drivers/md/raid10.c

   1 /*
   2  * raid10.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 2000-2004 Neil Brown
   5  *
   6  * RAID-10 support for md.
   7  *
   8  * Base on code in raid1.c.  See raid1.c for further copyright information.
   9  *
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2, or (at your option)
  14  * any later version.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * (for example /usr/src/linux/COPYING); if not, write to the Free
  18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 #include <linux/slab.h>
  22 #include <linux/delay.h>
  23 #include <linux/blkdev.h>
  24 #include <linux/module.h>
  25 #include <linux/seq_file.h>
  26 #include <linux/ratelimit.h>
  27 #include "md.h"
  28 #include "raid10.h"
  29 #include "raid0.h"
  30 #include "bitmap.h"
  31
  32 /*
  33  * RAID10 provides a combination of RAID0 and RAID1 functionality.
  34  * The layout of data is defined by
  35  *    chunk_size
  36  *    raid_disks
  37  *    near_copies (stored in low byte of layout)
  38  *    far_copies (stored in second byte of layout)
  39  *    far_offset (stored in bit 16 of layout )
  40  *
  41  * The data to be stored is divided into chunks using chunksize.
  42  * Each device is divided into far_copies sections.
  43  * In each section, chunks are laid out in a style similar to raid0, but
  44  * near_copies copies of each chunk is stored (each on a different drive).
  45  * The starting device for each section is offset near_copies from the starting
  46  * device of the previous section.
  47  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
  48  * drive.
  49  * near_copies and far_copies must be at least one, and their product is at most
  50  * raid_disks.
  51  *
  52  * If far_offset is true, then the far_copies are handled a bit differently.
  53  * The copies are still in different stripes, but instead of be very far apart
  54  * on disk, there are adjacent stripes.
  55  */
  56
  57 /*
  58  * Number of guaranteed r10bios in case of extreme VM load:
  59  */
  60 #define NR_RAID10_BIOS 256
  61
  62 /* When there are this many requests queue to be written by
  63  * the raid10 thread, we become 'congested' to provide back-pressure
  64  * for writeback.
  65  */
  66 static int max_queued_requests = 1024;
  67
  68 static void allow_barrier(struct r10conf *conf);
  69 static void lower_barrier(struct r10conf *conf);
  70 static int enough(struct r10conf *conf, int ignore);
  71
  72 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
  73 {
  74         struct r10conf *conf = data;
  75         int size = offsetof(struct r10bio, devs[conf->copies]);
  76
  77         /* allocate a r10bio with room for raid_disks entries in the
  78          * bios array */
  79         return kzalloc(size, gfp_flags);
  80 }
  81
  82 static void r10bio_pool_free(void *r10_bio, void *data)
  83 {
  84         kfree(r10_bio);
  85 }
  86
  87 /* Maximum size of each resync request */
  88 #define RESYNC_BLOCK_SIZE (64*1024)
  89 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  90 /* amount of memory to reserve for resync requests */
  91 #define RESYNC_WINDOW (1024*1024)
  92 /* maximum number of concurrent requests, memory permitting */
  93 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
  94
  95 /*
  96  * When performing a resync, we need to read and compare, so
  97  * we need as many pages are there are copies.
  98  * When performing a recovery, we need 2 bios, one for read,
  99  * one for write (we recover only one drive per r10buf)
 100  *
 101  */
 102 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
 103 {
 104         struct r10conf *conf = data;
 105         struct page *page;
 106         struct r10bio *r10_bio;
 107         struct bio *bio;
 108         int i, j;
 109         int nalloc;
 110
 111         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 112         if (!r10_bio)
 113                 return NULL;
 114
 115         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
 116                 nalloc = conf->copies; /* resync */
 117         else
 118                 nalloc = 2; /* recovery */
 119
 120         /*
 121          * Allocate bios.
 122          */
 123         for (j = nalloc ; j-- ; ) {
 124                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 125                 if (!bio)
 126                         goto out_free_bio;
 127                 r10_bio->devs[j].bio = bio;
 128                 if (!conf->have_replacement)
 129                         continue;
 130                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 131                 if (!bio)
 132                         goto out_free_bio;
 133                 r10_bio->devs[j].repl_bio = bio;
 134         }
 135         /*
 136          * Allocate RESYNC_PAGES data pages and attach them
 137          * where needed.
 138          */
 139         for (j = 0 ; j < nalloc; j++) {
 140                 struct bio *rbio = r10_bio->devs[j].repl_bio;
 141                 bio = r10_bio->devs[j].bio;
 142                 for (i = 0; i < RESYNC_PAGES; i++) {
 143                         if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
 144                                                 &conf->mddev->recovery)) {
 145                                 /* we can share bv_page's during recovery */
 146                                 struct bio *rbio = r10_bio->devs[0].bio;
 147                                 page = rbio->bi_io_vec[i].bv_page;
 148                                 get_page(page);
 149                         } else
 150                                 page = alloc_page(gfp_flags);
 151                         if (unlikely(!page))
 152                                 goto out_free_pages;
 153
 154                         bio->bi_io_vec[i].bv_page = page;
 155                         if (rbio)
 156                                 rbio->bi_io_vec[i].bv_page = page;
 157                 }
 158         }
 159
 160         return r10_bio;
 161
 162 out_free_pages:
 163         for ( ; i > 0 ; i--)
 164                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
 165         while (j--)
 166                 for (i = 0; i < RESYNC_PAGES ; i++)
 167                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 168         j = -1;
 169 out_free_bio:
 170         while (++j < nalloc) {
 171                 bio_put(r10_bio->devs[j].bio);
 172                 if (r10_bio->devs[j].repl_bio)
 173                         bio_put(r10_bio->devs[j].repl_bio);
 174         }
 175         r10bio_pool_free(r10_bio, conf);
 176         return NULL;
 177 }
 178
 179 static void r10buf_pool_free(void *__r10_bio, void *data)
 180 {
 181         int i;
 182         struct r10conf *conf = data;
 183         struct r10bio *r10bio = __r10_bio;
 184         int j;
 185
 186         for (j=0; j < conf->copies; j++) {
 187                 struct bio *bio = r10bio->devs[j].bio;
 188                 if (bio) {
 189                         for (i = 0; i < RESYNC_PAGES; i++) {
 190                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 191                                 bio->bi_io_vec[i].bv_page = NULL;
 192                         }
 193                         bio_put(bio);
 194                 }
 195                 bio = r10bio->devs[j].repl_bio;
 196                 if (bio)
 197                         bio_put(bio);
 198         }
 199         r10bio_pool_free(r10bio, conf);
 200 }
 201
 202 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
 203 {
 204         int i;
 205
 206         for (i = 0; i < conf->copies; i++) {
 207                 struct bio **bio = & r10_bio->devs[i].bio;
 208                 if (!BIO_SPECIAL(*bio))
 209                         bio_put(*bio);
 210                 *bio = NULL;
 211                 bio = &r10_bio->devs[i].repl_bio;
 212                 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
 213                         bio_put(*bio);
 214                 *bio = NULL;
 215         }
 216 }
 217
 218 static void free_r10bio(struct r10bio *r10_bio)
 219 {
 220         struct r10conf *conf = r10_bio->mddev->private;
 221
 222         put_all_bios(conf, r10_bio);
 223         mempool_free(r10_bio, conf->r10bio_pool);
 224 }
 225
 226 static void put_buf(struct r10bio *r10_bio)
 227 {
 228         struct r10conf *conf = r10_bio->mddev->private;
 229
 230         mempool_free(r10_bio, conf->r10buf_pool);
 231
 232         lower_barrier(conf);
 233 }
 234
 235 static void reschedule_retry(struct r10bio *r10_bio)
 236 {
 237         unsigned long flags;
 238         struct mddev *mddev = r10_bio->mddev;
 239         struct r10conf *conf = mddev->private;
 240
 241         spin_lock_irqsave(&conf->device_lock, flags);
 242         list_add(&r10_bio->retry_list, &conf->retry_list);
 243         conf->nr_queued ++;
 244         spin_unlock_irqrestore(&conf->device_lock, flags);
 245
 246         /* wake up frozen array... */
 247         wake_up(&conf->wait_barrier);
 248
 249         md_wakeup_thread(mddev->thread);
 250 }
 251
 252 /*
 253  * raid_end_bio_io() is called when we have finished servicing a mirrored
 254  * operation and are ready to return a success/failure code to the buffer
 255  * cache layer.
 256  */
 257 static void raid_end_bio_io(struct r10bio *r10_bio)
 258 {
 259         struct bio *bio = r10_bio->master_bio;
 260         int done;
 261         struct r10conf *conf = r10_bio->mddev->private;
 262
 263         if (bio->bi_phys_segments) {
 264                 unsigned long flags;
 265                 spin_lock_irqsave(&conf->device_lock, flags);
 266                 bio->bi_phys_segments--;
 267                 done = (bio->bi_phys_segments == 0);
 268                 spin_unlock_irqrestore(&conf->device_lock, flags);
 269         } else
 270                 done = 1;
 271         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
 272                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 273         if (done) {
 274                 bio_endio(bio, 0);
 275                 /*
 276                  * Wake up any possible resync thread that waits for the device
 277                  * to go idle.
 278                  */
 279                 allow_barrier(conf);
 280         }
 281         free_r10bio(r10_bio);
 282 }
 283
 284 /*
 285  * Update disk head position estimator based on IRQ completion info.
 286  */
 287 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
 288 {
 289         struct r10conf *conf = r10_bio->mddev->private;
 290
 291         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 292                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 293 }
 294
 295 /*
 296  * Find the disk number which triggered given bio
 297  */
 298 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
 299                          struct bio *bio, int *slotp, int *replp)
 300 {
 301         int slot;
 302         int repl = 0;
 303
 304         for (slot = 0; slot < conf->copies; slot++) {
 305                 if (r10_bio->devs[slot].bio == bio)
 306                         break;
 307                 if (r10_bio->devs[slot].repl_bio == bio) {
 308                         repl = 1;
 309                         break;
 310                 }
 311         }
 312
 313         BUG_ON(slot == conf->copies);
 314         update_head_pos(slot, r10_bio);
 315
 316         if (slotp)
 317                 *slotp = slot;
 318         if (replp)
 319                 *replp = repl;
 320         return r10_bio->devs[slot].devnum;
 321 }
 322
 323 static void raid10_end_read_request(struct bio *bio, int error)
 324 {
 325         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 326         struct r10bio *r10_bio = bio->bi_private;
 327         int slot, dev;
 328         struct md_rdev *rdev;
 329         struct r10conf *conf = r10_bio->mddev->private;
 330
 331
 332         slot = r10_bio->read_slot;
 333         dev = r10_bio->devs[slot].devnum;
 334         rdev = r10_bio->devs[slot].rdev;
 335         /*
 336          * this branch is our 'one mirror IO has finished' event handler:
 337          */
 338         update_head_pos(slot, r10_bio);
 339
 340         if (uptodate) {
 341                 /*
 342                  * Set R10BIO_Uptodate in our master bio, so that
 343                  * we will return a good error code to the higher
 344                  * levels even if IO on some other mirrored buffer fails.
 345                  *
 346                  * The 'master' represents the composite IO operation to
 347                  * user-side. So if something waits for IO, then it will
 348                  * wait for the 'master' bio.
 349                  */
 350                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 351         } else {
 352                 /* If all other devices that store this block have
 353                  * failed, we want to return the error upwards rather
 354                  * than fail the last device.  Here we redefine
 355                  * "uptodate" to mean "Don't want to retry"
 356                  */
 357                 unsigned long flags;
 358                 spin_lock_irqsave(&conf->device_lock, flags);
 359                 if (!enough(conf, rdev->raid_disk))
 360                         uptodate = 1;
 361                 spin_unlock_irqrestore(&conf->device_lock, flags);
 362         }
 363         if (uptodate) {
 364                 raid_end_bio_io(r10_bio);
 365                 rdev_dec_pending(rdev, conf->mddev);
 366         } else {
 367                 /*
 368                  * oops, read error - keep the refcount on the rdev
 369                  */
 370                 char b[BDEVNAME_SIZE];
 371                 printk_ratelimited(KERN_ERR
 372                                    "md/raid10:%s: %s: rescheduling sector %llu\n",
 373                                    mdname(conf->mddev),
 374                                    bdevname(rdev->bdev, b),
 375                                    (unsigned long long)r10_bio->sector);
 376                 set_bit(R10BIO_ReadError, &r10_bio->state);
 377                 reschedule_retry(r10_bio);
 378         }
 379 }
 380
 381 static void close_write(struct r10bio *r10_bio)
 382 {
 383         /* clear the bitmap if all writes complete successfully */
 384         bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 385                         r10_bio->sectors,
 386                         !test_bit(R10BIO_Degraded, &r10_bio->state),
 387                         0);
 388         md_write_end(r10_bio->mddev);
 389 }
 390
 391 static void one_write_done(struct r10bio *r10_bio)
 392 {
 393         if (atomic_dec_and_test(&r10_bio->remaining)) {
 394                 if (test_bit(R10BIO_WriteError, &r10_bio->state))
 395                         reschedule_retry(r10_bio);
 396                 else {
 397                         close_write(r10_bio);
 398                         if (test_bit(R10BIO_MadeGood, &r10_bio->state))
 399                                 reschedule_retry(r10_bio);
 400                         else
 401                                 raid_end_bio_io(r10_bio);
 402                 }
 403         }
 404 }
 405
 406 static void raid10_end_write_request(struct bio *bio, int error)
 407 {
 408         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 409         struct r10bio *r10_bio = bio->bi_private;
 410         int dev;
 411         int dec_rdev = 1;
 412         struct r10conf *conf = r10_bio->mddev->private;
 413         int slot, repl;
 414         struct md_rdev *rdev = NULL;
 415
 416         dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
 417
 418         if (repl)
 419                 rdev = conf->mirrors[dev].replacement;
 420         if (!rdev) {
 421                 smp_rmb();
 422                 repl = 0;
 423                 rdev = conf->mirrors[dev].rdev;
 424         }
 425         /*
 426          * this branch is our 'one mirror IO has finished' event handler:
 427          */
 428         if (!uptodate) {
 429                 if (repl)
 430                         /* Never record new bad blocks to replacement,
 431                          * just fail it.
 432                          */
 433                         md_error(rdev->mddev, rdev);
 434                 else {
 435                         set_bit(WriteErrorSeen, &rdev->flags);
 436                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
 437                                 set_bit(MD_RECOVERY_NEEDED,
 438                                         &rdev->mddev->recovery);
 439                         set_bit(R10BIO_WriteError, &r10_bio->state);
 440                         dec_rdev = 0;
 441                 }
 442         } else {
 443                 /*
 444                  * Set R10BIO_Uptodate in our master bio, so that
 445                  * we will return a good error code for to the higher
 446                  * levels even if IO on some other mirrored buffer fails.
 447                  *
 448                  * The 'master' represents the composite IO operation to
 449                  * user-side. So if something waits for IO, then it will
 450                  * wait for the 'master' bio.
