drivers/md/raid10.c

   1 /*
   2  * raid10.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 2000-2004 Neil Brown
   5  *
   6  * RAID-10 support for md.
   7  *
   8  * Base on code in raid1.c.  See raid1.c for further copyright information.
   9  *
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2, or (at your option)
  14  * any later version.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * (for example /usr/src/linux/COPYING); if not, write to the Free
  18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 #include <linux/slab.h>
  22 #include <linux/delay.h>
  23 #include <linux/blkdev.h>
  24 #include <linux/module.h>
  25 #include <linux/seq_file.h>
  26 #include <linux/ratelimit.h>
  27 #include <linux/kthread.h>
  28 #include <trace/events/block.h>
  29 #include "md.h"
  30 #include "raid10.h"
  31 #include "raid0.h"
  32 #include "bitmap.h"
  33
  34 /*
  35  * RAID10 provides a combination of RAID0 and RAID1 functionality.
  36  * The layout of data is defined by
  37  *    chunk_size
  38  *    raid_disks
  39  *    near_copies (stored in low byte of layout)
  40  *    far_copies (stored in second byte of layout)
  41  *    far_offset (stored in bit 16 of layout )
  42  *    use_far_sets (stored in bit 17 of layout )
  43  *    use_far_sets_bugfixed (stored in bit 18 of layout )
  44  *
  45  * The data to be stored is divided into chunks using chunksize.  Each device
  46  * is divided into far_copies sections.   In each section, chunks are laid out
  47  * in a style similar to raid0, but near_copies copies of each chunk is stored
  48  * (each on a different drive).  The starting device for each section is offset
  49  * near_copies from the starting device of the previous section.  Thus there
  50  * are (near_copies * far_copies) of each chunk, and each is on a different
  51  * drive.  near_copies and far_copies must be at least one, and their product
  52  * is at most raid_disks.
  53  *
  54  * If far_offset is true, then the far_copies are handled a bit differently.
  55  * The copies are still in different stripes, but instead of being very far
  56  * apart on disk, there are adjacent stripes.
  57  *
  58  * The far and offset algorithms are handled slightly differently if
  59  * 'use_far_sets' is true.  In this case, the array's devices are grouped into
  60  * sets that are (near_copies * far_copies) in size.  The far copied stripes
  61  * are still shifted by 'near_copies' devices, but this shifting stays confined
  62  * to the set rather than the entire array.  This is done to improve the number
  63  * of device combinations that can fail without causing the array to fail.
  64  * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
  65  * on a device):
  66  *    A B C D    A B C D E
  67  *      ...         ...
  68  *    D A B C    E A B C D
  69  * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
  70  *    [A B] [C D]    [A B] [C D E]
  71  *    |...| |...|    |...| | ... |
  72  *    [B A] [D C]    [B A] [E C D]
  73  */
  74
  75 /*
  76  * Number of guaranteed r10bios in case of extreme VM load:
  77  */
  78 #define NR_RAID10_BIOS 256
  79
  80 /* when we get a read error on a read-only array, we redirect to another
  81  * device without failing the first device, or trying to over-write to
  82  * correct the read error.  To keep track of bad blocks on a per-bio
  83  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
  84  */
  85 #define IO_BLOCKED ((struct bio *)1)
  86 /* When we successfully write to a known bad-block, we need to remove the
  87  * bad-block marking which must be done from process context.  So we record
  88  * the success by setting devs[n].bio to IO_MADE_GOOD
  89  */
  90 #define IO_MADE_GOOD ((struct bio *)2)
  91
  92 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
  93
  94 /* When there are this many requests queued to be written by
  95  * the raid10 thread, we become 'congested' to provide back-pressure
  96  * for writeback.
  97  */
  98 static int max_queued_requests = 1024;
  99
 100 static void allow_barrier(struct r10conf *conf);
 101 static void lower_barrier(struct r10conf *conf);
 102 static int _enough(struct r10conf *conf, int previous, int ignore);
 103 static int enough(struct r10conf *conf, int ignore);
 104 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
 105                                 int *skipped);
 106 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
 107 static void end_reshape_write(struct bio *bio);
 108 static void end_reshape(struct r10conf *conf);
 109
 110 #define raid10_log(md, fmt, args...)                            \
 111         do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
 112
 113 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
 114 {
 115         struct r10conf *conf = data;
 116         int size = offsetof(struct r10bio, devs[conf->copies]);
 117
 118         /* allocate a r10bio with room for raid_disks entries in the
 119          * bios array */
 120         return kzalloc(size, gfp_flags);
 121 }
 122
 123 static void r10bio_pool_free(void *r10_bio, void *data)
 124 {
 125         kfree(r10_bio);
 126 }
 127
 128 /* Maximum size of each resync request */
 129 #define RESYNC_BLOCK_SIZE (64*1024)
 130 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
 131 /* amount of memory to reserve for resync requests */
 132 #define RESYNC_WINDOW (1024*1024)
 133 /* maximum number of concurrent requests, memory permitting */
 134 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
 135
 136 /*
 137  * When performing a resync, we need to read and compare, so
 138  * we need as many pages are there are copies.
 139  * When performing a recovery, we need 2 bios, one for read,
 140  * one for write (we recover only one drive per r10buf)
 141  *
 142  */
 143 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
 144 {
 145         struct r10conf *conf = data;
 146         struct page *page;
 147         struct r10bio *r10_bio;
 148         struct bio *bio;
 149         int i, j;
 150         int nalloc;
 151
 152         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 153         if (!r10_bio)
 154                 return NULL;
 155
 156         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
 157             test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
 158                 nalloc = conf->copies; /* resync */
 159         else
 160                 nalloc = 2; /* recovery */
 161
 162         /*
 163          * Allocate bios.
 164          */
 165         for (j = nalloc ; j-- ; ) {
 166                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 167                 if (!bio)
 168                         goto out_free_bio;
 169                 r10_bio->devs[j].bio = bio;
 170                 if (!conf->have_replacement)
 171                         continue;
 172                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 173                 if (!bio)
 174                         goto out_free_bio;
 175                 r10_bio->devs[j].repl_bio = bio;
 176         }
 177         /*
 178          * Allocate RESYNC_PAGES data pages and attach them
 179          * where needed.
 180          */
 181         for (j = 0 ; j < nalloc; j++) {
 182                 struct bio *rbio = r10_bio->devs[j].repl_bio;
 183                 bio = r10_bio->devs[j].bio;
 184                 for (i = 0; i < RESYNC_PAGES; i++) {
 185                         if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
 186                                                &conf->mddev->recovery)) {
 187                                 /* we can share bv_page's during recovery
 188                                  * and reshape */
 189                                 struct bio *rbio = r10_bio->devs[0].bio;
 190                                 page = rbio->bi_io_vec[i].bv_page;
 191                                 get_page(page);
 192                         } else
 193                                 page = alloc_page(gfp_flags);
 194                         if (unlikely(!page))
 195                                 goto out_free_pages;
 196
 197                         bio->bi_io_vec[i].bv_page = page;
 198                         if (rbio)
 199                                 rbio->bi_io_vec[i].bv_page = page;
 200                 }
 201         }
 202
 203         return r10_bio;
 204
 205 out_free_pages:
 206         for ( ; i > 0 ; i--)
 207                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
 208         while (j--)
 209                 for (i = 0; i < RESYNC_PAGES ; i++)
 210                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 211         j = 0;
 212 out_free_bio:
 213         for ( ; j < nalloc; j++) {
 214                 if (r10_bio->devs[j].bio)
 215                         bio_put(r10_bio->devs[j].bio);
 216                 if (r10_bio->devs[j].repl_bio)
 217                         bio_put(r10_bio->devs[j].repl_bio);
 218         }
 219         r10bio_pool_free(r10_bio, conf);
 220         return NULL;
 221 }
 222
 223 static void r10buf_pool_free(void *__r10_bio, void *data)
 224 {
 225         int i;
 226         struct r10conf *conf = data;
 227         struct r10bio *r10bio = __r10_bio;
 228         int j;
 229
 230         for (j=0; j < conf->copies; j++) {
 231                 struct bio *bio = r10bio->devs[j].bio;
 232                 if (bio) {
 233                         for (i = 0; i < RESYNC_PAGES; i++) {
 234                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 235                                 bio->bi_io_vec[i].bv_page = NULL;
 236                         }
 237                         bio_put(bio);
 238                 }
 239                 bio = r10bio->devs[j].repl_bio;
 240                 if (bio)
 241                         bio_put(bio);
 242         }
 243         r10bio_pool_free(r10bio, conf);
 244 }
 245
 246 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
 247 {
 248         int i;
 249
 250         for (i = 0; i < conf->copies; i++) {
 251                 struct bio **bio = & r10_bio->devs[i].bio;
 252                 if (!BIO_SPECIAL(*bio))
 253                         bio_put(*bio);
 254                 *bio = NULL;
 255                 bio = &r10_bio->devs[i].repl_bio;
 256                 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
 257                         bio_put(*bio);
 258                 *bio = NULL;
 259         }
 260 }
 261
 262 static void free_r10bio(struct r10bio *r10_bio)
 263 {
 264         struct r10conf *conf = r10_bio->mddev->private;
 265
 266         put_all_bios(conf, r10_bio);
 267         mempool_free(r10_bio, conf->r10bio_pool);
 268 }
 269
 270 static void put_buf(struct r10bio *r10_bio)
 271 {
 272         struct r10conf *conf = r10_bio->mddev->private;
 273
 274         mempool_free(r10_bio, conf->r10buf_pool);
 275
 276         lower_barrier(conf);
 277 }
 278
 279 static void reschedule_retry(struct r10bio *r10_bio)
 280 {
 281         unsigned long flags;
 282         struct mddev *mddev = r10_bio->mddev;
 283         struct r10conf *conf = mddev->private;
 284
 285         spin_lock_irqsave(&conf->device_lock, flags);
 286         list_add(&r10_bio->retry_list, &conf->retry_list);
 287         conf->nr_queued ++;
 288         spin_unlock_irqrestore(&conf->device_lock, flags);
 289
 290         /* wake up frozen array... */
 291         wake_up(&conf->wait_barrier);
 292
 293         md_wakeup_thread(mddev->thread);
 294 }
 295
 296 /*
 297  * raid_end_bio_io() is called when we have finished servicing a mirrored
 298  * operation and are ready to return a success/failure code to the buffer
 299  * cache layer.
 300  */
 301 static void raid_end_bio_io(struct r10bio *r10_bio)
 302 {
 303         struct bio *bio = r10_bio->master_bio;
 304         int done;
 305         struct r10conf *conf = r10_bio->mddev->private;
 306
 307         if (bio->bi_phys_segments) {
 308                 unsigned long flags;
 309                 spin_lock_irqsave(&conf->device_lock, flags);
 310                 bio->bi_phys_segments--;
 311                 done = (bio->bi_phys_segments == 0);
 312                 spin_unlock_irqrestore(&conf->device_lock, flags);
 313         } else
 314                 done = 1;
 315         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
 316                 bio->bi_error = -EIO;
 317         if (done) {
 318                 bio_endio(bio);
 319                 /*
 320                  * Wake up any possible resync thread that waits for the device
 321                  * to go idle.
 322                  */
 323                 allow_barrier(conf);
 324         }
 325         free_r10bio(r10_bio);
 326 }
 327
 328 /*
 329  * Update disk head position estimator based on IRQ completion info.
 330  */
 331 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
 332 {
 333         struct r10conf *conf = r10_bio->mddev->private;
 334
 335         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 336                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 337 }
 338
 339 /*
 340  * Find the disk number which triggered given bio
 341  */
 342 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
 343                          struct bio *bio, int *slotp, int *replp)
 344 {
 345         int slot;
 346         int repl = 0;
 347
 348         for (slot = 0; slot < conf->copies; slot++) {
 349                 if (r10_bio->devs[slot].bio == bio)
 350                         break;
 351                 if (r10_bio->devs[slot].repl_bio == bio) {
 352                         repl = 1;
 353                         break;
 354                 }
 355         }
 356
 357         BUG_ON(slot == conf->copies);
 358         update_head_pos(slot, r10_bio);
 359
 360         if (slotp)
 361                 *slotp = slot;
 362         if (replp)
 363                 *replp = repl;
 364         return r10_bio->devs[slot].devnum;
 365 }
 366
 367 static void raid10_end_read_request(struct bio *bio)
 368 {
 369         int uptodate = !bio->bi_error;
 370         struct r10bio *r10_bio = bio->bi_private;
 371         int slot, dev;
 372         struct md_rdev *rdev;
 373         struct r10conf *conf = r10_bio->mddev->private;
 374
 375         slot = r10_bio->read_slot;
 376         dev = r10_bio->devs[slot].devnum;
 377         rdev = r10_bio->devs[slot].rdev;
 378         /*
 379          * this branch is our 'one mirror IO has finished' event handler:
 380          */
 381         update_head_pos(slot, r10_bio);
 382
 383         if (uptodate) {
 384                 /*
 385                  * Set R10BIO_Uptodate in our master bio, so that
 386                  * we will return a good error code to the higher
 387                  * levels even if IO on some other mirrored buffer fails.
 388                  *
 389                  * The 'master' represents the composite IO operation to
 390                  * user-side. So if something waits for IO, then it will
 391                  * wait for the 'master' bio.
 392                  */
 393                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 394         } else {
 395                 /* If all other devices that store this block have
 396                  * failed, we want to return the error upwards rather
 397                  * than fail the last device.  Here we redefine
 398                  * "uptodate" to mean "Don't want to retry"
 399                  */
 400                 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
 401                              rdev->raid_disk))
 402                         uptodate = 1;
 403         }
 404         if (uptodate) {
 405                 raid_end_bio_io(r10_bio);
 406                 rdev_dec_pending(rdev, conf->mddev);
 407         } else {
 408                 /*
 409                  * oops, read error - keep the refcount on the rdev
 410                  */
 411                 char b[BDEVNAME_SIZE];
 412                 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
 413                                    mdname(conf->mddev),
 414                                    bdevname(rdev->bdev, b),
 415                                    (unsigned long long)r10_bio->sector);
 416                 set_bit(R10BIO_ReadError, &r10_bio->state);
 417                 reschedule_retry(r10_bio);
 418         }
 419 }
 420
 421 static void close_write(struct r10bio *r10_bio)
 422 {
 423         /* clear the bitmap if all writes complete successfully */
 424         bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 425                         r10_bio->sectors,
 426                         !test_bit(R10BIO_Degraded, &r10_bio->state),
 427                         0);
 428         md_write_end(r10_bio->mddev);
 429 }
 430
 431 static void one_write_done(struct r10bio *r10_bio)
 432 {
 433         if (atomic_dec_and_test(&r10_bio->remaining)) {
 434                 if (test_bit(R10BIO_WriteError, &r10_bio->state))
 435                         reschedule_retry(r10_bio);
 436                 else {
 437                         close_write(r10_bio);
 438                         if (test_bit(R10BIO_MadeGood, &r10_bio->state))
 439                                 reschedule_retry(r10_bio);
 440                         else
 441                                 raid_end_bio_io(r10_bio);
 442                 }
 443         }
 444 }
 445
 446 static void raid10_end_write_request(struct bio *bio)
 447 {
 448         struct r10bio *r10_bio = bio->bi_private;
 449         int dev;
 450         int dec_rdev = 1;
 451         struct r10conf *conf = r10_bio->mddev->private;
 452         int slot, repl;
 453         struct md_rdev *rdev = NULL;
 454         struct bio *to_put = NULL;
 455         bool discard_error;
 456
 457         discard_error = bio->bi_error && bio_op(bio) == REQ_OP_DISCARD;
 458
 459         dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
 460
 461         if (repl)
 462                 rdev = conf->mirrors[dev].replacement;
 463         if (!rdev) {
 464                 smp_rmb();
 465                 repl = 0;
 466                 rdev = conf->mirrors[dev].rdev;
 467         }
 468         /*
 469          * this branch is our 'one mirror IO has finished' event handler:
 470          */
 471         if (bio->bi_error && !discard_error) {
 472                 if (repl)
 473                         /* Never record new bad blocks to replacement,
 474                          * just fail it.
 475                          */
 476                         md_error(rdev->mddev, rdev);
 477                 else {
 478                         set_bit(WriteErrorSeen, &rdev->flags);
 479                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
 480                                 set_bit(MD_RECOVERY_NEEDED,
 481                                         &rdev->mddev->recovery);
 482
 483                         dec_rdev = 0;
 484                         if (test_bit(FailFast, &rdev->flags) &&
 485                             (bio->bi_opf & MD_FAILFAST)) {
 486                                 md_error(rdev->mddev, rdev);
 487                                 if (!test_bit(Faulty, &rdev->flags))
 488                                         /* This is the only remaining device,
 489                                          * We need to retry the write without
 490                                          * FailFast
 491                                          */
 492                                         set_bit(R10BIO_WriteError, &r10_bio->state);
 493                                 else {
 494                                         r10_bio->devs[slot].bio = NULL;
 495                                         to_put = bio;
 496                                         dec_rdev = 1;
 497                                 }
 498                         } else
 499                                 set_bit(R10BIO_WriteError, &r10_bio->state);
 500                 }
 501         } else {
 502                 /*
 503                  * Set R10BIO_Uptodate in our master bio, so that
 504                  * we will return a good error code for to the higher
 505                  * levels even if IO on some other mirrored buffer fails.
 506                  *
 507                  * The 'master' represents the composite IO operation to
 508                  * user-side. So if something waits for IO, then it will
 509                  * wait for the 'master' bio.
 510                  */
 511                 sector_t first_bad;
 512                 int bad_sectors;
 513
 514                 /*
 515                  * Do not set R10BIO_Uptodate if the current device is
 516                  * rebuilding or Faulty. This is because we cannot use
 517                  * such device for properly reading the data back (we could
 518                  * potentially use it, if the current write would have felt
 519                  * before rdev->recovery_offset, but for simplicity we don't
 520                  * check this here.
 521                  */
 522                 if (test_bit(In_sync, &rdev->flags) &&
 523                     !test_bit(Faulty, &rdev->flags))
 524                         set_bit(R10BIO_Uptodate, &r10_bio->state);
 525
 526                 /* Maybe we can clear some bad blocks. */
 527                 if (is_badblock(rdev,
 528                                 r10_bio->devs[slot].addr,
 529                                 r10_bio->sectors,
 530                                 &first_bad, &bad_sectors) && !discard_error) {
 531                         bio_put(bio);
 532                         if (repl)
 533                                 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
 534                         else
 535                                 r10_bio->devs[slot].bio = IO_MADE_GOOD;
 536                         dec_rdev = 0;
 537                         set_bit(R10BIO_MadeGood, &r10_bio->state);
 538                 }
 539         }
 540
 541         /*
 542          *
 543          * Let's see if all mirrored write operations have finished
 544          * already.
 545          */
 546         one_write_done(r10_bio);
 547         if (dec_rdev)
 548                 rdev_dec_pending(rdev, conf->mddev);
 549         if (to_put)
 550                 bio_put(to_put);
 551 }
 552
 553 /*
 554  * RAID10 layout manager
 555  * As well as the chunksize and raid_disks count, there are two
 556  * parameters: near_copies and far_copies.
 557  * near_copies * far_copies must be <= raid_disks.
 558  * Normally one of these will be 1.
 559  * If both are 1, we get raid0.
 560  * If near_copies == raid_disks, we get raid1.
 561  *
 562  * Chunks are laid out in raid0 style with near_copies copies of the
 563  * first chunk, followed by near_copies copies of the next chunk and
 564  * so on.
 565  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 566  * as described above, we start again with a device offset of near_copies.
 567  * So we effectively have another copy of the whole array further down all
 568  * the drives, but with blocks on different drives.
 569  * With this layout, and block is never stored twice on the one device.
 570  *
 571  * raid10_find_phys finds the sector offset of a given virtual sector
 572  * on each device that it is on.
