drivers/md/raid1.c

   1 /*
   2  * raid1.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5  *
   6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7  *
   8  * RAID-1 management functions.
   9  *
  10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11  *
  12  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14  *
  15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  16  * bitmapped intelligence in resync:
  17  *
  18  *      - bitmap marked during normal i/o
  19  *      - bitmap used to skip nondirty blocks during sync
  20  *
  21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  22  * - persistent bitmap code
  23  *
  24  * This program is free software; you can redistribute it and/or modify
  25  * it under the terms of the GNU General Public License as published by
  26  * the Free Software Foundation; either version 2, or (at your option)
  27  * any later version.
  28  *
  29  * You should have received a copy of the GNU General Public License
  30  * (for example /usr/src/linux/COPYING); if not, write to the Free
  31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  32  */
  33
  34 #include <linux/delay.h>
  35 #include <linux/blkdev.h>
  36 #include <linux/seq_file.h>
  37 #include "md.h"
  38 #include "dm-bio-list.h"
  39 #include "raid1.h"
  40 #include "bitmap.h"
  41
  42 #define DEBUG 0
  43 #if DEBUG
  44 #define PRINTK(x...) printk(x)
  45 #else
  46 #define PRINTK(x...)
  47 #endif
  48
  49 /*
  50  * Number of guaranteed r1bios in case of extreme VM load:
  51  */
  52 #define NR_RAID1_BIOS 256
  53
  54
  55 static void unplug_slaves(mddev_t *mddev);
  56
  57 static void allow_barrier(conf_t *conf);
  58 static void lower_barrier(conf_t *conf);
  59
  60 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  61 {
  62         struct pool_info *pi = data;
  63         r1bio_t *r1_bio;
  64         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
  65
  66         /* allocate a r1bio with room for raid_disks entries in the bios array */
  67         r1_bio = kzalloc(size, gfp_flags);
  68         if (!r1_bio)
  69                 unplug_slaves(pi->mddev);
  70
  71         return r1_bio;
  72 }
  73
  74 static void r1bio_pool_free(void *r1_bio, void *data)
  75 {
  76         kfree(r1_bio);
  77 }
  78
  79 #define RESYNC_BLOCK_SIZE (64*1024)
  80 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  81 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  82 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  83 #define RESYNC_WINDOW (2048*1024)
  84
  85 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  86 {
  87         struct pool_info *pi = data;
  88         struct page *page;
  89         r1bio_t *r1_bio;
  90         struct bio *bio;
  91         int i, j;
  92
  93         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  94         if (!r1_bio) {
  95                 unplug_slaves(pi->mddev);
  96                 return NULL;
  97         }
  98
  99         /*
 100          * Allocate bios : 1 for reading, n-1 for writing
 101          */
 102         for (j = pi->raid_disks ; j-- ; ) {
 103                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 104                 if (!bio)
 105                         goto out_free_bio;
 106                 r1_bio->bios[j] = bio;
 107         }
 108         /*
 109          * Allocate RESYNC_PAGES data pages and attach them to
 110          * the first bio.
 111          * If this is a user-requested check/repair, allocate
 112          * RESYNC_PAGES for each bio.
 113          */
 114         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 115                 j = pi->raid_disks;
 116         else
 117                 j = 1;
 118         while(j--) {
 119                 bio = r1_bio->bios[j];
 120                 for (i = 0; i < RESYNC_PAGES; i++) {
 121                         page = alloc_page(gfp_flags);
 122                         if (unlikely(!page))
 123                                 goto out_free_pages;
 124
 125                         bio->bi_io_vec[i].bv_page = page;
 126                         bio->bi_vcnt = i+1;
 127                 }
 128         }
 129         /* If not user-requests, copy the page pointers to all bios */
 130         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 131                 for (i=0; i<RESYNC_PAGES ; i++)
 132                         for (j=1; j<pi->raid_disks; j++)
 133                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 134                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 135         }
 136
 137         r1_bio->master_bio = NULL;
 138
 139         return r1_bio;
 140
 141 out_free_pages:
 142         for (j=0 ; j < pi->raid_disks; j++)
 143                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 144                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 145         j = -1;
 146 out_free_bio:
 147         while ( ++j < pi->raid_disks )
 148                 bio_put(r1_bio->bios[j]);
 149         r1bio_pool_free(r1_bio, data);
 150         return NULL;
 151 }
 152
 153 static void r1buf_pool_free(void *__r1_bio, void *data)
 154 {
 155         struct pool_info *pi = data;
 156         int i,j;
 157         r1bio_t *r1bio = __r1_bio;
 158
 159         for (i = 0; i < RESYNC_PAGES; i++)
 160                 for (j = pi->raid_disks; j-- ;) {
 161                         if (j == 0 ||
 162                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 163                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 164                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 165                 }
 166         for (i=0 ; i < pi->raid_disks; i++)
 167                 bio_put(r1bio->bios[i]);
 168
 169         r1bio_pool_free(r1bio, data);
 170 }
 171
 172 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 173 {
 174         int i;
 175
 176         for (i = 0; i < conf->raid_disks; i++) {
 177                 struct bio **bio = r1_bio->bios + i;
 178                 if (*bio && *bio != IO_BLOCKED)
 179                         bio_put(*bio);
 180                 *bio = NULL;
 181         }
 182 }
 183
 184 static void free_r1bio(r1bio_t *r1_bio)
 185 {
 186         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 187
 188         /*
 189          * Wake up any possible resync thread that waits for the device
 190          * to go idle.
 191          */
 192         allow_barrier(conf);
 193
 194         put_all_bios(conf, r1_bio);
 195         mempool_free(r1_bio, conf->r1bio_pool);
 196 }
 197
 198 static void put_buf(r1bio_t *r1_bio)
 199 {
 200         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 201         int i;
 202
 203         for (i=0; i<conf->raid_disks; i++) {
 204                 struct bio *bio = r1_bio->bios[i];
 205                 if (bio->bi_end_io)
 206                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 207         }
 208
 209         mempool_free(r1_bio, conf->r1buf_pool);
 210
 211         lower_barrier(conf);
 212 }
 213
 214 static void reschedule_retry(r1bio_t *r1_bio)
 215 {
 216         unsigned long flags;
 217         mddev_t *mddev = r1_bio->mddev;
 218         conf_t *conf = mddev_to_conf(mddev);
 219
 220         spin_lock_irqsave(&conf->device_lock, flags);
 221         list_add(&r1_bio->retry_list, &conf->retry_list);
 222         conf->nr_queued ++;
 223         spin_unlock_irqrestore(&conf->device_lock, flags);
 224
 225         wake_up(&conf->wait_barrier);
 226         md_wakeup_thread(mddev->thread);
 227 }
 228
 229 /*
 230  * raid_end_bio_io() is called when we have finished servicing a mirrored
 231  * operation and are ready to return a success/failure code to the buffer
 232  * cache layer.
 233  */
 234 static void raid_end_bio_io(r1bio_t *r1_bio)
 235 {
 236         struct bio *bio = r1_bio->master_bio;
 237
 238         /* if nobody has done the final endio yet, do it now */
 239         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 240                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
 241                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 242                         (unsigned long long) bio->bi_sector,
 243                         (unsigned long long) bio->bi_sector +
 244                                 (bio->bi_size >> 9) - 1);
 245
 246                 bio_endio(bio,
 247                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 248         }
 249         free_r1bio(r1_bio);
 250 }
 251
 252 /*
 253  * Update disk head position estimator based on IRQ completion info.
 254  */
 255 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 256 {
 257         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 258
 259         conf->mirrors[disk].head_position =
 260                 r1_bio->sector + (r1_bio->sectors);
 261 }
 262
 263 static void raid1_end_read_request(struct bio *bio, int error)
 264 {
 265         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 266         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 267         int mirror;
 268         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 269
 270         mirror = r1_bio->read_disk;
 271         /*
 272          * this branch is our 'one mirror IO has finished' event handler:
 273          */
 274         update_head_pos(mirror, r1_bio);
 275
 276         if (uptodate)
 277                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 278         else {
 279                 /* If all other devices have failed, we want to return
 280                  * the error upwards rather than fail the last device.
 281                  * Here we redefine "uptodate" to mean "Don't want to retry"
 282                  */
 283                 unsigned long flags;
 284                 spin_lock_irqsave(&conf->device_lock, flags);
 285                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 286                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 287                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 288                         uptodate = 1;
 289                 spin_unlock_irqrestore(&conf->device_lock, flags);
 290         }
 291
 292         if (uptodate)
 293                 raid_end_bio_io(r1_bio);
 294         else {
 295                 /*
 296                  * oops, read error:
 297                  */
 298                 char b[BDEVNAME_SIZE];
 299                 if (printk_ratelimit())
 300                         printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
 301                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 302                 reschedule_retry(r1_bio);
 303         }
 304
 305         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 306 }
 307
 308 static void raid1_end_write_request(struct bio *bio, int error)
 309 {
 310         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 311         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 312         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 313         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 314         struct bio *to_put = NULL;
 315
 316
 317         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 318                 if (r1_bio->bios[mirror] == bio)
 319                         break;
 320
 321         if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
 322                 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
 323                 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
 324                 r1_bio->mddev->barriers_work = 0;
 325                 /* Don't rdev_dec_pending in this branch - keep it for the retry */
 326         } else {
 327                 /*
 328                  * this branch is our 'one mirror IO has finished' event handler:
 329                  */
 330                 r1_bio->bios[mirror] = NULL;
 331                 to_put = bio;
 332                 if (!uptodate) {
 333                         md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 334                         /* an I/O failed, we can't clear the bitmap */
 335                         set_bit(R1BIO_Degraded, &r1_bio->state);
 336                 } else
 337                         /*
 338                          * Set R1BIO_Uptodate in our master bio, so that
 339                          * we will return a good error code for to the higher
 340                          * levels even if IO on some other mirrored buffer fails.
