drivers/md/raid1.c

   1 /*
   2  * raid1.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5  *
   6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7  *
   8  * RAID-1 management functions.
   9  *
  10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11  *
  12  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14  *
  15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  16  * bitmapped intelligence in resync:
  17  *
  18  *      - bitmap marked during normal i/o
  19  *      - bitmap used to skip nondirty blocks during sync
  20  *
  21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  22  * - persistent bitmap code
  23  *
  24  * This program is free software; you can redistribute it and/or modify
  25  * it under the terms of the GNU General Public License as published by
  26  * the Free Software Foundation; either version 2, or (at your option)
  27  * any later version.
  28  *
  29  * You should have received a copy of the GNU General Public License
  30  * (for example /usr/src/linux/COPYING); if not, write to the Free
  31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  32  */
  33
  34 #include <linux/slab.h>
  35 #include <linux/delay.h>
  36 #include <linux/blkdev.h>
  37 #include <linux/seq_file.h>
  38 #include "md.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 && pi->mddev)
  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_kmalloc(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 = r1_bio->mddev->private;
 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 = r1_bio->mddev->private;
 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->private;
 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 = r1_bio->mddev->private;
 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 = bio->bi_private;
 267         int mirror;
 268         conf_t *conf = r1_bio->mddev->private;
 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 "md/raid1:%s: %s: rescheduling sector %llu\n",
 301                                mdname(conf->mddev),
 302                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 303                 reschedule_retry(r1_bio);
 304         }
 305
 306         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 307 }
 308
 309 static void r1_bio_write_done(r1bio_t *r1_bio, int vcnt, struct bio_vec *bv,
 310                               int behind)
 311 {
 312         if (atomic_dec_and_test(&r1_bio->remaining))
 313         {
 314                 /* it really is the end of this request */
 315                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 316                         /* free extra copy of the data pages */
 317                         int i = vcnt;
 318                         while (i--)
 319                                 safe_put_page(bv[i].bv_page);
 320                 }
 321                 /* clear the bitmap if all writes complete successfully */
 322                 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 323                                 r1_bio->sectors,
 324                                 !test_bit(R1BIO_Degraded, &r1_bio->state),
 325                                 behind);
 326                 md_write_end(r1_bio->mddev);
 327                 raid_end_bio_io(r1_bio);
 328         }
 329 }
 330
 331 static void raid1_end_write_request(struct bio *bio, int error)
 332 {
 333         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 334         r1bio_t *r1_bio = bio->bi_private;
 335         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 336         conf_t *conf = r1_bio->mddev->private;
 337         struct bio *to_put = NULL;
 338
 339
 340         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 341                 if (r1_bio->bios[mirror] == bio)
 342                         break;
 343
 344         /*
 345          * 'one mirror IO has finished' event handler:
 346          */
 347         r1_bio->bios[mirror] = NULL;
 348         to_put = bio;
 349         if (!uptodate) {
 350                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 351                 /* an I/O failed, we can't clear the bitmap */
 352                 set_bit(R1BIO_Degraded, &r1_bio->state);
 353         } else
 354                 /*
 355                  * Set R1BIO_Uptodate in our master bio, so that we
 356                  * will return a good error code for to the higher
 357                  * levels even if IO on some other mirrored buffer
 358                  * fails.
 359                  *
 360                  * The 'master' represents the composite IO operation
 361                  * to user-side. So if something waits for IO, then it
 362                  * will wait for the 'master' bio.
 363                  */
 364                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 365
 366         update_head_pos(mirror, r1_bio);
 367
 368         if (behind) {
 369                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 370                         atomic_dec(&r1_bio->behind_remaining);
 371
 372                 /*
 373                  * In behind mode, we ACK the master bio once the I/O
 374                  * has safely reached all non-writemostly
 375                  * disks. Setting the Returned bit ensures that this
 376                  * gets done only once -- we don't ever want to return
 377                  * -EIO here, instead we'll wait
 378                  */
 379                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 380                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 381                         /* Maybe we can return now */
 382                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 383                                 struct bio *mbio = r1_bio->master_bio;
 384                                 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 385                                        (unsigned long long) mbio->bi_sector,
 386                                        (unsigned long long) mbio->bi_sector +
 387                                        (mbio->bi_size >> 9) - 1);
 388                                 bio_endio(mbio, 0);
 389                         }
 390                 }
 391         }
 392         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 393
 394         /*
 395          * Let's see if all mirrored write operations have finished
 396          * already.
 397          */
 398         r1_bio_write_done(r1_bio, bio->bi_vcnt, bio->bi_io_vec, behind);
 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 sector_t this_sector = r1_bio->sector;
 422         const int sectors = r1_bio->sectors;
 423         int new_disk = -1;
 424         int start_disk;
 425         int i;
 426         sector_t new_distance, current_distance;
 427         mdk_rdev_t *rdev;
 428         int choose_first;
 429
 430         rcu_read_lock();
 431         /*
 432          * Check if we can balance. We can balance on the whole
 433          * device if no resync is going on, or below the resync window.
 434          * We take the first readable disk when above the resync window.
 435          */
 436  retry:
 437         if (conf->mddev->recovery_cp < MaxSector &&
 438             (this_sector + sectors >= conf->next_resync)) {
 439                 choose_first = 1;
 440                 start_disk = 0;
 441         } else {
 442                 choose_first = 0;
 443                 start_disk = conf->last_used;
 444         }
 445
 446         /* make sure the disk is operational */
 447         for (i = 0 ; i < conf->raid_disks ; i++) {
 448                 int disk = start_disk + i;
 449                 if (disk >= conf->raid_disks)
 450                         disk -= conf->raid_disks;
 451
 452                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 453                 if (r1_bio->bios[disk] == IO_BLOCKED
 454                     || rdev == NULL
 455                     || !test_bit(In_sync, &rdev->flags))
 456                         continue;
 457
 458                 new_disk = disk;
 459                 if (!test_bit(WriteMostly, &rdev->flags))
 460                         break;
 461         }
 462
 463         if (new_disk < 0 || choose_first)
 464                 goto rb_out;
 465
 466         /*
 467          * Don't change to another disk for sequential reads:
 468          */
 469         if (conf->next_seq_sect == this_sector)
 470                 goto rb_out;
 471         if (this_sector == conf->mirrors[new_disk].head_position)
 472                 goto rb_out;
 473
 474         current_distance = abs(this_sector
 475                                - conf->mirrors[new_disk].head_position);
 476
 477         /* look for a better disk - i.e. head is closer */
 478         start_disk = new_disk;
 479         for (i = 1; i < conf->raid_disks; i++) {
 480                 int disk = start_disk + 1;
 481                 if (disk >= conf->raid_disks)
 482                         disk -= conf->raid_disks;
 483
 484                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 485                 if (r1_bio->bios[disk] == IO_BLOCKED
 486                     || rdev == NULL
 487                     || !test_bit(In_sync, &rdev->flags)
 488                     || test_bit(WriteMostly, &rdev->flags))
 489                         continue;
 490
 491                 if (!atomic_read(&rdev->nr_pending)) {
 492                         new_disk = disk;
 493                         break;
 494                 }
 495                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 496                 if (new_distance < current_distance) {
 497                         current_distance = new_distance;
 498                         new_disk = disk;
 499                 }
 500         }
 501
 502  rb_out:
 503         if (new_disk >= 0) {
 504                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 505                 if (!rdev)
 506                         goto retry;
 507                 atomic_inc(&rdev->nr_pending);
 508                 if (!test_bit(In_sync, &rdev->flags)) {
 509                         /* cannot risk returning a device that failed
 510                          * before we inc'ed nr_pending
 511                          */
 512                         rdev_dec_pending(rdev, conf->mddev);
 513                         goto retry;
 514                 }
 515                 conf->next_seq_sect = this_sector + sectors;
 516                 conf->last_used = new_disk;
 517         }
 518         rcu_read_unlock();
 519
 520         return new_disk;
 521 }
 522
 523 static void unplug_slaves(mddev_t *mddev)
 524 {
 525         conf_t *conf = mddev->private;
 526         int i;
 527
 528         rcu_read_lock();
 529         for (i=0; i<mddev->raid_disks; i++) {
 530                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 531                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 532                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 533
 534                         atomic_inc(&rdev->nr_pending);
 535                         rcu_read_unlock();
 536
 537                         blk_unplug(r_queue);
 538
 539                         rdev_dec_pending(rdev, mddev);
 540                         rcu_read_lock();
 541                 }
 542         }
 543         rcu_read_unlock();
 544 }
 545
 546 static void raid1_unplug(struct request_queue *q)
 547 {
 548         mddev_t *mddev = q->queuedata;
 549
 550         unplug_slaves(mddev);
 551         md_wakeup_thread(mddev->thread);
 552 }
 553
 554 static int raid1_congested(void *data, int bits)
 555 {
 556         mddev_t *mddev = data;
 557         conf_t *conf = mddev->private;
 558         int i, ret = 0;
 559
 560         if (mddev_congested(mddev, bits))
 561                 return 1;
 562
 563         rcu_read_lock();
 564         for (i = 0; i < mddev->raid_disks; i++) {
 565                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 566                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 567                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 568
 569                         /* Note the '|| 1' - when read_balance prefers
 570                          * non-congested targets, it can be removed
 571                          */
 572                         if ((bits & (1<<BDI_async_congested)) || 1)
 573                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 574                         else
 575                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 576                 }
 577         }
 578         rcu_read_unlock();
 579         return ret;
 580 }
 581
 582
 583 static int flush_pending_writes(conf_t *conf)
 584 {
 585         /* Any writes that have been queued but are awaiting
 586          * bitmap updates get flushed here.
