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
  51 static void unplug_slaves(mddev_t *mddev);
  52
  53 static void allow_barrier(conf_t *conf);
  54 static void lower_barrier(conf_t *conf);
  55
  56 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  57 {
  58         struct pool_info *pi = data;
  59         r1bio_t *r1_bio;
  60         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
  61
  62         /* allocate a r1bio with room for raid_disks entries in the bios array */
  63         r1_bio = kzalloc(size, gfp_flags);
  64         if (!r1_bio)
  65                 unplug_slaves(pi->mddev);
  66
  67         return r1_bio;
  68 }
  69
  70 static void r1bio_pool_free(void *r1_bio, void *data)
  71 {
  72         kfree(r1_bio);
  73 }
  74
  75 #define RESYNC_BLOCK_SIZE (64*1024)
  76 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  77 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  78 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  79 #define RESYNC_WINDOW (2048*1024)
  80
  81 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  82 {
  83         struct pool_info *pi = data;
  84         struct page *page;
  85         r1bio_t *r1_bio;
  86         struct bio *bio;
  87         int i, j;
  88
  89         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  90         if (!r1_bio) {
  91                 unplug_slaves(pi->mddev);
  92                 return NULL;
  93         }
  94
  95         /*
  96          * Allocate bios : 1 for reading, n-1 for writing
  97          */
  98         for (j = pi->raid_disks ; j-- ; ) {
  99                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 100                 if (!bio)
 101                         goto out_free_bio;
 102                 r1_bio->bios[j] = bio;
 103         }
 104         /*
 105          * Allocate RESYNC_PAGES data pages and attach them to
 106          * the first bio.
 107          * If this is a user-requested check/repair, allocate
 108          * RESYNC_PAGES for each bio.
 109          */
 110         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 111                 j = pi->raid_disks;
 112         else
 113                 j = 1;
 114         while(j--) {
 115                 bio = r1_bio->bios[j];
 116                 for (i = 0; i < RESYNC_PAGES; i++) {
 117                         page = alloc_page(gfp_flags);
 118                         if (unlikely(!page))
 119                                 goto out_free_pages;
 120
 121                         bio->bi_io_vec[i].bv_page = page;
 122                 }
 123         }
 124         /* If not user-requests, copy the page pointers to all bios */
 125         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 126                 for (i=0; i<RESYNC_PAGES ; i++)
 127                         for (j=1; j<pi->raid_disks; j++)
 128                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 129                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 130         }
 131
 132         r1_bio->master_bio = NULL;
 133
 134         return r1_bio;
 135
 136 out_free_pages:
 137         for (i=0; i < RESYNC_PAGES ; i++)
 138                 for (j=0 ; j < pi->raid_disks; j++)
 139                         safe_put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 140         j = -1;
 141 out_free_bio:
 142         while ( ++j < pi->raid_disks )
 143                 bio_put(r1_bio->bios[j]);
 144         r1bio_pool_free(r1_bio, data);
 145         return NULL;
 146 }
 147
 148 static void r1buf_pool_free(void *__r1_bio, void *data)
 149 {
 150         struct pool_info *pi = data;
 151         int i,j;
 152         r1bio_t *r1bio = __r1_bio;
 153
 154         for (i = 0; i < RESYNC_PAGES; i++)
 155                 for (j = pi->raid_disks; j-- ;) {
 156                         if (j == 0 ||
 157                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 158                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 159                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 160                 }
 161         for (i=0 ; i < pi->raid_disks; i++)
 162                 bio_put(r1bio->bios[i]);
 163
 164         r1bio_pool_free(r1bio, data);
 165 }
 166
 167 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 168 {
 169         int i;
 170
 171         for (i = 0; i < conf->raid_disks; i++) {
 172                 struct bio **bio = r1_bio->bios + i;
 173                 if (*bio && *bio != IO_BLOCKED)
 174                         bio_put(*bio);
 175                 *bio = NULL;
 176         }
 177 }
 178
 179 static void free_r1bio(r1bio_t *r1_bio)
 180 {
 181         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 182
 183         /*
 184          * Wake up any possible resync thread that waits for the device
 185          * to go idle.
 186          */
 187         allow_barrier(conf);
 188
 189         put_all_bios(conf, r1_bio);
 190         mempool_free(r1_bio, conf->r1bio_pool);
 191 }
 192
 193 static void put_buf(r1bio_t *r1_bio)
 194 {
 195         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 196         int i;
 197
 198         for (i=0; i<conf->raid_disks; i++) {
 199                 struct bio *bio = r1_bio->bios[i];
 200                 if (bio->bi_end_io)
 201                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 202         }
 203
 204         mempool_free(r1_bio, conf->r1buf_pool);
 205
 206         lower_barrier(conf);
 207 }
 208
 209 static void reschedule_retry(r1bio_t *r1_bio)
 210 {
 211         unsigned long flags;
 212         mddev_t *mddev = r1_bio->mddev;
 213         conf_t *conf = mddev_to_conf(mddev);
 214
 215         spin_lock_irqsave(&conf->device_lock, flags);
 216         list_add(&r1_bio->retry_list, &conf->retry_list);
 217         conf->nr_queued ++;
 218         spin_unlock_irqrestore(&conf->device_lock, flags);
 219
 220         wake_up(&conf->wait_barrier);
 221         md_wakeup_thread(mddev->thread);
 222 }
 223
 224 /*
 225  * raid_end_bio_io() is called when we have finished servicing a mirrored
 226  * operation and are ready to return a success/failure code to the buffer
 227  * cache layer.
 228  */
 229 static void raid_end_bio_io(r1bio_t *r1_bio)
 230 {
 231         struct bio *bio = r1_bio->master_bio;
 232
 233         /* if nobody has done the final endio yet, do it now */
 234         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 235                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
 236                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 237                         (unsigned long long) bio->bi_sector,
 238                         (unsigned long long) bio->bi_sector +
 239                                 (bio->bi_size >> 9) - 1);
 240
 241                 bio_endio(bio, bio->bi_size,
 242                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 243         }
 244         free_r1bio(r1_bio);
 245 }
 246
 247 /*
 248  * Update disk head position estimator based on IRQ completion info.
 249  */
 250 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 251 {
 252         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 253
 254         conf->mirrors[disk].head_position =
 255                 r1_bio->sector + (r1_bio->sectors);
 256 }
 257
 258 static int raid1_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
 259 {
 260         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 261         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 262         int mirror;
 263         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 264
 265         if (bio->bi_size)
 266                 return 1;
 267
 268         mirror = r1_bio->read_disk;
 269         /*
 270          * this branch is our 'one mirror IO has finished' event handler:
 271          */
 272         update_head_pos(mirror, r1_bio);
 273
 274         if (uptodate || conf->working_disks <= 1) {
 275                 /*
 276                  * Set R1BIO_Uptodate in our master bio, so that
 277                  * we will return a good error code for to the higher
 278                  * levels even if IO on some other mirrored buffer fails.
 279                  *
 280                  * The 'master' represents the composite IO operation to
 281                  * user-side. So if something waits for IO, then it will
 282                  * wait for the 'master' bio.
 283                  */
 284                 if (uptodate)
 285                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 286
 287                 raid_end_bio_io(r1_bio);
 288         } else {
 289                 /*
 290                  * oops, read error:
 291                  */
 292                 char b[BDEVNAME_SIZE];
 293                 if (printk_ratelimit())
 294                         printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
 295                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 296                 reschedule_retry(r1_bio);
 297         }
 298
 299         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 300         return 0;
 301 }
 302
 303 static int raid1_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
 304 {
 305         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 306         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 307         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 308         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 309         struct bio *to_put = NULL;
 310
 311         if (bio->bi_size)
 312                 return 1;
 313
 314         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 315                 if (r1_bio->bios[mirror] == bio)
 316                         break;
 317
 318         if (error == -ENOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
 319                 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
 320                 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
 321                 r1_bio->mddev->barriers_work = 0;
 322         } else {
 323                 /*
 324                  * this branch is our 'one mirror IO has finished' event handler:
 325                  */
 326                 r1_bio->bios[mirror] = NULL;
 327                 to_put = bio;
 328                 if (!uptodate) {
 329                         md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 330                         /* an I/O failed, we can't clear the bitmap */
 331                         set_bit(R1BIO_Degraded, &r1_bio->state);
 332                 } else
 333                         /*
 334                          * Set R1BIO_Uptodate in our master bio, so that
 335                          * we will return a good error code for to the higher
 336                          * levels even if IO on some other mirrored buffer fails.
 337                          *
 338                          * The 'master' represents the composite IO operation to
 339                          * user-side. So if something waits for IO, then it will
 340                          * wait for the 'master' bio.