 341                          *
 342                          * The 'master' represents the composite IO operation to
 343                          * user-side. So if something waits for IO, then it will
 344                          * wait for the 'master' bio.
 345                          */
 346                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 347
 348                 update_head_pos(mirror, r1_bio);
 349
 350                 if (behind) {
 351                         if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 352                                 atomic_dec(&r1_bio->behind_remaining);
 353
 354                         /* In behind mode, we ACK the master bio once the I/O has safely
 355                          * reached all non-writemostly disks. Setting the Returned bit
 356                          * ensures that this gets done only once -- we don't ever want to
 357                          * return -EIO here, instead we'll wait */
 358
 359                         if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 360                             test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 361                                 /* Maybe we can return now */
 362                                 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 363                                         struct bio *mbio = r1_bio->master_bio;
 364                                         PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 365                                                (unsigned long long) mbio->bi_sector,
 366                                                (unsigned long long) mbio->bi_sector +
 367                                                (mbio->bi_size >> 9) - 1);
 368                                         bio_endio(mbio, 0);
 369                                 }
 370                         }
 371                 }
 372                 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 373         }
 374         /*
 375          *
 376          * Let's see if all mirrored write operations have finished
 377          * already.
 378          */
 379         if (atomic_dec_and_test(&r1_bio->remaining)) {
 380                 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
 381                         reschedule_retry(r1_bio);
 382                 else {
 383                         /* it really is the end of this request */
 384                         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 385                                 /* free extra copy of the data pages */
 386                                 int i = bio->bi_vcnt;
 387                                 while (i--)
 388                                         safe_put_page(bio->bi_io_vec[i].bv_page);
 389                         }
 390                         /* clear the bitmap if all writes complete successfully */
 391                         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 392                                         r1_bio->sectors,
 393                                         !test_bit(R1BIO_Degraded, &r1_bio->state),
 394                                         behind);
 395                         md_write_end(r1_bio->mddev);
 396                         raid_end_bio_io(r1_bio);
 397                 }
 398         }
 399
 400         if (to_put)
 401                 bio_put(to_put);
 402 }
 403
 404
 405 /*
 406  * This routine returns the disk from which the requested read should
 407  * be done. There is a per-array 'next expected sequential IO' sector
 408  * number - if this matches on the next IO then we use the last disk.
 409  * There is also a per-disk 'last know head position' sector that is
 410  * maintained from IRQ contexts, both the normal and the resync IO
 411  * completion handlers update this position correctly. If there is no
 412  * perfect sequential match then we pick the disk whose head is closest.
 413  *
 414  * If there are 2 mirrors in the same 2 devices, performance degrades
 415  * because position is mirror, not device based.
 416  *
 417  * The rdev for the device selected will have nr_pending incremented.
 418  */
 419 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 420 {
 421         const unsigned long this_sector = r1_bio->sector;
 422         int new_disk = conf->last_used, disk = new_disk;
 423         int wonly_disk = -1;
 424         const int sectors = r1_bio->sectors;
 425         sector_t new_distance, current_distance;
 426         mdk_rdev_t *rdev;
 427
 428         rcu_read_lock();
 429         /*
 430          * Check if we can balance. We can balance on the whole
 431          * device if no resync is going on, or below the resync window.
 432          * We take the first readable disk when above the resync window.
 433          */
 434  retry:
 435         if (conf->mddev->recovery_cp < MaxSector &&
 436             (this_sector + sectors >= conf->next_resync)) {
 437                 /* Choose the first operation device, for consistancy */
 438                 new_disk = 0;
 439
 440                 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 441                      r1_bio->bios[new_disk] == IO_BLOCKED ||
 442                      !rdev || !test_bit(In_sync, &rdev->flags)
 443                              || test_bit(WriteMostly, &rdev->flags);
 444                      rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
 445
 446                         if (rdev && test_bit(In_sync, &rdev->flags) &&
 447                                 r1_bio->bios[new_disk] != IO_BLOCKED)
 448                                 wonly_disk = new_disk;
 449
 450                         if (new_disk == conf->raid_disks - 1) {
 451                                 new_disk = wonly_disk;
 452                                 break;
 453                         }
 454                 }
 455                 goto rb_out;
 456         }
 457
 458
 459         /* make sure the disk is operational */
 460         for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 461              r1_bio->bios[new_disk] == IO_BLOCKED ||
 462              !rdev || !test_bit(In_sync, &rdev->flags) ||
 463                      test_bit(WriteMostly, &rdev->flags);
 464              rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
 465
 466                 if (rdev && test_bit(In_sync, &rdev->flags) &&
 467                     r1_bio->bios[new_disk] != IO_BLOCKED)
 468                         wonly_disk = new_disk;
 469
 470                 if (new_disk <= 0)
 471                         new_disk = conf->raid_disks;
 472                 new_disk--;
 473                 if (new_disk == disk) {
 474                         new_disk = wonly_disk;
 475                         break;
 476                 }
 477         }
 478
 479         if (new_disk < 0)
 480                 goto rb_out;
 481
 482         disk = new_disk;
 483         /* now disk == new_disk == starting point for search */
 484
 485         /*
 486          * Don't change to another disk for sequential reads:
 487          */
 488         if (conf->next_seq_sect == this_sector)
 489                 goto rb_out;
 490         if (this_sector == conf->mirrors[new_disk].head_position)
 491                 goto rb_out;
 492
 493         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 494
 495         /* Find the disk whose head is closest */
 496
 497         do {
 498                 if (disk <= 0)
 499                         disk = conf->raid_disks;
 500                 disk--;
 501
 502                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 503
 504                 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
 505                     !test_bit(In_sync, &rdev->flags) ||
 506                     test_bit(WriteMostly, &rdev->flags))
 507                         continue;
 508
 509                 if (!atomic_read(&rdev->nr_pending)) {
 510                         new_disk = disk;
 511                         break;
 512                 }
 513                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 514                 if (new_distance < current_distance) {
 515                         current_distance = new_distance;
 516                         new_disk = disk;
 517                 }
 518         } while (disk != conf->last_used);
 519
 520  rb_out:
 521
 522
 523         if (new_disk >= 0) {
 524                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 525                 if (!rdev)
 526                         goto retry;
 527                 atomic_inc(&rdev->nr_pending);
 528                 if (!test_bit(In_sync, &rdev->flags)) {
 529                         /* cannot risk returning a device that failed
 530                          * before we inc'ed nr_pending
 531                          */
 532                         rdev_dec_pending(rdev, conf->mddev);
 533                         goto retry;
 534                 }
 535                 conf->next_seq_sect = this_sector + sectors;
 536                 conf->last_used = new_disk;
 537         }
 538         rcu_read_unlock();
 539
 540         return new_disk;
 541 }
 542
 543 static void unplug_slaves(mddev_t *mddev)
 544 {
 545         conf_t *conf = mddev_to_conf(mddev);
 546         int i;
 547
 548         rcu_read_lock();
 549         for (i=0; i<mddev->raid_disks; i++) {
 550                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 551                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 552                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 553
 554                         atomic_inc(&rdev->nr_pending);
 555                         rcu_read_unlock();
 556
 557                         blk_unplug(r_queue);
 558
 559                         rdev_dec_pending(rdev, mddev);
 560                         rcu_read_lock();
 561                 }
 562         }
 563         rcu_read_unlock();
 564 }
 565
 566 static void raid1_unplug(struct request_queue *q)
 567 {
 568         mddev_t *mddev = q->queuedata;
 569
 570         unplug_slaves(mddev);
 571         md_wakeup_thread(mddev->thread);
 572 }
 573
 574 static int raid1_congested(void *data, int bits)
 575 {
 576         mddev_t *mddev = data;
 577         conf_t *conf = mddev_to_conf(mddev);
 578         int i, ret = 0;
 579
 580         rcu_read_lock();
 581         for (i = 0; i < mddev->raid_disks; i++) {
 582                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 583                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 584                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 585
 586                         /* Note the '|| 1' - when read_balance prefers
 587                          * non-congested targets, it can be removed
 588                          */
 589                         if ((bits & (1<<BDI_async_congested)) || 1)
 590                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 591                         else
 592                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 593                 }
 594         }
 595         rcu_read_unlock();
 596         return ret;
 597 }
 598
 599
 600 static int flush_pending_writes(conf_t *conf)
 601 {
 602         /* Any writes that have been queued but are awaiting
 603          * bitmap updates get flushed here.
 604          * We return 1 if any requests were actually submitted.
 605          */
 606         int rv = 0;
 607
 608         spin_lock_irq(&conf->device_lock);
 609
 610         if (conf->pending_bio_list.head) {
 611                 struct bio *bio;
 612                 bio = bio_list_get(&conf->pending_bio_list);
 613                 blk_remove_plug(conf->mddev->queue);
 614                 spin_unlock_irq(&conf->device_lock);
 615                 /* flush any pending bitmap writes to
 616                  * disk before proceeding w/ I/O */
 617                 bitmap_unplug(conf->mddev->bitmap);
 618
 619                 while (bio) { /* submit pending writes */
 620                         struct bio *next = bio->bi_next;
 621                         bio->bi_next = NULL;
 622                         generic_make_request(bio);
 623                         bio = next;
 624                 }
 625                 rv = 1;
 626         } else
 627                 spin_unlock_irq(&conf->device_lock);
 628         return rv;
 629 }
 630
 631 /* Barriers....