 587          * We return 1 if any requests were actually submitted.
 588          */
 589         int rv = 0;
 590
 591         spin_lock_irq(&conf->device_lock);
 592
 593         if (conf->pending_bio_list.head) {
 594                 struct bio *bio;
 595                 bio = bio_list_get(&conf->pending_bio_list);
 596                 /* Only take the spinlock to quiet a warning */
 597                 spin_lock(conf->mddev->queue->queue_lock);
 598                 blk_remove_plug(conf->mddev->queue);
 599                 spin_unlock(conf->mddev->queue->queue_lock);
 600                 spin_unlock_irq(&conf->device_lock);
 601                 /* flush any pending bitmap writes to
 602                  * disk before proceeding w/ I/O */
 603                 bitmap_unplug(conf->mddev->bitmap);
 604
 605                 while (bio) { /* submit pending writes */
 606                         struct bio *next = bio->bi_next;
 607                         bio->bi_next = NULL;
 608                         generic_make_request(bio);
 609                         bio = next;
 610                 }
 611                 rv = 1;
 612         } else
 613                 spin_unlock_irq(&conf->device_lock);
 614         return rv;
 615 }
 616
 617 /* Barriers....
 618  * Sometimes we need to suspend IO while we do something else,
 619  * either some resync/recovery, or reconfigure the array.
 620  * To do this we raise a 'barrier'.
 621  * The 'barrier' is a counter that can be raised multiple times
 622  * to count how many activities are happening which preclude
 623  * normal IO.
 624  * We can only raise the barrier if there is no pending IO.
 625  * i.e. if nr_pending == 0.
 626  * We choose only to raise the barrier if no-one is waiting for the
 627  * barrier to go down.  This means that as soon as an IO request
 628  * is ready, no other operations which require a barrier will start
 629  * until the IO request has had a chance.
 630  *
 631  * So: regular IO calls 'wait_barrier'.  When that returns there
 632  *    is no backgroup IO happening,  It must arrange to call
 633  *    allow_barrier when it has finished its IO.
 634  * backgroup IO calls must call raise_barrier.  Once that returns
 635  *    there is no normal IO happeing.  It must arrange to call
 636  *    lower_barrier when the particular background IO completes.
 637  */
 638 #define RESYNC_DEPTH 32
 639
 640 static void raise_barrier(conf_t *conf)
 641 {
 642         spin_lock_irq(&conf->resync_lock);
 643
 644         /* Wait until no block IO is waiting */
 645         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 646                             conf->resync_lock,
 647                             raid1_unplug(conf->mddev->queue));
 648
 649         /* block any new IO from starting */
 650         conf->barrier++;
 651
 652         /* Now wait for all pending IO to complete */
 653         wait_event_lock_irq(conf->wait_barrier,
 654                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 655                             conf->resync_lock,
 656                             raid1_unplug(conf->mddev->queue));
 657
 658         spin_unlock_irq(&conf->resync_lock);
 659 }
 660
 661 static void lower_barrier(conf_t *conf)
 662 {
 663         unsigned long flags;
 664         BUG_ON(conf->barrier <= 0);
 665         spin_lock_irqsave(&conf->resync_lock, flags);
 666         conf->barrier--;
 667         spin_unlock_irqrestore(&conf->resync_lock, flags);
 668         wake_up(&conf->wait_barrier);
 669 }
 670
 671 static void wait_barrier(conf_t *conf)
 672 {
 673         spin_lock_irq(&conf->resync_lock);
 674         if (conf->barrier) {
 675                 conf->nr_waiting++;
 676                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 677                                     conf->resync_lock,
 678                                     raid1_unplug(conf->mddev->queue));
 679                 conf->nr_waiting--;
 680         }
 681         conf->nr_pending++;
 682         spin_unlock_irq(&conf->resync_lock);
 683 }
 684
 685 static void allow_barrier(conf_t *conf)
 686 {
 687         unsigned long flags;
 688         spin_lock_irqsave(&conf->resync_lock, flags);
 689         conf->nr_pending--;
 690         spin_unlock_irqrestore(&conf->resync_lock, flags);
 691         wake_up(&conf->wait_barrier);
 692 }
 693
 694 static void freeze_array(conf_t *conf)
 695 {
 696         /* stop syncio and normal IO and wait for everything to
 697          * go quite.
 698          * We increment barrier and nr_waiting, and then
 699          * wait until nr_pending match nr_queued+1
 700          * This is called in the context of one normal IO request
 701          * that has failed. Thus any sync request that might be pending
 702          * will be blocked by nr_pending, and we need to wait for
 703          * pending IO requests to complete or be queued for re-try.
 704          * Thus the number queued (nr_queued) plus this request (1)
 705          * must match the number of pending IOs (nr_pending) before
 706          * we continue.
 707          */
 708         spin_lock_irq(&conf->resync_lock);
 709         conf->barrier++;
 710         conf->nr_waiting++;
 711         wait_event_lock_irq(conf->wait_barrier,
 712                             conf->nr_pending == conf->nr_queued+1,
 713                             conf->resync_lock,
 714                             ({ flush_pending_writes(conf);
 715                                raid1_unplug(conf->mddev->queue); }));
 716         spin_unlock_irq(&conf->resync_lock);
 717 }
 718 static void unfreeze_array(conf_t *conf)
 719 {
 720         /* reverse the effect of the freeze */
 721         spin_lock_irq(&conf->resync_lock);
 722         conf->barrier--;
 723         conf->nr_waiting--;
 724         wake_up(&conf->wait_barrier);
 725         spin_unlock_irq(&conf->resync_lock);
 726 }
 727
 728
 729 /* duplicate the data pages for behind I/O
 730  * We return a list of bio_vec rather than just page pointers
 731  * as it makes freeing easier
 732  */
 733 static struct bio_vec *alloc_behind_pages(struct bio *bio)
 734 {
 735         int i;
 736         struct bio_vec *bvec;
 737         struct bio_vec *pages = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
 738                                         GFP_NOIO);
 739         if (unlikely(!pages))
 740                 goto do_sync_io;
 741
 742         bio_for_each_segment(bvec, bio, i) {
 743                 pages[i].bv_page = alloc_page(GFP_NOIO);
 744                 if (unlikely(!pages[i].bv_page))
 745                         goto do_sync_io;
 746                 memcpy(kmap(pages[i].bv_page) + bvec->bv_offset,
 747                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 748                 kunmap(pages[i].bv_page);
 749                 kunmap(bvec->bv_page);
 750         }
 751
 752         return pages;
 753
 754 do_sync_io:
 755         if (pages)
 756                 for (i = 0; i < bio->bi_vcnt && pages[i].bv_page; i++)
 757                         put_page(pages[i].bv_page);
 758         kfree(pages);
 759         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 760         return NULL;
 761 }
 762
 763 static int make_request(mddev_t *mddev, struct bio * bio)
 764 {
 765         conf_t *conf = mddev->private;
 766         mirror_info_t *mirror;
 767         r1bio_t *r1_bio;
 768         struct bio *read_bio;
 769         int i, targets = 0, disks;
 770         struct bitmap *bitmap;
 771         unsigned long flags;
 772         struct bio_vec *behind_pages = NULL;
 773         const int rw = bio_data_dir(bio);
 774         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 775         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
 776         mdk_rdev_t *blocked_rdev;
 777
 778         /*
 779          * Register the new request and wait if the reconstruction
 780          * thread has put up a bar for new requests.