 341                          */
 342                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 343
 344                 update_head_pos(mirror, r1_bio);
 345
 346                 if (behind) {
 347                         if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 348                                 atomic_dec(&r1_bio->behind_remaining);
 349
 350                         /* In behind mode, we ACK the master bio once the I/O has safely
 351                          * reached all non-writemostly disks. Setting the Returned bit
 352                          * ensures that this gets done only once -- we don't ever want to
 353                          * return -EIO here, instead we'll wait */
 354
 355                         if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 356                             test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 357                                 /* Maybe we can return now */
 358                                 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 359                                         struct bio *mbio = r1_bio->master_bio;
 360                                         PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 361                                                (unsigned long long) mbio->bi_sector,
 362                                                (unsigned long long) mbio->bi_sector +
 363                                                (mbio->bi_size >> 9) - 1);
 364                                         bio_endio(mbio, mbio->bi_size, 0);
 365                                 }
 366                         }
 367                 }
 368         }
 369         /*
 370          *
 371          * Let's see if all mirrored write operations have finished
 372          * already.
 373          */
 374         if (atomic_dec_and_test(&r1_bio->remaining)) {
 375                 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
 376                         reschedule_retry(r1_bio);
 377                         /* Don't dec_pending yet, we want to hold
 378                          * the reference over the retry
 379                          */
 380                         goto out;
 381                 }
 382                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 383                         /* free extra copy of the data pages */
 384                         int i = bio->bi_vcnt;
 385                         while (i--)
 386                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 387                 }
 388                 /* clear the bitmap if all writes complete successfully */
 389                 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 390                                 r1_bio->sectors,
 391                                 !test_bit(R1BIO_Degraded, &r1_bio->state),
 392                                 behind);
 393                 md_write_end(r1_bio->mddev);
 394                 raid_end_bio_io(r1_bio);
 395         }
 396
 397         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 398  out:
 399         if (to_put)
 400                 bio_put(to_put);
 401
 402         return 0;
 403 }
 404
 405
 406 /*
 407  * This routine returns the disk from which the requested read should
 408  * be done. There is a per-array 'next expected sequential IO' sector
 409  * number - if this matches on the next IO then we use the last disk.
 410  * There is also a per-disk 'last know head position' sector that is
 411  * maintained from IRQ contexts, both the normal and the resync IO
 412  * completion handlers update this position correctly. If there is no
 413  * perfect sequential match then we pick the disk whose head is closest.
 414  *
 415  * If there are 2 mirrors in the same 2 devices, performance degrades
 416  * because position is mirror, not device based.
 417  *
 418  * The rdev for the device selected will have nr_pending incremented.
 419  */
 420 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 421 {
 422         const unsigned long this_sector = r1_bio->sector;
 423         int new_disk = conf->last_used, disk = new_disk;
 424         int wonly_disk = -1;
 425         const int sectors = r1_bio->sectors;
 426         sector_t new_distance, current_distance;
 427         mdk_rdev_t *rdev;
 428
 429         rcu_read_lock();
 430         /*
 431          * Check if we can balance. We can balance on the whole
 432          * device if no resync is going on, or below the resync window.
 433          * We take the first readable disk when above the resync window.
 434          */
 435  retry:
 436         if (conf->mddev->recovery_cp < MaxSector &&
 437             (this_sector + sectors >= conf->next_resync)) {
 438                 /* Choose the first operation device, for consistancy */
 439                 new_disk = 0;
 440
 441                 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 442                      r1_bio->bios[new_disk] == IO_BLOCKED ||
 443                      !rdev || !test_bit(In_sync, &rdev->flags)
 444                              || test_bit(WriteMostly, &rdev->flags);
 445                      rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
 446
 447                         if (rdev && test_bit(In_sync, &rdev->flags) &&
 448                                 r1_bio->bios[new_disk] != IO_BLOCKED)
 449                                 wonly_disk = new_disk;
 450
 451                         if (new_disk == conf->raid_disks - 1) {
 452                                 new_disk = wonly_disk;
 453                                 break;
 454                         }
 455                 }
 456                 goto rb_out;
 457         }
 458
 459
 460         /* make sure the disk is operational */
 461         for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 462              r1_bio->bios[new_disk] == IO_BLOCKED ||
 463              !rdev || !test_bit(In_sync, &rdev->flags) ||
 464                      test_bit(WriteMostly, &rdev->flags);
 465              rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
 466
 467                 if (rdev && test_bit(In_sync, &rdev->flags) &&
 468                     r1_bio->bios[new_disk] != IO_BLOCKED)
 469                         wonly_disk = new_disk;
 470
 471                 if (new_disk <= 0)
 472                         new_disk = conf->raid_disks;
 473                 new_disk--;
 474                 if (new_disk == disk) {
 475                         new_disk = wonly_disk;
 476                         break;
 477                 }
 478         }
 479
 480         if (new_disk < 0)
 481                 goto rb_out;
 482
 483         disk = new_disk;
 484         /* now disk == new_disk == starting point for search */
 485
 486         /*
 487          * Don't change to another disk for sequential reads:
 488          */
 489         if (conf->next_seq_sect == this_sector)
 490                 goto rb_out;
 491         if (this_sector == conf->mirrors[new_disk].head_position)
 492                 goto rb_out;
 493
 494         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 495
 496         /* Find the disk whose head is closest */
 497
 498         do {
 499                 if (disk <= 0)
 500                         disk = conf->raid_disks;
 501                 disk--;
 502
 503                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 504
 505                 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
 506                     !test_bit(In_sync, &rdev->flags) ||
 507                     test_bit(WriteMostly, &rdev->flags))
 508                         continue;
 509
 510                 if (!atomic_read(&rdev->nr_pending)) {
 511                         new_disk = disk;
 512                         break;
 513                 }
 514                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 515                 if (new_distance < current_distance) {
 516                         current_distance = new_distance;
 517                         new_disk = disk;
 518                 }
 519         } while (disk != conf->last_used);
 520
 521  rb_out:
 522
 523
 524         if (new_disk >= 0) {
 525                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 526                 if (!rdev)
 527                         goto retry;
 528                 atomic_inc(&rdev->nr_pending);
 529                 if (!test_bit(In_sync, &rdev->flags)) {
 530                         /* cannot risk returning a device that failed
 531                          * before we inc'ed nr_pending
 532                          */
 533                         rdev_dec_pending(rdev, conf->mddev);
 534                         goto retry;
 535                 }
 536                 conf->next_seq_sect = this_sector + sectors;
 537                 conf->last_used = new_disk;
 538         }
 539         rcu_read_unlock();
 540
 541         return new_disk;
 542 }
 543
 544 static void unplug_slaves(mddev_t *mddev)
 545 {
 546         conf_t *conf = mddev_to_conf(mddev);
 547         int i;
 548
 549         rcu_read_lock();
 550         for (i=0; i<mddev->raid_disks; i++) {
 551                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 552                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 553                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 554
 555                         atomic_inc(&rdev->nr_pending);
 556                         rcu_read_unlock();
 557
 558                         if (r_queue->unplug_fn)
 559                                 r_queue->unplug_fn(r_queue);
 560
 561                         rdev_dec_pending(rdev, mddev);
 562                         rcu_read_lock();
 563                 }
 564         }
 565         rcu_read_unlock();
 566 }
 567
 568 static void raid1_unplug(request_queue_t *q)
 569 {
 570         mddev_t *mddev = q->queuedata;
 571
 572         unplug_slaves(mddev);
 573         md_wakeup_thread(mddev->thread);
 574 }
 575
 576 static int raid1_issue_flush(request_queue_t *q, struct gendisk *disk,
 577                              sector_t *error_sector)
 578 {
 579         mddev_t *mddev = q->queuedata;
 580         conf_t *conf = mddev_to_conf(mddev);
 581         int i, ret = 0;
 582
 583         rcu_read_lock();
 584         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
 585                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 586                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 587                         struct block_device *bdev = rdev->bdev;
 588                         request_queue_t *r_queue = bdev_get_queue(bdev);
 589
 590                         if (!r_queue->issue_flush_fn)
 591                                 ret = -EOPNOTSUPP;
 592                         else {
 593                                 atomic_inc(&rdev->nr_pending);
 594                                 rcu_read_unlock();
 595                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
 596                                                               error_sector);
 597                                 rdev_dec_pending(rdev, mddev);
 598                                 rcu_read_lock();
 599                         }
 600                 }
 601         }
 602         rcu_read_unlock();
 603         return ret;
 604 }
 605
 606 /* Barriers....
 607  * Sometimes we need to suspend IO while we do something else,
 608  * either some resync/recovery, or reconfigure the array.
 609  * To do this we raise a 'barrier'.
