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 "dm-bio-list.h"
  35 #include <linux/raid/raid1.h>
  36 #include <linux/raid/bitmap.h>
  37
  38 #define DEBUG 0
  39 #if DEBUG
  40 #define PRINTK(x...) printk(x)
  41 #else
  42 #define PRINTK(x...)
  43 #endif
  44
  45 /*
  46  * Number of guaranteed r1bios in case of extreme VM load:
  47  */
  48 #define NR_RAID1_BIOS 256
  49
  50 static mdk_personality_t raid1_personality;
  51
  52 static void unplug_slaves(mddev_t *mddev);
  53
  54
  55 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  56 {
  57         struct pool_info *pi = data;
  58         r1bio_t *r1_bio;
  59         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
  60
  61         /* allocate a r1bio with room for raid_disks entries in the bios array */
  62         r1_bio = kmalloc(size, gfp_flags);
  63         if (r1_bio)
  64                 memset(r1_bio, 0, size);
  65         else
  66                 unplug_slaves(pi->mddev);
  67
  68         return r1_bio;
  69 }
  70
  71 static void r1bio_pool_free(void *r1_bio, void *data)
  72 {
  73         kfree(r1_bio);
  74 }
  75
  76 #define RESYNC_BLOCK_SIZE (64*1024)
  77 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  78 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  79 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  80 #define RESYNC_WINDOW (2048*1024)
  81
  82 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  83 {
  84         struct pool_info *pi = data;
  85         struct page *page;
  86         r1bio_t *r1_bio;
  87         struct bio *bio;
  88         int i, j;
  89
  90         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  91         if (!r1_bio) {
  92                 unplug_slaves(pi->mddev);
  93                 return NULL;
  94         }
  95
  96         /*
  97          * Allocate bios : 1 for reading, n-1 for writing
  98          */
  99         for (j = pi->raid_disks ; j-- ; ) {
 100                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 101                 if (!bio)
 102                         goto out_free_bio;
 103                 r1_bio->bios[j] = bio;
 104         }
 105         /*
 106          * Allocate RESYNC_PAGES data pages and attach them to
 107          * the first bio;
 108          */
 109         bio = r1_bio->bios[0];
 110         for (i = 0; i < RESYNC_PAGES; i++) {
 111                 page = alloc_page(gfp_flags);
 112                 if (unlikely(!page))
 113                         goto out_free_pages;
 114
 115                 bio->bi_io_vec[i].bv_page = page;
 116         }
 117
 118         r1_bio->master_bio = NULL;
 119
 120         return r1_bio;
 121
 122 out_free_pages:
 123         for ( ; i > 0 ; i--)
 124                 __free_page(bio->bi_io_vec[i-1].bv_page);
 125 out_free_bio:
 126         while ( ++j < pi->raid_disks )
 127                 bio_put(r1_bio->bios[j]);
 128         r1bio_pool_free(r1_bio, data);
 129         return NULL;
 130 }
 131
 132 static void r1buf_pool_free(void *__r1_bio, void *data)
 133 {
 134         struct pool_info *pi = data;
 135         int i;
 136         r1bio_t *r1bio = __r1_bio;
 137         struct bio *bio = r1bio->bios[0];
 138
 139         for (i = 0; i < RESYNC_PAGES; i++) {
 140                 __free_page(bio->bi_io_vec[i].bv_page);
 141                 bio->bi_io_vec[i].bv_page = NULL;
 142         }
 143         for (i=0 ; i < pi->raid_disks; i++)
 144                 bio_put(r1bio->bios[i]);
 145
 146         r1bio_pool_free(r1bio, data);
 147 }
 148
 149 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 150 {
 151         int i;
 152
 153         for (i = 0; i < conf->raid_disks; i++) {
 154                 struct bio **bio = r1_bio->bios + i;
 155                 if (*bio)
 156                         bio_put(*bio);
 157                 *bio = NULL;
 158         }
 159 }
 160
 161 static inline void free_r1bio(r1bio_t *r1_bio)
 162 {
 163         unsigned long flags;
 164
 165         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 166
 167         /*
 168          * Wake up any possible resync thread that waits for the device
 169          * to go idle.
 170          */
 171         spin_lock_irqsave(&conf->resync_lock, flags);
 172         if (!--conf->nr_pending) {
 173                 wake_up(&conf->wait_idle);
 174                 wake_up(&conf->wait_resume);
 175         }
 176         spin_unlock_irqrestore(&conf->resync_lock, flags);
 177
 178         put_all_bios(conf, r1_bio);
 179         mempool_free(r1_bio, conf->r1bio_pool);
 180 }
 181
 182 static inline void put_buf(r1bio_t *r1_bio)
 183 {
 184         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 185         unsigned long flags;
 186
 187         mempool_free(r1_bio, conf->r1buf_pool);
 188
 189         spin_lock_irqsave(&conf->resync_lock, flags);
 190         if (!conf->barrier)
 191                 BUG();
 192         --conf->barrier;
 193         wake_up(&conf->wait_resume);
 194         wake_up(&conf->wait_idle);
 195
 196         if (!--conf->nr_pending) {
 197                 wake_up(&conf->wait_idle);
 198                 wake_up(&conf->wait_resume);
 199         }
 200         spin_unlock_irqrestore(&conf->resync_lock, flags);
 201 }
 202
 203 static void reschedule_retry(r1bio_t *r1_bio)
 204 {
 205         unsigned long flags;
 206         mddev_t *mddev = r1_bio->mddev;
 207         conf_t *conf = mddev_to_conf(mddev);
 208
 209         spin_lock_irqsave(&conf->device_lock, flags);
 210         list_add(&r1_bio->retry_list, &conf->retry_list);
 211         spin_unlock_irqrestore(&conf->device_lock, flags);
 212
 213         md_wakeup_thread(mddev->thread);
 214 }
 215
 216 /*
 217  * raid_end_bio_io() is called when we have finished servicing a mirrored
 218  * operation and are ready to return a success/failure code to the buffer
 219  * cache layer.
 220  */
 221 static void raid_end_bio_io(r1bio_t *r1_bio)
 222 {
 223         struct bio *bio = r1_bio->master_bio;
 224
 225         /* if nobody has done the final endio yet, do it now */
 226         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 227                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
 228                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 229                         (unsigned long long) bio->bi_sector,
 230                         (unsigned long long) bio->bi_sector +
 231                                 (bio->bi_size >> 9) - 1);
 232
 233                 bio_endio(bio, bio->bi_size,
 234                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 235         }
 236         free_r1bio(r1_bio);
 237 }
 238
 239 /*
 240  * Update disk head position estimator based on IRQ completion info.
 241  */
 242 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 243 {
 244         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 245
 246         conf->mirrors[disk].head_position =
 247                 r1_bio->sector + (r1_bio->sectors);
 248 }
 249
 250 static int raid1_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
 251 {
 252         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 253         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 254         int mirror;
 255         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 256
 257         if (bio->bi_size)
 258                 return 1;
 259
 260         mirror = r1_bio->read_disk;
 261         /*
 262          * this branch is our 'one mirror IO has finished' event handler:
 263          */
 264         if (!uptodate)
 265                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 266         else
 267                 /*
 268                  * Set R1BIO_Uptodate in our master bio, so that
 269                  * we will return a good error code for to the higher
 270                  * levels even if IO on some other mirrored buffer fails.
 271                  *
 272                  * The 'master' represents the composite IO operation to
 273                  * user-side. So if something waits for IO, then it will
 274                  * wait for the 'master' bio.
