drivers/md/raid10.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * raid10.c : Multiple Devices driver for Linux
   4  *
   5  * Copyright (C) 2000-2004 Neil Brown
   6  *
   7  * RAID-10 support for md.
   8  *
   9  * Base on code in raid1.c.  See raid1.c for further copyright information.
  10  */
  11
  12 #include <linux/slab.h>
  13 #include <linux/delay.h>
  14 #include <linux/blkdev.h>
  15 #include <linux/module.h>
  16 #include <linux/seq_file.h>
  17 #include <linux/ratelimit.h>
  18 #include <linux/kthread.h>
  19 #include <linux/raid/md_p.h>
  20 #include <trace/events/block.h>
  21 #include "md.h"
  22 #include "raid10.h"
  23 #include "raid0.h"
  24 #include "md-bitmap.h"
  25
  26 /*
  27  * RAID10 provides a combination of RAID0 and RAID1 functionality.
  28  * The layout of data is defined by
  29  *    chunk_size
  30  *    raid_disks
  31  *    near_copies (stored in low byte of layout)
  32  *    far_copies (stored in second byte of layout)
  33  *    far_offset (stored in bit 16 of layout )
  34  *    use_far_sets (stored in bit 17 of layout )
  35  *    use_far_sets_bugfixed (stored in bit 18 of layout )
  36  *
  37  * The data to be stored is divided into chunks using chunksize.  Each device
  38  * is divided into far_copies sections.   In each section, chunks are laid out
  39  * in a style similar to raid0, but near_copies copies of each chunk is stored
  40  * (each on a different drive).  The starting device for each section is offset
  41  * near_copies from the starting device of the previous section.  Thus there
  42  * are (near_copies * far_copies) of each chunk, and each is on a different
  43  * drive.  near_copies and far_copies must be at least one, and their product
  44  * is at most raid_disks.
  45  *
  46  * If far_offset is true, then the far_copies are handled a bit differently.
  47  * The copies are still in different stripes, but instead of being very far
  48  * apart on disk, there are adjacent stripes.
  49  *
  50  * The far and offset algorithms are handled slightly differently if
  51  * 'use_far_sets' is true.  In this case, the array's devices are grouped into
  52  * sets that are (near_copies * far_copies) in size.  The far copied stripes
  53  * are still shifted by 'near_copies' devices, but this shifting stays confined
  54  * to the set rather than the entire array.  This is done to improve the number
  55  * of device combinations that can fail without causing the array to fail.
  56  * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
  57  * on a device):
  58  *    A B C D    A B C D E
  59  *      ...         ...
  60  *    D A B C    E A B C D
  61  * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
  62  *    [A B] [C D]    [A B] [C D E]
  63  *    |...| |...|    |...| | ... |
  64  *    [B A] [D C]    [B A] [E C D]
  65  */
  66
  67 /*
  68  * Number of guaranteed r10bios in case of extreme VM load:
  69  */
  70 #define NR_RAID10_BIOS 256
  71
  72 /* when we get a read error on a read-only array, we redirect to another
  73  * device without failing the first device, or trying to over-write to
  74  * correct the read error.  To keep track of bad blocks on a per-bio
  75  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
  76  */
  77 #define IO_BLOCKED ((struct bio *)1)
  78 /* When we successfully write to a known bad-block, we need to remove the
  79  * bad-block marking which must be done from process context.  So we record
  80  * the success by setting devs[n].bio to IO_MADE_GOOD
  81  */
  82 #define IO_MADE_GOOD ((struct bio *)2)
  83
  84 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
  85
  86 /* When there are this many requests queued to be written by
  87  * the raid10 thread, we become 'congested' to provide back-pressure
  88  * for writeback.
  89  */
  90 static int max_queued_requests = 1024;
  91
  92 static void allow_barrier(struct r10conf *conf);
  93 static void lower_barrier(struct r10conf *conf);
  94 static int _enough(struct r10conf *conf, int previous, int ignore);
  95 static int enough(struct r10conf *conf, int ignore);
  96 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
  97                                 int *skipped);
  98 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
  99 static void end_reshape_write(struct bio *bio);
 100 static void end_reshape(struct r10conf *conf);
 101
 102 #define raid10_log(md, fmt, args...)                            \
 103         do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
 104
 105 #include "raid1-10.c"
 106
 107 /*
 108  * for resync bio, r10bio pointer can be retrieved from the per-bio
 109  * 'struct resync_pages'.
 110  */
 111 static inline struct r10bio *get_resync_r10bio(struct bio *bio)
 112 {
 113         return get_resync_pages(bio)->raid_bio;
 114 }
 115
 116 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
 117 {
 118         struct r10conf *conf = data;
 119         int size = offsetof(struct r10bio, devs[conf->copies]);
 120
 121         /* allocate a r10bio with room for raid_disks entries in the
 122          * bios array */
 123         return kzalloc(size, gfp_flags);
 124 }
 125
 126 static void r10bio_pool_free(void *r10_bio, void *data)
 127 {
 128         kfree(r10_bio);
 129 }
 130
 131 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
 132 /* amount of memory to reserve for resync requests */
 133 #define RESYNC_WINDOW (1024*1024)
 134 /* maximum number of concurrent requests, memory permitting */
 135 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
 136 #define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
 137 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
 138
 139 /*
 140  * When performing a resync, we need to read and compare, so
 141  * we need as many pages are there are copies.
 142  * When performing a recovery, we need 2 bios, one for read,
 143  * one for write (we recover only one drive per r10buf)
 144  *
 145  */
 146 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
 147 {
 148         struct r10conf *conf = data;
 149         struct r10bio *r10_bio;
 150         struct bio *bio;
 151         int j;
 152         int nalloc, nalloc_rp;
 153         struct resync_pages *rps;
 154
 155         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 156         if (!r10_bio)
 157                 return NULL;
 158
 159         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
 160             test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
 161                 nalloc = conf->copies; /* resync */
 162         else
 163                 nalloc = 2; /* recovery */
 164
 165         /* allocate once for all bios */
 166         if (!conf->have_replacement)
 167                 nalloc_rp = nalloc;
 168         else
 169                 nalloc_rp = nalloc * 2;
 170         rps = kmalloc_array(nalloc_rp, sizeof(struct resync_pages), gfp_flags);
 171         if (!rps)
 172                 goto out_free_r10bio;
 173
 174         /*
 175          * Allocate bios.
 176          */
 177         for (j = nalloc ; j-- ; ) {
 178                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 179                 if (!bio)
 180                         goto out_free_bio;
 181                 r10_bio->devs[j].bio = bio;
 182                 if (!conf->have_replacement)
 183                         continue;
 184                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 185                 if (!bio)
 186                         goto out_free_bio;
 187                 r10_bio->devs[j].repl_bio = bio;
 188         }
 189         /*
 190          * Allocate RESYNC_PAGES data pages and attach them
 191          * where needed.
 192          */
 193         for (j = 0; j < nalloc; j++) {
 194                 struct bio *rbio = r10_bio->devs[j].repl_bio;
 195                 struct resync_pages *rp, *rp_repl;
 196
 197                 rp = &rps[j];
 198                 if (rbio)
 199                         rp_repl = &rps[nalloc + j];
 200
 201                 bio = r10_bio->devs[j].bio;
 202
 203                 if (!j || test_bit(MD_RECOVERY_SYNC,
 204                                    &conf->mddev->recovery)) {
 205                         if (resync_alloc_pages(rp, gfp_flags))
 206                                 goto out_free_pages;
 207                 } else {
 208                         memcpy(rp, &rps[0], sizeof(*rp));
 209                         resync_get_all_pages(rp);
 210                 }
 211
 212                 rp->raid_bio = r10_bio;
 213                 bio->bi_private = rp;
 214                 if (rbio) {
 215                         memcpy(rp_repl, rp, sizeof(*rp));
 216                         rbio->bi_private = rp_repl;
 217                 }
 218         }
 219
 220         return r10_bio;
 221
 222 out_free_pages:
 223         while (--j >= 0)
 224                 resync_free_pages(&rps[j * 2]);
 225
 226         j = 0;
 227 out_free_bio:
 228         for ( ; j < nalloc; j++) {
 229                 if (r10_bio->devs[j].bio)
 230                         bio_put(r10_bio->devs[j].bio);
 231                 if (r10_bio->devs[j].repl_bio)
 232                         bio_put(r10_bio->devs[j].repl_bio);
 233         }
 234         kfree(rps);
 235 out_free_r10bio:
 236         r10bio_pool_free(r10_bio, conf);
 237         return NULL;
 238 }
 239
 240 static void r10buf_pool_free(void *__r10_bio, void *data)
 241 {
 242         struct r10conf *conf = data;
 243         struct r10bio *r10bio = __r10_bio;
 244         int j;
 245         struct resync_pages *rp = NULL;
 246
 247         for (j = conf->copies; j--; ) {
 248                 struct bio *bio = r10bio->devs[j].bio;
 249
 250                 if (bio) {
 251                         rp = get_resync_pages(bio);
 252                         resync_free_pages(rp);
 253                         bio_put(bio);
 254                 }
 255
 256                 bio = r10bio->devs[j].repl_bio;
 257                 if (bio)
 258                         bio_put(bio);
 259         }
 260
 261         /* resync pages array stored in the 1st bio's .bi_private */
 262         kfree(rp);
 263
 264         r10bio_pool_free(r10bio, conf);
 265 }
 266
 267 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
 268 {
 269         int i;
 270
 271         for (i = 0; i < conf->copies; i++) {
 272                 struct bio **bio = & r10_bio->devs[i].bio;
 273                 if (!BIO_SPECIAL(*bio))
 274                         bio_put(*bio);
 275                 *bio = NULL;
 276                 bio = &r10_bio->devs[i].repl_bio;
 277                 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
 278                         bio_put(*bio);
 279                 *bio = NULL;
 280         }
 281 }
 282
 283 static void free_r10bio(struct r10bio *r10_bio)
 284 {
 285         struct r10conf *conf = r10_bio->mddev->private;
 286
 287         put_all_bios(conf, r10_bio);
 288         mempool_free(r10_bio, &conf->r10bio_pool);
 289 }
 290
 291 static void put_buf(struct r10bio *r10_bio)
 292 {
 293         struct r10conf *conf = r10_bio->mddev->private;
 294
 295         mempool_free(r10_bio, &conf->r10buf_pool);
 296
 297         lower_barrier(conf);
 298 }
 299
 300 static void reschedule_retry(struct r10bio *r10_bio)
 301 {
 302         unsigned long flags;
 303         struct mddev *mddev = r10_bio->mddev;
 304         struct r10conf *conf = mddev->private;
 305
 306         spin_lock_irqsave(&conf->device_lock, flags);
 307         list_add(&r10_bio->retry_list, &conf->retry_list);
 308         conf->nr_queued ++;
 309         spin_unlock_irqrestore(&conf->device_lock, flags);
 310
 311         /* wake up frozen array... */
 312         wake_up(&conf->wait_barrier);
 313
 314         md_wakeup_thread(mddev->thread);
 315 }
 316
 317 /*
 318  * raid_end_bio_io() is called when we have finished servicing a mirrored
 319  * operation and are ready to return a success/failure code to the buffer
 320  * cache layer.
 321  */
 322 static void raid_end_bio_io(struct r10bio *r10_bio)
 323 {
 324         struct bio *bio = r10_bio->master_bio;
 325         struct r10conf *conf = r10_bio->mddev->private;
 326
 327         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
 328                 bio->bi_status = BLK_STS_IOERR;
 329
 330         bio_endio(bio);
 331         /*
 332          * Wake up any possible resync thread that waits for the device
 333          * to go idle.
 334          */
 335         allow_barrier(conf);
 336
 337         free_r10bio(r10_bio);
 338 }
 339
 340 /*
 341  * Update disk head position estimator based on IRQ completion info.
 342  */
 343 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
 344 {
 345         struct r10conf *conf = r10_bio->mddev->private;
 346
 347         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 348                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 349 }
 350
 351 /*
 352  * Find the disk number which triggered given bio
 353  */
 354 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
 355                          struct bio *bio, int *slotp, int *replp)
 356 {
 357         int slot;
 358         int repl = 0;
 359
 360         for (slot = 0; slot < conf->copies; slot++) {
 361                 if (r10_bio->devs[slot].bio == bio)
 362                         break;
 363                 if (r10_bio->devs[slot].repl_bio == bio) {
 364                         repl = 1;
 365                         break;
 366                 }
 367         }
 368
 369         BUG_ON(slot == conf->copies);
 370         update_head_pos(slot, r10_bio);
 371
 372         if (slotp)
 373                 *slotp = slot;
 374         if (replp)
 375                 *replp = repl;
 376         return r10_bio->devs[slot].devnum;
 377 }
 378
 379 static void raid10_end_read_request(struct bio *bio)
 380 {
 381         int uptodate = !bio->bi_status;
 382         struct r10bio *r10_bio = bio->bi_private;
 383         int slot;
 384         struct md_rdev *rdev;
 385         struct r10conf *conf = r10_bio->mddev->private;
 386
 387         slot = r10_bio->read_slot;
 388         rdev = r10_bio->devs[slot].rdev;
 389         /*
 390          * this branch is our 'one mirror IO has finished' event handler:
 391          */
 392         update_head_pos(slot, r10_bio);
 393
 394         if (uptodate) {
 395                 /*
 396                  * Set R10BIO_Uptodate in our master bio, so that
 397                  * we will return a good error code to the higher
 398                  * levels even if IO on some other mirrored buffer fails.
 399                  *
 400                  * The 'master' represents the composite IO operation to
 401                  * user-side. So if something waits for IO, then it will
 402                  * wait for the 'master' bio.
 403                  */
 404                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 405         } else {
 406                 /* If all other devices that store this block have
 407                  * failed, we want to return the error upwards rather
 408                  * than fail the last device.  Here we redefine
 409                  * "uptodate" to mean "Don't want to retry"
 410                  */
 411                 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
 412                              rdev->raid_disk))
 413                         uptodate = 1;
 414         }
 415         if (uptodate) {
 416                 raid_end_bio_io(r10_bio);
 417                 rdev_dec_pending(rdev, conf->mddev);
 418         } else {
 419                 /*
 420                  * oops, read error - keep the refcount on the rdev
 421                  */
 422                 char b[BDEVNAME_SIZE];
 423                 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
 424                                    mdname(conf->mddev),
 425                                    bdevname(rdev->bdev, b),
 426                                    (unsigned long long)r10_bio->sector);
 427                 set_bit(R10BIO_ReadError, &r10_bio->state);
 428                 reschedule_retry(r10_bio);
 429         }
 430 }
 431
 432 static void close_write(struct r10bio *r10_bio)
 433 {
 434         /* clear the bitmap if all writes complete successfully */
 435         md_bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 436                            r10_bio->sectors,
 437                            !test_bit(R10BIO_Degraded, &r10_bio->state),
 438                            0);
 439         md_write_end(r10_bio->mddev);
 440 }
 441
 442 static void one_write_done(struct r10bio *r10_bio)
 443 {
 444         if (atomic_dec_and_test(&r10_bio->remaining)) {
 445                 if (test_bit(R10BIO_WriteError, &r10_bio->state))
 446                         reschedule_retry(r10_bio);
 447                 else {
 448                         close_write(r10_bio);
 449                         if (test_bit(R10BIO_MadeGood, &r10_bio->state))
 450                                 reschedule_retry(r10_bio);
 451                         else
 452                                 raid_end_bio_io(r10_bio);
 453                 }
 454         }
 455 }
 456
 457 static void raid10_end_write_request(struct bio *bio)
 458 {
 459         struct r10bio *r10_bio = bio->bi_private;
 460         int dev;
 461         int dec_rdev = 1;
 462         struct r10conf *conf = r10_bio->mddev->private;
 463         int slot, repl;
 464         struct md_rdev *rdev = NULL;
 465         struct bio *to_put = NULL;
 466         bool discard_error;
 467
 468         discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 469
 470         dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
 471
 472         if (repl)
 473                 rdev = conf->mirrors[dev].replacement;
 474         if (!rdev) {
 475                 smp_rmb();
 476                 repl = 0;
 477                 rdev = conf->mirrors[dev].rdev;
 478         }
 479         /*
 480          * this branch is our 'one mirror IO has finished' event handler:
 481          */
 482         if (bio->bi_status && !discard_error) {
 483                 if (repl)
 484                         /* Never record new bad blocks to replacement,
 485                          * just fail it.
 486                          */
 487                         md_error(rdev->mddev, rdev);
 488                 else {
 489                         set_bit(WriteErrorSeen, &rdev->flags);
 490                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
 491                                 set_bit(MD_RECOVERY_NEEDED,
 492                                         &rdev->mddev->recovery);
 493
 494                         dec_rdev = 0;
 495                         if (test_bit(FailFast, &rdev->flags) &&
 496                             (bio->bi_opf & MD_FAILFAST)) {
 497                                 md_error(rdev->mddev, rdev);
 498                                 if (!test_bit(Faulty, &rdev->flags))
 499                                         /* This is the only remaining device,
 500                                          * We need to retry the write without
 501                                          * FailFast
 502                                          */
 503                                         set_bit(R10BIO_WriteError, &r10_bio->state);
 504                                 else {
 505                                         r10_bio->devs[slot].bio = NULL;
 506                                         to_put = bio;
 507                                         dec_rdev = 1;
 508                                 }
 509                         } else
 510                                 set_bit(R10BIO_WriteError, &r10_bio->state);
 511                 }
 512         } else {
 513                 /*
 514                  * Set R10BIO_Uptodate in our master bio, so that
 515                  * we will return a good error code for to the higher
 516                  * levels even if IO on some other mirrored buffer fails.
 517                  *
 518                  * The 'master' represents the composite IO operation to
 519                  * user-side. So if something waits for IO, then it will
 520                  * wait for the 'master' bio.
 521                  */
 522                 sector_t first_bad;
 523                 int bad_sectors;
 524
 525                 /*
 526                  * Do not set R10BIO_Uptodate if the current device is
 527                  * rebuilding or Faulty. This is because we cannot use
 528                  * such device for properly reading the data back (we could
 529                  * potentially use it, if the current write would have felt
 530                  * before rdev->recovery_offset, but for simplicity we don't
 531                  * check this here.
 532                  */
 533                 if (test_bit(In_sync, &rdev->flags) &&
 534                     !test_bit(Faulty, &rdev->flags))
 535                         set_bit(R10BIO_Uptodate, &r10_bio->state);
 536
 537                 /* Maybe we can clear some bad blocks. */
 538                 if (is_badblock(rdev,
 539                                 r10_bio->devs[slot].addr,
 540                                 r10_bio->sectors,
 541                                 &first_bad, &bad_sectors) && !discard_error) {
 542                         bio_put(bio);
 543                         if (repl)
 544                                 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
 545                         else
 546                                 r10_bio->devs[slot].bio = IO_MADE_GOOD;
 547                         dec_rdev = 0;
 548                         set_bit(R10BIO_MadeGood, &r10_bio->state);
 549                 }
 550         }
 551
 552         /*
 553          *
 554          * Let's see if all mirrored write operations have finished
 555          * already.
 556          */
 557         one_write_done(r10_bio);
 558         if (dec_rdev)
 559                 rdev_dec_pending(rdev, conf->mddev);
 560         if (to_put)
 561                 bio_put(to_put);
 562 }
 563
 564 /*
 565  * RAID10 layout manager
 566  * As well as the chunksize and raid_disks count, there are two
 567  * parameters: near_copies and far_copies.
 568  * near_copies * far_copies must be <= raid_disks.
 569  * Normally one of these will be 1.
 570  * If both are 1, we get raid0.
 571  * If near_copies == raid_disks, we get raid1.
 572  *
 573  * Chunks are laid out in raid0 style with near_copies copies of the
 574  * first chunk, followed by near_copies copies of the next chunk and
 575  * so on.
 576  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 577  * as described above, we start again with a device offset of near_copies.
 578  * So we effectively have another copy of the whole array further down all
 579  * the drives, but with blocks on different drives.
 580  * With this layout, and block is never stored twice on the one device.
 581  *
 582  * raid10_find_phys finds the sector offset of a given virtual sector
 583  * on each device that it is on.
