drivers/md/raid1.c

   1 /*
   2  * raid1.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5  *
   6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7  *
   8  * RAID-1 management functions.
   9  *
  10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11  *
  12  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14  *
  15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  16  * bitmapped intelligence in resync:
  17  *
  18  *      - bitmap marked during normal i/o
  19  *      - bitmap used to skip nondirty blocks during sync
  20  *
  21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  22  * - persistent bitmap code
  23  *
  24  * This program is free software; you can redistribute it and/or modify
  25  * it under the terms of the GNU General Public License as published by
  26  * the Free Software Foundation; either version 2, or (at your option)
  27  * any later version.
  28  *
  29  * You should have received a copy of the GNU General Public License
  30  * (for example /usr/src/linux/COPYING); if not, write to the Free
  31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  32  */
  33
  34 #include <linux/delay.h>
  35 #include <linux/blkdev.h>
  36 #include <linux/seq_file.h>
  37 #include "md.h"
  38 #include "raid1.h"
  39 #include "bitmap.h"
  40
  41 #define DEBUG 0
  42 #if DEBUG
  43 #define PRINTK(x...) printk(x)
  44 #else
  45 #define PRINTK(x...)
  46 #endif
  47
  48 /*
  49  * Number of guaranteed r1bios in case of extreme VM load:
  50  */
  51 #define NR_RAID1_BIOS 256
  52
  53
  54 static void unplug_slaves(mddev_t *mddev);
  55
  56 static void allow_barrier(conf_t *conf);
  57 static void lower_barrier(conf_t *conf);
  58
  59 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  60 {
  61         struct pool_info *pi = data;
  62         r1bio_t *r1_bio;
  63         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
  64
  65         /* allocate a r1bio with room for raid_disks entries in the bios array */
  66         r1_bio = kzalloc(size, gfp_flags);
  67         if (!r1_bio && pi->mddev)
  68                 unplug_slaves(pi->mddev);
  69
  70         return r1_bio;
  71 }
  72
  73 static void r1bio_pool_free(void *r1_bio, void *data)
  74 {
  75         kfree(r1_bio);
  76 }
  77
  78 #define RESYNC_BLOCK_SIZE (64*1024)
  79 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  80 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  81 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  82 #define RESYNC_WINDOW (2048*1024)
  83
  84 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  85 {
  86         struct pool_info *pi = data;
  87         struct page *page;
  88         r1bio_t *r1_bio;
  89         struct bio *bio;
  90         int i, j;
  91
  92         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  93         if (!r1_bio) {
  94                 unplug_slaves(pi->mddev);
  95                 return NULL;
  96         }
  97
  98         /*
  99          * Allocate bios : 1 for reading, n-1 for writing
 100          */
 101         for (j = pi->raid_disks ; j-- ; ) {
 102                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 103                 if (!bio)
 104                         goto out_free_bio;
 105                 r1_bio->bios[j] = bio;
 106         }
 107         /*
 108          * Allocate RESYNC_PAGES data pages and attach them to
 109          * the first bio.
 110          * If this is a user-requested check/repair, allocate
 111          * RESYNC_PAGES for each bio.
 112          */
 113         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 114                 j = pi->raid_disks;
 115         else
 116                 j = 1;
 117         while(j--) {
 118                 bio = r1_bio->bios[j];
 119                 for (i = 0; i < RESYNC_PAGES; i++) {
 120                         page = alloc_page(gfp_flags);
 121                         if (unlikely(!page))
 122                                 goto out_free_pages;
 123
 124                         bio->bi_io_vec[i].bv_page = page;
 125                         bio->bi_vcnt = i+1;
 126                 }
 127         }
 128         /* If not user-requests, copy the page pointers to all bios */
 129         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 130                 for (i=0; i<RESYNC_PAGES ; i++)
 131                         for (j=1; j<pi->raid_disks; j++)
 132                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 133                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 134         }
 135
 136         r1_bio->master_bio = NULL;
 137
 138         return r1_bio;
 139
 140 out_free_pages:
 141         for (j=0 ; j < pi->raid_disks; j++)
 142                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 143                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 144         j = -1;
 145 out_free_bio:
 146         while ( ++j < pi->raid_disks )
 147                 bio_put(r1_bio->bios[j]);
 148         r1bio_pool_free(r1_bio, data);
 149         return NULL;
 150 }
 151
 152 static void r1buf_pool_free(void *__r1_bio, void *data)
 153 {
 154         struct pool_info *pi = data;
 155         int i,j;
 156         r1bio_t *r1bio = __r1_bio;
 157
 158         for (i = 0; i < RESYNC_PAGES; i++)
 159                 for (j = pi->raid_disks; j-- ;) {
 160                         if (j == 0 ||
 161                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 162                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 163                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 164                 }
 165         for (i=0 ; i < pi->raid_disks; i++)
 166                 bio_put(r1bio->bios[i]);
 167
 168         r1bio_pool_free(r1bio, data);
 169 }
 170
 171 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 172 {
 173         int i;
 174
 175         for (i = 0; i < conf->raid_disks; i++) {
 176                 struct bio **bio = r1_bio->bios + i;
 177                 if (*bio && *bio != IO_BLOCKED)
 178                         bio_put(*bio);
 179                 *bio = NULL;
 180         }
 181 }
 182
 183 static void free_r1bio(r1bio_t *r1_bio)
 184 {
 185         conf_t *conf = r1_bio->mddev->private;
 186
 187         /*
 188          * Wake up any possible resync thread that waits for the device
 189          * to go idle.
 190          */
 191         allow_barrier(conf);
 192
 193         put_all_bios(conf, r1_bio);
 194         mempool_free(r1_bio, conf->r1bio_pool);
 195 }
 196
 197 static void put_buf(r1bio_t *r1_bio)
 198 {
 199         conf_t *conf = r1_bio->mddev->private;
 200         int i;
 201
 202         for (i=0; i<conf->raid_disks; i++) {
 203                 struct bio *bio = r1_bio->bios[i];
 204                 if (bio->bi_end_io)
 205                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 206         }
 207
 208         mempool_free(r1_bio, conf->r1buf_pool);
 209
 210         lower_barrier(conf);
 211 }
 212
 213 static void reschedule_retry(r1bio_t *r1_bio)
 214 {
 215         unsigned long flags;
 216         mddev_t *mddev = r1_bio->mddev;
 217         conf_t *conf = mddev->private;
 218
 219         spin_lock_irqsave(&conf->device_lock, flags);
 220         list_add(&r1_bio->retry_list, &conf->retry_list);
 221         conf->nr_queued ++;
 222         spin_unlock_irqrestore(&conf->device_lock, flags);
 223
 224         wake_up(&conf->wait_barrier);
 225         md_wakeup_thread(mddev->thread);
 226 }
 227
 228 /*
 229  * raid_end_bio_io() is called when we have finished servicing a mirrored
 230  * operation and are ready to return a success/failure code to the buffer
 231  * cache layer.
 232  */
 233 static void raid_end_bio_io(r1bio_t *r1_bio)
 234 {
 235         struct bio *bio = r1_bio->master_bio;
 236
 237         /* if nobody has done the final endio yet, do it now */
 238         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 239                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
 240                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 241                         (unsigned long long) bio->bi_sector,
 242                         (unsigned long long) bio->bi_sector +
 243                                 (bio->bi_size >> 9) - 1);
 244
 245                 bio_endio(bio,
 246                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 247         }
 248         free_r1bio(r1_bio);
 249 }
 250
 251 /*
 252  * Update disk head position estimator based on IRQ completion info.
 253  */
 254 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 255 {
 256         conf_t *conf = r1_bio->mddev->private;
 257
 258         conf->mirrors[disk].head_position =
 259                 r1_bio->sector + (r1_bio->sectors);
 260 }
 261
 262 static void raid1_end_read_request(struct bio *bio, int error)
 263 {
 264         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 265         r1bio_t *r1_bio = bio->bi_private;
 266         int mirror;
 267         conf_t *conf = r1_bio->mddev->private;
 268
 269         mirror = r1_bio->read_disk;
 270         /*
 271          * this branch is our 'one mirror IO has finished' event handler:
 272          */
 273         update_head_pos(mirror, r1_bio);
 274
 275         if (uptodate)
 276                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 277         else {
 278                 /* If all other devices have failed, we want to return
 279                  * the error upwards rather than fail the last device.
 280                  * Here we redefine "uptodate" to mean "Don't want to retry"
 281                  */
 282                 unsigned long flags;
 283                 spin_lock_irqsave(&conf->device_lock, flags);
 284                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 285                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 286                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 287                         uptodate = 1;
 288                 spin_unlock_irqrestore(&conf->device_lock, flags);
 289         }
 290
 291         if (uptodate)
 292                 raid_end_bio_io(r1_bio);
 293         else {
 294                 /*
 295                  * oops, read error:
 296                  */
 297                 char b[BDEVNAME_SIZE];
 298                 if (printk_ratelimit())
 299                         printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
 300                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 301                 reschedule_retry(r1_bio);
 302         }
 303
 304         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 305 }
 306
 307 static void raid1_end_write_request(struct bio *bio, int error)
 308 {
 309         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 310         r1bio_t *r1_bio = bio->bi_private;
 311         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 312         conf_t *conf = r1_bio->mddev->private;
 313         struct bio *to_put = NULL;
 314
 315
 316         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 317                 if (r1_bio->bios[mirror] == bio)
 318                         break;
 319
 320         if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
 321                 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
 322                 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
 323                 r1_bio->mddev->barriers_work = 0;
 324                 /* Don't rdev_dec_pending in this branch - keep it for the retry */
 325         } else {
 326                 /*
 327                  * this branch is our 'one mirror IO has finished' event handler:
 328                  */
 329                 r1_bio->bios[mirror] = NULL;
 330                 to_put = bio;
 331                 if (!uptodate) {
 332                         md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 333                         /* an I/O failed, we can't clear the bitmap */
 334                         set_bit(R1BIO_Degraded, &r1_bio->state);
 335                 } else
 336                         /*
 337                          * Set R1BIO_Uptodate in our master bio, so that
 338                          * we will return a good error code for to the higher
 339                          * levels even if IO on some other mirrored buffer fails.