 451                  */
 452                 sector_t first_bad;
 453                 int bad_sectors;
 454
 455                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 456
 457                 /* Maybe we can clear some bad blocks. */
 458                 if (is_badblock(rdev,
 459                                 r10_bio->devs[slot].addr,
 460                                 r10_bio->sectors,
 461                                 &first_bad, &bad_sectors)) {
 462                         bio_put(bio);
 463                         if (repl)
 464                                 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
 465                         else
 466                                 r10_bio->devs[slot].bio = IO_MADE_GOOD;
 467                         dec_rdev = 0;
 468                         set_bit(R10BIO_MadeGood, &r10_bio->state);
 469                 }
 470         }
 471
 472         /*
 473          *
 474          * Let's see if all mirrored write operations have finished
 475          * already.
 476          */
 477         one_write_done(r10_bio);
 478         if (dec_rdev)
 479                 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 480 }
 481
 482 /*
 483  * RAID10 layout manager
 484  * As well as the chunksize and raid_disks count, there are two
 485  * parameters: near_copies and far_copies.
 486  * near_copies * far_copies must be <= raid_disks.
 487  * Normally one of these will be 1.
 488  * If both are 1, we get raid0.
 489  * If near_copies == raid_disks, we get raid1.
 490  *
 491  * Chunks are laid out in raid0 style with near_copies copies of the
 492  * first chunk, followed by near_copies copies of the next chunk and
 493  * so on.
 494  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 495  * as described above, we start again with a device offset of near_copies.
 496  * So we effectively have another copy of the whole array further down all
 497  * the drives, but with blocks on different drives.
 498  * With this layout, and block is never stored twice on the one device.
 499  *
 500  * raid10_find_phys finds the sector offset of a given virtual sector
 501  * on each device that it is on.
 502  *
 503  * raid10_find_virt does the reverse mapping, from a device and a
 504  * sector offset to a virtual address
 505  */
 506
 507 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
 508 {
 509         int n,f;
 510         sector_t sector;
 511         sector_t chunk;
 512         sector_t stripe;
 513         int dev;
 514
 515         int slot = 0;
 516
 517         /* now calculate first sector/dev */
 518         chunk = r10bio->sector >> conf->chunk_shift;
 519         sector = r10bio->sector & conf->chunk_mask;
 520
 521         chunk *= conf->near_copies;
 522         stripe = chunk;
 523         dev = sector_div(stripe, conf->raid_disks);
 524         if (conf->far_offset)
 525                 stripe *= conf->far_copies;
 526
 527         sector += stripe << conf->chunk_shift;
 528
 529         /* and calculate all the others */
 530         for (n=0; n < conf->near_copies; n++) {
 531                 int d = dev;
 532                 sector_t s = sector;
 533                 r10bio->devs[slot].addr = sector;
 534                 r10bio->devs[slot].devnum = d;
 535                 slot++;
 536
 537                 for (f = 1; f < conf->far_copies; f++) {
 538                         d += conf->near_copies;
 539                         if (d >= conf->raid_disks)
 540                                 d -= conf->raid_disks;
 541                         s += conf->stride;
 542                         r10bio->devs[slot].devnum = d;
 543                         r10bio->devs[slot].addr = s;
 544                         slot++;
 545                 }
 546                 dev++;
 547                 if (dev >= conf->raid_disks) {
 548                         dev = 0;
 549                         sector += (conf->chunk_mask + 1);
 550                 }
 551         }
 552         BUG_ON(slot != conf->copies);
 553 }
 554
 555 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
 556 {
 557         sector_t offset, chunk, vchunk;
 558
 559         offset = sector & conf->chunk_mask;
 560         if (conf->far_offset) {
 561                 int fc;
 562                 chunk = sector >> conf->chunk_shift;
 563                 fc = sector_div(chunk, conf->far_copies);
 564                 dev -= fc * conf->near_copies;
 565                 if (dev < 0)
 566                         dev += conf->raid_disks;
 567         } else {
 568                 while (sector >= conf->stride) {
 569                         sector -= conf->stride;
 570                         if (dev < conf->near_copies)
 571                                 dev += conf->raid_disks - conf->near_copies;
 572                         else
 573                                 dev -= conf->near_copies;
 574                 }
 575                 chunk = sector >> conf->chunk_shift;
 576         }
 577         vchunk = chunk * conf->raid_disks + dev;
 578         sector_div(vchunk, conf->near_copies);
 579         return (vchunk << conf->chunk_shift) + offset;
 580 }
 581
 582 /**
 583  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 584  *      @q: request queue
 585  *      @bvm: properties of new bio
 586  *      @biovec: the request that could be merged to it.
 587  *
 588  *      Return amount of bytes we can accept at this offset
 589  *      If near_copies == raid_disk, there are no striping issues,
 590  *      but in that case, the function isn't called at all.
 591  */
 592 static int raid10_mergeable_bvec(struct request_queue *q,
 593                                  struct bvec_merge_data *bvm,
 594                                  struct bio_vec *biovec)
 595 {
 596         struct mddev *mddev = q->queuedata;
 597         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 598         int max;
 599         unsigned int chunk_sectors = mddev->chunk_sectors;
 600         unsigned int bio_sectors = bvm->bi_size >> 9;
 601
 602         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 603         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 604         if (max <= biovec->bv_len && bio_sectors == 0)
 605                 return biovec->bv_len;
 606         else
 607                 return max;
 608 }
 609
 610 /*
 611  * This routine returns the disk from which the requested read should
 612  * be done. There is a per-array 'next expected sequential IO' sector
 613  * number - if this matches on the next IO then we use the last disk.
 614  * There is also a per-disk 'last know head position' sector that is
 615  * maintained from IRQ contexts, both the normal and the resync IO
 616  * completion handlers update this position correctly. If there is no
 617  * perfect sequential match then we pick the disk whose head is closest.
 618  *
 619  * If there are 2 mirrors in the same 2 devices, performance degrades
 620  * because position is mirror, not device based.
 621  *
 622  * The rdev for the device selected will have nr_pending incremented.
 623  */
 624
 625 /*
 626  * FIXME: possibly should rethink readbalancing and do it differently
 627  * depending on near_copies / far_copies geometry.
 628  */
 629 static struct md_rdev *read_balance(struct r10conf *conf,
 630                                     struct r10bio *r10_bio,
 631                                     int *max_sectors)
 632 {
 633         const sector_t this_sector = r10_bio->sector;
 634         int disk, slot;
 635         int sectors = r10_bio->sectors;
 636         int best_good_sectors;
 637         sector_t new_distance, best_dist;
 638         struct md_rdev *rdev, *best_rdev;
 639         int do_balance;
 640         int best_slot;
 641
 642         raid10_find_phys(conf, r10_bio);
 643         rcu_read_lock();
 644 retry:
 645         sectors = r10_bio->sectors;
 646         best_slot = -1;
 647         best_rdev = NULL;
 648         best_dist = MaxSector;
 649         best_good_sectors = 0;
 650         do_balance = 1;
 651         /*
 652          * Check if we can balance. We can balance on the whole
 653          * device if no resync is going on (recovery is ok), or below
 654          * the resync window. We take the first readable disk when
 655          * above the resync window.
 656          */
 657         if (conf->mddev->recovery_cp < MaxSector
 658             && (this_sector + sectors >= conf->next_resync))
 659                 do_balance = 0;
 660
 661         for (slot = 0; slot < conf->copies ; slot++) {
 662                 sector_t first_bad;
 663                 int bad_sectors;
 664                 sector_t dev_sector;
 665
 666                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
 667                         continue;
 668                 disk = r10_bio->devs[slot].devnum;
 669                 rdev = rcu_dereference(conf->mirrors[disk].replacement);
 670                 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
 671                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 672                         rdev = rcu_dereference(conf->mirrors[disk].rdev);
 673                 if (rdev == NULL)
 674                         continue;
 675                 if (test_bit(Faulty, &rdev->flags))
 676                         continue;
 677                 if (!test_bit(In_sync, &rdev->flags) &&
 678                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 679                         continue;
 680
 681                 dev_sector = r10_bio->devs[slot].addr;
 682                 if (is_badblock(rdev, dev_sector, sectors,
 683                                 &first_bad, &bad_sectors)) {
 684                         if (best_dist < MaxSector)
 685                                 /* Already have a better slot */
 686                                 continue;
 687                         if (first_bad <= dev_sector) {
 688                                 /* Cannot read here.  If this is the
 689                                  * 'primary' device, then we must not read
 690                                  * beyond 'bad_sectors' from another device.
 691                                  */
 692                                 bad_sectors -= (dev_sector - first_bad);
 693                                 if (!do_balance && sectors > bad_sectors)
 694                                         sectors = bad_sectors;
 695                                 if (best_good_sectors > sectors)
 696                                         best_good_sectors = sectors;
 697                         } else {
 698                                 sector_t good_sectors =
 699                                         first_bad - dev_sector;
 700                                 if (good_sectors > best_good_sectors) {
 701                                         best_good_sectors = good_sectors;
 702                                         best_slot = slot;
 703                                         best_rdev = rdev;
 704                                 }
 705                                 if (!do_balance)
 706                                         /* Must read from here */
 707                                         break;
 708                         }
 709                         continue;
 710                 } else
 711                         best_good_sectors = sectors;
 712
 713                 if (!do_balance)
 714                         break;
 715
 716                 /* This optimisation is debatable, and completely destroys
 717                  * sequential read speed for 'far copies' arrays.  So only
 718                  * keep it for 'near' arrays, and review those later.
 719                  */
 720                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
 721                         break;
 722
 723                 /* for far > 1 always use the lowest address */
 724                 if (conf->far_copies > 1)
 725                         new_distance = r10_bio->devs[slot].addr;
 726                 else
 727                         new_distance = abs(r10_bio->devs[slot].addr -
 728                                            conf->mirrors[disk].head_position);
 729                 if (new_distance < best_dist) {
 730                         best_dist = new_distance;
 731                         best_slot = slot;
 732                         best_rdev = rdev;
 733                 }
 734         }
 735         if (slot >= conf->copies) {
 736                 slot = best_slot;
 737                 rdev = best_rdev;
 738         }
 739
 740         if (slot >= 0) {
 741                 atomic_inc(&rdev->nr_pending);
 742                 if (test_bit(Faulty, &rdev->flags)) {
 743                         /* Cannot risk returning a device that failed
 744                          * before we inc'ed nr_pending
 745                          */
 746                         rdev_dec_pending(rdev, conf->mddev);
 747                         goto retry;
 748                 }
 749                 r10_bio->read_slot = slot;
 750         } else
 751                 rdev = NULL;
 752         rcu_read_unlock();
 753         *max_sectors = best_good_sectors;
 754
 755         return rdev;
 756 }
 757
 758 static int raid10_congested(void *data, int bits)
 759 {
 760         struct mddev *mddev = data;
 761         struct r10conf *conf = mddev->private;
 762         int i, ret = 0;
 763
 764         if ((bits & (1 << BDI_async_congested)) &&
 765             conf->pending_count >= max_queued_requests)
 766                 return 1;
 767
 768         if (mddev_congested(mddev, bits))
 769                 return 1;
 770         rcu_read_lock();
 771         for (i = 0; i < conf->raid_disks && ret == 0; i++) {
 772                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 773                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 774                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 775
 776                         ret |= bdi_congested(&q->backing_dev_info, bits);
 777                 }
 778         }
 779         rcu_read_unlock();
 780         return ret;
 781 }
 782
 783 static void flush_pending_writes(struct r10conf *conf)
 784 {
 785         /* Any writes that have been queued but are awaiting
 786          * bitmap updates get flushed here.
 787          */
 788         spin_lock_irq(&conf->device_lock);
 789
 790         if (conf->pending_bio_list.head) {
 791                 struct bio *bio;
 792                 bio = bio_list_get(&conf->pending_bio_list);
 793                 conf->pending_count = 0;
 794                 spin_unlock_irq(&conf->device_lock);
 795                 /* flush any pending bitmap writes to disk
 796                  * before proceeding w/ I/O */
 797                 bitmap_unplug(conf->mddev->bitmap);
 798                 wake_up(&conf->wait_barrier);
 799
 800                 while (bio) { /* submit pending writes */
 801                         struct bio *next = bio->bi_next;
 802                         bio->bi_next = NULL;
 803                         generic_make_request(bio);
 804                         bio = next;
 805                 }
 806         } else
 807                 spin_unlock_irq(&conf->device_lock);
 808 }
 809
 810 /* Barriers....
 811  * Sometimes we need to suspend IO while we do something else,
 812  * either some resync/recovery, or reconfigure the array.
 813  * To do this we raise a 'barrier'.
 814  * The 'barrier' is a counter that can be raised multiple times
 815  * to count how many activities are happening which preclude
 816  * normal IO.
 817  * We can only raise the barrier if there is no pending IO.
 818  * i.e. if nr_pending == 0.
 819  * We choose only to raise the barrier if no-one is waiting for the
 820  * barrier to go down.  This means that as soon as an IO request
 821  * is ready, no other operations which require a barrier will start
 822  * until the IO request has had a chance.
 823  *
 824  * So: regular IO calls 'wait_barrier'.  When that returns there
 825  *    is no backgroup IO happening,  It must arrange to call
 826  *    allow_barrier when it has finished its IO.
 827  * backgroup IO calls must call raise_barrier.  Once that returns
 828  *    there is no normal IO happeing.  It must arrange to call
 829  *    lower_barrier when the particular background IO completes.
 830  */
 831
 832 static void raise_barrier(struct r10conf *conf, int force)
 833 {
 834         BUG_ON(force && !conf->barrier);
 835         spin_lock_irq(&conf->resync_lock);
 836
 837         /* Wait until no block IO is waiting (unless 'force') */
 838         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 839                             conf->resync_lock, );
 840
 841         /* block any new IO from starting */
 842         conf->barrier++;
 843
 844         /* Now wait for all pending IO to complete */
 845         wait_event_lock_irq(conf->wait_barrier,
 846                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 847                             conf->resync_lock, );
 848
 849         spin_unlock_irq(&conf->resync_lock);
 850 }
 851
 852 static void lower_barrier(struct r10conf *conf)
 853 {
 854         unsigned long flags;
 855         spin_lock_irqsave(&conf->resync_lock, flags);
 856         conf->barrier--;
 857         spin_unlock_irqrestore(&conf->resync_lock, flags);
 858         wake_up(&conf->wait_barrier);
 859 }
 860
 861 static void wait_barrier(struct r10conf *conf)
 862 {
 863         spin_lock_irq(&conf->resync_lock);
 864         if (conf->barrier) {
 865                 conf->nr_waiting++;
 866                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 867                                     conf->resync_lock,
 868                                     );
 869                 conf->nr_waiting--;
 870         }
 871         conf->nr_pending++;
 872         spin_unlock_irq(&conf->resync_lock);
 873 }
 874
 875 static void allow_barrier(struct r10conf *conf)
 876 {
 877         unsigned long flags;
 878         spin_lock_irqsave(&conf->resync_lock, flags);
 879         conf->nr_pending--;
 880         spin_unlock_irqrestore(&conf->resync_lock, flags);
 881         wake_up(&conf->wait_barrier);
 882 }
 883
 884 static void freeze_array(struct r10conf *conf)
 885 {
 886         /* stop syncio and normal IO and wait for everything to
 887          * go quiet.