 610  * The 'barrier' is a counter that can be raised multiple times
 611  * to count how many activities are happening which preclude
 612  * normal IO.
 613  * We can only raise the barrier if there is no pending IO.
 614  * i.e. if nr_pending == 0.
 615  * We choose only to raise the barrier if no-one is waiting for the
 616  * barrier to go down.  This means that as soon as an IO request
 617  * is ready, no other operations which require a barrier will start
 618  * until the IO request has had a chance.
 619  *
 620  * So: regular IO calls 'wait_barrier'.  When that returns there
 621  *    is no backgroup IO happening,  It must arrange to call
 622  *    allow_barrier when it has finished its IO.
 623  * backgroup IO calls must call raise_barrier.  Once that returns
 624  *    there is no normal IO happeing.  It must arrange to call
 625  *    lower_barrier when the particular background IO completes.
 626  */
 627 #define RESYNC_DEPTH 32
 628
 629 static void raise_barrier(conf_t *conf)
 630 {
 631         spin_lock_irq(&conf->resync_lock);
 632
 633         /* Wait until no block IO is waiting */
 634         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 635                             conf->resync_lock,
 636                             raid1_unplug(conf->mddev->queue));
 637
 638         /* block any new IO from starting */
 639         conf->barrier++;
 640
 641         /* No wait for all pending IO to complete */
 642         wait_event_lock_irq(conf->wait_barrier,
 643                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 644                             conf->resync_lock,
 645                             raid1_unplug(conf->mddev->queue));
 646
 647         spin_unlock_irq(&conf->resync_lock);
 648 }
 649
 650 static void lower_barrier(conf_t *conf)
 651 {
 652         unsigned long flags;
 653         spin_lock_irqsave(&conf->resync_lock, flags);
 654         conf->barrier--;
 655         spin_unlock_irqrestore(&conf->resync_lock, flags);
 656         wake_up(&conf->wait_barrier);
 657 }
 658
 659 static void wait_barrier(conf_t *conf)
 660 {
 661         spin_lock_irq(&conf->resync_lock);
 662         if (conf->barrier) {
 663                 conf->nr_waiting++;
 664                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 665                                     conf->resync_lock,
 666                                     raid1_unplug(conf->mddev->queue));
 667                 conf->nr_waiting--;
 668         }
 669         conf->nr_pending++;
 670         spin_unlock_irq(&conf->resync_lock);
 671 }
 672
 673 static void allow_barrier(conf_t *conf)
 674 {
 675         unsigned long flags;
 676         spin_lock_irqsave(&conf->resync_lock, flags);
 677         conf->nr_pending--;
 678         spin_unlock_irqrestore(&conf->resync_lock, flags);
 679         wake_up(&conf->wait_barrier);
 680 }
 681
 682 static void freeze_array(conf_t *conf)
 683 {
 684         /* stop syncio and normal IO and wait for everything to
 685          * go quite.
 686          * We increment barrier and nr_waiting, and then
 687          * wait until barrier+nr_pending match nr_queued+2
 688          */
 689         spin_lock_irq(&conf->resync_lock);
 690         conf->barrier++;
 691         conf->nr_waiting++;
 692         wait_event_lock_irq(conf->wait_barrier,
 693                             conf->barrier+conf->nr_pending == conf->nr_queued+2,
 694                             conf->resync_lock,
 695                             raid1_unplug(conf->mddev->queue));
 696         spin_unlock_irq(&conf->resync_lock);
 697 }
 698 static void unfreeze_array(conf_t *conf)
 699 {
 700         /* reverse the effect of the freeze */
 701         spin_lock_irq(&conf->resync_lock);
 702         conf->barrier--;
 703         conf->nr_waiting--;
 704         wake_up(&conf->wait_barrier);
 705         spin_unlock_irq(&conf->resync_lock);
 706 }
 707
 708
 709 /* duplicate the data pages for behind I/O */
 710 static struct page **alloc_behind_pages(struct bio *bio)
 711 {
 712         int i;
 713         struct bio_vec *bvec;
 714         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
 715                                         GFP_NOIO);
 716         if (unlikely(!pages))
 717                 goto do_sync_io;
 718
 719         bio_for_each_segment(bvec, bio, i) {
 720                 pages[i] = alloc_page(GFP_NOIO);
 721                 if (unlikely(!pages[i]))
 722                         goto do_sync_io;
 723                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 724                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 725                 kunmap(pages[i]);
 726                 kunmap(bvec->bv_page);
 727         }
 728
 729         return pages;
 730
 731 do_sync_io:
 732         if (pages)
 733                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
 734                         put_page(pages[i]);
 735         kfree(pages);
 736         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 737         return NULL;
 738 }
 739
 740 static int make_request(request_queue_t *q, struct bio * bio)
 741 {
 742         mddev_t *mddev = q->queuedata;
 743         conf_t *conf = mddev_to_conf(mddev);
 744         mirror_info_t *mirror;
 745         r1bio_t *r1_bio;
 746         struct bio *read_bio;
 747         int i, targets = 0, disks;
 748         mdk_rdev_t *rdev;
 749         struct bitmap *bitmap = mddev->bitmap;
 750         unsigned long flags;
 751         struct bio_list bl;
 752         struct page **behind_pages = NULL;
 753         const int rw = bio_data_dir(bio);
 754         const int do_sync = bio_sync(bio);
 755         int do_barriers;
 756
 757         if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
 758                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
 759                 return 0;
 760         }
 761
 762         /*
 763          * Register the new request and wait if the reconstruction
 764          * thread has put up a bar for new requests.
 765          * Continue immediately if no resync is active currently.
 766          */
 767         md_write_start(mddev, bio); /* wait on superblock update early */
 768
 769         wait_barrier(conf);
 770
 771         disk_stat_inc(mddev->gendisk, ios[rw]);
 772         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
 773
 774         /*
 775          * make_request() can abort the operation when READA is being
 776          * used and no empty request is available.
 777          *
 778          */
 779         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 780
 781         r1_bio->master_bio = bio;
 782         r1_bio->sectors = bio->bi_size >> 9;
 783         r1_bio->state = 0;
 784         r1_bio->mddev = mddev;
 785         r1_bio->sector = bio->bi_sector;
 786
 787         if (rw == READ) {
 788                 /*
 789                  * read balancing logic:
 790                  */
 791                 int rdisk = read_balance(conf, r1_bio);
 792
 793                 if (rdisk < 0) {
 794                         /* couldn't find anywhere to read from */
 795                         raid_end_bio_io(r1_bio);
 796                         return 0;
 797                 }
 798                 mirror = conf->mirrors + rdisk;
 799
 800                 r1_bio->read_disk = rdisk;
 801
 802                 read_bio = bio_clone(bio, GFP_NOIO);
 803
 804                 r1_bio->bios[rdisk] = read_bio;
 805
 806                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 807                 read_bio->bi_bdev = mirror->rdev->bdev;
 808                 read_bio->bi_end_io = raid1_end_read_request;
 809                 read_bio->bi_rw = READ | do_sync;
 810                 read_bio->bi_private = r1_bio;
 811
 812                 generic_make_request(read_bio);
 813                 return 0;
 814         }
 815
 816         /*
 817          * WRITE:
 818          */
 819         /* first select target devices under spinlock and
 820          * inc refcount on their rdev.  Record them by setting
 821          * bios[x] to bio
 822          */
 823         disks = conf->raid_disks;
 824 #if 0
 825         { static int first=1;
 826         if (first) printk("First Write sector %llu disks %d\n",
 827                           (unsigned long long)r1_bio->sector, disks);
 828         first = 0;
 829         }
 830 #endif
 831         rcu_read_lock();
 832         for (i = 0;  i < disks; i++) {
 833                 if ((rdev=rcu_dereference(conf->mirrors[i].rdev)) != NULL &&
 834                     !test_bit(Faulty, &rdev->flags)) {
 835                         atomic_inc(&rdev->nr_pending);
 836                         if (test_bit(Faulty, &rdev->flags)) {
 837                                 rdev_dec_pending(rdev, mddev);
 838                                 r1_bio->bios[i] = NULL;
 839                         } else
 840                                 r1_bio->bios[i] = bio;
 841                         targets++;
 842                 } else
 843                         r1_bio->bios[i] = NULL;
 844         }
 845         rcu_read_unlock();
 846
 847         BUG_ON(targets == 0); /* we never fail the last device */
 848
 849         if (targets < conf->raid_disks) {
 850                 /* array is degraded, we will not clear the bitmap
 851                  * on I/O completion (see raid1_end_write_request) */
 852                 set_bit(R1BIO_Degraded, &r1_bio->state);
 853         }
 854
 855         /* do behind I/O ? */
 856         if (bitmap &&
 857             atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
 858             (behind_pages = alloc_behind_pages(bio)) != NULL)