 584  *
 585  * raid10_find_virt does the reverse mapping, from a device and a
 586  * sector offset to a virtual address
 587  */
 588
 589 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
 590 {
 591         int n,f;
 592         sector_t sector;
 593         sector_t chunk;
 594         sector_t stripe;
 595         int dev;
 596         int slot = 0;
 597         int last_far_set_start, last_far_set_size;
 598
 599         last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
 600         last_far_set_start *= geo->far_set_size;
 601
 602         last_far_set_size = geo->far_set_size;
 603         last_far_set_size += (geo->raid_disks % geo->far_set_size);
 604
 605         /* now calculate first sector/dev */
 606         chunk = r10bio->sector >> geo->chunk_shift;
 607         sector = r10bio->sector & geo->chunk_mask;
 608
 609         chunk *= geo->near_copies;
 610         stripe = chunk;
 611         dev = sector_div(stripe, geo->raid_disks);
 612         if (geo->far_offset)
 613                 stripe *= geo->far_copies;
 614
 615         sector += stripe << geo->chunk_shift;
 616
 617         /* and calculate all the others */
 618         for (n = 0; n < geo->near_copies; n++) {
 619                 int d = dev;
 620                 int set;
 621                 sector_t s = sector;
 622                 r10bio->devs[slot].devnum = d;
 623                 r10bio->devs[slot].addr = s;
 624                 slot++;
 625
 626                 for (f = 1; f < geo->far_copies; f++) {
 627                         set = d / geo->far_set_size;
 628                         d += geo->near_copies;
 629
 630                         if ((geo->raid_disks % geo->far_set_size) &&
 631                             (d > last_far_set_start)) {
 632                                 d -= last_far_set_start;
 633                                 d %= last_far_set_size;
 634                                 d += last_far_set_start;
 635                         } else {
 636                                 d %= geo->far_set_size;
 637                                 d += geo->far_set_size * set;
 638                         }
 639                         s += geo->stride;
 640                         r10bio->devs[slot].devnum = d;
 641                         r10bio->devs[slot].addr = s;
 642                         slot++;
 643                 }
 644                 dev++;
 645                 if (dev >= geo->raid_disks) {
 646                         dev = 0;
 647                         sector += (geo->chunk_mask + 1);
 648                 }
 649         }
 650 }
 651
 652 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
 653 {
 654         struct geom *geo = &conf->geo;
 655
 656         if (conf->reshape_progress != MaxSector &&
 657             ((r10bio->sector >= conf->reshape_progress) !=
 658              conf->mddev->reshape_backwards)) {
 659                 set_bit(R10BIO_Previous, &r10bio->state);
 660                 geo = &conf->prev;
 661         } else
 662                 clear_bit(R10BIO_Previous, &r10bio->state);
 663
 664         __raid10_find_phys(geo, r10bio);
 665 }
 666
 667 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
 668 {
 669         sector_t offset, chunk, vchunk;
 670         /* Never use conf->prev as this is only called during resync
 671          * or recovery, so reshape isn't happening
 672          */
 673         struct geom *geo = &conf->geo;
 674         int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
 675         int far_set_size = geo->far_set_size;
 676         int last_far_set_start;
 677
 678         if (geo->raid_disks % geo->far_set_size) {
 679                 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
 680                 last_far_set_start *= geo->far_set_size;
 681
 682                 if (dev >= last_far_set_start) {
 683                         far_set_size = geo->far_set_size;
 684                         far_set_size += (geo->raid_disks % geo->far_set_size);
 685                         far_set_start = last_far_set_start;
 686                 }
 687         }
 688
 689         offset = sector & geo->chunk_mask;
 690         if (geo->far_offset) {
 691                 int fc;
 692                 chunk = sector >> geo->chunk_shift;
 693                 fc = sector_div(chunk, geo->far_copies);
 694                 dev -= fc * geo->near_copies;
 695                 if (dev < far_set_start)
 696                         dev += far_set_size;
 697         } else {
 698                 while (sector >= geo->stride) {
 699                         sector -= geo->stride;
 700                         if (dev < (geo->near_copies + far_set_start))
 701                                 dev += far_set_size - geo->near_copies;
 702                         else
 703                                 dev -= geo->near_copies;
 704                 }
 705                 chunk = sector >> geo->chunk_shift;
 706         }
 707         vchunk = chunk * geo->raid_disks + dev;
 708         sector_div(vchunk, geo->near_copies);
 709         return (vchunk << geo->chunk_shift) + offset;
 710 }
 711
 712 /*
 713  * This routine returns the disk from which the requested read should
 714  * be done. There is a per-array 'next expected sequential IO' sector
 715  * number - if this matches on the next IO then we use the last disk.
 716  * There is also a per-disk 'last know head position' sector that is
 717  * maintained from IRQ contexts, both the normal and the resync IO
 718  * completion handlers update this position correctly. If there is no
 719  * perfect sequential match then we pick the disk whose head is closest.
 720  *
 721  * If there are 2 mirrors in the same 2 devices, performance degrades
 722  * because position is mirror, not device based.
 723  *
 724  * The rdev for the device selected will have nr_pending incremented.
 725  */
 726
 727 /*
 728  * FIXME: possibly should rethink readbalancing and do it differently
 729  * depending on near_copies / far_copies geometry.
 730  */
 731 static struct md_rdev *read_balance(struct r10conf *conf,
 732                                     struct r10bio *r10_bio,
 733                                     int *max_sectors)
 734 {
 735         const sector_t this_sector = r10_bio->sector;
 736         int disk, slot;
 737         int sectors = r10_bio->sectors;
 738         int best_good_sectors;
 739         sector_t new_distance, best_dist;
 740         struct md_rdev *best_rdev, *rdev = NULL;
 741         int do_balance;
 742         int best_slot;
 743         struct geom *geo = &conf->geo;
 744
 745         raid10_find_phys(conf, r10_bio);
 746         rcu_read_lock();
 747         best_slot = -1;
 748         best_rdev = NULL;
 749         best_dist = MaxSector;
 750         best_good_sectors = 0;
 751         do_balance = 1;
 752         clear_bit(R10BIO_FailFast, &r10_bio->state);
 753         /*
 754          * Check if we can balance. We can balance on the whole
 755          * device if no resync is going on (recovery is ok), or below
 756          * the resync window. We take the first readable disk when
 757          * above the resync window.
 758          */
 759         if ((conf->mddev->recovery_cp < MaxSector
 760              && (this_sector + sectors >= conf->next_resync)) ||
 761             (mddev_is_clustered(conf->mddev) &&
 762              md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 763                                             this_sector + sectors)))
 764                 do_balance = 0;
 765
 766         for (slot = 0; slot < conf->copies ; slot++) {
 767                 sector_t first_bad;
 768                 int bad_sectors;
 769                 sector_t dev_sector;
 770
 771                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
 772                         continue;
 773                 disk = r10_bio->devs[slot].devnum;
 774                 rdev = rcu_dereference(conf->mirrors[disk].replacement);
 775                 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
 776                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 777                         rdev = rcu_dereference(conf->mirrors[disk].rdev);
 778                 if (rdev == NULL ||
 779                     test_bit(Faulty, &rdev->flags))
 780                         continue;
 781                 if (!test_bit(In_sync, &rdev->flags) &&
 782                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 783                         continue;
 784
 785                 dev_sector = r10_bio->devs[slot].addr;
 786                 if (is_badblock(rdev, dev_sector, sectors,
 787                                 &first_bad, &bad_sectors)) {
 788                         if (best_dist < MaxSector)
 789                                 /* Already have a better slot */
 790                                 continue;
 791                         if (first_bad <= dev_sector) {
 792                                 /* Cannot read here.  If this is the
 793                                  * 'primary' device, then we must not read
 794                                  * beyond 'bad_sectors' from another device.
 795                                  */
 796                                 bad_sectors -= (dev_sector - first_bad);
 797                                 if (!do_balance && sectors > bad_sectors)
 798                                         sectors = bad_sectors;
 799                                 if (best_good_sectors > sectors)
 800                                         best_good_sectors = sectors;
 801                         } else {
 802                                 sector_t good_sectors =
 803                                         first_bad - dev_sector;
 804                                 if (good_sectors > best_good_sectors) {
 805                                         best_good_sectors = good_sectors;
 806                                         best_slot = slot;
 807                                         best_rdev = rdev;
 808                                 }
 809                                 if (!do_balance)
 810                                         /* Must read from here */
 811                                         break;
 812                         }
 813                         continue;
 814                 } else
 815                         best_good_sectors = sectors;
 816
 817                 if (!do_balance)
 818                         break;
 819
 820                 if (best_slot >= 0)
 821                         /* At least 2 disks to choose from so failfast is OK */
 822                         set_bit(R10BIO_FailFast, &r10_bio->state);
 823                 /* This optimisation is debatable, and completely destroys
 824                  * sequential read speed for 'far copies' arrays.  So only
 825                  * keep it for 'near' arrays, and review those later.
 826                  */
 827                 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
 828                         new_distance = 0;
 829
 830                 /* for far > 1 always use the lowest address */
 831                 else if (geo->far_copies > 1)
 832                         new_distance = r10_bio->devs[slot].addr;
 833                 else
 834                         new_distance = abs(r10_bio->devs[slot].addr -
 835                                            conf->mirrors[disk].head_position);
 836                 if (new_distance < best_dist) {
 837                         best_dist = new_distance;
 838                         best_slot = slot;
 839                         best_rdev = rdev;
 840                 }
 841         }
 842         if (slot >= conf->copies) {
 843                 slot = best_slot;
 844                 rdev = best_rdev;
 845         }
 846
 847         if (slot >= 0) {
 848                 atomic_inc(&rdev->nr_pending);
 849                 r10_bio->read_slot = slot;
 850         } else
 851                 rdev = NULL;
 852         rcu_read_unlock();
 853         *max_sectors = best_good_sectors;
 854
 855         return rdev;
 856 }
 857
 858 static int raid10_congested(struct mddev *mddev, int bits)
 859 {
 860         struct r10conf *conf = mddev->private;
 861         int i, ret = 0;
 862
 863         if ((bits & (1 << WB_async_congested)) &&
 864             conf->pending_count >= max_queued_requests)
 865                 return 1;
 866
 867         rcu_read_lock();
 868         for (i = 0;
 869              (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
 870                      && ret == 0;
 871              i++) {
 872                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 873                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 874                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 875
 876                         ret |= bdi_congested(q->backing_dev_info, bits);
 877                 }
 878         }
 879         rcu_read_unlock();
 880         return ret;
 881 }
 882
 883 static void flush_pending_writes(struct r10conf *conf)
 884 {
 885         /* Any writes that have been queued but are awaiting
 886          * bitmap updates get flushed here.
 887          */
 888         spin_lock_irq(&conf->device_lock);
 889
 890         if (conf->pending_bio_list.head) {
 891                 struct blk_plug plug;
 892                 struct bio *bio;
 893
 894                 bio = bio_list_get(&conf->pending_bio_list);
 895                 conf->pending_count = 0;
 896                 spin_unlock_irq(&conf->device_lock);
 897
 898                 /*
 899                  * As this is called in a wait_event() loop (see freeze_array),
 900                  * current->state might be TASK_UNINTERRUPTIBLE which will
 901                  * cause a warning when we prepare to wait again.  As it is
 902                  * rare that this path is taken, it is perfectly safe to force
 903                  * us to go around the wait_event() loop again, so the warning
 904                  * is a false-positive. Silence the warning by resetting
 905                  * thread state
 906                  */
 907                 __set_current_state(TASK_RUNNING);
 908
 909                 blk_start_plug(&plug);
 910                 /* flush any pending bitmap writes to disk
 911                  * before proceeding w/ I/O */
 912                 md_bitmap_unplug(conf->mddev->bitmap);
 913                 wake_up(&conf->wait_barrier);
 914
 915                 while (bio) { /* submit pending writes */
 916                         struct bio *next = bio->bi_next;
 917                         struct md_rdev *rdev = (void*)bio->bi_disk;
 918                         bio->bi_next = NULL;
 919                         bio_set_dev(bio, rdev->bdev);
 920                         if (test_bit(Faulty, &rdev->flags)) {
 921                                 bio_io_error(bio);
 922                         } else if (unlikely((bio_op(bio) ==  REQ_OP_DISCARD) &&
 923                                             !blk_queue_discard(bio->bi_disk->queue)))
 924                                 /* Just ignore it */
 925                                 bio_endio(bio);
 926                         else
 927                                 generic_make_request(bio);
 928                         bio = next;
 929                 }
 930                 blk_finish_plug(&plug);
 931         } else
 932                 spin_unlock_irq(&conf->device_lock);
 933 }
 934
 935 /* Barriers....
 936  * Sometimes we need to suspend IO while we do something else,
 937  * either some resync/recovery, or reconfigure the array.
 938  * To do this we raise a 'barrier'.
 939  * The 'barrier' is a counter that can be raised multiple times
 940  * to count how many activities are happening which preclude
 941  * normal IO.
 942  * We can only raise the barrier if there is no pending IO.
 943  * i.e. if nr_pending == 0.
 944  * We choose only to raise the barrier if no-one is waiting for the
 945  * barrier to go down.  This means that as soon as an IO request
 946  * is ready, no other operations which require a barrier will start
 947  * until the IO request has had a chance.
 948  *
 949  * So: regular IO calls 'wait_barrier'.  When that returns there
 950  *    is no backgroup IO happening,  It must arrange to call
 951  *    allow_barrier when it has finished its IO.
 952  * backgroup IO calls must call raise_barrier.  Once that returns
 953  *    there is no normal IO happeing.  It must arrange to call
 954  *    lower_barrier when the particular background IO completes.
 955  */
 956
 957 static void raise_barrier(struct r10conf *conf, int force)
 958 {
 959         BUG_ON(force && !conf->barrier);
 960         spin_lock_irq(&conf->resync_lock);
 961
 962         /* Wait until no block IO is waiting (unless 'force') */
 963         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 964                             conf->resync_lock);
 965
 966         /* block any new IO from starting */
 967         conf->barrier++;
 968
 969         /* Now wait for all pending IO to complete */
 970         wait_event_lock_irq(conf->wait_barrier,
 971                             !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
 972                             conf->resync_lock);
 973
 974         spin_unlock_irq(&conf->resync_lock);
 975 }
 976
 977 static void lower_barrier(struct r10conf *conf)
 978 {
 979         unsigned long flags;
 980         spin_lock_irqsave(&conf->resync_lock, flags);
 981         conf->barrier--;
 982         spin_unlock_irqrestore(&conf->resync_lock, flags);
 983         wake_up(&conf->wait_barrier);
 984 }
 985
 986 static void wait_barrier(struct r10conf *conf)
 987 {
 988         spin_lock_irq(&conf->resync_lock);
 989         if (conf->barrier) {
 990                 conf->nr_waiting++;
 991                 /* Wait for the barrier to drop.
 992                  * However if there are already pending
 993                  * requests (preventing the barrier from
 994                  * rising completely), and the
 995                  * pre-process bio queue isn't empty,
 996                  * then don't wait, as we need to empty
 997                  * that queue to get the nr_pending
 998                  * count down.
 999                  */
1000                 raid10_log(conf->mddev, "wait barrier");
1001                 wait_event_lock_irq(conf->wait_barrier,
1002                                     !conf->barrier ||
1003                                     (atomic_read(&conf->nr_pending) &&
1004                                      current->bio_list &&
1005                                      (!bio_list_empty(&current->bio_list[0]) ||
1006                                       !bio_list_empty(&current->bio_list[1]))),
1007                                     conf->resync_lock);
1008                 conf->nr_waiting--;
1009                 if (!conf->nr_waiting)
1010                         wake_up(&conf->wait_barrier);
1011         }
1012         atomic_inc(&conf->nr_pending);
1013         spin_unlock_irq(&conf->resync_lock);
1014 }
1015
1016 static void allow_barrier(struct r10conf *conf)
1017 {
1018         if ((atomic_dec_and_test(&conf->nr_pending)) ||
1019                         (conf->array_freeze_pending))
1020                 wake_up(&conf->wait_barrier);
1021 }
1022
1023 static void freeze_array(struct r10conf *conf, int extra)
1024 {
1025         /* stop syncio and normal IO and wait for everything to
1026          * go quiet.
1027          * We increment barrier and nr_waiting, and then
1028          * wait until nr_pending match nr_queued+extra
1029          * This is called in the context of one normal IO request
1030          * that has failed. Thus any sync request that might be pending
1031          * will be blocked by nr_pending, and we need to wait for
1032          * pending IO requests to complete or be queued for re-try.
1033          * Thus the number queued (nr_queued) plus this request (extra)
1034          * must match the number of pending IOs (nr_pending) before
1035          * we continue.
1036          */
1037         spin_lock_irq(&conf->resync_lock);
1038         conf->array_freeze_pending++;
1039         conf->barrier++;
1040         conf->nr_waiting++;
1041         wait_event_lock_irq_cmd(conf->wait_barrier,
1042                                 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1043                                 conf->resync_lock,
1044                                 flush_pending_writes(conf));
1045
1046         conf->array_freeze_pending--;
1047         spin_unlock_irq(&conf->resync_lock);
1048 }
1049
1050 static void unfreeze_array(struct r10conf *conf)
1051 {
1052         /* reverse the effect of the freeze */
1053         spin_lock_irq(&conf->resync_lock);
1054         conf->barrier--;
1055         conf->nr_waiting--;
1056         wake_up(&conf->wait_barrier);
1057         spin_unlock_irq(&conf->resync_lock);
1058 }
1059
1060 static sector_t choose_data_offset(struct r10bio *r10_bio,
1061                                    struct md_rdev *rdev)
1062 {
1063         if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1064             test_bit(R10BIO_Previous, &r10_bio->state))
1065                 return rdev->data_offset;
1066         else
1067                 return rdev->new_data_offset;
1068 }
1069
1070 struct raid10_plug_cb {
1071         struct blk_plug_cb      cb;
1072         struct bio_list         pending;
1073         int                     pending_cnt;
1074 };
1075
1076 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1077 {
1078         struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1079                                                    cb);
1080         struct mddev *mddev = plug->cb.data;
1081         struct r10conf *conf = mddev->private;
1082         struct bio *bio;
1083
1084         if (from_schedule || current->bio_list) {
1085                 spin_lock_irq(&conf->device_lock);
1086                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1087                 conf->pending_count += plug->pending_cnt;
1088                 spin_unlock_irq(&conf->device_lock);
1089                 wake_up(&conf->wait_barrier);
1090                 md_wakeup_thread(mddev->thread);
1091                 kfree(plug);
1092                 return;
1093         }
1094
1095         /* we aren't scheduling, so we can do the write-out directly. */
1096         bio = bio_list_get(&plug->pending);
1097         md_bitmap_unplug(mddev->bitmap);
1098         wake_up(&conf->wait_barrier);
1099
1100         while (bio) { /* submit pending writes */
1101                 struct bio *next = bio->bi_next;
1102                 struct md_rdev *rdev = (void*)bio->bi_disk;
1103                 bio->bi_next = NULL;
1104                 bio_set_dev(bio, rdev->bdev);
1105                 if (test_bit(Faulty, &rdev->flags)) {
1106                         bio_io_error(bio);
1107                 } else if (unlikely((bio_op(bio) ==  REQ_OP_DISCARD) &&
1108                                     !blk_queue_discard(bio->bi_disk->queue)))
1109                         /* Just ignore it */
1110                         bio_endio(bio);
1111                 else
1112                         generic_make_request(bio);
1113                 bio = next;
1114         }
1115         kfree(plug);
1116 }
1117
1118 /*
1119  * 1. Register the new request and wait if the reconstruction thread has put
1120  * up a bar for new requests. Continue immediately if no resync is active
1121  * currently.
1122  * 2. If IO spans the reshape position.  Need to wait for reshape to pass.
1123  */
1124 static void regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1125                                  struct bio *bio, sector_t sectors)
1126 {
1127         wait_barrier(conf);
1128         while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1129             bio->bi_iter.bi_sector < conf->reshape_progress &&
1130             bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1131                 raid10_log(conf->mddev, "wait reshape");
1132                 allow_barrier(conf);
1133                 wait_event(conf->wait_barrier,
1134                            conf->reshape_progress <= bio->bi_iter.bi_sector ||
1135                            conf->reshape_progress >= bio->bi_iter.bi_sector +
1136                            sectors);
1137                 wait_barrier(conf);
1138         }
1139 }
1140
1141 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1142                                 struct r10bio *r10_bio)
1143 {
1144         struct r10conf *conf = mddev->private;
1145         struct bio *read_bio;
1146         const int op = bio_op(bio);
1147         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1148         int max_sectors;
1149         struct md_rdev *rdev;
1150         char b[BDEVNAME_SIZE];
1151         int slot = r10_bio->read_slot;
1152         struct md_rdev *err_rdev = NULL;
1153         gfp_t gfp = GFP_NOIO;
1154
1155         if (r10_bio->devs[slot].rdev) {
1156                 /*
1157                  * This is an error retry, but we cannot
1158                  * safely dereference the rdev in the r10_bio,
1159                  * we must use the one in conf.
1160                  * If it has already been disconnected (unlikely)
1161                  * we lose the device name in error messages.
1162                  */
1163                 int disk;
1164                 /*
1165                  * As we are blocking raid10, it is a little safer to
1166                  * use __GFP_HIGH.
