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 = (r1bio_t *)(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 = (r1bio_t *)(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         spin_lock_irqsave(&conf->resync_lock, flags);
 681         conf->barrier--;
 682         spin_unlock_irqrestore(&conf->resync_lock, flags);
 683         wake_up(&conf->wait_barrier);
 684 }
 685
 686 static void wait_barrier(conf_t *conf)
 687 {
 688         spin_lock_irq(&conf->resync_lock);
 689         if (conf->barrier) {
 690                 conf->nr_waiting++;
 691                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 692                                     conf->resync_lock,
 693                                     raid1_unplug(conf->mddev->queue));
 694                 conf->nr_waiting--;
 695         }
 696         conf->nr_pending++;
 697         spin_unlock_irq(&conf->resync_lock);
 698 }
 699
 700 static void allow_barrier(conf_t *conf)
 701 {
 702         unsigned long flags;
 703         spin_lock_irqsave(&conf->resync_lock, flags);
 704         conf->nr_pending--;
 705         spin_unlock_irqrestore(&conf->resync_lock, flags);
 706         wake_up(&conf->wait_barrier);
 707 }
 708
 709 static void freeze_array(conf_t *conf)
 710 {
 711         /* stop syncio and normal IO and wait for everything to
 712          * go quite.
 713          * We increment barrier and nr_waiting, and then
 714          * wait until nr_pending match nr_queued+1
 715          * This is called in the context of one normal IO request
 716          * that has failed. Thus any sync request that might be pending
 717          * will be blocked by nr_pending, and we need to wait for
 718          * pending IO requests to complete or be queued for re-try.
 719          * Thus the number queued (nr_queued) plus this request (1)
 720          * must match the number of pending IOs (nr_pending) before
 721          * we continue.
 722          */
 723         spin_lock_irq(&conf->resync_lock);
 724         conf->barrier++;
 725         conf->nr_waiting++;
 726         wait_event_lock_irq(conf->wait_barrier,
 727                             conf->nr_pending == conf->nr_queued+1,
 728                             conf->resync_lock,
 729                             ({ flush_pending_writes(conf);
 730                                raid1_unplug(conf->mddev->queue); }));
 731         spin_unlock_irq(&conf->resync_lock);
 732 }
 733 static void unfreeze_array(conf_t *conf)
 734 {
 735         /* reverse the effect of the freeze */
 736         spin_lock_irq(&conf->resync_lock);
 737         conf->barrier--;
 738         conf->nr_waiting--;
 739         wake_up(&conf->wait_barrier);
 740         spin_unlock_irq(&conf->resync_lock);
 741 }
 742
 743
 744 /* duplicate the data pages for behind I/O */
 745 static struct page **alloc_behind_pages(struct bio *bio)
 746 {
 747         int i;
 748         struct bio_vec *bvec;
 749         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
 750                                         GFP_NOIO);
 751         if (unlikely(!pages))
 752                 goto do_sync_io;
 753
 754         bio_for_each_segment(bvec, bio, i) {
 755                 pages[i] = alloc_page(GFP_NOIO);
 756                 if (unlikely(!pages[i]))
 757                         goto do_sync_io;
 758                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 759                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 760                 kunmap(pages[i]);
 761                 kunmap(bvec->bv_page);
 762         }
 763
 764         return pages;
 765
 766 do_sync_io:
 767         if (pages)
 768                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
 769                         put_page(pages[i]);
 770         kfree(pages);
 771         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 772         return NULL;
 773 }
 774
 775 static int make_request(struct request_queue *q, struct bio * bio)
 776 {
 777         mddev_t *mddev = q->queuedata;
 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         int cpu;
 790         bool do_barriers;
 791         mdk_rdev_t *blocked_rdev;
 792
 793         /*
 794          * Register the new request and wait if the reconstruction
 795          * thread has put up a bar for new requests.
 796          * Continue immediately if no resync is active currently.
 797          * We test barriers_work *after* md_write_start as md_write_start
 798          * may cause the first superblock write, and that will check out
 799          * if barriers work.
 800          */
 801
 802         md_write_start(mddev, bio); /* wait on superblock update early */
 803
 804         if (unlikely(!mddev->barriers_work &&
 805                      bio_rw_flagged(bio, BIO_RW_BARRIER))) {
 806                 if (rw == WRITE)
 807                         md_write_end(mddev);
 808                 bio_endio(bio, -EOPNOTSUPP);
 809                 return 0;
 810         }
 811
 812         wait_barrier(conf);
 813
 814         bitmap = mddev->bitmap;
 815
 816         cpu = part_stat_lock();
 817         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
 818         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
 819                       bio_sectors(bio));
 820         part_stat_unlock();
 821
 822         /*
 823          * make_request() can abort the operation when READA is being
 824          * used and no empty request is available.
 825          *
 826          */
 827         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 828
 829         r1_bio->master_bio = bio;
 830         r1_bio->sectors = bio->bi_size >> 9;
 831         r1_bio->state = 0;
 832         r1_bio->mddev = mddev;
 833         r1_bio->sector = bio->bi_sector;
 834
 835         if (rw == READ) {
 836                 /*
 837                  * read balancing logic:
 838                  */
 839                 int rdisk = read_balance(conf, r1_bio);
 840
 841                 if (rdisk < 0) {
 842                         /* couldn't find anywhere to read from */
 843                         raid_end_bio_io(r1_bio);
 844                         return 0;
 845                 }
 846                 mirror = conf->mirrors + rdisk;
 847
 848                 r1_bio->read_disk = rdisk;
 849
 850                 read_bio = bio_clone(bio, GFP_NOIO);
 851
 852                 r1_bio->bios[rdisk] = read_bio;
 853
 854                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 855                 read_bio->bi_bdev = mirror->rdev->bdev;
 856                 read_bio->bi_end_io = raid1_end_read_request;
 857                 read_bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
 858                 read_bio->bi_private = r1_bio;
 859
 860                 generic_make_request(read_bio);
 861                 return 0;
 862         }
 863
 864         /*
 865          * WRITE:
 866          */
 867         /* first select target devices under spinlock and
 868          * inc refcount on their rdev.  Record them by setting
 869          * bios[x] to bio
 870          */
 871         disks = conf->raid_disks;
 872 #if 0
 873         { static int first=1;
 874         if (first) printk("First Write sector %llu disks %d\n",
 875                           (unsigned long long)r1_bio->sector, disks);
 876         first = 0;
 877         }
 878 #endif
 879  retry_write:
 880         blocked_rdev = NULL;
 881         rcu_read_lock();
 882         for (i = 0;  i < disks; i++) {
 883                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 884                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 885                         atomic_inc(&rdev->nr_pending);
 886                         blocked_rdev = rdev;
 887                         break;
 888                 }
 889                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 890                         atomic_inc(&rdev->nr_pending);
 891                         if (test_bit(Faulty, &rdev->flags)) {
 892                                 rdev_dec_pending(rdev, mddev);
 893                                 r1_bio->bios[i] = NULL;