 859                 set_bit(R1BIO_BehindIO, &r1_bio->state);
 860
 861         atomic_set(&r1_bio->remaining, 0);
 862         atomic_set(&r1_bio->behind_remaining, 0);
 863
 864         do_barriers = bio_barrier(bio);
 865         if (do_barriers)
 866                 set_bit(R1BIO_Barrier, &r1_bio->state);
 867
 868         bio_list_init(&bl);
 869         for (i = 0; i < disks; i++) {
 870                 struct bio *mbio;
 871                 if (!r1_bio->bios[i])
 872                         continue;
 873
 874                 mbio = bio_clone(bio, GFP_NOIO);
 875                 r1_bio->bios[i] = mbio;
 876
 877                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 878                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 879                 mbio->bi_end_io = raid1_end_write_request;
 880                 mbio->bi_rw = WRITE | do_barriers | do_sync;
 881                 mbio->bi_private = r1_bio;
 882
 883                 if (behind_pages) {
 884                         struct bio_vec *bvec;
 885                         int j;
 886
 887                         /* Yes, I really want the '__' version so that
 888                          * we clear any unused pointer in the io_vec, rather
 889                          * than leave them unchanged.  This is important
 890                          * because when we come to free the pages, we won't
 891                          * know the originial bi_idx, so we just free
 892                          * them all
 893                          */
 894                         __bio_for_each_segment(bvec, mbio, j, 0)
 895                                 bvec->bv_page = behind_pages[j];
 896                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 897                                 atomic_inc(&r1_bio->behind_remaining);
 898                 }
 899
 900                 atomic_inc(&r1_bio->remaining);
 901
 902                 bio_list_add(&bl, mbio);
 903         }
 904         kfree(behind_pages); /* the behind pages are attached to the bios now */
 905
 906         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
 907                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 908         spin_lock_irqsave(&conf->device_lock, flags);
 909         bio_list_merge(&conf->pending_bio_list, &bl);
 910         bio_list_init(&bl);
 911
 912         blk_plug_device(mddev->queue);
 913         spin_unlock_irqrestore(&conf->device_lock, flags);
 914
 915         if (do_sync)
 916                 md_wakeup_thread(mddev->thread);
 917 #if 0
 918         while ((bio = bio_list_pop(&bl)) != NULL)
 919                 generic_make_request(bio);
 920 #endif
 921
 922         return 0;
 923 }
 924
 925 static void status(struct seq_file *seq, mddev_t *mddev)
 926 {
 927         conf_t *conf = mddev_to_conf(mddev);
 928         int i;
 929
 930         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 931                                                 conf->working_disks);
 932         for (i = 0; i < conf->raid_disks; i++)
 933                 seq_printf(seq, "%s",
 934                               conf->mirrors[i].rdev &&
 935                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
 936         seq_printf(seq, "]");
 937 }
 938
 939
 940 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 941 {
 942         char b[BDEVNAME_SIZE];
 943         conf_t *conf = mddev_to_conf(mddev);
 944
 945         /*
 946          * If it is not operational, then we have already marked it as dead
 947          * else if it is the last working disks, ignore the error, let the
 948          * next level up know.
 949          * else mark the drive as failed
 950          */
 951         if (test_bit(In_sync, &rdev->flags)
 952             && conf->working_disks == 1)
 953                 /*
 954                  * Don't fail the drive, act as though we were just a
 955                  * normal single drive
 956                  */
 957                 return;
 958         if (test_bit(In_sync, &rdev->flags)) {
 959                 mddev->degraded++;
 960                 conf->working_disks--;
 961                 /*
 962                  * if recovery is running, make sure it aborts.
 963                  */
 964                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 965         }
 966         clear_bit(In_sync, &rdev->flags);
 967         set_bit(Faulty, &rdev->flags);
 968         mddev->sb_dirty = 1;
 969         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
 970                 "       Operation continuing on %d devices\n",
 971                 bdevname(rdev->bdev,b), conf->working_disks);
 972 }
 973
 974 static void print_conf(conf_t *conf)
 975 {
 976         int i;
 977         mirror_info_t *tmp;
 978
 979         printk("RAID1 conf printout:\n");
 980         if (!conf) {
 981                 printk("(!conf)\n");
 982                 return;
 983         }
 984         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 985                 conf->raid_disks);
 986
 987         for (i = 0; i < conf->raid_disks; i++) {
 988                 char b[BDEVNAME_SIZE];
 989                 tmp = conf->mirrors + i;
 990                 if (tmp->rdev)
 991                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
 992                                 i, !test_bit(In_sync, &tmp->rdev->flags), !test_bit(Faulty, &tmp->rdev->flags),
 993                                 bdevname(tmp->rdev->bdev,b));
 994         }
 995 }
 996
 997 static void close_sync(conf_t *conf)
 998 {
 999         wait_barrier(conf);
1000         allow_barrier(conf);
1001
1002         mempool_destroy(conf->r1buf_pool);
1003         conf->r1buf_pool = NULL;
1004 }
1005
1006 static int raid1_spare_active(mddev_t *mddev)
1007 {
1008         int i;
1009         conf_t *conf = mddev->private;
1010         mirror_info_t *tmp;
1011
1012         /*
1013          * Find all failed disks within the RAID1 configuration
1014          * and mark them readable
1015          */
1016         for (i = 0; i < conf->raid_disks; i++) {
1017                 tmp = conf->mirrors + i;
1018                 if (tmp->rdev
1019                     && !test_bit(Faulty, &tmp->rdev->flags)
1020                     && !test_bit(In_sync, &tmp->rdev->flags)) {
1021                         conf->working_disks++;
1022                         mddev->degraded--;
1023                         set_bit(In_sync, &tmp->rdev->flags);
1024                 }
1025         }
1026
1027         print_conf(conf);
1028         return 0;
1029 }
1030
1031
1032 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1033 {
1034         conf_t *conf = mddev->private;
1035         int found = 0;
1036         int mirror = 0;
1037         mirror_info_t *p;
1038
1039         for (mirror=0; mirror < mddev->raid_disks; mirror++)
1040                 if ( !(p=conf->mirrors+mirror)->rdev) {
1041
1042                         blk_queue_stack_limits(mddev->queue,
1043                                                rdev->bdev->bd_disk->queue);
1044                         /* as we don't honour merge_bvec_fn, we must never risk
1045                          * violating it, so limit ->max_sector to one PAGE, as
1046                          * a one page request is never in violation.
1047                          */
1048                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1049                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1050                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1051
1052                         p->head_position = 0;
1053                         rdev->raid_disk = mirror;
1054                         found = 1;
1055                         /* As all devices are equivalent, we don't need a full recovery
1056                          * if this was recently any drive of the array
1057                          */
1058                         if (rdev->saved_raid_disk < 0)
1059                                 conf->fullsync = 1;
1060                         rcu_assign_pointer(p->rdev, rdev);
1061                         break;
1062                 }
1063
1064         print_conf(conf);
1065         return found;
1066 }
1067
1068 static int raid1_remove_disk(mddev_t *mddev, int number)
1069 {
1070         conf_t *conf = mddev->private;
1071         int err = 0;
1072         mdk_rdev_t *rdev;
1073         mirror_info_t *p = conf->mirrors+ number;
1074
1075         print_conf(conf);
1076         rdev = p->rdev;
1077         if (rdev) {
1078                 if (test_bit(In_sync, &rdev->flags) ||
1079                     atomic_read(&rdev->nr_pending)) {
1080                         err = -EBUSY;
1081                         goto abort;
1082                 }
1083                 p->rdev = NULL;
1084                 synchronize_rcu();
1085                 if (atomic_read(&rdev->nr_pending)) {
1086                         /* lost the race, try later */
1087                         err = -EBUSY;
1088                         p->rdev = rdev;
1089                 }
1090         }
1091 abort:
1092
1093         print_conf(conf);
1094         return err;
1095 }
1096
1097
1098 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1099 {
1100         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1101         int i;
1102
1103         if (bio->bi_size)
1104                 return 1;
1105
1106         for (i=r1_bio->mddev->raid_disks; i--; )
1107                 if (r1_bio->bios[i] == bio)
1108                         break;
1109         BUG_ON(i < 0);
1110         update_head_pos(i, r1_bio);
1111         /*
1112          * we have read a block, now it needs to be re-written,
1113          * or re-read if the read failed.