1167                  */
1168                 gfp = GFP_NOIO | __GFP_HIGH;
1169
1170                 rcu_read_lock();
1171                 disk = r10_bio->devs[slot].devnum;
1172                 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1173                 if (err_rdev)
1174                         bdevname(err_rdev->bdev, b);
1175                 else {
1176                         strcpy(b, "???");
1177                         /* This never gets dereferenced */
1178                         err_rdev = r10_bio->devs[slot].rdev;
1179                 }
1180                 rcu_read_unlock();
1181         }
1182
1183         regular_request_wait(mddev, conf, bio, r10_bio->sectors);
1184         rdev = read_balance(conf, r10_bio, &max_sectors);
1185         if (!rdev) {
1186                 if (err_rdev) {
1187                         pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1188                                             mdname(mddev), b,
1189                                             (unsigned long long)r10_bio->sector);
1190                 }
1191                 raid_end_bio_io(r10_bio);
1192                 return;
1193         }
1194         if (err_rdev)
1195                 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1196                                    mdname(mddev),
1197                                    bdevname(rdev->bdev, b),
1198                                    (unsigned long long)r10_bio->sector);
1199         if (max_sectors < bio_sectors(bio)) {
1200                 struct bio *split = bio_split(bio, max_sectors,
1201                                               gfp, &conf->bio_split);
1202                 bio_chain(split, bio);
1203                 allow_barrier(conf);
1204                 generic_make_request(bio);
1205                 wait_barrier(conf);
1206                 bio = split;
1207                 r10_bio->master_bio = bio;
1208                 r10_bio->sectors = max_sectors;
1209         }
1210         slot = r10_bio->read_slot;
1211
1212         read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1213
1214         r10_bio->devs[slot].bio = read_bio;
1215         r10_bio->devs[slot].rdev = rdev;
1216
1217         read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1218                 choose_data_offset(r10_bio, rdev);
1219         bio_set_dev(read_bio, rdev->bdev);
1220         read_bio->bi_end_io = raid10_end_read_request;
1221         bio_set_op_attrs(read_bio, op, do_sync);
1222         if (test_bit(FailFast, &rdev->flags) &&
1223             test_bit(R10BIO_FailFast, &r10_bio->state))
1224                 read_bio->bi_opf |= MD_FAILFAST;
1225         read_bio->bi_private = r10_bio;
1226
1227         if (mddev->gendisk)
1228                 trace_block_bio_remap(read_bio->bi_disk->queue,
1229                                       read_bio, disk_devt(mddev->gendisk),
1230                                       r10_bio->sector);
1231         generic_make_request(read_bio);
1232         return;
1233 }
1234
1235 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1236                                   struct bio *bio, bool replacement,
1237                                   int n_copy)
1238 {
1239         const int op = bio_op(bio);
1240         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1241         const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1242         unsigned long flags;
1243         struct blk_plug_cb *cb;
1244         struct raid10_plug_cb *plug = NULL;
1245         struct r10conf *conf = mddev->private;
1246         struct md_rdev *rdev;
1247         int devnum = r10_bio->devs[n_copy].devnum;
1248         struct bio *mbio;
1249
1250         if (replacement) {
1251                 rdev = conf->mirrors[devnum].replacement;
1252                 if (rdev == NULL) {
1253                         /* Replacement just got moved to main 'rdev' */
1254                         smp_mb();
1255                         rdev = conf->mirrors[devnum].rdev;
1256                 }
1257         } else
1258                 rdev = conf->mirrors[devnum].rdev;
1259
1260         mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1261         if (replacement)
1262                 r10_bio->devs[n_copy].repl_bio = mbio;
1263         else
1264                 r10_bio->devs[n_copy].bio = mbio;
1265
1266         mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1267                                    choose_data_offset(r10_bio, rdev));
1268         bio_set_dev(mbio, rdev->bdev);
1269         mbio->bi_end_io = raid10_end_write_request;
1270         bio_set_op_attrs(mbio, op, do_sync | do_fua);
1271         if (!replacement && test_bit(FailFast,
1272                                      &conf->mirrors[devnum].rdev->flags)
1273                          && enough(conf, devnum))
1274                 mbio->bi_opf |= MD_FAILFAST;
1275         mbio->bi_private = r10_bio;
1276
1277         if (conf->mddev->gendisk)
1278                 trace_block_bio_remap(mbio->bi_disk->queue,
1279                                       mbio, disk_devt(conf->mddev->gendisk),
1280                                       r10_bio->sector);
1281         /* flush_pending_writes() needs access to the rdev so...*/
1282         mbio->bi_disk = (void *)rdev;
1283
1284         atomic_inc(&r10_bio->remaining);
1285
1286         cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1287         if (cb)
1288                 plug = container_of(cb, struct raid10_plug_cb, cb);
1289         else
1290                 plug = NULL;
1291         if (plug) {
1292                 bio_list_add(&plug->pending, mbio);
1293                 plug->pending_cnt++;
1294         } else {
1295                 spin_lock_irqsave(&conf->device_lock, flags);
1296                 bio_list_add(&conf->pending_bio_list, mbio);
1297                 conf->pending_count++;
1298                 spin_unlock_irqrestore(&conf->device_lock, flags);
1299                 md_wakeup_thread(mddev->thread);
1300         }
1301 }
1302
1303 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1304                                  struct r10bio *r10_bio)
1305 {
1306         struct r10conf *conf = mddev->private;
1307         int i;
1308         struct md_rdev *blocked_rdev;
1309         sector_t sectors;
1310         int max_sectors;
1311
1312         if ((mddev_is_clustered(mddev) &&
1313              md_cluster_ops->area_resyncing(mddev, WRITE,
1314                                             bio->bi_iter.bi_sector,
1315                                             bio_end_sector(bio)))) {
1316                 DEFINE_WAIT(w);
1317                 for (;;) {
1318                         prepare_to_wait(&conf->wait_barrier,
1319                                         &w, TASK_IDLE);
1320                         if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1321                                  bio->bi_iter.bi_sector, bio_end_sector(bio)))
1322                                 break;
1323                         schedule();
1324                 }
1325                 finish_wait(&conf->wait_barrier, &w);
1326         }
1327
1328         sectors = r10_bio->sectors;
1329         regular_request_wait(mddev, conf, bio, sectors);
1330         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1331             (mddev->reshape_backwards
1332              ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1333                 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1334              : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1335                 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1336                 /* Need to update reshape_position in metadata */
1337                 mddev->reshape_position = conf->reshape_progress;
1338                 set_mask_bits(&mddev->sb_flags, 0,
1339                               BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1340                 md_wakeup_thread(mddev->thread);
1341                 raid10_log(conf->mddev, "wait reshape metadata");
1342                 wait_event(mddev->sb_wait,
1343                            !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1344
1345                 conf->reshape_safe = mddev->reshape_position;
1346         }
1347
1348         if (conf->pending_count >= max_queued_requests) {
1349                 md_wakeup_thread(mddev->thread);
1350                 raid10_log(mddev, "wait queued");
1351                 wait_event(conf->wait_barrier,
1352                            conf->pending_count < max_queued_requests);
1353         }
1354         /* first select target devices under rcu_lock and
1355          * inc refcount on their rdev.  Record them by setting
1356          * bios[x] to bio
1357          * If there are known/acknowledged bad blocks on any device
1358          * on which we have seen a write error, we want to avoid
1359          * writing to those blocks.  This potentially requires several
1360          * writes to write around the bad blocks.  Each set of writes
1361          * gets its own r10_bio with a set of bios attached.
1362          */
1363
1364         r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1365         raid10_find_phys(conf, r10_bio);
1366 retry_write:
1367         blocked_rdev = NULL;
1368         rcu_read_lock();
1369         max_sectors = r10_bio->sectors;
1370
1371         for (i = 0;  i < conf->copies; i++) {
1372                 int d = r10_bio->devs[i].devnum;
1373                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1374                 struct md_rdev *rrdev = rcu_dereference(
1375                         conf->mirrors[d].replacement);
1376                 if (rdev == rrdev)
1377                         rrdev = NULL;
1378                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1379                         atomic_inc(&rdev->nr_pending);
1380                         blocked_rdev = rdev;
1381                         break;
1382                 }
1383                 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1384                         atomic_inc(&rrdev->nr_pending);
1385                         blocked_rdev = rrdev;
1386                         break;
1387                 }
1388                 if (rdev && (test_bit(Faulty, &rdev->flags)))
1389                         rdev = NULL;
1390                 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1391                         rrdev = NULL;
1392
1393                 r10_bio->devs[i].bio = NULL;
1394                 r10_bio->devs[i].repl_bio = NULL;
1395
1396                 if (!rdev && !rrdev) {
1397                         set_bit(R10BIO_Degraded, &r10_bio->state);
1398                         continue;
1399                 }
1400                 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1401                         sector_t first_bad;
1402                         sector_t dev_sector = r10_bio->devs[i].addr;
1403                         int bad_sectors;
1404                         int is_bad;
1405
1406                         is_bad = is_badblock(rdev, dev_sector, max_sectors,
1407                                              &first_bad, &bad_sectors);
1408                         if (is_bad < 0) {
1409                                 /* Mustn't write here until the bad block
1410                                  * is acknowledged
1411                                  */
1412                                 atomic_inc(&rdev->nr_pending);
1413                                 set_bit(BlockedBadBlocks, &rdev->flags);
1414                                 blocked_rdev = rdev;
1415                                 break;
1416                         }
1417                         if (is_bad && first_bad <= dev_sector) {
1418                                 /* Cannot write here at all */
1419                                 bad_sectors -= (dev_sector - first_bad);
1420                                 if (bad_sectors < max_sectors)
1421                                         /* Mustn't write more than bad_sectors
1422                                          * to other devices yet
1423                                          */
1424                                         max_sectors = bad_sectors;
1425                                 /* We don't set R10BIO_Degraded as that
1426                                  * only applies if the disk is missing,
1427                                  * so it might be re-added, and we want to
1428                                  * know to recover this chunk.
1429                                  * In this case the device is here, and the
1430                                  * fact that this chunk is not in-sync is
1431                                  * recorded in the bad block log.
1432                                  */
1433                                 continue;
1434                         }
1435                         if (is_bad) {
1436                                 int good_sectors = first_bad - dev_sector;
1437                                 if (good_sectors < max_sectors)
1438                                         max_sectors = good_sectors;
1439                         }
1440                 }
1441                 if (rdev) {
1442                         r10_bio->devs[i].bio = bio;
1443                         atomic_inc(&rdev->nr_pending);
1444                 }
1445                 if (rrdev) {
1446                         r10_bio->devs[i].repl_bio = bio;
1447                         atomic_inc(&rrdev->nr_pending);
1448                 }
1449         }
1450         rcu_read_unlock();
1451
1452         if (unlikely(blocked_rdev)) {
1453                 /* Have to wait for this device to get unblocked, then retry */
1454                 int j;
1455                 int d;
1456
1457                 for (j = 0; j < i; j++) {
1458                         if (r10_bio->devs[j].bio) {
1459                                 d = r10_bio->devs[j].devnum;
1460                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1461                         }
1462                         if (r10_bio->devs[j].repl_bio) {
1463                                 struct md_rdev *rdev;
1464                                 d = r10_bio->devs[j].devnum;
1465                                 rdev = conf->mirrors[d].replacement;
1466                                 if (!rdev) {
1467                                         /* Race with remove_disk */
1468                                         smp_mb();
1469                                         rdev = conf->mirrors[d].rdev;
1470                                 }
1471                                 rdev_dec_pending(rdev, mddev);
1472                         }
1473                 }
1474                 allow_barrier(conf);
1475                 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1476                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1477                 wait_barrier(conf);
1478                 goto retry_write;
1479         }
1480
1481         if (max_sectors < r10_bio->sectors)
1482                 r10_bio->sectors = max_sectors;
1483
1484         if (r10_bio->sectors < bio_sectors(bio)) {
1485                 struct bio *split = bio_split(bio, r10_bio->sectors,
1486                                               GFP_NOIO, &conf->bio_split);
1487                 bio_chain(split, bio);
1488                 allow_barrier(conf);
1489                 generic_make_request(bio);
1490                 wait_barrier(conf);
1491                 bio = split;
1492                 r10_bio->master_bio = bio;
1493         }
1494
1495         atomic_set(&r10_bio->remaining, 1);
1496         md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1497
1498         for (i = 0; i < conf->copies; i++) {
1499                 if (r10_bio->devs[i].bio)
1500                         raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1501                 if (r10_bio->devs[i].repl_bio)
1502                         raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1503         }
1504         one_write_done(r10_bio);
1505 }
1506
1507 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1508 {
1509         struct r10conf *conf = mddev->private;
1510         struct r10bio *r10_bio;
1511
1512         r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1513
1514         r10_bio->master_bio = bio;
1515         r10_bio->sectors = sectors;
1516
1517         r10_bio->mddev = mddev;
1518         r10_bio->sector = bio->bi_iter.bi_sector;
1519         r10_bio->state = 0;
1520         memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1521
1522         if (bio_data_dir(bio) == READ)
1523                 raid10_read_request(mddev, bio, r10_bio);
1524         else
1525                 raid10_write_request(mddev, bio, r10_bio);
1526 }
1527
1528 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1529 {
1530         struct r10conf *conf = mddev->private;
1531         sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1532         int chunk_sects = chunk_mask + 1;
1533         int sectors = bio_sectors(bio);
1534
1535         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1536                 md_flush_request(mddev, bio);
1537                 return true;
1538         }
1539
1540         if (!md_write_start(mddev, bio))
1541                 return false;
1542
1543         /*
1544          * If this request crosses a chunk boundary, we need to split
1545          * it.
1546          */
1547         if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1548                      sectors > chunk_sects
1549                      && (conf->geo.near_copies < conf->geo.raid_disks
1550                          || conf->prev.near_copies <
1551                          conf->prev.raid_disks)))
1552                 sectors = chunk_sects -
1553                         (bio->bi_iter.bi_sector &
1554                          (chunk_sects - 1));
1555         __make_request(mddev, bio, sectors);
1556
1557         /* In case raid10d snuck in to freeze_array */
1558         wake_up(&conf->wait_barrier);
1559         return true;
1560 }
1561
1562 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1563 {
1564         struct r10conf *conf = mddev->private;
1565         int i;
1566
1567         if (conf->geo.near_copies < conf->geo.raid_disks)
1568                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1569         if (conf->geo.near_copies > 1)
1570                 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1571         if (conf->geo.far_copies > 1) {
1572                 if (conf->geo.far_offset)
1573                         seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1574                 else
1575                         seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1576                 if (conf->geo.far_set_size != conf->geo.raid_disks)
1577                         seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1578         }
1579         seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1580                                         conf->geo.raid_disks - mddev->degraded);
1581         rcu_read_lock();
1582         for (i = 0; i < conf->geo.raid_disks; i++) {
1583                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1584                 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1585         }
1586         rcu_read_unlock();
1587         seq_printf(seq, "]");
1588 }
1589
1590 /* check if there are enough drives for
1591  * every block to appear on atleast one.
1592  * Don't consider the device numbered 'ignore'
1593  * as we might be about to remove it.
1594  */
1595 static int _enough(struct r10conf *conf, int previous, int ignore)
1596 {
1597         int first = 0;
1598         int has_enough = 0;
1599         int disks, ncopies;
1600         if (previous) {
1601                 disks = conf->prev.raid_disks;
1602                 ncopies = conf->prev.near_copies;
1603         } else {
1604                 disks = conf->geo.raid_disks;
1605                 ncopies = conf->geo.near_copies;
1606         }
1607
1608         rcu_read_lock();
1609         do {
1610                 int n = conf->copies;
1611                 int cnt = 0;
1612                 int this = first;
1613                 while (n--) {
1614                         struct md_rdev *rdev;
1615                         if (this != ignore &&
1616                             (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1617                             test_bit(In_sync, &rdev->flags))
1618                                 cnt++;
1619                         this = (this+1) % disks;
1620                 }
1621                 if (cnt == 0)
1622                         goto out;
1623                 first = (first + ncopies) % disks;
1624         } while (first != 0);
1625         has_enough = 1;
1626 out:
1627         rcu_read_unlock();
1628         return has_enough;
1629 }
1630
1631 static int enough(struct r10conf *conf, int ignore)
1632 {
1633         /* when calling 'enough', both 'prev' and 'geo' must
1634          * be stable.
1635          * This is ensured if ->reconfig_mutex or ->device_lock
1636          * is held.
1637          */
1638         return _enough(conf, 0, ignore) &&
1639                 _enough(conf, 1, ignore);
1640 }
1641
1642 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1643 {
1644         char b[BDEVNAME_SIZE];
1645         struct r10conf *conf = mddev->private;
1646         unsigned long flags;
1647
1648         /*
1649          * If it is not operational, then we have already marked it as dead
1650          * else if it is the last working disks, ignore the error, let the
1651          * next level up know.
1652          * else mark the drive as failed
1653          */
1654         spin_lock_irqsave(&conf->device_lock, flags);
1655         if (test_bit(In_sync, &rdev->flags)
1656             && !enough(conf, rdev->raid_disk)) {
1657                 /*
1658                  * Don't fail the drive, just return an IO error.
1659                  */
1660                 spin_unlock_irqrestore(&conf->device_lock, flags);
1661                 return;
1662         }
1663         if (test_and_clear_bit(In_sync, &rdev->flags))
1664                 mddev->degraded++;
1665         /*
1666          * If recovery is running, make sure it aborts.
1667          */
1668         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1669         set_bit(Blocked, &rdev->flags);
1670         set_bit(Faulty, &rdev->flags);
1671         set_mask_bits(&mddev->sb_flags, 0,
1672                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1673         spin_unlock_irqrestore(&conf->device_lock, flags);
1674         pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1675                 "md/raid10:%s: Operation continuing on %d devices.\n",
1676                 mdname(mddev), bdevname(rdev->bdev, b),
1677                 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1678 }
1679
1680 static void print_conf(struct r10conf *conf)
1681 {
1682         int i;
1683         struct md_rdev *rdev;
1684
1685         pr_debug("RAID10 conf printout:\n");
1686         if (!conf) {
1687                 pr_debug("(!conf)\n");
1688                 return;
1689         }
1690         pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1691                  conf->geo.raid_disks);
1692
1693         /* This is only called with ->reconfix_mutex held, so
1694          * rcu protection of rdev is not needed */
1695         for (i = 0; i < conf->geo.raid_disks; i++) {
1696                 char b[BDEVNAME_SIZE];
1697                 rdev = conf->mirrors[i].rdev;
1698                 if (rdev)
1699                         pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1700                                  i, !test_bit(In_sync, &rdev->flags),
1701                                  !test_bit(Faulty, &rdev->flags),
1702                                  bdevname(rdev->bdev,b));
1703         }
1704 }
1705
1706 static void close_sync(struct r10conf *conf)
1707 {
1708         wait_barrier(conf);
1709         allow_barrier(conf);
1710
1711         mempool_exit(&conf->r10buf_pool);
1712 }
1713
1714 static int raid10_spare_active(struct mddev *mddev)
1715 {
1716         int i;
1717         struct r10conf *conf = mddev->private;
1718         struct raid10_info *tmp;
1719         int count = 0;
1720         unsigned long flags;
1721
1722         /*
1723          * Find all non-in_sync disks within the RAID10 configuration
1724          * and mark them in_sync
1725          */
1726         for (i = 0; i < conf->geo.raid_disks; i++) {
1727                 tmp = conf->mirrors + i;
1728                 if (tmp->replacement
1729                     && tmp->replacement->recovery_offset == MaxSector
1730                     && !test_bit(Faulty, &tmp->replacement->flags)
1731                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1732                         /* Replacement has just become active */
1733                         if (!tmp->rdev
1734                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1735                                 count++;
1736                         if (tmp->rdev) {
1737                                 /* Replaced device not technically faulty,
1738                                  * but we need to be sure it gets removed
1739                                  * and never re-added.
1740                                  */
1741                                 set_bit(Faulty, &tmp->rdev->flags);
1742                                 sysfs_notify_dirent_safe(
1743                                         tmp->rdev->sysfs_state);
1744                         }
1745                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1746                 } else if (tmp->rdev
1747                            && tmp->rdev->recovery_offset == MaxSector
1748                            && !test_bit(Faulty, &tmp->rdev->flags)
1749                            && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1750                         count++;
1751                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1752                 }
1753         }
1754         spin_lock_irqsave(&conf->device_lock, flags);
1755         mddev->degraded -= count;
1756         spin_unlock_irqrestore(&conf->device_lock, flags);
1757
1758         print_conf(conf);
1759         return count;
1760 }
1761
1762 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1763 {
1764         struct r10conf *conf = mddev->private;
1765         int err = -EEXIST;
1766         int mirror;
1767         int first = 0;
1768         int last = conf->geo.raid_disks - 1;
1769
1770         if (mddev->recovery_cp < MaxSector)
1771                 /* only hot-add to in-sync arrays, as recovery is
1772                  * very different from resync
1773                  */
1774                 return -EBUSY;
1775         if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1776                 return -EINVAL;
1777
1778         if (md_integrity_add_rdev(rdev, mddev))
1779                 return -ENXIO;
1780
1781         if (rdev->raid_disk >= 0)
1782                 first = last = rdev->raid_disk;
1783
1784         if (rdev->saved_raid_disk >= first &&
1785             rdev->saved_raid_disk < conf->geo.raid_disks &&
1786             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1787                 mirror = rdev->saved_raid_disk;
1788         else
1789                 mirror = first;
1790         for ( ; mirror <= last ; mirror++) {
1791                 struct raid10_info *p = &conf->mirrors[mirror];
1792                 if (p->recovery_disabled == mddev->recovery_disabled)
1793                         continue;
1794                 if (p->rdev) {
1795                         if (!test_bit(WantReplacement, &p->rdev->flags) ||
1796                             p->replacement != NULL)
1797                                 continue;
1798                         clear_bit(In_sync, &rdev->flags);
1799                         set_bit(Replacement, &rdev->flags);
1800                         rdev->raid_disk = mirror;
1801                         err = 0;
1802                         if (mddev->gendisk)
1803                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1804                                                   rdev->data_offset << 9);
1805                         conf->fullsync = 1;
1806                         rcu_assign_pointer(p->replacement, rdev);
1807                         break;
1808                 }
1809
1810                 if (mddev->gendisk)
1811                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1812                                           rdev->data_offset << 9);
1813
1814                 p->head_position = 0;
1815                 p->recovery_disabled = mddev->recovery_disabled - 1;
1816                 rdev->raid_disk = mirror;
1817                 err = 0;
1818                 if (rdev->saved_raid_disk != mirror)
1819                         conf->fullsync = 1;
1820                 rcu_assign_pointer(p->rdev, rdev);
1821                 break;
1822         }
1823         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1824                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1825
1826         print_conf(conf);
1827         return err;
1828 }
1829
1830 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1831 {
1832         struct r10conf *conf = mddev->private;
1833         int err = 0;
1834         int number = rdev->raid_disk;
1835         struct md_rdev **rdevp;
1836         struct raid10_info *p = conf->mirrors + number;
1837
1838         print_conf(conf);
1839         if (rdev == p->rdev)
1840                 rdevp = &p->rdev;
1841         else if (rdev == p->replacement)
1842                 rdevp = &p->replacement;
1843         else
1844                 return 0;
1845
1846         if (test_bit(In_sync, &rdev->flags) ||
1847             atomic_read(&rdev->nr_pending)) {
1848                 err = -EBUSY;
1849                 goto abort;
1850         }
1851         /* Only remove non-faulty devices if recovery
1852          * is not possible.