 340                          *
 341                          * The 'master' represents the composite IO operation to
 342                          * user-side. So if something waits for IO, then it will
 343                          * wait for the 'master' bio.
 344                          */
 345                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 346
 347                 update_head_pos(mirror, r1_bio);
 348
 349                 if (behind) {
 350                         if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 351                                 atomic_dec(&r1_bio->behind_remaining);
 352
 353                         /* In behind mode, we ACK the master bio once the I/O has safely
 354                          * reached all non-writemostly disks. Setting the Returned bit
 355                          * ensures that this gets done only once -- we don't ever want to
 356                          * return -EIO here, instead we'll wait */
 357
 358                         if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 359                             test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 360                                 /* Maybe we can return now */
 361                                 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 362                                         struct bio *mbio = r1_bio->master_bio;
 363                                         PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 364                                                (unsigned long long) mbio->bi_sector,
 365                                                (unsigned long long) mbio->bi_sector +
 366                                                (mbio->bi_size >> 9) - 1);
 367                                         bio_endio(mbio, 0);
 368                                 }
 369                         }
 370                 }
 371                 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 372         }
 373         /*
 374          *
 375          * Let's see if all mirrored write operations have finished
 376          * already.
 377          */
 378         if (atomic_dec_and_test(&r1_bio->remaining)) {
 379                 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
 380                         reschedule_retry(r1_bio);
 381                 else {
 382                         /* it really is the end of this request */
 383                         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 384                                 /* free extra copy of the data pages */
 385                                 int i = bio->bi_vcnt;
 386                                 while (i--)
 387                                         safe_put_page(bio->bi_io_vec[i].bv_page);
 388                         }
 389                         /* clear the bitmap if all writes complete successfully */
 390                         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 391                                         r1_bio->sectors,
 392                                         !test_bit(R1BIO_Degraded, &r1_bio->state),
 393                                         behind);
 394                         md_write_end(r1_bio->mddev);
 395                         raid_end_bio_io(r1_bio);
 396                 }
 397         }
 398
 399         if (to_put)
 400                 bio_put(to_put);
 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->private;
 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                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 552
 553                         atomic_inc(&rdev->nr_pending);
 554                         rcu_read_unlock();
 555
 556                         blk_unplug(r_queue);
 557
 558                         rdev_dec_pending(rdev, mddev);
 559                         rcu_read_lock();
 560                 }
 561         }
 562         rcu_read_unlock();
 563 }
 564
 565 static void raid1_unplug(struct request_queue *q)
 566 {
 567         mddev_t *mddev = q->queuedata;
 568
 569         unplug_slaves(mddev);
 570         md_wakeup_thread(mddev->thread);
 571 }
 572
 573 static int raid1_congested(void *data, int bits)
 574 {
 575         mddev_t *mddev = data;
 576         conf_t *conf = mddev->private;
 577         int i, ret = 0;
 578
 579         if (mddev_congested(mddev, bits))
 580                 return 1;
 581
 582         rcu_read_lock();
 583         for (i = 0; i < mddev->raid_disks; i++) {
 584                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 585                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 586                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 587
 588                         /* Note the '|| 1' - when read_balance prefers
 589                          * non-congested targets, it can be removed
 590                          */
 591                         if ((bits & (1<<BDI_async_congested)) || 1)
 592                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 593                         else
 594                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 595                 }
 596         }
 597         rcu_read_unlock();
 598         return ret;
 599 }
 600
 601
 602 static int flush_pending_writes(conf_t *conf)
 603 {
 604         /* Any writes that have been queued but are awaiting
 605          * bitmap updates get flushed here.
 606          * We return 1 if any requests were actually submitted.
 607          */
 608         int rv = 0;
 609
 610         spin_lock_irq(&conf->device_lock);
 611
 612         if (conf->pending_bio_list.head) {
 613                 struct bio *bio;
 614                 bio = bio_list_get(&conf->pending_bio_list);
 615                 blk_remove_plug(conf->mddev->queue);
 616                 spin_unlock_irq(&conf->device_lock);
 617                 /* flush any pending bitmap writes to
 618                  * disk before proceeding w/ I/O */
 619                 bitmap_unplug(conf->mddev->bitmap);
 620
 621                 while (bio) { /* submit pending writes */
 622                         struct bio *next = bio->bi_next;
 623                         bio->bi_next = NULL;
 624                         generic_make_request(bio);
 625                         bio = next;
 626                 }
 627                 rv = 1;
 628         } else
 629                 spin_unlock_irq(&conf->device_lock);
 630         return rv;
 631 }
 632
 633 /* Barriers....
 634  * Sometimes we need to suspend IO while we do something else,
 635  * either some resync/recovery, or reconfigure the array.
 636  * To do this we raise a 'barrier'.
 637  * The 'barrier' is a counter that can be raised multiple times
 638  * to count how many activities are happening which preclude
 639  * normal IO.
 640  * We can only raise the barrier if there is no pending IO.
 641  * i.e. if nr_pending == 0.
 642  * We choose only to raise the barrier if no-one is waiting for the
 643  * barrier to go down.  This means that as soon as an IO request
 644  * is ready, no other operations which require a barrier will start
 645  * until the IO request has had a chance.
 646  *
 647  * So: regular IO calls 'wait_barrier'.  When that returns there
 648  *    is no backgroup IO happening,  It must arrange to call
 649  *    allow_barrier when it has finished its IO.
 650  * backgroup IO calls must call raise_barrier.  Once that returns
 651  *    there is no normal IO happeing.  It must arrange to call
 652  *    lower_barrier when the particular background IO completes.
 653  */
 654 #define RESYNC_DEPTH 32
 655
 656 static void raise_barrier(conf_t *conf)
 657 {
 658         spin_lock_irq(&conf->resync_lock);
 659
 660         /* Wait until no block IO is waiting */
 661         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 662                             conf->resync_lock,
 663                             raid1_unplug(conf->mddev->queue));
 664
 665         /* block any new IO from starting */
 666         conf->barrier++;
 667
 668         /* No wait for all pending IO to complete */
 669         wait_event_lock_irq(conf->wait_barrier,
 670                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 671                             conf->resync_lock,
 672                             raid1_unplug(conf->mddev->queue));
 673
 674         spin_unlock_irq(&conf->resync_lock);
 675 }
 676
 677 static void lower_barrier(conf_t *conf)
 678 {
 679         unsigned long flags;
 680         BUG_ON(conf->barrier <= 0);
 681         spin_lock_irqsave(&conf->resync_lock, flags);
 682         conf->barrier--;
 683         spin_unlock_irqrestore(&conf->resync_lock, flags);
 684         wake_up(&conf->wait_barrier);
 685 }
 686
 687 static void wait_barrier(conf_t *conf)
 688 {
 689         spin_lock_irq(&conf->resync_lock);
 690         if (conf->barrier) {
 691                 conf->nr_waiting++;
 692                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 693                                     conf->resync_lock,
 694                                     raid1_unplug(conf->mddev->queue));
 695                 conf->nr_waiting--;
 696         }
 697         conf->nr_pending++;
 698         spin_unlock_irq(&conf->resync_lock);
 699 }
 700
 701 static void allow_barrier(conf_t *conf)
 702 {
 703         unsigned long flags;
 704         spin_lock_irqsave(&conf->resync_lock, flags);
 705         conf->nr_pending--;
 706         spin_unlock_irqrestore(&conf->resync_lock, flags);
 707         wake_up(&conf->wait_barrier);
 708 }
 709
 710 static void freeze_array(conf_t *conf)
 711 {
 712         /* stop syncio and normal IO and wait for everything to
 713          * go quite.
 714          * We increment barrier and nr_waiting, and then
 715          * wait until nr_pending match nr_queued+1
 716          * This is called in the context of one normal IO request
 717          * that has failed. Thus any sync request that might be pending
 718          * will be blocked by nr_pending, and we need to wait for
 719          * pending IO requests to complete or be queued for re-try.
 720          * Thus the number queued (nr_queued) plus this request (1)
 721          * must match the number of pending IOs (nr_pending) before
 722          * we continue.
