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_alloc(gfp_flags, RESYNC_PAGES);
 104                 if (!bio)
 105                         goto out_free_bio;
 106                 r1_bio->bios[j] = bio;
 107         }
 108         /*
 109          * Allocate RESYNC_PAGES data pages and attach them to
 110          * the first bio.
 111          * If this is a user-requested check/repair, allocate
 112          * RESYNC_PAGES for each bio.
 113          */
 114         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 115                 j = pi->raid_disks;
 116         else
 117                 j = 1;
 118         while(j--) {
 119                 bio = r1_bio->bios[j];
 120                 for (i = 0; i < RESYNC_PAGES; i++) {
 121                         page = alloc_page(gfp_flags);
 122                         if (unlikely(!page))
 123                                 goto out_free_pages;
 124
 125                         bio->bi_io_vec[i].bv_page = page;
 126                         bio->bi_vcnt = i+1;
 127                 }
 128         }
 129         /* If not user-requests, copy the page pointers to all bios */
 130         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 131                 for (i=0; i<RESYNC_PAGES ; i++)
 132                         for (j=1; j<pi->raid_disks; j++)
 133                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 134                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 135         }
 136
 137         r1_bio->master_bio = NULL;
 138
 139         return r1_bio;
 140
 141 out_free_pages:
 142         for (j=0 ; j < pi->raid_disks; j++)
 143                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 144                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 145         j = -1;
 146 out_free_bio:
 147         while ( ++j < pi->raid_disks )
 148                 bio_put(r1_bio->bios[j]);
 149         r1bio_pool_free(r1_bio, data);
 150         return NULL;
 151 }
 152
 153 static void r1buf_pool_free(void *__r1_bio, void *data)
 154 {
 155         struct pool_info *pi = data;
 156         int i,j;
 157         r1bio_t *r1bio = __r1_bio;
 158
 159         for (i = 0; i < RESYNC_PAGES; i++)
 160                 for (j = pi->raid_disks; j-- ;) {
 161                         if (j == 0 ||
 162                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 163                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 164                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 165                 }
 166         for (i=0 ; i < pi->raid_disks; i++)
 167                 bio_put(r1bio->bios[i]);
 168
 169         r1bio_pool_free(r1bio, data);
 170 }
 171
 172 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 173 {
 174         int i;
 175
 176         for (i = 0; i < conf->raid_disks; i++) {
 177                 struct bio **bio = r1_bio->bios + i;
 178                 if (*bio && *bio != IO_BLOCKED)
 179                         bio_put(*bio);
 180                 *bio = NULL;
 181         }
 182 }
 183
 184 static void free_r1bio(r1bio_t *r1_bio)
 185 {
 186         conf_t *conf = 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 = (r1bio_t *)(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 "raid1: %s: rescheduling sector %llu\n",
 301                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 302                 reschedule_retry(r1_bio);
 303         }
 304
 305         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 306 }
 307
 308 static void raid1_end_write_request(struct bio *bio, int error)
 309 {
 310         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 311         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 312         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 313         conf_t *conf = r1_bio->mddev->private;
 314         struct bio *to_put = NULL;
 315
 316
 317         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 318                 if (r1_bio->bios[mirror] == bio)
 319                         break;
 320
 321         if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
 322                 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
 323                 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
 324                 r1_bio->mddev->barriers_work = 0;
 325                 /* Don't rdev_dec_pending in this branch - keep it for the retry */
 326         } else {
 327                 /*
 328                  * this branch is our 'one mirror IO has finished' event handler:
 329                  */
 330                 r1_bio->bios[mirror] = NULL;
 331                 to_put = bio;
 332                 if (!uptodate) {
 333                         md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 334                         /* an I/O failed, we can't clear the bitmap */
 335                         set_bit(R1BIO_Degraded, &r1_bio->state);
 336                 } else
 337                         /*
 338                          * Set R1BIO_Uptodate in our master bio, so that
 339                          * we will return a good error code for to the higher
 340                          * levels even if IO on some other mirrored buffer fails.
 341                          *
 342                          * The 'master' represents the composite IO operation to
 343                          * user-side. So if something waits for IO, then it will
 344                          * wait for the 'master' bio.
 345                          */
 346                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 347
 348                 update_head_pos(mirror, r1_bio);
 349
 350                 if (behind) {
 351                         if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 352                                 atomic_dec(&r1_bio->behind_remaining);
 353
 354                         /* In behind mode, we ACK the master bio once the I/O has safely
 355                          * reached all non-writemostly disks. Setting the Returned bit
 356                          * ensures that this gets done only once -- we don't ever want to
 357                          * return -EIO here, instead we'll wait */
 358
 359                         if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 360                             test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 361                                 /* Maybe we can return now */
 362                                 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 363                                         struct bio *mbio = r1_bio->master_bio;
 364                                         PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 365                                                (unsigned long long) mbio->bi_sector,
 366                                                (unsigned long long) mbio->bi_sector +
 367                                                (mbio->bi_size >> 9) - 1);
 368                                         bio_endio(mbio, 0);
 369                                 }
 370                         }
 371                 }
 372                 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 373         }
 374         /*
 375          *
 376          * Let's see if all mirrored write operations have finished
 377          * already.
 378          */
 379         if (atomic_dec_and_test(&r1_bio->remaining)) {
 380                 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
 381                         reschedule_retry(r1_bio);
 382                 else {
 383                         /* it really is the end of this request */
 384                         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 385                                 /* free extra copy of the data pages */
 386                                 int i = bio->bi_vcnt;
 387                                 while (i--)
 388                                         safe_put_page(bio->bi_io_vec[i].bv_page);
 389                         }
 390                         /* clear the bitmap if all writes complete successfully */
 391                         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 392                                         r1_bio->sectors,
 393                                         !test_bit(R1BIO_Degraded, &r1_bio->state),
 394                                         behind);
 395                         md_write_end(r1_bio->mddev);
 396                         raid_end_bio_io(r1_bio);
 397                 }
 398         }
 399
 400         if (to_put)
 401                 bio_put(to_put);
 402 }
 403
 404
 405 /*
 406  * This routine returns the disk from which the requested read should
 407  * be done. There is a per-array 'next expected sequential IO' sector
 408  * number - if this matches on the next IO then we use the last disk.
 409  * There is also a per-disk 'last know head position' sector that is
 410  * maintained from IRQ contexts, both the normal and the resync IO
 411  * completion handlers update this position correctly. If there is no
 412  * perfect sequential match then we pick the disk whose head is closest.
 413  *
 414  * If there are 2 mirrors in the same 2 devices, performance degrades
 415  * because position is mirror, not device based.
 416  *
 417  * The rdev for the device selected will have nr_pending incremented.
 418  */
 419 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 420 {
 421         const unsigned long this_sector = r1_bio->sector;
 422         int new_disk = conf->last_used, disk = new_disk;
 423         int wonly_disk = -1;
 424         const int sectors = r1_bio->sectors;
 425         sector_t new_distance, current_distance;
 426         mdk_rdev_t *rdev;
 427
 428         rcu_read_lock();
 429         /*
 430          * Check if we can balance. We can balance on the whole
 431          * device if no resync is going on, or below the resync window.
 432          * We take the first readable disk when above the resync window.