1853          */
1854         if (!test_bit(Faulty, &rdev->flags) &&
1855             mddev->recovery_disabled != p->recovery_disabled &&
1856             (!p->replacement || p->replacement == rdev) &&
1857             number < conf->geo.raid_disks &&
1858             enough(conf, -1)) {
1859                 err = -EBUSY;
1860                 goto abort;
1861         }
1862         *rdevp = NULL;
1863         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1864                 synchronize_rcu();
1865                 if (atomic_read(&rdev->nr_pending)) {
1866                         /* lost the race, try later */
1867                         err = -EBUSY;
1868                         *rdevp = rdev;
1869                         goto abort;
1870                 }
1871         }
1872         if (p->replacement) {
1873                 /* We must have just cleared 'rdev' */
1874                 p->rdev = p->replacement;
1875                 clear_bit(Replacement, &p->replacement->flags);
1876                 smp_mb(); /* Make sure other CPUs may see both as identical
1877                            * but will never see neither -- if they are careful.
1878                            */
1879                 p->replacement = NULL;
1880         }
1881
1882         clear_bit(WantReplacement, &rdev->flags);
1883         err = md_integrity_register(mddev);
1884
1885 abort:
1886
1887         print_conf(conf);
1888         return err;
1889 }
1890
1891 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1892 {
1893         struct r10conf *conf = r10_bio->mddev->private;
1894
1895         if (!bio->bi_status)
1896                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1897         else
1898                 /* The write handler will notice the lack of
1899                  * R10BIO_Uptodate and record any errors etc
1900                  */
1901                 atomic_add(r10_bio->sectors,
1902                            &conf->mirrors[d].rdev->corrected_errors);
1903
1904         /* for reconstruct, we always reschedule after a read.
1905          * for resync, only after all reads
1906          */
1907         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1908         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1909             atomic_dec_and_test(&r10_bio->remaining)) {
1910                 /* we have read all the blocks,
1911                  * do the comparison in process context in raid10d
1912                  */
1913                 reschedule_retry(r10_bio);
1914         }
1915 }
1916
1917 static void end_sync_read(struct bio *bio)
1918 {
1919         struct r10bio *r10_bio = get_resync_r10bio(bio);
1920         struct r10conf *conf = r10_bio->mddev->private;
1921         int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1922
1923         __end_sync_read(r10_bio, bio, d);
1924 }
1925
1926 static void end_reshape_read(struct bio *bio)
1927 {
1928         /* reshape read bio isn't allocated from r10buf_pool */
1929         struct r10bio *r10_bio = bio->bi_private;
1930
1931         __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1932 }
1933
1934 static void end_sync_request(struct r10bio *r10_bio)
1935 {
1936         struct mddev *mddev = r10_bio->mddev;
1937
1938         while (atomic_dec_and_test(&r10_bio->remaining)) {
1939                 if (r10_bio->master_bio == NULL) {
1940                         /* the primary of several recovery bios */
1941                         sector_t s = r10_bio->sectors;
1942                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1943                             test_bit(R10BIO_WriteError, &r10_bio->state))
1944                                 reschedule_retry(r10_bio);
1945                         else
1946                                 put_buf(r10_bio);
1947                         md_done_sync(mddev, s, 1);
1948                         break;
1949                 } else {
1950                         struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1951                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1952                             test_bit(R10BIO_WriteError, &r10_bio->state))
1953                                 reschedule_retry(r10_bio);
1954                         else
1955                                 put_buf(r10_bio);
1956                         r10_bio = r10_bio2;
1957                 }
1958         }
1959 }
1960
1961 static void end_sync_write(struct bio *bio)
1962 {
1963         struct r10bio *r10_bio = get_resync_r10bio(bio);
1964         struct mddev *mddev = r10_bio->mddev;
1965         struct r10conf *conf = mddev->private;
1966         int d;
1967         sector_t first_bad;
1968         int bad_sectors;
1969         int slot;
1970         int repl;
1971         struct md_rdev *rdev = NULL;
1972
1973         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1974         if (repl)
1975                 rdev = conf->mirrors[d].replacement;
1976         else
1977                 rdev = conf->mirrors[d].rdev;
1978
1979         if (bio->bi_status) {
1980                 if (repl)
1981                         md_error(mddev, rdev);
1982                 else {
1983                         set_bit(WriteErrorSeen, &rdev->flags);
1984                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
1985                                 set_bit(MD_RECOVERY_NEEDED,
1986                                         &rdev->mddev->recovery);
1987                         set_bit(R10BIO_WriteError, &r10_bio->state);
1988                 }
1989         } else if (is_badblock(rdev,
1990                              r10_bio->devs[slot].addr,
1991                              r10_bio->sectors,
1992                              &first_bad, &bad_sectors))
1993                 set_bit(R10BIO_MadeGood, &r10_bio->state);
1994
1995         rdev_dec_pending(rdev, mddev);
1996
1997         end_sync_request(r10_bio);
1998 }
1999
2000 /*
2001  * Note: sync and recover and handled very differently for raid10
2002  * This code is for resync.
2003  * For resync, we read through virtual addresses and read all blocks.
2004  * If there is any error, we schedule a write.  The lowest numbered
2005  * drive is authoritative.
2006  * However requests come for physical address, so we need to map.
2007  * For every physical address there are raid_disks/copies virtual addresses,
2008  * which is always are least one, but is not necessarly an integer.
2009  * This means that a physical address can span multiple chunks, so we may
2010  * have to submit multiple io requests for a single sync request.
2011  */
2012 /*
2013  * We check if all blocks are in-sync and only write to blocks that
2014  * aren't in sync
2015  */
2016 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2017 {
2018         struct r10conf *conf = mddev->private;
2019         int i, first;
2020         struct bio *tbio, *fbio;
2021         int vcnt;
2022         struct page **tpages, **fpages;
2023
2024         atomic_set(&r10_bio->remaining, 1);
2025
2026         /* find the first device with a block */
2027         for (i=0; i<conf->copies; i++)
2028                 if (!r10_bio->devs[i].bio->bi_status)
2029                         break;
2030
2031         if (i == conf->copies)
2032                 goto done;
2033
2034         first = i;
2035         fbio = r10_bio->devs[i].bio;
2036         fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2037         fbio->bi_iter.bi_idx = 0;
2038         fpages = get_resync_pages(fbio)->pages;
2039
2040         vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2041         /* now find blocks with errors */
2042         for (i=0 ; i < conf->copies ; i++) {
2043                 int  j, d;
2044                 struct md_rdev *rdev;
2045                 struct resync_pages *rp;
2046
2047                 tbio = r10_bio->devs[i].bio;
2048
2049                 if (tbio->bi_end_io != end_sync_read)
2050                         continue;
2051                 if (i == first)
2052                         continue;
2053
2054                 tpages = get_resync_pages(tbio)->pages;
2055                 d = r10_bio->devs[i].devnum;
2056                 rdev = conf->mirrors[d].rdev;
2057                 if (!r10_bio->devs[i].bio->bi_status) {
2058                         /* We know that the bi_io_vec layout is the same for
2059                          * both 'first' and 'i', so we just compare them.
2060                          * All vec entries are PAGE_SIZE;
2061                          */
2062                         int sectors = r10_bio->sectors;
2063                         for (j = 0; j < vcnt; j++) {
2064                                 int len = PAGE_SIZE;
2065                                 if (sectors < (len / 512))
2066                                         len = sectors * 512;
2067                                 if (memcmp(page_address(fpages[j]),
2068                                            page_address(tpages[j]),
2069                                            len))
2070                                         break;
2071                                 sectors -= len/512;
2072                         }
2073                         if (j == vcnt)
2074                                 continue;
2075                         atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2076                         if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2077                                 /* Don't fix anything. */
2078                                 continue;
2079                 } else if (test_bit(FailFast, &rdev->flags)) {
2080                         /* Just give up on this device */
2081                         md_error(rdev->mddev, rdev);
2082                         continue;
2083                 }
2084                 /* Ok, we need to write this bio, either to correct an
2085                  * inconsistency or to correct an unreadable block.
2086                  * First we need to fixup bv_offset, bv_len and
2087                  * bi_vecs, as the read request might have corrupted these
2088                  */
2089                 rp = get_resync_pages(tbio);
2090                 bio_reset(tbio);
2091
2092                 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2093
2094                 rp->raid_bio = r10_bio;
2095                 tbio->bi_private = rp;
2096                 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2097                 tbio->bi_end_io = end_sync_write;
2098                 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2099
2100                 bio_copy_data(tbio, fbio);
2101
2102                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2103                 atomic_inc(&r10_bio->remaining);
2104                 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2105
2106                 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2107                         tbio->bi_opf |= MD_FAILFAST;
2108                 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2109                 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2110                 generic_make_request(tbio);
2111         }
2112
2113         /* Now write out to any replacement devices
2114          * that are active
2115          */
2116         for (i = 0; i < conf->copies; i++) {
2117                 int d;
2118
2119                 tbio = r10_bio->devs[i].repl_bio;
2120                 if (!tbio || !tbio->bi_end_io)
2121                         continue;
2122                 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2123                     && r10_bio->devs[i].bio != fbio)
2124                         bio_copy_data(tbio, fbio);
2125                 d = r10_bio->devs[i].devnum;
2126                 atomic_inc(&r10_bio->remaining);
2127                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2128                              bio_sectors(tbio));
2129                 generic_make_request(tbio);
2130         }
2131
2132 done:
2133         if (atomic_dec_and_test(&r10_bio->remaining)) {
2134                 md_done_sync(mddev, r10_bio->sectors, 1);
2135                 put_buf(r10_bio);
2136         }
2137 }
2138
2139 /*
2140  * Now for the recovery code.
2141  * Recovery happens across physical sectors.
2142  * We recover all non-is_sync drives by finding the virtual address of
2143  * each, and then choose a working drive that also has that virt address.
2144  * There is a separate r10_bio for each non-in_sync drive.
2145  * Only the first two slots are in use. The first for reading,
2146  * The second for writing.
2147  *
2148  */
2149 static void fix_recovery_read_error(struct r10bio *r10_bio)
2150 {
2151         /* We got a read error during recovery.
2152          * We repeat the read in smaller page-sized sections.
2153          * If a read succeeds, write it to the new device or record
2154          * a bad block if we cannot.
2155          * If a read fails, record a bad block on both old and
2156          * new devices.
2157          */
2158         struct mddev *mddev = r10_bio->mddev;
2159         struct r10conf *conf = mddev->private;
2160         struct bio *bio = r10_bio->devs[0].bio;
2161         sector_t sect = 0;
2162         int sectors = r10_bio->sectors;
2163         int idx = 0;
2164         int dr = r10_bio->devs[0].devnum;
2165         int dw = r10_bio->devs[1].devnum;
2166         struct page **pages = get_resync_pages(bio)->pages;
2167
2168         while (sectors) {
2169                 int s = sectors;
2170                 struct md_rdev *rdev;
2171                 sector_t addr;
2172                 int ok;
2173
2174                 if (s > (PAGE_SIZE>>9))
2175                         s = PAGE_SIZE >> 9;
2176
2177                 rdev = conf->mirrors[dr].rdev;
2178                 addr = r10_bio->devs[0].addr + sect,
2179                 ok = sync_page_io(rdev,
2180                                   addr,
2181                                   s << 9,
2182                                   pages[idx],
2183                                   REQ_OP_READ, 0, false);
2184                 if (ok) {
2185                         rdev = conf->mirrors[dw].rdev;
2186                         addr = r10_bio->devs[1].addr + sect;
2187                         ok = sync_page_io(rdev,
2188                                           addr,
2189                                           s << 9,
2190                                           pages[idx],
2191                                           REQ_OP_WRITE, 0, false);
2192                         if (!ok) {
2193                                 set_bit(WriteErrorSeen, &rdev->flags);
2194                                 if (!test_and_set_bit(WantReplacement,
2195                                                       &rdev->flags))
2196                                         set_bit(MD_RECOVERY_NEEDED,
2197                                                 &rdev->mddev->recovery);
2198                         }
2199                 }
2200                 if (!ok) {
2201                         /* We don't worry if we cannot set a bad block -
2202                          * it really is bad so there is no loss in not
2203                          * recording it yet
2204                          */
2205                         rdev_set_badblocks(rdev, addr, s, 0);
2206
2207                         if (rdev != conf->mirrors[dw].rdev) {
2208                                 /* need bad block on destination too */
2209                                 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2210                                 addr = r10_bio->devs[1].addr + sect;
2211                                 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2212                                 if (!ok) {
2213                                         /* just abort the recovery */
2214                                         pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2215                                                   mdname(mddev));
2216
2217                                         conf->mirrors[dw].recovery_disabled
2218                                                 = mddev->recovery_disabled;
2219                                         set_bit(MD_RECOVERY_INTR,
2220                                                 &mddev->recovery);
2221                                         break;
2222                                 }
2223                         }
2224                 }
2225
2226                 sectors -= s;
2227                 sect += s;
2228                 idx++;
2229         }
2230 }
2231
2232 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2233 {
2234         struct r10conf *conf = mddev->private;
2235         int d;
2236         struct bio *wbio, *wbio2;
2237
2238         if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2239                 fix_recovery_read_error(r10_bio);
2240                 end_sync_request(r10_bio);
2241                 return;
2242         }
2243
2244         /*
2245          * share the pages with the first bio
2246          * and submit the write request
2247          */
2248         d = r10_bio->devs[1].devnum;
2249         wbio = r10_bio->devs[1].bio;
2250         wbio2 = r10_bio->devs[1].repl_bio;
2251         /* Need to test wbio2->bi_end_io before we call
2252          * generic_make_request as if the former is NULL,
2253          * the latter is free to free wbio2.
2254          */
2255         if (wbio2 && !wbio2->bi_end_io)
2256                 wbio2 = NULL;
2257         if (wbio->bi_end_io) {
2258                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2259                 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2260                 generic_make_request(wbio);
2261         }
2262         if (wbio2) {
2263                 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2264                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2265                              bio_sectors(wbio2));
2266                 generic_make_request(wbio2);
2267         }
2268 }
2269
2270 /*
2271  * Used by fix_read_error() to decay the per rdev read_errors.
2272  * We halve the read error count for every hour that has elapsed
2273  * since the last recorded read error.
2274  *
2275  */
2276 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2277 {
2278         long cur_time_mon;
2279         unsigned long hours_since_last;
2280         unsigned int read_errors = atomic_read(&rdev->read_errors);
2281
2282         cur_time_mon = ktime_get_seconds();
2283
2284         if (rdev->last_read_error == 0) {
2285                 /* first time we've seen a read error */
2286                 rdev->last_read_error = cur_time_mon;
2287                 return;
2288         }
2289
2290         hours_since_last = (long)(cur_time_mon -
2291                             rdev->last_read_error) / 3600;
2292
2293         rdev->last_read_error = cur_time_mon;
2294
2295         /*
2296          * if hours_since_last is > the number of bits in read_errors
2297          * just set read errors to 0. We do this to avoid
2298          * overflowing the shift of read_errors by hours_since_last.
2299          */
2300         if (hours_since_last >= 8 * sizeof(read_errors))
2301                 atomic_set(&rdev->read_errors, 0);
2302         else
2303                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2304 }
2305
2306 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2307                             int sectors, struct page *page, int rw)
2308 {
2309         sector_t first_bad;
2310         int bad_sectors;
2311
2312         if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2313             && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2314                 return -1;
2315         if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2316                 /* success */
2317                 return 1;
2318         if (rw == WRITE) {
2319                 set_bit(WriteErrorSeen, &rdev->flags);
2320                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2321                         set_bit(MD_RECOVERY_NEEDED,
2322                                 &rdev->mddev->recovery);
2323         }
2324         /* need to record an error - either for the block or the device */
2325         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2326                 md_error(rdev->mddev, rdev);
2327         return 0;
2328 }
2329
2330 /*
2331  * This is a kernel thread which:
2332  *
2333  *      1.      Retries failed read operations on working mirrors.
2334  *      2.      Updates the raid superblock when problems encounter.
2335  *      3.      Performs writes following reads for array synchronising.
2336  */
2337
2338 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2339 {
2340         int sect = 0; /* Offset from r10_bio->sector */
2341         int sectors = r10_bio->sectors;
2342         struct md_rdev *rdev;
2343         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2344         int d = r10_bio->devs[r10_bio->read_slot].devnum;
2345
2346         /* still own a reference to this rdev, so it cannot
2347          * have been cleared recently.
2348          */
2349         rdev = conf->mirrors[d].rdev;
2350
2351         if (test_bit(Faulty, &rdev->flags))
2352                 /* drive has already been failed, just ignore any
2353                    more fix_read_error() attempts */
2354                 return;
2355
2356         check_decay_read_errors(mddev, rdev);
2357         atomic_inc(&rdev->read_errors);
2358         if (atomic_read(&rdev->read_errors) > max_read_errors) {
2359                 char b[BDEVNAME_SIZE];
2360                 bdevname(rdev->bdev, b);
2361
2362                 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2363                           mdname(mddev), b,
2364                           atomic_read(&rdev->read_errors), max_read_errors);
2365                 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2366                           mdname(mddev), b);
2367                 md_error(mddev, rdev);
2368                 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2369                 return;
2370         }
2371
2372         while(sectors) {
2373                 int s = sectors;
2374                 int sl = r10_bio->read_slot;
2375                 int success = 0;
2376                 int start;
2377
2378                 if (s > (PAGE_SIZE>>9))
2379                         s = PAGE_SIZE >> 9;
2380
2381                 rcu_read_lock();
2382                 do {
2383                         sector_t first_bad;
2384                         int bad_sectors;
2385
2386                         d = r10_bio->devs[sl].devnum;
2387                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2388                         if (rdev &&
2389                             test_bit(In_sync, &rdev->flags) &&
2390                             !test_bit(Faulty, &rdev->flags) &&
2391                             is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2392                                         &first_bad, &bad_sectors) == 0) {
2393                                 atomic_inc(&rdev->nr_pending);
2394                                 rcu_read_unlock();
2395                                 success = sync_page_io(rdev,
2396                                                        r10_bio->devs[sl].addr +
2397                                                        sect,
2398                                                        s<<9,
2399                                                        conf->tmppage,
2400                                                        REQ_OP_READ, 0, false);
2401                                 rdev_dec_pending(rdev, mddev);
2402                                 rcu_read_lock();
2403                                 if (success)
2404                                         break;
2405                         }
2406                         sl++;
2407                         if (sl == conf->copies)
2408                                 sl = 0;
2409                 } while (!success && sl != r10_bio->read_slot);
2410                 rcu_read_unlock();
2411
2412                 if (!success) {
2413                         /* Cannot read from anywhere, just mark the block
2414                          * as bad on the first device to discourage future
2415                          * reads.
2416                          */
2417                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2418                         rdev = conf->mirrors[dn].rdev;
2419
2420                         if (!rdev_set_badblocks(
2421                                     rdev,
2422                                     r10_bio->devs[r10_bio->read_slot].addr
2423                                     + sect,
2424                                     s, 0)) {
2425                                 md_error(mddev, rdev);
2426                                 r10_bio->devs[r10_bio->read_slot].bio
2427                                         = IO_BLOCKED;
2428                         }
2429                         break;
2430                 }
2431
2432                 start = sl;
2433                 /* write it back and re-read */
2434                 rcu_read_lock();
2435                 while (sl != r10_bio->read_slot) {
2436                         char b[BDEVNAME_SIZE];
2437
2438                         if (sl==0)
2439                                 sl = conf->copies;
2440                         sl--;
2441                         d = r10_bio->devs[sl].devnum;
2442                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2443                         if (!rdev ||
2444                             test_bit(Faulty, &rdev->flags) ||
2445                             !test_bit(In_sync, &rdev->flags))
2446                                 continue;
2447
2448                         atomic_inc(&rdev->nr_pending);
2449                         rcu_read_unlock();
2450                         if (r10_sync_page_io(rdev,
2451                                              r10_bio->devs[sl].addr +
2452                                              sect,
2453                                              s, conf->tmppage, WRITE)
2454                             == 0) {
2455                                 /* Well, this device is dead */
2456                                 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2457                                           mdname(mddev), s,
2458                                           (unsigned long long)(
2459                                                   sect +
2460                                                   choose_data_offset(r10_bio,
2461                                                                      rdev)),
2462                                           bdevname(rdev->bdev, b));
2463                                 pr_notice("md/raid10:%s: %s: failing drive\n",
2464                                           mdname(mddev),
2465                                           bdevname(rdev->bdev, b));
2466                         }
2467                         rdev_dec_pending(rdev, mddev);
2468                         rcu_read_lock();
2469                 }
2470                 sl = start;
2471                 while (sl != r10_bio->read_slot) {
2472                         char b[BDEVNAME_SIZE];
2473
2474                         if (sl==0)
2475                                 sl = conf->copies;
2476                         sl--;
2477                         d = r10_bio->devs[sl].devnum;
2478                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2479                         if (!rdev ||
2480                             test_bit(Faulty, &rdev->flags) ||
2481                             !test_bit(In_sync, &rdev->flags))
2482                                 continue;
2483
2484                         atomic_inc(&rdev->nr_pending);
2485                         rcu_read_unlock();
2486                         switch (r10_sync_page_io(rdev,
2487                                              r10_bio->devs[sl].addr +
2488                                              sect,
2489                                              s, conf->tmppage,
2490                                                  READ)) {
2491                         case 0:
2492                                 /* Well, this device is dead */
2493                                 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2494                                        mdname(mddev), s,
2495                                        (unsigned long long)(
2496                                                sect +
2497                                                choose_data_offset(r10_bio, rdev)),
2498                                        bdevname(rdev->bdev, b));
2499                                 pr_notice("md/raid10:%s: %s: failing drive\n",
2500                                        mdname(mddev),
2501                                        bdevname(rdev->bdev, b));
2502                                 break;
2503                         case 1:
2504                                 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2505                                        mdname(mddev), s,
2506                                        (unsigned long long)(
2507                                                sect +
2508                                                choose_data_offset(r10_bio, rdev)),
2509                                        bdevname(rdev->bdev, b));
2510                                 atomic_add(s, &rdev->corrected_errors);
2511                         }
2512
2513                         rdev_dec_pending(rdev, mddev);
2514                         rcu_read_lock();
2515                 }
2516                 rcu_read_unlock();
2517
2518                 sectors -= s;
2519                 sect += s;
2520         }
2521 }
2522
2523 static int narrow_write_error(struct r10bio *r10_bio, int i)
2524 {
2525         struct bio *bio = r10_bio->master_bio;
2526         struct mddev *mddev = r10_bio->mddev;
2527         struct r10conf *conf = mddev->private;
2528         struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2529         /* bio has the data to be written to slot 'i' where
2530          * we just recently had a write error.
