Pileus Git - ~andy/linux/blob - drivers/md/raid1.c

   1 /*
   2  * raid1.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5  *
   6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7  *
   8  * RAID-1 management functions.
   9  *
  10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11  *
  12  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14  *
  15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  16  * bitmapped intelligence in resync:
  17  *
  18  *      - bitmap marked during normal i/o
  19  *      - bitmap used to skip nondirty blocks during sync
  20  *
  21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  22  * - persistent bitmap code
  23  *
  24  * This program is free software; you can redistribute it and/or modify
  25  * it under the terms of the GNU General Public License as published by
  26  * the Free Software Foundation; either version 2, or (at your option)
  27  * any later version.
  28  *
  29  * You should have received a copy of the GNU General Public License
  30  * (for example /usr/src/linux/COPYING); if not, write to the Free
  31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  32  */
  33
  34 #include <linux/slab.h>
  35 #include <linux/delay.h>
  36 #include <linux/blkdev.h>
  37 #include <linux/module.h>
  38 #include <linux/seq_file.h>
  39 #include <linux/ratelimit.h>
  40 #include "md.h"
  41 #include "raid1.h"
  42 #include "bitmap.h"
  43
  44 /*
  45  * Number of guaranteed r1bios in case of extreme VM load:
  46  */
  47 #define NR_RAID1_BIOS 256
  48
  49 /* when we get a read error on a read-only array, we redirect to another
  50  * device without failing the first device, or trying to over-write to
  51  * correct the read error.  To keep track of bad blocks on a per-bio
  52  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
  53  */
  54 #define IO_BLOCKED ((struct bio *)1)
  55 /* When we successfully write to a known bad-block, we need to remove the
  56  * bad-block marking which must be done from process context.  So we record
  57  * the success by setting devs[n].bio to IO_MADE_GOOD
  58  */
  59 #define IO_MADE_GOOD ((struct bio *)2)
  60
  61 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
  62
  63 /* When there are this many requests queue to be written by
  64  * the raid1 thread, we become 'congested' to provide back-pressure
  65  * for writeback.
  66  */
  67 static int max_queued_requests = 1024;
  68
  69 static void allow_barrier(struct r1conf *conf);
  70 static void lower_barrier(struct r1conf *conf);
  71
  72 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  73 {
  74         struct pool_info *pi = data;
  75         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
  76
  77         /* allocate a r1bio with room for raid_disks entries in the bios array */
  78         return kzalloc(size, gfp_flags);
  79 }
  80
  81 static void r1bio_pool_free(void *r1_bio, void *data)
  82 {
  83         kfree(r1_bio);
  84 }
  85
  86 #define RESYNC_BLOCK_SIZE (64*1024)
  87 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  88 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  89 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  90 #define RESYNC_WINDOW (2048*1024)
  91
  92 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  93 {
  94         struct pool_info *pi = data;
  95         struct page *page;
  96         struct r1bio *r1_bio;
  97         struct bio *bio;
  98         int i, j;
  99
 100         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 101         if (!r1_bio)
 102                 return NULL;
 103
 104         /*
 105          * Allocate bios : 1 for reading, n-1 for writing
 106          */
 107         for (j = pi->raid_disks ; j-- ; ) {
 108                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 109                 if (!bio)
 110                         goto out_free_bio;
 111                 r1_bio->bios[j] = bio;
 112         }
 113         /*
 114          * Allocate RESYNC_PAGES data pages and attach them to
 115          * the first bio.
 116          * If this is a user-requested check/repair, allocate
 117          * RESYNC_PAGES for each bio.
 118          */
 119         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 120                 j = pi->raid_disks;
 121         else
 122                 j = 1;
 123         while(j--) {
 124                 bio = r1_bio->bios[j];
 125                 for (i = 0; i < RESYNC_PAGES; i++) {
 126                         page = alloc_page(gfp_flags);
 127                         if (unlikely(!page))
 128                                 goto out_free_pages;
 129
 130                         bio->bi_io_vec[i].bv_page = page;
 131                         bio->bi_vcnt = i+1;
 132                 }
 133         }
 134         /* If not user-requests, copy the page pointers to all bios */
 135         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 136                 for (i=0; i<RESYNC_PAGES ; i++)
 137                         for (j=1; j<pi->raid_disks; j++)
 138                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 139                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 140         }
 141
 142         r1_bio->master_bio = NULL;
 143
 144         return r1_bio;
 145
 146 out_free_pages:
 147         for (j=0 ; j < pi->raid_disks; j++)
 148                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 149                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 150         j = -1;
 151 out_free_bio:
 152         while (++j < pi->raid_disks)
 153                 bio_put(r1_bio->bios[j]);
 154         r1bio_pool_free(r1_bio, data);
 155         return NULL;
 156 }
 157
 158 static void r1buf_pool_free(void *__r1_bio, void *data)
 159 {
 160         struct pool_info *pi = data;
 161         int i,j;
 162         struct r1bio *r1bio = __r1_bio;
 163
 164         for (i = 0; i < RESYNC_PAGES; i++)
 165                 for (j = pi->raid_disks; j-- ;) {
 166                         if (j == 0 ||
 167                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 168                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 169                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 170                 }
 171         for (i=0 ; i < pi->raid_disks; i++)
 172                 bio_put(r1bio->bios[i]);
 173
 174         r1bio_pool_free(r1bio, data);
 175 }
 176
 177 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 178 {
 179         int i;
 180
 181         for (i = 0; i < conf->raid_disks * 2; i++) {
 182                 struct bio **bio = r1_bio->bios + i;
 183                 if (!BIO_SPECIAL(*bio))
 184                         bio_put(*bio);
 185                 *bio = NULL;
 186         }
 187 }
 188
 189 static void free_r1bio(struct r1bio *r1_bio)
 190 {
 191         struct r1conf *conf = r1_bio->mddev->private;
 192
 193         put_all_bios(conf, r1_bio);
 194         mempool_free(r1_bio, conf->r1bio_pool);
 195 }
 196
 197 static void put_buf(struct r1bio *r1_bio)
 198 {
 199         struct r1conf *conf = r1_bio->mddev->private;
 200         int i;
 201
 202         for (i = 0; i < conf->raid_disks * 2; 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(struct r1bio *r1_bio)
 214 {
 215         unsigned long flags;
 216         struct mddev *mddev = r1_bio->mddev;
 217         struct r1conf *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 call_bio_endio(struct r1bio *r1_bio)
 234 {
 235         struct bio *bio = r1_bio->master_bio;
 236         int done;
 237         struct r1conf *conf = r1_bio->mddev->private;
 238
 239         if (bio->bi_phys_segments) {
 240                 unsigned long flags;
 241                 spin_lock_irqsave(&conf->device_lock, flags);
 242                 bio->bi_phys_segments--;
 243                 done = (bio->bi_phys_segments == 0);
 244                 spin_unlock_irqrestore(&conf->device_lock, flags);
 245         } else
 246                 done = 1;
 247
 248         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 249                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 250         if (done) {
 251                 bio_endio(bio, 0);
 252                 /*
 253                  * Wake up any possible resync thread that waits for the device
 254                  * to go idle.
 255                  */
 256                 allow_barrier(conf);
 257         }
 258 }
 259
 260 static void raid_end_bio_io(struct r1bio *r1_bio)
 261 {
 262         struct bio *bio = r1_bio->master_bio;
 263
 264         /* if nobody has done the final endio yet, do it now */
 265         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 266                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 267                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 268                          (unsigned long long) bio->bi_sector,
 269                          (unsigned long long) bio->bi_sector +
 270                          (bio->bi_size >> 9) - 1);
 271
 272                 call_bio_endio(r1_bio);
 273         }
 274         free_r1bio(r1_bio);
 275 }
 276
 277 /*
 278  * Update disk head position estimator based on IRQ completion info.
 279  */
 280 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 281 {
 282         struct r1conf *conf = r1_bio->mddev->private;
 283
 284         conf->mirrors[disk].head_position =
 285                 r1_bio->sector + (r1_bio->sectors);
 286 }
 287
 288 /*
 289  * Find the disk number which triggered given bio
 290  */
 291 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 292 {
 293         int mirror;
 294         struct r1conf *conf = r1_bio->mddev->private;
 295         int raid_disks = conf->raid_disks;
 296
 297         for (mirror = 0; mirror < raid_disks * 2; mirror++)
 298                 if (r1_bio->bios[mirror] == bio)
 299                         break;
 300
 301         BUG_ON(mirror == raid_disks * 2);
 302         update_head_pos(mirror, r1_bio);
 303
 304         return mirror;
 305 }
 306
 307 static void raid1_end_read_request(struct bio *bio, int error)
 308 {
 309         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 310         struct r1bio *r1_bio = bio->bi_private;
 311         int mirror;
 312         struct r1conf *conf = r1_bio->mddev->private;
 313
 314         mirror = r1_bio->read_disk;
 315         /*
 316          * this branch is our 'one mirror IO has finished' event handler:
 317          */
 318         update_head_pos(mirror, r1_bio);
 319
 320         if (uptodate)
 321                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 322         else {
 323                 /* If all other devices have failed, we want to return
 324                  * the error upwards rather than fail the last device.
 325                  * Here we redefine "uptodate" to mean "Don't want to retry"
 326                  */
 327                 unsigned long flags;
 328                 spin_lock_irqsave(&conf->device_lock, flags);
 329                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 330                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 331                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 332                         uptodate = 1;
 333                 spin_unlock_irqrestore(&conf->device_lock, flags);
 334         }
 335
 336         if (uptodate)
 337                 raid_end_bio_io(r1_bio);
 338         else {
 339                 /*
 340                  * oops, read error:
 341                  */
 342                 char b[BDEVNAME_SIZE];
 343                 printk_ratelimited(
 344                         KERN_ERR "md/raid1:%s: %s: "
 345                         "rescheduling sector %llu\n",
 346                         mdname(conf->mddev),
 347                         bdevname(conf->mirrors[mirror].rdev->bdev,
 348                                  b),
 349                         (unsigned long long)r1_bio->sector);
 350                 set_bit(R1BIO_ReadError, &r1_bio->state);
 351                 reschedule_retry(r1_bio);
 352         }
 353
 354         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 355 }
 356
 357 static void close_write(struct r1bio *r1_bio)
 358 {
 359         /* it really is the end of this request */
 360         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 361                 /* free extra copy of the data pages */
 362                 int i = r1_bio->behind_page_count;
 363                 while (i--)
 364                         safe_put_page(r1_bio->behind_bvecs[i].bv_page);
 365                 kfree(r1_bio->behind_bvecs);
 366                 r1_bio->behind_bvecs = NULL;
 367         }
 368         /* clear the bitmap if all writes complete successfully */
 369         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 370                         r1_bio->sectors,
 371                         !test_bit(R1BIO_Degraded, &r1_bio->state),
 372                         test_bit(R1BIO_BehindIO, &r1_bio->state));
 373         md_write_end(r1_bio->mddev);
 374 }
 375
 376 static void r1_bio_write_done(struct r1bio *r1_bio)
 377 {
 378         if (!atomic_dec_and_test(&r1_bio->remaining))
 379                 return;
 380
 381         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 382                 reschedule_retry(r1_bio);
 383         else {
 384                 close_write(r1_bio);
 385                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 386                         reschedule_retry(r1_bio);
 387                 else
 388                         raid_end_bio_io(r1_bio);
 389         }
 390 }
 391
 392 static void raid1_end_write_request(struct bio *bio, int error)
 393 {
 394         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 395         struct r1bio *r1_bio = bio->bi_private;
 396         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 397         struct r1conf *conf = r1_bio->mddev->private;
 398         struct bio *to_put = NULL;
 399
 400         mirror = find_bio_disk(r1_bio, bio);
 401
 402         /*
 403          * 'one mirror IO has finished' event handler:
 404          */
 405         if (!uptodate) {
 406                 set_bit(WriteErrorSeen,
 407                         &conf->mirrors[mirror].rdev->flags);
 408                 if (!test_and_set_bit(WantReplacement,
 409                                       &conf->mirrors[mirror].rdev->flags))
 410                         set_bit(MD_RECOVERY_NEEDED, &
 411                                 conf->mddev->recovery);
 412
 413                 set_bit(R1BIO_WriteError, &r1_bio->state);
 414         } else {
 415                 /*
 416                  * Set R1BIO_Uptodate in our master bio, so that we
 417                  * will return a good error code for to the higher
 418                  * levels even if IO on some other mirrored buffer
 419                  * fails.
