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 static void allow_barrier(conf_t *conf);
  55 static void lower_barrier(conf_t *conf);
  56
  57 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  58 {
  59         struct pool_info *pi = data;
  60         r1bio_t *r1_bio;
  61         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
  62
  63         /* allocate a r1bio with room for raid_disks entries in the bios array */
  64         r1_bio = kmalloc(size, gfp_flags);
  65         if (r1_bio)
  66                 memset(r1_bio, 0, size);
  67         else
  68                 unplug_slaves(pi->mddev);
  69
  70         return r1_bio;
  71 }
  72
  73 static void r1bio_pool_free(void *r1_bio, void *data)
  74 {
  75         kfree(r1_bio);
  76 }
  77
  78 #define RESYNC_BLOCK_SIZE (64*1024)
  79 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  80 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  81 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  82 #define RESYNC_WINDOW (2048*1024)
  83
  84 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  85 {
  86         struct pool_info *pi = data;
  87         struct page *page;
  88         r1bio_t *r1_bio;
  89         struct bio *bio;
  90         int i, j;
  91
  92         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  93         if (!r1_bio) {
  94                 unplug_slaves(pi->mddev);
  95                 return NULL;
  96         }
  97
  98         /*
  99          * Allocate bios : 1 for reading, n-1 for writing
 100          */
 101         for (j = pi->raid_disks ; j-- ; ) {
 102                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 103                 if (!bio)
 104                         goto out_free_bio;
 105                 r1_bio->bios[j] = bio;
 106         }
 107         /*
 108          * Allocate RESYNC_PAGES data pages and attach them to
 109          * the first bio;
 110          */
 111         bio = r1_bio->bios[0];
 112         for (i = 0; i < RESYNC_PAGES; i++) {
 113                 page = alloc_page(gfp_flags);
 114                 if (unlikely(!page))
 115                         goto out_free_pages;
 116
 117                 bio->bi_io_vec[i].bv_page = page;
 118         }
 119
 120         r1_bio->master_bio = NULL;
 121
 122         return r1_bio;
 123
 124 out_free_pages:
 125         for ( ; i > 0 ; i--)
 126                 __free_page(bio->bi_io_vec[i-1].bv_page);
 127 out_free_bio:
 128         while ( ++j < pi->raid_disks )
 129                 bio_put(r1_bio->bios[j]);
 130         r1bio_pool_free(r1_bio, data);
 131         return NULL;
 132 }
 133
 134 static void r1buf_pool_free(void *__r1_bio, void *data)
 135 {
 136         struct pool_info *pi = data;
 137         int i;
 138         r1bio_t *r1bio = __r1_bio;
 139         struct bio *bio = r1bio->bios[0];
 140
 141         for (i = 0; i < RESYNC_PAGES; i++) {
 142                 __free_page(bio->bi_io_vec[i].bv_page);
 143                 bio->bi_io_vec[i].bv_page = NULL;
 144         }
 145         for (i=0 ; i < pi->raid_disks; i++)
 146                 bio_put(r1bio->bios[i]);
 147
 148         r1bio_pool_free(r1bio, data);
 149 }
 150
 151 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 152 {
 153         int i;
 154
 155         for (i = 0; i < conf->raid_disks; i++) {
 156                 struct bio **bio = r1_bio->bios + i;
 157                 if (*bio)
 158                         bio_put(*bio);
 159                 *bio = NULL;
 160         }
 161 }
 162
 163 static inline void free_r1bio(r1bio_t *r1_bio)
 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         allow_barrier(conf);
 172
 173         put_all_bios(conf, r1_bio);
 174         mempool_free(r1_bio, conf->r1bio_pool);
 175 }
 176
 177 static inline void put_buf(r1bio_t *r1_bio)
 178 {
 179         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 180
 181         mempool_free(r1_bio, conf->r1buf_pool);
 182
 183         lower_barrier(conf);
 184 }
 185
 186 static void reschedule_retry(r1bio_t *r1_bio)
 187 {
 188         unsigned long flags;
 189         mddev_t *mddev = r1_bio->mddev;
 190         conf_t *conf = mddev_to_conf(mddev);
 191
 192         spin_lock_irqsave(&conf->device_lock, flags);
 193         list_add(&r1_bio->retry_list, &conf->retry_list);
 194         spin_unlock_irqrestore(&conf->device_lock, flags);
 195
 196         wake_up(&conf->wait_barrier);
 197         md_wakeup_thread(mddev->thread);
 198 }
 199
 200 /*
 201  * raid_end_bio_io() is called when we have finished servicing a mirrored
 202  * operation and are ready to return a success/failure code to the buffer
 203  * cache layer.
 204  */
 205 static void raid_end_bio_io(r1bio_t *r1_bio)
 206 {
 207         struct bio *bio = r1_bio->master_bio;
 208
 209         /* if nobody has done the final endio yet, do it now */
 210         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 211                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
 212                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 213                         (unsigned long long) bio->bi_sector,
 214                         (unsigned long long) bio->bi_sector +
 215                                 (bio->bi_size >> 9) - 1);
 216
 217                 bio_endio(bio, bio->bi_size,
 218                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 219         }
 220         free_r1bio(r1_bio);
 221 }
 222
 223 /*
 224  * Update disk head position estimator based on IRQ completion info.
 225  */
 226 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 227 {
 228         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 229
 230         conf->mirrors[disk].head_position =
 231                 r1_bio->sector + (r1_bio->sectors);
 232 }
 233
 234 static int raid1_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
 235 {
 236         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 237         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 238         int mirror;
 239         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 240
 241         if (bio->bi_size)
 242                 return 1;
 243
 244         mirror = r1_bio->read_disk;
 245         /*
 246          * this branch is our 'one mirror IO has finished' event handler:
 247          */
 248         if (!uptodate)
 249                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 250         else
 251                 /*
 252                  * Set R1BIO_Uptodate in our master bio, so that
 253                  * we will return a good error code for to the higher
 254                  * levels even if IO on some other mirrored buffer fails.
 255                  *
 256                  * The 'master' represents the composite IO operation to
 257                  * user-side. So if something waits for IO, then it will
 258                  * wait for the 'master' bio.
 259                  */
 260                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 261
 262         update_head_pos(mirror, r1_bio);
 263
 264         /*
 265          * we have only one bio on the read side
 266          */
 267         if (uptodate)
 268                 raid_end_bio_io(r1_bio);
 269         else {
 270                 /*
 271                  * oops, read error:
 272                  */
 273                 char b[BDEVNAME_SIZE];
 274                 if (printk_ratelimit())
 275                         printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
 276                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 277                 reschedule_retry(r1_bio);
 278         }
 279
 280         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 281         return 0;
 282 }
 283
 284 static int raid1_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
 285 {
 286         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 287         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 288         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 289         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 290
 291         if (bio->bi_size)
 292                 return 1;
 293
 294         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 295                 if (r1_bio->bios[mirror] == bio)
 296                         break;
 297
 298         if (error == -ENOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
 299                 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
 300                 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
 301                 r1_bio->mddev->barriers_work = 0;
 302         } else {
 303                 /*
 304                  * this branch is our 'one mirror IO has finished' event handler:
 305                  */
 306                 r1_bio->bios[mirror] = NULL;
 307                 if (!uptodate) {
 308                         md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 309                         /* an I/O failed, we can't clear the bitmap */
 310                         set_bit(R1BIO_Degraded, &r1_bio->state);
 311                 } else
 312                         /*
 313                          * Set R1BIO_Uptodate in our master bio, so that
 314                          * we will return a good error code for to the higher
 315                          * levels even if IO on some other mirrored buffer fails.