 888          * We increment barrier and nr_waiting, and then
 889          * wait until nr_pending match nr_queued+1
 890          * This is called in the context of one normal IO request
 891          * that has failed. Thus any sync request that might be pending
 892          * will be blocked by nr_pending, and we need to wait for
 893          * pending IO requests to complete or be queued for re-try.
 894          * Thus the number queued (nr_queued) plus this request (1)
 895          * must match the number of pending IOs (nr_pending) before
 896          * we continue.
 897          */
 898         spin_lock_irq(&conf->resync_lock);
 899         conf->barrier++;
 900         conf->nr_waiting++;
 901         wait_event_lock_irq(conf->wait_barrier,
 902                             conf->nr_pending == conf->nr_queued+1,
 903                             conf->resync_lock,
 904                             flush_pending_writes(conf));
 905
 906         spin_unlock_irq(&conf->resync_lock);
 907 }
 908
 909 static void unfreeze_array(struct r10conf *conf)
 910 {
 911         /* reverse the effect of the freeze */
 912         spin_lock_irq(&conf->resync_lock);
 913         conf->barrier--;
 914         conf->nr_waiting--;
 915         wake_up(&conf->wait_barrier);
 916         spin_unlock_irq(&conf->resync_lock);
 917 }
 918
 919 static void make_request(struct mddev *mddev, struct bio * bio)
 920 {
 921         struct r10conf *conf = mddev->private;
 922         struct r10bio *r10_bio;
 923         struct bio *read_bio;
 924         int i;
 925         int chunk_sects = conf->chunk_mask + 1;
 926         const int rw = bio_data_dir(bio);
 927         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 928         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
 929         unsigned long flags;
 930         struct md_rdev *blocked_rdev;
 931         int plugged;
 932         int sectors_handled;
 933         int max_sectors;
 934
 935         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
 936                 md_flush_request(mddev, bio);
 937                 return;
 938         }
 939
 940         /* If this request crosses a chunk boundary, we need to
 941          * split it.  This will only happen for 1 PAGE (or less) requests.
 942          */
 943         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 944                       > chunk_sects &&
 945                     conf->near_copies < conf->raid_disks)) {
 946                 struct bio_pair *bp;
 947                 /* Sanity check -- queue functions should prevent this happening */
 948                 if (bio->bi_vcnt != 1 ||
 949                     bio->bi_idx != 0)
 950                         goto bad_map;
 951                 /* This is a one page bio that upper layers
 952                  * refuse to split for us, so we need to split it.
 953                  */
 954                 bp = bio_split(bio,
 955                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 956
 957                 /* Each of these 'make_request' calls will call 'wait_barrier'.
 958                  * If the first succeeds but the second blocks due to the resync
 959                  * thread raising the barrier, we will deadlock because the
 960                  * IO to the underlying device will be queued in generic_make_request
 961                  * and will never complete, so will never reduce nr_pending.
 962                  * So increment nr_waiting here so no new raise_barriers will
 963                  * succeed, and so the second wait_barrier cannot block.
 964                  */
 965                 spin_lock_irq(&conf->resync_lock);
 966                 conf->nr_waiting++;
 967                 spin_unlock_irq(&conf->resync_lock);
 968
 969                 make_request(mddev, &bp->bio1);
 970                 make_request(mddev, &bp->bio2);
 971
 972                 spin_lock_irq(&conf->resync_lock);
 973                 conf->nr_waiting--;
 974                 wake_up(&conf->wait_barrier);
 975                 spin_unlock_irq(&conf->resync_lock);
 976
 977                 bio_pair_release(bp);
 978                 return;
 979         bad_map:
 980                 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
 981                        " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
 982                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 983
 984                 bio_io_error(bio);
 985                 return;
 986         }
 987
 988         md_write_start(mddev, bio);
 989
 990         /*
 991          * Register the new request and wait if the reconstruction
 992          * thread has put up a bar for new requests.
 993          * Continue immediately if no resync is active currently.
 994          */
 995         wait_barrier(conf);
 996
 997         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 998
 999         r10_bio->master_bio = bio;
1000         r10_bio->sectors = bio->bi_size >> 9;
1001
1002         r10_bio->mddev = mddev;
1003         r10_bio->sector = bio->bi_sector;
1004         r10_bio->state = 0;
1005
1006         /* We might need to issue multiple reads to different
1007          * devices if there are bad blocks around, so we keep
1008          * track of the number of reads in bio->bi_phys_segments.
1009          * If this is 0, there is only one r10_bio and no locking
1010          * will be needed when the request completes.  If it is
1011          * non-zero, then it is the number of not-completed requests.
1012          */
1013         bio->bi_phys_segments = 0;
1014         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1015
1016         if (rw == READ) {
1017                 /*
1018                  * read balancing logic:
1019                  */
1020                 struct md_rdev *rdev;
1021                 int slot;
1022
1023 read_again:
1024                 rdev = read_balance(conf, r10_bio, &max_sectors);
1025                 if (!rdev) {
1026                         raid_end_bio_io(r10_bio);
1027                         return;
1028                 }
1029                 slot = r10_bio->read_slot;
1030
1031                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1032                 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1033                             max_sectors);
1034
1035                 r10_bio->devs[slot].bio = read_bio;
1036                 r10_bio->devs[slot].rdev = rdev;
1037
1038                 read_bio->bi_sector = r10_bio->devs[slot].addr +
1039                         rdev->data_offset;
1040                 read_bio->bi_bdev = rdev->bdev;
1041                 read_bio->bi_end_io = raid10_end_read_request;
1042                 read_bio->bi_rw = READ | do_sync;
1043                 read_bio->bi_private = r10_bio;
1044
1045                 if (max_sectors < r10_bio->sectors) {
1046                         /* Could not read all from this device, so we will
1047                          * need another r10_bio.
1048                          */
1049                         sectors_handled = (r10_bio->sectors + max_sectors
1050                                            - bio->bi_sector);
1051                         r10_bio->sectors = max_sectors;
1052                         spin_lock_irq(&conf->device_lock);
1053                         if (bio->bi_phys_segments == 0)
1054                                 bio->bi_phys_segments = 2;
1055                         else
1056                                 bio->bi_phys_segments++;
1057                         spin_unlock(&conf->device_lock);
1058                         /* Cannot call generic_make_request directly
1059                          * as that will be queued in __generic_make_request
1060                          * and subsequent mempool_alloc might block
1061                          * waiting for it.  so hand bio over to raid10d.
1062                          */
1063                         reschedule_retry(r10_bio);
1064
1065                         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1066
1067                         r10_bio->master_bio = bio;
1068                         r10_bio->sectors = ((bio->bi_size >> 9)
1069                                             - sectors_handled);
1070                         r10_bio->state = 0;
1071                         r10_bio->mddev = mddev;
1072                         r10_bio->sector = bio->bi_sector + sectors_handled;
1073                         goto read_again;
1074                 } else
1075                         generic_make_request(read_bio);
1076                 return;
1077         }
1078
1079         /*
1080          * WRITE:
1081          */
1082         if (conf->pending_count >= max_queued_requests) {
1083                 md_wakeup_thread(mddev->thread);
1084                 wait_event(conf->wait_barrier,
1085                            conf->pending_count < max_queued_requests);
1086         }
1087         /* first select target devices under rcu_lock and
1088          * inc refcount on their rdev.  Record them by setting
1089          * bios[x] to bio
1090          * If there are known/acknowledged bad blocks on any device
1091          * on which we have seen a write error, we want to avoid
1092          * writing to those blocks.  This potentially requires several
1093          * writes to write around the bad blocks.  Each set of writes
1094          * gets its own r10_bio with a set of bios attached.  The number
1095          * of r10_bios is recored in bio->bi_phys_segments just as with
1096          * the read case.
1097          */
1098         plugged = mddev_check_plugged(mddev);
1099
1100         r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1101         raid10_find_phys(conf, r10_bio);
1102 retry_write:
1103         blocked_rdev = NULL;
1104         rcu_read_lock();
1105         max_sectors = r10_bio->sectors;
1106
1107         for (i = 0;  i < conf->copies; i++) {
1108                 int d = r10_bio->devs[i].devnum;
1109                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1110                 struct md_rdev *rrdev = rcu_dereference(
1111                         conf->mirrors[d].replacement);
1112                 if (rdev == rrdev)
1113                         rrdev = NULL;
1114                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1115                         atomic_inc(&rdev->nr_pending);
1116                         blocked_rdev = rdev;
1117                         break;
1118                 }
1119                 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1120                         atomic_inc(&rrdev->nr_pending);
1121                         blocked_rdev = rrdev;
1122                         break;
1123                 }
1124                 if (rrdev && test_bit(Faulty, &rrdev->flags))
1125                         rrdev = NULL;
1126
1127                 r10_bio->devs[i].bio = NULL;
1128                 r10_bio->devs[i].repl_bio = NULL;
1129                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1130                         set_bit(R10BIO_Degraded, &r10_bio->state);
1131                         continue;
1132                 }
1133                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1134                         sector_t first_bad;
1135                         sector_t dev_sector = r10_bio->devs[i].addr;
1136                         int bad_sectors;
1137                         int is_bad;
1138
1139                         is_bad = is_badblock(rdev, dev_sector,
1140                                              max_sectors,
1141                                              &first_bad, &bad_sectors);
1142                         if (is_bad < 0) {
1143                                 /* Mustn't write here until the bad block
1144                                  * is acknowledged
1145                                  */
1146                                 atomic_inc(&rdev->nr_pending);
1147                                 set_bit(BlockedBadBlocks, &rdev->flags);
1148                                 blocked_rdev = rdev;
1149                                 break;
1150                         }
1151                         if (is_bad && first_bad <= dev_sector) {
1152                                 /* Cannot write here at all */
1153                                 bad_sectors -= (dev_sector - first_bad);
1154                                 if (bad_sectors < max_sectors)
1155                                         /* Mustn't write more than bad_sectors
1156                                          * to other devices yet
1157                                          */
1158                                         max_sectors = bad_sectors;
1159                                 /* We don't set R10BIO_Degraded as that
1160                                  * only applies if the disk is missing,
1161                                  * so it might be re-added, and we want to
1162                                  * know to recover this chunk.
1163                                  * In this case the device is here, and the
1164                                  * fact that this chunk is not in-sync is
1165                                  * recorded in the bad block log.
1166                                  */
1167                                 continue;
1168                         }
1169                         if (is_bad) {
1170                                 int good_sectors = first_bad - dev_sector;
1171                                 if (good_sectors < max_sectors)
1172                                         max_sectors = good_sectors;
1173                         }
1174                 }
1175                 r10_bio->devs[i].bio = bio;
1176                 atomic_inc(&rdev->nr_pending);
1177                 if (rrdev) {
1178                         r10_bio->devs[i].repl_bio = bio;
1179                         atomic_inc(&rrdev->nr_pending);
1180                 }
1181         }
1182         rcu_read_unlock();
1183
1184         if (unlikely(blocked_rdev)) {
1185                 /* Have to wait for this device to get unblocked, then retry */
1186                 int j;
1187                 int d;
1188
1189                 for (j = 0; j < i; j++) {
1190                         if (r10_bio->devs[j].bio) {
1191                                 d = r10_bio->devs[j].devnum;
1192                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1193                         }
1194                         if (r10_bio->devs[j].repl_bio) {
1195                                 struct md_rdev *rdev;
1196                                 d = r10_bio->devs[j].devnum;
1197                                 rdev = conf->mirrors[d].replacement;
1198                                 if (!rdev) {
1199                                         /* Race with remove_disk */
1200                                         smp_mb();
1201                                         rdev = conf->mirrors[d].rdev;
1202                                 }
1203                                 rdev_dec_pending(rdev, mddev);
1204                         }
1205                 }
1206                 allow_barrier(conf);
1207                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1208                 wait_barrier(conf);
1209                 goto retry_write;
1210         }
1211
1212         if (max_sectors < r10_bio->sectors) {
1213                 /* We are splitting this into multiple parts, so
1214                  * we need to prepare for allocating another r10_bio.
1215                  */
1216                 r10_bio->sectors = max_sectors;
1217                 spin_lock_irq(&conf->device_lock);
1218                 if (bio->bi_phys_segments == 0)
1219                         bio->bi_phys_segments = 2;
1220                 else
1221                         bio->bi_phys_segments++;
1222                 spin_unlock_irq(&conf->device_lock);
1223         }
1224         sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1225
1226         atomic_set(&r10_bio->remaining, 1);
1227         bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1228
1229         for (i = 0; i < conf->copies; i++) {
1230                 struct bio *mbio;
1231                 int d = r10_bio->devs[i].devnum;
1232                 if (!r10_bio->devs[i].bio)
1233                         continue;
1234
1235                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1236                 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1237                             max_sectors);
1238                 r10_bio->devs[i].bio = mbio;
1239
1240                 mbio->bi_sector = (r10_bio->devs[i].addr+
1241                                    conf->mirrors[d].rdev->data_offset);
1242                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1243                 mbio->bi_end_io = raid10_end_write_request;
1244                 mbio->bi_rw = WRITE | do_sync | do_fua;
1245                 mbio->bi_private = r10_bio;
1246
1247                 atomic_inc(&r10_bio->remaining);
1248                 spin_lock_irqsave(&conf->device_lock, flags);
1249                 bio_list_add(&conf->pending_bio_list, mbio);
1250                 conf->pending_count++;
1251                 spin_unlock_irqrestore(&conf->device_lock, flags);
1252
1253                 if (!r10_bio->devs[i].repl_bio)
1254                         continue;
1255
1256                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1257                 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1258                             max_sectors);
1259                 r10_bio->devs[i].repl_bio = mbio;
1260
1261                 /* We are actively writing to the original device
1262                  * so it cannot disappear, so the replacement cannot
1263                  * become NULL here
1264                  */
1265                 mbio->bi_sector = (r10_bio->devs[i].addr+
1266                                    conf->mirrors[d].replacement->data_offset);
1267                 mbio->bi_bdev = conf->mirrors[d].replacement->bdev;
1268                 mbio->bi_end_io = raid10_end_write_request;
1269                 mbio->bi_rw = WRITE | do_sync | do_fua;
1270                 mbio->bi_private = r10_bio;
1271
1272                 atomic_inc(&r10_bio->remaining);
1273                 spin_lock_irqsave(&conf->device_lock, flags);
1274                 bio_list_add(&conf->pending_bio_list, mbio);
1275                 conf->pending_count++;
1276                 spin_unlock_irqrestore(&conf->device_lock, flags);
1277         }
1278
1279         /* Don't remove the bias on 'remaining' (one_write_done) until
1280          * after checking if we need to go around again.
1281          */
1282
1283         if (sectors_handled < (bio->bi_size >> 9)) {
1284                 one_write_done(r10_bio);
1285                 /* We need another r10_bio.  It has already been counted
1286                  * in bio->bi_phys_segments.