2531          * We repeatedly clone the bio and trim down to one block,
2532          * then try the write.  Where the write fails we record
2533          * a bad block.
2534          * It is conceivable that the bio doesn't exactly align with
2535          * blocks.  We must handle this.
2536          *
2537          * We currently own a reference to the rdev.
2538          */
2539
2540         int block_sectors;
2541         sector_t sector;
2542         int sectors;
2543         int sect_to_write = r10_bio->sectors;
2544         int ok = 1;
2545
2546         if (rdev->badblocks.shift < 0)
2547                 return 0;
2548
2549         block_sectors = roundup(1 << rdev->badblocks.shift,
2550                                 bdev_logical_block_size(rdev->bdev) >> 9);
2551         sector = r10_bio->sector;
2552         sectors = ((r10_bio->sector + block_sectors)
2553                    & ~(sector_t)(block_sectors - 1))
2554                 - sector;
2555
2556         while (sect_to_write) {
2557                 struct bio *wbio;
2558                 sector_t wsector;
2559                 if (sectors > sect_to_write)
2560                         sectors = sect_to_write;
2561                 /* Write at 'sector' for 'sectors' */
2562                 wbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
2563                 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2564                 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2565                 wbio->bi_iter.bi_sector = wsector +
2566                                    choose_data_offset(r10_bio, rdev);
2567                 bio_set_dev(wbio, rdev->bdev);
2568                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2569
2570                 if (submit_bio_wait(wbio) < 0)
2571                         /* Failure! */
2572                         ok = rdev_set_badblocks(rdev, wsector,
2573                                                 sectors, 0)
2574                                 && ok;
2575
2576                 bio_put(wbio);
2577                 sect_to_write -= sectors;
2578                 sector += sectors;
2579                 sectors = block_sectors;
2580         }
2581         return ok;
2582 }
2583
2584 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2585 {
2586         int slot = r10_bio->read_slot;
2587         struct bio *bio;
2588         struct r10conf *conf = mddev->private;
2589         struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2590
2591         /* we got a read error. Maybe the drive is bad.  Maybe just
2592          * the block and we can fix it.
2593          * We freeze all other IO, and try reading the block from
2594          * other devices.  When we find one, we re-write
2595          * and check it that fixes the read error.
2596          * This is all done synchronously while the array is
2597          * frozen.
2598          */
2599         bio = r10_bio->devs[slot].bio;
2600         bio_put(bio);
2601         r10_bio->devs[slot].bio = NULL;
2602
2603         if (mddev->ro)
2604                 r10_bio->devs[slot].bio = IO_BLOCKED;
2605         else if (!test_bit(FailFast, &rdev->flags)) {
2606                 freeze_array(conf, 1);
2607                 fix_read_error(conf, mddev, r10_bio);
2608                 unfreeze_array(conf);
2609         } else
2610                 md_error(mddev, rdev);
2611
2612         rdev_dec_pending(rdev, mddev);
2613         allow_barrier(conf);
2614         r10_bio->state = 0;
2615         raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2616 }
2617
2618 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2619 {
2620         /* Some sort of write request has finished and it
2621          * succeeded in writing where we thought there was a
2622          * bad block.  So forget the bad block.
2623          * Or possibly if failed and we need to record
2624          * a bad block.
2625          */
2626         int m;
2627         struct md_rdev *rdev;
2628
2629         if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2630             test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2631                 for (m = 0; m < conf->copies; m++) {
2632                         int dev = r10_bio->devs[m].devnum;
2633                         rdev = conf->mirrors[dev].rdev;
2634                         if (r10_bio->devs[m].bio == NULL ||
2635                                 r10_bio->devs[m].bio->bi_end_io == NULL)
2636                                 continue;
2637                         if (!r10_bio->devs[m].bio->bi_status) {
2638                                 rdev_clear_badblocks(
2639                                         rdev,
2640                                         r10_bio->devs[m].addr,
2641                                         r10_bio->sectors, 0);
2642                         } else {
2643                                 if (!rdev_set_badblocks(
2644                                             rdev,
2645                                             r10_bio->devs[m].addr,
2646                                             r10_bio->sectors, 0))
2647                                         md_error(conf->mddev, rdev);
2648                         }
2649                         rdev = conf->mirrors[dev].replacement;
2650                         if (r10_bio->devs[m].repl_bio == NULL ||
2651                                 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2652                                 continue;
2653
2654                         if (!r10_bio->devs[m].repl_bio->bi_status) {
2655                                 rdev_clear_badblocks(
2656                                         rdev,
2657                                         r10_bio->devs[m].addr,
2658                                         r10_bio->sectors, 0);
2659                         } else {
2660                                 if (!rdev_set_badblocks(
2661                                             rdev,
2662                                             r10_bio->devs[m].addr,
2663                                             r10_bio->sectors, 0))
2664                                         md_error(conf->mddev, rdev);
2665                         }
2666                 }
2667                 put_buf(r10_bio);
2668         } else {
2669                 bool fail = false;
2670                 for (m = 0; m < conf->copies; m++) {
2671                         int dev = r10_bio->devs[m].devnum;
2672                         struct bio *bio = r10_bio->devs[m].bio;
2673                         rdev = conf->mirrors[dev].rdev;
2674                         if (bio == IO_MADE_GOOD) {
2675                                 rdev_clear_badblocks(
2676                                         rdev,
2677                                         r10_bio->devs[m].addr,
2678                                         r10_bio->sectors, 0);
2679                                 rdev_dec_pending(rdev, conf->mddev);
2680                         } else if (bio != NULL && bio->bi_status) {
2681                                 fail = true;
2682                                 if (!narrow_write_error(r10_bio, m)) {
2683                                         md_error(conf->mddev, rdev);
2684                                         set_bit(R10BIO_Degraded,
2685                                                 &r10_bio->state);
2686                                 }
2687                                 rdev_dec_pending(rdev, conf->mddev);
2688                         }
2689                         bio = r10_bio->devs[m].repl_bio;
2690                         rdev = conf->mirrors[dev].replacement;
2691                         if (rdev && bio == IO_MADE_GOOD) {
2692                                 rdev_clear_badblocks(
2693                                         rdev,
2694                                         r10_bio->devs[m].addr,
2695                                         r10_bio->sectors, 0);
2696                                 rdev_dec_pending(rdev, conf->mddev);
2697                         }
2698                 }
2699                 if (fail) {
2700                         spin_lock_irq(&conf->device_lock);
2701                         list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2702                         conf->nr_queued++;
2703                         spin_unlock_irq(&conf->device_lock);
2704                         /*
2705                          * In case freeze_array() is waiting for condition
2706                          * nr_pending == nr_queued + extra to be true.
2707                          */
2708                         wake_up(&conf->wait_barrier);
2709                         md_wakeup_thread(conf->mddev->thread);
2710                 } else {
2711                         if (test_bit(R10BIO_WriteError,
2712                                      &r10_bio->state))
2713                                 close_write(r10_bio);
2714                         raid_end_bio_io(r10_bio);
2715                 }
2716         }
2717 }
2718
2719 static void raid10d(struct md_thread *thread)
2720 {
2721         struct mddev *mddev = thread->mddev;
2722         struct r10bio *r10_bio;
2723         unsigned long flags;
2724         struct r10conf *conf = mddev->private;
2725         struct list_head *head = &conf->retry_list;
2726         struct blk_plug plug;
2727
2728         md_check_recovery(mddev);
2729
2730         if (!list_empty_careful(&conf->bio_end_io_list) &&
2731             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2732                 LIST_HEAD(tmp);
2733                 spin_lock_irqsave(&conf->device_lock, flags);
2734                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2735                         while (!list_empty(&conf->bio_end_io_list)) {
2736                                 list_move(conf->bio_end_io_list.prev, &tmp);
2737                                 conf->nr_queued--;
2738                         }
2739                 }
2740                 spin_unlock_irqrestore(&conf->device_lock, flags);
2741                 while (!list_empty(&tmp)) {
2742                         r10_bio = list_first_entry(&tmp, struct r10bio,
2743                                                    retry_list);
2744                         list_del(&r10_bio->retry_list);
2745                         if (mddev->degraded)
2746                                 set_bit(R10BIO_Degraded, &r10_bio->state);
2747
2748                         if (test_bit(R10BIO_WriteError,
2749                                      &r10_bio->state))
2750                                 close_write(r10_bio);
2751                         raid_end_bio_io(r10_bio);
2752                 }
2753         }
2754
2755         blk_start_plug(&plug);
2756         for (;;) {
2757
2758                 flush_pending_writes(conf);
2759
2760                 spin_lock_irqsave(&conf->device_lock, flags);
2761                 if (list_empty(head)) {
2762                         spin_unlock_irqrestore(&conf->device_lock, flags);
2763                         break;
2764                 }
2765                 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2766                 list_del(head->prev);
2767                 conf->nr_queued--;
2768                 spin_unlock_irqrestore(&conf->device_lock, flags);
2769
2770                 mddev = r10_bio->mddev;
2771                 conf = mddev->private;
2772                 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2773                     test_bit(R10BIO_WriteError, &r10_bio->state))
2774                         handle_write_completed(conf, r10_bio);
2775                 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2776                         reshape_request_write(mddev, r10_bio);
2777                 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2778                         sync_request_write(mddev, r10_bio);
2779                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2780                         recovery_request_write(mddev, r10_bio);
2781                 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2782                         handle_read_error(mddev, r10_bio);
2783                 else
2784                         WARN_ON_ONCE(1);
2785
2786                 cond_resched();
2787                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2788                         md_check_recovery(mddev);
2789         }
2790         blk_finish_plug(&plug);
2791 }
2792
2793 static int init_resync(struct r10conf *conf)
2794 {
2795         int ret, buffs, i;
2796
2797         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2798         BUG_ON(mempool_initialized(&conf->r10buf_pool));
2799         conf->have_replacement = 0;
2800         for (i = 0; i < conf->geo.raid_disks; i++)
2801                 if (conf->mirrors[i].replacement)
2802                         conf->have_replacement = 1;
2803         ret = mempool_init(&conf->r10buf_pool, buffs,
2804                            r10buf_pool_alloc, r10buf_pool_free, conf);
2805         if (ret)
2806                 return ret;
2807         conf->next_resync = 0;
2808         return 0;
2809 }
2810
2811 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2812 {
2813         struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
2814         struct rsync_pages *rp;
2815         struct bio *bio;
2816         int nalloc;
2817         int i;
2818
2819         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2820             test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2821                 nalloc = conf->copies; /* resync */
2822         else
2823                 nalloc = 2; /* recovery */
2824
2825         for (i = 0; i < nalloc; i++) {
2826                 bio = r10bio->devs[i].bio;
2827                 rp = bio->bi_private;
2828                 bio_reset(bio);
2829                 bio->bi_private = rp;
2830                 bio = r10bio->devs[i].repl_bio;
2831                 if (bio) {
2832                         rp = bio->bi_private;
2833                         bio_reset(bio);
2834                         bio->bi_private = rp;
2835                 }
2836         }
2837         return r10bio;
2838 }
2839
2840 /*
2841  * Set cluster_sync_high since we need other nodes to add the
2842  * range [cluster_sync_low, cluster_sync_high] to suspend list.
2843  */
2844 static void raid10_set_cluster_sync_high(struct r10conf *conf)
2845 {
2846         sector_t window_size;
2847         int extra_chunk, chunks;
2848
2849         /*
2850          * First, here we define "stripe" as a unit which across
2851          * all member devices one time, so we get chunks by use
2852          * raid_disks / near_copies. Otherwise, if near_copies is
2853          * close to raid_disks, then resync window could increases
2854          * linearly with the increase of raid_disks, which means
2855          * we will suspend a really large IO window while it is not
2856          * necessary. If raid_disks is not divisible by near_copies,
2857          * an extra chunk is needed to ensure the whole "stripe" is
2858          * covered.
2859          */
2860
2861         chunks = conf->geo.raid_disks / conf->geo.near_copies;
2862         if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2863                 extra_chunk = 0;
2864         else
2865                 extra_chunk = 1;
2866         window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2867
2868         /*
2869          * At least use a 32M window to align with raid1's resync window
2870          */
2871         window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2872                         CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2873
2874         conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2875 }
2876
2877 /*
2878  * perform a "sync" on one "block"
2879  *
2880  * We need to make sure that no normal I/O request - particularly write
2881  * requests - conflict with active sync requests.
2882  *
2883  * This is achieved by tracking pending requests and a 'barrier' concept
2884  * that can be installed to exclude normal IO requests.
2885  *
2886  * Resync and recovery are handled very differently.
2887  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2888  *
2889  * For resync, we iterate over virtual addresses, read all copies,
2890  * and update if there are differences.  If only one copy is live,
2891  * skip it.
2892  * For recovery, we iterate over physical addresses, read a good
2893  * value for each non-in_sync drive, and over-write.
2894  *
2895  * So, for recovery we may have several outstanding complex requests for a
2896  * given address, one for each out-of-sync device.  We model this by allocating
2897  * a number of r10_bio structures, one for each out-of-sync device.
2898  * As we setup these structures, we collect all bio's together into a list
2899  * which we then process collectively to add pages, and then process again
2900  * to pass to generic_make_request.
2901  *
2902  * The r10_bio structures are linked using a borrowed master_bio pointer.
2903  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2904  * has its remaining count decremented to 0, the whole complex operation
2905  * is complete.
2906  *
2907  */
2908
2909 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2910                              int *skipped)
2911 {
2912         struct r10conf *conf = mddev->private;
2913         struct r10bio *r10_bio;
2914         struct bio *biolist = NULL, *bio;
2915         sector_t max_sector, nr_sectors;
2916         int i;
2917         int max_sync;
2918         sector_t sync_blocks;
2919         sector_t sectors_skipped = 0;
2920         int chunks_skipped = 0;
2921         sector_t chunk_mask = conf->geo.chunk_mask;
2922         int page_idx = 0;
2923
2924         if (!mempool_initialized(&conf->r10buf_pool))
2925                 if (init_resync(conf))
2926                         return 0;
2927
2928         /*
2929          * Allow skipping a full rebuild for incremental assembly
2930          * of a clean array, like RAID1 does.
2931          */
2932         if (mddev->bitmap == NULL &&
2933             mddev->recovery_cp == MaxSector &&
2934             mddev->reshape_position == MaxSector &&
2935             !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2936             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2937             !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2938             conf->fullsync == 0) {
2939                 *skipped = 1;
2940                 return mddev->dev_sectors - sector_nr;
2941         }
2942
2943  skipped:
2944         max_sector = mddev->dev_sectors;
2945         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2946             test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2947                 max_sector = mddev->resync_max_sectors;
2948         if (sector_nr >= max_sector) {
2949                 conf->cluster_sync_low = 0;
2950                 conf->cluster_sync_high = 0;
2951
2952                 /* If we aborted, we need to abort the
2953                  * sync on the 'current' bitmap chucks (there can
2954                  * be several when recovering multiple devices).
2955                  * as we may have started syncing it but not finished.
2956                  * We can find the current address in
2957                  * mddev->curr_resync, but for recovery,
2958                  * we need to convert that to several
2959                  * virtual addresses.
2960                  */
2961                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2962                         end_reshape(conf);
2963                         close_sync(conf);
2964                         return 0;
2965                 }
2966
2967                 if (mddev->curr_resync < max_sector) { /* aborted */
2968                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2969                                 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2970                                                    &sync_blocks, 1);
2971                         else for (i = 0; i < conf->geo.raid_disks; i++) {
2972                                 sector_t sect =
2973                                         raid10_find_virt(conf, mddev->curr_resync, i);
2974                                 md_bitmap_end_sync(mddev->bitmap, sect,
2975                                                    &sync_blocks, 1);
2976                         }
2977                 } else {
2978                         /* completed sync */
2979                         if ((!mddev->bitmap || conf->fullsync)
2980                             && conf->have_replacement
2981                             && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2982                                 /* Completed a full sync so the replacements
2983                                  * are now fully recovered.
2984                                  */
2985                                 rcu_read_lock();
2986                                 for (i = 0; i < conf->geo.raid_disks; i++) {
2987                                         struct md_rdev *rdev =
2988                                                 rcu_dereference(conf->mirrors[i].replacement);
2989                                         if (rdev)
2990                                                 rdev->recovery_offset = MaxSector;
2991                                 }
2992                                 rcu_read_unlock();
2993                         }
2994                         conf->fullsync = 0;
2995                 }
2996                 md_bitmap_close_sync(mddev->bitmap);
2997                 close_sync(conf);
2998                 *skipped = 1;
2999                 return sectors_skipped;
3000         }
3001
3002         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3003                 return reshape_request(mddev, sector_nr, skipped);
3004
3005         if (chunks_skipped >= conf->geo.raid_disks) {
3006                 /* if there has been nothing to do on any drive,
3007                  * then there is nothing to do at all..
3008                  */
3009                 *skipped = 1;
3010                 return (max_sector - sector_nr) + sectors_skipped;
3011         }
3012
3013         if (max_sector > mddev->resync_max)
3014                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3015
3016         /* make sure whole request will fit in a chunk - if chunks
3017          * are meaningful
3018          */
3019         if (conf->geo.near_copies < conf->geo.raid_disks &&
3020             max_sector > (sector_nr | chunk_mask))
3021                 max_sector = (sector_nr | chunk_mask) + 1;
3022
3023         /*
3024          * If there is non-resync activity waiting for a turn, then let it
3025          * though before starting on this new sync request.
3026          */
3027         if (conf->nr_waiting)
3028                 schedule_timeout_uninterruptible(1);
3029
3030         /* Again, very different code for resync and recovery.
3031          * Both must result in an r10bio with a list of bios that
3032          * have bi_end_io, bi_sector, bi_disk set,
3033          * and bi_private set to the r10bio.
3034          * For recovery, we may actually create several r10bios
3035          * with 2 bios in each, that correspond to the bios in the main one.
3036          * In this case, the subordinate r10bios link back through a
3037          * borrowed master_bio pointer, and the counter in the master
3038          * includes a ref from each subordinate.
3039          */
3040         /* First, we decide what to do and set ->bi_end_io
3041          * To end_sync_read if we want to read, and
3042          * end_sync_write if we will want to write.
3043          */
3044
3045         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3046         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3047                 /* recovery... the complicated one */
3048                 int j;
3049                 r10_bio = NULL;
3050
3051                 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3052                         int still_degraded;
3053                         struct r10bio *rb2;
3054                         sector_t sect;
3055                         int must_sync;
3056                         int any_working;
3057                         int need_recover = 0;
3058                         int need_replace = 0;
3059                         struct raid10_info *mirror = &conf->mirrors[i];
3060                         struct md_rdev *mrdev, *mreplace;
3061
3062                         rcu_read_lock();
3063                         mrdev = rcu_dereference(mirror->rdev);
3064                         mreplace = rcu_dereference(mirror->replacement);
3065
3066                         if (mrdev != NULL &&
3067                             !test_bit(Faulty, &mrdev->flags) &&
3068                             !test_bit(In_sync, &mrdev->flags))
3069                                 need_recover = 1;
3070                         if (mreplace != NULL &&
3071                             !test_bit(Faulty, &mreplace->flags))
3072                                 need_replace = 1;
3073
3074                         if (!need_recover && !need_replace) {
3075                                 rcu_read_unlock();
3076                                 continue;
3077                         }
3078
3079                         still_degraded = 0;
3080                         /* want to reconstruct this device */
3081                         rb2 = r10_bio;
3082                         sect = raid10_find_virt(conf, sector_nr, i);
3083                         if (sect >= mddev->resync_max_sectors) {
3084                                 /* last stripe is not complete - don't
3085                                  * try to recover this sector.