 894                         } else
 895                                 r1_bio->bios[i] = bio;
 896                         targets++;
 897                 } else
 898                         r1_bio->bios[i] = NULL;
 899         }
 900         rcu_read_unlock();
 901
 902         if (unlikely(blocked_rdev)) {
 903                 /* Wait for this device to become unblocked */
 904                 int j;
 905
 906                 for (j = 0; j < i; j++)
 907                         if (r1_bio->bios[j])
 908                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
 909
 910                 allow_barrier(conf);
 911                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 912                 wait_barrier(conf);
 913                 goto retry_write;
 914         }
 915
 916         BUG_ON(targets == 0); /* we never fail the last device */
 917
 918         if (targets < conf->raid_disks) {
 919                 /* array is degraded, we will not clear the bitmap
 920                  * on I/O completion (see raid1_end_write_request) */
 921                 set_bit(R1BIO_Degraded, &r1_bio->state);
 922         }
 923
 924         /* do behind I/O ? */
 925         if (bitmap &&
 926             atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
 927             (behind_pages = alloc_behind_pages(bio)) != NULL)
 928                 set_bit(R1BIO_BehindIO, &r1_bio->state);
 929
 930         atomic_set(&r1_bio->remaining, 0);
 931         atomic_set(&r1_bio->behind_remaining, 0);
 932
 933         do_barriers = bio_rw_flagged(bio, BIO_RW_BARRIER);
 934         if (do_barriers)
 935                 set_bit(R1BIO_Barrier, &r1_bio->state);
 936
 937         bio_list_init(&bl);
 938         for (i = 0; i < disks; i++) {
 939                 struct bio *mbio;
 940                 if (!r1_bio->bios[i])
 941                         continue;
 942
 943                 mbio = bio_clone(bio, GFP_NOIO);
 944                 r1_bio->bios[i] = mbio;
 945
 946                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 947                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 948                 mbio->bi_end_io = raid1_end_write_request;
 949                 mbio->bi_rw = WRITE | (do_barriers << BIO_RW_BARRIER) |
 950                         (do_sync << BIO_RW_SYNCIO);
 951                 mbio->bi_private = r1_bio;
 952
 953                 if (behind_pages) {
 954                         struct bio_vec *bvec;
 955                         int j;
 956
 957                         /* Yes, I really want the '__' version so that
 958                          * we clear any unused pointer in the io_vec, rather
 959                          * than leave them unchanged.  This is important
 960                          * because when we come to free the pages, we won't
 961                          * know the originial bi_idx, so we just free
 962                          * them all
 963                          */
 964                         __bio_for_each_segment(bvec, mbio, j, 0)
 965                                 bvec->bv_page = behind_pages[j];
 966                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 967                                 atomic_inc(&r1_bio->behind_remaining);
 968                 }
 969
 970                 atomic_inc(&r1_bio->remaining);
 971
 972                 bio_list_add(&bl, mbio);
 973         }
 974         kfree(behind_pages); /* the behind pages are attached to the bios now */
 975
 976         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
 977                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 978         spin_lock_irqsave(&conf->device_lock, flags);
 979         bio_list_merge(&conf->pending_bio_list, &bl);
 980         bio_list_init(&bl);
 981
 982         blk_plug_device(mddev->queue);
 983         spin_unlock_irqrestore(&conf->device_lock, flags);
 984
 985         /* In case raid1d snuck into freeze_array */
 986         wake_up(&conf->wait_barrier);
 987
 988         if (do_sync)
 989                 md_wakeup_thread(mddev->thread);
 990 #if 0
 991         while ((bio = bio_list_pop(&bl)) != NULL)
 992                 generic_make_request(bio);
 993 #endif
 994
 995         return 0;
 996 }
 997
 998 static void status(struct seq_file *seq, mddev_t *mddev)
 999 {
1000         conf_t *conf = mddev->private;
1001         int i;
1002
1003         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1004                    conf->raid_disks - mddev->degraded);
1005         rcu_read_lock();
1006         for (i = 0; i < conf->raid_disks; i++) {
1007                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1008                 seq_printf(seq, "%s",
1009                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1010         }
1011         rcu_read_unlock();
1012         seq_printf(seq, "]");
1013 }
1014
1015
1016 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1017 {
1018         char b[BDEVNAME_SIZE];
1019         conf_t *conf = mddev->private;
1020
1021         /*
1022          * If it is not operational, then we have already marked it as dead
1023          * else if it is the last working disks, ignore the error, let the
1024          * next level up know.
1025          * else mark the drive as failed
1026          */
1027         if (test_bit(In_sync, &rdev->flags)
1028             && (conf->raid_disks - mddev->degraded) == 1) {
1029                 /*
1030                  * Don't fail the drive, act as though we were just a
1031                  * normal single drive.
1032                  * However don't try a recovery from this drive as
1033                  * it is very likely to fail.
1034                  */
1035                 mddev->recovery_disabled = 1;
1036                 return;
1037         }
1038         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1039                 unsigned long flags;
1040                 spin_lock_irqsave(&conf->device_lock, flags);
1041                 mddev->degraded++;
1042                 set_bit(Faulty, &rdev->flags);
1043                 spin_unlock_irqrestore(&conf->device_lock, flags);
1044                 /*
1045                  * if recovery is running, make sure it aborts.
1046                  */
1047                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1048         } else
1049                 set_bit(Faulty, &rdev->flags);
1050         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1051         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device.\n"
1052                 "raid1: Operation continuing on %d devices.\n",
1053                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1054 }
1055
1056 static void print_conf(conf_t *conf)
1057 {
1058         int i;
1059
1060         printk("RAID1 conf printout:\n");
1061         if (!conf) {
1062                 printk("(!conf)\n");
1063                 return;
1064         }
1065         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1066                 conf->raid_disks);
1067
1068         rcu_read_lock();
1069         for (i = 0; i < conf->raid_disks; i++) {
1070                 char b[BDEVNAME_SIZE];
1071                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1072                 if (rdev)
1073                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1074                                i, !test_bit(In_sync, &rdev->flags),
1075                                !test_bit(Faulty, &rdev->flags),
1076                                bdevname(rdev->bdev,b));
1077         }
1078         rcu_read_unlock();
1079 }
1080
1081 static void close_sync(conf_t *conf)
1082 {
1083         wait_barrier(conf);
1084         allow_barrier(conf);
1085
1086         mempool_destroy(conf->r1buf_pool);
1087         conf->r1buf_pool = NULL;
1088 }
1089
1090 static int raid1_spare_active(mddev_t *mddev)
1091 {
1092         int i;
1093         conf_t *conf = mddev->private;
1094
1095         /*
1096          * Find all failed disks within the RAID1 configuration
1097          * and mark them readable.
1098          * Called under mddev lock, so rcu protection not needed.