 632  * Sometimes we need to suspend IO while we do something else,
 633  * either some resync/recovery, or reconfigure the array.
 634  * To do this we raise a 'barrier'.
 635  * The 'barrier' is a counter that can be raised multiple times
 636  * to count how many activities are happening which preclude
 637  * normal IO.
 638  * We can only raise the barrier if there is no pending IO.
 639  * i.e. if nr_pending == 0.
 640  * We choose only to raise the barrier if no-one is waiting for the
 641  * barrier to go down.  This means that as soon as an IO request
 642  * is ready, no other operations which require a barrier will start
 643  * until the IO request has had a chance.
 644  *
 645  * So: regular IO calls 'wait_barrier'.  When that returns there
 646  *    is no backgroup IO happening,  It must arrange to call
 647  *    allow_barrier when it has finished its IO.
 648  * backgroup IO calls must call raise_barrier.  Once that returns
 649  *    there is no normal IO happeing.  It must arrange to call
 650  *    lower_barrier when the particular background IO completes.
 651  */
 652 #define RESYNC_DEPTH 32
 653
 654 static void raise_barrier(conf_t *conf)
 655 {
 656         spin_lock_irq(&conf->resync_lock);
 657
 658         /* Wait until no block IO is waiting */
 659         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 660                             conf->resync_lock,
 661                             raid1_unplug(conf->mddev->queue));
 662
 663         /* block any new IO from starting */
 664         conf->barrier++;
 665
 666         /* No wait for all pending IO to complete */
 667         wait_event_lock_irq(conf->wait_barrier,
 668                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 669                             conf->resync_lock,
 670                             raid1_unplug(conf->mddev->queue));
 671
 672         spin_unlock_irq(&conf->resync_lock);
 673 }
 674
 675 static void lower_barrier(conf_t *conf)
 676 {
 677         unsigned long flags;
 678         spin_lock_irqsave(&conf->resync_lock, flags);
 679         conf->barrier--;
 680         spin_unlock_irqrestore(&conf->resync_lock, flags);
 681         wake_up(&conf->wait_barrier);
 682 }
 683
 684 static void wait_barrier(conf_t *conf)
 685 {
 686         spin_lock_irq(&conf->resync_lock);
 687         if (conf->barrier) {
 688                 conf->nr_waiting++;
 689                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 690                                     conf->resync_lock,
 691                                     raid1_unplug(conf->mddev->queue));
 692                 conf->nr_waiting--;
 693         }
 694         conf->nr_pending++;
 695         spin_unlock_irq(&conf->resync_lock);
 696 }
 697
 698 static void allow_barrier(conf_t *conf)
 699 {
 700         unsigned long flags;
 701         spin_lock_irqsave(&conf->resync_lock, flags);
 702         conf->nr_pending--;
 703         spin_unlock_irqrestore(&conf->resync_lock, flags);
 704         wake_up(&conf->wait_barrier);
 705 }
 706
 707 static void freeze_array(conf_t *conf)
 708 {
 709         /* stop syncio and normal IO and wait for everything to
 710          * go quite.
 711          * We increment barrier and nr_waiting, and then
 712          * wait until nr_pending match nr_queued+1
 713          * This is called in the context of one normal IO request
 714          * that has failed. Thus any sync request that might be pending
 715          * will be blocked by nr_pending, and we need to wait for
 716          * pending IO requests to complete or be queued for re-try.
 717          * Thus the number queued (nr_queued) plus this request (1)
 718          * must match the number of pending IOs (nr_pending) before
 719          * we continue.
 720          */
 721         spin_lock_irq(&conf->resync_lock);
 722         conf->barrier++;
 723         conf->nr_waiting++;
 724         wait_event_lock_irq(conf->wait_barrier,
 725                             conf->nr_pending == conf->nr_queued+1,
 726                             conf->resync_lock,
 727                             ({ flush_pending_writes(conf);
 728                                raid1_unplug(conf->mddev->queue); }));
 729         spin_unlock_irq(&conf->resync_lock);
 730 }
 731 static void unfreeze_array(conf_t *conf)
 732 {
 733         /* reverse the effect of the freeze */
 734         spin_lock_irq(&conf->resync_lock);
 735         conf->barrier--;
 736         conf->nr_waiting--;
 737         wake_up(&conf->wait_barrier);
 738         spin_unlock_irq(&conf->resync_lock);
 739 }
 740
 741
 742 /* duplicate the data pages for behind I/O */
 743 static struct page **alloc_behind_pages(struct bio *bio)
 744 {
 745         int i;
 746         struct bio_vec *bvec;
 747         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
 748                                         GFP_NOIO);
 749         if (unlikely(!pages))
 750                 goto do_sync_io;
 751
 752         bio_for_each_segment(bvec, bio, i) {
 753                 pages[i] = alloc_page(GFP_NOIO);
 754                 if (unlikely(!pages[i]))
 755                         goto do_sync_io;
 756                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 757                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 758                 kunmap(pages[i]);
 759                 kunmap(bvec->bv_page);
 760         }
 761
 762         return pages;
 763
 764 do_sync_io:
 765         if (pages)
 766                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
 767                         put_page(pages[i]);
 768         kfree(pages);
 769         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 770         return NULL;
 771 }
 772
 773 static int make_request(struct request_queue *q, struct bio * bio)
 774 {
 775         mddev_t *mddev = q->queuedata;
 776         conf_t *conf = mddev_to_conf(mddev);
 777         mirror_info_t *mirror;
 778         r1bio_t *r1_bio;
 779         struct bio *read_bio;
 780         int i, targets = 0, disks;
 781         struct bitmap *bitmap;
 782         unsigned long flags;
 783         struct bio_list bl;
 784         struct page **behind_pages = NULL;
 785         const int rw = bio_data_dir(bio);
 786         const int do_sync = bio_sync(bio);
 787         int cpu, do_barriers;
 788         mdk_rdev_t *blocked_rdev;
 789
 790         /*
 791          * Register the new request and wait if the reconstruction
 792          * thread has put up a bar for new requests.
 793          * Continue immediately if no resync is active currently.
 794          * We test barriers_work *after* md_write_start as md_write_start
 795          * may cause the first superblock write, and that will check out
 796          * if barriers work.
 797          */
 798
 799         md_write_start(mddev, bio); /* wait on superblock update early */
 800
 801         if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
 802                 if (rw == WRITE)
 803                         md_write_end(mddev);
 804                 bio_endio(bio, -EOPNOTSUPP);
 805                 return 0;
 806         }
 807
 808         wait_barrier(conf);
 809
 810         bitmap = mddev->bitmap;
 811
 812         cpu = part_stat_lock();
 813         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
 814         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
 815                       bio_sectors(bio));
 816         part_stat_unlock();
 817
 818         /*
 819          * make_request() can abort the operation when READA is being
 820          * used and no empty request is available.