 723          */
 724         spin_lock_irq(&conf->resync_lock);
 725         conf->barrier++;
 726         conf->nr_waiting++;
 727         wait_event_lock_irq(conf->wait_barrier,
 728                             conf->nr_pending == conf->nr_queued+1,
 729                             conf->resync_lock,
 730                             ({ flush_pending_writes(conf);
 731                                raid1_unplug(conf->mddev->queue); }));
 732         spin_unlock_irq(&conf->resync_lock);
 733 }
 734 static void unfreeze_array(conf_t *conf)
 735 {
 736         /* reverse the effect of the freeze */
 737         spin_lock_irq(&conf->resync_lock);
 738         conf->barrier--;
 739         conf->nr_waiting--;
 740         wake_up(&conf->wait_barrier);
 741         spin_unlock_irq(&conf->resync_lock);
 742 }
 743
 744
 745 /* duplicate the data pages for behind I/O */
 746 static struct page **alloc_behind_pages(struct bio *bio)
 747 {
 748         int i;
 749         struct bio_vec *bvec;
 750         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
 751                                         GFP_NOIO);
 752         if (unlikely(!pages))
 753                 goto do_sync_io;
 754
 755         bio_for_each_segment(bvec, bio, i) {
 756                 pages[i] = alloc_page(GFP_NOIO);
 757                 if (unlikely(!pages[i]))
 758                         goto do_sync_io;
 759                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 760                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 761                 kunmap(pages[i]);
 762                 kunmap(bvec->bv_page);
 763         }
 764
 765         return pages;
 766
 767 do_sync_io:
 768         if (pages)
 769                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
 770                         put_page(pages[i]);
 771         kfree(pages);
 772         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 773         return NULL;
 774 }
 775
 776 static int make_request(mddev_t *mddev, struct bio * bio)
 777 {
 778         conf_t *conf = mddev->private;
 779         mirror_info_t *mirror;
 780         r1bio_t *r1_bio;
 781         struct bio *read_bio;
 782         int i, targets = 0, disks;
 783         struct bitmap *bitmap;
 784         unsigned long flags;
 785         struct bio_list bl;
 786         struct page **behind_pages = NULL;
 787         const int rw = bio_data_dir(bio);
 788         const bool do_sync = bio_rw_flagged(bio, BIO_RW_SYNCIO);
 789         bool do_barriers;
 790         mdk_rdev_t *blocked_rdev;
 791
 792         /*
 793          * Register the new request and wait if the reconstruction
 794          * thread has put up a bar for new requests.
 795          * Continue immediately if no resync is active currently.
 796          * We test barriers_work *after* md_write_start as md_write_start
 797          * may cause the first superblock write, and that will check out
 798          * if barriers work.
 799          */
 800
 801         md_write_start(mddev, bio); /* wait on superblock update early */
 802
 803         if (bio_data_dir(bio) == WRITE &&
 804             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
 805             bio->bi_sector < mddev->suspend_hi) {
 806                 /* As the suspend_* range is controlled by
 807                  * userspace, we want an interruptible
 808                  * wait.
 809                  */
 810                 DEFINE_WAIT(w);
 811                 for (;;) {
 812                         flush_signals(current);
 813                         prepare_to_wait(&conf->wait_barrier,
 814                                         &w, TASK_INTERRUPTIBLE);
 815                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
 816                             bio->bi_sector >= mddev->suspend_hi)
 817                                 break;
 818                         schedule();
 819                 }
 820                 finish_wait(&conf->wait_barrier, &w);
 821         }
 822         if (unlikely(!mddev->barriers_work &&
 823                      bio_rw_flagged(bio, BIO_RW_BARRIER))) {
 824                 if (rw == WRITE)
 825                         md_write_end(mddev);
 826                 bio_endio(bio, -EOPNOTSUPP);
 827                 return 0;
 828         }
 829
 830         wait_barrier(conf);
 831
 832         bitmap = mddev->bitmap;
 833
 834         /*
 835          * make_request() can abort the operation when READA is being
 836          * used and no empty request is available.
 837          *
 838          */
 839         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 840
 841         r1_bio->master_bio = bio;
 842         r1_bio->sectors = bio->bi_size >> 9;
 843         r1_bio->state = 0;
 844         r1_bio->mddev = mddev;
 845         r1_bio->sector = bio->bi_sector;
 846
 847         if (rw == READ) {
 848                 /*
 849                  * read balancing logic:
 850                  */
 851                 int rdisk = read_balance(conf, r1_bio);
 852
 853                 if (rdisk < 0) {
 854                         /* couldn't find anywhere to read from */
 855                         raid_end_bio_io(r1_bio);
 856                         return 0;
 857                 }
 858                 mirror = conf->mirrors + rdisk;
 859
 860                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
 861                     bitmap) {
 862                         /* Reading from a write-mostly device must
 863                          * take care not to over-take any writes
 864                          * that are 'behind'
 865                          */
 866                         wait_event(bitmap->behind_wait,
 867                                    atomic_read(&bitmap->behind_writes) == 0);
 868                 }
 869                 r1_bio->read_disk = rdisk;
 870
 871                 read_bio = bio_clone(bio, GFP_NOIO);
 872
 873                 r1_bio->bios[rdisk] = read_bio;
 874
 875                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 876                 read_bio->bi_bdev = mirror->rdev->bdev;
 877                 read_bio->bi_end_io = raid1_end_read_request;
 878                 read_bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
 879                 read_bio->bi_private = r1_bio;
 880
 881                 generic_make_request(read_bio);
 882                 return 0;
 883         }
 884
 885         /*
 886          * WRITE:
 887          */
 888         /* first select target devices under spinlock and
 889          * inc refcount on their rdev.  Record them by setting
 890          * bios[x] to bio
 891          */
 892         disks = conf->raid_disks;
 893 #if 0
 894         { static int first=1;
 895         if (first) printk("First Write sector %llu disks %d\n",
 896                           (unsigned long long)r1_bio->sector, disks);
 897         first = 0;
 898         }
 899 #endif
 900  retry_write:
 901         blocked_rdev = NULL;
 902         rcu_read_lock();
 903         for (i = 0;  i < disks; i++) {
 904                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 905                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 906                         atomic_inc(&rdev->nr_pending);
 907                         blocked_rdev = rdev;
 908                         break;
 909                 }
 910                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 911                         atomic_inc(&rdev->nr_pending);
 912                         if (test_bit(Faulty, &rdev->flags)) {
 913                                 rdev_dec_pending(rdev, mddev);
 914                                 r1_bio->bios[i] = NULL;
 915                         } else {
 916                                 r1_bio->bios[i] = bio;
 917                                 targets++;
 918                         }
 919                 } else
 920                         r1_bio->bios[i] = NULL;
 921         }
 922         rcu_read_unlock();
 923
 924         if (unlikely(blocked_rdev)) {
 925                 /* Wait for this device to become unblocked */
 926                 int j;
 927
 928                 for (j = 0; j < i; j++)
 929                         if (r1_bio->bios[j])
 930                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
 931
 932                 allow_barrier(conf);
 933                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 934                 wait_barrier(conf);
 935                 goto retry_write;
 936         }
 937
 938         BUG_ON(targets == 0); /* we never fail the last device */
 939
 940         if (targets < conf->raid_disks) {
 941                 /* array is degraded, we will not clear the bitmap
 942                  * on I/O completion (see raid1_end_write_request) */
 943                 set_bit(R1BIO_Degraded, &r1_bio->state);
 944         }
 945
 946         /* do behind I/O ?
 947          * Not if there are too many, or cannot allocate memory,
 948          * or a reader on WriteMostly is waiting for behind writes
 949          * to flush */
 950         if (bitmap &&
 951             (atomic_read(&bitmap->behind_writes)
 952              < mddev->bitmap_info.max_write_behind) &&
 953             !waitqueue_active(&bitmap->behind_wait) &&
 954             (behind_pages = alloc_behind_pages(bio)) != NULL)
 955                 set_bit(R1BIO_BehindIO, &r1_bio->state);
 956
 957         atomic_set(&r1_bio->remaining, 0);
 958         atomic_set(&r1_bio->behind_remaining, 0);
 959
 960         do_barriers = bio_rw_flagged(bio, BIO_RW_BARRIER);
 961         if (do_barriers)
 962                 set_bit(R1BIO_Barrier, &r1_bio->state);
 963
 964         bio_list_init(&bl);
 965         for (i = 0; i < disks; i++) {
 966                 struct bio *mbio;
 967                 if (!r1_bio->bios[i])
 968                         continue;
 969
 970                 mbio = bio_clone(bio, GFP_NOIO);
 971                 r1_bio->bios[i] = mbio;
 972
 973                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 974                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 975                 mbio->bi_end_io = raid1_end_write_request;
 976                 mbio->bi_rw = WRITE | (do_barriers << BIO_RW_BARRIER) |
 977                         (do_sync << BIO_RW_SYNCIO);
 978                 mbio->bi_private = r1_bio;
 979
 980                 if (behind_pages) {
 981                         struct bio_vec *bvec;
 982                         int j;
 983
 984                         /* Yes, I really want the '__' version so that
 985                          * we clear any unused pointer in the io_vec, rather
 986                          * than leave them unchanged.  This is important
 987                          * because when we come to free the pages, we won't
 988                          * know the originial bi_idx, so we just free
 989                          * them all
 990                          */
 991                         __bio_for_each_segment(bvec, mbio, j, 0)
 992                                 bvec->bv_page = behind_pages[j];
 993                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 994                                 atomic_inc(&r1_bio->behind_remaining);
 995                 }
 996
 997                 atomic_inc(&r1_bio->remaining);
 998
 999                 bio_list_add(&bl, mbio);
1000         }
1001         kfree(behind_pages); /* the behind pages are attached to the bios now */
1002
1003         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
1004                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
1005         spin_lock_irqsave(&conf->device_lock, flags);
1006         bio_list_merge(&conf->pending_bio_list, &bl);
1007         bio_list_init(&bl);
1008
1009         blk_plug_device(mddev->queue);
1010         spin_unlock_irqrestore(&conf->device_lock, flags);
1011
1012         /* In case raid1d snuck into freeze_array */
1013         wake_up(&conf->wait_barrier);
1014
1015         if (do_sync)
1016                 md_wakeup_thread(mddev->thread);
1017 #if 0
1018         while ((bio = bio_list_pop(&bl)) != NULL)
1019                 generic_make_request(bio);
1020 #endif
1021
1022         return 0;
1023 }
1024
1025 static void status(struct seq_file *seq, mddev_t *mddev)
1026 {
1027         conf_t *conf = mddev->private;
1028         int i;
1029
1030         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1031                    conf->raid_disks - mddev->degraded);
1032         rcu_read_lock();
1033         for (i = 0; i < conf->raid_disks; i++) {
1034                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1035                 seq_printf(seq, "%s",
1036                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1037         }
1038         rcu_read_unlock();
1039         seq_printf(seq, "]");
1040 }
1041
1042
1043 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1044 {
1045         char b[BDEVNAME_SIZE];
1046         conf_t *conf = mddev->private;
1047
1048         /*
1049          * If it is not operational, then we have already marked it as dead
1050          * else if it is the last working disks, ignore the error, let the
1051          * next level up know.