 420                  *
 421                  * The 'master' represents the composite IO operation
 422                  * to user-side. So if something waits for IO, then it
 423                  * will wait for the 'master' bio.
 424                  */
 425                 sector_t first_bad;
 426                 int bad_sectors;
 427
 428                 r1_bio->bios[mirror] = NULL;
 429                 to_put = bio;
 430                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 431
 432                 /* Maybe we can clear some bad blocks. */
 433                 if (is_badblock(conf->mirrors[mirror].rdev,
 434                                 r1_bio->sector, r1_bio->sectors,
 435                                 &first_bad, &bad_sectors)) {
 436                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 437                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 438                 }
 439         }
 440
 441         if (behind) {
 442                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 443                         atomic_dec(&r1_bio->behind_remaining);
 444
 445                 /*
 446                  * In behind mode, we ACK the master bio once the I/O
 447                  * has safely reached all non-writemostly
 448                  * disks. Setting the Returned bit ensures that this
 449                  * gets done only once -- we don't ever want to return
 450                  * -EIO here, instead we'll wait
 451                  */
 452                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 453                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 454                         /* Maybe we can return now */
 455                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 456                                 struct bio *mbio = r1_bio->master_bio;
 457                                 pr_debug("raid1: behind end write sectors"
 458                                          " %llu-%llu\n",
 459                                          (unsigned long long) mbio->bi_sector,
 460                                          (unsigned long long) mbio->bi_sector +
 461                                          (mbio->bi_size >> 9) - 1);
 462                                 call_bio_endio(r1_bio);
 463                         }
 464                 }
 465         }
 466         if (r1_bio->bios[mirror] == NULL)
 467                 rdev_dec_pending(conf->mirrors[mirror].rdev,
 468                                  conf->mddev);
 469
 470         /*
 471          * Let's see if all mirrored write operations have finished
 472          * already.
 473          */
 474         r1_bio_write_done(r1_bio);
 475
 476         if (to_put)
 477                 bio_put(to_put);
 478 }
 479
 480
 481 /*
 482  * This routine returns the disk from which the requested read should
 483  * be done. There is a per-array 'next expected sequential IO' sector
 484  * number - if this matches on the next IO then we use the last disk.
 485  * There is also a per-disk 'last know head position' sector that is
 486  * maintained from IRQ contexts, both the normal and the resync IO
 487  * completion handlers update this position correctly. If there is no
 488  * perfect sequential match then we pick the disk whose head is closest.
 489  *
 490  * If there are 2 mirrors in the same 2 devices, performance degrades
 491  * because position is mirror, not device based.
 492  *
 493  * The rdev for the device selected will have nr_pending incremented.
 494  */
 495 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 496 {
 497         const sector_t this_sector = r1_bio->sector;
 498         int sectors;
 499         int best_good_sectors;
 500         int best_disk;
 501         int disk;
 502         sector_t best_dist;
 503         struct md_rdev *rdev;
 504         int choose_first;
 505
 506         rcu_read_lock();
 507         /*
 508          * Check if we can balance. We can balance on the whole
 509          * device if no resync is going on, or below the resync window.
 510          * We take the first readable disk when above the resync window.
 511          */
 512  retry:
 513         sectors = r1_bio->sectors;
 514         best_disk = -1;
 515         best_dist = MaxSector;
 516         best_good_sectors = 0;
 517
 518         if (conf->mddev->recovery_cp < MaxSector &&
 519             (this_sector + sectors >= conf->next_resync))
 520                 choose_first = 1;
 521         else
 522                 choose_first = 0;
 523
 524         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 525                 sector_t dist;
 526                 sector_t first_bad;
 527                 int bad_sectors;
 528
 529                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 530                 if (r1_bio->bios[disk] == IO_BLOCKED
 531                     || rdev == NULL
 532                     || test_bit(Unmerged, &rdev->flags)
 533                     || test_bit(Faulty, &rdev->flags))
 534                         continue;
 535                 if (!test_bit(In_sync, &rdev->flags) &&
 536                     rdev->recovery_offset < this_sector + sectors)
 537                         continue;
 538                 if (test_bit(WriteMostly, &rdev->flags)) {
 539                         /* Don't balance among write-mostly, just
 540                          * use the first as a last resort */
 541                         if (best_disk < 0) {
 542                                 if (is_badblock(rdev, this_sector, sectors,
 543                                                 &first_bad, &bad_sectors)) {
 544                                         if (first_bad < this_sector)
 545                                                 /* Cannot use this */
 546                                                 continue;
 547                                         best_good_sectors = first_bad - this_sector;
 548                                 } else
 549                                         best_good_sectors = sectors;
 550                                 best_disk = disk;
 551                         }
 552                         continue;
 553                 }
 554                 /* This is a reasonable device to use.  It might
 555                  * even be best.
 556                  */
 557                 if (is_badblock(rdev, this_sector, sectors,
 558                                 &first_bad, &bad_sectors)) {
 559                         if (best_dist < MaxSector)
 560                                 /* already have a better device */
 561                                 continue;
 562                         if (first_bad <= this_sector) {
 563                                 /* cannot read here. If this is the 'primary'
 564                                  * device, then we must not read beyond
 565                                  * bad_sectors from another device..
 566                                  */
 567                                 bad_sectors -= (this_sector - first_bad);
 568                                 if (choose_first && sectors > bad_sectors)
 569                                         sectors = bad_sectors;
 570                                 if (best_good_sectors > sectors)
 571                                         best_good_sectors = sectors;
 572
 573                         } else {
 574                                 sector_t good_sectors = first_bad - this_sector;
 575                                 if (good_sectors > best_good_sectors) {
 576                                         best_good_sectors = good_sectors;
 577                                         best_disk = disk;
 578                                 }
 579                                 if (choose_first)
 580                                         break;
 581                         }
 582                         continue;
 583                 } else
 584                         best_good_sectors = sectors;
 585
 586                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 587                 if (choose_first
 588                     /* Don't change to another disk for sequential reads */
 589                     || conf->mirrors[disk].next_seq_sect == this_sector
 590                     || dist == 0
 591                     /* If device is idle, use it */
 592                     || atomic_read(&rdev->nr_pending) == 0) {
 593                         best_disk = disk;
 594                         break;
 595                 }
 596                 if (dist < best_dist) {
 597                         best_dist = dist;
 598                         best_disk = disk;
 599                 }
 600         }
 601
 602         if (best_disk >= 0) {
 603                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 604                 if (!rdev)
 605                         goto retry;
 606                 atomic_inc(&rdev->nr_pending);
 607                 if (test_bit(Faulty, &rdev->flags)) {
 608                         /* cannot risk returning a device that failed
 609                          * before we inc'ed nr_pending
 610                          */
 611                         rdev_dec_pending(rdev, conf->mddev);
 612                         goto retry;
 613                 }
 614                 sectors = best_good_sectors;
 615                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 616         }
 617         rcu_read_unlock();
 618         *max_sectors = sectors;
 619
 620         return best_disk;
 621 }
 622
 623 static int raid1_mergeable_bvec(struct request_queue *q,
 624                                 struct bvec_merge_data *bvm,
 625                                 struct bio_vec *biovec)
 626 {
 627         struct mddev *mddev = q->queuedata;
 628         struct r1conf *conf = mddev->private;
 629         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 630         int max = biovec->bv_len;
 631
 632         if (mddev->merge_check_needed) {
 633                 int disk;
 634                 rcu_read_lock();
 635                 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
 636                         struct md_rdev *rdev = rcu_dereference(
 637                                 conf->mirrors[disk].rdev);
 638                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
 639                                 struct request_queue *q =
 640                                         bdev_get_queue(rdev->bdev);
 641                                 if (q->merge_bvec_fn) {
 642                                         bvm->bi_sector = sector +
 643                                                 rdev->data_offset;
 644                                         bvm->bi_bdev = rdev->bdev;
 645                                         max = min(max, q->merge_bvec_fn(
 646                                                           q, bvm, biovec));
 647                                 }
 648                         }
 649                 }
 650                 rcu_read_unlock();
 651         }
 652         return max;
 653
 654 }
 655
 656 int md_raid1_congested(struct mddev *mddev, int bits)
 657 {
 658         struct r1conf *conf = mddev->private;
 659         int i, ret = 0;
 660
 661         if ((bits & (1 << BDI_async_congested)) &&
 662             conf->pending_count >= max_queued_requests)
 663                 return 1;
 664
 665         rcu_read_lock();
 666         for (i = 0; i < conf->raid_disks * 2; i++) {
 667                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 668                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 669                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 670
 671                         BUG_ON(!q);
 672
 673                         /* Note the '|| 1' - when read_balance prefers
 674                          * non-congested targets, it can be removed
 675                          */
 676                         if ((bits & (1<<BDI_async_congested)) || 1)
 677                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 678                         else
 679                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 680                 }
 681         }
 682         rcu_read_unlock();
 683         return ret;
 684 }
 685 EXPORT_SYMBOL_GPL(md_raid1_congested);
 686
 687 static int raid1_congested(void *data, int bits)
 688 {
 689         struct mddev *mddev = data;
 690
 691         return mddev_congested(mddev, bits) ||
 692                 md_raid1_congested(mddev, bits);
 693 }
 694
 695 static void flush_pending_writes(struct r1conf *conf)
 696 {
 697         /* Any writes that have been queued but are awaiting
 698          * bitmap updates get flushed here.
 699          */
 700         spin_lock_irq(&conf->device_lock);
 701
 702         if (conf->pending_bio_list.head) {
 703                 struct bio *bio;
 704                 bio = bio_list_get(&conf->pending_bio_list);
 705                 conf->pending_count = 0;
 706                 spin_unlock_irq(&conf->device_lock);
 707                 /* flush any pending bitmap writes to
 708                  * disk before proceeding w/ I/O */
 709                 bitmap_unplug(conf->mddev->bitmap);
 710                 wake_up(&conf->wait_barrier);
 711
 712                 while (bio) { /* submit pending writes */
 713                         struct bio *next = bio->bi_next;
 714                         bio->bi_next = NULL;
 715                         generic_make_request(bio);
 716                         bio = next;
 717                 }
 718         } else
 719                 spin_unlock_irq(&conf->device_lock);
 720 }
 721
 722 /* Barriers....
 723  * Sometimes we need to suspend IO while we do something else,
 724  * either some resync/recovery, or reconfigure the array.
 725  * To do this we raise a 'barrier'.
 726  * The 'barrier' is a counter that can be raised multiple times
 727  * to count how many activities are happening which preclude
 728  * normal IO.
 729  * We can only raise the barrier if there is no pending IO.
 730  * i.e. if nr_pending == 0.
 731  * We choose only to raise the barrier if no-one is waiting for the
 732  * barrier to go down.  This means that as soon as an IO request
 733  * is ready, no other operations which require a barrier will start
 734  * until the IO request has had a chance.
 735  *
 736  * So: regular IO calls 'wait_barrier'.  When that returns there
 737  *    is no backgroup IO happening,  It must arrange to call
 738  *    allow_barrier when it has finished its IO.
 739  * backgroup IO calls must call raise_barrier.  Once that returns
 740  *    there is no normal IO happeing.  It must arrange to call
 741  *    lower_barrier when the particular background IO completes.
 742  */
 743 #define RESYNC_DEPTH 32
 744
 745 static void raise_barrier(struct r1conf *conf)
 746 {
 747         spin_lock_irq(&conf->resync_lock);
 748
 749         /* Wait until no block IO is waiting */
 750         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 751                             conf->resync_lock, );
 752
 753         /* block any new IO from starting */
 754         conf->barrier++;
 755
 756         /* Now wait for all pending IO to complete */
 757         wait_event_lock_irq(conf->wait_barrier,
 758                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 759                             conf->resync_lock, );
 760
 761         spin_unlock_irq(&conf->resync_lock);
 762 }
 763
 764 static void lower_barrier(struct r1conf *conf)
 765 {
 766         unsigned long flags;
 767         BUG_ON(conf->barrier <= 0);
 768         spin_lock_irqsave(&conf->resync_lock, flags);
 769         conf->barrier--;
 770         spin_unlock_irqrestore(&conf->resync_lock, flags);
 771         wake_up(&conf->wait_barrier);
 772 }
 773
 774 static void wait_barrier(struct r1conf *conf)
 775 {
 776         spin_lock_irq(&conf->resync_lock);
 777         if (conf->barrier) {
 778                 conf->nr_waiting++;
 779                 /* Wait for the barrier to drop.