1287                  */
1288                 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1289
1290                 r10_bio->master_bio = bio;
1291                 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1292
1293                 r10_bio->mddev = mddev;
1294                 r10_bio->sector = bio->bi_sector + sectors_handled;
1295                 r10_bio->state = 0;
1296                 goto retry_write;
1297         }
1298         one_write_done(r10_bio);
1299
1300         /* In case raid10d snuck in to freeze_array */
1301         wake_up(&conf->wait_barrier);
1302
1303         if (do_sync || !mddev->bitmap || !plugged)
1304                 md_wakeup_thread(mddev->thread);
1305 }
1306
1307 static void status(struct seq_file *seq, struct mddev *mddev)
1308 {
1309         struct r10conf *conf = mddev->private;
1310         int i;
1311
1312         if (conf->near_copies < conf->raid_disks)
1313                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1314         if (conf->near_copies > 1)
1315                 seq_printf(seq, " %d near-copies", conf->near_copies);
1316         if (conf->far_copies > 1) {
1317                 if (conf->far_offset)
1318                         seq_printf(seq, " %d offset-copies", conf->far_copies);
1319                 else
1320                         seq_printf(seq, " %d far-copies", conf->far_copies);
1321         }
1322         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1323                                         conf->raid_disks - mddev->degraded);
1324         for (i = 0; i < conf->raid_disks; i++)
1325                 seq_printf(seq, "%s",
1326                               conf->mirrors[i].rdev &&
1327                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1328         seq_printf(seq, "]");
1329 }
1330
1331 /* check if there are enough drives for
1332  * every block to appear on atleast one.
1333  * Don't consider the device numbered 'ignore'
1334  * as we might be about to remove it.
1335  */
1336 static int enough(struct r10conf *conf, int ignore)
1337 {
1338         int first = 0;
1339
1340         do {
1341                 int n = conf->copies;
1342                 int cnt = 0;
1343                 while (n--) {
1344                         if (conf->mirrors[first].rdev &&
1345                             first != ignore)
1346                                 cnt++;
1347                         first = (first+1) % conf->raid_disks;
1348                 }
1349                 if (cnt == 0)
1350                         return 0;
1351         } while (first != 0);
1352         return 1;
1353 }
1354
1355 static void error(struct mddev *mddev, struct md_rdev *rdev)
1356 {
1357         char b[BDEVNAME_SIZE];
1358         struct r10conf *conf = mddev->private;
1359
1360         /*
1361          * If it is not operational, then we have already marked it as dead
1362          * else if it is the last working disks, ignore the error, let the
1363          * next level up know.
1364          * else mark the drive as failed
1365          */
1366         if (test_bit(In_sync, &rdev->flags)
1367             && !enough(conf, rdev->raid_disk))
1368                 /*
1369                  * Don't fail the drive, just return an IO error.
1370                  */
1371                 return;
1372         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1373                 unsigned long flags;
1374                 spin_lock_irqsave(&conf->device_lock, flags);
1375                 mddev->degraded++;
1376                 spin_unlock_irqrestore(&conf->device_lock, flags);
1377                 /*
1378                  * if recovery is running, make sure it aborts.
1379                  */
1380                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1381         }
1382         set_bit(Blocked, &rdev->flags);
1383         set_bit(Faulty, &rdev->flags);
1384         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1385         printk(KERN_ALERT
1386                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1387                "md/raid10:%s: Operation continuing on %d devices.\n",
1388                mdname(mddev), bdevname(rdev->bdev, b),
1389                mdname(mddev), conf->raid_disks - mddev->degraded);
1390 }
1391
1392 static void print_conf(struct r10conf *conf)
1393 {
1394         int i;
1395         struct mirror_info *tmp;
1396
1397         printk(KERN_DEBUG "RAID10 conf printout:\n");
1398         if (!conf) {
1399                 printk(KERN_DEBUG "(!conf)\n");
1400                 return;
1401         }
1402         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1403                 conf->raid_disks);
1404
1405         for (i = 0; i < conf->raid_disks; i++) {
1406                 char b[BDEVNAME_SIZE];
1407                 tmp = conf->mirrors + i;
1408                 if (tmp->rdev)
1409                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1410                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1411                                 !test_bit(Faulty, &tmp->rdev->flags),
1412                                 bdevname(tmp->rdev->bdev,b));
1413         }
1414 }
1415
1416 static void close_sync(struct r10conf *conf)
1417 {
1418         wait_barrier(conf);
1419         allow_barrier(conf);
1420
1421         mempool_destroy(conf->r10buf_pool);
1422         conf->r10buf_pool = NULL;
1423 }
1424
1425 static int raid10_spare_active(struct mddev *mddev)
1426 {
1427         int i;
1428         struct r10conf *conf = mddev->private;
1429         struct mirror_info *tmp;
1430         int count = 0;
1431         unsigned long flags;
1432
1433         /*
1434          * Find all non-in_sync disks within the RAID10 configuration
1435          * and mark them in_sync
1436          */
1437         for (i = 0; i < conf->raid_disks; i++) {
1438                 tmp = conf->mirrors + i;
1439                 if (tmp->replacement
1440                     && tmp->replacement->recovery_offset == MaxSector
1441                     && !test_bit(Faulty, &tmp->replacement->flags)
1442                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1443                         /* Replacement has just become active */
1444                         if (!tmp->rdev
1445                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1446                                 count++;
1447                         if (tmp->rdev) {
1448                                 /* Replaced device not technically faulty,
1449                                  * but we need to be sure it gets removed
1450                                  * and never re-added.
1451                                  */
1452                                 set_bit(Faulty, &tmp->rdev->flags);
1453                                 sysfs_notify_dirent_safe(
1454                                         tmp->rdev->sysfs_state);
1455                         }
1456                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1457                 } else if (tmp->rdev
1458                            && !test_bit(Faulty, &tmp->rdev->flags)
1459                            && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1460                         count++;
1461                         sysfs_notify_dirent(tmp->rdev->sysfs_state);
1462                 }
1463         }
1464         spin_lock_irqsave(&conf->device_lock, flags);
1465         mddev->degraded -= count;
1466         spin_unlock_irqrestore(&conf->device_lock, flags);
1467
1468         print_conf(conf);
1469         return count;
1470 }
1471
1472
1473 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1474 {
1475         struct r10conf *conf = mddev->private;
1476         int err = -EEXIST;
1477         int mirror;
1478         int first = 0;
1479         int last = conf->raid_disks - 1;
1480
1481         if (mddev->recovery_cp < MaxSector)
1482                 /* only hot-add to in-sync arrays, as recovery is
1483                  * very different from resync
1484                  */
1485                 return -EBUSY;
1486         if (!enough(conf, -1))
1487                 return -EINVAL;
1488
1489         if (rdev->raid_disk >= 0)
1490                 first = last = rdev->raid_disk;
1491
1492         if (rdev->saved_raid_disk >= first &&
1493             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1494                 mirror = rdev->saved_raid_disk;
1495         else
1496                 mirror = first;
1497         for ( ; mirror <= last ; mirror++) {
1498                 struct mirror_info *p = &conf->mirrors[mirror];
1499                 if (p->recovery_disabled == mddev->recovery_disabled)
1500                         continue;
1501                 if (p->rdev) {
1502                         if (!test_bit(WantReplacement, &p->rdev->flags) ||
1503                             p->replacement != NULL)
1504                                 continue;
1505                         clear_bit(In_sync, &rdev->flags);
1506                         set_bit(Replacement, &rdev->flags);
1507                         rdev->raid_disk = mirror;
1508                         err = 0;
1509                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1510                                           rdev->data_offset << 9);
1511                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1512                                 blk_queue_max_segments(mddev->queue, 1);
1513                                 blk_queue_segment_boundary(mddev->queue,
1514                                                            PAGE_CACHE_SIZE - 1);
1515                         }
1516                         conf->fullsync = 1;
1517                         rcu_assign_pointer(p->replacement, rdev);
1518                         break;
1519                 }
1520
1521                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1522                                   rdev->data_offset << 9);
1523                 /* as we don't honour merge_bvec_fn, we must
1524                  * never risk violating it, so limit
1525                  * ->max_segments to one lying with a single
1526                  * page, as a one page request is never in
1527                  * violation.
1528                  */
1529                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1530                         blk_queue_max_segments(mddev->queue, 1);
1531                         blk_queue_segment_boundary(mddev->queue,
1532                                                    PAGE_CACHE_SIZE - 1);
1533                 }
1534
1535                 p->head_position = 0;
1536                 p->recovery_disabled = mddev->recovery_disabled - 1;
1537                 rdev->raid_disk = mirror;
1538                 err = 0;
1539                 if (rdev->saved_raid_disk != mirror)
1540                         conf->fullsync = 1;
1541                 rcu_assign_pointer(p->rdev, rdev);
1542                 break;
1543         }
1544
1545         md_integrity_add_rdev(rdev, mddev);
1546         print_conf(conf);
1547         return err;
1548 }
1549
1550 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1551 {
1552         struct r10conf *conf = mddev->private;
1553         int err = 0;
1554         int number = rdev->raid_disk;
1555         struct md_rdev **rdevp;
1556         struct mirror_info *p = conf->mirrors + number;
1557
1558         print_conf(conf);
1559         if (rdev == p->rdev)
1560                 rdevp = &p->rdev;
1561         else if (rdev == p->replacement)
1562                 rdevp = &p->replacement;
1563         else
1564                 return 0;
1565
1566         if (test_bit(In_sync, &rdev->flags) ||
1567             atomic_read(&rdev->nr_pending)) {
1568                 err = -EBUSY;
1569                 goto abort;
1570         }
1571         /* Only remove faulty devices if recovery
1572          * is not possible.
1573          */
1574         if (!test_bit(Faulty, &rdev->flags) &&
1575             mddev->recovery_disabled != p->recovery_disabled &&
1576             (!p->replacement || p->replacement == rdev) &&
1577             enough(conf, -1)) {
1578                 err = -EBUSY;
1579                 goto abort;
1580         }
1581         *rdevp = NULL;
1582         synchronize_rcu();
1583         if (atomic_read(&rdev->nr_pending)) {
1584                 /* lost the race, try later */
1585                 err = -EBUSY;
1586                 *rdevp = rdev;
1587                 goto abort;
1588         } else if (p->replacement) {
1589                 /* We must have just cleared 'rdev' */
1590                 p->rdev = p->replacement;
1591                 clear_bit(Replacement, &p->replacement->flags);
1592                 smp_mb(); /* Make sure other CPUs may see both as identical
1593                            * but will never see neither -- if they are careful.
1594                            */
1595                 p->replacement = NULL;
1596                 clear_bit(WantReplacement, &rdev->flags);
1597         } else
1598                 /* We might have just remove the Replacement as faulty
1599                  * Clear the flag just in case
1600                  */
1601                 clear_bit(WantReplacement, &rdev->flags);
1602
1603         err = md_integrity_register(mddev);
1604
1605 abort:
1606
1607         print_conf(conf);
1608         return err;
1609 }
1610
1611
1612 static void end_sync_read(struct bio *bio, int error)
1613 {
1614         struct r10bio *r10_bio = bio->bi_private;
1615         struct r10conf *conf = r10_bio->mddev->private;
1616         int d;
1617
1618         d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1619
1620         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1621                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1622         else
1623                 /* The write handler will notice the lack of
1624                  * R10BIO_Uptodate and record any errors etc
1625                  */
1626                 atomic_add(r10_bio->sectors,
1627                            &conf->mirrors[d].rdev->corrected_errors);
1628
1629         /* for reconstruct, we always reschedule after a read.
1630          * for resync, only after all reads
1631          */
1632         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1633         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1634             atomic_dec_and_test(&r10_bio->remaining)) {
1635                 /* we have read all the blocks,
1636                  * do the comparison in process context in raid10d
1637                  */
1638                 reschedule_retry(r10_bio);
1639         }
1640 }
1641
1642 static void end_sync_request(struct r10bio *r10_bio)
1643 {
1644         struct mddev *mddev = r10_bio->mddev;
1645
1646         while (atomic_dec_and_test(&r10_bio->remaining)) {
1647                 if (r10_bio->master_bio == NULL) {
1648                         /* the primary of several recovery bios */
1649                         sector_t s = r10_bio->sectors;
1650                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1651                             test_bit(R10BIO_WriteError, &r10_bio->state))
1652                                 reschedule_retry(r10_bio);
1653                         else
1654                                 put_buf(r10_bio);
1655                         md_done_sync(mddev, s, 1);
1656                         break;
1657                 } else {
1658                         struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1659                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1660                             test_bit(R10BIO_WriteError, &r10_bio->state))
1661                                 reschedule_retry(r10_bio);
1662                         else
1663                                 put_buf(r10_bio);
1664                         r10_bio = r10_bio2;
1665                 }
1666         }
1667 }
1668
1669 static void end_sync_write(struct bio *bio, int error)
1670 {
1671         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1672         struct r10bio *r10_bio = bio->bi_private;
1673         struct mddev *mddev = r10_bio->mddev;
1674         struct r10conf *conf = mddev->private;
1675         int d;
1676         sector_t first_bad;
1677         int bad_sectors;
1678         int slot;
1679         int repl;
1680         struct md_rdev *rdev = NULL;
1681
1682         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1683         if (repl)
1684                 rdev = conf->mirrors[d].replacement;
1685         if (!rdev) {
1686                 smp_mb();
1687                 rdev = conf->mirrors[d].rdev;
1688         }
1689
1690         if (!uptodate) {
1691                 if (repl)
1692                         md_error(mddev, rdev);
1693                 else {
1694                         set_bit(WriteErrorSeen, &rdev->flags);
1695                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
1696                                 set_bit(MD_RECOVERY_NEEDED,
1697                                         &rdev->mddev->recovery);
1698                         set_bit(R10BIO_WriteError, &r10_bio->state);
1699                 }
1700         } else if (is_badblock(rdev,
1701                              r10_bio->devs[slot].addr,
1702                              r10_bio->sectors,
1703                              &first_bad, &bad_sectors))
1704                 set_bit(R10BIO_MadeGood, &r10_bio->state);
1705
1706         rdev_dec_pending(rdev, mddev);
1707
1708         end_sync_request(r10_bio);
1709 }
1710
1711 /*
1712  * Note: sync and recover and handled very differently for raid10
1713  * This code is for resync.
1714  * For resync, we read through virtual addresses and read all blocks.
1715  * If there is any error, we schedule a write.  The lowest numbered
1716  * drive is authoritative.
1717  * However requests come for physical address, so we need to map.
1718  * For every physical address there are raid_disks/copies virtual addresses,
1719  * which is always are least one, but is not necessarly an integer.
1720  * This means that a physical address can span multiple chunks, so we may
1721  * have to submit multiple io requests for a single sync request.
1722  */
1723 /*
1724  * We check if all blocks are in-sync and only write to blocks that
1725  * aren't in sync
1726  */
1727 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1728 {
1729         struct r10conf *conf = mddev->private;
1730         int i, first;
1731         struct bio *tbio, *fbio;
1732
1733         atomic_set(&r10_bio->remaining, 1);
1734
1735         /* find the first device with a block */
1736         for (i=0; i<conf->copies; i++)
1737                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1738                         break;
1739
1740         if (i == conf->copies)
1741                 goto done;
1742
1743         first = i;
1744         fbio = r10_bio->devs[i].bio;
1745
1746         /* now find blocks with errors */
1747         for (i=0 ; i < conf->copies ; i++) {
1748                 int  j, d;
1749                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1750
1751                 tbio = r10_bio->devs[i].bio;
1752
1753                 if (tbio->bi_end_io != end_sync_read)
1754                         continue;
1755                 if (i == first)
1756                         continue;
1757                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1758                         /* We know that the bi_io_vec layout is the same for
1759                          * both 'first' and 'i', so we just compare them.