 433          */
 434  retry:
 435         if (conf->mddev->recovery_cp < MaxSector &&
 436             (this_sector + sectors >= conf->next_resync)) {
 437                 /* Choose the first operation device, for consistancy */
 438                 new_disk = 0;
 439
 440                 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 441                      r1_bio->bios[new_disk] == IO_BLOCKED ||
 442                      !rdev || !test_bit(In_sync, &rdev->flags)
 443                              || test_bit(WriteMostly, &rdev->flags);
 444                      rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
 445
 446                         if (rdev && test_bit(In_sync, &rdev->flags) &&
 447                                 r1_bio->bios[new_disk] != IO_BLOCKED)
 448                                 wonly_disk = new_disk;
 449
 450                         if (new_disk == conf->raid_disks - 1) {
 451                                 new_disk = wonly_disk;
 452                                 break;
 453                         }
 454                 }
 455                 goto rb_out;
 456         }
 457
 458
 459         /* make sure the disk is operational */
 460         for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 461              r1_bio->bios[new_disk] == IO_BLOCKED ||
 462              !rdev || !test_bit(In_sync, &rdev->flags) ||
 463                      test_bit(WriteMostly, &rdev->flags);
 464              rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
 465
 466                 if (rdev && test_bit(In_sync, &rdev->flags) &&
 467                     r1_bio->bios[new_disk] != IO_BLOCKED)
 468                         wonly_disk = new_disk;
 469
 470                 if (new_disk <= 0)
 471                         new_disk = conf->raid_disks;
 472                 new_disk--;
 473                 if (new_disk == disk) {
 474                         new_disk = wonly_disk;
 475                         break;
 476                 }
 477         }
 478
 479         if (new_disk < 0)
 480                 goto rb_out;
 481
 482         disk = new_disk;
 483         /* now disk == new_disk == starting point for search */
 484
 485         /*
 486          * Don't change to another disk for sequential reads:
 487          */
 488         if (conf->next_seq_sect == this_sector)
 489                 goto rb_out;
 490         if (this_sector == conf->mirrors[new_disk].head_position)
 491                 goto rb_out;
 492
 493         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 494
 495         /* Find the disk whose head is closest */
 496
 497         do {
 498                 if (disk <= 0)
 499                         disk = conf->raid_disks;
 500                 disk--;
 501
 502                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 503
 504                 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
 505                     !test_bit(In_sync, &rdev->flags) ||
 506                     test_bit(WriteMostly, &rdev->flags))
 507                         continue;
 508
 509                 if (!atomic_read(&rdev->nr_pending)) {
 510                         new_disk = disk;
 511                         break;
 512                 }
 513                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 514                 if (new_distance < current_distance) {
 515                         current_distance = new_distance;
 516                         new_disk = disk;
 517                 }
 518         } while (disk != conf->last_used);
 519
 520  rb_out:
 521
 522
 523         if (new_disk >= 0) {
 524                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 525                 if (!rdev)
 526                         goto retry;
 527                 atomic_inc(&rdev->nr_pending);
 528                 if (!test_bit(In_sync, &rdev->flags)) {
 529                         /* cannot risk returning a device that failed
 530                          * before we inc'ed nr_pending
 531                          */
 532                         rdev_dec_pending(rdev, conf->mddev);
 533                         goto retry;
 534                 }
 535                 conf->next_seq_sect = this_sector + sectors;
 536                 conf->last_used = new_disk;
 537         }
 538         rcu_read_unlock();
 539
 540         return new_disk;
 541 }
 542
 543 static void unplug_slaves(mddev_t *mddev)
 544 {
 545         conf_t *conf = mddev->private;
 546         int i;
 547
 548         rcu_read_lock();
 549         for (i=0; i<mddev->raid_disks; i++) {
 550                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 551                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 552                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 553
 554                         atomic_inc(&rdev->nr_pending);
 555                         rcu_read_unlock();
 556
 557                         blk_unplug(r_queue);
 558
 559                         rdev_dec_pending(rdev, mddev);
 560                         rcu_read_lock();
 561                 }
 562         }
 563         rcu_read_unlock();
 564 }
 565
 566 static void raid1_unplug(struct request_queue *q)
 567 {
 568         mddev_t *mddev = q->queuedata;
 569
 570         unplug_slaves(mddev);
 571         md_wakeup_thread(mddev->thread);
 572 }
 573
 574 static int raid1_congested(void *data, int bits)
 575 {
 576         mddev_t *mddev = data;
 577         conf_t *conf = mddev->private;
 578         int i, ret = 0;
 579
 580         if (mddev_congested(mddev, bits))
 581                 return 1;
 582
 583         rcu_read_lock();
 584         for (i = 0; i < mddev->raid_disks; i++) {
 585                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 586                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 587                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 588
 589                         /* Note the '|| 1' - when read_balance prefers
 590                          * non-congested targets, it can be removed
 591                          */
 592                         if ((bits & (1<<BDI_async_congested)) || 1)
 593                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 594                         else
 595                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 596                 }
 597         }
 598         rcu_read_unlock();
 599         return ret;
 600 }
 601
 602
 603 static int flush_pending_writes(conf_t *conf)
 604 {
 605         /* Any writes that have been queued but are awaiting
 606          * bitmap updates get flushed here.
 607          * We return 1 if any requests were actually submitted.
 608          */
 609         int rv = 0;
 610
 611         spin_lock_irq(&conf->device_lock);
 612
 613         if (conf->pending_bio_list.head) {
 614                 struct bio *bio;
 615                 bio = bio_list_get(&conf->pending_bio_list);
 616                 blk_remove_plug(conf->mddev->queue);
 617                 spin_unlock_irq(&conf->device_lock);
 618                 /* flush any pending bitmap writes to
 619                  * disk before proceeding w/ I/O */
 620                 bitmap_unplug(conf->mddev->bitmap);
 621
 622                 while (bio) { /* submit pending writes */
 623                         struct bio *next = bio->bi_next;
 624                         bio->bi_next = NULL;
 625                         generic_make_request(bio);
 626                         bio = next;
 627                 }
 628                 rv = 1;
 629         } else
 630                 spin_unlock_irq(&conf->device_lock);
 631         return rv;
 632 }
 633
 634 /* Barriers....
 635  * Sometimes we need to suspend IO while we do something else,
 636  * either some resync/recovery, or reconfigure the array.
 637  * To do this we raise a 'barrier'.
 638  * The 'barrier' is a counter that can be raised multiple times
 639  * to count how many activities are happening which preclude
 640  * normal IO.
 641  * We can only raise the barrier if there is no pending IO.
 642  * i.e. if nr_pending == 0.
 643  * We choose only to raise the barrier if no-one is waiting for the
 644  * barrier to go down.  This means that as soon as an IO request
 645  * is ready, no other operations which require a barrier will start
 646  * until the IO request has had a chance.
 647  *
 648  * So: regular IO calls 'wait_barrier'.  When that returns there
 649  *    is no backgroup IO happening,  It must arrange to call
 650  *    allow_barrier when it has finished its IO.
 651  * backgroup IO calls must call raise_barrier.  Once that returns
 652  *    there is no normal IO happeing.  It must arrange to call
 653  *    lower_barrier when the particular background IO completes.
 654  */
 655 #define RESYNC_DEPTH 32
 656
 657 static void raise_barrier(conf_t *conf)
 658 {
 659         spin_lock_irq(&conf->resync_lock);
 660
 661         /* Wait until no block IO is waiting */
 662         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 663                             conf->resync_lock,
 664                             raid1_unplug(conf->mddev->queue));
 665
 666         /* block any new IO from starting */
 667         conf->barrier++;
 668
 669         /* No wait for all pending IO to complete */
 670         wait_event_lock_irq(conf->wait_barrier,
 671                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 672                             conf->resync_lock,
 673                             raid1_unplug(conf->mddev->queue));
 674
 675         spin_unlock_irq(&conf->resync_lock);
 676 }
 677
 678 static void lower_barrier(conf_t *conf)
 679 {
 680         unsigned long flags;
 681         BUG_ON(conf->barrier <= 0);
 682         spin_lock_irqsave(&conf->resync_lock, flags);
 683         conf->barrier--;
 684         spin_unlock_irqrestore(&conf->resync_lock, flags);
 685         wake_up(&conf->wait_barrier);
 686 }
 687
 688 static void wait_barrier(conf_t *conf)
 689 {
 690         spin_lock_irq(&conf->resync_lock);
 691         if (conf->barrier) {
 692                 conf->nr_waiting++;
 693                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 694                                     conf->resync_lock,
 695                                     raid1_unplug(conf->mddev->queue));
 696                 conf->nr_waiting--;
 697         }
 698         conf->nr_pending++;
 699         spin_unlock_irq(&conf->resync_lock);
 700 }
 701
 702 static void allow_barrier(conf_t *conf)
 703 {
 704         unsigned long flags;
 705         spin_lock_irqsave(&conf->resync_lock, flags);
 706         conf->nr_pending--;
 707         spin_unlock_irqrestore(&conf->resync_lock, flags);
 708         wake_up(&conf->wait_barrier);
 709 }
 710
 711 static void freeze_array(conf_t *conf)
 712 {
 713         /* stop syncio and normal IO and wait for everything to
 714          * go quite.
 715          * We increment barrier and nr_waiting, and then
 716          * wait until nr_pending match nr_queued+1
 717          * This is called in the context of one normal IO request
 718          * that has failed. Thus any sync request that might be pending
 719          * will be blocked by nr_pending, and we need to wait for
 720          * pending IO requests to complete or be queued for re-try.
 721          * Thus the number queued (nr_queued) plus this request (1)
 722          * must match the number of pending IOs (nr_pending) before
 723          * we continue.
 724          */
 725         spin_lock_irq(&conf->resync_lock);
 726         conf->barrier++;
 727         conf->nr_waiting++;
 728         wait_event_lock_irq(conf->wait_barrier,
 729                             conf->nr_pending == conf->nr_queued+1,
 730                             conf->resync_lock,
 731                             ({ flush_pending_writes(conf);
 732                                raid1_unplug(conf->mddev->queue); }));
 733         spin_unlock_irq(&conf->resync_lock);
 734 }
 735 static void unfreeze_array(conf_t *conf)
 736 {
 737         /* reverse the effect of the freeze */
 738         spin_lock_irq(&conf->resync_lock);
 739         conf->barrier--;
 740         conf->nr_waiting--;
 741         wake_up(&conf->wait_barrier);
 742         spin_unlock_irq(&conf->resync_lock);
 743 }
 744
 745
 746 /* duplicate the data pages for behind I/O */
 747 static struct page **alloc_behind_pages(struct bio *bio)
 748 {
 749         int i;
 750         struct bio_vec *bvec;
 751         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
 752                                         GFP_NOIO);
 753         if (unlikely(!pages))
 754                 goto do_sync_io;
 755
 756         bio_for_each_segment(bvec, bio, i) {
 757                 pages[i] = alloc_page(GFP_NOIO);
 758                 if (unlikely(!pages[i]))
 759                         goto do_sync_io;
 760                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 761                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 762                 kunmap(pages[i]);
 763                 kunmap(bvec->bv_page);
 764         }
 765
 766         return pages;
 767
 768 do_sync_io:
 769         if (pages)
 770                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
 771                         put_page(pages[i]);
 772         kfree(pages);
 773         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 774         return NULL;
 775 }
 776
 777 static int make_request(struct request_queue *q, struct bio * bio)
 778 {
 779         mddev_t *mddev = q->queuedata;
 780         conf_t *conf = mddev->private;
 781         mirror_info_t *mirror;
 782         r1bio_t *r1_bio;
 783         struct bio *read_bio;
 784         int i, targets = 0, disks;
 785         struct bitmap *bitmap;
 786         unsigned long flags;
 787         struct bio_list bl;
 788         struct page **behind_pages = NULL;
 789         const int rw = bio_data_dir(bio);
 790         const bool do_sync = bio_rw_flagged(bio, BIO_RW_SYNCIO);
 791         int cpu;
 792         bool do_barriers;
 793         mdk_rdev_t *blocked_rdev;
 794
 795         /*
 796          * Register the new request and wait if the reconstruction
 797          * thread has put up a bar for new requests.