3086                                  */
3087                                 rcu_read_unlock();
3088                                 continue;
3089                         }
3090                         if (mreplace && test_bit(Faulty, &mreplace->flags))
3091                                 mreplace = NULL;
3092                         /* Unless we are doing a full sync, or a replacement
3093                          * we only need to recover the block if it is set in
3094                          * the bitmap
3095                          */
3096                         must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3097                                                          &sync_blocks, 1);
3098                         if (sync_blocks < max_sync)
3099                                 max_sync = sync_blocks;
3100                         if (!must_sync &&
3101                             mreplace == NULL &&
3102                             !conf->fullsync) {
3103                                 /* yep, skip the sync_blocks here, but don't assume
3104                                  * that there will never be anything to do here
3105                                  */
3106                                 chunks_skipped = -1;
3107                                 rcu_read_unlock();
3108                                 continue;
3109                         }
3110                         atomic_inc(&mrdev->nr_pending);
3111                         if (mreplace)
3112                                 atomic_inc(&mreplace->nr_pending);
3113                         rcu_read_unlock();
3114
3115                         r10_bio = raid10_alloc_init_r10buf(conf);
3116                         r10_bio->state = 0;
3117                         raise_barrier(conf, rb2 != NULL);
3118                         atomic_set(&r10_bio->remaining, 0);
3119
3120                         r10_bio->master_bio = (struct bio*)rb2;
3121                         if (rb2)
3122                                 atomic_inc(&rb2->remaining);
3123                         r10_bio->mddev = mddev;
3124                         set_bit(R10BIO_IsRecover, &r10_bio->state);
3125                         r10_bio->sector = sect;
3126
3127                         raid10_find_phys(conf, r10_bio);
3128
3129                         /* Need to check if the array will still be
3130                          * degraded
3131                          */
3132                         rcu_read_lock();
3133                         for (j = 0; j < conf->geo.raid_disks; j++) {
3134                                 struct md_rdev *rdev = rcu_dereference(
3135                                         conf->mirrors[j].rdev);
3136                                 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3137                                         still_degraded = 1;
3138                                         break;
3139                                 }
3140                         }
3141
3142                         must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3143                                                          &sync_blocks, still_degraded);
3144
3145                         any_working = 0;
3146                         for (j=0; j<conf->copies;j++) {
3147                                 int k;
3148                                 int d = r10_bio->devs[j].devnum;
3149                                 sector_t from_addr, to_addr;
3150                                 struct md_rdev *rdev =
3151                                         rcu_dereference(conf->mirrors[d].rdev);
3152                                 sector_t sector, first_bad;
3153                                 int bad_sectors;
3154                                 if (!rdev ||
3155                                     !test_bit(In_sync, &rdev->flags))
3156                                         continue;
3157                                 /* This is where we read from */
3158                                 any_working = 1;
3159                                 sector = r10_bio->devs[j].addr;
3160
3161                                 if (is_badblock(rdev, sector, max_sync,
3162                                                 &first_bad, &bad_sectors)) {
3163                                         if (first_bad > sector)
3164                                                 max_sync = first_bad - sector;
3165                                         else {
3166                                                 bad_sectors -= (sector
3167                                                                 - first_bad);
3168                                                 if (max_sync > bad_sectors)
3169                                                         max_sync = bad_sectors;
3170                                                 continue;
3171                                         }
3172                                 }
3173                                 bio = r10_bio->devs[0].bio;
3174                                 bio->bi_next = biolist;
3175                                 biolist = bio;
3176                                 bio->bi_end_io = end_sync_read;
3177                                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3178                                 if (test_bit(FailFast, &rdev->flags))
3179                                         bio->bi_opf |= MD_FAILFAST;
3180                                 from_addr = r10_bio->devs[j].addr;
3181                                 bio->bi_iter.bi_sector = from_addr +
3182                                         rdev->data_offset;
3183                                 bio_set_dev(bio, rdev->bdev);
3184                                 atomic_inc(&rdev->nr_pending);
3185                                 /* and we write to 'i' (if not in_sync) */
3186
3187                                 for (k=0; k<conf->copies; k++)
3188                                         if (r10_bio->devs[k].devnum == i)
3189                                                 break;
3190                                 BUG_ON(k == conf->copies);
3191                                 to_addr = r10_bio->devs[k].addr;
3192                                 r10_bio->devs[0].devnum = d;
3193                                 r10_bio->devs[0].addr = from_addr;
3194                                 r10_bio->devs[1].devnum = i;
3195                                 r10_bio->devs[1].addr = to_addr;
3196
3197                                 if (need_recover) {
3198                                         bio = r10_bio->devs[1].bio;
3199                                         bio->bi_next = biolist;
3200                                         biolist = bio;
3201                                         bio->bi_end_io = end_sync_write;
3202                                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3203                                         bio->bi_iter.bi_sector = to_addr
3204                                                 + mrdev->data_offset;
3205                                         bio_set_dev(bio, mrdev->bdev);
3206                                         atomic_inc(&r10_bio->remaining);
3207                                 } else
3208                                         r10_bio->devs[1].bio->bi_end_io = NULL;
3209
3210                                 /* and maybe write to replacement */
3211                                 bio = r10_bio->devs[1].repl_bio;
3212                                 if (bio)
3213                                         bio->bi_end_io = NULL;
3214                                 /* Note: if need_replace, then bio
3215                                  * cannot be NULL as r10buf_pool_alloc will
3216                                  * have allocated it.
3217                                  */
3218                                 if (!need_replace)
3219                                         break;
3220                                 bio->bi_next = biolist;
3221                                 biolist = bio;
3222                                 bio->bi_end_io = end_sync_write;
3223                                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3224                                 bio->bi_iter.bi_sector = to_addr +
3225                                         mreplace->data_offset;
3226                                 bio_set_dev(bio, mreplace->bdev);
3227                                 atomic_inc(&r10_bio->remaining);
3228                                 break;
3229                         }
3230                         rcu_read_unlock();
3231                         if (j == conf->copies) {
3232                                 /* Cannot recover, so abort the recovery or
3233                                  * record a bad block */
3234                                 if (any_working) {
3235                                         /* problem is that there are bad blocks
3236                                          * on other device(s)
3237                                          */
3238                                         int k;
3239                                         for (k = 0; k < conf->copies; k++)
3240                                                 if (r10_bio->devs[k].devnum == i)
3241                                                         break;
3242                                         if (!test_bit(In_sync,
3243                                                       &mrdev->flags)
3244                                             && !rdev_set_badblocks(
3245                                                     mrdev,
3246                                                     r10_bio->devs[k].addr,
3247                                                     max_sync, 0))
3248                                                 any_working = 0;
3249                                         if (mreplace &&
3250                                             !rdev_set_badblocks(
3251                                                     mreplace,
3252                                                     r10_bio->devs[k].addr,
3253                                                     max_sync, 0))
3254                                                 any_working = 0;
3255                                 }
3256                                 if (!any_working)  {
3257                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
3258                                                               &mddev->recovery))
3259                                                 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3260                                                        mdname(mddev));
3261                                         mirror->recovery_disabled
3262                                                 = mddev->recovery_disabled;
3263                                 }
3264                                 put_buf(r10_bio);
3265                                 if (rb2)
3266                                         atomic_dec(&rb2->remaining);
3267                                 r10_bio = rb2;
3268                                 rdev_dec_pending(mrdev, mddev);
3269                                 if (mreplace)
3270                                         rdev_dec_pending(mreplace, mddev);
3271                                 break;
3272                         }
3273                         rdev_dec_pending(mrdev, mddev);
3274                         if (mreplace)
3275                                 rdev_dec_pending(mreplace, mddev);
3276                         if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3277                                 /* Only want this if there is elsewhere to
3278                                  * read from. 'j' is currently the first
3279                                  * readable copy.
3280                                  */
3281                                 int targets = 1;
3282                                 for (; j < conf->copies; j++) {
3283                                         int d = r10_bio->devs[j].devnum;
3284                                         if (conf->mirrors[d].rdev &&
3285                                             test_bit(In_sync,
3286                                                       &conf->mirrors[d].rdev->flags))
3287                                                 targets++;
3288                                 }
3289                                 if (targets == 1)
3290                                         r10_bio->devs[0].bio->bi_opf
3291                                                 &= ~MD_FAILFAST;
3292                         }
3293                 }
3294                 if (biolist == NULL) {
3295                         while (r10_bio) {
3296                                 struct r10bio *rb2 = r10_bio;
3297                                 r10_bio = (struct r10bio*) rb2->master_bio;
3298                                 rb2->master_bio = NULL;
3299                                 put_buf(rb2);
3300                         }
3301                         goto giveup;
3302                 }
3303         } else {
3304                 /* resync. Schedule a read for every block at this virt offset */
3305                 int count = 0;
3306
3307                 /*
3308                  * Since curr_resync_completed could probably not update in
3309                  * time, and we will set cluster_sync_low based on it.
3310                  * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3311                  * safety reason, which ensures curr_resync_completed is
3312                  * updated in bitmap_cond_end_sync.
3313                  */
3314                 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3315                                         mddev_is_clustered(mddev) &&
3316                                         (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3317
3318                 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3319                                           &sync_blocks, mddev->degraded) &&
3320                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3321                                                  &mddev->recovery)) {
3322                         /* We can skip this block */
3323                         *skipped = 1;
3324                         return sync_blocks + sectors_skipped;
3325                 }
3326                 if (sync_blocks < max_sync)
3327                         max_sync = sync_blocks;
3328                 r10_bio = raid10_alloc_init_r10buf(conf);
3329                 r10_bio->state = 0;
3330
3331                 r10_bio->mddev = mddev;
3332                 atomic_set(&r10_bio->remaining, 0);
3333                 raise_barrier(conf, 0);
3334                 conf->next_resync = sector_nr;
3335
3336                 r10_bio->master_bio = NULL;
3337                 r10_bio->sector = sector_nr;
3338                 set_bit(R10BIO_IsSync, &r10_bio->state);
3339                 raid10_find_phys(conf, r10_bio);
3340                 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3341
3342                 for (i = 0; i < conf->copies; i++) {
3343                         int d = r10_bio->devs[i].devnum;
3344                         sector_t first_bad, sector;
3345                         int bad_sectors;
3346                         struct md_rdev *rdev;
3347
3348                         if (r10_bio->devs[i].repl_bio)
3349                                 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3350
3351                         bio = r10_bio->devs[i].bio;
3352                         bio->bi_status = BLK_STS_IOERR;
3353                         rcu_read_lock();
3354                         rdev = rcu_dereference(conf->mirrors[d].rdev);
3355                         if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3356                                 rcu_read_unlock();
3357                                 continue;
3358                         }
3359                         sector = r10_bio->devs[i].addr;
3360                         if (is_badblock(rdev, sector, max_sync,
3361                                         &first_bad, &bad_sectors)) {
3362                                 if (first_bad > sector)
3363                                         max_sync = first_bad - sector;
3364                                 else {
3365                                         bad_sectors -= (sector - first_bad);
3366                                         if (max_sync > bad_sectors)
3367                                                 max_sync = bad_sectors;
3368                                         rcu_read_unlock();
3369                                         continue;
3370                                 }
3371                         }
3372                         atomic_inc(&rdev->nr_pending);
3373                         atomic_inc(&r10_bio->remaining);
3374                         bio->bi_next = biolist;
3375                         biolist = bio;
3376                         bio->bi_end_io = end_sync_read;
3377                         bio_set_op_attrs(bio, REQ_OP_READ, 0);
3378                         if (test_bit(FailFast, &rdev->flags))
3379                                 bio->bi_opf |= MD_FAILFAST;
3380                         bio->bi_iter.bi_sector = sector + rdev->data_offset;
3381                         bio_set_dev(bio, rdev->bdev);
3382                         count++;
3383
3384                         rdev = rcu_dereference(conf->mirrors[d].replacement);
3385                         if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3386                                 rcu_read_unlock();
3387                                 continue;
3388                         }
3389                         atomic_inc(&rdev->nr_pending);
3390
3391                         /* Need to set up for writing to the replacement */
3392                         bio = r10_bio->devs[i].repl_bio;
3393                         bio->bi_status = BLK_STS_IOERR;
3394
3395                         sector = r10_bio->devs[i].addr;
3396                         bio->bi_next = biolist;
3397                         biolist = bio;
3398                         bio->bi_end_io = end_sync_write;
3399                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3400                         if (test_bit(FailFast, &rdev->flags))
3401                                 bio->bi_opf |= MD_FAILFAST;
3402                         bio->bi_iter.bi_sector = sector + rdev->data_offset;
3403                         bio_set_dev(bio, rdev->bdev);
3404                         count++;
3405                         rcu_read_unlock();
3406                 }
3407
3408                 if (count < 2) {
3409                         for (i=0; i<conf->copies; i++) {
3410                                 int d = r10_bio->devs[i].devnum;
3411                                 if (r10_bio->devs[i].bio->bi_end_io)
3412                                         rdev_dec_pending(conf->mirrors[d].rdev,
3413                                                          mddev);
3414                                 if (r10_bio->devs[i].repl_bio &&
3415                                     r10_bio->devs[i].repl_bio->bi_end_io)
3416                                         rdev_dec_pending(
3417                                                 conf->mirrors[d].replacement,
3418                                                 mddev);
3419                         }
3420                         put_buf(r10_bio);
3421                         biolist = NULL;
3422                         goto giveup;
3423                 }
3424         }
3425
3426         nr_sectors = 0;
3427         if (sector_nr + max_sync < max_sector)
3428                 max_sector = sector_nr + max_sync;
3429         do {
3430                 struct page *page;
3431                 int len = PAGE_SIZE;
3432                 if (sector_nr + (len>>9) > max_sector)
3433                         len = (max_sector - sector_nr) << 9;
3434                 if (len == 0)
3435                         break;
3436                 for (bio= biolist ; bio ; bio=bio->bi_next) {
3437                         struct resync_pages *rp = get_resync_pages(bio);
3438                         page = resync_fetch_page(rp, page_idx);
3439                         /*
3440                          * won't fail because the vec table is big enough
3441                          * to hold all these pages
3442                          */
3443                         bio_add_page(bio, page, len, 0);
3444                 }
3445                 nr_sectors += len>>9;
3446                 sector_nr += len>>9;
3447         } while (++page_idx < RESYNC_PAGES);
3448         r10_bio->sectors = nr_sectors;
3449
3450         if (mddev_is_clustered(mddev) &&
3451             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3452                 /* It is resync not recovery */
3453                 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3454                         conf->cluster_sync_low = mddev->curr_resync_completed;
3455                         raid10_set_cluster_sync_high(conf);
3456                         /* Send resync message */
3457                         md_cluster_ops->resync_info_update(mddev,
3458                                                 conf->cluster_sync_low,
3459                                                 conf->cluster_sync_high);
3460                 }
3461         } else if (mddev_is_clustered(mddev)) {
3462                 /* This is recovery not resync */
3463                 sector_t sect_va1, sect_va2;
3464                 bool broadcast_msg = false;
3465
3466                 for (i = 0; i < conf->geo.raid_disks; i++) {
3467                         /*
3468                          * sector_nr is a device address for recovery, so we
3469                          * need translate it to array address before compare
3470                          * with cluster_sync_high.
3471                          */
3472                         sect_va1 = raid10_find_virt(conf, sector_nr, i);
3473
3474                         if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3475                                 broadcast_msg = true;
3476                                 /*
3477                                  * curr_resync_completed is similar as
3478                                  * sector_nr, so make the translation too.
3479                                  */
3480                                 sect_va2 = raid10_find_virt(conf,
3481                                         mddev->curr_resync_completed, i);
3482
3483                                 if (conf->cluster_sync_low == 0 ||
3484                                     conf->cluster_sync_low > sect_va2)
3485                                         conf->cluster_sync_low = sect_va2;
3486                         }
3487                 }
3488                 if (broadcast_msg) {
3489                         raid10_set_cluster_sync_high(conf);
3490                         md_cluster_ops->resync_info_update(mddev,
3491                                                 conf->cluster_sync_low,
3492                                                 conf->cluster_sync_high);
3493                 }
3494         }
3495
3496         while (biolist) {
3497                 bio = biolist;
3498                 biolist = biolist->bi_next;
3499
3500                 bio->bi_next = NULL;
3501                 r10_bio = get_resync_r10bio(bio);
3502                 r10_bio->sectors = nr_sectors;
3503
3504                 if (bio->bi_end_io == end_sync_read) {
3505                         md_sync_acct_bio(bio, nr_sectors);
3506                         bio->bi_status = 0;
3507                         generic_make_request(bio);
3508                 }
3509         }
3510
3511         if (sectors_skipped)
3512                 /* pretend they weren't skipped, it makes
3513                  * no important difference in this case
3514                  */
3515                 md_done_sync(mddev, sectors_skipped, 1);
3516
3517         return sectors_skipped + nr_sectors;
3518  giveup:
3519         /* There is nowhere to write, so all non-sync
3520          * drives must be failed or in resync, all drives
3521          * have a bad block, so try the next chunk...
3522          */
3523         if (sector_nr + max_sync < max_sector)
3524                 max_sector = sector_nr + max_sync;
3525
3526         sectors_skipped += (max_sector - sector_nr);
3527         chunks_skipped ++;
3528         sector_nr = max_sector;
3529         goto skipped;
3530 }
3531
3532 static sector_t
3533 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3534 {
3535         sector_t size;
3536         struct r10conf *conf = mddev->private;
3537
3538         if (!raid_disks)
3539                 raid_disks = min(conf->geo.raid_disks,
3540                                  conf->prev.raid_disks);
3541         if (!sectors)
3542                 sectors = conf->dev_sectors;
3543
3544         size = sectors >> conf->geo.chunk_shift;
3545         sector_div(size, conf->geo.far_copies);
3546         size = size * raid_disks;
3547         sector_div(size, conf->geo.near_copies);
3548
3549         return size << conf->geo.chunk_shift;
3550 }
3551
3552 static void calc_sectors(struct r10conf *conf, sector_t size)
3553 {
3554         /* Calculate the number of sectors-per-device that will
3555          * actually be used, and set conf->dev_sectors and
3556          * conf->stride
3557          */
3558
3559         size = size >> conf->geo.chunk_shift;
3560         sector_div(size, conf->geo.far_copies);
3561         size = size * conf->geo.raid_disks;
3562         sector_div(size, conf->geo.near_copies);
3563         /* 'size' is now the number of chunks in the array */
3564         /* calculate "used chunks per device" */
3565         size = size * conf->copies;
3566
3567         /* We need to round up when dividing by raid_disks to
3568          * get the stride size.