1760                          * All vec entries are PAGE_SIZE;
1761                          */
1762                         for (j = 0; j < vcnt; j++)
1763                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1764                                            page_address(tbio->bi_io_vec[j].bv_page),
1765                                            PAGE_SIZE))
1766                                         break;
1767                         if (j == vcnt)
1768                                 continue;
1769                         mddev->resync_mismatches += r10_bio->sectors;
1770                         if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1771                                 /* Don't fix anything. */
1772                                 continue;
1773                 }
1774                 /* Ok, we need to write this bio, either to correct an
1775                  * inconsistency or to correct an unreadable block.
1776                  * First we need to fixup bv_offset, bv_len and
1777                  * bi_vecs, as the read request might have corrupted these
1778                  */
1779                 tbio->bi_vcnt = vcnt;
1780                 tbio->bi_size = r10_bio->sectors << 9;
1781                 tbio->bi_idx = 0;
1782                 tbio->bi_phys_segments = 0;
1783                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1784                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1785                 tbio->bi_next = NULL;
1786                 tbio->bi_rw = WRITE;
1787                 tbio->bi_private = r10_bio;
1788                 tbio->bi_sector = r10_bio->devs[i].addr;
1789
1790                 for (j=0; j < vcnt ; j++) {
1791                         tbio->bi_io_vec[j].bv_offset = 0;
1792                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1793
1794                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1795                                page_address(fbio->bi_io_vec[j].bv_page),
1796                                PAGE_SIZE);
1797                 }
1798                 tbio->bi_end_io = end_sync_write;
1799
1800                 d = r10_bio->devs[i].devnum;
1801                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1802                 atomic_inc(&r10_bio->remaining);
1803                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1804
1805                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1806                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1807                 generic_make_request(tbio);
1808         }
1809
1810         /* Now write out to any replacement devices
1811          * that are active
1812          */
1813         for (i = 0; i < conf->copies; i++) {
1814                 int j, d;
1815                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1816
1817                 tbio = r10_bio->devs[i].repl_bio;
1818                 if (!tbio || !tbio->bi_end_io)
1819                         continue;
1820                 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
1821                     && r10_bio->devs[i].bio != fbio)
1822                         for (j = 0; j < vcnt; j++)
1823                                 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1824                                        page_address(fbio->bi_io_vec[j].bv_page),
1825                                        PAGE_SIZE);
1826                 d = r10_bio->devs[i].devnum;
1827                 atomic_inc(&r10_bio->remaining);
1828                 md_sync_acct(conf->mirrors[d].replacement->bdev,
1829                              tbio->bi_size >> 9);
1830                 generic_make_request(tbio);
1831         }
1832
1833 done:
1834         if (atomic_dec_and_test(&r10_bio->remaining)) {
1835                 md_done_sync(mddev, r10_bio->sectors, 1);
1836                 put_buf(r10_bio);
1837         }
1838 }
1839
1840 /*
1841  * Now for the recovery code.
1842  * Recovery happens across physical sectors.
1843  * We recover all non-is_sync drives by finding the virtual address of
1844  * each, and then choose a working drive that also has that virt address.
1845  * There is a separate r10_bio for each non-in_sync drive.
1846  * Only the first two slots are in use. The first for reading,
1847  * The second for writing.
1848  *
1849  */
1850 static void fix_recovery_read_error(struct r10bio *r10_bio)
1851 {
1852         /* We got a read error during recovery.
1853          * We repeat the read in smaller page-sized sections.
1854          * If a read succeeds, write it to the new device or record
1855          * a bad block if we cannot.
1856          * If a read fails, record a bad block on both old and
1857          * new devices.
1858          */
1859         struct mddev *mddev = r10_bio->mddev;
1860         struct r10conf *conf = mddev->private;
1861         struct bio *bio = r10_bio->devs[0].bio;
1862         sector_t sect = 0;
1863         int sectors = r10_bio->sectors;
1864         int idx = 0;
1865         int dr = r10_bio->devs[0].devnum;
1866         int dw = r10_bio->devs[1].devnum;
1867
1868         while (sectors) {
1869                 int s = sectors;
1870                 struct md_rdev *rdev;
1871                 sector_t addr;
1872                 int ok;
1873
1874                 if (s > (PAGE_SIZE>>9))
1875                         s = PAGE_SIZE >> 9;
1876
1877                 rdev = conf->mirrors[dr].rdev;
1878                 addr = r10_bio->devs[0].addr + sect,
1879                 ok = sync_page_io(rdev,
1880                                   addr,
1881                                   s << 9,
1882                                   bio->bi_io_vec[idx].bv_page,
1883                                   READ, false);
1884                 if (ok) {
1885                         rdev = conf->mirrors[dw].rdev;
1886                         addr = r10_bio->devs[1].addr + sect;
1887                         ok = sync_page_io(rdev,
1888                                           addr,
1889                                           s << 9,
1890                                           bio->bi_io_vec[idx].bv_page,
1891                                           WRITE, false);
1892                         if (!ok) {
1893                                 set_bit(WriteErrorSeen, &rdev->flags);
1894                                 if (!test_and_set_bit(WantReplacement,
1895                                                       &rdev->flags))
1896                                         set_bit(MD_RECOVERY_NEEDED,
1897                                                 &rdev->mddev->recovery);
1898                         }
1899                 }
1900                 if (!ok) {
1901                         /* We don't worry if we cannot set a bad block -
1902                          * it really is bad so there is no loss in not
1903                          * recording it yet
1904                          */
1905                         rdev_set_badblocks(rdev, addr, s, 0);
1906
1907                         if (rdev != conf->mirrors[dw].rdev) {
1908                                 /* need bad block on destination too */
1909                                 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
1910                                 addr = r10_bio->devs[1].addr + sect;
1911                                 ok = rdev_set_badblocks(rdev2, addr, s, 0);
1912                                 if (!ok) {
1913                                         /* just abort the recovery */
1914                                         printk(KERN_NOTICE
1915                                                "md/raid10:%s: recovery aborted"
1916                                                " due to read error\n",
1917                                                mdname(mddev));
1918
1919                                         conf->mirrors[dw].recovery_disabled
1920                                                 = mddev->recovery_disabled;
1921                                         set_bit(MD_RECOVERY_INTR,
1922                                                 &mddev->recovery);
1923                                         break;
1924                                 }
1925                         }
1926                 }
1927
1928                 sectors -= s;
1929                 sect += s;
1930                 idx++;
1931         }
1932 }
1933
1934 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1935 {
1936         struct r10conf *conf = mddev->private;
1937         int d;
1938         struct bio *wbio, *wbio2;
1939
1940         if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
1941                 fix_recovery_read_error(r10_bio);
1942                 end_sync_request(r10_bio);
1943                 return;
1944         }
1945
1946         /*
1947          * share the pages with the first bio
1948          * and submit the write request
1949          */
1950         d = r10_bio->devs[1].devnum;
1951         wbio = r10_bio->devs[1].bio;
1952         wbio2 = r10_bio->devs[1].repl_bio;
1953         if (wbio->bi_end_io) {
1954                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1955                 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1956                 generic_make_request(wbio);
1957         }
1958         if (wbio2 && wbio2->bi_end_io) {
1959                 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
1960                 md_sync_acct(conf->mirrors[d].replacement->bdev,
1961                              wbio2->bi_size >> 9);
1962                 generic_make_request(wbio2);
1963         }
1964 }
1965
1966
1967 /*
1968  * Used by fix_read_error() to decay the per rdev read_errors.
1969  * We halve the read error count for every hour that has elapsed
1970  * since the last recorded read error.
1971  *
1972  */
1973 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
1974 {
1975         struct timespec cur_time_mon;
1976         unsigned long hours_since_last;
1977         unsigned int read_errors = atomic_read(&rdev->read_errors);
1978
1979         ktime_get_ts(&cur_time_mon);
1980
1981         if (rdev->last_read_error.tv_sec == 0 &&
1982             rdev->last_read_error.tv_nsec == 0) {
1983                 /* first time we've seen a read error */
1984                 rdev->last_read_error = cur_time_mon;
1985                 return;
1986         }
1987
1988         hours_since_last = (cur_time_mon.tv_sec -
1989                             rdev->last_read_error.tv_sec) / 3600;
1990
1991         rdev->last_read_error = cur_time_mon;
1992
1993         /*
1994          * if hours_since_last is > the number of bits in read_errors
1995          * just set read errors to 0. We do this to avoid
1996          * overflowing the shift of read_errors by hours_since_last.
1997          */
1998         if (hours_since_last >= 8 * sizeof(read_errors))
1999                 atomic_set(&rdev->read_errors, 0);
2000         else
2001                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2002 }
2003
2004 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2005                             int sectors, struct page *page, int rw)
2006 {
2007         sector_t first_bad;
2008         int bad_sectors;
2009
2010         if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2011             && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2012                 return -1;
2013         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2014                 /* success */
2015                 return 1;
2016         if (rw == WRITE) {
2017                 set_bit(WriteErrorSeen, &rdev->flags);
2018                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2019                         set_bit(MD_RECOVERY_NEEDED,
2020                                 &rdev->mddev->recovery);
2021         }
2022         /* need to record an error - either for the block or the device */
2023         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2024                 md_error(rdev->mddev, rdev);
2025         return 0;
2026 }
2027
2028 /*
2029  * This is a kernel thread which:
2030  *
2031  *      1.      Retries failed read operations on working mirrors.
2032  *      2.      Updates the raid superblock when problems encounter.
2033  *      3.      Performs writes following reads for array synchronising.
2034  */
2035
2036 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2037 {
2038         int sect = 0; /* Offset from r10_bio->sector */
2039         int sectors = r10_bio->sectors;
2040         struct md_rdev*rdev;
2041         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2042         int d = r10_bio->devs[r10_bio->read_slot].devnum;
2043
2044         /* still own a reference to this rdev, so it cannot
2045          * have been cleared recently.
2046          */
2047         rdev = conf->mirrors[d].rdev;
2048
2049         if (test_bit(Faulty, &rdev->flags))
2050                 /* drive has already been failed, just ignore any
2051                    more fix_read_error() attempts */
2052                 return;
2053
2054         check_decay_read_errors(mddev, rdev);
2055         atomic_inc(&rdev->read_errors);
2056         if (atomic_read(&rdev->read_errors) > max_read_errors) {
2057                 char b[BDEVNAME_SIZE];
2058                 bdevname(rdev->bdev, b);
2059
2060                 printk(KERN_NOTICE
2061                        "md/raid10:%s: %s: Raid device exceeded "
2062                        "read_error threshold [cur %d:max %d]\n",
2063                        mdname(mddev), b,
2064                        atomic_read(&rdev->read_errors), max_read_errors);
2065                 printk(KERN_NOTICE
2066                        "md/raid10:%s: %s: Failing raid device\n",
2067                        mdname(mddev), b);
2068                 md_error(mddev, conf->mirrors[d].rdev);
2069                 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2070                 return;
2071         }
2072
2073         while(sectors) {
2074                 int s = sectors;
2075                 int sl = r10_bio->read_slot;
2076                 int success = 0;
2077                 int start;
2078
2079                 if (s > (PAGE_SIZE>>9))
2080                         s = PAGE_SIZE >> 9;
2081
2082                 rcu_read_lock();
2083                 do {
2084                         sector_t first_bad;
2085                         int bad_sectors;
2086
2087                         d = r10_bio->devs[sl].devnum;
2088                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2089                         if (rdev &&
2090                             test_bit(In_sync, &rdev->flags) &&
2091                             is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2092                                         &first_bad, &bad_sectors) == 0) {
2093                                 atomic_inc(&rdev->nr_pending);
2094                                 rcu_read_unlock();
2095                                 success = sync_page_io(rdev,
2096                                                        r10_bio->devs[sl].addr +
2097                                                        sect,
2098                                                        s<<9,
2099                                                        conf->tmppage, READ, false);
2100                                 rdev_dec_pending(rdev, mddev);
2101                                 rcu_read_lock();
2102                                 if (success)
2103                                         break;
2104                         }
2105                         sl++;
2106                         if (sl == conf->copies)
2107                                 sl = 0;
2108                 } while (!success && sl != r10_bio->read_slot);
2109                 rcu_read_unlock();
2110
2111                 if (!success) {
2112                         /* Cannot read from anywhere, just mark the block
2113                          * as bad on the first device to discourage future
2114                          * reads.
2115                          */
2116                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2117                         rdev = conf->mirrors[dn].rdev;
2118
2119                         if (!rdev_set_badblocks(
2120                                     rdev,
2121                                     r10_bio->devs[r10_bio->read_slot].addr
2122                                     + sect,
2123                                     s, 0)) {
2124                                 md_error(mddev, rdev);
2125                                 r10_bio->devs[r10_bio->read_slot].bio
2126                                         = IO_BLOCKED;
2127                         }
2128                         break;
2129                 }
2130
2131                 start = sl;
2132                 /* write it back and re-read */
2133                 rcu_read_lock();
2134                 while (sl != r10_bio->read_slot) {
2135                         char b[BDEVNAME_SIZE];
2136
2137                         if (sl==0)
2138                                 sl = conf->copies;
2139                         sl--;
2140                         d = r10_bio->devs[sl].devnum;
2141                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2142                         if (!rdev ||
2143                             !test_bit(In_sync, &rdev->flags))
2144                                 continue;
2145
2146                         atomic_inc(&rdev->nr_pending);
2147                         rcu_read_unlock();
2148                         if (r10_sync_page_io(rdev,
2149                                              r10_bio->devs[sl].addr +
2150                                              sect,
2151                                              s<<9, conf->tmppage, WRITE)
2152                             == 0) {
2153                                 /* Well, this device is dead */
2154                                 printk(KERN_NOTICE
2155                                        "md/raid10:%s: read correction "
2156                                        "write failed"
2157                                        " (%d sectors at %llu on %s)\n",
2158                                        mdname(mddev), s,
2159                                        (unsigned long long)(
2160                                                sect + rdev->data_offset),
2161                                        bdevname(rdev->bdev, b));
2162                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2163                                        "drive\n",
2164                                        mdname(mddev),
2165                                        bdevname(rdev->bdev, b));
2166                         }
2167                         rdev_dec_pending(rdev, mddev);
2168                         rcu_read_lock();
2169                 }
2170                 sl = start;
2171                 while (sl != r10_bio->read_slot) {
2172                         char b[BDEVNAME_SIZE];
2173
2174                         if (sl==0)
2175                                 sl = conf->copies;
2176                         sl--;
2177                         d = r10_bio->devs[sl].devnum;
2178                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2179                         if (!rdev ||
2180                             !test_bit(In_sync, &rdev->flags))
2181                                 continue;
2182
2183                         atomic_inc(&rdev->nr_pending);
2184                         rcu_read_unlock();
2185                         switch (r10_sync_page_io(rdev,
2186                                              r10_bio->devs[sl].addr +
2187                                              sect,
2188                                              s<<9, conf->tmppage,
2189                                                  READ)) {
2190                         case 0:
2191                                 /* Well, this device is dead */
2192                                 printk(KERN_NOTICE
2193                                        "md/raid10:%s: unable to read back "
2194                                        "corrected sectors"
2195                                        " (%d sectors at %llu on %s)\n",
2196                                        mdname(mddev), s,
2197                                        (unsigned long long)(
2198                                                sect + rdev->data_offset),
2199                                        bdevname(rdev->bdev, b));
2200                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2201                                        "drive\n",
2202                                        mdname(mddev),
2203                                        bdevname(rdev->bdev, b));
2204                                 break;
2205                         case 1:
2206                                 printk(KERN_INFO
2207                                        "md/raid10:%s: read error corrected"
2208                                        " (%d sectors at %llu on %s)\n",
2209                                        mdname(mddev), s,
2210                                        (unsigned long long)(
2211                                                sect + rdev->data_offset),
2212                                        bdevname(rdev->bdev, b));
2213                                 atomic_add(s, &rdev->corrected_errors);
2214                         }
2215
2216                         rdev_dec_pending(rdev, mddev);
2217                         rcu_read_lock();
2218                 }
2219                 rcu_read_unlock();
2220
2221                 sectors -= s;
2222                 sect += s;
2223         }
2224 }
2225
2226 static void bi_complete(struct bio *bio, int error)
2227 {
2228         complete((struct completion *)bio->bi_private);
2229 }
2230
2231 static int submit_bio_wait(int rw, struct bio *bio)
2232 {
2233         struct completion event;
2234         rw |= REQ_SYNC;
2235
2236         init_completion(&event);
2237         bio->bi_private = &event;
2238         bio->bi_end_io = bi_complete;
2239         submit_bio(rw, bio);
2240         wait_for_completion(&event);
2241
2242         return test_bit(BIO_UPTODATE, &bio->bi_flags);
2243 }
2244
2245 static int narrow_write_error(struct r10bio *r10_bio, int i)
2246 {
2247         struct bio *bio = r10_bio->master_bio;
2248         struct mddev *mddev = r10_bio->mddev;
2249         struct r10conf *conf = mddev->private;
2250         struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2251         /* bio has the data to be written to slot 'i' where
2252          * we just recently had a write error.
