kernel/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 "md.h"
  29 #include "raid10.h"
  30 #include "raid0.h"
  31 #include "bitmap.h"
  32
  33 /*
  34  * RAID10 provides a combination of RAID0 and RAID1 functionality.
  35  * The layout of data is defined by
  36  *    chunk_size
  37  *    raid_disks
  38  *    near_copies (stored in low byte of layout)
  39  *    far_copies (stored in second byte of layout)
  40  *    far_offset (stored in bit 16 of layout )
  41  *    use_far_sets (stored in bit 17 of layout )
  42  *
  43  * The data to be stored is divided into chunks using chunksize.  Each device
  44  * is divided into far_copies sections.   In each section, chunks are laid out
  45  * in a style similar to raid0, but near_copies copies of each chunk is stored
  46  * (each on a different drive).  The starting device for each section is offset
  47  * near_copies from the starting device of the previous section.  Thus there
  48  * are (near_copies * far_copies) of each chunk, and each is on a different
  49  * drive.  near_copies and far_copies must be at least one, and their product
  50  * is at most raid_disks.
  51  *
  52  * If far_offset is true, then the far_copies are handled a bit differently.
  53  * The copies are still in different stripes, but instead of being very far
  54  * apart on disk, there are adjacent stripes.
  55  *
  56  * The far and offset algorithms are handled slightly differently if
  57  * 'use_far_sets' is true.  In this case, the array's devices are grouped into
  58  * sets that are (near_copies * far_copies) in size.  The far copied stripes
  59  * are still shifted by 'near_copies' devices, but this shifting stays confined
  60  * to the set rather than the entire array.  This is done to improve the number
  61  * of device combinations that can fail without causing the array to fail.
  62  * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
  63  * on a device):
  64  *    A B C D    A B C D E
  65  *      ...         ...
  66  *    D A B C    E A B C D
  67  * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
  68  *    [A B] [C D]    [A B] [C D E]
  69  *    |...| |...|    |...| | ... |
  70  *    [B A] [D C]    [B A] [E C D]
  71  */
  72
  73 /*
  74  * Number of guaranteed r10bios in case of extreme VM load:
  75  */
  76 #define NR_RAID10_BIOS 256
  77
  78 /* when we get a read error on a read-only array, we redirect to another
  79  * device without failing the first device, or trying to over-write to
  80  * correct the read error.  To keep track of bad blocks on a per-bio
  81  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
  82  */
  83 #define IO_BLOCKED ((struct bio *)1)
  84 /* When we successfully write to a known bad-block, we need to remove the
  85  * bad-block marking which must be done from process context.  So we record
  86  * the success by setting devs[n].bio to IO_MADE_GOOD
  87  */
  88 #define IO_MADE_GOOD ((struct bio *)2)
  89
  90 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
  91
  92 /* When there are this many requests queued to be written by
  93  * the raid10 thread, we become 'congested' to provide back-pressure
  94  * for writeback.
  95  */
  96 static int max_queued_requests = 1024;
  97
  98 static void allow_barrier(struct r10conf *conf);
  99 static void lower_barrier(struct r10conf *conf);
 100 static int _enough(struct r10conf *conf, int previous, int ignore);
 101 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
 102                                 int *skipped);
 103 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
 104 static void end_reshape_write(struct bio *bio, int error);
 105 static void end_reshape(struct r10conf *conf);
 106
 107 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
 108 {
 109         struct r10conf *conf = data;
 110         int size = offsetof(struct r10bio, devs[conf->copies]);
 111
 112         /* allocate a r10bio with room for raid_disks entries in the
 113          * bios array */
 114         return kzalloc(size, gfp_flags);
 115 }
 116
 117 static void r10bio_pool_free(void *r10_bio, void *data)
 118 {
 119         kfree(r10_bio);
 120 }
 121
 122 /* Maximum size of each resync request */
 123 #define RESYNC_BLOCK_SIZE (64*1024)
 124 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
 125 /* amount of memory to reserve for resync requests */
 126 #define RESYNC_WINDOW (1024*1024)
 127 /* maximum number of concurrent requests, memory permitting */
 128 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
 129
 130 /*
 131  * When performing a resync, we need to read and compare, so
 132  * we need as many pages are there are copies.
 133  * When performing a recovery, we need 2 bios, one for read,
 134  * one for write (we recover only one drive per r10buf)
 135  *
 136  */
 137 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
 138 {
 139         struct r10conf *conf = data;
 140         struct page *page;
 141         struct r10bio *r10_bio;
 142         struct bio *bio;
 143         int i, j;
 144         int nalloc;
 145
 146         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 147         if (!r10_bio)
 148                 return NULL;
 149
 150         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
 151             test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
 152                 nalloc = conf->copies; /* resync */
 153         else
 154                 nalloc = 2; /* recovery */
 155
 156         /*
 157          * Allocate bios.
 158          */
 159         for (j = nalloc ; j-- ; ) {
 160                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 161                 if (!bio)
 162                         goto out_free_bio;
 163                 r10_bio->devs[j].bio = bio;
 164                 if (!conf->have_replacement)
 165                         continue;
 166                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 167                 if (!bio)
 168                         goto out_free_bio;
 169                 r10_bio->devs[j].repl_bio = bio;
 170         }
 171         /*
 172          * Allocate RESYNC_PAGES data pages and attach them
 173          * where needed.
 174          */
 175         for (j = 0 ; j < nalloc; j++) {
 176                 struct bio *rbio = r10_bio->devs[j].repl_bio;
 177                 bio = r10_bio->devs[j].bio;
 178                 for (i = 0; i < RESYNC_PAGES; i++) {
 179                         if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
 180                                                &conf->mddev->recovery)) {
 181                                 /* we can share bv_page's during recovery
 182                                  * and reshape */
 183                                 struct bio *rbio = r10_bio->devs[0].bio;
 184                                 page = rbio->bi_io_vec[i].bv_page;
 185                                 get_page(page);
 186                         } else
 187                                 page = alloc_page(gfp_flags);
 188                         if (unlikely(!page))
 189                                 goto out_free_pages;
 190
 191                         bio->bi_io_vec[i].bv_page = page;
 192                         if (rbio)
 193                                 rbio->bi_io_vec[i].bv_page = page;
 194                 }
 195         }
 196
 197         return r10_bio;
 198
 199 out_free_pages:
 200         for ( ; i > 0 ; i--)
 201                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
 202         while (j--)
 203                 for (i = 0; i < RESYNC_PAGES ; i++)
 204                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 205         j = 0;
 206 out_free_bio:
 207         for ( ; j < nalloc; j++) {
 208                 if (r10_bio->devs[j].bio)
 209                         bio_put(r10_bio->devs[j].bio);
 210                 if (r10_bio->devs[j].repl_bio)
 211                         bio_put(r10_bio->devs[j].repl_bio);
 212         }
 213         r10bio_pool_free(r10_bio, conf);
 214         return NULL;
 215 }
 216
 217 static void r10buf_pool_free(void *__r10_bio, void *data)
 218 {
 219         int i;
 220         struct r10conf *conf = data;
 221         struct r10bio *r10bio = __r10_bio;
 222         int j;
 223
 224         for (j=0; j < conf->copies; j++) {
 225                 struct bio *bio = r10bio->devs[j].bio;
 226                 if (bio) {
 227                         for (i = 0; i < RESYNC_PAGES; i++) {
 228                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 229                                 bio->bi_io_vec[i].bv_page = NULL;
 230                         }
 231                         bio_put(bio);
 232                 }
 233                 bio = r10bio->devs[j].repl_bio;
 234                 if (bio)
 235                         bio_put(bio);
 236         }
 237         r10bio_pool_free(r10bio, conf);
 238 }
 239
 240 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
 241 {
 242         int i;
 243
 244         for (i = 0; i < conf->copies; i++) {
 245                 struct bio **bio = & r10_bio->devs[i].bio;
 246                 if (!BIO_SPECIAL(*bio))
 247                         bio_put(*bio);
 248                 *bio = NULL;
 249                 bio = &r10_bio->devs[i].repl_bio;
 250                 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
 251                         bio_put(*bio);
 252                 *bio = NULL;
 253         }
 254 }
 255
 256 static void free_r10bio(struct r10bio *r10_bio)
 257 {
 258         struct r10conf *conf = r10_bio->mddev->private;
 259
 260         put_all_bios(conf, r10_bio);
 261         mempool_free(r10_bio, conf->r10bio_pool);
 262 }
 263
 264 static void put_buf(struct r10bio *r10_bio)
 265 {
 266         struct r10conf *conf = r10_bio->mddev->private;
 267
 268         mempool_free(r10_bio, conf->r10buf_pool);
 269
 270         lower_barrier(conf);
 271 }
 272
 273 static void reschedule_retry(struct r10bio *r10_bio)
 274 {
 275         unsigned long flags;
 276         struct mddev *mddev = r10_bio->mddev;
 277         struct r10conf *conf = mddev->private;
 278
 279         spin_lock_irqsave(&conf->device_lock, flags);
 280         list_add(&r10_bio->retry_list, &conf->retry_list);
 281         conf->nr_queued ++;
 282         spin_unlock_irqrestore(&conf->device_lock, flags);
 283
 284         /* wake up frozen array... */
 285         wake_up(&conf->wait_barrier);
 286
 287         md_wakeup_thread(mddev->thread);
 288 }
 289
 290 /*
 291  * raid_end_bio_io() is called when we have finished servicing a mirrored
 292  * operation and are ready to return a success/failure code to the buffer
 293  * cache layer.
 294  */
 295 static void raid_end_bio_io(struct r10bio *r10_bio)
 296 {
 297         struct bio *bio = r10_bio->master_bio;
 298         int done;
 299         struct r10conf *conf = r10_bio->mddev->private;
 300
 301         if (bio->bi_phys_segments) {
 302                 unsigned long flags;
 303                 spin_lock_irqsave(&conf->device_lock, flags);
 304                 bio->bi_phys_segments--;
 305                 done = (bio->bi_phys_segments == 0);
 306                 spin_unlock_irqrestore(&conf->device_lock, flags);
 307         } else
 308                 done = 1;
 309         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
 310                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 311         if (done) {
 312                 bio_endio(bio, 0);
 313                 /*
 314                  * Wake up any possible resync thread that waits for the device
 315                  * to go idle.
 316                  */
 317                 allow_barrier(conf);
 318         }
 319         free_r10bio(r10_bio);
 320 }
 321
 322 /*
 323  * Update disk head position estimator based on IRQ completion info.
 324  */
 325 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
 326 {
 327         struct r10conf *conf = r10_bio->mddev->private;
 328
 329         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 330                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 331 }
 332
 333 /*
 334  * Find the disk number which triggered given bio
 335  */
 336 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
 337                          struct bio *bio, int *slotp, int *replp)
 338 {
 339         int slot;
 340         int repl = 0;
 341
 342         for (slot = 0; slot < conf->copies; slot++) {
 343                 if (r10_bio->devs[slot].bio == bio)
 344                         break;
 345                 if (r10_bio->devs[slot].repl_bio == bio) {
 346                         repl = 1;
 347                         break;
 348                 }
 349         }
 350
 351         BUG_ON(slot == conf->copies);
 352         update_head_pos(slot, r10_bio);
 353
 354         if (slotp)
 355                 *slotp = slot;
 356         if (replp)
 357                 *replp = repl;
 358         return r10_bio->devs[slot].devnum;
 359 }
 360
 361 static void raid10_end_read_request(struct bio *bio, int error)
 362 {
 363         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 364         struct r10bio *r10_bio = bio->bi_private;
 365         int slot, dev;
 366         struct md_rdev *rdev;
 367         struct r10conf *conf = r10_bio->mddev->private;
 368
 369         slot = r10_bio->read_slot;
 370         dev = r10_bio->devs[slot].devnum;
 371         rdev = r10_bio->devs[slot].rdev;
 372         /*
 373          * this branch is our 'one mirror IO has finished' event handler:
 374          */
 375         update_head_pos(slot, r10_bio);
 376
 377         if (uptodate) {
 378                 /*
 379                  * Set R10BIO_Uptodate in our master bio, so that
 380                  * we will return a good error code to the higher
 381                  * levels even if IO on some other mirrored buffer fails.
