drivers/md/raid1.c

   1 /*
   2  * raid1.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5  *
   6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7  *
   8  * RAID-1 management functions.
   9  *
  10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11  *
  12  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14  *
  15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  16  * bitmapped intelligence in resync:
  17  *
  18  *      - bitmap marked during normal i/o
  19  *      - bitmap used to skip nondirty blocks during sync
  20  *
  21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  22  * - persistent bitmap code
  23  *
  24  * This program is free software; you can redistribute it and/or modify
  25  * it under the terms of the GNU General Public License as published by
  26  * the Free Software Foundation; either version 2, or (at your option)
  27  * any later version.
  28  *
  29  * You should have received a copy of the GNU General Public License
  30  * (for example /usr/src/linux/COPYING); if not, write to the Free
  31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  32  */
  33
  34 #include "dm-bio-list.h"
  35 #include <linux/raid/raid1.h>
  36 #include <linux/raid/bitmap.h>
  37
  38 #define DEBUG 0
  39 #if DEBUG
  40 #define PRINTK(x...) printk(x)
  41 #else
  42 #define PRINTK(x...)
  43 #endif
  44
  45 /*
  46  * Number of guaranteed r1bios in case of extreme VM load:
  47  */
  48 #define NR_RAID1_BIOS 256
  49
  50
  51 static void unplug_slaves(mddev_t *mddev);
  52
  53 static void allow_barrier(conf_t *conf);
  54 static void lower_barrier(conf_t *conf);
  55
  56 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  57 {
  58         struct pool_info *pi = data;
  59         r1bio_t *r1_bio;
  60         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
  61
  62         /* allocate a r1bio with room for raid_disks entries in the bios array */
  63         r1_bio = kzalloc(size, gfp_flags);
  64         if (!r1_bio)
  65                 unplug_slaves(pi->mddev);
  66
  67         return r1_bio;
  68 }
  69
  70 static void r1bio_pool_free(void *r1_bio, void *data)
  71 {
  72         kfree(r1_bio);
  73 }
  74
  75 #define RESYNC_BLOCK_SIZE (64*1024)
  76 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  77 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  78 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  79 #define RESYNC_WINDOW (2048*1024)
  80
  81 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  82 {
  83         struct pool_info *pi = data;
  84         struct page *page;
  85         r1bio_t *r1_bio;
  86         struct bio *bio;
  87         int i, j;
  88
  89         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  90         if (!r1_bio) {
  91                 unplug_slaves(pi->mddev);
  92                 return NULL;
  93         }
  94
  95         /*
  96          * Allocate bios : 1 for reading, n-1 for writing
  97          */
  98         for (j = pi->raid_disks ; j-- ; ) {
  99                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 100                 if (!bio)
 101                         goto out_free_bio;
 102                 r1_bio->bios[j] = bio;
 103         }
 104         /*
 105          * Allocate RESYNC_PAGES data pages and attach them to
 106          * the first bio.
 107          * If this is a user-requested check/repair, allocate
 108          * RESYNC_PAGES for each bio.
 109          */
 110         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 111                 j = pi->raid_disks;
 112         else
 113                 j = 1;
 114         while(j--) {
 115                 bio = r1_bio->bios[j];
 116                 for (i = 0; i < RESYNC_PAGES; i++) {
 117                         page = alloc_page(gfp_flags);
 118                         if (unlikely(!page))
 119                                 goto out_free_pages;
 120
 121                         bio->bi_io_vec[i].bv_page = page;
 122                 }
 123         }
 124         /* If not user-requests, copy the page pointers to all bios */
 125         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 126                 for (i=0; i<RESYNC_PAGES ; i++)
 127                         for (j=1; j<pi->raid_disks; j++)
 128                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 129                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 130         }
 131
 132         r1_bio->master_bio = NULL;
 133
 134         return r1_bio;
 135
 136 out_free_pages:
 137         for (i=0; i < RESYNC_PAGES ; i++)
 138                 for (j=0 ; j < pi->raid_disks; j++)
 139                         safe_put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 140         j = -1;
 141 out_free_bio:
 142         while ( ++j < pi->raid_disks )
 143                 bio_put(r1_bio->bios[j]);
 144         r1bio_pool_free(r1_bio, data);
 145         return NULL;
 146 }
 147
 148 static void r1buf_pool_free(void *__r1_bio, void *data)
 149 {
 150         struct pool_info *pi = data;
 151         int i,j;
 152         r1bio_t *r1bio = __r1_bio;
 153
 154         for (i = 0; i < RESYNC_PAGES; i++)
 155                 for (j = pi->raid_disks; j-- ;) {
 156                         if (j == 0 ||
 157                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 158                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 159                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 160                 }
 161         for (i=0 ; i < pi->raid_disks; i++)
 162                 bio_put(r1bio->bios[i]);
 163
 164         r1bio_pool_free(r1bio, data);
 165 }
 166
 167 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 168 {
 169         int i;
 170
 171         for (i = 0; i < conf->raid_disks; i++) {
 172                 struct bio **bio = r1_bio->bios + i;
 173                 if (*bio && *bio != IO_BLOCKED)
 174                         bio_put(*bio);
 175                 *bio = NULL;
 176         }
 177 }
 178
 179 static void free_r1bio(r1bio_t *r1_bio)
 180 {
 181         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 182
 183         /*
 184          * Wake up any possible resync thread that waits for the device
 185          * to go idle.
 186          */
 187         allow_barrier(conf);
 188
 189         put_all_bios(conf, r1_bio);
 190         mempool_free(r1_bio, conf->r1bio_pool);
 191 }
 192
 193 static void put_buf(r1bio_t *r1_bio)
 194 {
 195         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 196         int i;
 197
 198         for (i=0; i<conf->raid_disks; i++) {
 199                 struct bio *bio = r1_bio->bios[i];
 200                 if (bio->bi_end_io)
 201                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 202         }
 203
 204         mempool_free(r1_bio, conf->r1buf_pool);
 205
 206         lower_barrier(conf);
 207 }
 208
 209 static void reschedule_retry(r1bio_t *r1_bio)
 210 {
 211         unsigned long flags;
 212         mddev_t *mddev = r1_bio->mddev;
 213         conf_t *conf = mddev_to_conf(mddev);
 214
 215         spin_lock_irqsave(&conf->device_lock, flags);
 216         list_add(&r1_bio->retry_list, &conf->retry_list);
 217         conf->nr_queued ++;
 218         spin_unlock_irqrestore(&conf->device_lock, flags);
 219
 220         wake_up(&conf->wait_barrier);
 221         md_wakeup_thread(mddev->thread);
 222 }
 223
 224 /*
 225  * raid_end_bio_io() is called when we have finished servicing a mirrored
 226  * operation and are ready to return a success/failure code to the buffer
 227  * cache layer.
 228  */
 229 static void raid_end_bio_io(r1bio_t *r1_bio)
 230 {
 231         struct bio *bio = r1_bio->master_bio;
 232
 233         /* if nobody has done the final endio yet, do it now */
 234         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 235                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
 236                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 237                         (unsigned long long) bio->bi_sector,
 238                         (unsigned long long) bio->bi_sector +
 239                                 (bio->bi_size >> 9) - 1);
 240
 241                 bio_endio(bio, bio->bi_size,
 242                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 243         }
 244         free_r1bio(r1_bio);
 245 }
 246
 247 /*
 248  * Update disk head position estimator based on IRQ completion info.
 249  */
 250 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 251 {
 252         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 253
 254         conf->mirrors[disk].head_position =
 255                 r1_bio->sector + (r1_bio->sectors);
 256 }
 257
 258 static int raid1_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
 259 {
 260         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 261         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 262         int mirror;
 263         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 264
 265         if (bio->bi_size)
 266                 return 1;
 267
 268         mirror = r1_bio->read_disk;
 269         /*
 270          * this branch is our 'one mirror IO has finished' event handler:
 271          */
 272         update_head_pos(mirror, r1_bio);
 273
 274         if (uptodate || (conf->raid_disks - conf->mddev->degraded) <= 1) {
 275                 /*
 276                  * Set R1BIO_Uptodate in our master bio, so that
 277                  * we will return a good error code for to the higher
 278                  * levels even if IO on some other mirrored buffer fails.
 279                  *
 280                  * The 'master' represents the composite IO operation to
 281                  * user-side. So if something waits for IO, then it will
 282                  * wait for the 'master' bio.