 798          * Continue immediately if no resync is active currently.
 799          * We test barriers_work *after* md_write_start as md_write_start
 800          * may cause the first superblock write, and that will check out
 801          * if barriers work.
 802          */
 803
 804         md_write_start(mddev, bio); /* wait on superblock update early */
 805
 806         if (bio_data_dir(bio) == WRITE &&
 807             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
 808             bio->bi_sector < mddev->suspend_hi) {
 809                 /* As the suspend_* range is controlled by
 810                  * userspace, we want an interruptible
 811                  * wait.
 812                  */
 813                 DEFINE_WAIT(w);
 814                 for (;;) {
 815                         flush_signals(current);
 816                         prepare_to_wait(&conf->wait_barrier,
 817                                         &w, TASK_INTERRUPTIBLE);
 818                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
 819                             bio->bi_sector >= mddev->suspend_hi)
 820                                 break;
 821                         schedule();
 822                 }
 823                 finish_wait(&conf->wait_barrier, &w);
 824         }
 825         if (unlikely(!mddev->barriers_work &&
 826                      bio_rw_flagged(bio, BIO_RW_BARRIER))) {
 827                 if (rw == WRITE)
 828                         md_write_end(mddev);
 829                 bio_endio(bio, -EOPNOTSUPP);
 830                 return 0;
 831         }
 832
 833         wait_barrier(conf);
 834
 835         bitmap = mddev->bitmap;
 836
 837         cpu = part_stat_lock();
 838         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
 839         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
 840                       bio_sectors(bio));
 841         part_stat_unlock();
 842
 843         /*
 844          * make_request() can abort the operation when READA is being
 845          * used and no empty request is available.
 846          *
 847          */
 848         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 849
 850         r1_bio->master_bio = bio;
 851         r1_bio->sectors = bio->bi_size >> 9;
 852         r1_bio->state = 0;
 853         r1_bio->mddev = mddev;
 854         r1_bio->sector = bio->bi_sector;
 855
 856         if (rw == READ) {
 857                 /*
 858                  * read balancing logic:
 859                  */
 860                 int rdisk = read_balance(conf, r1_bio);
 861
 862                 if (rdisk < 0) {
 863                         /* couldn't find anywhere to read from */
 864                         raid_end_bio_io(r1_bio);
 865                         return 0;
 866                 }
 867                 mirror = conf->mirrors + rdisk;
 868
 869                 r1_bio->read_disk = rdisk;
 870
 871                 read_bio = bio_clone(bio, GFP_NOIO);
 872
 873                 r1_bio->bios[rdisk] = read_bio;
 874
 875                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 876                 read_bio->bi_bdev = mirror->rdev->bdev;
 877                 read_bio->bi_end_io = raid1_end_read_request;
 878                 read_bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
 879                 read_bio->bi_private = r1_bio;
 880
 881                 generic_make_request(read_bio);
 882                 return 0;
 883         }
 884
 885         /*
 886          * WRITE:
 887          */
 888         /* first select target devices under spinlock and
 889          * inc refcount on their rdev.  Record them by setting
 890          * bios[x] to bio
 891          */
 892         disks = conf->raid_disks;
 893 #if 0
 894         { static int first=1;
 895         if (first) printk("First Write sector %llu disks %d\n",
 896                           (unsigned long long)r1_bio->sector, disks);
 897         first = 0;
 898         }
 899 #endif
 900  retry_write:
 901         blocked_rdev = NULL;
 902         rcu_read_lock();
 903         for (i = 0;  i < disks; i++) {
 904                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 905                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 906                         atomic_inc(&rdev->nr_pending);
 907                         blocked_rdev = rdev;
 908                         break;
 909                 }
 910                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 911                         atomic_inc(&rdev->nr_pending);
 912                         if (test_bit(Faulty, &rdev->flags)) {
 913                                 rdev_dec_pending(rdev, mddev);
 914                                 r1_bio->bios[i] = NULL;
 915                         } else
 916                                 r1_bio->bios[i] = bio;
 917                         targets++;
 918                 } else
 919                         r1_bio->bios[i] = NULL;
 920         }
 921         rcu_read_unlock();
 922
 923         if (unlikely(blocked_rdev)) {
 924                 /* Wait for this device to become unblocked */
 925                 int j;
 926
 927                 for (j = 0; j < i; j++)
 928                         if (r1_bio->bios[j])
 929                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
 930
 931                 allow_barrier(conf);
 932                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 933                 wait_barrier(conf);
 934                 goto retry_write;
 935         }
 936
 937         BUG_ON(targets == 0); /* we never fail the last device */
 938
 939         if (targets < conf->raid_disks) {
 940                 /* array is degraded, we will not clear the bitmap
 941                  * on I/O completion (see raid1_end_write_request) */
 942                 set_bit(R1BIO_Degraded, &r1_bio->state);
 943         }
 944
 945         /* do behind I/O ? */
 946         if (bitmap &&
 947             (atomic_read(&bitmap->behind_writes)
 948              < mddev->bitmap_info.max_write_behind) &&
 949             (behind_pages = alloc_behind_pages(bio)) != NULL)
 950                 set_bit(R1BIO_BehindIO, &r1_bio->state);
 951
 952         atomic_set(&r1_bio->remaining, 0);
 953         atomic_set(&r1_bio->behind_remaining, 0);
 954
 955         do_barriers = bio_rw_flagged(bio, BIO_RW_BARRIER);
 956         if (do_barriers)
 957                 set_bit(R1BIO_Barrier, &r1_bio->state);
 958
 959         bio_list_init(&bl);
 960         for (i = 0; i < disks; i++) {
 961                 struct bio *mbio;
 962                 if (!r1_bio->bios[i])
 963                         continue;
 964
 965                 mbio = bio_clone(bio, GFP_NOIO);
 966                 r1_bio->bios[i] = mbio;
 967
 968                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 969                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 970                 mbio->bi_end_io = raid1_end_write_request;
 971                 mbio->bi_rw = WRITE | (do_barriers << BIO_RW_BARRIER) |
 972                         (do_sync << BIO_RW_SYNCIO);
 973                 mbio->bi_private = r1_bio;
 974
 975                 if (behind_pages) {
 976                         struct bio_vec *bvec;
 977                         int j;
 978
 979                         /* Yes, I really want the '__' version so that
 980                          * we clear any unused pointer in the io_vec, rather
 981                          * than leave them unchanged.  This is important
 982                          * because when we come to free the pages, we won't
 983                          * know the originial bi_idx, so we just free
 984                          * them all
 985                          */
 986                         __bio_for_each_segment(bvec, mbio, j, 0)
 987                                 bvec->bv_page = behind_pages[j];
 988                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 989                                 atomic_inc(&r1_bio->behind_remaining);
 990                 }
 991
 992                 atomic_inc(&r1_bio->remaining);
 993
 994                 bio_list_add(&bl, mbio);
 995         }
 996         kfree(behind_pages); /* the behind pages are attached to the bios now */
 997
 998         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
 999                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
1000         spin_lock_irqsave(&conf->device_lock, flags);
1001         bio_list_merge(&conf->pending_bio_list, &bl);
1002         bio_list_init(&bl);
1003
1004         blk_plug_device(mddev->queue);
1005         spin_unlock_irqrestore(&conf->device_lock, flags);
1006
1007         /* In case raid1d snuck into freeze_array */
1008         wake_up(&conf->wait_barrier);
1009
1010         if (do_sync)
1011                 md_wakeup_thread(mddev->thread);
1012 #if 0
1013         while ((bio = bio_list_pop(&bl)) != NULL)
1014                 generic_make_request(bio);
1015 #endif
1016
1017         return 0;
1018 }
1019
1020 static void status(struct seq_file *seq, mddev_t *mddev)
1021 {
1022         conf_t *conf = mddev->private;
1023         int i;
1024
1025         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1026                    conf->raid_disks - mddev->degraded);
1027         rcu_read_lock();
1028         for (i = 0; i < conf->raid_disks; i++) {
1029                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1030                 seq_printf(seq, "%s",
1031                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1032         }
1033         rcu_read_unlock();
1034         seq_printf(seq, "]");
1035 }
1036
1037
1038 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1039 {
1040         char b[BDEVNAME_SIZE];
1041         conf_t *conf = mddev->private;
1042
1043         /*
1044          * If it is not operational, then we have already marked it as dead
1045          * else if it is the last working disks, ignore the error, let the
1046          * next level up know.