 382                  *
 383                  * The 'master' represents the composite IO operation to
 384                  * user-side. So if something waits for IO, then it will
 385                  * wait for the 'master' bio.
 386                  */
 387                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 388         } else {
 389                 /* If all other devices that store this block have
 390                  * failed, we want to return the error upwards rather
 391                  * than fail the last device.  Here we redefine
 392                  * "uptodate" to mean "Don't want to retry"
 393                  */
 394                 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
 395                              rdev->raid_disk))
 396                         uptodate = 1;
 397         }
 398         if (uptodate) {
 399                 raid_end_bio_io(r10_bio);
 400                 rdev_dec_pending(rdev, conf->mddev);
 401         } else {
 402                 /*
 403                  * oops, read error - keep the refcount on the rdev
 404                  */
 405                 char b[BDEVNAME_SIZE];
 406                 printk_ratelimited(KERN_ERR
 407                                    "md/raid10:%s: %s: rescheduling sector %llu\n",
 408                                    mdname(conf->mddev),
 409                                    bdevname(rdev->bdev, b),
 410                                    (unsigned long long)r10_bio->sector);
 411                 set_bit(R10BIO_ReadError, &r10_bio->state);
 412                 reschedule_retry(r10_bio);
 413         }
 414 }
 415
 416 static void close_write(struct r10bio *r10_bio)
 417 {
 418         /* clear the bitmap if all writes complete successfully */
 419         bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 420                         r10_bio->sectors,
 421                         !test_bit(R10BIO_Degraded, &r10_bio->state),
 422                         0);
 423         md_write_end(r10_bio->mddev);
 424 }
 425
 426 static void one_write_done(struct r10bio *r10_bio)
 427 {
 428         if (atomic_dec_and_test(&r10_bio->remaining)) {
 429                 if (test_bit(R10BIO_WriteError, &r10_bio->state))
 430                         reschedule_retry(r10_bio);
 431                 else {
 432                         close_write(r10_bio);
 433                         if (test_bit(R10BIO_MadeGood, &r10_bio->state))
 434                                 reschedule_retry(r10_bio);
 435                         else
 436                                 raid_end_bio_io(r10_bio);
 437                 }
 438         }
 439 }
 440
 441 static void raid10_end_write_request(struct bio *bio, int error)
 442 {
 443         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 444         struct r10bio *r10_bio = bio->bi_private;
 445         int dev;
 446         int dec_rdev = 1;
 447         struct r10conf *conf = r10_bio->mddev->private;
 448         int slot, repl;
 449         struct md_rdev *rdev = NULL;
 450
 451         dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
 452
 453         if (repl)
 454                 rdev = conf->mirrors[dev].replacement;
 455         if (!rdev) {
 456                 smp_rmb();
 457                 repl = 0;
 458                 rdev = conf->mirrors[dev].rdev;
 459         }
 460         /*
 461          * this branch is our 'one mirror IO has finished' event handler:
 462          */
 463         if (!uptodate) {
 464                 if (repl)
 465                         /* Never record new bad blocks to replacement,
 466                          * just fail it.
 467                          */
 468                         md_error(rdev->mddev, rdev);
 469                 else {
 470                         set_bit(WriteErrorSeen, &rdev->flags);
 471                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
 472                                 set_bit(MD_RECOVERY_NEEDED,
 473                                         &rdev->mddev->recovery);
 474                         set_bit(R10BIO_WriteError, &r10_bio->state);
 475                         dec_rdev = 0;
 476                 }
 477         } else {
 478                 /*
 479                  * Set R10BIO_Uptodate in our master bio, so that
 480                  * we will return a good error code for to the higher
 481                  * levels even if IO on some other mirrored buffer fails.
 482                  *
 483                  * The 'master' represents the composite IO operation to
 484                  * user-side. So if something waits for IO, then it will
 485                  * wait for the 'master' bio.
 486                  */
 487                 sector_t first_bad;
 488                 int bad_sectors;
 489
 490                 /*
 491                  * Do not set R10BIO_Uptodate if the current device is
 492                  * rebuilding or Faulty. This is because we cannot use
 493                  * such device for properly reading the data back (we could
 494                  * potentially use it, if the current write would have felt
 495                  * before rdev->recovery_offset, but for simplicity we don't
 496                  * check this here.
 497                  */
 498                 if (test_bit(In_sync, &rdev->flags) &&
 499                     !test_bit(Faulty, &rdev->flags))
 500                         set_bit(R10BIO_Uptodate, &r10_bio->state);
 501
 502                 /* Maybe we can clear some bad blocks. */
 503                 if (is_badblock(rdev,
 504                                 r10_bio->devs[slot].addr,
 505                                 r10_bio->sectors,
 506                                 &first_bad, &bad_sectors)) {
 507                         bio_put(bio);
 508                         if (repl)
 509                                 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
 510                         else
 511                                 r10_bio->devs[slot].bio = IO_MADE_GOOD;
 512                         dec_rdev = 0;
 513                         set_bit(R10BIO_MadeGood, &r10_bio->state);
 514                 }
 515         }
 516
 517         /*
 518          *
 519          * Let's see if all mirrored write operations have finished
 520          * already.
 521          */
 522         one_write_done(r10_bio);
 523         if (dec_rdev)
 524                 rdev_dec_pending(rdev, conf->mddev);
 525 }
 526
 527 /*
 528  * RAID10 layout manager
 529  * As well as the chunksize and raid_disks count, there are two
 530  * parameters: near_copies and far_copies.
 531  * near_copies * far_copies must be <= raid_disks.
 532  * Normally one of these will be 1.
 533  * If both are 1, we get raid0.
 534  * If near_copies == raid_disks, we get raid1.
 535  *
 536  * Chunks are laid out in raid0 style with near_copies copies of the
 537  * first chunk, followed by near_copies copies of the next chunk and
 538  * so on.
 539  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 540  * as described above, we start again with a device offset of near_copies.
 541  * So we effectively have another copy of the whole array further down all
 542  * the drives, but with blocks on different drives.
 543  * With this layout, and block is never stored twice on the one device.
 544  *
 545  * raid10_find_phys finds the sector offset of a given virtual sector
 546  * on each device that it is on.
 547  *
 548  * raid10_find_virt does the reverse mapping, from a device and a
 549  * sector offset to a virtual address
 550  */
 551
 552 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
 553 {
 554         int n,f;
 555         sector_t sector;
 556         sector_t chunk;
 557         sector_t stripe;
 558         int dev;
 559         int slot = 0;
 560         int last_far_set_start, last_far_set_size;
 561
 562         last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
 563         last_far_set_start *= geo->far_set_size;
 564
 565         last_far_set_size = geo->far_set_size;
 566         last_far_set_size += (geo->raid_disks % geo->far_set_size);
 567
 568         /* now calculate first sector/dev */
 569         chunk = r10bio->sector >> geo->chunk_shift;
 570         sector = r10bio->sector & geo->chunk_mask;
 571
 572         chunk *= geo->near_copies;
 573         stripe = chunk;
 574         dev = sector_div(stripe, geo->raid_disks);
 575         if (geo->far_offset)
 576                 stripe *= geo->far_copies;
 577
 578         sector += stripe << geo->chunk_shift;
 579
 580         /* and calculate all the others */
 581         for (n = 0; n < geo->near_copies; n++) {
 582                 int d = dev;
 583                 int set;
 584                 sector_t s = sector;
 585                 r10bio->devs[slot].devnum = d;
 586                 r10bio->devs[slot].addr = s;
 587                 slot++;
 588
 589                 for (f = 1; f < geo->far_copies; f++) {
 590                         set = d / geo->far_set_size;
 591                         d += geo->near_copies;
 592
 593                         if ((geo->raid_disks % geo->far_set_size) &&
 594                             (d > last_far_set_start)) {
 595                                 d -= last_far_set_start;
 596                                 d %= last_far_set_size;
 597                                 d += last_far_set_start;
 598                         } else {
 599                                 d %= geo->far_set_size;
 600                                 d += geo->far_set_size * set;
 601                         }
 602                         s += geo->stride;
 603                         r10bio->devs[slot].devnum = d;
 604                         r10bio->devs[slot].addr = s;
 605                         slot++;
 606                 }
 607                 dev++;
 608                 if (dev >= geo->raid_disks) {
 609                         dev = 0;
 610                         sector += (geo->chunk_mask + 1);
 611                 }
 612         }
 613 }
 614
 615 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
 616 {
 617         struct geom *geo = &conf->geo;
 618
 619         if (conf->reshape_progress != MaxSector &&
 620             ((r10bio->sector >= conf->reshape_progress) !=
 621              conf->mddev->reshape_backwards)) {
 622                 set_bit(R10BIO_Previous, &r10bio->state);
 623                 geo = &conf->prev;
 624         } else
 625                 clear_bit(R10BIO_Previous, &r10bio->state);
 626
 627         __raid10_find_phys(geo, r10bio);
 628 }
 629
 630 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
 631 {
 632         sector_t offset, chunk, vchunk;
 633         /* Never use conf->prev as this is only called during resync
 634          * or recovery, so reshape isn't happening
 635          */
 636         struct geom *geo = &conf->geo;
 637         int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
 638         int far_set_size = geo->far_set_size;
 639         int last_far_set_start;
 640
 641         if (geo->raid_disks % geo->far_set_size) {
 642                 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
 643                 last_far_set_start *= geo->far_set_size;
 644
 645                 if (dev >= last_far_set_start) {
 646                         far_set_size = geo->far_set_size;
 647                         far_set_size += (geo->raid_disks % geo->far_set_size);
 648                         far_set_start = last_far_set_start;
 649                 }
 650         }
 651
 652         offset = sector & geo->chunk_mask;
 653         if (geo->far_offset) {
 654                 int fc;
 655                 chunk = sector >> geo->chunk_shift;
 656                 fc = sector_div(chunk, geo->far_copies);
 657                 dev -= fc * geo->near_copies;
 658                 if (dev < far_set_start)
 659                         dev += far_set_size;
 660         } else {
 661                 while (sector >= geo->stride) {
 662                         sector -= geo->stride;
 663                         if (dev < (geo->near_copies + far_set_start))
 664                                 dev += far_set_size - geo->near_copies;
 665                         else
 666                                 dev -= geo->near_copies;
 667                 }
 668                 chunk = sector >> geo->chunk_shift;
 669         }
 670         vchunk = chunk * geo->raid_disks + dev;
 671         sector_div(vchunk, geo->near_copies);
 672         return (vchunk << geo->chunk_shift) + offset;
 673 }
 674
 675 /**
 676  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 677  *      @mddev: the md device
 678  *      @bvm: properties of new bio
 679  *      @biovec: the request that could be merged to it.
 680  *
 681  *      Return amount of bytes we can accept at this offset
 682  *      This requires checking for end-of-chunk if near_copies != raid_disks,
 683  *      and for subordinate merge_bvec_fns if merge_check_needed.