 275                  */
 276                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 277
 278         update_head_pos(mirror, r1_bio);
 279
 280         /*
 281          * we have only one bio on the read side
 282          */
 283         if (uptodate)
 284                 raid_end_bio_io(r1_bio);
 285         else {
 286                 /*
 287                  * oops, read error:
 288                  */
 289                 char b[BDEVNAME_SIZE];
 290                 if (printk_ratelimit())
 291                         printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
 292                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 293                 reschedule_retry(r1_bio);
 294         }
 295
 296         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 297         return 0;
 298 }
 299
 300 static int raid1_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
 301 {
 302         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 303         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 304         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 305         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 306
 307         if (bio->bi_size)
 308                 return 1;
 309
 310         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 311                 if (r1_bio->bios[mirror] == bio)
 312                         break;
 313
 314         if (error == -ENOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
 315                 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
 316                 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
 317                 r1_bio->mddev->barriers_work = 0;
 318         } else {
 319                 /*
 320                  * this branch is our 'one mirror IO has finished' event handler:
 321                  */
 322                 r1_bio->bios[mirror] = NULL;
 323                 if (!uptodate) {
 324                         md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 325                         /* an I/O failed, we can't clear the bitmap */
 326                         set_bit(R1BIO_Degraded, &r1_bio->state);
 327                 } else
 328                         /*
 329                          * Set R1BIO_Uptodate in our master bio, so that
 330                          * we will return a good error code for to the higher
 331                          * levels even if IO on some other mirrored buffer fails.
 332                          *
 333                          * The 'master' represents the composite IO operation to
 334                          * user-side. So if something waits for IO, then it will
 335                          * wait for the 'master' bio.
 336                          */
 337                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 338
 339                 update_head_pos(mirror, r1_bio);
 340
 341                 if (behind) {
 342                         if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 343                                 atomic_dec(&r1_bio->behind_remaining);
 344
 345                         /* In behind mode, we ACK the master bio once the I/O has safely
 346                          * reached all non-writemostly disks. Setting the Returned bit
 347                          * ensures that this gets done only once -- we don't ever want to
 348                          * return -EIO here, instead we'll wait */
 349
 350                         if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 351                             test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 352                                 /* Maybe we can return now */
 353                                 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 354                                         struct bio *mbio = r1_bio->master_bio;
 355                                         PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 356                                                (unsigned long long) mbio->bi_sector,
 357                                                (unsigned long long) mbio->bi_sector +
 358                                                (mbio->bi_size >> 9) - 1);
 359                                         bio_endio(mbio, mbio->bi_size, 0);
 360                                 }
 361                         }
 362                 }
 363         }
 364         /*
 365          *
 366          * Let's see if all mirrored write operations have finished
 367          * already.
 368          */
 369         if (atomic_dec_and_test(&r1_bio->remaining)) {
 370                 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
 371                         reschedule_retry(r1_bio);
 372                         /* Don't dec_pending yet, we want to hold
 373                          * the reference over the retry
 374                          */
 375                         return 0;
 376                 }
 377                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 378                         /* free extra copy of the data pages */
 379                         int i = bio->bi_vcnt;
 380                         while (i--)
 381                                 __free_page(bio->bi_io_vec[i].bv_page);
 382                 }
 383                 /* clear the bitmap if all writes complete successfully */
 384                 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 385                                 r1_bio->sectors,
 386                                 !test_bit(R1BIO_Degraded, &r1_bio->state),
 387                                 behind);
 388                 md_write_end(r1_bio->mddev);
 389                 raid_end_bio_io(r1_bio);
 390         }
 391
 392         if (r1_bio->bios[mirror]==NULL)
 393                 bio_put(bio);
 394
 395         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 396         return 0;
 397 }
 398
 399
 400 /*
 401  * This routine returns the disk from which the requested read should
 402  * be done. There is a per-array 'next expected sequential IO' sector
 403  * number - if this matches on the next IO then we use the last disk.
 404  * There is also a per-disk 'last know head position' sector that is
 405  * maintained from IRQ contexts, both the normal and the resync IO
 406  * completion handlers update this position correctly. If there is no
 407  * perfect sequential match then we pick the disk whose head is closest.
 408  *
 409  * If there are 2 mirrors in the same 2 devices, performance degrades
 410  * because position is mirror, not device based.
 411  *
 412  * The rdev for the device selected will have nr_pending incremented.
 413  */
 414 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 415 {
 416         const unsigned long this_sector = r1_bio->sector;
 417         int new_disk = conf->last_used, disk = new_disk;
 418         int wonly_disk = -1;
 419         const int sectors = r1_bio->sectors;
 420         sector_t new_distance, current_distance;
 421         mdk_rdev_t *rdev;
 422
 423         rcu_read_lock();
 424         /*
 425          * Check if we can balance. We can balance on the whole
 426          * device if no resync is going on, or below the resync window.
 427          * We take the first readable disk when above the resync window.
 428          */
 429  retry:
 430         if (conf->mddev->recovery_cp < MaxSector &&
 431             (this_sector + sectors >= conf->next_resync)) {
 432                 /* Choose the first operation device, for consistancy */
 433                 new_disk = 0;
 434
 435                 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 436                      !rdev || !test_bit(In_sync, &rdev->flags)
 437                              || test_bit(WriteMostly, &rdev->flags);
 438                      rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
 439
 440                         if (rdev && test_bit(In_sync, &rdev->flags))
 441                                 wonly_disk = new_disk;
 442
 443                         if (new_disk == conf->raid_disks - 1) {
 444                                 new_disk = wonly_disk;
 445                                 break;
 446                         }
 447                 }
 448                 goto rb_out;
 449         }
 450
 451
 452         /* make sure the disk is operational */
 453         for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 454              !rdev || !test_bit(In_sync, &rdev->flags) ||
 455                      test_bit(WriteMostly, &rdev->flags);
 456              rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
 457
 458                 if (rdev && test_bit(In_sync, &rdev->flags))
 459                         wonly_disk = new_disk;
 460
 461                 if (new_disk <= 0)