 780                  * However if there are already pending
 781                  * requests (preventing the barrier from
 782                  * rising completely), and the
 783                  * pre-process bio queue isn't empty,
 784                  * then don't wait, as we need to empty
 785                  * that queue to get the nr_pending
 786                  * count down.
 787                  */
 788                 wait_event_lock_irq(conf->wait_barrier,
 789                                     !conf->barrier ||
 790                                     (conf->nr_pending &&
 791                                      current->bio_list &&
 792                                      !bio_list_empty(current->bio_list)),
 793                                     conf->resync_lock,
 794                         );
 795                 conf->nr_waiting--;
 796         }
 797         conf->nr_pending++;
 798         spin_unlock_irq(&conf->resync_lock);
 799 }
 800
 801 static void allow_barrier(struct r1conf *conf)
 802 {
 803         unsigned long flags;
 804         spin_lock_irqsave(&conf->resync_lock, flags);
 805         conf->nr_pending--;
 806         spin_unlock_irqrestore(&conf->resync_lock, flags);
 807         wake_up(&conf->wait_barrier);
 808 }
 809
 810 static void freeze_array(struct r1conf *conf)
 811 {
 812         /* stop syncio and normal IO and wait for everything to
 813          * go quite.
 814          * We increment barrier and nr_waiting, and then
 815          * wait until nr_pending match nr_queued+1
 816          * This is called in the context of one normal IO request
 817          * that has failed. Thus any sync request that might be pending
 818          * will be blocked by nr_pending, and we need to wait for
 819          * pending IO requests to complete or be queued for re-try.
 820          * Thus the number queued (nr_queued) plus this request (1)
 821          * must match the number of pending IOs (nr_pending) before
 822          * we continue.
 823          */
 824         spin_lock_irq(&conf->resync_lock);
 825         conf->barrier++;
 826         conf->nr_waiting++;
 827         wait_event_lock_irq(conf->wait_barrier,
 828                             conf->nr_pending == conf->nr_queued+1,
 829                             conf->resync_lock,
 830                             flush_pending_writes(conf));
 831         spin_unlock_irq(&conf->resync_lock);
 832 }
 833 static void unfreeze_array(struct r1conf *conf)
 834 {
 835         /* reverse the effect of the freeze */
 836         spin_lock_irq(&conf->resync_lock);
 837         conf->barrier--;
 838         conf->nr_waiting--;
 839         wake_up(&conf->wait_barrier);
 840         spin_unlock_irq(&conf->resync_lock);
 841 }
 842
 843
 844 /* duplicate the data pages for behind I/O
 845  */
 846 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
 847 {
 848         int i;
 849         struct bio_vec *bvec;
 850         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
 851                                         GFP_NOIO);
 852         if (unlikely(!bvecs))
 853                 return;
 854
 855         bio_for_each_segment(bvec, bio, i) {
 856                 bvecs[i] = *bvec;
 857                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
 858                 if (unlikely(!bvecs[i].bv_page))
 859                         goto do_sync_io;
 860                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
 861                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 862                 kunmap(bvecs[i].bv_page);
 863                 kunmap(bvec->bv_page);
 864         }
 865         r1_bio->behind_bvecs = bvecs;
 866         r1_bio->behind_page_count = bio->bi_vcnt;
 867         set_bit(R1BIO_BehindIO, &r1_bio->state);
 868         return;
 869
 870 do_sync_io:
 871         for (i = 0; i < bio->bi_vcnt; i++)
 872                 if (bvecs[i].bv_page)
 873                         put_page(bvecs[i].bv_page);
 874         kfree(bvecs);
 875         pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 876 }
 877
 878 static void make_request(struct mddev *mddev, struct bio * bio)
 879 {
 880         struct r1conf *conf = mddev->private;
 881         struct raid1_info *mirror;
 882         struct r1bio *r1_bio;
 883         struct bio *read_bio;
 884         int i, disks;
 885         struct bitmap *bitmap;
 886         unsigned long flags;
 887         const int rw = bio_data_dir(bio);
 888         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 889         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
 890         struct md_rdev *blocked_rdev;
 891         int first_clone;
 892         int sectors_handled;
 893         int max_sectors;
 894
 895         /*
 896          * Register the new request and wait if the reconstruction
 897          * thread has put up a bar for new requests.
 898          * Continue immediately if no resync is active currently.
 899          */
 900
 901         md_write_start(mddev, bio); /* wait on superblock update early */
 902
 903         if (bio_data_dir(bio) == WRITE &&
 904             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
 905             bio->bi_sector < mddev->suspend_hi) {
 906                 /* As the suspend_* range is controlled by
 907                  * userspace, we want an interruptible
 908                  * wait.
 909                  */
 910                 DEFINE_WAIT(w);
 911                 for (;;) {
 912                         flush_signals(current);
 913                         prepare_to_wait(&conf->wait_barrier,
 914                                         &w, TASK_INTERRUPTIBLE);
 915                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
 916                             bio->bi_sector >= mddev->suspend_hi)
 917                                 break;
 918                         schedule();
 919                 }
 920                 finish_wait(&conf->wait_barrier, &w);
 921         }
 922
 923         wait_barrier(conf);
 924
 925         bitmap = mddev->bitmap;
 926
 927         /*
 928          * make_request() can abort the operation when READA is being
 929          * used and no empty request is available.
 930          *
 931          */
 932         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 933
 934         r1_bio->master_bio = bio;
 935         r1_bio->sectors = bio->bi_size >> 9;
 936         r1_bio->state = 0;
 937         r1_bio->mddev = mddev;
 938         r1_bio->sector = bio->bi_sector;
 939
 940         /* We might need to issue multiple reads to different
 941          * devices if there are bad blocks around, so we keep
 942          * track of the number of reads in bio->bi_phys_segments.
 943          * If this is 0, there is only one r1_bio and no locking
 944          * will be needed when requests complete.  If it is
 945          * non-zero, then it is the number of not-completed requests.
 946          */
 947         bio->bi_phys_segments = 0;
 948         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
 949
 950         if (rw == READ) {
 951                 /*
 952                  * read balancing logic:
 953                  */
 954                 int rdisk;
 955
 956 read_again:
 957                 rdisk = read_balance(conf, r1_bio, &max_sectors);
 958
 959                 if (rdisk < 0) {
 960                         /* couldn't find anywhere to read from */
 961                         raid_end_bio_io(r1_bio);
 962                         return;
 963                 }
 964                 mirror = conf->mirrors + rdisk;
 965
 966                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
 967                     bitmap) {
 968                         /* Reading from a write-mostly device must
 969                          * take care not to over-take any writes
 970                          * that are 'behind'
 971                          */
 972                         wait_event(bitmap->behind_wait,
 973                                    atomic_read(&bitmap->behind_writes) == 0);
 974                 }
 975                 r1_bio->read_disk = rdisk;
 976
 977                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 978                 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
 979                             max_sectors);
 980
 981                 r1_bio->bios[rdisk] = read_bio;
 982
 983                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 984                 read_bio->bi_bdev = mirror->rdev->bdev;
 985                 read_bio->bi_end_io = raid1_end_read_request;
 986                 read_bio->bi_rw = READ | do_sync;
 987                 read_bio->bi_private = r1_bio;
 988
 989                 if (max_sectors < r1_bio->sectors) {
 990                         /* could not read all from this device, so we will
 991                          * need another r1_bio.
 992                          */
 993
 994                         sectors_handled = (r1_bio->sector + max_sectors
 995                                            - bio->bi_sector);
 996                         r1_bio->sectors = max_sectors;
 997                         spin_lock_irq(&conf->device_lock);
 998                         if (bio->bi_phys_segments == 0)
 999                                 bio->bi_phys_segments = 2;
1000                         else
1001                                 bio->bi_phys_segments++;
1002                         spin_unlock_irq(&conf->device_lock);
1003                         /* Cannot call generic_make_request directly
1004                          * as that will be queued in __make_request
1005                          * and subsequent mempool_alloc might block waiting
1006                          * for it.  So hand bio over to raid1d.
1007                          */
1008                         reschedule_retry(r1_bio);
1009
1010                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1011
1012                         r1_bio->master_bio = bio;
1013                         r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1014                         r1_bio->state = 0;
1015                         r1_bio->mddev = mddev;
1016                         r1_bio->sector = bio->bi_sector + sectors_handled;
1017                         goto read_again;
1018                 } else
1019                         generic_make_request(read_bio);
1020                 return;
1021         }
1022
1023         /*
1024          * WRITE:
1025          */
1026         if (conf->pending_count >= max_queued_requests) {
1027                 md_wakeup_thread(mddev->thread);
1028                 wait_event(conf->wait_barrier,
1029                            conf->pending_count < max_queued_requests);
1030         }
1031         /* first select target devices under rcu_lock and
1032          * inc refcount on their rdev.  Record them by setting
1033          * bios[x] to bio
1034          * If there are known/acknowledged bad blocks on any device on
1035          * which we have seen a write error, we want to avoid writing those
1036          * blocks.
1037          * This potentially requires several writes to write around
1038          * the bad blocks.  Each set of writes gets it's own r1bio
1039          * with a set of bios attached.
1040          */
1041
1042         disks = conf->raid_disks * 2;
1043  retry_write:
1044         blocked_rdev = NULL;
1045         rcu_read_lock();
1046         max_sectors = r1_bio->sectors;
1047         for (i = 0;  i < disks; i++) {
1048                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1049                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1050                         atomic_inc(&rdev->nr_pending);
1051                         blocked_rdev = rdev;
1052                         break;
1053                 }
1054                 r1_bio->bios[i] = NULL;
1055                 if (!rdev || test_bit(Faulty, &rdev->flags)
1056                     || test_bit(Unmerged, &rdev->flags)) {
1057                         if (i < conf->raid_disks)
1058                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1059                         continue;
1060                 }
1061
1062                 atomic_inc(&rdev->nr_pending);
1063                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1064                         sector_t first_bad;
1065                         int bad_sectors;
1066                         int is_bad;
1067
1068                         is_bad = is_badblock(rdev, r1_bio->sector,
1069                                              max_sectors,
1070                                              &first_bad, &bad_sectors);
1071                         if (is_bad < 0) {
1072                                 /* mustn't write here until the bad block is
1073                                  * acknowledged*/
1074                                 set_bit(BlockedBadBlocks, &rdev->flags);
1075                                 blocked_rdev = rdev;
1076                                 break;
1077                         }
1078                         if (is_bad && first_bad <= r1_bio->sector) {
1079                                 /* Cannot write here at all */
1080                                 bad_sectors -= (r1_bio->sector - first_bad);
1081                                 if (bad_sectors < max_sectors)
1082                                         /* mustn't write more than bad_sectors
1083                                          * to other devices yet
1084                                          */
1085                                         max_sectors = bad_sectors;
1086                                 rdev_dec_pending(rdev, mddev);
1087                                 /* We don't set R1BIO_Degraded as that
1088                                  * only applies if the disk is
1089                                  * missing, so it might be re-added,
1090                                  * and we want to know to recover this
1091                                  * chunk.
1092                                  * In this case the device is here,
1093                                  * and the fact that this chunk is not
1094                                  * in-sync is recorded in the bad
1095                                  * block log
1096                                  */
1097                                 continue;
1098                         }
1099                         if (is_bad) {
1100                                 int good_sectors = first_bad - r1_bio->sector;
1101                                 if (good_sectors < max_sectors)
1102                                         max_sectors = good_sectors;
1103                         }
1104                 }
1105                 r1_bio->bios[i] = bio;
1106         }
1107         rcu_read_unlock();
1108
1109         if (unlikely(blocked_rdev)) {
1110                 /* Wait for this device to become unblocked */
1111                 int j;
1112
1113                 for (j = 0; j < i; j++)
1114                         if (r1_bio->bios[j])
1115                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1116                 r1_bio->state = 0;
1117                 allow_barrier(conf);
1118                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1119                 wait_barrier(conf);
1120                 goto retry_write;
1121         }
1122
1123         if (max_sectors < r1_bio->sectors) {
1124                 /* We are splitting this write into multiple parts, so
1125                  * we need to prepare for allocating another r1_bio.