 316                          *
 317                          * The 'master' represents the composite IO operation to
 318                          * user-side. So if something waits for IO, then it will
 319                          * wait for the 'master' bio.
 320                          */
 321                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 322
 323                 update_head_pos(mirror, r1_bio);
 324
 325                 if (behind) {
 326                         if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 327                                 atomic_dec(&r1_bio->behind_remaining);
 328
 329                         /* In behind mode, we ACK the master bio once the I/O has safely
 330                          * reached all non-writemostly disks. Setting the Returned bit
 331                          * ensures that this gets done only once -- we don't ever want to
 332                          * return -EIO here, instead we'll wait */
 333
 334                         if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 335                             test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 336                                 /* Maybe we can return now */
 337                                 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 338                                         struct bio *mbio = r1_bio->master_bio;
 339                                         PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 340                                                (unsigned long long) mbio->bi_sector,
 341                                                (unsigned long long) mbio->bi_sector +
 342                                                (mbio->bi_size >> 9) - 1);
 343                                         bio_endio(mbio, mbio->bi_size, 0);
 344                                 }
 345                         }
 346                 }
 347         }
 348         /*
 349          *
 350          * Let's see if all mirrored write operations have finished
 351          * already.
 352          */
 353         if (atomic_dec_and_test(&r1_bio->remaining)) {
 354                 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
 355                         reschedule_retry(r1_bio);
 356                         /* Don't dec_pending yet, we want to hold
 357                          * the reference over the retry
 358                          */
 359                         return 0;
 360                 }
 361                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 362                         /* free extra copy of the data pages */
 363                         int i = bio->bi_vcnt;
 364                         while (i--)
 365                                 __free_page(bio->bi_io_vec[i].bv_page);
 366                 }
 367                 /* clear the bitmap if all writes complete successfully */
 368                 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 369                                 r1_bio->sectors,
 370                                 !test_bit(R1BIO_Degraded, &r1_bio->state),
 371                                 behind);
 372                 md_write_end(r1_bio->mddev);
 373                 raid_end_bio_io(r1_bio);
 374         }
 375
 376         if (r1_bio->bios[mirror]==NULL)
 377                 bio_put(bio);
 378
 379         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 380         return 0;
 381 }
 382
 383
 384 /*
 385  * This routine returns the disk from which the requested read should
 386  * be done. There is a per-array 'next expected sequential IO' sector
 387  * number - if this matches on the next IO then we use the last disk.
 388  * There is also a per-disk 'last know head position' sector that is
 389  * maintained from IRQ contexts, both the normal and the resync IO
 390  * completion handlers update this position correctly. If there is no
 391  * perfect sequential match then we pick the disk whose head is closest.
 392  *
 393  * If there are 2 mirrors in the same 2 devices, performance degrades
 394  * because position is mirror, not device based.
 395  *
 396  * The rdev for the device selected will have nr_pending incremented.
 397  */
 398 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 399 {
 400         const unsigned long this_sector = r1_bio->sector;
 401         int new_disk = conf->last_used, disk = new_disk;
 402         int wonly_disk = -1;
 403         const int sectors = r1_bio->sectors;
 404         sector_t new_distance, current_distance;
 405         mdk_rdev_t *rdev;
 406
 407         rcu_read_lock();
 408         /*
 409          * Check if we can balance. We can balance on the whole
 410          * device if no resync is going on, or below the resync window.
 411          * We take the first readable disk when above the resync window.
 412          */
 413  retry:
 414         if (conf->mddev->recovery_cp < MaxSector &&
 415             (this_sector + sectors >= conf->next_resync)) {
 416                 /* Choose the first operation device, for consistancy */
 417                 new_disk = 0;
 418
 419                 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 420                      !rdev || !test_bit(In_sync, &rdev->flags)
 421                              || test_bit(WriteMostly, &rdev->flags);
 422                      rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
 423
 424                         if (rdev && test_bit(In_sync, &rdev->flags))
 425                                 wonly_disk = new_disk;
 426
 427                         if (new_disk == conf->raid_disks - 1) {
 428                                 new_disk = wonly_disk;
 429                                 break;
 430                         }
 431                 }
 432                 goto rb_out;
 433         }
 434
 435
 436         /* make sure the disk is operational */
 437         for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 438              !rdev || !test_bit(In_sync, &rdev->flags) ||
 439                      test_bit(WriteMostly, &rdev->flags);
 440              rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
 441
 442                 if (rdev && test_bit(In_sync, &rdev->flags))
 443                         wonly_disk = new_disk;
 444
 445                 if (new_disk <= 0)
 446                         new_disk = conf->raid_disks;
 447                 new_disk--;
 448                 if (new_disk == disk) {
 449                         new_disk = wonly_disk;
 450                         break;
 451                 }
 452         }
 453
 454         if (new_disk < 0)
 455                 goto rb_out;
 456
 457         disk = new_disk;
 458         /* now disk == new_disk == starting point for search */
 459
 460         /*
 461          * Don't change to another disk for sequential reads:
 462          */
 463         if (conf->next_seq_sect == this_sector)
 464                 goto rb_out;
 465         if (this_sector == conf->mirrors[new_disk].head_position)
 466                 goto rb_out;
 467
 468         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 469
 470         /* Find the disk whose head is closest */
 471
 472         do {
 473                 if (disk <= 0)
 474                         disk = conf->raid_disks;
 475                 disk--;
 476
 477                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 478
 479                 if (!rdev ||
 480                     !test_bit(In_sync, &rdev->flags) ||
 481                     test_bit(WriteMostly, &rdev->flags))
 482                         continue;
 483
 484                 if (!atomic_read(&rdev->nr_pending)) {
 485                         new_disk = disk;
 486                         break;
 487                 }
 488                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 489                 if (new_distance < current_distance) {
 490                         current_distance = new_distance;
 491                         new_disk = disk;
 492                 }
 493         } while (disk != conf->last_used);
 494
 495  rb_out:
 496
 497
 498         if (new_disk >= 0) {
 499                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 500                 if (!rdev)
 501                         goto retry;
 502                 atomic_inc(&rdev->nr_pending);
 503                 if (!test_bit(In_sync, &rdev->flags)) {
 504                         /* cannot risk returning a device that failed
 505                          * before we inc'ed nr_pending
 506                          */
 507                         atomic_dec(&rdev->nr_pending);
 508                         goto retry;
 509                 }
 510                 conf->next_seq_sect = this_sector + sectors;
 511                 conf->last_used = new_disk;
 512         }
 513         rcu_read_unlock();
 514
 515         return new_disk;
 516 }
 517
 518 static void unplug_slaves(mddev_t *mddev)
 519 {
 520         conf_t *conf = mddev_to_conf(mddev);
 521         int i;
 522
 523         rcu_read_lock();
 524         for (i=0; i<mddev->raid_disks; i++) {
 525                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 526                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 527                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 528
 529                         atomic_inc(&rdev->nr_pending);
 530                         rcu_read_unlock();
 531
 532                         if (r_queue->unplug_fn)
 533                                 r_queue->unplug_fn(r_queue);
 534
 535                         rdev_dec_pending(rdev, mddev);
 536                         rcu_read_lock();
 537                 }
 538         }
 539         rcu_read_unlock();
 540 }
 541
 542 static void raid1_unplug(request_queue_t *q)
 543 {
 544         mddev_t *mddev = q->queuedata;
 545
 546         unplug_slaves(mddev);
 547         md_wakeup_thread(mddev->thread);
 548 }
 549
 550 static int raid1_issue_flush(request_queue_t *q, struct gendisk *disk,
 551                              sector_t *error_sector)
 552 {
 553         mddev_t *mddev = q->queuedata;
 554         conf_t *conf = mddev_to_conf(mddev);
 555         int i, ret = 0;
 556
 557         rcu_read_lock();
 558         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
 559                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 560                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 561                         struct block_device *bdev = rdev->bdev;
 562                         request_queue_t *r_queue = bdev_get_queue(bdev);
 563
 564                         if (!r_queue->issue_flush_fn)
 565                                 ret = -EOPNOTSUPP;
 566                         else {
 567                                 atomic_inc(&rdev->nr_pending);
 568                                 rcu_read_unlock();
 569                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
 570                                                               error_sector);
 571                                 rdev_dec_pending(rdev, mddev);
 572                                 rcu_read_lock();
 573                         }
 574                 }
 575         }
 576         rcu_read_unlock();
 577         return ret;
 578 }
 579
 580 /* Barriers....