1099          */
1100         for (i = 0; i < conf->raid_disks; i++) {
1101                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1102                 if (rdev
1103                     && !test_bit(Faulty, &rdev->flags)
1104                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1105                         unsigned long flags;
1106                         spin_lock_irqsave(&conf->device_lock, flags);
1107                         mddev->degraded--;
1108                         spin_unlock_irqrestore(&conf->device_lock, flags);
1109                 }
1110         }
1111
1112         print_conf(conf);
1113         return 0;
1114 }
1115
1116
1117 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1118 {
1119         conf_t *conf = mddev->private;
1120         int err = -EEXIST;
1121         int mirror = 0;
1122         mirror_info_t *p;
1123         int first = 0;
1124         int last = mddev->raid_disks - 1;
1125
1126         if (rdev->raid_disk >= 0)
1127                 first = last = rdev->raid_disk;
1128
1129         for (mirror = first; mirror <= last; mirror++)
1130                 if ( !(p=conf->mirrors+mirror)->rdev) {
1131
1132                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1133                                           rdev->data_offset << 9);
1134                         /* as we don't honour merge_bvec_fn, we must never risk
1135                          * violating it, so limit ->max_sector to one PAGE, as
1136                          * a one page request is never in violation.
1137                          */
1138                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1139                             queue_max_sectors(mddev->queue) > (PAGE_SIZE>>9))
1140                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1141
1142                         p->head_position = 0;
1143                         rdev->raid_disk = mirror;
1144                         err = 0;
1145                         /* As all devices are equivalent, we don't need a full recovery
1146                          * if this was recently any drive of the array
1147                          */
1148                         if (rdev->saved_raid_disk < 0)
1149                                 conf->fullsync = 1;
1150                         rcu_assign_pointer(p->rdev, rdev);
1151                         break;
1152                 }
1153         md_integrity_add_rdev(rdev, mddev);
1154         print_conf(conf);
1155         return err;
1156 }
1157
1158 static int raid1_remove_disk(mddev_t *mddev, int number)
1159 {
1160         conf_t *conf = mddev->private;
1161         int err = 0;
1162         mdk_rdev_t *rdev;
1163         mirror_info_t *p = conf->mirrors+ number;
1164
1165         print_conf(conf);
1166         rdev = p->rdev;
1167         if (rdev) {
1168                 if (test_bit(In_sync, &rdev->flags) ||
1169                     atomic_read(&rdev->nr_pending)) {
1170                         err = -EBUSY;
1171                         goto abort;
1172                 }
1173                 /* Only remove non-faulty devices is recovery
1174                  * is not possible.
1175                  */
1176                 if (!test_bit(Faulty, &rdev->flags) &&
1177                     mddev->degraded < conf->raid_disks) {
1178                         err = -EBUSY;
1179                         goto abort;
1180                 }
1181                 p->rdev = NULL;
1182                 synchronize_rcu();
1183                 if (atomic_read(&rdev->nr_pending)) {
1184                         /* lost the race, try later */
1185                         err = -EBUSY;
1186                         p->rdev = rdev;
1187                         goto abort;
1188                 }
1189                 md_integrity_register(mddev);
1190         }
1191 abort:
1192
1193         print_conf(conf);
1194         return err;
1195 }
1196
1197
1198 static void end_sync_read(struct bio *bio, int error)
1199 {
1200         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1201         int i;
1202
1203         for (i=r1_bio->mddev->raid_disks; i--; )
1204                 if (r1_bio->bios[i] == bio)
1205                         break;
1206         BUG_ON(i < 0);
1207         update_head_pos(i, r1_bio);
1208         /*
1209          * we have read a block, now it needs to be re-written,
1210          * or re-read if the read failed.
1211          * We don't do much here, just schedule handling by raid1d
1212          */
1213         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1214                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1215
1216         if (atomic_dec_and_test(&r1_bio->remaining))
1217                 reschedule_retry(r1_bio);
1218 }
1219
1220 static void end_sync_write(struct bio *bio, int error)
1221 {
1222         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1223         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1224         mddev_t *mddev = r1_bio->mddev;
1225         conf_t *conf = mddev->private;
1226         int i;
1227         int mirror=0;
1228
1229         for (i = 0; i < conf->raid_disks; i++)
1230                 if (r1_bio->bios[i] == bio) {
1231                         mirror = i;
1232                         break;
1233                 }
1234         if (!uptodate) {
1235                 int sync_blocks = 0;
1236                 sector_t s = r1_bio->sector;
1237                 long sectors_to_go = r1_bio->sectors;
1238                 /* make sure these bits doesn't get cleared. */
1239                 do {
1240                         bitmap_end_sync(mddev->bitmap, s,
1241                                         &sync_blocks, 1);
1242                         s += sync_blocks;
1243                         sectors_to_go -= sync_blocks;
1244                 } while (sectors_to_go > 0);
1245                 md_error(mddev, conf->mirrors[mirror].rdev);
1246         }
1247
1248         update_head_pos(mirror, r1_bio);
1249
1250         if (atomic_dec_and_test(&r1_bio->remaining)) {
1251                 sector_t s = r1_bio->sectors;
1252                 put_buf(r1_bio);
1253                 md_done_sync(mddev, s, uptodate);
1254         }
1255 }
1256
1257 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1258 {
1259         conf_t *conf = mddev->private;
1260         int i;
1261         int disks = conf->raid_disks;
1262         struct bio *bio, *wbio;
1263
1264         bio = r1_bio->bios[r1_bio->read_disk];
1265
1266
1267         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1268                 /* We have read all readable devices.  If we haven't
1269                  * got the block, then there is no hope left.
1270                  * If we have, then we want to do a comparison
1271                  * and skip the write if everything is the same.
1272                  * If any blocks failed to read, then we need to
1273                  * attempt an over-write
1274                  */
1275                 int primary;
1276                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1277                         for (i=0; i<mddev->raid_disks; i++)
1278                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1279                                         md_error(mddev, conf->mirrors[i].rdev);
1280
1281                         md_done_sync(mddev, r1_bio->sectors, 1);
1282                         put_buf(r1_bio);
1283                         return;
1284                 }
1285                 for (primary=0; primary<mddev->raid_disks; primary++)
1286                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1287                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1288                                 r1_bio->bios[primary]->bi_end_io = NULL;
1289                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1290                                 break;
1291                         }
1292                 r1_bio->read_disk = primary;
1293                 for (i=0; i<mddev->raid_disks; i++)
1294                         if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1295                                 int j;
1296                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1297                                 struct bio *pbio = r1_bio->bios[primary];
1298                                 struct bio *sbio = r1_bio->bios[i];
1299
1300                                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1301                                         for (j = vcnt; j-- ; ) {
1302                                                 struct page *p, *s;
1303                                                 p = pbio->bi_io_vec[j].bv_page;
1304                                                 s = sbio->bi_io_vec[j].bv_page;
1305                                                 if (memcmp(page_address(p),
1306                                                            page_address(s),
1307                                                            PAGE_SIZE))
1308                                                         break;
1309                                         }
1310                                 } else
1311                                         j = 0;
1312                                 if (j >= 0)
1313                                         mddev->resync_mismatches += r1_bio->sectors;
1314                                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1315                                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1316                                         sbio->bi_end_io = NULL;
1317                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1318                                 } else {
1319                                         /* fixup the bio for reuse */
1320                                         int size;
1321                                         sbio->bi_vcnt = vcnt;
1322                                         sbio->bi_size = r1_bio->sectors << 9;
1323                                         sbio->bi_idx = 0;
1324                                         sbio->bi_phys_segments = 0;
1325                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1326                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1327                                         sbio->bi_next = NULL;
1328                                         sbio->bi_sector = r1_bio->sector +
1329                                                 conf->mirrors[i].rdev->data_offset;
1330                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1331                                         size = sbio->bi_size;
1332                                         for (j = 0; j < vcnt ; j++) {
1333                                                 struct bio_vec *bi;
1334                                                 bi = &sbio->bi_io_vec[j];
1335                                                 bi->bv_offset = 0;
1336                                                 if (size > PAGE_SIZE)
1337                                                         bi->bv_len = PAGE_SIZE;
1338                                                 else
1339                                                         bi->bv_len = size;
1340                                                 size -= PAGE_SIZE;
1341                                                 memcpy(page_address(bi->bv_page),
1342                                                        page_address(pbio->bi_io_vec[j].bv_page),
1343                                                        PAGE_SIZE);
1344                                         }
1345
1346                                 }
1347                         }
1348         }
1349         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1350                 /* ouch - failed to read all of that.