 462                         new_disk = conf->raid_disks;
 463                 new_disk--;
 464                 if (new_disk == disk) {
 465                         new_disk = wonly_disk;
 466                         break;
 467                 }
 468         }
 469
 470         if (new_disk < 0)
 471                 goto rb_out;
 472
 473         disk = new_disk;
 474         /* now disk == new_disk == starting point for search */
 475
 476         /*
 477          * Don't change to another disk for sequential reads:
 478          */
 479         if (conf->next_seq_sect == this_sector)
 480                 goto rb_out;
 481         if (this_sector == conf->mirrors[new_disk].head_position)
 482                 goto rb_out;
 483
 484         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 485
 486         /* Find the disk whose head is closest */
 487
 488         do {
 489                 if (disk <= 0)
 490                         disk = conf->raid_disks;
 491                 disk--;
 492
 493                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 494
 495                 if (!rdev ||
 496                     !test_bit(In_sync, &rdev->flags) ||
 497                     test_bit(WriteMostly, &rdev->flags))
 498                         continue;
 499
 500                 if (!atomic_read(&rdev->nr_pending)) {
 501                         new_disk = disk;
 502                         break;
 503                 }
 504                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 505                 if (new_distance < current_distance) {
 506                         current_distance = new_distance;
 507                         new_disk = disk;
 508                 }
 509         } while (disk != conf->last_used);
 510
 511  rb_out:
 512
 513
 514         if (new_disk >= 0) {
 515                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 516                 if (!rdev)
 517                         goto retry;
 518                 atomic_inc(&rdev->nr_pending);
 519                 if (!test_bit(In_sync, &rdev->flags)) {
 520                         /* cannot risk returning a device that failed
 521                          * before we inc'ed nr_pending
 522                          */
 523                         atomic_dec(&rdev->nr_pending);
 524                         goto retry;
 525                 }
 526                 conf->next_seq_sect = this_sector + sectors;
 527                 conf->last_used = new_disk;
 528         }
 529         rcu_read_unlock();
 530
 531         return new_disk;
 532 }
 533
 534 static void unplug_slaves(mddev_t *mddev)
 535 {
 536         conf_t *conf = mddev_to_conf(mddev);
 537         int i;
 538
 539         rcu_read_lock();
 540         for (i=0; i<mddev->raid_disks; i++) {
 541                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 542                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 543                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 544
 545                         atomic_inc(&rdev->nr_pending);
 546                         rcu_read_unlock();
 547
 548                         if (r_queue->unplug_fn)
 549                                 r_queue->unplug_fn(r_queue);
 550
 551                         rdev_dec_pending(rdev, mddev);
 552                         rcu_read_lock();
 553                 }
 554         }
 555         rcu_read_unlock();
 556 }
 557
 558 static void raid1_unplug(request_queue_t *q)
 559 {
 560         mddev_t *mddev = q->queuedata;
 561
 562         unplug_slaves(mddev);
 563         md_wakeup_thread(mddev->thread);
 564 }
 565
 566 static int raid1_issue_flush(request_queue_t *q, struct gendisk *disk,
 567                              sector_t *error_sector)
 568 {
 569         mddev_t *mddev = q->queuedata;
 570         conf_t *conf = mddev_to_conf(mddev);
 571         int i, ret = 0;
 572
 573         rcu_read_lock();
 574         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
 575                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 576                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 577                         struct block_device *bdev = rdev->bdev;
 578                         request_queue_t *r_queue = bdev_get_queue(bdev);
 579
 580                         if (!r_queue->issue_flush_fn)
 581                                 ret = -EOPNOTSUPP;
 582                         else {
 583                                 atomic_inc(&rdev->nr_pending);
 584                                 rcu_read_unlock();
 585                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
 586                                                               error_sector);
 587                                 rdev_dec_pending(rdev, mddev);
 588                                 rcu_read_lock();
 589                         }
 590                 }
 591         }
 592         rcu_read_unlock();
 593         return ret;
 594 }
 595
 596 /*
 597  * Throttle resync depth, so that we can both get proper overlapping of
 598  * requests, but are still able to handle normal requests quickly.
 599  */
 600 #define RESYNC_DEPTH 32
 601
 602 static void device_barrier(conf_t *conf, sector_t sect)
 603 {
 604         spin_lock_irq(&conf->resync_lock);
 605         wait_event_lock_irq(conf->wait_idle, !waitqueue_active(&conf->wait_resume),
 606                             conf->resync_lock, raid1_unplug(conf->mddev->queue));
 607
 608         if (!conf->barrier++) {
 609                 wait_event_lock_irq(conf->wait_idle, !conf->nr_pending,
 610                                     conf->resync_lock, raid1_unplug(conf->mddev->queue));
 611                 if (conf->nr_pending)
 612                         BUG();
 613         }
 614         wait_event_lock_irq(conf->wait_resume, conf->barrier < RESYNC_DEPTH,
 615                             conf->resync_lock, raid1_unplug(conf->mddev->queue));
 616         conf->next_resync = sect;
 617         spin_unlock_irq(&conf->resync_lock);
 618 }
 619
 620 /* duplicate the data pages for behind I/O */
 621 static struct page **alloc_behind_pages(struct bio *bio)
 622 {
 623         int i;
 624         struct bio_vec *bvec;
 625         struct page **pages = kmalloc(bio->bi_vcnt * sizeof(struct page *),
 626                                         GFP_NOIO);
 627         if (unlikely(!pages))
 628                 goto do_sync_io;
 629
 630         memset(pages, 0, bio->bi_vcnt * sizeof(struct page *));
 631
 632         bio_for_each_segment(bvec, bio, i) {
 633                 pages[i] = alloc_page(GFP_NOIO);
 634                 if (unlikely(!pages[i]))
 635                         goto do_sync_io;
 636                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 637                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 638                 kunmap(pages[i]);
 639                 kunmap(bvec->bv_page);
 640         }
 641
 642         return pages;
 643
 644 do_sync_io:
 645         if (pages)
 646                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
 647                         __free_page(pages[i]);
 648         kfree(pages);
 649         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 650         return NULL;
 651 }
 652
 653 static int make_request(request_queue_t *q, struct bio * bio)
 654 {
 655         mddev_t *mddev = q->queuedata;
 656         conf_t *conf = mddev_to_conf(mddev);
 657         mirror_info_t *mirror;
 658         r1bio_t *r1_bio;
 659         struct bio *read_bio;
 660         int i, targets = 0, disks;
 661         mdk_rdev_t *rdev;
 662         struct bitmap *bitmap = mddev->bitmap;
 663         unsigned long flags;
 664         struct bio_list bl;
 665         struct page **behind_pages = NULL;
 666         const int rw = bio_data_dir(bio);
 667         int do_barriers;
 668
 669         if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
 670                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
 671                 return 0;
 672         }
 673
 674         /*
 675          * Register the new request and wait if the reconstruction
 676          * thread has put up a bar for new requests.
 677          * Continue immediately if no resync is active currently.
 678          */
 679         md_write_start(mddev, bio); /* wait on superblock update early */
 680
 681         spin_lock_irq(&conf->resync_lock);
 682         wait_event_lock_irq(conf->wait_resume, !conf->barrier, conf->resync_lock, );
 683         conf->nr_pending++;
 684         spin_unlock_irq(&conf->resync_lock);
 685
 686         disk_stat_inc(mddev->gendisk, ios[rw]);
 687         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
 688
 689         /*
 690          * make_request() can abort the operation when READA is being
 691          * used and no empty request is available.