 821          *
 822          */
 823         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 824
 825         r1_bio->master_bio = bio;
 826         r1_bio->sectors = bio->bi_size >> 9;
 827         r1_bio->state = 0;
 828         r1_bio->mddev = mddev;
 829         r1_bio->sector = bio->bi_sector;
 830
 831         if (rw == READ) {
 832                 /*
 833                  * read balancing logic:
 834                  */
 835                 int rdisk = read_balance(conf, r1_bio);
 836
 837                 if (rdisk < 0) {
 838                         /* couldn't find anywhere to read from */
 839                         raid_end_bio_io(r1_bio);
 840                         return 0;
 841                 }
 842                 mirror = conf->mirrors + rdisk;
 843
 844                 r1_bio->read_disk = rdisk;
 845
 846                 read_bio = bio_clone(bio, GFP_NOIO);
 847
 848                 r1_bio->bios[rdisk] = read_bio;
 849
 850                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 851                 read_bio->bi_bdev = mirror->rdev->bdev;
 852                 read_bio->bi_end_io = raid1_end_read_request;
 853                 read_bio->bi_rw = READ | do_sync;
 854                 read_bio->bi_private = r1_bio;
 855
 856                 generic_make_request(read_bio);
 857                 return 0;
 858         }
 859
 860         /*
 861          * WRITE:
 862          */
 863         /* first select target devices under spinlock and
 864          * inc refcount on their rdev.  Record them by setting
 865          * bios[x] to bio
 866          */
 867         disks = conf->raid_disks;
 868 #if 0
 869         { static int first=1;
 870         if (first) printk("First Write sector %llu disks %d\n",
 871                           (unsigned long long)r1_bio->sector, disks);
 872         first = 0;
 873         }
 874 #endif
 875  retry_write:
 876         blocked_rdev = NULL;
 877         rcu_read_lock();
 878         for (i = 0;  i < disks; i++) {
 879                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 880                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 881                         atomic_inc(&rdev->nr_pending);
 882                         blocked_rdev = rdev;
 883                         break;
 884                 }
 885                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 886                         atomic_inc(&rdev->nr_pending);
 887                         if (test_bit(Faulty, &rdev->flags)) {
 888                                 rdev_dec_pending(rdev, mddev);
 889                                 r1_bio->bios[i] = NULL;
 890                         } else
 891                                 r1_bio->bios[i] = bio;
 892                         targets++;
 893                 } else
 894                         r1_bio->bios[i] = NULL;
 895         }
 896         rcu_read_unlock();
 897
 898         if (unlikely(blocked_rdev)) {
 899                 /* Wait for this device to become unblocked */
 900                 int j;
 901
 902                 for (j = 0; j < i; j++)
 903                         if (r1_bio->bios[j])
 904                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
 905
 906                 allow_barrier(conf);
 907                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 908                 wait_barrier(conf);
 909                 goto retry_write;
 910         }
 911
 912         BUG_ON(targets == 0); /* we never fail the last device */
 913
 914         if (targets < conf->raid_disks) {
 915                 /* array is degraded, we will not clear the bitmap
 916                  * on I/O completion (see raid1_end_write_request) */
 917                 set_bit(R1BIO_Degraded, &r1_bio->state);
 918         }
 919
 920         /* do behind I/O ? */
 921         if (bitmap &&
 922             atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
 923             (behind_pages = alloc_behind_pages(bio)) != NULL)
 924                 set_bit(R1BIO_BehindIO, &r1_bio->state);
 925
 926         atomic_set(&r1_bio->remaining, 0);
 927         atomic_set(&r1_bio->behind_remaining, 0);
 928
 929         do_barriers = bio_barrier(bio);
 930         if (do_barriers)
 931                 set_bit(R1BIO_Barrier, &r1_bio->state);
 932
 933         bio_list_init(&bl);
 934         for (i = 0; i < disks; i++) {
 935                 struct bio *mbio;
 936                 if (!r1_bio->bios[i])
 937                         continue;
 938
 939                 mbio = bio_clone(bio, GFP_NOIO);
 940                 r1_bio->bios[i] = mbio;
 941
 942                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 943                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 944                 mbio->bi_end_io = raid1_end_write_request;
 945                 mbio->bi_rw = WRITE | do_barriers | do_sync;
 946                 mbio->bi_private = r1_bio;
 947
 948                 if (behind_pages) {
 949                         struct bio_vec *bvec;
 950                         int j;
 951
 952                         /* Yes, I really want the '__' version so that
 953                          * we clear any unused pointer in the io_vec, rather
 954                          * than leave them unchanged.  This is important
 955                          * because when we come to free the pages, we won't
 956                          * know the originial bi_idx, so we just free
 957                          * them all
 958                          */
 959                         __bio_for_each_segment(bvec, mbio, j, 0)
 960                                 bvec->bv_page = behind_pages[j];
 961                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 962                                 atomic_inc(&r1_bio->behind_remaining);
 963                 }
 964
 965                 atomic_inc(&r1_bio->remaining);
 966
 967                 bio_list_add(&bl, mbio);
 968         }
 969         kfree(behind_pages); /* the behind pages are attached to the bios now */
 970
 971         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
 972                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 973         spin_lock_irqsave(&conf->device_lock, flags);
 974         bio_list_merge(&conf->pending_bio_list, &bl);
 975         bio_list_init(&bl);
 976
 977         blk_plug_device(mddev->queue);
 978         spin_unlock_irqrestore(&conf->device_lock, flags);
 979
 980         /* In case raid1d snuck into freeze_array */
 981         wake_up(&conf->wait_barrier);
 982
 983         if (do_sync)
 984                 md_wakeup_thread(mddev->thread);
 985 #if 0
 986         while ((bio = bio_list_pop(&bl)) != NULL)
 987                 generic_make_request(bio);
 988 #endif
 989
 990         return 0;
 991 }
 992
 993 static void status(struct seq_file *seq, mddev_t *mddev)
 994 {
 995         conf_t *conf = mddev_to_conf(mddev);
 996         int i;
 997
 998         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 999                    conf->raid_disks - mddev->degraded);
1000         rcu_read_lock();
1001         for (i = 0; i < conf->raid_disks; i++) {
1002                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1003                 seq_printf(seq, "%s",
1004                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1005         }
1006         rcu_read_unlock();
1007         seq_printf(seq, "]");
1008 }
1009
1010
1011 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1012 {
1013         char b[BDEVNAME_SIZE];
1014         conf_t *conf = mddev_to_conf(mddev);
1015
1016         /*
1017          * If it is not operational, then we have already marked it as dead
1018          * else if it is the last working disks, ignore the error, let the
1019          * next level up know.
1020          * else mark the drive as failed
1021          */
1022         if (test_bit(In_sync, &rdev->flags)
1023             && (conf->raid_disks - mddev->degraded) == 1) {
1024                 /*
1025                  * Don't fail the drive, act as though we were just a
1026                  * normal single drive.
1027                  * However don't try a recovery from this drive as
1028                  * it is very likely to fail.
1029                  */
1030                 mddev->recovery_disabled = 1;
1031                 return;
1032         }
1033         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1034                 unsigned long flags;
1035                 spin_lock_irqsave(&conf->device_lock, flags);
1036                 mddev->degraded++;
1037                 set_bit(Faulty, &rdev->flags);
1038                 spin_unlock_irqrestore(&conf->device_lock, flags);
1039                 /*
1040                  * if recovery is running, make sure it aborts.
1041                  */
1042                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1043         } else
1044                 set_bit(Faulty, &rdev->flags);
1045         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1046         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device.\n"
1047                 "raid1: Operation continuing on %d devices.\n",
1048                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1049 }
1050
1051 static void print_conf(conf_t *conf)
1052 {
1053         int i;
1054
1055         printk("RAID1 conf printout:\n");
1056         if (!conf) {
1057                 printk("(!conf)\n");
1058                 return;
1059         }
1060         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1061                 conf->raid_disks);
1062
1063         rcu_read_lock();
1064         for (i = 0; i < conf->raid_disks; i++) {
1065                 char b[BDEVNAME_SIZE];
1066                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1067                 if (rdev)
1068                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1069                                i, !test_bit(In_sync, &rdev->flags),
1070                                !test_bit(Faulty, &rdev->flags),
1071                                bdevname(rdev->bdev,b));
1072         }
1073         rcu_read_unlock();
1074 }
1075
1076 static void close_sync(conf_t *conf)
1077 {
1078         wait_barrier(conf);
1079         allow_barrier(conf);
1080
1081         mempool_destroy(conf->r1buf_pool);
1082         conf->r1buf_pool = NULL;
1083 }
1084
1085 static int raid1_spare_active(mddev_t *mddev)
1086 {
1087         int i;
1088         conf_t *conf = mddev->private;
1089
1090         /*
1091          * Find all failed disks within the RAID1 configuration
1092          * and mark them readable.
1093          * Called under mddev lock, so rcu protection not needed.
1094          */
1095         for (i = 0; i < conf->raid_disks; i++) {
1096                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1097                 if (rdev
1098                     && !test_bit(Faulty, &rdev->flags)
1099                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1100                         unsigned long flags;
1101                         spin_lock_irqsave(&conf->device_lock, flags);
1102                         mddev->degraded--;
1103                         spin_unlock_irqrestore(&conf->device_lock, flags);
1104                 }
1105         }
1106
1107         print_conf(conf);
1108         return 0;
1109 }
1110
1111
1112 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1113 {
1114         conf_t *conf = mddev->private;
1115         int err = -EEXIST;
1116         int mirror = 0;
1117         mirror_info_t *p;
1118         int first = 0;
1119         int last = mddev->raid_disks - 1;
1120
1121         if (rdev->raid_disk >= 0)
1122                 first = last = rdev->raid_disk;
1123
1124         for (mirror = first; mirror <= last; mirror++)
1125                 if ( !(p=conf->mirrors+mirror)->rdev) {
1126
1127                         blk_queue_stack_limits(mddev->queue,
1128                                                rdev->bdev->bd_disk->queue);
1129                         /* as we don't honour merge_bvec_fn, we must never risk
1130                          * violating it, so limit ->max_sector to one PAGE, as
1131                          * a one page request is never in violation.
1132                          */
1133                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1134                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1135                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1136
1137                         p->head_position = 0;
1138                         rdev->raid_disk = mirror;
1139                         err = 0;
1140                         /* As all devices are equivalent, we don't need a full recovery
1141                          * if this was recently any drive of the array
1142                          */
1143                         if (rdev->saved_raid_disk < 0)
1144                                 conf->fullsync = 1;
1145                         rcu_assign_pointer(p->rdev, rdev);
1146                         break;
1147                 }
1148
1149         print_conf(conf);
1150         return err;
1151 }
1152
1153 static int raid1_remove_disk(mddev_t *mddev, int number)
1154 {
1155         conf_t *conf = mddev->private;
1156         int err = 0;
1157         mdk_rdev_t *rdev;
1158         mirror_info_t *p = conf->mirrors+ number;
1159
1160         print_conf(conf);
1161         rdev = p->rdev;
1162         if (rdev) {
1163                 if (test_bit(In_sync, &rdev->flags) ||
1164                     atomic_read(&rdev->nr_pending)) {
1165                         err = -EBUSY;
1166                         goto abort;
1167                 }
1168                 /* Only remove non-faulty devices is recovery
1169                  * is not possible.
1170                  */
1171                 if (!test_bit(Faulty, &rdev->flags) &&
1172                     mddev->degraded < conf->raid_disks) {
1173                         err = -EBUSY;
1174                         goto abort;
1175                 }
1176                 p->rdev = NULL;
1177                 synchronize_rcu();
1178                 if (atomic_read(&rdev->nr_pending)) {
1179                         /* lost the race, try later */
1180                         err = -EBUSY;
1181                         p->rdev = rdev;
1182                 }
1183         }
1184 abort:
1185
1186         print_conf(conf);
1187         return err;
1188 }
1189
1190
1191 static void end_sync_read(struct bio *bio, int error)
1192 {
1193         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1194         int i;
1195
1196         for (i=r1_bio->mddev->raid_disks; i--; )
1197                 if (r1_bio->bios[i] == bio)
1198                         break;
1199         BUG_ON(i < 0);
1200         update_head_pos(i, r1_bio);
1201         /*
1202          * we have read a block, now it needs to be re-written,
1203          * or re-read if the read failed.