1351                  * Try some synchronous reads of other devices to get
1352                  * good data, much like with normal read errors.  Only
1353                  * read into the pages we already have so we don't
1354                  * need to re-issue the read request.
1355                  * We don't need to freeze the array, because being in an
1356                  * active sync request, there is no normal IO, and
1357                  * no overlapping syncs.
1358                  */
1359                 sector_t sect = r1_bio->sector;
1360                 int sectors = r1_bio->sectors;
1361                 int idx = 0;
1362
1363                 while(sectors) {
1364                         int s = sectors;
1365                         int d = r1_bio->read_disk;
1366                         int success = 0;
1367                         mdk_rdev_t *rdev;
1368
1369                         if (s > (PAGE_SIZE>>9))
1370                                 s = PAGE_SIZE >> 9;
1371                         do {
1372                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1373                                         /* No rcu protection needed here devices
1374                                          * can only be removed when no resync is
1375                                          * active, and resync is currently active
1376                                          */
1377                                         rdev = conf->mirrors[d].rdev;
1378                                         if (sync_page_io(rdev->bdev,
1379                                                          sect + rdev->data_offset,
1380                                                          s<<9,
1381                                                          bio->bi_io_vec[idx].bv_page,
1382                                                          READ)) {
1383                                                 success = 1;
1384                                                 break;
1385                                         }
1386                                 }
1387                                 d++;
1388                                 if (d == conf->raid_disks)
1389                                         d = 0;
1390                         } while (!success && d != r1_bio->read_disk);
1391
1392                         if (success) {
1393                                 int start = d;
1394                                 /* write it back and re-read */
1395                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1396                                 while (d != r1_bio->read_disk) {
1397                                         if (d == 0)
1398                                                 d = conf->raid_disks;
1399                                         d--;
1400                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1401                                                 continue;
1402                                         rdev = conf->mirrors[d].rdev;
1403                                         atomic_add(s, &rdev->corrected_errors);
1404                                         if (sync_page_io(rdev->bdev,
1405                                                          sect + rdev->data_offset,
1406                                                          s<<9,
1407                                                          bio->bi_io_vec[idx].bv_page,
1408                                                          WRITE) == 0)
1409                                                 md_error(mddev, rdev);
1410                                 }
1411                                 d = start;
1412                                 while (d != r1_bio->read_disk) {
1413                                         if (d == 0)
1414                                                 d = conf->raid_disks;
1415                                         d--;
1416                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1417                                                 continue;
1418                                         rdev = conf->mirrors[d].rdev;
1419                                         if (sync_page_io(rdev->bdev,
1420                                                          sect + rdev->data_offset,
1421                                                          s<<9,
1422                                                          bio->bi_io_vec[idx].bv_page,
1423                                                          READ) == 0)
1424                                                 md_error(mddev, rdev);
1425                                 }
1426                         } else {
1427                                 char b[BDEVNAME_SIZE];
1428                                 /* Cannot read from anywhere, array is toast */
1429                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1430                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1431                                        " for block %llu\n",
1432                                        bdevname(bio->bi_bdev,b),
1433                                        (unsigned long long)r1_bio->sector);
1434                                 md_done_sync(mddev, r1_bio->sectors, 0);
1435                                 put_buf(r1_bio);
1436                                 return;
1437                         }
1438                         sectors -= s;
1439                         sect += s;
1440                         idx ++;
1441                 }
1442         }
1443
1444         /*
1445          * schedule writes
1446          */
1447         atomic_set(&r1_bio->remaining, 1);
1448         for (i = 0; i < disks ; i++) {
1449                 wbio = r1_bio->bios[i];
1450                 if (wbio->bi_end_io == NULL ||
1451                     (wbio->bi_end_io == end_sync_read &&
1452                      (i == r1_bio->read_disk ||
1453                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1454                         continue;
1455
1456                 wbio->bi_rw = WRITE;
1457                 wbio->bi_end_io = end_sync_write;
1458                 atomic_inc(&r1_bio->remaining);
1459                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1460
1461                 generic_make_request(wbio);
1462         }
1463
1464         if (atomic_dec_and_test(&r1_bio->remaining)) {
1465                 /* if we're here, all write(s) have completed, so clean up */
1466                 md_done_sync(mddev, r1_bio->sectors, 1);
1467                 put_buf(r1_bio);
1468         }
1469 }
1470
1471 /*
1472  * This is a kernel thread which:
1473  *
1474  *      1.      Retries failed read operations on working mirrors.
1475  *      2.      Updates the raid superblock when problems encounter.
1476  *      3.      Performs writes following reads for array syncronising.
1477  */
1478
1479 static void fix_read_error(conf_t *conf, int read_disk,
1480                            sector_t sect, int sectors)
1481 {
1482         mddev_t *mddev = conf->mddev;
1483         while(sectors) {
1484                 int s = sectors;
1485                 int d = read_disk;
1486                 int success = 0;
1487                 int start;
1488                 mdk_rdev_t *rdev;
1489
1490                 if (s > (PAGE_SIZE>>9))
1491                         s = PAGE_SIZE >> 9;
1492
1493                 do {
1494                         /* Note: no rcu protection needed here
1495                          * as this is synchronous in the raid1d thread
1496                          * which is the thread that might remove
1497                          * a device.  If raid1d ever becomes multi-threaded....