 573  *
 574  * raid10_find_virt does the reverse mapping, from a device and a
 575  * sector offset to a virtual address
 576  */
 577
 578 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
 579 {
 580         int n,f;
 581         sector_t sector;
 582         sector_t chunk;
 583         sector_t stripe;
 584         int dev;
 585         int slot = 0;
 586         int last_far_set_start, last_far_set_size;
 587
 588         last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
 589         last_far_set_start *= geo->far_set_size;
 590
 591         last_far_set_size = geo->far_set_size;
 592         last_far_set_size += (geo->raid_disks % geo->far_set_size);
 593
 594         /* now calculate first sector/dev */
 595         chunk = r10bio->sector >> geo->chunk_shift;
 596         sector = r10bio->sector & geo->chunk_mask;
 597
 598         chunk *= geo->near_copies;
 599         stripe = chunk;
 600         dev = sector_div(stripe, geo->raid_disks);
 601         if (geo->far_offset)
 602                 stripe *= geo->far_copies;
 603
 604         sector += stripe << geo->chunk_shift;
 605
 606         /* and calculate all the others */
 607         for (n = 0; n < geo->near_copies; n++) {
 608                 int d = dev;
 609                 int set;
 610                 sector_t s = sector;
 611                 r10bio->devs[slot].devnum = d;
 612                 r10bio->devs[slot].addr = s;
 613                 slot++;
 614
 615                 for (f = 1; f < geo->far_copies; f++) {
 616                         set = d / geo->far_set_size;
 617                         d += geo->near_copies;
 618
 619                         if ((geo->raid_disks % geo->far_set_size) &&
 620                             (d > last_far_set_start)) {
 621                                 d -= last_far_set_start;
 622                                 d %= last_far_set_size;
 623                                 d += last_far_set_start;
 624                         } else {
 625                                 d %= geo->far_set_size;
 626                                 d += geo->far_set_size * set;
 627                         }
 628                         s += geo->stride;
 629                         r10bio->devs[slot].devnum = d;
 630                         r10bio->devs[slot].addr = s;
 631                         slot++;
 632                 }
 633                 dev++;
 634                 if (dev >= geo->raid_disks) {
 635                         dev = 0;
 636                         sector += (geo->chunk_mask + 1);
 637                 }
 638         }
 639 }
 640
 641 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
 642 {
 643         struct geom *geo = &conf->geo;
 644
 645         if (conf->reshape_progress != MaxSector &&
 646             ((r10bio->sector >= conf->reshape_progress) !=
 647              conf->mddev->reshape_backwards)) {
 648                 set_bit(R10BIO_Previous, &r10bio->state);
 649                 geo = &conf->prev;
 650         } else
 651                 clear_bit(R10BIO_Previous, &r10bio->state);
 652
 653         __raid10_find_phys(geo, r10bio);
 654 }
 655
 656 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
 657 {
 658         sector_t offset, chunk, vchunk;
 659         /* Never use conf->prev as this is only called during resync
 660          * or recovery, so reshape isn't happening
 661          */
 662         struct geom *geo = &conf->geo;
 663         int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
 664         int far_set_size = geo->far_set_size;
 665         int last_far_set_start;
 666
 667         if (geo->raid_disks % geo->far_set_size) {
 668                 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
 669                 last_far_set_start *= geo->far_set_size;
 670
 671                 if (dev >= last_far_set_start) {
 672                         far_set_size = geo->far_set_size;
 673                         far_set_size += (geo->raid_disks % geo->far_set_size);
 674                         far_set_start = last_far_set_start;
 675                 }
 676         }
 677
 678         offset = sector & geo->chunk_mask;
 679         if (geo->far_offset) {
 680                 int fc;
 681                 chunk = sector >> geo->chunk_shift;
 682                 fc = sector_div(chunk, geo->far_copies);
 683                 dev -= fc * geo->near_copies;
 684                 if (dev < far_set_start)
 685                         dev += far_set_size;
 686         } else {
 687                 while (sector >= geo->stride) {
 688                         sector -= geo->stride;
 689                         if (dev < (geo->near_copies + far_set_start))
 690                                 dev += far_set_size - geo->near_copies;
 691                         else
 692                                 dev -= geo->near_copies;
 693                 }
 694                 chunk = sector >> geo->chunk_shift;
 695         }
 696         vchunk = chunk * geo->raid_disks + dev;
 697         sector_div(vchunk, geo->near_copies);
 698         return (vchunk << geo->chunk_shift) + offset;
 699 }
 700
 701 /*
 702  * This routine returns the disk from which the requested read should
 703  * be done. There is a per-array 'next expected sequential IO' sector
 704  * number - if this matches on the next IO then we use the last disk.
 705  * There is also a per-disk 'last know head position' sector that is
 706  * maintained from IRQ contexts, both the normal and the resync IO
 707  * completion handlers update this position correctly. If there is no
 708  * perfect sequential match then we pick the disk whose head is closest.
 709  *
 710  * If there are 2 mirrors in the same 2 devices, performance degrades
 711  * because position is mirror, not device based.
 712  *
 713  * The rdev for the device selected will have nr_pending incremented.
 714  */
 715
 716 /*
 717  * FIXME: possibly should rethink readbalancing and do it differently
 718  * depending on near_copies / far_copies geometry.
 719  */
 720 static struct md_rdev *read_balance(struct r10conf *conf,
 721                                     struct r10bio *r10_bio,
 722                                     int *max_sectors)
 723 {
 724         const sector_t this_sector = r10_bio->sector;
 725         int disk, slot;
 726         int sectors = r10_bio->sectors;
 727         int best_good_sectors;
 728         sector_t new_distance, best_dist;
 729         struct md_rdev *best_rdev, *rdev = NULL;
 730         int do_balance;
 731         int best_slot;
 732         struct geom *geo = &conf->geo;
 733
 734         raid10_find_phys(conf, r10_bio);
 735         rcu_read_lock();
 736         sectors = r10_bio->sectors;
 737         best_slot = -1;
 738         best_rdev = NULL;
 739         best_dist = MaxSector;
 740         best_good_sectors = 0;
 741         do_balance = 1;
 742         clear_bit(R10BIO_FailFast, &r10_bio->state);
 743         /*
 744          * Check if we can balance. We can balance on the whole
 745          * device if no resync is going on (recovery is ok), or below
 746          * the resync window. We take the first readable disk when
 747          * above the resync window.
 748          */
 749         if (conf->mddev->recovery_cp < MaxSector
 750             && (this_sector + sectors >= conf->next_resync))
 751                 do_balance = 0;
 752
 753         for (slot = 0; slot < conf->copies ; slot++) {
 754                 sector_t first_bad;
 755                 int bad_sectors;
 756                 sector_t dev_sector;
 757
 758                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
 759                         continue;
 760                 disk = r10_bio->devs[slot].devnum;
 761                 rdev = rcu_dereference(conf->mirrors[disk].replacement);
 762                 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
 763                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 764                         rdev = rcu_dereference(conf->mirrors[disk].rdev);
 765                 if (rdev == NULL ||
 766                     test_bit(Faulty, &rdev->flags))
 767                         continue;
 768                 if (!test_bit(In_sync, &rdev->flags) &&
 769                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 770                         continue;
 771
 772                 dev_sector = r10_bio->devs[slot].addr;
 773                 if (is_badblock(rdev, dev_sector, sectors,
 774                                 &first_bad, &bad_sectors)) {
 775                         if (best_dist < MaxSector)
 776                                 /* Already have a better slot */
 777                                 continue;
 778                         if (first_bad <= dev_sector) {
 779                                 /* Cannot read here.  If this is the
 780                                  * 'primary' device, then we must not read
 781                                  * beyond 'bad_sectors' from another device.
 782                                  */
 783                                 bad_sectors -= (dev_sector - first_bad);
 784                                 if (!do_balance && sectors > bad_sectors)
 785                                         sectors = bad_sectors;
 786                                 if (best_good_sectors > sectors)
 787                                         best_good_sectors = sectors;
 788                         } else {
 789                                 sector_t good_sectors =
 790                                         first_bad - dev_sector;
 791                                 if (good_sectors > best_good_sectors) {
 792                                         best_good_sectors = good_sectors;
 793                                         best_slot = slot;
 794                                         best_rdev = rdev;
 795                                 }
 796                                 if (!do_balance)
 797                                         /* Must read from here */
 798                                         break;
 799                         }
 800                         continue;
 801                 } else
 802                         best_good_sectors = sectors;
 803
 804                 if (!do_balance)
 805                         break;
 806
 807                 if (best_slot >= 0)
 808                         /* At least 2 disks to choose from so failfast is OK */
 809                         set_bit(R10BIO_FailFast, &r10_bio->state);
 810                 /* This optimisation is debatable, and completely destroys
 811                  * sequential read speed for 'far copies' arrays.  So only
 812                  * keep it for 'near' arrays, and review those later.
 813                  */
 814                 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
 815                         new_distance = 0;
 816
 817                 /* for far > 1 always use the lowest address */
 818                 else if (geo->far_copies > 1)
 819                         new_distance = r10_bio->devs[slot].addr;
 820                 else
 821                         new_distance = abs(r10_bio->devs[slot].addr -
 822                                            conf->mirrors[disk].head_position);
 823                 if (new_distance < best_dist) {
 824                         best_dist = new_distance;
 825                         best_slot = slot;
 826                         best_rdev = rdev;
 827                 }
 828         }
 829         if (slot >= conf->copies) {
 830                 slot = best_slot;
 831                 rdev = best_rdev;
 832         }
 833
 834         if (slot >= 0) {
 835                 atomic_inc(&rdev->nr_pending);
 836                 r10_bio->read_slot = slot;
 837         } else
 838                 rdev = NULL;
 839         rcu_read_unlock();
 840         *max_sectors = best_good_sectors;
 841
 842         return rdev;
 843 }
 844
 845 static int raid10_congested(struct mddev *mddev, int bits)
 846 {
 847         struct r10conf *conf = mddev->private;
 848         int i, ret = 0;
 849
 850         if ((bits & (1 << WB_async_congested)) &&
 851             conf->pending_count >= max_queued_requests)
 852                 return 1;
 853
 854         rcu_read_lock();
 855         for (i = 0;
 856              (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
 857                      && ret == 0;
 858              i++) {
 859                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 860                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 861                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 862
 863                         ret |= bdi_congested(q->backing_dev_info, bits);
 864                 }
 865         }
 866         rcu_read_unlock();
 867         return ret;
 868 }
 869
 870 static void flush_pending_writes(struct r10conf *conf)
 871 {
 872         /* Any writes that have been queued but are awaiting
 873          * bitmap updates get flushed here.
 874          */
 875         spin_lock_irq(&conf->device_lock);
 876
 877         if (conf->pending_bio_list.head) {
 878                 struct bio *bio;
 879                 bio = bio_list_get(&conf->pending_bio_list);
 880                 conf->pending_count = 0;
 881                 spin_unlock_irq(&conf->device_lock);
 882                 /* flush any pending bitmap writes to disk
 883                  * before proceeding w/ I/O */
 884                 bitmap_unplug(conf->mddev->bitmap);
 885                 wake_up(&conf->wait_barrier);
 886
 887                 while (bio) { /* submit pending writes */
 888                         struct bio *next = bio->bi_next;
 889                         struct md_rdev *rdev = (void*)bio->bi_bdev;
 890                         bio->bi_next = NULL;
 891                         bio->bi_bdev = rdev->bdev;
 892                         if (test_bit(Faulty, &rdev->flags)) {
 893                                 bio->bi_error = -EIO;
 894                                 bio_endio(bio);
 895                         } else if (unlikely((bio_op(bio) ==  REQ_OP_DISCARD) &&
 896                                             !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 897                                 /* Just ignore it */
 898                                 bio_endio(bio);
 899                         else
 900                                 generic_make_request(bio);
 901                         bio = next;
 902                 }
 903         } else
 904                 spin_unlock_irq(&conf->device_lock);
 905 }
 906
 907 /* Barriers....
 908  * Sometimes we need to suspend IO while we do something else,
 909  * either some resync/recovery, or reconfigure the array.
 910  * To do this we raise a 'barrier'.
 911  * The 'barrier' is a counter that can be raised multiple times
 912  * to count how many activities are happening which preclude
 913  * normal IO.
 914  * We can only raise the barrier if there is no pending IO.
 915  * i.e. if nr_pending == 0.
 916  * We choose only to raise the barrier if no-one is waiting for the
 917  * barrier to go down.  This means that as soon as an IO request
 918  * is ready, no other operations which require a barrier will start
 919  * until the IO request has had a chance.
 920  *
 921  * So: regular IO calls 'wait_barrier'.  When that returns there
 922  *    is no backgroup IO happening,  It must arrange to call
 923  *    allow_barrier when it has finished its IO.
 924  * backgroup IO calls must call raise_barrier.  Once that returns
 925  *    there is no normal IO happeing.  It must arrange to call
 926  *    lower_barrier when the particular background IO completes.
 927  */
 928
 929 static void raise_barrier(struct r10conf *conf, int force)
 930 {
 931         BUG_ON(force && !conf->barrier);
 932         spin_lock_irq(&conf->resync_lock);
 933
 934         /* Wait until no block IO is waiting (unless 'force') */
 935         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 936                             conf->resync_lock);
 937
 938         /* block any new IO from starting */
 939         conf->barrier++;
 940
 941         /* Now wait for all pending IO to complete */
 942         wait_event_lock_irq(conf->wait_barrier,
 943                             !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
 944                             conf->resync_lock);
 945
 946         spin_unlock_irq(&conf->resync_lock);
 947 }
 948
 949 static void lower_barrier(struct r10conf *conf)
 950 {
 951         unsigned long flags;
 952         spin_lock_irqsave(&conf->resync_lock, flags);
 953         conf->barrier--;
 954         spin_unlock_irqrestore(&conf->resync_lock, flags);
 955         wake_up(&conf->wait_barrier);
 956 }
 957
 958 static void wait_barrier(struct r10conf *conf)
 959 {
 960         spin_lock_irq(&conf->resync_lock);
 961         if (conf->barrier) {
 962                 conf->nr_waiting++;
 963                 /* Wait for the barrier to drop.
 964                  * However if there are already pending
 965                  * requests (preventing the barrier from
 966                  * rising completely), and the
 967                  * pre-process bio queue isn't empty,
 968                  * then don't wait, as we need to empty
 969                  * that queue to get the nr_pending
 970                  * count down.
 971                  */
 972                 raid10_log(conf->mddev, "wait barrier");
 973                 wait_event_lock_irq(conf->wait_barrier,
 974                                     !conf->barrier ||
 975                                     (atomic_read(&conf->nr_pending) &&
 976                                      current->bio_list &&
 977                                      !bio_list_empty(current->bio_list)),
 978                                     conf->resync_lock);
 979                 conf->nr_waiting--;
 980                 if (!conf->nr_waiting)
 981                         wake_up(&conf->wait_barrier);
 982         }
 983         atomic_inc(&conf->nr_pending);
 984         spin_unlock_irq(&conf->resync_lock);
 985 }
 986
 987 static void allow_barrier(struct r10conf *conf)
 988 {
 989         if ((atomic_dec_and_test(&conf->nr_pending)) ||
 990                         (conf->array_freeze_pending))
 991                 wake_up(&conf->wait_barrier);
 992 }
 993
 994 static void freeze_array(struct r10conf *conf, int extra)
 995 {
 996         /* stop syncio and normal IO and wait for everything to
 997          * go quiet.
 998          * We increment barrier and nr_waiting, and then
 999          * wait until nr_pending match nr_queued+extra
1000          * This is called in the context of one normal IO request
1001          * that has failed. Thus any sync request that might be pending
1002          * will be blocked by nr_pending, and we need to wait for
1003          * pending IO requests to complete or be queued for re-try.
1004          * Thus the number queued (nr_queued) plus this request (extra)
1005          * must match the number of pending IOs (nr_pending) before
1006          * we continue.
1007          */
1008         spin_lock_irq(&conf->resync_lock);
1009         conf->array_freeze_pending++;
1010         conf->barrier++;
1011         conf->nr_waiting++;
1012         wait_event_lock_irq_cmd(conf->wait_barrier,
1013                                 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1014                                 conf->resync_lock,
1015                                 flush_pending_writes(conf));
1016
1017         conf->array_freeze_pending--;
1018         spin_unlock_irq(&conf->resync_lock);
1019 }
1020
1021 static void unfreeze_array(struct r10conf *conf)
1022 {
1023         /* reverse the effect of the freeze */
1024         spin_lock_irq(&conf->resync_lock);
1025         conf->barrier--;
1026         conf->nr_waiting--;
1027         wake_up(&conf->wait_barrier);
1028         spin_unlock_irq(&conf->resync_lock);
1029 }
1030
1031 static sector_t choose_data_offset(struct r10bio *r10_bio,
1032                                    struct md_rdev *rdev)
1033 {
1034         if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1035             test_bit(R10BIO_Previous, &r10_bio->state))
1036                 return rdev->data_offset;
1037         else
1038                 return rdev->new_data_offset;
1039 }
1040
1041 struct raid10_plug_cb {
1042         struct blk_plug_cb      cb;
1043         struct bio_list         pending;
1044         int                     pending_cnt;
1045 };
1046
1047 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1048 {
1049         struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1050                                                    cb);
1051         struct mddev *mddev = plug->cb.data;
1052         struct r10conf *conf = mddev->private;
1053         struct bio *bio;
1054
1055         if (from_schedule || current->bio_list) {
1056                 spin_lock_irq(&conf->device_lock);
1057                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1058                 conf->pending_count += plug->pending_cnt;
1059                 spin_unlock_irq(&conf->device_lock);
1060                 wake_up(&conf->wait_barrier);
1061                 md_wakeup_thread(mddev->thread);
1062                 kfree(plug);
1063                 return;
1064         }
1065
1066         /* we aren't scheduling, so we can do the write-out directly. */
1067         bio = bio_list_get(&plug->pending);
1068         bitmap_unplug(mddev->bitmap);
1069         wake_up(&conf->wait_barrier);
1070
1071         while (bio) { /* submit pending writes */
1072                 struct bio *next = bio->bi_next;
1073                 struct md_rdev *rdev = (void*)bio->bi_bdev;
1074                 bio->bi_next = NULL;
1075                 bio->bi_bdev = rdev->bdev;
1076                 if (test_bit(Faulty, &rdev->flags)) {
1077                         bio->bi_error = -EIO;
1078                         bio_endio(bio);
1079                 } else if (unlikely((bio_op(bio) ==  REQ_OP_DISCARD) &&
1080                                     !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1081                         /* Just ignore it */
1082                         bio_endio(bio);
1083                 else
1084                         generic_make_request(bio);
1085                 bio = next;
1086         }
1087         kfree(plug);
1088 }
1089
1090 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1091                                 struct r10bio *r10_bio)
1092 {
1093         struct r10conf *conf = mddev->private;
1094         struct bio *read_bio;
1095         const int op = bio_op(bio);
1096         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1097         int sectors_handled;
1098         int max_sectors;
1099         sector_t sectors;
1100         struct md_rdev *rdev;
1101         int slot;
1102
1103         /*
1104          * Register the new request and wait if the reconstruction
1105          * thread has put up a bar for new requests.
1106          * Continue immediately if no resync is active currently.
1107          */
1108         wait_barrier(conf);
1109
1110         sectors = bio_sectors(bio);
1111         while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1112             bio->bi_iter.bi_sector < conf->reshape_progress &&
1113             bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1114                 /*
1115                  * IO spans the reshape position.  Need to wait for reshape to
1116                  * pass
1117                  */
1118                 raid10_log(conf->mddev, "wait reshape");
1119                 allow_barrier(conf);
1120                 wait_event(conf->wait_barrier,
1121                            conf->reshape_progress <= bio->bi_iter.bi_sector ||
1122                            conf->reshape_progress >= bio->bi_iter.bi_sector +
1123                            sectors);
1124                 wait_barrier(conf);
1125         }
1126
1127 read_again:
1128         rdev = read_balance(conf, r10_bio, &max_sectors);
1129         if (!rdev) {
1130                 raid_end_bio_io(r10_bio);
1131                 return;
1132         }
1133         slot = r10_bio->read_slot;
1134
1135         read_bio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1136         bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1137                  max_sectors);
1138
1139         r10_bio->devs[slot].bio = read_bio;
1140         r10_bio->devs[slot].rdev = rdev;
1141
1142         read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1143                 choose_data_offset(r10_bio, rdev);
1144         read_bio->bi_bdev = rdev->bdev;
1145         read_bio->bi_end_io = raid10_end_read_request;
1146         bio_set_op_attrs(read_bio, op, do_sync);
1147         if (test_bit(FailFast, &rdev->flags) &&
1148             test_bit(R10BIO_FailFast, &r10_bio->state))
1149                 read_bio->bi_opf |= MD_FAILFAST;
1150         read_bio->bi_private = r10_bio;
1151
1152         if (mddev->gendisk)
1153                 trace_block_bio_remap(bdev_get_queue(read_bio->bi_bdev),
1154                                       read_bio, disk_devt(mddev->gendisk),
1155                                       r10_bio->sector);
1156         if (max_sectors < r10_bio->sectors) {
1157                 /*
1158                  * Could not read all from this device, so we will need another
1159                  * r10_bio.
1160                  */
1161                 sectors_handled = (r10_bio->sector + max_sectors
1162                                    - bio->bi_iter.bi_sector);
1163                 r10_bio->sectors = max_sectors;
1164                 spin_lock_irq(&conf->device_lock);
1165                 if (bio->bi_phys_segments == 0)
1166                         bio->bi_phys_segments = 2;
1167                 else
1168                         bio->bi_phys_segments++;
1169                 spin_unlock_irq(&conf->device_lock);
1170                 /*
1171                  * Cannot call generic_make_request directly as that will be
1172                  * queued in __generic_make_request and subsequent
1173                  * mempool_alloc might block waiting for it.  so hand bio over
1174                  * to raid10d.
1175                  */
1176                 reschedule_retry(r10_bio);
1177
1178                 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1179
1180                 r10_bio->master_bio = bio;
1181                 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1182                 r10_bio->state = 0;
1183                 r10_bio->mddev = mddev;
1184                 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1185                 goto read_again;
1186         } else
1187                 generic_make_request(read_bio);
1188         return;
1189 }
1190
1191 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1192                                  struct r10bio *r10_bio)
1193 {
1194         struct r10conf *conf = mddev->private;
1195         int i;
1196         const int op = bio_op(bio);
1197         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1198         const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1199         unsigned long flags;
1200         struct md_rdev *blocked_rdev;
1201         struct blk_plug_cb *cb;
1202         struct raid10_plug_cb *plug = NULL;
1203         sector_t sectors;
1204         int sectors_handled;
1205         int max_sectors;
1206
1207         md_write_start(mddev, bio);
1208
1209         /*
1210          * Register the new request and wait if the reconstruction
1211          * thread has put up a bar for new requests.