 692          *
 693          */
 694         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 695
 696         r1_bio->master_bio = bio;
 697         r1_bio->sectors = bio->bi_size >> 9;
 698         r1_bio->state = 0;
 699         r1_bio->mddev = mddev;
 700         r1_bio->sector = bio->bi_sector;
 701
 702         if (rw == READ) {
 703                 /*
 704                  * read balancing logic:
 705                  */
 706                 int rdisk = read_balance(conf, r1_bio);
 707
 708                 if (rdisk < 0) {
 709                         /* couldn't find anywhere to read from */
 710                         raid_end_bio_io(r1_bio);
 711                         return 0;
 712                 }
 713                 mirror = conf->mirrors + rdisk;
 714
 715                 r1_bio->read_disk = rdisk;
 716
 717                 read_bio = bio_clone(bio, GFP_NOIO);
 718
 719                 r1_bio->bios[rdisk] = read_bio;
 720
 721                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 722                 read_bio->bi_bdev = mirror->rdev->bdev;
 723                 read_bio->bi_end_io = raid1_end_read_request;
 724                 read_bio->bi_rw = READ;
 725                 read_bio->bi_private = r1_bio;
 726
 727                 generic_make_request(read_bio);
 728                 return 0;
 729         }
 730
 731         /*
 732          * WRITE:
 733          */
 734         /* first select target devices under spinlock and
 735          * inc refcount on their rdev.  Record them by setting
 736          * bios[x] to bio
 737          */
 738         disks = conf->raid_disks;
 739 #if 0
 740         { static int first=1;
 741         if (first) printk("First Write sector %llu disks %d\n",
 742                           (unsigned long long)r1_bio->sector, disks);
 743         first = 0;
 744         }
 745 #endif
 746         rcu_read_lock();
 747         for (i = 0;  i < disks; i++) {
 748                 if ((rdev=rcu_dereference(conf->mirrors[i].rdev)) != NULL &&
 749                     !test_bit(Faulty, &rdev->flags)) {
 750                         atomic_inc(&rdev->nr_pending);
 751                         if (test_bit(Faulty, &rdev->flags)) {
 752                                 atomic_dec(&rdev->nr_pending);
 753                                 r1_bio->bios[i] = NULL;
 754                         } else
 755                                 r1_bio->bios[i] = bio;
 756                         targets++;
 757                 } else
 758                         r1_bio->bios[i] = NULL;
 759         }
 760         rcu_read_unlock();
 761
 762         BUG_ON(targets == 0); /* we never fail the last device */
 763
 764         if (targets < conf->raid_disks) {
 765                 /* array is degraded, we will not clear the bitmap
 766                  * on I/O completion (see raid1_end_write_request) */
 767                 set_bit(R1BIO_Degraded, &r1_bio->state);
 768         }
 769
 770         /* do behind I/O ? */
 771         if (bitmap &&
 772             atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
 773             (behind_pages = alloc_behind_pages(bio)) != NULL)
 774                 set_bit(R1BIO_BehindIO, &r1_bio->state);
 775
 776         atomic_set(&r1_bio->remaining, 0);
 777         atomic_set(&r1_bio->behind_remaining, 0);
 778
 779         do_barriers = bio->bi_rw & BIO_RW_BARRIER;
 780         if (do_barriers)
 781                 set_bit(R1BIO_Barrier, &r1_bio->state);
 782
 783         bio_list_init(&bl);
 784         for (i = 0; i < disks; i++) {
 785                 struct bio *mbio;
 786                 if (!r1_bio->bios[i])
 787                         continue;
 788
 789                 mbio = bio_clone(bio, GFP_NOIO);
 790                 r1_bio->bios[i] = mbio;
 791
 792                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 793                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 794                 mbio->bi_end_io = raid1_end_write_request;
 795                 mbio->bi_rw = WRITE | do_barriers;
 796                 mbio->bi_private = r1_bio;
 797
 798                 if (behind_pages) {
 799                         struct bio_vec *bvec;
 800                         int j;
 801
 802                         /* Yes, I really want the '__' version so that
 803                          * we clear any unused pointer in the io_vec, rather
 804                          * than leave them unchanged.  This is important
 805                          * because when we come to free the pages, we won't
 806                          * know the originial bi_idx, so we just free
 807                          * them all
 808                          */
 809                         __bio_for_each_segment(bvec, mbio, j, 0)
 810                                 bvec->bv_page = behind_pages[j];
 811                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 812                                 atomic_inc(&r1_bio->behind_remaining);
 813                 }
 814
 815                 atomic_inc(&r1_bio->remaining);
 816
 817                 bio_list_add(&bl, mbio);
 818         }
 819         kfree(behind_pages); /* the behind pages are attached to the bios now */
 820
 821         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
 822                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 823         spin_lock_irqsave(&conf->device_lock, flags);
 824         bio_list_merge(&conf->pending_bio_list, &bl);
 825         bio_list_init(&bl);
 826
 827         blk_plug_device(mddev->queue);
 828         spin_unlock_irqrestore(&conf->device_lock, flags);
 829
 830 #if 0
 831         while ((bio = bio_list_pop(&bl)) != NULL)
 832                 generic_make_request(bio);
 833 #endif
 834
 835         return 0;
 836 }
 837
 838 static void status(struct seq_file *seq, mddev_t *mddev)
 839 {
 840         conf_t *conf = mddev_to_conf(mddev);
 841         int i;
 842
 843         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 844                                                 conf->working_disks);
 845         for (i = 0; i < conf->raid_disks; i++)
 846                 seq_printf(seq, "%s",
 847                               conf->mirrors[i].rdev &&
 848                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
 849         seq_printf(seq, "]");
 850 }
 851
 852
 853 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 854 {
 855         char b[BDEVNAME_SIZE];
 856         conf_t *conf = mddev_to_conf(mddev);
 857
 858         /*
 859          * If it is not operational, then we have already marked it as dead
 860          * else if it is the last working disks, ignore the error, let the
 861          * next level up know.
 862          * else mark the drive as failed
 863          */
 864         if (test_bit(In_sync, &rdev->flags)
 865             && conf->working_disks == 1)
 866                 /*
 867                  * Don't fail the drive, act as though we were just a
 868                  * normal single drive
 869                  */
 870                 return;
 871         if (test_bit(In_sync, &rdev->flags)) {
 872                 mddev->degraded++;
 873                 conf->working_disks--;
 874                 /*
 875                  * if recovery is running, make sure it aborts.
 876                  */
 877                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 878         }
 879         clear_bit(In_sync, &rdev->flags);
 880         set_bit(Faulty, &rdev->flags);
 881         mddev->sb_dirty = 1;
 882         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
 883                 "       Operation continuing on %d devices\n",
 884                 bdevname(rdev->bdev,b), conf->working_disks);
 885 }
 886
 887 static void print_conf(conf_t *conf)
 888 {
 889         int i;
 890         mirror_info_t *tmp;
 891
 892         printk("RAID1 conf printout:\n");
 893         if (!conf) {
 894                 printk("(!conf)\n");
 895                 return;
 896         }
 897         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 898                 conf->raid_disks);
 899
 900         for (i = 0; i < conf->raid_disks; i++) {
 901                 char b[BDEVNAME_SIZE];
 902                 tmp = conf->mirrors + i;
 903                 if (tmp->rdev)
 904                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
 905                                 i, !test_bit(In_sync, &tmp->rdev->flags), !test_bit(Faulty, &tmp->rdev->flags),
 906                                 bdevname(tmp->rdev->bdev,b));
 907         }
 908 }
 909
 910 static void close_sync(conf_t *conf)
 911 {
 912         spin_lock_irq(&conf->resync_lock);
 913         wait_event_lock_irq(conf->wait_resume, !conf->barrier,
 914                             conf->resync_lock,  raid1_unplug(conf->mddev->queue));
 915         spin_unlock_irq(&conf->resync_lock);
 916
 917         if (conf->barrier) BUG();
 918         if (waitqueue_active(&conf->wait_idle)) BUG();
 919
 920         mempool_destroy(conf->r1buf_pool);
 921         conf->r1buf_pool = NULL;
 922 }
 923
 924 static int raid1_spare_active(mddev_t *mddev)
 925 {
 926         int i;
 927         conf_t *conf = mddev->private;
 928         mirror_info_t *tmp;
 929
 930         /*
 931          * Find all failed disks within the RAID1 configuration
 932          * and mark them readable
 933          */
 934         for (i = 0; i < conf->raid_disks; i++) {
 935                 tmp = conf->mirrors + i;
 936                 if (tmp->rdev
 937                     && !test_bit(Faulty, &tmp->rdev->flags)
 938                     && !test_bit(In_sync, &tmp->rdev->flags)) {
 939                         conf->working_disks++;
 940                         mddev->degraded--;
 941                         set_bit(In_sync, &tmp->rdev->flags);
 942                 }
 943         }
 944
 945         print_conf(conf);
 946         return 0;
 947 }
 948
 949
 950 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 951 {
 952         conf_t *conf = mddev->private;
 953         int found = 0;
 954         int mirror = 0;
 955         mirror_info_t *p;
 956
 957         for (mirror=0; mirror < mddev->raid_disks; mirror++)
 958                 if ( !(p=conf->mirrors+mirror)->rdev) {
 959
 960                         blk_queue_stack_limits(mddev->queue,
 961                                                rdev->bdev->bd_disk->queue);
 962                         /* as we don't honour merge_bvec_fn, we must never risk
 963                          * violating it, so limit ->max_sector to one PAGE, as
 964                          * a one page request is never in violation.