 781          * Continue immediately if no resync is active currently.
 782          */
 783
 784         md_write_start(mddev, bio); /* wait on superblock update early */
 785
 786         if (bio_data_dir(bio) == WRITE &&
 787             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
 788             bio->bi_sector < mddev->suspend_hi) {
 789                 /* As the suspend_* range is controlled by
 790                  * userspace, we want an interruptible
 791                  * wait.
 792                  */
 793                 DEFINE_WAIT(w);
 794                 for (;;) {
 795                         flush_signals(current);
 796                         prepare_to_wait(&conf->wait_barrier,
 797                                         &w, TASK_INTERRUPTIBLE);
 798                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
 799                             bio->bi_sector >= mddev->suspend_hi)
 800                                 break;
 801                         schedule();
 802                 }
 803                 finish_wait(&conf->wait_barrier, &w);
 804         }
 805
 806         wait_barrier(conf);
 807
 808         bitmap = mddev->bitmap;
 809
 810         /*
 811          * make_request() can abort the operation when READA is being
 812          * used and no empty request is available.
 813          *
 814          */
 815         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 816
 817         r1_bio->master_bio = bio;
 818         r1_bio->sectors = bio->bi_size >> 9;
 819         r1_bio->state = 0;
 820         r1_bio->mddev = mddev;
 821         r1_bio->sector = bio->bi_sector;
 822
 823         if (rw == READ) {
 824                 /*
 825                  * read balancing logic:
 826                  */
 827                 int rdisk = read_balance(conf, r1_bio);
 828
 829                 if (rdisk < 0) {
 830                         /* couldn't find anywhere to read from */
 831                         raid_end_bio_io(r1_bio);
 832                         return 0;
 833                 }
 834                 mirror = conf->mirrors + rdisk;
 835
 836                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
 837                     bitmap) {
 838                         /* Reading from a write-mostly device must
 839                          * take care not to over-take any writes
 840                          * that are 'behind'
 841                          */
 842                         wait_event(bitmap->behind_wait,
 843                                    atomic_read(&bitmap->behind_writes) == 0);
 844                 }
 845                 r1_bio->read_disk = rdisk;
 846
 847                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 848
 849                 r1_bio->bios[rdisk] = read_bio;
 850
 851                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 852                 read_bio->bi_bdev = mirror->rdev->bdev;
 853                 read_bio->bi_end_io = raid1_end_read_request;
 854                 read_bio->bi_rw = READ | do_sync;
 855                 read_bio->bi_private = r1_bio;
 856
 857                 generic_make_request(read_bio);
 858                 return 0;
 859         }
 860
 861         /*
 862          * WRITE:
 863          */
 864         /* first select target devices under spinlock and
 865          * inc refcount on their rdev.  Record them by setting
 866          * bios[x] to bio
 867          */
 868         disks = conf->raid_disks;
 869  retry_write:
 870         blocked_rdev = NULL;
 871         rcu_read_lock();
 872         for (i = 0;  i < disks; i++) {
 873                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 874                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 875                         atomic_inc(&rdev->nr_pending);
 876                         blocked_rdev = rdev;
 877                         break;
 878                 }
 879                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 880                         atomic_inc(&rdev->nr_pending);
 881                         if (test_bit(Faulty, &rdev->flags)) {
 882                                 rdev_dec_pending(rdev, mddev);
 883                                 r1_bio->bios[i] = NULL;
 884                         } else {
 885                                 r1_bio->bios[i] = bio;
 886                                 targets++;
 887                         }
 888                 } else
 889                         r1_bio->bios[i] = NULL;
 890         }
 891         rcu_read_unlock();
 892
 893         if (unlikely(blocked_rdev)) {
 894                 /* Wait for this device to become unblocked */
 895                 int j;
 896
 897                 for (j = 0; j < i; j++)
 898                         if (r1_bio->bios[j])
 899                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
 900
 901                 allow_barrier(conf);
 902                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 903                 wait_barrier(conf);
 904                 goto retry_write;
 905         }
 906
 907         BUG_ON(targets == 0); /* we never fail the last device */
 908
 909         if (targets < conf->raid_disks) {
 910                 /* array is degraded, we will not clear the bitmap
 911                  * on I/O completion (see raid1_end_write_request) */
 912                 set_bit(R1BIO_Degraded, &r1_bio->state);
 913         }
 914
 915         /* do behind I/O ?
 916          * Not if there are too many, or cannot allocate memory,
 917          * or a reader on WriteMostly is waiting for behind writes
 918          * to flush */
 919         if (bitmap &&
 920             (atomic_read(&bitmap->behind_writes)
 921              < mddev->bitmap_info.max_write_behind) &&
 922             !waitqueue_active(&bitmap->behind_wait) &&
 923             (behind_pages = alloc_behind_pages(bio)) != NULL)
 924                 set_bit(R1BIO_BehindIO, &r1_bio->state);
 925
 926         atomic_set(&r1_bio->remaining, 1);
 927         atomic_set(&r1_bio->behind_remaining, 0);
 928
 929         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
 930                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 931         for (i = 0; i < disks; i++) {
 932                 struct bio *mbio;
 933                 if (!r1_bio->bios[i])
 934                         continue;
 935
 936                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 937                 r1_bio->bios[i] = mbio;
 938
 939                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 940                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 941                 mbio->bi_end_io = raid1_end_write_request;
 942                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
 943                 mbio->bi_private = r1_bio;
 944
 945                 if (behind_pages) {
 946                         struct bio_vec *bvec;
 947                         int j;
 948
 949                         /* Yes, I really want the '__' version so that
 950                          * we clear any unused pointer in the io_vec, rather
 951                          * than leave them unchanged.  This is important
 952                          * because when we come to free the pages, we won't
 953                          * know the original bi_idx, so we just free
 954                          * them all
 955                          */
 956                         __bio_for_each_segment(bvec, mbio, j, 0)
 957                                 bvec->bv_page = behind_pages[j].bv_page;
 958                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 959                                 atomic_inc(&r1_bio->behind_remaining);
 960                 }
 961
 962                 atomic_inc(&r1_bio->remaining);
 963                 spin_lock_irqsave(&conf->device_lock, flags);
 964                 bio_list_add(&conf->pending_bio_list, mbio);
 965                 blk_plug_device_unlocked(mddev->queue);
 966                 spin_unlock_irqrestore(&conf->device_lock, flags);
 967         }
 968         r1_bio_write_done(r1_bio, bio->bi_vcnt, behind_pages, behind_pages != NULL);
 969         kfree(behind_pages); /* the behind pages are attached to the bios now */
 970
 971         /* In case raid1d snuck in to freeze_array */
 972         wake_up(&conf->wait_barrier);
 973
 974         if (do_sync)
 975                 md_wakeup_thread(mddev->thread);
 976
 977         return 0;
 978 }
 979
 980 static void status(struct seq_file *seq, mddev_t *mddev)
 981 {
 982         conf_t *conf = mddev->private;
 983         int i;
 984
 985         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 986                    conf->raid_disks - mddev->degraded);
 987         rcu_read_lock();
 988         for (i = 0; i < conf->raid_disks; i++) {
 989                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 990                 seq_printf(seq, "%s",
 991                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
 992         }
 993         rcu_read_unlock();
 994         seq_printf(seq, "]");
 995 }
 996
 997
 998 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 999 {
1000         char b[BDEVNAME_SIZE];
1001         conf_t *conf = mddev->private;
1002
1003         /*
1004          * If it is not operational, then we have already marked it as dead
1005          * else if it is the last working disks, ignore the error, let the
1006          * next level up know.
1007          * else mark the drive as failed
1008          */
1009         if (test_bit(In_sync, &rdev->flags)
1010             && (conf->raid_disks - mddev->degraded) == 1) {
1011                 /*
1012                  * Don't fail the drive, act as though we were just a
1013                  * normal single drive.
1014                  * However don't try a recovery from this drive as
1015                  * it is very likely to fail.
