state);
 315                 reschedule_retry(r1_bio);
 316         }
 317
 318         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 319 }
 320
 321 static void r1_bio_write_done(r1bio_t *r1_bio)
 322 {
 323         if (atomic_dec_and_test(&r1_bio->remaining))
 324         {
 325                 /* it really is the end of this request */
 326                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 327                         /* free extra copy of the data pages */
 328                         int i = r1_bio->behind_page_count;
 329                         while (i--)
 330                                 safe_put_page(r1_bio->behind_pages[i]);
 331                         kfree(r1_bio->behind_pages);
 332                         r1_bio->behind_pages = NULL;
 333                 }
 334                 /* clear the bitmap if all writes complete successfully */
 335                 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 336                                 r1_bio->sectors,
 337                                 !test_bit(R1BIO_Degraded, &r1_bio->state),
 338                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 339                 md_write_end(r1_bio->mddev);
 340                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 341                         reschedule_retry(r1_bio);
 342                 else
 343                         raid_end_bio_io(r1_bio);
 344         }
 345 }
 346
 347 static void raid1_end_write_request(struct bio *bio, int error)
 348 {
 349         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 350         r1bio_t *r1_bio = bio->bi_private;
 351         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 352         conf_t *conf = r1_bio->mddev->private;
 353         struct bio *to_put = NULL;
 354
 355
 356         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 357                 if (r1_bio->bios[mirror] == bio)
 358                         break;
 359
 360         /*
 361          * 'one mirror IO has finished' event handler:
 362          */
 363         r1_bio->bios[mirror] = NULL;
 364         to_put = bio;
 365         if (!uptodate) {
 366                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 367                 /* an I/O failed, we can't clear the bitmap */
 368                 set_bit(R1BIO_Degraded, &r1_bio->state);
 369         } else {
 370                 /*
 371                  * Set R1BIO_Uptodate in our master bio, so that we
 372                  * will return a good error code for to the higher
 373                  * levels even if IO on some other mirrored buffer
 374                  * fails.
 375                  *
 376                  * The 'master' represents the composite IO operation
 377                  * to user-side. So if something waits for IO, then it
 378                  * will wait for the 'master' bio.
 379                  */
 380                 sector_t first_bad;
 381                 int bad_sectors;
 382
 383                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 384
 385                 /* Maybe we can clear some bad blocks. */
 386                 if (is_badblock(conf->mirrors[mirror].rdev,
 387                                 r1_bio->sector, r1_bio->sectors,
 388                                 &first_bad, &bad_sectors)) {
 389                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 390                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 391                 }
 392         }
 393
 394         update_head_pos(mirror, r1_bio);
 395
 396         if (behind) {
 397                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 398                         atomic_dec(&r1_bio->behind_remaining);
 399
 400                 /*
 401                  * In behind mode, we ACK the master bio once the I/O
 402                  * has safely reached all non-writemostly
 403                  * disks. Setting the Returned bit ensures that this
 404                  * gets done only once -- we don't ever want to return
 405                  * -EIO here, instead we'll wait
 406                  */
 407                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 408                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 409                         /* Maybe we can return now */
 410                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 411                                 struct bio *mbio = r1_bio->master_bio;
 412                                 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 413                                        (unsigned long long) mbio->bi_sector,
 414                                        (unsigned long long) mbio->bi_sector +
 415                                        (mbio->bi_size >> 9) - 1);
 416                                 call_bio_endio(r1_bio);
 417                         }
 418                 }
 419         }
 420         if (r1_bio->bios[mirror] == NULL)
 421                 rdev_dec_pending(conf->mirrors[mirror].rdev,
 422                                  conf->mddev);
 423
 424         /*
 425          * Let's see if all mirrored write operations have finished
 426          * already.
 427          */
 428         r1_bio_write_done(r1_bio);
 429
 430         if (to_put)
 431                 bio_put(to_put);
 432 }
 433
 434
 435 /*
 436  * This routine returns the disk from which the requested read should
 437  * be done. There is a per-array 'next expected sequential IO' sector
 438  * number - if this matches on the next IO then we use the last disk.
 439  * There is also a per-disk 'last know head position' sector that is
 440  * maintained from IRQ contexts, both the normal and the resync IO
 441  * completion handlers update this position correctly. If there is no
 442  * perfect sequential match then we pick the disk whose head is closest.
 443  *
 444  * If there are 2 mirrors in the same 2 devices, performance degrades
 445  * because position is mirror, not device based.
 446  *
 447  * The rdev for the device selected will have nr_pending incremented.
 448  */
 449 static int read_balance(conf_t *conf, r1bio_t *r1_bio, int *max_sectors)
 450 {
 451         const sector_t this_sector = r1_bio->sector;
 452         int sectors;
 453         int best_good_sectors;
 454         int start_disk;
 455         int best_disk;
 456         int i;
 457         sector_t best_dist;
 458         mdk_rdev_t *rdev;
 459         int choose_first;
 460
 461         rcu_read_lock();
 462         /*
 463          * Check if we can balance. We can balance on the whole
 464          * device if no resync is going on, or below the resync window.
 465          * We take the first readable disk when above the resync window.
 466          */
 467  retry:
 468         sectors = r1_bio->sectors;
 469         best_disk = -1;
 470         best_dist = MaxSector;
 471         best_good_sectors = 0;
 472
 473         if (conf->mddev->recovery_cp < MaxSector &&
 474             (this_sector + sectors >= conf->next_resync)) {
 475                 choose_first = 1;
 476                 start_disk = 0;
 477         } else {
 478                 choose_first = 0;
 479                 start_disk = conf->last_used;
 480         }
 481
 482         for (i = 0 ; i < conf->raid_disks ; i++) {
 483                 sector_t dist;
 484                 sector_t first_bad;
 485                 int bad_sectors;
 486
 487                 int disk = start_disk + i;
 488                 if (disk >= conf->raid_disks)
 489                         disk -= conf->raid_disks;
 490
 491                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 492                 if (r1_bio->bios[disk] == IO_BLOCKED
 493                     || rdev == NULL
 494                     || test_bit(Faulty, &rdev->flags))
 495                         continue;
 496                 if (!test_bit(In_sync, &rdev->flags) &&
 497                     rdev->recovery_offset < this_sector + sectors)
 498                         continue;
 499                 if (test_bit(WriteMostly, &rdev->flags)) {
 500                         /* Don't balance among write-mostly, just
 501                          * use the first as a last resort */
 502                         if (best_disk < 0)
 503                                 best_disk = disk;
 504                         continue;
 505                 }
 506                 /* This is a reasonable device to use.  It might
 507                  * even be best.
 508                  */
 509                 if (is_badblock(rdev, this_sector, sectors,
 510                                 &first_bad, &bad_sectors)) {
 511                         if (best_dist < MaxSector)
 512                                 /* already have a better device */
 513                                 continue;
 514                         if (first_bad <= this_sector) {
 515                                 /* cannot read here. If this is the 'primary'
 516                                  * device, then we must not read beyond
 517                                  * bad_sectors from another device..
 518                                  */
 519                                 bad_sectors -= (this_sector - first_bad);
 520                                 if (choose_first && sectors > bad_sectors)
 521                                         sectors = bad_sectors;
 522                                 if (best_good_sectors > sectors)
 523                                         best_good_sectors = sectors;
 524
 525                         } else {
 526                                 sector_t good_sectors = first_bad - this_sector;
 527                                 if (good_sectors > best_good_sectors) {
 528                                         best_good_sectors = good_sectors;
 529                                         best_disk = disk;
 530                                 }
 531                                 if (choose_first)
 532                                         break;
 533                         }
 534                         continue;
 535                 } else
 536                         best_good_sectors = sectors;
 537
 538                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 539                 if (choose_first
 540                     /* Don't change to another disk for sequential reads */
 541                     || conf->next_seq_sect == this_sector
 542                     || dist == 0
 543                     /* If device is idle, use it */
 544                     || atomic_read(&rdev->nr_pending) == 0) {
 545                         best_disk = disk;
 546                         break;
 547                 }
 548                 if (dist < best_dist) {
 549                         best_dist = dist;
 550                         best_disk = disk;
 551                 }
 552         }
 553
 554         if (best_disk >= 0) {
 555                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 556                 if (!rdev)
 557                         goto retry;
 558                 atomic_inc(&rdev->nr_pending);
 559                 if (test_bit(Faulty, &rdev->flags)) {
 560                         /* cannot risk returning a device that failed
 561                          * before we inc'ed nr_pending
 562                          */
 563                         rdev_dec_pending(rdev, conf->mddev);
 564                         goto retry;
 565                 }
 566                 sectors = best_good_sectors;
 567                 conf->next_seq_sect = this_sector + sectors;
 568                 conf->last_used = best_disk;
 569         }
 570         rcu_read_unlock();
 571         *max_sectors = sectors;
 572
 573         return best_disk;
 574 }
 575
 576 int md_raid1_congested(mddev_t *mddev, int bits)
 577 {
 578         conf_t *conf = mddev->private;
 579         int i, ret = 0;
 580
 581         rcu_read_lock();
 582         for (i = 0; i < mddev->raid_disks; i++) {
 583                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 584                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 585                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 586
 587                         BUG_ON(!q);
 588
 589                         /* Note the '|| 1' - when read_balance prefers
 590                          * non-congested targets, it can be removed
 591                          */
 592                         if ((bits & (1<<BDI_async_congested)) || 1)
 593                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 594                         else
 595                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 596                 }
 597         }
 598         rcu_read_unlock();
 599         return ret;
 600 }
 601 EXPORT_SYMBOL_GPL(md_raid1_congested);
 602
 603 static int raid1_congested(void *data, int bits)
 604 {
 605         mddev_t *mddev = data;
 606
 607         return mddev_congested(mddev, bits) ||
 608                 md_raid1_congested(mddev, bits);
 609 }
 610
 611 static void flush_pending_writes(conf_t *conf)
 612 {
 613         /* Any writes that have been queued but are awaiting
 614          * bitmap updates get flushed here.
