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 r1bio *r1_bio;
  96         struct bio *bio;
  97         int i, j;
  98
  99         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 100         if (!r1_bio)
 101                 return NULL;
 102
 103         /*
 104          * Allocate bios : 1 for reading, n-1 for writing
 105          */
 106         for (j = pi->raid_disks ; j-- ; ) {
 107                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 108                 if (!bio)
 109                         goto out_free_bio;
 110                 r1_bio->bios[j] = bio;
 111         }
 112         /*
 113          * Allocate RESYNC_PAGES data pages and attach them to
 114          * the first bio.
 115          * If this is a user-requested check/repair, allocate
 116          * RESYNC_PAGES for each bio.
 117          */
 118         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 119                 j = pi->raid_disks;
 120         else
 121                 j = 1;
 122         while(j--) {
 123                 bio = r1_bio->bios[j];
 124                 bio->bi_vcnt = RESYNC_PAGES;
 125
 126                 if (bio_alloc_pages(bio, gfp_flags))
 127                         goto out_free_bio;
 128         }
 129         /* If not user-requests, copy the page pointers to all bios */
 130         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 131                 for (i=0; i<RESYNC_PAGES ; i++)
 132                         for (j=1; j<pi->raid_disks; j++)
 133                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 134                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 135         }
 136
 137         r1_bio->master_bio = NULL;
 138
 139         return r1_bio;
 140
 141 out_free_bio:
 142         while (++j < pi->raid_disks)
 143                 bio_put(r1_bio->bios[j]);
 144         r1bio_pool_free(r1_bio, data);
 145         return NULL;
 146 }
 147
 148 static void r1buf_pool_free(void *__r1_bio, void *data)
 149 {
 150         struct pool_info *pi = data;
 151         int i,j;
 152         struct r1bio *r1bio = __r1_bio;
 153
 154         for (i = 0; i < RESYNC_PAGES; i++)
 155                 for (j = pi->raid_disks; j-- ;) {
 156                         if (j == 0 ||
 157                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 158                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 159                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 160                 }
 161         for (i=0 ; i < pi->raid_disks; i++)
 162                 bio_put(r1bio->bios[i]);
 163
 164         r1bio_pool_free(r1bio, data);
 165 }
 166
 167 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 168 {
 169         int i;
 170
 171         for (i = 0; i < conf->raid_disks * 2; i++) {
 172                 struct bio **bio = r1_bio->bios + i;
 173                 if (!BIO_SPECIAL(*bio))
 174                         bio_put(*bio);
 175                 *bio = NULL;
 176         }
 177 }
 178
 179 static void free_r1bio(struct r1bio *r1_bio)
 180 {
 181         struct r1conf *conf = r1_bio->mddev->private;
 182
 183         put_all_bios(conf, r1_bio);
 184         mempool_free(r1_bio, conf->r1bio_pool);
 185 }
 186
 187 static void put_buf(struct r1bio *r1_bio)
 188 {
 189         struct r1conf *conf = r1_bio->mddev->private;
 190         int i;
 191
 192         for (i = 0; i < conf->raid_disks * 2; i++) {
 193                 struct bio *bio = r1_bio->bios[i];
 194                 if (bio->bi_end_io)
 195                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 196         }
 197
 198         mempool_free(r1_bio, conf->r1buf_pool);
 199
 200         lower_barrier(conf);
 201 }
 202
 203 static void reschedule_retry(struct r1bio *r1_bio)
 204 {
 205         unsigned long flags;
 206         struct mddev *mddev = r1_bio->mddev;
 207         struct r1conf *conf = mddev->private;
 208
 209         spin_lock_irqsave(&conf->device_lock, flags);
 210         list_add(&r1_bio->retry_list, &conf->retry_list);
 211         conf->nr_queued ++;
 212         spin_unlock_irqrestore(&conf->device_lock, flags);
 213
 214         wake_up(&conf->wait_barrier);
 215         md_wakeup_thread(mddev->thread);
 216 }
 217
 218 /*
 219  * raid_end_bio_io() is called when we have finished servicing a mirrored
 220  * operation and are ready to return a success/failure code to the buffer
 221  * cache layer.
 222  */
 223 static void call_bio_endio(struct r1bio *r1_bio)
 224 {
 225         struct bio *bio = r1_bio->master_bio;
 226         int done;
 227         struct r1conf *conf = r1_bio->mddev->private;
 228
 229         if (bio->bi_phys_segments) {
 230                 unsigned long flags;
 231                 spin_lock_irqsave(&conf->device_lock, flags);
 232                 bio->bi_phys_segments--;
 233                 done = (bio->bi_phys_segments == 0);
 234                 spin_unlock_irqrestore(&conf->device_lock, flags);
 235         } else
 236                 done = 1;
 237
 238         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 239                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 240         if (done) {
 241                 bio_endio(bio, 0);
 242                 /*
 243                  * Wake up any possible resync thread that waits for the device
 244                  * to go idle.
 245                  */
 246                 allow_barrier(conf);
 247         }
 248 }
 249
 250 static void raid_end_bio_io(struct r1bio *r1_bio)
 251 {
 252         struct bio *bio = r1_bio->master_bio;
 253
 254         /* if nobody has done the final endio yet, do it now */
 255         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 256                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 257                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 258                          (unsigned long long) bio->bi_sector,
 259                          (unsigned long long) bio->bi_sector +
 260                          bio_sectors(bio) - 1);
 261
 262                 call_bio_endio(r1_bio);
 263         }
 264         free_r1bio(r1_bio);
 265 }
 266
 267 /*
 268  * Update disk head position estimator based on IRQ completion info.
 269  */
 270 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 271 {
 272         struct r1conf *conf = r1_bio->mddev->private;
 273
 274         conf->mirrors[disk].head_position =
 275                 r1_bio->sector + (r1_bio->sectors);
 276 }
 277
 278 /*
 279  * Find the disk number which triggered given bio
 280  */
 281 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 282 {
 283         int mirror;
 284         struct r1conf *conf = r1_bio->mddev->private;
 285         int raid_disks = conf->raid_disks;
 286
 287         for (mirror = 0; mirror < raid_disks * 2; mirror++)
 288                 if (r1_bio->bios[mirror] == bio)
 289                         break;
 290
 291         BUG_ON(mirror == raid_disks * 2);
 292         update_head_pos(mirror, r1_bio);
 293
 294         return mirror;
 295 }
 296
 297 static void raid1_end_read_request(struct bio *bio, int error)
 298 {
 299         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 300         struct r1bio *r1_bio = bio->bi_private;
 301         int mirror;
 302         struct r1conf *conf = r1_bio->mddev->private;
 303
 304         mirror = r1_bio->read_disk;
 305         /*
 306          * this branch is our 'one mirror IO has finished' event handler:
 307          */
 308         update_head_pos(mirror, r1_bio);
 309
 310         if (uptodate)
 311                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 312         else {
 313                 /* If all other devices have failed, we want to return
 314                  * the error upwards rather than fail the last device.
 315                  * Here we redefine "uptodate" to mean "Don't want to retry"
 316                  */
 317                 unsigned long flags;
 318                 spin_lock_irqsave(&conf->device_lock, flags);
 319                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 320                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 321                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 322                         uptodate = 1;
 323                 spin_unlock_irqrestore(&conf->device_lock, flags);
 324         }
 325
 326         if (uptodate) {
 327                 raid_end_bio_io(r1_bio);
 328                 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 329         } else {
 330                 /*
 331                  * oops, read error:
 332                  */
 333                 char b[BDEVNAME_SIZE];
 334                 printk_ratelimited(
 335                         KERN_ERR "md/raid1:%s: %s: "
 336                         "rescheduling sector %llu\n",
 337                         mdname(conf->mddev),
 338                         bdevname(conf->mirrors[mirror].rdev->bdev,
 339                                  b),
 340                         (unsigned long long)r1_bio->sector);
 341                 set_bit(R1BIO_ReadError, &r1_bio->state);
 342                 reschedule_retry(r1_bio);
 343                 /* don't drop the reference on read_disk yet */
 344         }
 345 }
 346
 347 static void close_write(struct r1bio *r1_bio)
 348 {
 349         /* it really is the end of this request */
 350         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 351                 /* free extra copy of the data pages */
 352                 int i = r1_bio->behind_page_count;
 353                 while (i--)
 354                         safe_put_page(r1_bio->behind_bvecs[i].bv_page);
 355                 kfree(r1_bio->behind_bvecs);
 356                 r1_bio->behind_bvecs = NULL;
 357         }
 358         /* clear the bitmap if all writes complete successfully */
 359         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 360                         r1_bio->sectors,
 361                         !test_bit(R1BIO_Degraded, &r1_bio->state),
 362                         test_bit(R1BIO_BehindIO, &r1_bio->state));
 363         md_write_end(r1_bio->mddev);
 364 }
 365
 366 static void r1_bio_write_done(struct r1bio *r1_bio)
 367 {
 368         if (!atomic_dec_and_test(&r1_bio->remaining))
 369                 return;
 370
 371         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 372                 reschedule_retry(r1_bio);
 373         else {
 374                 close_write(r1_bio);
 375                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 376                         reschedule_retry(r1_bio);
 377                 else
 378                         raid_end_bio_io(r1_bio);
 379         }
 380 }
 381
 382 static void raid1_end_write_request(struct bio *bio, int error)
 383 {
 384         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 385         struct r1bio *r1_bio = bio->bi_private;
 386         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 387         struct r1conf *conf = r1_bio->mddev->private;
 388         struct bio *to_put = NULL;
 389
 390         mirror = find_bio_disk(r1_bio, bio);
 391
 392         /*
 393          * 'one mirror IO has finished' event handler:
 394          */
 395         if (!uptodate) {
 396                 set_bit(WriteErrorSeen,
 397                         &conf->mirrors[mirror].rdev->flags);
 398                 if (!test_and_set_bit(WantReplacement,
 399                                       &conf->mirrors[mirror].rdev->flags))
 400                         set_bit(MD_RECOVERY_NEEDED, &
 401                                 conf->mddev->recovery);
 402
 403                 set_bit(R1BIO_WriteError, &r1_bio->state);
 404         } else {
 405                 /*
 406                  * Set R1BIO_Uptodate in our master bio, so that we
 407                  * will return a good error code for to the higher
 408                  * levels even if IO on some other mirrored buffer
 409                  * fails.
 410                  *
 411                  * The 'master' represents the composite IO operation
 412                  * to user-side. So if something waits for IO, then it
 413                  * will wait for the 'master' bio.
 414                  */
 415                 sector_t first_bad;
 416                 int bad_sectors;
 417
 418                 r1_bio->bios[mirror] = NULL;
 419                 to_put = bio;
 420                 /*
 421                  * Do not set R1BIO_Uptodate if the current device is
 422                  * rebuilding or Faulty. This is because we cannot use
 423                  * such device for properly reading the data back (we could
 424                  * potentially use it, if the current write would have felt
 425                  * before rdev->recovery_offset, but for simplicity we don't
 426                  * check this here.
 427                  */
 428                 if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) &&
 429                     !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))
 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                                          bio_sectors(mbio) - 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, best_dist_disk, best_pending_disk;
 501         int has_nonrot_disk;
 502         int disk;
 503         sector_t best_dist;
 504         unsigned int min_pending;
 505         struct md_rdev *rdev;
 506         int choose_first;
 507         int choose_next_idle;
 508
 509         rcu_read_lock();
 510         /*
 511          * Check if we can balance. We can balance on the whole
 512          * device if no resync is going on, or below the resync window.
 513          * We take the first readable disk when above the resync window.
