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