1047          * else mark the drive as failed
1048          */
1049         if (test_bit(In_sync, &rdev->flags)
1050             && (conf->raid_disks - mddev->degraded) == 1) {
1051                 /*
1052                  * Don't fail the drive, act as though we were just a
1053                  * normal single drive.
1054                  * However don't try a recovery from this drive as
1055                  * it is very likely to fail.
1056                  */
1057                 mddev->recovery_disabled = 1;
1058                 return;
1059         }
1060         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1061                 unsigned long flags;
1062                 spin_lock_irqsave(&conf->device_lock, flags);
1063                 mddev->degraded++;
1064                 set_bit(Faulty, &rdev->flags);
1065                 spin_unlock_irqrestore(&conf->device_lock, flags);
1066                 /*
1067                  * if recovery is running, make sure it aborts.
1068                  */
1069                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1070         } else
1071                 set_bit(Faulty, &rdev->flags);
1072         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1073         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device.\n"
1074                 "raid1: Operation continuing on %d devices.\n",
1075                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1076 }
1077
1078 static void print_conf(conf_t *conf)
1079 {
1080         int i;
1081
1082         printk("RAID1 conf printout:\n");
1083         if (!conf) {
1084                 printk("(!conf)\n");
1085                 return;
1086         }
1087         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1088                 conf->raid_disks);
1089
1090         rcu_read_lock();
1091         for (i = 0; i < conf->raid_disks; i++) {
1092                 char b[BDEVNAME_SIZE];
1093                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1094                 if (rdev)
1095                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1096                                i, !test_bit(In_sync, &rdev->flags),
1097                                !test_bit(Faulty, &rdev->flags),
1098                                bdevname(rdev->bdev,b));
1099         }
1100         rcu_read_unlock();
1101 }
1102
1103 static void close_sync(conf_t *conf)
1104 {
1105         wait_barrier(conf);
1106         allow_barrier(conf);
1107
1108         mempool_destroy(conf->r1buf_pool);
1109         conf->r1buf_pool = NULL;
1110 }
1111
1112 static int raid1_spare_active(mddev_t *mddev)
1113 {
1114         int i;
1115         conf_t *conf = mddev->private;
1116
1117         /*
1118          * Find all failed disks within the RAID1 configuration
1119          * and mark them readable.
1120          * Called under mddev lock, so rcu protection not needed.
1121          */
1122         for (i = 0; i < conf->raid_disks; i++) {
1123                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1124                 if (rdev
1125                     && !test_bit(Faulty, &rdev->flags)
1126                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1127                         unsigned long flags;
1128                         spin_lock_irqsave(&conf->device_lock, flags);
1129                         mddev->degraded--;
1130                         spin_unlock_irqrestore(&conf->device_lock, flags);
1131                 }
1132         }
1133
1134         print_conf(conf);
1135         return 0;
1136 }
1137
1138
1139 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1140 {
1141         conf_t *conf = mddev->private;
1142         int err = -EEXIST;
1143         int mirror = 0;
1144         mirror_info_t *p;
1145         int first = 0;
1146         int last = mddev->raid_disks - 1;
1147
1148         if (rdev->raid_disk >= 0)
1149                 first = last = rdev->raid_disk;
1150
1151         for (mirror = first; mirror <= last; mirror++)
1152                 if ( !(p=conf->mirrors+mirror)->rdev) {
1153
1154                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1155                                           rdev->data_offset << 9);
1156                         /* as we don't honour merge_bvec_fn, we must
1157                          * never risk violating it, so limit
1158                          * ->max_segments to one lying with a single
1159                          * page, as a one page request is never in
1160                          * violation.
1161                          */
1162                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1163                                 blk_queue_max_segments(mddev->queue, 1);
1164                                 blk_queue_segment_boundary(mddev->queue,
1165                                                            PAGE_CACHE_SIZE - 1);
1166                         }
1167
1168                         p->head_position = 0;
1169                         rdev->raid_disk = mirror;
1170                         err = 0;
1171                         /* As all devices are equivalent, we don't need a full recovery
1172                          * if this was recently any drive of the array
1173                          */
1174                         if (rdev->saved_raid_disk < 0)
1175                                 conf->fullsync = 1;
1176                         rcu_assign_pointer(p->rdev, rdev);
1177                         break;
1178                 }
1179         md_integrity_add_rdev(rdev, mddev);
1180         print_conf(conf);
1181         return err;
1182 }
1183
1184 static int raid1_remove_disk(mddev_t *mddev, int number)
1185 {
1186         conf_t *conf = mddev->private;
1187         int err = 0;
1188         mdk_rdev_t *rdev;
1189         mirror_info_t *p = conf->mirrors+ number;
1190
1191         print_conf(conf);
1192         rdev = p->rdev;
1193         if (rdev) {
1194                 if (test_bit(In_sync, &rdev->flags) ||
1195                     atomic_read(&rdev->nr_pending)) {
1196                         err = -EBUSY;
1197                         goto abort;
1198                 }
1199                 /* Only remove non-faulty devices is recovery
1200                  * is not possible.
1201                  */
1202                 if (!test_bit(Faulty, &rdev->flags) &&
1203                     mddev->degraded < conf->raid_disks) {
1204                         err = -EBUSY;
1205                         goto abort;
1206                 }
1207                 p->rdev = NULL;
1208                 synchronize_rcu();
1209                 if (atomic_read(&rdev->nr_pending)) {
1210                         /* lost the race, try later */
1211                         err = -EBUSY;
1212                         p->rdev = rdev;
1213                         goto abort;
1214                 }
1215                 md_integrity_register(mddev);
1216         }
1217 abort:
1218
1219         print_conf(conf);
1220         return err;
1221 }
1222
1223
1224 static void end_sync_read(struct bio *bio, int error)
1225 {
1226         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1227         int i;
1228
1229         for (i=r1_bio->mddev->raid_disks; i--; )
1230                 if (r1_bio->bios[i] == bio)
1231                         break;
1232         BUG_ON(i < 0);
1233         update_head_pos(i, r1_bio);
1234         /*
1235          * we have read a block, now it needs to be re-written,
1236          * or re-read if the read failed.
1237          * We don't do much here, just schedule handling by raid1d
1238          */
1239         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1240                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1241
1242         if (atomic_dec_and_test(&r1_bio->remaining))
1243                 reschedule_retry(r1_bio);
1244 }
1245
1246 static void end_sync_write(struct bio *bio, int error)
1247 {
1248         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1249         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1250         mddev_t *mddev = r1_bio->mddev;
1251         conf_t *conf = mddev->private;
1252         int i;
1253         int mirror=0;
1254
1255         for (i = 0; i < conf->raid_disks; i++)
1256                 if (r1_bio->bios[i] == bio) {
1257                         mirror = i;
1258                         break;
1259                 }
1260         if (!uptodate) {
1261                 int sync_blocks = 0;
1262                 sector_t s = r1_bio->sector;
1263                 long sectors_to_go = r1_bio->sectors;
1264                 /* make sure these bits doesn't get cleared. */
1265                 do {
1266                         bitmap_end_sync(mddev->bitmap, s,
1267                                         &sync_blocks, 1);
1268                         s += sync_blocks;
1269                         sectors_to_go -= sync_blocks;
1270                 } while (sectors_to_go > 0);
1271                 md_error(mddev, conf->mirrors[mirror].rdev);
1272         }
1273
1274         update_head_pos(mirror, r1_bio);
1275
1276         if (atomic_dec_and_test(&r1_bio->remaining)) {
1277                 sector_t s = r1_bio->sectors;
1278                 put_buf(r1_bio);
1279                 md_done_sync(mddev, s, uptodate);
1280         }
1281 }
1282
1283 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1284 {
1285         conf_t *conf = mddev->private;
1286         int i;
1287         int disks = conf->raid_disks;
1288         struct bio *bio, *wbio;
1289
1290         bio = r1_bio->bios[r1_bio->read_disk];
1291
1292
1293         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1294                 /* We have read all readable devices.  If we haven't
1295                  * got the block, then there is no hope left.
1296                  * If we have, then we want to do a comparison
1297                  * and skip the write if everything is the same.