 514          */
 515  retry:
 516         sectors = r1_bio->sectors;
 517         best_disk = -1;
 518         best_dist_disk = -1;
 519         best_dist = MaxSector;
 520         best_pending_disk = -1;
 521         min_pending = UINT_MAX;
 522         best_good_sectors = 0;
 523         has_nonrot_disk = 0;
 524         choose_next_idle = 0;
 525
 526         if (conf->mddev->recovery_cp < MaxSector &&
 527             (this_sector + sectors >= conf->next_resync))
 528                 choose_first = 1;
 529         else
 530                 choose_first = 0;
 531
 532         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 533                 sector_t dist;
 534                 sector_t first_bad;
 535                 int bad_sectors;
 536                 unsigned int pending;
 537                 bool nonrot;
 538
 539                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 540                 if (r1_bio->bios[disk] == IO_BLOCKED
 541                     || rdev == NULL
 542                     || test_bit(Unmerged, &rdev->flags)
 543                     || test_bit(Faulty, &rdev->flags))
 544                         continue;
 545                 if (!test_bit(In_sync, &rdev->flags) &&
 546                     rdev->recovery_offset < this_sector + sectors)
 547                         continue;
 548                 if (test_bit(WriteMostly, &rdev->flags)) {
 549                         /* Don't balance among write-mostly, just
 550                          * use the first as a last resort */
 551                         if (best_disk < 0) {
 552                                 if (is_badblock(rdev, this_sector, sectors,
 553                                                 &first_bad, &bad_sectors)) {
 554                                         if (first_bad < this_sector)
 555                                                 /* Cannot use this */
 556                                                 continue;
 557                                         best_good_sectors = first_bad - this_sector;
 558                                 } else
 559                                         best_good_sectors = sectors;
 560                                 best_disk = disk;
 561                         }
 562                         continue;
 563                 }
 564                 /* This is a reasonable device to use.  It might
 565                  * even be best.
 566                  */
 567                 if (is_badblock(rdev, this_sector, sectors,
 568                                 &first_bad, &bad_sectors)) {
 569                         if (best_dist < MaxSector)
 570                                 /* already have a better device */
 571                                 continue;
 572                         if (first_bad <= this_sector) {
 573                                 /* cannot read here. If this is the 'primary'
 574                                  * device, then we must not read beyond
 575                                  * bad_sectors from another device..
 576                                  */
 577                                 bad_sectors -= (this_sector - first_bad);
 578                                 if (choose_first && sectors > bad_sectors)
 579                                         sectors = bad_sectors;
 580                                 if (best_good_sectors > sectors)
 581                                         best_good_sectors = sectors;
 582
 583                         } else {
 584                                 sector_t good_sectors = first_bad - this_sector;
 585                                 if (good_sectors > best_good_sectors) {
 586                                         best_good_sectors = good_sectors;
 587                                         best_disk = disk;
 588                                 }
 589                                 if (choose_first)
 590                                         break;
 591                         }
 592                         continue;
 593                 } else
 594                         best_good_sectors = sectors;
 595
 596                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 597                 has_nonrot_disk |= nonrot;
 598                 pending = atomic_read(&rdev->nr_pending);
 599                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 600                 if (choose_first) {
 601                         best_disk = disk;
 602                         break;
 603                 }
 604                 /* Don't change to another disk for sequential reads */
 605                 if (conf->mirrors[disk].next_seq_sect == this_sector
 606                     || dist == 0) {
 607                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 608                         struct raid1_info *mirror = &conf->mirrors[disk];
 609
 610                         best_disk = disk;
 611                         /*
 612                          * If buffered sequential IO size exceeds optimal
 613                          * iosize, check if there is idle disk. If yes, choose
 614                          * the idle disk. read_balance could already choose an
 615                          * idle disk before noticing it's a sequential IO in
 616                          * this disk. This doesn't matter because this disk
 617                          * will idle, next time it will be utilized after the
 618                          * first disk has IO size exceeds optimal iosize. In
 619                          * this way, iosize of the first disk will be optimal
 620                          * iosize at least. iosize of the second disk might be
 621                          * small, but not a big deal since when the second disk
 622                          * starts IO, the first disk is likely still busy.
 623                          */
 624                         if (nonrot && opt_iosize > 0 &&
 625                             mirror->seq_start != MaxSector &&
 626                             mirror->next_seq_sect > opt_iosize &&
 627                             mirror->next_seq_sect - opt_iosize >=
 628                             mirror->seq_start) {
 629                                 choose_next_idle = 1;
 630                                 continue;
 631                         }
 632                         break;
 633                 }
 634                 /* If device is idle, use it */
 635                 if (pending == 0) {
 636                         best_disk = disk;
 637                         break;
 638                 }
 639
 640                 if (choose_next_idle)
 641                         continue;
 642
 643                 if (min_pending > pending) {
 644                         min_pending = pending;
 645                         best_pending_disk = disk;
 646                 }
 647
 648                 if (dist < best_dist) {
 649                         best_dist = dist;
 650                         best_dist_disk = disk;
 651                 }
 652         }
 653
 654         /*
 655          * If all disks are rotational, choose the closest disk. If any disk is
 656          * non-rotational, choose the disk with less pending request even the
 657          * disk is rotational, which might/might not be optimal for raids with
 658          * mixed ratation/non-rotational disks depending on workload.
 659          */
 660         if (best_disk == -1) {
 661                 if (has_nonrot_disk)
 662                         best_disk = best_pending_disk;
 663                 else
 664                         best_disk = best_dist_disk;
 665         }
 666
 667         if (best_disk >= 0) {
 668                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 669                 if (!rdev)
 670                         goto retry;
 671                 atomic_inc(&rdev->nr_pending);
 672                 if (test_bit(Faulty, &rdev->flags)) {
 673                         /* cannot risk returning a device that failed
 674                          * before we inc'ed nr_pending
 675                          */
 676                         rdev_dec_pending(rdev, conf->mddev);
 677                         goto retry;
 678                 }
 679                 sectors = best_good_sectors;
 680
 681                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 682                         conf->mirrors[best_disk].seq_start = this_sector;
 683
 684                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 685         }
 686         rcu_read_unlock();
 687         *max_sectors = sectors;
 688
 689         return best_disk;
 690 }
 691
 692 static int raid1_mergeable_bvec(struct request_queue *q,
 693                                 struct bvec_merge_data *bvm,
 694                                 struct bio_vec *biovec)
 695 {
 696         struct mddev *mddev = q->queuedata;
 697         struct r1conf *conf = mddev->private;
 698         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 699         int max = biovec->bv_len;
 700
 701         if (mddev->merge_check_needed) {
 702                 int disk;
 703                 rcu_read_lock();
 704                 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
 705                         struct md_rdev *rdev = rcu_dereference(
 706                                 conf->mirrors[disk].rdev);
 707                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
 708                                 struct request_queue *q =
 709                                         bdev_get_queue(rdev->bdev);
 710                                 if (q->merge_bvec_fn) {
 711                                         bvm->bi_sector = sector +
 712                                                 rdev->data_offset;
 713                                         bvm->bi_bdev = rdev->bdev;
 714                                         max = min(max, q->merge_bvec_fn(
 715                                                           q, bvm, biovec));
 716                                 }
 717                         }
 718                 }
 719                 rcu_read_unlock();
 720         }
 721         return max;
 722
 723 }
 724
 725 int md_raid1_congested(struct mddev *mddev, int bits)
 726 {
 727         struct r1conf *conf = mddev->private;
 728         int i, ret = 0;
 729
 730         if ((bits & (1 << BDI_async_congested)) &&
 731             conf->pending_count >= max_queued_requests)
 732                 return 1;
 733
 734         rcu_read_lock();
 735         for (i = 0; i < conf->raid_disks * 2; i++) {
 736                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 737                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 738                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 739
 740                         BUG_ON(!q);
 741
 742                         /* Note the '|| 1' - when read_balance prefers
 743                          * non-congested targets, it can be removed
 744                          */
 745                         if ((bits & (1<<BDI_async_congested)) || 1)
 746                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 747                         else
 748                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 749                 }
 750         }
 751         rcu_read_unlock();
 752         return ret;
 753 }
 754 EXPORT_SYMBOL_GPL(md_raid1_congested);
 755
 756 static int raid1_congested(void *data, int bits)
 757 {
 758         struct mddev *mddev = data;
 759
 760         return mddev_congested(mddev, bits) ||
 761                 md_raid1_congested(mddev, bits);
 762 }
 763
 764 static void flush_pending_writes(struct r1conf *conf)
 765 {
 766         /* Any writes that have been queued but are awaiting
 767          * bitmap updates get flushed here.
 768          */
 769         spin_lock_irq(&conf->device_lock);
 770
 771         if (conf->pending_bio_list.head) {
 772                 struct bio *bio;
 773                 bio = bio_list_get(&conf->pending_bio_list);
 774                 conf->pending_count = 0;
 775                 spin_unlock_irq(&conf->device_lock);
 776                 /* flush any pending bitmap writes to
 777                  * disk before proceeding w/ I/O */
 778                 bitmap_unplug(conf->mddev->bitmap);
 779                 wake_up(&conf->wait_barrier);
 780
 781                 while (bio) { /* submit pending writes */
 782                         struct bio *next = bio->bi_next;
 783                         bio->bi_next = NULL;
 784                         if (unlikely((bio->bi_rw & REQ_DISCARD) &&
 785                             !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 786                                 /* Just ignore it */
 787                                 bio_endio(bio, 0);
 788                         else
 789                                 generic_make_request(bio);
 790                         bio = next;
 791                 }
 792         } else
 793                 spin_unlock_irq(&conf->device_lock);
 794 }
 795
 796 /* Barriers....
 797  * Sometimes we need to suspend IO while we do something else,
 798  * either some resync/recovery, or reconfigure the array.
 799  * To do this we raise a 'barrier'.
 800  * The 'barrier' is a counter that can be raised multiple times
 801  * to count how many activities are happening which preclude
 802  * normal IO.
 803  * We can only raise the barrier if there is no pending IO.
 804  * i.e. if nr_pending == 0.
 805  * We choose only to raise the barrier if no-one is waiting for the
 806  * barrier to go down.  This means that as soon as an IO request
 807  * is ready, no other operations which require a barrier will start
 808  * until the IO request has had a chance.
 809  *
 810  * So: regular IO calls 'wait_barrier'.  When that returns there
 811  *    is no backgroup IO happening,  It must arrange to call
 812  *    allow_barrier when it has finished its IO.
 813  * backgroup IO calls must call raise_barrier.  Once that returns
 814  *    there is no normal IO happeing.  It must arrange to call
 815  *    lower_barrier when the particular background IO completes.
 816  */
 817 #define RESYNC_DEPTH 32
 818
 819 static void raise_barrier(struct r1conf *conf)
 820 {
 821         spin_lock_irq(&conf->resync_lock);
 822
 823         /* Wait until no block IO is waiting */
 824         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 825                             conf->resync_lock);
 826
 827         /* block any new IO from starting */
 828         conf->barrier++;
 829
 830         /* Now wait for all pending IO to complete */
 831         wait_event_lock_irq(conf->wait_barrier,
 832                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 833                             conf->resync_lock);
 834
 835         spin_unlock_irq(&conf->resync_lock);
 836 }
 837
 838 static void lower_barrier(struct r1conf *conf)
 839 {
 840         unsigned long flags;
 841         BUG_ON(conf->barrier <= 0);
 842         spin_lock_irqsave(&conf->resync_lock, flags);
 843         conf->barrier--;
 844         spin_unlock_irqrestore(&conf->resync_lock, flags);
 845         wake_up(&conf->wait_barrier);
 846 }
 847
 848 static void wait_barrier(struct r1conf *conf)
 849 {
 850         spin_lock_irq(&conf->resync_lock);
 851         if (conf->barrier) {
 852                 conf->nr_waiting++;
 853                 /* Wait for the barrier to drop.
 854                  * However if there are already pending
 855                  * requests (preventing the barrier from
 856                  * rising completely), and the
 857                  * pre-process bio queue isn't empty,
 858                  * then don't wait, as we need to empty
 859                  * that queue to get the nr_pending
 860                  * count down.
 861                  */
 862                 wait_event_lock_irq(conf->wait_barrier,
 863                                     !conf->barrier ||
 864                                     (conf->nr_pending &&
 865                                      current->bio_list &&
 866                                      !bio_list_empty(current->bio_list)),
 867                                     conf->resync_lock);
 868                 conf->nr_waiting--;
 869         }
 870         conf->nr_pending++;
 871         spin_unlock_irq(&conf->resync_lock);
 872 }
 873
 874 static void allow_barrier(struct r1conf *conf)
 875 {
 876         unsigned long flags;
 877         spin_lock_irqsave(&conf->resync_lock, flags);
 878         conf->nr_pending--;
 879         spin_unlock_irqrestore(&conf->resync_lock, flags);
 880         wake_up(&conf->wait_barrier);
 881 }
 882
 883 static void freeze_array(struct r1conf *conf, int extra)
 884 {
 885         /* stop syncio and normal IO and wait for everything to
 886          * go quite.
 887          * We increment barrier and nr_waiting, and then
 888          * wait until nr_pending match nr_queued+extra
 889          * This is called in the context of one normal IO request
 890          * that has failed. Thus any sync request that might be pending
 891          * will be blocked by nr_pending, and we need to wait for
 892          * pending IO requests to complete or be queued for re-try.