 615          */
 616         spin_lock_irq(&conf->device_lock);
 617
 618         if (conf->pending_bio_list.head) {
 619                 struct bio *bio;
 620                 bio = bio_list_get(&conf->pending_bio_list);
 621                 spin_unlock_irq(&conf->device_lock);
 622                 /* flush any pending bitmap writes to
 623                  * disk before proceeding w/ I/O */
 624                 bitmap_unplug(conf->mddev->bitmap);
 625
 626                 while (bio) { /* submit pending writes */
 627                         struct bio *next = bio->bi_next;
 628                         bio->bi_next = NULL;
 629                         generic_make_request(bio);
 630                         bio = next;
 631                 }
 632         } else
 633                 spin_unlock_irq(&conf->device_lock);
 634 }
 635
 636 /* Barriers....
 637  * Sometimes we need to suspend IO while we do something else,
 638  * either some resync/recovery, or reconfigure the array.
 639  * To do this we raise a 'barrier'.
 640  * The 'barrier' is a counter that can be raised multiple times
 641  * to count how many activities are happening which preclude
 642  * normal IO.
 643  * We can only raise the barrier if there is no pending IO.
 644  * i.e. if nr_pending == 0.
 645  * We choose only to raise the barrier if no-one is waiting for the
 646  * barrier to go down.  This means that as soon as an IO request
 647  * is ready, no other operations which require a barrier will start
 648  * until the IO request has had a chance.
 649  *
 650  * So: regular IO calls 'wait_barrier'.  When that returns there
 651  *    is no backgroup IO happening,  It must arrange to call
 652  *    allow_barrier when it has finished its IO.
 653  * backgroup IO calls must call raise_barrier.  Once that returns
 654  *    there is no normal IO happeing.  It must arrange to call
 655  *    lower_barrier when the particular background IO completes.
 656  */
 657 #define RESYNC_DEPTH 32
 658
 659 static void raise_barrier(conf_t *conf)
 660 {
 661         spin_lock_irq(&conf->resync_lock);
 662
 663         /* Wait until no block IO is waiting */
 664         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 665                             conf->resync_lock, );
 666
 667         /* block any new IO from starting */
 668         conf->barrier++;
 669
 670         /* Now wait for all pending IO to complete */
 671         wait_event_lock_irq(conf->wait_barrier,
 672                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 673                             conf->resync_lock, );
 674
 675         spin_unlock_irq(&conf->resync_lock);
 676 }
 677
 678 static void lower_barrier(conf_t *conf)
 679 {
 680         unsigned long flags;
 681         BUG_ON(conf->barrier <= 0);
 682         spin_lock_irqsave(&conf->resync_lock, flags);
 683         conf->barrier--;
 684         spin_unlock_irqrestore(&conf->resync_lock, flags);
 685         wake_up(&conf->wait_barrier);
 686 }
 687
 688 static void wait_barrier(conf_t *conf)
 689 {
 690         spin_lock_irq(&conf->resync_lock);
 691         if (conf->barrier) {
 692                 conf->nr_waiting++;
 693                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 694                                     conf->resync_lock,
 695                                     );
 696                 conf->nr_waiting--;
 697         }
 698         conf->nr_pending++;
 699         spin_unlock_irq(&conf->resync_lock);
 700 }
 701
 702 static void allow_barrier(conf_t *conf)
 703 {
 704         unsigned long flags;
 705         spin_lock_irqsave(&conf->resync_lock, flags);
 706         conf->nr_pending--;
 707         spin_unlock_irqrestore(&conf->resync_lock, flags);
 708         wake_up(&conf->wait_barrier);
 709 }
 710
 711 static void freeze_array(conf_t *conf)
 712 {
 713         /* stop syncio and normal IO and wait for everything to
 714          * go quite.
 715          * We increment barrier and nr_waiting, and then
 716          * wait until nr_pending match nr_queued+1
 717          * This is called in the context of one normal IO request
 718          * that has failed. Thus any sync request that might be pending
 719          * will be blocked by nr_pending, and we need to wait for
 720          * pending IO requests to complete or be queued for re-try.
 721          * Thus the number queued (nr_queued) plus this request (1)
 722          * must match the number of pending IOs (nr_pending) before
 723          * we continue.
 724          */
 725         spin_lock_irq(&conf->resync_lock);
 726         conf->barrier++;
 727         conf->nr_waiting++;
 728         wait_event_lock_irq(conf->wait_barrier,
 729                             conf->nr_pending == conf->nr_queued+1,
 730                             conf->resync_lock,
 731                             flush_pending_writes(conf));
 732         spin_unlock_irq(&conf->resync_lock);
 733 }
 734 static void unfreeze_array(conf_t *conf)
 735 {
 736         /* reverse the effect of the freeze */
 737         spin_lock_irq(&conf->resync_lock);
 738         conf->barrier--;
 739         conf->nr_waiting--;
 740         wake_up(&conf->wait_barrier);
 741         spin_unlock_irq(&conf->resync_lock);
 742 }
 743
 744
 745 /* duplicate the data pages for behind I/O
 746  */
 747 static void alloc_behind_pages(struct bio *bio, r1bio_t *r1_bio)
 748 {
 749         int i;
 750         struct bio_vec *bvec;
 751         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page*),
 752                                         GFP_NOIO);
 753         if (unlikely(!pages))
 754                 return;
 755
 756         bio_for_each_segment(bvec, bio, i) {
 757                 pages[i] = alloc_page(GFP_NOIO);
 758                 if (unlikely(!pages[i]))
 759                         goto do_sync_io;
 760                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 761                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 762                 kunmap(pages[i]);
 763                 kunmap(bvec->bv_page);
 764         }
 765         r1_bio->behind_pages = pages;
 766         r1_bio->behind_page_count = bio->bi_vcnt;
 767         set_bit(R1BIO_BehindIO, &r1_bio->state);
 768         return;
 769
 770 do_sync_io:
 771         for (i = 0; i < bio->bi_vcnt; i++)
 772                 if (pages[i])
 773                         put_page(pages[i]);
 774         kfree(pages);
 775         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 776 }
 777
 778 static int make_request(mddev_t *mddev, struct bio * bio)
 779 {
 780         conf_t *conf = mddev->private;
 781         mirror_info_t *mirror;
 782         r1bio_t *r1_bio;
 783         struct bio *read_bio;
 784         int i, disks;
 785         struct bitmap *bitmap;
 786         unsigned long flags;
 787         const int rw = bio_data_dir(bio);
 788         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 789         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
 790         mdk_rdev_t *blocked_rdev;
 791         int plugged;
 792         int first_clone;
 793         int sectors_handled;
 794         int max_sectors;
 795
 796         /*
 797          * Register the new request and wait if the reconstruction
 798          * thread has put up a bar for new requests.
 799          * Continue immediately if no resync is active currently.
 800          */
 801
 802         md_write_start(mddev, bio); /* wait on superblock update early */
 803
 804         if (bio_data_dir(bio) == WRITE &&
 805             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
 806             bio->bi_sector < mddev->suspend_hi) {
 807                 /* As the suspend_* range is controlled by
 808                  * userspace, we want an interruptible
 809                  * wait.
 810                  */
 811                 DEFINE_WAIT(w);
 812                 for (;;) {
 813                         flush_signals(current);
 814                         prepare_to_wait(&conf->wait_barrier,
 815                                         &w, TASK_INTERRUPTIBLE);
 816                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
 817                             bio->bi_sector >= mddev->suspend_hi)
 818                                 break;
 819                         schedule();
 820                 }
 821                 finish_wait(&conf->wait_barrier, &w);
 822         }
 823
 824         wait_barrier(conf);
 825
 826         bitmap = mddev->bitmap;
 827
 828         /*
 829          * make_request() can abort the operation when READA is being
 830          * used and no empty request is available.
 831          *
 832          */
 833         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 834
 835         r1_bio->master_bio = bio;
 836         r1_bio->sectors = bio->bi_size >> 9;
 837         r1_bio->state = 0;
 838         r1_bio->mddev = mddev;
 839         r1_bio->sector = bio->bi_sector;
 840
 841         /* We might need to issue multiple reads to different
 842          * devices if there are bad blocks around, so we keep
 843          * track of the number of reads in bio->bi_phys_segments.
 844          * If this is 0, there is only one r1_bio and no locking
 845          * will be needed when requests complete.  If it is
 846          * non-zero, then it is the number of not-completed requests.
 847          */
 848         bio->bi_phys_segments = 0;
 849         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
 850
 851         if (rw == READ) {
 852                 /*
 853                  * read balancing logic:
 854                  */
 855                 int rdisk;
 856
 857 read_again:
 858                 rdisk = read_balance(conf, r1_bio, &max_sectors);
 859
 860                 if (rdisk < 0) {
 861                         /* couldn't find anywhere to read from */
 862                         raid_end_bio_io(r1_bio);
 863                         return 0;
 864                 }
 865                 mirror = conf->mirrors + rdisk;
 866
 867                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
 868                     bitmap) {
 869                         /* Reading from a write-mostly device must
 870                          * take care not to over-take any writes
 871                          * that are 'behind'
 872                          */
 873                         wait_event(bitmap->behind_wait,
 874                                    atomic_read(&bitmap->behind_writes) == 0);
 875                 }
 876                 r1_bio->read_disk = rdisk;
 877
 878                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 879                 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
 880                             max_sectors);
 881
 882                 r1_bio->bios[rdisk] = read_bio;
 883
 884                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 885                 read_bio->bi_bdev = mirror->rdev->bdev;
 886                 read_bio->bi_end_io = raid1_end_read_request;
 887                 read_bio->bi_rw = READ | do_sync;
 888                 read_bio->bi_private = r1_bio;
 889
 890                 if (max_sectors < r1_bio->sectors) {
 891                         /* could not read all from this device, so we will
 892                          * need another r1_bio.