 684  */
 685 static int raid10_mergeable_bvec(struct mddev *mddev,
 686                                  struct bvec_merge_data *bvm,
 687                                  struct bio_vec *biovec)
 688 {
 689         struct r10conf *conf = mddev->private;
 690         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 691         int max;
 692         unsigned int chunk_sectors;
 693         unsigned int bio_sectors = bvm->bi_size >> 9;
 694         struct geom *geo = &conf->geo;
 695
 696         chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
 697         if (conf->reshape_progress != MaxSector &&
 698             ((sector >= conf->reshape_progress) !=
 699              conf->mddev->reshape_backwards))
 700                 geo = &conf->prev;
 701
 702         if (geo->near_copies < geo->raid_disks) {
 703                 max = (chunk_sectors - ((sector & (chunk_sectors - 1))
 704                                         + bio_sectors)) << 9;
 705                 if (max < 0)
 706                         /* bio_add cannot handle a negative return */
 707                         max = 0;
 708                 if (max <= biovec->bv_len && bio_sectors == 0)
 709                         return biovec->bv_len;
 710         } else
 711                 max = biovec->bv_len;
 712
 713         if (mddev->merge_check_needed) {
 714                 struct {
 715                         struct r10bio r10_bio;
 716                         struct r10dev devs[conf->copies];
 717                 } on_stack;
 718                 struct r10bio *r10_bio = &on_stack.r10_bio;
 719                 int s;
 720                 if (conf->reshape_progress != MaxSector) {
 721                         /* Cannot give any guidance during reshape */
 722                         if (max <= biovec->bv_len && bio_sectors == 0)
 723                                 return biovec->bv_len;
 724                         return 0;
 725                 }
 726                 r10_bio->sector = sector;
 727                 raid10_find_phys(conf, r10_bio);
 728                 rcu_read_lock();
 729                 for (s = 0; s < conf->copies; s++) {
 730                         int disk = r10_bio->devs[s].devnum;
 731                         struct md_rdev *rdev = rcu_dereference(
 732                                 conf->mirrors[disk].rdev);
 733                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
 734                                 struct request_queue *q =
 735                                         bdev_get_queue(rdev->bdev);
 736                                 if (q->merge_bvec_fn) {
 737                                         bvm->bi_sector = r10_bio->devs[s].addr
 738                                                 + rdev->data_offset;
 739                                         bvm->bi_bdev = rdev->bdev;
 740                                         max = min(max, q->merge_bvec_fn(
 741                                                           q, bvm, biovec));
 742                                 }
 743                         }
 744                         rdev = rcu_dereference(conf->mirrors[disk].replacement);
 745                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
 746                                 struct request_queue *q =
 747                                         bdev_get_queue(rdev->bdev);
 748                                 if (q->merge_bvec_fn) {
 749                                         bvm->bi_sector = r10_bio->devs[s].addr
 750                                                 + rdev->data_offset;
 751                                         bvm->bi_bdev = rdev->bdev;
 752                                         max = min(max, q->merge_bvec_fn(
 753                                                           q, bvm, biovec));
 754                                 }
 755                         }
 756                 }
 757                 rcu_read_unlock();
 758         }
 759         return max;
 760 }
 761
 762 /*
 763  * This routine returns the disk from which the requested read should
 764  * be done. There is a per-array 'next expected sequential IO' sector
 765  * number - if this matches on the next IO then we use the last disk.
 766  * There is also a per-disk 'last know head position' sector that is
 767  * maintained from IRQ contexts, both the normal and the resync IO
 768  * completion handlers update this position correctly. If there is no
 769  * perfect sequential match then we pick the disk whose head is closest.
 770  *
 771  * If there are 2 mirrors in the same 2 devices, performance degrades
 772  * because position is mirror, not device based.
 773  *
 774  * The rdev for the device selected will have nr_pending incremented.
 775  */
 776
 777 /*
 778  * FIXME: possibly should rethink readbalancing and do it differently
 779  * depending on near_copies / far_copies geometry.
 780  */
 781 static struct md_rdev *read_balance(struct r10conf *conf,
 782                                     struct r10bio *r10_bio,
 783                                     int *max_sectors)
 784 {
 785         const sector_t this_sector = r10_bio->sector;
 786         int disk, slot;
 787         int sectors = r10_bio->sectors;
 788         int best_good_sectors;
 789         sector_t new_distance, best_dist;
 790         struct md_rdev *best_rdev, *rdev = NULL;
 791         int do_balance;
 792         int best_slot;
 793         struct geom *geo = &conf->geo;
 794
 795         raid10_find_phys(conf, r10_bio);
 796         rcu_read_lock();
 797 retry:
 798         sectors = r10_bio->sectors;
 799         best_slot = -1;
 800         best_rdev = NULL;
 801         best_dist = MaxSector;
 802         best_good_sectors = 0;
 803         do_balance = 1;
 804         /*
 805          * Check if we can balance. We can balance on the whole
 806          * device if no resync is going on (recovery is ok), or below
 807          * the resync window. We take the first readable disk when
 808          * above the resync window.
 809          */
 810         if (conf->mddev->recovery_cp < MaxSector
 811             && (this_sector + sectors >= conf->next_resync))
 812                 do_balance = 0;
 813
 814         for (slot = 0; slot < conf->copies ; slot++) {
 815                 sector_t first_bad;
 816                 int bad_sectors;
 817                 sector_t dev_sector;
 818
 819                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
 820                         continue;
 821                 disk = r10_bio->devs[slot].devnum;
 822                 rdev = rcu_dereference(conf->mirrors[disk].replacement);
 823                 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
 824                     test_bit(Unmerged, &rdev->flags) ||
 825                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 826                         rdev = rcu_dereference(conf->mirrors[disk].rdev);
 827                 if (rdev == NULL ||
 828                     test_bit(Faulty, &rdev->flags) ||
 829                     test_bit(Unmerged, &rdev->flags))
 830                         continue;
 831                 if (!test_bit(In_sync, &rdev->flags) &&
 832                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 833                         continue;
 834
 835                 dev_sector = r10_bio->devs[slot].addr;
 836                 if (is_badblock(rdev, dev_sector, sectors,
 837                                 &first_bad, &bad_sectors)) {
 838                         if (best_dist < MaxSector)
 839                                 /* Already have a better slot */
 840                                 continue;
 841                         if (first_bad <= dev_sector) {
 842                                 /* Cannot read here.  If this is the
 843                                  * 'primary' device, then we must not read
 844                                  * beyond 'bad_sectors' from another device.
 845                                  */
 846                                 bad_sectors -= (dev_sector - first_bad);
 847                                 if (!do_balance && sectors > bad_sectors)
 848                                         sectors = bad_sectors;
 849                                 if (best_good_sectors > sectors)
 850                                         best_good_sectors = sectors;
 851                         } else {
 852                                 sector_t good_sectors =
 853                                         first_bad - dev_sector;
 854                                 if (good_sectors > best_good_sectors) {
 855                                         best_good_sectors = good_sectors;
 856                                         best_slot = slot;
 857                                         best_rdev = rdev;
 858                                 }
 859                                 if (!do_balance)
 860                                         /* Must read from here */
 861                                         break;
 862                         }
 863                         continue;
 864                 } else
 865                         best_good_sectors = sectors;
 866
 867                 if (!do_balance)
 868                         break;
 869
 870                 /* This optimisation is debatable, and completely destroys
 871                  * sequential read speed for 'far copies' arrays.  So only
 872                  * keep it for 'near' arrays, and review those later.
 873                  */
 874                 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
 875                         break;
 876
 877                 /* for far > 1 always use the lowest address */
 878                 if (geo->far_copies > 1)
 879                         new_distance = r10_bio->devs[slot].addr;
 880                 else
 881                         new_distance = abs(r10_bio->devs[slot].addr -
 882                                            conf->mirrors[disk].head_position);
 883                 if (new_distance < best_dist) {
 884                         best_dist = new_distance;
 885                         best_slot = slot;
 886                         best_rdev = rdev;
 887                 }
 888         }
 889         if (slot >= conf->copies) {
 890                 slot = best_slot;
 891                 rdev = best_rdev;
 892         }
 893
 894         if (slot >= 0) {
 895                 atomic_inc(&rdev->nr_pending);
 896                 if (test_bit(Faulty, &rdev->flags)) {
 897                         /* Cannot risk returning a device that failed
 898                          * before we inc'ed nr_pending
 899                          */
 900                         rdev_dec_pending(rdev, conf->mddev);
 901                         goto retry;
 902                 }
 903                 r10_bio->read_slot = slot;
 904         } else
 905                 rdev = NULL;
 906         rcu_read_unlock();
 907         *max_sectors = best_good_sectors;
 908
 909         return rdev;
 910 }
 911
 912 static int raid10_congested(struct mddev *mddev, int bits)
 913 {
 914         struct r10conf *conf = mddev->private;
 915         int i, ret = 0;
 916
 917         if ((bits & (1 << BDI_async_congested)) &&
 918             conf->pending_count >= max_queued_requests)
 919                 return 1;
 920
 921         rcu_read_lock();
 922         for (i = 0;
 923              (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
 924                      && ret == 0;
 925              i++) {
 926                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 927                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 928                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 929
 930                         ret |= bdi_congested(&q->backing_dev_info, bits);
 931                 }
 932         }
 933         rcu_read_unlock();
 934         return ret;
 935 }
 936
 937 static void flush_pending_writes(struct r10conf *conf)
 938 {
 939         /* Any writes that have been queued but are awaiting
 940          * bitmap updates get flushed here.
 941          */
 942         spin_lock_irq(&conf->device_lock);
 943
 944         if (conf->pending_bio_list.head) {
 945                 struct bio *bio;
 946                 bio = bio_list_get(&conf->pending_bio_list);
 947                 conf->pending_count = 0;
 948                 spin_unlock_irq(&conf->device_lock);
 949                 /* flush any pending bitmap writes to disk
 950                  * before proceeding w/ I/O */
 951                 bitmap_unplug(conf->mddev->bitmap);
 952                 wake_up(&conf->wait_barrier);
 953
 954                 while (bio) { /* submit pending writes */
 955                         struct bio *next = bio->bi_next;
 956                         bio->bi_next = NULL;
 957                         if (unlikely((bio->bi_rw & REQ_DISCARD) &&
 958                             !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 959                                 /* Just ignore it */
 960                                 bio_endio(bio, 0);
 961                         else
 962                                 generic_make_request(bio);
 963                         bio = next;
 964                 }
 965         } else
 966                 spin_unlock_irq(&conf->device_lock);
 967 }
 968
 969 /* Barriers....
 970  * Sometimes we need to suspend IO while we do something else,
 971  * either some resync/recovery, or reconfigure the array.
 972  * To do this we raise a 'barrier'.
 973  * The 'barrier' is a counter that can be raised multiple times
 974  * to count how many activities are happening which preclude
 975  * normal IO.
 976  * We can only raise the barrier if there is no pending IO.
 977  * i.e. if nr_pending == 0.
 978  * We choose only to raise the barrier if no-one is waiting for the
 979  * barrier to go down.  This means that as soon as an IO request
 980  * is ready, no other operations which require a barrier will start
 981  * until the IO request has had a chance.
 982  *
 983  * So: regular IO calls 'wait_barrier'.  When that returns there
 984  *    is no backgroup IO happening,  It must arrange to call
 985  *    allow_barrier when it has finished its IO.
 986  * backgroup IO calls must call raise_barrier.  Once that returns
 987  *    there is no normal IO happeing.  It must arrange to call
 988  *    lower_barrier when the particular background IO completes.
 989  */
 990
 991 static void raise_barrier(struct r10conf *conf, int force)
 992 {
 993         BUG_ON(force && !conf->barrier);
 994         spin_lock_irq(&conf->resync_lock);
 995
 996         /* Wait until no block IO is waiting (unless 'force') */
 997         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 998                             conf->resync_lock);
 999
1000         /* block any new IO from starting */
1001         conf->barrier++;
1002
1003         /* Now wait for all pending IO to complete */
1004         wait_event_lock_irq(conf->wait_barrier,
1005                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
1006                             conf->resync_lock);
1007
1008         spin_unlock_irq(&conf->resync_lock);
1009 }
1010
1011 static void lower_barrier(struct r10conf *conf)
1012 {
1013         unsigned long flags;
1014         spin_lock_irqsave(&conf->resync_lock, flags);
1015         conf->barrier--;
1016         spin_unlock_irqrestore(&conf->resync_lock, flags);
1017         wake_up(&conf->wait_barrier);
1018 }
1019
1020 static void wait_barrier(struct r10conf *conf)
1021 {
1022         spin_lock_irq(&conf->resync_lock);
1023         if (conf->barrier) {
1024                 conf->nr_waiting++;
1025                 /* Wait for the barrier to drop.
1026                  * However if there are already pending
1027                  * requests (preventing the barrier from
1028                  * rising completely), and the
1029                  * pre-process bio queue isn't empty,
1030                  * then don't wait, as we need to empty
1031                  * that queue to get the nr_pending
1032                  * count down.