1016                  */
1017                 mddev->recovery_disabled = 1;
1018                 return;
1019         }
1020         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1021                 unsigned long flags;
1022                 spin_lock_irqsave(&conf->device_lock, flags);
1023                 mddev->degraded++;
1024                 set_bit(Faulty, &rdev->flags);
1025                 spin_unlock_irqrestore(&conf->device_lock, flags);
1026                 /*
1027                  * if recovery is running, make sure it aborts.
1028                  */
1029                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1030         } else
1031                 set_bit(Faulty, &rdev->flags);
1032         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1033         printk(KERN_ALERT
1034                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1035                "md/raid1:%s: Operation continuing on %d devices.\n",
1036                mdname(mddev), bdevname(rdev->bdev, b),
1037                mdname(mddev), conf->raid_disks - mddev->degraded);
1038 }
1039
1040 static void print_conf(conf_t *conf)
1041 {
1042         int i;
1043
1044         printk(KERN_DEBUG "RAID1 conf printout:\n");
1045         if (!conf) {
1046                 printk(KERN_DEBUG "(!conf)\n");
1047                 return;
1048         }
1049         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1050                 conf->raid_disks);
1051
1052         rcu_read_lock();
1053         for (i = 0; i < conf->raid_disks; i++) {
1054                 char b[BDEVNAME_SIZE];
1055                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1056                 if (rdev)
1057                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1058                                i, !test_bit(In_sync, &rdev->flags),
1059                                !test_bit(Faulty, &rdev->flags),
1060                                bdevname(rdev->bdev,b));
1061         }
1062         rcu_read_unlock();
1063 }
1064
1065 static void close_sync(conf_t *conf)
1066 {
1067         wait_barrier(conf);
1068         allow_barrier(conf);
1069
1070         mempool_destroy(conf->r1buf_pool);
1071         conf->r1buf_pool = NULL;
1072 }
1073
1074 static int raid1_spare_active(mddev_t *mddev)
1075 {
1076         int i;
1077         conf_t *conf = mddev->private;
1078         int count = 0;
1079         unsigned long flags;
1080
1081         /*
1082          * Find all failed disks within the RAID1 configuration
1083          * and mark them readable.
1084          * Called under mddev lock, so rcu protection not needed.
1085          */
1086         for (i = 0; i < conf->raid_disks; i++) {
1087                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1088                 if (rdev
1089                     && !test_bit(Faulty, &rdev->flags)
1090                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1091                         count++;
1092                         sysfs_notify_dirent(rdev->sysfs_state);
1093                 }
1094         }
1095         spin_lock_irqsave(&conf->device_lock, flags);
1096         mddev->degraded -= count;
1097         spin_unlock_irqrestore(&conf->device_lock, flags);
1098
1099         print_conf(conf);
1100         return count;
1101 }
1102
1103
1104 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1105 {
1106         conf_t *conf = mddev->private;
1107         int err = -EEXIST;
1108         int mirror = 0;
1109         mirror_info_t *p;
1110         int first = 0;
1111         int last = mddev->raid_disks - 1;
1112
1113         if (rdev->raid_disk >= 0)
1114                 first = last = rdev->raid_disk;
1115
1116         for (mirror = first; mirror <= last; mirror++)
1117                 if ( !(p=conf->mirrors+mirror)->rdev) {
1118
1119                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1120                                           rdev->data_offset << 9);
1121                         /* as we don't honour merge_bvec_fn, we must
1122                          * never risk violating it, so limit
1123                          * ->max_segments to one lying with a single
1124                          * page, as a one page request is never in
1125                          * violation.
1126                          */
1127                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1128                                 blk_queue_max_segments(mddev->queue, 1);
1129                                 blk_queue_segment_boundary(mddev->queue,
1130                                                            PAGE_CACHE_SIZE - 1);
1131                         }
1132
1133                         p->head_position = 0;
1134                         rdev->raid_disk = mirror;
1135                         err = 0;
1136                         /* As all devices are equivalent, we don't need a full recovery
1137                          * if this was recently any drive of the array
1138                          */
1139                         if (rdev->saved_raid_disk < 0)
1140                                 conf->fullsync = 1;
1141                         rcu_assign_pointer(p->rdev, rdev);
1142                         break;
1143                 }
1144         md_integrity_add_rdev(rdev, mddev);
1145         print_conf(conf);
1146         return err;
1147 }
1148
1149 static int raid1_remove_disk(mddev_t *mddev, int number)
1150 {
1151         conf_t *conf = mddev->private;
1152         int err = 0;
1153         mdk_rdev_t *rdev;
1154         mirror_info_t *p = conf->mirrors+ number;
1155
1156         print_conf(conf);
1157         rdev = p->rdev;
1158         if (rdev) {
1159                 if (test_bit(In_sync, &rdev->flags) ||
1160                     atomic_read(&rdev->nr_pending)) {
1161                         err = -EBUSY;
1162                         goto abort;
1163                 }
1164                 /* Only remove non-faulty devices if recovery
1165                  * is not possible.
1166                  */
1167                 if (!test_bit(Faulty, &rdev->flags) &&
1168                     !mddev->recovery_disabled &&
1169                     mddev->degraded < conf->raid_disks) {
1170                         err = -EBUSY;
1171                         goto abort;
1172                 }
1173                 p->rdev = NULL;
1174                 synchronize_rcu();
1175                 if (atomic_read(&rdev->nr_pending)) {
1176                         /* lost the race, try later */
1177                         err = -EBUSY;
1178                         p->rdev = rdev;
1179                         goto abort;
1180                 }
1181                 md_integrity_register(mddev);
1182         }
1183 abort:
1184
1185         print_conf(conf);
1186         return err;
1187 }
1188
1189
1190 static void end_sync_read(struct bio *bio, int error)
1191 {
1192         r1bio_t *r1_bio = bio->bi_private;
1193         int i;
1194
1195         for (i=r1_bio->mddev->raid_disks; i--; )
1196                 if (r1_bio->bios[i] == bio)
1197                         break;
1198         BUG_ON(i < 0);
1199         update_head_pos(i, r1_bio);
1200         /*
1201          * we have read a block, now it needs to be re-written,
1202          * or re-read if the read failed.
1203          * We don't do much here, just schedule handling by raid1d
1204          */
1205         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1206                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1207
1208         if (atomic_dec_and_test(&r1_bio->remaining))
1209                 reschedule_retry(r1_bio);
1210 }
1211
1212 static void end_sync_write(struct bio *bio, int error)
1213 {
1214         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1215         r1bio_t *r1_bio = bio->bi_private;
1216         mddev_t *mddev = r1_bio->mddev;
1217         conf_t *conf = mddev->private;
1218         int i;
1219         int mirror=0;
1220
1221         for (i = 0; i < conf->raid_disks; i++)
1222                 if (r1_bio->bios[i] == bio) {
1223                         mirror = i;
1224                         break;
1225                 }
1226         if (!uptodate) {
1227                 sector_t sync_blocks = 0;
1228                 sector_t s = r1_bio->sector;
1229                 long sectors_to_go = r1_bio->sectors;
1230                 /* make sure these bits doesn't get cleared. */
1231                 do {
1232                         bitmap_end_sync(mddev->bitmap, s,
1233                                         &sync_blocks, 1);
1234                         s += sync_blocks;
1235                         sectors_to_go -= sync_blocks;
1236                 } while (sectors_to_go > 0);
1237                 md_error(mddev, conf->mirrors[mirror].rdev);
1238         }
1239
1240         update_head_pos(mirror, r1_bio);
1241
1242         if (atomic_dec_and_test(&r1_bio->remaining)) {
1243                 sector_t s = r1_bio->sectors;
1244                 put_buf(r1_bio);
1245                 md_done_sync(mddev, s, uptodate);
1246         }
1247 }
1248
1249 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1250 {
1251         conf_t *conf = mddev->private;
1252         int i;
1253         int disks = conf->raid_disks;
1254         struct bio *bio, *wbio;
1255
1256         bio = r1_bio->bios[r1_bio->read_disk];
1257
1258
1259         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1260                 /* We have read all readable devices.  If we haven't
1261                  * got the block, then there is no hope left.
1262                  * If we have, then we want to do a comparison
1263                  * and skip the write if everything is the same.