1204          * We don't do much here, just schedule handling by raid1d
1205          */
1206         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1207                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1208
1209         if (atomic_dec_and_test(&r1_bio->remaining))
1210                 reschedule_retry(r1_bio);
1211 }
1212
1213 static void end_sync_write(struct bio *bio, int error)
1214 {
1215         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1216         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1217         mddev_t *mddev = r1_bio->mddev;
1218         conf_t *conf = mddev_to_conf(mddev);
1219         int i;
1220         int mirror=0;
1221
1222         for (i = 0; i < conf->raid_disks; i++)
1223                 if (r1_bio->bios[i] == bio) {
1224                         mirror = i;
1225                         break;
1226                 }
1227         if (!uptodate) {
1228                 int sync_blocks = 0;
1229                 sector_t s = r1_bio->sector;
1230                 long sectors_to_go = r1_bio->sectors;
1231                 /* make sure these bits doesn't get cleared. */
1232                 do {
1233                         bitmap_end_sync(mddev->bitmap, s,
1234                                         &sync_blocks, 1);
1235                         s += sync_blocks;
1236                         sectors_to_go -= sync_blocks;
1237                 } while (sectors_to_go > 0);
1238                 md_error(mddev, conf->mirrors[mirror].rdev);
1239         }
1240
1241         update_head_pos(mirror, r1_bio);
1242
1243         if (atomic_dec_and_test(&r1_bio->remaining)) {
1244                 sector_t s = r1_bio->sectors;
1245                 put_buf(r1_bio);
1246                 md_done_sync(mddev, s, uptodate);
1247         }
1248 }
1249
1250 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1251 {
1252         conf_t *conf = mddev_to_conf(mddev);
1253         int i;
1254         int disks = conf->raid_disks;
1255         struct bio *bio, *wbio;
1256
1257         bio = r1_bio->bios[r1_bio->read_disk];
1258
1259
1260         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1261                 /* We have read all readable devices.  If we haven't
1262                  * got the block, then there is no hope left.
1263                  * If we have, then we want to do a comparison
1264                  * and skip the write if everything is the same.
1265                  * If any blocks failed to read, then we need to
1266                  * attempt an over-write
1267                  */
1268                 int primary;
1269                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1270                         for (i=0; i<mddev->raid_disks; i++)
1271                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1272                                         md_error(mddev, conf->mirrors[i].rdev);
1273
1274                         md_done_sync(mddev, r1_bio->sectors, 1);
1275                         put_buf(r1_bio);
1276                         return;
1277                 }
1278                 for (primary=0; primary<mddev->raid_disks; primary++)
1279                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1280                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1281                                 r1_bio->bios[primary]->bi_end_io = NULL;
1282                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1283                                 break;
1284                         }
1285                 r1_bio->read_disk = primary;
1286                 for (i=0; i<mddev->raid_disks; i++)
1287                         if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1288                                 int j;
1289                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1290                                 struct bio *pbio = r1_bio->bios[primary];
1291                                 struct bio *sbio = r1_bio->bios[i];
1292
1293                                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1294                                         for (j = vcnt; j-- ; ) {
1295                                                 struct page *p, *s;
1296                                                 p = pbio->bi_io_vec[j].bv_page;
1297                                                 s = sbio->bi_io_vec[j].bv_page;
1298                                                 if (memcmp(page_address(p),
1299                                                            page_address(s),
1300                                                            PAGE_SIZE))
1301                                                         break;
1302                                         }
1303                                 } else
1304                                         j = 0;
1305                                 if (j >= 0)
1306                                         mddev->resync_mismatches += r1_bio->sectors;
1307                                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1308                                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1309                                         sbio->bi_end_io = NULL;
1310                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1311                                 } else {
1312                                         /* fixup the bio for reuse */
1313                                         int size;
1314                                         sbio->bi_vcnt = vcnt;
1315                                         sbio->bi_size = r1_bio->sectors << 9;
1316                                         sbio->bi_idx = 0;
1317                                         sbio->bi_phys_segments = 0;
1318                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1319                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1320                                         sbio->bi_next = NULL;
1321                                         sbio->bi_sector = r1_bio->sector +
1322                                                 conf->mirrors[i].rdev->data_offset;
1323                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1324                                         size = sbio->bi_size;
1325                                         for (j = 0; j < vcnt ; j++) {
1326                                                 struct bio_vec *bi;
1327                                                 bi = &sbio->bi_io_vec[j];
1328                                                 bi->bv_offset = 0;
1329                                                 if (size > PAGE_SIZE)
1330                                                         bi->bv_len = PAGE_SIZE;
1331                                                 else
1332                                                         bi->bv_len = size;
1333                                                 size -= PAGE_SIZE;
1334                                                 memcpy(page_address(bi->bv_page),
1335                                                        page_address(pbio->bi_io_vec[j].bv_page),
1336                                                        PAGE_SIZE);
1337                                         }
1338
1339                                 }
1340                         }
1341         }
1342         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1343                 /* ouch - failed to read all of that.
1344                  * Try some synchronous reads of other devices to get
1345                  * good data, much like with normal read errors.  Only
1346                  * read into the pages we already have so we don't
1347                  * need to re-issue the read request.
1348                  * We don't need to freeze the array, because being in an
1349                  * active sync request, there is no normal IO, and
1350                  * no overlapping syncs.
1351                  */
1352                 sector_t sect = r1_bio->sector;
1353                 int sectors = r1_bio->sectors;
1354                 int idx = 0;
1355
1356                 while(sectors) {
1357                         int s = sectors;
1358                         int d = r1_bio->read_disk;
1359                         int success = 0;
1360                         mdk_rdev_t *rdev;
1361
1362                         if (s > (PAGE_SIZE>>9))
1363                                 s = PAGE_SIZE >> 9;
1364                         do {
1365                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1366                                         /* No rcu protection needed here devices
1367                                          * can only be removed when no resync is
1368                                          * active, and resync is currently active
1369                                          */
1370                                         rdev = conf->mirrors[d].rdev;
1371                                         if (sync_page_io(rdev->bdev,
1372                                                          sect + rdev->data_offset,
1373                                                          s<<9,
1374                                                          bio->bi_io_vec[idx].bv_page,
1375                                                          READ)) {
1376                                                 success = 1;
1377                                                 break;
1378                                         }
1379                                 }
1380                                 d++;
1381                                 if (d == conf->raid_disks)
1382                                         d = 0;
1383                         } while (!success && d != r1_bio->read_disk);
1384
1385                         if (success) {
1386                                 int start = d;
1387                                 /* write it back and re-read */
1388                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1389                                 while (d != r1_bio->read_disk) {
1390                                         if (d == 0)
1391                                                 d = conf->raid_disks;
1392                                         d--;
1393                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1394                                                 continue;
1395                                         rdev = conf->mirrors[d].rdev;
1396                                         atomic_add(s, &rdev->corrected_errors);
1397                                         if (sync_page_io(rdev->bdev,
1398                                                          sect + rdev->data_offset,
1399                                                          s<<9,
1400                                                          bio->bi_io_vec[idx].bv_page,
1401                                                          WRITE) == 0)
1402                                                 md_error(mddev, rdev);
1403                                 }
1404                                 d = start;
1405                                 while (d != r1_bio->read_disk) {
1406                                         if (d == 0)
1407                                                 d = conf->raid_disks;
1408                                         d--;
1409                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1410                                                 continue;
1411                                         rdev = conf->mirrors[d].rdev;
1412                                         if (sync_page_io(rdev->bdev,
1413                                                          sect + rdev->data_offset,
1414                                                          s<<9,
1415                                                          bio->bi_io_vec[idx].bv_page,
1416                                                          READ) == 0)
1417                                                 md_error(mddev, rdev);
1418                                 }
1419                         } else {
1420                                 char b[BDEVNAME_SIZE];
1421                                 /* Cannot read from anywhere, array is toast */
1422                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1423                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1424                                        " for block %llu\n",
1425                                        bdevname(bio->bi_bdev,b),
1426                                        (unsigned long long)r1_bio->sector);
1427                                 md_done_sync(mddev, r1_bio->sectors, 0);
1428                                 put_buf(r1_bio);
1429                                 return;
1430                         }
1431                         sectors -= s;
1432                         sect += s;
1433                         idx ++;
1434                 }
1435         }
1436
1437         /*
1438          * schedule writes
1439          */
1440         atomic_set(&r1_bio->remaining, 1);
1441         for (i = 0; i < disks ; i++) {
1442                 wbio = r1_bio->bios[i];
1443                 if (wbio->bi_end_io == NULL ||
1444                     (wbio->bi_end_io == end_sync_read &&
1445                      (i == r1_bio->read_disk ||
1446                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1447                         continue;
1448
1449                 wbio->bi_rw = WRITE;
1450                 wbio->bi_end_io = end_sync_write;
1451                 atomic_inc(&r1_bio->remaining);
1452                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1453
1454                 generic_make_request(wbio);
1455         }
1456
1457         if (atomic_dec_and_test(&r1_bio->remaining)) {
1458                 /* if we're here, all write(s) have completed, so clean up */
1459                 md_done_sync(mddev, r1_bio->sectors, 1);
1460                 put_buf(r1_bio);
1461         }
1462 }
1463
1464 /*
1465  * This is a kernel thread which:
1466  *
1467  *      1.      Retries failed read operations on working mirrors.
1468  *      2.      Updates the raid superblock when problems encounter.
1469  *      3.      Performs writes following reads for array syncronising.