 283                  */
 284                 if (uptodate)
 285                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 286
 287                 raid_end_bio_io(r1_bio);
 288         } else {
 289                 /*
 290                  * oops, read error:
 291                  */
 292                 char b[BDEVNAME_SIZE];
 293                 if (printk_ratelimit())
 294                         printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
 295                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 296                 reschedule_retry(r1_bio);
 297         }
 298
 299         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 300         return 0;
 301 }
 302
 303 static int raid1_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
 304 {
 305         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 306         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 307         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 308         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 309         struct bio *to_put = NULL;
 310
 311         if (bio->bi_size)
 312                 return 1;
 313
 314         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 315                 if (r1_bio->bios[mirror] == bio)
 316                         break;
 317
 318         if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
 319                 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
 320                 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
 321                 r1_bio->mddev->barriers_work = 0;
 322                 /* Don't rdev_dec_pending in this branch - keep it for the retry */
 323         } else {
 324                 /*
 325                  * this branch is our 'one mirror IO has finished' event handler:
 326                  */
 327                 r1_bio->bios[mirror] = NULL;
 328                 to_put = bio;
 329                 if (!uptodate) {
 330                         md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 331                         /* an I/O failed, we can't clear the bitmap */
 332                         set_bit(R1BIO_Degraded, &r1_bio->state);
 333                 } else
 334                         /*
 335                          * Set R1BIO_Uptodate in our master bio, so that
 336                          * we will return a good error code for to the higher
 337                          * levels even if IO on some other mirrored buffer fails.
 338                          *
 339                          * The 'master' represents the composite IO operation to
 340                          * user-side. So if something waits for IO, then it will
 341                          * wait for the 'master' bio.
 342                          */
 343                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 344
 345                 update_head_pos(mirror, r1_bio);
 346
 347                 if (behind) {
 348                         if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 349                                 atomic_dec(&r1_bio->behind_remaining);
 350
 351                         /* In behind mode, we ACK the master bio once the I/O has safely
 352                          * reached all non-writemostly disks. Setting the Returned bit
 353                          * ensures that this gets done only once -- we don't ever want to
 354                          * return -EIO here, instead we'll wait */
 355
 356                         if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 357                             test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 358                                 /* Maybe we can return now */
 359                                 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 360                                         struct bio *mbio = r1_bio->master_bio;
 361                                         PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 362                                                (unsigned long long) mbio->bi_sector,
 363                                                (unsigned long long) mbio->bi_sector +
 364                                                (mbio->bi_size >> 9) - 1);
 365                                         bio_endio(mbio, mbio->bi_size, 0);
 366                                 }
 367                         }
 368                 }
 369                 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 370         }
 371         /*
 372          *
 373          * Let's see if all mirrored write operations have finished
 374          * already.
 375          */
 376         if (atomic_dec_and_test(&r1_bio->remaining)) {
 377                 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
 378                         reschedule_retry(r1_bio);
 379                 else {
 380                         /* it really is the end of this request */
 381                         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 382                                 /* free extra copy of the data pages */
 383                                 int i = bio->bi_vcnt;
 384                                 while (i--)
 385                                         safe_put_page(bio->bi_io_vec[i].bv_page);
 386                         }
 387                         /* clear the bitmap if all writes complete successfully */
 388                         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 389                                         r1_bio->sectors,
 390                                         !test_bit(R1BIO_Degraded, &r1_bio->state),
 391                                         behind);
 392                         md_write_end(r1_bio->mddev);
 393                         raid_end_bio_io(r1_bio);
 394                 }
 395         }
 396
 397         if (to_put)
 398                 bio_put(to_put);
 399
 400         return 0;
 401 }
 402
 403
 404 /*
 405  * This routine returns the disk from which the requested read should
 406  * be done. There is a per-array 'next expected sequential IO' sector
 407  * number - if this matches on the next IO then we use the last disk.
 408  * There is also a per-disk 'last know head position' sector that is
 409  * maintained from IRQ contexts, both the normal and the resync IO
 410  * completion handlers update this position correctly. If there is no
 411  * perfect sequential match then we pick the disk whose head is closest.
 412  *
 413  * If there are 2 mirrors in the same 2 devices, performance degrades
 414  * because position is mirror, not device based.
 415  *
 416  * The rdev for the device selected will have nr_pending incremented.
 417  */
 418 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 419 {
 420         const unsigned long this_sector = r1_bio->sector;
 421         int new_disk = conf->last_used, disk = new_disk;
 422         int wonly_disk = -1;
 423         const int sectors = r1_bio->sectors;
 424         sector_t new_distance, current_distance;
 425         mdk_rdev_t *rdev;
 426
 427         rcu_read_lock();
 428         /*
 429          * Check if we can balance. We can balance on the whole
 430          * device if no resync is going on, or below the resync window.
 431          * We take the first readable disk when above the resync window.
 432          */
 433  retry:
 434         if (conf->mddev->recovery_cp < MaxSector &&
 435             (this_sector + sectors >= conf->next_resync)) {
 436                 /* Choose the first operation device, for consistancy */
 437                 new_disk = 0;
 438
 439                 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 440                      r1_bio->bios[new_disk] == IO_BLOCKED ||
 441                      !rdev || !test_bit(In_sync, &rdev->flags)
 442                              || test_bit(WriteMostly, &rdev->flags);
 443                      rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
 444
 445                         if (rdev && test_bit(In_sync, &rdev->flags) &&
 446                                 r1_bio->bios[new_disk] != IO_BLOCKED)
 447                                 wonly_disk = new_disk;
 448
 449                         if (new_disk == conf->raid_disks - 1) {
 450                                 new_disk = wonly_disk;
 451                                 break;
 452                         }
 453                 }
 454                 goto rb_out;
 455         }
 456
 457
 458         /* make sure the disk is operational */
 459         for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 460              r1_bio->bios[new_disk] == IO_BLOCKED ||
 461              !rdev || !test_bit(In_sync, &rdev->flags) ||
 462                      test_bit(WriteMostly, &rdev->flags);
 463              rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
 464
 465                 if (rdev && test_bit(In_sync, &rdev->flags) &&
 466                     r1_bio->bios[new_disk] != IO_BLOCKED)
 467                         wonly_disk = new_disk;
 468
 469                 if (new_disk <= 0)
 470                         new_disk = conf->raid_disks;
 471                 new_disk--;
 472                 if (new_disk == disk) {
 473                         new_disk = wonly_disk;
 474                         break;
 475                 }
 476         }
 477
 478         if (new_disk < 0)
 479                 goto rb_out;
 480
 481         disk = new_disk;
 482         /* now disk == new_disk == starting point for search */
 483
 484         /*
 485          * Don't change to another disk for sequential reads:
 486          */
 487         if (conf->next_seq_sect == this_sector)
 488                 goto rb_out;
 489         if (this_sector == conf->mirrors[new_disk].head_position)
 490                 goto rb_out;
 491
 492         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 493
 494         /* Find the disk whose head is closest */
 495
 496         do {
 497                 if (disk <= 0)
 498                         disk = conf->raid_disks;
 499                 disk--;
 500
 501                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 502
 503                 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
 504                     !test_bit(In_sync, &rdev->flags) ||
 505                     test_bit(WriteMostly, &rdev->flags))
 506                         continue;
 507
 508                 if (!atomic_read(&rdev->nr_pending)) {
 509                         new_disk = disk;
 510                         break;
 511                 }
 512                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 513                 if (new_distance < current_distance) {
 514                         current_distance = new_distance;
 515                         new_disk = disk;
 516                 }
 517         } while (disk != conf->last_used);
 518
 519  rb_out:
 520
 521
 522         if (new_disk >= 0) {
 523                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 524                 if (!rdev)
 525                         goto retry;
 526                 atomic_inc(&rdev->nr_pending);
 527                 if (!test_bit(In_sync, &rdev->flags)) {
 528                         /* cannot risk returning a device that failed
 529                          * before we inc'ed nr_pending
 530                          */
 531                         rdev_dec_pending(rdev, conf->mddev);
 532                         goto retry;
 533                 }
 534                 conf->next_seq_sect = this_sector + sectors;
 535                 conf->last_used = new_disk;
 536         }
 537         rcu_read_unlock();
 538
 539         return new_disk;
 540 }
 541
 542 static void unplug_slaves(mddev_t *mddev)
 543 {
 544         conf_t *conf = mddev_to_conf(mddev);
 545         int i;
 546
 547         rcu_read_lock();
 548         for (i=0; i<mddev->raid_disks; i++) {
 549                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 550                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 551                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 552
 553                         atomic_inc(&rdev->nr_pending);
 554                         rcu_read_unlock();
 555
 556                         if (r_queue->unplug_fn)
 557                                 r_queue->unplug_fn(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(request_queue_t *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_issue_flush(request_queue_t *q, struct gendisk *disk,
 575                              sector_t *error_sector)
 576 {
 577         mddev_t *mddev = q->queuedata;
 578         conf_t *conf = mddev_to_conf(mddev);
 579         int i, ret = 0;
 580
 581         rcu_read_lock();
 582         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
 583                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 584                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 585                         struct block_device *bdev = rdev->bdev;
 586                         request_queue_t *r_queue = bdev_get_queue(bdev);
 587
 588                         if (!r_queue->issue_flush_fn)
 589                                 ret = -EOPNOTSUPP;
 590                         else {
 591                                 atomic_inc(&rdev->nr_pending);
 592                                 rcu_read_unlock();
 593                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
 594                                                               error_sector);
 595                                 rdev_dec_pending(rdev, mddev);
 596                                 rcu_read_lock();
 597                         }
 598                 }
 599         }
 600         rcu_read_unlock();
 601         return ret;
 602 }
 603
 604 /* Barriers....