 893          * Thus the number queued (nr_queued) plus this request (extra)
 894          * must match the number of pending IOs (nr_pending) before
 895          * we continue.
 896          */
 897         spin_lock_irq(&conf->resync_lock);
 898         conf->barrier++;
 899         conf->nr_waiting++;
 900         wait_event_lock_irq_cmd(conf->wait_barrier,
 901                                 conf->nr_pending == conf->nr_queued+extra,
 902                                 conf->resync_lock,
 903                                 flush_pending_writes(conf));
 904         spin_unlock_irq(&conf->resync_lock);
 905 }
 906 static void unfreeze_array(struct r1conf *conf)
 907 {
 908         /* reverse the effect of the freeze */
 909         spin_lock_irq(&conf->resync_lock);
 910         conf->barrier--;
 911         conf->nr_waiting--;
 912         wake_up(&conf->wait_barrier);
 913         spin_unlock_irq(&conf->resync_lock);
 914 }
 915
 916
 917 /* duplicate the data pages for behind I/O
 918  */
 919 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
 920 {
 921         int i;
 922         struct bio_vec *bvec;
 923         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
 924                                         GFP_NOIO);
 925         if (unlikely(!bvecs))
 926                 return;
 927
 928         bio_for_each_segment_all(bvec, bio, i) {
 929                 bvecs[i] = *bvec;
 930                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
 931                 if (unlikely(!bvecs[i].bv_page))
 932                         goto do_sync_io;
 933                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
 934                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 935                 kunmap(bvecs[i].bv_page);
 936                 kunmap(bvec->bv_page);
 937         }
 938         r1_bio->behind_bvecs = bvecs;
 939         r1_bio->behind_page_count = bio->bi_vcnt;
 940         set_bit(R1BIO_BehindIO, &r1_bio->state);
 941         return;
 942
 943 do_sync_io:
 944         for (i = 0; i < bio->bi_vcnt; i++)
 945                 if (bvecs[i].bv_page)
 946                         put_page(bvecs[i].bv_page);
 947         kfree(bvecs);
 948         pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 949 }
 950
 951 struct raid1_plug_cb {
 952         struct blk_plug_cb      cb;
 953         struct bio_list         pending;
 954         int                     pending_cnt;
 955 };
 956
 957 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
 958 {
 959         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
 960                                                   cb);
 961         struct mddev *mddev = plug->cb.data;
 962         struct r1conf *conf = mddev->private;
 963         struct bio *bio;
 964
 965         if (from_schedule || current->bio_list) {
 966                 spin_lock_irq(&conf->device_lock);
 967                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
 968                 conf->pending_count += plug->pending_cnt;
 969                 spin_unlock_irq(&conf->device_lock);
 970                 wake_up(&conf->wait_barrier);
 971                 md_wakeup_thread(mddev->thread);
 972                 kfree(plug);
 973                 return;
 974         }
 975
 976         /* we aren't scheduling, so we can do the write-out directly. */
 977         bio = bio_list_get(&plug->pending);
 978         bitmap_unplug(mddev->bitmap);
 979         wake_up(&conf->wait_barrier);
 980
 981         while (bio) { /* submit pending writes */
 982                 struct bio *next = bio->bi_next;
 983                 bio->bi_next = NULL;
 984                 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
 985                     !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 986                         /* Just ignore it */
 987                         bio_endio(bio, 0);
 988                 else
 989                         generic_make_request(bio);
 990                 bio = next;
 991         }
 992         kfree(plug);
 993 }
 994
 995 static void make_request(struct mddev *mddev, struct bio * bio)
 996 {
 997         struct r1conf *conf = mddev->private;
 998         struct raid1_info *mirror;
 999         struct r1bio *r1_bio;
1000         struct bio *read_bio;
1001         int i, disks;
1002         struct bitmap *bitmap;
1003         unsigned long flags;
1004         const int rw = bio_data_dir(bio);
1005         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1006         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1007         const unsigned long do_discard = (bio->bi_rw
1008                                           & (REQ_DISCARD | REQ_SECURE));
1009         const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1010         struct md_rdev *blocked_rdev;
1011         struct blk_plug_cb *cb;
1012         struct raid1_plug_cb *plug = NULL;
1013         int first_clone;
1014         int sectors_handled;
1015         int max_sectors;
1016
1017         /*
1018          * Register the new request and wait if the reconstruction
1019          * thread has put up a bar for new requests.
1020          * Continue immediately if no resync is active currently.
1021          */
1022
1023         md_write_start(mddev, bio); /* wait on superblock update early */
1024
1025         if (bio_data_dir(bio) == WRITE &&
1026             bio_end_sector(bio) > mddev->suspend_lo &&
1027             bio->bi_sector < mddev->suspend_hi) {
1028                 /* As the suspend_* range is controlled by
1029                  * userspace, we want an interruptible
1030                  * wait.
1031                  */
1032                 DEFINE_WAIT(w);
1033                 for (;;) {
1034                         flush_signals(current);
1035                         prepare_to_wait(&conf->wait_barrier,
1036                                         &w, TASK_INTERRUPTIBLE);
1037                         if (bio_end_sector(bio) <= mddev->suspend_lo ||
1038                             bio->bi_sector >= mddev->suspend_hi)
1039                                 break;
1040                         schedule();
1041                 }
1042                 finish_wait(&conf->wait_barrier, &w);
1043         }
1044
1045         wait_barrier(conf);
1046
1047         bitmap = mddev->bitmap;
1048
1049         /*
1050          * make_request() can abort the operation when READA is being
1051          * used and no empty request is available.
1052          *
1053          */
1054         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1055
1056         r1_bio->master_bio = bio;
1057         r1_bio->sectors = bio_sectors(bio);
1058         r1_bio->state = 0;
1059         r1_bio->mddev = mddev;
1060         r1_bio->sector = bio->bi_sector;
1061
1062         /* We might need to issue multiple reads to different
1063          * devices if there are bad blocks around, so we keep
1064          * track of the number of reads in bio->bi_phys_segments.
1065          * If this is 0, there is only one r1_bio and no locking
1066          * will be needed when requests complete.  If it is
1067          * non-zero, then it is the number of not-completed requests.
1068          */
1069         bio->bi_phys_segments = 0;
1070         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1071
1072         if (rw == READ) {
1073                 /*
1074                  * read balancing logic:
1075                  */
1076                 int rdisk;
1077
1078 read_again:
1079                 rdisk = read_balance(conf, r1_bio, &max_sectors);
1080
1081                 if (rdisk < 0) {
1082                         /* couldn't find anywhere to read from */
1083                         raid_end_bio_io(r1_bio);
1084                         return;
1085                 }
1086                 mirror = conf->mirrors + rdisk;
1087
1088                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1089                     bitmap) {
1090                         /* Reading from a write-mostly device must
1091                          * take care not to over-take any writes
1092                          * that are 'behind'
1093                          */
1094                         wait_event(bitmap->behind_wait,
1095                                    atomic_read(&bitmap->behind_writes) == 0);
1096                 }
1097                 r1_bio->read_disk = rdisk;
1098
1099                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1100                 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
1101                             max_sectors);
1102
1103                 r1_bio->bios[rdisk] = read_bio;
1104
1105                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
1106                 read_bio->bi_bdev = mirror->rdev->bdev;
1107                 read_bio->bi_end_io = raid1_end_read_request;
1108                 read_bio->bi_rw = READ | do_sync;
1109                 read_bio->bi_private = r1_bio;
1110
1111                 if (max_sectors < r1_bio->sectors) {
1112                         /* could not read all from this device, so we will
1113                          * need another r1_bio.
1114                          */
1115
1116                         sectors_handled = (r1_bio->sector + max_sectors
1117                                            - bio->bi_sector);
1118                         r1_bio->sectors = max_sectors;
1119                         spin_lock_irq(&conf->device_lock);
1120                         if (bio->bi_phys_segments == 0)
1121                                 bio->bi_phys_segments = 2;
1122                         else
1123                                 bio->bi_phys_segments++;
1124                         spin_unlock_irq(&conf->device_lock);
1125                         /* Cannot call generic_make_request directly
1126                          * as that will be queued in __make_request
1127                          * and subsequent mempool_alloc might block waiting
1128                          * for it.  So hand bio over to raid1d.
1129                          */
1130                         reschedule_retry(r1_bio);
1131
1132                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1133
1134                         r1_bio->master_bio = bio;
1135                         r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1136                         r1_bio->state = 0;
1137                         r1_bio->mddev = mddev;
1138                         r1_bio->sector = bio->bi_sector + sectors_handled;
1139                         goto read_again;
1140                 } else
1141                         generic_make_request(read_bio);
1142                 return;
1143         }
1144
1145         /*
1146          * WRITE:
1147          */
1148         if (conf->pending_count >= max_queued_requests) {
1149                 md_wakeup_thread(mddev->thread);
1150                 wait_event(conf->wait_barrier,
1151                            conf->pending_count < max_queued_requests);
1152         }
1153         /* first select target devices under rcu_lock and
1154          * inc refcount on their rdev.  Record them by setting
1155          * bios[x] to bio
1156          * If there are known/acknowledged bad blocks on any device on
1157          * which we have seen a write error, we want to avoid writing those
1158          * blocks.
1159          * This potentially requires several writes to write around
1160          * the bad blocks.  Each set of writes gets it's own r1bio
1161          * with a set of bios attached.
1162          */
1163
1164         disks = conf->raid_disks * 2;
1165  retry_write:
1166         blocked_rdev = NULL;
1167         rcu_read_lock();
1168         max_sectors = r1_bio->sectors;
1169         for (i = 0;  i < disks; i++) {
1170                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1171                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1172                         atomic_inc(&rdev->nr_pending);
1173                         blocked_rdev = rdev;
1174                         break;
1175                 }
1176                 r1_bio->bios[i] = NULL;
1177                 if (!rdev || test_bit(Faulty, &rdev->flags)
1178                     || test_bit(Unmerged, &rdev->flags)) {
1179                         if (i < conf->raid_disks)
1180                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1181                         continue;
1182                 }
1183
1184                 atomic_inc(&rdev->nr_pending);
1185                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1186                         sector_t first_bad;
1187                         int bad_sectors;
1188                         int is_bad;
1189
1190                         is_bad = is_badblock(rdev, r1_bio->sector,
1191                                              max_sectors,
1192                                              &first_bad, &bad_sectors);
1193                         if (is_bad < 0) {
1194                                 /* mustn't write here until the bad block is
1195                                  * acknowledged*/
1196                                 set_bit(BlockedBadBlocks, &rdev->flags);
1197                                 blocked_rdev = rdev;
1198                                 break;
1199                         }
1200                         if (is_bad && first_bad <= r1_bio->sector) {
1201                                 /* Cannot write here at all */
1202                                 bad_sectors -= (r1_bio->sector - first_bad);
1203                                 if (bad_sectors < max_sectors)
1204                                         /* mustn't write more than bad_sectors
1205                                          * to other devices yet
1206                                          */
1207                                         max_sectors = bad_sectors;
1208                                 rdev_dec_pending(rdev, mddev);
1209                                 /* We don't set R1BIO_Degraded as that
1210                                  * only applies if the disk is
1211                                  * missing, so it might be re-added,
1212                                  * and we want to know to recover this
1213                                  * chunk.
1214                                  * In this case the device is here,
1215                                  * and the fact that this chunk is not
1216                                  * in-sync is recorded in the bad
1217                                  * block log
1218                                  */
1219                                 continue;
1220                         }
1221                         if (is_bad) {
1222                                 int good_sectors = first_bad - r1_bio->sector;
1223                                 if (good_sectors < max_sectors)
1224                                         max_sectors = good_sectors;
1225                         }
1226                 }
1227                 r1_bio->bios[i] = bio;
1228         }
1229         rcu_read_unlock();
1230
1231         if (unlikely(blocked_rdev)) {
1232                 /* Wait for this device to become unblocked */
1233                 int j;
1234
1235                 for (j = 0; j < i; j++)
1236                         if (r1_bio->bios[j])
1237                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1238                 r1_bio->state = 0;
1239                 allow_barrier(conf);
1240                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1241                 wait_barrier(conf);
1242                 goto retry_write;
1243         }
1244
1245         if (max_sectors < r1_bio->sectors) {
1246                 /* We are splitting this write into multiple parts, so
1247                  * we need to prepare for allocating another r1_bio.