3569          */
3570         size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3571
3572         conf->dev_sectors = size << conf->geo.chunk_shift;
3573
3574         if (conf->geo.far_offset)
3575                 conf->geo.stride = 1 << conf->geo.chunk_shift;
3576         else {
3577                 sector_div(size, conf->geo.far_copies);
3578                 conf->geo.stride = size << conf->geo.chunk_shift;
3579         }
3580 }
3581
3582 enum geo_type {geo_new, geo_old, geo_start};
3583 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3584 {
3585         int nc, fc, fo;
3586         int layout, chunk, disks;
3587         switch (new) {
3588         case geo_old:
3589                 layout = mddev->layout;
3590                 chunk = mddev->chunk_sectors;
3591                 disks = mddev->raid_disks - mddev->delta_disks;
3592                 break;
3593         case geo_new:
3594                 layout = mddev->new_layout;
3595                 chunk = mddev->new_chunk_sectors;
3596                 disks = mddev->raid_disks;
3597                 break;
3598         default: /* avoid 'may be unused' warnings */
3599         case geo_start: /* new when starting reshape - raid_disks not
3600                          * updated yet. */
3601                 layout = mddev->new_layout;
3602                 chunk = mddev->new_chunk_sectors;
3603                 disks = mddev->raid_disks + mddev->delta_disks;
3604                 break;
3605         }
3606         if (layout >> 19)
3607                 return -1;
3608         if (chunk < (PAGE_SIZE >> 9) ||
3609             !is_power_of_2(chunk))
3610                 return -2;
3611         nc = layout & 255;
3612         fc = (layout >> 8) & 255;
3613         fo = layout & (1<<16);
3614         geo->raid_disks = disks;
3615         geo->near_copies = nc;
3616         geo->far_copies = fc;
3617         geo->far_offset = fo;
3618         switch (layout >> 17) {
3619         case 0: /* original layout.  simple but not always optimal */
3620                 geo->far_set_size = disks;
3621                 break;
3622         case 1: /* "improved" layout which was buggy.  Hopefully no-one is
3623                  * actually using this, but leave code here just in case.*/
3624                 geo->far_set_size = disks/fc;
3625                 WARN(geo->far_set_size < fc,
3626                      "This RAID10 layout does not provide data safety - please backup and create new array\n");
3627                 break;
3628         case 2: /* "improved" layout fixed to match documentation */
3629                 geo->far_set_size = fc * nc;
3630                 break;
3631         default: /* Not a valid layout */
3632                 return -1;
3633         }
3634         geo->chunk_mask = chunk - 1;
3635         geo->chunk_shift = ffz(~chunk);
3636         return nc*fc;
3637 }
3638
3639 static struct r10conf *setup_conf(struct mddev *mddev)
3640 {
3641         struct r10conf *conf = NULL;
3642         int err = -EINVAL;
3643         struct geom geo;
3644         int copies;
3645
3646         copies = setup_geo(&geo, mddev, geo_new);
3647
3648         if (copies == -2) {
3649                 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3650                         mdname(mddev), PAGE_SIZE);
3651                 goto out;
3652         }
3653
3654         if (copies < 2 || copies > mddev->raid_disks) {
3655                 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3656                         mdname(mddev), mddev->new_layout);
3657                 goto out;
3658         }
3659
3660         err = -ENOMEM;
3661         conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3662         if (!conf)
3663                 goto out;
3664
3665         /* FIXME calc properly */
3666         conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3667                                 sizeof(struct raid10_info),
3668                                 GFP_KERNEL);
3669         if (!conf->mirrors)
3670                 goto out;
3671
3672         conf->tmppage = alloc_page(GFP_KERNEL);
3673         if (!conf->tmppage)
3674                 goto out;
3675
3676         conf->geo = geo;
3677         conf->copies = copies;
3678         err = mempool_init(&conf->r10bio_pool, NR_RAID10_BIOS, r10bio_pool_alloc,
3679                            r10bio_pool_free, conf);
3680         if (err)
3681                 goto out;
3682
3683         err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3684         if (err)
3685                 goto out;
3686
3687         calc_sectors(conf, mddev->dev_sectors);
3688         if (mddev->reshape_position == MaxSector) {
3689                 conf->prev = conf->geo;
3690                 conf->reshape_progress = MaxSector;
3691         } else {
3692                 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3693                         err = -EINVAL;
3694                         goto out;
3695                 }
3696                 conf->reshape_progress = mddev->reshape_position;
3697                 if (conf->prev.far_offset)
3698                         conf->prev.stride = 1 << conf->prev.chunk_shift;
3699                 else
3700                         /* far_copies must be 1 */
3701                         conf->prev.stride = conf->dev_sectors;
3702         }
3703         conf->reshape_safe = conf->reshape_progress;
3704         spin_lock_init(&conf->device_lock);
3705         INIT_LIST_HEAD(&conf->retry_list);
3706         INIT_LIST_HEAD(&conf->bio_end_io_list);
3707
3708         spin_lock_init(&conf->resync_lock);
3709         init_waitqueue_head(&conf->wait_barrier);
3710         atomic_set(&conf->nr_pending, 0);
3711
3712         err = -ENOMEM;
3713         conf->thread = md_register_thread(raid10d, mddev, "raid10");
3714         if (!conf->thread)
3715                 goto out;
3716
3717         conf->mddev = mddev;
3718         return conf;
3719
3720  out:
3721         if (conf) {
3722                 mempool_exit(&conf->r10bio_pool);
3723                 kfree(conf->mirrors);
3724                 safe_put_page(conf->tmppage);
3725                 bioset_exit(&conf->bio_split);
3726                 kfree(conf);
3727         }
3728         return ERR_PTR(err);
3729 }
3730
3731 static int raid10_run(struct mddev *mddev)
3732 {
3733         struct r10conf *conf;
3734         int i, disk_idx, chunk_size;
3735         struct raid10_info *disk;
3736         struct md_rdev *rdev;
3737         sector_t size;
3738         sector_t min_offset_diff = 0;
3739         int first = 1;
3740         bool discard_supported = false;
3741
3742         if (mddev_init_writes_pending(mddev) < 0)
3743                 return -ENOMEM;
3744
3745         if (mddev->private == NULL) {
3746                 conf = setup_conf(mddev);
3747                 if (IS_ERR(conf))
3748                         return PTR_ERR(conf);
3749                 mddev->private = conf;
3750         }
3751         conf = mddev->private;
3752         if (!conf)
3753                 goto out;
3754
3755         if (mddev_is_clustered(conf->mddev)) {
3756                 int fc, fo;
3757
3758                 fc = (mddev->layout >> 8) & 255;
3759                 fo = mddev->layout & (1<<16);
3760                 if (fc > 1 || fo > 0) {
3761                         pr_err("only near layout is supported by clustered"
3762                                 " raid10\n");
3763                         goto out_free_conf;
3764                 }
3765         }
3766
3767         mddev->thread = conf->thread;
3768         conf->thread = NULL;
3769
3770         chunk_size = mddev->chunk_sectors << 9;
3771         if (mddev->queue) {
3772                 blk_queue_max_discard_sectors(mddev->queue,
3773                                               mddev->chunk_sectors);
3774                 blk_queue_max_write_same_sectors(mddev->queue, 0);
3775                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3776                 blk_queue_io_min(mddev->queue, chunk_size);
3777                 if (conf->geo.raid_disks % conf->geo.near_copies)
3778                         blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3779                 else
3780                         blk_queue_io_opt(mddev->queue, chunk_size *
3781                                          (conf->geo.raid_disks / conf->geo.near_copies));
3782         }
3783
3784         rdev_for_each(rdev, mddev) {
3785                 long long diff;
3786
3787                 disk_idx = rdev->raid_disk;
3788                 if (disk_idx < 0)
3789                         continue;
3790                 if (disk_idx >= conf->geo.raid_disks &&
3791                     disk_idx >= conf->prev.raid_disks)
3792                         continue;
3793                 disk = conf->mirrors + disk_idx;
3794
3795                 if (test_bit(Replacement, &rdev->flags)) {
3796                         if (disk->replacement)
3797                                 goto out_free_conf;
3798                         disk->replacement = rdev;
3799                 } else {
3800                         if (disk->rdev)
3801                                 goto out_free_conf;
3802                         disk->rdev = rdev;
3803                 }
3804                 diff = (rdev->new_data_offset - rdev->data_offset);
3805                 if (!mddev->reshape_backwards)
3806                         diff = -diff;
3807                 if (diff < 0)
3808                         diff = 0;
3809                 if (first || diff < min_offset_diff)
3810                         min_offset_diff = diff;
3811
3812                 if (mddev->gendisk)
3813                         disk_stack_limits(mddev->gendisk, rdev->bdev,
3814                                           rdev->data_offset << 9);
3815
3816                 disk->head_position = 0;
3817
3818                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3819                         discard_supported = true;
3820                 first = 0;
3821         }
3822
3823         if (mddev->queue) {
3824                 if (discard_supported)
3825                         blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3826                                                 mddev->queue);
3827                 else
3828                         blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3829                                                   mddev->queue);
3830         }
3831         /* need to check that every block has at least one working mirror */
3832         if (!enough(conf, -1)) {
3833                 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3834                        mdname(mddev));
3835                 goto out_free_conf;
3836         }
3837
3838         if (conf->reshape_progress != MaxSector) {
3839                 /* must ensure that shape change is supported */
3840                 if (conf->geo.far_copies != 1 &&
3841                     conf->geo.far_offset == 0)
3842                         goto out_free_conf;
3843                 if (conf->prev.far_copies != 1 &&
3844                     conf->prev.far_offset == 0)
3845                         goto out_free_conf;
3846         }
3847
3848         mddev->degraded = 0;
3849         for (i = 0;
3850              i < conf->geo.raid_disks
3851                      || i < conf->prev.raid_disks;
3852              i++) {
3853
3854                 disk = conf->mirrors + i;
3855
3856                 if (!disk->rdev && disk->replacement) {
3857                         /* The replacement is all we have - use it */
3858                         disk->rdev = disk->replacement;
3859                         disk->replacement = NULL;
3860                         clear_bit(Replacement, &disk->rdev->flags);
3861                 }
3862
3863                 if (!disk->rdev ||
3864                     !test_bit(In_sync, &disk->rdev->flags)) {
3865                         disk->head_position = 0;
3866                         mddev->degraded++;
3867                         if (disk->rdev &&
3868                             disk->rdev->saved_raid_disk < 0)
3869                                 conf->fullsync = 1;
3870                 }
3871
3872                 if (disk->replacement &&
3873                     !test_bit(In_sync, &disk->replacement->flags) &&
3874                     disk->replacement->saved_raid_disk < 0) {
3875                         conf->fullsync = 1;
3876                 }
3877
3878                 disk->recovery_disabled = mddev->recovery_disabled - 1;
3879         }
3880
3881         if (mddev->recovery_cp != MaxSector)
3882                 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3883                           mdname(mddev));
3884         pr_info("md/raid10:%s: active with %d out of %d devices\n",
3885                 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3886                 conf->geo.raid_disks);
3887         /*
3888          * Ok, everything is just fine now
3889          */
3890         mddev->dev_sectors = conf->dev_sectors;
3891         size = raid10_size(mddev, 0, 0);
3892         md_set_array_sectors(mddev, size);
3893         mddev->resync_max_sectors = size;
3894         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3895
3896         if (mddev->queue) {
3897                 int stripe = conf->geo.raid_disks *
3898                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3899
3900                 /* Calculate max read-ahead size.
3901                  * We need to readahead at least twice a whole stripe....
3902                  * maybe...
3903                  */
3904                 stripe /= conf->geo.near_copies;
3905                 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3906                         mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3907         }
3908
3909         if (md_integrity_register(mddev))
3910                 goto out_free_conf;
3911
3912         if (conf->reshape_progress != MaxSector) {
3913                 unsigned long before_length, after_length;
3914
3915                 before_length = ((1 << conf->prev.chunk_shift) *
3916                                  conf->prev.far_copies);
3917                 after_length = ((1 << conf->geo.chunk_shift) *
3918                                 conf->geo.far_copies);
3919
3920                 if (max(before_length, after_length) > min_offset_diff) {
3921                         /* This cannot work */
3922                         pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3923                         goto out_free_conf;
3924                 }
3925                 conf->offset_diff = min_offset_diff;
3926
3927                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3928                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3929                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3930                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3931                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3932                                                         "reshape");
3933                 if (!mddev->sync_thread)
3934                         goto out_free_conf;
3935         }
3936
3937         return 0;
3938
3939 out_free_conf:
3940         md_unregister_thread(&mddev->thread);
3941         mempool_exit(&conf->r10bio_pool);
3942         safe_put_page(conf->tmppage);
3943         kfree(conf->mirrors);
3944         kfree(conf);
3945         mddev->private = NULL;
3946 out:
3947         return -EIO;
3948 }
3949
3950 static void raid10_free(struct mddev *mddev, void *priv)
3951 {
3952         struct r10conf *conf = priv;
3953
3954         mempool_exit(&conf->r10bio_pool);
3955         safe_put_page(conf->tmppage);
3956         kfree(conf->mirrors);
3957         kfree(conf->mirrors_old);
3958         kfree(conf->mirrors_new);
3959         bioset_exit(&conf->bio_split);
3960         kfree(conf);
3961 }
3962
3963 static void raid10_quiesce(struct mddev *mddev, int quiesce)
3964 {
3965         struct r10conf *conf = mddev->private;
3966
3967         if (quiesce)
3968                 raise_barrier(conf, 0);
3969         else
3970                 lower_barrier(conf);
3971 }
3972
3973 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3974 {
3975         /* Resize of 'far' arrays is not supported.
3976          * For 'near' and 'offset' arrays we can set the
3977          * number of sectors used to be an appropriate multiple
3978          * of the chunk size.
3979          * For 'offset', this is far_copies*chunksize.
3980          * For 'near' the multiplier is the LCM of
3981          * near_copies and raid_disks.
3982          * So if far_copies > 1 && !far_offset, fail.
3983          * Else find LCM(raid_disks, near_copy)*far_copies and
3984          * multiply by chunk_size.  Then round to this number.
3985          * This is mostly done by raid10_size()
3986          */
3987         struct r10conf *conf = mddev->private;
3988         sector_t oldsize, size;
3989
3990         if (mddev->reshape_position != MaxSector)
3991                 return -EBUSY;
3992
3993         if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3994                 return -EINVAL;
3995
3996         oldsize = raid10_size(mddev, 0, 0);
3997         size = raid10_size(mddev, sectors, 0);
3998         if (mddev->external_size &&
3999             mddev->array_sectors > size)
4000                 return -EINVAL;
4001         if (mddev->bitmap) {
4002                 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
4003                 if (ret)
4004                         return ret;
4005         }
4006         md_set_array_sectors(mddev, size);
4007         if (sectors > mddev->dev_sectors &&
4008             mddev->recovery_cp > oldsize) {
4009                 mddev->recovery_cp = oldsize;
4010                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4011         }
4012         calc_sectors(conf, sectors);
4013         mddev->dev_sectors = conf->dev_sectors;
4014         mddev->resync_max_sectors = size;
4015         return 0;
4016 }
4017
4018 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4019 {
4020         struct md_rdev *rdev;
4021         struct r10conf *conf;
4022
4023         if (mddev->degraded > 0) {
4024                 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4025                         mdname(mddev));
4026                 return ERR_PTR(-EINVAL);
4027         }
4028         sector_div(size, devs);
4029
4030         /* Set new parameters */
4031         mddev->new_level = 10;
4032         /* new layout: far_copies = 1, near_copies = 2 */
4033         mddev->new_layout = (1<<8) + 2;
4034         mddev->new_chunk_sectors = mddev->chunk_sectors;
4035         mddev->delta_disks = mddev->raid_disks;
4036         mddev->raid_disks *= 2;
4037         /* make sure it will be not marked as dirty */
4038         mddev->recovery_cp = MaxSector;
4039         mddev->dev_sectors = size;
4040
4041         conf = setup_conf(mddev);
4042         if (!IS_ERR(conf)) {
4043                 rdev_for_each(rdev, mddev)
4044                         if (rdev->raid_disk >= 0) {
4045                                 rdev->new_raid_disk = rdev->raid_disk * 2;
4046                                 rdev->sectors = size;
4047                         }
4048                 conf->barrier = 1;
4049         }
4050
4051         return conf;
4052 }
4053
4054 static void *raid10_takeover(struct mddev *mddev)
4055 {
4056         struct r0conf *raid0_conf;
4057
4058         /* raid10 can take over:
4059          *  raid0 - providing it has only two drives
4060          */
4061         if (mddev->level == 0) {
4062                 /* for raid0 takeover only one zone is supported */
4063                 raid0_conf = mddev->private;
4064                 if (raid0_conf->nr_strip_zones > 1) {
4065                         pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4066                                 mdname(mddev));
4067                         return ERR_PTR(-EINVAL);
4068                 }
4069                 return raid10_takeover_raid0(mddev,
4070                         raid0_conf->strip_zone->zone_end,
4071                         raid0_conf->strip_zone->nb_dev);
4072         }
4073         return ERR_PTR(-EINVAL);
4074 }
4075
4076 static int raid10_check_reshape(struct mddev *mddev)
4077 {
4078         /* Called when there is a request to change
4079          * - layout (to ->new_layout)
4080          * - chunk size (to ->new_chunk_sectors)
4081          * - raid_disks (by delta_disks)
4082          * or when trying to restart a reshape that was ongoing.
4083          *
4084          * We need to validate the request and possibly allocate
4085          * space if that might be an issue later.
4086          *
4087          * Currently we reject any reshape of a 'far' mode array,
4088          * allow chunk size to change if new is generally acceptable,
4089          * allow raid_disks to increase, and allow
4090          * a switch between 'near' mode and 'offset' mode.
4091          */
4092         struct r10conf *conf = mddev->private;
4093         struct geom geo;
4094
4095         if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4096                 return -EINVAL;
4097
4098         if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4099                 /* mustn't change number of copies */
4100                 return -EINVAL;
4101         if (geo.far_copies > 1 && !geo.far_offset)
4102                 /* Cannot switch to 'far' mode */
4103                 return -EINVAL;
4104
4105         if (mddev->array_sectors & geo.chunk_mask)
4106                         /* not factor of array size */
4107                         return -EINVAL;
4108
4109         if (!enough(conf, -1))
4110                 return -EINVAL;
4111
4112         kfree(conf->mirrors_new);
4113         conf->mirrors_new = NULL;
4114         if (mddev->delta_disks > 0) {
4115                 /* allocate new 'mirrors' list */
4116                 conf->mirrors_new =
4117                         kcalloc(mddev->raid_disks + mddev->delta_disks,
4118                                 sizeof(struct raid10_info),
4119                                 GFP_KERNEL);
4120                 if (!conf->mirrors_new)
4121                         return -ENOMEM;
4122         }
4123         return 0;
4124 }
4125
4126 /*
4127  * Need to check if array has failed when deciding whether to:
4128  *  - start an array
4129  *  - remove non-faulty devices
4130  *  - add a spare
4131  *  - allow a reshape
4132  * This determination is simple when no reshape is happening.
4133  * However if there is a reshape, we need to carefully check
4134  * both the before and after sections.
4135  * This is because some failed devices may only affect one
4136  * of the two sections, and some non-in_sync devices may
4137  * be insync in the section most affected by failed devices.
4138  */
4139 static int calc_degraded(struct r10conf *conf)
4140 {
4141         int degraded, degraded2;
4142         int i;
4143
4144         rcu_read_lock();
4145         degraded = 0;
4146         /* 'prev' section first */
4147         for (i = 0; i < conf->prev.raid_disks; i++) {
4148                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4149                 if (!rdev || test_bit(Faulty, &rdev->flags))
4150                         degraded++;
4151                 else if (!test_bit(In_sync, &rdev->flags))
4152                         /* When we can reduce the number of devices in
4153                          * an array, this might not contribute to
4154                          * 'degraded'.  It does now.
4155                          */
4156                         degraded++;
4157         }
4158         rcu_read_unlock();
4159         if (conf->geo.raid_disks == conf->prev.raid_disks)
4160                 return degraded;
4161         rcu_read_lock();
4162         degraded2 = 0;
4163         for (i = 0; i < conf->geo.raid_disks; i++) {
4164                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4165                 if (!rdev || test_bit(Faulty, &rdev->flags))
4166                         degraded2++;
4167                 else if (!test_bit(In_sync, &rdev->flags)) {
4168                         /* If reshape is increasing the number of devices,
4169                          * this section has already been recovered, so
4170                          * it doesn't contribute to degraded.
4171                          * else it does.
4172                          */
4173                         if (conf->geo.raid_disks <= conf->prev.raid_disks)
4174                                 degraded2++;
4175                 }
4176         }
4177         rcu_read_unlock();
4178         if (degraded2 > degraded)
4179                 return degraded2;
4180         return degraded;
4181 }
4182
4183 static int raid10_start_reshape(struct mddev *mddev)
4184 {
4185         /* A 'reshape' has been requested. This commits
4186          * the various 'new' fields and sets MD_RECOVER_RESHAPE
4187          * This also checks if there are enough spares and adds them
4188          * to the array.
4189          * We currently require enough spares to make the final
4190          * array non-degraded.  We also require that the difference
4191          * between old and new data_offset - on each device - is
4192          * enough that we never risk over-writing.
4193          */
4194
4195         unsigned long before_length, after_length;
4196         sector_t min_offset_diff = 0;
4197         int first = 1;
4198         struct geom new;
4199         struct r10conf *conf = mddev->private;
4200         struct md_rdev *rdev;
4201         int spares = 0;
4202         int ret;
4203
4204         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4205                 return -EBUSY;
4206
4207         if (setup_geo(&new, mddev, geo_start) != conf->copies)
4208                 return -EINVAL;
4209
4210         before_length = ((1 << conf->prev.chunk_shift) *
4211                          conf->prev.far_copies);
4212         after_length = ((1 << conf->geo.chunk_shift) *
4213                         conf->geo.far_copies);
4214
4215         rdev_for_each(rdev, mddev) {
4216                 if (!test_bit(In_sync, &rdev->flags)
4217                     && !test_bit(Faulty, &rdev->flags))
4218                         spares++;
4219                 if (rdev->raid_disk >= 0) {
4220                         long long diff = (rdev->new_data_offset
4221                                           - rdev->data_offset);
4222                         if (!mddev->reshape_backwards)
4223                                 diff = -diff;
4224                         if (diff < 0)
4225                                 diff = 0;
4226                         if (first || diff < min_offset_diff)
4227                                 min_offset_diff = diff;
4228                         first = 0;
4229                 }
4230         }
4231
4232         if (max(before_length, after_length) > min_offset_diff)
4233                 return -EINVAL;
4234
4235         if (spares < mddev->delta_disks)
4236                 return -EINVAL;
4237
4238         conf->offset_diff = min_offset_diff;
4239         spin_lock_irq(&conf->device_lock);
4240         if (conf->mirrors_new) {
4241                 memcpy(conf->mirrors_new, conf->mirrors,
4242                        sizeof(struct raid10_info)*conf->prev.raid_disks);
4243                 smp_mb();
4244                 kfree(conf->mirrors_old);
4245                 conf->mirrors_old = conf->mirrors;
4246                 conf->mirrors = conf->mirrors_new;
4247                 conf->mirrors_new = NULL;
4248         }
4249         setup_geo(&conf->geo, mddev, geo_start);
4250         smp_mb();
4251         if (mddev->reshape_backwards) {
4252                 sector_t size = raid10_size(mddev, 0, 0);
4253                 if (size < mddev->array_sectors) {
4254                         spin_unlock_irq(&conf->device_lock);
4255                         pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4256                                 mdname(mddev));
4257                         return -EINVAL;
4258                 }
4259                 mddev->resync_max_sectors = size;
4260                 conf->reshape_progress = size;
4261         } else
4262                 conf->reshape_progress = 0;
4263         conf->reshape_safe = conf->reshape_progress;
4264         spin_unlock_irq(&conf->device_lock);
4265
4266         if (mddev->delta_disks && mddev->bitmap) {
4267                 struct mdp_superblock_1 *sb = NULL;
4268                 sector_t oldsize, newsize;
4269
4270                 oldsize = raid10_size(mddev, 0, 0);
4271                 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4272
4273                 if (!mddev_is_clustered(mddev)) {
4274                         ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4275                         if (ret)
4276                                 goto abort;
4277                         else
4278                                 goto out;
4279                 }
4280
4281                 rdev_for_each(rdev, mddev) {
4282                         if (rdev->raid_disk > -1 &&
4283                             !test_bit(Faulty, &rdev->flags))
4284                                 sb = page_address(rdev->sb_page);
4285                 }
4286
4287                 /*
4288                  * some node is already performing reshape, and no need to
4289                  * call md_bitmap_resize again since it should be called when
4290                  * receiving BITMAP_RESIZE msg
4291                  */
4292                 if ((sb && (le32_to_cpu(sb->feature_map) &
4293                             MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4294                         goto out;
4295
4296                 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4297                 if (ret)
4298                         goto abort;
4299
4300                 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4301                 if (ret) {
4302                         md_bitmap_resize(mddev->bitmap, oldsize, 0, 0);
4303                         goto abort;
4304                 }
4305         }
4306 out:
4307         if (mddev->delta_disks > 0) {
4308                 rdev_for_each(rdev, mddev)
4309                         if (rdev->raid_disk < 0 &&
4310                             !test_bit(Faulty, &rdev->flags)) {
4311                                 if (raid10_add_disk(mddev, rdev) == 0) {
4312                                         if (rdev->raid_disk >=
4313                                             conf->prev.raid_disks)
4314                                                 set_bit(In_sync, &rdev->flags);
4315                                         else
4316                                                 rdev->recovery_offset = 0;
4317
4318                                         if (sysfs_link_rdev(mddev, rdev))
4319                                                 /* Failure here  is OK */;
4320                                 }
4321                         } else if (rdev->raid_disk >= conf->prev.raid_disks
4322                                    && !test_bit(Faulty, &rdev->flags)) {
4323                                 /* This is a spare that was manually added */
4324                                 set_bit(In_sync, &rdev->flags);
4325                         }
4326         }
4327         /* When a reshape changes the number of devices,
4328          * ->degraded is measured against the larger of the
4329          * pre and  post numbers.