 581  * Sometimes we need to suspend IO while we do something else,
 582  * either some resync/recovery, or reconfigure the array.
 583  * To do this we raise a 'barrier'.
 584  * The 'barrier' is a counter that can be raised multiple times
 585  * to count how many activities are happening which preclude
 586  * normal IO.
 587  * We can only raise the barrier if there is no pending IO.
 588  * i.e. if nr_pending == 0.
 589  * We choose only to raise the barrier if no-one is waiting for the
 590  * barrier to go down.  This means that as soon as an IO request
 591  * is ready, no other operations which require a barrier will start
 592  * until the IO request has had a chance.
 593  *
 594  * So: regular IO calls 'wait_barrier'.  When that returns there
 595  *    is no backgroup IO happening,  It must arrange to call
 596  *    allow_barrier when it has finished its IO.
 597  * backgroup IO calls must call raise_barrier.  Once that returns
 598  *    there is no normal IO happeing.  It must arrange to call
 599  *    lower_barrier when the particular background IO completes.
 600  */
 601 #define RESYNC_DEPTH 32
 602
 603 static void raise_barrier(conf_t *conf)
 604 {
 605         spin_lock_irq(&conf->resync_lock);
 606
 607         /* Wait until no block IO is waiting */
 608         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 609                             conf->resync_lock,
 610                             raid1_unplug(conf->mddev->queue));
 611
 612         /* block any new IO from starting */
 613         conf->barrier++;
 614
 615         /* No wait for all pending IO to complete */
 616         wait_event_lock_irq(conf->wait_barrier,
 617                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 618                             conf->resync_lock,
 619                             raid1_unplug(conf->mddev->queue));
 620
 621         spin_unlock_irq(&conf->resync_lock);
 622 }
 623
 624 static void lower_barrier(conf_t *conf)
 625 {
 626         unsigned long flags;
 627         spin_lock_irqsave(&conf->resync_lock, flags);
 628         conf->barrier--;
 629         spin_unlock_irqrestore(&conf->resync_lock, flags);
 630         wake_up(&conf->wait_barrier);
 631 }
 632
 633 static void wait_barrier(conf_t *conf)
 634 {
 635         spin_lock_irq(&conf->resync_lock);
 636         if (conf->barrier) {
 637                 conf->nr_waiting++;
 638                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 639                                     conf->resync_lock,
 640                                     raid1_unplug(conf->mddev->queue));
 641                 conf->nr_waiting--;
 642         }
 643         conf->nr_pending++;
 644         spin_unlock_irq(&conf->resync_lock);
 645 }
 646
 647 static void allow_barrier(conf_t *conf)
 648 {
 649         unsigned long flags;
 650         spin_lock_irqsave(&conf->resync_lock, flags);
 651         conf->nr_pending--;
 652         spin_unlock_irqrestore(&conf->resync_lock, flags);
 653         wake_up(&conf->wait_barrier);
 654 }
 655
 656
 657 /* duplicate the data pages for behind I/O */
 658 static struct page **alloc_behind_pages(struct bio *bio)
 659 {
 660         int i;
 661         struct bio_vec *bvec;
 662         struct page **pages = kmalloc(bio->bi_vcnt * sizeof(struct page *),
 663                                         GFP_NOIO);
 664         if (unlikely(!pages))
 665                 goto do_sync_io;
 666
 667         memset(pages, 0, bio->bi_vcnt * sizeof(struct page *));
 668
 669         bio_for_each_segment(bvec, bio, i) {
 670                 pages[i] = alloc_page(GFP_NOIO);
 671                 if (unlikely(!pages[i]))
 672                         goto do_sync_io;
 673                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 674                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 675                 kunmap(pages[i]);
 676                 kunmap(bvec->bv_page);
 677         }
 678
 679         return pages;
 680
 681 do_sync_io:
 682         if (pages)
 683                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
 684                         __free_page(pages[i]);
 685         kfree(pages);
 686         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 687         return NULL;
 688 }
 689
 690 static int make_request(request_queue_t *q, struct bio * bio)
 691 {
 692         mddev_t *mddev = q->queuedata;
 693         conf_t *conf = mddev_to_conf(mddev);
 694         mirror_info_t *mirror;
 695         r1bio_t *r1_bio;
 696         struct bio *read_bio;
 697         int i, targets = 0, disks;
 698         mdk_rdev_t *rdev;
 699         struct bitmap *bitmap = mddev->bitmap;
 700         unsigned long flags;
 701         struct bio_list bl;
 702         struct page **behind_pages = NULL;
 703         const int rw = bio_data_dir(bio);
 704         int do_barriers;
 705
 706         if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
 707                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
 708                 return 0;
 709         }
 710
 711         /*
 712          * Register the new request and wait if the reconstruction
 713          * thread has put up a bar for new requests.
 714          * Continue immediately if no resync is active currently.
 715          */
 716         md_write_start(mddev, bio); /* wait on superblock update early */
 717
 718         wait_barrier(conf);
 719
 720         disk_stat_inc(mddev->gendisk, ios[rw]);
 721         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
 722
 723         /*
 724          * make_request() can abort the operation when READA is being
 725          * used and no empty request is available.