1248                  */
1249                 r1_bio->sectors = max_sectors;
1250                 spin_lock_irq(&conf->device_lock);
1251                 if (bio->bi_phys_segments == 0)
1252                         bio->bi_phys_segments = 2;
1253                 else
1254                         bio->bi_phys_segments++;
1255                 spin_unlock_irq(&conf->device_lock);
1256         }
1257         sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1258
1259         atomic_set(&r1_bio->remaining, 1);
1260         atomic_set(&r1_bio->behind_remaining, 0);
1261
1262         first_clone = 1;
1263         for (i = 0; i < disks; i++) {
1264                 struct bio *mbio;
1265                 if (!r1_bio->bios[i])
1266                         continue;
1267
1268                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1269                 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1270
1271                 if (first_clone) {
1272                         /* do behind I/O ?
1273                          * Not if there are too many, or cannot
1274                          * allocate memory, or a reader on WriteMostly
1275                          * is waiting for behind writes to flush */
1276                         if (bitmap &&
1277                             (atomic_read(&bitmap->behind_writes)
1278                              < mddev->bitmap_info.max_write_behind) &&
1279                             !waitqueue_active(&bitmap->behind_wait))
1280                                 alloc_behind_pages(mbio, r1_bio);
1281
1282                         bitmap_startwrite(bitmap, r1_bio->sector,
1283                                           r1_bio->sectors,
1284                                           test_bit(R1BIO_BehindIO,
1285                                                    &r1_bio->state));
1286                         first_clone = 0;
1287                 }
1288                 if (r1_bio->behind_bvecs) {
1289                         struct bio_vec *bvec;
1290                         int j;
1291
1292                         /*
1293                          * We trimmed the bio, so _all is legit
1294                          */
1295                         bio_for_each_segment_all(bvec, mbio, j)
1296                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1297                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1298                                 atomic_inc(&r1_bio->behind_remaining);
1299                 }
1300
1301                 r1_bio->bios[i] = mbio;
1302
1303                 mbio->bi_sector = (r1_bio->sector +
1304                                    conf->mirrors[i].rdev->data_offset);
1305                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1306                 mbio->bi_end_io = raid1_end_write_request;
1307                 mbio->bi_rw =
1308                         WRITE | do_flush_fua | do_sync | do_discard | do_same;
1309                 mbio->bi_private = r1_bio;
1310
1311                 atomic_inc(&r1_bio->remaining);
1312
1313                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1314                 if (cb)
1315                         plug = container_of(cb, struct raid1_plug_cb, cb);
1316                 else
1317                         plug = NULL;
1318                 spin_lock_irqsave(&conf->device_lock, flags);
1319                 if (plug) {
1320                         bio_list_add(&plug->pending, mbio);
1321                         plug->pending_cnt++;
1322                 } else {
1323                         bio_list_add(&conf->pending_bio_list, mbio);
1324                         conf->pending_count++;
1325                 }
1326                 spin_unlock_irqrestore(&conf->device_lock, flags);
1327                 if (!plug)
1328                         md_wakeup_thread(mddev->thread);
1329         }
1330         /* Mustn't call r1_bio_write_done before this next test,
1331          * as it could result in the bio being freed.
1332          */
1333         if (sectors_handled < bio_sectors(bio)) {
1334                 r1_bio_write_done(r1_bio);
1335                 /* We need another r1_bio.  It has already been counted
1336                  * in bio->bi_phys_segments
1337                  */
1338                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1339                 r1_bio->master_bio = bio;
1340                 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1341                 r1_bio->state = 0;
1342                 r1_bio->mddev = mddev;
1343                 r1_bio->sector = bio->bi_sector + sectors_handled;
1344                 goto retry_write;
1345         }
1346
1347         r1_bio_write_done(r1_bio);
1348
1349         /* In case raid1d snuck in to freeze_array */
1350         wake_up(&conf->wait_barrier);
1351 }
1352
1353 static void status(struct seq_file *seq, struct mddev *mddev)
1354 {
1355         struct r1conf *conf = mddev->private;
1356         int i;
1357
1358         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1359                    conf->raid_disks - mddev->degraded);
1360         rcu_read_lock();
1361         for (i = 0; i < conf->raid_disks; i++) {
1362                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1363                 seq_printf(seq, "%s",
1364                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1365         }
1366         rcu_read_unlock();
1367         seq_printf(seq, "]");
1368 }
1369
1370
1371 static void error(struct mddev *mddev, struct md_rdev *rdev)
1372 {
1373         char b[BDEVNAME_SIZE];
1374         struct r1conf *conf = mddev->private;
1375
1376         /*
1377          * If it is not operational, then we have already marked it as dead
1378          * else if it is the last working disks, ignore the error, let the
1379          * next level up know.
1380          * else mark the drive as failed
1381          */
1382         if (test_bit(In_sync, &rdev->flags)
1383             && (conf->raid_disks - mddev->degraded) == 1) {
1384                 /*
1385                  * Don't fail the drive, act as though we were just a
1386                  * normal single drive.
1387                  * However don't try a recovery from this drive as
1388                  * it is very likely to fail.
1389                  */
1390                 conf->recovery_disabled = mddev->recovery_disabled;
1391                 return;
1392         }
1393         set_bit(Blocked, &rdev->flags);
1394         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1395                 unsigned long flags;
1396                 spin_lock_irqsave(&conf->device_lock, flags);
1397                 mddev->degraded++;
1398                 set_bit(Faulty, &rdev->flags);
1399                 spin_unlock_irqrestore(&conf->device_lock, flags);
1400                 /*
1401                  * if recovery is running, make sure it aborts.
1402                  */
1403                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1404         } else
1405                 set_bit(Faulty, &rdev->flags);
1406         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1407         printk(KERN_ALERT
1408                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1409                "md/raid1:%s: Operation continuing on %d devices.\n",
1410                mdname(mddev), bdevname(rdev->bdev, b),
1411                mdname(mddev), conf->raid_disks - mddev->degraded);
1412 }
1413
1414 static void print_conf(struct r1conf *conf)
1415 {
1416         int i;
1417
1418         printk(KERN_DEBUG "RAID1 conf printout:\n");
1419         if (!conf) {
1420                 printk(KERN_DEBUG "(!conf)\n");
1421                 return;
1422         }
1423         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1424                 conf->raid_disks);
1425
1426         rcu_read_lock();
1427         for (i = 0; i < conf->raid_disks; i++) {
1428                 char b[BDEVNAME_SIZE];
1429                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1430                 if (rdev)
1431                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1432                                i, !test_bit(In_sync, &rdev->flags),
1433                                !test_bit(Faulty, &rdev->flags),
1434                                bdevname(rdev->bdev,b));
1435         }
1436         rcu_read_unlock();
1437 }
1438
1439 static void close_sync(struct r1conf *conf)
1440 {
1441         wait_barrier(conf);
1442         allow_barrier(conf);
1443
1444         mempool_destroy(conf->r1buf_pool);
1445         conf->r1buf_pool = NULL;
1446 }
1447
1448 static int raid1_spare_active(struct mddev *mddev)
1449 {
1450         int i;
1451         struct r1conf *conf = mddev->private;
1452         int count = 0;
1453         unsigned long flags;
1454
1455         /*
1456          * Find all failed disks within the RAID1 configuration
1457          * and mark them readable.
1458          * Called under mddev lock, so rcu protection not needed.
1459          */
1460         for (i = 0; i < conf->raid_disks; i++) {
1461                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1462                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1463                 if (repl
1464                     && repl->recovery_offset == MaxSector
1465                     && !test_bit(Faulty, &repl->flags)
1466                     && !test_and_set_bit(In_sync, &repl->flags)) {
1467                         /* replacement has just become active */
1468                         if (!rdev ||
1469                             !test_and_clear_bit(In_sync, &rdev->flags))
1470                                 count++;
1471                         if (rdev) {
1472                                 /* Replaced device not technically
1473                                  * faulty, but we need to be sure
1474                                  * it gets removed and never re-added
1475                                  */
1476                                 set_bit(Faulty, &rdev->flags);
1477                                 sysfs_notify_dirent_safe(
1478                                         rdev->sysfs_state);
1479                         }
1480                 }
1481                 if (rdev
1482                     && !test_bit(Faulty, &rdev->flags)
1483                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1484                         count++;
1485                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1486                 }
1487         }
1488         spin_lock_irqsave(&conf->device_lock, flags);
1489         mddev->degraded -= count;
1490         spin_unlock_irqrestore(&conf->device_lock, flags);
1491
1492         print_conf(conf);
1493         return count;
1494 }
1495
1496
1497 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1498 {
1499         struct r1conf *conf = mddev->private;
1500         int err = -EEXIST;
1501         int mirror = 0;
1502         struct raid1_info *p;
1503         int first = 0;
1504         int last = conf->raid_disks - 1;
1505         struct request_queue *q = bdev_get_queue(rdev->bdev);
1506
1507         if (mddev->recovery_disabled == conf->recovery_disabled)
1508                 return -EBUSY;
1509
1510         if (rdev->raid_disk >= 0)
1511                 first = last = rdev->raid_disk;
1512
1513         if (q->merge_bvec_fn) {
1514                 set_bit(Unmerged, &rdev->flags);
1515                 mddev->merge_check_needed = 1;
1516         }
1517
1518         for (mirror = first; mirror <= last; mirror++) {
1519                 p = conf->mirrors+mirror;
1520                 if (!p->rdev) {
1521
1522                         if (mddev->gendisk)
1523                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1524                                                   rdev->data_offset << 9);
1525
1526                         p->head_position = 0;
1527                         rdev->raid_disk = mirror;
1528                         err = 0;
1529                         /* As all devices are equivalent, we don't need a full recovery
1530                          * if this was recently any drive of the array
1531                          */
1532                         if (rdev->saved_raid_disk < 0)
1533                                 conf->fullsync = 1;
1534                         rcu_assign_pointer(p->rdev, rdev);
1535                         break;
1536                 }
1537                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1538                     p[conf->raid_disks].rdev == NULL) {
1539                         /* Add this device as a replacement */
1540                         clear_bit(In_sync, &rdev->flags);
1541                         set_bit(Replacement, &rdev->flags);
1542                         rdev->raid_disk = mirror;
1543                         err = 0;
1544                         conf->fullsync = 1;
1545                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1546                         break;
1547                 }
1548         }
1549         if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1550                 /* Some requests might not have seen this new
1551                  * merge_bvec_fn.  We must wait for them to complete
1552                  * before merging the device fully.
1553                  * First we make sure any code which has tested
1554                  * our function has submitted the request, then
1555                  * we wait for all outstanding requests to complete.
1556                  */
1557                 synchronize_sched();
1558                 freeze_array(conf, 0);
1559                 unfreeze_array(conf);
1560                 clear_bit(Unmerged, &rdev->flags);
1561         }
1562         md_integrity_add_rdev(rdev, mddev);
1563         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1564                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1565         print_conf(conf);
1566         return err;
1567 }
1568
1569 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1570 {
1571         struct r1conf *conf = mddev->private;
1572         int err = 0;
1573         int number = rdev->raid_disk;
1574         struct raid1_info *p = conf->mirrors + number;
1575
1576         if (rdev != p->rdev)
1577                 p = conf->mirrors + conf->raid_disks + number;
1578
1579         print_conf(conf);
1580         if (rdev == p->rdev) {
1581                 if (test_bit(In_sync, &rdev->flags) ||
1582                     atomic_read(&rdev->nr_pending)) {
1583                         err = -EBUSY;
1584                         goto abort;
1585                 }
1586                 /* Only remove non-faulty devices if recovery
1587                  * is not possible.
1588                  */
1589                 if (!test_bit(Faulty, &rdev->flags) &&
1590                     mddev->recovery_disabled != conf->recovery_disabled &&
1591                     mddev->degraded < conf->raid_disks) {
1592                         err = -EBUSY;
1593                         goto abort;
1594                 }
1595                 p->rdev = NULL;
1596                 synchronize_rcu();
1597                 if (atomic_read(&rdev->nr_pending)) {
1598                         /* lost the race, try later */
1599                         err = -EBUSY;
1600                         p->rdev = rdev;
1601                         goto abort;
1602                 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1603                         /* We just removed a device that is being replaced.
1604                          * Move down the replacement.  We drain all IO before
1605                          * doing this to avoid confusion.