1298                  * If any blocks failed to read, then we need to
1299                  * attempt an over-write
1300                  */
1301                 int primary;
1302                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1303                         for (i=0; i<mddev->raid_disks; i++)
1304                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1305                                         md_error(mddev, conf->mirrors[i].rdev);
1306
1307                         md_done_sync(mddev, r1_bio->sectors, 1);
1308                         put_buf(r1_bio);
1309                         return;
1310                 }
1311                 for (primary=0; primary<mddev->raid_disks; primary++)
1312                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1313                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1314                                 r1_bio->bios[primary]->bi_end_io = NULL;
1315                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1316                                 break;
1317                         }
1318                 r1_bio->read_disk = primary;
1319                 for (i=0; i<mddev->raid_disks; i++)
1320                         if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1321                                 int j;
1322                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1323                                 struct bio *pbio = r1_bio->bios[primary];
1324                                 struct bio *sbio = r1_bio->bios[i];
1325
1326                                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1327                                         for (j = vcnt; j-- ; ) {
1328                                                 struct page *p, *s;
1329                                                 p = pbio->bi_io_vec[j].bv_page;
1330                                                 s = sbio->bi_io_vec[j].bv_page;
1331                                                 if (memcmp(page_address(p),
1332                                                            page_address(s),
1333                                                            PAGE_SIZE))
1334                                                         break;
1335                                         }
1336                                 } else
1337                                         j = 0;
1338                                 if (j >= 0)
1339                                         mddev->resync_mismatches += r1_bio->sectors;
1340                                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1341                                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1342                                         sbio->bi_end_io = NULL;
1343                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1344                                 } else {
1345                                         /* fixup the bio for reuse */
1346                                         int size;
1347                                         sbio->bi_vcnt = vcnt;
1348                                         sbio->bi_size = r1_bio->sectors << 9;
1349                                         sbio->bi_idx = 0;
1350                                         sbio->bi_phys_segments = 0;
1351                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1352                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1353                                         sbio->bi_next = NULL;
1354                                         sbio->bi_sector = r1_bio->sector +
1355                                                 conf->mirrors[i].rdev->data_offset;
1356                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1357                                         size = sbio->bi_size;
1358                                         for (j = 0; j < vcnt ; j++) {
1359                                                 struct bio_vec *bi;
1360                                                 bi = &sbio->bi_io_vec[j];
1361                                                 bi->bv_offset = 0;
1362                                                 if (size > PAGE_SIZE)
1363                                                         bi->bv_len = PAGE_SIZE;
1364                                                 else
1365                                                         bi->bv_len = size;
1366                                                 size -= PAGE_SIZE;
1367                                                 memcpy(page_address(bi->bv_page),
1368                                                        page_address(pbio->bi_io_vec[j].bv_page),
1369                                                        PAGE_SIZE);
1370                                         }
1371
1372                                 }
1373                         }
1374         }
1375         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1376                 /* ouch - failed to read all of that.
1377                  * Try some synchronous reads of other devices to get
1378                  * good data, much like with normal read errors.  Only
1379                  * read into the pages we already have so we don't
1380                  * need to re-issue the read request.
1381                  * We don't need to freeze the array, because being in an
1382                  * active sync request, there is no normal IO, and
1383                  * no overlapping syncs.
1384                  */
1385                 sector_t sect = r1_bio->sector;
1386                 int sectors = r1_bio->sectors;
1387                 int idx = 0;
1388
1389                 while(sectors) {
1390                         int s = sectors;
1391                         int d = r1_bio->read_disk;
1392                         int success = 0;
1393                         mdk_rdev_t *rdev;
1394
1395                         if (s > (PAGE_SIZE>>9))
1396                                 s = PAGE_SIZE >> 9;
1397                         do {
1398                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1399                                         /* No rcu protection needed here devices
1400                                          * can only be removed when no resync is
1401                                          * active, and resync is currently active
1402                                          */
1403                                         rdev = conf->mirrors[d].rdev;
1404                                         if (sync_page_io(rdev->bdev,
1405                                                          sect + rdev->data_offset,
1406                                                          s<<9,
1407                                                          bio->bi_io_vec[idx].bv_page,
1408                                                          READ)) {
1409                                                 success = 1;
1410                                                 break;
1411                                         }
1412                                 }
1413                                 d++;
1414                                 if (d == conf->raid_disks)
1415                                         d = 0;
1416                         } while (!success && d != r1_bio->read_disk);
1417
1418                         if (success) {
1419                                 int start = d;
1420                                 /* write it back and re-read */
1421                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1422                                 while (d != r1_bio->read_disk) {
1423                                         if (d == 0)
1424                                                 d = conf->raid_disks;
1425                                         d--;
1426                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1427                                                 continue;
1428                                         rdev = conf->mirrors[d].rdev;
1429                                         atomic_add(s, &rdev->corrected_errors);
1430                                         if (sync_page_io(rdev->bdev,
1431                                                          sect + rdev->data_offset,
1432                                                          s<<9,
1433                                                          bio->bi_io_vec[idx].bv_page,
1434                                                          WRITE) == 0)
1435                                                 md_error(mddev, rdev);
1436                                 }
1437                                 d = start;
1438                                 while (d != r1_bio->read_disk) {
1439                                         if (d == 0)
1440                                                 d = conf->raid_disks;
1441                                         d--;
1442                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1443                                                 continue;
1444                                         rdev = conf->mirrors[d].rdev;
1445                                         if (sync_page_io(rdev->bdev,
1446                                                          sect + rdev->data_offset,
1447                                                          s<<9,
1448                                                          bio->bi_io_vec[idx].bv_page,
1449                                                          READ) == 0)
1450                                                 md_error(mddev, rdev);
1451                                 }
1452                         } else {
1453                                 char b[BDEVNAME_SIZE];
1454                                 /* Cannot read from anywhere, array is toast */
1455                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1456                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1457                                        " for block %llu\n",
1458                                        bdevname(bio->bi_bdev,b),
1459                                        (unsigned long long)r1_bio->sector);
1460                                 md_done_sync(mddev, r1_bio->sectors, 0);
1461                                 put_buf(r1_bio);
1462                                 return;
1463                         }
1464                         sectors -= s;
1465                         sect += s;
1466                         idx ++;
1467                 }
1468         }
1469
1470         /*
1471          * schedule writes
1472          */
1473         atomic_set(&r1_bio->remaining, 1);
1474         for (i = 0; i < disks ; i++) {
1475                 wbio = r1_bio->bios[i];
1476                 if (wbio->bi_end_io == NULL ||
1477                     (wbio->bi_end_io == end_sync_read &&
1478                      (i == r1_bio->read_disk ||
1479                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1480                         continue;
1481
1482                 wbio->bi_rw = WRITE;
1483                 wbio->bi_end_io = end_sync_write;
1484                 atomic_inc(&r1_bio->remaining);
1485                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1486
1487                 generic_make_request(wbio);
1488         }
1489
1490         if (atomic_dec_and_test(&r1_bio->remaining)) {
1491                 /* if we're here, all write(s) have completed, so clean up */
1492                 md_done_sync(mddev, r1_bio->sectors, 1);
1493                 put_buf(r1_bio);
1494         }
1495 }
1496
1497 /*
1498  * This is a kernel thread which:
1499  *
1500  *      1.      Retries failed read operations on working mirrors.
1501  *      2.      Updates the raid superblock when problems encounter.
1502  *      3.      Performs writes following reads for array syncronising.
1503  */
1504
1505 static void fix_read_error(conf_t *conf, int read_disk,
1506                            sector_t sect, int sectors)
1507 {
1508         mddev_t *mddev = conf->mddev;
1509         while(sectors) {
1510                 int s = sectors;
1511                 int d = read_disk;
1512                 int success = 0;
1513                 int start;
1514                 mdk_rdev_t *rdev;
1515
1516                 if (s > (PAGE_SIZE>>9))
1517                         s = PAGE_SIZE >> 9;
1518
1519                 do {
1520                         /* Note: no rcu protection needed here
1521                          * as this is synchronous in the raid1d thread
1522                          * which is the thread that might remove
1523                          * a device.  If raid1d ever becomes multi-threaded....
1524                          */
1525                         rdev = conf->mirrors[d].rdev;
1526                         if (rdev &&
1527                             test_bit(In_sync, &rdev->flags) &&
1528                             sync_page_io(rdev->bdev,
1529                                          sect + rdev->data_offset,
1530                                          s<<9,
1531                                          conf->tmppage, READ))
1532                                 success = 1;
1533                         else {
1534                                 d++;
1535                                 if (d == conf->raid_disks)
1536                                         d = 0;
1537                         }
1538                 } while (!success && d != read_disk);
1539
1540                 if (!success) {
1541                         /* Cannot read from anywhere -- bye bye array */
1542                         md_error(mddev, conf->mirrors[read_disk].rdev);
1543                         break;
1544                 }
1545                 /* write it back and re-read */
1546                 start = d;
1547                 while (d != read_disk) {
1548                         if (d==0)
1549                                 d = conf->raid_disks;
1550                         d--;
1551                         rdev = conf->mirrors[d].rdev;
1552                         if (rdev &&
1553                             test_bit(In_sync, &rdev->flags)) {
1554                                 if (sync_page_io(rdev->bdev,
1555                                                  sect + rdev->data_offset,
1556                                                  s<<9, conf->tmppage, WRITE)
1557                                     == 0)
1558                                         /* Well, this device is dead */
1559                                         md_error(mddev, rdev);
1560                         }
1561                 }
1562                 d = start;
1563                 while (d != read_disk) {
1564                         char b[BDEVNAME_SIZE];
1565                         if (d==0)
1566                                 d = conf->raid_disks;
1567                         d--;
1568                         rdev = conf->mirrors[d].rdev;
1569                         if (rdev &&
1570                             test_bit(In_sync, &rdev->flags)) {
1571                                 if (sync_page_io(rdev->bdev,
1572                                                  sect + rdev->data_offset,
1573                                                  s<<9, conf->tmppage, READ)
1574                                     == 0)
1575                                         /* Well, this device is dead */
1576                                         md_error(mddev, rdev);
1577                                 else {
1578                                         atomic_add(s, &rdev->corrected_errors);
1579                                         printk(KERN_INFO
1580                                                "raid1:%s: read error corrected "
1581                                                "(%d sectors at %llu on %s)\n",
1582                                                mdname(mddev), s,
1583                                                (unsigned long long)(sect +
1584                                                    rdev->data_offset),
1585                                                bdevname(rdev->bdev, b));
1586                                 }
1587                         }
1588                 }
1589                 sectors -= s;
1590                 sect += s;
1591         }
1592 }
1593
1594 static void raid1d(mddev_t *mddev)
1595 {
1596         r1bio_t *r1_bio;
1597         struct bio *bio;
1598         unsigned long flags;
1599         conf_t *conf = mddev->private;
1600         struct list_head *head = &conf->retry_list;
1601         int unplug=0;
1602         mdk_rdev_t *rdev;
1603
1604         md_check_recovery(mddev);
1605
1606         for (;;) {
1607                 char b[BDEVNAME_SIZE];
1608
1609                 unplug += flush_pending_writes(conf);
1610
1611                 spin_lock_irqsave(&conf->device_lock, flags);
1612                 if (list_empty(head)) {
1613                         spin_unlock_irqrestore(&conf->device_lock, flags);
1614                         break;
1615                 }
1616                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1617                 list_del(head->prev);
1618                 conf->nr_queued--;
1619                 spin_unlock_irqrestore(&conf->device_lock, flags);
1620
1621                 mddev = r1_bio->mddev;
1622                 conf = mddev->private;
1623                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1624                         sync_request_write(mddev, r1_bio);
1625                         unplug = 1;
1626                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1627                         /* some requests in the r1bio were BIO_RW_BARRIER
1628                          * requests which failed with -EOPNOTSUPP.  Hohumm..