1212          * Continue immediately if no resync is active currently.
1213          */
1214         wait_barrier(conf);
1215
1216         sectors = bio_sectors(bio);
1217         while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1218             bio->bi_iter.bi_sector < conf->reshape_progress &&
1219             bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1220                 /*
1221                  * IO spans the reshape position.  Need to wait for reshape to
1222                  * pass
1223                  */
1224                 raid10_log(conf->mddev, "wait reshape");
1225                 allow_barrier(conf);
1226                 wait_event(conf->wait_barrier,
1227                            conf->reshape_progress <= bio->bi_iter.bi_sector ||
1228                            conf->reshape_progress >= bio->bi_iter.bi_sector +
1229                            sectors);
1230                 wait_barrier(conf);
1231         }
1232
1233         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1234             (mddev->reshape_backwards
1235              ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1236                 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1237              : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1238                 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1239                 /* Need to update reshape_position in metadata */
1240                 mddev->reshape_position = conf->reshape_progress;
1241                 set_mask_bits(&mddev->sb_flags, 0,
1242                               BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1243                 md_wakeup_thread(mddev->thread);
1244                 raid10_log(conf->mddev, "wait reshape metadata");
1245                 wait_event(mddev->sb_wait,
1246                            !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1247
1248                 conf->reshape_safe = mddev->reshape_position;
1249         }
1250
1251         if (conf->pending_count >= max_queued_requests) {
1252                 md_wakeup_thread(mddev->thread);
1253                 raid10_log(mddev, "wait queued");
1254                 wait_event(conf->wait_barrier,
1255                            conf->pending_count < max_queued_requests);
1256         }
1257         /* first select target devices under rcu_lock and
1258          * inc refcount on their rdev.  Record them by setting
1259          * bios[x] to bio
1260          * If there are known/acknowledged bad blocks on any device
1261          * on which we have seen a write error, we want to avoid
1262          * writing to those blocks.  This potentially requires several
1263          * writes to write around the bad blocks.  Each set of writes
1264          * gets its own r10_bio with a set of bios attached.  The number
1265          * of r10_bios is recored in bio->bi_phys_segments just as with
1266          * the read case.
1267          */
1268
1269         r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1270         raid10_find_phys(conf, r10_bio);
1271 retry_write:
1272         blocked_rdev = NULL;
1273         rcu_read_lock();
1274         max_sectors = r10_bio->sectors;
1275
1276         for (i = 0;  i < conf->copies; i++) {
1277                 int d = r10_bio->devs[i].devnum;
1278                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1279                 struct md_rdev *rrdev = rcu_dereference(
1280                         conf->mirrors[d].replacement);
1281                 if (rdev == rrdev)
1282                         rrdev = NULL;
1283                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1284                         atomic_inc(&rdev->nr_pending);
1285                         blocked_rdev = rdev;
1286                         break;
1287                 }
1288                 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1289                         atomic_inc(&rrdev->nr_pending);
1290                         blocked_rdev = rrdev;
1291                         break;
1292                 }
1293                 if (rdev && (test_bit(Faulty, &rdev->flags)))
1294                         rdev = NULL;
1295                 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1296                         rrdev = NULL;
1297
1298                 r10_bio->devs[i].bio = NULL;
1299                 r10_bio->devs[i].repl_bio = NULL;
1300
1301                 if (!rdev && !rrdev) {
1302                         set_bit(R10BIO_Degraded, &r10_bio->state);
1303                         continue;
1304                 }
1305                 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1306                         sector_t first_bad;
1307                         sector_t dev_sector = r10_bio->devs[i].addr;
1308                         int bad_sectors;
1309                         int is_bad;
1310
1311                         is_bad = is_badblock(rdev, dev_sector, max_sectors,
1312                                              &first_bad, &bad_sectors);
1313                         if (is_bad < 0) {
1314                                 /* Mustn't write here until the bad block
1315                                  * is acknowledged
1316                                  */
1317                                 atomic_inc(&rdev->nr_pending);
1318                                 set_bit(BlockedBadBlocks, &rdev->flags);
1319                                 blocked_rdev = rdev;
1320                                 break;
1321                         }
1322                         if (is_bad && first_bad <= dev_sector) {
1323                                 /* Cannot write here at all */
1324                                 bad_sectors -= (dev_sector - first_bad);
1325                                 if (bad_sectors < max_sectors)
1326                                         /* Mustn't write more than bad_sectors
1327                                          * to other devices yet
1328                                          */
1329                                         max_sectors = bad_sectors;
1330                                 /* We don't set R10BIO_Degraded as that
1331                                  * only applies if the disk is missing,
1332                                  * so it might be re-added, and we want to
1333                                  * know to recover this chunk.
1334                                  * In this case the device is here, and the
1335                                  * fact that this chunk is not in-sync is
1336                                  * recorded in the bad block log.
1337                                  */
1338                                 continue;
1339                         }
1340                         if (is_bad) {
1341                                 int good_sectors = first_bad - dev_sector;
1342                                 if (good_sectors < max_sectors)
1343                                         max_sectors = good_sectors;
1344                         }
1345                 }
1346                 if (rdev) {
1347                         r10_bio->devs[i].bio = bio;
1348                         atomic_inc(&rdev->nr_pending);
1349                 }
1350                 if (rrdev) {
1351                         r10_bio->devs[i].repl_bio = bio;
1352                         atomic_inc(&rrdev->nr_pending);
1353                 }
1354         }
1355         rcu_read_unlock();
1356
1357         if (unlikely(blocked_rdev)) {
1358                 /* Have to wait for this device to get unblocked, then retry */
1359                 int j;
1360                 int d;
1361
1362                 for (j = 0; j < i; j++) {
1363                         if (r10_bio->devs[j].bio) {
1364                                 d = r10_bio->devs[j].devnum;
1365                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1366                         }
1367                         if (r10_bio->devs[j].repl_bio) {
1368                                 struct md_rdev *rdev;
1369                                 d = r10_bio->devs[j].devnum;
1370                                 rdev = conf->mirrors[d].replacement;
1371                                 if (!rdev) {
1372                                         /* Race with remove_disk */
1373                                         smp_mb();
1374                                         rdev = conf->mirrors[d].rdev;
1375                                 }
1376                                 rdev_dec_pending(rdev, mddev);
1377                         }
1378                 }
1379                 allow_barrier(conf);
1380                 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1381                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1382                 wait_barrier(conf);
1383                 goto retry_write;
1384         }
1385
1386         if (max_sectors < r10_bio->sectors) {
1387                 /* We are splitting this into multiple parts, so
1388                  * we need to prepare for allocating another r10_bio.
1389                  */
1390                 r10_bio->sectors = max_sectors;
1391                 spin_lock_irq(&conf->device_lock);
1392                 if (bio->bi_phys_segments == 0)
1393                         bio->bi_phys_segments = 2;
1394                 else
1395                         bio->bi_phys_segments++;
1396                 spin_unlock_irq(&conf->device_lock);
1397         }
1398         sectors_handled = r10_bio->sector + max_sectors -
1399                 bio->bi_iter.bi_sector;
1400
1401         atomic_set(&r10_bio->remaining, 1);
1402         bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1403
1404         for (i = 0; i < conf->copies; i++) {
1405                 struct bio *mbio;
1406                 int d = r10_bio->devs[i].devnum;
1407                 if (r10_bio->devs[i].bio) {
1408                         struct md_rdev *rdev = conf->mirrors[d].rdev;
1409                         mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1410                         bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1411                                  max_sectors);
1412                         r10_bio->devs[i].bio = mbio;
1413
1414                         mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
1415                                            choose_data_offset(r10_bio, rdev));
1416                         mbio->bi_bdev = rdev->bdev;
1417                         mbio->bi_end_io = raid10_end_write_request;
1418                         bio_set_op_attrs(mbio, op, do_sync | do_fua);
1419                         if (test_bit(FailFast, &conf->mirrors[d].rdev->flags) &&
1420                             enough(conf, d))
1421                                 mbio->bi_opf |= MD_FAILFAST;
1422                         mbio->bi_private = r10_bio;
1423
1424                         if (conf->mddev->gendisk)
1425                                 trace_block_bio_remap(bdev_get_queue(mbio->bi_bdev),
1426                                                       mbio, disk_devt(conf->mddev->gendisk),
1427                                                       r10_bio->sector);
1428                         /* flush_pending_writes() needs access to the rdev so...*/
1429                         mbio->bi_bdev = (void*)rdev;
1430
1431                         atomic_inc(&r10_bio->remaining);
1432
1433                         cb = blk_check_plugged(raid10_unplug, mddev,
1434                                                sizeof(*plug));
1435                         if (cb)
1436                                 plug = container_of(cb, struct raid10_plug_cb,
1437                                                     cb);
1438                         else
1439                                 plug = NULL;
1440                         spin_lock_irqsave(&conf->device_lock, flags);
1441                         if (plug) {
1442                                 bio_list_add(&plug->pending, mbio);
1443                                 plug->pending_cnt++;
1444                         } else {
1445                                 bio_list_add(&conf->pending_bio_list, mbio);
1446                                 conf->pending_count++;
1447                         }
1448                         spin_unlock_irqrestore(&conf->device_lock, flags);
1449                         if (!plug)
1450                                 md_wakeup_thread(mddev->thread);
1451                 }
1452
1453                 if (r10_bio->devs[i].repl_bio) {
1454                         struct md_rdev *rdev = conf->mirrors[d].replacement;
1455                         if (rdev == NULL) {
1456                                 /* Replacement just got moved to main 'rdev' */
1457                                 smp_mb();
1458                                 rdev = conf->mirrors[d].rdev;
1459                         }
1460                         mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1461                         bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1462                                  max_sectors);
1463                         r10_bio->devs[i].repl_bio = mbio;
1464
1465                         mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr +
1466                                            choose_data_offset(r10_bio, rdev));
1467                         mbio->bi_bdev = rdev->bdev;
1468                         mbio->bi_end_io = raid10_end_write_request;
1469                         bio_set_op_attrs(mbio, op, do_sync | do_fua);
1470                         mbio->bi_private = r10_bio;
1471
1472                         if (conf->mddev->gendisk)
1473                                 trace_block_bio_remap(bdev_get_queue(mbio->bi_bdev),
1474                                                       mbio, disk_devt(conf->mddev->gendisk),
1475                                                       r10_bio->sector);
1476                         /* flush_pending_writes() needs access to the rdev so...*/
1477                         mbio->bi_bdev = (void*)rdev;
1478
1479                         atomic_inc(&r10_bio->remaining);
1480                         spin_lock_irqsave(&conf->device_lock, flags);
1481                         bio_list_add(&conf->pending_bio_list, mbio);
1482                         conf->pending_count++;
1483                         spin_unlock_irqrestore(&conf->device_lock, flags);
1484                         if (!mddev_check_plugged(mddev))
1485                                 md_wakeup_thread(mddev->thread);
1486                 }
1487         }
1488
1489         /* Don't remove the bias on 'remaining' (one_write_done) until
1490          * after checking if we need to go around again.
1491          */
1492
1493         if (sectors_handled < bio_sectors(bio)) {
1494                 one_write_done(r10_bio);
1495                 /* We need another r10_bio.  It has already been counted
1496                  * in bio->bi_phys_segments.
1497                  */
1498                 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1499
1500                 r10_bio->master_bio = bio;
1501                 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1502
1503                 r10_bio->mddev = mddev;
1504                 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1505                 r10_bio->state = 0;
1506                 goto retry_write;
1507         }
1508         one_write_done(r10_bio);
1509 }
1510
1511 static void __make_request(struct mddev *mddev, struct bio *bio)
1512 {
1513         struct r10conf *conf = mddev->private;
1514         struct r10bio *r10_bio;
1515
1516         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1517
1518         r10_bio->master_bio = bio;
1519         r10_bio->sectors = bio_sectors(bio);
1520
1521         r10_bio->mddev = mddev;
1522         r10_bio->sector = bio->bi_iter.bi_sector;
1523         r10_bio->state = 0;
1524
1525         /*
1526          * We might need to issue multiple reads to different devices if there
1527          * are bad blocks around, so we keep track of the number of reads in
1528          * bio->bi_phys_segments.  If this is 0, there is only one r10_bio and
1529          * no locking will be needed when the request completes.  If it is
1530          * non-zero, then it is the number of not-completed requests.
1531          */
1532         bio->bi_phys_segments = 0;
1533         bio_clear_flag(bio, BIO_SEG_VALID);
1534
1535         if (bio_data_dir(bio) == READ)
1536                 raid10_read_request(mddev, bio, r10_bio);
1537         else
1538                 raid10_write_request(mddev, bio, r10_bio);
1539 }
1540
1541 static void raid10_make_request(struct mddev *mddev, struct bio *bio)
1542 {
1543         struct r10conf *conf = mddev->private;
1544         sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1545         int chunk_sects = chunk_mask + 1;
1546
1547         struct bio *split;
1548
1549         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1550                 md_flush_request(mddev, bio);
1551                 return;
1552         }
1553
1554         do {
1555
1556                 /*
1557                  * If this request crosses a chunk boundary, we need to split
1558                  * it.
1559                  */
1560                 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1561                              bio_sectors(bio) > chunk_sects
1562                              && (conf->geo.near_copies < conf->geo.raid_disks
1563                                  || conf->prev.near_copies <
1564                                  conf->prev.raid_disks))) {
1565                         split = bio_split(bio, chunk_sects -
1566                                           (bio->bi_iter.bi_sector &
1567                                            (chunk_sects - 1)),
1568                                           GFP_NOIO, fs_bio_set);
1569                         bio_chain(split, bio);
1570                 } else {
1571                         split = bio;
1572                 }
1573
1574                 __make_request(mddev, split);
1575         } while (split != bio);
1576
1577         /* In case raid10d snuck in to freeze_array */
1578         wake_up(&conf->wait_barrier);
1579 }
1580
1581 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1582 {
1583         struct r10conf *conf = mddev->private;
1584         int i;
1585
1586         if (conf->geo.near_copies < conf->geo.raid_disks)
1587                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1588         if (conf->geo.near_copies > 1)
1589                 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1590         if (conf->geo.far_copies > 1) {
1591                 if (conf->geo.far_offset)
1592                         seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1593                 else
1594                         seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1595                 if (conf->geo.far_set_size != conf->geo.raid_disks)
1596                         seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1597         }
1598         seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1599                                         conf->geo.raid_disks - mddev->degraded);
1600         rcu_read_lock();
1601         for (i = 0; i < conf->geo.raid_disks; i++) {
1602                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1603                 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1604         }
1605         rcu_read_unlock();
1606         seq_printf(seq, "]");
1607 }
1608
1609 /* check if there are enough drives for
1610  * every block to appear on atleast one.
1611  * Don't consider the device numbered 'ignore'
1612  * as we might be about to remove it.
1613  */
1614 static int _enough(struct r10conf *conf, int previous, int ignore)
1615 {
1616         int first = 0;
1617         int has_enough = 0;
1618         int disks, ncopies;
1619         if (previous) {
1620                 disks = conf->prev.raid_disks;
1621                 ncopies = conf->prev.near_copies;
1622         } else {
1623                 disks = conf->geo.raid_disks;
1624                 ncopies = conf->geo.near_copies;
1625         }
1626
1627         rcu_read_lock();
1628         do {
1629                 int n = conf->copies;
1630                 int cnt = 0;
1631                 int this = first;
1632                 while (n--) {
1633                         struct md_rdev *rdev;
1634                         if (this != ignore &&
1635                             (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1636                             test_bit(In_sync, &rdev->flags))
1637                                 cnt++;
1638                         this = (this+1) % disks;
1639                 }
1640                 if (cnt == 0)
1641                         goto out;
1642                 first = (first + ncopies) % disks;
1643         } while (first != 0);
1644         has_enough = 1;
1645 out:
1646         rcu_read_unlock();
1647         return has_enough;
1648 }
1649
1650 static int enough(struct r10conf *conf, int ignore)
1651 {
1652         /* when calling 'enough', both 'prev' and 'geo' must
1653          * be stable.
1654          * This is ensured if ->reconfig_mutex or ->device_lock
1655          * is held.
1656          */
1657         return _enough(conf, 0, ignore) &&
1658                 _enough(conf, 1, ignore);
1659 }
1660
1661 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1662 {
1663         char b[BDEVNAME_SIZE];
1664         struct r10conf *conf = mddev->private;
1665         unsigned long flags;
1666
1667         /*
1668          * If it is not operational, then we have already marked it as dead
1669          * else if it is the last working disks, ignore the error, let the
1670          * next level up know.
1671          * else mark the drive as failed
1672          */
1673         spin_lock_irqsave(&conf->device_lock, flags);
1674         if (test_bit(In_sync, &rdev->flags)
1675             && !enough(conf, rdev->raid_disk)) {
1676                 /*
1677                  * Don't fail the drive, just return an IO error.
1678                  */
1679                 spin_unlock_irqrestore(&conf->device_lock, flags);
1680                 return;
1681         }
1682         if (test_and_clear_bit(In_sync, &rdev->flags))
1683                 mddev->degraded++;
1684         /*
1685          * If recovery is running, make sure it aborts.
1686          */
1687         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1688         set_bit(Blocked, &rdev->flags);
1689         set_bit(Faulty, &rdev->flags);
1690         set_mask_bits(&mddev->sb_flags, 0,
1691                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1692         spin_unlock_irqrestore(&conf->device_lock, flags);
1693         pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1694                 "md/raid10:%s: Operation continuing on %d devices.\n",
1695                 mdname(mddev), bdevname(rdev->bdev, b),
1696                 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1697 }
1698
1699 static void print_conf(struct r10conf *conf)
1700 {
1701         int i;
1702         struct md_rdev *rdev;
1703
1704         pr_debug("RAID10 conf printout:\n");
1705         if (!conf) {
1706                 pr_debug("(!conf)\n");
1707                 return;
1708         }
1709         pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1710                  conf->geo.raid_disks);
1711
1712         /* This is only called with ->reconfix_mutex held, so
1713          * rcu protection of rdev is not needed */
1714         for (i = 0; i < conf->geo.raid_disks; i++) {
1715                 char b[BDEVNAME_SIZE];
1716                 rdev = conf->mirrors[i].rdev;
1717                 if (rdev)
1718                         pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1719                                  i, !test_bit(In_sync, &rdev->flags),
1720                                  !test_bit(Faulty, &rdev->flags),
1721                                  bdevname(rdev->bdev,b));
1722         }
1723 }
1724
1725 static void close_sync(struct r10conf *conf)
1726 {
1727         wait_barrier(conf);
1728         allow_barrier(conf);
1729
1730         mempool_destroy(conf->r10buf_pool);
1731         conf->r10buf_pool = NULL;
1732 }
1733
1734 static int raid10_spare_active(struct mddev *mddev)
1735 {
1736         int i;
1737         struct r10conf *conf = mddev->private;
1738         struct raid10_info *tmp;
1739         int count = 0;
1740         unsigned long flags;
1741
1742         /*
1743          * Find all non-in_sync disks within the RAID10 configuration
1744          * and mark them in_sync
1745          */
1746         for (i = 0; i < conf->geo.raid_disks; i++) {
1747                 tmp = conf->mirrors + i;
1748                 if (tmp->replacement
1749                     && tmp->replacement->recovery_offset == MaxSector
1750                     && !test_bit(Faulty, &tmp->replacement->flags)
1751                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1752                         /* Replacement has just become active */
1753                         if (!tmp->rdev
1754                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1755                                 count++;
1756                         if (tmp->rdev) {
1757                                 /* Replaced device not technically faulty,
1758                                  * but we need to be sure it gets removed
1759                                  * and never re-added.