1606                          */
1607                         struct md_rdev *repl =
1608                                 conf->mirrors[conf->raid_disks + number].rdev;
1609                         freeze_array(conf, 0);
1610                         clear_bit(Replacement, &repl->flags);
1611                         p->rdev = repl;
1612                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1613                         unfreeze_array(conf);
1614                         clear_bit(WantReplacement, &rdev->flags);
1615                 } else
1616                         clear_bit(WantReplacement, &rdev->flags);
1617                 err = md_integrity_register(mddev);
1618         }
1619 abort:
1620
1621         print_conf(conf);
1622         return err;
1623 }
1624
1625
1626 static void end_sync_read(struct bio *bio, int error)
1627 {
1628         struct r1bio *r1_bio = bio->bi_private;
1629
1630         update_head_pos(r1_bio->read_disk, r1_bio);
1631
1632         /*
1633          * we have read a block, now it needs to be re-written,
1634          * or re-read if the read failed.
1635          * We don't do much here, just schedule handling by raid1d
1636          */
1637         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1638                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1639
1640         if (atomic_dec_and_test(&r1_bio->remaining))
1641                 reschedule_retry(r1_bio);
1642 }
1643
1644 static void end_sync_write(struct bio *bio, int error)
1645 {
1646         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1647         struct r1bio *r1_bio = bio->bi_private;
1648         struct mddev *mddev = r1_bio->mddev;
1649         struct r1conf *conf = mddev->private;
1650         int mirror=0;
1651         sector_t first_bad;
1652         int bad_sectors;
1653
1654         mirror = find_bio_disk(r1_bio, bio);
1655
1656         if (!uptodate) {
1657                 sector_t sync_blocks = 0;
1658                 sector_t s = r1_bio->sector;
1659                 long sectors_to_go = r1_bio->sectors;
1660                 /* make sure these bits doesn't get cleared. */
1661                 do {
1662                         bitmap_end_sync(mddev->bitmap, s,
1663                                         &sync_blocks, 1);
1664                         s += sync_blocks;
1665                         sectors_to_go -= sync_blocks;
1666                 } while (sectors_to_go > 0);
1667                 set_bit(WriteErrorSeen,
1668                         &conf->mirrors[mirror].rdev->flags);
1669                 if (!test_and_set_bit(WantReplacement,
1670                                       &conf->mirrors[mirror].rdev->flags))
1671                         set_bit(MD_RECOVERY_NEEDED, &
1672                                 mddev->recovery);
1673                 set_bit(R1BIO_WriteError, &r1_bio->state);
1674         } else if (is_badblock(conf->mirrors[mirror].rdev,
1675                                r1_bio->sector,
1676                                r1_bio->sectors,
1677                                &first_bad, &bad_sectors) &&
1678                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1679                                 r1_bio->sector,
1680                                 r1_bio->sectors,
1681                                 &first_bad, &bad_sectors)
1682                 )
1683                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1684
1685         if (atomic_dec_and_test(&r1_bio->remaining)) {
1686                 int s = r1_bio->sectors;
1687                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1688                     test_bit(R1BIO_WriteError, &r1_bio->state))
1689                         reschedule_retry(r1_bio);
1690                 else {
1691                         put_buf(r1_bio);
1692                         md_done_sync(mddev, s, uptodate);
1693                 }
1694         }
1695 }
1696
1697 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1698                             int sectors, struct page *page, int rw)
1699 {
1700         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1701                 /* success */
1702                 return 1;
1703         if (rw == WRITE) {
1704                 set_bit(WriteErrorSeen, &rdev->flags);
1705                 if (!test_and_set_bit(WantReplacement,
1706                                       &rdev->flags))
1707                         set_bit(MD_RECOVERY_NEEDED, &
1708                                 rdev->mddev->recovery);
1709         }
1710         /* need to record an error - either for the block or the device */
1711         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1712                 md_error(rdev->mddev, rdev);
1713         return 0;
1714 }
1715
1716 static int fix_sync_read_error(struct r1bio *r1_bio)
1717 {
1718         /* Try some synchronous reads of other devices to get
1719          * good data, much like with normal read errors.  Only
1720          * read into the pages we already have so we don't
1721          * need to re-issue the read request.
1722          * We don't need to freeze the array, because being in an
1723          * active sync request, there is no normal IO, and
1724          * no overlapping syncs.
1725          * We don't need to check is_badblock() again as we
1726          * made sure that anything with a bad block in range
1727          * will have bi_end_io clear.
1728          */
1729         struct mddev *mddev = r1_bio->mddev;
1730         struct r1conf *conf = mddev->private;
1731         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1732         sector_t sect = r1_bio->sector;
1733         int sectors = r1_bio->sectors;
1734         int idx = 0;
1735
1736         while(sectors) {
1737                 int s = sectors;
1738                 int d = r1_bio->read_disk;
1739                 int success = 0;
1740                 struct md_rdev *rdev;
1741                 int start;
1742
1743                 if (s > (PAGE_SIZE>>9))
1744                         s = PAGE_SIZE >> 9;
1745                 do {
1746                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1747                                 /* No rcu protection needed here devices
1748                                  * can only be removed when no resync is
1749                                  * active, and resync is currently active
1750                                  */
1751                                 rdev = conf->mirrors[d].rdev;
1752                                 if (sync_page_io(rdev, sect, s<<9,
1753                                                  bio->bi_io_vec[idx].bv_page,
1754                                                  READ, false)) {
1755                                         success = 1;
1756                                         break;
1757                                 }
1758                         }
1759                         d++;
1760                         if (d == conf->raid_disks * 2)
1761                                 d = 0;
1762                 } while (!success && d != r1_bio->read_disk);
1763
1764                 if (!success) {
1765                         char b[BDEVNAME_SIZE];
1766                         int abort = 0;
1767                         /* Cannot read from anywhere, this block is lost.
1768                          * Record a bad block on each device.  If that doesn't
1769                          * work just disable and interrupt the recovery.
1770                          * Don't fail devices as that won't really help.
1771                          */
1772                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1773                                " for block %llu\n",
1774                                mdname(mddev),
1775                                bdevname(bio->bi_bdev, b),
1776                                (unsigned long long)r1_bio->sector);
1777                         for (d = 0; d < conf->raid_disks * 2; d++) {
1778                                 rdev = conf->mirrors[d].rdev;
1779                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1780                                         continue;
1781                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1782                                         abort = 1;
1783                         }
1784                         if (abort) {
1785                                 conf->recovery_disabled =
1786                                         mddev->recovery_disabled;
1787                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1788                                 md_done_sync(mddev, r1_bio->sectors, 0);
1789                                 put_buf(r1_bio);
1790                                 return 0;
1791                         }
1792                         /* Try next page */
1793                         sectors -= s;
1794                         sect += s;
1795                         idx++;
1796                         continue;
1797                 }
1798
1799                 start = d;
1800                 /* write it back and re-read */
1801                 while (d != r1_bio->read_disk) {
1802                         if (d == 0)
1803                                 d = conf->raid_disks * 2;
1804                         d--;
1805                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1806                                 continue;
1807                         rdev = conf->mirrors[d].rdev;
1808                         if (r1_sync_page_io(rdev, sect, s,
1809                                             bio->bi_io_vec[idx].bv_page,
1810                                             WRITE) == 0) {
1811                                 r1_bio->bios[d]->bi_end_io = NULL;
1812                                 rdev_dec_pending(rdev, mddev);
1813                         }
1814                 }
1815                 d = start;
1816                 while (d != r1_bio->read_disk) {
1817                         if (d == 0)
1818                                 d = conf->raid_disks * 2;
1819                         d--;
1820                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1821                                 continue;
1822                         rdev = conf->mirrors[d].rdev;
1823                         if (r1_sync_page_io(rdev, sect, s,
1824                                             bio->bi_io_vec[idx].bv_page,
1825                                             READ) != 0)
1826                                 atomic_add(s, &rdev->corrected_errors);
1827                 }
1828                 sectors -= s;
1829                 sect += s;
1830                 idx ++;
1831         }
1832         set_bit(R1BIO_Uptodate, &r1_bio->state);
1833         set_bit(BIO_UPTODATE, &bio->bi_flags);
1834         return 1;
1835 }
1836
1837 static int process_checks(struct r1bio *r1_bio)
1838 {
1839         /* We have read all readable devices.  If we haven't
1840          * got the block, then there is no hope left.
1841          * If we have, then we want to do a comparison
1842          * and skip the write if everything is the same.
1843          * If any blocks failed to read, then we need to
1844          * attempt an over-write
1845          */
1846         struct mddev *mddev = r1_bio->mddev;
1847         struct r1conf *conf = mddev->private;
1848         int primary;
1849         int i;
1850         int vcnt;
1851
1852         for (primary = 0; primary < conf->raid_disks * 2; primary++)
1853                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1854                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1855                         r1_bio->bios[primary]->bi_end_io = NULL;
1856                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1857                         break;
1858                 }
1859         r1_bio->read_disk = primary;
1860         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1861         for (i = 0; i < conf->raid_disks * 2; i++) {
1862                 int j;
1863                 struct bio *pbio = r1_bio->bios[primary];
1864                 struct bio *sbio = r1_bio->bios[i];
1865                 int size;
1866
1867                 if (sbio->bi_end_io != end_sync_read)
1868                         continue;
1869
1870                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1871                         for (j = vcnt; j-- ; ) {
1872                                 struct page *p, *s;
1873                                 p = pbio->bi_io_vec[j].bv_page;
1874                                 s = sbio->bi_io_vec[j].bv_page;
1875                                 if (memcmp(page_address(p),
1876                                            page_address(s),
1877                                            sbio->bi_io_vec[j].bv_len))
1878                                         break;
1879                         }
1880                 } else
1881                         j = 0;
1882                 if (j >= 0)
1883                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
1884                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1885                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1886                         /* No need to write to this device. */
1887                         sbio->bi_end_io = NULL;
1888                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1889                         continue;
1890                 }
1891                 /* fixup the bio for reuse */
1892                 bio_reset(sbio);
1893                 sbio->bi_vcnt = vcnt;
1894                 sbio->bi_size = r1_bio->sectors << 9;
1895                 sbio->bi_sector = r1_bio->sector +
1896                         conf->mirrors[i].rdev->data_offset;
1897                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1898                 sbio->bi_end_io = end_sync_read;
1899                 sbio->bi_private = r1_bio;
1900
1901                 size = sbio->bi_size;
1902                 for (j = 0; j < vcnt ; j++) {
1903                         struct bio_vec *bi;
1904                         bi = &sbio->bi_io_vec[j];
1905                         bi->bv_offset = 0;
1906                         if (size > PAGE_SIZE)
1907                                 bi->bv_len = PAGE_SIZE;
1908                         else
1909                                 bi->bv_len = size;
1910                         size -= PAGE_SIZE;
1911                 }
1912
1913                 bio_copy_data(sbio, pbio);
1914         }
1915         return 0;
1916 }
1917
1918 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1919 {
1920         struct r1conf *conf = mddev->private;
1921         int i;
1922         int disks = conf->raid_disks * 2;
1923         struct bio *bio, *wbio;
1924
1925         bio = r1_bio->bios[r1_bio->read_disk];
1926
1927         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1928                 /* ouch - failed to read all of that. */
1929                 if (!fix_sync_read_error(r1_bio))
1930                         return;
1931
1932         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1933                 if (process_checks(r1_bio) < 0)
1934                         return;
1935         /*
1936          * schedule writes
1937          */
1938         atomic_set(&r1_bio->remaining, 1);
1939         for (i = 0; i < disks ; i++) {
1940                 wbio = r1_bio->bios[i];
1941                 if (wbio->bi_end_io == NULL ||
1942                     (wbio->bi_end_io == end_sync_read &&
1943                      (i == r1_bio->read_disk ||
1944                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1945                         continue;
1946
1947                 wbio->bi_rw = WRITE;
1948                 wbio->bi_end_io = end_sync_write;
1949                 atomic_inc(&r1_bio->remaining);
1950                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
1951
1952                 generic_make_request(wbio);
1953         }
1954
1955         if (atomic_dec_and_test(&r1_bio->remaining)) {
1956                 /* if we're here, all write(s) have completed, so clean up */
1957                 int s = r1_bio->sectors;
1958                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1959                     test_bit(R1BIO_WriteError, &r1_bio->state))
1960                         reschedule_retry(r1_bio);
1961                 else {
1962                         put_buf(r1_bio);
1963                         md_done_sync(mddev, s, 1);
1964                 }
1965         }
1966 }
1967
1968 /*
1969  * This is a kernel thread which:
1970  *
1971  *      1.      Retries failed read operations on working mirrors.