 893                          */
 894
 895                         sectors_handled = (r1_bio->sector + max_sectors
 896                                            - bio->bi_sector);
 897                         r1_bio->sectors = max_sectors;
 898                         spin_lock_irq(&conf->device_lock);
 899                         if (bio->bi_phys_segments == 0)
 900                                 bio->bi_phys_segments = 2;
 901                         else
 902                                 bio->bi_phys_segments++;
 903                         spin_unlock_irq(&conf->device_lock);
 904                         /* Cannot call generic_make_request directly
 905                          * as that will be queued in __make_request
 906                          * and subsequent mempool_alloc might block waiting
 907                          * for it.  So hand bio over to raid1d.
 908                          */
 909                         reschedule_retry(r1_bio);
 910
 911                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 912
 913                         r1_bio->master_bio = bio;
 914                         r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
 915                         r1_bio->state = 0;
 916                         r1_bio->mddev = mddev;
 917                         r1_bio->sector = bio->bi_sector + sectors_handled;
 918                         goto read_again;
 919                 } else
 920                         generic_make_request(read_bio);
 921                 return 0;
 922         }
 923
 924         /*
 925          * WRITE:
 926          */
 927         /* first select target devices under rcu_lock and
 928          * inc refcount on their rdev.  Record them by setting
 929          * bios[x] to bio
 930          * If there are known/acknowledged bad blocks on any device on
 931          * which we have seen a write error, we want to avoid writing those
 932          * blocks.
 933          * This potentially requires several writes to write around
 934          * the bad blocks.  Each set of writes gets it's own r1bio
 935          * with a set of bios attached.
 936          */
 937         plugged = mddev_check_plugged(mddev);
 938
 939         disks = conf->raid_disks;
 940  retry_write:
 941         blocked_rdev = NULL;
 942         rcu_read_lock();
 943         max_sectors = r1_bio->sectors;
 944         for (i = 0;  i < disks; i++) {
 945                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 946                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 947                         atomic_inc(&rdev->nr_pending);
 948                         blocked_rdev = rdev;
 949                         break;
 950                 }
 951                 r1_bio->bios[i] = NULL;
 952                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
 953                         set_bit(R1BIO_Degraded, &r1_bio->state);
 954                         continue;
 955                 }
 956
 957                 atomic_inc(&rdev->nr_pending);
 958                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
 959                         sector_t first_bad;
 960                         int bad_sectors;
 961                         int is_bad;
 962
 963                         is_bad = is_badblock(rdev, r1_bio->sector,
 964                                              max_sectors,
 965                                              &first_bad, &bad_sectors);
 966                         if (is_bad < 0) {
 967                                 /* mustn't write here until the bad block is
 968                                  * acknowledged*/
 969                                 set_bit(BlockedBadBlocks, &rdev->flags);
 970                                 blocked_rdev = rdev;
 971                                 break;
 972                         }
 973                         if (is_bad && first_bad <= r1_bio->sector) {
 974                                 /* Cannot write here at all */
 975                                 bad_sectors -= (r1_bio->sector - first_bad);
 976                                 if (bad_sectors < max_sectors)
 977                                         /* mustn't write more than bad_sectors
 978                                          * to other devices yet
 979                                          */
 980                                         max_sectors = bad_sectors;
 981                                 rdev_dec_pending(rdev, mddev);
 982                                 /* We don't set R1BIO_Degraded as that
 983                                  * only applies if the disk is
 984                                  * missing, so it might be re-added,
 985                                  * and we want to know to recover this
 986                                  * chunk.
 987                                  * In this case the device is here,
 988                                  * and the fact that this chunk is not
 989                                  * in-sync is recorded in the bad
 990                                  * block log
 991                                  */
 992                                 continue;
 993                         }
 994                         if (is_bad) {
 995                                 int good_sectors = first_bad - r1_bio->sector;
 996                                 if (good_sectors < max_sectors)
 997                                         max_sectors = good_sectors;
 998                         }
 999                 }
1000                 r1_bio->bios[i] = bio;
1001         }
1002         rcu_read_unlock();
1003
1004         if (unlikely(blocked_rdev)) {
1005                 /* Wait for this device to become unblocked */
1006                 int j;
1007
1008                 for (j = 0; j < i; j++)
1009                         if (r1_bio->bios[j])
1010                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1011                 r1_bio->state = 0;
1012                 allow_barrier(conf);
1013                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1014                 wait_barrier(conf);
1015                 goto retry_write;
1016         }
1017
1018         if (max_sectors < r1_bio->sectors) {
1019                 /* We are splitting this write into multiple parts, so
1020                  * we need to prepare for allocating another r1_bio.
1021                  */
1022                 r1_bio->sectors = max_sectors;
1023                 spin_lock_irq(&conf->device_lock);
1024                 if (bio->bi_phys_segments == 0)
1025                         bio->bi_phys_segments = 2;
1026                 else
1027                         bio->bi_phys_segments++;
1028                 spin_unlock_irq(&conf->device_lock);
1029         }
1030         sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1031
1032         atomic_set(&r1_bio->remaining, 1);
1033         atomic_set(&r1_bio->behind_remaining, 0);
1034
1035         first_clone = 1;
1036         for (i = 0; i < disks; i++) {
1037                 struct bio *mbio;
1038                 if (!r1_bio->bios[i])
1039                         continue;
1040
1041                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1042                 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1043
1044                 if (first_clone) {
1045                         /* do behind I/O ?
1046                          * Not if there are too many, or cannot
1047                          * allocate memory, or a reader on WriteMostly
1048                          * is waiting for behind writes to flush */
1049                         if (bitmap &&
1050                             (atomic_read(&bitmap->behind_writes)
1051                              < mddev->bitmap_info.max_write_behind) &&
1052                             !waitqueue_active(&bitmap->behind_wait))
1053                                 alloc_behind_pages(mbio, r1_bio);
1054
1055                         bitmap_startwrite(bitmap, r1_bio->sector,
1056                                           r1_bio->sectors,
1057                                           test_bit(R1BIO_BehindIO,
1058                                                    &r1_bio->state));
1059                         first_clone = 0;
1060                 }
1061                 if (r1_bio->behind_pages) {
1062                         struct bio_vec *bvec;
1063                         int j;
1064
1065                         /* Yes, I really want the '__' version so that
1066                          * we clear any unused pointer in the io_vec, rather
1067                          * than leave them unchanged.  This is important
1068                          * because when we come to free the pages, we won't
1069                          * know the original bi_idx, so we just free
1070                          * them all
1071                          */
1072                         __bio_for_each_segment(bvec, mbio, j, 0)
1073                                 bvec->bv_page = r1_bio->behind_pages[j];
1074                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1075                                 atomic_inc(&r1_bio->behind_remaining);
1076                 }
1077
1078                 r1_bio->bios[i] = mbio;
1079
1080                 mbio->bi_sector = (r1_bio->sector +
1081                                    conf->mirrors[i].rdev->data_offset);
1082                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1083                 mbio->bi_end_io = raid1_end_write_request;
1084                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
1085                 mbio->bi_private = r1_bio;
1086
1087                 atomic_inc(&r1_bio->remaining);
1088                 spin_lock_irqsave(&conf->device_lock, flags);
1089                 bio_list_add(&conf->pending_bio_list, mbio);
1090                 spin_unlock_irqrestore(&conf->device_lock, flags);
1091         }
1092         r1_bio_write_done(r1_bio);
1093
1094         /* In case raid1d snuck in to freeze_array */
1095         wake_up(&conf->wait_barrier);
1096
1097         if (sectors_handled < (bio->bi_size >> 9)) {
1098                 /* We need another r1_bio.  It has already been counted
1099                  * in bio->bi_phys_segments
1100                  */
1101                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1102                 r1_bio->master_bio = bio;
1103                 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1104                 r1_bio->state = 0;
1105                 r1_bio->mddev = mddev;
1106                 r1_bio->sector = bio->bi_sector + sectors_handled;
1107                 goto retry_write;
1108         }
1109
1110         if (do_sync || !bitmap || !plugged)
1111                 md_wakeup_thread(mddev->thread);
1112
1113         return 0;
1114 }
1115
1116 static void status(struct seq_file *seq, mddev_t *mddev)
1117 {
1118         conf_t *conf = mddev->private;
1119         int i;
1120
1121         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1122                    conf->raid_disks - mddev->degraded);
1123         rcu_read_lock();
1124         for (i = 0; i < conf->raid_disks; i++) {
1125                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1126                 seq_printf(seq, "%s",
1127                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1128         }
1129         rcu_read_unlock();
1130         seq_printf(seq, "]");
1131 }
1132
1133
1134 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1135 {
1136         char b[BDEVNAME_SIZE];
1137         conf_t *conf = mddev->private;
1138
1139         /*
1140          * If it is not operational, then we have already marked it as dead
1141          * else if it is the last working disks, ignore the error, let the
1142          * next level up know.
1143          * else mark the drive as failed
1144          */
1145         if (test_bit(In_sync, &rdev->flags)
1146             && (conf->raid_disks - mddev->degraded) == 1) {
1147                 /*
1148                  * Don't fail the drive, act as though we were just a
1149                  * normal single drive.
1150                  * However don't try a recovery from this drive as
1151                  * it is very likely to fail.
1152                  */
1153                 conf->recovery_disabled = mddev->recovery_disabled;
1154                 return;
1155         }
1156         set_bit(Blocked, &rdev->flags);
1157         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1158                 unsigned long flags;
1159                 spin_lock_irqsave(&conf->device_lock, flags);
1160                 mddev->degraded++;
1161                 set_bit(Faulty, &rdev->flags);
1162                 spin_unlock_irqrestore(&conf->device_lock, flags);
1163                 /*
1164                  * if recovery is running, make sure it aborts.