1470  */
1471
1472 static void fix_read_error(conf_t *conf, int read_disk,
1473                            sector_t sect, int sectors)
1474 {
1475         mddev_t *mddev = conf->mddev;
1476         while(sectors) {
1477                 int s = sectors;
1478                 int d = read_disk;
1479                 int success = 0;
1480                 int start;
1481                 mdk_rdev_t *rdev;
1482
1483                 if (s > (PAGE_SIZE>>9))
1484                         s = PAGE_SIZE >> 9;
1485
1486                 do {
1487                         /* Note: no rcu protection needed here
1488                          * as this is synchronous in the raid1d thread
1489                          * which is the thread that might remove
1490                          * a device.  If raid1d ever becomes multi-threaded....
1491                          */
1492                         rdev = conf->mirrors[d].rdev;
1493                         if (rdev &&
1494                             test_bit(In_sync, &rdev->flags) &&
1495                             sync_page_io(rdev->bdev,
1496                                          sect + rdev->data_offset,
1497                                          s<<9,
1498                                          conf->tmppage, READ))
1499                                 success = 1;
1500                         else {
1501                                 d++;
1502                                 if (d == conf->raid_disks)
1503                                         d = 0;
1504                         }
1505                 } while (!success && d != read_disk);
1506
1507                 if (!success) {
1508                         /* Cannot read from anywhere -- bye bye array */
1509                         md_error(mddev, conf->mirrors[read_disk].rdev);
1510                         break;
1511                 }
1512                 /* write it back and re-read */
1513                 start = d;
1514                 while (d != read_disk) {
1515                         if (d==0)
1516                                 d = conf->raid_disks;
1517                         d--;
1518                         rdev = conf->mirrors[d].rdev;
1519                         if (rdev &&
1520                             test_bit(In_sync, &rdev->flags)) {
1521                                 if (sync_page_io(rdev->bdev,
1522                                                  sect + rdev->data_offset,
1523                                                  s<<9, conf->tmppage, WRITE)
1524                                     == 0)
1525                                         /* Well, this device is dead */
1526                                         md_error(mddev, rdev);
1527                         }
1528                 }
1529                 d = start;
1530                 while (d != read_disk) {
1531                         char b[BDEVNAME_SIZE];
1532                         if (d==0)
1533                                 d = conf->raid_disks;
1534                         d--;
1535                         rdev = conf->mirrors[d].rdev;
1536                         if (rdev &&
1537                             test_bit(In_sync, &rdev->flags)) {
1538                                 if (sync_page_io(rdev->bdev,
1539                                                  sect + rdev->data_offset,
1540                                                  s<<9, conf->tmppage, READ)
1541                                     == 0)
1542                                         /* Well, this device is dead */
1543                                         md_error(mddev, rdev);
1544                                 else {
1545                                         atomic_add(s, &rdev->corrected_errors);
1546                                         printk(KERN_INFO
1547                                                "raid1:%s: read error corrected "
1548                                                "(%d sectors at %llu on %s)\n",
1549                                                mdname(mddev), s,
1550                                                (unsigned long long)(sect +
1551                                                    rdev->data_offset),
1552                                                bdevname(rdev->bdev, b));
1553                                 }
1554                         }
1555                 }
1556                 sectors -= s;
1557                 sect += s;
1558         }
1559 }
1560
1561 static void raid1d(mddev_t *mddev)
1562 {
1563         r1bio_t *r1_bio;
1564         struct bio *bio;
1565         unsigned long flags;
1566         conf_t *conf = mddev_to_conf(mddev);
1567         struct list_head *head = &conf->retry_list;
1568         int unplug=0;
1569         mdk_rdev_t *rdev;
1570
1571         md_check_recovery(mddev);
1572
1573         for (;;) {
1574                 char b[BDEVNAME_SIZE];
1575
1576                 unplug += flush_pending_writes(conf);
1577
1578                 spin_lock_irqsave(&conf->device_lock, flags);
1579                 if (list_empty(head)) {
1580                         spin_unlock_irqrestore(&conf->device_lock, flags);
1581                         break;
1582                 }
1583                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1584                 list_del(head->prev);
1585                 conf->nr_queued--;
1586                 spin_unlock_irqrestore(&conf->device_lock, flags);
1587
1588                 mddev = r1_bio->mddev;
1589                 conf = mddev_to_conf(mddev);
1590                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1591                         sync_request_write(mddev, r1_bio);
1592                         unplug = 1;
1593                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1594                         /* some requests in the r1bio were BIO_RW_BARRIER
1595                          * requests which failed with -EOPNOTSUPP.  Hohumm..
1596                          * Better resubmit without the barrier.
1597                          * We know which devices to resubmit for, because
1598                          * all others have had their bios[] entry cleared.
1599                          * We already have a nr_pending reference on these rdevs.
1600                          */
1601                         int i;
1602                         const int do_sync = bio_sync(r1_bio->master_bio);
1603                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1604                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1605                         for (i=0; i < conf->raid_disks; i++)
1606                                 if (r1_bio->bios[i])
1607                                         atomic_inc(&r1_bio->remaining);
1608                         for (i=0; i < conf->raid_disks; i++)
1609                                 if (r1_bio->bios[i]) {
1610                                         struct bio_vec *bvec;
1611                                         int j;
1612
1613                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1614                                         /* copy pages from the failed bio, as
1615                                          * this might be a write-behind device */
1616                                         __bio_for_each_segment(bvec, bio, j, 0)
1617                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1618                                         bio_put(r1_bio->bios[i]);
1619                                         bio->bi_sector = r1_bio->sector +
1620                                                 conf->mirrors[i].rdev->data_offset;
1621                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1622                                         bio->bi_end_io = raid1_end_write_request;
1623                                         bio->bi_rw = WRITE | do_sync;
1624                                         bio->bi_private = r1_bio;
1625                                         r1_bio->bios[i] = bio;
1626                                         generic_make_request(bio);
1627                                 }
1628                 } else {
1629                         int disk;
1630
1631                         /* we got a read error. Maybe the drive is bad.  Maybe just
1632                          * the block and we can fix it.
1633                          * We freeze all other IO, and try reading the block from
1634                          * other devices.  When we find one, we re-write
1635                          * and check it that fixes the read error.
1636                          * This is all done synchronously while the array is
1637                          * frozen
1638                          */
1639                         if (mddev->ro == 0) {
1640                                 freeze_array(conf);
1641                                 fix_read_error(conf, r1_bio->read_disk,
1642                                                r1_bio->sector,
1643                                                r1_bio->sectors);
1644                                 unfreeze_array(conf);
1645                         }
1646
1647                         bio = r1_bio->bios[r1_bio->read_disk];
1648                         if ((disk=read_balance(conf, r1_bio)) == -1 ||
1649                             disk == r1_bio->read_disk) {
1650                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1651                                        " read error for block %llu\n",
1652                                        bdevname(bio->bi_bdev,b),
1653                                        (unsigned long long)r1_bio->sector);
1654                                 raid_end_bio_io(r1_bio);
1655                         } else {
1656                                 const int do_sync = bio_sync(r1_bio->master_bio);
1657                                 r1_bio->bios[r1_bio->read_disk] =
1658                                         mddev->ro ? IO_BLOCKED : NULL;
1659                                 r1_bio->read_disk = disk;
1660                                 bio_put(bio);
1661                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1662                                 r1_bio->bios[r1_bio->read_disk] = bio;
1663                                 rdev = conf->mirrors[disk].rdev;
1664                                 if (printk_ratelimit())
1665                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1666                                                " another mirror\n",
1667                                                bdevname(rdev->bdev,b),
1668                                                (unsigned long long)r1_bio->sector);
1669                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1670                                 bio->bi_bdev = rdev->bdev;
1671                                 bio->bi_end_io = raid1_end_read_request;
1672                                 bio->bi_rw = READ | do_sync;
1673                                 bio->bi_private = r1_bio;
1674                                 unplug = 1;
1675                                 generic_make_request(bio);
1676                         }
1677                 }
1678         }
1679         if (unplug)
1680                 unplug_slaves(mddev);
1681 }
1682
1683
1684 static int init_resync(conf_t *conf)
1685 {
1686         int buffs;
1687
1688         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1689         BUG_ON(conf->r1buf_pool);
1690         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1691                                           conf->poolinfo);
1692         if (!conf->r1buf_pool)
1693                 return -ENOMEM;
1694         conf->next_resync = 0;
1695         return 0;
1696 }
1697
1698 /*
1699  * perform a "sync" on one "block"
1700  *
1701  * We need to make sure that no normal I/O request - particularly write
1702  * requests - conflict with active sync requests.
1703  *
1704  * This is achieved by tracking pending requests and a 'barrier' concept
1705  * that can be installed to exclude normal IO requests.
1706  */
1707
1708 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1709 {
1710         conf_t *conf = mddev_to_conf(mddev);
1711         r1bio_t *r1_bio;
1712         struct bio *bio;
1713         sector_t max_sector, nr_sectors;
1714         int disk = -1;
1715         int i;
1716         int wonly = -1;
1717         int write_targets = 0, read_targets = 0;
1718         int sync_blocks;
1719         int still_degraded = 0;
1720
1721         if (!conf->r1buf_pool)
1722         {
1723 /*
1724                 printk("sync start - bitmap %p\n", mddev->bitmap);
1725 */
1726                 if (init_resync(conf))
1727                         return 0;
1728         }
1729
1730         max_sector = mddev->dev_sectors;
1731         if (sector_nr >= max_sector) {
1732                 /* If we aborted, we need to abort the
1733                  * sync on the 'current' bitmap chunk (there will
1734                  * only be one in raid1 resync.