1972  *      2.      Updates the raid superblock when problems encounter.
1973  *      3.      Performs writes following reads for array synchronising.
1974  */
1975
1976 static void fix_read_error(struct r1conf *conf, int read_disk,
1977                            sector_t sect, int sectors)
1978 {
1979         struct mddev *mddev = conf->mddev;
1980         while(sectors) {
1981                 int s = sectors;
1982                 int d = read_disk;
1983                 int success = 0;
1984                 int start;
1985                 struct md_rdev *rdev;
1986
1987                 if (s > (PAGE_SIZE>>9))
1988                         s = PAGE_SIZE >> 9;
1989
1990                 do {
1991                         /* Note: no rcu protection needed here
1992                          * as this is synchronous in the raid1d thread
1993                          * which is the thread that might remove
1994                          * a device.  If raid1d ever becomes multi-threaded....
1995                          */
1996                         sector_t first_bad;
1997                         int bad_sectors;
1998
1999                         rdev = conf->mirrors[d].rdev;
2000                         if (rdev &&
2001                             (test_bit(In_sync, &rdev->flags) ||
2002                              (!test_bit(Faulty, &rdev->flags) &&
2003                               rdev->recovery_offset >= sect + s)) &&
2004                             is_badblock(rdev, sect, s,
2005                                         &first_bad, &bad_sectors) == 0 &&
2006                             sync_page_io(rdev, sect, s<<9,
2007                                          conf->tmppage, READ, false))
2008                                 success = 1;
2009                         else {
2010                                 d++;
2011                                 if (d == conf->raid_disks * 2)
2012                                         d = 0;
2013                         }
2014                 } while (!success && d != read_disk);
2015
2016                 if (!success) {
2017                         /* Cannot read from anywhere - mark it bad */
2018                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2019                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2020                                 md_error(mddev, rdev);
2021                         break;
2022                 }
2023                 /* write it back and re-read */
2024                 start = d;
2025                 while (d != read_disk) {
2026                         if (d==0)
2027                                 d = conf->raid_disks * 2;
2028                         d--;
2029                         rdev = conf->mirrors[d].rdev;
2030                         if (rdev &&
2031                             test_bit(In_sync, &rdev->flags))
2032                                 r1_sync_page_io(rdev, sect, s,
2033                                                 conf->tmppage, WRITE);
2034                 }
2035                 d = start;
2036                 while (d != read_disk) {
2037                         char b[BDEVNAME_SIZE];
2038                         if (d==0)
2039                                 d = conf->raid_disks * 2;
2040                         d--;
2041                         rdev = conf->mirrors[d].rdev;
2042                         if (rdev &&
2043                             test_bit(In_sync, &rdev->flags)) {
2044                                 if (r1_sync_page_io(rdev, sect, s,
2045                                                     conf->tmppage, READ)) {
2046                                         atomic_add(s, &rdev->corrected_errors);
2047                                         printk(KERN_INFO
2048                                                "md/raid1:%s: read error corrected "
2049                                                "(%d sectors at %llu on %s)\n",
2050                                                mdname(mddev), s,
2051                                                (unsigned long long)(sect +
2052                                                    rdev->data_offset),
2053                                                bdevname(rdev->bdev, b));
2054                                 }
2055                         }
2056                 }
2057                 sectors -= s;
2058                 sect += s;
2059         }
2060 }
2061
2062 static int narrow_write_error(struct r1bio *r1_bio, int i)
2063 {
2064         struct mddev *mddev = r1_bio->mddev;
2065         struct r1conf *conf = mddev->private;
2066         struct md_rdev *rdev = conf->mirrors[i].rdev;
2067
2068         /* bio has the data to be written to device 'i' where
2069          * we just recently had a write error.
2070          * We repeatedly clone the bio and trim down to one block,
2071          * then try the write.  Where the write fails we record
2072          * a bad block.
2073          * It is conceivable that the bio doesn't exactly align with
2074          * blocks.  We must handle this somehow.
2075          *
2076          * We currently own a reference on the rdev.
2077          */
2078
2079         int block_sectors;
2080         sector_t sector;
2081         int sectors;
2082         int sect_to_write = r1_bio->sectors;
2083         int ok = 1;
2084
2085         if (rdev->badblocks.shift < 0)
2086                 return 0;
2087
2088         block_sectors = 1 << rdev->badblocks.shift;
2089         sector = r1_bio->sector;
2090         sectors = ((sector + block_sectors)
2091                    & ~(sector_t)(block_sectors - 1))
2092                 - sector;
2093
2094         while (sect_to_write) {
2095                 struct bio *wbio;
2096                 if (sectors > sect_to_write)
2097                         sectors = sect_to_write;
2098                 /* Write at 'sector' for 'sectors'*/
2099
2100                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2101                         unsigned vcnt = r1_bio->behind_page_count;
2102                         struct bio_vec *vec = r1_bio->behind_bvecs;
2103
2104                         while (!vec->bv_page) {
2105                                 vec++;
2106                                 vcnt--;
2107                         }
2108
2109                         wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2110                         memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2111
2112                         wbio->bi_vcnt = vcnt;
2113                 } else {
2114                         wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2115                 }
2116
2117                 wbio->bi_rw = WRITE;
2118                 wbio->bi_sector = r1_bio->sector;
2119                 wbio->bi_size = r1_bio->sectors << 9;
2120
2121                 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2122                 wbio->bi_sector += rdev->data_offset;
2123                 wbio->bi_bdev = rdev->bdev;
2124                 if (submit_bio_wait(WRITE, wbio) == 0)
2125                         /* failure! */
2126                         ok = rdev_set_badblocks(rdev, sector,
2127                                                 sectors, 0)
2128                                 && ok;
2129
2130                 bio_put(wbio);
2131                 sect_to_write -= sectors;
2132                 sector += sectors;
2133                 sectors = block_sectors;
2134         }
2135         return ok;
2136 }
2137
2138 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2139 {
2140         int m;
2141         int s = r1_bio->sectors;
2142         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2143                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2144                 struct bio *bio = r1_bio->bios[m];
2145                 if (bio->bi_end_io == NULL)
2146                         continue;
2147                 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2148                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2149                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2150                 }
2151                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2152                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2153                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2154                                 md_error(conf->mddev, rdev);
2155                 }
2156         }
2157         put_buf(r1_bio);
2158         md_done_sync(conf->mddev, s, 1);
2159 }
2160
2161 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2162 {
2163         int m;
2164         for (m = 0; m < conf->raid_disks * 2 ; m++)
2165                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2166                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2167                         rdev_clear_badblocks(rdev,
2168                                              r1_bio->sector,
2169                                              r1_bio->sectors, 0);
2170                         rdev_dec_pending(rdev, conf->mddev);
2171                 } else if (r1_bio->bios[m] != NULL) {
2172                         /* This drive got a write error.  We need to
2173                          * narrow down and record precise write
2174                          * errors.
2175                          */
2176                         if (!narrow_write_error(r1_bio, m)) {
2177                                 md_error(conf->mddev,
2178                                          conf->mirrors[m].rdev);
2179                                 /* an I/O failed, we can't clear the bitmap */
2180                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2181                         }
2182                         rdev_dec_pending(conf->mirrors[m].rdev,
2183                                          conf->mddev);
2184                 }
2185         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2186                 close_write(r1_bio);
2187         raid_end_bio_io(r1_bio);
2188 }
2189
2190 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2191 {
2192         int disk;
2193         int max_sectors;
2194         struct mddev *mddev = conf->mddev;
2195         struct bio *bio;
2196         char b[BDEVNAME_SIZE];
2197         struct md_rdev *rdev;
2198
2199         clear_bit(R1BIO_ReadError, &r1_bio->state);
2200         /* we got a read error. Maybe the drive is bad.  Maybe just
2201          * the block and we can fix it.
2202          * We freeze all other IO, and try reading the block from
2203          * other devices.  When we find one, we re-write
2204          * and check it that fixes the read error.
2205          * This is all done synchronously while the array is
2206          * frozen
2207          */
2208         if (mddev->ro == 0) {
2209                 freeze_array(conf, 1);
2210                 fix_read_error(conf, r1_bio->read_disk,
2211                                r1_bio->sector, r1_bio->sectors);
2212                 unfreeze_array(conf);
2213         } else
2214                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2215         rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2216
2217         bio = r1_bio->bios[r1_bio->read_disk];
2218         bdevname(bio->bi_bdev, b);
2219 read_more:
2220         disk = read_balance(conf, r1_bio, &max_sectors);
2221         if (disk == -1) {
2222                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2223                        " read error for block %llu\n",
2224                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
2225                 raid_end_bio_io(r1_bio);
2226         } else {
2227                 const unsigned long do_sync
2228                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
2229                 if (bio) {
2230                         r1_bio->bios[r1_bio->read_disk] =
2231                                 mddev->ro ? IO_BLOCKED : NULL;
2232                         bio_put(bio);
2233                 }
2234                 r1_bio->read_disk = disk;
2235                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2236                 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2237                 r1_bio->bios[r1_bio->read_disk] = bio;
2238                 rdev = conf->mirrors[disk].rdev;
2239                 printk_ratelimited(KERN_ERR
2240                                    "md/raid1:%s: redirecting sector %llu"
2241                                    " to other mirror: %s\n",
2242                                    mdname(mddev),
2243                                    (unsigned long long)r1_bio->sector,
2244                                    bdevname(rdev->bdev, b));
2245                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
2246                 bio->bi_bdev = rdev->bdev;
2247                 bio->bi_end_io = raid1_end_read_request;
2248                 bio->bi_rw = READ | do_sync;
2249                 bio->bi_private = r1_bio;
2250                 if (max_sectors < r1_bio->sectors) {
2251                         /* Drat - have to split this up more */
2252                         struct bio *mbio = r1_bio->master_bio;
2253                         int sectors_handled = (r1_bio->sector + max_sectors
2254                                                - mbio->bi_sector);
2255                         r1_bio->sectors = max_sectors;
2256                         spin_lock_irq(&conf->device_lock);
2257                         if (mbio->bi_phys_segments == 0)
2258                                 mbio->bi_phys_segments = 2;
2259                         else
2260                                 mbio->bi_phys_segments++;
2261                         spin_unlock_irq(&conf->device_lock);
2262                         generic_make_request(bio);
2263                         bio = NULL;
2264
2265                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2266
2267                         r1_bio->master_bio = mbio;
2268                         r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2269                         r1_bio->state = 0;
2270                         set_bit(R1BIO_ReadError, &r1_bio->state);
2271                         r1_bio->mddev = mddev;
2272                         r1_bio->sector = mbio->bi_sector + sectors_handled;
2273
2274                         goto read_more;
2275                 } else
2276                         generic_make_request(bio);
2277         }
2278 }
2279
2280 static void raid1d(struct md_thread *thread)
2281 {
2282         struct mddev *mddev = thread->mddev;
2283         struct r1bio *r1_bio;
2284         unsigned long flags;
2285         struct r1conf *conf = mddev->private;
2286         struct list_head *head = &conf->retry_list;
2287         struct blk_plug plug;
2288
2289         md_check_recovery(mddev);
2290
2291         blk_start_plug(&plug);
2292         for (;;) {
2293
2294                 flush_pending_writes(conf);
2295
2296                 spin_lock_irqsave(&conf->device_lock, flags);
2297                 if (list_empty(head)) {
2298                         spin_unlock_irqrestore(&conf->device_lock, flags);
2299                         break;
2300                 }
2301                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2302                 list_del(head->prev);
2303                 conf->nr_queued--;
2304                 spin_unlock_irqrestore(&conf->device_lock, flags);
2305
2306                 mddev = r1_bio->mddev;
2307                 conf = mddev->private;
2308                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2309                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2310                             test_bit(R1BIO_WriteError, &r1_bio->state))
2311                                 handle_sync_write_finished(conf, r1_bio);
2312                         else
2313                                 sync_request_write(mddev, r1_bio);
2314                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2315                            test_bit(R1BIO_WriteError, &r1_bio->state))
2316                         handle_write_finished(conf, r1_bio);
2317                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2318                         handle_read_error(conf, r1_bio);
2319                 else
2320                         /* just a partial read to be scheduled from separate
2321                          * context
2322                          */
2323                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2324
2325                 cond_resched();
2326                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2327                         md_check_recovery(mddev);
2328         }
2329         blk_finish_plug(&plug);
2330 }
2331
2332
2333 static int init_resync(struct r1conf *conf)
2334 {
2335         int buffs;
2336
2337         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2338         BUG_ON(conf->r1buf_pool);
2339         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2340                                           conf->poolinfo);
2341         if (!conf->r1buf_pool)
2342                 return -ENOMEM;
2343         conf->next_resync = 0;
2344         return 0;
2345 }
2346
2347 /*
2348  * perform a "sync" on one "block"
2349  *
2350  * We need to make sure that no normal I/O request - particularly write
2351  * requests - conflict with active sync requests.