1114          * We don't do much here, just schedule handling by raid1d
1115          */
1116         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1117                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1118
1119         if (atomic_dec_and_test(&r1_bio->remaining))
1120                 reschedule_retry(r1_bio);
1121         return 0;
1122 }
1123
1124 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1125 {
1126         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1127         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1128         mddev_t *mddev = r1_bio->mddev;
1129         conf_t *conf = mddev_to_conf(mddev);
1130         int i;
1131         int mirror=0;
1132
1133         if (bio->bi_size)
1134                 return 1;
1135
1136         for (i = 0; i < conf->raid_disks; i++)
1137                 if (r1_bio->bios[i] == bio) {
1138                         mirror = i;
1139                         break;
1140                 }
1141         if (!uptodate)
1142                 md_error(mddev, conf->mirrors[mirror].rdev);
1143
1144         update_head_pos(mirror, r1_bio);
1145
1146         if (atomic_dec_and_test(&r1_bio->remaining)) {
1147                 md_done_sync(mddev, r1_bio->sectors, uptodate);
1148                 put_buf(r1_bio);
1149         }
1150         return 0;
1151 }
1152
1153 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1154 {
1155         conf_t *conf = mddev_to_conf(mddev);
1156         int i;
1157         int disks = conf->raid_disks;
1158         struct bio *bio, *wbio;
1159
1160         bio = r1_bio->bios[r1_bio->read_disk];
1161
1162
1163         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1164                 /* We have read all readable devices.  If we haven't
1165                  * got the block, then there is no hope left.
1166                  * If we have, then we want to do a comparison
1167                  * and skip the write if everything is the same.
1168                  * If any blocks failed to read, then we need to
1169                  * attempt an over-write
1170                  */
1171                 int primary;
1172                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1173                         for (i=0; i<mddev->raid_disks; i++)
1174                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1175                                         md_error(mddev, conf->mirrors[i].rdev);
1176
1177                         md_done_sync(mddev, r1_bio->sectors, 1);
1178                         put_buf(r1_bio);
1179                         return;
1180                 }
1181                 for (primary=0; primary<mddev->raid_disks; primary++)
1182                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1183                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1184                                 r1_bio->bios[primary]->bi_end_io = NULL;
1185                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1186                                 break;
1187                         }
1188                 r1_bio->read_disk = primary;
1189                 for (i=0; i<mddev->raid_disks; i++)
1190                         if (r1_bio->bios[i]->bi_end_io == end_sync_read &&
1191                             test_bit(BIO_UPTODATE, &r1_bio->bios[i]->bi_flags)) {
1192                                 int j;
1193                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1194                                 struct bio *pbio = r1_bio->bios[primary];
1195                                 struct bio *sbio = r1_bio->bios[i];
1196                                 for (j = vcnt; j-- ; )
1197                                         if (memcmp(page_address(pbio->bi_io_vec[j].bv_page),
1198                                                    page_address(sbio->bi_io_vec[j].bv_page),
1199                                                    PAGE_SIZE))
1200                                                 break;
1201                                 if (j >= 0)
1202                                         mddev->resync_mismatches += r1_bio->sectors;
1203                                 if (j < 0 || test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
1204                                         sbio->bi_end_io = NULL;
1205                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1206                                 } else {
1207                                         /* fixup the bio for reuse */
1208                                         sbio->bi_vcnt = vcnt;
1209                                         sbio->bi_size = r1_bio->sectors << 9;
1210                                         sbio->bi_idx = 0;
1211                                         sbio->bi_phys_segments = 0;
1212                                         sbio->bi_hw_segments = 0;
1213                                         sbio->bi_hw_front_size = 0;
1214                                         sbio->bi_hw_back_size = 0;
1215                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1216                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1217                                         sbio->bi_next = NULL;
1218                                         sbio->bi_sector = r1_bio->sector +
1219                                                 conf->mirrors[i].rdev->data_offset;
1220                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1221                                         for (j = 0; j < vcnt ; j++)
1222                                                 memcpy(page_address(sbio->bi_io_vec[j].bv_page),
1223                                                        page_address(pbio->bi_io_vec[j].bv_page),
1224                                                        PAGE_SIZE);
1225
1226                                 }
1227                         }
1228         }
1229         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1230                 /* ouch - failed to read all of that.
1231                  * Try some synchronous reads of other devices to get
1232                  * good data, much like with normal read errors.  Only
1233                  * read into the pages we already have so they we don't
1234                  * need to re-issue the read request.
1235                  * We don't need to freeze the array, because being in an
1236                  * active sync request, there is no normal IO, and
1237                  * no overlapping syncs.
1238                  */
1239                 sector_t sect = r1_bio->sector;
1240                 int sectors = r1_bio->sectors;
1241                 int idx = 0;
1242
1243                 while(sectors) {
1244                         int s = sectors;
1245                         int d = r1_bio->read_disk;
1246                         int success = 0;
1247                         mdk_rdev_t *rdev;
1248
1249                         if (s > (PAGE_SIZE>>9))
1250                                 s = PAGE_SIZE >> 9;
1251                         do {
1252                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1253                                         rdev = conf->mirrors[d].rdev;
1254                                         if (sync_page_io(rdev->bdev,
1255                                                          sect + rdev->data_offset,
1256                                                          s<<9,
1257                                                          bio->bi_io_vec[idx].bv_page,
1258                                                          READ)) {
1259                                                 success = 1;
1260                                                 break;
1261                                         }
1262                                 }
1263                                 d++;
1264                                 if (d == conf->raid_disks)
1265                                         d = 0;
1266                         } while (!success && d != r1_bio->read_disk);
1267
1268                         if (success) {
1269                                 int start = d;
1270                                 /* write it back and re-read */
1271                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1272                                 while (d != r1_bio->read_disk) {
1273                                         if (d == 0)
1274                                                 d = conf->raid_disks;
1275                                         d--;
1276                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1277                                                 continue;
1278                                         rdev = conf->mirrors[d].rdev;
1279                                         atomic_add(s, &rdev->corrected_errors);
1280                                         if (sync_page_io(rdev->bdev,
1281                                                          sect + rdev->data_offset,
1282                                                          s<<9,
1283                                                          bio->bi_io_vec[idx].bv_page,
1284                                                          WRITE) == 0)
1285                                                 md_error(mddev, rdev);
1286                                 }
1287                                 d = start;
1288                                 while (d != r1_bio->read_disk) {
1289                                         if (d == 0)
1290                                                 d = conf->raid_disks;
1291                                         d--;
1292                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1293                                                 continue;
1294                                         rdev = conf->mirrors[d].rdev;
1295                                         if (sync_page_io(rdev->bdev,
1296                                                          sect + rdev->data_offset,
1297                                                          s<<9,
1298                                                          bio->bi_io_vec[idx].bv_page,
1299                                                          READ) == 0)
1300                                                 md_error(mddev, rdev);
1301                                 }
1302                         } else {
1303                                 char b[BDEVNAME_SIZE];
1304                                 /* Cannot read from anywhere, array is toast */
1305                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1306                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1307                                        " for block %llu\n",
1308                                        bdevname(bio->bi_bdev,b),
1309                                        (unsigned long long)r1_bio->sector);
1310                                 md_done_sync(mddev, r1_bio->sectors, 0);
1311                                 put_buf(r1_bio);
1312                                 return;
1313                         }
1314                         sectors -= s;
1315                         sect += s;
1316                         idx ++;
1317                 }
1318         }
1319
1320         /*
1321          * schedule writes
1322          */
1323         atomic_set(&r1_bio->remaining, 1);
1324         for (i = 0; i < disks ; i++) {
1325                 wbio = r1_bio->bios[i];
1326                 if (wbio->bi_end_io == NULL ||
1327                     (wbio->bi_end_io == end_sync_read &&
1328                      (i == r1_bio->read_disk ||
1329                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1330                         continue;
1331
1332                 wbio->bi_rw = WRITE;
1333                 wbio->bi_end_io = end_sync_write;
1334                 atomic_inc(&r1_bio->remaining);
1335                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1336
1337                 generic_make_request(wbio);
1338         }
1339
1340         if (atomic_dec_and_test(&r1_bio->remaining)) {
1341                 /* if we're here, all write(s) have completed, so clean up */
1342                 md_done_sync(mddev, r1_bio->sectors, 1);
1343                 put_buf(r1_bio);
1344         }
1345 }
1346
1347 /*
1348  * This is a kernel thread which:
1349  *
1350  *      1.      Retries failed read operations on working mirrors.
1351  *      2.      Updates the raid superblock when problems encounter.
1352  *      3.      Performs writes following reads for array syncronising.