1165                  */
1166                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1167         } else
1168                 set_bit(Faulty, &rdev->flags);
1169         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1170         printk(KERN_ALERT
1171                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1172                "md/raid1:%s: Operation continuing on %d devices.\n",
1173                mdname(mddev), bdevname(rdev->bdev, b),
1174                mdname(mddev), conf->raid_disks - mddev->degraded);
1175 }
1176
1177 static void print_conf(conf_t *conf)
1178 {
1179         int i;
1180
1181         printk(KERN_DEBUG "RAID1 conf printout:\n");
1182         if (!conf) {
1183                 printk(KERN_DEBUG "(!conf)\n");
1184                 return;
1185         }
1186         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1187                 conf->raid_disks);
1188
1189         rcu_read_lock();
1190         for (i = 0; i < conf->raid_disks; i++) {
1191                 char b[BDEVNAME_SIZE];
1192                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1193                 if (rdev)
1194                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1195                                i, !test_bit(In_sync, &rdev->flags),
1196                                !test_bit(Faulty, &rdev->flags),
1197                                bdevname(rdev->bdev,b));
1198         }
1199         rcu_read_unlock();
1200 }
1201
1202 static void close_sync(conf_t *conf)
1203 {
1204         wait_barrier(conf);
1205         allow_barrier(conf);
1206
1207         mempool_destroy(conf->r1buf_pool);
1208         conf->r1buf_pool = NULL;
1209 }
1210
1211 static int raid1_spare_active(mddev_t *mddev)
1212 {
1213         int i;
1214         conf_t *conf = mddev->private;
1215         int count = 0;
1216         unsigned long flags;
1217
1218         /*
1219          * Find all failed disks within the RAID1 configuration
1220          * and mark them readable.
1221          * Called under mddev lock, so rcu protection not needed.
1222          */
1223         for (i = 0; i < conf->raid_disks; i++) {
1224                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1225                 if (rdev
1226                     && !test_bit(Faulty, &rdev->flags)
1227                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1228                         count++;
1229                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1230                 }
1231         }
1232         spin_lock_irqsave(&conf->device_lock, flags);
1233         mddev->degraded -= count;
1234         spin_unlock_irqrestore(&conf->device_lock, flags);
1235
1236         print_conf(conf);
1237         return count;
1238 }
1239
1240
1241 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1242 {
1243         conf_t *conf = mddev->private;
1244         int err = -EEXIST;
1245         int mirror = 0;
1246         mirror_info_t *p;
1247         int first = 0;
1248         int last = mddev->raid_disks - 1;
1249
1250         if (mddev->recovery_disabled == conf->recovery_disabled)
1251                 return -EBUSY;
1252
1253         if (rdev->raid_disk >= 0)
1254                 first = last = rdev->raid_disk;
1255
1256         for (mirror = first; mirror <= last; mirror++)
1257                 if ( !(p=conf->mirrors+mirror)->rdev) {
1258
1259                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1260                                           rdev->data_offset << 9);
1261                         /* as we don't honour merge_bvec_fn, we must
1262                          * never risk violating it, so limit
1263                          * ->max_segments to one lying with a single
1264                          * page, as a one page request is never in
1265                          * violation.
1266                          */
1267                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1268                                 blk_queue_max_segments(mddev->queue, 1);
1269                                 blk_queue_segment_boundary(mddev->queue,
1270                                                            PAGE_CACHE_SIZE - 1);
1271                         }
1272
1273                         p->head_position = 0;
1274                         rdev->raid_disk = mirror;
1275                         err = 0;
1276                         /* As all devices are equivalent, we don't need a full recovery
1277                          * if this was recently any drive of the array
1278                          */
1279                         if (rdev->saved_raid_disk < 0)
1280                                 conf->fullsync = 1;
1281                         rcu_assign_pointer(p->rdev, rdev);
1282                         break;
1283                 }
1284         md_integrity_add_rdev(rdev, mddev);
1285         print_conf(conf);
1286         return err;
1287 }
1288
1289 static int raid1_remove_disk(mddev_t *mddev, int number)
1290 {
1291         conf_t *conf = mddev->private;
1292         int err = 0;
1293         mdk_rdev_t *rdev;
1294         mirror_info_t *p = conf->mirrors+ number;
1295
1296         print_conf(conf);
1297         rdev = p->rdev;
1298         if (rdev) {
1299                 if (test_bit(In_sync, &rdev->flags) ||
1300                     atomic_read(&rdev->nr_pending)) {
1301                         err = -EBUSY;
1302                         goto abort;
1303                 }
1304                 /* Only remove non-faulty devices if recovery
1305                  * is not possible.
1306                  */
1307                 if (!test_bit(Faulty, &rdev->flags) &&
1308                     mddev->recovery_disabled != conf->recovery_disabled &&
1309                     mddev->degraded < conf->raid_disks) {
1310                         err = -EBUSY;
1311                         goto abort;
1312                 }
1313                 p->rdev = NULL;
1314                 synchronize_rcu();
1315                 if (atomic_read(&rdev->nr_pending)) {
1316                         /* lost the race, try later */
1317                         err = -EBUSY;
1318                         p->rdev = rdev;
1319                         goto abort;
1320                 }
1321                 err = md_integrity_register(mddev);
1322         }
1323 abort:
1324
1325         print_conf(conf);
1326         return err;
1327 }
1328
1329
1330 static void end_sync_read(struct bio *bio, int error)
1331 {
1332         r1bio_t *r1_bio = bio->bi_private;
1333         int i;
1334
1335         for (i=r1_bio->mddev->raid_disks; i--; )
1336                 if (r1_bio->bios[i] == bio)
1337                         break;
1338         BUG_ON(i < 0);
1339         update_head_pos(i, r1_bio);
1340         /*
1341          * we have read a block, now it needs to be re-written,
1342          * or re-read if the read failed.
1343          * We don't do much here, just schedule handling by raid1d
1344          */
1345         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1346                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1347
1348         if (atomic_dec_and_test(&r1_bio->remaining))
1349                 reschedule_retry(r1_bio);
1350 }
1351
1352 static void end_sync_write(struct bio *bio, int error)
1353 {
1354         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1355         r1bio_t *r1_bio = bio->bi_private;
1356         mddev_t *mddev = r1_bio->mddev;
1357         conf_t *conf = mddev->private;
1358         int i;
1359         int mirror=0;
1360         sector_t first_bad;
1361         int bad_sectors;
1362
1363         for (i = 0; i < conf->raid_disks; i++)
1364                 if (r1_bio->bios[i] == bio) {
1365                         mirror = i;
1366                         break;
1367                 }
1368         if (!uptodate) {
1369                 sector_t sync_blocks = 0;
1370                 sector_t s = r1_bio->sector;
1371                 long sectors_to_go = r1_bio->sectors;
1372                 /* make sure these bits doesn't get cleared. */
1373                 do {
1374                         bitmap_end_sync(mddev->bitmap, s,
1375                                         &sync_blocks, 1);
1376                         s += sync_blocks;
1377                         sectors_to_go -= sync_blocks;
1378                 } while (sectors_to_go > 0);
1379                 md_error(mddev, conf->mirrors[mirror].rdev);
1380         } else if (is_badblock(conf->mirrors[mirror].rdev,
1381                                r1_bio->sector,
1382                                r1_bio->sectors,
1383                                &first_bad, &bad_sectors))
1384                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1385
1386         update_head_pos(mirror, r1_bio);
1387
1388         if (atomic_dec_and_test(&r1_bio->remaining)) {
1389                 int s = r1_bio->sectors;
1390                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
1391                         reschedule_retry(r1_bio);
1392                 else {
1393                         put_buf(r1_bio);
1394                         md_done_sync(mddev, s, uptodate);
1395                 }
1396         }
1397 }
1398
1399 static int fix_sync_read_error(r1bio_t *r1_bio)
1400 {
1401         /* Try some synchronous reads of other devices to get
1402          * good data, much like with normal read errors.  Only
1403          * read into the pages we already have so we don't
1404          * need to re-issue the read request.
1405          * We don't need to freeze the array, because being in an
1406          * active sync request, there is no normal IO, and
1407          * no overlapping syncs.
1408          * We don't need to check is_badblock() again as we
1409          * made sure that anything with a bad block in range
1410          * will have bi_end_io clear.
1411          */
1412         mddev_t *mddev = r1_bio->mddev;
1413         conf_t *conf = mddev->private;
1414         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1415         sector_t sect = r1_bio->sector;
1416         int sectors = r1_bio->sectors;
1417         int idx = 0;
1418
1419         while(sectors) {
1420                 int s = sectors;
1421                 int d = r1_bio->read_disk;
1422                 int success = 0;
1423                 mdk_rdev_t *rdev;
1424                 int start;
1425
1426                 if (s > (PAGE_SIZE>>9))
1427                         s = PAGE_SIZE >> 9;
1428                 do {
1429                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1430                                 /* No rcu protection needed here devices
1431                                  * can only be removed when no resync is
1432                                  * active, and resync is currently active
1433                                  */
1434                                 rdev = conf->mirrors[d].rdev;
1435                                 if (sync_page_io(rdev, sect, s<<9,
1436                                                  bio->bi_io_vec[idx].bv_page,
1437                                                  READ, false)) {
1438                                         success = 1;
1439                                         break;
1440                                 }
1441                         }
1442                         d++;
1443                         if (d == conf->raid_disks)
1444                                 d = 0;
1445                 } while (!success && d != r1_bio->read_disk);
1446
1447                 if (!success) {
1448                         char b[BDEVNAME_SIZE];
1449                         /* Cannot read from anywhere, array is toast */
1450                         md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1451                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1452                                " for block %llu\n",
1453                                mdname(mddev),
1454                                bdevname(bio->bi_bdev, b),
1455                                (unsigned long long)r1_bio->sector);
1456                         md_done_sync(mddev, r1_bio->sectors, 0);
1457                         put_buf(r1_bio);
1458                         return 0;
1459                 }
1460
1461                 start = d;
1462                 /* write it back and re-read */
1463                 while (d != r1_bio->read_disk) {
1464                         if (d == 0)
1465                                 d = conf->raid_disks;
1466                         d--;
1467                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1468                                 continue;
1469                         rdev = conf->mirrors[d].rdev;
1470                         if (sync_page_io(rdev, sect, s<<9,
1471                                          bio->bi_io_vec[idx].bv_page,
1472                                          WRITE, false) == 0) {
1473                                 r1_bio->bios[d]->bi_end_io = NULL;
1474                                 rdev_dec_pending(rdev, mddev);
1475                                 md_error(mddev, rdev);
1476                         }
1477                 }
1478                 d = start;
1479                 while (d != r1_bio->read_disk) {
1480                         if (d == 0)
1481                                 d = conf->raid_disks;
1482                         d--;
1483                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1484                                 continue;
1485                         rdev = conf->mirrors[d].rdev;
1486                         if (sync_page_io(rdev, sect, s<<9,
1487                                          bio->bi_io_vec[idx].bv_page,
1488                                          READ, false) == 0)
1489                                 md_error(mddev, rdev);
1490                         else
1491                                 atomic_add(s, &rdev->corrected_errors);
1492                 }
1493                 sectors -= s;
1494                 sect += s;
1495                 idx ++;
1496         }
1497         set_bit(R1BIO_Uptodate, &r1_bio->state);
1498         set_bit(BIO_UPTODATE, &bio->bi_flags);
1499         return 1;
1500 }
1501
1502 static int process_checks(r1bio_t *r1_bio)
1503 {
1504         /* We have read all readable devices.  If we haven't
1505          * got the block, then there is no hope left.