1052          * else mark the drive as failed
1053          */
1054         if (test_bit(In_sync, &rdev->flags)
1055             && (conf->raid_disks - mddev->degraded) == 1) {
1056                 /*
1057                  * Don't fail the drive, act as though we were just a
1058                  * normal single drive.
1059                  * However don't try a recovery from this drive as
1060                  * it is very likely to fail.
1061                  */
1062                 mddev->recovery_disabled = 1;
1063                 return;
1064         }
1065         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1066                 unsigned long flags;
1067                 spin_lock_irqsave(&conf->device_lock, flags);
1068                 mddev->degraded++;
1069                 set_bit(Faulty, &rdev->flags);
1070                 spin_unlock_irqrestore(&conf->device_lock, flags);
1071                 /*
1072                  * if recovery is running, make sure it aborts.
1073                  */
1074                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1075         } else
1076                 set_bit(Faulty, &rdev->flags);
1077         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1078         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device.\n"
1079                 "raid1: Operation continuing on %d devices.\n",
1080                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1081 }
1082
1083 static void print_conf(conf_t *conf)
1084 {
1085         int i;
1086
1087         printk("RAID1 conf printout:\n");
1088         if (!conf) {
1089                 printk("(!conf)\n");
1090                 return;
1091         }
1092         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1093                 conf->raid_disks);
1094
1095         rcu_read_lock();
1096         for (i = 0; i < conf->raid_disks; i++) {
1097                 char b[BDEVNAME_SIZE];
1098                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1099                 if (rdev)
1100                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1101                                i, !test_bit(In_sync, &rdev->flags),
1102                                !test_bit(Faulty, &rdev->flags),
1103                                bdevname(rdev->bdev,b));
1104         }
1105         rcu_read_unlock();
1106 }
1107
1108 static void close_sync(conf_t *conf)
1109 {
1110         wait_barrier(conf);
1111         allow_barrier(conf);
1112
1113         mempool_destroy(conf->r1buf_pool);
1114         conf->r1buf_pool = NULL;
1115 }
1116
1117 static int raid1_spare_active(mddev_t *mddev)
1118 {
1119         int i;
1120         conf_t *conf = mddev->private;
1121
1122         /*
1123          * Find all failed disks within the RAID1 configuration
1124          * and mark them readable.
1125          * Called under mddev lock, so rcu protection not needed.
1126          */
1127         for (i = 0; i < conf->raid_disks; i++) {
1128                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1129                 if (rdev
1130                     && !test_bit(Faulty, &rdev->flags)
1131                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1132                         unsigned long flags;
1133                         spin_lock_irqsave(&conf->device_lock, flags);
1134                         mddev->degraded--;
1135                         spin_unlock_irqrestore(&conf->device_lock, flags);
1136                 }
1137         }
1138
1139         print_conf(conf);
1140         return 0;
1141 }
1142
1143
1144 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1145 {
1146         conf_t *conf = mddev->private;
1147         int err = -EEXIST;
1148         int mirror = 0;
1149         mirror_info_t *p;
1150         int first = 0;
1151         int last = mddev->raid_disks - 1;
1152
1153         if (rdev->raid_disk >= 0)
1154                 first = last = rdev->raid_disk;
1155
1156         for (mirror = first; mirror <= last; mirror++)
1157                 if ( !(p=conf->mirrors+mirror)->rdev) {
1158
1159                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1160                                           rdev->data_offset << 9);
1161                         /* as we don't honour merge_bvec_fn, we must
1162                          * never risk violating it, so limit
1163                          * ->max_segments to one lying with a single
1164                          * page, as a one page request is never in
1165                          * violation.
1166                          */
1167                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1168                                 blk_queue_max_segments(mddev->queue, 1);
1169                                 blk_queue_segment_boundary(mddev->queue,
1170                                                            PAGE_CACHE_SIZE - 1);
1171                         }
1172
1173                         p->head_position = 0;
1174                         rdev->raid_disk = mirror;
1175                         err = 0;
1176                         /* As all devices are equivalent, we don't need a full recovery
1177                          * if this was recently any drive of the array
1178                          */
1179                         if (rdev->saved_raid_disk < 0)
1180                                 conf->fullsync = 1;
1181                         rcu_assign_pointer(p->rdev, rdev);
1182                         break;
1183                 }
1184         md_integrity_add_rdev(rdev, mddev);
1185         print_conf(conf);
1186         return err;
1187 }
1188
1189 static int raid1_remove_disk(mddev_t *mddev, int number)
1190 {
1191         conf_t *conf = mddev->private;
1192         int err = 0;
1193         mdk_rdev_t *rdev;
1194         mirror_info_t *p = conf->mirrors+ number;
1195
1196         print_conf(conf);
1197         rdev = p->rdev;
1198         if (rdev) {
1199                 if (test_bit(In_sync, &rdev->flags) ||
1200                     atomic_read(&rdev->nr_pending)) {
1201                         err = -EBUSY;
1202                         goto abort;
1203                 }
1204                 /* Only remove non-faulty devices is recovery
1205                  * is not possible.
1206                  */
1207                 if (!test_bit(Faulty, &rdev->flags) &&
1208                     mddev->degraded < conf->raid_disks) {
1209                         err = -EBUSY;
1210                         goto abort;
1211                 }
1212                 p->rdev = NULL;
1213                 synchronize_rcu();
1214                 if (atomic_read(&rdev->nr_pending)) {
1215                         /* lost the race, try later */
1216                         err = -EBUSY;
1217                         p->rdev = rdev;
1218                         goto abort;
1219                 }
1220                 md_integrity_register(mddev);
1221         }
1222 abort:
1223
1224         print_conf(conf);
1225         return err;
1226 }
1227
1228
1229 static void end_sync_read(struct bio *bio, int error)
1230 {
1231         r1bio_t *r1_bio = bio->bi_private;
1232         int i;
1233
1234         for (i=r1_bio->mddev->raid_disks; i--; )
1235                 if (r1_bio->bios[i] == bio)
1236                         break;
1237         BUG_ON(i < 0);
1238         update_head_pos(i, r1_bio);
1239         /*
1240          * we have read a block, now it needs to be re-written,
1241          * or re-read if the read failed.
1242          * We don't do much here, just schedule handling by raid1d
1243          */
1244         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1245                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1246
1247         if (atomic_dec_and_test(&r1_bio->remaining))
1248                 reschedule_retry(r1_bio);
1249 }
1250
1251 static void end_sync_write(struct bio *bio, int error)
1252 {
1253         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1254         r1bio_t *r1_bio = bio->bi_private;
1255         mddev_t *mddev = r1_bio->mddev;
1256         conf_t *conf = mddev->private;
1257         int i;
1258         int mirror=0;
1259
1260         for (i = 0; i < conf->raid_disks; i++)
1261                 if (r1_bio->bios[i] == bio) {
1262                         mirror = i;
1263                         break;
1264                 }
1265         if (!uptodate) {
1266                 int sync_blocks = 0;
1267                 sector_t s = r1_bio->sector;
1268                 long sectors_to_go = r1_bio->sectors;
1269                 /* make sure these bits doesn't get cleared. */
1270                 do {
1271                         bitmap_end_sync(mddev->bitmap, s,
1272                                         &sync_blocks, 1);
1273                         s += sync_blocks;
1274                         sectors_to_go -= sync_blocks;
1275                 } while (sectors_to_go > 0);
1276                 md_error(mddev, conf->mirrors[mirror].rdev);
1277         }
1278
1279         update_head_pos(mirror, r1_bio);
1280
1281         if (atomic_dec_and_test(&r1_bio->remaining)) {
1282                 sector_t s = r1_bio->sectors;
1283                 put_buf(r1_bio);
1284                 md_done_sync(mddev, s, uptodate);
1285         }
1286 }
1287
1288 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1289 {
1290         conf_t *conf = mddev->private;
1291         int i;
1292         int disks = conf->raid_disks;
1293         struct bio *bio, *wbio;
1294
1295         bio = r1_bio->bios[r1_bio->read_disk];
1296
1297
1298         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1299                 /* We have read all readable devices.  If we haven't
1300                  * got the block, then there is no hope left.
1301                  * If we have, then we want to do a comparison
1302                  * and skip the write if everything is the same.