 726          *
 727          */
 728         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 729
 730         r1_bio->master_bio = bio;
 731         r1_bio->sectors = bio->bi_size >> 9;
 732         r1_bio->state = 0;
 733         r1_bio->mddev = mddev;
 734         r1_bio->sector = bio->bi_sector;
 735
 736         if (rw == READ) {
 737                 /*
 738                  * read balancing logic:
 739                  */
 740                 int rdisk = read_balance(conf, r1_bio);
 741
 742                 if (rdisk < 0) {
 743                         /* couldn't find anywhere to read from */
 744                         raid_end_bio_io(r1_bio);
 745                         return 0;
 746                 }
 747                 mirror = conf->mirrors + rdisk;
 748
 749                 r1_bio->read_disk = rdisk;
 750
 751                 read_bio = bio_clone(bio, GFP_NOIO);
 752
 753                 r1_bio->bios[rdisk] = read_bio;
 754
 755                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 756                 read_bio->bi_bdev = mirror->rdev->bdev;
 757                 read_bio->bi_end_io = raid1_end_read_request;
 758                 read_bio->bi_rw = READ;
 759                 read_bio->bi_private = r1_bio;
 760
 761                 generic_make_request(read_bio);
 762                 return 0;
 763         }
 764
 765         /*
 766          * WRITE:
 767          */
 768         /* first select target devices under spinlock and
 769          * inc refcount on their rdev.  Record them by setting
 770          * bios[x] to bio
 771          */
 772         disks = conf->raid_disks;
 773 #if 0
 774         { static int first=1;
 775         if (first) printk("First Write sector %llu disks %d\n",
 776                           (unsigned long long)r1_bio->sector, disks);
 777         first = 0;
 778         }
 779 #endif
 780         rcu_read_lock();
 781         for (i = 0;  i < disks; i++) {
 782                 if ((rdev=rcu_dereference(conf->mirrors[i].rdev)) != NULL &&
 783                     !test_bit(Faulty, &rdev->flags)) {
 784                         atomic_inc(&rdev->nr_pending);
 785                         if (test_bit(Faulty, &rdev->flags)) {
 786                                 atomic_dec(&rdev->nr_pending);
 787                                 r1_bio->bios[i] = NULL;
 788                         } else
 789                                 r1_bio->bios[i] = bio;
 790                         targets++;
 791                 } else
 792                         r1_bio->bios[i] = NULL;
 793         }
 794         rcu_read_unlock();
 795
 796         BUG_ON(targets == 0); /* we never fail the last device */
 797
 798         if (targets < conf->raid_disks) {
 799                 /* array is degraded, we will not clear the bitmap
 800                  * on I/O completion (see raid1_end_write_request) */
 801                 set_bit(R1BIO_Degraded, &r1_bio->state);
 802         }
 803
 804         /* do behind I/O ? */
 805         if (bitmap &&
 806             atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
 807             (behind_pages = alloc_behind_pages(bio)) != NULL)
 808                 set_bit(R1BIO_BehindIO, &r1_bio->state);
 809
 810         atomic_set(&r1_bio->remaining, 0);
 811         atomic_set(&r1_bio->behind_remaining, 0);
 812
 813         do_barriers = bio->bi_rw & BIO_RW_BARRIER;
 814         if (do_barriers)
 815                 set_bit(R1BIO_Barrier, &r1_bio->state);
 816
 817         bio_list_init(&bl);
 818         for (i = 0; i < disks; i++) {
 819                 struct bio *mbio;
 820                 if (!r1_bio->bios[i])
 821                         continue;
 822
 823                 mbio = bio_clone(bio, GFP_NOIO);
 824                 r1_bio->bios[i] = mbio;
 825
 826                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 827                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 828                 mbio->bi_end_io = raid1_end_write_request;
 829                 mbio->bi_rw = WRITE | do_barriers;
 830                 mbio->bi_private = r1_bio;
 831
 832                 if (behind_pages) {
 833                         struct bio_vec *bvec;
 834                         int j;
 835
 836                         /* Yes, I really want the '__' version so that
 837                          * we clear any unused pointer in the io_vec, rather
 838                          * than leave them unchanged.  This is important
 839                          * because when we come to free the pages, we won't
 840                          * know the originial bi_idx, so we just free
 841                          * them all
 842                          */
 843                         __bio_for_each_segment(bvec, mbio, j, 0)
 844                                 bvec->bv_page = behind_pages[j];
 845                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 846                                 atomic_inc(&r1_bio->behind_remaining);
 847                 }
 848
 849                 atomic_inc(&r1_bio->remaining);
 850
 851                 bio_list_add(&bl, mbio);
 852         }
 853         kfree(behind_pages); /* the behind pages are attached to the bios now */
 854
 855         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
 856                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 857         spin_lock_irqsave(&conf->device_lock, flags);
 858         bio_list_merge(&conf->pending_bio_list, &bl);
 859         bio_list_init(&bl);
 860
 861         blk_plug_device(mddev->queue);
 862         spin_unlock_irqrestore(&conf->device_lock, flags);
 863
 864 #if 0
 865         while ((bio = bio_list_pop(&bl)) != NULL)
 866                 generic_make_request(bio);
 867 #endif
 868
 869         return 0;
 870 }
 871
 872 static void status(struct seq_file *seq, mddev_t *mddev)
 873 {
 874         conf_t *conf = mddev_to_conf(mddev);
 875         int i;
 876
 877         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 878                                                 conf->working_disks);
 879         for (i = 0; i < conf->raid_disks; i++)
 880                 seq_printf(seq, "%s",
 881                               conf->mirrors[i].rdev &&
 882                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
 883         seq_printf(seq, "]");
 884 }
 885
 886
 887 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 888 {
 889         char b[BDEVNAME_SIZE];
 890         conf_t *conf = mddev_to_conf(mddev);
 891
 892         /*
 893          * If it is not operational, then we have already marked it as dead
 894          * else if it is the last working disks, ignore the error, let the
 895          * next level up know.
 896          * else mark the drive as failed
 897          */
 898         if (test_bit(In_sync, &rdev->flags)
 899             && conf->working_disks == 1)
 900                 /*
 901                  * Don't fail the drive, act as though we were just a
 902                  * normal single drive
 903                  */
 904                 return;
 905         if (test_bit(In_sync, &rdev->flags)) {
 906                 mddev->degraded++;
 907                 conf->working_disks--;
 908                 /*
 909                  * if recovery is running, make sure it aborts.
 910                  */
 911                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 912         }
 913         clear_bit(In_sync, &rdev->flags);
 914         set_bit(Faulty, &rdev->flags);
 915         mddev->sb_dirty = 1;
 916         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
 917                 "       Operation continuing on %d devices\n",
 918                 bdevname(rdev->bdev,b), conf->working_disks);
 919 }
 920
 921 static void print_conf(conf_t *conf)
 922 {
 923         int i;
 924         mirror_info_t *tmp;
 925
 926         printk("RAID1 conf printout:\n");
 927         if (!conf) {
 928                 printk("(!conf)\n");
 929                 return;
 930         }
 931         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 932                 conf->raid_disks);
 933
 934         for (i = 0; i < conf->raid_disks; i++) {
 935                 char b[BDEVNAME_SIZE];
 936                 tmp = conf->mirrors + i;
 937                 if (tmp->rdev)
 938                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
 939                                 i, !test_bit(In_sync, &tmp->rdev->flags), !test_bit(Faulty, &tmp->rdev->flags),
 940                                 bdevname(tmp->rdev->bdev,b));
 941         }
 942 }
 943
 944 static void close_sync(conf_t *conf)
 945 {
 946         wait_barrier(conf);
 947         allow_barrier(conf);
 948
 949         mempool_destroy(conf->r1buf_pool);
 950         conf->r1buf_pool = NULL;
 951 }
 952
 953 static int raid1_spare_active(mddev_t *mddev)
 954 {
 955         int i;
 956         conf_t *conf = mddev->private;
 957         mirror_info_t *tmp;
 958
 959         /*
 960          * Find all failed disks within the RAID1 configuration
 961          * and mark them readable
 962          */
 963         for (i = 0; i < conf->raid_disks; i++) {
 964                 tmp = conf->mirrors + i;
 965                 if (tmp->rdev
 966                     && !test_bit(Faulty, &tmp->rdev->flags)
 967                     && !test_bit(In_sync, &tmp->rdev->flags)) {
 968                         conf->working_disks++;
 969                         mddev->degraded--;
 970                         set_bit(In_sync, &tmp->rdev->flags);
 971                 }
 972         }
 973
 974         print_conf(conf);
 975         return 0;
 976 }
 977
 978
 979 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 980 {
 981         conf_t *conf = mddev->private;
 982         int found = 0;
 983         int mirror = 0;
 984         mirror_info_t *p;
 985
 986         for (mirror=0; mirror < mddev->raid_disks; mirror++)
 987                 if ( !(p=conf->mirrors+mirror)->rdev) {
 988
 989                         blk_queue_stack_limits(mddev->queue,
 990                                                rdev->bdev->bd_disk->queue);
 991                         /* as we don't honour merge_bvec_fn, we must never risk
 992                          * violating it, so limit ->max_sector to one PAGE, as
 993                          * a one page request is never in violation.