1760                                  */
1761                                 set_bit(Faulty, &tmp->rdev->flags);
1762                                 sysfs_notify_dirent_safe(
1763                                         tmp->rdev->sysfs_state);
1764                         }
1765                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1766                 } else if (tmp->rdev
1767                            && tmp->rdev->recovery_offset == MaxSector
1768                            && !test_bit(Faulty, &tmp->rdev->flags)
1769                            && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1770                         count++;
1771                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1772                 }
1773         }
1774         spin_lock_irqsave(&conf->device_lock, flags);
1775         mddev->degraded -= count;
1776         spin_unlock_irqrestore(&conf->device_lock, flags);
1777
1778         print_conf(conf);
1779         return count;
1780 }
1781
1782 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1783 {
1784         struct r10conf *conf = mddev->private;
1785         int err = -EEXIST;
1786         int mirror;
1787         int first = 0;
1788         int last = conf->geo.raid_disks - 1;
1789
1790         if (mddev->recovery_cp < MaxSector)
1791                 /* only hot-add to in-sync arrays, as recovery is
1792                  * very different from resync
1793                  */
1794                 return -EBUSY;
1795         if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1796                 return -EINVAL;
1797
1798         if (md_integrity_add_rdev(rdev, mddev))
1799                 return -ENXIO;
1800
1801         if (rdev->raid_disk >= 0)
1802                 first = last = rdev->raid_disk;
1803
1804         if (rdev->saved_raid_disk >= first &&
1805             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1806                 mirror = rdev->saved_raid_disk;
1807         else
1808                 mirror = first;
1809         for ( ; mirror <= last ; mirror++) {
1810                 struct raid10_info *p = &conf->mirrors[mirror];
1811                 if (p->recovery_disabled == mddev->recovery_disabled)
1812                         continue;
1813                 if (p->rdev) {
1814                         if (!test_bit(WantReplacement, &p->rdev->flags) ||
1815                             p->replacement != NULL)
1816                                 continue;
1817                         clear_bit(In_sync, &rdev->flags);
1818                         set_bit(Replacement, &rdev->flags);
1819                         rdev->raid_disk = mirror;
1820                         err = 0;
1821                         if (mddev->gendisk)
1822                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1823                                                   rdev->data_offset << 9);
1824                         conf->fullsync = 1;
1825                         rcu_assign_pointer(p->replacement, rdev);
1826                         break;
1827                 }
1828
1829                 if (mddev->gendisk)
1830                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1831                                           rdev->data_offset << 9);
1832
1833                 p->head_position = 0;
1834                 p->recovery_disabled = mddev->recovery_disabled - 1;
1835                 rdev->raid_disk = mirror;
1836                 err = 0;
1837                 if (rdev->saved_raid_disk != mirror)
1838                         conf->fullsync = 1;
1839                 rcu_assign_pointer(p->rdev, rdev);
1840                 break;
1841         }
1842         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1843                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1844
1845         print_conf(conf);
1846         return err;
1847 }
1848
1849 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1850 {
1851         struct r10conf *conf = mddev->private;
1852         int err = 0;
1853         int number = rdev->raid_disk;
1854         struct md_rdev **rdevp;
1855         struct raid10_info *p = conf->mirrors + number;
1856
1857         print_conf(conf);
1858         if (rdev == p->rdev)
1859                 rdevp = &p->rdev;
1860         else if (rdev == p->replacement)
1861                 rdevp = &p->replacement;
1862         else
1863                 return 0;
1864
1865         if (test_bit(In_sync, &rdev->flags) ||
1866             atomic_read(&rdev->nr_pending)) {
1867                 err = -EBUSY;
1868                 goto abort;
1869         }
1870         /* Only remove non-faulty devices if recovery
1871          * is not possible.
1872          */
1873         if (!test_bit(Faulty, &rdev->flags) &&
1874             mddev->recovery_disabled != p->recovery_disabled &&
1875             (!p->replacement || p->replacement == rdev) &&
1876             number < conf->geo.raid_disks &&
1877             enough(conf, -1)) {
1878                 err = -EBUSY;
1879                 goto abort;
1880         }
1881         *rdevp = NULL;
1882         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1883                 synchronize_rcu();
1884                 if (atomic_read(&rdev->nr_pending)) {
1885                         /* lost the race, try later */
1886                         err = -EBUSY;
1887                         *rdevp = rdev;
1888                         goto abort;
1889                 }
1890         }
1891         if (p->replacement) {
1892                 /* We must have just cleared 'rdev' */
1893                 p->rdev = p->replacement;
1894                 clear_bit(Replacement, &p->replacement->flags);
1895                 smp_mb(); /* Make sure other CPUs may see both as identical
1896                            * but will never see neither -- if they are careful.
1897                            */
1898                 p->replacement = NULL;
1899                 clear_bit(WantReplacement, &rdev->flags);
1900         } else
1901                 /* We might have just remove the Replacement as faulty
1902                  * Clear the flag just in case
1903                  */
1904                 clear_bit(WantReplacement, &rdev->flags);
1905
1906         err = md_integrity_register(mddev);
1907
1908 abort:
1909
1910         print_conf(conf);
1911         return err;
1912 }
1913
1914 static void end_sync_read(struct bio *bio)
1915 {
1916         struct r10bio *r10_bio = bio->bi_private;
1917         struct r10conf *conf = r10_bio->mddev->private;
1918         int d;
1919
1920         if (bio == r10_bio->master_bio) {
1921                 /* this is a reshape read */
1922                 d = r10_bio->read_slot; /* really the read dev */
1923         } else
1924                 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1925
1926         if (!bio->bi_error)
1927                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1928         else
1929                 /* The write handler will notice the lack of
1930                  * R10BIO_Uptodate and record any errors etc
1931                  */
1932                 atomic_add(r10_bio->sectors,
1933                            &conf->mirrors[d].rdev->corrected_errors);
1934
1935         /* for reconstruct, we always reschedule after a read.
1936          * for resync, only after all reads
1937          */
1938         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1939         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1940             atomic_dec_and_test(&r10_bio->remaining)) {
1941                 /* we have read all the blocks,
1942                  * do the comparison in process context in raid10d
1943                  */
1944                 reschedule_retry(r10_bio);
1945         }
1946 }
1947
1948 static void end_sync_request(struct r10bio *r10_bio)
1949 {
1950         struct mddev *mddev = r10_bio->mddev;
1951
1952         while (atomic_dec_and_test(&r10_bio->remaining)) {
1953                 if (r10_bio->master_bio == NULL) {
1954                         /* the primary of several recovery bios */
1955                         sector_t s = r10_bio->sectors;
1956                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1957                             test_bit(R10BIO_WriteError, &r10_bio->state))
1958                                 reschedule_retry(r10_bio);
1959                         else
1960                                 put_buf(r10_bio);
1961                         md_done_sync(mddev, s, 1);
1962                         break;
1963                 } else {
1964                         struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1965                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1966                             test_bit(R10BIO_WriteError, &r10_bio->state))
1967                                 reschedule_retry(r10_bio);
1968                         else
1969                                 put_buf(r10_bio);
1970                         r10_bio = r10_bio2;
1971                 }
1972         }
1973 }
1974
1975 static void end_sync_write(struct bio *bio)
1976 {
1977         struct r10bio *r10_bio = bio->bi_private;
1978         struct mddev *mddev = r10_bio->mddev;
1979         struct r10conf *conf = mddev->private;
1980         int d;
1981         sector_t first_bad;
1982         int bad_sectors;
1983         int slot;
1984         int repl;
1985         struct md_rdev *rdev = NULL;
1986
1987         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1988         if (repl)
1989                 rdev = conf->mirrors[d].replacement;
1990         else
1991                 rdev = conf->mirrors[d].rdev;
1992
1993         if (bio->bi_error) {
1994                 if (repl)
1995                         md_error(mddev, rdev);
1996                 else {
1997                         set_bit(WriteErrorSeen, &rdev->flags);
1998                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
1999                                 set_bit(MD_RECOVERY_NEEDED,
2000                                         &rdev->mddev->recovery);
2001                         set_bit(R10BIO_WriteError, &r10_bio->state);
2002                 }
2003         } else if (is_badblock(rdev,
2004                              r10_bio->devs[slot].addr,
2005                              r10_bio->sectors,
2006                              &first_bad, &bad_sectors))
2007                 set_bit(R10BIO_MadeGood, &r10_bio->state);
2008
2009         rdev_dec_pending(rdev, mddev);
2010
2011         end_sync_request(r10_bio);
2012 }
2013
2014 /*
2015  * Note: sync and recover and handled very differently for raid10
2016  * This code is for resync.
2017  * For resync, we read through virtual addresses and read all blocks.
2018  * If there is any error, we schedule a write.  The lowest numbered
2019  * drive is authoritative.
2020  * However requests come for physical address, so we need to map.
2021  * For every physical address there are raid_disks/copies virtual addresses,
2022  * which is always are least one, but is not necessarly an integer.
2023  * This means that a physical address can span multiple chunks, so we may
2024  * have to submit multiple io requests for a single sync request.
2025  */
2026 /*
2027  * We check if all blocks are in-sync and only write to blocks that
2028  * aren't in sync
2029  */
2030 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2031 {
2032         struct r10conf *conf = mddev->private;
2033         int i, first;
2034         struct bio *tbio, *fbio;
2035         int vcnt;
2036
2037         atomic_set(&r10_bio->remaining, 1);
2038
2039         /* find the first device with a block */
2040         for (i=0; i<conf->copies; i++)
2041                 if (!r10_bio->devs[i].bio->bi_error)
2042                         break;
2043
2044         if (i == conf->copies)
2045                 goto done;
2046
2047         first = i;
2048         fbio = r10_bio->devs[i].bio;
2049         fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2050         fbio->bi_iter.bi_idx = 0;
2051
2052         vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2053         /* now find blocks with errors */
2054         for (i=0 ; i < conf->copies ; i++) {
2055                 int  j, d;
2056                 struct md_rdev *rdev;
2057
2058                 tbio = r10_bio->devs[i].bio;
2059
2060                 if (tbio->bi_end_io != end_sync_read)
2061                         continue;
2062                 if (i == first)
2063                         continue;
2064                 d = r10_bio->devs[i].devnum;
2065                 rdev = conf->mirrors[d].rdev;
2066                 if (!r10_bio->devs[i].bio->bi_error) {
2067                         /* We know that the bi_io_vec layout is the same for
2068                          * both 'first' and 'i', so we just compare them.
2069                          * All vec entries are PAGE_SIZE;
2070                          */
2071                         int sectors = r10_bio->sectors;
2072                         for (j = 0; j < vcnt; j++) {
2073                                 int len = PAGE_SIZE;
2074                                 if (sectors < (len / 512))
2075                                         len = sectors * 512;
2076                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2077                                            page_address(tbio->bi_io_vec[j].bv_page),
2078                                            len))
2079                                         break;
2080                                 sectors -= len/512;
2081                         }
2082                         if (j == vcnt)
2083                                 continue;
2084                         atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2085                         if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2086                                 /* Don't fix anything. */
2087                                 continue;
2088                 } else if (test_bit(FailFast, &rdev->flags)) {
2089                         /* Just give up on this device */
2090                         md_error(rdev->mddev, rdev);
2091                         continue;
2092                 }
2093                 /* Ok, we need to write this bio, either to correct an
2094                  * inconsistency or to correct an unreadable block.
2095                  * First we need to fixup bv_offset, bv_len and
2096                  * bi_vecs, as the read request might have corrupted these
2097                  */
2098                 bio_reset(tbio);
2099
2100                 tbio->bi_vcnt = vcnt;
2101                 tbio->bi_iter.bi_size = fbio->bi_iter.bi_size;
2102                 tbio->bi_private = r10_bio;
2103                 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2104                 tbio->bi_end_io = end_sync_write;
2105                 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2106
2107                 bio_copy_data(tbio, fbio);
2108
2109                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2110                 atomic_inc(&r10_bio->remaining);
2111                 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2112
2113                 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2114                         tbio->bi_opf |= MD_FAILFAST;
2115                 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2116                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2117                 generic_make_request(tbio);
2118         }
2119
2120         /* Now write out to any replacement devices
2121          * that are active
2122          */
2123         for (i = 0; i < conf->copies; i++) {
2124                 int d;
2125
2126                 tbio = r10_bio->devs[i].repl_bio;
2127                 if (!tbio || !tbio->bi_end_io)
2128                         continue;
2129                 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2130                     && r10_bio->devs[i].bio != fbio)
2131                         bio_copy_data(tbio, fbio);
2132                 d = r10_bio->devs[i].devnum;
2133                 atomic_inc(&r10_bio->remaining);
2134                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2135                              bio_sectors(tbio));
2136                 generic_make_request(tbio);
2137         }
2138
2139 done:
2140         if (atomic_dec_and_test(&r10_bio->remaining)) {
2141                 md_done_sync(mddev, r10_bio->sectors, 1);
2142                 put_buf(r10_bio);
2143         }
2144 }
2145
2146 /*
2147  * Now for the recovery code.
2148  * Recovery happens across physical sectors.
2149  * We recover all non-is_sync drives by finding the virtual address of
2150  * each, and then choose a working drive that also has that virt address.
2151  * There is a separate r10_bio for each non-in_sync drive.
2152  * Only the first two slots are in use. The first for reading,
2153  * The second for writing.
2154  *
2155  */
2156 static void fix_recovery_read_error(struct r10bio *r10_bio)
2157 {
2158         /* We got a read error during recovery.
2159          * We repeat the read in smaller page-sized sections.
2160          * If a read succeeds, write it to the new device or record
2161          * a bad block if we cannot.
2162          * If a read fails, record a bad block on both old and
2163          * new devices.
2164          */
2165         struct mddev *mddev = r10_bio->mddev;
2166         struct r10conf *conf = mddev->private;
2167         struct bio *bio = r10_bio->devs[0].bio;
2168         sector_t sect = 0;
2169         int sectors = r10_bio->sectors;
2170         int idx = 0;
2171         int dr = r10_bio->devs[0].devnum;
2172         int dw = r10_bio->devs[1].devnum;
2173
2174         while (sectors) {
2175                 int s = sectors;
2176                 struct md_rdev *rdev;
2177                 sector_t addr;
2178                 int ok;
2179
2180                 if (s > (PAGE_SIZE>>9))
2181                         s = PAGE_SIZE >> 9;
2182
2183                 rdev = conf->mirrors[dr].rdev;
2184                 addr = r10_bio->devs[0].addr + sect,
2185                 ok = sync_page_io(rdev,
2186                                   addr,
2187                                   s << 9,
2188                                   bio->bi_io_vec[idx].bv_page,
2189                                   REQ_OP_READ, 0, false);
2190                 if (ok) {
2191                         rdev = conf->mirrors[dw].rdev;
2192                         addr = r10_bio->devs[1].addr + sect;
2193                         ok = sync_page_io(rdev,
2194                                           addr,
2195                                           s << 9,
2196                                           bio->bi_io_vec[idx].bv_page,
2197                                           REQ_OP_WRITE, 0, false);
2198                         if (!ok) {
2199                                 set_bit(WriteErrorSeen, &rdev->flags);
2200                                 if (!test_and_set_bit(WantReplacement,
2201                                                       &rdev->flags))
2202                                         set_bit(MD_RECOVERY_NEEDED,
2203                                                 &rdev->mddev->recovery);
2204                         }
2205                 }
2206                 if (!ok) {
2207                         /* We don't worry if we cannot set a bad block -
2208                          * it really is bad so there is no loss in not
2209                          * recording it yet
2210                          */
2211                         rdev_set_badblocks(rdev, addr, s, 0);
2212
2213                         if (rdev != conf->mirrors[dw].rdev) {
2214                                 /* need bad block on destination too */
2215                                 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2216                                 addr = r10_bio->devs[1].addr + sect;
2217                                 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2218                                 if (!ok) {
2219                                         /* just abort the recovery */
2220                                         pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2221                                                   mdname(mddev));
2222
2223                                         conf->mirrors[dw].recovery_disabled
2224                                                 = mddev->recovery_disabled;
2225                                         set_bit(MD_RECOVERY_INTR,
2226                                                 &mddev->recovery);
2227                                         break;
2228                                 }
2229                         }
2230                 }
2231
2232                 sectors -= s;
2233                 sect += s;
2234                 idx++;
2235         }
2236 }
2237
2238 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2239 {
2240         struct r10conf *conf = mddev->private;
2241         int d;
2242         struct bio *wbio, *wbio2;
2243
2244         if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2245                 fix_recovery_read_error(r10_bio);
2246                 end_sync_request(r10_bio);
2247                 return;
2248         }
2249
2250         /*
2251          * share the pages with the first bio
2252          * and submit the write request
2253          */
2254         d = r10_bio->devs[1].devnum;
2255         wbio = r10_bio->devs[1].bio;
2256         wbio2 = r10_bio->devs[1].repl_bio;
2257         /* Need to test wbio2->bi_end_io before we call
2258          * generic_make_request as if the former is NULL,
2259          * the latter is free to free wbio2.
2260          */
2261         if (wbio2 && !wbio2->bi_end_io)
2262                 wbio2 = NULL;
2263         if (wbio->bi_end_io) {
2264                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2265                 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2266                 generic_make_request(wbio);
2267         }
2268         if (wbio2) {
2269                 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2270                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2271                              bio_sectors(wbio2));
2272                 generic_make_request(wbio2);
2273         }
2274 }
2275
2276 /*
2277  * Used by fix_read_error() to decay the per rdev read_errors.
2278  * We halve the read error count for every hour that has elapsed
2279  * since the last recorded read error.
2280  *
2281  */
2282 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2283 {
2284         long cur_time_mon;
2285         unsigned long hours_since_last;
2286         unsigned int read_errors = atomic_read(&rdev->read_errors);
2287
2288         cur_time_mon = ktime_get_seconds();
2289
2290         if (rdev->last_read_error == 0) {
2291                 /* first time we've seen a read error */
2292                 rdev->last_read_error = cur_time_mon;
2293                 return;
2294         }
2295
2296         hours_since_last = (long)(cur_time_mon -
2297                             rdev->last_read_error) / 3600;
2298
2299         rdev->last_read_error = cur_time_mon;
2300
2301         /*
2302          * if hours_since_last is > the number of bits in read_errors
2303          * just set read errors to 0. We do this to avoid
2304          * overflowing the shift of read_errors by hours_since_last.
2305          */
2306         if (hours_since_last >= 8 * sizeof(read_errors))
2307                 atomic_set(&rdev->read_errors, 0);
2308         else
2309                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2310 }
2311
2312 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2313                             int sectors, struct page *page, int rw)
2314 {
2315         sector_t first_bad;
2316         int bad_sectors;
2317
2318         if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2319             && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2320                 return -1;
2321         if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2322                 /* success */
2323                 return 1;
2324         if (rw == WRITE) {
2325                 set_bit(WriteErrorSeen, &rdev->flags);
2326                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2327                         set_bit(MD_RECOVERY_NEEDED,
2328                                 &rdev->mddev->recovery);
2329         }
2330         /* need to record an error - either for the block or the device */
2331         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2332                 md_error(rdev->mddev, rdev);
2333         return 0;
2334 }
2335
2336 /*
2337  * This is a kernel thread which:
2338  *
2339  *      1.      Retries failed read operations on working mirrors.
2340  *      2.      Updates the raid superblock when problems encounter.
2341  *      3.      Performs writes following reads for array synchronising.
2342  */
2343
2344 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2345 {
2346         int sect = 0; /* Offset from r10_bio->sector */
2347         int sectors = r10_bio->sectors;
2348         struct md_rdev*rdev;
2349         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2350         int d = r10_bio->devs[r10_bio->read_slot].devnum;
2351
2352         /* still own a reference to this rdev, so it cannot
2353          * have been cleared recently.
2354          */
2355         rdev = conf->mirrors[d].rdev;
2356
2357         if (test_bit(Faulty, &rdev->flags))
2358                 /* drive has already been failed, just ignore any
2359                    more fix_read_error() attempts */
2360                 return;
2361
2362         check_decay_read_errors(mddev, rdev);
2363         atomic_inc(&rdev->read_errors);
2364         if (atomic_read(&rdev->read_errors) > max_read_errors) {
2365                 char b[BDEVNAME_SIZE];
2366                 bdevname(rdev->bdev, b);
2367
2368                 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2369                           mdname(mddev), b,
2370                           atomic_read(&rdev->read_errors), max_read_errors);
2371                 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2372                           mdname(mddev), b);
2373                 md_error(mddev, rdev);
2374                 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2375                 return;
2376         }
2377
2378         while(sectors) {
2379                 int s = sectors;
2380                 int sl = r10_bio->read_slot;
2381                 int success = 0;
2382                 int start;
2383
2384                 if (s > (PAGE_SIZE>>9))
2385                         s = PAGE_SIZE >> 9;
2386
2387                 rcu_read_lock();
2388                 do {
2389                         sector_t first_bad;
2390                         int bad_sectors;
2391
2392                         d = r10_bio->devs[sl].devnum;
2393                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2394                         if (rdev &&
2395                             test_bit(In_sync, &rdev->flags) &&
2396                             !test_bit(Faulty, &rdev->flags) &&
2397                             is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2398                                         &first_bad, &bad_sectors) == 0) {
2399                                 atomic_inc(&rdev->nr_pending);
2400                                 rcu_read_unlock();
2401                                 success = sync_page_io(rdev,
2402                                                        r10_bio->devs[sl].addr +
2403                                                        sect,
2404                                                        s<<9,
2405                                                        conf->tmppage,
2406                                                        REQ_OP_READ, 0, false);
2407                                 rdev_dec_pending(rdev, mddev);
2408                                 rcu_read_lock();
2409                                 if (success)
2410                                         break;
2411                         }
2412                         sl++;
2413                         if (sl == conf->copies)
2414                                 sl = 0;
2415                 } while (!success && sl != r10_bio->read_slot);
2416                 rcu_read_unlock();
2417
2418                 if (!success) {
2419                         /* Cannot read from anywhere, just mark the block
2420                          * as bad on the first device to discourage future
2421                          * reads.