1629                          * Better resubmit without the barrier.
1630                          * We know which devices to resubmit for, because
1631                          * all others have had their bios[] entry cleared.
1632                          * We already have a nr_pending reference on these rdevs.
1633                          */
1634                         int i;
1635                         const bool do_sync = bio_rw_flagged(r1_bio->master_bio, BIO_RW_SYNCIO);
1636                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1637                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1638                         for (i=0; i < conf->raid_disks; i++)
1639                                 if (r1_bio->bios[i])
1640                                         atomic_inc(&r1_bio->remaining);
1641                         for (i=0; i < conf->raid_disks; i++)
1642                                 if (r1_bio->bios[i]) {
1643                                         struct bio_vec *bvec;
1644                                         int j;
1645
1646                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1647                                         /* copy pages from the failed bio, as
1648                                          * this might be a write-behind device */
1649                                         __bio_for_each_segment(bvec, bio, j, 0)
1650                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1651                                         bio_put(r1_bio->bios[i]);
1652                                         bio->bi_sector = r1_bio->sector +
1653                                                 conf->mirrors[i].rdev->data_offset;
1654                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1655                                         bio->bi_end_io = raid1_end_write_request;
1656                                         bio->bi_rw = WRITE |
1657                                                 (do_sync << BIO_RW_SYNCIO);
1658                                         bio->bi_private = r1_bio;
1659                                         r1_bio->bios[i] = bio;
1660                                         generic_make_request(bio);
1661                                 }
1662                 } else {
1663                         int disk;
1664
1665                         /* we got a read error. Maybe the drive is bad.  Maybe just
1666                          * the block and we can fix it.
1667                          * We freeze all other IO, and try reading the block from
1668                          * other devices.  When we find one, we re-write
1669                          * and check it that fixes the read error.
1670                          * This is all done synchronously while the array is
1671                          * frozen
1672                          */
1673                         if (mddev->ro == 0) {
1674                                 freeze_array(conf);
1675                                 fix_read_error(conf, r1_bio->read_disk,
1676                                                r1_bio->sector,
1677                                                r1_bio->sectors);
1678                                 unfreeze_array(conf);
1679                         } else
1680                                 md_error(mddev,
1681                                          conf->mirrors[r1_bio->read_disk].rdev);
1682
1683                         bio = r1_bio->bios[r1_bio->read_disk];
1684                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1685                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1686                                        " read error for block %llu\n",
1687                                        bdevname(bio->bi_bdev,b),
1688                                        (unsigned long long)r1_bio->sector);
1689                                 raid_end_bio_io(r1_bio);
1690                         } else {
1691                                 const bool do_sync = bio_rw_flagged(r1_bio->master_bio, BIO_RW_SYNCIO);
1692                                 r1_bio->bios[r1_bio->read_disk] =
1693                                         mddev->ro ? IO_BLOCKED : NULL;
1694                                 r1_bio->read_disk = disk;
1695                                 bio_put(bio);
1696                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1697                                 r1_bio->bios[r1_bio->read_disk] = bio;
1698                                 rdev = conf->mirrors[disk].rdev;
1699                                 if (printk_ratelimit())
1700                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1701                                                " another mirror\n",
1702                                                bdevname(rdev->bdev,b),
1703                                                (unsigned long long)r1_bio->sector);
1704                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1705                                 bio->bi_bdev = rdev->bdev;
1706                                 bio->bi_end_io = raid1_end_read_request;
1707                                 bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
1708                                 bio->bi_private = r1_bio;
1709                                 unplug = 1;
1710                                 generic_make_request(bio);
1711                         }
1712                 }
1713                 cond_resched();
1714         }
1715         if (unplug)
1716                 unplug_slaves(mddev);
1717 }
1718
1719
1720 static int init_resync(conf_t *conf)
1721 {
1722         int buffs;
1723
1724         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1725         BUG_ON(conf->r1buf_pool);
1726         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1727                                           conf->poolinfo);
1728         if (!conf->r1buf_pool)
1729                 return -ENOMEM;
1730         conf->next_resync = 0;
1731         return 0;
1732 }
1733
1734 /*
1735  * perform a "sync" on one "block"
1736  *
1737  * We need to make sure that no normal I/O request - particularly write
1738  * requests - conflict with active sync requests.
1739  *
1740  * This is achieved by tracking pending requests and a 'barrier' concept
1741  * that can be installed to exclude normal IO requests.
1742  */
1743
1744 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1745 {
1746         conf_t *conf = mddev->private;
1747         r1bio_t *r1_bio;
1748         struct bio *bio;
1749         sector_t max_sector, nr_sectors;
1750         int disk = -1;
1751         int i;
1752         int wonly = -1;
1753         int write_targets = 0, read_targets = 0;
1754         int sync_blocks;
1755         int still_degraded = 0;
1756
1757         if (!conf->r1buf_pool)
1758         {
1759 /*
1760                 printk("sync start - bitmap %p\n", mddev->bitmap);
1761 */
1762                 if (init_resync(conf))
1763                         return 0;
1764         }
1765
1766         max_sector = mddev->dev_sectors;
1767         if (sector_nr >= max_sector) {
1768                 /* If we aborted, we need to abort the
1769                  * sync on the 'current' bitmap chunk (there will
1770                  * only be one in raid1 resync.
1771                  * We can find the current addess in mddev->curr_resync
1772                  */
1773                 if (mddev->curr_resync < max_sector) /* aborted */
1774                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1775                                                 &sync_blocks, 1);
1776                 else /* completed sync */
1777                         conf->fullsync = 0;
1778
1779                 bitmap_close_sync(mddev->bitmap);
1780                 close_sync(conf);
1781                 return 0;
1782         }
1783
1784         if (mddev->bitmap == NULL &&
1785             mddev->recovery_cp == MaxSector &&
1786             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1787             conf->fullsync == 0) {
1788                 *skipped = 1;
1789                 return max_sector - sector_nr;
1790         }
1791         /* before building a request, check if we can skip these blocks..
1792          * This call the bitmap_start_sync doesn't actually record anything
1793          */
1794         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1795             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1796                 /* We can skip this block, and probably several more */
1797                 *skipped = 1;
1798                 return sync_blocks;
1799         }
1800         /*
1801          * If there is non-resync activity waiting for a turn,
1802          * and resync is going fast enough,
1803          * then let it though before starting on this new sync request.
1804          */
1805         if (!go_faster && conf->nr_waiting)
1806                 msleep_interruptible(1000);
1807
1808         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1809         raise_barrier(conf);
1810
1811         conf->next_resync = sector_nr;
1812
1813         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1814         rcu_read_lock();
1815         /*
1816          * If we get a correctably read error during resync or recovery,
1817          * we might want to read from a different device.  So we
1818          * flag all drives that could conceivably be read from for READ,
1819          * and any others (which will be non-In_sync devices) for WRITE.
1820          * If a read fails, we try reading from something else for which READ
1821          * is OK.