 994                          */
 995                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 996                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
 997                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
 998
 999                         p->head_position = 0;
1000                         rdev->raid_disk = mirror;
1001                         found = 1;
1002                         /* As all devices are equivalent, we don't need a full recovery
1003                          * if this was recently any drive of the array
1004                          */
1005                         if (rdev->saved_raid_disk < 0)
1006                                 conf->fullsync = 1;
1007                         rcu_assign_pointer(p->rdev, rdev);
1008                         break;
1009                 }
1010
1011         print_conf(conf);
1012         return found;
1013 }
1014
1015 static int raid1_remove_disk(mddev_t *mddev, int number)
1016 {
1017         conf_t *conf = mddev->private;
1018         int err = 0;
1019         mdk_rdev_t *rdev;
1020         mirror_info_t *p = conf->mirrors+ number;
1021
1022         print_conf(conf);
1023         rdev = p->rdev;
1024         if (rdev) {
1025                 if (test_bit(In_sync, &rdev->flags) ||
1026                     atomic_read(&rdev->nr_pending)) {
1027                         err = -EBUSY;
1028                         goto abort;
1029                 }
1030                 p->rdev = NULL;
1031                 synchronize_rcu();
1032                 if (atomic_read(&rdev->nr_pending)) {
1033                         /* lost the race, try later */
1034                         err = -EBUSY;
1035                         p->rdev = rdev;
1036                 }
1037         }
1038 abort:
1039
1040         print_conf(conf);
1041         return err;
1042 }
1043
1044
1045 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1046 {
1047         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1048         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1049         conf_t *conf = mddev_to_conf(r1_bio->mddev);
1050
1051         if (bio->bi_size)
1052                 return 1;
1053
1054         if (r1_bio->bios[r1_bio->read_disk] != bio)
1055                 BUG();
1056         update_head_pos(r1_bio->read_disk, r1_bio);
1057         /*
1058          * we have read a block, now it needs to be re-written,
1059          * or re-read if the read failed.
1060          * We don't do much here, just schedule handling by raid1d
1061          */
1062         if (!uptodate) {
1063                 md_error(r1_bio->mddev,
1064                          conf->mirrors[r1_bio->read_disk].rdev);
1065         } else
1066                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1067         rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
1068         reschedule_retry(r1_bio);
1069         return 0;
1070 }
1071
1072 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1073 {
1074         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1075         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1076         mddev_t *mddev = r1_bio->mddev;
1077         conf_t *conf = mddev_to_conf(mddev);
1078         int i;
1079         int mirror=0;
1080
1081         if (bio->bi_size)
1082                 return 1;
1083
1084         for (i = 0; i < conf->raid_disks; i++)
1085                 if (r1_bio->bios[i] == bio) {
1086                         mirror = i;
1087                         break;
1088                 }
1089         if (!uptodate)
1090                 md_error(mddev, conf->mirrors[mirror].rdev);
1091
1092         update_head_pos(mirror, r1_bio);
1093
1094         if (atomic_dec_and_test(&r1_bio->remaining)) {
1095                 md_done_sync(mddev, r1_bio->sectors, uptodate);
1096                 put_buf(r1_bio);
1097         }
1098         rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
1099         return 0;
1100 }
1101
1102 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1103 {
1104         conf_t *conf = mddev_to_conf(mddev);
1105         int i;
1106         int disks = conf->raid_disks;
1107         struct bio *bio, *wbio;
1108
1109         bio = r1_bio->bios[r1_bio->read_disk];
1110
1111 /*
1112         if (r1_bio->sector == 0) printk("First sync write startss\n");
1113 */
1114         /*
1115          * schedule writes
1116          */
1117         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1118                 /*
1119                  * There is no point trying a read-for-reconstruct as
1120                  * reconstruct is about to be aborted
1121                  */
1122                 char b[BDEVNAME_SIZE];
1123                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1124                         " for block %llu\n",
1125                         bdevname(bio->bi_bdev,b),
1126                         (unsigned long long)r1_bio->sector);
1127                 md_done_sync(mddev, r1_bio->sectors, 0);
1128                 put_buf(r1_bio);
1129                 return;
1130         }
1131
1132         atomic_set(&r1_bio->remaining, 1);
1133         for (i = 0; i < disks ; i++) {
1134                 wbio = r1_bio->bios[i];
1135                 if (wbio->bi_end_io != end_sync_write)
1136                         continue;
1137
1138                 atomic_inc(&conf->mirrors[i].rdev->nr_pending);
1139                 atomic_inc(&r1_bio->remaining);
1140                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1141
1142                 generic_make_request(wbio);
1143         }
1144
1145         if (atomic_dec_and_test(&r1_bio->remaining)) {
1146                 /* if we're here, all write(s) have completed, so clean up */
1147                 md_done_sync(mddev, r1_bio->sectors, 1);
1148                 put_buf(r1_bio);
1149         }
1150 }
1151
1152 /*
1153  * This is a kernel thread which:
1154  *
1155  *      1.      Retries failed read operations on working mirrors.
1156  *      2.      Updates the raid superblock when problems encounter.
1157  *      3.      Performs writes following reads for array syncronising.
1158  */
1159
1160 static void raid1d(mddev_t *mddev)
1161 {
1162         r1bio_t *r1_bio;
1163         struct bio *bio;
1164         unsigned long flags;
1165         conf_t *conf = mddev_to_conf(mddev);
1166         struct list_head *head = &conf->retry_list;
1167         int unplug=0;
1168         mdk_rdev_t *rdev;
1169
1170         md_check_recovery(mddev);
1171
1172         for (;;) {
1173                 char b[BDEVNAME_SIZE];
1174                 spin_lock_irqsave(&conf->device_lock, flags);
1175
1176                 if (conf->pending_bio_list.head) {
1177                         bio = bio_list_get(&conf->pending_bio_list);
1178                         blk_remove_plug(mddev->queue);
1179                         spin_unlock_irqrestore(&conf->device_lock, flags);
1180                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1181                         if (bitmap_unplug(mddev->bitmap) != 0)
1182                                 printk("%s: bitmap file write failed!\n", mdname(mddev));
1183
1184                         while (bio) { /* submit pending writes */
1185                                 struct bio *next = bio->bi_next;
1186                                 bio->bi_next = NULL;
1187                                 generic_make_request(bio);
1188                                 bio = next;
1189                         }
1190                         unplug = 1;
1191
1192                         continue;
1193                 }
1194
1195                 if (list_empty(head))
1196                         break;
1197                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1198                 list_del(head->prev);
1199                 spin_unlock_irqrestore(&conf->device_lock, flags);
1200
1201                 mddev = r1_bio->mddev;
1202                 conf = mddev_to_conf(mddev);
1203                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1204                         sync_request_write(mddev, r1_bio);
1205                         unplug = 1;
1206                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1207                         /* some requests in the r1bio were BIO_RW_BARRIER
1208                          * requests which failed with -ENOTSUPP.  Hohumm..