1033                  */
1034                 wait_event_lock_irq(conf->wait_barrier,
1035                                     !conf->barrier ||
1036                                     (conf->nr_pending &&
1037                                      current->bio_list &&
1038                                      !bio_list_empty(current->bio_list)),
1039                                     conf->resync_lock);
1040                 conf->nr_waiting--;
1041         }
1042         conf->nr_pending++;
1043         spin_unlock_irq(&conf->resync_lock);
1044 }
1045
1046 static void allow_barrier(struct r10conf *conf)
1047 {
1048         unsigned long flags;
1049         spin_lock_irqsave(&conf->resync_lock, flags);
1050         conf->nr_pending--;
1051         spin_unlock_irqrestore(&conf->resync_lock, flags);
1052         wake_up(&conf->wait_barrier);
1053 }
1054
1055 static void freeze_array(struct r10conf *conf, int extra)
1056 {
1057         /* stop syncio and normal IO and wait for everything to
1058          * go quiet.
1059          * We increment barrier and nr_waiting, and then
1060          * wait until nr_pending match nr_queued+extra
1061          * This is called in the context of one normal IO request
1062          * that has failed. Thus any sync request that might be pending
1063          * will be blocked by nr_pending, and we need to wait for
1064          * pending IO requests to complete or be queued for re-try.
1065          * Thus the number queued (nr_queued) plus this request (extra)
1066          * must match the number of pending IOs (nr_pending) before
1067          * we continue.
1068          */
1069         spin_lock_irq(&conf->resync_lock);
1070         conf->barrier++;
1071         conf->nr_waiting++;
1072         wait_event_lock_irq_cmd(conf->wait_barrier,
1073                                 conf->nr_pending == conf->nr_queued+extra,
1074                                 conf->resync_lock,
1075                                 flush_pending_writes(conf));
1076
1077         spin_unlock_irq(&conf->resync_lock);
1078 }
1079
1080 static void unfreeze_array(struct r10conf *conf)
1081 {
1082         /* reverse the effect of the freeze */
1083         spin_lock_irq(&conf->resync_lock);
1084         conf->barrier--;
1085         conf->nr_waiting--;
1086         wake_up(&conf->wait_barrier);
1087         spin_unlock_irq(&conf->resync_lock);
1088 }
1089
1090 static sector_t choose_data_offset(struct r10bio *r10_bio,
1091                                    struct md_rdev *rdev)
1092 {
1093         if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1094             test_bit(R10BIO_Previous, &r10_bio->state))
1095                 return rdev->data_offset;
1096         else
1097                 return rdev->new_data_offset;
1098 }
1099
1100 struct raid10_plug_cb {
1101         struct blk_plug_cb      cb;
1102         struct bio_list         pending;
1103         int                     pending_cnt;
1104 };
1105
1106 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1107 {
1108         struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1109                                                    cb);
1110         struct mddev *mddev = plug->cb.data;
1111         struct r10conf *conf = mddev->private;
1112         struct bio *bio;
1113
1114         if (from_schedule || current->bio_list) {
1115                 spin_lock_irq(&conf->device_lock);
1116                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1117                 conf->pending_count += plug->pending_cnt;
1118                 spin_unlock_irq(&conf->device_lock);
1119                 wake_up(&conf->wait_barrier);
1120                 md_wakeup_thread(mddev->thread);
1121                 kfree(plug);
1122                 return;
1123         }
1124
1125         /* we aren't scheduling, so we can do the write-out directly. */
1126         bio = bio_list_get(&plug->pending);
1127         bitmap_unplug(mddev->bitmap);
1128         wake_up(&conf->wait_barrier);
1129
1130         while (bio) { /* submit pending writes */
1131                 struct bio *next = bio->bi_next;
1132                 bio->bi_next = NULL;
1133                 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1134                     !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1135                         /* Just ignore it */
1136                         bio_endio(bio, 0);
1137                 else
1138                         generic_make_request(bio);
1139                 bio = next;
1140         }
1141         kfree(plug);
1142 }
1143
1144 static void __make_request(struct mddev *mddev, struct bio *bio)
1145 {
1146         struct r10conf *conf = mddev->private;
1147         struct r10bio *r10_bio;
1148         struct bio *read_bio;
1149         int i;
1150         const int rw = bio_data_dir(bio);
1151         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1152         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1153         const unsigned long do_discard = (bio->bi_rw
1154                                           & (REQ_DISCARD | REQ_SECURE));
1155         const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1156         unsigned long flags;
1157         struct md_rdev *blocked_rdev;
1158         struct blk_plug_cb *cb;
1159         struct raid10_plug_cb *plug = NULL;
1160         int sectors_handled;
1161         int max_sectors;
1162         int sectors;
1163
1164         /*
1165          * Register the new request and wait if the reconstruction
1166          * thread has put up a bar for new requests.
1167          * Continue immediately if no resync is active currently.
1168          */
1169         wait_barrier(conf);
1170
1171         sectors = bio_sectors(bio);
1172         while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1173             bio->bi_iter.bi_sector < conf->reshape_progress &&
1174             bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1175                 /* IO spans the reshape position.  Need to wait for
1176                  * reshape to pass
1177                  */
1178                 allow_barrier(conf);
1179                 wait_event(conf->wait_barrier,
1180                            conf->reshape_progress <= bio->bi_iter.bi_sector ||
1181                            conf->reshape_progress >= bio->bi_iter.bi_sector +
1182                            sectors);
1183                 wait_barrier(conf);
1184         }
1185         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1186             bio_data_dir(bio) == WRITE &&
1187             (mddev->reshape_backwards
1188              ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1189                 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1190              : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1191                 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1192                 /* Need to update reshape_position in metadata */
1193                 mddev->reshape_position = conf->reshape_progress;
1194                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1195                 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1196                 md_wakeup_thread(mddev->thread);
1197                 wait_event(mddev->sb_wait,
1198                            !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1199
1200                 conf->reshape_safe = mddev->reshape_position;
1201         }
1202
1203         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1204
1205         r10_bio->master_bio = bio;
1206         r10_bio->sectors = sectors;
1207
1208         r10_bio->mddev = mddev;
1209         r10_bio->sector = bio->bi_iter.bi_sector;
1210         r10_bio->state = 0;
1211
1212         /* We might need to issue multiple reads to different
1213          * devices if there are bad blocks around, so we keep
1214          * track of the number of reads in bio->bi_phys_segments.
1215          * If this is 0, there is only one r10_bio and no locking
1216          * will be needed when the request completes.  If it is
1217          * non-zero, then it is the number of not-completed requests.
1218          */
1219         bio->bi_phys_segments = 0;
1220         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1221
1222         if (rw == READ) {
1223                 /*
1224                  * read balancing logic:
1225                  */
1226                 struct md_rdev *rdev;
1227                 int slot;
1228
1229 read_again:
1230                 rdev = read_balance(conf, r10_bio, &max_sectors);
1231                 if (!rdev) {
1232                         raid_end_bio_io(r10_bio);
1233                         return;
1234                 }
1235                 slot = r10_bio->read_slot;
1236
1237                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1238                 bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1239                          max_sectors);
1240
1241                 r10_bio->devs[slot].bio = read_bio;
1242                 r10_bio->devs[slot].rdev = rdev;
1243
1244                 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1245                         choose_data_offset(r10_bio, rdev);
1246                 read_bio->bi_bdev = rdev->bdev;
1247                 read_bio->bi_end_io = raid10_end_read_request;
1248                 read_bio->bi_rw = READ | do_sync;
1249                 read_bio->bi_private = r10_bio;
1250
1251                 if (max_sectors < r10_bio->sectors) {
1252                         /* Could not read all from this device, so we will
1253                          * need another r10_bio.
1254                          */
1255                         sectors_handled = (r10_bio->sector + max_sectors
1256                                            - bio->bi_iter.bi_sector);
1257                         r10_bio->sectors = max_sectors;
1258                         spin_lock_irq(&conf->device_lock);
1259                         if (bio->bi_phys_segments == 0)
1260                                 bio->bi_phys_segments = 2;
1261                         else
1262                                 bio->bi_phys_segments++;
1263                         spin_unlock_irq(&conf->device_lock);
1264                         /* Cannot call generic_make_request directly
1265                          * as that will be queued in __generic_make_request
1266                          * and subsequent mempool_alloc might block
1267                          * waiting for it.  so hand bio over to raid10d.
1268                          */
1269                         reschedule_retry(r10_bio);
1270
1271                         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1272
1273                         r10_bio->master_bio = bio;
1274                         r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1275                         r10_bio->state = 0;
1276                         r10_bio->mddev = mddev;
1277                         r10_bio->sector = bio->bi_iter.bi_sector +
1278                                 sectors_handled;
1279                         goto read_again;
1280                 } else
1281                         generic_make_request(read_bio);
1282                 return;
1283         }
1284
1285         /*
1286          * WRITE:
1287          */
1288         if (conf->pending_count >= max_queued_requests) {
1289                 md_wakeup_thread(mddev->thread);
1290                 wait_event(conf->wait_barrier,
1291                            conf->pending_count < max_queued_requests);
1292         }
1293         /* first select target devices under rcu_lock and
1294          * inc refcount on their rdev.  Record them by setting
1295          * bios[x] to bio
1296          * If there are known/acknowledged bad blocks on any device
1297          * on which we have seen a write error, we want to avoid
1298          * writing to those blocks.  This potentially requires several
1299          * writes to write around the bad blocks.  Each set of writes
1300          * gets its own r10_bio with a set of bios attached.  The number
1301          * of r10_bios is recored in bio->bi_phys_segments just as with
1302          * the read case.
1303          */
1304
1305         r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1306         raid10_find_phys(conf, r10_bio);
1307 retry_write:
1308         blocked_rdev = NULL;
1309         rcu_read_lock();
1310         max_sectors = r10_bio->sectors;
1311
1312         for (i = 0;  i < conf->copies; i++) {
1313                 int d = r10_bio->devs[i].devnum;
1314                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1315                 struct md_rdev *rrdev = rcu_dereference(
1316                         conf->mirrors[d].replacement);
1317                 if (rdev == rrdev)
1318                         rrdev = NULL;
1319                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1320                         atomic_inc(&rdev->nr_pending);
1321                         blocked_rdev = rdev;
1322                         break;
1323                 }
1324                 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1325                         atomic_inc(&rrdev->nr_pending);
1326                         blocked_rdev = rrdev;
1327                         break;
1328                 }
1329                 if (rdev && (test_bit(Faulty, &rdev->flags)
1330                              || test_bit(Unmerged, &rdev->flags)))
1331                         rdev = NULL;
1332                 if (rrdev && (test_bit(Faulty, &rrdev->flags)
1333                               || test_bit(Unmerged, &rrdev->flags)))
1334                         rrdev = NULL;
1335
1336                 r10_bio->devs[i].bio = NULL;
1337                 r10_bio->devs[i].repl_bio = NULL;
1338
1339                 if (!rdev && !rrdev) {
1340                         set_bit(R10BIO_Degraded, &r10_bio->state);
1341                         continue;
1342                 }
1343                 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1344                         sector_t first_bad;
1345                         sector_t dev_sector = r10_bio->devs[i].addr;
1346                         int bad_sectors;
1347                         int is_bad;
1348
1349                         is_bad = is_badblock(rdev, dev_sector,
1350                                              max_sectors,
1351                                              &first_bad, &bad_sectors);
1352                         if (is_bad < 0) {
1353                                 /* Mustn't write here until the bad block
1354                                  * is acknowledged
1355                                  */
1356                                 atomic_inc(&rdev->nr_pending);
1357                                 set_bit(BlockedBadBlocks, &rdev->flags);
1358                                 blocked_rdev = rdev;
1359                                 break;
1360                         }
1361                         if (is_bad && first_bad <= dev_sector) {
1362                                 /* Cannot write here at all */
1363                                 bad_sectors -= (dev_sector - first_bad);
1364                                 if (bad_sectors < max_sectors)
1365                                         /* Mustn't write more than bad_sectors
1366                                          * to other devices yet
1367                                          */
1368                                         max_sectors = bad_sectors;
1369                                 /* We don't set R10BIO_Degraded as that
1370                                  * only applies if the disk is missing,
1371                                  * so it might be re-added, and we want to
1372                                  * know to recover this chunk.
1373                                  * In this case the device is here, and the
1374                                  * fact that this chunk is not in-sync is
1375                                  * recorded in the bad block log.