2352  *
2353  * This is achieved by tracking pending requests and a 'barrier' concept
2354  * that can be installed to exclude normal IO requests.
2355  */
2356
2357 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2358 {
2359         struct r1conf *conf = mddev->private;
2360         struct r1bio *r1_bio;
2361         struct bio *bio;
2362         sector_t max_sector, nr_sectors;
2363         int disk = -1;
2364         int i;
2365         int wonly = -1;
2366         int write_targets = 0, read_targets = 0;
2367         sector_t sync_blocks;
2368         int still_degraded = 0;
2369         int good_sectors = RESYNC_SECTORS;
2370         int min_bad = 0; /* number of sectors that are bad in all devices */
2371
2372         if (!conf->r1buf_pool)
2373                 if (init_resync(conf))
2374                         return 0;
2375
2376         max_sector = mddev->dev_sectors;
2377         if (sector_nr >= max_sector) {
2378                 /* If we aborted, we need to abort the
2379                  * sync on the 'current' bitmap chunk (there will
2380                  * only be one in raid1 resync.
2381                  * We can find the current addess in mddev->curr_resync
2382                  */
2383                 if (mddev->curr_resync < max_sector) /* aborted */
2384                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2385                                                 &sync_blocks, 1);
2386                 else /* completed sync */
2387                         conf->fullsync = 0;
2388
2389                 bitmap_close_sync(mddev->bitmap);
2390                 close_sync(conf);
2391                 return 0;
2392         }
2393
2394         if (mddev->bitmap == NULL &&
2395             mddev->recovery_cp == MaxSector &&
2396             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2397             conf->fullsync == 0) {
2398                 *skipped = 1;
2399                 return max_sector - sector_nr;
2400         }
2401         /* before building a request, check if we can skip these blocks..
2402          * This call the bitmap_start_sync doesn't actually record anything
2403          */
2404         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2405             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2406                 /* We can skip this block, and probably several more */
2407                 *skipped = 1;
2408                 return sync_blocks;
2409         }
2410         /*
2411          * If there is non-resync activity waiting for a turn,
2412          * and resync is going fast enough,
2413          * then let it though before starting on this new sync request.
2414          */
2415         if (!go_faster && conf->nr_waiting)
2416                 msleep_interruptible(1000);
2417
2418         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2419         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2420         raise_barrier(conf);
2421
2422         conf->next_resync = sector_nr;
2423
2424         rcu_read_lock();
2425         /*
2426          * If we get a correctably read error during resync or recovery,
2427          * we might want to read from a different device.  So we
2428          * flag all drives that could conceivably be read from for READ,
2429          * and any others (which will be non-In_sync devices) for WRITE.
2430          * If a read fails, we try reading from something else for which READ
2431          * is OK.
2432          */
2433
2434         r1_bio->mddev = mddev;
2435         r1_bio->sector = sector_nr;
2436         r1_bio->state = 0;
2437         set_bit(R1BIO_IsSync, &r1_bio->state);
2438
2439         for (i = 0; i < conf->raid_disks * 2; i++) {
2440                 struct md_rdev *rdev;
2441                 bio = r1_bio->bios[i];
2442                 bio_reset(bio);
2443
2444                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2445                 if (rdev == NULL ||
2446                     test_bit(Faulty, &rdev->flags)) {
2447                         if (i < conf->raid_disks)
2448                                 still_degraded = 1;
2449                 } else if (!test_bit(In_sync, &rdev->flags)) {
2450                         bio->bi_rw = WRITE;
2451                         bio->bi_end_io = end_sync_write;
2452                         write_targets ++;
2453                 } else {
2454                         /* may need to read from here */
2455                         sector_t first_bad = MaxSector;
2456                         int bad_sectors;
2457
2458                         if (is_badblock(rdev, sector_nr, good_sectors,
2459                                         &first_bad, &bad_sectors)) {
2460                                 if (first_bad > sector_nr)
2461                                         good_sectors = first_bad - sector_nr;
2462                                 else {
2463                                         bad_sectors -= (sector_nr - first_bad);
2464                                         if (min_bad == 0 ||
2465                                             min_bad > bad_sectors)
2466                                                 min_bad = bad_sectors;
2467                                 }
2468                         }
2469                         if (sector_nr < first_bad) {
2470                                 if (test_bit(WriteMostly, &rdev->flags)) {
2471                                         if (wonly < 0)
2472                                                 wonly = i;
2473                                 } else {
2474                                         if (disk < 0)
2475                                                 disk = i;
2476                                 }
2477                                 bio->bi_rw = READ;
2478                                 bio->bi_end_io = end_sync_read;
2479                                 read_targets++;
2480                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2481                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2482                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2483                                 /*
2484                                  * The device is suitable for reading (InSync),
2485                                  * but has bad block(s) here. Let's try to correct them,
2486                                  * if we are doing resync or repair. Otherwise, leave
2487                                  * this device alone for this sync request.
2488                                  */
2489                                 bio->bi_rw = WRITE;
2490                                 bio->bi_end_io = end_sync_write;
2491                                 write_targets++;
2492                         }
2493                 }
2494                 if (bio->bi_end_io) {
2495                         atomic_inc(&rdev->nr_pending);
2496                         bio->bi_sector = sector_nr + rdev->data_offset;
2497                         bio->bi_bdev = rdev->bdev;
2498                         bio->bi_private = r1_bio;
2499                 }
2500         }
2501         rcu_read_unlock();
2502         if (disk < 0)
2503                 disk = wonly;
2504         r1_bio->read_disk = disk;
2505
2506         if (read_targets == 0 && min_bad > 0) {
2507                 /* These sectors are bad on all InSync devices, so we
2508                  * need to mark them bad on all write targets
2509                  */
2510                 int ok = 1;
2511                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2512                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2513                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2514                                 ok = rdev_set_badblocks(rdev, sector_nr,
2515                                                         min_bad, 0
2516                                         ) && ok;
2517                         }
2518                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2519                 *skipped = 1;
2520                 put_buf(r1_bio);
2521
2522                 if (!ok) {
2523                         /* Cannot record the badblocks, so need to
2524                          * abort the resync.
2525                          * If there are multiple read targets, could just
2526                          * fail the really bad ones ???
2527                          */
2528                         conf->recovery_disabled = mddev->recovery_disabled;
2529                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2530                         return 0;
2531                 } else
2532                         return min_bad;
2533
2534         }
2535         if (min_bad > 0 && min_bad < good_sectors) {
2536                 /* only resync enough to reach the next bad->good
2537                  * transition */
2538                 good_sectors = min_bad;
2539         }
2540
2541         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2542                 /* extra read targets are also write targets */
2543                 write_targets += read_targets-1;
2544
2545         if (write_targets == 0 || read_targets == 0) {
2546                 /* There is nowhere to write, so all non-sync
2547                  * drives must be failed - so we are finished
2548                  */
2549                 sector_t rv;
2550                 if (min_bad > 0)
2551                         max_sector = sector_nr + min_bad;
2552                 rv = max_sector - sector_nr;
2553                 *skipped = 1;
2554                 put_buf(r1_bio);
2555                 return rv;
2556         }
2557
2558         if (max_sector > mddev->resync_max)
2559                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2560         if (max_sector > sector_nr + good_sectors)
2561                 max_sector = sector_nr + good_sectors;
2562         nr_sectors = 0;
2563         sync_blocks = 0;
2564         do {
2565                 struct page *page;
2566                 int len = PAGE_SIZE;
2567                 if (sector_nr + (len>>9) > max_sector)
2568                         len = (max_sector - sector_nr) << 9;
2569                 if (len == 0)
2570                         break;
2571                 if (sync_blocks == 0) {
2572                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2573                                                &sync_blocks, still_degraded) &&
2574                             !conf->fullsync &&
2575                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2576                                 break;
2577                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2578                         if ((len >> 9) > sync_blocks)
2579                                 len = sync_blocks<<9;
2580                 }
2581
2582                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2583                         bio = r1_bio->bios[i];
2584                         if (bio->bi_end_io) {
2585                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2586                                 if (bio_add_page(bio, page, len, 0) == 0) {
2587                                         /* stop here */
2588                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2589                                         while (i > 0) {
2590                                                 i--;
2591                                                 bio = r1_bio->bios[i];
2592                                                 if (bio->bi_end_io==NULL)
2593                                                         continue;
2594                                                 /* remove last page from this bio */
2595                                                 bio->bi_vcnt--;
2596                                                 bio->bi_size -= len;
2597                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2598                                         }
2599                                         goto bio_full;
2600                                 }
2601                         }
2602                 }
2603                 nr_sectors += len>>9;
2604                 sector_nr += len>>9;
2605                 sync_blocks -= (len>>9);
2606         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2607  bio_full:
2608         r1_bio->sectors = nr_sectors;
2609
2610         /* For a user-requested sync, we read all readable devices and do a
2611          * compare
2612          */
2613         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2614                 atomic_set(&r1_bio->remaining, read_targets);
2615                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2616                         bio = r1_bio->bios[i];
2617                         if (bio->bi_end_io == end_sync_read) {
2618                                 read_targets--;
2619                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2620                                 generic_make_request(bio);
2621                         }
2622                 }
2623         } else {
2624                 atomic_set(&r1_bio->remaining, 1);
2625                 bio = r1_bio->bios[r1_bio->read_disk];
2626                 md_sync_acct(bio->bi_bdev, nr_sectors);
2627                 generic_make_request(bio);
2628
2629         }
2630         return nr_sectors;
2631 }
2632
2633 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2634 {
2635         if (sectors)
2636                 return sectors;
2637
2638         return mddev->dev_sectors;
2639 }
2640
2641 static struct r1conf *setup_conf(struct mddev *mddev)
2642 {
2643         struct r1conf *conf;
2644         int i;
2645         struct raid1_info *disk;
2646         struct md_rdev *rdev;
2647         int err = -ENOMEM;
2648
2649         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2650         if (!conf)
2651                 goto abort;
2652
2653         conf->mirrors = kzalloc(sizeof(struct raid1_info)
2654                                 * mddev->raid_disks * 2,
2655                                  GFP_KERNEL);
2656         if (!conf->mirrors)
2657                 goto abort;
2658
2659         conf->tmppage = alloc_page(GFP_KERNEL);
2660         if (!conf->tmppage)
2661                 goto abort;
2662
2663         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2664         if (!conf->poolinfo)
2665                 goto abort;
2666         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2667         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2668                                           r1bio_pool_free,
2669                                           conf->poolinfo);
2670         if (!conf->r1bio_pool)
2671                 goto abort;
2672
2673         conf->poolinfo->mddev = mddev;
2674
2675         err = -EINVAL;
2676         spin_lock_init(&conf->device_lock);
2677         rdev_for_each(rdev, mddev) {
2678                 struct request_queue *q;
2679                 int disk_idx = rdev->raid_disk;
2680                 if (disk_idx >= mddev->raid_disks
2681                     || disk_idx < 0)
2682                         continue;
2683                 if (test_bit(Replacement, &rdev->flags))
2684                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
2685                 else
2686                         disk = conf->mirrors + disk_idx;
2687
2688                 if (disk->rdev)
2689                         goto abort;
2690                 disk->rdev = rdev;