1126                  */
1127                 r1_bio->sectors = max_sectors;
1128                 spin_lock_irq(&conf->device_lock);
1129                 if (bio->bi_phys_segments == 0)
1130                         bio->bi_phys_segments = 2;
1131                 else
1132                         bio->bi_phys_segments++;
1133                 spin_unlock_irq(&conf->device_lock);
1134         }
1135         sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1136
1137         atomic_set(&r1_bio->remaining, 1);
1138         atomic_set(&r1_bio->behind_remaining, 0);
1139
1140         first_clone = 1;
1141         for (i = 0; i < disks; i++) {
1142                 struct bio *mbio;
1143                 if (!r1_bio->bios[i])
1144                         continue;
1145
1146                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1147                 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1148
1149                 if (first_clone) {
1150                         /* do behind I/O ?
1151                          * Not if there are too many, or cannot
1152                          * allocate memory, or a reader on WriteMostly
1153                          * is waiting for behind writes to flush */
1154                         if (bitmap &&
1155                             (atomic_read(&bitmap->behind_writes)
1156                              < mddev->bitmap_info.max_write_behind) &&
1157                             !waitqueue_active(&bitmap->behind_wait))
1158                                 alloc_behind_pages(mbio, r1_bio);
1159
1160                         bitmap_startwrite(bitmap, r1_bio->sector,
1161                                           r1_bio->sectors,
1162                                           test_bit(R1BIO_BehindIO,
1163                                                    &r1_bio->state));
1164                         first_clone = 0;
1165                 }
1166                 if (r1_bio->behind_bvecs) {
1167                         struct bio_vec *bvec;
1168                         int j;
1169
1170                         /* Yes, I really want the '__' version so that
1171                          * we clear any unused pointer in the io_vec, rather
1172                          * than leave them unchanged.  This is important
1173                          * because when we come to free the pages, we won't
1174                          * know the original bi_idx, so we just free
1175                          * them all
1176                          */
1177                         __bio_for_each_segment(bvec, mbio, j, 0)
1178                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1179                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1180                                 atomic_inc(&r1_bio->behind_remaining);
1181                 }
1182
1183                 r1_bio->bios[i] = mbio;
1184
1185                 mbio->bi_sector = (r1_bio->sector +
1186                                    conf->mirrors[i].rdev->data_offset);
1187                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1188                 mbio->bi_end_io = raid1_end_write_request;
1189                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
1190                 mbio->bi_private = r1_bio;
1191
1192                 atomic_inc(&r1_bio->remaining);
1193                 spin_lock_irqsave(&conf->device_lock, flags);
1194                 bio_list_add(&conf->pending_bio_list, mbio);
1195                 conf->pending_count++;
1196                 spin_unlock_irqrestore(&conf->device_lock, flags);
1197                 if (!mddev_check_plugged(mddev))
1198                         md_wakeup_thread(mddev->thread);
1199         }
1200         /* Mustn't call r1_bio_write_done before this next test,
1201          * as it could result in the bio being freed.
1202          */
1203         if (sectors_handled < (bio->bi_size >> 9)) {
1204                 r1_bio_write_done(r1_bio);
1205                 /* We need another r1_bio.  It has already been counted
1206                  * in bio->bi_phys_segments
1207                  */
1208                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1209                 r1_bio->master_bio = bio;
1210                 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1211                 r1_bio->state = 0;
1212                 r1_bio->mddev = mddev;
1213                 r1_bio->sector = bio->bi_sector + sectors_handled;
1214                 goto retry_write;
1215         }
1216
1217         r1_bio_write_done(r1_bio);
1218
1219         /* In case raid1d snuck in to freeze_array */
1220         wake_up(&conf->wait_barrier);
1221 }
1222
1223 static void status(struct seq_file *seq, struct mddev *mddev)
1224 {
1225         struct r1conf *conf = mddev->private;
1226         int i;
1227
1228         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1229                    conf->raid_disks - mddev->degraded);
1230         rcu_read_lock();
1231         for (i = 0; i < conf->raid_disks; i++) {
1232                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1233                 seq_printf(seq, "%s",
1234                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1235         }
1236         rcu_read_unlock();
1237         seq_printf(seq, "]");
1238 }
1239
1240
1241 static void error(struct mddev *mddev, struct md_rdev *rdev)
1242 {
1243         char b[BDEVNAME_SIZE];
1244         struct r1conf *conf = mddev->private;
1245
1246         /*
1247          * If it is not operational, then we have already marked it as dead
1248          * else if it is the last working disks, ignore the error, let the
1249          * next level up know.
1250          * else mark the drive as failed
1251          */
1252         if (test_bit(In_sync, &rdev->flags)
1253             && (conf->raid_disks - mddev->degraded) == 1) {
1254                 /*
1255                  * Don't fail the drive, act as though we were just a
1256                  * normal single drive.
1257                  * However don't try a recovery from this drive as
1258                  * it is very likely to fail.
1259                  */
1260                 conf->recovery_disabled = mddev->recovery_disabled;
1261                 return;
1262         }
1263         set_bit(Blocked, &rdev->flags);
1264         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1265                 unsigned long flags;
1266                 spin_lock_irqsave(&conf->device_lock, flags);
1267                 mddev->degraded++;
1268                 set_bit(Faulty, &rdev->flags);
1269                 spin_unlock_irqrestore(&conf->device_lock, flags);
1270                 /*
1271                  * if recovery is running, make sure it aborts.
1272                  */
1273                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1274         } else
1275                 set_bit(Faulty, &rdev->flags);
1276         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1277         printk(KERN_ALERT
1278                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1279                "md/raid1:%s: Operation continuing on %d devices.\n",
1280                mdname(mddev), bdevname(rdev->bdev, b),
1281                mdname(mddev), conf->raid_disks - mddev->degraded);
1282 }
1283
1284 static void print_conf(struct r1conf *conf)
1285 {
1286         int i;
1287
1288         printk(KERN_DEBUG "RAID1 conf printout:\n");
1289         if (!conf) {
1290                 printk(KERN_DEBUG "(!conf)\n");
1291                 return;
1292         }
1293         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1294                 conf->raid_disks);
1295
1296         rcu_read_lock();
1297         for (i = 0; i < conf->raid_disks; i++) {
1298                 char b[BDEVNAME_SIZE];
1299                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1300                 if (rdev)
1301                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1302                                i, !test_bit(In_sync, &rdev->flags),
1303                                !test_bit(Faulty, &rdev->flags),
1304                                bdevname(rdev->bdev,b));
1305         }
1306         rcu_read_unlock();
1307 }
1308
1309 static void close_sync(struct r1conf *conf)
1310 {
1311         wait_barrier(conf);
1312         allow_barrier(conf);
1313
1314         mempool_destroy(conf->r1buf_pool);
1315         conf->r1buf_pool = NULL;
1316 }
1317
1318 static int raid1_spare_active(struct mddev *mddev)
1319 {
1320         int i;
1321         struct r1conf *conf = mddev->private;
1322         int count = 0;
1323         unsigned long flags;
1324
1325         /*
1326          * Find all failed disks within the RAID1 configuration
1327          * and mark them readable.
1328          * Called under mddev lock, so rcu protection not needed.
1329          */
1330         for (i = 0; i < conf->raid_disks; i++) {
1331                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1332                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1333                 if (repl
1334                     && repl->recovery_offset == MaxSector
1335                     && !test_bit(Faulty, &repl->flags)
1336                     && !test_and_set_bit(In_sync, &repl->flags)) {
1337                         /* replacement has just become active */
1338                         if (!rdev ||
1339                             !test_and_clear_bit(In_sync, &rdev->flags))
1340                                 count++;
1341                         if (rdev) {
1342                                 /* Replaced device not technically
1343                                  * faulty, but we need to be sure
1344                                  * it gets removed and never re-added
1345                                  */
1346                                 set_bit(Faulty, &rdev->flags);
1347                                 sysfs_notify_dirent_safe(
1348                                         rdev->sysfs_state);
1349                         }
1350                 }
1351                 if (rdev
1352                     && !test_bit(Faulty, &rdev->flags)
1353                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1354                         count++;
1355                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1356                 }
1357         }
1358         spin_lock_irqsave(&conf->device_lock, flags);
1359         mddev->degraded -= count;
1360         spin_unlock_irqrestore(&conf->device_lock, flags);
1361
1362         print_conf(conf);
1363         return count;
1364 }
1365
1366
1367 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1368 {
1369         struct r1conf *conf = mddev->private;
1370         int err = -EEXIST;
1371         int mirror = 0;
1372         struct raid1_info *p;
1373         int first = 0;
1374         int last = conf->raid_disks - 1;
1375         struct request_queue *q = bdev_get_queue(rdev->bdev);
1376
1377         if (mddev->recovery_disabled == conf->recovery_disabled)
1378                 return -EBUSY;
1379
1380         if (rdev->raid_disk >= 0)
1381                 first = last = rdev->raid_disk;
1382
1383         if (q->merge_bvec_fn) {
1384                 set_bit(Unmerged, &rdev->flags);
1385                 mddev->merge_check_needed = 1;
1386         }
1387
1388         for (mirror = first; mirror <= last; mirror++) {
1389                 p = conf->mirrors+mirror;
1390                 if (!p->rdev) {
1391
1392                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1393                                           rdev->data_offset << 9);
1394
1395                         p->head_position = 0;
1396                         rdev->raid_disk = mirror;
1397                         err = 0;
1398                         /* As all devices are equivalent, we don't need a full recovery
1399                          * if this was recently any drive of the array
1400                          */
1401                         if (rdev->saved_raid_disk < 0)
1402                                 conf->fullsync = 1;
1403                         rcu_assign_pointer(p->rdev, rdev);
1404                         break;
1405                 }
1406                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1407                     p[conf->raid_disks].rdev == NULL) {
1408                         /* Add this device as a replacement */
1409                         clear_bit(In_sync, &rdev->flags);
1410                         set_bit(Replacement, &rdev->flags);
1411                         rdev->raid_disk = mirror;
1412                         err = 0;
1413                         conf->fullsync = 1;
1414                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1415                         break;
1416                 }
1417         }
1418         if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1419                 /* Some requests might not have seen this new
1420                  * merge_bvec_fn.  We must wait for them to complete
1421                  * before merging the device fully.
1422                  * First we make sure any code which has tested
1423                  * our function has submitted the request, then
1424                  * we wait for all outstanding requests to complete.
1425                  */
1426                 synchronize_sched();
1427                 raise_barrier(conf);
1428                 lower_barrier(conf);
1429                 clear_bit(Unmerged, &rdev->flags);
1430         }
1431         md_integrity_add_rdev(rdev, mddev);
1432         print_conf(conf);
1433         return err;
1434 }
1435
1436 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1437 {
1438         struct r1conf *conf = mddev->private;
1439         int err = 0;
1440         int number = rdev->raid_disk;
1441         struct raid1_info *p = conf->mirrors + number;
1442
1443         if (rdev != p->rdev)
1444                 p = conf->mirrors + conf->raid_disks + number;
1445
1446         print_conf(conf);
1447         if (rdev == p->rdev) {
1448                 if (test_bit(In_sync, &rdev->flags) ||
1449                     atomic_read(&rdev->nr_pending)) {
1450                         err = -EBUSY;
1451                         goto abort;
1452                 }
1453                 /* Only remove non-faulty devices if recovery
1454                  * is not possible.
1455                  */
1456                 if (!test_bit(Faulty, &rdev->flags) &&
1457                     mddev->recovery_disabled != conf->recovery_disabled &&
1458                     mddev->degraded < conf->raid_disks) {
1459                         err = -EBUSY;
1460                         goto abort;
1461                 }
1462                 p->rdev = NULL;
1463                 synchronize_rcu();
1464                 if (atomic_read(&rdev->nr_pending)) {
1465                         /* lost the race, try later */
1466                         err = -EBUSY;
1467                         p->rdev = rdev;
1468                         goto abort;
1469                 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1470                         /* We just removed a device that is being replaced.
1471                          * Move down the replacement.  We drain all IO before
1472                          * doing this to avoid confusion.
1473                          */
1474                         struct md_rdev *repl =
1475                                 conf->mirrors[conf->raid_disks + number].rdev;
1476                         raise_barrier(conf);
1477                         clear_bit(Replacement, &repl->flags);
1478                         p->rdev = repl;
1479                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1480                         lower_barrier(conf);
1481                         clear_bit(WantReplacement, &rdev->flags);
1482                 } else
1483                         clear_bit(WantReplacement, &rdev->flags);
1484                 err = md_integrity_register(mddev);
1485         }
1486 abort:
1487
1488         print_conf(conf);
1489         return err;
1490 }
1491
1492
1493 static void end_sync_read(struct bio *bio, int error)
1494 {
1495         struct r1bio *r1_bio = bio->bi_private;
1496
1497         update_head_pos(r1_bio->read_disk, r1_bio);
1498
1499         /*
1500          * we have read a block, now it needs to be re-written,
1501          * or re-read if the read failed.