 605  * Sometimes we need to suspend IO while we do something else,
 606  * either some resync/recovery, or reconfigure the array.
 607  * To do this we raise a 'barrier'.
 608  * The 'barrier' is a counter that can be raised multiple times
 609  * to count how many activities are happening which preclude
 610  * normal IO.
 611  * We can only raise the barrier if there is no pending IO.
 612  * i.e. if nr_pending == 0.
 613  * We choose only to raise the barrier if no-one is waiting for the
 614  * barrier to go down.  This means that as soon as an IO request
 615  * is ready, no other operations which require a barrier will start
 616  * until the IO request has had a chance.
 617  *
 618  * So: regular IO calls 'wait_barrier'.  When that returns there
 619  *    is no backgroup IO happening,  It must arrange to call
 620  *    allow_barrier when it has finished its IO.
 621  * backgroup IO calls must call raise_barrier.  Once that returns
 622  *    there is no normal IO happeing.  It must arrange to call
 623  *    lower_barrier when the particular background IO completes.
 624  */
 625 #define RESYNC_DEPTH 32
 626
 627 static void raise_barrier(conf_t *conf)
 628 {
 629         spin_lock_irq(&conf->resync_lock);
 630
 631         /* Wait until no block IO is waiting */
 632         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 633                             conf->resync_lock,
 634                             raid1_unplug(conf->mddev->queue));
 635
 636         /* block any new IO from starting */
 637         conf->barrier++;
 638
 639         /* No wait for all pending IO to complete */
 640         wait_event_lock_irq(conf->wait_barrier,
 641                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 642                             conf->resync_lock,
 643                             raid1_unplug(conf->mddev->queue));
 644
 645         spin_unlock_irq(&conf->resync_lock);
 646 }
 647
 648 static void lower_barrier(conf_t *conf)
 649 {
 650         unsigned long flags;
 651         spin_lock_irqsave(&conf->resync_lock, flags);
 652         conf->barrier--;
 653         spin_unlock_irqrestore(&conf->resync_lock, flags);
 654         wake_up(&conf->wait_barrier);
 655 }
 656
 657 static void wait_barrier(conf_t *conf)
 658 {
 659         spin_lock_irq(&conf->resync_lock);
 660         if (conf->barrier) {
 661                 conf->nr_waiting++;
 662                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 663                                     conf->resync_lock,
 664                                     raid1_unplug(conf->mddev->queue));
 665                 conf->nr_waiting--;
 666         }
 667         conf->nr_pending++;
 668         spin_unlock_irq(&conf->resync_lock);
 669 }
 670
 671 static void allow_barrier(conf_t *conf)
 672 {
 673         unsigned long flags;
 674         spin_lock_irqsave(&conf->resync_lock, flags);
 675         conf->nr_pending--;
 676         spin_unlock_irqrestore(&conf->resync_lock, flags);
 677         wake_up(&conf->wait_barrier);
 678 }
 679
 680 static void freeze_array(conf_t *conf)
 681 {
 682         /* stop syncio and normal IO and wait for everything to
 683          * go quite.
 684          * We increment barrier and nr_waiting, and then
 685          * wait until barrier+nr_pending match nr_queued+2
 686          */
 687         spin_lock_irq(&conf->resync_lock);
 688         conf->barrier++;
 689         conf->nr_waiting++;
 690         wait_event_lock_irq(conf->wait_barrier,
 691                             conf->barrier+conf->nr_pending == conf->nr_queued+2,
 692                             conf->resync_lock,
 693                             raid1_unplug(conf->mddev->queue));
 694         spin_unlock_irq(&conf->resync_lock);
 695 }
 696 static void unfreeze_array(conf_t *conf)
 697 {
 698         /* reverse the effect of the freeze */
 699         spin_lock_irq(&conf->resync_lock);
 700         conf->barrier--;
 701         conf->nr_waiting--;
 702         wake_up(&conf->wait_barrier);
 703         spin_unlock_irq(&conf->resync_lock);
 704 }
 705
 706
 707 /* duplicate the data pages for behind I/O */
 708 static struct page **alloc_behind_pages(struct bio *bio)
 709 {
 710         int i;
 711         struct bio_vec *bvec;
 712         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
 713                                         GFP_NOIO);
 714         if (unlikely(!pages))
 715                 goto do_sync_io;
 716
 717         bio_for_each_segment(bvec, bio, i) {
 718                 pages[i] = alloc_page(GFP_NOIO);
 719                 if (unlikely(!pages[i]))
 720                         goto do_sync_io;
 721                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 722                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 723                 kunmap(pages[i]);
 724                 kunmap(bvec->bv_page);
 725         }
 726
 727         return pages;
 728
 729 do_sync_io:
 730         if (pages)
 731                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
 732                         put_page(pages[i]);
 733         kfree(pages);
 734         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 735         return NULL;
 736 }
 737
 738 static int make_request(request_queue_t *q, struct bio * bio)
 739 {
 740         mddev_t *mddev = q->queuedata;
 741         conf_t *conf = mddev_to_conf(mddev);
 742         mirror_info_t *mirror;
 743         r1bio_t *r1_bio;
 744         struct bio *read_bio;
 745         int i, targets = 0, disks;
 746         mdk_rdev_t *rdev;
 747         struct bitmap *bitmap = mddev->bitmap;
 748         unsigned long flags;
 749         struct bio_list bl;
 750         struct page **behind_pages = NULL;
 751         const int rw = bio_data_dir(bio);
 752         int do_barriers;
 753
 754         /*
 755          * Register the new request and wait if the reconstruction
 756          * thread has put up a bar for new requests.
 757          * Continue immediately if no resync is active currently.
 758          * We test barriers_work *after* md_write_start as md_write_start
 759          * may cause the first superblock write, and that will check out
 760          * if barriers work.
 761          */
 762
 763         md_write_start(mddev, bio); /* wait on superblock update early */
 764
 765         if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
 766                 if (rw == WRITE)
 767                         md_write_end(mddev);
 768                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
 769                 return 0;
 770         }
 771
 772         wait_barrier(conf);
 773
 774         disk_stat_inc(mddev->gendisk, ios[rw]);
 775         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
 776
 777         /*
 778          * make_request() can abort the operation when READA is being
 779          * used and no empty request is available.
 780          *
 781          */
 782         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 783
 784         r1_bio->master_bio = bio;
 785         r1_bio->sectors = bio->bi_size >> 9;
 786         r1_bio->state = 0;
 787         r1_bio->mddev = mddev;
 788         r1_bio->sector = bio->bi_sector;
 789
 790         if (rw == READ) {
 791                 /*
 792                  * read balancing logic:
 793                  */
 794                 int rdisk = read_balance(conf, r1_bio);
 795
 796                 if (rdisk < 0) {
 797                         /* couldn't find anywhere to read from */
 798                         raid_end_bio_io(r1_bio);
 799                         return 0;
 800                 }
 801                 mirror = conf->mirrors + rdisk;
 802
 803                 r1_bio->read_disk = rdisk;
 804
 805                 read_bio = bio_clone(bio, GFP_NOIO);
 806
 807                 r1_bio->bios[rdisk] = read_bio;
 808
 809                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 810                 read_bio->bi_bdev = mirror->rdev->bdev;
 811                 read_bio->bi_end_io = raid1_end_read_request;
 812                 read_bio->bi_rw = READ;
 813                 read_bio->bi_private = r1_bio;
 814
 815                 generic_make_request(read_bio);
 816                 return 0;
 817         }
 818
 819         /*
 820          * WRITE:
 821          */
 822         /* first select target devices under spinlock and
 823          * inc refcount on their rdev.  Record them by setting
 824          * bios[x] to bio
 825          */
 826         disks = conf->raid_disks;
 827 #if 0
 828         { static int first=1;
 829         if (first) printk("First Write sector %llu disks %d\n",
 830                           (unsigned long long)r1_bio->sector, disks);
 831         first = 0;
 832         }
 833 #endif
 834         rcu_read_lock();
 835         for (i = 0;  i < disks; i++) {
 836                 if ((rdev=rcu_dereference(conf->mirrors[i].rdev)) != NULL &&
 837                     !test_bit(Faulty, &rdev->flags)) {
 838                         atomic_inc(&rdev->nr_pending);
 839                         if (test_bit(Faulty, &rdev->flags)) {
 840                                 rdev_dec_pending(rdev, mddev);
 841                                 r1_bio->bios[i] = NULL;
 842                         } else
 843                                 r1_bio->bios[i] = bio;
 844                         targets++;
 845                 } else
 846                         r1_bio->bios[i] = NULL;
 847         }
 848         rcu_read_unlock();
 849
 850         BUG_ON(targets == 0); /* we never fail the last device */
 851
 852         if (targets < conf->raid_disks) {
 853                 /* array is degraded, we will not clear the bitmap
 854                  * on I/O completion (see raid1_end_write_request) */
 855                 set_bit(R1BIO_Degraded, &r1_bio->state);
 856         }
 857
 858         /* do behind I/O ? */
 859         if (bitmap &&
 860             atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
 861             (behind_pages = alloc_behind_pages(bio)) != NULL)