1735                  * We can find the current addess in mddev->curr_resync
1736                  */
1737                 if (mddev->curr_resync < max_sector) /* aborted */
1738                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1739                                                 &sync_blocks, 1);
1740                 else /* completed sync */
1741                         conf->fullsync = 0;
1742
1743                 bitmap_close_sync(mddev->bitmap);
1744                 close_sync(conf);
1745                 return 0;
1746         }
1747
1748         if (mddev->bitmap == NULL &&
1749             mddev->recovery_cp == MaxSector &&
1750             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1751             conf->fullsync == 0) {
1752                 *skipped = 1;
1753                 return max_sector - sector_nr;
1754         }
1755         /* before building a request, check if we can skip these blocks..
1756          * This call the bitmap_start_sync doesn't actually record anything
1757          */
1758         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1759             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1760                 /* We can skip this block, and probably several more */
1761                 *skipped = 1;
1762                 return sync_blocks;
1763         }
1764         /*
1765          * If there is non-resync activity waiting for a turn,
1766          * and resync is going fast enough,
1767          * then let it though before starting on this new sync request.
1768          */
1769         if (!go_faster && conf->nr_waiting)
1770                 msleep_interruptible(1000);
1771
1772         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1773         raise_barrier(conf);
1774
1775         conf->next_resync = sector_nr;
1776
1777         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1778         rcu_read_lock();
1779         /*
1780          * If we get a correctably read error during resync or recovery,
1781          * we might want to read from a different device.  So we
1782          * flag all drives that could conceivably be read from for READ,
1783          * and any others (which will be non-In_sync devices) for WRITE.
1784          * If a read fails, we try reading from something else for which READ
1785          * is OK.
1786          */
1787
1788         r1_bio->mddev = mddev;
1789         r1_bio->sector = sector_nr;
1790         r1_bio->state = 0;
1791         set_bit(R1BIO_IsSync, &r1_bio->state);
1792
1793         for (i=0; i < conf->raid_disks; i++) {
1794                 mdk_rdev_t *rdev;
1795                 bio = r1_bio->bios[i];
1796
1797                 /* take from bio_init */
1798                 bio->bi_next = NULL;
1799                 bio->bi_flags |= 1 << BIO_UPTODATE;
1800                 bio->bi_rw = READ;
1801                 bio->bi_vcnt = 0;
1802                 bio->bi_idx = 0;
1803                 bio->bi_phys_segments = 0;
1804                 bio->bi_size = 0;
1805                 bio->bi_end_io = NULL;
1806                 bio->bi_private = NULL;
1807
1808                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1809                 if (rdev == NULL ||
1810                            test_bit(Faulty, &rdev->flags)) {
1811                         still_degraded = 1;
1812                         continue;
1813                 } else if (!test_bit(In_sync, &rdev->flags)) {
1814                         bio->bi_rw = WRITE;
1815                         bio->bi_end_io = end_sync_write;
1816                         write_targets ++;
1817                 } else {
1818                         /* may need to read from here */
1819                         bio->bi_rw = READ;
1820                         bio->bi_end_io = end_sync_read;
1821                         if (test_bit(WriteMostly, &rdev->flags)) {
1822                                 if (wonly < 0)
1823                                         wonly = i;
1824                         } else {
1825                                 if (disk < 0)
1826                                         disk = i;
1827                         }
1828                         read_targets++;
1829                 }
1830                 atomic_inc(&rdev->nr_pending);
1831                 bio->bi_sector = sector_nr + rdev->data_offset;
1832                 bio->bi_bdev = rdev->bdev;
1833                 bio->bi_private = r1_bio;
1834         }
1835         rcu_read_unlock();
1836         if (disk < 0)
1837                 disk = wonly;
1838         r1_bio->read_disk = disk;
1839
1840         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1841                 /* extra read targets are also write targets */
1842                 write_targets += read_targets-1;
1843
1844         if (write_targets == 0 || read_targets == 0) {
1845                 /* There is nowhere to write, so all non-sync
1846                  * drives must be failed - so we are finished
1847                  */
1848                 sector_t rv = max_sector - sector_nr;
1849                 *skipped = 1;
1850                 put_buf(r1_bio);
1851                 return rv;
1852         }
1853
1854         if (max_sector > mddev->resync_max)
1855                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1856         nr_sectors = 0;
1857         sync_blocks = 0;
1858         do {
1859                 struct page *page;
1860                 int len = PAGE_SIZE;
1861                 if (sector_nr + (len>>9) > max_sector)
1862                         len = (max_sector - sector_nr) << 9;
1863                 if (len == 0)
1864                         break;
1865                 if (sync_blocks == 0) {
1866                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1867                                                &sync_blocks, still_degraded) &&
1868                             !conf->fullsync &&
1869                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1870                                 break;
1871                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1872                         if (len > (sync_blocks<<9))
1873                                 len = sync_blocks<<9;
1874                 }
1875
1876                 for (i=0 ; i < conf->raid_disks; i++) {
1877                         bio = r1_bio->bios[i];
1878                         if (bio->bi_end_io) {
1879                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1880                                 if (bio_add_page(bio, page, len, 0) == 0) {
1881                                         /* stop here */
1882                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1883                                         while (i > 0) {
1884                                                 i--;
1885                                                 bio = r1_bio->bios[i];
1886                                                 if (bio->bi_end_io==NULL)
1887                                                         continue;
1888                                                 /* remove last page from this bio */
1889                                                 bio->bi_vcnt--;
1890                                                 bio->bi_size -= len;
1891                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1892                                         }
1893                                         goto bio_full;
1894                                 }
1895                         }
1896                 }
1897                 nr_sectors += len>>9;
1898                 sector_nr += len>>9;
1899                 sync_blocks -= (len>>9);
1900         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1901  bio_full:
1902         r1_bio->sectors = nr_sectors;
1903
1904         /* For a user-requested sync, we read all readable devices and do a
1905          * compare
1906          */
1907         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1908                 atomic_set(&r1_bio->remaining, read_targets);
1909                 for (i=0; i<conf->raid_disks; i++) {
1910                         bio = r1_bio->bios[i];
1911                         if (bio->bi_end_io == end_sync_read) {
1912                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1913                                 generic_make_request(bio);
1914                         }
1915                 }
1916         } else {
1917                 atomic_set(&r1_bio->remaining, 1);
1918                 bio = r1_bio->bios[r1_bio->read_disk];
1919                 md_sync_acct(bio->bi_bdev, nr_sectors);
1920                 generic_make_request(bio);
1921
1922         }
1923         return nr_sectors;
1924 }
1925
1926 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1927 {
1928         if (sectors)
1929                 return sectors;
1930
1931         return mddev->dev_sectors;
1932 }
1933
1934 static int run(mddev_t *mddev)
1935 {
1936         conf_t *conf;
1937         int i, j, disk_idx;
1938         mirror_info_t *disk;
1939         mdk_rdev_t *rdev;
1940
1941         if (mddev->level != 1) {
1942                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1943                        mdname(mddev), mddev->level);
1944                 goto out;
1945         }
1946         if (mddev->reshape_position != MaxSector) {
1947                 printk("raid1: %s: reshape_position set but not supported\n",
1948                        mdname(mddev));
1949                 goto out;
1950         }
1951         /*
1952          * copy the already verified devices into our private RAID1
1953          * bookkeeping area. [whatever we allocate in run(),
1954          * should be freed in stop()]
1955          */
1956         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1957         mddev->private = conf;
1958         if (!conf)
1959                 goto out_no_mem;
1960
1961         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1962                                  GFP_KERNEL);
1963         if (!conf->mirrors)
1964                 goto out_no_mem;
1965
1966         conf->tmppage = alloc_page(GFP_KERNEL);
1967         if (!conf->tmppage)
1968                 goto out_no_mem;
1969
1970         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1971         if (!conf->poolinfo)
1972                 goto out_no_mem;
1973         conf->poolinfo->mddev = mddev;
1974         conf->poolinfo->raid_disks = mddev->raid_disks;
1975         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1976                                           r1bio_pool_free,
1977                                           conf->poolinfo);
1978         if (!conf->r1bio_pool)
1979                 goto out_no_mem;
1980
1981         spin_lock_init(&conf->device_lock);
1982         mddev->queue->queue_lock = &conf->device_lock;
1983
1984         list_for_each_entry(rdev, &mddev->disks, same_set) {
1985                 disk_idx = rdev->raid_disk;
1986                 if (disk_idx >= mddev->raid_disks
1987                     || disk_idx < 0)
1988                         continue;
1989                 disk = conf->mirrors + disk_idx;
1990
1991                 disk->rdev = rdev;
1992
1993                 blk_queue_stack_limits(mddev->queue,
1994                                        rdev->bdev->bd_disk->queue);
1995                 /* as we don't honour merge_bvec_fn, we must never risk
1996                  * violating it, so limit ->max_sector to one PAGE, as
1997                  * a one page request is never in violation.