 965                          */
 966                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 967                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
 968                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
 969
 970                         p->head_position = 0;
 971                         rdev->raid_disk = mirror;
 972                         found = 1;
 973                         /* As all devices are equivalent, we don't need a full recovery
 974                          * if this was recently any drive of the array
 975                          */
 976                         if (rdev->saved_raid_disk < 0)
 977                                 conf->fullsync = 1;
 978                         rcu_assign_pointer(p->rdev, rdev);
 979                         break;
 980                 }
 981
 982         print_conf(conf);
 983         return found;
 984 }
 985
 986 static int raid1_remove_disk(mddev_t *mddev, int number)
 987 {
 988         conf_t *conf = mddev->private;
 989         int err = 0;
 990         mdk_rdev_t *rdev;
 991         mirror_info_t *p = conf->mirrors+ number;
 992
 993         print_conf(conf);
 994         rdev = p->rdev;
 995         if (rdev) {
 996                 if (test_bit(In_sync, &rdev->flags) ||
 997                     atomic_read(&rdev->nr_pending)) {
 998                         err = -EBUSY;
 999                         goto abort;
1000                 }
1001                 p->rdev = NULL;
1002                 synchronize_rcu();
1003                 if (atomic_read(&rdev->nr_pending)) {
1004                         /* lost the race, try later */
1005                         err = -EBUSY;
1006                         p->rdev = rdev;
1007                 }
1008         }
1009 abort:
1010
1011         print_conf(conf);
1012         return err;
1013 }
1014
1015
1016 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1017 {
1018         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1019         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1020         conf_t *conf = mddev_to_conf(r1_bio->mddev);
1021
1022         if (bio->bi_size)
1023                 return 1;
1024
1025         if (r1_bio->bios[r1_bio->read_disk] != bio)
1026                 BUG();
1027         update_head_pos(r1_bio->read_disk, r1_bio);
1028         /*
1029          * we have read a block, now it needs to be re-written,
1030          * or re-read if the read failed.
1031          * We don't do much here, just schedule handling by raid1d
1032          */
1033         if (!uptodate) {
1034                 md_error(r1_bio->mddev,
1035                          conf->mirrors[r1_bio->read_disk].rdev);
1036         } else
1037                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1038         rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
1039         reschedule_retry(r1_bio);
1040         return 0;
1041 }
1042
1043 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1044 {
1045         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1046         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1047         mddev_t *mddev = r1_bio->mddev;
1048         conf_t *conf = mddev_to_conf(mddev);
1049         int i;
1050         int mirror=0;
1051
1052         if (bio->bi_size)
1053                 return 1;
1054
1055         for (i = 0; i < conf->raid_disks; i++)
1056                 if (r1_bio->bios[i] == bio) {
1057                         mirror = i;
1058                         break;
1059                 }
1060         if (!uptodate)
1061                 md_error(mddev, conf->mirrors[mirror].rdev);
1062
1063         update_head_pos(mirror, r1_bio);
1064
1065         if (atomic_dec_and_test(&r1_bio->remaining)) {
1066                 md_done_sync(mddev, r1_bio->sectors, uptodate);
1067                 put_buf(r1_bio);
1068         }
1069         rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
1070         return 0;
1071 }
1072
1073 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1074 {
1075         conf_t *conf = mddev_to_conf(mddev);
1076         int i;
1077         int disks = conf->raid_disks;
1078         struct bio *bio, *wbio;
1079
1080         bio = r1_bio->bios[r1_bio->read_disk];
1081
1082 /*
1083         if (r1_bio->sector == 0) printk("First sync write startss\n");
1084 */
1085         /*
1086          * schedule writes
1087          */
1088         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1089                 /*
1090                  * There is no point trying a read-for-reconstruct as
1091                  * reconstruct is about to be aborted
1092                  */
1093                 char b[BDEVNAME_SIZE];
1094                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1095                         " for block %llu\n",
1096                         bdevname(bio->bi_bdev,b),
1097                         (unsigned long long)r1_bio->sector);
1098                 md_done_sync(mddev, r1_bio->sectors, 0);
1099                 put_buf(r1_bio);
1100                 return;
1101         }
1102
1103         atomic_set(&r1_bio->remaining, 1);
1104         for (i = 0; i < disks ; i++) {
1105                 wbio = r1_bio->bios[i];
1106                 if (wbio->bi_end_io != end_sync_write)
1107                         continue;
1108
1109                 atomic_inc(&conf->mirrors[i].rdev->nr_pending);
1110                 atomic_inc(&r1_bio->remaining);
1111                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1112
1113                 generic_make_request(wbio);
1114         }
1115
1116         if (atomic_dec_and_test(&r1_bio->remaining)) {
1117                 /* if we're here, all write(s) have completed, so clean up */
1118                 md_done_sync(mddev, r1_bio->sectors, 1);
1119                 put_buf(r1_bio);
1120         }
1121 }
1122
1123 /*
1124  * This is a kernel thread which:
1125  *
1126  *      1.      Retries failed read operations on working mirrors.
1127  *      2.      Updates the raid superblock when problems encounter.
1128  *      3.      Performs writes following reads for array syncronising.
1129  */
1130
1131 static void raid1d(mddev_t *mddev)
1132 {
1133         r1bio_t *r1_bio;
1134         struct bio *bio;
1135         unsigned long flags;
1136         conf_t *conf = mddev_to_conf(mddev);
1137         struct list_head *head = &conf->retry_list;
1138         int unplug=0;
1139         mdk_rdev_t *rdev;
1140
1141         md_check_recovery(mddev);
1142
1143         for (;;) {
1144                 char b[BDEVNAME_SIZE];
1145                 spin_lock_irqsave(&conf->device_lock, flags);
1146
1147                 if (conf->pending_bio_list.head) {
1148                         bio = bio_list_get(&conf->pending_bio_list);
1149                         blk_remove_plug(mddev->queue);
1150                         spin_unlock_irqrestore(&conf->device_lock, flags);
1151                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1152                         if (bitmap_unplug(mddev->bitmap) != 0)
1153                                 printk("%s: bitmap file write failed!\n", mdname(mddev));
1154
1155                         while (bio) { /* submit pending writes */
1156                                 struct bio *next = bio->bi_next;
1157                                 bio->bi_next = NULL;
1158                                 generic_make_request(bio);
1159                                 bio = next;
1160                         }
1161                         unplug = 1;
1162
1163                         continue;
1164                 }
1165
1166                 if (list_empty(head))
1167                         break;
1168                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1169                 list_del(head->prev);
1170                 spin_unlock_irqrestore(&conf->device_lock, flags);
1171
1172                 mddev = r1_bio->mddev;
1173                 conf = mddev_to_conf(mddev);
1174                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1175                         sync_request_write(mddev, r1_bio);
1176                         unplug = 1;
1177                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1178                         /* some requests in the r1bio were BIO_RW_BARRIER
1179                          * requests which failed with -ENOTSUPP.  Hohumm..