1264                  * If any blocks failed to read, then we need to
1265                  * attempt an over-write
1266                  */
1267                 int primary;
1268                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1269                         for (i=0; i<mddev->raid_disks; i++)
1270                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1271                                         md_error(mddev, conf->mirrors[i].rdev);
1272
1273                         md_done_sync(mddev, r1_bio->sectors, 1);
1274                         put_buf(r1_bio);
1275                         return;
1276                 }
1277                 for (primary=0; primary<mddev->raid_disks; primary++)
1278                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1279                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1280                                 r1_bio->bios[primary]->bi_end_io = NULL;
1281                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1282                                 break;
1283                         }
1284                 r1_bio->read_disk = primary;
1285                 for (i=0; i<mddev->raid_disks; i++)
1286                         if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1287                                 int j;
1288                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1289                                 struct bio *pbio = r1_bio->bios[primary];
1290                                 struct bio *sbio = r1_bio->bios[i];
1291
1292                                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1293                                         for (j = vcnt; j-- ; ) {
1294                                                 struct page *p, *s;
1295                                                 p = pbio->bi_io_vec[j].bv_page;
1296                                                 s = sbio->bi_io_vec[j].bv_page;
1297                                                 if (memcmp(page_address(p),
1298                                                            page_address(s),
1299                                                            PAGE_SIZE))
1300                                                         break;
1301                                         }
1302                                 } else
1303                                         j = 0;
1304                                 if (j >= 0)
1305                                         mddev->resync_mismatches += r1_bio->sectors;
1306                                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1307                                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1308                                         sbio->bi_end_io = NULL;
1309                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1310                                 } else {
1311                                         /* fixup the bio for reuse */
1312                                         int size;
1313                                         sbio->bi_vcnt = vcnt;
1314                                         sbio->bi_size = r1_bio->sectors << 9;
1315                                         sbio->bi_idx = 0;
1316                                         sbio->bi_phys_segments = 0;
1317                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1318                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1319                                         sbio->bi_next = NULL;
1320                                         sbio->bi_sector = r1_bio->sector +
1321                                                 conf->mirrors[i].rdev->data_offset;
1322                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1323                                         size = sbio->bi_size;
1324                                         for (j = 0; j < vcnt ; j++) {
1325                                                 struct bio_vec *bi;
1326                                                 bi = &sbio->bi_io_vec[j];
1327                                                 bi->bv_offset = 0;
1328                                                 if (size > PAGE_SIZE)
1329                                                         bi->bv_len = PAGE_SIZE;
1330                                                 else
1331                                                         bi->bv_len = size;
1332                                                 size -= PAGE_SIZE;
1333                                                 memcpy(page_address(bi->bv_page),
1334                                                        page_address(pbio->bi_io_vec[j].bv_page),
1335                                                        PAGE_SIZE);
1336                                         }
1337
1338                                 }
1339                         }
1340         }
1341         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1342                 /* ouch - failed to read all of that.
1343                  * Try some synchronous reads of other devices to get
1344                  * good data, much like with normal read errors.  Only
1345                  * read into the pages we already have so we don't
1346                  * need to re-issue the read request.
1347                  * We don't need to freeze the array, because being in an
1348                  * active sync request, there is no normal IO, and
1349                  * no overlapping syncs.
1350                  */
1351                 sector_t sect = r1_bio->sector;
1352                 int sectors = r1_bio->sectors;
1353                 int idx = 0;
1354
1355                 while(sectors) {
1356                         int s = sectors;
1357                         int d = r1_bio->read_disk;
1358                         int success = 0;
1359                         mdk_rdev_t *rdev;
1360
1361                         if (s > (PAGE_SIZE>>9))
1362                                 s = PAGE_SIZE >> 9;
1363                         do {
1364                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1365                                         /* No rcu protection needed here devices
1366                                          * can only be removed when no resync is
1367                                          * active, and resync is currently active
1368                                          */
1369                                         rdev = conf->mirrors[d].rdev;
1370                                         if (sync_page_io(rdev,
1371                                                          sect,
1372                                                          s<<9,
1373                                                          bio->bi_io_vec[idx].bv_page,
1374                                                          READ, false)) {
1375                                                 success = 1;
1376                                                 break;
1377                                         }
1378                                 }
1379                                 d++;
1380                                 if (d == conf->raid_disks)
1381                                         d = 0;
1382                         } while (!success && d != r1_bio->read_disk);
1383
1384                         if (success) {
1385                                 int start = d;
1386                                 /* write it back and re-read */
1387                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1388                                 while (d != r1_bio->read_disk) {
1389                                         if (d == 0)
1390                                                 d = conf->raid_disks;
1391                                         d--;
1392                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1393                                                 continue;
1394                                         rdev = conf->mirrors[d].rdev;
1395                                         atomic_add(s, &rdev->corrected_errors);
1396                                         if (sync_page_io(rdev,
1397                                                          sect,
1398                                                          s<<9,
1399                                                          bio->bi_io_vec[idx].bv_page,
1400                                                          WRITE, false) == 0)
1401                                                 md_error(mddev, rdev);
1402                                 }
1403                                 d = start;
1404                                 while (d != r1_bio->read_disk) {
1405                                         if (d == 0)
1406                                                 d = conf->raid_disks;
1407                                         d--;
1408                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1409                                                 continue;
1410                                         rdev = conf->mirrors[d].rdev;
1411                                         if (sync_page_io(rdev,
1412                                                          sect,
1413                                                          s<<9,
1414                                                          bio->bi_io_vec[idx].bv_page,
1415                                                          READ, false) == 0)
1416                                                 md_error(mddev, rdev);
1417                                 }
1418                         } else {
1419                                 char b[BDEVNAME_SIZE];
1420                                 /* Cannot read from anywhere, array is toast */
1421                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1422                                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1423                                        " for block %llu\n",
1424                                        mdname(mddev),
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, sect, s<<9,
1496                                          conf->tmppage, READ, false))
1497                                 success = 1;
1498                         else {
1499                                 d++;
1500                                 if (d == conf->raid_disks)
1501                                         d = 0;
1502                         }
1503                 } while (!success && d != read_disk);
1504
1505                 if (!success) {
1506                         /* Cannot read from anywhere -- bye bye array */
1507                         md_error(mddev, conf->mirrors[read_disk].rdev);
1508                         break;
1509                 }
1510                 /* write it back and re-read */
1511                 start = d;
1512                 while (d != read_disk) {
1513                         if (d==0)
1514                                 d = conf->raid_disks;
1515                         d--;
1516                         rdev = conf->mirrors[d].rdev;
1517                         if (rdev &&
1518                             test_bit(In_sync, &rdev->flags)) {
1519                                 if (sync_page_io(rdev, sect, s<<9,
1520                                                  conf->tmppage, WRITE, false)
1521                                     == 0)
1522                                         /* Well, this device is dead */
1523                                         md_error(mddev, rdev);
1524                         }
1525                 }
1526                 d = start;
1527                 while (d != read_disk) {
1528                         char b[BDEVNAME_SIZE];
1529                         if (d==0)
1530                                 d = conf->raid_disks;
1531                         d--;
1532                         rdev = conf->mirrors[d].rdev;
1533                         if (rdev &&
1534                             test_bit(In_sync, &rdev->flags)) {
1535                                 if (sync_page_io(rdev, sect, s<<9,
1536                                                  conf->tmppage, READ, false)
1537                                     == 0)
1538                                         /* Well, this device is dead */
1539                                         md_error(mddev, rdev);
1540                                 else {
1541                                         atomic_add(s, &rdev->corrected_errors);
1542                                         printk(KERN_INFO
1543                                                "md/raid1:%s: read error corrected "
1544                                                "(%d sectors at %llu on %s)\n",
1545                                                mdname(mddev), s,
1546                                                (unsigned long long)(sect +
1547                                                    rdev->data_offset),
1548                                                bdevname(rdev->bdev, b));
1549                                 }
1550                         }
1551                 }
1552                 sectors -= s;
1553                 sect += s;
1554         }
1555 }
1556
1557 static void raid1d(mddev_t *mddev)
1558 {
1559         r1bio_t *r1_bio;
1560         struct bio *bio;
1561         unsigned long flags;
1562         conf_t *conf = mddev->private;
1563         struct list_head *head = &conf->retry_list;
1564         int unplug=0;
1565         mdk_rdev_t *rdev;
1566
1567         md_check_recovery(mddev);
1568
1569         for (;;) {
1570                 char b[BDEVNAME_SIZE];
1571
1572                 unplug += flush_pending_writes(conf);
1573
1574                 spin_lock_irqsave(&conf->device_lock, flags);
1575                 if (list_empty(head)) {
1576                         spin_unlock_irqrestore(&conf->device_lock, flags);
1577                         break;
1578                 }
1579                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1580                 list_del(head->prev);
1581                 conf->nr_queued--;
1582                 spin_unlock_irqrestore(&conf->device_lock, flags);
1583
1584                 mddev = r1_bio->mddev;
1585                 conf = mddev->private;
1586                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1587                         sync_request_write(mddev, r1_bio);
1588                         unplug = 1;
1589                 } else {
1590                         int disk;
1591
1592                         /* we got a read error. Maybe the drive is bad.  Maybe just
1593                          * the block and we can fix it.