1498                          */
1499                         rdev = conf->mirrors[d].rdev;
1500                         if (rdev &&
1501                             test_bit(In_sync, &rdev->flags) &&
1502                             sync_page_io(rdev->bdev,
1503                                          sect + rdev->data_offset,
1504                                          s<<9,
1505                                          conf->tmppage, READ))
1506                                 success = 1;
1507                         else {
1508                                 d++;
1509                                 if (d == conf->raid_disks)
1510                                         d = 0;
1511                         }
1512                 } while (!success && d != read_disk);
1513
1514                 if (!success) {
1515                         /* Cannot read from anywhere -- bye bye array */
1516                         md_error(mddev, conf->mirrors[read_disk].rdev);
1517                         break;
1518                 }
1519                 /* write it back and re-read */
1520                 start = d;
1521                 while (d != read_disk) {
1522                         if (d==0)
1523                                 d = conf->raid_disks;
1524                         d--;
1525                         rdev = conf->mirrors[d].rdev;
1526                         if (rdev &&
1527                             test_bit(In_sync, &rdev->flags)) {
1528                                 if (sync_page_io(rdev->bdev,
1529                                                  sect + rdev->data_offset,
1530                                                  s<<9, conf->tmppage, WRITE)
1531                                     == 0)
1532                                         /* Well, this device is dead */
1533                                         md_error(mddev, rdev);
1534                         }
1535                 }
1536                 d = start;
1537                 while (d != read_disk) {
1538                         char b[BDEVNAME_SIZE];
1539                         if (d==0)
1540                                 d = conf->raid_disks;
1541                         d--;
1542                         rdev = conf->mirrors[d].rdev;
1543                         if (rdev &&
1544                             test_bit(In_sync, &rdev->flags)) {
1545                                 if (sync_page_io(rdev->bdev,
1546                                                  sect + rdev->data_offset,
1547                                                  s<<9, conf->tmppage, READ)
1548                                     == 0)
1549                                         /* Well, this device is dead */
1550                                         md_error(mddev, rdev);
1551                                 else {
1552                                         atomic_add(s, &rdev->corrected_errors);
1553                                         printk(KERN_INFO
1554                                                "raid1:%s: read error corrected "
1555                                                "(%d sectors at %llu on %s)\n",
1556                                                mdname(mddev), s,
1557                                                (unsigned long long)(sect +
1558                                                    rdev->data_offset),
1559                                                bdevname(rdev->bdev, b));
1560                                 }
1561                         }
1562                 }
1563                 sectors -= s;
1564                 sect += s;
1565         }
1566 }
1567
1568 static void raid1d(mddev_t *mddev)
1569 {
1570         r1bio_t *r1_bio;
1571         struct bio *bio;
1572         unsigned long flags;
1573         conf_t *conf = mddev->private;
1574         struct list_head *head = &conf->retry_list;
1575         int unplug=0;
1576         mdk_rdev_t *rdev;
1577
1578         md_check_recovery(mddev);
1579
1580         for (;;) {
1581                 char b[BDEVNAME_SIZE];
1582
1583                 unplug += flush_pending_writes(conf);
1584
1585                 spin_lock_irqsave(&conf->device_lock, flags);
1586                 if (list_empty(head)) {
1587                         spin_unlock_irqrestore(&conf->device_lock, flags);
1588                         break;
1589                 }
1590                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1591                 list_del(head->prev);
1592                 conf->nr_queued--;
1593                 spin_unlock_irqrestore(&conf->device_lock, flags);
1594
1595                 mddev = r1_bio->mddev;
1596                 conf = mddev->private;
1597                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1598                         sync_request_write(mddev, r1_bio);
1599                         unplug = 1;
1600                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1601                         /* some requests in the r1bio were BIO_RW_BARRIER
1602                          * requests which failed with -EOPNOTSUPP.  Hohumm..
1603                          * Better resubmit without the barrier.
1604                          * We know which devices to resubmit for, because
1605                          * all others have had their bios[] entry cleared.
1606                          * We already have a nr_pending reference on these rdevs.
1607                          */
1608                         int i;
1609                         const bool do_sync = bio_rw_flagged(r1_bio->master_bio, BIO_RW_SYNCIO);
1610                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1611                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1612                         for (i=0; i < conf->raid_disks; i++)
1613                                 if (r1_bio->bios[i])
1614                                         atomic_inc(&r1_bio->remaining);
1615                         for (i=0; i < conf->raid_disks; i++)
1616                                 if (r1_bio->bios[i]) {
1617                                         struct bio_vec *bvec;
1618                                         int j;
1619
1620                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1621                                         /* copy pages from the failed bio, as
1622                                          * this might be a write-behind device */
1623                                         __bio_for_each_segment(bvec, bio, j, 0)
1624                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1625                                         bio_put(r1_bio->bios[i]);
1626                                         bio->bi_sector = r1_bio->sector +
1627                                                 conf->mirrors[i].rdev->data_offset;
1628                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1629                                         bio->bi_end_io = raid1_end_write_request;
1630                                         bio->bi_rw = WRITE |
1631                                                 (do_sync << BIO_RW_SYNCIO);
1632                                         bio->bi_private = r1_bio;
1633                                         r1_bio->bios[i] = bio;
1634                                         generic_make_request(bio);
1635                                 }
1636                 } else {
1637                         int disk;
1638
1639                         /* we got a read error. Maybe the drive is bad.  Maybe just
1640                          * the block and we can fix it.
1641                          * We freeze all other IO, and try reading the block from
1642                          * other devices.  When we find one, we re-write
1643                          * and check it that fixes the read error.
1644                          * This is all done synchronously while the array is
1645                          * frozen
1646                          */
1647                         if (mddev->ro == 0) {
1648                                 freeze_array(conf);
1649                                 fix_read_error(conf, r1_bio->read_disk,
1650                                                r1_bio->sector,
1651                                                r1_bio->sectors);
1652                                 unfreeze_array(conf);
1653                         } else
1654                                 md_error(mddev,
1655                                          conf->mirrors[r1_bio->read_disk].rdev);
1656
1657                         bio = r1_bio->bios[r1_bio->read_disk];
1658                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1659                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1660                                        " read error for block %llu\n",
1661                                        bdevname(bio->bi_bdev,b),
1662                                        (unsigned long long)r1_bio->sector);
1663                                 raid_end_bio_io(r1_bio);
1664                         } else {
1665                                 const bool do_sync = bio_rw_flagged(r1_bio->master_bio, BIO_RW_SYNCIO);
1666                                 r1_bio->bios[r1_bio->read_disk] =
1667                                         mddev->ro ? IO_BLOCKED : NULL;
1668                                 r1_bio->read_disk = disk;
1669                                 bio_put(bio);
1670                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1671                                 r1_bio->bios[r1_bio->read_disk] = bio;
1672                                 rdev = conf->mirrors[disk].rdev;
1673                                 if (printk_ratelimit())
1674                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1675                                                " another mirror\n",
1676                                                bdevname(rdev->bdev,b),
1677                                                (unsigned long long)r1_bio->sector);
1678                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1679                                 bio->bi_bdev = rdev->bdev;
1680                                 bio->bi_end_io = raid1_end_read_request;
1681                                 bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
1682                                 bio->bi_private = r1_bio;
1683                                 unplug = 1;
1684                                 generic_make_request(bio);
1685                         }
1686                 }
1687                 cond_resched();
1688         }
1689         if (unplug)
1690                 unplug_slaves(mddev);
1691 }
1692
1693
1694 static int init_resync(conf_t *conf)
1695 {
1696         int buffs;
1697
1698         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1699         BUG_ON(conf->r1buf_pool);
1700         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1701                                           conf->poolinfo);
1702         if (!conf->r1buf_pool)
1703                 return -ENOMEM;
1704         conf->next_resync = 0;
1705         return 0;
1706 }
1707
1708 /*
1709  * perform a "sync" on one "block"
1710  *
1711  * We need to make sure that no normal I/O request - particularly write
1712  * requests - conflict with active sync requests.