 862                 set_bit(R1BIO_BehindIO, &r1_bio->state);
 863
 864         atomic_set(&r1_bio->remaining, 0);
 865         atomic_set(&r1_bio->behind_remaining, 0);
 866
 867         do_barriers = bio_barrier(bio);
 868         if (do_barriers)
 869                 set_bit(R1BIO_Barrier, &r1_bio->state);
 870
 871         bio_list_init(&bl);
 872         for (i = 0; i < disks; i++) {
 873                 struct bio *mbio;
 874                 if (!r1_bio->bios[i])
 875                         continue;
 876
 877                 mbio = bio_clone(bio, GFP_NOIO);
 878                 r1_bio->bios[i] = mbio;
 879
 880                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 881                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 882                 mbio->bi_end_io = raid1_end_write_request;
 883                 mbio->bi_rw = WRITE | do_barriers;
 884                 mbio->bi_private = r1_bio;
 885
 886                 if (behind_pages) {
 887                         struct bio_vec *bvec;
 888                         int j;
 889
 890                         /* Yes, I really want the '__' version so that
 891                          * we clear any unused pointer in the io_vec, rather
 892                          * than leave them unchanged.  This is important
 893                          * because when we come to free the pages, we won't
 894                          * know the originial bi_idx, so we just free
 895                          * them all
 896                          */
 897                         __bio_for_each_segment(bvec, mbio, j, 0)
 898                                 bvec->bv_page = behind_pages[j];
 899                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 900                                 atomic_inc(&r1_bio->behind_remaining);
 901                 }
 902
 903                 atomic_inc(&r1_bio->remaining);
 904
 905                 bio_list_add(&bl, mbio);
 906         }
 907         kfree(behind_pages); /* the behind pages are attached to the bios now */
 908
 909         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
 910                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 911         spin_lock_irqsave(&conf->device_lock, flags);
 912         bio_list_merge(&conf->pending_bio_list, &bl);
 913         bio_list_init(&bl);
 914
 915         blk_plug_device(mddev->queue);
 916         spin_unlock_irqrestore(&conf->device_lock, flags);
 917
 918 #if 0
 919         while ((bio = bio_list_pop(&bl)) != NULL)
 920                 generic_make_request(bio);
 921 #endif
 922
 923         return 0;
 924 }
 925
 926 static void status(struct seq_file *seq, mddev_t *mddev)
 927 {
 928         conf_t *conf = mddev_to_conf(mddev);
 929         int i;
 930
 931         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 932                    conf->raid_disks - mddev->degraded);
 933         rcu_read_lock();
 934         for (i = 0; i < conf->raid_disks; i++) {
 935                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 936                 seq_printf(seq, "%s",
 937                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
 938         }
 939         rcu_read_unlock();
 940         seq_printf(seq, "]");
 941 }
 942
 943
 944 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 945 {
 946         char b[BDEVNAME_SIZE];
 947         conf_t *conf = mddev_to_conf(mddev);
 948
 949         /*
 950          * If it is not operational, then we have already marked it as dead
 951          * else if it is the last working disks, ignore the error, let the
 952          * next level up know.
 953          * else mark the drive as failed
 954          */
 955         if (test_bit(In_sync, &rdev->flags)
 956             && (conf->raid_disks - mddev->degraded) == 1)
 957                 /*
 958                  * Don't fail the drive, act as though we were just a
 959                  * normal single drive
 960                  */
 961                 return;
 962         if (test_and_clear_bit(In_sync, &rdev->flags)) {
 963                 unsigned long flags;
 964                 spin_lock_irqsave(&conf->device_lock, flags);
 965                 mddev->degraded++;
 966                 spin_unlock_irqrestore(&conf->device_lock, flags);
 967                 /*
 968                  * if recovery is running, make sure it aborts.
 969                  */
 970                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 971         }
 972         set_bit(Faulty, &rdev->flags);
 973         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 974         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
 975                 "       Operation continuing on %d devices\n",
 976                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
 977 }
 978
 979 static void print_conf(conf_t *conf)
 980 {
 981         int i;
 982
 983         printk("RAID1 conf printout:\n");
 984         if (!conf) {
 985                 printk("(!conf)\n");
 986                 return;
 987         }
 988         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
 989                 conf->raid_disks);
 990
 991         rcu_read_lock();
 992         for (i = 0; i < conf->raid_disks; i++) {
 993                 char b[BDEVNAME_SIZE];
 994                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 995                 if (rdev)
 996                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
 997                                i, !test_bit(In_sync, &rdev->flags),
 998                                !test_bit(Faulty, &rdev->flags),
 999                                bdevname(rdev->bdev,b));
1000         }
1001         rcu_read_unlock();
1002 }
1003
1004 static void close_sync(conf_t *conf)
1005 {
1006         wait_barrier(conf);
1007         allow_barrier(conf);
1008
1009         mempool_destroy(conf->r1buf_pool);
1010         conf->r1buf_pool = NULL;
1011 }
1012
1013 static int raid1_spare_active(mddev_t *mddev)
1014 {
1015         int i;
1016         conf_t *conf = mddev->private;
1017
1018         /*
1019          * Find all failed disks within the RAID1 configuration
1020          * and mark them readable.
1021          * Called under mddev lock, so rcu protection not needed.
1022          */
1023         for (i = 0; i < conf->raid_disks; i++) {
1024                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1025                 if (rdev
1026                     && !test_bit(Faulty, &rdev->flags)
1027                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1028                         unsigned long flags;
1029                         spin_lock_irqsave(&conf->device_lock, flags);
1030                         mddev->degraded--;
1031                         spin_unlock_irqrestore(&conf->device_lock, flags);
1032                 }
1033         }
1034
1035         print_conf(conf);
1036         return 0;
1037 }
1038
1039
1040 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1041 {
1042         conf_t *conf = mddev->private;
1043         int found = 0;
1044         int mirror = 0;
1045         mirror_info_t *p;
1046
1047         for (mirror=0; mirror < mddev->raid_disks; mirror++)
1048                 if ( !(p=conf->mirrors+mirror)->rdev) {
1049
1050                         blk_queue_stack_limits(mddev->queue,
1051                                                rdev->bdev->bd_disk->queue);
1052                         /* as we don't honour merge_bvec_fn, we must never risk
1053                          * violating it, so limit ->max_sector to one PAGE, as
1054                          * a one page request is never in violation.
1055                          */
1056                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1057                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1058                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1059
1060                         p->head_position = 0;
1061                         rdev->raid_disk = mirror;
1062                         found = 1;
1063                         /* As all devices are equivalent, we don't need a full recovery
1064                          * if this was recently any drive of the array
1065                          */
1066                         if (rdev->saved_raid_disk < 0)
1067                                 conf->fullsync = 1;
1068                         rcu_assign_pointer(p->rdev, rdev);
1069                         break;
1070                 }
1071
1072         print_conf(conf);
1073         return found;
1074 }
1075
1076 static int raid1_remove_disk(mddev_t *mddev, int number)
1077 {
1078         conf_t *conf = mddev->private;
1079         int err = 0;
1080         mdk_rdev_t *rdev;
1081         mirror_info_t *p = conf->mirrors+ number;
1082
1083         print_conf(conf);
1084         rdev = p->rdev;
1085         if (rdev) {
1086                 if (test_bit(In_sync, &rdev->flags) ||
1087                     atomic_read(&rdev->nr_pending)) {
1088                         err = -EBUSY;
1089                         goto abort;
1090                 }
1091                 p->rdev = NULL;
1092                 synchronize_rcu();
1093                 if (atomic_read(&rdev->nr_pending)) {
1094                         /* lost the race, try later */
1095                         err = -EBUSY;
1096                         p->rdev = rdev;
1097                 }
1098         }
1099 abort:
1100
1101         print_conf(conf);
1102         return err;
1103 }
1104
1105
1106 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1107 {
1108         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1109         int i;
1110
1111         if (bio->bi_size)
1112                 return 1;
1113
1114         for (i=r1_bio->mddev->raid_disks; i--; )
1115                 if (r1_bio->bios[i] == bio)
1116                         break;
1117         BUG_ON(i < 0);
1118         update_head_pos(i, r1_bio);
1119         /*
1120          * we have read a block, now it needs to be re-written,
1121          * or re-read if the read failed.