4330          */
4331         spin_lock_irq(&conf->device_lock);
4332         mddev->degraded = calc_degraded(conf);
4333         spin_unlock_irq(&conf->device_lock);
4334         mddev->raid_disks = conf->geo.raid_disks;
4335         mddev->reshape_position = conf->reshape_progress;
4336         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4337
4338         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4339         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4340         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4341         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4342         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4343
4344         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4345                                                 "reshape");
4346         if (!mddev->sync_thread) {
4347                 ret = -EAGAIN;
4348                 goto abort;
4349         }
4350         conf->reshape_checkpoint = jiffies;
4351         md_wakeup_thread(mddev->sync_thread);
4352         md_new_event(mddev);
4353         return 0;
4354
4355 abort:
4356         mddev->recovery = 0;
4357         spin_lock_irq(&conf->device_lock);
4358         conf->geo = conf->prev;
4359         mddev->raid_disks = conf->geo.raid_disks;
4360         rdev_for_each(rdev, mddev)
4361                 rdev->new_data_offset = rdev->data_offset;
4362         smp_wmb();
4363         conf->reshape_progress = MaxSector;
4364         conf->reshape_safe = MaxSector;
4365         mddev->reshape_position = MaxSector;
4366         spin_unlock_irq(&conf->device_lock);
4367         return ret;
4368 }
4369
4370 /* Calculate the last device-address that could contain
4371  * any block from the chunk that includes the array-address 's'
4372  * and report the next address.
4373  * i.e. the address returned will be chunk-aligned and after
4374  * any data that is in the chunk containing 's'.
4375  */
4376 static sector_t last_dev_address(sector_t s, struct geom *geo)
4377 {
4378         s = (s | geo->chunk_mask) + 1;
4379         s >>= geo->chunk_shift;
4380         s *= geo->near_copies;
4381         s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4382         s *= geo->far_copies;
4383         s <<= geo->chunk_shift;
4384         return s;
4385 }
4386
4387 /* Calculate the first device-address that could contain
4388  * any block from the chunk that includes the array-address 's'.
4389  * This too will be the start of a chunk
4390  */
4391 static sector_t first_dev_address(sector_t s, struct geom *geo)
4392 {
4393         s >>= geo->chunk_shift;
4394         s *= geo->near_copies;
4395         sector_div(s, geo->raid_disks);
4396         s *= geo->far_copies;
4397         s <<= geo->chunk_shift;
4398         return s;
4399 }
4400
4401 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4402                                 int *skipped)
4403 {
4404         /* We simply copy at most one chunk (smallest of old and new)
4405          * at a time, possibly less if that exceeds RESYNC_PAGES,
4406          * or we hit a bad block or something.
4407          * This might mean we pause for normal IO in the middle of
4408          * a chunk, but that is not a problem as mddev->reshape_position
4409          * can record any location.
4410          *
4411          * If we will want to write to a location that isn't
4412          * yet recorded as 'safe' (i.e. in metadata on disk) then
4413          * we need to flush all reshape requests and update the metadata.
4414          *
4415          * When reshaping forwards (e.g. to more devices), we interpret
4416          * 'safe' as the earliest block which might not have been copied
4417          * down yet.  We divide this by previous stripe size and multiply
4418          * by previous stripe length to get lowest device offset that we
4419          * cannot write to yet.
4420          * We interpret 'sector_nr' as an address that we want to write to.
4421          * From this we use last_device_address() to find where we might
4422          * write to, and first_device_address on the  'safe' position.
4423          * If this 'next' write position is after the 'safe' position,
4424          * we must update the metadata to increase the 'safe' position.
4425          *
4426          * When reshaping backwards, we round in the opposite direction
4427          * and perform the reverse test:  next write position must not be
4428          * less than current safe position.
4429          *
4430          * In all this the minimum difference in data offsets
4431          * (conf->offset_diff - always positive) allows a bit of slack,
4432          * so next can be after 'safe', but not by more than offset_diff
4433          *
4434          * We need to prepare all the bios here before we start any IO
4435          * to ensure the size we choose is acceptable to all devices.
4436          * The means one for each copy for write-out and an extra one for
4437          * read-in.
4438          * We store the read-in bio in ->master_bio and the others in
4439          * ->devs[x].bio and ->devs[x].repl_bio.
4440          */
4441         struct r10conf *conf = mddev->private;
4442         struct r10bio *r10_bio;
4443         sector_t next, safe, last;
4444         int max_sectors;
4445         int nr_sectors;
4446         int s;
4447         struct md_rdev *rdev;
4448         int need_flush = 0;
4449         struct bio *blist;
4450         struct bio *bio, *read_bio;
4451         int sectors_done = 0;
4452         struct page **pages;
4453
4454         if (sector_nr == 0) {
4455                 /* If restarting in the middle, skip the initial sectors */
4456                 if (mddev->reshape_backwards &&
4457                     conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4458                         sector_nr = (raid10_size(mddev, 0, 0)
4459                                      - conf->reshape_progress);
4460                 } else if (!mddev->reshape_backwards &&
4461                            conf->reshape_progress > 0)
4462                         sector_nr = conf->reshape_progress;
4463                 if (sector_nr) {
4464                         mddev->curr_resync_completed = sector_nr;
4465                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4466                         *skipped = 1;
4467                         return sector_nr;
4468                 }
4469         }
4470
4471         /* We don't use sector_nr to track where we are up to
4472          * as that doesn't work well for ->reshape_backwards.
4473          * So just use ->reshape_progress.
4474          */
4475         if (mddev->reshape_backwards) {
4476                 /* 'next' is the earliest device address that we might
4477                  * write to for this chunk in the new layout
4478                  */
4479                 next = first_dev_address(conf->reshape_progress - 1,
4480                                          &conf->geo);
4481
4482                 /* 'safe' is the last device address that we might read from
4483                  * in the old layout after a restart
4484                  */
4485                 safe = last_dev_address(conf->reshape_safe - 1,
4486                                         &conf->prev);
4487
4488                 if (next + conf->offset_diff < safe)
4489                         need_flush = 1;
4490
4491                 last = conf->reshape_progress - 1;
4492                 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4493                                                & conf->prev.chunk_mask);
4494                 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4495                         sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4496         } else {
4497                 /* 'next' is after the last device address that we
4498                  * might write to for this chunk in the new layout
4499                  */
4500                 next = last_dev_address(conf->reshape_progress, &conf->geo);
4501
4502                 /* 'safe' is the earliest device address that we might
4503                  * read from in the old layout after a restart
4504                  */
4505                 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4506
4507                 /* Need to update metadata if 'next' might be beyond 'safe'
4508                  * as that would possibly corrupt data
4509                  */
4510                 if (next > safe + conf->offset_diff)
4511                         need_flush = 1;
4512
4513                 sector_nr = conf->reshape_progress;
4514                 last  = sector_nr | (conf->geo.chunk_mask
4515                                      & conf->prev.chunk_mask);
4516
4517                 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4518                         last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4519         }
4520
4521         if (need_flush ||
4522             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4523                 /* Need to update reshape_position in metadata */
4524                 wait_barrier(conf);
4525                 mddev->reshape_position = conf->reshape_progress;
4526                 if (mddev->reshape_backwards)
4527                         mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4528                                 - conf->reshape_progress;
4529                 else
4530                         mddev->curr_resync_completed = conf->reshape_progress;
4531                 conf->reshape_checkpoint = jiffies;
4532                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4533                 md_wakeup_thread(mddev->thread);
4534                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4535                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4536                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4537                         allow_barrier(conf);
4538                         return sectors_done;
4539                 }
4540                 conf->reshape_safe = mddev->reshape_position;
4541                 allow_barrier(conf);
4542         }
4543
4544         raise_barrier(conf, 0);
4545 read_more:
4546         /* Now schedule reads for blocks from sector_nr to last */
4547         r10_bio = raid10_alloc_init_r10buf(conf);
4548         r10_bio->state = 0;
4549         raise_barrier(conf, 1);
4550         atomic_set(&r10_bio->remaining, 0);
4551         r10_bio->mddev = mddev;
4552         r10_bio->sector = sector_nr;
4553         set_bit(R10BIO_IsReshape, &r10_bio->state);
4554         r10_bio->sectors = last - sector_nr + 1;
4555         rdev = read_balance(conf, r10_bio, &max_sectors);
4556         BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4557
4558         if (!rdev) {
4559                 /* Cannot read from here, so need to record bad blocks
4560                  * on all the target devices.
4561                  */
4562                 // FIXME
4563                 mempool_free(r10_bio, &conf->r10buf_pool);
4564                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4565                 return sectors_done;
4566         }
4567
4568         read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4569
4570         bio_set_dev(read_bio, rdev->bdev);
4571         read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4572                                + rdev->data_offset);
4573         read_bio->bi_private = r10_bio;
4574         read_bio->bi_end_io = end_reshape_read;
4575         bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4576         read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4577         read_bio->bi_status = 0;
4578         read_bio->bi_vcnt = 0;
4579         read_bio->bi_iter.bi_size = 0;
4580         r10_bio->master_bio = read_bio;
4581         r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4582
4583         /*
4584          * Broadcast RESYNC message to other nodes, so all nodes would not
4585          * write to the region to avoid conflict.
4586         */
4587         if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4588                 struct mdp_superblock_1 *sb = NULL;
4589                 int sb_reshape_pos = 0;
4590
4591                 conf->cluster_sync_low = sector_nr;
4592                 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4593                 sb = page_address(rdev->sb_page);
4594                 if (sb) {
4595                         sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4596                         /*
4597                          * Set cluster_sync_low again if next address for array
4598                          * reshape is less than cluster_sync_low. Since we can't
4599                          * update cluster_sync_low until it has finished reshape.
4600                          */
4601                         if (sb_reshape_pos < conf->cluster_sync_low)
4602                                 conf->cluster_sync_low = sb_reshape_pos;
4603                 }
4604
4605                 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4606                                                           conf->cluster_sync_high);
4607         }
4608
4609         /* Now find the locations in the new layout */
4610         __raid10_find_phys(&conf->geo, r10_bio);
4611
4612         blist = read_bio;
4613         read_bio->bi_next = NULL;
4614
4615         rcu_read_lock();
4616         for (s = 0; s < conf->copies*2; s++) {
4617                 struct bio *b;
4618                 int d = r10_bio->devs[s/2].devnum;
4619                 struct md_rdev *rdev2;
4620                 if (s&1) {
4621                         rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4622                         b = r10_bio->devs[s/2].repl_bio;
4623                 } else {
4624                         rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4625                         b = r10_bio->devs[s/2].bio;
4626                 }
4627                 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4628                         continue;
4629
4630                 bio_set_dev(b, rdev2->bdev);
4631                 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4632                         rdev2->new_data_offset;
4633                 b->bi_end_io = end_reshape_write;
4634                 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4635                 b->bi_next = blist;
4636                 blist = b;
4637         }
4638
4639         /* Now add as many pages as possible to all of these bios. */
4640
4641         nr_sectors = 0;
4642         pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4643         for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4644                 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4645                 int len = (max_sectors - s) << 9;
4646                 if (len > PAGE_SIZE)
4647                         len = PAGE_SIZE;
4648                 for (bio = blist; bio ; bio = bio->bi_next) {
4649                         /*
4650                          * won't fail because the vec table is big enough
4651                          * to hold all these pages
4652                          */
4653                         bio_add_page(bio, page, len, 0);
4654                 }
4655                 sector_nr += len >> 9;
4656                 nr_sectors += len >> 9;
4657         }
4658         rcu_read_unlock();
4659         r10_bio->sectors = nr_sectors;
4660
4661         /* Now submit the read */
4662         md_sync_acct_bio(read_bio, r10_bio->sectors);
4663         atomic_inc(&r10_bio->remaining);
4664         read_bio->bi_next = NULL;
4665         generic_make_request(read_bio);
4666         sectors_done += nr_sectors;
4667         if (sector_nr <= last)
4668                 goto read_more;
4669
4670         lower_barrier(conf);
4671
4672         /* Now that we have done the whole section we can
4673          * update reshape_progress
4674          */
4675         if (mddev->reshape_backwards)
4676                 conf->reshape_progress -= sectors_done;
4677         else
4678                 conf->reshape_progress += sectors_done;
4679
4680         return sectors_done;
4681 }
4682
4683 static void end_reshape_request(struct r10bio *r10_bio);
4684 static int handle_reshape_read_error(struct mddev *mddev,
4685                                      struct r10bio *r10_bio);
4686 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4687 {
4688         /* Reshape read completed.  Hopefully we have a block
4689          * to write out.
4690          * If we got a read error then we do sync 1-page reads from
4691          * elsewhere until we find the data - or give up.
4692          */
4693         struct r10conf *conf = mddev->private;
4694         int s;
4695
4696         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4697                 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4698                         /* Reshape has been aborted */
4699                         md_done_sync(mddev, r10_bio->sectors, 0);
4700                         return;
4701                 }
4702
4703         /* We definitely have the data in the pages, schedule the
4704          * writes.
4705          */
4706         atomic_set(&r10_bio->remaining, 1);
4707         for (s = 0; s < conf->copies*2; s++) {
4708                 struct bio *b;
4709                 int d = r10_bio->devs[s/2].devnum;
4710                 struct md_rdev *rdev;
4711                 rcu_read_lock();
4712                 if (s&1) {
4713                         rdev = rcu_dereference(conf->mirrors[d].replacement);
4714                         b = r10_bio->devs[s/2].repl_bio;
4715                 } else {
4716                         rdev = rcu_dereference(conf->mirrors[d].rdev);
4717                         b = r10_bio->devs[s/2].bio;
4718                 }
4719                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4720                         rcu_read_unlock();
4721                         continue;
4722                 }
4723                 atomic_inc(&rdev->nr_pending);
4724                 rcu_read_unlock();
4725                 md_sync_acct_bio(b, r10_bio->sectors);
4726                 atomic_inc(&r10_bio->remaining);
4727                 b->bi_next = NULL;
4728                 generic_make_request(b);
4729         }
4730         end_reshape_request(r10_bio);
4731 }
4732
4733 static void end_reshape(struct r10conf *conf)
4734 {
4735         if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4736                 return;
4737
4738         spin_lock_irq(&conf->device_lock);
4739         conf->prev = conf->geo;
4740         md_finish_reshape(conf->mddev);
4741         smp_wmb();
4742         conf->reshape_progress = MaxSector;
4743         conf->reshape_safe = MaxSector;
4744         spin_unlock_irq(&conf->device_lock);
4745
4746         /* read-ahead size must cover two whole stripes, which is
4747          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4748          */
4749         if (conf->mddev->queue) {
4750                 int stripe = conf->geo.raid_disks *
4751                         ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4752                 stripe /= conf->geo.near_copies;
4753                 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4754                         conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4755         }
4756         conf->fullsync = 0;
4757 }
4758
4759 static void raid10_update_reshape_pos(struct mddev *mddev)
4760 {
4761         struct r10conf *conf = mddev->private;
4762         sector_t lo, hi;
4763
4764         md_cluster_ops->resync_info_get(mddev, &lo, &hi);
4765         if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
4766             || mddev->reshape_position == MaxSector)
4767                 conf->reshape_progress = mddev->reshape_position;
4768         else
4769                 WARN_ON_ONCE(1);
4770 }
4771
4772 static int handle_reshape_read_error(struct mddev *mddev,
4773                                      struct r10bio *r10_bio)
4774 {
4775         /* Use sync reads to get the blocks from somewhere else */
4776         int sectors = r10_bio->sectors;
4777         struct r10conf *conf = mddev->private;
4778         struct r10bio *r10b;
4779         int slot = 0;
4780         int idx = 0;
4781         struct page **pages;
4782
4783         r10b = kmalloc(sizeof(*r10b) +
4784                sizeof(struct r10dev) * conf->copies, GFP_NOIO);
4785         if (!r10b) {
4786                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4787                 return -ENOMEM;
4788         }
4789
4790         /* reshape IOs share pages from .devs[0].bio */
4791         pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4792
4793         r10b->sector = r10_bio->sector;
4794         __raid10_find_phys(&conf->prev, r10b);
4795
4796         while (sectors) {
4797                 int s = sectors;
4798                 int success = 0;
4799                 int first_slot = slot;
4800
4801                 if (s > (PAGE_SIZE >> 9))
4802                         s = PAGE_SIZE >> 9;
4803
4804                 rcu_read_lock();
4805                 while (!success) {
4806                         int d = r10b->devs[slot].devnum;
4807                         struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4808                         sector_t addr;
4809                         if (rdev == NULL ||
4810                             test_bit(Faulty, &rdev->flags) ||
4811                             !test_bit(In_sync, &rdev->flags))
4812                                 goto failed;
4813
4814                         addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4815                         atomic_inc(&rdev->nr_pending);
4816                         rcu_read_unlock();
4817                         success = sync_page_io(rdev,
4818                                                addr,
4819                                                s << 9,
4820                                                pages[idx],
4821                                                REQ_OP_READ, 0, false);
4822                         rdev_dec_pending(rdev, mddev);
4823                         rcu_read_lock();
4824                         if (success)
4825                                 break;
4826                 failed:
4827                         slot++;
4828                         if (slot >= conf->copies)
4829                                 slot = 0;
4830                         if (slot == first_slot)
4831                                 break;
4832                 }
4833                 rcu_read_unlock();
4834                 if (!success) {
4835                         /* couldn't read this block, must give up */
4836                         set_bit(MD_RECOVERY_INTR,
4837                                 &mddev->recovery);
4838                         kfree(r10b);
4839                         return -EIO;
4840                 }
4841                 sectors -= s;
4842                 idx++;
4843         }
4844         kfree(r10b);
4845         return 0;
4846 }
4847
4848 static void end_reshape_write(struct bio *bio)
4849 {
4850         struct r10bio *r10_bio = get_resync_r10bio(bio);
4851         struct mddev *mddev = r10_bio->mddev;
4852         struct r10conf *conf = mddev->private;
4853         int d;
4854         int slot;
4855         int repl;
4856         struct md_rdev *rdev = NULL;
4857
4858         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4859         if (repl)
4860                 rdev = conf->mirrors[d].replacement;
4861         if (!rdev) {
4862                 smp_mb();
4863                 rdev = conf->mirrors[d].rdev;
4864         }
4865
4866         if (bio->bi_status) {
4867                 /* FIXME should record badblock */
4868                 md_error(mddev, rdev);
4869         }
4870
4871         rdev_dec_pending(rdev, mddev);
4872         end_reshape_request(r10_bio);
4873 }
4874
4875 static void end_reshape_request(struct r10bio *r10_bio)
4876 {
4877         if (!atomic_dec_and_test(&r10_bio->remaining))
4878                 return;
4879         md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4880         bio_put(r10_bio->master_bio);
4881         put_buf(r10_bio);
4882 }
4883
4884 static void raid10_finish_reshape(struct mddev *mddev)
4885 {
4886         struct r10conf *conf = mddev->private;
4887
4888         if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4889                 return;
4890
4891         if (mddev->delta_disks > 0) {
4892                 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4893                         mddev->recovery_cp = mddev->resync_max_sectors;
4894                         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4895                 }
4896                 mddev->resync_max_sectors = mddev->array_sectors;
4897         } else {
4898                 int d;
4899                 rcu_read_lock();
4900                 for (d = conf->geo.raid_disks ;
4901                      d < conf->geo.raid_disks - mddev->delta_disks;
4902                      d++) {
4903                         struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4904                         if (rdev)
4905                                 clear_bit(In_sync, &rdev->flags);
4906                         rdev = rcu_dereference(conf->mirrors[d].replacement);
4907                         if (rdev)
4908                                 clear_bit(In_sync, &rdev->flags);
4909                 }
4910                 rcu_read_unlock();
4911         }
4912         mddev->layout = mddev->new_layout;
4913         mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4914         mddev->reshape_position = MaxSector;
4915         mddev->delta_disks = 0;
4916         mddev->reshape_backwards = 0;
4917 }
4918
4919 static struct md_personality raid10_personality =
4920 {
4921         .name           = "raid10",
4922         .level          = 10,
4923         .owner          = THIS_MODULE,
4924         .make_request   = raid10_make_request,
4925         .run            = raid10_run,
4926         .free           = raid10_free,
4927         .status         = raid10_status,
4928         .error_handler  = raid10_error,
4929         .hot_add_disk   = raid10_add_disk,
4930         .hot_remove_disk= raid10_remove_disk,
4931         .spare_active   = raid10_spare_active,
4932         .sync_request   = raid10_sync_request,
4933         .quiesce        = raid10_quiesce,
4934         .size           = raid10_size,
4935         .resize         = raid10_resize,
4936         .takeover       = raid10_takeover,
4937         .check_reshape  = raid10_check_reshape,
4938         .start_reshape  = raid10_start_reshape,
4939         .finish_reshape = raid10_finish_reshape,
4940         .update_reshape_pos = raid10_update_reshape_pos,
4941         .congested      = raid10_congested,
4942 };
4943
4944 static int __init raid_init(void)
4945 {
4946         return register_md_personality(&raid10_personality);
4947 }
4948
4949 static void raid_exit(void)
4950 {
4951         unregister_md_personality(&raid10_personality);
4952 }
4953
4954 module_init(raid_init);
4955 module_exit(raid_exit);
4956 MODULE_LICENSE("GPL");
4957 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4958 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4959 MODULE_ALIAS("md-raid10");
4960 MODULE_ALIAS("md-level-10");
4961
4962 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);