1303                  * If any blocks failed to read, then we need to
1304                  * attempt an over-write
1305                  */
1306                 int primary;
1307                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1308                         for (i=0; i<mddev->raid_disks; i++)
1309                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1310                                         md_error(mddev, conf->mirrors[i].rdev);
1311
1312                         md_done_sync(mddev, r1_bio->sectors, 1);
1313                         put_buf(r1_bio);
1314                         return;
1315                 }
1316                 for (primary=0; primary<mddev->raid_disks; primary++)
1317                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1318                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1319                                 r1_bio->bios[primary]->bi_end_io = NULL;
1320                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1321                                 break;
1322                         }
1323                 r1_bio->read_disk = primary;
1324                 for (i=0; i<mddev->raid_disks; i++)
1325                         if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1326                                 int j;
1327                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1328                                 struct bio *pbio = r1_bio->bios[primary];
1329                                 struct bio *sbio = r1_bio->bios[i];
1330
1331                                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1332                                         for (j = vcnt; j-- ; ) {
1333                                                 struct page *p, *s;
1334                                                 p = pbio->bi_io_vec[j].bv_page;
1335                                                 s = sbio->bi_io_vec[j].bv_page;
1336                                                 if (memcmp(page_address(p),
1337                                                            page_address(s),
1338                                                            PAGE_SIZE))
1339                                                         break;
1340                                         }
1341                                 } else
1342                                         j = 0;
1343                                 if (j >= 0)
1344                                         mddev->resync_mismatches += r1_bio->sectors;
1345                                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1346                                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1347                                         sbio->bi_end_io = NULL;
1348                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1349                                 } else {
1350                                         /* fixup the bio for reuse */
1351                                         int size;
1352                                         sbio->bi_vcnt = vcnt;
1353                                         sbio->bi_size = r1_bio->sectors << 9;
1354                                         sbio->bi_idx = 0;
1355                                         sbio->bi_phys_segments = 0;
1356                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1357                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1358                                         sbio->bi_next = NULL;
1359                                         sbio->bi_sector = r1_bio->sector +
1360                                                 conf->mirrors[i].rdev->data_offset;
1361                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1362                                         size = sbio->bi_size;
1363                                         for (j = 0; j < vcnt ; j++) {
1364                                                 struct bio_vec *bi;
1365                                                 bi = &sbio->bi_io_vec[j];
1366                                                 bi->bv_offset = 0;
1367                                                 if (size > PAGE_SIZE)
1368                                                         bi->bv_len = PAGE_SIZE;
1369                                                 else
1370                                                         bi->bv_len = size;
1371                                                 size -= PAGE_SIZE;
1372                                                 memcpy(page_address(bi->bv_page),
1373                                                        page_address(pbio->bi_io_vec[j].bv_page),
1374                                                        PAGE_SIZE);
1375                                         }
1376
1377                                 }
1378                         }
1379         }
1380         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1381                 /* ouch - failed to read all of that.
1382                  * Try some synchronous reads of other devices to get
1383                  * good data, much like with normal read errors.  Only
1384                  * read into the pages we already have so we don't
1385                  * need to re-issue the read request.
1386                  * We don't need to freeze the array, because being in an
1387                  * active sync request, there is no normal IO, and
1388                  * no overlapping syncs.
1389                  */
1390                 sector_t sect = r1_bio->sector;
1391                 int sectors = r1_bio->sectors;
1392                 int idx = 0;
1393
1394                 while(sectors) {
1395                         int s = sectors;
1396                         int d = r1_bio->read_disk;
1397                         int success = 0;
1398                         mdk_rdev_t *rdev;
1399
1400                         if (s > (PAGE_SIZE>>9))
1401                                 s = PAGE_SIZE >> 9;
1402                         do {
1403                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1404                                         /* No rcu protection needed here devices
1405                                          * can only be removed when no resync is
1406                                          * active, and resync is currently active
1407                                          */
1408                                         rdev = conf->mirrors[d].rdev;
1409                                         if (sync_page_io(rdev->bdev,
1410                                                          sect + rdev->data_offset,
1411                                                          s<<9,
1412                                                          bio->bi_io_vec[idx].bv_page,
1413                                                          READ)) {
1414                                                 success = 1;
1415                                                 break;
1416                                         }
1417                                 }
1418                                 d++;
1419                                 if (d == conf->raid_disks)
1420                                         d = 0;
1421                         } while (!success && d != r1_bio->read_disk);
1422
1423                         if (success) {
1424                                 int start = d;
1425                                 /* write it back and re-read */
1426                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1427                                 while (d != r1_bio->read_disk) {
1428                                         if (d == 0)
1429                                                 d = conf->raid_disks;
1430                                         d--;
1431                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1432                                                 continue;
1433                                         rdev = conf->mirrors[d].rdev;
1434                                         atomic_add(s, &rdev->corrected_errors);
1435                                         if (sync_page_io(rdev->bdev,
1436                                                          sect + rdev->data_offset,
1437                                                          s<<9,
1438                                                          bio->bi_io_vec[idx].bv_page,
1439                                                          WRITE) == 0)
1440                                                 md_error(mddev, rdev);
1441                                 }
1442                                 d = start;
1443                                 while (d != r1_bio->read_disk) {
1444                                         if (d == 0)
1445                                                 d = conf->raid_disks;
1446                                         d--;
1447                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1448                                                 continue;
1449                                         rdev = conf->mirrors[d].rdev;
1450                                         if (sync_page_io(rdev->bdev,
1451                                                          sect + rdev->data_offset,
1452                                                          s<<9,
1453                                                          bio->bi_io_vec[idx].bv_page,
1454                                                          READ) == 0)
1455                                                 md_error(mddev, rdev);
1456                                 }
1457                         } else {
1458                                 char b[BDEVNAME_SIZE];
1459                                 /* Cannot read from anywhere, array is toast */
1460                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1461                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1462                                        " for block %llu\n",
1463                                        bdevname(bio->bi_bdev,b),
1464                                        (unsigned long long)r1_bio->sector);
1465                                 md_done_sync(mddev, r1_bio->sectors, 0);
1466                                 put_buf(r1_bio);
1467                                 return;
1468                         }
1469                         sectors -= s;
1470                         sect += s;
1471                         idx ++;
1472                 }
1473         }
1474
1475         /*
1476          * schedule writes
1477          */
1478         atomic_set(&r1_bio->remaining, 1);
1479         for (i = 0; i < disks ; i++) {
1480                 wbio = r1_bio->bios[i];
1481                 if (wbio->bi_end_io == NULL ||
1482                     (wbio->bi_end_io == end_sync_read &&
1483                      (i == r1_bio->read_disk ||
1484                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1485                         continue;
1486
1487                 wbio->bi_rw = WRITE;
1488                 wbio->bi_end_io = end_sync_write;
1489                 atomic_inc(&r1_bio->remaining);
1490                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1491
1492                 generic_make_request(wbio);
1493         }
1494
1495         if (atomic_dec_and_test(&r1_bio->remaining)) {
1496                 /* if we're here, all write(s) have completed, so clean up */
1497                 md_done_sync(mddev, r1_bio->sectors, 1);
1498                 put_buf(r1_bio);
1499         }
1500 }
1501
1502 /*
1503  * This is a kernel thread which:
1504  *
1505  *      1.      Retries failed read operations on working mirrors.
1506  *      2.      Updates the raid superblock when problems encounter.
1507  *      3.      Performs writes following reads for array syncronising.
1508  */
1509
1510 static void fix_read_error(conf_t *conf, int read_disk,
1511                            sector_t sect, int sectors)
1512 {
1513         mddev_t *mddev = conf->mddev;
1514         while(sectors) {
1515                 int s = sectors;
1516                 int d = read_disk;
1517                 int success = 0;
1518                 int start;
1519                 mdk_rdev_t *rdev;
1520
1521                 if (s > (PAGE_SIZE>>9))
1522                         s = PAGE_SIZE >> 9;
1523
1524                 do {
1525                         /* Note: no rcu protection needed here
1526                          * as this is synchronous in the raid1d thread
1527                          * which is the thread that might remove
1528                          * a device.  If raid1d ever becomes multi-threaded....