1122          * We don't do much here, just schedule handling by raid1d
1123          */
1124         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1125                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1126
1127         if (atomic_dec_and_test(&r1_bio->remaining))
1128                 reschedule_retry(r1_bio);
1129         return 0;
1130 }
1131
1132 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1133 {
1134         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1135         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1136         mddev_t *mddev = r1_bio->mddev;
1137         conf_t *conf = mddev_to_conf(mddev);
1138         int i;
1139         int mirror=0;
1140
1141         if (bio->bi_size)
1142                 return 1;
1143
1144         for (i = 0; i < conf->raid_disks; i++)
1145                 if (r1_bio->bios[i] == bio) {
1146                         mirror = i;
1147                         break;
1148                 }
1149         if (!uptodate) {
1150                 int sync_blocks = 0;
1151                 sector_t s = r1_bio->sector;
1152                 long sectors_to_go = r1_bio->sectors;
1153                 /* make sure these bits doesn't get cleared. */
1154                 do {
1155                         bitmap_end_sync(mddev->bitmap, s,
1156                                         &sync_blocks, 1);
1157                         s += sync_blocks;
1158                         sectors_to_go -= sync_blocks;
1159                 } while (sectors_to_go > 0);
1160                 md_error(mddev, conf->mirrors[mirror].rdev);
1161         }
1162
1163         update_head_pos(mirror, r1_bio);
1164
1165         if (atomic_dec_and_test(&r1_bio->remaining)) {
1166                 md_done_sync(mddev, r1_bio->sectors, uptodate);
1167                 put_buf(r1_bio);
1168         }
1169         return 0;
1170 }
1171
1172 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1173 {
1174         conf_t *conf = mddev_to_conf(mddev);
1175         int i;
1176         int disks = conf->raid_disks;
1177         struct bio *bio, *wbio;
1178
1179         bio = r1_bio->bios[r1_bio->read_disk];
1180
1181
1182         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1183                 /* We have read all readable devices.  If we haven't
1184                  * got the block, then there is no hope left.
1185                  * If we have, then we want to do a comparison
1186                  * and skip the write if everything is the same.
1187                  * If any blocks failed to read, then we need to
1188                  * attempt an over-write
1189                  */
1190                 int primary;
1191                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1192                         for (i=0; i<mddev->raid_disks; i++)
1193                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1194                                         md_error(mddev, conf->mirrors[i].rdev);
1195
1196                         md_done_sync(mddev, r1_bio->sectors, 1);
1197                         put_buf(r1_bio);
1198                         return;
1199                 }
1200                 for (primary=0; primary<mddev->raid_disks; primary++)
1201                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1202                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1203                                 r1_bio->bios[primary]->bi_end_io = NULL;
1204                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1205                                 break;
1206                         }
1207                 r1_bio->read_disk = primary;
1208                 for (i=0; i<mddev->raid_disks; i++)
1209                         if (r1_bio->bios[i]->bi_end_io == end_sync_read &&
1210                             test_bit(BIO_UPTODATE, &r1_bio->bios[i]->bi_flags)) {
1211                                 int j;
1212                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1213                                 struct bio *pbio = r1_bio->bios[primary];
1214                                 struct bio *sbio = r1_bio->bios[i];
1215                                 for (j = vcnt; j-- ; )
1216                                         if (memcmp(page_address(pbio->bi_io_vec[j].bv_page),
1217                                                    page_address(sbio->bi_io_vec[j].bv_page),
1218                                                    PAGE_SIZE))
1219                                                 break;
1220                                 if (j >= 0)
1221                                         mddev->resync_mismatches += r1_bio->sectors;
1222                                 if (j < 0 || test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
1223                                         sbio->bi_end_io = NULL;
1224                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1225                                 } else {
1226                                         /* fixup the bio for reuse */
1227                                         sbio->bi_vcnt = vcnt;
1228                                         sbio->bi_size = r1_bio->sectors << 9;
1229                                         sbio->bi_idx = 0;
1230                                         sbio->bi_phys_segments = 0;
1231                                         sbio->bi_hw_segments = 0;
1232                                         sbio->bi_hw_front_size = 0;
1233                                         sbio->bi_hw_back_size = 0;
1234                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1235                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1236                                         sbio->bi_next = NULL;
1237                                         sbio->bi_sector = r1_bio->sector +
1238                                                 conf->mirrors[i].rdev->data_offset;
1239                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1240                                 }
1241                         }
1242         }
1243         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1244                 /* ouch - failed to read all of that.
1245                  * Try some synchronous reads of other devices to get
1246                  * good data, much like with normal read errors.  Only
1247                  * read into the pages we already have so we don't
1248                  * need to re-issue the read request.
1249                  * We don't need to freeze the array, because being in an
1250                  * active sync request, there is no normal IO, and
1251                  * no overlapping syncs.
1252                  */
1253                 sector_t sect = r1_bio->sector;
1254                 int sectors = r1_bio->sectors;
1255                 int idx = 0;
1256
1257                 while(sectors) {
1258                         int s = sectors;
1259                         int d = r1_bio->read_disk;
1260                         int success = 0;
1261                         mdk_rdev_t *rdev;
1262
1263                         if (s > (PAGE_SIZE>>9))
1264                                 s = PAGE_SIZE >> 9;
1265                         do {
1266                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1267                                         /* No rcu protection needed here devices
1268                                          * can only be removed when no resync is
1269                                          * active, and resync is currently active
1270                                          */
1271                                         rdev = conf->mirrors[d].rdev;
1272                                         if (sync_page_io(rdev->bdev,
1273                                                          sect + rdev->data_offset,
1274                                                          s<<9,
1275                                                          bio->bi_io_vec[idx].bv_page,
1276                                                          READ)) {
1277                                                 success = 1;
1278                                                 break;
1279                                         }
1280                                 }
1281                                 d++;
1282                                 if (d == conf->raid_disks)
1283                                         d = 0;
1284                         } while (!success && d != r1_bio->read_disk);
1285
1286                         if (success) {
1287                                 int start = d;
1288                                 /* write it back and re-read */
1289                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1290                                 while (d != r1_bio->read_disk) {
1291                                         if (d == 0)
1292                                                 d = conf->raid_disks;
1293                                         d--;
1294                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1295                                                 continue;
1296                                         rdev = conf->mirrors[d].rdev;
1297                                         atomic_add(s, &rdev->corrected_errors);
1298                                         if (sync_page_io(rdev->bdev,
1299                                                          sect + rdev->data_offset,
1300                                                          s<<9,
1301                                                          bio->bi_io_vec[idx].bv_page,
1302                                                          WRITE) == 0)
1303                                                 md_error(mddev, rdev);
1304                                 }
1305                                 d = start;
1306                                 while (d != r1_bio->read_disk) {
1307                                         if (d == 0)
1308                                                 d = conf->raid_disks;
1309                                         d--;
1310                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1311                                                 continue;
1312                                         rdev = conf->mirrors[d].rdev;
1313                                         if (sync_page_io(rdev->bdev,
1314                                                          sect + rdev->data_offset,
1315                                                          s<<9,
1316                                                          bio->bi_io_vec[idx].bv_page,
1317                                                          READ) == 0)
1318                                                 md_error(mddev, rdev);
1319                                 }
1320                         } else {
1321                                 char b[BDEVNAME_SIZE];
1322                                 /* Cannot read from anywhere, array is toast */
1323                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1324                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1325                                        " for block %llu\n",
1326                                        bdevname(bio->bi_bdev,b),
1327                                        (unsigned long long)r1_bio->sector);
1328                                 md_done_sync(mddev, r1_bio->sectors, 0);
1329                                 put_buf(r1_bio);
1330                                 return;
1331                         }
1332                         sectors -= s;
1333                         sect += s;
1334                         idx ++;
1335                 }
1336         }
1337
1338         /*
1339          * schedule writes
1340          */
1341         atomic_set(&r1_bio->remaining, 1);
1342         for (i = 0; i < disks ; i++) {
1343                 wbio = r1_bio->bios[i];
1344                 if (wbio->bi_end_io == NULL ||
1345                     (wbio->bi_end_io == end_sync_read &&
1346                      (i == r1_bio->read_disk ||
1347                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1348                         continue;
1349
1350                 wbio->bi_rw = WRITE;
1351                 wbio->bi_end_io = end_sync_write;
1352                 atomic_inc(&r1_bio->remaining);
1353                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1354
1355                 generic_make_request(wbio);
1356         }
1357
1358         if (atomic_dec_and_test(&r1_bio->remaining)) {
1359                 /* if we're here, all write(s) have completed, so clean up */
1360                 md_done_sync(mddev, r1_bio->sectors, 1);
1361                 put_buf(r1_bio);
1362         }
1363 }
1364
1365 /*
1366  * This is a kernel thread which:
1367  *
1368  *      1.      Retries failed read operations on working mirrors.
1369  *      2.      Updates the raid superblock when problems encounter.
1370  *      3.      Performs writes following reads for array syncronising.
1371  */
1372
1373 static void fix_read_error(conf_t *conf, int read_disk,
1374                            sector_t sect, int sectors)
1375 {
1376         mddev_t *mddev = conf->mddev;
1377         while(sectors) {
1378                 int s = sectors;
1379                 int d = read_disk;
1380                 int success = 0;
1381                 int start;
1382                 mdk_rdev_t *rdev;
1383
1384                 if (s > (PAGE_SIZE>>9))
1385                         s = PAGE_SIZE >> 9;
1386
1387                 do {
1388                         /* Note: no rcu protection needed here
1389                          * as this is synchronous in the raid1d thread
1390                          * which is the thread that might remove
1391                          * a device.  If raid1d ever becomes multi-threaded....