1180                          * Better resubmit without the barrier.
1181                          * We know which devices to resubmit for, because
1182                          * all others have had their bios[] entry cleared.
1183                          */
1184                         int i;
1185                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1186                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1187                         for (i=0; i < conf->raid_disks; i++)
1188                                 if (r1_bio->bios[i]) {
1189                                         struct bio_vec *bvec;
1190                                         int j;
1191
1192                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1193                                         /* copy pages from the failed bio, as
1194                                          * this might be a write-behind device */
1195                                         __bio_for_each_segment(bvec, bio, j, 0)
1196                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1197                                         bio_put(r1_bio->bios[i]);
1198                                         bio->bi_sector = r1_bio->sector +
1199                                                 conf->mirrors[i].rdev->data_offset;
1200                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1201                                         bio->bi_end_io = raid1_end_write_request;
1202                                         bio->bi_rw = WRITE;
1203                                         bio->bi_private = r1_bio;
1204                                         r1_bio->bios[i] = bio;
1205                                         generic_make_request(bio);
1206                                 }
1207                 } else {
1208                         int disk;
1209                         bio = r1_bio->bios[r1_bio->read_disk];
1210                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1211                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1212                                        " read error for block %llu\n",
1213                                        bdevname(bio->bi_bdev,b),
1214                                        (unsigned long long)r1_bio->sector);
1215                                 raid_end_bio_io(r1_bio);
1216                         } else {
1217                                 r1_bio->bios[r1_bio->read_disk] = NULL;
1218                                 r1_bio->read_disk = disk;
1219                                 bio_put(bio);
1220                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1221                                 r1_bio->bios[r1_bio->read_disk] = bio;
1222                                 rdev = conf->mirrors[disk].rdev;
1223                                 if (printk_ratelimit())
1224                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1225                                                " another mirror\n",
1226                                                bdevname(rdev->bdev,b),
1227                                                (unsigned long long)r1_bio->sector);
1228                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1229                                 bio->bi_bdev = rdev->bdev;
1230                                 bio->bi_end_io = raid1_end_read_request;
1231                                 bio->bi_rw = READ;
1232                                 bio->bi_private = r1_bio;
1233                                 unplug = 1;
1234                                 generic_make_request(bio);
1235                         }
1236                 }
1237         }
1238         spin_unlock_irqrestore(&conf->device_lock, flags);
1239         if (unplug)
1240                 unplug_slaves(mddev);
1241 }
1242
1243
1244 static int init_resync(conf_t *conf)
1245 {
1246         int buffs;
1247
1248         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1249         if (conf->r1buf_pool)
1250                 BUG();
1251         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1252                                           conf->poolinfo);
1253         if (!conf->r1buf_pool)
1254                 return -ENOMEM;
1255         conf->next_resync = 0;
1256         return 0;
1257 }
1258
1259 /*
1260  * perform a "sync" on one "block"
1261  *
1262  * We need to make sure that no normal I/O request - particularly write
1263  * requests - conflict with active sync requests.
1264  *
1265  * This is achieved by tracking pending requests and a 'barrier' concept
1266  * that can be installed to exclude normal IO requests.
1267  */
1268
1269 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1270 {
1271         conf_t *conf = mddev_to_conf(mddev);
1272         mirror_info_t *mirror;
1273         r1bio_t *r1_bio;
1274         struct bio *bio;
1275         sector_t max_sector, nr_sectors;
1276         int disk;
1277         int i;
1278         int wonly;
1279         int write_targets = 0;
1280         int sync_blocks;
1281         int still_degraded = 0;
1282
1283         if (!conf->r1buf_pool)
1284         {
1285 /*
1286                 printk("sync start - bitmap %p\n", mddev->bitmap);
1287 */
1288                 if (init_resync(conf))
1289                         return 0;
1290         }
1291
1292         max_sector = mddev->size << 1;
1293         if (sector_nr >= max_sector) {
1294                 /* If we aborted, we need to abort the
1295                  * sync on the 'current' bitmap chunk (there will
1296                  * only be one in raid1 resync.
1297                  * We can find the current addess in mddev->curr_resync
1298                  */
1299                 if (mddev->curr_resync < max_sector) /* aborted */
1300                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1301                                                 &sync_blocks, 1);
1302                 else /* completed sync */
1303                         conf->fullsync = 0;
1304
1305                 bitmap_close_sync(mddev->bitmap);
1306                 close_sync(conf);
1307                 return 0;
1308         }
1309
1310         /* before building a request, check if we can skip these blocks..
1311          * This call the bitmap_start_sync doesn't actually record anything
1312          */
1313         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1314             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1315                 /* We can skip this block, and probably several more */
1316                 *skipped = 1;
1317                 return sync_blocks;
1318         }
1319         /*
1320          * If there is non-resync activity waiting for us then
1321          * put in a delay to throttle resync.
1322          */
1323         if (!go_faster && waitqueue_active(&conf->wait_resume))
1324                 msleep_interruptible(1000);
1325         device_barrier(conf, sector_nr + RESYNC_SECTORS);
1326
1327         /*
1328          * If reconstructing, and >1 working disc,
1329          * could dedicate one to rebuild and others to
1330          * service read requests ..
1331          */
1332         disk = conf->last_used;
1333         /* make sure disk is operational */
1334         wonly = disk;
1335         while (conf->mirrors[disk].rdev == NULL ||
1336                !test_bit(In_sync, &conf->mirrors[disk].rdev->flags) ||
1337                test_bit(WriteMostly, &conf->mirrors[disk].rdev->flags)
1338                 ) {
1339                 if (conf->mirrors[disk].rdev  &&
1340                     test_bit(In_sync, &conf->mirrors[disk].rdev->flags))
1341                         wonly = disk;
1342                 if (disk <= 0)
1343                         disk = conf->raid_disks;
1344                 disk--;
1345                 if (disk == conf->last_used) {
1346                         disk = wonly;
1347                         break;
1348                 }
1349         }
1350         conf->last_used = disk;
1351         atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
1352
1353
1354         mirror = conf->mirrors + disk;
1355
1356         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1357
1358         spin_lock_irq(&conf->resync_lock);
1359         conf->nr_pending++;
1360         spin_unlock_irq(&conf->resync_lock);
1361
1362         r1_bio->mddev = mddev;
1363         r1_bio->sector = sector_nr;
1364         r1_bio->state = 0;
1365         set_bit(R1BIO_IsSync, &r1_bio->state);
1366         r1_bio->read_disk = disk;
1367
1368         for (i=0; i < conf->raid_disks; i++) {
1369                 bio = r1_bio->bios[i];
1370
1371                 /* take from bio_init */
1372                 bio->bi_next = NULL;