1594                          * We freeze all other IO, and try reading the block from
1595                          * other devices.  When we find one, we re-write
1596                          * and check it that fixes the read error.
1597                          * This is all done synchronously while the array is
1598                          * frozen
1599                          */
1600                         if (mddev->ro == 0) {
1601                                 freeze_array(conf);
1602                                 fix_read_error(conf, r1_bio->read_disk,
1603                                                r1_bio->sector,
1604                                                r1_bio->sectors);
1605                                 unfreeze_array(conf);
1606                         } else
1607                                 md_error(mddev,
1608                                          conf->mirrors[r1_bio->read_disk].rdev);
1609
1610                         bio = r1_bio->bios[r1_bio->read_disk];
1611                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1612                                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1613                                        " read error for block %llu\n",
1614                                        mdname(mddev),
1615                                        bdevname(bio->bi_bdev,b),
1616                                        (unsigned long long)r1_bio->sector);
1617                                 raid_end_bio_io(r1_bio);
1618                         } else {
1619                                 const unsigned long do_sync = r1_bio->master_bio->bi_rw & REQ_SYNC;
1620                                 r1_bio->bios[r1_bio->read_disk] =
1621                                         mddev->ro ? IO_BLOCKED : NULL;
1622                                 r1_bio->read_disk = disk;
1623                                 bio_put(bio);
1624                                 bio = bio_clone_mddev(r1_bio->master_bio,
1625                                                       GFP_NOIO, mddev);
1626                                 r1_bio->bios[r1_bio->read_disk] = bio;
1627                                 rdev = conf->mirrors[disk].rdev;
1628                                 if (printk_ratelimit())
1629                                         printk(KERN_ERR "md/raid1:%s: redirecting sector %llu to"
1630                                                " other mirror: %s\n",
1631                                                mdname(mddev),
1632                                                (unsigned long long)r1_bio->sector,
1633                                                bdevname(rdev->bdev,b));
1634                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1635                                 bio->bi_bdev = rdev->bdev;
1636                                 bio->bi_end_io = raid1_end_read_request;
1637                                 bio->bi_rw = READ | do_sync;
1638                                 bio->bi_private = r1_bio;
1639                                 unplug = 1;
1640                                 generic_make_request(bio);
1641                         }
1642                 }
1643                 cond_resched();
1644         }
1645         if (unplug)
1646                 unplug_slaves(mddev);
1647 }
1648
1649
1650 static int init_resync(conf_t *conf)
1651 {
1652         int buffs;
1653
1654         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1655         BUG_ON(conf->r1buf_pool);
1656         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1657                                           conf->poolinfo);
1658         if (!conf->r1buf_pool)
1659                 return -ENOMEM;
1660         conf->next_resync = 0;
1661         return 0;
1662 }
1663
1664 /*
1665  * perform a "sync" on one "block"
1666  *
1667  * We need to make sure that no normal I/O request - particularly write
1668  * requests - conflict with active sync requests.
1669  *
1670  * This is achieved by tracking pending requests and a 'barrier' concept
1671  * that can be installed to exclude normal IO requests.
1672  */
1673
1674 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1675 {
1676         conf_t *conf = mddev->private;
1677         r1bio_t *r1_bio;
1678         struct bio *bio;
1679         sector_t max_sector, nr_sectors;
1680         int disk = -1;
1681         int i;
1682         int wonly = -1;
1683         int write_targets = 0, read_targets = 0;
1684         sector_t sync_blocks;
1685         int still_degraded = 0;
1686
1687         if (!conf->r1buf_pool)
1688                 if (init_resync(conf))
1689                         return 0;
1690
1691         max_sector = mddev->dev_sectors;
1692         if (sector_nr >= max_sector) {
1693                 /* If we aborted, we need to abort the
1694                  * sync on the 'current' bitmap chunk (there will
1695                  * only be one in raid1 resync.
1696                  * We can find the current addess in mddev->curr_resync
1697                  */
1698                 if (mddev->curr_resync < max_sector) /* aborted */
1699                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1700                                                 &sync_blocks, 1);
1701                 else /* completed sync */
1702                         conf->fullsync = 0;
1703
1704                 bitmap_close_sync(mddev->bitmap);
1705                 close_sync(conf);
1706                 return 0;
1707         }
1708
1709         if (mddev->bitmap == NULL &&
1710             mddev->recovery_cp == MaxSector &&
1711             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1712             conf->fullsync == 0) {
1713                 *skipped = 1;
1714                 return max_sector - sector_nr;
1715         }
1716         /* before building a request, check if we can skip these blocks..
1717          * This call the bitmap_start_sync doesn't actually record anything
1718          */
1719         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1720             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1721                 /* We can skip this block, and probably several more */
1722                 *skipped = 1;
1723                 return sync_blocks;
1724         }
1725         /*
1726          * If there is non-resync activity waiting for a turn,
1727          * and resync is going fast enough,
1728          * then let it though before starting on this new sync request.
1729          */
1730         if (!go_faster && conf->nr_waiting)
1731                 msleep_interruptible(1000);
1732
1733         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1734         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1735         raise_barrier(conf);
1736
1737         conf->next_resync = sector_nr;
1738
1739         rcu_read_lock();
1740         /*
1741          * If we get a correctably read error during resync or recovery,
1742          * we might want to read from a different device.  So we
1743          * flag all drives that could conceivably be read from for READ,
1744          * and any others (which will be non-In_sync devices) for WRITE.
1745          * If a read fails, we try reading from something else for which READ
1746          * is OK.