1376                                  */
1377                                 continue;
1378                         }
1379                         if (is_bad) {
1380                                 int good_sectors = first_bad - dev_sector;
1381                                 if (good_sectors < max_sectors)
1382                                         max_sectors = good_sectors;
1383                         }
1384                 }
1385                 if (rdev) {
1386                         r10_bio->devs[i].bio = bio;
1387                         atomic_inc(&rdev->nr_pending);
1388                 }
1389                 if (rrdev) {
1390                         r10_bio->devs[i].repl_bio = bio;
1391                         atomic_inc(&rrdev->nr_pending);
1392                 }
1393         }
1394         rcu_read_unlock();
1395
1396         if (unlikely(blocked_rdev)) {
1397                 /* Have to wait for this device to get unblocked, then retry */
1398                 int j;
1399                 int d;
1400
1401                 for (j = 0; j < i; j++) {
1402                         if (r10_bio->devs[j].bio) {
1403                                 d = r10_bio->devs[j].devnum;
1404                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1405                         }
1406                         if (r10_bio->devs[j].repl_bio) {
1407                                 struct md_rdev *rdev;
1408                                 d = r10_bio->devs[j].devnum;
1409                                 rdev = conf->mirrors[d].replacement;
1410                                 if (!rdev) {
1411                                         /* Race with remove_disk */
1412                                         smp_mb();
1413                                         rdev = conf->mirrors[d].rdev;
1414                                 }
1415                                 rdev_dec_pending(rdev, mddev);
1416                         }
1417                 }
1418                 allow_barrier(conf);
1419                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1420                 wait_barrier(conf);
1421                 goto retry_write;
1422         }
1423
1424         if (max_sectors < r10_bio->sectors) {
1425                 /* We are splitting this into multiple parts, so
1426                  * we need to prepare for allocating another r10_bio.
1427                  */
1428                 r10_bio->sectors = max_sectors;
1429                 spin_lock_irq(&conf->device_lock);
1430                 if (bio->bi_phys_segments == 0)
1431                         bio->bi_phys_segments = 2;
1432                 else
1433                         bio->bi_phys_segments++;
1434                 spin_unlock_irq(&conf->device_lock);
1435         }
1436         sectors_handled = r10_bio->sector + max_sectors -
1437                 bio->bi_iter.bi_sector;
1438
1439         atomic_set(&r10_bio->remaining, 1);
1440         bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1441
1442         for (i = 0; i < conf->copies; i++) {
1443                 struct bio *mbio;
1444                 int d = r10_bio->devs[i].devnum;
1445                 if (r10_bio->devs[i].bio) {
1446                         struct md_rdev *rdev = conf->mirrors[d].rdev;
1447                         mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1448                         bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1449                                  max_sectors);
1450                         r10_bio->devs[i].bio = mbio;
1451
1452                         mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
1453                                            choose_data_offset(r10_bio,
1454                                                               rdev));
1455                         mbio->bi_bdev = rdev->bdev;
1456                         mbio->bi_end_io = raid10_end_write_request;
1457                         mbio->bi_rw =
1458                                 WRITE | do_sync | do_fua | do_discard | do_same;
1459                         mbio->bi_private = r10_bio;
1460
1461                         atomic_inc(&r10_bio->remaining);
1462
1463                         cb = blk_check_plugged(raid10_unplug, mddev,
1464                                                sizeof(*plug));
1465                         if (cb)
1466                                 plug = container_of(cb, struct raid10_plug_cb,
1467                                                     cb);
1468                         else
1469                                 plug = NULL;
1470                         spin_lock_irqsave(&conf->device_lock, flags);
1471                         if (plug) {
1472                                 bio_list_add(&plug->pending, mbio);
1473                                 plug->pending_cnt++;
1474                         } else {
1475                                 bio_list_add(&conf->pending_bio_list, mbio);
1476                                 conf->pending_count++;
1477                         }
1478                         spin_unlock_irqrestore(&conf->device_lock, flags);
1479                         if (!plug)
1480                                 md_wakeup_thread(mddev->thread);
1481                 }
1482
1483                 if (r10_bio->devs[i].repl_bio) {
1484                         struct md_rdev *rdev = conf->mirrors[d].replacement;
1485                         if (rdev == NULL) {
1486                                 /* Replacement just got moved to main 'rdev' */
1487                                 smp_mb();
1488                                 rdev = conf->mirrors[d].rdev;
1489                         }
1490                         mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1491                         bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1492                                  max_sectors);
1493                         r10_bio->devs[i].repl_bio = mbio;
1494
1495                         mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr +
1496                                            choose_data_offset(
1497                                                    r10_bio, rdev));
1498                         mbio->bi_bdev = rdev->bdev;
1499                         mbio->bi_end_io = raid10_end_write_request;
1500                         mbio->bi_rw =
1501                                 WRITE | do_sync | do_fua | do_discard | do_same;
1502                         mbio->bi_private = r10_bio;
1503
1504                         atomic_inc(&r10_bio->remaining);
1505                         spin_lock_irqsave(&conf->device_lock, flags);
1506                         bio_list_add(&conf->pending_bio_list, mbio);
1507                         conf->pending_count++;
1508                         spin_unlock_irqrestore(&conf->device_lock, flags);
1509                         if (!mddev_check_plugged(mddev))
1510                                 md_wakeup_thread(mddev->thread);
1511                 }
1512         }
1513
1514         /* Don't remove the bias on 'remaining' (one_write_done) until
1515          * after checking if we need to go around again.
1516          */
1517
1518         if (sectors_handled < bio_sectors(bio)) {
1519                 one_write_done(r10_bio);
1520                 /* We need another r10_bio.  It has already been counted
1521                  * in bio->bi_phys_segments.
1522                  */
1523                 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1524
1525                 r10_bio->master_bio = bio;
1526                 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1527
1528                 r10_bio->mddev = mddev;
1529                 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1530                 r10_bio->state = 0;
1531                 goto retry_write;
1532         }
1533         one_write_done(r10_bio);
1534 }
1535
1536 static void make_request(struct mddev *mddev, struct bio *bio)
1537 {
1538         struct r10conf *conf = mddev->private;
1539         sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1540         int chunk_sects = chunk_mask + 1;
1541
1542         struct bio *split;
1543
1544         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1545                 md_flush_request(mddev, bio);
1546                 return;
1547         }
1548
1549         md_write_start(mddev, bio);
1550
1551         do {
1552
1553                 /*
1554                  * If this request crosses a chunk boundary, we need to split
1555                  * it.
1556                  */
1557                 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1558                              bio_sectors(bio) > chunk_sects
1559                              && (conf->geo.near_copies < conf->geo.raid_disks
1560                                  || conf->prev.near_copies <
1561                                  conf->prev.raid_disks))) {
1562                         split = bio_split(bio, chunk_sects -
1563                                           (bio->bi_iter.bi_sector &
1564                                            (chunk_sects - 1)),
1565                                           GFP_NOIO, fs_bio_set);
1566                         bio_chain(split, bio);
1567                 } else {
1568                         split = bio;
1569                 }
1570
1571                 __make_request(mddev, split);
1572         } while (split != bio);
1573
1574         /* In case raid10d snuck in to freeze_array */
1575         wake_up(&conf->wait_barrier);
1576 }
1577
1578 static void status(struct seq_file *seq, struct mddev *mddev)
1579 {
1580         struct r10conf *conf = mddev->private;
1581         int i;
1582
1583         if (conf->geo.near_copies < conf->geo.raid_disks)
1584                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1585         if (conf->geo.near_copies > 1)
1586                 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1587         if (conf->geo.far_copies > 1) {
1588                 if (conf->geo.far_offset)
1589                         seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1590                 else
1591                         seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1592         }
1593         seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1594                                         conf->geo.raid_disks - mddev->degraded);
1595         for (i = 0; i < conf->geo.raid_disks; i++)
1596                 seq_printf(seq, "%s",
1597                               conf->mirrors[i].rdev &&
1598                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1599         seq_printf(seq, "]");
1600 }
1601
1602 /* check if there are enough drives for
1603  * every block to appear on atleast one.
1604  * Don't consider the device numbered 'ignore'
1605  * as we might be about to remove it.
1606  */
1607 static int _enough(struct r10conf *conf, int previous, int ignore)
1608 {
1609         int first = 0;
1610         int has_enough = 0;
1611         int disks, ncopies;
1612         if (previous) {
1613                 disks = conf->prev.raid_disks;
1614                 ncopies = conf->prev.near_copies;
1615         } else {
1616                 disks = conf->geo.raid_disks;
1617                 ncopies = conf->geo.near_copies;
1618         }
1619
1620         rcu_read_lock();
1621         do {
1622                 int n = conf->copies;
1623                 int cnt = 0;
1624                 int this = first;
1625                 while (n--) {
1626                         struct md_rdev *rdev;
1627                         if (this != ignore &&
1628                             (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1629                             test_bit(In_sync, &rdev->flags))
1630                                 cnt++;
1631                         this = (this+1) % disks;
1632                 }
1633                 if (cnt == 0)
1634                         goto out;
1635                 first = (first + ncopies) % disks;
1636         } while (first != 0);
1637         has_enough = 1;
1638 out:
1639         rcu_read_unlock();
1640         return has_enough;
1641 }
1642
1643 static int enough(struct r10conf *conf, int ignore)
1644 {
1645         /* when calling 'enough', both 'prev' and 'geo' must
1646          * be stable.
1647          * This is ensured if ->reconfig_mutex or ->device_lock
1648          * is held.
1649          */
1650         return _enough(conf, 0, ignore) &&
1651                 _enough(conf, 1, ignore);
1652 }
1653
1654 static void error(struct mddev *mddev, struct md_rdev *rdev)
1655 {
1656         char b[BDEVNAME_SIZE];
1657         struct r10conf *conf = mddev->private;
1658         unsigned long flags;
1659
1660         /*
1661          * If it is not operational, then we have already marked it as dead
1662          * else if it is the last working disks, ignore the error, let the
1663          * next level up know.
1664          * else mark the drive as failed
1665          */
1666         spin_lock_irqsave(&conf->device_lock, flags);
1667         if (test_bit(In_sync, &rdev->flags)
1668             && !enough(conf, rdev->raid_disk)) {
1669                 /*
1670                  * Don't fail the drive, just return an IO error.
1671                  */
1672                 spin_unlock_irqrestore(&conf->device_lock, flags);
1673                 return;
1674         }
1675         if (test_and_clear_bit(In_sync, &rdev->flags))
1676                 mddev->degraded++;
1677         /*
1678          * If recovery is running, make sure it aborts.
1679          */
1680         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1681         set_bit(Blocked, &rdev->flags);
1682         set_bit(Faulty, &rdev->flags);
1683         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1684         spin_unlock_irqrestore(&conf->device_lock, flags);
1685         printk(KERN_ALERT
1686                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1687                "md/raid10:%s: Operation continuing on %d devices.\n",
1688                mdname(mddev), bdevname(rdev->bdev, b),
1689                mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1690 }
1691
1692 static void print_conf(struct r10conf *conf)
1693 {
1694         int i;
1695         struct raid10_info *tmp;
1696
1697         printk(KERN_DEBUG "RAID10 conf printout:\n");
1698         if (!conf) {
1699                 printk(KERN_DEBUG "(!conf)\n");
1700                 return;
1701         }
1702         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1703                 conf->geo.raid_disks);
1704
1705         for (i = 0; i < conf->geo.raid_disks; i++) {
1706                 char b[BDEVNAME_SIZE];
1707                 tmp = conf->mirrors + i;
1708                 if (tmp->rdev)
1709                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1710                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1711                                 !test_bit(Faulty, &tmp->rdev->flags),
1712                                 bdevname(tmp->rdev->bdev,b));
1713         }
1714 }
1715
1716 static void close_sync(struct r10conf *conf)
1717 {
1718         wait_barrier(conf);
1719         allow_barrier(conf);
1720
1721         mempool_destroy(conf->r10buf_pool);
1722         conf->r10buf_pool = NULL;
1723 }
1724
1725 static int raid10_spare_active(struct mddev *mddev)
1726 {
1727         int i;
1728         struct r10conf *conf = mddev->private;
1729         struct raid10_info *tmp;
1730         int count = 0;
1731         unsigned long flags;
1732
1733         /*
1734          * Find all non-in_sync disks within the RAID10 configuration
1735          * and mark them in_sync
1736          */
1737         for (i = 0; i < conf->geo.raid_disks; i++) {
1738                 tmp = conf->mirrors + i;
1739                 if (tmp->replacement
1740                     && tmp->replacement->recovery_offset == MaxSector
1741                     && !test_bit(Faulty, &tmp->replacement->flags)
1742                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1743                         /* Replacement has just become active */
1744                         if (!tmp->rdev
1745                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1746                                 count++;
1747                         if (tmp->rdev) {
1748                                 /* Replaced device not technically faulty,
1749                                  * but we need to be sure it gets removed
1750                                  * and never re-added.