1392                          */
1393                         rdev = conf->mirrors[d].rdev;
1394                         if (rdev &&
1395                             test_bit(In_sync, &rdev->flags) &&
1396                             sync_page_io(rdev->bdev,
1397                                          sect + rdev->data_offset,
1398                                          s<<9,
1399                                          conf->tmppage, READ))
1400                                 success = 1;
1401                         else {
1402                                 d++;
1403                                 if (d == conf->raid_disks)
1404                                         d = 0;
1405                         }
1406                 } while (!success && d != read_disk);
1407
1408                 if (!success) {
1409                         /* Cannot read from anywhere -- bye bye array */
1410                         md_error(mddev, conf->mirrors[read_disk].rdev);
1411                         break;
1412                 }
1413                 /* write it back and re-read */
1414                 start = d;
1415                 while (d != read_disk) {
1416                         if (d==0)
1417                                 d = conf->raid_disks;
1418                         d--;
1419                         rdev = conf->mirrors[d].rdev;
1420                         if (rdev &&
1421                             test_bit(In_sync, &rdev->flags)) {
1422                                 if (sync_page_io(rdev->bdev,
1423                                                  sect + rdev->data_offset,
1424                                                  s<<9, conf->tmppage, WRITE)
1425                                     == 0)
1426                                         /* Well, this device is dead */
1427                                         md_error(mddev, rdev);
1428                         }
1429                 }
1430                 d = start;
1431                 while (d != read_disk) {
1432                         char b[BDEVNAME_SIZE];
1433                         if (d==0)
1434                                 d = conf->raid_disks;
1435                         d--;
1436                         rdev = conf->mirrors[d].rdev;
1437                         if (rdev &&
1438                             test_bit(In_sync, &rdev->flags)) {
1439                                 if (sync_page_io(rdev->bdev,
1440                                                  sect + rdev->data_offset,
1441                                                  s<<9, conf->tmppage, READ)
1442                                     == 0)
1443                                         /* Well, this device is dead */
1444                                         md_error(mddev, rdev);
1445                                 else {
1446                                         atomic_add(s, &rdev->corrected_errors);
1447                                         printk(KERN_INFO
1448                                                "raid1:%s: read error corrected "
1449                                                "(%d sectors at %llu on %s)\n",
1450                                                mdname(mddev), s,
1451                                                (unsigned long long)sect +
1452                                                    rdev->data_offset,
1453                                                bdevname(rdev->bdev, b));
1454                                 }
1455                         }
1456                 }
1457                 sectors -= s;
1458                 sect += s;
1459         }
1460 }
1461
1462 static void raid1d(mddev_t *mddev)
1463 {
1464         r1bio_t *r1_bio;
1465         struct bio *bio;
1466         unsigned long flags;
1467         conf_t *conf = mddev_to_conf(mddev);
1468         struct list_head *head = &conf->retry_list;
1469         int unplug=0;
1470         mdk_rdev_t *rdev;
1471
1472         md_check_recovery(mddev);
1473
1474         for (;;) {
1475                 char b[BDEVNAME_SIZE];
1476                 spin_lock_irqsave(&conf->device_lock, flags);
1477
1478                 if (conf->pending_bio_list.head) {
1479                         bio = bio_list_get(&conf->pending_bio_list);
1480                         blk_remove_plug(mddev->queue);
1481                         spin_unlock_irqrestore(&conf->device_lock, flags);
1482                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1483                         if (bitmap_unplug(mddev->bitmap) != 0)
1484                                 printk("%s: bitmap file write failed!\n", mdname(mddev));
1485
1486                         while (bio) { /* submit pending writes */
1487                                 struct bio *next = bio->bi_next;
1488                                 bio->bi_next = NULL;
1489                                 generic_make_request(bio);
1490                                 bio = next;
1491                         }
1492                         unplug = 1;
1493
1494                         continue;
1495                 }
1496
1497                 if (list_empty(head))
1498                         break;
1499                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1500                 list_del(head->prev);
1501                 conf->nr_queued--;
1502                 spin_unlock_irqrestore(&conf->device_lock, flags);
1503
1504                 mddev = r1_bio->mddev;
1505                 conf = mddev_to_conf(mddev);
1506                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1507                         sync_request_write(mddev, r1_bio);
1508                         unplug = 1;
1509                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1510                         /* some requests in the r1bio were BIO_RW_BARRIER
1511                          * requests which failed with -EOPNOTSUPP.  Hohumm..
1512                          * Better resubmit without the barrier.
1513                          * We know which devices to resubmit for, because
1514                          * all others have had their bios[] entry cleared.
1515                          * We already have a nr_pending reference on these rdevs.
1516                          */
1517                         int i;
1518                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1519                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1520                         for (i=0; i < conf->raid_disks; i++)
1521                                 if (r1_bio->bios[i])
1522                                         atomic_inc(&r1_bio->remaining);
1523                         for (i=0; i < conf->raid_disks; i++)
1524                                 if (r1_bio->bios[i]) {
1525                                         struct bio_vec *bvec;
1526                                         int j;
1527
1528                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1529                                         /* copy pages from the failed bio, as
1530                                          * this might be a write-behind device */
1531                                         __bio_for_each_segment(bvec, bio, j, 0)
1532                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1533                                         bio_put(r1_bio->bios[i]);
1534                                         bio->bi_sector = r1_bio->sector +
1535                                                 conf->mirrors[i].rdev->data_offset;
1536                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1537                                         bio->bi_end_io = raid1_end_write_request;
1538                                         bio->bi_rw = WRITE;
1539                                         bio->bi_private = r1_bio;
1540                                         r1_bio->bios[i] = bio;
1541                                         generic_make_request(bio);
1542                                 }
1543                 } else {
1544                         int disk;
1545
1546                         /* we got a read error. Maybe the drive is bad.  Maybe just
1547                          * the block and we can fix it.
1548                          * We freeze all other IO, and try reading the block from
1549                          * other devices.  When we find one, we re-write
1550                          * and check it that fixes the read error.
1551                          * This is all done synchronously while the array is
1552                          * frozen
1553                          */
1554                         if (mddev->ro == 0) {
1555                                 freeze_array(conf);
1556                                 fix_read_error(conf, r1_bio->read_disk,
1557                                                r1_bio->sector,
1558                                                r1_bio->sectors);
1559                                 unfreeze_array(conf);
1560                         }
1561
1562                         bio = r1_bio->bios[r1_bio->read_disk];
1563                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1564                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1565                                        " read error for block %llu\n",
1566                                        bdevname(bio->bi_bdev,b),
1567                                        (unsigned long long)r1_bio->sector);
1568                                 raid_end_bio_io(r1_bio);
1569                         } else {
1570                                 r1_bio->bios[r1_bio->read_disk] =
1571                                         mddev->ro ? IO_BLOCKED : NULL;
1572                                 r1_bio->read_disk = disk;
1573                                 bio_put(bio);
1574                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1575                                 r1_bio->bios[r1_bio->read_disk] = bio;
1576                                 rdev = conf->mirrors[disk].rdev;
1577                                 if (printk_ratelimit())
1578                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1579                                                " another mirror\n",
1580                                                bdevname(rdev->bdev,b),
1581                                                (unsigned long long)r1_bio->sector);
1582                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1583                                 bio->bi_bdev = rdev->bdev;
1584                                 bio->bi_end_io = raid1_end_read_request;
1585                                 bio->bi_rw = READ;
1586                                 bio->bi_private = r1_bio;
1587                                 unplug = 1;
1588                                 generic_make_request(bio);
1589                         }
1590                 }
1591         }
1592         spin_unlock_irqrestore(&conf->device_lock, flags);
1593         if (unplug)
1594                 unplug_slaves(mddev);
1595 }
1596
1597
1598 static int init_resync(conf_t *conf)
1599 {
1600         int buffs;
1601
1602         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1603         BUG_ON(conf->r1buf_pool);
1604         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1605                                           conf->poolinfo);
1606         if (!conf->r1buf_pool)
1607                 return -ENOMEM;
1608         conf->next_resync = 0;
1609         return 0;
1610 }
1611
1612 /*
1613  * perform a "sync" on one "block"
1614  *
1615  * We need to make sure that no normal I/O request - particularly write
1616  * requests - conflict with active sync requests.
1617  *
1618  * This is achieved by tracking pending requests and a 'barrier' concept
1619  * that can be installed to exclude normal IO requests.
1620  */
1621
1622 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1623 {
1624         conf_t *conf = mddev_to_conf(mddev);
1625         r1bio_t *r1_bio;
1626         struct bio *bio;
1627         sector_t max_sector, nr_sectors;
1628         int disk = -1;
1629         int i;
1630         int wonly = -1;
1631         int write_targets = 0, read_targets = 0;
1632         int sync_blocks;
1633         int still_degraded = 0;
1634
1635         if (!conf->r1buf_pool)
1636         {
1637 /*
1638                 printk("sync start - bitmap %p\n", mddev->bitmap);
1639 */
1640                 if (init_resync(conf))
1641                         return 0;
1642         }
1643
1644         max_sector = mddev->size << 1;
1645         if (sector_nr >= max_sector) {
1646                 /* If we aborted, we need to abort the
1647                  * sync on the 'current' bitmap chunk (there will
1648                  * only be one in raid1 resync.