2253          * We repeatedly clone the bio and trim down to one block,
2254          * then try the write.  Where the write fails we record
2255          * a bad block.
2256          * It is conceivable that the bio doesn't exactly align with
2257          * blocks.  We must handle this.
2258          *
2259          * We currently own a reference to the rdev.
2260          */
2261
2262         int block_sectors;
2263         sector_t sector;
2264         int sectors;
2265         int sect_to_write = r10_bio->sectors;
2266         int ok = 1;
2267
2268         if (rdev->badblocks.shift < 0)
2269                 return 0;
2270
2271         block_sectors = 1 << rdev->badblocks.shift;
2272         sector = r10_bio->sector;
2273         sectors = ((r10_bio->sector + block_sectors)
2274                    & ~(sector_t)(block_sectors - 1))
2275                 - sector;
2276
2277         while (sect_to_write) {
2278                 struct bio *wbio;
2279                 if (sectors > sect_to_write)
2280                         sectors = sect_to_write;
2281                 /* Write at 'sector' for 'sectors' */
2282                 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2283                 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2284                 wbio->bi_sector = (r10_bio->devs[i].addr+
2285                                    rdev->data_offset+
2286                                    (sector - r10_bio->sector));
2287                 wbio->bi_bdev = rdev->bdev;
2288                 if (submit_bio_wait(WRITE, wbio) == 0)
2289                         /* Failure! */
2290                         ok = rdev_set_badblocks(rdev, sector,
2291                                                 sectors, 0)
2292                                 && ok;
2293
2294                 bio_put(wbio);
2295                 sect_to_write -= sectors;
2296                 sector += sectors;
2297                 sectors = block_sectors;
2298         }
2299         return ok;
2300 }
2301
2302 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2303 {
2304         int slot = r10_bio->read_slot;
2305         struct bio *bio;
2306         struct r10conf *conf = mddev->private;
2307         struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2308         char b[BDEVNAME_SIZE];
2309         unsigned long do_sync;
2310         int max_sectors;
2311
2312         /* we got a read error. Maybe the drive is bad.  Maybe just
2313          * the block and we can fix it.
2314          * We freeze all other IO, and try reading the block from
2315          * other devices.  When we find one, we re-write
2316          * and check it that fixes the read error.
2317          * This is all done synchronously while the array is
2318          * frozen.
2319          */
2320         bio = r10_bio->devs[slot].bio;
2321         bdevname(bio->bi_bdev, b);
2322         bio_put(bio);
2323         r10_bio->devs[slot].bio = NULL;
2324
2325         if (mddev->ro == 0) {
2326                 freeze_array(conf);
2327                 fix_read_error(conf, mddev, r10_bio);
2328                 unfreeze_array(conf);
2329         } else
2330                 r10_bio->devs[slot].bio = IO_BLOCKED;
2331
2332         rdev_dec_pending(rdev, mddev);
2333
2334 read_more:
2335         rdev = read_balance(conf, r10_bio, &max_sectors);
2336         if (rdev == NULL) {
2337                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2338                        " read error for block %llu\n",
2339                        mdname(mddev), b,
2340                        (unsigned long long)r10_bio->sector);
2341                 raid_end_bio_io(r10_bio);
2342                 return;
2343         }
2344
2345         do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2346         slot = r10_bio->read_slot;
2347         printk_ratelimited(
2348                 KERN_ERR
2349                 "md/raid10:%s: %s: redirecting"
2350                 "sector %llu to another mirror\n",
2351                 mdname(mddev),
2352                 bdevname(rdev->bdev, b),
2353                 (unsigned long long)r10_bio->sector);
2354         bio = bio_clone_mddev(r10_bio->master_bio,
2355                               GFP_NOIO, mddev);
2356         md_trim_bio(bio,
2357                     r10_bio->sector - bio->bi_sector,
2358                     max_sectors);
2359         r10_bio->devs[slot].bio = bio;
2360         r10_bio->devs[slot].rdev = rdev;
2361         bio->bi_sector = r10_bio->devs[slot].addr
2362                 + rdev->data_offset;
2363         bio->bi_bdev = rdev->bdev;
2364         bio->bi_rw = READ | do_sync;
2365         bio->bi_private = r10_bio;
2366         bio->bi_end_io = raid10_end_read_request;
2367         if (max_sectors < r10_bio->sectors) {
2368                 /* Drat - have to split this up more */
2369                 struct bio *mbio = r10_bio->master_bio;
2370                 int sectors_handled =
2371                         r10_bio->sector + max_sectors
2372                         - mbio->bi_sector;
2373                 r10_bio->sectors = max_sectors;
2374                 spin_lock_irq(&conf->device_lock);
2375                 if (mbio->bi_phys_segments == 0)
2376                         mbio->bi_phys_segments = 2;
2377                 else
2378                         mbio->bi_phys_segments++;
2379                 spin_unlock_irq(&conf->device_lock);
2380                 generic_make_request(bio);
2381
2382                 r10_bio = mempool_alloc(conf->r10bio_pool,
2383                                         GFP_NOIO);
2384                 r10_bio->master_bio = mbio;
2385                 r10_bio->sectors = (mbio->bi_size >> 9)
2386                         - sectors_handled;
2387                 r10_bio->state = 0;
2388                 set_bit(R10BIO_ReadError,
2389                         &r10_bio->state);
2390                 r10_bio->mddev = mddev;
2391                 r10_bio->sector = mbio->bi_sector
2392                         + sectors_handled;
2393
2394                 goto read_more;
2395         } else
2396                 generic_make_request(bio);
2397 }
2398
2399 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2400 {
2401         /* Some sort of write request has finished and it
2402          * succeeded in writing where we thought there was a
2403          * bad block.  So forget the bad block.
2404          * Or possibly if failed and we need to record
2405          * a bad block.
2406          */
2407         int m;
2408         struct md_rdev *rdev;
2409
2410         if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2411             test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2412                 for (m = 0; m < conf->copies; m++) {
2413                         int dev = r10_bio->devs[m].devnum;
2414                         rdev = conf->mirrors[dev].rdev;
2415                         if (r10_bio->devs[m].bio == NULL)
2416                                 continue;
2417                         if (test_bit(BIO_UPTODATE,
2418                                      &r10_bio->devs[m].bio->bi_flags)) {
2419                                 rdev_clear_badblocks(
2420                                         rdev,
2421                                         r10_bio->devs[m].addr,
2422                                         r10_bio->sectors);
2423                         } else {
2424                                 if (!rdev_set_badblocks(
2425                                             rdev,
2426                                             r10_bio->devs[m].addr,
2427                                             r10_bio->sectors, 0))
2428                                         md_error(conf->mddev, rdev);
2429                         }
2430                         rdev = conf->mirrors[dev].replacement;
2431                         if (r10_bio->devs[m].repl_bio == NULL)
2432                                 continue;
2433                         if (test_bit(BIO_UPTODATE,
2434                                      &r10_bio->devs[m].repl_bio->bi_flags)) {
2435                                 rdev_clear_badblocks(
2436                                         rdev,
2437                                         r10_bio->devs[m].addr,
2438                                         r10_bio->sectors);
2439                         } else {
2440                                 if (!rdev_set_badblocks(
2441                                             rdev,
2442                                             r10_bio->devs[m].addr,
2443                                             r10_bio->sectors, 0))
2444                                         md_error(conf->mddev, rdev);
2445                         }
2446                 }
2447                 put_buf(r10_bio);
2448         } else {
2449                 for (m = 0; m < conf->copies; m++) {
2450                         int dev = r10_bio->devs[m].devnum;
2451                         struct bio *bio = r10_bio->devs[m].bio;
2452                         rdev = conf->mirrors[dev].rdev;
2453                         if (bio == IO_MADE_GOOD) {
2454                                 rdev_clear_badblocks(
2455                                         rdev,
2456                                         r10_bio->devs[m].addr,
2457                                         r10_bio->sectors);
2458                                 rdev_dec_pending(rdev, conf->mddev);
2459                         } else if (bio != NULL &&
2460                                    !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2461                                 if (!narrow_write_error(r10_bio, m)) {
2462                                         md_error(conf->mddev, rdev);
2463                                         set_bit(R10BIO_Degraded,
2464                                                 &r10_bio->state);
2465                                 }
2466                                 rdev_dec_pending(rdev, conf->mddev);
2467                         }
2468                         bio = r10_bio->devs[m].repl_bio;
2469                         rdev = conf->mirrors[dev].replacement;
2470                         if (rdev && bio == IO_MADE_GOOD) {
2471                                 rdev_clear_badblocks(
2472                                         rdev,
2473                                         r10_bio->devs[m].addr,
2474                                         r10_bio->sectors);
2475                                 rdev_dec_pending(rdev, conf->mddev);
2476                         }
2477                 }
2478                 if (test_bit(R10BIO_WriteError,
2479                              &r10_bio->state))
2480                         close_write(r10_bio);
2481                 raid_end_bio_io(r10_bio);
2482         }
2483 }
2484
2485 static void raid10d(struct mddev *mddev)
2486 {
2487         struct r10bio *r10_bio;
2488         unsigned long flags;
2489         struct r10conf *conf = mddev->private;
2490         struct list_head *head = &conf->retry_list;
2491         struct blk_plug plug;
2492
2493         md_check_recovery(mddev);
2494
2495         blk_start_plug(&plug);
2496         for (;;) {
2497
2498                 flush_pending_writes(conf);
2499
2500                 spin_lock_irqsave(&conf->device_lock, flags);
2501                 if (list_empty(head)) {
2502                         spin_unlock_irqrestore(&conf->device_lock, flags);
2503                         break;
2504                 }
2505                 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2506                 list_del(head->prev);
2507                 conf->nr_queued--;
2508                 spin_unlock_irqrestore(&conf->device_lock, flags);
2509
2510                 mddev = r10_bio->mddev;
2511                 conf = mddev->private;
2512                 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2513                     test_bit(R10BIO_WriteError, &r10_bio->state))
2514                         handle_write_completed(conf, r10_bio);
2515                 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2516                         sync_request_write(mddev, r10_bio);
2517                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2518                         recovery_request_write(mddev, r10_bio);
2519                 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2520                         handle_read_error(mddev, r10_bio);
2521                 else {
2522                         /* just a partial read to be scheduled from a
2523                          * separate context
2524                          */
2525                         int slot = r10_bio->read_slot;
2526                         generic_make_request(r10_bio->devs[slot].bio);
2527                 }
2528
2529                 cond_resched();
2530                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2531                         md_check_recovery(mddev);
2532         }
2533         blk_finish_plug(&plug);
2534 }
2535
2536
2537 static int init_resync(struct r10conf *conf)
2538 {
2539         int buffs;
2540         int i;
2541
2542         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2543         BUG_ON(conf->r10buf_pool);
2544         conf->have_replacement = 0;
2545         for (i = 0; i < conf->raid_disks; i++)
2546                 if (conf->mirrors[i].replacement)
2547                         conf->have_replacement = 1;
2548         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2549         if (!conf->r10buf_pool)
2550                 return -ENOMEM;
2551         conf->next_resync = 0;
2552         return 0;
2553 }
2554
2555 /*
2556  * perform a "sync" on one "block"
2557  *
2558  * We need to make sure that no normal I/O request - particularly write
2559  * requests - conflict with active sync requests.
2560  *
2561  * This is achieved by tracking pending requests and a 'barrier' concept
2562  * that can be installed to exclude normal IO requests.
2563  *
2564  * Resync and recovery are handled very differently.
2565  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2566  *
2567  * For resync, we iterate over virtual addresses, read all copies,
2568  * and update if there are differences.  If only one copy is live,
2569  * skip it.
2570  * For recovery, we iterate over physical addresses, read a good
2571  * value for each non-in_sync drive, and over-write.
2572  *
2573  * So, for recovery we may have several outstanding complex requests for a
2574  * given address, one for each out-of-sync device.  We model this by allocating
2575  * a number of r10_bio structures, one for each out-of-sync device.
2576  * As we setup these structures, we collect all bio's together into a list
2577  * which we then process collectively to add pages, and then process again
2578  * to pass to generic_make_request.
2579  *
2580  * The r10_bio structures are linked using a borrowed master_bio pointer.
2581  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2582  * has its remaining count decremented to 0, the whole complex operation
2583  * is complete.
2584  *
2585  */
2586
2587 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2588                              int *skipped, int go_faster)
2589 {
2590         struct r10conf *conf = mddev->private;
2591         struct r10bio *r10_bio;
2592         struct bio *biolist = NULL, *bio;
2593         sector_t max_sector, nr_sectors;
2594         int i;
2595         int max_sync;
2596         sector_t sync_blocks;
2597         sector_t sectors_skipped = 0;
2598         int chunks_skipped = 0;
2599
2600         if (!conf->r10buf_pool)
2601                 if (init_resync(conf))
2602                         return 0;
2603
2604  skipped:
2605         max_sector = mddev->dev_sectors;
2606         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2607                 max_sector = mddev->resync_max_sectors;
2608         if (sector_nr >= max_sector) {
2609                 /* If we aborted, we need to abort the
2610                  * sync on the 'current' bitmap chucks (there can
2611                  * be several when recovering multiple devices).
2612                  * as we may have started syncing it but not finished.
2613                  * We can find the current address in
2614                  * mddev->curr_resync, but for recovery,
2615                  * we need to convert that to several
2616                  * virtual addresses.