1751                                  */
1752                                 set_bit(Faulty, &tmp->rdev->flags);
1753                                 sysfs_notify_dirent_safe(
1754                                         tmp->rdev->sysfs_state);
1755                         }
1756                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1757                 } else if (tmp->rdev
1758                            && tmp->rdev->recovery_offset == MaxSector
1759                            && !test_bit(Faulty, &tmp->rdev->flags)
1760                            && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1761                         count++;
1762                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1763                 }
1764         }
1765         spin_lock_irqsave(&conf->device_lock, flags);
1766         mddev->degraded -= count;
1767         spin_unlock_irqrestore(&conf->device_lock, flags);
1768
1769         print_conf(conf);
1770         return count;
1771 }
1772
1773 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1774 {
1775         struct r10conf *conf = mddev->private;
1776         int err = -EEXIST;
1777         int mirror;
1778         int first = 0;
1779         int last = conf->geo.raid_disks - 1;
1780         struct request_queue *q = bdev_get_queue(rdev->bdev);
1781
1782         if (mddev->recovery_cp < MaxSector)
1783                 /* only hot-add to in-sync arrays, as recovery is
1784                  * very different from resync
1785                  */
1786                 return -EBUSY;
1787         if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1788                 return -EINVAL;
1789
1790         if (rdev->raid_disk >= 0)
1791                 first = last = rdev->raid_disk;
1792
1793         if (q->merge_bvec_fn) {
1794                 set_bit(Unmerged, &rdev->flags);
1795                 mddev->merge_check_needed = 1;
1796         }
1797
1798         if (rdev->saved_raid_disk >= first &&
1799             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1800                 mirror = rdev->saved_raid_disk;
1801         else
1802                 mirror = first;
1803         for ( ; mirror <= last ; mirror++) {
1804                 struct raid10_info *p = &conf->mirrors[mirror];
1805                 if (p->recovery_disabled == mddev->recovery_disabled)
1806                         continue;
1807                 if (p->rdev) {
1808                         if (!test_bit(WantReplacement, &p->rdev->flags) ||
1809                             p->replacement != NULL)
1810                                 continue;
1811                         clear_bit(In_sync, &rdev->flags);
1812                         set_bit(Replacement, &rdev->flags);
1813                         rdev->raid_disk = mirror;
1814                         err = 0;
1815                         if (mddev->gendisk)
1816                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1817                                                   rdev->data_offset << 9);
1818                         conf->fullsync = 1;
1819                         rcu_assign_pointer(p->replacement, rdev);
1820                         break;
1821                 }
1822
1823                 if (mddev->gendisk)
1824                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1825                                           rdev->data_offset << 9);
1826
1827                 p->head_position = 0;
1828                 p->recovery_disabled = mddev->recovery_disabled - 1;
1829                 rdev->raid_disk = mirror;
1830                 err = 0;
1831                 if (rdev->saved_raid_disk != mirror)
1832                         conf->fullsync = 1;
1833                 rcu_assign_pointer(p->rdev, rdev);
1834                 break;
1835         }
1836         if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1837                 /* Some requests might not have seen this new
1838                  * merge_bvec_fn.  We must wait for them to complete
1839                  * before merging the device fully.
1840                  * First we make sure any code which has tested
1841                  * our function has submitted the request, then
1842                  * we wait for all outstanding requests to complete.
1843                  */
1844                 synchronize_sched();
1845                 freeze_array(conf, 0);
1846                 unfreeze_array(conf);
1847                 clear_bit(Unmerged, &rdev->flags);
1848         }
1849         md_integrity_add_rdev(rdev, mddev);
1850         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1851                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1852
1853         print_conf(conf);
1854         return err;
1855 }
1856
1857 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1858 {
1859         struct r10conf *conf = mddev->private;
1860         int err = 0;
1861         int number = rdev->raid_disk;
1862         struct md_rdev **rdevp;
1863         struct raid10_info *p = conf->mirrors + number;
1864
1865         print_conf(conf);
1866         if (rdev == p->rdev)
1867                 rdevp = &p->rdev;
1868         else if (rdev == p->replacement)
1869                 rdevp = &p->replacement;
1870         else
1871                 return 0;
1872
1873         if (test_bit(In_sync, &rdev->flags) ||
1874             atomic_read(&rdev->nr_pending)) {
1875                 err = -EBUSY;
1876                 goto abort;
1877         }
1878         /* Only remove faulty devices if recovery
1879          * is not possible.
1880          */
1881         if (!test_bit(Faulty, &rdev->flags) &&
1882             mddev->recovery_disabled != p->recovery_disabled &&
1883             (!p->replacement || p->replacement == rdev) &&
1884             number < conf->geo.raid_disks &&
1885             enough(conf, -1)) {
1886                 err = -EBUSY;
1887                 goto abort;
1888         }
1889         *rdevp = NULL;
1890         synchronize_rcu();
1891         if (atomic_read(&rdev->nr_pending)) {
1892                 /* lost the race, try later */
1893                 err = -EBUSY;
1894                 *rdevp = rdev;
1895                 goto abort;
1896         } else if (p->replacement) {
1897                 /* We must have just cleared 'rdev' */
1898                 p->rdev = p->replacement;
1899                 clear_bit(Replacement, &p->replacement->flags);
1900                 smp_mb(); /* Make sure other CPUs may see both as identical
1901                            * but will never see neither -- if they are careful.
1902                            */
1903                 p->replacement = NULL;
1904                 clear_bit(WantReplacement, &rdev->flags);
1905         } else
1906                 /* We might have just remove the Replacement as faulty
1907                  * Clear the flag just in case
1908                  */
1909                 clear_bit(WantReplacement, &rdev->flags);
1910
1911         err = md_integrity_register(mddev);
1912
1913 abort:
1914
1915         print_conf(conf);
1916         return err;
1917 }
1918
1919 static void end_sync_read(struct bio *bio, int error)
1920 {
1921         struct r10bio *r10_bio = bio->bi_private;
1922         struct r10conf *conf = r10_bio->mddev->private;
1923         int d;
1924
1925         if (bio == r10_bio->master_bio) {
1926                 /* this is a reshape read */
1927                 d = r10_bio->read_slot; /* really the read dev */
1928         } else
1929                 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1930
1931         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1932                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1933         else
1934                 /* The write handler will notice the lack of
1935                  * R10BIO_Uptodate and record any errors etc
1936                  */
1937                 atomic_add(r10_bio->sectors,
1938                            &conf->mirrors[d].rdev->corrected_errors);
1939
1940         /* for reconstruct, we always reschedule after a read.
1941          * for resync, only after all reads
1942          */
1943         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1944         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1945             atomic_dec_and_test(&r10_bio->remaining)) {
1946                 /* we have read all the blocks,
1947                  * do the comparison in process context in raid10d
1948                  */
1949                 reschedule_retry(r10_bio);
1950         }
1951 }
1952
1953 static void end_sync_request(struct r10bio *r10_bio)
1954 {
1955         struct mddev *mddev = r10_bio->mddev;
1956
1957         while (atomic_dec_and_test(&r10_bio->remaining)) {
1958                 if (r10_bio->master_bio == NULL) {
1959                         /* the primary of several recovery bios */
1960                         sector_t s = r10_bio->sectors;
1961                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1962                             test_bit(R10BIO_WriteError, &r10_bio->state))
1963                                 reschedule_retry(r10_bio);
1964                         else
1965                                 put_buf(r10_bio);
1966                         md_done_sync(mddev, s, 1);
1967                         break;
1968                 } else {
1969                         struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1970                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1971                             test_bit(R10BIO_WriteError, &r10_bio->state))
1972                                 reschedule_retry(r10_bio);
1973                         else
1974                                 put_buf(r10_bio);
1975                         r10_bio = r10_bio2;
1976                 }
1977         }
1978 }
1979
1980 static void end_sync_write(struct bio *bio, int error)
1981 {
1982         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1983         struct r10bio *r10_bio = bio->bi_private;
1984         struct mddev *mddev = r10_bio->mddev;
1985         struct r10conf *conf = mddev->private;
1986         int d;
1987         sector_t first_bad;
1988         int bad_sectors;
1989         int slot;
1990         int repl;
1991         struct md_rdev *rdev = NULL;
1992
1993         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1994         if (repl)
1995                 rdev = conf->mirrors[d].replacement;
1996         else
1997                 rdev = conf->mirrors[d].rdev;
1998
1999         if (!uptodate) {
2000                 if (repl)
2001                         md_error(mddev, rdev);
2002                 else {
2003                         set_bit(WriteErrorSeen, &rdev->flags);
2004                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2005                                 set_bit(MD_RECOVERY_NEEDED,
2006                                         &rdev->mddev->recovery);
2007                         set_bit(R10BIO_WriteError, &r10_bio->state);
2008                 }
2009         } else if (is_badblock(rdev,
2010                              r10_bio->devs[slot].addr,
2011                              r10_bio->sectors,
2012                              &first_bad, &bad_sectors))
2013                 set_bit(R10BIO_MadeGood, &r10_bio->state);
2014
2015         rdev_dec_pending(rdev, mddev);
2016
2017         end_sync_request(r10_bio);
2018 }
2019
2020 /*
2021  * Note: sync and recover and handled very differently for raid10
2022  * This code is for resync.
2023  * For resync, we read through virtual addresses and read all blocks.
2024  * If there is any error, we schedule a write.  The lowest numbered
2025  * drive is authoritative.
2026  * However requests come for physical address, so we need to map.
2027  * For every physical address there are raid_disks/copies virtual addresses,
2028  * which is always are least one, but is not necessarly an integer.
2029  * This means that a physical address can span multiple chunks, so we may
2030  * have to submit multiple io requests for a single sync request.
2031  */
2032 /*
2033  * We check if all blocks are in-sync and only write to blocks that
2034  * aren't in sync
2035  */
2036 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2037 {
2038         struct r10conf *conf = mddev->private;
2039         int i, first;
2040         struct bio *tbio, *fbio;
2041         int vcnt;
2042
2043         atomic_set(&r10_bio->remaining, 1);
2044
2045         /* find the first device with a block */
2046         for (i=0; i<conf->copies; i++)
2047                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
2048                         break;
2049
2050         if (i == conf->copies)
2051                 goto done;
2052
2053         first = i;
2054         fbio = r10_bio->devs[i].bio;
2055
2056         vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2057         /* now find blocks with errors */
2058         for (i=0 ; i < conf->copies ; i++) {
2059                 int  j, d;
2060
2061                 tbio = r10_bio->devs[i].bio;
2062
2063                 if (tbio->bi_end_io != end_sync_read)
2064                         continue;
2065                 if (i == first)
2066                         continue;
2067                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
2068                         /* We know that the bi_io_vec layout is the same for
2069                          * both 'first' and 'i', so we just compare them.
2070                          * All vec entries are PAGE_SIZE;
2071                          */
2072                         int sectors = r10_bio->sectors;
2073                         for (j = 0; j < vcnt; j++) {
2074                                 int len = PAGE_SIZE;
2075                                 if (sectors < (len / 512))
2076                                         len = sectors * 512;
2077                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2078                                            page_address(tbio->bi_io_vec[j].bv_page),
2079                                            len))
2080                                         break;
2081                                 sectors -= len/512;
2082                         }
2083                         if (j == vcnt)
2084                                 continue;
2085                         atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2086                         if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2087                                 /* Don't fix anything. */
2088                                 continue;
2089                 }
2090                 /* Ok, we need to write this bio, either to correct an
2091                  * inconsistency or to correct an unreadable block.