1209                          * Better resubmit without the barrier.
1210                          * We know which devices to resubmit for, because
1211                          * all others have had their bios[] entry cleared.
1212                          */
1213                         int i;
1214                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1215                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1216                         for (i=0; i < conf->raid_disks; i++)
1217                                 if (r1_bio->bios[i]) {
1218                                         struct bio_vec *bvec;
1219                                         int j;
1220
1221                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1222                                         /* copy pages from the failed bio, as
1223                                          * this might be a write-behind device */
1224                                         __bio_for_each_segment(bvec, bio, j, 0)
1225                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1226                                         bio_put(r1_bio->bios[i]);
1227                                         bio->bi_sector = r1_bio->sector +
1228                                                 conf->mirrors[i].rdev->data_offset;
1229                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1230                                         bio->bi_end_io = raid1_end_write_request;
1231                                         bio->bi_rw = WRITE;
1232                                         bio->bi_private = r1_bio;
1233                                         r1_bio->bios[i] = bio;
1234                                         generic_make_request(bio);
1235                                 }
1236                 } else {
1237                         int disk;
1238                         bio = r1_bio->bios[r1_bio->read_disk];
1239                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1240                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1241                                        " read error for block %llu\n",
1242                                        bdevname(bio->bi_bdev,b),
1243                                        (unsigned long long)r1_bio->sector);
1244                                 raid_end_bio_io(r1_bio);
1245                         } else {
1246                                 r1_bio->bios[r1_bio->read_disk] = NULL;
1247                                 r1_bio->read_disk = disk;
1248                                 bio_put(bio);
1249                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1250                                 r1_bio->bios[r1_bio->read_disk] = bio;
1251                                 rdev = conf->mirrors[disk].rdev;
1252                                 if (printk_ratelimit())
1253                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1254                                                " another mirror\n",
1255                                                bdevname(rdev->bdev,b),
1256                                                (unsigned long long)r1_bio->sector);
1257                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1258                                 bio->bi_bdev = rdev->bdev;
1259                                 bio->bi_end_io = raid1_end_read_request;
1260                                 bio->bi_rw = READ;
1261                                 bio->bi_private = r1_bio;
1262                                 unplug = 1;
1263                                 generic_make_request(bio);
1264                         }
1265                 }
1266         }
1267         spin_unlock_irqrestore(&conf->device_lock, flags);
1268         if (unplug)
1269                 unplug_slaves(mddev);
1270 }
1271
1272
1273 static int init_resync(conf_t *conf)
1274 {
1275         int buffs;
1276
1277         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1278         if (conf->r1buf_pool)
1279                 BUG();
1280         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1281                                           conf->poolinfo);
1282         if (!conf->r1buf_pool)
1283                 return -ENOMEM;
1284         conf->next_resync = 0;
1285         return 0;
1286 }
1287
1288 /*
1289  * perform a "sync" on one "block"
1290  *
1291  * We need to make sure that no normal I/O request - particularly write
1292  * requests - conflict with active sync requests.
1293  *
1294  * This is achieved by tracking pending requests and a 'barrier' concept
1295  * that can be installed to exclude normal IO requests.
1296  */
1297
1298 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1299 {
1300         conf_t *conf = mddev_to_conf(mddev);
1301         mirror_info_t *mirror;
1302         r1bio_t *r1_bio;
1303         struct bio *bio;
1304         sector_t max_sector, nr_sectors;
1305         int disk;
1306         int i;
1307         int wonly;
1308         int write_targets = 0;
1309         int sync_blocks;
1310         int still_degraded = 0;
1311
1312         if (!conf->r1buf_pool)
1313         {
1314 /*
1315                 printk("sync start - bitmap %p\n", mddev->bitmap);
1316 */
1317                 if (init_resync(conf))
1318                         return 0;
1319         }
1320
1321         max_sector = mddev->size << 1;
1322         if (sector_nr >= max_sector) {
1323                 /* If we aborted, we need to abort the
1324                  * sync on the 'current' bitmap chunk (there will
1325                  * only be one in raid1 resync.
1326                  * We can find the current addess in mddev->curr_resync
1327                  */
1328                 if (mddev->curr_resync < max_sector) /* aborted */
1329                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1330                                                 &sync_blocks, 1);
1331                 else /* completed sync */
1332                         conf->fullsync = 0;
1333
1334                 bitmap_close_sync(mddev->bitmap);
1335                 close_sync(conf);
1336                 return 0;
1337         }
1338
1339         /* before building a request, check if we can skip these blocks..
1340          * This call the bitmap_start_sync doesn't actually record anything
1341          */
1342         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1343             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1344                 /* We can skip this block, and probably several more */
1345                 *skipped = 1;
1346                 return sync_blocks;
1347         }
1348         /*
1349          * If there is non-resync activity waiting for a turn,
1350          * and resync is going fast enough,
1351          * then let it though before starting on this new sync request.
1352          */
1353         if (!go_faster && conf->nr_waiting)
1354                 msleep_interruptible(1000);
1355
1356         raise_barrier(conf);
1357
1358         conf->next_resync = sector_nr;
1359
1360         /*
1361          * If reconstructing, and >1 working disc,
1362          * could dedicate one to rebuild and others to
1363          * service read requests ..