1373                 bio->bi_flags |= 1 << BIO_UPTODATE;
1374                 bio->bi_rw = 0;
1375                 bio->bi_vcnt = 0;
1376                 bio->bi_idx = 0;
1377                 bio->bi_phys_segments = 0;
1378                 bio->bi_hw_segments = 0;
1379                 bio->bi_size = 0;
1380                 bio->bi_end_io = NULL;
1381                 bio->bi_private = NULL;
1382
1383                 if (i == disk) {
1384                         bio->bi_rw = READ;
1385                         bio->bi_end_io = end_sync_read;
1386                 } else if (conf->mirrors[i].rdev == NULL ||
1387                            test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
1388                         still_degraded = 1;
1389                         continue;
1390                 } else if (!test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
1391                            sector_nr + RESYNC_SECTORS > mddev->recovery_cp   ||
1392                            test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1393                         bio->bi_rw = WRITE;
1394                         bio->bi_end_io = end_sync_write;
1395                         write_targets ++;
1396                 } else
1397                         /* no need to read or write here */
1398                         continue;
1399                 bio->bi_sector = sector_nr + conf->mirrors[i].rdev->data_offset;
1400                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1401                 bio->bi_private = r1_bio;
1402         }
1403
1404         if (write_targets == 0) {
1405                 /* There is nowhere to write, so all non-sync
1406                  * drives must be failed - so we are finished
1407                  */
1408                 sector_t rv = max_sector - sector_nr;
1409                 *skipped = 1;
1410                 put_buf(r1_bio);
1411                 rdev_dec_pending(conf->mirrors[disk].rdev, mddev);
1412                 return rv;
1413         }
1414
1415         nr_sectors = 0;
1416         sync_blocks = 0;
1417         do {
1418                 struct page *page;
1419                 int len = PAGE_SIZE;
1420                 if (sector_nr + (len>>9) > max_sector)
1421                         len = (max_sector - sector_nr) << 9;
1422                 if (len == 0)
1423                         break;
1424                 if (sync_blocks == 0) {
1425                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1426                                                &sync_blocks, still_degraded) &&
1427                             !conf->fullsync &&
1428                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1429                                 break;
1430                         if (sync_blocks < (PAGE_SIZE>>9))
1431                                 BUG();
1432                         if (len > (sync_blocks<<9))
1433                                 len = sync_blocks<<9;
1434                 }
1435
1436                 for (i=0 ; i < conf->raid_disks; i++) {
1437                         bio = r1_bio->bios[i];
1438                         if (bio->bi_end_io) {
1439                                 page = r1_bio->bios[0]->bi_io_vec[bio->bi_vcnt].bv_page;
1440                                 if (bio_add_page(bio, page, len, 0) == 0) {
1441                                         /* stop here */
1442                                         r1_bio->bios[0]->bi_io_vec[bio->bi_vcnt].bv_page = page;
1443                                         while (i > 0) {
1444                                                 i--;
1445                                                 bio = r1_bio->bios[i];
1446                                                 if (bio->bi_end_io==NULL)
1447                                                         continue;
1448                                                 /* remove last page from this bio */
1449                                                 bio->bi_vcnt--;
1450                                                 bio->bi_size -= len;
1451                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1452                                         }
1453                                         goto bio_full;
1454                                 }
1455                         }
1456                 }
1457                 nr_sectors += len>>9;
1458                 sector_nr += len>>9;
1459                 sync_blocks -= (len>>9);
1460         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1461  bio_full:
1462         bio = r1_bio->bios[disk];
1463         r1_bio->sectors = nr_sectors;
1464
1465         md_sync_acct(mirror->rdev->bdev, nr_sectors);
1466
1467         generic_make_request(bio);
1468
1469         return nr_sectors;
1470 }
1471
1472 static int run(mddev_t *mddev)
1473 {
1474         conf_t *conf;
1475         int i, j, disk_idx;
1476         mirror_info_t *disk;
1477         mdk_rdev_t *rdev;
1478         struct list_head *tmp;
1479
1480         if (mddev->level != 1) {
1481                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1482                        mdname(mddev), mddev->level);
1483                 goto out;
1484         }
1485         /*
1486          * copy the already verified devices into our private RAID1
1487          * bookkeeping area. [whatever we allocate in run(),
1488          * should be freed in stop()]
1489          */
1490         conf = kmalloc(sizeof(conf_t), GFP_KERNEL);
1491         mddev->private = conf;
1492         if (!conf)
1493                 goto out_no_mem;
1494
1495         memset(conf, 0, sizeof(*conf));
1496         conf->mirrors = kmalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1497                                  GFP_KERNEL);
1498         if (!conf->mirrors)
1499                 goto out_no_mem;
1500
1501         memset(conf->mirrors, 0, sizeof(struct mirror_info)*mddev->raid_disks);
1502
1503         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1504         if (!conf->poolinfo)
1505                 goto out_no_mem;
1506         conf->poolinfo->mddev = mddev;
1507         conf->poolinfo->raid_disks = mddev->raid_disks;
1508         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1509                                           r1bio_pool_free,
1510                                           conf->poolinfo);
1511         if (!conf->r1bio_pool)
1512                 goto out_no_mem;
1513
1514         ITERATE_RDEV(mddev, rdev, tmp) {
1515                 disk_idx = rdev->raid_disk;
1516                 if (disk_idx >= mddev->raid_disks
1517                     || disk_idx < 0)
1518                         continue;
1519                 disk = conf->mirrors + disk_idx;
1520
1521                 disk->rdev = rdev;
1522
1523                 blk_queue_stack_limits(mddev->queue,
1524                                        rdev->bdev->bd_disk->queue);
1525                 /* as we don't honour merge_bvec_fn, we must never risk
1526                  * violating it, so limit ->max_sector to one PAGE, as
1527                  * a one page request is never in violation.
1528                  */
1529                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1530                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1531                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1532
1533                 disk->head_position = 0;
1534                 if (!test_bit(Faulty, &rdev->flags) && test_bit(In_sync, &rdev->flags))
1535                         conf->working_disks++;
1536         }
1537         conf->raid_disks = mddev->raid_disks;
1538         conf->mddev = mddev;
1539         spin_lock_init(&conf->device_lock);
1540         INIT_LIST_HEAD(&conf->retry_list);
1541         if (conf->working_disks == 1)
1542                 mddev->recovery_cp = MaxSector;
1543
1544         spin_lock_init(&conf->resync_lock);
1545         init_waitqueue_head(&conf->wait_idle);
1546         init_waitqueue_head(&conf->wait_resume);
1547
1548         bio_list_init(&conf->pending_bio_list);
1549         bio_list_init(&conf->flushing_bio_list);
1550
1551         if (!conf->working_disks) {
1552                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1553                         mdname(mddev));
1554                 goto out_free_conf;
1555         }
1556
1557         mddev->degraded = 0;
1558         for (i = 0; i < conf->raid_disks; i++) {
1559
1560                 disk = conf->mirrors + i;
1561
1562                 if (!disk->rdev) {
1563                         disk->head_position = 0;
1564                         mddev->degraded++;
1565                 }
1566         }
1567
1568         /*
1569          * find the first working one and use it as a starting point
1570          * to read balancing.