2092                  * First we need to fixup bv_offset, bv_len and
2093                  * bi_vecs, as the read request might have corrupted these
2094                  */
2095                 bio_reset(tbio);
2096
2097                 tbio->bi_vcnt = vcnt;
2098                 tbio->bi_iter.bi_size = r10_bio->sectors << 9;
2099                 tbio->bi_rw = WRITE;
2100                 tbio->bi_private = r10_bio;
2101                 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2102
2103                 for (j=0; j < vcnt ; j++) {
2104                         tbio->bi_io_vec[j].bv_offset = 0;
2105                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
2106
2107                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2108                                page_address(fbio->bi_io_vec[j].bv_page),
2109                                PAGE_SIZE);
2110                 }
2111                 tbio->bi_end_io = end_sync_write;
2112
2113                 d = r10_bio->devs[i].devnum;
2114                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2115                 atomic_inc(&r10_bio->remaining);
2116                 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2117
2118                 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2119                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2120                 generic_make_request(tbio);
2121         }
2122
2123         /* Now write out to any replacement devices
2124          * that are active
2125          */
2126         for (i = 0; i < conf->copies; i++) {
2127                 int j, d;
2128
2129                 tbio = r10_bio->devs[i].repl_bio;
2130                 if (!tbio || !tbio->bi_end_io)
2131                         continue;
2132                 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2133                     && r10_bio->devs[i].bio != fbio)
2134                         for (j = 0; j < vcnt; j++)
2135                                 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2136                                        page_address(fbio->bi_io_vec[j].bv_page),
2137                                        PAGE_SIZE);
2138                 d = r10_bio->devs[i].devnum;
2139                 atomic_inc(&r10_bio->remaining);
2140                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2141                              bio_sectors(tbio));
2142                 generic_make_request(tbio);
2143         }
2144
2145 done:
2146         if (atomic_dec_and_test(&r10_bio->remaining)) {
2147                 md_done_sync(mddev, r10_bio->sectors, 1);
2148                 put_buf(r10_bio);
2149         }
2150 }
2151
2152 /*
2153  * Now for the recovery code.
2154  * Recovery happens across physical sectors.
2155  * We recover all non-is_sync drives by finding the virtual address of
2156  * each, and then choose a working drive that also has that virt address.
2157  * There is a separate r10_bio for each non-in_sync drive.
2158  * Only the first two slots are in use. The first for reading,
2159  * The second for writing.
2160  *
2161  */
2162 static void fix_recovery_read_error(struct r10bio *r10_bio)
2163 {
2164         /* We got a read error during recovery.
2165          * We repeat the read in smaller page-sized sections.
2166          * If a read succeeds, write it to the new device or record
2167          * a bad block if we cannot.
2168          * If a read fails, record a bad block on both old and
2169          * new devices.
2170          */
2171         struct mddev *mddev = r10_bio->mddev;
2172         struct r10conf *conf = mddev->private;
2173         struct bio *bio = r10_bio->devs[0].bio;
2174         sector_t sect = 0;
2175         int sectors = r10_bio->sectors;
2176         int idx = 0;
2177         int dr = r10_bio->devs[0].devnum;
2178         int dw = r10_bio->devs[1].devnum;
2179
2180         while (sectors) {
2181                 int s = sectors;
2182                 struct md_rdev *rdev;
2183                 sector_t addr;
2184                 int ok;
2185
2186                 if (s > (PAGE_SIZE>>9))
2187                         s = PAGE_SIZE >> 9;
2188
2189                 rdev = conf->mirrors[dr].rdev;
2190                 addr = r10_bio->devs[0].addr + sect,
2191                 ok = sync_page_io(rdev,
2192                                   addr,
2193                                   s << 9,
2194                                   bio->bi_io_vec[idx].bv_page,
2195                                   READ, false);
2196                 if (ok) {
2197                         rdev = conf->mirrors[dw].rdev;
2198                         addr = r10_bio->devs[1].addr + sect;
2199                         ok = sync_page_io(rdev,
2200                                           addr,
2201                                           s << 9,
2202                                           bio->bi_io_vec[idx].bv_page,
2203                                           WRITE, false);
2204                         if (!ok) {
2205                                 set_bit(WriteErrorSeen, &rdev->flags);
2206                                 if (!test_and_set_bit(WantReplacement,
2207                                                       &rdev->flags))
2208                                         set_bit(MD_RECOVERY_NEEDED,
2209                                                 &rdev->mddev->recovery);
2210                         }
2211                 }
2212                 if (!ok) {
2213                         /* We don't worry if we cannot set a bad block -
2214                          * it really is bad so there is no loss in not
2215                          * recording it yet
2216                          */
2217                         rdev_set_badblocks(rdev, addr, s, 0);
2218
2219                         if (rdev != conf->mirrors[dw].rdev) {
2220                                 /* need bad block on destination too */
2221                                 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2222                                 addr = r10_bio->devs[1].addr + sect;
2223                                 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2224                                 if (!ok) {
2225                                         /* just abort the recovery */
2226                                         printk(KERN_NOTICE
2227                                                "md/raid10:%s: recovery aborted"
2228                                                " due to read error\n",
2229                                                mdname(mddev));
2230
2231                                         conf->mirrors[dw].recovery_disabled
2232                                                 = mddev->recovery_disabled;
2233                                         set_bit(MD_RECOVERY_INTR,
2234                                                 &mddev->recovery);
2235                                         break;
2236                                 }
2237                         }
2238                 }
2239
2240                 sectors -= s;
2241                 sect += s;
2242                 idx++;
2243         }
2244 }
2245
2246 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2247 {
2248         struct r10conf *conf = mddev->private;
2249         int d;
2250         struct bio *wbio, *wbio2;
2251
2252         if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2253                 fix_recovery_read_error(r10_bio);
2254                 end_sync_request(r10_bio);
2255                 return;
2256         }
2257
2258         /*
2259          * share the pages with the first bio
2260          * and submit the write request
2261          */
2262         d = r10_bio->devs[1].devnum;
2263         wbio = r10_bio->devs[1].bio;
2264         wbio2 = r10_bio->devs[1].repl_bio;
2265         /* Need to test wbio2->bi_end_io before we call
2266          * generic_make_request as if the former is NULL,
2267          * the latter is free to free wbio2.
2268          */
2269         if (wbio2 && !wbio2->bi_end_io)
2270                 wbio2 = NULL;
2271         if (wbio->bi_end_io) {
2272                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2273                 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2274                 generic_make_request(wbio);
2275         }
2276         if (wbio2) {
2277                 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2278                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2279                              bio_sectors(wbio2));
2280                 generic_make_request(wbio2);
2281         }
2282 }
2283
2284 /*
2285  * Used by fix_read_error() to decay the per rdev read_errors.
2286  * We halve the read error count for every hour that has elapsed
2287  * since the last recorded read error.
2288  *
2289  */
2290 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2291 {
2292         struct timespec cur_time_mon;
2293         unsigned long hours_since_last;
2294         unsigned int read_errors = atomic_read(&rdev->read_errors);
2295
2296         ktime_get_ts(&cur_time_mon);
2297
2298         if (rdev->last_read_error.tv_sec == 0 &&
2299             rdev->last_read_error.tv_nsec == 0) {
2300                 /* first time we've seen a read error */
2301                 rdev->last_read_error = cur_time_mon;
2302                 return;
2303         }
2304
2305         hours_since_last = (cur_time_mon.tv_sec -
2306                             rdev->last_read_error.tv_sec) / 3600;
2307
2308         rdev->last_read_error = cur_time_mon;
2309
2310         /*
2311          * if hours_since_last is > the number of bits in read_errors
2312          * just set read errors to 0. We do this to avoid
2313          * overflowing the shift of read_errors by hours_since_last.
2314          */
2315         if (hours_since_last >= 8 * sizeof(read_errors))
2316                 atomic_set(&rdev->read_errors, 0);
2317         else
2318                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2319 }
2320
2321 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2322                             int sectors, struct page *page, int rw)
2323 {
2324         sector_t first_bad;
2325         int bad_sectors;
2326
2327         if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2328             && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2329                 return -1;
2330         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2331                 /* success */
2332                 return 1;
2333         if (rw == WRITE) {
2334                 set_bit(WriteErrorSeen, &rdev->flags);
2335                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2336                         set_bit(MD_RECOVERY_NEEDED,
2337                                 &rdev->mddev->recovery);
2338         }
2339         /* need to record an error - either for the block or the device */
2340         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2341                 md_error(rdev->mddev, rdev);
2342         return 0;
2343 }
2344
2345 /*
2346  * This is a kernel thread which:
2347  *
2348  *      1.      Retries failed read operations on working mirrors.
2349  *      2.      Updates the raid superblock when problems encounter.
2350  *      3.      Performs writes following reads for array synchronising.
2351  */
2352
2353 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2354 {
2355         int sect = 0; /* Offset from r10_bio->sector */
2356         int sectors = r10_bio->sectors;
2357         struct md_rdev*rdev;
2358         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2359         int d = r10_bio->devs[r10_bio->read_slot].devnum;
2360
2361         /* still own a reference to this rdev, so it cannot
2362          * have been cleared recently.
2363          */
2364         rdev = conf->mirrors[d].rdev;
2365
2366         if (test_bit(Faulty, &rdev->flags))
2367                 /* drive has already been failed, just ignore any
2368                    more fix_read_error() attempts */
2369                 return;
2370
2371         check_decay_read_errors(mddev, rdev);
2372         atomic_inc(&rdev->read_errors);
2373         if (atomic_read(&rdev->read_errors) > max_read_errors) {
2374                 char b[BDEVNAME_SIZE];
2375                 bdevname(rdev->bdev, b);
2376
2377                 printk(KERN_NOTICE
2378                        "md/raid10:%s: %s: Raid device exceeded "
2379                        "read_error threshold [cur %d:max %d]\n",
2380                        mdname(mddev), b,
2381                        atomic_read(&rdev->read_errors), max_read_errors);
2382                 printk(KERN_NOTICE
2383                        "md/raid10:%s: %s: Failing raid device\n",
2384                        mdname(mddev), b);
2385                 md_error(mddev, conf->mirrors[d].rdev);
2386                 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2387                 return;
2388         }
2389
2390         while(sectors) {
2391                 int s = sectors;
2392                 int sl = r10_bio->read_slot;
2393                 int success = 0;
2394                 int start;
2395
2396                 if (s > (PAGE_SIZE>>9))
2397                         s = PAGE_SIZE >> 9;
2398
2399                 rcu_read_lock();
2400                 do {
2401                         sector_t first_bad;
2402                         int bad_sectors;
2403
2404                         d = r10_bio->devs[sl].devnum;
2405                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2406                         if (rdev &&
2407                             !test_bit(Unmerged, &rdev->flags) &&
2408                             test_bit(In_sync, &rdev->flags) &&
2409                             is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2410                                         &first_bad, &bad_sectors) == 0) {
2411                                 atomic_inc(&rdev->nr_pending);
2412                                 rcu_read_unlock();
2413                                 success = sync_page_io(rdev,
2414                                                        r10_bio->devs[sl].addr +
2415                                                        sect,
2416                                                        s<<9,
2417                                                        conf->tmppage, READ, false);
2418                                 rdev_dec_pending(rdev, mddev);
2419                                 rcu_read_lock();
2420                                 if (success)
2421                                         break;
2422                         }
2423                         sl++;
2424                         if (sl == conf->copies)
2425                                 sl = 0;
2426                 } while (!success && sl != r10_bio->read_slot);
2427                 rcu_read_unlock();
2428
2429                 if (!success) {
2430                         /* Cannot read from anywhere, just mark the block
2431                          * as bad on the first device to discourage future
2432                          * reads.