1502          * We don't do much here, just schedule handling by raid1d
1503          */
1504         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1505                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1506
1507         if (atomic_dec_and_test(&r1_bio->remaining))
1508                 reschedule_retry(r1_bio);
1509 }
1510
1511 static void end_sync_write(struct bio *bio, int error)
1512 {
1513         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1514         struct r1bio *r1_bio = bio->bi_private;
1515         struct mddev *mddev = r1_bio->mddev;
1516         struct r1conf *conf = mddev->private;
1517         int mirror=0;
1518         sector_t first_bad;
1519         int bad_sectors;
1520
1521         mirror = find_bio_disk(r1_bio, bio);
1522
1523         if (!uptodate) {
1524                 sector_t sync_blocks = 0;
1525                 sector_t s = r1_bio->sector;
1526                 long sectors_to_go = r1_bio->sectors;
1527                 /* make sure these bits doesn't get cleared. */
1528                 do {
1529                         bitmap_end_sync(mddev->bitmap, s,
1530                                         &sync_blocks, 1);
1531                         s += sync_blocks;
1532                         sectors_to_go -= sync_blocks;
1533                 } while (sectors_to_go > 0);
1534                 set_bit(WriteErrorSeen,
1535                         &conf->mirrors[mirror].rdev->flags);
1536                 if (!test_and_set_bit(WantReplacement,
1537                                       &conf->mirrors[mirror].rdev->flags))
1538                         set_bit(MD_RECOVERY_NEEDED, &
1539                                 mddev->recovery);
1540                 set_bit(R1BIO_WriteError, &r1_bio->state);
1541         } else if (is_badblock(conf->mirrors[mirror].rdev,
1542                                r1_bio->sector,
1543                                r1_bio->sectors,
1544                                &first_bad, &bad_sectors) &&
1545                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1546                                 r1_bio->sector,
1547                                 r1_bio->sectors,
1548                                 &first_bad, &bad_sectors)
1549                 )
1550                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1551
1552         if (atomic_dec_and_test(&r1_bio->remaining)) {
1553                 int s = r1_bio->sectors;
1554                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1555                     test_bit(R1BIO_WriteError, &r1_bio->state))
1556                         reschedule_retry(r1_bio);
1557                 else {
1558                         put_buf(r1_bio);
1559                         md_done_sync(mddev, s, uptodate);
1560                 }
1561         }
1562 }
1563
1564 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1565                             int sectors, struct page *page, int rw)
1566 {
1567         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1568                 /* success */
1569                 return 1;
1570         if (rw == WRITE) {
1571                 set_bit(WriteErrorSeen, &rdev->flags);
1572                 if (!test_and_set_bit(WantReplacement,
1573                                       &rdev->flags))
1574                         set_bit(MD_RECOVERY_NEEDED, &
1575                                 rdev->mddev->recovery);
1576         }
1577         /* need to record an error - either for the block or the device */
1578         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1579                 md_error(rdev->mddev, rdev);
1580         return 0;
1581 }
1582
1583 static int fix_sync_read_error(struct r1bio *r1_bio)
1584 {
1585         /* Try some synchronous reads of other devices to get
1586          * good data, much like with normal read errors.  Only
1587          * read into the pages we already have so we don't
1588          * need to re-issue the read request.
1589          * We don't need to freeze the array, because being in an
1590          * active sync request, there is no normal IO, and
1591          * no overlapping syncs.
1592          * We don't need to check is_badblock() again as we
1593          * made sure that anything with a bad block in range
1594          * will have bi_end_io clear.
1595          */
1596         struct mddev *mddev = r1_bio->mddev;
1597         struct r1conf *conf = mddev->private;
1598         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1599         sector_t sect = r1_bio->sector;
1600         int sectors = r1_bio->sectors;
1601         int idx = 0;
1602
1603         while(sectors) {
1604                 int s = sectors;
1605                 int d = r1_bio->read_disk;
1606                 int success = 0;
1607                 struct md_rdev *rdev;
1608                 int start;
1609
1610                 if (s > (PAGE_SIZE>>9))
1611                         s = PAGE_SIZE >> 9;
1612                 do {
1613                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1614                                 /* No rcu protection needed here devices
1615                                  * can only be removed when no resync is
1616                                  * active, and resync is currently active
1617                                  */
1618                                 rdev = conf->mirrors[d].rdev;
1619                                 if (sync_page_io(rdev, sect, s<<9,
1620                                                  bio->bi_io_vec[idx].bv_page,
1621                                                  READ, false)) {
1622                                         success = 1;
1623                                         break;
1624                                 }
1625                         }
1626                         d++;
1627                         if (d == conf->raid_disks * 2)
1628                                 d = 0;
1629                 } while (!success && d != r1_bio->read_disk);
1630
1631                 if (!success) {
1632                         char b[BDEVNAME_SIZE];
1633                         int abort = 0;
1634                         /* Cannot read from anywhere, this block is lost.
1635                          * Record a bad block on each device.  If that doesn't
1636                          * work just disable and interrupt the recovery.
1637                          * Don't fail devices as that won't really help.
1638                          */
1639                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1640                                " for block %llu\n",
1641                                mdname(mddev),
1642                                bdevname(bio->bi_bdev, b),
1643                                (unsigned long long)r1_bio->sector);
1644                         for (d = 0; d < conf->raid_disks * 2; d++) {
1645                                 rdev = conf->mirrors[d].rdev;
1646                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1647                                         continue;
1648                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1649                                         abort = 1;
1650                         }
1651                         if (abort) {
1652                                 conf->recovery_disabled =
1653                                         mddev->recovery_disabled;
1654                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1655                                 md_done_sync(mddev, r1_bio->sectors, 0);
1656                                 put_buf(r1_bio);
1657                                 return 0;
1658                         }
1659                         /* Try next page */
1660                         sectors -= s;
1661                         sect += s;
1662                         idx++;
1663                         continue;
1664                 }
1665
1666                 start = d;
1667                 /* write it back and re-read */
1668                 while (d != r1_bio->read_disk) {
1669                         if (d == 0)
1670                                 d = conf->raid_disks * 2;
1671                         d--;
1672                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1673                                 continue;
1674                         rdev = conf->mirrors[d].rdev;
1675                         if (r1_sync_page_io(rdev, sect, s,
1676                                             bio->bi_io_vec[idx].bv_page,
1677                                             WRITE) == 0) {
1678                                 r1_bio->bios[d]->bi_end_io = NULL;
1679                                 rdev_dec_pending(rdev, mddev);
1680                         }
1681                 }
1682                 d = start;
1683                 while (d != r1_bio->read_disk) {
1684                         if (d == 0)
1685                                 d = conf->raid_disks * 2;
1686                         d--;
1687                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1688                                 continue;
1689                         rdev = conf->mirrors[d].rdev;
1690                         if (r1_sync_page_io(rdev, sect, s,
1691                                             bio->bi_io_vec[idx].bv_page,
1692                                             READ) != 0)
1693                                 atomic_add(s, &rdev->corrected_errors);
1694                 }
1695                 sectors -= s;
1696                 sect += s;
1697                 idx ++;
1698         }
1699         set_bit(R1BIO_Uptodate, &r1_bio->state);
1700         set_bit(BIO_UPTODATE, &bio->bi_flags);
1701         return 1;
1702 }
1703
1704 static int process_checks(struct r1bio *r1_bio)
1705 {
1706         /* We have read all readable devices.  If we haven't
1707          * got the block, then there is no hope left.
1708          * If we have, then we want to do a comparison
1709          * and skip the write if everything is the same.
1710          * If any blocks failed to read, then we need to
1711          * attempt an over-write
1712          */
1713         struct mddev *mddev = r1_bio->mddev;
1714         struct r1conf *conf = mddev->private;
1715         int primary;
1716         int i;
1717         int vcnt;
1718
1719         for (primary = 0; primary < conf->raid_disks * 2; primary++)
1720                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1721                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1722                         r1_bio->bios[primary]->bi_end_io = NULL;
1723                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1724                         break;
1725                 }
1726         r1_bio->read_disk = primary;
1727         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1728         for (i = 0; i < conf->raid_disks * 2; i++) {
1729                 int j;
1730                 struct bio *pbio = r1_bio->bios[primary];
1731                 struct bio *sbio = r1_bio->bios[i];
1732                 int size;
1733
1734                 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1735                         continue;
1736
1737                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1738                         for (j = vcnt; j-- ; ) {
1739                                 struct page *p, *s;
1740                                 p = pbio->bi_io_vec[j].bv_page;
1741                                 s = sbio->bi_io_vec[j].bv_page;
1742                                 if (memcmp(page_address(p),
1743                                            page_address(s),
1744                                            sbio->bi_io_vec[j].bv_len))
1745                                         break;
1746                         }
1747                 } else
1748                         j = 0;
1749                 if (j >= 0)
1750                         mddev->resync_mismatches += r1_bio->sectors;
1751                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1752                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1753                         /* No need to write to this device. */
1754                         sbio->bi_end_io = NULL;
1755                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1756                         continue;
1757                 }
1758                 /* fixup the bio for reuse */
1759                 sbio->bi_vcnt = vcnt;
1760                 sbio->bi_size = r1_bio->sectors << 9;
1761                 sbio->bi_idx = 0;
1762                 sbio->bi_phys_segments = 0;
1763                 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1764                 sbio->bi_flags |= 1 << BIO_UPTODATE;
1765                 sbio->bi_next = NULL;
1766                 sbio->bi_sector = r1_bio->sector +
1767                         conf->mirrors[i].rdev->data_offset;
1768                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1769                 size = sbio->bi_size;
1770                 for (j = 0; j < vcnt ; j++) {
1771                         struct bio_vec *bi;
1772                         bi = &sbio->bi_io_vec[j];
1773                         bi->bv_offset = 0;
1774                         if (size > PAGE_SIZE)
1775                                 bi->bv_len = PAGE_SIZE;
1776                         else
1777                                 bi->bv_len = size;
1778                         size -= PAGE_SIZE;
1779                         memcpy(page_address(bi->bv_page),
1780                                page_address(pbio->bi_io_vec[j].bv_page),
1781                                PAGE_SIZE);
1782                 }
1783         }
1784         return 0;
1785 }
1786
1787 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1788 {
1789         struct r1conf *conf = mddev->private;
1790         int i;
1791         int disks = conf->raid_disks * 2;
1792         struct bio *bio, *wbio;
1793
1794         bio = r1_bio->bios[r1_bio->read_disk];
1795
1796         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1797                 /* ouch - failed to read all of that. */
1798                 if (!fix_sync_read_error(r1_bio))
1799                         return;
1800
1801         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1802                 if (process_checks(r1_bio) < 0)
1803                         return;
1804         /*
1805          * schedule writes
1806          */
1807         atomic_set(&r1_bio->remaining, 1);
1808         for (i = 0; i < disks ; i++) {
1809                 wbio = r1_bio->bios[i];
1810                 if (wbio->bi_end_io == NULL ||
1811                     (wbio->bi_end_io == end_sync_read &&
1812                      (i == r1_bio->read_disk ||
1813                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1814                         continue;
1815
1816                 wbio->bi_rw = WRITE;
1817                 wbio->bi_end_io = end_sync_write;
1818                 atomic_inc(&r1_bio->remaining);
1819                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1820
1821                 generic_make_request(wbio);
1822         }
1823
1824         if (atomic_dec_and_test(&r1_bio->remaining)) {
1825                 /* if we're here, all write(s) have completed, so clean up */
1826                 int s = r1_bio->sectors;
1827                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1828                     test_bit(R1BIO_WriteError, &r1_bio->state))
1829                         reschedule_retry(r1_bio);
1830                 else {
1831                         put_buf(r1_bio);
1832                         md_done_sync(mddev, s, 1);
1833                 }
1834         }
1835 }
1836
1837 /*
1838  * This is a kernel thread which:
1839  *
1840  *      1.      Retries failed read operations on working mirrors.
1841  *      2.      Updates the raid superblock when problems encounter.
1842  *      3.      Performs writes following reads for array synchronising.