1506          * If we have, then we want to do a comparison
1507          * and skip the write if everything is the same.
1508          * If any blocks failed to read, then we need to
1509          * attempt an over-write
1510          */
1511         mddev_t *mddev = r1_bio->mddev;
1512         conf_t *conf = mddev->private;
1513         int primary;
1514         int i;
1515
1516         for (primary = 0; primary < conf->raid_disks; primary++)
1517                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1518                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1519                         r1_bio->bios[primary]->bi_end_io = NULL;
1520                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1521                         break;
1522                 }
1523         r1_bio->read_disk = primary;
1524         for (i = 0; i < conf->raid_disks; i++) {
1525                 int j;
1526                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1527                 struct bio *pbio = r1_bio->bios[primary];
1528                 struct bio *sbio = r1_bio->bios[i];
1529                 int size;
1530
1531                 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1532                         continue;
1533
1534                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1535                         for (j = vcnt; j-- ; ) {
1536                                 struct page *p, *s;
1537                                 p = pbio->bi_io_vec[j].bv_page;
1538                                 s = sbio->bi_io_vec[j].bv_page;
1539                                 if (memcmp(page_address(p),
1540                                            page_address(s),
1541                                            PAGE_SIZE))
1542                                         break;
1543                         }
1544                 } else
1545                         j = 0;
1546                 if (j >= 0)
1547                         mddev->resync_mismatches += r1_bio->sectors;
1548                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1549                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1550                         /* No need to write to this device. */
1551                         sbio->bi_end_io = NULL;
1552                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1553                         continue;
1554                 }
1555                 /* fixup the bio for reuse */
1556                 sbio->bi_vcnt = vcnt;
1557                 sbio->bi_size = r1_bio->sectors << 9;
1558                 sbio->bi_idx = 0;
1559                 sbio->bi_phys_segments = 0;
1560                 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1561                 sbio->bi_flags |= 1 << BIO_UPTODATE;
1562                 sbio->bi_next = NULL;
1563                 sbio->bi_sector = r1_bio->sector +
1564                         conf->mirrors[i].rdev->data_offset;
1565                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1566                 size = sbio->bi_size;
1567                 for (j = 0; j < vcnt ; j++) {
1568                         struct bio_vec *bi;
1569                         bi = &sbio->bi_io_vec[j];
1570                         bi->bv_offset = 0;
1571                         if (size > PAGE_SIZE)
1572                                 bi->bv_len = PAGE_SIZE;
1573                         else
1574                                 bi->bv_len = size;
1575                         size -= PAGE_SIZE;
1576                         memcpy(page_address(bi->bv_page),
1577                                page_address(pbio->bi_io_vec[j].bv_page),
1578                                PAGE_SIZE);
1579                 }
1580         }
1581         return 0;
1582 }
1583
1584 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1585 {
1586         conf_t *conf = mddev->private;
1587         int i;
1588         int disks = conf->raid_disks;
1589         struct bio *bio, *wbio;
1590
1591         bio = r1_bio->bios[r1_bio->read_disk];
1592
1593         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1594                 /* ouch - failed to read all of that. */
1595                 if (!fix_sync_read_error(r1_bio))
1596                         return;
1597
1598         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1599                 if (process_checks(r1_bio) < 0)
1600                         return;
1601         /*
1602          * schedule writes
1603          */
1604         atomic_set(&r1_bio->remaining, 1);
1605         for (i = 0; i < disks ; i++) {
1606                 wbio = r1_bio->bios[i];
1607                 if (wbio->bi_end_io == NULL ||
1608                     (wbio->bi_end_io == end_sync_read &&
1609                      (i == r1_bio->read_disk ||
1610                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1611                         continue;
1612
1613                 wbio->bi_rw = WRITE;
1614                 wbio->bi_end_io = end_sync_write;
1615                 atomic_inc(&r1_bio->remaining);
1616                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1617
1618                 generic_make_request(wbio);
1619         }
1620
1621         if (atomic_dec_and_test(&r1_bio->remaining)) {
1622                 /* if we're here, all write(s) have completed, so clean up */
1623                 md_done_sync(mddev, r1_bio->sectors, 1);
1624                 put_buf(r1_bio);
1625         }
1626 }
1627
1628 /*
1629  * This is a kernel thread which:
1630  *
1631  *      1.      Retries failed read operations on working mirrors.
1632  *      2.      Updates the raid superblock when problems encounter.
1633  *      3.      Performs writes following reads for array synchronising.
1634  */
1635
1636 static void fix_read_error(conf_t *conf, int read_disk,
1637                            sector_t sect, int sectors)
1638 {
1639         mddev_t *mddev = conf->mddev;
1640         while(sectors) {
1641                 int s = sectors;
1642                 int d = read_disk;
1643                 int success = 0;
1644                 int start;
1645                 mdk_rdev_t *rdev;
1646
1647                 if (s > (PAGE_SIZE>>9))
1648                         s = PAGE_SIZE >> 9;
1649
1650                 do {
1651                         /* Note: no rcu protection needed here
1652                          * as this is synchronous in the raid1d thread
1653                          * which is the thread that might remove
1654                          * a device.  If raid1d ever becomes multi-threaded....
1655                          */
1656                         sector_t first_bad;
1657                         int bad_sectors;
1658
1659                         rdev = conf->mirrors[d].rdev;
1660                         if (rdev &&
1661                             test_bit(In_sync, &rdev->flags) &&
1662                             is_badblock(rdev, sect, s,
1663                                         &first_bad, &bad_sectors) == 0 &&
1664                             sync_page_io(rdev, sect, s<<9,
1665                                          conf->tmppage, READ, false))
1666                                 success = 1;
1667                         else {
1668                                 d++;
1669                                 if (d == conf->raid_disks)
1670                                         d = 0;
1671                         }
1672                 } while (!success && d != read_disk);
1673
1674                 if (!success) {
1675                         /* Cannot read from anywhere -- bye bye array */
1676                         md_error(mddev, conf->mirrors[read_disk].rdev);
1677                         break;
1678                 }
1679                 /* write it back and re-read */
1680                 start = d;
1681                 while (d != read_disk) {
1682                         if (d==0)
1683                                 d = conf->raid_disks;
1684                         d--;
1685                         rdev = conf->mirrors[d].rdev;
1686                         if (rdev &&
1687                             test_bit(In_sync, &rdev->flags)) {
1688                                 if (sync_page_io(rdev, sect, s<<9,
1689                                                  conf->tmppage, WRITE, false)
1690                                     == 0)
1691                                         /* Well, this device is dead */
1692                                         md_error(mddev, rdev);
1693                         }
1694                 }
1695                 d = start;
1696                 while (d != read_disk) {
1697                         char b[BDEVNAME_SIZE];
1698                         if (d==0)
1699                                 d = conf->raid_disks;
1700                         d--;
1701                         rdev = conf->mirrors[d].rdev;
1702                         if (rdev &&
1703                             test_bit(In_sync, &rdev->flags)) {
1704                                 if (sync_page_io(rdev, sect, s<<9,
1705                                                  conf->tmppage, READ, false)
1706                                     == 0)
1707                                         /* Well, this device is dead */
1708                                         md_error(mddev, rdev);
1709                                 else {
1710                                         atomic_add(s, &rdev->corrected_errors);
1711                                         printk(KERN_INFO
1712                                                "md/raid1:%s: read error corrected "
1713                                                "(%d sectors at %llu on %s)\n",
1714                                                mdname(mddev), s,
1715                                                (unsigned long long)(sect +
1716                                                    rdev->data_offset),
1717                                                bdevname(rdev->bdev, b));
1718                                 }
1719                         }
1720                 }
1721                 sectors -= s;
1722                 sect += s;
1723         }
1724 }
1725
1726 static void raid1d(mddev_t *mddev)
1727 {
1728         r1bio_t *r1_bio;
1729         struct bio *bio;
1730         unsigned long flags;
1731         conf_t *conf = mddev->private;
1732         struct list_head *head = &conf->retry_list;
1733         mdk_rdev_t *rdev;
1734         struct blk_plug plug;
1735
1736         md_check_recovery(mddev);
1737
1738         blk_start_plug(&plug);
1739         for (;;) {
1740                 char b[BDEVNAME_SIZE];
1741
1742                 if (atomic_read(&mddev->plug_cnt) == 0)
1743                         flush_pending_writes(conf);
1744
1745                 spin_lock_irqsave(&conf->device_lock, flags);
1746                 if (list_empty(head)) {
1747                         spin_unlock_irqrestore(&conf->device_lock, flags);
1748                         break;
1749                 }
1750                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1751                 list_del(head->prev);
1752                 conf->nr_queued--;
1753                 spin_unlock_irqrestore(&conf->device_lock, flags);
1754
1755                 mddev = r1_bio->mddev;
1756                 conf = mddev->private;
1757                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1758                         if (test_bit(R1BIO_MadeGood, &r1_bio->state)) {
1759                                 int m;
1760                                 int s = r1_bio->sectors;
1761                                 for (m = 0; m < conf->raid_disks ; m++) {
1762                                         struct bio *bio = r1_bio->bios[m];
1763                                         if (bio->bi_end_io != NULL &&
1764                                             test_bit(BIO_UPTODATE,
1765                                                      &bio->bi_flags)) {
1766                                                 rdev = conf->mirrors[m].rdev;
1767                                                 rdev_clear_badblocks(
1768                                                         rdev,
1769                                                         r1_bio->sector,
1770                                                         r1_bio->sectors);
1771                                         }
1772                                 }
1773                                 put_buf(r1_bio);
1774                                 md_done_sync(mddev, s, 1);
1775                         } else
1776                                 sync_request_write(mddev, r1_bio);
1777                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state)) {
1778                         int m;
1779                         for (m = 0; m < conf->raid_disks ; m++)
1780                                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
1781                                         rdev = conf->mirrors[m].rdev;
1782                                         rdev_clear_badblocks(
1783                                                 rdev,
1784                                                 r1_bio->sector,
1785                                                 r1_bio->sectors);
1786                                         rdev_dec_pending(rdev, mddev);
1787                                 }
1788                         raid_end_bio_io(r1_bio);
1789                 } else if (test_bit(R1BIO_ReadError, &r1_bio->state)) {
1790                         int disk;
1791                         int max_sectors;
1792
1793                         clear_bit(R1BIO_ReadError, &r1_bio->state);
1794                         /* we got a read error. Maybe the drive is bad.  Maybe just
1795                          * the block and we can fix it.