1998                  */
1999                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2000                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
2001                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
2002
2003                 disk->head_position = 0;
2004         }
2005         conf->raid_disks = mddev->raid_disks;
2006         conf->mddev = mddev;
2007         INIT_LIST_HEAD(&conf->retry_list);
2008
2009         spin_lock_init(&conf->resync_lock);
2010         init_waitqueue_head(&conf->wait_barrier);
2011
2012         bio_list_init(&conf->pending_bio_list);
2013         bio_list_init(&conf->flushing_bio_list);
2014
2015
2016         mddev->degraded = 0;
2017         for (i = 0; i < conf->raid_disks; i++) {
2018
2019                 disk = conf->mirrors + i;
2020
2021                 if (!disk->rdev ||
2022                     !test_bit(In_sync, &disk->rdev->flags)) {
2023                         disk->head_position = 0;
2024                         mddev->degraded++;
2025                         if (disk->rdev)
2026                                 conf->fullsync = 1;
2027                 }
2028         }
2029         if (mddev->degraded == conf->raid_disks) {
2030                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
2031                         mdname(mddev));
2032                 goto out_free_conf;
2033         }
2034         if (conf->raid_disks - mddev->degraded == 1)
2035                 mddev->recovery_cp = MaxSector;
2036
2037         /*
2038          * find the first working one and use it as a starting point
2039          * to read balancing.
2040          */
2041         for (j = 0; j < conf->raid_disks &&
2042                      (!conf->mirrors[j].rdev ||
2043                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
2044                 /* nothing */;
2045         conf->last_used = j;
2046
2047
2048         mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
2049         if (!mddev->thread) {
2050                 printk(KERN_ERR
2051                        "raid1: couldn't allocate thread for %s\n",
2052                        mdname(mddev));
2053                 goto out_free_conf;
2054         }
2055
2056         printk(KERN_INFO
2057                 "raid1: raid set %s active with %d out of %d mirrors\n",
2058                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2059                 mddev->raid_disks);
2060         /*
2061          * Ok, everything is just fine now
2062          */
2063         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2064
2065         mddev->queue->unplug_fn = raid1_unplug;
2066         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2067         mddev->queue->backing_dev_info.congested_data = mddev;
2068
2069         return 0;
2070
2071 out_no_mem:
2072         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
2073                mdname(mddev));
2074
2075 out_free_conf:
2076         if (conf) {
2077                 if (conf->r1bio_pool)
2078                         mempool_destroy(conf->r1bio_pool);
2079                 kfree(conf->mirrors);
2080                 safe_put_page(conf->tmppage);
2081                 kfree(conf->poolinfo);
2082                 kfree(conf);
2083                 mddev->private = NULL;
2084         }
2085 out:
2086         return -EIO;
2087 }
2088
2089 static int stop(mddev_t *mddev)
2090 {
2091         conf_t *conf = mddev_to_conf(mddev);
2092         struct bitmap *bitmap = mddev->bitmap;
2093         int behind_wait = 0;
2094
2095         /* wait for behind writes to complete */
2096         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2097                 behind_wait++;
2098                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
2099                 set_current_state(TASK_UNINTERRUPTIBLE);
2100                 schedule_timeout(HZ); /* wait a second */
2101                 /* need to kick something here to make sure I/O goes? */
2102         }
2103
2104         raise_barrier(conf);
2105         lower_barrier(conf);
2106
2107         md_unregister_thread(mddev->thread);
2108         mddev->thread = NULL;
2109         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2110         if (conf->r1bio_pool)
2111                 mempool_destroy(conf->r1bio_pool);
2112         kfree(conf->mirrors);
2113         kfree(conf->poolinfo);
2114         kfree(conf);
2115         mddev->private = NULL;
2116         return 0;
2117 }
2118
2119 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2120 {
2121         /* no resync is happening, and there is enough space
2122          * on all devices, so we can resize.
2123          * We need to make sure resync covers any new space.
2124          * If the array is shrinking we should possibly wait until
2125          * any io in the removed space completes, but it hardly seems
2126          * worth it.
2127          */
2128         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2129         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2130                 return -EINVAL;
2131         set_capacity(mddev->gendisk, mddev->array_sectors);
2132         mddev->changed = 1;
2133         if (sectors > mddev->dev_sectors &&
2134             mddev->recovery_cp == MaxSector) {
2135                 mddev->recovery_cp = mddev->dev_sectors;
2136                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2137         }
2138         mddev->dev_sectors = sectors;
2139         mddev->resync_max_sectors = sectors;
2140         return 0;
2141 }
2142
2143 static int raid1_reshape(mddev_t *mddev)
2144 {
2145         /* We need to:
2146          * 1/ resize the r1bio_pool
2147          * 2/ resize conf->mirrors
2148          *
2149          * We allocate a new r1bio_pool if we can.
2150          * Then raise a device barrier and wait until all IO stops.
2151          * Then resize conf->mirrors and swap in the new r1bio pool.
2152          *
2153          * At the same time, we "pack" the devices so that all the missing
2154          * devices have the higher raid_disk numbers.
2155          */
2156         mempool_t *newpool, *oldpool;
2157         struct pool_info *newpoolinfo;
2158         mirror_info_t *newmirrors;
2159         conf_t *conf = mddev_to_conf(mddev);
2160         int cnt, raid_disks;
2161         unsigned long flags;
2162         int d, d2, err;
2163
2164         /* Cannot change chunk_size, layout, or level */
2165         if (mddev->chunk_size != mddev->new_chunk ||
2166             mddev->layout != mddev->new_layout ||
2167             mddev->level != mddev->new_level) {
2168                 mddev->new_chunk = mddev->chunk_size;
2169                 mddev->new_layout = mddev->layout;
2170                 mddev->new_level = mddev->level;
2171                 return -EINVAL;
2172         }
2173
2174         err = md_allow_write(mddev);
2175         if (err)
2176                 return err;
2177
2178         raid_disks = mddev->raid_disks + mddev->delta_disks;
2179
2180         if (raid_disks < conf->raid_disks) {
2181                 cnt=0;
2182                 for (d= 0; d < conf->raid_disks; d++)
2183                         if (conf->mirrors[d].rdev)
2184                                 cnt++;
2185                 if (cnt > raid_disks)
2186                         return -EBUSY;
2187         }
2188
2189         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2190         if (!newpoolinfo)
2191                 return -ENOMEM;
2192         newpoolinfo->mddev = mddev;
2193         newpoolinfo->raid_disks = raid_disks;
2194
2195         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2196                                  r1bio_pool_free, newpoolinfo);
2197         if (!newpool) {
2198                 kfree(newpoolinfo);
2199                 return -ENOMEM;
2200         }
2201         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2202         if (!newmirrors) {
2203                 kfree(newpoolinfo);
2204                 mempool_destroy(newpool);
2205                 return -ENOMEM;
2206         }
2207
2208         raise_barrier(conf);
2209
2210         /* ok, everything is stopped */
2211         oldpool = conf->r1bio_pool;
2212         conf->r1bio_pool = newpool;
2213
2214         for (d = d2 = 0; d < conf->raid_disks; d++) {
2215                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2216                 if (rdev && rdev->raid_disk != d2) {
2217                         char nm[20];
2218                         sprintf(nm, "rd%d", rdev->raid_disk);
2219                         sysfs_remove_link(&mddev->kobj, nm);
2220                         rdev->raid_disk = d2;
2221                         sprintf(nm, "rd%d", rdev->raid_disk);
2222                         sysfs_remove_link(&mddev->kobj, nm);
2223                         if (sysfs_create_link(&mddev->kobj,
2224                                               &rdev->kobj, nm))
2225                                 printk(KERN_WARNING
2226                                        "md/raid1: cannot register "
2227                                        "%s for %s\n",
2228                                        nm, mdname(mddev));
2229                 }
2230                 if (rdev)
2231                         newmirrors[d2++].rdev = rdev;
2232         }
2233         kfree(conf->mirrors);
2234         conf->mirrors = newmirrors;
2235         kfree(conf->poolinfo);
2236         conf->poolinfo = newpoolinfo;
2237
2238         spin_lock_irqsave(&conf->device_lock, flags);
2239         mddev->degraded += (raid_disks - conf->raid_disks);
2240         spin_unlock_irqrestore(&conf->device_lock, flags);
2241         conf->raid_disks = mddev->raid_disks = raid_disks;
2242         mddev->delta_disks = 0;
2243
2244         conf->last_used = 0; /* just make sure it is in-range */
2245         lower_barrier(conf);
2246
2247         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2248         md_wakeup_thread(mddev->thread);
2249
2250         mempool_destroy(oldpool);
2251         return 0;
2252 }
2253
2254 static void raid1_quiesce(mddev_t *mddev, int state)
2255 {
2256         conf_t *conf = mddev_to_conf(mddev);
2257
2258         switch(state) {
2259         case 1:
2260                 raise_barrier(conf);
2261                 break;
2262         case 0:
2263                 lower_barrier(conf);
2264                 break;
2265         }
2266 }
2267
2268
2269 static struct mdk_personality raid1_personality =
2270 {
2271         .name           = "raid1",
2272         .level          = 1,
2273         .owner          = THIS_MODULE,
2274         .make_request   = make_request,
2275         .run            = run,
2276         .stop           = stop,
2277         .status         = status,
2278         .error_handler  = error,
2279         .hot_add_disk   = raid1_add_disk,
2280         .hot_remove_disk= raid1_remove_disk,
2281         .spare_active   = raid1_spare_active,
2282         .sync_request   = sync_request,
2283         .resize         = raid1_resize,
2284         .size           = raid1_size,
2285         .check_reshape  = raid1_reshape,
2286         .quiesce        = raid1_quiesce,
2287 };
2288
2289 static int __init raid_init(void)
2290 {
2291         return register_md_personality(&raid1_personality);
2292 }
2293
2294 static void raid_exit(void)
2295 {
2296         unregister_md_personality(&raid1_personality);
2297 }
2298
2299 module_init(raid_init);
2300 module_exit(raid_exit);
2301 MODULE_LICENSE("GPL");
2302 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2303 MODULE_ALIAS("md-raid1");
2304 MODULE_ALIAS("md-level-1");