1364          */
1365         disk = conf->last_used;
1366         /* make sure disk is operational */
1367         wonly = disk;
1368         while (conf->mirrors[disk].rdev == NULL ||
1369                !test_bit(In_sync, &conf->mirrors[disk].rdev->flags) ||
1370                test_bit(WriteMostly, &conf->mirrors[disk].rdev->flags)
1371                 ) {
1372                 if (conf->mirrors[disk].rdev  &&
1373                     test_bit(In_sync, &conf->mirrors[disk].rdev->flags))
1374                         wonly = disk;
1375                 if (disk <= 0)
1376                         disk = conf->raid_disks;
1377                 disk--;
1378                 if (disk == conf->last_used) {
1379                         disk = wonly;
1380                         break;
1381                 }
1382         }
1383         conf->last_used = disk;
1384         atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
1385
1386
1387         mirror = conf->mirrors + disk;
1388
1389         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1390
1391         r1_bio->mddev = mddev;
1392         r1_bio->sector = sector_nr;
1393         r1_bio->state = 0;
1394         set_bit(R1BIO_IsSync, &r1_bio->state);
1395         r1_bio->read_disk = disk;
1396
1397         for (i=0; i < conf->raid_disks; i++) {
1398                 bio = r1_bio->bios[i];
1399
1400                 /* take from bio_init */
1401                 bio->bi_next = NULL;
1402                 bio->bi_flags |= 1 << BIO_UPTODATE;
1403                 bio->bi_rw = 0;
1404                 bio->bi_vcnt = 0;
1405                 bio->bi_idx = 0;
1406                 bio->bi_phys_segments = 0;
1407                 bio->bi_hw_segments = 0;
1408                 bio->bi_size = 0;
1409                 bio->bi_end_io = NULL;
1410                 bio->bi_private = NULL;
1411
1412                 if (i == disk) {
1413                         bio->bi_rw = READ;
1414                         bio->bi_end_io = end_sync_read;
1415                 } else if (conf->mirrors[i].rdev == NULL ||
1416                            test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
1417                         still_degraded = 1;
1418                         continue;
1419                 } else if (!test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
1420                            sector_nr + RESYNC_SECTORS > mddev->recovery_cp   ||
1421                            test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1422                         bio->bi_rw = WRITE;
1423                         bio->bi_end_io = end_sync_write;
1424                         write_targets ++;
1425                 } else
1426                         /* no need to read or write here */
1427                         continue;
1428                 bio->bi_sector = sector_nr + conf->mirrors[i].rdev->data_offset;
1429                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1430                 bio->bi_private = r1_bio;
1431         }
1432
1433         if (write_targets == 0) {
1434                 /* There is nowhere to write, so all non-sync
1435                  * drives must be failed - so we are finished
1436                  */
1437                 sector_t rv = max_sector - sector_nr;
1438                 *skipped = 1;
1439                 put_buf(r1_bio);
1440                 rdev_dec_pending(conf->mirrors[disk].rdev, mddev);
1441                 return rv;
1442         }
1443
1444         nr_sectors = 0;
1445         sync_blocks = 0;
1446         do {
1447                 struct page *page;
1448                 int len = PAGE_SIZE;
1449                 if (sector_nr + (len>>9) > max_sector)
1450                         len = (max_sector - sector_nr) << 9;
1451                 if (len == 0)
1452                         break;
1453                 if (sync_blocks == 0) {
1454                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1455                                                &sync_blocks, still_degraded) &&
1456                             !conf->fullsync &&
1457                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1458                                 break;
1459                         if (sync_blocks < (PAGE_SIZE>>9))
1460                                 BUG();
1461                         if (len > (sync_blocks<<9))
1462                                 len = sync_blocks<<9;
1463                 }
1464
1465                 for (i=0 ; i < conf->raid_disks; i++) {
1466                         bio = r1_bio->bios[i];
1467                         if (bio->bi_end_io) {
1468                                 page = r1_bio->bios[0]->bi_io_vec[bio->bi_vcnt].bv_page;
1469                                 if (bio_add_page(bio, page, len, 0) == 0) {
1470                                         /* stop here */
1471                                         r1_bio->bios[0]->bi_io_vec[bio->bi_vcnt].bv_page = page;
1472                                         while (i > 0) {
1473                                                 i--;
1474                                                 bio = r1_bio->bios[i];
1475                                                 if (bio->bi_end_io==NULL)
1476                                                         continue;
1477                                                 /* remove last page from this bio */
1478                                                 bio->bi_vcnt--;
1479                                                 bio->bi_size -= len;
1480                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1481                                         }
1482                                         goto bio_full;
1483                                 }
1484                         }
1485                 }
1486                 nr_sectors += len>>9;
1487                 sector_nr += len>>9;
1488                 sync_blocks -= (len>>9);
1489         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1490  bio_full:
1491         bio = r1_bio->bios[disk];
1492         r1_bio->sectors = nr_sectors;
1493
1494         md_sync_acct(mirror->rdev->bdev, nr_sectors);
1495
1496         generic_make_request(bio);
1497
1498         return nr_sectors;
1499 }
1500
1501 static int run(mddev_t *mddev)
1502 {
1503         conf_t *conf;
1504         int i, j, disk_idx;
1505         mirror_info_t *disk;
1506         mdk_rdev_t *rdev;
1507         struct list_head *tmp;
1508
1509         if (mddev->level != 1) {
1510                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1511                        mdname(mddev), mddev->level);
1512                 goto out;
1513         }
1514         /*
1515          * copy the already verified devices into our private RAID1
1516          * bookkeeping area. [whatever we allocate in run(),
1517          * should be freed in stop()]
1518          */
1519         conf = kmalloc(sizeof(conf_t), GFP_KERNEL);
1520         mddev->private = conf;
1521         if (!conf)
1522                 goto out_no_mem;
1523
1524         memset(conf, 0, sizeof(*conf));
1525         conf->mirrors = kmalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1526                                  GFP_KERNEL);
1527         if (!conf->mirrors)
1528                 goto out_no_mem;
1529
1530         memset(conf->mirrors, 0, sizeof(struct mirror_info)*mddev->raid_disks);
1531
1532         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1533         if (!conf->poolinfo)
1534                 goto out_no_mem;
1535         conf->poolinfo->mddev = mddev;
1536         conf->poolinfo->raid_disks = mddev->raid_disks;
1537         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1538                                           r1bio_pool_free,
1539                                           conf->poolinfo);
1540         if (!conf->r1bio_pool)
1541                 goto out_no_mem;
1542
1543         ITERATE_RDEV(mddev, rdev, tmp) {
1544                 disk_idx = rdev->raid_disk;
1545                 if (disk_idx >= mddev->raid_disks
1546                     || disk_idx < 0)
1547                         continue;
1548                 disk = conf->mirrors + disk_idx;
1549
1550                 disk->rdev = rdev;
1551
1552                 blk_queue_stack_limits(mddev->queue,
1553                                        rdev->bdev->bd_disk->queue);
1554                 /* as we don't honour merge_bvec_fn, we must never risk
1555                  * violating it, so limit ->max_sector to one PAGE, as
1556                  * a one page request is never in violation.
1557                  */
1558                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1559                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1560                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1561
1562                 disk->head_position = 0;
1563                 if (!test_bit(Faulty, &rdev->flags) && test_bit(In_sync, &rdev->flags))
1564                         conf->working_disks++;
1565         }
1566         conf->raid_disks = mddev->raid_disks;
1567         conf->mddev = mddev;
1568         spin_lock_init(&conf->device_lock);
1569         INIT_LIST_HEAD(&conf->retry_list);
1570         if (conf->working_disks == 1)
1571                 mddev->recovery_cp = MaxSector;
1572
1573         spin_lock_init(&conf->resync_lock);
1574         init_waitqueue_head(&conf->wait_barrier);
1575
1576         bio_list_init(&conf->pending_bio_list);
1577         bio_list_init(&conf->flushing_bio_list);
1578
1579         if (!conf->working_disks) {
1580                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1581                         mdname(mddev));
1582                 goto out_free_conf;
1583         }
1584
1585         mddev->degraded = 0;
1586         for (i = 0; i < conf->raid_disks; i++) {
1587
1588                 disk = conf->mirrors + i;
1589
1590                 if (!disk->rdev) {
1591                         disk->head_position = 0;
1592                         mddev->degraded++;
1593                 }
1594         }
1595
1596         /*
1597          * find the first working one and use it as a starting point
1598          * to read balancing.