1529                          */
1530                         rdev = conf->mirrors[d].rdev;
1531                         if (rdev &&
1532                             test_bit(In_sync, &rdev->flags) &&
1533                             sync_page_io(rdev->bdev,
1534                                          sect + rdev->data_offset,
1535                                          s<<9,
1536                                          conf->tmppage, READ))
1537                                 success = 1;
1538                         else {
1539                                 d++;
1540                                 if (d == conf->raid_disks)
1541                                         d = 0;
1542                         }
1543                 } while (!success && d != read_disk);
1544
1545                 if (!success) {
1546                         /* Cannot read from anywhere -- bye bye array */
1547                         md_error(mddev, conf->mirrors[read_disk].rdev);
1548                         break;
1549                 }
1550                 /* write it back and re-read */
1551                 start = d;
1552                 while (d != read_disk) {
1553                         if (d==0)
1554                                 d = conf->raid_disks;
1555                         d--;
1556                         rdev = conf->mirrors[d].rdev;
1557                         if (rdev &&
1558                             test_bit(In_sync, &rdev->flags)) {
1559                                 if (sync_page_io(rdev->bdev,
1560                                                  sect + rdev->data_offset,
1561                                                  s<<9, conf->tmppage, WRITE)
1562                                     == 0)
1563                                         /* Well, this device is dead */
1564                                         md_error(mddev, rdev);
1565                         }
1566                 }
1567                 d = start;
1568                 while (d != read_disk) {
1569                         char b[BDEVNAME_SIZE];
1570                         if (d==0)
1571                                 d = conf->raid_disks;
1572                         d--;
1573                         rdev = conf->mirrors[d].rdev;
1574                         if (rdev &&
1575                             test_bit(In_sync, &rdev->flags)) {
1576                                 if (sync_page_io(rdev->bdev,
1577                                                  sect + rdev->data_offset,
1578                                                  s<<9, conf->tmppage, READ)
1579                                     == 0)
1580                                         /* Well, this device is dead */
1581                                         md_error(mddev, rdev);
1582                                 else {
1583                                         atomic_add(s, &rdev->corrected_errors);
1584                                         printk(KERN_INFO
1585                                                "raid1:%s: read error corrected "
1586                                                "(%d sectors at %llu on %s)\n",
1587                                                mdname(mddev), s,
1588                                                (unsigned long long)(sect +
1589                                                    rdev->data_offset),
1590                                                bdevname(rdev->bdev, b));
1591                                 }
1592                         }
1593                 }
1594                 sectors -= s;
1595                 sect += s;
1596         }
1597 }
1598
1599 static void raid1d(mddev_t *mddev)
1600 {
1601         r1bio_t *r1_bio;
1602         struct bio *bio;
1603         unsigned long flags;
1604         conf_t *conf = mddev->private;
1605         struct list_head *head = &conf->retry_list;
1606         int unplug=0;
1607         mdk_rdev_t *rdev;
1608
1609         md_check_recovery(mddev);
1610
1611         for (;;) {
1612                 char b[BDEVNAME_SIZE];
1613
1614                 unplug += flush_pending_writes(conf);
1615
1616                 spin_lock_irqsave(&conf->device_lock, flags);
1617                 if (list_empty(head)) {
1618                         spin_unlock_irqrestore(&conf->device_lock, flags);
1619                         break;
1620                 }
1621                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1622                 list_del(head->prev);
1623                 conf->nr_queued--;
1624                 spin_unlock_irqrestore(&conf->device_lock, flags);
1625
1626                 mddev = r1_bio->mddev;
1627                 conf = mddev->private;
1628                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1629                         sync_request_write(mddev, r1_bio);
1630                         unplug = 1;
1631                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1632                         /* some requests in the r1bio were BIO_RW_BARRIER
1633                          * requests which failed with -EOPNOTSUPP.  Hohumm..
1634                          * Better resubmit without the barrier.
1635                          * We know which devices to resubmit for, because
1636                          * all others have had their bios[] entry cleared.
1637                          * We already have a nr_pending reference on these rdevs.
1638                          */
1639                         int i;
1640                         const bool do_sync = bio_rw_flagged(r1_bio->master_bio, BIO_RW_SYNCIO);
1641                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1642                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1643                         for (i=0; i < conf->raid_disks; i++)
1644                                 if (r1_bio->bios[i])
1645                                         atomic_inc(&r1_bio->remaining);
1646                         for (i=0; i < conf->raid_disks; i++)
1647                                 if (r1_bio->bios[i]) {
1648                                         struct bio_vec *bvec;
1649                                         int j;
1650
1651                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1652                                         /* copy pages from the failed bio, as
1653                                          * this might be a write-behind device */
1654                                         __bio_for_each_segment(bvec, bio, j, 0)
1655                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1656                                         bio_put(r1_bio->bios[i]);
1657                                         bio->bi_sector = r1_bio->sector +
1658                                                 conf->mirrors[i].rdev->data_offset;
1659                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1660                                         bio->bi_end_io = raid1_end_write_request;
1661                                         bio->bi_rw = WRITE |
1662                                                 (do_sync << BIO_RW_SYNCIO);
1663                                         bio->bi_private = r1_bio;
1664                                         r1_bio->bios[i] = bio;
1665                                         generic_make_request(bio);
1666                                 }
1667                 } else {
1668                         int disk;
1669
1670                         /* we got a read error. Maybe the drive is bad.  Maybe just
1671                          * the block and we can fix it.
1672                          * We freeze all other IO, and try reading the block from
1673                          * other devices.  When we find one, we re-write
1674                          * and check it that fixes the read error.
1675                          * This is all done synchronously while the array is
1676                          * frozen
1677                          */
1678                         if (mddev->ro == 0) {
1679                                 freeze_array(conf);
1680                                 fix_read_error(conf, r1_bio->read_disk,
1681                                                r1_bio->sector,
1682                                                r1_bio->sectors);
1683                                 unfreeze_array(conf);
1684                         } else
1685                                 md_error(mddev,
1686                                          conf->mirrors[r1_bio->read_disk].rdev);
1687
1688                         bio = r1_bio->bios[r1_bio->read_disk];
1689                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1690                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1691                                        " read error for block %llu\n",
1692                                        bdevname(bio->bi_bdev,b),
1693                                        (unsigned long long)r1_bio->sector);
1694                                 raid_end_bio_io(r1_bio);
1695                         } else {
1696                                 const bool do_sync = bio_rw_flagged(r1_bio->master_bio, BIO_RW_SYNCIO);
1697                                 r1_bio->bios[r1_bio->read_disk] =
1698                                         mddev->ro ? IO_BLOCKED : NULL;
1699                                 r1_bio->read_disk = disk;
1700                                 bio_put(bio);
1701                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1702                                 r1_bio->bios[r1_bio->read_disk] = bio;
1703                                 rdev = conf->mirrors[disk].rdev;
1704                                 if (printk_ratelimit())
1705                                         printk(KERN_ERR "raid1: redirecting sector %llu to"
1706                                                " other mirror: %s\n",
1707                                                (unsigned long long)r1_bio->sector,
1708                                                bdevname(rdev->bdev,b));
1709                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1710                                 bio->bi_bdev = rdev->bdev;
1711                                 bio->bi_end_io = raid1_end_read_request;
1712                                 bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
1713                                 bio->bi_private = r1_bio;
1714                                 unplug = 1;
1715                                 generic_make_request(bio);
1716                         }
1717                 }
1718                 cond_resched();
1719         }
1720         if (unplug)
1721                 unplug_slaves(mddev);
1722 }
1723
1724
1725 static int init_resync(conf_t *conf)
1726 {
1727         int buffs;
1728
1729         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1730         BUG_ON(conf->r1buf_pool);
1731         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1732                                           conf->poolinfo);
1733         if (!conf->r1buf_pool)
1734                 return -ENOMEM;
1735         conf->next_resync = 0;
1736         return 0;
1737 }
1738
1739 /*
1740  * perform a "sync" on one "block"
1741  *
1742  * We need to make sure that no normal I/O request - particularly write
1743  * requests - conflict with active sync requests.
1744  *
1745  * This is achieved by tracking pending requests and a 'barrier' concept
1746  * that can be installed to exclude normal IO requests.
1747  */
1748
1749 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1750 {
1751         conf_t *conf = mddev->private;
1752         r1bio_t *r1_bio;
1753         struct bio *bio;
1754         sector_t max_sector, nr_sectors;
1755         int disk = -1;
1756         int i;
1757         int wonly = -1;
1758         int write_targets = 0, read_targets = 0;
1759         int sync_blocks;
1760         int still_degraded = 0;
1761
1762         if (!conf->r1buf_pool)
1763         {
1764 /*
1765                 printk("sync start - bitmap %p\n", mddev->bitmap);
1766 */
1767                 if (init_resync(conf))
1768                         return 0;
1769         }
1770
1771         max_sector = mddev->dev_sectors;
1772         if (sector_nr >= max_sector) {
1773                 /* If we aborted, we need to abort the
1774                  * sync on the 'current' bitmap chunk (there will
1775                  * only be one in raid1 resync.
1776                  * We can find the current addess in mddev->curr_resync
1777                  */
1778                 if (mddev->curr_resync < max_sector) /* aborted */
1779                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1780                                                 &sync_blocks, 1);
1781                 else /* completed sync */
1782                         conf->fullsync = 0;
1783
1784                 bitmap_close_sync(mddev->bitmap);
1785                 close_sync(conf);
1786                 return 0;
1787         }
1788
1789         if (mddev->bitmap == NULL &&
1790             mddev->recovery_cp == MaxSector &&
1791             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1792             conf->fullsync == 0) {
1793                 *skipped = 1;
1794                 return max_sector - sector_nr;
1795         }
1796         /* before building a request, check if we can skip these blocks..
1797          * This call the bitmap_start_sync doesn't actually record anything
1798          */
1799         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1800             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1801                 /* We can skip this block, and probably several more */
1802                 *skipped = 1;
1803                 return sync_blocks;
1804         }
1805         /*
1806          * If there is non-resync activity waiting for a turn,
1807          * and resync is going fast enough,
1808          * then let it though before starting on this new sync request.
1809          */
1810         if (!go_faster && conf->nr_waiting)
1811                 msleep_interruptible(1000);
1812
1813         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1814         raise_barrier(conf);
1815
1816         conf->next_resync = sector_nr;
1817
1818         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1819         rcu_read_lock();
1820         /*
1821          * If we get a correctably read error during resync or recovery,
1822          * we might want to read from a different device.  So we
1823          * flag all drives that could conceivably be read from for READ,
1824          * and any others (which will be non-In_sync devices) for WRITE.
1825          * If a read fails, we try reading from something else for which READ
1826          * is OK.