1843  */
1844
1845 static void fix_read_error(struct r1conf *conf, int read_disk,
1846                            sector_t sect, int sectors)
1847 {
1848         struct mddev *mddev = conf->mddev;
1849         while(sectors) {
1850                 int s = sectors;
1851                 int d = read_disk;
1852                 int success = 0;
1853                 int start;
1854                 struct md_rdev *rdev;
1855
1856                 if (s > (PAGE_SIZE>>9))
1857                         s = PAGE_SIZE >> 9;
1858
1859                 do {
1860                         /* Note: no rcu protection needed here
1861                          * as this is synchronous in the raid1d thread
1862                          * which is the thread that might remove
1863                          * a device.  If raid1d ever becomes multi-threaded....
1864                          */
1865                         sector_t first_bad;
1866                         int bad_sectors;
1867
1868                         rdev = conf->mirrors[d].rdev;
1869                         if (rdev &&
1870                             (test_bit(In_sync, &rdev->flags) ||
1871                              (!test_bit(Faulty, &rdev->flags) &&
1872                               rdev->recovery_offset >= sect + s)) &&
1873                             is_badblock(rdev, sect, s,
1874                                         &first_bad, &bad_sectors) == 0 &&
1875                             sync_page_io(rdev, sect, s<<9,
1876                                          conf->tmppage, READ, false))
1877                                 success = 1;
1878                         else {
1879                                 d++;
1880                                 if (d == conf->raid_disks * 2)
1881                                         d = 0;
1882                         }
1883                 } while (!success && d != read_disk);
1884
1885                 if (!success) {
1886                         /* Cannot read from anywhere - mark it bad */
1887                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
1888                         if (!rdev_set_badblocks(rdev, sect, s, 0))
1889                                 md_error(mddev, rdev);
1890                         break;
1891                 }
1892                 /* write it back and re-read */
1893                 start = d;
1894                 while (d != read_disk) {
1895                         if (d==0)
1896                                 d = conf->raid_disks * 2;
1897                         d--;
1898                         rdev = conf->mirrors[d].rdev;
1899                         if (rdev &&
1900                             test_bit(In_sync, &rdev->flags))
1901                                 r1_sync_page_io(rdev, sect, s,
1902                                                 conf->tmppage, WRITE);
1903                 }
1904                 d = start;
1905                 while (d != read_disk) {
1906                         char b[BDEVNAME_SIZE];
1907                         if (d==0)
1908                                 d = conf->raid_disks * 2;
1909                         d--;
1910                         rdev = conf->mirrors[d].rdev;
1911                         if (rdev &&
1912                             test_bit(In_sync, &rdev->flags)) {
1913                                 if (r1_sync_page_io(rdev, sect, s,
1914                                                     conf->tmppage, READ)) {
1915                                         atomic_add(s, &rdev->corrected_errors);
1916                                         printk(KERN_INFO
1917                                                "md/raid1:%s: read error corrected "
1918                                                "(%d sectors at %llu on %s)\n",
1919                                                mdname(mddev), s,
1920                                                (unsigned long long)(sect +
1921                                                    rdev->data_offset),
1922                                                bdevname(rdev->bdev, b));
1923                                 }
1924                         }
1925                 }
1926                 sectors -= s;
1927                 sect += s;
1928         }
1929 }
1930
1931 static void bi_complete(struct bio *bio, int error)
1932 {
1933         complete((struct completion *)bio->bi_private);
1934 }
1935
1936 static int submit_bio_wait(int rw, struct bio *bio)
1937 {
1938         struct completion event;
1939         rw |= REQ_SYNC;
1940
1941         init_completion(&event);
1942         bio->bi_private = &event;
1943         bio->bi_end_io = bi_complete;
1944         submit_bio(rw, bio);
1945         wait_for_completion(&event);
1946
1947         return test_bit(BIO_UPTODATE, &bio->bi_flags);
1948 }
1949
1950 static int narrow_write_error(struct r1bio *r1_bio, int i)
1951 {
1952         struct mddev *mddev = r1_bio->mddev;
1953         struct r1conf *conf = mddev->private;
1954         struct md_rdev *rdev = conf->mirrors[i].rdev;
1955         int vcnt, idx;
1956         struct bio_vec *vec;
1957
1958         /* bio has the data to be written to device 'i' where
1959          * we just recently had a write error.
1960          * We repeatedly clone the bio and trim down to one block,
1961          * then try the write.  Where the write fails we record
1962          * a bad block.
1963          * It is conceivable that the bio doesn't exactly align with
1964          * blocks.  We must handle this somehow.
1965          *
1966          * We currently own a reference on the rdev.
1967          */
1968
1969         int block_sectors;
1970         sector_t sector;
1971         int sectors;
1972         int sect_to_write = r1_bio->sectors;
1973         int ok = 1;
1974
1975         if (rdev->badblocks.shift < 0)
1976                 return 0;
1977
1978         block_sectors = 1 << rdev->badblocks.shift;
1979         sector = r1_bio->sector;
1980         sectors = ((sector + block_sectors)
1981                    & ~(sector_t)(block_sectors - 1))
1982                 - sector;
1983
1984         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
1985                 vcnt = r1_bio->behind_page_count;
1986                 vec = r1_bio->behind_bvecs;
1987                 idx = 0;
1988                 while (vec[idx].bv_page == NULL)
1989                         idx++;
1990         } else {
1991                 vcnt = r1_bio->master_bio->bi_vcnt;
1992                 vec = r1_bio->master_bio->bi_io_vec;
1993                 idx = r1_bio->master_bio->bi_idx;
1994         }
1995         while (sect_to_write) {
1996                 struct bio *wbio;
1997                 if (sectors > sect_to_write)
1998                         sectors = sect_to_write;
1999                 /* Write at 'sector' for 'sectors'*/
2000
2001                 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2002                 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2003                 wbio->bi_sector = r1_bio->sector;
2004                 wbio->bi_rw = WRITE;
2005                 wbio->bi_vcnt = vcnt;
2006                 wbio->bi_size = r1_bio->sectors << 9;
2007                 wbio->bi_idx = idx;
2008
2009                 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2010                 wbio->bi_sector += rdev->data_offset;
2011                 wbio->bi_bdev = rdev->bdev;
2012                 if (submit_bio_wait(WRITE, wbio) == 0)
2013                         /* failure! */
2014                         ok = rdev_set_badblocks(rdev, sector,
2015                                                 sectors, 0)
2016                                 && ok;
2017
2018                 bio_put(wbio);
2019                 sect_to_write -= sectors;
2020                 sector += sectors;
2021                 sectors = block_sectors;
2022         }
2023         return ok;
2024 }
2025
2026 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2027 {
2028         int m;
2029         int s = r1_bio->sectors;
2030         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2031                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2032                 struct bio *bio = r1_bio->bios[m];
2033                 if (bio->bi_end_io == NULL)
2034                         continue;
2035                 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2036                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2037                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2038                 }
2039                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2040                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2041                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2042                                 md_error(conf->mddev, rdev);
2043                 }
2044         }
2045         put_buf(r1_bio);
2046         md_done_sync(conf->mddev, s, 1);
2047 }
2048
2049 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2050 {
2051         int m;
2052         for (m = 0; m < conf->raid_disks * 2 ; m++)
2053                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2054                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2055                         rdev_clear_badblocks(rdev,
2056                                              r1_bio->sector,
2057                                              r1_bio->sectors, 0);
2058                         rdev_dec_pending(rdev, conf->mddev);
2059                 } else if (r1_bio->bios[m] != NULL) {
2060                         /* This drive got a write error.  We need to
2061                          * narrow down and record precise write
2062                          * errors.
2063                          */
2064                         if (!narrow_write_error(r1_bio, m)) {
2065                                 md_error(conf->mddev,
2066                                          conf->mirrors[m].rdev);
2067                                 /* an I/O failed, we can't clear the bitmap */
2068                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2069                         }
2070                         rdev_dec_pending(conf->mirrors[m].rdev,
2071                                          conf->mddev);
2072                 }
2073         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2074                 close_write(r1_bio);
2075         raid_end_bio_io(r1_bio);
2076 }
2077
2078 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2079 {
2080         int disk;
2081         int max_sectors;
2082         struct mddev *mddev = conf->mddev;
2083         struct bio *bio;
2084         char b[BDEVNAME_SIZE];
2085         struct md_rdev *rdev;
2086
2087         clear_bit(R1BIO_ReadError, &r1_bio->state);
2088         /* we got a read error. Maybe the drive is bad.  Maybe just
2089          * the block and we can fix it.
2090          * We freeze all other IO, and try reading the block from
2091          * other devices.  When we find one, we re-write
2092          * and check it that fixes the read error.
2093          * This is all done synchronously while the array is
2094          * frozen
2095          */
2096         if (mddev->ro == 0) {
2097                 freeze_array(conf);
2098                 fix_read_error(conf, r1_bio->read_disk,
2099                                r1_bio->sector, r1_bio->sectors);
2100                 unfreeze_array(conf);
2101         } else
2102                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2103
2104         bio = r1_bio->bios[r1_bio->read_disk];
2105         bdevname(bio->bi_bdev, b);
2106 read_more:
2107         disk = read_balance(conf, r1_bio, &max_sectors);
2108         if (disk == -1) {
2109                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2110                        " read error for block %llu\n",
2111                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
2112                 raid_end_bio_io(r1_bio);
2113         } else {
2114                 const unsigned long do_sync
2115                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
2116                 if (bio) {
2117                         r1_bio->bios[r1_bio->read_disk] =
2118                                 mddev->ro ? IO_BLOCKED : NULL;
2119                         bio_put(bio);
2120                 }
2121                 r1_bio->read_disk = disk;
2122                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2123                 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2124                 r1_bio->bios[r1_bio->read_disk] = bio;
2125                 rdev = conf->mirrors[disk].rdev;
2126                 printk_ratelimited(KERN_ERR
2127                                    "md/raid1:%s: redirecting sector %llu"
2128                                    " to other mirror: %s\n",
2129                                    mdname(mddev),
2130                                    (unsigned long long)r1_bio->sector,
2131                                    bdevname(rdev->bdev, b));
2132                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
2133                 bio->bi_bdev = rdev->bdev;
2134                 bio->bi_end_io = raid1_end_read_request;
2135                 bio->bi_rw = READ | do_sync;
2136                 bio->bi_private = r1_bio;
2137                 if (max_sectors < r1_bio->sectors) {
2138                         /* Drat - have to split this up more */
2139                         struct bio *mbio = r1_bio->master_bio;
2140                         int sectors_handled = (r1_bio->sector + max_sectors
2141                                                - mbio->bi_sector);
2142                         r1_bio->sectors = max_sectors;
2143                         spin_lock_irq(&conf->device_lock);
2144                         if (mbio->bi_phys_segments == 0)
2145                                 mbio->bi_phys_segments = 2;
2146                         else
2147                                 mbio->bi_phys_segments++;
2148                         spin_unlock_irq(&conf->device_lock);
2149                         generic_make_request(bio);
2150                         bio = NULL;
2151
2152                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2153
2154                         r1_bio->master_bio = mbio;
2155                         r1_bio->sectors = (mbio->bi_size >> 9)
2156                                           - sectors_handled;
2157                         r1_bio->state = 0;
2158                         set_bit(R1BIO_ReadError, &r1_bio->state);
2159                         r1_bio->mddev = mddev;
2160                         r1_bio->sector = mbio->bi_sector + sectors_handled;
2161
2162                         goto read_more;
2163                 } else
2164                         generic_make_request(bio);
2165         }
2166 }
2167
2168 static void raid1d(struct mddev *mddev)
2169 {
2170         struct r1bio *r1_bio;
2171         unsigned long flags;
2172         struct r1conf *conf = mddev->private;
2173         struct list_head *head = &conf->retry_list;
2174         struct blk_plug plug;
2175
2176         md_check_recovery(mddev);
2177
2178         blk_start_plug(&plug);
2179         for (;;) {
2180
2181                 if (atomic_read(&mddev->plug_cnt) == 0)
2182                         flush_pending_writes(conf);
2183
2184                 spin_lock_irqsave(&conf->device_lock, flags);
2185                 if (list_empty(head)) {
2186                         spin_unlock_irqrestore(&conf->device_lock, flags);
2187                         break;
2188                 }
2189                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2190                 list_del(head->prev);
2191                 conf->nr_queued--;
2192                 spin_unlock_irqrestore(&conf->device_lock, flags);
2193
2194                 mddev = r1_bio->mddev;
2195                 conf = mddev->private;
2196                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2197                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2198                             test_bit(R1BIO_WriteError, &r1_bio->state))
2199                                 handle_sync_write_finished(conf, r1_bio);
2200                         else
2201                                 sync_request_write(mddev, r1_bio);
2202                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2203                            test_bit(R1BIO_WriteError, &r1_bio->state))
2204                         handle_write_finished(conf, r1_bio);
2205                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2206                         handle_read_error(conf, r1_bio);
2207                 else
2208                         /* just a partial read to be scheduled from separate
2209                          * context
2210                          */
2211                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2212
2213                 cond_resched();
2214                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2215                         md_check_recovery(mddev);
2216         }
2217         blk_finish_plug(&plug);
2218 }
2219
2220
2221 static int init_resync(struct r1conf *conf)
2222 {
2223         int buffs;
2224
2225         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2226         BUG_ON(conf->r1buf_pool);
2227         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2228                                           conf->poolinfo);
2229         if (!conf->r1buf_pool)
2230                 return -ENOMEM;
2231         conf->next_resync = 0;
2232         return 0;
2233 }
2234
2235 /*
2236  * perform a "sync" on one "block"
2237  *
2238  * We need to make sure that no normal I/O request - particularly write
2239  * requests - conflict with active sync requests.