1747          */
1748
1749         r1_bio->mddev = mddev;
1750         r1_bio->sector = sector_nr;
1751         r1_bio->state = 0;
1752         set_bit(R1BIO_IsSync, &r1_bio->state);
1753
1754         for (i=0; i < conf->raid_disks; i++) {
1755                 mdk_rdev_t *rdev;
1756                 bio = r1_bio->bios[i];
1757
1758                 /* take from bio_init */
1759                 bio->bi_next = NULL;
1760                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
1761                 bio->bi_flags |= 1 << BIO_UPTODATE;
1762                 bio->bi_comp_cpu = -1;
1763                 bio->bi_rw = READ;
1764                 bio->bi_vcnt = 0;
1765                 bio->bi_idx = 0;
1766                 bio->bi_phys_segments = 0;
1767                 bio->bi_size = 0;
1768                 bio->bi_end_io = NULL;
1769                 bio->bi_private = NULL;
1770
1771                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1772                 if (rdev == NULL ||
1773                            test_bit(Faulty, &rdev->flags)) {
1774                         still_degraded = 1;
1775                         continue;
1776                 } else if (!test_bit(In_sync, &rdev->flags)) {
1777                         bio->bi_rw = WRITE;
1778                         bio->bi_end_io = end_sync_write;
1779                         write_targets ++;
1780                 } else {
1781                         /* may need to read from here */
1782                         bio->bi_rw = READ;
1783                         bio->bi_end_io = end_sync_read;
1784                         if (test_bit(WriteMostly, &rdev->flags)) {
1785                                 if (wonly < 0)
1786                                         wonly = i;
1787                         } else {
1788                                 if (disk < 0)
1789                                         disk = i;
1790                         }
1791                         read_targets++;
1792                 }
1793                 atomic_inc(&rdev->nr_pending);
1794                 bio->bi_sector = sector_nr + rdev->data_offset;
1795                 bio->bi_bdev = rdev->bdev;
1796                 bio->bi_private = r1_bio;
1797         }
1798         rcu_read_unlock();
1799         if (disk < 0)
1800                 disk = wonly;
1801         r1_bio->read_disk = disk;
1802
1803         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1804                 /* extra read targets are also write targets */
1805                 write_targets += read_targets-1;
1806
1807         if (write_targets == 0 || read_targets == 0) {
1808                 /* There is nowhere to write, so all non-sync
1809                  * drives must be failed - so we are finished
1810                  */
1811                 sector_t rv = max_sector - sector_nr;
1812                 *skipped = 1;
1813                 put_buf(r1_bio);
1814                 return rv;
1815         }
1816
1817         if (max_sector > mddev->resync_max)
1818                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1819         nr_sectors = 0;
1820         sync_blocks = 0;
1821         do {
1822                 struct page *page;
1823                 int len = PAGE_SIZE;
1824                 if (sector_nr + (len>>9) > max_sector)
1825                         len = (max_sector - sector_nr) << 9;
1826                 if (len == 0)
1827                         break;
1828                 if (sync_blocks == 0) {
1829                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1830                                                &sync_blocks, still_degraded) &&
1831                             !conf->fullsync &&
1832                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1833                                 break;
1834                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1835                         if ((len >> 9) > sync_blocks)
1836                                 len = sync_blocks<<9;
1837                 }
1838
1839                 for (i=0 ; i < conf->raid_disks; i++) {
1840                         bio = r1_bio->bios[i];
1841                         if (bio->bi_end_io) {
1842                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1843                                 if (bio_add_page(bio, page, len, 0) == 0) {
1844                                         /* stop here */
1845                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1846                                         while (i > 0) {
1847                                                 i--;
1848                                                 bio = r1_bio->bios[i];
1849                                                 if (bio->bi_end_io==NULL)
1850                                                         continue;
1851                                                 /* remove last page from this bio */
1852                                                 bio->bi_vcnt--;
1853                                                 bio->bi_size -= len;
1854                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1855                                         }
1856                                         goto bio_full;
1857                                 }
1858                         }
1859                 }
1860                 nr_sectors += len>>9;
1861                 sector_nr += len>>9;
1862                 sync_blocks -= (len>>9);
1863         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1864  bio_full:
1865         r1_bio->sectors = nr_sectors;
1866
1867         /* For a user-requested sync, we read all readable devices and do a
1868          * compare
1869          */
1870         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1871                 atomic_set(&r1_bio->remaining, read_targets);
1872                 for (i=0; i<conf->raid_disks; i++) {
1873                         bio = r1_bio->bios[i];
1874                         if (bio->bi_end_io == end_sync_read) {
1875                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1876                                 generic_make_request(bio);
1877                         }
1878                 }
1879         } else {
1880                 atomic_set(&r1_bio->remaining, 1);
1881                 bio = r1_bio->bios[r1_bio->read_disk];
1882                 md_sync_acct(bio->bi_bdev, nr_sectors);
1883                 generic_make_request(bio);
1884
1885         }
1886         return nr_sectors;
1887 }
1888
1889 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1890 {
1891         if (sectors)
1892                 return sectors;
1893
1894         return mddev->dev_sectors;
1895 }
1896
1897 static conf_t *setup_conf(mddev_t *mddev)
1898 {
1899         conf_t *conf;
1900         int i;
1901         mirror_info_t *disk;
1902         mdk_rdev_t *rdev;
1903         int err = -ENOMEM;
1904
1905         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1906         if (!conf)
1907                 goto abort;
1908
1909         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1910                                  GFP_KERNEL);
1911         if (!conf->mirrors)
1912                 goto abort;
1913
1914         conf->tmppage = alloc_page(GFP_KERNEL);
1915         if (!conf->tmppage)
1916                 goto abort;
1917
1918         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1919         if (!conf->poolinfo)
1920                 goto abort;
1921         conf->poolinfo->raid_disks = mddev->raid_disks;
1922         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1923                                           r1bio_pool_free,
1924                                           conf->poolinfo);
1925         if (!conf->r1bio_pool)
1926                 goto abort;
1927
1928         conf->poolinfo->mddev = mddev;
1929
1930         spin_lock_init(&conf->device_lock);
1931         list_for_each_entry(rdev, &mddev->disks, same_set) {
1932                 int disk_idx = rdev->raid_disk;
1933                 if (disk_idx >= mddev->raid_disks
1934                     || disk_idx < 0)
1935                         continue;
1936                 disk = conf->mirrors + disk_idx;
1937
1938                 disk->rdev = rdev;
1939
1940                 disk->head_position = 0;
1941         }
1942         conf->raid_disks = mddev->raid_disks;
1943         conf->mddev = mddev;
1944         INIT_LIST_HEAD(&conf->retry_list);
1945
1946         spin_lock_init(&conf->resync_lock);
1947         init_waitqueue_head(&conf->wait_barrier);
1948
1949         bio_list_init(&conf->pending_bio_list);
1950
1951         conf->last_used = -1;
1952         for (i = 0; i < conf->raid_disks; i++) {
1953
1954                 disk = conf->mirrors + i;
1955
1956                 if (!disk->rdev ||
1957                     !test_bit(In_sync, &disk->rdev->flags)) {
1958                         disk->head_position = 0;
1959                         if (disk->rdev)
1960                                 conf->fullsync = 1;
1961                 } else if (conf->last_used < 0)
1962                         /*
1963                          * The first working device is used as a
1964                          * starting point to read balancing.
1965                          */
1966                         conf->last_used = i;
1967         }
1968
1969         err = -EIO;
1970         if (conf->last_used < 0) {
1971                 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
1972                        mdname(mddev));
1973                 goto abort;
1974         }
1975         err = -ENOMEM;
1976         conf->thread = md_register_thread(raid1d, mddev, NULL);
1977         if (!conf->thread) {
1978                 printk(KERN_ERR
1979                        "md/raid1:%s: couldn't allocate thread\n",
1980                        mdname(mddev));
1981                 goto abort;
1982         }
1983
1984         return conf;
1985
1986  abort:
1987         if (conf) {
1988                 if (conf->r1bio_pool)
1989                         mempool_destroy(conf->r1bio_pool);
1990                 kfree(conf->mirrors);
1991                 safe_put_page(conf->tmppage);
1992                 kfree(conf->poolinfo);
1993                 kfree(conf);
1994         }
1995         return ERR_PTR(err);
1996 }
1997
1998 static int run(mddev_t *mddev)
1999 {
2000         conf_t *conf;
2001         int i;
2002         mdk_rdev_t *rdev;
2003
2004         if (mddev->level != 1) {
2005                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2006                        mdname(mddev), mddev->level);
2007                 return -EIO;
2008         }
2009         if (mddev->reshape_position != MaxSector) {
2010                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2011                        mdname(mddev));
2012                 return -EIO;
2013         }
2014         /*
2015          * copy the already verified devices into our private RAID1
2016          * bookkeeping area. [whatever we allocate in run(),
2017          * should be freed in stop()]
2018          */
2019         if (mddev->private == NULL)
2020                 conf = setup_conf(mddev);
2021         else
2022                 conf = mddev->private;
2023
2024         if (IS_ERR(conf))
2025                 return PTR_ERR(conf);
2026
2027         list_for_each_entry(rdev, &mddev->disks, same_set) {
2028                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2029                                   rdev->data_offset << 9);
2030                 /* as we don't honour merge_bvec_fn, we must never risk
2031                  * violating it, so limit ->max_segments to 1 lying within
2032                  * a single page, as a one page request is never in violation.