2422                          */
2423                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2424                         rdev = conf->mirrors[dn].rdev;
2425
2426                         if (!rdev_set_badblocks(
2427                                     rdev,
2428                                     r10_bio->devs[r10_bio->read_slot].addr
2429                                     + sect,
2430                                     s, 0)) {
2431                                 md_error(mddev, rdev);
2432                                 r10_bio->devs[r10_bio->read_slot].bio
2433                                         = IO_BLOCKED;
2434                         }
2435                         break;
2436                 }
2437
2438                 start = sl;
2439                 /* write it back and re-read */
2440                 rcu_read_lock();
2441                 while (sl != r10_bio->read_slot) {
2442                         char b[BDEVNAME_SIZE];
2443
2444                         if (sl==0)
2445                                 sl = conf->copies;
2446                         sl--;
2447                         d = r10_bio->devs[sl].devnum;
2448                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2449                         if (!rdev ||
2450                             test_bit(Faulty, &rdev->flags) ||
2451                             !test_bit(In_sync, &rdev->flags))
2452                                 continue;
2453
2454                         atomic_inc(&rdev->nr_pending);
2455                         rcu_read_unlock();
2456                         if (r10_sync_page_io(rdev,
2457                                              r10_bio->devs[sl].addr +
2458                                              sect,
2459                                              s, conf->tmppage, WRITE)
2460                             == 0) {
2461                                 /* Well, this device is dead */
2462                                 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2463                                           mdname(mddev), s,
2464                                           (unsigned long long)(
2465                                                   sect +
2466                                                   choose_data_offset(r10_bio,
2467                                                                      rdev)),
2468                                           bdevname(rdev->bdev, b));
2469                                 pr_notice("md/raid10:%s: %s: failing drive\n",
2470                                           mdname(mddev),
2471                                           bdevname(rdev->bdev, b));
2472                         }
2473                         rdev_dec_pending(rdev, mddev);
2474                         rcu_read_lock();
2475                 }
2476                 sl = start;
2477                 while (sl != r10_bio->read_slot) {
2478                         char b[BDEVNAME_SIZE];
2479
2480                         if (sl==0)
2481                                 sl = conf->copies;
2482                         sl--;
2483                         d = r10_bio->devs[sl].devnum;
2484                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2485                         if (!rdev ||
2486                             test_bit(Faulty, &rdev->flags) ||
2487                             !test_bit(In_sync, &rdev->flags))
2488                                 continue;
2489
2490                         atomic_inc(&rdev->nr_pending);
2491                         rcu_read_unlock();
2492                         switch (r10_sync_page_io(rdev,
2493                                              r10_bio->devs[sl].addr +
2494                                              sect,
2495                                              s, conf->tmppage,
2496                                                  READ)) {
2497                         case 0:
2498                                 /* Well, this device is dead */
2499                                 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2500                                        mdname(mddev), s,
2501                                        (unsigned long long)(
2502                                                sect +
2503                                                choose_data_offset(r10_bio, rdev)),
2504                                        bdevname(rdev->bdev, b));
2505                                 pr_notice("md/raid10:%s: %s: failing drive\n",
2506                                        mdname(mddev),
2507                                        bdevname(rdev->bdev, b));
2508                                 break;
2509                         case 1:
2510                                 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2511                                        mdname(mddev), s,
2512                                        (unsigned long long)(
2513                                                sect +
2514                                                choose_data_offset(r10_bio, rdev)),
2515                                        bdevname(rdev->bdev, b));
2516                                 atomic_add(s, &rdev->corrected_errors);
2517                         }
2518
2519                         rdev_dec_pending(rdev, mddev);
2520                         rcu_read_lock();
2521                 }
2522                 rcu_read_unlock();
2523
2524                 sectors -= s;
2525                 sect += s;
2526         }
2527 }
2528
2529 static int narrow_write_error(struct r10bio *r10_bio, int i)
2530 {
2531         struct bio *bio = r10_bio->master_bio;
2532         struct mddev *mddev = r10_bio->mddev;
2533         struct r10conf *conf = mddev->private;
2534         struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2535         /* bio has the data to be written to slot 'i' where
2536          * we just recently had a write error.
2537          * We repeatedly clone the bio and trim down to one block,
2538          * then try the write.  Where the write fails we record
2539          * a bad block.
2540          * It is conceivable that the bio doesn't exactly align with
2541          * blocks.  We must handle this.
2542          *
2543          * We currently own a reference to the rdev.
2544          */
2545
2546         int block_sectors;
2547         sector_t sector;
2548         int sectors;
2549         int sect_to_write = r10_bio->sectors;
2550         int ok = 1;
2551
2552         if (rdev->badblocks.shift < 0)
2553                 return 0;
2554
2555         block_sectors = roundup(1 << rdev->badblocks.shift,
2556                                 bdev_logical_block_size(rdev->bdev) >> 9);
2557         sector = r10_bio->sector;
2558         sectors = ((r10_bio->sector + block_sectors)
2559                    & ~(sector_t)(block_sectors - 1))
2560                 - sector;
2561
2562         while (sect_to_write) {
2563                 struct bio *wbio;
2564                 sector_t wsector;
2565                 if (sectors > sect_to_write)
2566                         sectors = sect_to_write;
2567                 /* Write at 'sector' for 'sectors' */
2568                 wbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
2569                 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2570                 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2571                 wbio->bi_iter.bi_sector = wsector +
2572                                    choose_data_offset(r10_bio, rdev);
2573                 wbio->bi_bdev = rdev->bdev;
2574                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2575
2576                 if (submit_bio_wait(wbio) < 0)
2577                         /* Failure! */
2578                         ok = rdev_set_badblocks(rdev, wsector,
2579                                                 sectors, 0)
2580                                 && ok;
2581
2582                 bio_put(wbio);
2583                 sect_to_write -= sectors;
2584                 sector += sectors;
2585                 sectors = block_sectors;
2586         }
2587         return ok;
2588 }
2589
2590 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2591 {
2592         int slot = r10_bio->read_slot;
2593         struct bio *bio;
2594         struct r10conf *conf = mddev->private;
2595         struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2596         char b[BDEVNAME_SIZE];
2597         unsigned long do_sync;
2598         int max_sectors;
2599         dev_t bio_dev;
2600         sector_t bio_last_sector;
2601
2602         /* we got a read error. Maybe the drive is bad.  Maybe just
2603          * the block and we can fix it.
2604          * We freeze all other IO, and try reading the block from
2605          * other devices.  When we find one, we re-write
2606          * and check it that fixes the read error.
2607          * This is all done synchronously while the array is
2608          * frozen.
2609          */
2610         bio = r10_bio->devs[slot].bio;
2611         bdevname(bio->bi_bdev, b);
2612         bio_dev = bio->bi_bdev->bd_dev;
2613         bio_last_sector = r10_bio->devs[slot].addr + rdev->data_offset + r10_bio->sectors;
2614         bio_put(bio);
2615         r10_bio->devs[slot].bio = NULL;
2616
2617         if (mddev->ro)
2618                 r10_bio->devs[slot].bio = IO_BLOCKED;
2619         else if (!test_bit(FailFast, &rdev->flags)) {
2620                 freeze_array(conf, 1);
2621                 fix_read_error(conf, mddev, r10_bio);
2622                 unfreeze_array(conf);
2623         } else
2624                 md_error(mddev, rdev);
2625
2626         rdev_dec_pending(rdev, mddev);
2627
2628 read_more:
2629         rdev = read_balance(conf, r10_bio, &max_sectors);
2630         if (rdev == NULL) {
2631                 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
2632                                     mdname(mddev), b,
2633                                     (unsigned long long)r10_bio->sector);
2634                 raid_end_bio_io(r10_bio);
2635                 return;
2636         }
2637
2638         do_sync = (r10_bio->master_bio->bi_opf & REQ_SYNC);
2639         slot = r10_bio->read_slot;
2640         pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
2641                            mdname(mddev),
2642                            bdevname(rdev->bdev, b),
2643                            (unsigned long long)r10_bio->sector);
2644         bio = bio_clone_fast(r10_bio->master_bio, GFP_NOIO, mddev->bio_set);
2645         bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2646         r10_bio->devs[slot].bio = bio;
2647         r10_bio->devs[slot].rdev = rdev;
2648         bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2649                 + choose_data_offset(r10_bio, rdev);
2650         bio->bi_bdev = rdev->bdev;
2651         bio_set_op_attrs(bio, REQ_OP_READ, do_sync);
2652         if (test_bit(FailFast, &rdev->flags) &&
2653             test_bit(R10BIO_FailFast, &r10_bio->state))
2654                 bio->bi_opf |= MD_FAILFAST;
2655         bio->bi_private = r10_bio;
2656         bio->bi_end_io = raid10_end_read_request;
2657         trace_block_bio_remap(bdev_get_queue(bio->bi_bdev),
2658                               bio, bio_dev,
2659                               bio_last_sector - r10_bio->sectors);
2660
2661         if (max_sectors < r10_bio->sectors) {
2662                 /* Drat - have to split this up more */
2663                 struct bio *mbio = r10_bio->master_bio;
2664                 int sectors_handled =
2665                         r10_bio->sector + max_sectors
2666                         - mbio->bi_iter.bi_sector;
2667                 r10_bio->sectors = max_sectors;
2668                 spin_lock_irq(&conf->device_lock);
2669                 if (mbio->bi_phys_segments == 0)
2670                         mbio->bi_phys_segments = 2;
2671                 else
2672                         mbio->bi_phys_segments++;
2673                 spin_unlock_irq(&conf->device_lock);
2674                 generic_make_request(bio);
2675
2676                 r10_bio = mempool_alloc(conf->r10bio_pool,
2677                                         GFP_NOIO);
2678                 r10_bio->master_bio = mbio;
2679                 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2680                 r10_bio->state = 0;
2681                 set_bit(R10BIO_ReadError,
2682                         &r10_bio->state);
2683                 r10_bio->mddev = mddev;
2684                 r10_bio->sector = mbio->bi_iter.bi_sector
2685                         + sectors_handled;
2686
2687                 goto read_more;
2688         } else
2689                 generic_make_request(bio);
2690 }
2691
2692 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2693 {
2694         /* Some sort of write request has finished and it
2695          * succeeded in writing where we thought there was a
2696          * bad block.  So forget the bad block.
2697          * Or possibly if failed and we need to record
2698          * a bad block.
2699          */
2700         int m;
2701         struct md_rdev *rdev;
2702
2703         if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2704             test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2705                 for (m = 0; m < conf->copies; m++) {
2706                         int dev = r10_bio->devs[m].devnum;
2707                         rdev = conf->mirrors[dev].rdev;
2708                         if (r10_bio->devs[m].bio == NULL)
2709                                 continue;
2710                         if (!r10_bio->devs[m].bio->bi_error) {
2711                                 rdev_clear_badblocks(
2712                                         rdev,
2713                                         r10_bio->devs[m].addr,
2714                                         r10_bio->sectors, 0);
2715                         } else {
2716                                 if (!rdev_set_badblocks(
2717                                             rdev,
2718                                             r10_bio->devs[m].addr,
2719                                             r10_bio->sectors, 0))
2720                                         md_error(conf->mddev, rdev);
2721                         }
2722                         rdev = conf->mirrors[dev].replacement;
2723                         if (r10_bio->devs[m].repl_bio == NULL)
2724                                 continue;
2725
2726                         if (!r10_bio->devs[m].repl_bio->bi_error) {
2727                                 rdev_clear_badblocks(
2728                                         rdev,
2729                                         r10_bio->devs[m].addr,
2730                                         r10_bio->sectors, 0);
2731                         } else {
2732                                 if (!rdev_set_badblocks(
2733                                             rdev,
2734                                             r10_bio->devs[m].addr,
2735                                             r10_bio->sectors, 0))
2736                                         md_error(conf->mddev, rdev);
2737                         }
2738                 }
2739                 put_buf(r10_bio);
2740         } else {
2741                 bool fail = false;
2742                 for (m = 0; m < conf->copies; m++) {
2743                         int dev = r10_bio->devs[m].devnum;
2744                         struct bio *bio = r10_bio->devs[m].bio;
2745                         rdev = conf->mirrors[dev].rdev;
2746                         if (bio == IO_MADE_GOOD) {
2747                                 rdev_clear_badblocks(
2748                                         rdev,
2749                                         r10_bio->devs[m].addr,
2750                                         r10_bio->sectors, 0);
2751                                 rdev_dec_pending(rdev, conf->mddev);
2752                         } else if (bio != NULL && bio->bi_error) {
2753                                 fail = true;
2754                                 if (!narrow_write_error(r10_bio, m)) {
2755                                         md_error(conf->mddev, rdev);
2756                                         set_bit(R10BIO_Degraded,
2757                                                 &r10_bio->state);
2758                                 }
2759                                 rdev_dec_pending(rdev, conf->mddev);
2760                         }
2761                         bio = r10_bio->devs[m].repl_bio;
2762                         rdev = conf->mirrors[dev].replacement;
2763                         if (rdev && bio == IO_MADE_GOOD) {
2764                                 rdev_clear_badblocks(
2765                                         rdev,
2766                                         r10_bio->devs[m].addr,
2767                                         r10_bio->sectors, 0);
2768                                 rdev_dec_pending(rdev, conf->mddev);
2769                         }
2770                 }
2771                 if (fail) {
2772                         spin_lock_irq(&conf->device_lock);
2773                         list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2774                         conf->nr_queued++;
2775                         spin_unlock_irq(&conf->device_lock);
2776                         md_wakeup_thread(conf->mddev->thread);
2777                 } else {
2778                         if (test_bit(R10BIO_WriteError,
2779                                      &r10_bio->state))
2780                                 close_write(r10_bio);
2781                         raid_end_bio_io(r10_bio);
2782                 }
2783         }
2784 }
2785
2786 static void raid10d(struct md_thread *thread)
2787 {
2788         struct mddev *mddev = thread->mddev;
2789         struct r10bio *r10_bio;
2790         unsigned long flags;
2791         struct r10conf *conf = mddev->private;
2792         struct list_head *head = &conf->retry_list;
2793         struct blk_plug plug;
2794
2795         md_check_recovery(mddev);
2796
2797         if (!list_empty_careful(&conf->bio_end_io_list) &&
2798             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2799                 LIST_HEAD(tmp);
2800                 spin_lock_irqsave(&conf->device_lock, flags);
2801                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2802                         while (!list_empty(&conf->bio_end_io_list)) {
2803                                 list_move(conf->bio_end_io_list.prev, &tmp);
2804                                 conf->nr_queued--;
2805                         }
2806                 }
2807                 spin_unlock_irqrestore(&conf->device_lock, flags);
2808                 while (!list_empty(&tmp)) {
2809                         r10_bio = list_first_entry(&tmp, struct r10bio,
2810                                                    retry_list);
2811                         list_del(&r10_bio->retry_list);
2812                         if (mddev->degraded)
2813                                 set_bit(R10BIO_Degraded, &r10_bio->state);
2814
2815                         if (test_bit(R10BIO_WriteError,
2816                                      &r10_bio->state))
2817                                 close_write(r10_bio);
2818                         raid_end_bio_io(r10_bio);
2819                 }
2820         }
2821
2822         blk_start_plug(&plug);
2823         for (;;) {
2824
2825                 flush_pending_writes(conf);
2826
2827                 spin_lock_irqsave(&conf->device_lock, flags);
2828                 if (list_empty(head)) {
2829                         spin_unlock_irqrestore(&conf->device_lock, flags);
2830                         break;
2831                 }
2832                 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2833                 list_del(head->prev);
2834                 conf->nr_queued--;
2835                 spin_unlock_irqrestore(&conf->device_lock, flags);
2836
2837                 mddev = r10_bio->mddev;
2838                 conf = mddev->private;
2839                 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2840                     test_bit(R10BIO_WriteError, &r10_bio->state))
2841                         handle_write_completed(conf, r10_bio);
2842                 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2843                         reshape_request_write(mddev, r10_bio);
2844                 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2845                         sync_request_write(mddev, r10_bio);
2846                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2847                         recovery_request_write(mddev, r10_bio);
2848                 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2849                         handle_read_error(mddev, r10_bio);
2850                 else {
2851                         /* just a partial read to be scheduled from a
2852                          * separate context
2853                          */
2854                         int slot = r10_bio->read_slot;
2855                         generic_make_request(r10_bio->devs[slot].bio);
2856                 }
2857
2858                 cond_resched();
2859                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2860                         md_check_recovery(mddev);
2861         }
2862         blk_finish_plug(&plug);
2863 }
2864
2865 static int init_resync(struct r10conf *conf)
2866 {
2867         int buffs;
2868         int i;
2869
2870         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2871         BUG_ON(conf->r10buf_pool);
2872         conf->have_replacement = 0;
2873         for (i = 0; i < conf->geo.raid_disks; i++)
2874                 if (conf->mirrors[i].replacement)
2875                         conf->have_replacement = 1;
2876         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2877         if (!conf->r10buf_pool)
2878                 return -ENOMEM;
2879         conf->next_resync = 0;
2880         return 0;
2881 }
2882
2883 /*
2884  * perform a "sync" on one "block"
2885  *
2886  * We need to make sure that no normal I/O request - particularly write
2887  * requests - conflict with active sync requests.
2888  *
2889  * This is achieved by tracking pending requests and a 'barrier' concept
2890  * that can be installed to exclude normal IO requests.
2891  *
2892  * Resync and recovery are handled very differently.
2893  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2894  *
2895  * For resync, we iterate over virtual addresses, read all copies,
2896  * and update if there are differences.  If only one copy is live,
2897  * skip it.
2898  * For recovery, we iterate over physical addresses, read a good
2899  * value for each non-in_sync drive, and over-write.
2900  *
2901  * So, for recovery we may have several outstanding complex requests for a
2902  * given address, one for each out-of-sync device.  We model this by allocating
2903  * a number of r10_bio structures, one for each out-of-sync device.
2904  * As we setup these structures, we collect all bio's together into a list
2905  * which we then process collectively to add pages, and then process again
2906  * to pass to generic_make_request.
2907  *
2908  * The r10_bio structures are linked using a borrowed master_bio pointer.
2909  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2910  * has its remaining count decremented to 0, the whole complex operation
2911  * is complete.
2912  *
2913  */
2914
2915 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2916                              int *skipped)
2917 {
2918         struct r10conf *conf = mddev->private;
2919         struct r10bio *r10_bio;
2920         struct bio *biolist = NULL, *bio;
2921         sector_t max_sector, nr_sectors;
2922         int i;
2923         int max_sync;
2924         sector_t sync_blocks;
2925         sector_t sectors_skipped = 0;
2926         int chunks_skipped = 0;
2927         sector_t chunk_mask = conf->geo.chunk_mask;
2928
2929         if (!conf->r10buf_pool)
2930                 if (init_resync(conf))
2931                         return 0;
2932
2933         /*
2934          * Allow skipping a full rebuild for incremental assembly
2935          * of a clean array, like RAID1 does.
2936          */
2937         if (mddev->bitmap == NULL &&
2938             mddev->recovery_cp == MaxSector &&
2939             mddev->reshape_position == MaxSector &&
2940             !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2941             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2942             !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2943             conf->fullsync == 0) {
2944                 *skipped = 1;
2945                 return mddev->dev_sectors - sector_nr;
2946         }
2947
2948  skipped:
2949         max_sector = mddev->dev_sectors;
2950         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2951             test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2952                 max_sector = mddev->resync_max_sectors;
2953         if (sector_nr >= max_sector) {
2954                 /* If we aborted, we need to abort the
2955                  * sync on the 'current' bitmap chucks (there can
2956                  * be several when recovering multiple devices).
2957                  * as we may have started syncing it but not finished.
2958                  * We can find the current address in
2959                  * mddev->curr_resync, but for recovery,
2960                  * we need to convert that to several
2961                  * virtual addresses.
2962                  */
2963                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2964                         end_reshape(conf);
2965                         close_sync(conf);
2966                         return 0;
2967                 }
2968
2969                 if (mddev->curr_resync < max_sector) { /* aborted */
2970                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2971                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2972                                                 &sync_blocks, 1);
2973                         else for (i = 0; i < conf->geo.raid_disks; i++) {
2974                                 sector_t sect =
2975                                         raid10_find_virt(conf, mddev->curr_resync, i);
2976                                 bitmap_end_sync(mddev->bitmap, sect,
2977                                                 &sync_blocks, 1);
2978                         }
2979                 } else {
2980                         /* completed sync */
2981                         if ((!mddev->bitmap || conf->fullsync)
2982                             && conf->have_replacement
2983                             && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2984                                 /* Completed a full sync so the replacements
2985                                  * are now fully recovered.
2986                                  */
2987                                 rcu_read_lock();
2988                                 for (i = 0; i < conf->geo.raid_disks; i++) {
2989                                         struct md_rdev *rdev =
2990                                                 rcu_dereference(conf->mirrors[i].replacement);
2991                                         if (rdev)
2992                                                 rdev->recovery_offset = MaxSector;
2993                                 }
2994                                 rcu_read_unlock();
2995                         }
2996                         conf->fullsync = 0;
2997                 }
2998                 bitmap_close_sync(mddev->bitmap);
2999                 close_sync(conf);
3000                 *skipped = 1;
3001                 return sectors_skipped;
3002         }
3003
3004         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3005                 return reshape_request(mddev, sector_nr, skipped);
3006
3007         if (chunks_skipped >= conf->geo.raid_disks) {
3008                 /* if there has been nothing to do on any drive,
3009                  * then there is nothing to do at all..