1353  */
1354
1355 static void raid1d(mddev_t *mddev)
1356 {
1357         r1bio_t *r1_bio;
1358         struct bio *bio;
1359         unsigned long flags;
1360         conf_t *conf = mddev_to_conf(mddev);
1361         struct list_head *head = &conf->retry_list;
1362         int unplug=0;
1363         mdk_rdev_t *rdev;
1364
1365         md_check_recovery(mddev);
1366
1367         for (;;) {
1368                 char b[BDEVNAME_SIZE];
1369                 spin_lock_irqsave(&conf->device_lock, flags);
1370
1371                 if (conf->pending_bio_list.head) {
1372                         bio = bio_list_get(&conf->pending_bio_list);
1373                         blk_remove_plug(mddev->queue);
1374                         spin_unlock_irqrestore(&conf->device_lock, flags);
1375                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1376                         if (bitmap_unplug(mddev->bitmap) != 0)
1377                                 printk("%s: bitmap file write failed!\n", mdname(mddev));
1378
1379                         while (bio) { /* submit pending writes */
1380                                 struct bio *next = bio->bi_next;
1381                                 bio->bi_next = NULL;
1382                                 generic_make_request(bio);
1383                                 bio = next;
1384                         }
1385                         unplug = 1;
1386
1387                         continue;
1388                 }
1389
1390                 if (list_empty(head))
1391                         break;
1392                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1393                 list_del(head->prev);
1394                 conf->nr_queued--;
1395                 spin_unlock_irqrestore(&conf->device_lock, flags);
1396
1397                 mddev = r1_bio->mddev;
1398                 conf = mddev_to_conf(mddev);
1399                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1400                         sync_request_write(mddev, r1_bio);
1401                         unplug = 1;
1402                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1403                         /* some requests in the r1bio were BIO_RW_BARRIER
1404                          * requests which failed with -ENOTSUPP.  Hohumm..
1405                          * Better resubmit without the barrier.
1406                          * We know which devices to resubmit for, because
1407                          * all others have had their bios[] entry cleared.
1408                          */
1409                         int i;
1410                         const int do_sync = bio_sync(r1_bio->master_bio);
1411                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1412                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1413                         for (i=0; i < conf->raid_disks; i++)
1414                                 if (r1_bio->bios[i]) {
1415                                         struct bio_vec *bvec;
1416                                         int j;
1417
1418                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1419                                         /* copy pages from the failed bio, as
1420                                          * this might be a write-behind device */
1421                                         __bio_for_each_segment(bvec, bio, j, 0)
1422                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1423                                         bio_put(r1_bio->bios[i]);
1424                                         bio->bi_sector = r1_bio->sector +
1425                                                 conf->mirrors[i].rdev->data_offset;
1426                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1427                                         bio->bi_end_io = raid1_end_write_request;
1428                                         bio->bi_rw = WRITE | do_sync;
1429                                         bio->bi_private = r1_bio;
1430                                         r1_bio->bios[i] = bio;
1431                                         generic_make_request(bio);
1432                                 }
1433                 } else {
1434                         int disk;
1435
1436                         /* we got a read error. Maybe the drive is bad.  Maybe just
1437                          * the block and we can fix it.
1438                          * We freeze all other IO, and try reading the block from
1439                          * other devices.  When we find one, we re-write
1440                          * and check it that fixes the read error.
1441                          * This is all done synchronously while the array is
1442                          * frozen
1443                          */
1444                         sector_t sect = r1_bio->sector;
1445                         int sectors = r1_bio->sectors;
1446                         freeze_array(conf);
1447                         if (mddev->ro == 0) while(sectors) {
1448                                 int s = sectors;
1449                                 int d = r1_bio->read_disk;
1450                                 int success = 0;
1451
1452                                 if (s > (PAGE_SIZE>>9))
1453                                         s = PAGE_SIZE >> 9;
1454
1455                                 do {
1456                                         rdev = conf->mirrors[d].rdev;
1457                                         if (rdev &&
1458                                             test_bit(In_sync, &rdev->flags) &&
1459                                             sync_page_io(rdev->bdev,
1460                                                          sect + rdev->data_offset,
1461                                                          s<<9,
1462                                                          conf->tmppage, READ))
1463                                                 success = 1;
1464                                         else {
1465                                                 d++;
1466                                                 if (d == conf->raid_disks)
1467                                                         d = 0;
1468                                         }
1469                                 } while (!success && d != r1_bio->read_disk);
1470
1471                                 if (success) {
1472                                         /* write it back and re-read */
1473                                         int start = d;
1474                                         while (d != r1_bio->read_disk) {
1475                                                 if (d==0)
1476                                                         d = conf->raid_disks;
1477                                                 d--;
1478                                                 rdev = conf->mirrors[d].rdev;
1479                                                 if (rdev &&
1480                                                     test_bit(In_sync, &rdev->flags)) {
1481                                                         if (sync_page_io(rdev->bdev,
1482                                                                          sect + rdev->data_offset,
1483                                                                          s<<9, conf->tmppage, WRITE) == 0)
1484                                                                 /* Well, this device is dead */
1485                                                                 md_error(mddev, rdev);
1486                                                 }
1487                                         }
1488                                         d = start;
1489                                         while (d != r1_bio->read_disk) {
1490                                                 if (d==0)
1491                                                         d = conf->raid_disks;
1492                                                 d--;
1493                                                 rdev = conf->mirrors[d].rdev;
1494                                                 if (rdev &&
1495                                                     test_bit(In_sync, &rdev->flags)) {
1496                                                         if (sync_page_io(rdev->bdev,
1497                                                                          sect + rdev->data_offset,
1498                                                                          s<<9, conf->tmppage, READ) == 0)
1499                                                                 /* Well, this device is dead */
1500                                                                 md_error(mddev, rdev);
1501                                                         else
1502                                                                 atomic_add(s, &rdev->corrected_errors);
1503
1504                                                 }
1505                                         }
1506                                 } else {
1507                                         /* Cannot read from anywhere -- bye bye array */
1508                                         md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1509                                         break;
1510                                 }
1511                                 sectors -= s;
1512                                 sect += s;
1513                         }
1514
1515                         unfreeze_array(conf);
1516
1517                         bio = r1_bio->bios[r1_bio->read_disk];
1518                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1519                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1520                                        " read error for block %llu\n",
1521                                        bdevname(bio->bi_bdev,b),
1522                                        (unsigned long long)r1_bio->sector);
1523                                 raid_end_bio_io(r1_bio);
1524                         } else {
1525                                 const int do_sync = bio_sync(r1_bio->master_bio);
1526                                 r1_bio->bios[r1_bio->read_disk] =
1527                                         mddev->ro ? IO_BLOCKED : NULL;
1528                                 r1_bio->read_disk = disk;
1529                                 bio_put(bio);
1530                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1531                                 r1_bio->bios[r1_bio->read_disk] = bio;
1532                                 rdev = conf->mirrors[disk].rdev;
1533                                 if (printk_ratelimit())
1534                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1535                                                " another mirror\n",
1536                                                bdevname(rdev->bdev,b),
1537                                                (unsigned long long)r1_bio->sector);
1538                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1539                                 bio->bi_bdev = rdev->bdev;
1540                                 bio->bi_end_io = raid1_end_read_request;
1541                                 bio->bi_rw = READ | do_sync;
1542                                 bio->bi_private = r1_bio;
1543                                 unplug = 1;
1544                                 generic_make_request(bio);
1545                         }
1546                 }
1547         }
1548         spin_unlock_irqrestore(&conf->device_lock, flags);
1549         if (unplug)
1550                 unplug_slaves(mddev);
1551 }
1552
1553
1554 static int init_resync(conf_t *conf)
1555 {
1556         int buffs;
1557
1558         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1559         if (conf->r1buf_pool)
1560                 BUG();
1561         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1562                                           conf->poolinfo);
1563         if (!conf->r1buf_pool)
1564                 return -ENOMEM;
1565         conf->next_resync = 0;
1566         return 0;
1567 }
1568
1569 /*
1570  * perform a "sync" on one "block"
1571  *
1572  * We need to make sure that no normal I/O request - particularly write
1573  * requests - conflict with active sync requests.
1574  *
1575  * This is achieved by tracking pending requests and a 'barrier' concept
1576  * that can be installed to exclude normal IO requests.
1577  */
1578
1579 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1580 {
1581         conf_t *conf = mddev_to_conf(mddev);
1582         r1bio_t *r1_bio;
1583         struct bio *bio;
1584         sector_t max_sector, nr_sectors;
1585         int disk = -1;
1586         int i;
1587         int wonly = -1;
1588         int write_targets = 0, read_targets = 0;
1589         int sync_blocks;
1590         int still_degraded = 0;
1591
1592         if (!conf->r1buf_pool)
1593         {
1594 /*
1595                 printk("sync start - bitmap %p\n", mddev->bitmap);
1596 */
1597                 if (init_resync(conf))
1598                         return 0;
1599         }
1600
1601         max_sector = mddev->size << 1;
1602         if (sector_nr >= max_sector) {
1603                 /* If we aborted, we need to abort the
1604                  * sync on the 'current' bitmap chunk (there will
1605                  * only be one in raid1 resync.