1827          */
1828
1829         r1_bio->mddev = mddev;
1830         r1_bio->sector = sector_nr;
1831         r1_bio->state = 0;
1832         set_bit(R1BIO_IsSync, &r1_bio->state);
1833
1834         for (i=0; i < conf->raid_disks; i++) {
1835                 mdk_rdev_t *rdev;
1836                 bio = r1_bio->bios[i];
1837
1838                 /* take from bio_init */
1839                 bio->bi_next = NULL;
1840                 bio->bi_flags |= 1 << BIO_UPTODATE;
1841                 bio->bi_rw = READ;
1842                 bio->bi_vcnt = 0;
1843                 bio->bi_idx = 0;
1844                 bio->bi_phys_segments = 0;
1845                 bio->bi_size = 0;
1846                 bio->bi_end_io = NULL;
1847                 bio->bi_private = NULL;
1848
1849                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1850                 if (rdev == NULL ||
1851                            test_bit(Faulty, &rdev->flags)) {
1852                         still_degraded = 1;
1853                         continue;
1854                 } else if (!test_bit(In_sync, &rdev->flags)) {
1855                         bio->bi_rw = WRITE;
1856                         bio->bi_end_io = end_sync_write;
1857                         write_targets ++;
1858                 } else {
1859                         /* may need to read from here */
1860                         bio->bi_rw = READ;
1861                         bio->bi_end_io = end_sync_read;
1862                         if (test_bit(WriteMostly, &rdev->flags)) {
1863                                 if (wonly < 0)
1864                                         wonly = i;
1865                         } else {
1866                                 if (disk < 0)
1867                                         disk = i;
1868                         }
1869                         read_targets++;
1870                 }
1871                 atomic_inc(&rdev->nr_pending);
1872                 bio->bi_sector = sector_nr + rdev->data_offset;
1873                 bio->bi_bdev = rdev->bdev;
1874                 bio->bi_private = r1_bio;
1875         }
1876         rcu_read_unlock();
1877         if (disk < 0)
1878                 disk = wonly;
1879         r1_bio->read_disk = disk;
1880
1881         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1882                 /* extra read targets are also write targets */
1883                 write_targets += read_targets-1;
1884
1885         if (write_targets == 0 || read_targets == 0) {
1886                 /* There is nowhere to write, so all non-sync
1887                  * drives must be failed - so we are finished
1888                  */
1889                 sector_t rv = max_sector - sector_nr;
1890                 *skipped = 1;
1891                 put_buf(r1_bio);
1892                 return rv;
1893         }
1894
1895         if (max_sector > mddev->resync_max)
1896                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1897         nr_sectors = 0;
1898         sync_blocks = 0;
1899         do {
1900                 struct page *page;
1901                 int len = PAGE_SIZE;
1902                 if (sector_nr + (len>>9) > max_sector)
1903                         len = (max_sector - sector_nr) << 9;
1904                 if (len == 0)
1905                         break;
1906                 if (sync_blocks == 0) {
1907                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1908                                                &sync_blocks, still_degraded) &&
1909                             !conf->fullsync &&
1910                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1911                                 break;
1912                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1913                         if (len > (sync_blocks<<9))
1914                                 len = sync_blocks<<9;
1915                 }
1916
1917                 for (i=0 ; i < conf->raid_disks; i++) {
1918                         bio = r1_bio->bios[i];
1919                         if (bio->bi_end_io) {
1920                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1921                                 if (bio_add_page(bio, page, len, 0) == 0) {
1922                                         /* stop here */
1923                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1924                                         while (i > 0) {
1925                                                 i--;
1926                                                 bio = r1_bio->bios[i];
1927                                                 if (bio->bi_end_io==NULL)
1928                                                         continue;
1929                                                 /* remove last page from this bio */
1930                                                 bio->bi_vcnt--;
1931                                                 bio->bi_size -= len;
1932                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1933                                         }
1934                                         goto bio_full;
1935                                 }
1936                         }
1937                 }
1938                 nr_sectors += len>>9;
1939                 sector_nr += len>>9;
1940                 sync_blocks -= (len>>9);
1941         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1942  bio_full:
1943         r1_bio->sectors = nr_sectors;
1944
1945         /* For a user-requested sync, we read all readable devices and do a
1946          * compare
1947          */
1948         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1949                 atomic_set(&r1_bio->remaining, read_targets);
1950                 for (i=0; i<conf->raid_disks; i++) {
1951                         bio = r1_bio->bios[i];
1952                         if (bio->bi_end_io == end_sync_read) {
1953                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1954                                 generic_make_request(bio);
1955                         }
1956                 }
1957         } else {
1958                 atomic_set(&r1_bio->remaining, 1);
1959                 bio = r1_bio->bios[r1_bio->read_disk];
1960                 md_sync_acct(bio->bi_bdev, nr_sectors);
1961                 generic_make_request(bio);
1962
1963         }
1964         return nr_sectors;
1965 }
1966
1967 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1968 {
1969         if (sectors)
1970                 return sectors;
1971
1972         return mddev->dev_sectors;
1973 }
1974
1975 static conf_t *setup_conf(mddev_t *mddev)
1976 {
1977         conf_t *conf;
1978         int i;
1979         mirror_info_t *disk;
1980         mdk_rdev_t *rdev;
1981         int err = -ENOMEM;
1982
1983         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1984         if (!conf)
1985                 goto abort;
1986
1987         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1988                                  GFP_KERNEL);
1989         if (!conf->mirrors)
1990                 goto abort;
1991
1992         conf->tmppage = alloc_page(GFP_KERNEL);
1993         if (!conf->tmppage)
1994                 goto abort;
1995
1996         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1997         if (!conf->poolinfo)
1998                 goto abort;
1999         conf->poolinfo->raid_disks = mddev->raid_disks;
2000         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2001                                           r1bio_pool_free,
2002                                           conf->poolinfo);
2003         if (!conf->r1bio_pool)
2004                 goto abort;
2005
2006         conf->poolinfo->mddev = mddev;
2007
2008         spin_lock_init(&conf->device_lock);
2009         list_for_each_entry(rdev, &mddev->disks, same_set) {
2010                 int disk_idx = rdev->raid_disk;
2011                 if (disk_idx >= mddev->raid_disks
2012                     || disk_idx < 0)
2013                         continue;
2014                 disk = conf->mirrors + disk_idx;
2015
2016                 disk->rdev = rdev;
2017
2018                 disk->head_position = 0;
2019         }
2020         conf->raid_disks = mddev->raid_disks;
2021         conf->mddev = mddev;
2022         INIT_LIST_HEAD(&conf->retry_list);
2023
2024         spin_lock_init(&conf->resync_lock);
2025         init_waitqueue_head(&conf->wait_barrier);
2026
2027         bio_list_init(&conf->pending_bio_list);
2028         bio_list_init(&conf->flushing_bio_list);
2029
2030         conf->last_used = -1;
2031         for (i = 0; i < conf->raid_disks; i++) {
2032
2033                 disk = conf->mirrors + i;
2034
2035                 if (!disk->rdev ||
2036                     !test_bit(In_sync, &disk->rdev->flags)) {
2037                         disk->head_position = 0;
2038                         if (disk->rdev)
2039                                 conf->fullsync = 1;
2040                 } else if (conf->last_used < 0)
2041                         /*
2042                          * The first working device is used as a
2043                          * starting point to read balancing.
2044                          */
2045                         conf->last_used = i;
2046         }
2047
2048         err = -EIO;
2049         if (conf->last_used < 0) {
2050                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
2051                        mdname(mddev));
2052                 goto abort;
2053         }
2054         err = -ENOMEM;
2055         conf->thread = md_register_thread(raid1d, mddev, NULL);
2056         if (!conf->thread) {
2057                 printk(KERN_ERR
2058                        "raid1: couldn't allocate thread for %s\n",
2059                        mdname(mddev));
2060                 goto abort;
2061         }
2062
2063         return conf;
2064
2065  abort:
2066         if (conf) {
2067                 if (conf->r1bio_pool)
2068                         mempool_destroy(conf->r1bio_pool);
2069                 kfree(conf->mirrors);
2070                 safe_put_page(conf->tmppage);
2071                 kfree(conf->poolinfo);
2072                 kfree(conf);
2073         }
2074         return ERR_PTR(err);
2075 }
2076
2077 static int run(mddev_t *mddev)
2078 {
2079         conf_t *conf;
2080         int i;
2081         mdk_rdev_t *rdev;
2082
2083         if (mddev->level != 1) {
2084                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
2085                        mdname(mddev), mddev->level);
2086                 return -EIO;
2087         }
2088         if (mddev->reshape_position != MaxSector) {
2089                 printk("raid1: %s: reshape_position set but not supported\n",
2090                        mdname(mddev));
2091                 return -EIO;
2092         }
2093         /*
2094          * copy the already verified devices into our private RAID1
2095          * bookkeeping area. [whatever we allocate in run(),
2096          * should be freed in stop()]
2097          */
2098         if (mddev->private == NULL)
2099                 conf = setup_conf(mddev);
2100         else
2101                 conf = mddev->private;
2102
2103         if (IS_ERR(conf))
2104                 return PTR_ERR(conf);
2105
2106         mddev->queue->queue_lock = &conf->device_lock;
2107         list_for_each_entry(rdev, &mddev->disks, same_set) {
2108                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2109                                   rdev->data_offset << 9);
2110                 /* as we don't honour merge_bvec_fn, we must never risk
2111                  * violating it, so limit ->max_segments to 1 lying within
2112                  * a single page, as a one page request is never in violation.