1713  *
1714  * This is achieved by tracking pending requests and a 'barrier' concept
1715  * that can be installed to exclude normal IO requests.
1716  */
1717
1718 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1719 {
1720         conf_t *conf = mddev->private;
1721         r1bio_t *r1_bio;
1722         struct bio *bio;
1723         sector_t max_sector, nr_sectors;
1724         int disk = -1;
1725         int i;
1726         int wonly = -1;
1727         int write_targets = 0, read_targets = 0;
1728         int sync_blocks;
1729         int still_degraded = 0;
1730
1731         if (!conf->r1buf_pool)
1732         {
1733 /*
1734                 printk("sync start - bitmap %p\n", mddev->bitmap);
1735 */
1736                 if (init_resync(conf))
1737                         return 0;
1738         }
1739
1740         max_sector = mddev->dev_sectors;
1741         if (sector_nr >= max_sector) {
1742                 /* If we aborted, we need to abort the
1743                  * sync on the 'current' bitmap chunk (there will
1744                  * only be one in raid1 resync.
1745                  * We can find the current addess in mddev->curr_resync
1746                  */
1747                 if (mddev->curr_resync < max_sector) /* aborted */
1748                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1749                                                 &sync_blocks, 1);
1750                 else /* completed sync */
1751                         conf->fullsync = 0;
1752
1753                 bitmap_close_sync(mddev->bitmap);
1754                 close_sync(conf);
1755                 return 0;
1756         }
1757
1758         if (mddev->bitmap == NULL &&
1759             mddev->recovery_cp == MaxSector &&
1760             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1761             conf->fullsync == 0) {
1762                 *skipped = 1;
1763                 return max_sector - sector_nr;
1764         }
1765         /* before building a request, check if we can skip these blocks..
1766          * This call the bitmap_start_sync doesn't actually record anything
1767          */
1768         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1769             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1770                 /* We can skip this block, and probably several more */
1771                 *skipped = 1;
1772                 return sync_blocks;
1773         }
1774         /*
1775          * If there is non-resync activity waiting for a turn,
1776          * and resync is going fast enough,
1777          * then let it though before starting on this new sync request.
1778          */
1779         if (!go_faster && conf->nr_waiting)
1780                 msleep_interruptible(1000);
1781
1782         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1783         raise_barrier(conf);
1784
1785         conf->next_resync = sector_nr;
1786
1787         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1788         rcu_read_lock();
1789         /*
1790          * If we get a correctably read error during resync or recovery,
1791          * we might want to read from a different device.  So we
1792          * flag all drives that could conceivably be read from for READ,
1793          * and any others (which will be non-In_sync devices) for WRITE.
1794          * If a read fails, we try reading from something else for which READ
1795          * is OK.
1796          */
1797
1798         r1_bio->mddev = mddev;
1799         r1_bio->sector = sector_nr;
1800         r1_bio->state = 0;
1801         set_bit(R1BIO_IsSync, &r1_bio->state);
1802
1803         for (i=0; i < conf->raid_disks; i++) {
1804                 mdk_rdev_t *rdev;
1805                 bio = r1_bio->bios[i];
1806
1807                 /* take from bio_init */
1808                 bio->bi_next = NULL;
1809                 bio->bi_flags |= 1 << BIO_UPTODATE;
1810                 bio->bi_rw = READ;
1811                 bio->bi_vcnt = 0;
1812                 bio->bi_idx = 0;
1813                 bio->bi_phys_segments = 0;
1814                 bio->bi_size = 0;
1815                 bio->bi_end_io = NULL;
1816                 bio->bi_private = NULL;
1817
1818                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1819                 if (rdev == NULL ||
1820                            test_bit(Faulty, &rdev->flags)) {
1821                         still_degraded = 1;
1822                         continue;
1823                 } else if (!test_bit(In_sync, &rdev->flags)) {
1824                         bio->bi_rw = WRITE;
1825                         bio->bi_end_io = end_sync_write;
1826                         write_targets ++;
1827                 } else {
1828                         /* may need to read from here */
1829                         bio->bi_rw = READ;
1830                         bio->bi_end_io = end_sync_read;
1831                         if (test_bit(WriteMostly, &rdev->flags)) {
1832                                 if (wonly < 0)
1833                                         wonly = i;
1834                         } else {
1835                                 if (disk < 0)
1836                                         disk = i;
1837                         }
1838                         read_targets++;
1839                 }
1840                 atomic_inc(&rdev->nr_pending);
1841                 bio->bi_sector = sector_nr + rdev->data_offset;
1842                 bio->bi_bdev = rdev->bdev;
1843                 bio->bi_private = r1_bio;
1844         }
1845         rcu_read_unlock();
1846         if (disk < 0)
1847                 disk = wonly;
1848         r1_bio->read_disk = disk;
1849
1850         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1851                 /* extra read targets are also write targets */
1852                 write_targets += read_targets-1;
1853
1854         if (write_targets == 0 || read_targets == 0) {
1855                 /* There is nowhere to write, so all non-sync
1856                  * drives must be failed - so we are finished
1857                  */
1858                 sector_t rv = max_sector - sector_nr;
1859                 *skipped = 1;
1860                 put_buf(r1_bio);
1861                 return rv;
1862         }
1863
1864         if (max_sector > mddev->resync_max)
1865                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1866         nr_sectors = 0;
1867         sync_blocks = 0;
1868         do {
1869                 struct page *page;
1870                 int len = PAGE_SIZE;
1871                 if (sector_nr + (len>>9) > max_sector)
1872                         len = (max_sector - sector_nr) << 9;
1873                 if (len == 0)
1874                         break;
1875                 if (sync_blocks == 0) {
1876                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1877                                                &sync_blocks, still_degraded) &&
1878                             !conf->fullsync &&
1879                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1880                                 break;
1881                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1882                         if (len > (sync_blocks<<9))
1883                                 len = sync_blocks<<9;
1884                 }
1885
1886                 for (i=0 ; i < conf->raid_disks; i++) {
1887                         bio = r1_bio->bios[i];
1888                         if (bio->bi_end_io) {
1889                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1890                                 if (bio_add_page(bio, page, len, 0) == 0) {
1891                                         /* stop here */
1892                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1893                                         while (i > 0) {
1894                                                 i--;
1895                                                 bio = r1_bio->bios[i];
1896                                                 if (bio->bi_end_io==NULL)
1897                                                         continue;
1898                                                 /* remove last page from this bio */
1899                                                 bio->bi_vcnt--;
1900                                                 bio->bi_size -= len;
1901                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1902                                         }
1903                                         goto bio_full;
1904                                 }
1905                         }
1906                 }
1907                 nr_sectors += len>>9;
1908                 sector_nr += len>>9;
1909                 sync_blocks -= (len>>9);
1910         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1911  bio_full:
1912         r1_bio->sectors = nr_sectors;
1913
1914         /* For a user-requested sync, we read all readable devices and do a
1915          * compare
1916          */
1917         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1918                 atomic_set(&r1_bio->remaining, read_targets);
1919                 for (i=0; i<conf->raid_disks; i++) {
1920                         bio = r1_bio->bios[i];
1921                         if (bio->bi_end_io == end_sync_read) {
1922                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1923                                 generic_make_request(bio);
1924                         }
1925                 }
1926         } else {
1927                 atomic_set(&r1_bio->remaining, 1);
1928                 bio = r1_bio->bios[r1_bio->read_disk];
1929                 md_sync_acct(bio->bi_bdev, nr_sectors);
1930                 generic_make_request(bio);
1931
1932         }
1933         return nr_sectors;
1934 }
1935
1936 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1937 {
1938         if (sectors)
1939                 return sectors;
1940
1941         return mddev->dev_sectors;
1942 }
1943
1944 static int run(mddev_t *mddev)
1945 {
1946         conf_t *conf;
1947         int i, j, disk_idx;
1948         mirror_info_t *disk;
1949         mdk_rdev_t *rdev;
1950
1951         if (mddev->level != 1) {
1952                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1953                        mdname(mddev), mddev->level);
1954                 goto out;
1955         }
1956         if (mddev->reshape_position != MaxSector) {
1957                 printk("raid1: %s: reshape_position set but not supported\n",
1958                        mdname(mddev));
1959                 goto out;
1960         }
1961         /*
1962          * copy the already verified devices into our private RAID1
1963          * bookkeeping area. [whatever we allocate in run(),
1964          * should be freed in stop()]
1965          */
1966         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1967         mddev->private = conf;
1968         if (!conf)
1969                 goto out_no_mem;
1970
1971         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1972                                  GFP_KERNEL);
1973         if (!conf->mirrors)
1974                 goto out_no_mem;
1975
1976         conf->tmppage = alloc_page(GFP_KERNEL);
1977         if (!conf->tmppage)
1978                 goto out_no_mem;
1979
1980         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1981         if (!conf->poolinfo)
1982                 goto out_no_mem;
1983         conf->poolinfo->mddev = NULL;
1984         conf->poolinfo->raid_disks = mddev->raid_disks;
1985         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1986                                           r1bio_pool_free,
1987                                           conf->poolinfo);
1988         if (!conf->r1bio_pool)
1989                 goto out_no_mem;
1990         conf->poolinfo->mddev = mddev;
1991
1992         spin_lock_init(&conf->device_lock);
1993         mddev->queue->queue_lock = &conf->device_lock;
1994
1995         list_for_each_entry(rdev, &mddev->disks, same_set) {
1996                 disk_idx = rdev->raid_disk;
1997                 if (disk_idx >= mddev->raid_disks
1998                     || disk_idx < 0)
1999                         continue;
2000                 disk = conf->mirrors + disk_idx;
2001
2002                 disk->rdev = rdev;
2003                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2004                                   rdev->data_offset << 9);
2005                 /* as we don't honour merge_bvec_fn, we must never risk
2006                  * violating it, so limit ->max_sector to one PAGE, as
2007                  * a one page request is never in violation.
2008                  */
2009                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2010                     queue_max_sectors(mddev->queue) > (PAGE_SIZE>>9))
2011                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
2012
2013                 disk->head_position = 0;
2014         }
2015         conf->raid_disks = mddev->raid_disks;
2016         conf->mddev = mddev;
2017         INIT_LIST_HEAD(&conf->retry_list);
2018
2019         spin_lock_init(&conf->resync_lock);
2020         init_waitqueue_head(&conf->wait_barrier);
2021
2022         bio_list_init(&conf->pending_bio_list);
2023         bio_list_init(&conf->flushing_bio_list);
2024
2025
2026         mddev->degraded = 0;
2027         for (i = 0; i < conf->raid_disks; i++) {
2028
2029                 disk = conf->mirrors + i;
2030
2031                 if (!disk->rdev ||
2032                     !test_bit(In_sync, &disk->rdev->flags)) {
2033                         disk->head_position = 0;
2034                         mddev->degraded++;
2035                         if (disk->rdev)
2036                                 conf->fullsync = 1;
2037                 }
2038         }
2039         if (mddev->degraded == conf->raid_disks) {
2040                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
2041                         mdname(mddev));
2042                 goto out_free_conf;
2043         }
2044         if (conf->raid_disks - mddev->degraded == 1)
2045                 mddev->recovery_cp = MaxSector;
2046
2047         /*
2048          * find the first working one and use it as a starting point
2049          * to read balancing.
2050          */
2051         for (j = 0; j < conf->raid_disks &&
2052                      (!conf->mirrors[j].rdev ||
2053                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
2054                 /* nothing */;
2055         conf->last_used = j;
2056
2057
2058         mddev->thread = md_register_thread(raid1d, mddev, NULL);
2059         if (!mddev->thread) {
2060                 printk(KERN_ERR
2061                        "raid1: couldn't allocate thread for %s\n",
2062                        mdname(mddev));
2063                 goto out_free_conf;
2064         }
2065
2066         if (mddev->recovery_cp != MaxSector)
2067                 printk(KERN_NOTICE "raid1: %s is not clean"
2068                        " -- starting background reconstruction\n",
2069                        mdname(mddev));
2070         printk(KERN_INFO
2071                 "raid1: raid set %s active with %d out of %d mirrors\n",
2072                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2073                 mddev->raid_disks);
2074         /*
2075          * Ok, everything is just fine now
2076          */
2077         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2078
2079         mddev->queue->unplug_fn = raid1_unplug;
2080         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2081         mddev->queue->backing_dev_info.congested_data = mddev;
2082         md_integrity_register(mddev);
2083         return 0;
2084
2085 out_no_mem:
2086         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
2087                mdname(mddev));
2088
2089 out_free_conf:
2090         if (conf) {
2091                 if (conf->r1bio_pool)
2092                         mempool_destroy(conf->r1bio_pool);
2093                 kfree(conf->mirrors);
2094                 safe_put_page(conf->tmppage);
2095                 kfree(conf->poolinfo);
2096                 kfree(conf);
2097                 mddev->private = NULL;
2098         }
2099 out:
2100         return -EIO;
2101 }
2102
2103 static int stop(mddev_t *mddev)
2104 {
2105         conf_t *conf = mddev->private;
2106         struct bitmap *bitmap = mddev->bitmap;
2107         int behind_wait = 0;
2108
2109         /* wait for behind writes to complete */
2110         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2111                 behind_wait++;
2112                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
2113                 set_current_state(TASK_UNINTERRUPTIBLE);
2114                 schedule_timeout(HZ); /* wait a second */
2115                 /* need to kick something here to make sure I/O goes? */
2116         }
2117
2118         raise_barrier(conf);
2119         lower_barrier(conf);
2120
2121         md_unregister_thread(mddev->thread);
2122         mddev->thread = NULL;
2123         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2124         if (conf->r1bio_pool)
2125                 mempool_destroy(conf->r1bio_pool);
2126         kfree(conf->mirrors);
2127         kfree(conf->poolinfo);
2128         kfree(conf);
2129         mddev->private = NULL;
2130         return 0;
2131 }
2132
2133 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2134 {
2135         /* no resync is happening, and there is enough space
2136          * on all devices, so we can resize.