1606                  * We can find the current addess in mddev->curr_resync
1607                  */
1608                 if (mddev->curr_resync < max_sector) /* aborted */
1609                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1610                                                 &sync_blocks, 1);
1611                 else /* completed sync */
1612                         conf->fullsync = 0;
1613
1614                 bitmap_close_sync(mddev->bitmap);
1615                 close_sync(conf);
1616                 return 0;
1617         }
1618
1619         /* before building a request, check if we can skip these blocks..
1620          * This call the bitmap_start_sync doesn't actually record anything
1621          */
1622         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1623             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1624                 /* We can skip this block, and probably several more */
1625                 *skipped = 1;
1626                 return sync_blocks;
1627         }
1628         /*
1629          * If there is non-resync activity waiting for a turn,
1630          * and resync is going fast enough,
1631          * then let it though before starting on this new sync request.
1632          */
1633         if (!go_faster && conf->nr_waiting)
1634                 msleep_interruptible(1000);
1635
1636         raise_barrier(conf);
1637
1638         conf->next_resync = sector_nr;
1639
1640         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1641         rcu_read_lock();
1642         /*
1643          * If we get a correctably read error during resync or recovery,
1644          * we might want to read from a different device.  So we
1645          * flag all drives that could conceivably be read from for READ,
1646          * and any others (which will be non-In_sync devices) for WRITE.
1647          * If a read fails, we try reading from something else for which READ
1648          * is OK.
1649          */
1650
1651         r1_bio->mddev = mddev;
1652         r1_bio->sector = sector_nr;
1653         r1_bio->state = 0;
1654         set_bit(R1BIO_IsSync, &r1_bio->state);
1655
1656         for (i=0; i < conf->raid_disks; i++) {
1657                 mdk_rdev_t *rdev;
1658                 bio = r1_bio->bios[i];
1659
1660                 /* take from bio_init */
1661                 bio->bi_next = NULL;
1662                 bio->bi_flags |= 1 << BIO_UPTODATE;
1663                 bio->bi_rw = 0;
1664                 bio->bi_vcnt = 0;
1665                 bio->bi_idx = 0;
1666                 bio->bi_phys_segments = 0;
1667                 bio->bi_hw_segments = 0;
1668                 bio->bi_size = 0;
1669                 bio->bi_end_io = NULL;
1670                 bio->bi_private = NULL;
1671
1672                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1673                 if (rdev == NULL ||
1674                            test_bit(Faulty, &rdev->flags)) {
1675                         still_degraded = 1;
1676                         continue;
1677                 } else if (!test_bit(In_sync, &rdev->flags)) {
1678                         bio->bi_rw = WRITE;
1679                         bio->bi_end_io = end_sync_write;
1680                         write_targets ++;
1681                 } else {
1682                         /* may need to read from here */
1683                         bio->bi_rw = READ;
1684                         bio->bi_end_io = end_sync_read;
1685                         if (test_bit(WriteMostly, &rdev->flags)) {
1686                                 if (wonly < 0)
1687                                         wonly = i;
1688                         } else {
1689                                 if (disk < 0)
1690                                         disk = i;
1691                         }
1692                         read_targets++;
1693                 }
1694                 atomic_inc(&rdev->nr_pending);
1695                 bio->bi_sector = sector_nr + rdev->data_offset;
1696                 bio->bi_bdev = rdev->bdev;
1697                 bio->bi_private = r1_bio;
1698         }
1699         rcu_read_unlock();
1700         if (disk < 0)
1701                 disk = wonly;
1702         r1_bio->read_disk = disk;
1703
1704         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1705                 /* extra read targets are also write targets */
1706                 write_targets += read_targets-1;
1707
1708         if (write_targets == 0 || read_targets == 0) {
1709                 /* There is nowhere to write, so all non-sync
1710                  * drives must be failed - so we are finished
1711                  */
1712                 sector_t rv = max_sector - sector_nr;
1713                 *skipped = 1;
1714                 put_buf(r1_bio);
1715                 return rv;
1716         }
1717
1718         nr_sectors = 0;
1719         sync_blocks = 0;
1720         do {
1721                 struct page *page;
1722                 int len = PAGE_SIZE;
1723                 if (sector_nr + (len>>9) > max_sector)
1724                         len = (max_sector - sector_nr) << 9;
1725                 if (len == 0)
1726                         break;
1727                 if (sync_blocks == 0) {
1728                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1729                                                &sync_blocks, still_degraded) &&
1730                             !conf->fullsync &&
1731                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1732                                 break;
1733                         if (sync_blocks < (PAGE_SIZE>>9))
1734                                 BUG();
1735                         if (len > (sync_blocks<<9))
1736                                 len = sync_blocks<<9;
1737                 }
1738
1739                 for (i=0 ; i < conf->raid_disks; i++) {
1740                         bio = r1_bio->bios[i];
1741                         if (bio->bi_end_io) {
1742                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1743                                 if (bio_add_page(bio, page, len, 0) == 0) {
1744                                         /* stop here */
1745                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1746                                         while (i > 0) {
1747                                                 i--;
1748                                                 bio = r1_bio->bios[i];
1749                                                 if (bio->bi_end_io==NULL)
1750                                                         continue;
1751                                                 /* remove last page from this bio */
1752                                                 bio->bi_vcnt--;
1753                                                 bio->bi_size -= len;
1754                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1755                                         }
1756                                         goto bio_full;
1757                                 }
1758                         }
1759                 }
1760                 nr_sectors += len>>9;
1761                 sector_nr += len>>9;
1762                 sync_blocks -= (len>>9);
1763         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1764  bio_full:
1765         r1_bio->sectors = nr_sectors;
1766
1767         /* For a user-requested sync, we read all readable devices and do a
1768          * compare
1769          */
1770         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1771                 atomic_set(&r1_bio->remaining, read_targets);
1772                 for (i=0; i<conf->raid_disks; i++) {
1773                         bio = r1_bio->bios[i];
1774                         if (bio->bi_end_io == end_sync_read) {
1775                                 md_sync_acct(conf->mirrors[i].rdev->bdev, nr_sectors);
1776                                 generic_make_request(bio);
1777                         }
1778                 }
1779         } else {
1780                 atomic_set(&r1_bio->remaining, 1);
1781                 bio = r1_bio->bios[r1_bio->read_disk];
1782                 md_sync_acct(conf->mirrors[r1_bio->read_disk].rdev->bdev,
1783                              nr_sectors);
1784                 generic_make_request(bio);
1785
1786         }
1787
1788         return nr_sectors;
1789 }
1790
1791 static int run(mddev_t *mddev)
1792 {
1793         conf_t *conf;
1794         int i, j, disk_idx;
1795         mirror_info_t *disk;
1796         mdk_rdev_t *rdev;
1797         struct list_head *tmp;
1798
1799         if (mddev->level != 1) {
1800                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1801                        mdname(mddev), mddev->level);
1802                 goto out;
1803         }
1804         /*
1805          * copy the already verified devices into our private RAID1
1806          * bookkeeping area. [whatever we allocate in run(),
1807          * should be freed in stop()]
1808          */
1809         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1810         mddev->private = conf;
1811         if (!conf)
1812                 goto out_no_mem;
1813
1814         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1815                                  GFP_KERNEL);
1816         if (!conf->mirrors)
1817                 goto out_no_mem;
1818
1819         conf->tmppage = alloc_page(GFP_KERNEL);
1820         if (!conf->tmppage)
1821                 goto out_no_mem;
1822
1823         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1824         if (!conf->poolinfo)
1825                 goto out_no_mem;
1826         conf->poolinfo->mddev = mddev;
1827         conf->poolinfo->raid_disks = mddev->raid_disks;
1828         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1829                                           r1bio_pool_free,
1830                                           conf->poolinfo);
1831         if (!conf->r1bio_pool)
1832                 goto out_no_mem;
1833
1834         ITERATE_RDEV(mddev, rdev, tmp) {
1835                 disk_idx = rdev->raid_disk;
1836                 if (disk_idx >= mddev->raid_disks
1837                     || disk_idx < 0)
1838                         continue;
1839                 disk = conf->mirrors + disk_idx;
1840
1841                 disk->rdev = rdev;
1842
1843                 blk_queue_stack_limits(mddev->queue,
1844                                        rdev->bdev->bd_disk->queue);
1845                 /* as we don't honour merge_bvec_fn, we must never risk
1846                  * violating it, so limit ->max_sector to one PAGE, as
1847                  * a one page request is never in violation.