2617                  */
2618                 if (mddev->curr_resync < max_sector) { /* aborted */
2619                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2620                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2621                                                 &sync_blocks, 1);
2622                         else for (i=0; i<conf->raid_disks; i++) {
2623                                 sector_t sect =
2624                                         raid10_find_virt(conf, mddev->curr_resync, i);
2625                                 bitmap_end_sync(mddev->bitmap, sect,
2626                                                 &sync_blocks, 1);
2627                         }
2628                 } else {
2629                         /* completed sync */
2630                         if ((!mddev->bitmap || conf->fullsync)
2631                             && conf->have_replacement
2632                             && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2633                                 /* Completed a full sync so the replacements
2634                                  * are now fully recovered.
2635                                  */
2636                                 for (i = 0; i < conf->raid_disks; i++)
2637                                         if (conf->mirrors[i].replacement)
2638                                                 conf->mirrors[i].replacement
2639                                                         ->recovery_offset
2640                                                         = MaxSector;
2641                         }
2642                         conf->fullsync = 0;
2643                 }
2644                 bitmap_close_sync(mddev->bitmap);
2645                 close_sync(conf);
2646                 *skipped = 1;
2647                 return sectors_skipped;
2648         }
2649         if (chunks_skipped >= conf->raid_disks) {
2650                 /* if there has been nothing to do on any drive,
2651                  * then there is nothing to do at all..
2652                  */
2653                 *skipped = 1;
2654                 return (max_sector - sector_nr) + sectors_skipped;
2655         }
2656
2657         if (max_sector > mddev->resync_max)
2658                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2659
2660         /* make sure whole request will fit in a chunk - if chunks
2661          * are meaningful
2662          */
2663         if (conf->near_copies < conf->raid_disks &&
2664             max_sector > (sector_nr | conf->chunk_mask))
2665                 max_sector = (sector_nr | conf->chunk_mask) + 1;
2666         /*
2667          * If there is non-resync activity waiting for us then
2668          * put in a delay to throttle resync.
2669          */
2670         if (!go_faster && conf->nr_waiting)
2671                 msleep_interruptible(1000);
2672
2673         /* Again, very different code for resync and recovery.
2674          * Both must result in an r10bio with a list of bios that
2675          * have bi_end_io, bi_sector, bi_bdev set,
2676          * and bi_private set to the r10bio.
2677          * For recovery, we may actually create several r10bios
2678          * with 2 bios in each, that correspond to the bios in the main one.
2679          * In this case, the subordinate r10bios link back through a
2680          * borrowed master_bio pointer, and the counter in the master
2681          * includes a ref from each subordinate.
2682          */
2683         /* First, we decide what to do and set ->bi_end_io
2684          * To end_sync_read if we want to read, and
2685          * end_sync_write if we will want to write.
2686          */
2687
2688         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2689         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2690                 /* recovery... the complicated one */
2691                 int j;
2692                 r10_bio = NULL;
2693
2694                 for (i=0 ; i<conf->raid_disks; i++) {
2695                         int still_degraded;
2696                         struct r10bio *rb2;
2697                         sector_t sect;
2698                         int must_sync;
2699                         int any_working;
2700                         struct mirror_info *mirror = &conf->mirrors[i];
2701
2702                         if ((mirror->rdev == NULL ||
2703                              test_bit(In_sync, &mirror->rdev->flags))
2704                             &&
2705                             (mirror->replacement == NULL ||
2706                              test_bit(Faulty,
2707                                       &mirror->replacement->flags)))
2708                                 continue;
2709
2710                         still_degraded = 0;
2711                         /* want to reconstruct this device */
2712                         rb2 = r10_bio;
2713                         sect = raid10_find_virt(conf, sector_nr, i);
2714                         /* Unless we are doing a full sync, or a replacement
2715                          * we only need to recover the block if it is set in
2716                          * the bitmap
2717                          */
2718                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2719                                                       &sync_blocks, 1);
2720                         if (sync_blocks < max_sync)
2721                                 max_sync = sync_blocks;
2722                         if (!must_sync &&
2723                             mirror->replacement == NULL &&
2724                             !conf->fullsync) {
2725                                 /* yep, skip the sync_blocks here, but don't assume
2726                                  * that there will never be anything to do here
2727                                  */
2728                                 chunks_skipped = -1;
2729                                 continue;
2730                         }
2731
2732                         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2733                         raise_barrier(conf, rb2 != NULL);
2734                         atomic_set(&r10_bio->remaining, 0);
2735
2736                         r10_bio->master_bio = (struct bio*)rb2;
2737                         if (rb2)
2738                                 atomic_inc(&rb2->remaining);
2739                         r10_bio->mddev = mddev;
2740                         set_bit(R10BIO_IsRecover, &r10_bio->state);
2741                         r10_bio->sector = sect;
2742
2743                         raid10_find_phys(conf, r10_bio);
2744
2745                         /* Need to check if the array will still be
2746                          * degraded
2747                          */
2748                         for (j=0; j<conf->raid_disks; j++)
2749                                 if (conf->mirrors[j].rdev == NULL ||
2750                                     test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2751                                         still_degraded = 1;
2752                                         break;
2753                                 }
2754
2755                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2756                                                       &sync_blocks, still_degraded);
2757
2758                         any_working = 0;
2759                         for (j=0; j<conf->copies;j++) {
2760                                 int k;
2761                                 int d = r10_bio->devs[j].devnum;
2762                                 sector_t from_addr, to_addr;
2763                                 struct md_rdev *rdev;
2764                                 sector_t sector, first_bad;
2765                                 int bad_sectors;
2766                                 if (!conf->mirrors[d].rdev ||
2767                                     !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2768                                         continue;
2769                                 /* This is where we read from */
2770                                 any_working = 1;
2771                                 rdev = conf->mirrors[d].rdev;
2772                                 sector = r10_bio->devs[j].addr;
2773
2774                                 if (is_badblock(rdev, sector, max_sync,
2775                                                 &first_bad, &bad_sectors)) {
2776                                         if (first_bad > sector)
2777                                                 max_sync = first_bad - sector;
2778                                         else {
2779                                                 bad_sectors -= (sector
2780                                                                 - first_bad);
2781                                                 if (max_sync > bad_sectors)
2782                                                         max_sync = bad_sectors;
2783                                                 continue;
2784                                         }
2785                                 }
2786                                 bio = r10_bio->devs[0].bio;
2787                                 bio->bi_next = biolist;
2788                                 biolist = bio;
2789                                 bio->bi_private = r10_bio;
2790                                 bio->bi_end_io = end_sync_read;
2791                                 bio->bi_rw = READ;
2792                                 from_addr = r10_bio->devs[j].addr;
2793                                 bio->bi_sector = from_addr + rdev->data_offset;
2794                                 bio->bi_bdev = rdev->bdev;
2795                                 atomic_inc(&rdev->nr_pending);
2796                                 /* and we write to 'i' (if not in_sync) */
2797
2798                                 for (k=0; k<conf->copies; k++)
2799                                         if (r10_bio->devs[k].devnum == i)
2800                                                 break;
2801                                 BUG_ON(k == conf->copies);
2802                                 to_addr = r10_bio->devs[k].addr;
2803                                 r10_bio->devs[0].devnum = d;
2804                                 r10_bio->devs[0].addr = from_addr;
2805                                 r10_bio->devs[1].devnum = i;
2806                                 r10_bio->devs[1].addr = to_addr;
2807
2808                                 rdev = mirror->rdev;
2809                                 if (!test_bit(In_sync, &rdev->flags)) {
2810                                         bio = r10_bio->devs[1].bio;
2811                                         bio->bi_next = biolist;
2812                                         biolist = bio;
2813                                         bio->bi_private = r10_bio;
2814                                         bio->bi_end_io = end_sync_write;
2815                                         bio->bi_rw = WRITE;
2816                                         bio->bi_sector = to_addr
2817                                                 + rdev->data_offset;
2818                                         bio->bi_bdev = rdev->bdev;
2819                                         atomic_inc(&r10_bio->remaining);
2820                                 } else
2821                                         r10_bio->devs[1].bio->bi_end_io = NULL;
2822
2823                                 /* and maybe write to replacement */
2824                                 bio = r10_bio->devs[1].repl_bio;
2825                                 if (bio)
2826                                         bio->bi_end_io = NULL;
2827                                 rdev = mirror->replacement;
2828                                 /* Note: if rdev != NULL, then bio
2829                                  * cannot be NULL as r10buf_pool_alloc will
2830                                  * have allocated it.
2831                                  * So the second test here is pointless.
2832                                  * But it keeps semantic-checkers happy, and
2833                                  * this comment keeps human reviewers
2834                                  * happy.
2835                                  */
2836                                 if (rdev == NULL || bio == NULL ||
2837                                     test_bit(Faulty, &rdev->flags))
2838                                         break;
2839                                 bio->bi_next = biolist;
2840                                 biolist = bio;
2841                                 bio->bi_private = r10_bio;
2842                                 bio->bi_end_io = end_sync_write;
2843                                 bio->bi_rw = WRITE;
2844                                 bio->bi_sector = to_addr + rdev->data_offset;
2845                                 bio->bi_bdev = rdev->bdev;
2846                                 atomic_inc(&r10_bio->remaining);
2847                                 break;
2848                         }
2849                         if (j == conf->copies) {
2850                                 /* Cannot recover, so abort the recovery or
2851                                  * record a bad block */
2852                                 put_buf(r10_bio);
2853                                 if (rb2)
2854                                         atomic_dec(&rb2->remaining);
2855                                 r10_bio = rb2;
2856                                 if (any_working) {
2857                                         /* problem is that there are bad blocks
2858                                          * on other device(s)
2859                                          */
2860                                         int k;
2861                                         for (k = 0; k < conf->copies; k++)
2862                                                 if (r10_bio->devs[k].devnum == i)
2863                                                         break;
2864                                         if (!test_bit(In_sync,
2865                                                       &mirror->rdev->flags)
2866                                             && !rdev_set_badblocks(
2867                                                     mirror->rdev,
2868                                                     r10_bio->devs[k].addr,
2869                                                     max_sync, 0))
2870                                                 any_working = 0;
2871                                         if (mirror->replacement &&
2872                                             !rdev_set_badblocks(
2873                                                     mirror->replacement,
2874                                                     r10_bio->devs[k].addr,
2875                                                     max_sync, 0))
2876                                                 any_working = 0;
2877                                 }
2878                                 if (!any_working)  {
2879                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
2880                                                               &mddev->recovery))
2881                                                 printk(KERN_INFO "md/raid10:%s: insufficient "
2882                                                        "working devices for recovery.\n",
2883                                                        mdname(mddev));
2884                                         mirror->recovery_disabled
2885                                                 = mddev->recovery_disabled;
2886                                 }
2887                                 break;
2888                         }
2889                 }
2890                 if (biolist == NULL) {
2891                         while (r10_bio) {
2892                                 struct r10bio *rb2 = r10_bio;
2893                                 r10_bio = (struct r10bio*) rb2->master_bio;
2894                                 rb2->master_bio = NULL;
2895                                 put_buf(rb2);
2896                         }
2897                         goto giveup;
2898                 }
2899         } else {
2900                 /* resync. Schedule a read for every block at this virt offset */
2901                 int count = 0;
2902
2903                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2904
2905                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2906                                        &sync_blocks, mddev->degraded) &&
2907                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
2908                                                  &mddev->recovery)) {
2909                         /* We can skip this block */
2910                         *skipped = 1;
2911                         return sync_blocks + sectors_skipped;
2912                 }
2913                 if (sync_blocks < max_sync)
2914                         max_sync = sync_blocks;
2915                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2916
2917                 r10_bio->mddev = mddev;
2918                 atomic_set(&r10_bio->remaining, 0);
2919                 raise_barrier(conf, 0);
2920                 conf->next_resync = sector_nr;
2921
2922                 r10_bio->master_bio = NULL;
2923                 r10_bio->sector = sector_nr;
2924                 set_bit(R10BIO_IsSync, &r10_bio->state);
2925                 raid10_find_phys(conf, r10_bio);
2926                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2927
2928                 for (i=0; i<conf->copies; i++) {
2929                         int d = r10_bio->devs[i].devnum;
2930                         sector_t first_bad, sector;
2931                         int bad_sectors;
2932
2933                         if (r10_bio->devs[i].repl_bio)
2934                                 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
2935
2936                         bio = r10_bio->devs[i].bio;
2937                         bio->bi_end_io = NULL;
2938                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
2939                         if (conf->mirrors[d].rdev == NULL ||
2940                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2941                                 continue;
2942                         sector = r10_bio->devs[i].addr;
2943                         if (is_badblock(conf->mirrors[d].rdev,
2944                                         sector, max_sync,
2945                                         &first_bad, &bad_sectors)) {
2946                                 if (first_bad > sector)
2947                                         max_sync = first_bad - sector;
2948                                 else {
2949                                         bad_sectors -= (sector - first_bad);
2950                                         if (max_sync > bad_sectors)
2951                                                 max_sync = max_sync;
2952                                         continue;
2953                                 }
2954                         }
2955                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2956                         atomic_inc(&r10_bio->remaining);
2957                         bio->bi_next = biolist;
2958                         biolist = bio;
2959                         bio->bi_private = r10_bio;
2960                         bio->bi_end_io = end_sync_read;
2961                         bio->bi_rw = READ;
2962                         bio->bi_sector = sector +
2963                                 conf->mirrors[d].rdev->data_offset;
2964                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2965                         count++;
2966
2967                         if (conf->mirrors[d].replacement == NULL ||
2968                             test_bit(Faulty,
2969                                      &conf->mirrors[d].replacement->flags))
2970                                 continue;
2971
2972                         /* Need to set up for writing to the replacement */
2973                         bio = r10_bio->devs[i].repl_bio;
2974                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
2975
2976                         sector = r10_bio->devs[i].addr;
2977                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2978                         bio->bi_next = biolist;
2979                         biolist = bio;
2980                         bio->bi_private = r10_bio;
2981                         bio->bi_end_io = end_sync_write;
2982                         bio->bi_rw = WRITE;
2983                         bio->bi_sector = sector +
2984                                 conf->mirrors[d].replacement->data_offset;
2985                         bio->bi_bdev = conf->mirrors[d].replacement->bdev;
2986                         count++;
2987                 }
2988
2989                 if (count < 2) {
2990                         for (i=0; i<conf->copies; i++) {
2991                                 int d = r10_bio->devs[i].devnum;
2992                                 if (r10_bio->devs[i].bio->bi_end_io)
2993                                         rdev_dec_pending(conf->mirrors[d].rdev,
2994                                                          mddev);
2995                                 if (r10_bio->devs[i].repl_bio &&
2996                                     r10_bio->devs[i].repl_bio->bi_end_io)
2997                                         rdev_dec_pending(
2998                                                 conf->mirrors[d].replacement,
2999                                                 mddev);
3000                         }
3001                         put_buf(r10_bio);
3002                         biolist = NULL;
3003                         goto giveup;
3004                 }
3005         }
3006
3007         for (bio = biolist; bio ; bio=bio->bi_next) {
3008
3009                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
3010                 if (bio->bi_end_io)
3011                         bio->bi_flags |= 1 << BIO_UPTODATE;
3012                 bio->bi_vcnt = 0;
3013                 bio->bi_idx = 0;
3014                 bio->bi_phys_segments = 0;
3015                 bio->bi_size = 0;
3016         }
3017
3018         nr_sectors = 0;
3019         if (sector_nr + max_sync < max_sector)
3020                 max_sector = sector_nr + max_sync;
3021         do {
3022                 struct page *page;
3023                 int len = PAGE_SIZE;
3024                 if (sector_nr + (len>>9) > max_sector)
3025                         len = (max_sector - sector_nr) << 9;
3026                 if (len == 0)
3027                         break;
3028                 for (bio= biolist ; bio ; bio=bio->bi_next) {
3029                         struct bio *bio2;
3030                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3031                         if (bio_add_page(bio, page, len, 0))
3032                                 continue;
3033
3034                         /* stop here */
3035                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3036                         for (bio2 = biolist;
3037                              bio2 && bio2 != bio;
3038                              bio2 = bio2->bi_next) {
3039                                 /* remove last page from this bio */
3040                                 bio2->bi_vcnt--;
3041                                 bio2->bi_size -= len;
3042                                 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3043                         }
3044                         goto bio_full;
3045                 }
3046                 nr_sectors += len>>9;
3047                 sector_nr += len>>9;
3048         } while (biolist->bi_vcnt < RESYNC_PAGES);
3049  bio_full:
3050         r10_bio->sectors = nr_sectors;
3051
3052         while (biolist) {
3053                 bio = biolist;
3054                 biolist = biolist->bi_next;
3055
3056                 bio->bi_next = NULL;
3057                 r10_bio = bio->bi_private;
3058                 r10_bio->sectors = nr_sectors;
3059
3060                 if (bio->bi_end_io == end_sync_read) {
3061                         md_sync_acct(bio->bi_bdev, nr_sectors);
3062                         generic_make_request(bio);
3063                 }
3064         }
3065
3066         if (sectors_skipped)
3067                 /* pretend they weren't skipped, it makes
3068                  * no important difference in this case
3069                  */
3070                 md_done_sync(mddev, sectors_skipped, 1);
3071
3072         return sectors_skipped + nr_sectors;
3073  giveup:
3074         /* There is nowhere to write, so all non-sync
3075          * drives must be failed or in resync, all drives
3076          * have a bad block, so try the next chunk...