2137          * We need to make sure resync covers any new space.
2138          * If the array is shrinking we should possibly wait until
2139          * any io in the removed space completes, but it hardly seems
2140          * worth it.
2141          */
2142         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2143         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2144                 return -EINVAL;
2145         set_capacity(mddev->gendisk, mddev->array_sectors);
2146         mddev->changed = 1;
2147         revalidate_disk(mddev->gendisk);
2148         if (sectors > mddev->dev_sectors &&
2149             mddev->recovery_cp == MaxSector) {
2150                 mddev->recovery_cp = mddev->dev_sectors;
2151                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2152         }
2153         mddev->dev_sectors = sectors;
2154         mddev->resync_max_sectors = sectors;
2155         return 0;
2156 }
2157
2158 static int raid1_reshape(mddev_t *mddev)
2159 {
2160         /* We need to:
2161          * 1/ resize the r1bio_pool
2162          * 2/ resize conf->mirrors
2163          *
2164          * We allocate a new r1bio_pool if we can.
2165          * Then raise a device barrier and wait until all IO stops.
2166          * Then resize conf->mirrors and swap in the new r1bio pool.
2167          *
2168          * At the same time, we "pack" the devices so that all the missing
2169          * devices have the higher raid_disk numbers.
2170          */
2171         mempool_t *newpool, *oldpool;
2172         struct pool_info *newpoolinfo;
2173         mirror_info_t *newmirrors;
2174         conf_t *conf = mddev->private;
2175         int cnt, raid_disks;
2176         unsigned long flags;
2177         int d, d2, err;
2178
2179         /* Cannot change chunk_size, layout, or level */
2180         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2181             mddev->layout != mddev->new_layout ||
2182             mddev->level != mddev->new_level) {
2183                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2184                 mddev->new_layout = mddev->layout;
2185                 mddev->new_level = mddev->level;
2186                 return -EINVAL;
2187         }
2188
2189         err = md_allow_write(mddev);
2190         if (err)
2191                 return err;
2192
2193         raid_disks = mddev->raid_disks + mddev->delta_disks;
2194
2195         if (raid_disks < conf->raid_disks) {
2196                 cnt=0;
2197                 for (d= 0; d < conf->raid_disks; d++)
2198                         if (conf->mirrors[d].rdev)
2199                                 cnt++;
2200                 if (cnt > raid_disks)
2201                         return -EBUSY;
2202         }
2203
2204         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2205         if (!newpoolinfo)
2206                 return -ENOMEM;
2207         newpoolinfo->mddev = mddev;
2208         newpoolinfo->raid_disks = raid_disks;
2209
2210         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2211                                  r1bio_pool_free, newpoolinfo);
2212         if (!newpool) {
2213                 kfree(newpoolinfo);
2214                 return -ENOMEM;
2215         }
2216         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2217         if (!newmirrors) {
2218                 kfree(newpoolinfo);
2219                 mempool_destroy(newpool);
2220                 return -ENOMEM;
2221         }
2222
2223         raise_barrier(conf);
2224
2225         /* ok, everything is stopped */
2226         oldpool = conf->r1bio_pool;
2227         conf->r1bio_pool = newpool;
2228
2229         for (d = d2 = 0; d < conf->raid_disks; d++) {
2230                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2231                 if (rdev && rdev->raid_disk != d2) {
2232                         char nm[20];
2233                         sprintf(nm, "rd%d", rdev->raid_disk);
2234                         sysfs_remove_link(&mddev->kobj, nm);
2235                         rdev->raid_disk = d2;
2236                         sprintf(nm, "rd%d", rdev->raid_disk);
2237                         sysfs_remove_link(&mddev->kobj, nm);
2238                         if (sysfs_create_link(&mddev->kobj,
2239                                               &rdev->kobj, nm))
2240                                 printk(KERN_WARNING
2241                                        "md/raid1: cannot register "
2242                                        "%s for %s\n",
2243                                        nm, mdname(mddev));
2244                 }
2245                 if (rdev)
2246                         newmirrors[d2++].rdev = rdev;
2247         }
2248         kfree(conf->mirrors);
2249         conf->mirrors = newmirrors;
2250         kfree(conf->poolinfo);
2251         conf->poolinfo = newpoolinfo;
2252
2253         spin_lock_irqsave(&conf->device_lock, flags);
2254         mddev->degraded += (raid_disks - conf->raid_disks);
2255         spin_unlock_irqrestore(&conf->device_lock, flags);
2256         conf->raid_disks = mddev->raid_disks = raid_disks;
2257         mddev->delta_disks = 0;
2258
2259         conf->last_used = 0; /* just make sure it is in-range */
2260         lower_barrier(conf);
2261
2262         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2263         md_wakeup_thread(mddev->thread);
2264
2265         mempool_destroy(oldpool);
2266         return 0;
2267 }
2268
2269 static void raid1_quiesce(mddev_t *mddev, int state)
2270 {
2271         conf_t *conf = mddev->private;
2272
2273         switch(state) {
2274         case 1:
2275                 raise_barrier(conf);
2276                 break;
2277         case 0:
2278                 lower_barrier(conf);
2279                 break;
2280         }
2281 }
2282
2283
2284 static struct mdk_personality raid1_personality =
2285 {
2286         .name           = "raid1",
2287         .level          = 1,
2288         .owner          = THIS_MODULE,
2289         .make_request   = make_request,
2290         .run            = run,
2291         .stop           = stop,
2292         .status         = status,
2293         .error_handler  = error,
2294         .hot_add_disk   = raid1_add_disk,
2295         .hot_remove_disk= raid1_remove_disk,
2296         .spare_active   = raid1_spare_active,
2297         .sync_request   = sync_request,
2298         .resize         = raid1_resize,
2299         .size           = raid1_size,
2300         .check_reshape  = raid1_reshape,
2301         .quiesce        = raid1_quiesce,
2302 };
2303
2304 static int __init raid_init(void)
2305 {
2306         return register_md_personality(&raid1_personality);
2307 }
2308
2309 static void raid_exit(void)
2310 {
2311         unregister_md_personality(&raid1_personality);
2312 }
2313
2314 module_init(raid_init);
2315 module_exit(raid_exit);
2316 MODULE_LICENSE("GPL");
2317 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2318 MODULE_ALIAS("md-raid1");
2319 MODULE_ALIAS("md-level-1");