1848                  */
1849                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1850                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1851                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1852
1853                 disk->head_position = 0;
1854                 if (!test_bit(Faulty, &rdev->flags) && test_bit(In_sync, &rdev->flags))
1855                         conf->working_disks++;
1856         }
1857         conf->raid_disks = mddev->raid_disks;
1858         conf->mddev = mddev;
1859         spin_lock_init(&conf->device_lock);
1860         INIT_LIST_HEAD(&conf->retry_list);
1861         if (conf->working_disks == 1)
1862                 mddev->recovery_cp = MaxSector;
1863
1864         spin_lock_init(&conf->resync_lock);
1865         init_waitqueue_head(&conf->wait_barrier);
1866
1867         bio_list_init(&conf->pending_bio_list);
1868         bio_list_init(&conf->flushing_bio_list);
1869
1870         if (!conf->working_disks) {
1871                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1872                         mdname(mddev));
1873                 goto out_free_conf;
1874         }
1875
1876         mddev->degraded = 0;
1877         for (i = 0; i < conf->raid_disks; i++) {
1878
1879                 disk = conf->mirrors + i;
1880
1881                 if (!disk->rdev) {
1882                         disk->head_position = 0;
1883                         mddev->degraded++;
1884                 }
1885         }
1886
1887         /*
1888          * find the first working one and use it as a starting point
1889          * to read balancing.
1890          */
1891         for (j = 0; j < conf->raid_disks &&
1892                      (!conf->mirrors[j].rdev ||
1893                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
1894                 /* nothing */;
1895         conf->last_used = j;
1896
1897
1898         mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1899         if (!mddev->thread) {
1900                 printk(KERN_ERR
1901                        "raid1: couldn't allocate thread for %s\n",
1902                        mdname(mddev));
1903                 goto out_free_conf;
1904         }
1905
1906         printk(KERN_INFO
1907                 "raid1: raid set %s active with %d out of %d mirrors\n",
1908                 mdname(mddev), mddev->raid_disks - mddev->degraded,
1909                 mddev->raid_disks);
1910         /*
1911          * Ok, everything is just fine now
1912          */
1913         mddev->array_size = mddev->size;
1914
1915         mddev->queue->unplug_fn = raid1_unplug;
1916         mddev->queue->issue_flush_fn = raid1_issue_flush;
1917
1918         return 0;
1919
1920 out_no_mem:
1921         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1922                mdname(mddev));
1923
1924 out_free_conf:
1925         if (conf) {
1926                 if (conf->r1bio_pool)
1927                         mempool_destroy(conf->r1bio_pool);
1928                 kfree(conf->mirrors);
1929                 safe_put_page(conf->tmppage);
1930                 kfree(conf->poolinfo);
1931                 kfree(conf);
1932                 mddev->private = NULL;
1933         }
1934 out:
1935         return -EIO;
1936 }
1937
1938 static int stop(mddev_t *mddev)
1939 {
1940         conf_t *conf = mddev_to_conf(mddev);
1941         struct bitmap *bitmap = mddev->bitmap;
1942         int behind_wait = 0;
1943
1944         /* wait for behind writes to complete */
1945         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
1946                 behind_wait++;
1947                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
1948                 set_current_state(TASK_UNINTERRUPTIBLE);
1949                 schedule_timeout(HZ); /* wait a second */
1950                 /* need to kick something here to make sure I/O goes? */
1951         }
1952
1953         md_unregister_thread(mddev->thread);
1954         mddev->thread = NULL;
1955         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1956         if (conf->r1bio_pool)
1957                 mempool_destroy(conf->r1bio_pool);
1958         kfree(conf->mirrors);
1959         kfree(conf->poolinfo);
1960         kfree(conf);
1961         mddev->private = NULL;
1962         return 0;
1963 }
1964
1965 static int raid1_resize(mddev_t *mddev, sector_t sectors)
1966 {
1967         /* no resync is happening, and there is enough space
1968          * on all devices, so we can resize.
1969          * We need to make sure resync covers any new space.
1970          * If the array is shrinking we should possibly wait until
1971          * any io in the removed space completes, but it hardly seems
1972          * worth it.
1973          */
1974         mddev->array_size = sectors>>1;
1975         set_capacity(mddev->gendisk, mddev->array_size << 1);
1976         mddev->changed = 1;
1977         if (mddev->array_size > mddev->size && mddev->recovery_cp == MaxSector) {
1978                 mddev->recovery_cp = mddev->size << 1;
1979                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1980         }
1981         mddev->size = mddev->array_size;
1982         mddev->resync_max_sectors = sectors;
1983         return 0;
1984 }
1985
1986 static int raid1_reshape(mddev_t *mddev, int raid_disks)
1987 {
1988         /* We need to:
1989          * 1/ resize the r1bio_pool
1990          * 2/ resize conf->mirrors
1991          *
1992          * We allocate a new r1bio_pool if we can.
1993          * Then raise a device barrier and wait until all IO stops.
1994          * Then resize conf->mirrors and swap in the new r1bio pool.
1995          *
1996          * At the same time, we "pack" the devices so that all the missing
1997          * devices have the higher raid_disk numbers.
1998          */
1999         mempool_t *newpool, *oldpool;
2000         struct pool_info *newpoolinfo;
2001         mirror_info_t *newmirrors;
2002         conf_t *conf = mddev_to_conf(mddev);
2003         int cnt;
2004
2005         int d, d2;
2006
2007         if (raid_disks < conf->raid_disks) {
2008                 cnt=0;
2009                 for (d= 0; d < conf->raid_disks; d++)
2010                         if (conf->mirrors[d].rdev)
2011                                 cnt++;
2012                 if (cnt > raid_disks)
2013                         return -EBUSY;
2014         }
2015
2016         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2017         if (!newpoolinfo)
2018                 return -ENOMEM;
2019         newpoolinfo->mddev = mddev;
2020         newpoolinfo->raid_disks = raid_disks;
2021
2022         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2023                                  r1bio_pool_free, newpoolinfo);
2024         if (!newpool) {
2025                 kfree(newpoolinfo);
2026                 return -ENOMEM;
2027         }
2028         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2029         if (!newmirrors) {
2030                 kfree(newpoolinfo);
2031                 mempool_destroy(newpool);
2032                 return -ENOMEM;
2033         }
2034
2035         raise_barrier(conf);
2036
2037         /* ok, everything is stopped */
2038         oldpool = conf->r1bio_pool;
2039         conf->r1bio_pool = newpool;
2040
2041         for (d=d2=0; d < conf->raid_disks; d++)
2042                 if (conf->mirrors[d].rdev) {
2043                         conf->mirrors[d].rdev->raid_disk = d2;
2044                         newmirrors[d2++].rdev = conf->mirrors[d].rdev;
2045                 }
2046         kfree(conf->mirrors);
2047         conf->mirrors = newmirrors;
2048         kfree(conf->poolinfo);
2049         conf->poolinfo = newpoolinfo;
2050
2051         mddev->degraded += (raid_disks - conf->raid_disks);
2052         conf->raid_disks = mddev->raid_disks = raid_disks;
2053
2054         conf->last_used = 0; /* just make sure it is in-range */
2055         lower_barrier(conf);
2056
2057         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2058         md_wakeup_thread(mddev->thread);
2059
2060         mempool_destroy(oldpool);
2061         return 0;
2062 }
2063
2064 static void raid1_quiesce(mddev_t *mddev, int state)
2065 {
2066         conf_t *conf = mddev_to_conf(mddev);
2067
2068         switch(state) {
2069         case 1:
2070                 raise_barrier(conf);
2071                 break;
2072         case 0:
2073                 lower_barrier(conf);
2074                 break;
2075         }
2076 }
2077
2078
2079 static struct mdk_personality raid1_personality =
2080 {
2081         .name           = "raid1",
2082         .level          = 1,
2083         .owner          = THIS_MODULE,
2084         .make_request   = make_request,
2085         .run            = run,
2086         .stop           = stop,
2087         .status         = status,
2088         .error_handler  = error,
2089         .hot_add_disk   = raid1_add_disk,
2090         .hot_remove_disk= raid1_remove_disk,
2091         .spare_active   = raid1_spare_active,
2092         .sync_request   = sync_request,
2093         .resize         = raid1_resize,
2094         .reshape        = raid1_reshape,
2095         .quiesce        = raid1_quiesce,
2096 };
2097
2098 static int __init raid_init(void)
2099 {
2100         return register_md_personality(&raid1_personality);
2101 }
2102
2103 static void raid_exit(void)
2104 {
2105         unregister_md_personality(&raid1_personality);
2106 }
2107
2108 module_init(raid_init);
2109 module_exit(raid_exit);
2110 MODULE_LICENSE("GPL");
2111 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2112 MODULE_ALIAS("md-raid1");
2113 MODULE_ALIAS("md-level-1");