1796                          * We freeze all other IO, and try reading the block from
1797                          * other devices.  When we find one, we re-write
1798                          * and check it that fixes the read error.
1799                          * This is all done synchronously while the array is
1800                          * frozen
1801                          */
1802                         if (mddev->ro == 0) {
1803                                 freeze_array(conf);
1804                                 fix_read_error(conf, r1_bio->read_disk,
1805                                                r1_bio->sector,
1806                                                r1_bio->sectors);
1807                                 unfreeze_array(conf);
1808                         } else
1809                                 md_error(mddev,
1810                                          conf->mirrors[r1_bio->read_disk].rdev);
1811
1812                         bio = r1_bio->bios[r1_bio->read_disk];
1813                         bdevname(bio->bi_bdev, b);
1814 read_more:
1815                         disk = read_balance(conf, r1_bio, &max_sectors);
1816                         if (disk == -1) {
1817                                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1818                                        " read error for block %llu\n",
1819                                        mdname(mddev), b,
1820                                        (unsigned long long)r1_bio->sector);
1821                                 raid_end_bio_io(r1_bio);
1822                         } else {
1823                                 const unsigned long do_sync = r1_bio->master_bio->bi_rw & REQ_SYNC;
1824                                 if (bio) {
1825                                         r1_bio->bios[r1_bio->read_disk] =
1826                                                 mddev->ro ? IO_BLOCKED : NULL;
1827                                         bio_put(bio);
1828                                 }
1829                                 r1_bio->read_disk = disk;
1830                                 bio = bio_clone_mddev(r1_bio->master_bio,
1831                                                       GFP_NOIO, mddev);
1832                                 md_trim_bio(bio,
1833                                             r1_bio->sector - bio->bi_sector,
1834                                             max_sectors);
1835                                 r1_bio->bios[r1_bio->read_disk] = bio;
1836                                 rdev = conf->mirrors[disk].rdev;
1837                                 printk_ratelimited(
1838                                         KERN_ERR
1839                                         "md/raid1:%s: redirecting sector %llu"
1840                                         " to other mirror: %s\n",
1841                                         mdname(mddev),
1842                                         (unsigned long long)r1_bio->sector,
1843                                         bdevname(rdev->bdev, b));
1844                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1845                                 bio->bi_bdev = rdev->bdev;
1846                                 bio->bi_end_io = raid1_end_read_request;
1847                                 bio->bi_rw = READ | do_sync;
1848                                 bio->bi_private = r1_bio;
1849                                 if (max_sectors < r1_bio->sectors) {
1850                                         /* Drat - have to split this up more */
1851                                         struct bio *mbio = r1_bio->master_bio;
1852                                         int sectors_handled =
1853                                                 r1_bio->sector + max_sectors
1854                                                 - mbio->bi_sector;
1855                                         r1_bio->sectors = max_sectors;
1856                                         spin_lock_irq(&conf->device_lock);
1857                                         if (mbio->bi_phys_segments == 0)
1858                                                 mbio->bi_phys_segments = 2;
1859                                         else
1860                                                 mbio->bi_phys_segments++;
1861                                         spin_unlock_irq(&conf->device_lock);
1862                                         generic_make_request(bio);
1863                                         bio = NULL;
1864
1865                                         r1_bio = mempool_alloc(conf->r1bio_pool,
1866                                                                GFP_NOIO);
1867
1868                                         r1_bio->master_bio = mbio;
1869                                         r1_bio->sectors = (mbio->bi_size >> 9)
1870                                                 - sectors_handled;
1871                                         r1_bio->state = 0;
1872                                         set_bit(R1BIO_ReadError,
1873                                                 &r1_bio->state);
1874                                         r1_bio->mddev = mddev;
1875                                         r1_bio->sector = mbio->bi_sector
1876                                                 + sectors_handled;
1877
1878                                         goto read_more;
1879                                 } else
1880                                         generic_make_request(bio);
1881                         }
1882                 } else {
1883                         /* just a partial read to be scheduled from separate
1884                          * context
1885                          */
1886                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
1887                 }
1888                 cond_resched();
1889                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
1890                         md_check_recovery(mddev);
1891         }
1892         blk_finish_plug(&plug);
1893 }
1894
1895
1896 static int init_resync(conf_t *conf)
1897 {
1898         int buffs;
1899
1900         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1901         BUG_ON(conf->r1buf_pool);
1902         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1903                                           conf->poolinfo);
1904         if (!conf->r1buf_pool)
1905                 return -ENOMEM;
1906         conf->next_resync = 0;
1907         return 0;
1908 }
1909
1910 /*
1911  * perform a "sync" on one "block"
1912  *
1913  * We need to make sure that no normal I/O request - particularly write
1914  * requests - conflict with active sync requests.
1915  *
1916  * This is achieved by tracking pending requests and a 'barrier' concept
1917  * that can be installed to exclude normal IO requests.
1918  */
1919
1920 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1921 {
1922         conf_t *conf = mddev->private;
1923         r1bio_t *r1_bio;
1924         struct bio *bio;
1925         sector_t max_sector, nr_sectors;
1926         int disk = -1;
1927         int i;
1928         int wonly = -1;
1929         int write_targets = 0, read_targets = 0;
1930         sector_t sync_blocks;
1931         int still_degraded = 0;
1932         int good_sectors = RESYNC_SECTORS;
1933         int min_bad = 0; /* number of sectors that are bad in all devices */
1934
1935         if (!conf->r1buf_pool)
1936                 if (init_resync(conf))
1937                         return 0;
1938
1939         max_sector = mddev->dev_sectors;
1940         if (sector_nr >= max_sector) {
1941                 /* If we aborted, we need to abort the
1942                  * sync on the 'current' bitmap chunk (there will
1943                  * only be one in raid1 resync.
1944                  * We can find the current addess in mddev->curr_resync
1945                  */
1946                 if (mddev->curr_resync < max_sector) /* aborted */
1947                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1948                                                 &sync_blocks, 1);
1949                 else /* completed sync */
1950                         conf->fullsync = 0;
1951
1952                 bitmap_close_sync(mddev->bitmap);
1953                 close_sync(conf);
1954                 return 0;
1955         }
1956
1957         if (mddev->bitmap == NULL &&
1958             mddev->recovery_cp == MaxSector &&
1959             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1960             conf->fullsync == 0) {
1961                 *skipped = 1;
1962                 return max_sector - sector_nr;
1963         }
1964         /* before building a request, check if we can skip these blocks..
1965          * This call the bitmap_start_sync doesn't actually record anything
1966          */
1967         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1968             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1969                 /* We can skip this block, and probably several more */
1970                 *skipped = 1;
1971                 return sync_blocks;
1972         }
1973         /*
1974          * If there is non-resync activity waiting for a turn,
1975          * and resync is going fast enough,
1976          * then let it though before starting on this new sync request.
1977          */
1978         if (!go_faster && conf->nr_waiting)
1979                 msleep_interruptible(1000);
1980
1981         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1982         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1983         raise_barrier(conf);
1984
1985         conf->next_resync = sector_nr;
1986
1987         rcu_read_lock();
1988         /*
1989          * If we get a correctably read error during resync or recovery,
1990          * we might want to read from a different device.  So we
1991          * flag all drives that could conceivably be read from for READ,
1992          * and any others (which will be non-In_sync devices) for WRITE.
1993          * If a read fails, we try reading from something else for which READ
1994          * is OK.