1599          */
1600         for (j = 0; j < conf->raid_disks &&
1601                      (!conf->mirrors[j].rdev ||
1602                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
1603                 /* nothing */;
1604         conf->last_used = j;
1605
1606
1607         mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1608         if (!mddev->thread) {
1609                 printk(KERN_ERR
1610                        "raid1: couldn't allocate thread for %s\n",
1611                        mdname(mddev));
1612                 goto out_free_conf;
1613         }
1614         if (mddev->bitmap) mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
1615
1616         printk(KERN_INFO
1617                 "raid1: raid set %s active with %d out of %d mirrors\n",
1618                 mdname(mddev), mddev->raid_disks - mddev->degraded,
1619                 mddev->raid_disks);
1620         /*
1621          * Ok, everything is just fine now
1622          */
1623         mddev->array_size = mddev->size;
1624
1625         mddev->queue->unplug_fn = raid1_unplug;
1626         mddev->queue->issue_flush_fn = raid1_issue_flush;
1627
1628         return 0;
1629
1630 out_no_mem:
1631         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1632                mdname(mddev));
1633
1634 out_free_conf:
1635         if (conf) {
1636                 if (conf->r1bio_pool)
1637                         mempool_destroy(conf->r1bio_pool);
1638                 kfree(conf->mirrors);
1639                 kfree(conf->poolinfo);
1640                 kfree(conf);
1641                 mddev->private = NULL;
1642         }
1643 out:
1644         return -EIO;
1645 }
1646
1647 static int stop(mddev_t *mddev)
1648 {
1649         conf_t *conf = mddev_to_conf(mddev);
1650         struct bitmap *bitmap = mddev->bitmap;
1651         int behind_wait = 0;
1652
1653         /* wait for behind writes to complete */
1654         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
1655                 behind_wait++;
1656                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
1657                 set_current_state(TASK_UNINTERRUPTIBLE);
1658                 schedule_timeout(HZ); /* wait a second */
1659                 /* need to kick something here to make sure I/O goes? */
1660         }
1661
1662         md_unregister_thread(mddev->thread);
1663         mddev->thread = NULL;
1664         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1665         if (conf->r1bio_pool)
1666                 mempool_destroy(conf->r1bio_pool);
1667         kfree(conf->mirrors);
1668         kfree(conf->poolinfo);
1669         kfree(conf);
1670         mddev->private = NULL;
1671         return 0;
1672 }
1673
1674 static int raid1_resize(mddev_t *mddev, sector_t sectors)
1675 {
1676         /* no resync is happening, and there is enough space
1677          * on all devices, so we can resize.
1678          * We need to make sure resync covers any new space.
1679          * If the array is shrinking we should possibly wait until
1680          * any io in the removed space completes, but it hardly seems
1681          * worth it.
1682          */
1683         mddev->array_size = sectors>>1;
1684         set_capacity(mddev->gendisk, mddev->array_size << 1);
1685         mddev->changed = 1;
1686         if (mddev->array_size > mddev->size && mddev->recovery_cp == MaxSector) {
1687                 mddev->recovery_cp = mddev->size << 1;
1688                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1689         }
1690         mddev->size = mddev->array_size;
1691         mddev->resync_max_sectors = sectors;
1692         return 0;
1693 }
1694
1695 static int raid1_reshape(mddev_t *mddev, int raid_disks)
1696 {
1697         /* We need to:
1698          * 1/ resize the r1bio_pool
1699          * 2/ resize conf->mirrors
1700          *
1701          * We allocate a new r1bio_pool if we can.
1702          * Then raise a device barrier and wait until all IO stops.
1703          * Then resize conf->mirrors and swap in the new r1bio pool.
1704          *
1705          * At the same time, we "pack" the devices so that all the missing
1706          * devices have the higher raid_disk numbers.
1707          */
1708         mempool_t *newpool, *oldpool;
1709         struct pool_info *newpoolinfo;
1710         mirror_info_t *newmirrors;
1711         conf_t *conf = mddev_to_conf(mddev);
1712         int cnt;
1713
1714         int d, d2;
1715
1716         if (raid_disks < conf->raid_disks) {
1717                 cnt=0;
1718                 for (d= 0; d < conf->raid_disks; d++)
1719                         if (conf->mirrors[d].rdev)
1720                                 cnt++;
1721                 if (cnt > raid_disks)
1722                         return -EBUSY;
1723         }
1724
1725         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
1726         if (!newpoolinfo)
1727                 return -ENOMEM;
1728         newpoolinfo->mddev = mddev;
1729         newpoolinfo->raid_disks = raid_disks;
1730
1731         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1732                                  r1bio_pool_free, newpoolinfo);
1733         if (!newpool) {
1734                 kfree(newpoolinfo);
1735                 return -ENOMEM;
1736         }
1737         newmirrors = kmalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
1738         if (!newmirrors) {
1739                 kfree(newpoolinfo);
1740                 mempool_destroy(newpool);
1741                 return -ENOMEM;
1742         }
1743         memset(newmirrors, 0, sizeof(struct mirror_info)*raid_disks);
1744
1745         raise_barrier(conf);
1746
1747         /* ok, everything is stopped */
1748         oldpool = conf->r1bio_pool;
1749         conf->r1bio_pool = newpool;
1750
1751         for (d=d2=0; d < conf->raid_disks; d++)
1752                 if (conf->mirrors[d].rdev) {
1753                         conf->mirrors[d].rdev->raid_disk = d2;
1754                         newmirrors[d2++].rdev = conf->mirrors[d].rdev;
1755                 }
1756         kfree(conf->mirrors);
1757         conf->mirrors = newmirrors;
1758         kfree(conf->poolinfo);
1759         conf->poolinfo = newpoolinfo;
1760
1761         mddev->degraded += (raid_disks - conf->raid_disks);
1762         conf->raid_disks = mddev->raid_disks = raid_disks;
1763
1764         conf->last_used = 0; /* just make sure it is in-range */
1765         lower_barrier(conf);
1766
1767         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1768         md_wakeup_thread(mddev->thread);
1769
1770         mempool_destroy(oldpool);
1771         return 0;
1772 }
1773
1774 static void raid1_quiesce(mddev_t *mddev, int state)
1775 {
1776         conf_t *conf = mddev_to_conf(mddev);
1777
1778         switch(state) {
1779         case 1:
1780                 raise_barrier(conf);
1781                 break;
1782         case 0:
1783                 lower_barrier(conf);
1784                 break;
1785         }
1786         if (mddev->thread) {
1787                 if (mddev->bitmap)
1788                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
1789                 else
1790                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1791                 md_wakeup_thread(mddev->thread);
1792         }
1793 }
1794
1795
1796 static mdk_personality_t raid1_personality =
1797 {
1798         .name           = "raid1",
1799         .owner          = THIS_MODULE,
1800         .make_request   = make_request,
1801         .run            = run,
1802         .stop           = stop,
1803         .status         = status,
1804         .error_handler  = error,
1805         .hot_add_disk   = raid1_add_disk,
1806         .hot_remove_disk= raid1_remove_disk,
1807         .spare_active   = raid1_spare_active,
1808         .sync_request   = sync_request,
1809         .resize         = raid1_resize,
1810         .reshape        = raid1_reshape,
1811         .quiesce        = raid1_quiesce,
1812 };
1813
1814 static int __init raid_init(void)
1815 {
1816         return register_md_personality(RAID1, &raid1_personality);
1817 }
1818
1819 static void raid_exit(void)
1820 {
1821         unregister_md_personality(RAID1);
1822 }
1823
1824 module_init(raid_init);
1825 module_exit(raid_exit);
1826 MODULE_LICENSE("GPL");
1827 MODULE_ALIAS("md-personality-3"); /* RAID1 */