1822          */
1823
1824         r1_bio->mddev = mddev;
1825         r1_bio->sector = sector_nr;
1826         r1_bio->state = 0;
1827         set_bit(R1BIO_IsSync, &r1_bio->state);
1828
1829         for (i=0; i < conf->raid_disks; i++) {
1830                 mdk_rdev_t *rdev;
1831                 bio = r1_bio->bios[i];
1832
1833                 /* take from bio_init */
1834                 bio->bi_next = NULL;
1835                 bio->bi_flags |= 1 << BIO_UPTODATE;
1836                 bio->bi_rw = READ;
1837                 bio->bi_vcnt = 0;
1838                 bio->bi_idx = 0;
1839                 bio->bi_phys_segments = 0;
1840                 bio->bi_size = 0;
1841                 bio->bi_end_io = NULL;
1842                 bio->bi_private = NULL;
1843
1844                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1845                 if (rdev == NULL ||
1846                            test_bit(Faulty, &rdev->flags)) {
1847                         still_degraded = 1;
1848                         continue;
1849                 } else if (!test_bit(In_sync, &rdev->flags)) {
1850                         bio->bi_rw = WRITE;
1851                         bio->bi_end_io = end_sync_write;
1852                         write_targets ++;
1853                 } else {
1854                         /* may need to read from here */
1855                         bio->bi_rw = READ;
1856                         bio->bi_end_io = end_sync_read;
1857                         if (test_bit(WriteMostly, &rdev->flags)) {
1858                                 if (wonly < 0)
1859                                         wonly = i;
1860                         } else {
1861                                 if (disk < 0)
1862                                         disk = i;
1863                         }
1864                         read_targets++;
1865                 }
1866                 atomic_inc(&rdev->nr_pending);
1867                 bio->bi_sector = sector_nr + rdev->data_offset;
1868                 bio->bi_bdev = rdev->bdev;
1869                 bio->bi_private = r1_bio;
1870         }
1871         rcu_read_unlock();
1872         if (disk < 0)
1873                 disk = wonly;
1874         r1_bio->read_disk = disk;
1875
1876         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1877                 /* extra read targets are also write targets */
1878                 write_targets += read_targets-1;
1879
1880         if (write_targets == 0 || read_targets == 0) {
1881                 /* There is nowhere to write, so all non-sync
1882                  * drives must be failed - so we are finished
1883                  */
1884                 sector_t rv = max_sector - sector_nr;
1885                 *skipped = 1;
1886                 put_buf(r1_bio);
1887                 return rv;
1888         }
1889
1890         if (max_sector > mddev->resync_max)
1891                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1892         nr_sectors = 0;
1893         sync_blocks = 0;
1894         do {
1895                 struct page *page;
1896                 int len = PAGE_SIZE;
1897                 if (sector_nr + (len>>9) > max_sector)
1898                         len = (max_sector - sector_nr) << 9;
1899                 if (len == 0)
1900                         break;
1901                 if (sync_blocks == 0) {
1902                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1903                                                &sync_blocks, still_degraded) &&
1904                             !conf->fullsync &&
1905                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1906                                 break;
1907                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1908                         if (len > (sync_blocks<<9))
1909                                 len = sync_blocks<<9;
1910                 }
1911
1912                 for (i=0 ; i < conf->raid_disks; i++) {
1913                         bio = r1_bio->bios[i];
1914                         if (bio->bi_end_io) {
1915                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1916                                 if (bio_add_page(bio, page, len, 0) == 0) {
1917                                         /* stop here */
1918                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1919                                         while (i > 0) {
1920                                                 i--;
1921                                                 bio = r1_bio->bios[i];
1922                                                 if (bio->bi_end_io==NULL)
1923                                                         continue;
1924                                                 /* remove last page from this bio */
1925                                                 bio->bi_vcnt--;
1926                                                 bio->bi_size -= len;
1927                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1928                                         }
1929                                         goto bio_full;
1930                                 }
1931                         }
1932                 }
1933                 nr_sectors += len>>9;
1934                 sector_nr += len>>9;
1935                 sync_blocks -= (len>>9);
1936         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1937  bio_full:
1938         r1_bio->sectors = nr_sectors;
1939
1940         /* For a user-requested sync, we read all readable devices and do a
1941          * compare
1942          */
1943         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1944                 atomic_set(&r1_bio->remaining, read_targets);
1945                 for (i=0; i<conf->raid_disks; i++) {
1946                         bio = r1_bio->bios[i];
1947                         if (bio->bi_end_io == end_sync_read) {
1948                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1949                                 generic_make_request(bio);
1950                         }
1951                 }
1952         } else {
1953                 atomic_set(&r1_bio->remaining, 1);
1954                 bio = r1_bio->bios[r1_bio->read_disk];
1955                 md_sync_acct(bio->bi_bdev, nr_sectors);
1956                 generic_make_request(bio);
1957
1958         }
1959         return nr_sectors;
1960 }
1961
1962 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1963 {
1964         if (sectors)
1965                 return sectors;
1966
1967         return mddev->dev_sectors;
1968 }
1969
1970 static conf_t *setup_conf(mddev_t *mddev)
1971 {
1972         conf_t *conf;
1973         int i;
1974         mirror_info_t *disk;
1975         mdk_rdev_t *rdev;
1976         int err = -ENOMEM;
1977
1978         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1979         if (!conf)
1980                 goto abort;
1981
1982         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1983                                  GFP_KERNEL);
1984         if (!conf->mirrors)
1985                 goto abort;
1986
1987         conf->tmppage = alloc_page(GFP_KERNEL);
1988         if (!conf->tmppage)
1989                 goto abort;
1990
1991         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1992         if (!conf->poolinfo)
1993                 goto abort;
1994         conf->poolinfo->raid_disks = mddev->raid_disks;
1995         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1996                                           r1bio_pool_free,
1997                                           conf->poolinfo);
1998         if (!conf->r1bio_pool)
1999                 goto abort;
2000
2001         conf->poolinfo->mddev = mddev;
2002
2003         spin_lock_init(&conf->device_lock);
2004         list_for_each_entry(rdev, &mddev->disks, same_set) {
2005                 int disk_idx = rdev->raid_disk;
2006                 if (disk_idx >= mddev->raid_disks
2007                     || disk_idx < 0)
2008                         continue;
2009                 disk = conf->mirrors + disk_idx;
2010
2011                 disk->rdev = rdev;
2012
2013                 disk->head_position = 0;
2014         }
2015         conf->raid_disks = mddev->raid_disks;
2016         conf->mddev = mddev;
2017         INIT_LIST_HEAD(&conf->retry_list);
2018
2019         spin_lock_init(&conf->resync_lock);
2020         init_waitqueue_head(&conf->wait_barrier);
2021
2022         bio_list_init(&conf->pending_bio_list);
2023         bio_list_init(&conf->flushing_bio_list);
2024
2025         conf->last_used = -1;
2026         for (i = 0; i < conf->raid_disks; i++) {
2027
2028                 disk = conf->mirrors + i;
2029
2030                 if (!disk->rdev ||
2031                     !test_bit(In_sync, &disk->rdev->flags)) {
2032                         disk->head_position = 0;
2033                         if (disk->rdev)
2034                                 conf->fullsync = 1;
2035                 } else if (conf->last_used < 0)
2036                         /*
2037                          * The first working device is used as a
2038                          * starting point to read balancing.
2039                          */
2040                         conf->last_used = i;
2041         }
2042
2043         err = -EIO;
2044         if (conf->last_used < 0) {
2045                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
2046                        mdname(mddev));
2047                 goto abort;
2048         }
2049         err = -ENOMEM;
2050         conf->thread = md_register_thread(raid1d, mddev, NULL);
2051         if (!conf->thread) {
2052                 printk(KERN_ERR
2053                        "raid1: couldn't allocate thread for %s\n",
2054                        mdname(mddev));
2055                 goto abort;
2056         }
2057
2058         return conf;
2059
2060  abort:
2061         if (conf) {
2062                 if (conf->r1bio_pool)
2063                         mempool_destroy(conf->r1bio_pool);
2064                 kfree(conf->mirrors);
2065                 safe_put_page(conf->tmppage);
2066                 kfree(conf->poolinfo);
2067                 kfree(conf);
2068         }
2069         return ERR_PTR(err);
2070 }
2071
2072 static int run(mddev_t *mddev)
2073 {
2074         conf_t *conf;
2075         int i;
2076         mdk_rdev_t *rdev;
2077
2078         if (mddev->level != 1) {
2079                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
2080                        mdname(mddev), mddev->level);
2081                 return -EIO;
2082         }
2083         if (mddev->reshape_position != MaxSector) {
2084                 printk("raid1: %s: reshape_position set but not supported\n",
2085                        mdname(mddev));
2086                 return -EIO;
2087         }
2088         /*
2089          * copy the already verified devices into our private RAID1
2090          * bookkeeping area. [whatever we allocate in run(),
2091          * should be freed in stop()]
2092          */
2093         if (mddev->private == NULL)
2094                 conf = setup_conf(mddev);
2095         else
2096                 conf = mddev->private;
2097
2098         if (IS_ERR(conf))
2099                 return PTR_ERR(conf);
2100
2101         mddev->queue->queue_lock = &conf->device_lock;
2102         list_for_each_entry(rdev, &mddev->disks, same_set) {
2103                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2104                                   rdev->data_offset << 9);
2105                 /* as we don't honour merge_bvec_fn, we must never risk
2106                  * violating it, so limit ->max_segments to 1 lying within
2107                  * a single page, as a one page request is never in violation.