1649                  * We can find the current addess in mddev->curr_resync
1650                  */
1651                 if (mddev->curr_resync < max_sector) /* aborted */
1652                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1653                                                 &sync_blocks, 1);
1654                 else /* completed sync */
1655                         conf->fullsync = 0;
1656
1657                 bitmap_close_sync(mddev->bitmap);
1658                 close_sync(conf);
1659                 return 0;
1660         }
1661
1662         if (mddev->bitmap == NULL &&
1663             mddev->recovery_cp == MaxSector &&
1664             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1665             conf->fullsync == 0) {
1666                 *skipped = 1;
1667                 return max_sector - sector_nr;
1668         }
1669         /* before building a request, check if we can skip these blocks..
1670          * This call the bitmap_start_sync doesn't actually record anything
1671          */
1672         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1673             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1674                 /* We can skip this block, and probably several more */
1675                 *skipped = 1;
1676                 return sync_blocks;
1677         }
1678         /*
1679          * If there is non-resync activity waiting for a turn,
1680          * and resync is going fast enough,
1681          * then let it though before starting on this new sync request.
1682          */
1683         if (!go_faster && conf->nr_waiting)
1684                 msleep_interruptible(1000);
1685
1686         raise_barrier(conf);
1687
1688         conf->next_resync = sector_nr;
1689
1690         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1691         rcu_read_lock();
1692         /*
1693          * If we get a correctably read error during resync or recovery,
1694          * we might want to read from a different device.  So we
1695          * flag all drives that could conceivably be read from for READ,
1696          * and any others (which will be non-In_sync devices) for WRITE.
1697          * If a read fails, we try reading from something else for which READ
1698          * is OK.
1699          */
1700
1701         r1_bio->mddev = mddev;
1702         r1_bio->sector = sector_nr;
1703         r1_bio->state = 0;
1704         set_bit(R1BIO_IsSync, &r1_bio->state);
1705
1706         for (i=0; i < conf->raid_disks; i++) {
1707                 mdk_rdev_t *rdev;
1708                 bio = r1_bio->bios[i];
1709
1710                 /* take from bio_init */
1711                 bio->bi_next = NULL;
1712                 bio->bi_flags |= 1 << BIO_UPTODATE;
1713                 bio->bi_rw = 0;
1714                 bio->bi_vcnt = 0;
1715                 bio->bi_idx = 0;
1716                 bio->bi_phys_segments = 0;
1717                 bio->bi_hw_segments = 0;
1718                 bio->bi_size = 0;
1719                 bio->bi_end_io = NULL;
1720                 bio->bi_private = NULL;
1721
1722                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1723                 if (rdev == NULL ||
1724                            test_bit(Faulty, &rdev->flags)) {
1725                         still_degraded = 1;
1726                         continue;
1727                 } else if (!test_bit(In_sync, &rdev->flags)) {
1728                         bio->bi_rw = WRITE;
1729                         bio->bi_end_io = end_sync_write;
1730                         write_targets ++;
1731                 } else {
1732                         /* may need to read from here */
1733                         bio->bi_rw = READ;
1734                         bio->bi_end_io = end_sync_read;
1735                         if (test_bit(WriteMostly, &rdev->flags)) {
1736                                 if (wonly < 0)
1737                                         wonly = i;
1738                         } else {
1739                                 if (disk < 0)
1740                                         disk = i;
1741                         }
1742                         read_targets++;
1743                 }
1744                 atomic_inc(&rdev->nr_pending);
1745                 bio->bi_sector = sector_nr + rdev->data_offset;
1746                 bio->bi_bdev = rdev->bdev;
1747                 bio->bi_private = r1_bio;
1748         }
1749         rcu_read_unlock();
1750         if (disk < 0)
1751                 disk = wonly;
1752         r1_bio->read_disk = disk;
1753
1754         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1755                 /* extra read targets are also write targets */
1756                 write_targets += read_targets-1;
1757
1758         if (write_targets == 0 || read_targets == 0) {
1759                 /* There is nowhere to write, so all non-sync
1760                  * drives must be failed - so we are finished
1761                  */
1762                 sector_t rv = max_sector - sector_nr;
1763                 *skipped = 1;
1764                 put_buf(r1_bio);
1765                 return rv;
1766         }
1767
1768         nr_sectors = 0;
1769         sync_blocks = 0;
1770         do {
1771                 struct page *page;
1772                 int len = PAGE_SIZE;
1773                 if (sector_nr + (len>>9) > max_sector)
1774                         len = (max_sector - sector_nr) << 9;
1775                 if (len == 0)
1776                         break;
1777                 if (sync_blocks == 0) {
1778                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1779                                                &sync_blocks, still_degraded) &&
1780                             !conf->fullsync &&
1781                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1782                                 break;
1783                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1784                         if (len > (sync_blocks<<9))
1785                                 len = sync_blocks<<9;
1786                 }
1787
1788                 for (i=0 ; i < conf->raid_disks; i++) {
1789                         bio = r1_bio->bios[i];
1790                         if (bio->bi_end_io) {
1791                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1792                                 if (bio_add_page(bio, page, len, 0) == 0) {
1793                                         /* stop here */
1794                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1795                                         while (i > 0) {
1796                                                 i--;
1797                                                 bio = r1_bio->bios[i];
1798                                                 if (bio->bi_end_io==NULL)
1799                                                         continue;
1800                                                 /* remove last page from this bio */
1801                                                 bio->bi_vcnt--;
1802                                                 bio->bi_size -= len;
1803                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1804                                         }
1805                                         goto bio_full;
1806                                 }
1807                         }
1808                 }
1809                 nr_sectors += len>>9;
1810                 sector_nr += len>>9;
1811                 sync_blocks -= (len>>9);
1812         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1813  bio_full:
1814         r1_bio->sectors = nr_sectors;
1815
1816         /* For a user-requested sync, we read all readable devices and do a
1817          * compare
1818          */
1819         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1820                 atomic_set(&r1_bio->remaining, read_targets);
1821                 for (i=0; i<conf->raid_disks; i++) {
1822                         bio = r1_bio->bios[i];
1823                         if (bio->bi_end_io == end_sync_read) {
1824                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1825                                 generic_make_request(bio);
1826                         }
1827                 }
1828         } else {
1829                 atomic_set(&r1_bio->remaining, 1);
1830                 bio = r1_bio->bios[r1_bio->read_disk];
1831                 md_sync_acct(bio->bi_bdev, nr_sectors);
1832                 generic_make_request(bio);
1833
1834         }
1835         return nr_sectors;
1836 }
1837
1838 static int run(mddev_t *mddev)
1839 {
1840         conf_t *conf;
1841         int i, j, disk_idx;
1842         mirror_info_t *disk;
1843         mdk_rdev_t *rdev;
1844         struct list_head *tmp;
1845
1846         if (mddev->level != 1) {
1847                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1848                        mdname(mddev), mddev->level);
1849                 goto out;
1850         }
1851         if (mddev->reshape_position != MaxSector) {
1852                 printk("raid1: %s: reshape_position set but not supported\n",
1853                        mdname(mddev));
1854                 goto out;
1855         }
1856         /*
1857          * copy the already verified devices into our private RAID1
1858          * bookkeeping area. [whatever we allocate in run(),
1859          * should be freed in stop()]
1860          */
1861         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1862         mddev->private = conf;
1863         if (!conf)
1864                 goto out_no_mem;
1865
1866         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1867                                  GFP_KERNEL);
1868         if (!conf->mirrors)
1869                 goto out_no_mem;
1870
1871         conf->tmppage = alloc_page(GFP_KERNEL);
1872         if (!conf->tmppage)
1873                 goto out_no_mem;
1874
1875         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1876         if (!conf->poolinfo)
1877                 goto out_no_mem;
1878         conf->poolinfo->mddev = mddev;
1879         conf->poolinfo->raid_disks = mddev->raid_disks;
1880         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1881                                           r1bio_pool_free,
1882                                           conf->poolinfo);
1883         if (!conf->r1bio_pool)
1884                 goto out_no_mem;
1885
1886         ITERATE_RDEV(mddev, rdev, tmp) {
1887                 disk_idx = rdev->raid_disk;
1888                 if (disk_idx >= mddev->raid_disks
1889                     || disk_idx < 0)
1890                         continue;
1891                 disk = conf->mirrors + disk_idx;
1892
1893                 disk->rdev = rdev;
1894
1895                 blk_queue_stack_limits(mddev->queue,
1896                                        rdev->bdev->bd_disk->queue);
1897                 /* as we don't honour merge_bvec_fn, we must never risk
1898                  * violating it, so limit ->max_sector to one PAGE, as
1899                  * a one page request is never in violation.