3077          */
3078         if (sector_nr + max_sync < max_sector)
3079                 max_sector = sector_nr + max_sync;
3080
3081         sectors_skipped += (max_sector - sector_nr);
3082         chunks_skipped ++;
3083         sector_nr = max_sector;
3084         goto skipped;
3085 }
3086
3087 static sector_t
3088 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3089 {
3090         sector_t size;
3091         struct r10conf *conf = mddev->private;
3092
3093         if (!raid_disks)
3094                 raid_disks = conf->raid_disks;
3095         if (!sectors)
3096                 sectors = conf->dev_sectors;
3097
3098         size = sectors >> conf->chunk_shift;
3099         sector_div(size, conf->far_copies);
3100         size = size * raid_disks;
3101         sector_div(size, conf->near_copies);
3102
3103         return size << conf->chunk_shift;
3104 }
3105
3106
3107 static struct r10conf *setup_conf(struct mddev *mddev)
3108 {
3109         struct r10conf *conf = NULL;
3110         int nc, fc, fo;
3111         sector_t stride, size;
3112         int err = -EINVAL;
3113
3114         if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
3115             !is_power_of_2(mddev->new_chunk_sectors)) {
3116                 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3117                        "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3118                        mdname(mddev), PAGE_SIZE);
3119                 goto out;
3120         }
3121
3122         nc = mddev->new_layout & 255;
3123         fc = (mddev->new_layout >> 8) & 255;
3124         fo = mddev->new_layout & (1<<16);
3125
3126         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
3127             (mddev->new_layout >> 17)) {
3128                 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3129                        mdname(mddev), mddev->new_layout);
3130                 goto out;
3131         }
3132
3133         err = -ENOMEM;
3134         conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3135         if (!conf)
3136                 goto out;
3137
3138         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
3139                                 GFP_KERNEL);
3140         if (!conf->mirrors)
3141                 goto out;
3142
3143         conf->tmppage = alloc_page(GFP_KERNEL);
3144         if (!conf->tmppage)
3145                 goto out;
3146
3147
3148         conf->raid_disks = mddev->raid_disks;
3149         conf->near_copies = nc;
3150         conf->far_copies = fc;
3151         conf->copies = nc*fc;
3152         conf->far_offset = fo;
3153         conf->chunk_mask = mddev->new_chunk_sectors - 1;
3154         conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
3155
3156         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3157                                            r10bio_pool_free, conf);
3158         if (!conf->r10bio_pool)
3159                 goto out;
3160
3161         size = mddev->dev_sectors >> conf->chunk_shift;
3162         sector_div(size, fc);
3163         size = size * conf->raid_disks;
3164         sector_div(size, nc);
3165         /* 'size' is now the number of chunks in the array */
3166         /* calculate "used chunks per device" in 'stride' */
3167         stride = size * conf->copies;
3168
3169         /* We need to round up when dividing by raid_disks to
3170          * get the stride size.
3171          */
3172         stride += conf->raid_disks - 1;
3173         sector_div(stride, conf->raid_disks);
3174
3175         conf->dev_sectors = stride << conf->chunk_shift;
3176
3177         if (fo)
3178                 stride = 1;
3179         else
3180                 sector_div(stride, fc);
3181         conf->stride = stride << conf->chunk_shift;
3182
3183
3184         spin_lock_init(&conf->device_lock);
3185         INIT_LIST_HEAD(&conf->retry_list);
3186
3187         spin_lock_init(&conf->resync_lock);
3188         init_waitqueue_head(&conf->wait_barrier);
3189
3190         conf->thread = md_register_thread(raid10d, mddev, NULL);
3191         if (!conf->thread)
3192                 goto out;
3193
3194         conf->mddev = mddev;
3195         return conf;
3196
3197  out:
3198         printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3199                mdname(mddev));
3200         if (conf) {
3201                 if (conf->r10bio_pool)
3202                         mempool_destroy(conf->r10bio_pool);
3203                 kfree(conf->mirrors);
3204                 safe_put_page(conf->tmppage);
3205                 kfree(conf);
3206         }
3207         return ERR_PTR(err);
3208 }
3209
3210 static int run(struct mddev *mddev)
3211 {
3212         struct r10conf *conf;
3213         int i, disk_idx, chunk_size;
3214         struct mirror_info *disk;
3215         struct md_rdev *rdev;
3216         sector_t size;
3217
3218         /*
3219          * copy the already verified devices into our private RAID10
3220          * bookkeeping area. [whatever we allocate in run(),
3221          * should be freed in stop()]
3222          */
3223
3224         if (mddev->private == NULL) {
3225                 conf = setup_conf(mddev);
3226                 if (IS_ERR(conf))
3227                         return PTR_ERR(conf);
3228                 mddev->private = conf;
3229         }
3230         conf = mddev->private;
3231         if (!conf)
3232                 goto out;
3233
3234         mddev->thread = conf->thread;
3235         conf->thread = NULL;
3236
3237         chunk_size = mddev->chunk_sectors << 9;
3238         blk_queue_io_min(mddev->queue, chunk_size);
3239         if (conf->raid_disks % conf->near_copies)
3240                 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
3241         else
3242                 blk_queue_io_opt(mddev->queue, chunk_size *
3243                                  (conf->raid_disks / conf->near_copies));
3244
3245         list_for_each_entry(rdev, &mddev->disks, same_set) {
3246
3247                 disk_idx = rdev->raid_disk;
3248                 if (disk_idx >= conf->raid_disks
3249                     || disk_idx < 0)
3250                         continue;
3251                 disk = conf->mirrors + disk_idx;
3252
3253                 if (test_bit(Replacement, &rdev->flags)) {
3254                         if (disk->replacement)
3255                                 goto out_free_conf;
3256                         disk->replacement = rdev;
3257                 } else {
3258                         if (disk->rdev)
3259                                 goto out_free_conf;
3260                         disk->rdev = rdev;
3261                 }
3262
3263                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3264                                   rdev->data_offset << 9);
3265                 /* as we don't honour merge_bvec_fn, we must never risk
3266                  * violating it, so limit max_segments to 1 lying
3267                  * within a single page.
3268                  */
3269                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
3270                         blk_queue_max_segments(mddev->queue, 1);
3271                         blk_queue_segment_boundary(mddev->queue,
3272                                                    PAGE_CACHE_SIZE - 1);
3273                 }
3274
3275                 disk->head_position = 0;
3276         }
3277         /* need to check that every block has at least one working mirror */
3278         if (!enough(conf, -1)) {
3279                 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3280                        mdname(mddev));
3281                 goto out_free_conf;
3282         }
3283
3284         mddev->degraded = 0;
3285         for (i = 0; i < conf->raid_disks; i++) {
3286
3287                 disk = conf->mirrors + i;
3288
3289                 if (!disk->rdev && disk->replacement) {
3290                         /* The replacement is all we have - use it */
3291                         disk->rdev = disk->replacement;
3292                         disk->replacement = NULL;
3293                         clear_bit(Replacement, &disk->rdev->flags);
3294                 }
3295
3296                 if (!disk->rdev ||
3297                     !test_bit(In_sync, &disk->rdev->flags)) {
3298                         disk->head_position = 0;
3299                         mddev->degraded++;
3300                         if (disk->rdev)
3301                                 conf->fullsync = 1;
3302                 }
3303                 disk->recovery_disabled = mddev->recovery_disabled - 1;
3304         }
3305
3306         if (mddev->recovery_cp != MaxSector)
3307                 printk(KERN_NOTICE "md/raid10:%s: not clean"
3308                        " -- starting background reconstruction\n",
3309                        mdname(mddev));
3310         printk(KERN_INFO
3311                 "md/raid10:%s: active with %d out of %d devices\n",
3312                 mdname(mddev), conf->raid_disks - mddev->degraded,
3313                 conf->raid_disks);
3314         /*
3315          * Ok, everything is just fine now
3316          */
3317         mddev->dev_sectors = conf->dev_sectors;
3318         size = raid10_size(mddev, 0, 0);
3319         md_set_array_sectors(mddev, size);
3320         mddev->resync_max_sectors = size;
3321
3322         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3323         mddev->queue->backing_dev_info.congested_data = mddev;
3324
3325         /* Calculate max read-ahead size.
3326          * We need to readahead at least twice a whole stripe....
3327          * maybe...
3328          */
3329         {
3330                 int stripe = conf->raid_disks *
3331                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3332                 stripe /= conf->near_copies;
3333                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
3334                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
3335         }
3336
3337         if (conf->near_copies < conf->raid_disks)
3338                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3339
3340         if (md_integrity_register(mddev))
3341                 goto out_free_conf;
3342
3343         return 0;
3344
3345 out_free_conf:
3346         md_unregister_thread(&mddev->thread);
3347         if (conf->r10bio_pool)
3348                 mempool_destroy(conf->r10bio_pool);
3349         safe_put_page(conf->tmppage);
3350         kfree(conf->mirrors);
3351         kfree(conf);
3352         mddev->private = NULL;
3353 out:
3354         return -EIO;
3355 }
3356
3357 static int stop(struct mddev *mddev)
3358 {
3359         struct r10conf *conf = mddev->private;
3360
3361         raise_barrier(conf, 0);
3362         lower_barrier(conf);
3363
3364         md_unregister_thread(&mddev->thread);
3365         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
3366         if (conf->r10bio_pool)
3367                 mempool_destroy(conf->r10bio_pool);
3368         kfree(conf->mirrors);
3369         kfree(conf);
3370         mddev->private = NULL;
3371         return 0;
3372 }
3373
3374 static void raid10_quiesce(struct mddev *mddev, int state)
3375 {
3376         struct r10conf *conf = mddev->private;
3377
3378         switch(state) {
3379         case 1:
3380                 raise_barrier(conf, 0);
3381                 break;
3382         case 0:
3383                 lower_barrier(conf);
3384                 break;
3385         }
3386 }
3387
3388 static void *raid10_takeover_raid0(struct mddev *mddev)
3389 {
3390         struct md_rdev *rdev;
3391         struct r10conf *conf;
3392
3393         if (mddev->degraded > 0) {
3394                 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3395                        mdname(mddev));
3396                 return ERR_PTR(-EINVAL);
3397         }
3398
3399         /* Set new parameters */
3400         mddev->new_level = 10;
3401         /* new layout: far_copies = 1, near_copies = 2 */
3402         mddev->new_layout = (1<<8) + 2;
3403         mddev->new_chunk_sectors = mddev->chunk_sectors;
3404         mddev->delta_disks = mddev->raid_disks;
3405         mddev->raid_disks *= 2;
3406         /* make sure it will be not marked as dirty */
3407         mddev->recovery_cp = MaxSector;
3408
3409         conf = setup_conf(mddev);
3410         if (!IS_ERR(conf)) {
3411                 list_for_each_entry(rdev, &mddev->disks, same_set)
3412                         if (rdev->raid_disk >= 0)
3413                                 rdev->new_raid_disk = rdev->raid_disk * 2;
3414                 conf->barrier = 1;
3415         }
3416
3417         return conf;
3418 }
3419
3420 static void *raid10_takeover(struct mddev *mddev)
3421 {
3422         struct r0conf *raid0_conf;
3423
3424         /* raid10 can take over:
3425          *  raid0 - providing it has only two drives
3426          */
3427         if (mddev->level == 0) {
3428                 /* for raid0 takeover only one zone is supported */
3429                 raid0_conf = mddev->private;
3430                 if (raid0_conf->nr_strip_zones > 1) {
3431                         printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3432                                " with more than one zone.\n",
3433                                mdname(mddev));
3434                         return ERR_PTR(-EINVAL);
3435                 }
3436                 return raid10_takeover_raid0(mddev);
3437         }
3438         return ERR_PTR(-EINVAL);
3439 }
3440
3441 static struct md_personality raid10_personality =
3442 {
3443         .name           = "raid10",
3444         .level          = 10,
3445         .owner          = THIS_MODULE,
3446         .make_request   = make_request,
3447         .run            = run,
3448         .stop           = stop,
3449         .status         = status,
3450         .error_handler  = error,
3451         .hot_add_disk   = raid10_add_disk,
3452         .hot_remove_disk= raid10_remove_disk,
3453         .spare_active   = raid10_spare_active,
3454         .sync_request   = sync_request,
3455         .quiesce        = raid10_quiesce,
3456         .size           = raid10_size,
3457         .takeover       = raid10_takeover,
3458 };
3459
3460 static int __init raid_init(void)
3461 {
3462         return register_md_personality(&raid10_personality);
3463 }
3464
3465 static void raid_exit(void)
3466 {
3467         unregister_md_personality(&raid10_personality);
3468 }
3469
3470 module_init(raid_init);
3471 module_exit(raid_exit);
3472 MODULE_LICENSE("GPL");
3473 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3474 MODULE_ALIAS("md-personality-9"); /* RAID10 */
3475 MODULE_ALIAS("md-raid10");
3476 MODULE_ALIAS("md-level-10");
3477
3478 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);