2240  *
2241  * This is achieved by tracking pending requests and a 'barrier' concept
2242  * that can be installed to exclude normal IO requests.
2243  */
2244
2245 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2246 {
2247         struct r1conf *conf = mddev->private;
2248         struct r1bio *r1_bio;
2249         struct bio *bio;
2250         sector_t max_sector, nr_sectors;
2251         int disk = -1;
2252         int i;
2253         int wonly = -1;
2254         int write_targets = 0, read_targets = 0;
2255         sector_t sync_blocks;
2256         int still_degraded = 0;
2257         int good_sectors = RESYNC_SECTORS;
2258         int min_bad = 0; /* number of sectors that are bad in all devices */
2259
2260         if (!conf->r1buf_pool)
2261                 if (init_resync(conf))
2262                         return 0;
2263
2264         max_sector = mddev->dev_sectors;
2265         if (sector_nr >= max_sector) {
2266                 /* If we aborted, we need to abort the
2267                  * sync on the 'current' bitmap chunk (there will
2268                  * only be one in raid1 resync.
2269                  * We can find the current addess in mddev->curr_resync
2270                  */
2271                 if (mddev->curr_resync < max_sector) /* aborted */
2272                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2273                                                 &sync_blocks, 1);
2274                 else /* completed sync */
2275                         conf->fullsync = 0;
2276
2277                 bitmap_close_sync(mddev->bitmap);
2278                 close_sync(conf);
2279                 return 0;
2280         }
2281
2282         if (mddev->bitmap == NULL &&
2283             mddev->recovery_cp == MaxSector &&
2284             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2285             conf->fullsync == 0) {
2286                 *skipped = 1;
2287                 return max_sector - sector_nr;
2288         }
2289         /* before building a request, check if we can skip these blocks..
2290          * This call the bitmap_start_sync doesn't actually record anything
2291          */
2292         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2293             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2294                 /* We can skip this block, and probably several more */
2295                 *skipped = 1;
2296                 return sync_blocks;
2297         }
2298         /*
2299          * If there is non-resync activity waiting for a turn,
2300          * and resync is going fast enough,
2301          * then let it though before starting on this new sync request.
2302          */
2303         if (!go_faster && conf->nr_waiting)
2304                 msleep_interruptible(1000);
2305
2306         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2307         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2308         raise_barrier(conf);
2309
2310         conf->next_resync = sector_nr;
2311
2312         rcu_read_lock();
2313         /*
2314          * If we get a correctably read error during resync or recovery,
2315          * we might want to read from a different device.  So we
2316          * flag all drives that could conceivably be read from for READ,
2317          * and any others (which will be non-In_sync devices) for WRITE.
2318          * If a read fails, we try reading from something else for which READ
2319          * is OK.
2320          */
2321
2322         r1_bio->mddev = mddev;
2323         r1_bio->sector = sector_nr;
2324         r1_bio->state = 0;
2325         set_bit(R1BIO_IsSync, &r1_bio->state);
2326
2327         for (i = 0; i < conf->raid_disks * 2; i++) {
2328                 struct md_rdev *rdev;
2329                 bio = r1_bio->bios[i];
2330
2331                 /* take from bio_init */
2332                 bio->bi_next = NULL;
2333                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
2334                 bio->bi_flags |= 1 << BIO_UPTODATE;
2335                 bio->bi_rw = READ;
2336                 bio->bi_vcnt = 0;
2337                 bio->bi_idx = 0;
2338                 bio->bi_phys_segments = 0;
2339                 bio->bi_size = 0;
2340                 bio->bi_end_io = NULL;
2341                 bio->bi_private = NULL;
2342
2343                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2344                 if (rdev == NULL ||
2345                     test_bit(Faulty, &rdev->flags)) {
2346                         if (i < conf->raid_disks)
2347                                 still_degraded = 1;
2348                 } else if (!test_bit(In_sync, &rdev->flags)) {
2349                         bio->bi_rw = WRITE;
2350                         bio->bi_end_io = end_sync_write;
2351                         write_targets ++;
2352                 } else {
2353                         /* may need to read from here */
2354                         sector_t first_bad = MaxSector;
2355                         int bad_sectors;
2356
2357                         if (is_badblock(rdev, sector_nr, good_sectors,
2358                                         &first_bad, &bad_sectors)) {
2359                                 if (first_bad > sector_nr)
2360                                         good_sectors = first_bad - sector_nr;
2361                                 else {
2362                                         bad_sectors -= (sector_nr - first_bad);
2363                                         if (min_bad == 0 ||
2364                                             min_bad > bad_sectors)
2365                                                 min_bad = bad_sectors;
2366                                 }
2367                         }
2368                         if (sector_nr < first_bad) {
2369                                 if (test_bit(WriteMostly, &rdev->flags)) {
2370                                         if (wonly < 0)
2371                                                 wonly = i;
2372                                 } else {
2373                                         if (disk < 0)
2374                                                 disk = i;
2375                                 }
2376                                 bio->bi_rw = READ;
2377                                 bio->bi_end_io = end_sync_read;
2378                                 read_targets++;
2379                         }
2380                 }
2381                 if (bio->bi_end_io) {
2382                         atomic_inc(&rdev->nr_pending);
2383                         bio->bi_sector = sector_nr + rdev->data_offset;
2384                         bio->bi_bdev = rdev->bdev;
2385                         bio->bi_private = r1_bio;
2386                 }
2387         }
2388         rcu_read_unlock();
2389         if (disk < 0)
2390                 disk = wonly;
2391         r1_bio->read_disk = disk;
2392
2393         if (read_targets == 0 && min_bad > 0) {
2394                 /* These sectors are bad on all InSync devices, so we
2395                  * need to mark them bad on all write targets
2396                  */
2397                 int ok = 1;
2398                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2399                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2400                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2401                                 ok = rdev_set_badblocks(rdev, sector_nr,
2402                                                         min_bad, 0
2403                                         ) && ok;
2404                         }
2405                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2406                 *skipped = 1;
2407                 put_buf(r1_bio);
2408
2409                 if (!ok) {
2410                         /* Cannot record the badblocks, so need to
2411                          * abort the resync.
2412                          * If there are multiple read targets, could just
2413                          * fail the really bad ones ???
2414                          */
2415                         conf->recovery_disabled = mddev->recovery_disabled;
2416                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2417                         return 0;
2418                 } else
2419                         return min_bad;
2420
2421         }
2422         if (min_bad > 0 && min_bad < good_sectors) {
2423                 /* only resync enough to reach the next bad->good
2424                  * transition */
2425                 good_sectors = min_bad;
2426         }
2427
2428         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2429                 /* extra read targets are also write targets */
2430                 write_targets += read_targets-1;
2431
2432         if (write_targets == 0 || read_targets == 0) {
2433                 /* There is nowhere to write, so all non-sync
2434                  * drives must be failed - so we are finished
2435                  */
2436                 sector_t rv = max_sector - sector_nr;
2437                 *skipped = 1;
2438                 put_buf(r1_bio);
2439                 return rv;
2440         }
2441
2442         if (max_sector > mddev->resync_max)
2443                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2444         if (max_sector > sector_nr + good_sectors)
2445                 max_sector = sector_nr + good_sectors;
2446         nr_sectors = 0;
2447         sync_blocks = 0;
2448         do {
2449                 struct page *page;
2450                 int len = PAGE_SIZE;
2451                 if (sector_nr + (len>>9) > max_sector)
2452                         len = (max_sector - sector_nr) << 9;
2453                 if (len == 0)
2454                         break;
2455                 if (sync_blocks == 0) {
2456                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2457                                                &sync_blocks, still_degraded) &&
2458                             !conf->fullsync &&
2459                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2460                                 break;
2461                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2462                         if ((len >> 9) > sync_blocks)
2463                                 len = sync_blocks<<9;
2464                 }
2465
2466                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2467                         bio = r1_bio->bios[i];
2468                         if (bio->bi_end_io) {
2469                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2470                                 if (bio_add_page(bio, page, len, 0) == 0) {
2471                                         /* stop here */
2472                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2473                                         while (i > 0) {
2474                                                 i--;
2475                                                 bio = r1_bio->bios[i];
2476                                                 if (bio->bi_end_io==NULL)
2477                                                         continue;
2478                                                 /* remove last page from this bio */
2479                                                 bio->bi_vcnt--;
2480                                                 bio->bi_size -= len;
2481                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2482                                         }
2483                                         goto bio_full;
2484                                 }
2485                         }
2486                 }
2487                 nr_sectors += len>>9;
2488                 sector_nr += len>>9;
2489                 sync_blocks -= (len>>9);
2490         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2491  bio_full:
2492         r1_bio->sectors = nr_sectors;
2493
2494         /* For a user-requested sync, we read all readable devices and do a
2495          * compare
2496          */
2497         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2498                 atomic_set(&r1_bio->remaining, read_targets);
2499                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2500                         bio = r1_bio->bios[i];
2501                         if (bio->bi_end_io == end_sync_read) {
2502                                 read_targets--;
2503                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2504                                 generic_make_request(bio);
2505                         }
2506                 }
2507         } else {
2508                 atomic_set(&r1_bio->remaining, 1);
2509                 bio = r1_bio->bios[r1_bio->read_disk];
2510                 md_sync_acct(bio->bi_bdev, nr_sectors);
2511                 generic_make_request(bio);
2512
2513         }
2514         return nr_sectors;
2515 }
2516
2517 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2518 {
2519         if (sectors)
2520                 return sectors;
2521
2522         return mddev->dev_sectors;
2523 }
2524
2525 static struct r1conf *setup_conf(struct mddev *mddev)
2526 {
2527         struct r1conf *conf;
2528         int i;
2529         struct raid1_info *disk;
2530         struct md_rdev *rdev;
2531         int err = -ENOMEM;
2532
2533         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2534         if (!conf)
2535                 goto abort;
2536
2537         conf->mirrors = kzalloc(sizeof(struct raid1_info)
2538                                 * mddev->raid_disks * 2,
2539                                  GFP_KERNEL);
2540         if (!conf->mirrors)
2541                 goto abort;
2542
2543         conf->tmppage = alloc_page(GFP_KERNEL);
2544         if (!conf->tmppage)
2545                 goto abort;
2546
2547         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2548         if (!conf->poolinfo)
2549                 goto abort;
2550         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2551         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2552                                           r1bio_pool_free,
2553                                           conf->poolinfo);
2554         if (!conf->r1bio_pool)
2555                 goto abort;
2556
2557         conf->poolinfo->mddev = mddev;
2558
2559         err = -EINVAL;
2560         spin_lock_init(&conf->device_lock);
2561         rdev_for_each(rdev, mddev) {
2562                 struct request_queue *q;
2563                 int disk_idx = rdev->raid_disk;
2564                 if (disk_idx >= mddev->raid_disks
2565                     || disk_idx < 0)
2566                         continue;
2567                 if (test_bit(Replacement, &rdev->flags))
2568                         disk = conf->mirrors + conf->raid_disks + disk_idx;
2569                 else
2570                         disk = conf->mirrors + disk_idx;