3010                  */
3011                 *skipped = 1;
3012                 return (max_sector - sector_nr) + sectors_skipped;
3013         }
3014
3015         if (max_sector > mddev->resync_max)
3016                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3017
3018         /* make sure whole request will fit in a chunk - if chunks
3019          * are meaningful
3020          */
3021         if (conf->geo.near_copies < conf->geo.raid_disks &&
3022             max_sector > (sector_nr | chunk_mask))
3023                 max_sector = (sector_nr | chunk_mask) + 1;
3024
3025         /*
3026          * If there is non-resync activity waiting for a turn, then let it
3027          * though before starting on this new sync request.
3028          */
3029         if (conf->nr_waiting)
3030                 schedule_timeout_uninterruptible(1);
3031
3032         /* Again, very different code for resync and recovery.
3033          * Both must result in an r10bio with a list of bios that
3034          * have bi_end_io, bi_sector, bi_bdev set,
3035          * and bi_private set to the r10bio.
3036          * For recovery, we may actually create several r10bios
3037          * with 2 bios in each, that correspond to the bios in the main one.
3038          * In this case, the subordinate r10bios link back through a
3039          * borrowed master_bio pointer, and the counter in the master
3040          * includes a ref from each subordinate.
3041          */
3042         /* First, we decide what to do and set ->bi_end_io
3043          * To end_sync_read if we want to read, and
3044          * end_sync_write if we will want to write.
3045          */
3046
3047         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3048         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3049                 /* recovery... the complicated one */
3050                 int j;
3051                 r10_bio = NULL;
3052
3053                 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3054                         int still_degraded;
3055                         struct r10bio *rb2;
3056                         sector_t sect;
3057                         int must_sync;
3058                         int any_working;
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                             (mreplace == NULL ||
3070                              test_bit(Faulty, &mreplace->flags))) {
3071                                 rcu_read_unlock();
3072                                 continue;
3073                         }
3074
3075                         still_degraded = 0;
3076                         /* want to reconstruct this device */
3077                         rb2 = r10_bio;
3078                         sect = raid10_find_virt(conf, sector_nr, i);
3079                         if (sect >= mddev->resync_max_sectors) {
3080                                 /* last stripe is not complete - don't
3081                                  * try to recover this sector.
3082                                  */
3083                                 rcu_read_unlock();
3084                                 continue;
3085                         }
3086                         if (mreplace && test_bit(Faulty, &mreplace->flags))
3087                                 mreplace = NULL;
3088                         /* Unless we are doing a full sync, or a replacement
3089                          * we only need to recover the block if it is set in
3090                          * the bitmap
3091                          */
3092                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
3093                                                       &sync_blocks, 1);
3094                         if (sync_blocks < max_sync)
3095                                 max_sync = sync_blocks;
3096                         if (!must_sync &&
3097                             mreplace == NULL &&
3098                             !conf->fullsync) {
3099                                 /* yep, skip the sync_blocks here, but don't assume
3100                                  * that there will never be anything to do here
3101                                  */
3102                                 chunks_skipped = -1;
3103                                 rcu_read_unlock();
3104                                 continue;
3105                         }
3106                         atomic_inc(&mrdev->nr_pending);
3107                         if (mreplace)
3108                                 atomic_inc(&mreplace->nr_pending);
3109                         rcu_read_unlock();
3110
3111                         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3112                         r10_bio->state = 0;
3113                         raise_barrier(conf, rb2 != NULL);
3114                         atomic_set(&r10_bio->remaining, 0);
3115
3116                         r10_bio->master_bio = (struct bio*)rb2;
3117                         if (rb2)
3118                                 atomic_inc(&rb2->remaining);
3119                         r10_bio->mddev = mddev;
3120                         set_bit(R10BIO_IsRecover, &r10_bio->state);
3121                         r10_bio->sector = sect;
3122
3123                         raid10_find_phys(conf, r10_bio);
3124
3125                         /* Need to check if the array will still be
3126                          * degraded
3127                          */
3128                         rcu_read_lock();
3129                         for (j = 0; j < conf->geo.raid_disks; j++) {
3130                                 struct md_rdev *rdev = rcu_dereference(
3131                                         conf->mirrors[j].rdev);
3132                                 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3133                                         still_degraded = 1;
3134                                         break;
3135                                 }
3136                         }
3137
3138                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
3139                                                       &sync_blocks, still_degraded);
3140
3141                         any_working = 0;
3142                         for (j=0; j<conf->copies;j++) {
3143                                 int k;
3144                                 int d = r10_bio->devs[j].devnum;
3145                                 sector_t from_addr, to_addr;
3146                                 struct md_rdev *rdev =
3147                                         rcu_dereference(conf->mirrors[d].rdev);
3148                                 sector_t sector, first_bad;
3149                                 int bad_sectors;
3150                                 if (!rdev ||
3151                                     !test_bit(In_sync, &rdev->flags))
3152                                         continue;
3153                                 /* This is where we read from */
3154                                 any_working = 1;
3155                                 sector = r10_bio->devs[j].addr;
3156
3157                                 if (is_badblock(rdev, sector, max_sync,
3158                                                 &first_bad, &bad_sectors)) {
3159                                         if (first_bad > sector)
3160                                                 max_sync = first_bad - sector;
3161                                         else {
3162                                                 bad_sectors -= (sector
3163                                                                 - first_bad);
3164                                                 if (max_sync > bad_sectors)
3165                                                         max_sync = bad_sectors;
3166                                                 continue;
3167                                         }
3168                                 }
3169                                 bio = r10_bio->devs[0].bio;
3170                                 bio_reset(bio);
3171                                 bio->bi_next = biolist;
3172                                 biolist = bio;
3173                                 bio->bi_private = r10_bio;
3174                                 bio->bi_end_io = end_sync_read;
3175                                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3176                                 if (test_bit(FailFast, &rdev->flags))
3177                                         bio->bi_opf |= MD_FAILFAST;
3178                                 from_addr = r10_bio->devs[j].addr;
3179                                 bio->bi_iter.bi_sector = from_addr +
3180                                         rdev->data_offset;
3181                                 bio->bi_bdev = rdev->bdev;
3182                                 atomic_inc(&rdev->nr_pending);
3183                                 /* and we write to 'i' (if not in_sync) */
3184
3185                                 for (k=0; k<conf->copies; k++)
3186                                         if (r10_bio->devs[k].devnum == i)
3187                                                 break;
3188                                 BUG_ON(k == conf->copies);
3189                                 to_addr = r10_bio->devs[k].addr;
3190                                 r10_bio->devs[0].devnum = d;
3191                                 r10_bio->devs[0].addr = from_addr;
3192                                 r10_bio->devs[1].devnum = i;
3193                                 r10_bio->devs[1].addr = to_addr;
3194
3195                                 if (!test_bit(In_sync, &mrdev->flags)) {
3196                                         bio = r10_bio->devs[1].bio;
3197                                         bio_reset(bio);
3198                                         bio->bi_next = biolist;
3199                                         biolist = bio;
3200                                         bio->bi_private = r10_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->bi_bdev = 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 mreplace != NULL, then bio
3215                                  * cannot be NULL as r10buf_pool_alloc will
3216                                  * have allocated it.
3217                                  * So the second test here is pointless.
3218                                  * But it keeps semantic-checkers happy, and
3219                                  * this comment keeps human reviewers
3220                                  * happy.
3221                                  */
3222                                 if (mreplace == NULL || bio == NULL ||
3223                                     test_bit(Faulty, &mreplace->flags))
3224                                         break;
3225                                 bio_reset(bio);
3226                                 bio->bi_next = biolist;
3227                                 biolist = bio;
3228                                 bio->bi_private = r10_bio;
3229                                 bio->bi_end_io = end_sync_write;
3230                                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3231                                 bio->bi_iter.bi_sector = to_addr +
3232                                         mreplace->data_offset;
3233                                 bio->bi_bdev = mreplace->bdev;
3234                                 atomic_inc(&r10_bio->remaining);
3235                                 break;
3236                         }
3237                         rcu_read_unlock();
3238                         if (j == conf->copies) {
3239                                 /* Cannot recover, so abort the recovery or
3240                                  * record a bad block */
3241                                 if (any_working) {
3242                                         /* problem is that there are bad blocks
3243                                          * on other device(s)
3244                                          */
3245                                         int k;
3246                                         for (k = 0; k < conf->copies; k++)
3247                                                 if (r10_bio->devs[k].devnum == i)
3248                                                         break;
3249                                         if (!test_bit(In_sync,
3250                                                       &mrdev->flags)
3251                                             && !rdev_set_badblocks(
3252                                                     mrdev,
3253                                                     r10_bio->devs[k].addr,
3254                                                     max_sync, 0))
3255                                                 any_working = 0;
3256                                         if (mreplace &&
3257                                             !rdev_set_badblocks(
3258                                                     mreplace,
3259                                                     r10_bio->devs[k].addr,
3260                                                     max_sync, 0))
3261                                                 any_working = 0;
3262                                 }
3263                                 if (!any_working)  {
3264                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
3265                                                               &mddev->recovery))
3266                                                 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3267                                                        mdname(mddev));
3268                                         mirror->recovery_disabled
3269                                                 = mddev->recovery_disabled;
3270                                 }
3271                                 put_buf(r10_bio);
3272                                 if (rb2)
3273                                         atomic_dec(&rb2->remaining);
3274                                 r10_bio = rb2;
3275                                 rdev_dec_pending(mrdev, mddev);
3276                                 if (mreplace)
3277                                         rdev_dec_pending(mreplace, mddev);
3278                                 break;
3279                         }
3280                         rdev_dec_pending(mrdev, mddev);
3281                         if (mreplace)
3282                                 rdev_dec_pending(mreplace, mddev);
3283                         if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3284                                 /* Only want this if there is elsewhere to
3285                                  * read from. 'j' is currently the first
3286                                  * readable copy.
3287                                  */
3288                                 int targets = 1;
3289                                 for (; j < conf->copies; j++) {
3290                                         int d = r10_bio->devs[j].devnum;
3291                                         if (conf->mirrors[d].rdev &&
3292                                             test_bit(In_sync,
3293                                                       &conf->mirrors[d].rdev->flags))
3294                                                 targets++;
3295                                 }
3296                                 if (targets == 1)
3297                                         r10_bio->devs[0].bio->bi_opf
3298                                                 &= ~MD_FAILFAST;
3299                         }
3300                 }
3301                 if (biolist == NULL) {
3302                         while (r10_bio) {
3303                                 struct r10bio *rb2 = r10_bio;
3304                                 r10_bio = (struct r10bio*) rb2->master_bio;
3305                                 rb2->master_bio = NULL;
3306                                 put_buf(rb2);
3307                         }
3308                         goto giveup;
3309                 }
3310         } else {
3311                 /* resync. Schedule a read for every block at this virt offset */
3312                 int count = 0;
3313
3314                 bitmap_cond_end_sync(mddev->bitmap, sector_nr, 0);
3315
3316                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3317                                        &sync_blocks, mddev->degraded) &&
3318                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3319                                                  &mddev->recovery)) {
3320                         /* We can skip this block */
3321                         *skipped = 1;
3322                         return sync_blocks + sectors_skipped;
3323                 }
3324                 if (sync_blocks < max_sync)
3325                         max_sync = sync_blocks;
3326                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3327                 r10_bio->state = 0;
3328
3329                 r10_bio->mddev = mddev;
3330                 atomic_set(&r10_bio->remaining, 0);
3331                 raise_barrier(conf, 0);
3332                 conf->next_resync = sector_nr;
3333
3334                 r10_bio->master_bio = NULL;
3335                 r10_bio->sector = sector_nr;
3336                 set_bit(R10BIO_IsSync, &r10_bio->state);
3337                 raid10_find_phys(conf, r10_bio);
3338                 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3339
3340                 for (i = 0; i < conf->copies; i++) {
3341                         int d = r10_bio->devs[i].devnum;
3342                         sector_t first_bad, sector;
3343                         int bad_sectors;
3344                         struct md_rdev *rdev;
3345
3346                         if (r10_bio->devs[i].repl_bio)
3347                                 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3348
3349                         bio = r10_bio->devs[i].bio;
3350                         bio_reset(bio);
3351                         bio->bi_error = -EIO;
3352                         rcu_read_lock();
3353                         rdev = rcu_dereference(conf->mirrors[d].rdev);
3354                         if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3355                                 rcu_read_unlock();
3356                                 continue;
3357                         }
3358                         sector = r10_bio->devs[i].addr;
3359                         if (is_badblock(rdev, sector, max_sync,
3360                                         &first_bad, &bad_sectors)) {
3361                                 if (first_bad > sector)
3362                                         max_sync = first_bad - sector;
3363                                 else {
3364                                         bad_sectors -= (sector - first_bad);
3365                                         if (max_sync > bad_sectors)
3366                                                 max_sync = bad_sectors;
3367                                         rcu_read_unlock();
3368                                         continue;
3369                                 }
3370                         }
3371                         atomic_inc(&rdev->nr_pending);
3372                         atomic_inc(&r10_bio->remaining);
3373                         bio->bi_next = biolist;
3374                         biolist = bio;
3375                         bio->bi_private = r10_bio;
3376                         bio->bi_end_io = end_sync_read;
3377                         bio_set_op_attrs(bio, REQ_OP_READ, 0);
3378                         if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
3379                                 bio->bi_opf |= MD_FAILFAST;
3380                         bio->bi_iter.bi_sector = sector + rdev->data_offset;
3381                         bio->bi_bdev = 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                         rcu_read_unlock();
3391
3392                         /* Need to set up for writing to the replacement */
3393                         bio = r10_bio->devs[i].repl_bio;
3394                         bio_reset(bio);
3395                         bio->bi_error = -EIO;
3396
3397                         sector = r10_bio->devs[i].addr;
3398                         bio->bi_next = biolist;
3399                         biolist = bio;
3400                         bio->bi_private = r10_bio;
3401                         bio->bi_end_io = end_sync_write;
3402                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3403                         if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
3404                                 bio->bi_opf |= MD_FAILFAST;
3405                         bio->bi_iter.bi_sector = sector + rdev->data_offset;
3406                         bio->bi_bdev = rdev->bdev;
3407                         count++;
3408                 }
3409
3410                 if (count < 2) {
3411                         for (i=0; i<conf->copies; i++) {
3412                                 int d = r10_bio->devs[i].devnum;
3413                                 if (r10_bio->devs[i].bio->bi_end_io)
3414                                         rdev_dec_pending(conf->mirrors[d].rdev,
3415                                                          mddev);
3416                                 if (r10_bio->devs[i].repl_bio &&
3417                                     r10_bio->devs[i].repl_bio->bi_end_io)
3418                                         rdev_dec_pending(
3419                                                 conf->mirrors[d].replacement,
3420                                                 mddev);
3421                         }
3422                         put_buf(r10_bio);
3423                         biolist = NULL;
3424                         goto giveup;
3425                 }
3426         }
3427
3428         nr_sectors = 0;
3429         if (sector_nr + max_sync < max_sector)
3430                 max_sector = sector_nr + max_sync;
3431         do {
3432                 struct page *page;
3433                 int len = PAGE_SIZE;
3434                 if (sector_nr + (len>>9) > max_sector)
3435                         len = (max_sector - sector_nr) << 9;
3436                 if (len == 0)
3437                         break;
3438                 for (bio= biolist ; bio ; bio=bio->bi_next) {
3439                         struct bio *bio2;
3440                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3441                         if (bio_add_page(bio, page, len, 0))
3442                                 continue;
3443
3444                         /* stop here */
3445                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3446                         for (bio2 = biolist;
3447                              bio2 && bio2 != bio;
3448                              bio2 = bio2->bi_next) {
3449                                 /* remove last page from this bio */
3450                                 bio2->bi_vcnt--;
3451                                 bio2->bi_iter.bi_size -= len;
3452                                 bio_clear_flag(bio2, BIO_SEG_VALID);
3453                         }
3454                         goto bio_full;
3455                 }
3456                 nr_sectors += len>>9;
3457                 sector_nr += len>>9;
3458         } while (biolist->bi_vcnt < RESYNC_PAGES);
3459  bio_full:
3460         r10_bio->sectors = nr_sectors;
3461
3462         while (biolist) {
3463                 bio = biolist;
3464                 biolist = biolist->bi_next;
3465
3466                 bio->bi_next = NULL;
3467                 r10_bio = bio->bi_private;
3468                 r10_bio->sectors = nr_sectors;
3469
3470                 if (bio->bi_end_io == end_sync_read) {
3471                         md_sync_acct(bio->bi_bdev, nr_sectors);
3472                         bio->bi_error = 0;
3473                         generic_make_request(bio);
3474                 }
3475         }
3476
3477         if (sectors_skipped)
3478                 /* pretend they weren't skipped, it makes
3479                  * no important difference in this case
3480                  */
3481                 md_done_sync(mddev, sectors_skipped, 1);
3482
3483         return sectors_skipped + nr_sectors;
3484  giveup:
3485         /* There is nowhere to write, so all non-sync
3486          * drives must be failed or in resync, all drives
3487          * have a bad block, so try the next chunk...
3488          */
3489         if (sector_nr + max_sync < max_sector)
3490                 max_sector = sector_nr + max_sync;
3491
3492         sectors_skipped += (max_sector - sector_nr);
3493         chunks_skipped ++;
3494         sector_nr = max_sector;
3495         goto skipped;
3496 }
3497
3498 static sector_t
3499 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3500 {
3501         sector_t size;
3502         struct r10conf *conf = mddev->private;
3503
3504         if (!raid_disks)
3505                 raid_disks = min(conf->geo.raid_disks,
3506                                  conf->prev.raid_disks);
3507         if (!sectors)
3508                 sectors = conf->dev_sectors;
3509
3510         size = sectors >> conf->geo.chunk_shift;
3511         sector_div(size, conf->geo.far_copies);
3512         size = size * raid_disks;
3513         sector_div(size, conf->geo.near_copies);
3514
3515         return size << conf->geo.chunk_shift;
3516 }
3517
3518 static void calc_sectors(struct r10conf *conf, sector_t size)
3519 {
3520         /* Calculate the number of sectors-per-device that will
3521          * actually be used, and set conf->dev_sectors and
3522          * conf->stride
3523          */
3524
3525         size = size >> conf->geo.chunk_shift;
3526         sector_div(size, conf->geo.far_copies);
3527         size = size * conf->geo.raid_disks;
3528         sector_div(size, conf->geo.near_copies);
3529         /* 'size' is now the number of chunks in the array */
3530         /* calculate "used chunks per device" */
3531         size = size * conf->copies;
3532
3533         /* We need to round up when dividing by raid_disks to
3534          * get the stride size.