2108                  */
2109                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2110                         blk_queue_max_segments(mddev->queue, 1);
2111                         blk_queue_segment_boundary(mddev->queue,
2112                                                    PAGE_CACHE_SIZE - 1);
2113                 }
2114         }
2115
2116         mddev->degraded = 0;
2117         for (i=0; i < conf->raid_disks; i++)
2118                 if (conf->mirrors[i].rdev == NULL ||
2119                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2120                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2121                         mddev->degraded++;
2122
2123         if (conf->raid_disks - mddev->degraded == 1)
2124                 mddev->recovery_cp = MaxSector;
2125
2126         if (mddev->recovery_cp != MaxSector)
2127                 printk(KERN_NOTICE "raid1: %s is not clean"
2128                        " -- starting background reconstruction\n",
2129                        mdname(mddev));
2130         printk(KERN_INFO
2131                 "raid1: raid set %s active with %d out of %d mirrors\n",
2132                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2133                 mddev->raid_disks);
2134
2135         /*
2136          * Ok, everything is just fine now
2137          */
2138         mddev->thread = conf->thread;
2139         conf->thread = NULL;
2140         mddev->private = conf;
2141
2142         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2143
2144         mddev->queue->unplug_fn = raid1_unplug;
2145         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2146         mddev->queue->backing_dev_info.congested_data = mddev;
2147         md_integrity_register(mddev);
2148         return 0;
2149 }
2150
2151 static int stop(mddev_t *mddev)
2152 {
2153         conf_t *conf = mddev->private;
2154         struct bitmap *bitmap = mddev->bitmap;
2155         int behind_wait = 0;
2156
2157         /* wait for behind writes to complete */
2158         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2159                 behind_wait++;
2160                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
2161                 set_current_state(TASK_UNINTERRUPTIBLE);
2162                 schedule_timeout(HZ); /* wait a second */
2163                 /* need to kick something here to make sure I/O goes? */
2164         }
2165
2166         raise_barrier(conf);
2167         lower_barrier(conf);
2168
2169         md_unregister_thread(mddev->thread);
2170         mddev->thread = NULL;
2171         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2172         if (conf->r1bio_pool)
2173                 mempool_destroy(conf->r1bio_pool);
2174         kfree(conf->mirrors);
2175         kfree(conf->poolinfo);
2176         kfree(conf);
2177         mddev->private = NULL;
2178         return 0;
2179 }
2180
2181 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2182 {
2183         /* no resync is happening, and there is enough space
2184          * on all devices, so we can resize.
2185          * We need to make sure resync covers any new space.
2186          * If the array is shrinking we should possibly wait until
2187          * any io in the removed space completes, but it hardly seems
2188          * worth it.
2189          */
2190         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2191         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2192                 return -EINVAL;
2193         set_capacity(mddev->gendisk, mddev->array_sectors);
2194         mddev->changed = 1;
2195         revalidate_disk(mddev->gendisk);
2196         if (sectors > mddev->dev_sectors &&
2197             mddev->recovery_cp == MaxSector) {
2198                 mddev->recovery_cp = mddev->dev_sectors;
2199                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2200         }
2201         mddev->dev_sectors = sectors;
2202         mddev->resync_max_sectors = sectors;
2203         return 0;
2204 }
2205
2206 static int raid1_reshape(mddev_t *mddev)
2207 {
2208         /* We need to:
2209          * 1/ resize the r1bio_pool
2210          * 2/ resize conf->mirrors
2211          *
2212          * We allocate a new r1bio_pool if we can.
2213          * Then raise a device barrier and wait until all IO stops.
2214          * Then resize conf->mirrors and swap in the new r1bio pool.
2215          *
2216          * At the same time, we "pack" the devices so that all the missing
2217          * devices have the higher raid_disk numbers.
2218          */
2219         mempool_t *newpool, *oldpool;
2220         struct pool_info *newpoolinfo;
2221         mirror_info_t *newmirrors;
2222         conf_t *conf = mddev->private;
2223         int cnt, raid_disks;
2224         unsigned long flags;
2225         int d, d2, err;
2226
2227         /* Cannot change chunk_size, layout, or level */
2228         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2229             mddev->layout != mddev->new_layout ||
2230             mddev->level != mddev->new_level) {
2231                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2232                 mddev->new_layout = mddev->layout;
2233                 mddev->new_level = mddev->level;
2234                 return -EINVAL;
2235         }
2236
2237         err = md_allow_write(mddev);
2238         if (err)
2239                 return err;
2240
2241         raid_disks = mddev->raid_disks + mddev->delta_disks;
2242
2243         if (raid_disks < conf->raid_disks) {
2244                 cnt=0;
2245                 for (d= 0; d < conf->raid_disks; d++)
2246                         if (conf->mirrors[d].rdev)
2247                                 cnt++;
2248                 if (cnt > raid_disks)
2249                         return -EBUSY;
2250         }
2251
2252         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2253         if (!newpoolinfo)
2254                 return -ENOMEM;
2255         newpoolinfo->mddev = mddev;
2256         newpoolinfo->raid_disks = raid_disks;
2257
2258         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2259                                  r1bio_pool_free, newpoolinfo);
2260         if (!newpool) {
2261                 kfree(newpoolinfo);
2262                 return -ENOMEM;
2263         }
2264         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2265         if (!newmirrors) {
2266                 kfree(newpoolinfo);
2267                 mempool_destroy(newpool);
2268                 return -ENOMEM;
2269         }
2270
2271         raise_barrier(conf);
2272
2273         /* ok, everything is stopped */
2274         oldpool = conf->r1bio_pool;
2275         conf->r1bio_pool = newpool;
2276
2277         for (d = d2 = 0; d < conf->raid_disks; d++) {
2278                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2279                 if (rdev && rdev->raid_disk != d2) {
2280                         char nm[20];
2281                         sprintf(nm, "rd%d", rdev->raid_disk);
2282                         sysfs_remove_link(&mddev->kobj, nm);
2283                         rdev->raid_disk = d2;
2284                         sprintf(nm, "rd%d", rdev->raid_disk);
2285                         sysfs_remove_link(&mddev->kobj, nm);
2286                         if (sysfs_create_link(&mddev->kobj,
2287                                               &rdev->kobj, nm))
2288                                 printk(KERN_WARNING
2289                                        "md/raid1: cannot register "
2290                                        "%s for %s\n",
2291                                        nm, mdname(mddev));
2292                 }
2293                 if (rdev)
2294                         newmirrors[d2++].rdev = rdev;
2295         }
2296         kfree(conf->mirrors);
2297         conf->mirrors = newmirrors;
2298         kfree(conf->poolinfo);
2299         conf->poolinfo = newpoolinfo;
2300
2301         spin_lock_irqsave(&conf->device_lock, flags);
2302         mddev->degraded += (raid_disks - conf->raid_disks);
2303         spin_unlock_irqrestore(&conf->device_lock, flags);
2304         conf->raid_disks = mddev->raid_disks = raid_disks;
2305         mddev->delta_disks = 0;
2306
2307         conf->last_used = 0; /* just make sure it is in-range */
2308         lower_barrier(conf);
2309
2310         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2311         md_wakeup_thread(mddev->thread);
2312
2313         mempool_destroy(oldpool);
2314         return 0;
2315 }
2316
2317 static void raid1_quiesce(mddev_t *mddev, int state)
2318 {
2319         conf_t *conf = mddev->private;
2320
2321         switch(state) {
2322         case 2: /* wake for suspend */
2323                 wake_up(&conf->wait_barrier);
2324                 break;
2325         case 1:
2326                 raise_barrier(conf);
2327                 break;
2328         case 0:
2329                 lower_barrier(conf);
2330                 break;
2331         }
2332 }
2333
2334 static void *raid1_takeover(mddev_t *mddev)
2335 {
2336         /* raid1 can take over:
2337          *  raid5 with 2 devices, any layout or chunk size
2338          */
2339         if (mddev->level == 5 && mddev->raid_disks == 2) {
2340                 conf_t *conf;
2341                 mddev->new_level = 1;
2342                 mddev->new_layout = 0;
2343                 mddev->new_chunk_sectors = 0;
2344                 conf = setup_conf(mddev);
2345                 if (!IS_ERR(conf))
2346                         conf->barrier = 1;
2347                 return conf;
2348         }
2349         return ERR_PTR(-EINVAL);
2350 }
2351
2352 static struct mdk_personality raid1_personality =
2353 {
2354         .name           = "raid1",
2355         .level          = 1,
2356         .owner          = THIS_MODULE,
2357         .make_request   = make_request,
2358         .run            = run,
2359         .stop           = stop,
2360         .status         = status,
2361         .error_handler  = error,
2362         .hot_add_disk   = raid1_add_disk,
2363         .hot_remove_disk= raid1_remove_disk,
2364         .spare_active   = raid1_spare_active,
2365         .sync_request   = sync_request,
2366         .resize         = raid1_resize,
2367         .size           = raid1_size,
2368         .check_reshape  = raid1_reshape,
2369         .quiesce        = raid1_quiesce,
2370         .takeover       = raid1_takeover,
2371 };
2372
2373 static int __init raid_init(void)
2374 {
2375         return register_md_personality(&raid1_personality);
2376 }
2377
2378 static void raid_exit(void)
2379 {
2380         unregister_md_personality(&raid1_personality);
2381 }
2382
2383 module_init(raid_init);
2384 module_exit(raid_exit);
2385 MODULE_LICENSE("GPL");
2386 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2387 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2388 MODULE_ALIAS("md-raid1");
2389 MODULE_ALIAS("md-level-1");