2691                 q = bdev_get_queue(rdev->bdev);
2692                 if (q->merge_bvec_fn)
2693                         mddev->merge_check_needed = 1;
2694
2695                 disk->head_position = 0;
2696                 disk->seq_start = MaxSector;
2697         }
2698         conf->raid_disks = mddev->raid_disks;
2699         conf->mddev = mddev;
2700         INIT_LIST_HEAD(&conf->retry_list);
2701
2702         spin_lock_init(&conf->resync_lock);
2703         init_waitqueue_head(&conf->wait_barrier);
2704
2705         bio_list_init(&conf->pending_bio_list);
2706         conf->pending_count = 0;
2707         conf->recovery_disabled = mddev->recovery_disabled - 1;
2708
2709         err = -EIO;
2710         for (i = 0; i < conf->raid_disks * 2; i++) {
2711
2712                 disk = conf->mirrors + i;
2713
2714                 if (i < conf->raid_disks &&
2715                     disk[conf->raid_disks].rdev) {
2716                         /* This slot has a replacement. */
2717                         if (!disk->rdev) {
2718                                 /* No original, just make the replacement
2719                                  * a recovering spare
2720                                  */
2721                                 disk->rdev =
2722                                         disk[conf->raid_disks].rdev;
2723                                 disk[conf->raid_disks].rdev = NULL;
2724                         } else if (!test_bit(In_sync, &disk->rdev->flags))
2725                                 /* Original is not in_sync - bad */
2726                                 goto abort;
2727                 }
2728
2729                 if (!disk->rdev ||
2730                     !test_bit(In_sync, &disk->rdev->flags)) {
2731                         disk->head_position = 0;
2732                         if (disk->rdev &&
2733                             (disk->rdev->saved_raid_disk < 0))
2734                                 conf->fullsync = 1;
2735                 }
2736         }
2737
2738         err = -ENOMEM;
2739         conf->thread = md_register_thread(raid1d, mddev, "raid1");
2740         if (!conf->thread) {
2741                 printk(KERN_ERR
2742                        "md/raid1:%s: couldn't allocate thread\n",
2743                        mdname(mddev));
2744                 goto abort;
2745         }
2746
2747         return conf;
2748
2749  abort:
2750         if (conf) {
2751                 if (conf->r1bio_pool)
2752                         mempool_destroy(conf->r1bio_pool);
2753                 kfree(conf->mirrors);
2754                 safe_put_page(conf->tmppage);
2755                 kfree(conf->poolinfo);
2756                 kfree(conf);
2757         }
2758         return ERR_PTR(err);
2759 }
2760
2761 static int stop(struct mddev *mddev);
2762 static int run(struct mddev *mddev)
2763 {
2764         struct r1conf *conf;
2765         int i;
2766         struct md_rdev *rdev;
2767         int ret;
2768         bool discard_supported = false;
2769
2770         if (mddev->level != 1) {
2771                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2772                        mdname(mddev), mddev->level);
2773                 return -EIO;
2774         }
2775         if (mddev->reshape_position != MaxSector) {
2776                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2777                        mdname(mddev));
2778                 return -EIO;
2779         }
2780         /*
2781          * copy the already verified devices into our private RAID1
2782          * bookkeeping area. [whatever we allocate in run(),
2783          * should be freed in stop()]
2784          */
2785         if (mddev->private == NULL)
2786                 conf = setup_conf(mddev);
2787         else
2788                 conf = mddev->private;
2789
2790         if (IS_ERR(conf))
2791                 return PTR_ERR(conf);
2792
2793         if (mddev->queue)
2794                 blk_queue_max_write_same_sectors(mddev->queue, 0);
2795
2796         rdev_for_each(rdev, mddev) {
2797                 if (!mddev->gendisk)
2798                         continue;
2799                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2800                                   rdev->data_offset << 9);
2801                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2802                         discard_supported = true;
2803         }
2804
2805         mddev->degraded = 0;
2806         for (i=0; i < conf->raid_disks; i++)
2807                 if (conf->mirrors[i].rdev == NULL ||
2808                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2809                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2810                         mddev->degraded++;
2811
2812         if (conf->raid_disks - mddev->degraded == 1)
2813                 mddev->recovery_cp = MaxSector;
2814
2815         if (mddev->recovery_cp != MaxSector)
2816                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2817                        " -- starting background reconstruction\n",
2818                        mdname(mddev));
2819         printk(KERN_INFO
2820                 "md/raid1:%s: active with %d out of %d mirrors\n",
2821                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2822                 mddev->raid_disks);
2823
2824         /*
2825          * Ok, everything is just fine now
2826          */
2827         mddev->thread = conf->thread;
2828         conf->thread = NULL;
2829         mddev->private = conf;
2830
2831         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2832
2833         if (mddev->queue) {
2834                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2835                 mddev->queue->backing_dev_info.congested_data = mddev;
2836                 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2837
2838                 if (discard_supported)
2839                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2840                                                 mddev->queue);
2841                 else
2842                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2843                                                   mddev->queue);
2844         }
2845
2846         ret =  md_integrity_register(mddev);
2847         if (ret)
2848                 stop(mddev);
2849         return ret;
2850 }
2851
2852 static int stop(struct mddev *mddev)
2853 {
2854         struct r1conf *conf = mddev->private;
2855         struct bitmap *bitmap = mddev->bitmap;
2856
2857         /* wait for behind writes to complete */
2858         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2859                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2860                        mdname(mddev));
2861                 /* need to kick something here to make sure I/O goes? */
2862                 wait_event(bitmap->behind_wait,
2863                            atomic_read(&bitmap->behind_writes) == 0);
2864         }
2865
2866         raise_barrier(conf);
2867         lower_barrier(conf);
2868
2869         md_unregister_thread(&mddev->thread);
2870         if (conf->r1bio_pool)
2871                 mempool_destroy(conf->r1bio_pool);
2872         kfree(conf->mirrors);
2873         safe_put_page(conf->tmppage);
2874         kfree(conf->poolinfo);
2875         kfree(conf);
2876         mddev->private = NULL;
2877         return 0;
2878 }
2879
2880 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2881 {
2882         /* no resync is happening, and there is enough space
2883          * on all devices, so we can resize.
2884          * We need to make sure resync covers any new space.
2885          * If the array is shrinking we should possibly wait until
2886          * any io in the removed space completes, but it hardly seems
2887          * worth it.
2888          */
2889         sector_t newsize = raid1_size(mddev, sectors, 0);
2890         if (mddev->external_size &&
2891             mddev->array_sectors > newsize)
2892                 return -EINVAL;
2893         if (mddev->bitmap) {
2894                 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2895                 if (ret)
2896                         return ret;
2897         }
2898         md_set_array_sectors(mddev, newsize);
2899         set_capacity(mddev->gendisk, mddev->array_sectors);
2900         revalidate_disk(mddev->gendisk);
2901         if (sectors > mddev->dev_sectors &&
2902             mddev->recovery_cp > mddev->dev_sectors) {
2903                 mddev->recovery_cp = mddev->dev_sectors;
2904                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2905         }
2906         mddev->dev_sectors = sectors;
2907         mddev->resync_max_sectors = sectors;
2908         return 0;
2909 }
2910
2911 static int raid1_reshape(struct mddev *mddev)
2912 {
2913         /* We need to:
2914          * 1/ resize the r1bio_pool
2915          * 2/ resize conf->mirrors
2916          *
2917          * We allocate a new r1bio_pool if we can.
2918          * Then raise a device barrier and wait until all IO stops.
2919          * Then resize conf->mirrors and swap in the new r1bio pool.
2920          *
2921          * At the same time, we "pack" the devices so that all the missing
2922          * devices have the higher raid_disk numbers.
2923          */
2924         mempool_t *newpool, *oldpool;
2925         struct pool_info *newpoolinfo;
2926         struct raid1_info *newmirrors;
2927         struct r1conf *conf = mddev->private;
2928         int cnt, raid_disks;
2929         unsigned long flags;
2930         int d, d2, err;
2931
2932         /* Cannot change chunk_size, layout, or level */
2933         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2934             mddev->layout != mddev->new_layout ||
2935             mddev->level != mddev->new_level) {
2936                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2937                 mddev->new_layout = mddev->layout;
2938                 mddev->new_level = mddev->level;
2939                 return -EINVAL;
2940         }
2941
2942         err = md_allow_write(mddev);
2943         if (err)
2944                 return err;
2945
2946         raid_disks = mddev->raid_disks + mddev->delta_disks;
2947
2948         if (raid_disks < conf->raid_disks) {
2949                 cnt=0;
2950                 for (d= 0; d < conf->raid_disks; d++)
2951                         if (conf->mirrors[d].rdev)
2952                                 cnt++;
2953                 if (cnt > raid_disks)
2954                         return -EBUSY;
2955         }
2956
2957         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2958         if (!newpoolinfo)
2959                 return -ENOMEM;
2960         newpoolinfo->mddev = mddev;
2961         newpoolinfo->raid_disks = raid_disks * 2;
2962
2963         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2964                                  r1bio_pool_free, newpoolinfo);
2965         if (!newpool) {
2966                 kfree(newpoolinfo);
2967                 return -ENOMEM;
2968         }
2969         newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
2970                              GFP_KERNEL);
2971         if (!newmirrors) {
2972                 kfree(newpoolinfo);
2973                 mempool_destroy(newpool);
2974                 return -ENOMEM;
2975         }
2976
2977         freeze_array(conf, 0);
2978
2979         /* ok, everything is stopped */
2980         oldpool = conf->r1bio_pool;
2981         conf->r1bio_pool = newpool;
2982
2983         for (d = d2 = 0; d < conf->raid_disks; d++) {
2984                 struct md_rdev *rdev = conf->mirrors[d].rdev;
2985                 if (rdev && rdev->raid_disk != d2) {
2986                         sysfs_unlink_rdev(mddev, rdev);
2987                         rdev->raid_disk = d2;
2988                         sysfs_unlink_rdev(mddev, rdev);
2989                         if (sysfs_link_rdev(mddev, rdev))
2990                                 printk(KERN_WARNING
2991                                        "md/raid1:%s: cannot register rd%d\n",
2992                                        mdname(mddev), rdev->raid_disk);
2993                 }
2994                 if (rdev)
2995                         newmirrors[d2++].rdev = rdev;
2996         }
2997         kfree(conf->mirrors);
2998         conf->mirrors = newmirrors;
2999         kfree(conf->poolinfo);
3000         conf->poolinfo = newpoolinfo;
3001
3002         spin_lock_irqsave(&conf->device_lock, flags);
3003         mddev->degraded += (raid_disks - conf->raid_disks);
3004         spin_unlock_irqrestore(&conf->device_lock, flags);
3005         conf->raid_disks = mddev->raid_disks = raid_disks;
3006         mddev->delta_disks = 0;
3007
3008         unfreeze_array(conf);
3009
3010         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3011         md_wakeup_thread(mddev->thread);
3012
3013         mempool_destroy(oldpool);
3014         return 0;
3015 }
3016
3017 static void raid1_quiesce(struct mddev *mddev, int state)
3018 {
3019         struct r1conf *conf = mddev->private;
3020
3021         switch(state) {
3022         case 2: /* wake for suspend */
3023                 wake_up(&conf->wait_barrier);
3024                 break;
3025         case 1:
3026                 raise_barrier(conf);
3027                 break;
3028         case 0:
3029                 lower_barrier(conf);
3030                 break;
3031         }
3032 }
3033
3034 static void *raid1_takeover(struct mddev *mddev)
3035 {
3036         /* raid1 can take over:
3037          *  raid5 with 2 devices, any layout or chunk size
3038          */
3039         if (mddev->level == 5 && mddev->raid_disks == 2) {
3040                 struct r1conf *conf;
3041                 mddev->new_level = 1;
3042                 mddev->new_layout = 0;
3043                 mddev->new_chunk_sectors = 0;
3044                 conf = setup_conf(mddev);
3045                 if (!IS_ERR(conf))
3046                         conf->barrier = 1;
3047                 return conf;
3048         }
3049         return ERR_PTR(-EINVAL);
3050 }
3051
3052 static struct md_personality raid1_personality =
3053 {
3054         .name           = "raid1",
3055         .level          = 1,
3056         .owner          = THIS_MODULE,
3057         .make_request   = make_request,
3058         .run            = run,
3059         .stop           = stop,
3060         .status         = status,
3061         .error_handler  = error,
3062         .hot_add_disk   = raid1_add_disk,
3063         .hot_remove_disk= raid1_remove_disk,
3064         .spare_active   = raid1_spare_active,
3065         .sync_request   = sync_request,
3066         .resize         = raid1_resize,
3067         .size           = raid1_size,
3068         .check_reshape  = raid1_reshape,
3069         .quiesce        = raid1_quiesce,
3070         .takeover       = raid1_takeover,
3071 };
3072
3073 static int __init raid_init(void)
3074 {
3075         return register_md_personality(&raid1_personality);
3076 }
3077
3078 static void raid_exit(void)
3079 {
3080         unregister_md_personality(&raid1_personality);
3081 }
3082
3083 module_init(raid_init);
3084 module_exit(raid_exit);
3085 MODULE_LICENSE("GPL");
3086 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3087 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3088 MODULE_ALIAS("md-raid1");
3089 MODULE_ALIAS("md-level-1");
3090
3091 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);