1995          */
1996
1997         r1_bio->mddev = mddev;
1998         r1_bio->sector = sector_nr;
1999         r1_bio->state = 0;
2000         set_bit(R1BIO_IsSync, &r1_bio->state);
2001
2002         for (i=0; i < conf->raid_disks; i++) {
2003                 mdk_rdev_t *rdev;
2004                 bio = r1_bio->bios[i];
2005
2006                 /* take from bio_init */
2007                 bio->bi_next = NULL;
2008                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
2009                 bio->bi_flags |= 1 << BIO_UPTODATE;
2010                 bio->bi_comp_cpu = -1;
2011                 bio->bi_rw = READ;
2012                 bio->bi_vcnt = 0;
2013                 bio->bi_idx = 0;
2014                 bio->bi_phys_segments = 0;
2015                 bio->bi_size = 0;
2016                 bio->bi_end_io = NULL;
2017                 bio->bi_private = NULL;
2018
2019                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2020                 if (rdev == NULL ||
2021                     test_bit(Faulty, &rdev->flags)) {
2022                         still_degraded = 1;
2023                 } else if (!test_bit(In_sync, &rdev->flags)) {
2024                         bio->bi_rw = WRITE;
2025                         bio->bi_end_io = end_sync_write;
2026                         write_targets ++;
2027                 } else {
2028                         /* may need to read from here */
2029                         sector_t first_bad = MaxSector;
2030                         int bad_sectors;
2031
2032                         if (is_badblock(rdev, sector_nr, good_sectors,
2033                                         &first_bad, &bad_sectors)) {
2034                                 if (first_bad > sector_nr)
2035                                         good_sectors = first_bad - sector_nr;
2036                                 else {
2037                                         bad_sectors -= (sector_nr - first_bad);
2038                                         if (min_bad == 0 ||
2039                                             min_bad > bad_sectors)
2040                                                 min_bad = bad_sectors;
2041                                 }
2042                         }
2043                         if (sector_nr < first_bad) {
2044                                 if (test_bit(WriteMostly, &rdev->flags)) {
2045                                         if (wonly < 0)
2046                                                 wonly = i;
2047                                 } else {
2048                                         if (disk < 0)
2049                                                 disk = i;
2050                                 }
2051                                 bio->bi_rw = READ;
2052                                 bio->bi_end_io = end_sync_read;
2053                                 read_targets++;
2054                         }
2055                 }
2056                 if (bio->bi_end_io) {
2057                         atomic_inc(&rdev->nr_pending);
2058                         bio->bi_sector = sector_nr + rdev->data_offset;
2059                         bio->bi_bdev = rdev->bdev;
2060                         bio->bi_private = r1_bio;
2061                 }
2062         }
2063         rcu_read_unlock();
2064         if (disk < 0)
2065                 disk = wonly;
2066         r1_bio->read_disk = disk;
2067
2068         if (read_targets == 0 && min_bad > 0) {
2069                 /* These sectors are bad on all InSync devices, so we
2070                  * need to mark them bad on all write targets
2071                  */
2072                 int ok = 1;
2073                 for (i = 0 ; i < conf->raid_disks ; i++)
2074                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2075                                 mdk_rdev_t *rdev =
2076                                         rcu_dereference(conf->mirrors[i].rdev);
2077                                 ok = rdev_set_badblocks(rdev, sector_nr,
2078                                                         min_bad, 0
2079                                         ) && ok;
2080                         }
2081                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2082                 *skipped = 1;
2083                 put_buf(r1_bio);
2084
2085                 if (!ok) {
2086                         /* Cannot record the badblocks, so need to
2087                          * abort the resync.
2088                          * If there are multiple read targets, could just
2089                          * fail the really bad ones ???
2090                          */
2091                         conf->recovery_disabled = mddev->recovery_disabled;
2092                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2093                         return 0;
2094                 } else
2095                         return min_bad;
2096
2097         }
2098         if (min_bad > 0 && min_bad < good_sectors) {
2099                 /* only resync enough to reach the next bad->good
2100                  * transition */
2101                 good_sectors = min_bad;
2102         }
2103
2104         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2105                 /* extra read targets are also write targets */
2106                 write_targets += read_targets-1;
2107
2108         if (write_targets == 0 || read_targets == 0) {
2109                 /* There is nowhere to write, so all non-sync
2110                  * drives must be failed - so we are finished
2111                  */
2112                 sector_t rv = max_sector - sector_nr;
2113                 *skipped = 1;
2114                 put_buf(r1_bio);
2115                 return rv;
2116         }
2117
2118         if (max_sector > mddev->resync_max)
2119                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2120         if (max_sector > sector_nr + good_sectors)
2121                 max_sector = sector_nr + good_sectors;
2122         nr_sectors = 0;
2123         sync_blocks = 0;
2124         do {
2125                 struct page *page;
2126                 int len = PAGE_SIZE;
2127                 if (sector_nr + (len>>9) > max_sector)
2128                         len = (max_sector - sector_nr) << 9;
2129                 if (len == 0)
2130                         break;
2131                 if (sync_blocks == 0) {
2132                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2133                                                &sync_blocks, still_degraded) &&
2134                             !conf->fullsync &&
2135                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2136                                 break;
2137                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2138                         if ((len >> 9) > sync_blocks)
2139                                 len = sync_blocks<<9;
2140                 }
2141
2142                 for (i=0 ; i < conf->raid_disks; i++) {
2143                         bio = r1_bio->bios[i];
2144                         if (bio->bi_end_io) {
2145                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2146                                 if (bio_add_page(bio, page, len, 0) == 0) {
2147                                         /* stop here */
2148                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2149                                         while (i > 0) {
2150                                                 i--;
2151                                                 bio = r1_bio->bios[i];
2152                                                 if (bio->bi_end_io==NULL)
2153                                                         continue;
2154                                                 /* remove last page from this bio */
2155                                                 bio->bi_vcnt--;
2156                                                 bio->bi_size -= len;
2157                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2158                                         }
2159                                         goto bio_full;
2160                                 }
2161                         }
2162                 }
2163                 nr_sectors += len>>9;
2164                 sector_nr += len>>9;
2165                 sync_blocks -= (len>>9);
2166         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2167  bio_full:
2168         r1_bio->sectors = nr_sectors;
2169
2170         /* For a user-requested sync, we read all readable devices and do a
2171          * compare
2172          */
2173         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2174                 atomic_set(&r1_bio->remaining, read_targets);
2175                 for (i=0; i<conf->raid_disks; i++) {
2176                         bio = r1_bio->bios[i];
2177                         if (bio->bi_end_io == end_sync_read) {
2178                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2179                                 generic_make_request(bio);
2180                         }
2181                 }
2182         } else {
2183                 atomic_set(&r1_bio->remaining, 1);
2184                 bio = r1_bio->bios[r1_bio->read_disk];
2185                 md_sync_acct(bio->bi_bdev, nr_sectors);
2186                 generic_make_request(bio);
2187
2188         }
2189         return nr_sectors;
2190 }
2191
2192 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2193 {
2194         if (sectors)
2195                 return sectors;
2196
2197         return mddev->dev_sectors;
2198 }
2199
2200 static conf_t *setup_conf(mddev_t *mddev)
2201 {
2202         conf_t *conf;
2203         int i;
2204         mirror_info_t *disk;
2205         mdk_rdev_t *rdev;
2206         int err = -ENOMEM;
2207
2208         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2209         if (!conf)
2210                 goto abort;
2211
2212         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2213                                  GFP_KERNEL);
2214         if (!conf->mirrors)
2215                 goto abort;
2216
2217         conf->tmppage = alloc_page(GFP_KERNEL);
2218         if (!conf->tmppage)
2219                 goto abort;
2220
2221         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2222         if (!conf->poolinfo)
2223                 goto abort;
2224         conf->poolinfo->raid_disks = mddev->raid_disks;
2225         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2226                                           r1bio_pool_free,
2227                                           conf->poolinfo);
2228         if (!conf->r1bio_pool)
2229                 goto abort;
2230
2231         conf->poolinfo->mddev = mddev;
2232
2233         spin_lock_init(&conf->device_lock);
2234         list_for_each_entry(rdev, &mddev->disks, same_set) {
2235                 int disk_idx = rdev->raid_disk;
2236                 if (disk_idx >= mddev->raid_disks
2237                     || disk_idx < 0)
2238                         continue;
2239                 disk = conf->mirrors + disk_idx;
2240
2241                 disk->rdev = rdev;
2242
2243                 disk->head_position = 0;
2244         }
2245         conf->raid_disks = mddev->raid_disks;
2246         conf->mddev = mddev;
2247         INIT_LIST_HEAD(&conf->retry_list);
2248
2249         spin_lock_init(&conf->resync_lock);
2250         init_waitqueue_head(&conf->wait_barrier);
2251
2252         bio_list_init(&conf->pending_bio_list);
2253
2254         conf->last_used = -1;
2255         for (i = 0; i < conf->raid_disks; i++) {
2256
2257                 disk = conf->mirrors + i;
2258
2259                 if (!disk->rdev ||
2260                     !test_bit(In_sync, &disk->rdev->flags)) {
2261                         disk->head_position = 0;
2262                         if (disk->rdev)
2263                                 conf->fullsync = 1;
2264                 } else if (conf->last_used < 0)
2265                         /*
2266                          * The first working device is used as a
2267                          * starting point to read balancing.
2268                          */
2269                         conf->last_used = i;
2270         }
2271
2272         err = -EIO;
2273         if (conf->last_used < 0) {
2274                 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
2275                        mdname(mddev));
2276                 goto abort;
2277         }
2278         err = -ENOMEM;
2279         conf->thread = md_register_thread(raid1d, mddev, NULL);
2280         if (!conf->thread) {
2281                 printk(KERN_ERR
2282                        "md/raid1:%s: couldn't allocate thread\n",
2283                        mdname(mddev));
2284                 goto abort;
2285         }
2286
2287         return conf;
2288
2289  abort:
2290         if (conf) {
2291                 if (conf->r1bio_pool)
2292                         mempool_destroy(conf->r1bio_pool);
2293                 kfree(conf->mirrors);
2294                 safe_put_page(conf->tmppage);
2295                 kfree(conf->poolinfo);
2296                 kfree(conf);
2297         }
2298         return ERR_PTR(err);
2299 }
2300
2301 static int run(mddev_t *mddev)
2302 {
2303         conf_t *conf;
2304         int i;
2305         mdk_rdev_t *rdev;
2306
2307         if (mddev->level != 1) {
2308                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2309                        mdname(mddev), mddev->level);
2310                 return -EIO;
2311         }
2312         if (mddev->reshape_position != MaxSector) {
2313                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2314                        mdname(mddev));
2315                 return -EIO;
2316         }
2317         /*
2318          * copy the already verified devices into our private RAID1
2319          * bookkeeping area. [whatever we allocate in run(),
2320          * should be freed in stop()]
2321          */
2322         if (mddev->private == NULL)
2323                 conf = setup_conf(mddev);
2324         else
2325                 conf = mddev->private;
2326
2327         if (IS_ERR(conf))
2328                 return PTR_ERR(conf);
2329
2330         list_for_each_entry(rdev, &mddev->disks, same_set) {
2331                 if (!mddev->gendisk)
2332                         continue;
2333                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2334                                   rdev->data_offset << 9);
2335                 /* as we don't honour merge_bvec_fn, we must never risk
2336                  * violating it, so limit ->max_segments to 1 lying within
2337                  * a single page, as a one page request is never in violation.