1571          */
1572         for (j = 0; j < conf->raid_disks &&
1573                      (!conf->mirrors[j].rdev ||
1574                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
1575                 /* nothing */;
1576         conf->last_used = j;
1577
1578
1579         mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1580         if (!mddev->thread) {
1581                 printk(KERN_ERR
1582                        "raid1: couldn't allocate thread for %s\n",
1583                        mdname(mddev));
1584                 goto out_free_conf;
1585         }
1586         if (mddev->bitmap) mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
1587
1588         printk(KERN_INFO
1589                 "raid1: raid set %s active with %d out of %d mirrors\n",
1590                 mdname(mddev), mddev->raid_disks - mddev->degraded,
1591                 mddev->raid_disks);
1592         /*
1593          * Ok, everything is just fine now
1594          */
1595         mddev->array_size = mddev->size;
1596
1597         mddev->queue->unplug_fn = raid1_unplug;
1598         mddev->queue->issue_flush_fn = raid1_issue_flush;
1599
1600         return 0;
1601
1602 out_no_mem:
1603         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1604                mdname(mddev));
1605
1606 out_free_conf:
1607         if (conf) {
1608                 if (conf->r1bio_pool)
1609                         mempool_destroy(conf->r1bio_pool);
1610                 kfree(conf->mirrors);
1611                 kfree(conf->poolinfo);
1612                 kfree(conf);
1613                 mddev->private = NULL;
1614         }
1615 out:
1616         return -EIO;
1617 }
1618
1619 static int stop(mddev_t *mddev)
1620 {
1621         conf_t *conf = mddev_to_conf(mddev);
1622         struct bitmap *bitmap = mddev->bitmap;
1623         int behind_wait = 0;
1624
1625         /* wait for behind writes to complete */
1626         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
1627                 behind_wait++;
1628                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
1629                 set_current_state(TASK_UNINTERRUPTIBLE);
1630                 schedule_timeout(HZ); /* wait a second */
1631                 /* need to kick something here to make sure I/O goes? */
1632         }
1633
1634         md_unregister_thread(mddev->thread);
1635         mddev->thread = NULL;
1636         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1637         if (conf->r1bio_pool)
1638                 mempool_destroy(conf->r1bio_pool);
1639         kfree(conf->mirrors);
1640         kfree(conf->poolinfo);
1641         kfree(conf);
1642         mddev->private = NULL;
1643         return 0;
1644 }
1645
1646 static int raid1_resize(mddev_t *mddev, sector_t sectors)
1647 {
1648         /* no resync is happening, and there is enough space
1649          * on all devices, so we can resize.
1650          * We need to make sure resync covers any new space.
1651          * If the array is shrinking we should possibly wait until
1652          * any io in the removed space completes, but it hardly seems
1653          * worth it.
1654          */
1655         mddev->array_size = sectors>>1;
1656         set_capacity(mddev->gendisk, mddev->array_size << 1);
1657         mddev->changed = 1;
1658         if (mddev->array_size > mddev->size && mddev->recovery_cp == MaxSector) {
1659                 mddev->recovery_cp = mddev->size << 1;
1660                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1661         }
1662         mddev->size = mddev->array_size;
1663         mddev->resync_max_sectors = sectors;
1664         return 0;
1665 }
1666
1667 static int raid1_reshape(mddev_t *mddev, int raid_disks)
1668 {
1669         /* We need to:
1670          * 1/ resize the r1bio_pool
1671          * 2/ resize conf->mirrors
1672          *
1673          * We allocate a new r1bio_pool if we can.
1674          * Then raise a device barrier and wait until all IO stops.
1675          * Then resize conf->mirrors and swap in the new r1bio pool.
1676          *
1677          * At the same time, we "pack" the devices so that all the missing
1678          * devices have the higher raid_disk numbers.
1679          */
1680         mempool_t *newpool, *oldpool;
1681         struct pool_info *newpoolinfo;
1682         mirror_info_t *newmirrors;
1683         conf_t *conf = mddev_to_conf(mddev);
1684         int cnt;
1685
1686         int d, d2;
1687
1688         if (raid_disks < conf->raid_disks) {
1689                 cnt=0;
1690                 for (d= 0; d < conf->raid_disks; d++)
1691                         if (conf->mirrors[d].rdev)
1692                                 cnt++;
1693                 if (cnt > raid_disks)
1694                         return -EBUSY;
1695         }
1696
1697         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
1698         if (!newpoolinfo)
1699                 return -ENOMEM;
1700         newpoolinfo->mddev = mddev;
1701         newpoolinfo->raid_disks = raid_disks;
1702
1703         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1704                                  r1bio_pool_free, newpoolinfo);
1705         if (!newpool) {
1706                 kfree(newpoolinfo);
1707                 return -ENOMEM;
1708         }
1709         newmirrors = kmalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
1710         if (!newmirrors) {
1711                 kfree(newpoolinfo);
1712                 mempool_destroy(newpool);
1713                 return -ENOMEM;
1714         }
1715         memset(newmirrors, 0, sizeof(struct mirror_info)*raid_disks);
1716
1717         spin_lock_irq(&conf->resync_lock);
1718         conf->barrier++;
1719         wait_event_lock_irq(conf->wait_idle, !conf->nr_pending,
1720                             conf->resync_lock, raid1_unplug(mddev->queue));
1721         spin_unlock_irq(&conf->resync_lock);
1722
1723         /* ok, everything is stopped */
1724         oldpool = conf->r1bio_pool;
1725         conf->r1bio_pool = newpool;
1726
1727         for (d=d2=0; d < conf->raid_disks; d++)
1728                 if (conf->mirrors[d].rdev) {
1729                         conf->mirrors[d].rdev->raid_disk = d2;
1730                         newmirrors[d2++].rdev = conf->mirrors[d].rdev;
1731                 }
1732         kfree(conf->mirrors);
1733         conf->mirrors = newmirrors;
1734         kfree(conf->poolinfo);
1735         conf->poolinfo = newpoolinfo;
1736
1737         mddev->degraded += (raid_disks - conf->raid_disks);
1738         conf->raid_disks = mddev->raid_disks = raid_disks;
1739
1740         conf->last_used = 0; /* just make sure it is in-range */
1741         spin_lock_irq(&conf->resync_lock);
1742         conf->barrier--;
1743         spin_unlock_irq(&conf->resync_lock);
1744         wake_up(&conf->wait_resume);
1745         wake_up(&conf->wait_idle);
1746
1747
1748         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1749         md_wakeup_thread(mddev->thread);
1750
1751         mempool_destroy(oldpool);
1752         return 0;
1753 }
1754
1755 static void raid1_quiesce(mddev_t *mddev, int state)
1756 {
1757         conf_t *conf = mddev_to_conf(mddev);
1758
1759         switch(state) {
1760         case 1:
1761                 spin_lock_irq(&conf->resync_lock);
1762                 conf->barrier++;
1763                 wait_event_lock_irq(conf->wait_idle, !conf->nr_pending,
1764                                     conf->resync_lock, raid1_unplug(mddev->queue));
1765                 spin_unlock_irq(&conf->resync_lock);
1766                 break;
1767         case 0:
1768                 spin_lock_irq(&conf->resync_lock);
1769                 conf->barrier--;
1770                 spin_unlock_irq(&conf->resync_lock);
1771                 wake_up(&conf->wait_resume);
1772                 wake_up(&conf->wait_idle);
1773                 break;
1774         }
1775         if (mddev->thread) {
1776                 if (mddev->bitmap)
1777                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
1778                 else
1779                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1780                 md_wakeup_thread(mddev->thread);
1781         }
1782 }
1783
1784
1785 static mdk_personality_t raid1_personality =
1786 {
1787         .name           = "raid1",
1788         .owner          = THIS_MODULE,
1789         .make_request   = make_request,
1790         .run            = run,
1791         .stop           = stop,
1792         .status         = status,
1793         .error_handler  = error,
1794         .hot_add_disk   = raid1_add_disk,
1795         .hot_remove_disk= raid1_remove_disk,
1796         .spare_active   = raid1_spare_active,
1797         .sync_request   = sync_request,
1798         .resize         = raid1_resize,
1799         .reshape        = raid1_reshape,
1800         .quiesce        = raid1_quiesce,
1801 };
1802
1803 static int __init raid_init(void)
1804 {
1805         return register_md_personality(RAID1, &raid1_personality);
1806 }
1807
1808 static void raid_exit(void)
1809 {
1810         unregister_md_personality(RAID1);
1811 }
1812
1813 module_init(raid_init);
1814 module_exit(raid_exit);
1815 MODULE_LICENSE("GPL");
1816 MODULE_ALIAS("md-personality-3"); /* RAID1 */