2113                  */
2114                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2115                         blk_queue_max_segments(mddev->queue, 1);
2116                         blk_queue_segment_boundary(mddev->queue,
2117                                                    PAGE_CACHE_SIZE - 1);
2118                 }
2119         }
2120
2121         mddev->degraded = 0;
2122         for (i=0; i < conf->raid_disks; i++)
2123                 if (conf->mirrors[i].rdev == NULL ||
2124                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2125                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2126                         mddev->degraded++;
2127
2128         if (conf->raid_disks - mddev->degraded == 1)
2129                 mddev->recovery_cp = MaxSector;
2130
2131         if (mddev->recovery_cp != MaxSector)
2132                 printk(KERN_NOTICE "raid1: %s is not clean"
2133                        " -- starting background reconstruction\n",
2134                        mdname(mddev));
2135         printk(KERN_INFO
2136                 "raid1: raid set %s active with %d out of %d mirrors\n",
2137                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2138                 mddev->raid_disks);
2139
2140         /*
2141          * Ok, everything is just fine now
2142          */
2143         mddev->thread = conf->thread;
2144         conf->thread = NULL;
2145         mddev->private = conf;
2146
2147         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2148
2149         mddev->queue->unplug_fn = raid1_unplug;
2150         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2151         mddev->queue->backing_dev_info.congested_data = mddev;
2152         md_integrity_register(mddev);
2153         return 0;
2154 }
2155
2156 static int stop(mddev_t *mddev)
2157 {
2158         conf_t *conf = mddev->private;
2159         struct bitmap *bitmap = mddev->bitmap;
2160
2161         /* wait for behind writes to complete */
2162         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2163                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop.\n", mdname(mddev));
2164                 /* need to kick something here to make sure I/O goes? */
2165                 wait_event(bitmap->behind_wait,
2166                            atomic_read(&bitmap->behind_writes) == 0);
2167         }
2168
2169         raise_barrier(conf);
2170         lower_barrier(conf);
2171
2172         md_unregister_thread(mddev->thread);
2173         mddev->thread = NULL;
2174         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2175         if (conf->r1bio_pool)
2176                 mempool_destroy(conf->r1bio_pool);
2177         kfree(conf->mirrors);
2178         kfree(conf->poolinfo);
2179         kfree(conf);
2180         mddev->private = NULL;
2181         return 0;
2182 }
2183
2184 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2185 {
2186         /* no resync is happening, and there is enough space
2187          * on all devices, so we can resize.
2188          * We need to make sure resync covers any new space.
2189          * If the array is shrinking we should possibly wait until
2190          * any io in the removed space completes, but it hardly seems
2191          * worth it.
2192          */
2193         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2194         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2195                 return -EINVAL;
2196         set_capacity(mddev->gendisk, mddev->array_sectors);
2197         revalidate_disk(mddev->gendisk);
2198         if (sectors > mddev->dev_sectors &&
2199             mddev->recovery_cp == MaxSector) {
2200                 mddev->recovery_cp = mddev->dev_sectors;
2201                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2202         }
2203         mddev->dev_sectors = sectors;
2204         mddev->resync_max_sectors = sectors;
2205         return 0;
2206 }
2207
2208 static int raid1_reshape(mddev_t *mddev)
2209 {
2210         /* We need to:
2211          * 1/ resize the r1bio_pool
2212          * 2/ resize conf->mirrors
2213          *
2214          * We allocate a new r1bio_pool if we can.
2215          * Then raise a device barrier and wait until all IO stops.
2216          * Then resize conf->mirrors and swap in the new r1bio pool.
2217          *
2218          * At the same time, we "pack" the devices so that all the missing
2219          * devices have the higher raid_disk numbers.
2220          */
2221         mempool_t *newpool, *oldpool;
2222         struct pool_info *newpoolinfo;
2223         mirror_info_t *newmirrors;
2224         conf_t *conf = mddev->private;
2225         int cnt, raid_disks;
2226         unsigned long flags;
2227         int d, d2, err;
2228
2229         /* Cannot change chunk_size, layout, or level */
2230         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2231             mddev->layout != mddev->new_layout ||
2232             mddev->level != mddev->new_level) {
2233                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2234                 mddev->new_layout = mddev->layout;
2235                 mddev->new_level = mddev->level;
2236                 return -EINVAL;
2237         }
2238
2239         err = md_allow_write(mddev);
2240         if (err)
2241                 return err;
2242
2243         raid_disks = mddev->raid_disks + mddev->delta_disks;
2244
2245         if (raid_disks < conf->raid_disks) {
2246                 cnt=0;
2247                 for (d= 0; d < conf->raid_disks; d++)
2248                         if (conf->mirrors[d].rdev)
2249                                 cnt++;
2250                 if (cnt > raid_disks)
2251                         return -EBUSY;
2252         }
2253
2254         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2255         if (!newpoolinfo)
2256                 return -ENOMEM;
2257         newpoolinfo->mddev = mddev;
2258         newpoolinfo->raid_disks = raid_disks;
2259
2260         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2261                                  r1bio_pool_free, newpoolinfo);
2262         if (!newpool) {
2263                 kfree(newpoolinfo);
2264                 return -ENOMEM;
2265         }
2266         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2267         if (!newmirrors) {
2268                 kfree(newpoolinfo);
2269                 mempool_destroy(newpool);
2270                 return -ENOMEM;
2271         }
2272
2273         raise_barrier(conf);
2274
2275         /* ok, everything is stopped */
2276         oldpool = conf->r1bio_pool;
2277         conf->r1bio_pool = newpool;
2278
2279         for (d = d2 = 0; d < conf->raid_disks; d++) {
2280                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2281                 if (rdev && rdev->raid_disk != d2) {
2282                         char nm[20];
2283                         sprintf(nm, "rd%d", rdev->raid_disk);
2284                         sysfs_remove_link(&mddev->kobj, nm);
2285                         rdev->raid_disk = d2;
2286                         sprintf(nm, "rd%d", rdev->raid_disk);
2287                         sysfs_remove_link(&mddev->kobj, nm);
2288                         if (sysfs_create_link(&mddev->kobj,
2289                                               &rdev->kobj, nm))
2290                                 printk(KERN_WARNING
2291                                        "md/raid1: cannot register "
2292                                        "%s for %s\n",
2293                                        nm, mdname(mddev));
2294                 }
2295                 if (rdev)
2296                         newmirrors[d2++].rdev = rdev;
2297         }
2298         kfree(conf->mirrors);
2299         conf->mirrors = newmirrors;
2300         kfree(conf->poolinfo);
2301         conf->poolinfo = newpoolinfo;
2302
2303         spin_lock_irqsave(&conf->device_lock, flags);
2304         mddev->degraded += (raid_disks - conf->raid_disks);
2305         spin_unlock_irqrestore(&conf->device_lock, flags);
2306         conf->raid_disks = mddev->raid_disks = raid_disks;
2307         mddev->delta_disks = 0;
2308
2309         conf->last_used = 0; /* just make sure it is in-range */
2310         lower_barrier(conf);
2311
2312         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2313         md_wakeup_thread(mddev->thread);
2314
2315         mempool_destroy(oldpool);
2316         return 0;
2317 }
2318
2319 static void raid1_quiesce(mddev_t *mddev, int state)
2320 {
2321         conf_t *conf = mddev->private;
2322
2323         switch(state) {
2324         case 2: /* wake for suspend */
2325                 wake_up(&conf->wait_barrier);
2326                 break;
2327         case 1:
2328                 raise_barrier(conf);
2329                 break;
2330         case 0:
2331                 lower_barrier(conf);
2332                 break;
2333         }
2334 }
2335
2336 static void *raid1_takeover(mddev_t *mddev)
2337 {
2338         /* raid1 can take over:
2339          *  raid5 with 2 devices, any layout or chunk size
2340          */
2341         if (mddev->level == 5 && mddev->raid_disks == 2) {
2342                 conf_t *conf;
2343                 mddev->new_level = 1;
2344                 mddev->new_layout = 0;
2345                 mddev->new_chunk_sectors = 0;
2346                 conf = setup_conf(mddev);
2347                 if (!IS_ERR(conf))
2348                         conf->barrier = 1;
2349                 return conf;
2350         }
2351         return ERR_PTR(-EINVAL);
2352 }
2353
2354 static struct mdk_personality raid1_personality =
2355 {
2356         .name           = "raid1",
2357         .level          = 1,
2358         .owner          = THIS_MODULE,
2359         .make_request   = make_request,
2360         .run            = run,
2361         .stop           = stop,
2362         .status         = status,
2363         .error_handler  = error,
2364         .hot_add_disk   = raid1_add_disk,
2365         .hot_remove_disk= raid1_remove_disk,
2366         .spare_active   = raid1_spare_active,
2367         .sync_request   = sync_request,
2368         .resize         = raid1_resize,
2369         .size           = raid1_size,
2370         .check_reshape  = raid1_reshape,
2371         .quiesce        = raid1_quiesce,
2372         .takeover       = raid1_takeover,
2373 };
2374
2375 static int __init raid_init(void)
2376 {
2377         return register_md_personality(&raid1_personality);
2378 }
2379
2380 static void raid_exit(void)
2381 {
2382         unregister_md_personality(&raid1_personality);
2383 }
2384
2385 module_init(raid_init);
2386 module_exit(raid_exit);
2387 MODULE_LICENSE("GPL");
2388 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2389 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2390 MODULE_ALIAS("md-raid1");
2391 MODULE_ALIAS("md-level-1");