1900                  */
1901                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1902                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1903                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1904
1905                 disk->head_position = 0;
1906         }
1907         conf->raid_disks = mddev->raid_disks;
1908         conf->mddev = mddev;
1909         spin_lock_init(&conf->device_lock);
1910         INIT_LIST_HEAD(&conf->retry_list);
1911
1912         spin_lock_init(&conf->resync_lock);
1913         init_waitqueue_head(&conf->wait_barrier);
1914
1915         bio_list_init(&conf->pending_bio_list);
1916         bio_list_init(&conf->flushing_bio_list);
1917
1918
1919         mddev->degraded = 0;
1920         for (i = 0; i < conf->raid_disks; i++) {
1921
1922                 disk = conf->mirrors + i;
1923
1924                 if (!disk->rdev ||
1925                     !test_bit(In_sync, &disk->rdev->flags)) {
1926                         disk->head_position = 0;
1927                         mddev->degraded++;
1928                 }
1929         }
1930         if (mddev->degraded == conf->raid_disks) {
1931                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1932                         mdname(mddev));
1933                 goto out_free_conf;
1934         }
1935         if (conf->raid_disks - mddev->degraded == 1)
1936                 mddev->recovery_cp = MaxSector;
1937
1938         /*
1939          * find the first working one and use it as a starting point
1940          * to read balancing.
1941          */
1942         for (j = 0; j < conf->raid_disks &&
1943                      (!conf->mirrors[j].rdev ||
1944                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
1945                 /* nothing */;
1946         conf->last_used = j;
1947
1948
1949         mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1950         if (!mddev->thread) {
1951                 printk(KERN_ERR
1952                        "raid1: couldn't allocate thread for %s\n",
1953                        mdname(mddev));
1954                 goto out_free_conf;
1955         }
1956
1957         printk(KERN_INFO
1958                 "raid1: raid set %s active with %d out of %d mirrors\n",
1959                 mdname(mddev), mddev->raid_disks - mddev->degraded,
1960                 mddev->raid_disks);
1961         /*
1962          * Ok, everything is just fine now
1963          */
1964         mddev->array_size = mddev->size;
1965
1966         mddev->queue->unplug_fn = raid1_unplug;
1967         mddev->queue->issue_flush_fn = raid1_issue_flush;
1968
1969         return 0;
1970
1971 out_no_mem:
1972         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1973                mdname(mddev));
1974
1975 out_free_conf:
1976         if (conf) {
1977                 if (conf->r1bio_pool)
1978                         mempool_destroy(conf->r1bio_pool);
1979                 kfree(conf->mirrors);
1980                 safe_put_page(conf->tmppage);
1981                 kfree(conf->poolinfo);
1982                 kfree(conf);
1983                 mddev->private = NULL;
1984         }
1985 out:
1986         return -EIO;
1987 }
1988
1989 static int stop(mddev_t *mddev)
1990 {
1991         conf_t *conf = mddev_to_conf(mddev);
1992         struct bitmap *bitmap = mddev->bitmap;
1993         int behind_wait = 0;
1994
1995         /* wait for behind writes to complete */
1996         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
1997                 behind_wait++;
1998                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
1999                 set_current_state(TASK_UNINTERRUPTIBLE);
2000                 schedule_timeout(HZ); /* wait a second */
2001                 /* need to kick something here to make sure I/O goes? */
2002         }
2003
2004         md_unregister_thread(mddev->thread);
2005         mddev->thread = NULL;
2006         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2007         if (conf->r1bio_pool)
2008                 mempool_destroy(conf->r1bio_pool);
2009         kfree(conf->mirrors);
2010         kfree(conf->poolinfo);
2011         kfree(conf);
2012         mddev->private = NULL;
2013         return 0;
2014 }
2015
2016 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2017 {
2018         /* no resync is happening, and there is enough space
2019          * on all devices, so we can resize.
2020          * We need to make sure resync covers any new space.
2021          * If the array is shrinking we should possibly wait until
2022          * any io in the removed space completes, but it hardly seems
2023          * worth it.
2024          */
2025         mddev->array_size = sectors>>1;
2026         set_capacity(mddev->gendisk, mddev->array_size << 1);
2027         mddev->changed = 1;
2028         if (mddev->array_size > mddev->size && mddev->recovery_cp == MaxSector) {
2029                 mddev->recovery_cp = mddev->size << 1;
2030                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2031         }
2032         mddev->size = mddev->array_size;
2033         mddev->resync_max_sectors = sectors;
2034         return 0;
2035 }
2036
2037 static int raid1_reshape(mddev_t *mddev)
2038 {
2039         /* We need to:
2040          * 1/ resize the r1bio_pool
2041          * 2/ resize conf->mirrors
2042          *
2043          * We allocate a new r1bio_pool if we can.
2044          * Then raise a device barrier and wait until all IO stops.
2045          * Then resize conf->mirrors and swap in the new r1bio pool.
2046          *
2047          * At the same time, we "pack" the devices so that all the missing
2048          * devices have the higher raid_disk numbers.
2049          */
2050         mempool_t *newpool, *oldpool;
2051         struct pool_info *newpoolinfo;
2052         mirror_info_t *newmirrors;
2053         conf_t *conf = mddev_to_conf(mddev);
2054         int cnt, raid_disks;
2055         unsigned long flags;
2056         int d, d2;
2057
2058         /* Cannot change chunk_size, layout, or level */
2059         if (mddev->chunk_size != mddev->new_chunk ||
2060             mddev->layout != mddev->new_layout ||
2061             mddev->level != mddev->new_level) {
2062                 mddev->new_chunk = mddev->chunk_size;
2063                 mddev->new_layout = mddev->layout;
2064                 mddev->new_level = mddev->level;
2065                 return -EINVAL;
2066         }
2067
2068         raid_disks = mddev->raid_disks + mddev->delta_disks;
2069
2070         if (raid_disks < conf->raid_disks) {
2071                 cnt=0;
2072                 for (d= 0; d < conf->raid_disks; d++)
2073                         if (conf->mirrors[d].rdev)
2074                                 cnt++;
2075                 if (cnt > raid_disks)
2076                         return -EBUSY;
2077         }
2078
2079         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2080         if (!newpoolinfo)
2081                 return -ENOMEM;
2082         newpoolinfo->mddev = mddev;
2083         newpoolinfo->raid_disks = raid_disks;
2084
2085         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2086                                  r1bio_pool_free, newpoolinfo);
2087         if (!newpool) {
2088                 kfree(newpoolinfo);
2089                 return -ENOMEM;
2090         }
2091         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2092         if (!newmirrors) {
2093                 kfree(newpoolinfo);
2094                 mempool_destroy(newpool);
2095                 return -ENOMEM;
2096         }
2097
2098         raise_barrier(conf);
2099
2100         /* ok, everything is stopped */
2101         oldpool = conf->r1bio_pool;
2102         conf->r1bio_pool = newpool;
2103
2104         for (d=d2=0; d < conf->raid_disks; d++)
2105                 if (conf->mirrors[d].rdev) {
2106                         conf->mirrors[d].rdev->raid_disk = d2;
2107                         newmirrors[d2++].rdev = conf->mirrors[d].rdev;
2108                 }
2109         kfree(conf->mirrors);
2110         conf->mirrors = newmirrors;
2111         kfree(conf->poolinfo);
2112         conf->poolinfo = newpoolinfo;
2113
2114         spin_lock_irqsave(&conf->device_lock, flags);
2115         mddev->degraded += (raid_disks - conf->raid_disks);
2116         spin_unlock_irqrestore(&conf->device_lock, flags);
2117         conf->raid_disks = mddev->raid_disks = raid_disks;
2118         mddev->delta_disks = 0;
2119
2120         conf->last_used = 0; /* just make sure it is in-range */
2121         lower_barrier(conf);
2122
2123         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2124         md_wakeup_thread(mddev->thread);
2125
2126         mempool_destroy(oldpool);
2127         return 0;
2128 }
2129
2130 static void raid1_quiesce(mddev_t *mddev, int state)
2131 {
2132         conf_t *conf = mddev_to_conf(mddev);
2133
2134         switch(state) {
2135         case 1:
2136                 raise_barrier(conf);
2137                 break;
2138         case 0:
2139                 lower_barrier(conf);
2140                 break;
2141         }
2142 }
2143
2144
2145 static struct mdk_personality raid1_personality =
2146 {
2147         .name           = "raid1",
2148         .level          = 1,
2149         .owner          = THIS_MODULE,
2150         .make_request   = make_request,
2151         .run            = run,
2152         .stop           = stop,
2153         .status         = status,
2154         .error_handler  = error,
2155         .hot_add_disk   = raid1_add_disk,
2156         .hot_remove_disk= raid1_remove_disk,
2157         .spare_active   = raid1_spare_active,
2158         .sync_request   = sync_request,
2159         .resize         = raid1_resize,
2160         .check_reshape  = raid1_reshape,
2161         .quiesce        = raid1_quiesce,
2162 };
2163
2164 static int __init raid_init(void)
2165 {
2166         return register_md_personality(&raid1_personality);
2167 }
2168
2169 static void raid_exit(void)
2170 {
2171         unregister_md_personality(&raid1_personality);
2172 }
2173
2174 module_init(raid_init);
2175 module_exit(raid_exit);
2176 MODULE_LICENSE("GPL");
2177 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2178 MODULE_ALIAS("md-raid1");
2179 MODULE_ALIAS("md-level-1");