3535          */
3536         size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3537
3538         conf->dev_sectors = size << conf->geo.chunk_shift;
3539
3540         if (conf->geo.far_offset)
3541                 conf->geo.stride = 1 << conf->geo.chunk_shift;
3542         else {
3543                 sector_div(size, conf->geo.far_copies);
3544                 conf->geo.stride = size << conf->geo.chunk_shift;
3545         }
3546 }
3547
3548 enum geo_type {geo_new, geo_old, geo_start};
3549 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3550 {
3551         int nc, fc, fo;
3552         int layout, chunk, disks;
3553         switch (new) {
3554         case geo_old:
3555                 layout = mddev->layout;
3556                 chunk = mddev->chunk_sectors;
3557                 disks = mddev->raid_disks - mddev->delta_disks;
3558                 break;
3559         case geo_new:
3560                 layout = mddev->new_layout;
3561                 chunk = mddev->new_chunk_sectors;
3562                 disks = mddev->raid_disks;
3563                 break;
3564         default: /* avoid 'may be unused' warnings */
3565         case geo_start: /* new when starting reshape - raid_disks not
3566                          * updated yet. */
3567                 layout = mddev->new_layout;
3568                 chunk = mddev->new_chunk_sectors;
3569                 disks = mddev->raid_disks + mddev->delta_disks;
3570                 break;
3571         }
3572         if (layout >> 19)
3573                 return -1;
3574         if (chunk < (PAGE_SIZE >> 9) ||
3575             !is_power_of_2(chunk))
3576                 return -2;
3577         nc = layout & 255;
3578         fc = (layout >> 8) & 255;
3579         fo = layout & (1<<16);
3580         geo->raid_disks = disks;
3581         geo->near_copies = nc;
3582         geo->far_copies = fc;
3583         geo->far_offset = fo;
3584         switch (layout >> 17) {
3585         case 0: /* original layout.  simple but not always optimal */
3586                 geo->far_set_size = disks;
3587                 break;
3588         case 1: /* "improved" layout which was buggy.  Hopefully no-one is
3589                  * actually using this, but leave code here just in case.*/
3590                 geo->far_set_size = disks/fc;
3591                 WARN(geo->far_set_size < fc,
3592                      "This RAID10 layout does not provide data safety - please backup and create new array\n");
3593                 break;
3594         case 2: /* "improved" layout fixed to match documentation */
3595                 geo->far_set_size = fc * nc;
3596                 break;
3597         default: /* Not a valid layout */
3598                 return -1;
3599         }
3600         geo->chunk_mask = chunk - 1;
3601         geo->chunk_shift = ffz(~chunk);
3602         return nc*fc;
3603 }
3604
3605 static struct r10conf *setup_conf(struct mddev *mddev)
3606 {
3607         struct r10conf *conf = NULL;
3608         int err = -EINVAL;
3609         struct geom geo;
3610         int copies;
3611
3612         copies = setup_geo(&geo, mddev, geo_new);
3613
3614         if (copies == -2) {
3615                 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3616                         mdname(mddev), PAGE_SIZE);
3617                 goto out;
3618         }
3619
3620         if (copies < 2 || copies > mddev->raid_disks) {
3621                 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3622                         mdname(mddev), mddev->new_layout);
3623                 goto out;
3624         }
3625
3626         err = -ENOMEM;
3627         conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3628         if (!conf)
3629                 goto out;
3630
3631         /* FIXME calc properly */
3632         conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3633                                                             max(0,-mddev->delta_disks)),
3634                                 GFP_KERNEL);
3635         if (!conf->mirrors)
3636                 goto out;
3637
3638         conf->tmppage = alloc_page(GFP_KERNEL);
3639         if (!conf->tmppage)
3640                 goto out;
3641
3642         conf->geo = geo;
3643         conf->copies = copies;
3644         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3645                                            r10bio_pool_free, conf);
3646         if (!conf->r10bio_pool)
3647                 goto out;
3648
3649         calc_sectors(conf, mddev->dev_sectors);
3650         if (mddev->reshape_position == MaxSector) {
3651                 conf->prev = conf->geo;
3652                 conf->reshape_progress = MaxSector;
3653         } else {
3654                 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3655                         err = -EINVAL;
3656                         goto out;
3657                 }
3658                 conf->reshape_progress = mddev->reshape_position;
3659                 if (conf->prev.far_offset)
3660                         conf->prev.stride = 1 << conf->prev.chunk_shift;
3661                 else
3662                         /* far_copies must be 1 */
3663                         conf->prev.stride = conf->dev_sectors;
3664         }
3665         conf->reshape_safe = conf->reshape_progress;
3666         spin_lock_init(&conf->device_lock);
3667         INIT_LIST_HEAD(&conf->retry_list);
3668         INIT_LIST_HEAD(&conf->bio_end_io_list);
3669
3670         spin_lock_init(&conf->resync_lock);
3671         init_waitqueue_head(&conf->wait_barrier);
3672         atomic_set(&conf->nr_pending, 0);
3673
3674         conf->thread = md_register_thread(raid10d, mddev, "raid10");
3675         if (!conf->thread)
3676                 goto out;
3677
3678         conf->mddev = mddev;
3679         return conf;
3680
3681  out:
3682         if (conf) {
3683                 mempool_destroy(conf->r10bio_pool);
3684                 kfree(conf->mirrors);
3685                 safe_put_page(conf->tmppage);
3686                 kfree(conf);
3687         }
3688         return ERR_PTR(err);
3689 }
3690
3691 static int raid10_run(struct mddev *mddev)
3692 {
3693         struct r10conf *conf;
3694         int i, disk_idx, chunk_size;
3695         struct raid10_info *disk;
3696         struct md_rdev *rdev;
3697         sector_t size;
3698         sector_t min_offset_diff = 0;
3699         int first = 1;
3700         bool discard_supported = false;
3701
3702         if (mddev->private == NULL) {
3703                 conf = setup_conf(mddev);
3704                 if (IS_ERR(conf))
3705                         return PTR_ERR(conf);
3706                 mddev->private = conf;
3707         }
3708         conf = mddev->private;
3709         if (!conf)
3710                 goto out;
3711
3712         mddev->thread = conf->thread;
3713         conf->thread = NULL;
3714
3715         chunk_size = mddev->chunk_sectors << 9;
3716         if (mddev->queue) {
3717                 blk_queue_max_discard_sectors(mddev->queue,
3718                                               mddev->chunk_sectors);
3719                 blk_queue_max_write_same_sectors(mddev->queue, 0);
3720                 blk_queue_io_min(mddev->queue, chunk_size);
3721                 if (conf->geo.raid_disks % conf->geo.near_copies)
3722                         blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3723                 else
3724                         blk_queue_io_opt(mddev->queue, chunk_size *
3725                                          (conf->geo.raid_disks / conf->geo.near_copies));
3726         }
3727
3728         rdev_for_each(rdev, mddev) {
3729                 long long diff;
3730                 struct request_queue *q;
3731
3732                 disk_idx = rdev->raid_disk;
3733                 if (disk_idx < 0)
3734                         continue;
3735                 if (disk_idx >= conf->geo.raid_disks &&
3736                     disk_idx >= conf->prev.raid_disks)
3737                         continue;
3738                 disk = conf->mirrors + disk_idx;
3739
3740                 if (test_bit(Replacement, &rdev->flags)) {
3741                         if (disk->replacement)
3742                                 goto out_free_conf;
3743                         disk->replacement = rdev;
3744                 } else {
3745                         if (disk->rdev)
3746                                 goto out_free_conf;
3747                         disk->rdev = rdev;
3748                 }
3749                 q = bdev_get_queue(rdev->bdev);
3750                 diff = (rdev->new_data_offset - rdev->data_offset);
3751                 if (!mddev->reshape_backwards)
3752                         diff = -diff;
3753                 if (diff < 0)
3754                         diff = 0;
3755                 if (first || diff < min_offset_diff)
3756                         min_offset_diff = diff;
3757
3758                 if (mddev->gendisk)
3759                         disk_stack_limits(mddev->gendisk, rdev->bdev,
3760                                           rdev->data_offset << 9);
3761
3762                 disk->head_position = 0;
3763
3764                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3765                         discard_supported = true;
3766         }
3767
3768         if (mddev->queue) {
3769                 if (discard_supported)
3770                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3771                                                 mddev->queue);
3772                 else
3773                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3774                                                   mddev->queue);
3775         }
3776         /* need to check that every block has at least one working mirror */
3777         if (!enough(conf, -1)) {
3778                 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3779                        mdname(mddev));
3780                 goto out_free_conf;
3781         }
3782
3783         if (conf->reshape_progress != MaxSector) {
3784                 /* must ensure that shape change is supported */
3785                 if (conf->geo.far_copies != 1 &&
3786                     conf->geo.far_offset == 0)
3787                         goto out_free_conf;
3788                 if (conf->prev.far_copies != 1 &&
3789                     conf->prev.far_offset == 0)
3790                         goto out_free_conf;
3791         }
3792
3793         mddev->degraded = 0;
3794         for (i = 0;
3795              i < conf->geo.raid_disks
3796                      || i < conf->prev.raid_disks;
3797              i++) {
3798
3799                 disk = conf->mirrors + i;
3800
3801                 if (!disk->rdev && disk->replacement) {
3802                         /* The replacement is all we have - use it */
3803                         disk->rdev = disk->replacement;
3804                         disk->replacement = NULL;
3805                         clear_bit(Replacement, &disk->rdev->flags);
3806                 }
3807
3808                 if (!disk->rdev ||
3809                     !test_bit(In_sync, &disk->rdev->flags)) {
3810                         disk->head_position = 0;
3811                         mddev->degraded++;
3812                         if (disk->rdev &&
3813                             disk->rdev->saved_raid_disk < 0)
3814                                 conf->fullsync = 1;
3815                 }
3816                 disk->recovery_disabled = mddev->recovery_disabled - 1;
3817         }
3818
3819         if (mddev->recovery_cp != MaxSector)
3820                 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3821                           mdname(mddev));
3822         pr_info("md/raid10:%s: active with %d out of %d devices\n",
3823                 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3824                 conf->geo.raid_disks);
3825         /*
3826          * Ok, everything is just fine now
3827          */
3828         mddev->dev_sectors = conf->dev_sectors;
3829         size = raid10_size(mddev, 0, 0);
3830         md_set_array_sectors(mddev, size);
3831         mddev->resync_max_sectors = size;
3832         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3833
3834         if (mddev->queue) {
3835                 int stripe = conf->geo.raid_disks *
3836                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3837
3838                 /* Calculate max read-ahead size.
3839                  * We need to readahead at least twice a whole stripe....
3840                  * maybe...
3841                  */
3842                 stripe /= conf->geo.near_copies;
3843                 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3844                         mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3845         }
3846
3847         if (md_integrity_register(mddev))
3848                 goto out_free_conf;
3849
3850         if (conf->reshape_progress != MaxSector) {
3851                 unsigned long before_length, after_length;
3852
3853                 before_length = ((1 << conf->prev.chunk_shift) *
3854                                  conf->prev.far_copies);
3855                 after_length = ((1 << conf->geo.chunk_shift) *
3856                                 conf->geo.far_copies);
3857
3858                 if (max(before_length, after_length) > min_offset_diff) {
3859                         /* This cannot work */
3860                         pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3861                         goto out_free_conf;
3862                 }
3863                 conf->offset_diff = min_offset_diff;
3864
3865                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3866                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3867                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3868                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3869                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3870                                                         "reshape");
3871         }
3872
3873         return 0;
3874
3875 out_free_conf:
3876         md_unregister_thread(&mddev->thread);
3877         mempool_destroy(conf->r10bio_pool);
3878         safe_put_page(conf->tmppage);
3879         kfree(conf->mirrors);
3880         kfree(conf);
3881         mddev->private = NULL;
3882 out:
3883         return -EIO;
3884 }
3885
3886 static void raid10_free(struct mddev *mddev, void *priv)
3887 {
3888         struct r10conf *conf = priv;
3889
3890         mempool_destroy(conf->r10bio_pool);
3891         safe_put_page(conf->tmppage);
3892         kfree(conf->mirrors);
3893         kfree(conf->mirrors_old);
3894         kfree(conf->mirrors_new);
3895         kfree(conf);
3896 }
3897
3898 static void raid10_quiesce(struct mddev *mddev, int state)
3899 {
3900         struct r10conf *conf = mddev->private;
3901
3902         switch(state) {
3903         case 1:
3904                 raise_barrier(conf, 0);
3905                 break;
3906         case 0:
3907                 lower_barrier(conf);
3908                 break;
3909         }
3910 }
3911
3912 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3913 {
3914         /* Resize of 'far' arrays is not supported.
3915          * For 'near' and 'offset' arrays we can set the
3916          * number of sectors used to be an appropriate multiple
3917          * of the chunk size.
3918          * For 'offset', this is far_copies*chunksize.
3919          * For 'near' the multiplier is the LCM of
3920          * near_copies and raid_disks.
3921          * So if far_copies > 1 && !far_offset, fail.
3922          * Else find LCM(raid_disks, near_copy)*far_copies and
3923          * multiply by chunk_size.  Then round to this number.
3924          * This is mostly done by raid10_size()
3925          */
3926         struct r10conf *conf = mddev->private;
3927         sector_t oldsize, size;
3928
3929         if (mddev->reshape_position != MaxSector)
3930                 return -EBUSY;
3931
3932         if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3933                 return -EINVAL;
3934
3935         oldsize = raid10_size(mddev, 0, 0);
3936         size = raid10_size(mddev, sectors, 0);
3937         if (mddev->external_size &&
3938             mddev->array_sectors > size)
3939                 return -EINVAL;
3940         if (mddev->bitmap) {
3941                 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3942                 if (ret)
3943                         return ret;
3944         }
3945         md_set_array_sectors(mddev, size);
3946         if (mddev->queue) {
3947                 set_capacity(mddev->gendisk, mddev->array_sectors);
3948                 revalidate_disk(mddev->gendisk);
3949         }
3950         if (sectors > mddev->dev_sectors &&
3951             mddev->recovery_cp > oldsize) {
3952                 mddev->recovery_cp = oldsize;
3953                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3954         }
3955         calc_sectors(conf, sectors);
3956         mddev->dev_sectors = conf->dev_sectors;
3957         mddev->resync_max_sectors = size;
3958         return 0;
3959 }
3960
3961 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3962 {
3963         struct md_rdev *rdev;
3964         struct r10conf *conf;
3965
3966         if (mddev->degraded > 0) {
3967                 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
3968                         mdname(mddev));
3969                 return ERR_PTR(-EINVAL);
3970         }
3971         sector_div(size, devs);
3972
3973         /* Set new parameters */
3974         mddev->new_level = 10;
3975         /* new layout: far_copies = 1, near_copies = 2 */
3976         mddev->new_layout = (1<<8) + 2;
3977         mddev->new_chunk_sectors = mddev->chunk_sectors;
3978         mddev->delta_disks = mddev->raid_disks;
3979         mddev->raid_disks *= 2;
3980         /* make sure it will be not marked as dirty */
3981         mddev->recovery_cp = MaxSector;
3982         mddev->dev_sectors = size;
3983
3984         conf = setup_conf(mddev);
3985         if (!IS_ERR(conf)) {
3986                 rdev_for_each(rdev, mddev)
3987                         if (rdev->raid_disk >= 0) {
3988                                 rdev->new_raid_disk = rdev->raid_disk * 2;
3989                                 rdev->sectors = size;
3990                         }
3991                 conf->barrier = 1;
3992         }
3993
3994         return conf;
3995 }
3996
3997 static void *raid10_takeover(struct mddev *mddev)
3998 {
3999         struct r0conf *raid0_conf;
4000
4001         /* raid10 can take over:
4002          *  raid0 - providing it has only two drives
4003          */
4004         if (mddev->level == 0) {
4005                 /* for raid0 takeover only one zone is supported */
4006                 raid0_conf = mddev->private;
4007                 if (raid0_conf->nr_strip_zones > 1) {
4008                         pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4009                                 mdname(mddev));
4010                         return ERR_PTR(-EINVAL);
4011                 }
4012                 return raid10_takeover_raid0(mddev,
4013                         raid0_conf->strip_zone->zone_end,
4014                         raid0_conf->strip_zone->nb_dev);
4015         }
4016         return ERR_PTR(-EINVAL);
4017 }
4018
4019 static int raid10_check_reshape(struct mddev *mddev)
4020 {
4021         /* Called when there is a request to change
4022          * - layout (to ->new_layout)
4023          * - chunk size (to ->new_chunk_sectors)
4024          * - raid_disks (by delta_disks)
4025          * or when trying to restart a reshape that was ongoing.
4026          *
4027          * We need to validate the request and possibly allocate
4028          * space if that might be an issue later.
4029          *
4030          * Currently we reject any reshape of a 'far' mode array,
4031          * allow chunk size to change if new is generally acceptable,
4032          * allow raid_disks to increase, and allow
4033          * a switch between 'near' mode and 'offset' mode.
4034          */
4035         struct r10conf *conf = mddev->private;
4036         struct geom geo;
4037
4038         if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4039                 return -EINVAL;
4040
4041         if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4042                 /* mustn't change number of copies */
4043                 return -EINVAL;
4044         if (geo.far_copies > 1 && !geo.far_offset)
4045                 /* Cannot switch to 'far' mode */
4046                 return -EINVAL;
4047
4048         if (mddev->array_sectors & geo.chunk_mask)
4049                         /* not factor of array size */
4050                         return -EINVAL;
4051
4052         if (!enough(conf, -1))
4053                 return -EINVAL;
4054
4055         kfree(conf->mirrors_new);
4056         conf->mirrors_new = NULL;
4057         if (mddev->delta_disks > 0) {
4058                 /* allocate new 'mirrors' list */
4059                 conf->mirrors_new = kzalloc(
4060                         sizeof(struct raid10_info)
4061                         *(mddev->raid_disks +
4062                           mddev->delta_disks),
4063                         GFP_KERNEL);
4064                 if (!conf->mirrors_new)
4065                         return -ENOMEM;
4066         }
4067         return 0;
4068 }
4069
4070 /*
4071  * Need to check if array has failed when deciding whether to:
4072  *  - start an array
4073  *  - remove non-faulty devices
4074  *  - add a spare
4075  *  - allow a reshape
4076  * This determination is simple when no reshape is happening.
4077  * However if there is a reshape, we need to carefully check
4078  * both the before and after sections.
4079  * This is because some failed devices may only affect one
4080  * of the two sections, and some non-in_sync devices may
4081  * be insync in the section most affected by failed devices.
4082  */
4083 static int calc_degraded(struct r10conf *conf)
4084 {
4085         int degraded, degraded2;
4086         int i;
4087
4088         rcu_read_lock();
4089         degraded = 0;
4090         /* 'prev' section first */
4091         for (i = 0; i < conf->prev.raid_disks; i++) {
4092                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4093                 if (!rdev || test_bit(Faulty, &rdev->flags))
4094                         degraded++;
4095                 else if (!test_bit(In_sync, &rdev->flags))
4096                         /* When we can reduce the number of devices in
4097                          * an array, this might not contribute to
4098                          * 'degraded'.  It does now.
4099                          */
4100                         degraded++;
4101         }
4102         rcu_read_unlock();
4103         if (conf->geo.raid_disks == conf->prev.raid_disks)
4104                 return degraded;
4105         rcu_read_lock();
4106         degraded2 = 0;
4107         for (i = 0; i < conf->geo.raid_disks; i++) {
4108                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4109                 if (!rdev || test_bit(Faulty, &rdev->flags))
4110                         degraded2++;
4111                 else if (!test_bit(In_sync, &rdev->flags)) {
4112                         /* If reshape is increasing the number of devices,
4113                          * this section has already been recovered, so
4114                          * it doesn't contribute to degraded.
4115                          * else it does.
4116                          */
4117                         if (conf->geo.raid_disks <= conf->prev.raid_disks)
4118                                 degraded2++;
4119                 }
4120         }
4121         rcu_read_unlock();
4122         if (degraded2 > degraded)
4123                 return degraded2;
4124         return degraded;
4125 }
4126
4127 static int raid10_start_reshape(struct mddev *mddev)
4128 {
4129         /* A 'reshape' has been requested. This commits
4130          * the various 'new' fields and sets MD_RECOVER_RESHAPE
4131          * This also checks if there are enough spares and adds them
4132          * to the array.
4133          * We currently require enough spares to make the final
4134          * array non-degraded.  We also require that the difference
4135          * between old and new data_offset - on each device - is
4136          * enough that we never risk over-writing.
4137          */
4138
4139         unsigned long before_length, after_length;
4140         sector_t min_offset_diff = 0;
4141         int first = 1;
4142         struct geom new;
4143         struct r10conf *conf = mddev->private;
4144         struct md_rdev *rdev;
4145         int spares = 0;
4146         int ret;
4147
4148         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4149                 return -EBUSY;
4150
4151         if (setup_geo(&new, mddev, geo_start) != conf->copies)
4152                 return -EINVAL;
4153
4154         before_length = ((1 << conf->prev.chunk_shift) *
4155                          conf->prev.far_copies);
4156         after_length = ((1 << conf->geo.chunk_shift) *
4157                         conf->geo.far_copies);
4158
4159         rdev_for_each(rdev, mddev) {
4160                 if (!test_bit(In_sync, &rdev->flags)
4161                     && !test_bit(Faulty, &rdev->flags))
4162                         spares++;
4163                 if (rdev->raid_disk >= 0) {
4164                         long long diff = (rdev->new_data_offset
4165                                           - rdev->data_offset);
4166                         if (!mddev->reshape_backwards)
4167                                 diff = -diff;
4168                         if (diff < 0)
4169                                 diff = 0;
4170                         if (first || diff < min_offset_diff)
4171                                 min_offset_diff = diff;
4172                 }
4173         }
4174
4175         if (max(before_length, after_length) > min_offset_diff)
4176                 return -EINVAL;
4177
4178         if (spares < mddev->delta_disks)
4179                 return -EINVAL;
4180
4181         conf->offset_diff = min_offset_diff;
4182         spin_lock_irq(&conf->device_lock);
4183         if (conf->mirrors_new) {
4184                 memcpy(conf->mirrors_new, conf->mirrors,
4185                        sizeof(struct raid10_info)*conf->prev.raid_disks);
4186                 smp_mb();
4187                 kfree(conf->mirrors_old);
4188                 conf->mirrors_old = conf->mirrors;
4189                 conf->mirrors = conf->mirrors_new;
4190                 conf->mirrors_new = NULL;
4191         }
4192         setup_geo(&conf->geo, mddev, geo_start);
4193         smp_mb();
4194         if (mddev->reshape_backwards) {
4195                 sector_t size = raid10_size(mddev, 0, 0);
4196                 if (size < mddev->array_sectors) {
4197                         spin_unlock_irq(&conf->device_lock);
4198                         pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4199                                 mdname(mddev));
4200                         return -EINVAL;
4201                 }
4202                 mddev->resync_max_sectors = size;
4203                 conf->reshape_progress = size;
4204         } else
4205                 conf->reshape_progress = 0;
4206         conf->reshape_safe = conf->reshape_progress;
4207         spin_unlock_irq(&conf->device_lock);
4208
4209         if (mddev->delta_disks && mddev->bitmap) {
4210                 ret = bitmap_resize(mddev->bitmap,
4211                                     raid10_size(mddev, 0,
4212                                                 conf->geo.raid_disks),
4213                                     0, 0);
4214                 if (ret)
4215                         goto abort;
4216         }
4217         if (mddev->delta_disks > 0) {
4218                 rdev_for_each(rdev, mddev)
4219                         if (rdev->raid_disk < 0 &&
4220                             !test_bit(Faulty, &rdev->flags)) {
4221                                 if (raid10_add_disk(mddev, rdev) == 0) {
4222                                         if (rdev->raid_disk >=
4223                                             conf->prev.raid_disks)
4224                                                 set_bit(In_sync, &rdev->flags);
4225                                         else
4226                                                 rdev->recovery_offset = 0;
4227
4228                                         if (sysfs_link_rdev(mddev, rdev))
4229                                                 /* Failure here  is OK */;
4230                                 }
4231                         } else if (rdev->raid_disk >= conf->prev.raid_disks
4232                                    && !test_bit(Faulty, &rdev->flags)) {
4233                                 /* This is a spare that was manually added */
4234                                 set_bit(In_sync, &rdev->flags);
4235                         }
4236         }
4237         /* When a reshape changes the number of devices,
4238          * ->degraded is measured against the larger of the
4239          * pre and  post numbers.