2571
2572                 if (disk->rdev)
2573                         goto abort;
2574                 disk->rdev = rdev;
2575                 q = bdev_get_queue(rdev->bdev);
2576                 if (q->merge_bvec_fn)
2577                         mddev->merge_check_needed = 1;
2578
2579                 disk->head_position = 0;
2580         }
2581         conf->raid_disks = mddev->raid_disks;
2582         conf->mddev = mddev;
2583         INIT_LIST_HEAD(&conf->retry_list);
2584
2585         spin_lock_init(&conf->resync_lock);
2586         init_waitqueue_head(&conf->wait_barrier);
2587
2588         bio_list_init(&conf->pending_bio_list);
2589         conf->pending_count = 0;
2590         conf->recovery_disabled = mddev->recovery_disabled - 1;
2591
2592         err = -EIO;
2593         for (i = 0; i < conf->raid_disks * 2; i++) {
2594
2595                 disk = conf->mirrors + i;
2596
2597                 if (i < conf->raid_disks &&
2598                     disk[conf->raid_disks].rdev) {
2599                         /* This slot has a replacement. */
2600                         if (!disk->rdev) {
2601                                 /* No original, just make the replacement
2602                                  * a recovering spare
2603                                  */
2604                                 disk->rdev =
2605                                         disk[conf->raid_disks].rdev;
2606                                 disk[conf->raid_disks].rdev = NULL;
2607                         } else if (!test_bit(In_sync, &disk->rdev->flags))
2608                                 /* Original is not in_sync - bad */
2609                                 goto abort;
2610                 }
2611
2612                 if (!disk->rdev ||
2613                     !test_bit(In_sync, &disk->rdev->flags)) {
2614                         disk->head_position = 0;
2615                         if (disk->rdev &&
2616                             (disk->rdev->saved_raid_disk < 0))
2617                                 conf->fullsync = 1;
2618                 }
2619         }
2620
2621         err = -ENOMEM;
2622         conf->thread = md_register_thread(raid1d, mddev, "raid1");
2623         if (!conf->thread) {
2624                 printk(KERN_ERR
2625                        "md/raid1:%s: couldn't allocate thread\n",
2626                        mdname(mddev));
2627                 goto abort;
2628         }
2629
2630         return conf;
2631
2632  abort:
2633         if (conf) {
2634                 if (conf->r1bio_pool)
2635                         mempool_destroy(conf->r1bio_pool);
2636                 kfree(conf->mirrors);
2637                 safe_put_page(conf->tmppage);
2638                 kfree(conf->poolinfo);
2639                 kfree(conf);
2640         }
2641         return ERR_PTR(err);
2642 }
2643
2644 static int stop(struct mddev *mddev);
2645 static int run(struct mddev *mddev)
2646 {
2647         struct r1conf *conf;
2648         int i;
2649         struct md_rdev *rdev;
2650         int ret;
2651
2652         if (mddev->level != 1) {
2653                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2654                        mdname(mddev), mddev->level);
2655                 return -EIO;
2656         }
2657         if (mddev->reshape_position != MaxSector) {
2658                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2659                        mdname(mddev));
2660                 return -EIO;
2661         }
2662         /*
2663          * copy the already verified devices into our private RAID1
2664          * bookkeeping area. [whatever we allocate in run(),
2665          * should be freed in stop()]
2666          */
2667         if (mddev->private == NULL)
2668                 conf = setup_conf(mddev);
2669         else
2670                 conf = mddev->private;
2671
2672         if (IS_ERR(conf))
2673                 return PTR_ERR(conf);
2674
2675         rdev_for_each(rdev, mddev) {
2676                 if (!mddev->gendisk)
2677                         continue;
2678                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2679                                   rdev->data_offset << 9);
2680         }
2681
2682         mddev->degraded = 0;
2683         for (i=0; i < conf->raid_disks; i++)
2684                 if (conf->mirrors[i].rdev == NULL ||
2685                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2686                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2687                         mddev->degraded++;
2688
2689         if (conf->raid_disks - mddev->degraded == 1)
2690                 mddev->recovery_cp = MaxSector;
2691
2692         if (mddev->recovery_cp != MaxSector)
2693                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2694                        " -- starting background reconstruction\n",
2695                        mdname(mddev));
2696         printk(KERN_INFO
2697                 "md/raid1:%s: active with %d out of %d mirrors\n",
2698                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2699                 mddev->raid_disks);
2700
2701         /*
2702          * Ok, everything is just fine now
2703          */
2704         mddev->thread = conf->thread;
2705         conf->thread = NULL;
2706         mddev->private = conf;
2707
2708         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2709
2710         if (mddev->queue) {
2711                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2712                 mddev->queue->backing_dev_info.congested_data = mddev;
2713                 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2714         }
2715
2716         ret =  md_integrity_register(mddev);
2717         if (ret)
2718                 stop(mddev);
2719         return ret;
2720 }
2721
2722 static int stop(struct mddev *mddev)
2723 {
2724         struct r1conf *conf = mddev->private;
2725         struct bitmap *bitmap = mddev->bitmap;
2726
2727         /* wait for behind writes to complete */
2728         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2729                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2730                        mdname(mddev));
2731                 /* need to kick something here to make sure I/O goes? */
2732                 wait_event(bitmap->behind_wait,
2733                            atomic_read(&bitmap->behind_writes) == 0);
2734         }
2735
2736         raise_barrier(conf);
2737         lower_barrier(conf);
2738
2739         md_unregister_thread(&mddev->thread);
2740         if (conf->r1bio_pool)
2741                 mempool_destroy(conf->r1bio_pool);
2742         kfree(conf->mirrors);
2743         kfree(conf->poolinfo);
2744         kfree(conf);
2745         mddev->private = NULL;
2746         return 0;
2747 }
2748
2749 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2750 {
2751         /* no resync is happening, and there is enough space
2752          * on all devices, so we can resize.
2753          * We need to make sure resync covers any new space.
2754          * If the array is shrinking we should possibly wait until
2755          * any io in the removed space completes, but it hardly seems
2756          * worth it.
2757          */
2758         sector_t newsize = raid1_size(mddev, sectors, 0);
2759         if (mddev->external_size &&
2760             mddev->array_sectors > newsize)
2761                 return -EINVAL;
2762         if (mddev->bitmap) {
2763                 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2764                 if (ret)
2765                         return ret;
2766         }
2767         md_set_array_sectors(mddev, newsize);
2768         set_capacity(mddev->gendisk, mddev->array_sectors);
2769         revalidate_disk(mddev->gendisk);
2770         if (sectors > mddev->dev_sectors &&
2771             mddev->recovery_cp > mddev->dev_sectors) {
2772                 mddev->recovery_cp = mddev->dev_sectors;
2773                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2774         }
2775         mddev->dev_sectors = sectors;
2776         mddev->resync_max_sectors = sectors;
2777         return 0;
2778 }
2779
2780 static int raid1_reshape(struct mddev *mddev)
2781 {
2782         /* We need to:
2783          * 1/ resize the r1bio_pool
2784          * 2/ resize conf->mirrors
2785          *
2786          * We allocate a new r1bio_pool if we can.
2787          * Then raise a device barrier and wait until all IO stops.
2788          * Then resize conf->mirrors and swap in the new r1bio pool.
2789          *
2790          * At the same time, we "pack" the devices so that all the missing
2791          * devices have the higher raid_disk numbers.
2792          */
2793         mempool_t *newpool, *oldpool;
2794         struct pool_info *newpoolinfo;
2795         struct raid1_info *newmirrors;
2796         struct r1conf *conf = mddev->private;
2797         int cnt, raid_disks;
2798         unsigned long flags;
2799         int d, d2, err;
2800
2801         /* Cannot change chunk_size, layout, or level */
2802         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2803             mddev->layout != mddev->new_layout ||
2804             mddev->level != mddev->new_level) {
2805                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2806                 mddev->new_layout = mddev->layout;
2807                 mddev->new_level = mddev->level;
2808                 return -EINVAL;
2809         }
2810
2811         err = md_allow_write(mddev);
2812         if (err)
2813                 return err;
2814
2815         raid_disks = mddev->raid_disks + mddev->delta_disks;
2816
2817         if (raid_disks < conf->raid_disks) {
2818                 cnt=0;
2819                 for (d= 0; d < conf->raid_disks; d++)
2820                         if (conf->mirrors[d].rdev)
2821                                 cnt++;
2822                 if (cnt > raid_disks)
2823                         return -EBUSY;
2824         }
2825
2826         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2827         if (!newpoolinfo)
2828                 return -ENOMEM;
2829         newpoolinfo->mddev = mddev;
2830         newpoolinfo->raid_disks = raid_disks * 2;
2831
2832         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2833                                  r1bio_pool_free, newpoolinfo);
2834         if (!newpool) {
2835                 kfree(newpoolinfo);
2836                 return -ENOMEM;
2837         }
2838         newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
2839                              GFP_KERNEL);
2840         if (!newmirrors) {
2841                 kfree(newpoolinfo);
2842                 mempool_destroy(newpool);
2843                 return -ENOMEM;
2844         }
2845
2846         raise_barrier(conf);
2847
2848         /* ok, everything is stopped */
2849         oldpool = conf->r1bio_pool;
2850         conf->r1bio_pool = newpool;
2851
2852         for (d = d2 = 0; d < conf->raid_disks; d++) {
2853                 struct md_rdev *rdev = conf->mirrors[d].rdev;
2854                 if (rdev && rdev->raid_disk != d2) {
2855                         sysfs_unlink_rdev(mddev, rdev);
2856                         rdev->raid_disk = d2;
2857                         sysfs_unlink_rdev(mddev, rdev);
2858                         if (sysfs_link_rdev(mddev, rdev))
2859                                 printk(KERN_WARNING
2860                                        "md/raid1:%s: cannot register rd%d\n",
2861                                        mdname(mddev), rdev->raid_disk);
2862                 }
2863                 if (rdev)
2864                         newmirrors[d2++].rdev = rdev;
2865         }
2866         kfree(conf->mirrors);
2867         conf->mirrors = newmirrors;
2868         kfree(conf->poolinfo);
2869         conf->poolinfo = newpoolinfo;
2870
2871         spin_lock_irqsave(&conf->device_lock, flags);
2872         mddev->degraded += (raid_disks - conf->raid_disks);
2873         spin_unlock_irqrestore(&conf->device_lock, flags);
2874         conf->raid_disks = mddev->raid_disks = raid_disks;
2875         mddev->delta_disks = 0;
2876
2877         lower_barrier(conf);
2878
2879         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2880         md_wakeup_thread(mddev->thread);
2881
2882         mempool_destroy(oldpool);
2883         return 0;
2884 }
2885
2886 static void raid1_quiesce(struct mddev *mddev, int state)
2887 {
2888         struct r1conf *conf = mddev->private;
2889
2890         switch(state) {
2891         case 2: /* wake for suspend */
2892                 wake_up(&conf->wait_barrier);
2893                 break;
2894         case 1:
2895                 raise_barrier(conf);
2896                 break;
2897         case 0:
2898                 lower_barrier(conf);
2899                 break;
2900         }
2901 }
2902
2903 static void *raid1_takeover(struct mddev *mddev)
2904 {
2905         /* raid1 can take over:
2906          *  raid5 with 2 devices, any layout or chunk size
2907          */
2908         if (mddev->level == 5 && mddev->raid_disks == 2) {
2909                 struct r1conf *conf;
2910                 mddev->new_level = 1;
2911                 mddev->new_layout = 0;
2912                 mddev->new_chunk_sectors = 0;
2913                 conf = setup_conf(mddev);
2914                 if (!IS_ERR(conf))
2915                         conf->barrier = 1;
2916                 return conf;
2917         }
2918         return ERR_PTR(-EINVAL);
2919 }
2920
2921 static struct md_personality raid1_personality =
2922 {
2923         .name           = "raid1",
2924         .level          = 1,
2925         .owner          = THIS_MODULE,
2926         .make_request   = make_request,
2927         .run            = run,
2928         .stop           = stop,
2929         .status         = status,
2930         .error_handler  = error,
2931         .hot_add_disk   = raid1_add_disk,
2932         .hot_remove_disk= raid1_remove_disk,
2933         .spare_active   = raid1_spare_active,
2934         .sync_request   = sync_request,
2935         .resize         = raid1_resize,
2936         .size           = raid1_size,
2937         .check_reshape  = raid1_reshape,
2938         .quiesce        = raid1_quiesce,
2939         .takeover       = raid1_takeover,
2940 };
2941
2942 static int __init raid_init(void)
2943 {
2944         return register_md_personality(&raid1_personality);
2945 }
2946
2947 static void raid_exit(void)
2948 {
2949         unregister_md_personality(&raid1_personality);
2950 }
2951
2952 module_init(raid_init);
2953 module_exit(raid_exit);
2954 MODULE_LICENSE("GPL");
2955 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2956 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2957 MODULE_ALIAS("md-raid1");
2958 MODULE_ALIAS("md-level-1");
2959
2960 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);