2338                  */
2339                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2340                         blk_queue_max_segments(mddev->queue, 1);
2341                         blk_queue_segment_boundary(mddev->queue,
2342                                                    PAGE_CACHE_SIZE - 1);
2343                 }
2344         }
2345
2346         mddev->degraded = 0;
2347         for (i=0; i < conf->raid_disks; i++)
2348                 if (conf->mirrors[i].rdev == NULL ||
2349                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2350                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2351                         mddev->degraded++;
2352
2353         if (conf->raid_disks - mddev->degraded == 1)
2354                 mddev->recovery_cp = MaxSector;
2355
2356         if (mddev->recovery_cp != MaxSector)
2357                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2358                        " -- starting background reconstruction\n",
2359                        mdname(mddev));
2360         printk(KERN_INFO
2361                 "md/raid1:%s: active with %d out of %d mirrors\n",
2362                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2363                 mddev->raid_disks);
2364
2365         /*
2366          * Ok, everything is just fine now
2367          */
2368         mddev->thread = conf->thread;
2369         conf->thread = NULL;
2370         mddev->private = conf;
2371
2372         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2373
2374         if (mddev->queue) {
2375                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2376                 mddev->queue->backing_dev_info.congested_data = mddev;
2377         }
2378         return md_integrity_register(mddev);
2379 }
2380
2381 static int stop(mddev_t *mddev)
2382 {
2383         conf_t *conf = mddev->private;
2384         struct bitmap *bitmap = mddev->bitmap;
2385
2386         /* wait for behind writes to complete */
2387         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2388                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2389                        mdname(mddev));
2390                 /* need to kick something here to make sure I/O goes? */
2391                 wait_event(bitmap->behind_wait,
2392                            atomic_read(&bitmap->behind_writes) == 0);
2393         }
2394
2395         raise_barrier(conf);
2396         lower_barrier(conf);
2397
2398         md_unregister_thread(mddev->thread);
2399         mddev->thread = NULL;
2400         if (conf->r1bio_pool)
2401                 mempool_destroy(conf->r1bio_pool);
2402         kfree(conf->mirrors);
2403         kfree(conf->poolinfo);
2404         kfree(conf);
2405         mddev->private = NULL;
2406         return 0;
2407 }
2408
2409 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2410 {
2411         /* no resync is happening, and there is enough space
2412          * on all devices, so we can resize.
2413          * We need to make sure resync covers any new space.
2414          * If the array is shrinking we should possibly wait until
2415          * any io in the removed space completes, but it hardly seems
2416          * worth it.
2417          */
2418         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2419         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2420                 return -EINVAL;
2421         set_capacity(mddev->gendisk, mddev->array_sectors);
2422         revalidate_disk(mddev->gendisk);
2423         if (sectors > mddev->dev_sectors &&
2424             mddev->recovery_cp > mddev->dev_sectors) {
2425                 mddev->recovery_cp = mddev->dev_sectors;
2426                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2427         }
2428         mddev->dev_sectors = sectors;
2429         mddev->resync_max_sectors = sectors;
2430         return 0;
2431 }
2432
2433 static int raid1_reshape(mddev_t *mddev)
2434 {
2435         /* We need to:
2436          * 1/ resize the r1bio_pool
2437          * 2/ resize conf->mirrors
2438          *
2439          * We allocate a new r1bio_pool if we can.
2440          * Then raise a device barrier and wait until all IO stops.
2441          * Then resize conf->mirrors and swap in the new r1bio pool.
2442          *
2443          * At the same time, we "pack" the devices so that all the missing
2444          * devices have the higher raid_disk numbers.
2445          */
2446         mempool_t *newpool, *oldpool;
2447         struct pool_info *newpoolinfo;
2448         mirror_info_t *newmirrors;
2449         conf_t *conf = mddev->private;
2450         int cnt, raid_disks;
2451         unsigned long flags;
2452         int d, d2, err;
2453
2454         /* Cannot change chunk_size, layout, or level */
2455         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2456             mddev->layout != mddev->new_layout ||
2457             mddev->level != mddev->new_level) {
2458                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2459                 mddev->new_layout = mddev->layout;
2460                 mddev->new_level = mddev->level;
2461                 return -EINVAL;
2462         }
2463
2464         err = md_allow_write(mddev);
2465         if (err)
2466                 return err;
2467
2468         raid_disks = mddev->raid_disks + mddev->delta_disks;
2469
2470         if (raid_disks < conf->raid_disks) {
2471                 cnt=0;
2472                 for (d= 0; d < conf->raid_disks; d++)
2473                         if (conf->mirrors[d].rdev)
2474                                 cnt++;
2475                 if (cnt > raid_disks)
2476                         return -EBUSY;
2477         }
2478
2479         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2480         if (!newpoolinfo)
2481                 return -ENOMEM;
2482         newpoolinfo->mddev = mddev;
2483         newpoolinfo->raid_disks = raid_disks;
2484
2485         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2486                                  r1bio_pool_free, newpoolinfo);
2487         if (!newpool) {
2488                 kfree(newpoolinfo);
2489                 return -ENOMEM;
2490         }
2491         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2492         if (!newmirrors) {
2493                 kfree(newpoolinfo);
2494                 mempool_destroy(newpool);
2495                 return -ENOMEM;
2496         }
2497
2498         raise_barrier(conf);
2499
2500         /* ok, everything is stopped */
2501         oldpool = conf->r1bio_pool;
2502         conf->r1bio_pool = newpool;
2503
2504         for (d = d2 = 0; d < conf->raid_disks; d++) {
2505                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2506                 if (rdev && rdev->raid_disk != d2) {
2507                         sysfs_unlink_rdev(mddev, rdev);
2508                         rdev->raid_disk = d2;
2509                         sysfs_unlink_rdev(mddev, rdev);
2510                         if (sysfs_link_rdev(mddev, rdev))
2511                                 printk(KERN_WARNING
2512                                        "md/raid1:%s: cannot register rd%d\n",
2513                                        mdname(mddev), rdev->raid_disk);
2514                 }
2515                 if (rdev)
2516                         newmirrors[d2++].rdev = rdev;
2517         }
2518         kfree(conf->mirrors);
2519         conf->mirrors = newmirrors;
2520         kfree(conf->poolinfo);
2521         conf->poolinfo = newpoolinfo;
2522
2523         spin_lock_irqsave(&conf->device_lock, flags);
2524         mddev->degraded += (raid_disks - conf->raid_disks);
2525         spin_unlock_irqrestore(&conf->device_lock, flags);
2526         conf->raid_disks = mddev->raid_disks = raid_disks;
2527         mddev->delta_disks = 0;
2528
2529         conf->last_used = 0; /* just make sure it is in-range */
2530         lower_barrier(conf);
2531
2532         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2533         md_wakeup_thread(mddev->thread);
2534
2535         mempool_destroy(oldpool);
2536         return 0;
2537 }
2538
2539 static void raid1_quiesce(mddev_t *mddev, int state)
2540 {
2541         conf_t *conf = mddev->private;
2542
2543         switch(state) {
2544         case 2: /* wake for suspend */
2545                 wake_up(&conf->wait_barrier);
2546                 break;
2547         case 1:
2548                 raise_barrier(conf);
2549                 break;
2550         case 0:
2551                 lower_barrier(conf);
2552                 break;
2553         }
2554 }
2555
2556 static void *raid1_takeover(mddev_t *mddev)
2557 {
2558         /* raid1 can take over:
2559          *  raid5 with 2 devices, any layout or chunk size
2560          */
2561         if (mddev->level == 5 && mddev->raid_disks == 2) {
2562                 conf_t *conf;
2563                 mddev->new_level = 1;
2564                 mddev->new_layout = 0;
2565                 mddev->new_chunk_sectors = 0;
2566                 conf = setup_conf(mddev);
2567                 if (!IS_ERR(conf))
2568                         conf->barrier = 1;
2569                 return conf;
2570         }
2571         return ERR_PTR(-EINVAL);
2572 }
2573
2574 static struct mdk_personality raid1_personality =
2575 {
2576         .name           = "raid1",
2577         .level          = 1,
2578         .owner          = THIS_MODULE,
2579         .make_request   = make_request,
2580         .run            = run,
2581         .stop           = stop,
2582         .status         = status,
2583         .error_handler  = error,
2584         .hot_add_disk   = raid1_add_disk,
2585         .hot_remove_disk= raid1_remove_disk,
2586         .spare_active   = raid1_spare_active,
2587         .sync_request   = sync_request,
2588         .resize         = raid1_resize,
2589         .size           = raid1_size,
2590         .check_reshape  = raid1_reshape,
2591         .quiesce        = raid1_quiesce,
2592         .takeover       = raid1_takeover,
2593 };
2594
2595 static int __init raid_init(void)
2596 {
2597         return register_md_personality(&raid1_personality);
2598 }
2599
2600 static void raid_exit(void)
2601 {
2602         unregister_md_personality(&raid1_personality);
2603 }
2604
2605 module_init(raid_init);
2606 module_exit(raid_exit);
2607 MODULE_LICENSE("GPL");
2608 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2609 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2610 MODULE_ALIAS("md-raid1");
2611 MODULE_ALIAS("md-level-1");