4240          */
4241         spin_lock_irq(&conf->device_lock);
4242         mddev->degraded = calc_degraded(conf);
4243         spin_unlock_irq(&conf->device_lock);
4244         mddev->raid_disks = conf->geo.raid_disks;
4245         mddev->reshape_position = conf->reshape_progress;
4246         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4247
4248         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4249         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4250         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4251         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4252         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4253
4254         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4255                                                 "reshape");
4256         if (!mddev->sync_thread) {
4257                 ret = -EAGAIN;
4258                 goto abort;
4259         }
4260         conf->reshape_checkpoint = jiffies;
4261         md_wakeup_thread(mddev->sync_thread);
4262         md_new_event(mddev);
4263         return 0;
4264
4265 abort:
4266         mddev->recovery = 0;
4267         spin_lock_irq(&conf->device_lock);
4268         conf->geo = conf->prev;
4269         mddev->raid_disks = conf->geo.raid_disks;
4270         rdev_for_each(rdev, mddev)
4271                 rdev->new_data_offset = rdev->data_offset;
4272         smp_wmb();
4273         conf->reshape_progress = MaxSector;
4274         conf->reshape_safe = MaxSector;
4275         mddev->reshape_position = MaxSector;
4276         spin_unlock_irq(&conf->device_lock);
4277         return ret;
4278 }
4279
4280 /* Calculate the last device-address that could contain
4281  * any block from the chunk that includes the array-address 's'
4282  * and report the next address.
4283  * i.e. the address returned will be chunk-aligned and after
4284  * any data that is in the chunk containing 's'.
4285  */
4286 static sector_t last_dev_address(sector_t s, struct geom *geo)
4287 {
4288         s = (s | geo->chunk_mask) + 1;
4289         s >>= geo->chunk_shift;
4290         s *= geo->near_copies;
4291         s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4292         s *= geo->far_copies;
4293         s <<= geo->chunk_shift;
4294         return s;
4295 }
4296
4297 /* Calculate the first device-address that could contain
4298  * any block from the chunk that includes the array-address 's'.
4299  * This too will be the start of a chunk
4300  */
4301 static sector_t first_dev_address(sector_t s, struct geom *geo)
4302 {
4303         s >>= geo->chunk_shift;
4304         s *= geo->near_copies;
4305         sector_div(s, geo->raid_disks);
4306         s *= geo->far_copies;
4307         s <<= geo->chunk_shift;
4308         return s;
4309 }
4310
4311 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4312                                 int *skipped)
4313 {
4314         /* We simply copy at most one chunk (smallest of old and new)
4315          * at a time, possibly less if that exceeds RESYNC_PAGES,
4316          * or we hit a bad block or something.
4317          * This might mean we pause for normal IO in the middle of
4318          * a chunk, but that is not a problem as mddev->reshape_position
4319          * can record any location.
4320          *
4321          * If we will want to write to a location that isn't
4322          * yet recorded as 'safe' (i.e. in metadata on disk) then
4323          * we need to flush all reshape requests and update the metadata.
4324          *
4325          * When reshaping forwards (e.g. to more devices), we interpret
4326          * 'safe' as the earliest block which might not have been copied
4327          * down yet.  We divide this by previous stripe size and multiply
4328          * by previous stripe length to get lowest device offset that we
4329          * cannot write to yet.
4330          * We interpret 'sector_nr' as an address that we want to write to.
4331          * From this we use last_device_address() to find where we might
4332          * write to, and first_device_address on the  'safe' position.
4333          * If this 'next' write position is after the 'safe' position,
4334          * we must update the metadata to increase the 'safe' position.
4335          *
4336          * When reshaping backwards, we round in the opposite direction
4337          * and perform the reverse test:  next write position must not be
4338          * less than current safe position.
4339          *
4340          * In all this the minimum difference in data offsets
4341          * (conf->offset_diff - always positive) allows a bit of slack,
4342          * so next can be after 'safe', but not by more than offset_diff
4343          *
4344          * We need to prepare all the bios here before we start any IO
4345          * to ensure the size we choose is acceptable to all devices.
4346          * The means one for each copy for write-out and an extra one for
4347          * read-in.
4348          * We store the read-in bio in ->master_bio and the others in
4349          * ->devs[x].bio and ->devs[x].repl_bio.
4350          */
4351         struct r10conf *conf = mddev->private;
4352         struct r10bio *r10_bio;
4353         sector_t next, safe, last;
4354         int max_sectors;
4355         int nr_sectors;
4356         int s;
4357         struct md_rdev *rdev;
4358         int need_flush = 0;
4359         struct bio *blist;
4360         struct bio *bio, *read_bio;
4361         int sectors_done = 0;
4362
4363         if (sector_nr == 0) {
4364                 /* If restarting in the middle, skip the initial sectors */
4365                 if (mddev->reshape_backwards &&
4366                     conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4367                         sector_nr = (raid10_size(mddev, 0, 0)
4368                                      - conf->reshape_progress);
4369                 } else if (!mddev->reshape_backwards &&
4370                            conf->reshape_progress > 0)
4371                         sector_nr = conf->reshape_progress;
4372                 if (sector_nr) {
4373                         mddev->curr_resync_completed = sector_nr;
4374                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4375                         *skipped = 1;
4376                         return sector_nr;
4377                 }
4378         }
4379
4380         /* We don't use sector_nr to track where we are up to
4381          * as that doesn't work well for ->reshape_backwards.
4382          * So just use ->reshape_progress.
4383          */
4384         if (mddev->reshape_backwards) {
4385                 /* 'next' is the earliest device address that we might
4386                  * write to for this chunk in the new layout
4387                  */
4388                 next = first_dev_address(conf->reshape_progress - 1,
4389                                          &conf->geo);
4390
4391                 /* 'safe' is the last device address that we might read from
4392                  * in the old layout after a restart
4393                  */
4394                 safe = last_dev_address(conf->reshape_safe - 1,
4395                                         &conf->prev);
4396
4397                 if (next + conf->offset_diff < safe)
4398                         need_flush = 1;
4399
4400                 last = conf->reshape_progress - 1;
4401                 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4402                                                & conf->prev.chunk_mask);
4403                 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4404                         sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4405         } else {
4406                 /* 'next' is after the last device address that we
4407                  * might write to for this chunk in the new layout
4408                  */
4409                 next = last_dev_address(conf->reshape_progress, &conf->geo);
4410
4411                 /* 'safe' is the earliest device address that we might
4412                  * read from in the old layout after a restart
4413                  */
4414                 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4415
4416                 /* Need to update metadata if 'next' might be beyond 'safe'
4417                  * as that would possibly corrupt data
4418                  */
4419                 if (next > safe + conf->offset_diff)
4420                         need_flush = 1;
4421
4422                 sector_nr = conf->reshape_progress;
4423                 last  = sector_nr | (conf->geo.chunk_mask
4424                                      & conf->prev.chunk_mask);
4425
4426                 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4427                         last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4428         }
4429
4430         if (need_flush ||
4431             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4432                 /* Need to update reshape_position in metadata */
4433                 wait_barrier(conf);
4434                 mddev->reshape_position = conf->reshape_progress;
4435                 if (mddev->reshape_backwards)
4436                         mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4437                                 - conf->reshape_progress;
4438                 else
4439                         mddev->curr_resync_completed = conf->reshape_progress;
4440                 conf->reshape_checkpoint = jiffies;
4441                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4442                 md_wakeup_thread(mddev->thread);
4443                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4444                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4445                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4446                         allow_barrier(conf);
4447                         return sectors_done;
4448                 }
4449                 conf->reshape_safe = mddev->reshape_position;
4450                 allow_barrier(conf);
4451         }
4452
4453 read_more:
4454         /* Now schedule reads for blocks from sector_nr to last */
4455         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4456         r10_bio->state = 0;
4457         raise_barrier(conf, sectors_done != 0);
4458         atomic_set(&r10_bio->remaining, 0);
4459         r10_bio->mddev = mddev;
4460         r10_bio->sector = sector_nr;
4461         set_bit(R10BIO_IsReshape, &r10_bio->state);
4462         r10_bio->sectors = last - sector_nr + 1;
4463         rdev = read_balance(conf, r10_bio, &max_sectors);
4464         BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4465
4466         if (!rdev) {
4467                 /* Cannot read from here, so need to record bad blocks
4468                  * on all the target devices.
4469                  */
4470                 // FIXME
4471                 mempool_free(r10_bio, conf->r10buf_pool);
4472                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4473                 return sectors_done;
4474         }
4475
4476         read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4477
4478         read_bio->bi_bdev = rdev->bdev;
4479         read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4480                                + rdev->data_offset);
4481         read_bio->bi_private = r10_bio;
4482         read_bio->bi_end_io = end_sync_read;
4483         bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4484         read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4485         read_bio->bi_error = 0;
4486         read_bio->bi_vcnt = 0;
4487         read_bio->bi_iter.bi_size = 0;
4488         r10_bio->master_bio = read_bio;
4489         r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4490
4491         /* Now find the locations in the new layout */
4492         __raid10_find_phys(&conf->geo, r10_bio);
4493
4494         blist = read_bio;
4495         read_bio->bi_next = NULL;
4496
4497         rcu_read_lock();
4498         for (s = 0; s < conf->copies*2; s++) {
4499                 struct bio *b;
4500                 int d = r10_bio->devs[s/2].devnum;
4501                 struct md_rdev *rdev2;
4502                 if (s&1) {
4503                         rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4504                         b = r10_bio->devs[s/2].repl_bio;
4505                 } else {
4506                         rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4507                         b = r10_bio->devs[s/2].bio;
4508                 }
4509                 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4510                         continue;
4511
4512                 bio_reset(b);
4513                 b->bi_bdev = rdev2->bdev;
4514                 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4515                         rdev2->new_data_offset;
4516                 b->bi_private = r10_bio;
4517                 b->bi_end_io = end_reshape_write;
4518                 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4519                 b->bi_next = blist;
4520                 blist = b;
4521         }
4522
4523         /* Now add as many pages as possible to all of these bios. */
4524
4525         nr_sectors = 0;
4526         for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4527                 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4528                 int len = (max_sectors - s) << 9;
4529                 if (len > PAGE_SIZE)
4530                         len = PAGE_SIZE;
4531                 for (bio = blist; bio ; bio = bio->bi_next) {
4532                         struct bio *bio2;
4533                         if (bio_add_page(bio, page, len, 0))
4534                                 continue;
4535
4536                         /* Didn't fit, must stop */
4537                         for (bio2 = blist;
4538                              bio2 && bio2 != bio;
4539                              bio2 = bio2->bi_next) {
4540                                 /* Remove last page from this bio */
4541                                 bio2->bi_vcnt--;
4542                                 bio2->bi_iter.bi_size -= len;
4543                                 bio_clear_flag(bio2, BIO_SEG_VALID);
4544                         }
4545                         goto bio_full;
4546                 }
4547                 sector_nr += len >> 9;
4548                 nr_sectors += len >> 9;
4549         }
4550 bio_full:
4551         rcu_read_unlock();
4552         r10_bio->sectors = nr_sectors;
4553
4554         /* Now submit the read */
4555         md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4556         atomic_inc(&r10_bio->remaining);
4557         read_bio->bi_next = NULL;
4558         generic_make_request(read_bio);
4559         sector_nr += nr_sectors;
4560         sectors_done += nr_sectors;
4561         if (sector_nr <= last)
4562                 goto read_more;
4563
4564         /* Now that we have done the whole section we can
4565          * update reshape_progress
4566          */
4567         if (mddev->reshape_backwards)
4568                 conf->reshape_progress -= sectors_done;
4569         else
4570                 conf->reshape_progress += sectors_done;
4571
4572         return sectors_done;
4573 }
4574
4575 static void end_reshape_request(struct r10bio *r10_bio);
4576 static int handle_reshape_read_error(struct mddev *mddev,
4577                                      struct r10bio *r10_bio);
4578 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4579 {
4580         /* Reshape read completed.  Hopefully we have a block
4581          * to write out.
4582          * If we got a read error then we do sync 1-page reads from
4583          * elsewhere until we find the data - or give up.
4584          */
4585         struct r10conf *conf = mddev->private;
4586         int s;
4587
4588         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4589                 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4590                         /* Reshape has been aborted */
4591                         md_done_sync(mddev, r10_bio->sectors, 0);
4592                         return;
4593                 }
4594
4595         /* We definitely have the data in the pages, schedule the
4596          * writes.
4597          */
4598         atomic_set(&r10_bio->remaining, 1);
4599         for (s = 0; s < conf->copies*2; s++) {
4600                 struct bio *b;
4601                 int d = r10_bio->devs[s/2].devnum;
4602                 struct md_rdev *rdev;
4603                 rcu_read_lock();
4604                 if (s&1) {
4605                         rdev = rcu_dereference(conf->mirrors[d].replacement);
4606                         b = r10_bio->devs[s/2].repl_bio;
4607                 } else {
4608                         rdev = rcu_dereference(conf->mirrors[d].rdev);
4609                         b = r10_bio->devs[s/2].bio;
4610                 }
4611                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4612                         rcu_read_unlock();
4613                         continue;
4614                 }
4615                 atomic_inc(&rdev->nr_pending);
4616                 rcu_read_unlock();
4617                 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4618                 atomic_inc(&r10_bio->remaining);
4619                 b->bi_next = NULL;
4620                 generic_make_request(b);
4621         }
4622         end_reshape_request(r10_bio);
4623 }
4624
4625 static void end_reshape(struct r10conf *conf)
4626 {
4627         if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4628                 return;
4629
4630         spin_lock_irq(&conf->device_lock);
4631         conf->prev = conf->geo;
4632         md_finish_reshape(conf->mddev);
4633         smp_wmb();
4634         conf->reshape_progress = MaxSector;
4635         conf->reshape_safe = MaxSector;
4636         spin_unlock_irq(&conf->device_lock);
4637
4638         /* read-ahead size must cover two whole stripes, which is
4639          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4640          */
4641         if (conf->mddev->queue) {
4642                 int stripe = conf->geo.raid_disks *
4643                         ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4644                 stripe /= conf->geo.near_copies;
4645                 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4646                         conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4647         }
4648         conf->fullsync = 0;
4649 }
4650
4651 static int handle_reshape_read_error(struct mddev *mddev,
4652                                      struct r10bio *r10_bio)
4653 {
4654         /* Use sync reads to get the blocks from somewhere else */
4655         int sectors = r10_bio->sectors;
4656         struct r10conf *conf = mddev->private;
4657         struct {
4658                 struct r10bio r10_bio;
4659                 struct r10dev devs[conf->copies];
4660         } on_stack;
4661         struct r10bio *r10b = &on_stack.r10_bio;
4662         int slot = 0;
4663         int idx = 0;
4664         struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4665
4666         r10b->sector = r10_bio->sector;
4667         __raid10_find_phys(&conf->prev, r10b);
4668
4669         while (sectors) {
4670                 int s = sectors;
4671                 int success = 0;
4672                 int first_slot = slot;
4673
4674                 if (s > (PAGE_SIZE >> 9))
4675                         s = PAGE_SIZE >> 9;
4676
4677                 rcu_read_lock();
4678                 while (!success) {
4679                         int d = r10b->devs[slot].devnum;
4680                         struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4681                         sector_t addr;
4682                         if (rdev == NULL ||
4683                             test_bit(Faulty, &rdev->flags) ||
4684                             !test_bit(In_sync, &rdev->flags))
4685                                 goto failed;
4686
4687                         addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4688                         atomic_inc(&rdev->nr_pending);
4689                         rcu_read_unlock();
4690                         success = sync_page_io(rdev,
4691                                                addr,
4692                                                s << 9,
4693                                                bvec[idx].bv_page,
4694                                                REQ_OP_READ, 0, false);
4695                         rdev_dec_pending(rdev, mddev);
4696                         rcu_read_lock();
4697                         if (success)
4698                                 break;
4699                 failed:
4700                         slot++;
4701                         if (slot >= conf->copies)
4702                                 slot = 0;
4703                         if (slot == first_slot)
4704                                 break;
4705                 }
4706                 rcu_read_unlock();
4707                 if (!success) {
4708                         /* couldn't read this block, must give up */
4709                         set_bit(MD_RECOVERY_INTR,
4710                                 &mddev->recovery);
4711                         return -EIO;
4712                 }
4713                 sectors -= s;
4714                 idx++;
4715         }
4716         return 0;
4717 }
4718
4719 static void end_reshape_write(struct bio *bio)
4720 {
4721         struct r10bio *r10_bio = bio->bi_private;
4722         struct mddev *mddev = r10_bio->mddev;
4723         struct r10conf *conf = mddev->private;
4724         int d;
4725         int slot;
4726         int repl;
4727         struct md_rdev *rdev = NULL;
4728
4729         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4730         if (repl)
4731                 rdev = conf->mirrors[d].replacement;
4732         if (!rdev) {
4733                 smp_mb();
4734                 rdev = conf->mirrors[d].rdev;
4735         }
4736
4737         if (bio->bi_error) {
4738                 /* FIXME should record badblock */
4739                 md_error(mddev, rdev);
4740         }
4741
4742         rdev_dec_pending(rdev, mddev);
4743         end_reshape_request(r10_bio);
4744 }
4745
4746 static void end_reshape_request(struct r10bio *r10_bio)
4747 {
4748         if (!atomic_dec_and_test(&r10_bio->remaining))
4749                 return;
4750         md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4751         bio_put(r10_bio->master_bio);
4752         put_buf(r10_bio);
4753 }
4754
4755 static void raid10_finish_reshape(struct mddev *mddev)
4756 {
4757         struct r10conf *conf = mddev->private;
4758
4759         if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4760                 return;
4761
4762         if (mddev->delta_disks > 0) {
4763                 sector_t size = raid10_size(mddev, 0, 0);
4764                 md_set_array_sectors(mddev, size);
4765                 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4766                         mddev->recovery_cp = mddev->resync_max_sectors;
4767                         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4768                 }
4769                 mddev->resync_max_sectors = size;
4770                 if (mddev->queue) {
4771                         set_capacity(mddev->gendisk, mddev->array_sectors);
4772                         revalidate_disk(mddev->gendisk);
4773                 }
4774         } else {
4775                 int d;
4776                 rcu_read_lock();
4777                 for (d = conf->geo.raid_disks ;
4778                      d < conf->geo.raid_disks - mddev->delta_disks;
4779                      d++) {
4780                         struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4781                         if (rdev)
4782                                 clear_bit(In_sync, &rdev->flags);
4783                         rdev = rcu_dereference(conf->mirrors[d].replacement);
4784                         if (rdev)
4785                                 clear_bit(In_sync, &rdev->flags);
4786                 }
4787                 rcu_read_unlock();
4788         }
4789         mddev->layout = mddev->new_layout;
4790         mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4791         mddev->reshape_position = MaxSector;
4792         mddev->delta_disks = 0;
4793         mddev->reshape_backwards = 0;
4794 }
4795
4796 static struct md_personality raid10_personality =
4797 {
4798         .name           = "raid10",
4799         .level          = 10,
4800         .owner          = THIS_MODULE,
4801         .make_request   = raid10_make_request,
4802         .run            = raid10_run,
4803         .free           = raid10_free,
4804         .status         = raid10_status,
4805         .error_handler  = raid10_error,
4806         .hot_add_disk   = raid10_add_disk,
4807         .hot_remove_disk= raid10_remove_disk,
4808         .spare_active   = raid10_spare_active,
4809         .sync_request   = raid10_sync_request,
4810         .quiesce        = raid10_quiesce,
4811         .size           = raid10_size,
4812         .resize         = raid10_resize,
4813         .takeover       = raid10_takeover,
4814         .check_reshape  = raid10_check_reshape,
4815         .start_reshape  = raid10_start_reshape,
4816         .finish_reshape = raid10_finish_reshape,
4817         .congested      = raid10_congested,
4818 };
4819
4820 static int __init raid_init(void)
4821 {
4822         return register_md_personality(&raid10_personality);
4823 }
4824
4825 static void raid_exit(void)
4826 {
4827         unregister_md_personality(&raid10_personality);
4828 }
4829
4830 module_init(raid_init);
4831 module_exit(raid_exit);
4832 MODULE_LICENSE("GPL");
4833 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4834 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4835 MODULE_ALIAS("md-raid10");
4836 MODULE_ALIAS("md-level-10");
4837
4838 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);