2033                  */
2034                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2035                         blk_queue_max_segments(mddev->queue, 1);
2036                         blk_queue_segment_boundary(mddev->queue,
2037                                                    PAGE_CACHE_SIZE - 1);
2038                 }
2039         }
2040
2041         mddev->degraded = 0;
2042         for (i=0; i < conf->raid_disks; i++)
2043                 if (conf->mirrors[i].rdev == NULL ||
2044                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2045                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2046                         mddev->degraded++;
2047
2048         if (conf->raid_disks - mddev->degraded == 1)
2049                 mddev->recovery_cp = MaxSector;
2050
2051         if (mddev->recovery_cp != MaxSector)
2052                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2053                        " -- starting background reconstruction\n",
2054                        mdname(mddev));
2055         printk(KERN_INFO
2056                 "md/raid1:%s: active with %d out of %d mirrors\n",
2057                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2058                 mddev->raid_disks);
2059
2060         /*
2061          * Ok, everything is just fine now
2062          */
2063         mddev->thread = conf->thread;
2064         conf->thread = NULL;
2065         mddev->private = conf;
2066
2067         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2068
2069         mddev->queue->unplug_fn = raid1_unplug;
2070         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2071         mddev->queue->backing_dev_info.congested_data = mddev;
2072         md_integrity_register(mddev);
2073         return 0;
2074 }
2075
2076 static int stop(mddev_t *mddev)
2077 {
2078         conf_t *conf = mddev->private;
2079         struct bitmap *bitmap = mddev->bitmap;
2080
2081         /* wait for behind writes to complete */
2082         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2083                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2084                        mdname(mddev));
2085                 /* need to kick something here to make sure I/O goes? */
2086                 wait_event(bitmap->behind_wait,
2087                            atomic_read(&bitmap->behind_writes) == 0);
2088         }
2089
2090         raise_barrier(conf);
2091         lower_barrier(conf);
2092
2093         md_unregister_thread(mddev->thread);
2094         mddev->thread = NULL;
2095         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2096         if (conf->r1bio_pool)
2097                 mempool_destroy(conf->r1bio_pool);
2098         kfree(conf->mirrors);
2099         kfree(conf->poolinfo);
2100         kfree(conf);
2101         mddev->private = NULL;
2102         return 0;
2103 }
2104
2105 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2106 {
2107         /* no resync is happening, and there is enough space
2108          * on all devices, so we can resize.
2109          * We need to make sure resync covers any new space.
2110          * If the array is shrinking we should possibly wait until
2111          * any io in the removed space completes, but it hardly seems
2112          * worth it.
2113          */
2114         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2115         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2116                 return -EINVAL;
2117         set_capacity(mddev->gendisk, mddev->array_sectors);
2118         revalidate_disk(mddev->gendisk);
2119         if (sectors > mddev->dev_sectors &&
2120             mddev->recovery_cp == MaxSector) {
2121                 mddev->recovery_cp = mddev->dev_sectors;
2122                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2123         }
2124         mddev->dev_sectors = sectors;
2125         mddev->resync_max_sectors = sectors;
2126         return 0;
2127 }
2128
2129 static int raid1_reshape(mddev_t *mddev)
2130 {
2131         /* We need to:
2132          * 1/ resize the r1bio_pool
2133          * 2/ resize conf->mirrors
2134          *
2135          * We allocate a new r1bio_pool if we can.
2136          * Then raise a device barrier and wait until all IO stops.
2137          * Then resize conf->mirrors and swap in the new r1bio pool.
2138          *
2139          * At the same time, we "pack" the devices so that all the missing
2140          * devices have the higher raid_disk numbers.
2141          */
2142         mempool_t *newpool, *oldpool;
2143         struct pool_info *newpoolinfo;
2144         mirror_info_t *newmirrors;
2145         conf_t *conf = mddev->private;
2146         int cnt, raid_disks;
2147         unsigned long flags;
2148         int d, d2, err;
2149
2150         /* Cannot change chunk_size, layout, or level */
2151         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2152             mddev->layout != mddev->new_layout ||
2153             mddev->level != mddev->new_level) {
2154                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2155                 mddev->new_layout = mddev->layout;
2156                 mddev->new_level = mddev->level;
2157                 return -EINVAL;
2158         }
2159
2160         err = md_allow_write(mddev);
2161         if (err)
2162                 return err;
2163
2164         raid_disks = mddev->raid_disks + mddev->delta_disks;
2165
2166         if (raid_disks < conf->raid_disks) {
2167                 cnt=0;
2168                 for (d= 0; d < conf->raid_disks; d++)
2169                         if (conf->mirrors[d].rdev)
2170                                 cnt++;
2171                 if (cnt > raid_disks)
2172                         return -EBUSY;
2173         }
2174
2175         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2176         if (!newpoolinfo)
2177                 return -ENOMEM;
2178         newpoolinfo->mddev = mddev;
2179         newpoolinfo->raid_disks = raid_disks;
2180
2181         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2182                                  r1bio_pool_free, newpoolinfo);
2183         if (!newpool) {
2184                 kfree(newpoolinfo);
2185                 return -ENOMEM;
2186         }
2187         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2188         if (!newmirrors) {
2189                 kfree(newpoolinfo);
2190                 mempool_destroy(newpool);
2191                 return -ENOMEM;
2192         }
2193
2194         raise_barrier(conf);
2195
2196         /* ok, everything is stopped */
2197         oldpool = conf->r1bio_pool;
2198         conf->r1bio_pool = newpool;
2199
2200         for (d = d2 = 0; d < conf->raid_disks; d++) {
2201                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2202                 if (rdev && rdev->raid_disk != d2) {
2203                         char nm[20];
2204                         sprintf(nm, "rd%d", rdev->raid_disk);
2205                         sysfs_remove_link(&mddev->kobj, nm);
2206                         rdev->raid_disk = d2;
2207                         sprintf(nm, "rd%d", rdev->raid_disk);
2208                         sysfs_remove_link(&mddev->kobj, nm);
2209                         if (sysfs_create_link(&mddev->kobj,
2210                                               &rdev->kobj, nm))
2211                                 printk(KERN_WARNING
2212                                        "md/raid1:%s: cannot register "
2213                                        "%s\n",
2214                                        mdname(mddev), nm);
2215                 }
2216                 if (rdev)
2217                         newmirrors[d2++].rdev = rdev;
2218         }
2219         kfree(conf->mirrors);
2220         conf->mirrors = newmirrors;
2221         kfree(conf->poolinfo);
2222         conf->poolinfo = newpoolinfo;
2223
2224         spin_lock_irqsave(&conf->device_lock, flags);
2225         mddev->degraded += (raid_disks - conf->raid_disks);
2226         spin_unlock_irqrestore(&conf->device_lock, flags);
2227         conf->raid_disks = mddev->raid_disks = raid_disks;
2228         mddev->delta_disks = 0;
2229
2230         conf->last_used = 0; /* just make sure it is in-range */
2231         lower_barrier(conf);
2232
2233         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2234         md_wakeup_thread(mddev->thread);
2235
2236         mempool_destroy(oldpool);
2237         return 0;
2238 }
2239
2240 static void raid1_quiesce(mddev_t *mddev, int state)
2241 {
2242         conf_t *conf = mddev->private;
2243
2244         switch(state) {
2245         case 2: /* wake for suspend */
2246                 wake_up(&conf->wait_barrier);
2247                 break;
2248         case 1:
2249                 raise_barrier(conf);
2250                 break;
2251         case 0:
2252                 lower_barrier(conf);
2253                 break;
2254         }
2255 }
2256
2257 static void *raid1_takeover(mddev_t *mddev)
2258 {
2259         /* raid1 can take over:
2260          *  raid5 with 2 devices, any layout or chunk size
2261          */
2262         if (mddev->level == 5 && mddev->raid_disks == 2) {
2263                 conf_t *conf;
2264                 mddev->new_level = 1;
2265                 mddev->new_layout = 0;
2266                 mddev->new_chunk_sectors = 0;
2267                 conf = setup_conf(mddev);
2268                 if (!IS_ERR(conf))
2269                         conf->barrier = 1;
2270                 return conf;
2271         }
2272         return ERR_PTR(-EINVAL);
2273 }
2274
2275 static struct mdk_personality raid1_personality =
2276 {
2277         .name           = "raid1",
2278         .level          = 1,
2279         .owner          = THIS_MODULE,
2280         .make_request   = make_request,
2281         .run            = run,
2282         .stop           = stop,
2283         .status         = status,
2284         .error_handler  = error,
2285         .hot_add_disk   = raid1_add_disk,
2286         .hot_remove_disk= raid1_remove_disk,
2287         .spare_active   = raid1_spare_active,
2288         .sync_request   = sync_request,
2289         .resize         = raid1_resize,
2290         .size           = raid1_size,
2291         .check_reshape  = raid1_reshape,
2292         .quiesce        = raid1_quiesce,
2293         .takeover       = raid1_takeover,
2294 };
2295
2296 static int __init raid_init(void)
2297 {
2298         return register_md_personality(&raid1_personality);
2299 }
2300
2301 static void raid_exit(void)
2302 {
2303         unregister_md_personality(&raid1_personality);
2304 }
2305
2306 module_init(raid_init);
2307 module_exit(raid_exit);
2308 MODULE_LICENSE("GPL");
2309 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2310 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2311 MODULE_ALIAS("md-raid1");
2312 MODULE_ALIAS("md-level-1");