2433                          */
2434                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2435                         rdev = conf->mirrors[dn].rdev;
2436
2437                         if (!rdev_set_badblocks(
2438                                     rdev,
2439                                     r10_bio->devs[r10_bio->read_slot].addr
2440                                     + sect,
2441                                     s, 0)) {
2442                                 md_error(mddev, rdev);
2443                                 r10_bio->devs[r10_bio->read_slot].bio
2444                                         = IO_BLOCKED;
2445                         }
2446                         break;
2447                 }
2448
2449                 start = sl;
2450                 /* write it back and re-read */
2451                 rcu_read_lock();
2452                 while (sl != r10_bio->read_slot) {
2453                         char b[BDEVNAME_SIZE];
2454
2455                         if (sl==0)
2456                                 sl = conf->copies;
2457                         sl--;
2458                         d = r10_bio->devs[sl].devnum;
2459                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2460                         if (!rdev ||
2461                             test_bit(Unmerged, &rdev->flags) ||
2462                             !test_bit(In_sync, &rdev->flags))
2463                                 continue;
2464
2465                         atomic_inc(&rdev->nr_pending);
2466                         rcu_read_unlock();
2467                         if (r10_sync_page_io(rdev,
2468                                              r10_bio->devs[sl].addr +
2469                                              sect,
2470                                              s, conf->tmppage, WRITE)
2471                             == 0) {
2472                                 /* Well, this device is dead */
2473                                 printk(KERN_NOTICE
2474                                        "md/raid10:%s: read correction "
2475                                        "write failed"
2476                                        " (%d sectors at %llu on %s)\n",
2477                                        mdname(mddev), s,
2478                                        (unsigned long long)(
2479                                                sect +
2480                                                choose_data_offset(r10_bio,
2481                                                                   rdev)),
2482                                        bdevname(rdev->bdev, b));
2483                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2484                                        "drive\n",
2485                                        mdname(mddev),
2486                                        bdevname(rdev->bdev, b));
2487                         }
2488                         rdev_dec_pending(rdev, mddev);
2489                         rcu_read_lock();
2490                 }
2491                 sl = start;
2492                 while (sl != r10_bio->read_slot) {
2493                         char b[BDEVNAME_SIZE];
2494
2495                         if (sl==0)
2496                                 sl = conf->copies;
2497                         sl--;
2498                         d = r10_bio->devs[sl].devnum;
2499                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2500                         if (!rdev ||
2501                             !test_bit(In_sync, &rdev->flags))
2502                                 continue;
2503
2504                         atomic_inc(&rdev->nr_pending);
2505                         rcu_read_unlock();
2506                         switch (r10_sync_page_io(rdev,
2507                                              r10_bio->devs[sl].addr +
2508                                              sect,
2509                                              s, conf->tmppage,
2510                                                  READ)) {
2511                         case 0:
2512                                 /* Well, this device is dead */
2513                                 printk(KERN_NOTICE
2514                                        "md/raid10:%s: unable to read back "
2515                                        "corrected sectors"
2516                                        " (%d sectors at %llu on %s)\n",
2517                                        mdname(mddev), s,
2518                                        (unsigned long long)(
2519                                                sect +
2520                                                choose_data_offset(r10_bio, rdev)),
2521                                        bdevname(rdev->bdev, b));
2522                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2523                                        "drive\n",
2524                                        mdname(mddev),
2525                                        bdevname(rdev->bdev, b));
2526                                 break;
2527                         case 1:
2528                                 printk(KERN_INFO
2529                                        "md/raid10:%s: read error corrected"
2530                                        " (%d sectors at %llu on %s)\n",
2531                                        mdname(mddev), s,
2532                                        (unsigned long long)(
2533                                                sect +
2534                                                choose_data_offset(r10_bio, rdev)),
2535                                        bdevname(rdev->bdev, b));
2536                                 atomic_add(s, &rdev->corrected_errors);
2537                         }
2538
2539                         rdev_dec_pending(rdev, mddev);
2540                         rcu_read_lock();
2541                 }
2542                 rcu_read_unlock();
2543
2544                 sectors -= s;
2545                 sect += s;
2546         }
2547 }
2548
2549 static int narrow_write_error(struct r10bio *r10_bio, int i)
2550 {
2551         struct bio *bio = r10_bio->master_bio;
2552         struct mddev *mddev = r10_bio->mddev;
2553         struct r10conf *conf = mddev->private;
2554         struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2555         /* bio has the data to be written to slot 'i' where
2556          * we just recently had a write error.
2557          * We repeatedly clone the bio and trim down to one block,
2558          * then try the write.  Where the write fails we record
2559          * a bad block.
2560          * It is conceivable that the bio doesn't exactly align with
2561          * blocks.  We must handle this.
2562          *
2563          * We currently own a reference to the rdev.
2564          */
2565
2566         int block_sectors;
2567         sector_t sector;
2568         int sectors;
2569         int sect_to_write = r10_bio->sectors;
2570         int ok = 1;
2571
2572         if (rdev->badblocks.shift < 0)
2573                 return 0;
2574
2575         block_sectors = roundup(1 << rdev->badblocks.shift,
2576                                 bdev_logical_block_size(rdev->bdev) >> 9);
2577         sector = r10_bio->sector;
2578         sectors = ((r10_bio->sector + block_sectors)
2579                    & ~(sector_t)(block_sectors - 1))
2580                 - sector;
2581
2582         while (sect_to_write) {
2583                 struct bio *wbio;
2584                 if (sectors > sect_to_write)
2585                         sectors = sect_to_write;
2586                 /* Write at 'sector' for 'sectors' */
2587                 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2588                 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2589                 wbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
2590                                    choose_data_offset(r10_bio, rdev) +
2591                                    (sector - r10_bio->sector));
2592                 wbio->bi_bdev = rdev->bdev;
2593                 if (submit_bio_wait(WRITE, wbio) == 0)
2594                         /* Failure! */
2595                         ok = rdev_set_badblocks(rdev, sector,
2596                                                 sectors, 0)
2597                                 && ok;
2598
2599                 bio_put(wbio);
2600                 sect_to_write -= sectors;
2601                 sector += sectors;
2602                 sectors = block_sectors;
2603         }
2604         return ok;
2605 }
2606
2607 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2608 {
2609         int slot = r10_bio->read_slot;
2610         struct bio *bio;
2611         struct r10conf *conf = mddev->private;
2612         struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2613         char b[BDEVNAME_SIZE];
2614         unsigned long do_sync;
2615         int max_sectors;
2616
2617         /* we got a read error. Maybe the drive is bad.  Maybe just
2618          * the block and we can fix it.
2619          * We freeze all other IO, and try reading the block from
2620          * other devices.  When we find one, we re-write
2621          * and check it that fixes the read error.
2622          * This is all done synchronously while the array is
2623          * frozen.
2624          */
2625         bio = r10_bio->devs[slot].bio;
2626         bdevname(bio->bi_bdev, b);
2627         bio_put(bio);
2628         r10_bio->devs[slot].bio = NULL;
2629
2630         if (mddev->ro == 0) {
2631                 freeze_array(conf, 1);
2632                 fix_read_error(conf, mddev, r10_bio);
2633                 unfreeze_array(conf);
2634         } else
2635                 r10_bio->devs[slot].bio = IO_BLOCKED;
2636
2637         rdev_dec_pending(rdev, mddev);
2638
2639 read_more:
2640         rdev = read_balance(conf, r10_bio, &max_sectors);
2641         if (rdev == NULL) {
2642                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2643                        " read error for block %llu\n",
2644                        mdname(mddev), b,
2645                        (unsigned long long)r10_bio->sector);
2646                 raid_end_bio_io(r10_bio);
2647                 return;
2648         }
2649
2650         do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2651         slot = r10_bio->read_slot;
2652         printk_ratelimited(
2653                 KERN_ERR
2654                 "md/raid10:%s: %s: redirecting "
2655                 "sector %llu to another mirror\n",
2656                 mdname(mddev),
2657                 bdevname(rdev->bdev, b),
2658                 (unsigned long long)r10_bio->sector);
2659         bio = bio_clone_mddev(r10_bio->master_bio,
2660                               GFP_NOIO, mddev);
2661         bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2662         r10_bio->devs[slot].bio = bio;
2663         r10_bio->devs[slot].rdev = rdev;
2664         bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2665                 + choose_data_offset(r10_bio, rdev);
2666         bio->bi_bdev = rdev->bdev;
2667         bio->bi_rw = READ | do_sync;
2668         bio->bi_private = r10_bio;
2669         bio->bi_end_io = raid10_end_read_request;
2670         if (max_sectors < r10_bio->sectors) {
2671                 /* Drat - have to split this up more */
2672                 struct bio *mbio = r10_bio->master_bio;
2673                 int sectors_handled =
2674                         r10_bio->sector + max_sectors
2675                         - mbio->bi_iter.bi_sector;
2676                 r10_bio->sectors = max_sectors;
2677                 spin_lock_irq(&conf->device_lock);
2678                 if (mbio->bi_phys_segments == 0)
2679                         mbio->bi_phys_segments = 2;
2680                 else
2681                         mbio->bi_phys_segments++;
2682                 spin_unlock_irq(&conf->device_lock);
2683                 generic_make_request(bio);
2684
2685                 r10_bio = mempool_alloc(conf->r10bio_pool,
2686                                         GFP_NOIO);
2687                 r10_bio->master_bio = mbio;
2688                 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2689                 r10_bio->state = 0;
2690                 set_bit(R10BIO_ReadError,
2691                         &r10_bio->state);
2692                 r10_bio->mddev = mddev;
2693                 r10_bio->sector = mbio->bi_iter.bi_sector
2694                         + sectors_handled;
2695
2696                 goto read_more;
2697         } else
2698                 generic_make_request(bio);
2699 }
2700
2701 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2702 {
2703         /* Some sort of write request has finished and it
2704          * succeeded in writing where we thought there was a
2705          * bad block.  So forget the bad block.
2706          * Or possibly if failed and we need to record
2707          * a bad block.
2708          */
2709         int m;
2710         struct md_rdev *rdev;
2711
2712         if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2713             test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2714                 for (m = 0; m < conf->copies; m++) {
2715                         int dev = r10_bio->devs[m].devnum;
2716                         rdev = conf->mirrors[dev].rdev;
2717                         if (r10_bio->devs[m].bio == NULL)
2718                                 continue;
2719                         if (test_bit(BIO_UPTODATE,
2720                                      &r10_bio->devs[m].bio->bi_flags)) {
2721                                 rdev_clear_badblocks(
2722                                         rdev,
2723                                         r10_bio->devs[m].addr,
2724                                         r10_bio->sectors, 0);
2725                         } else {
2726                                 if (!rdev_set_badblocks(
2727                                             rdev,
2728                                             r10_bio->devs[m].addr,
2729                                             r10_bio->sectors, 0))
2730                                         md_error(conf->mddev, rdev);
2731                         }
2732                         rdev = conf->mirrors[dev].replacement;
2733                         if (r10_bio->devs[m].repl_bio == NULL)
2734                                 continue;
2735                         if (test_bit(BIO_UPTODATE,
2736                                      &r10_bio->devs[m].repl_bio->bi_flags)) {
2737                                 rdev_clear_badblocks(
2738                                         rdev,
2739                                         r10_bio->devs[m].addr,
2740                                         r10_bio->sectors, 0);
2741                         } else {
2742                                 if (!rdev_set_badblocks(
2743                                             rdev,
2744                                             r10_bio->devs[m].addr,
2745                                             r10_bio->sectors, 0))
2746                                         md_error(conf->mddev, rdev);
2747                         }
2748                 }
2749                 put_buf(r10_bio);
2750         } else {
2751                 for (m = 0; m < conf->copies; m++) {
2752                         int dev = r10_bio->devs[m].devnum;
2753                         struct bio *bio = r10_bio->devs[m].bio;
2754                         rdev = conf->mirrors[dev].rdev;
2755                         if (bio == IO_MADE_GOOD) {
2756                                 rdev_clear_badblocks(
2757                                         rdev,
2758                                         r10_bio->devs[m].addr,
2759                                         r10_bio->sectors, 0);
2760                                 rdev_dec_pending(rdev, conf->mddev);
2761                         } else if (bio != NULL &&
2762                                    !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2763                                 if (!narrow_write_error(r10_bio, m)) {
2764                                         md_error(conf->mddev, rdev);
2765                                         set_bit(R10BIO_Degraded,
2766                                                 &r10_bio->state);
2767                                 }
2768                                 rdev_dec_pending(rdev, conf->mddev);
2769                         }
2770                         bio = r10_bio->devs[m].repl_bio;
2771                         rdev = conf->mirrors[dev].replacement;
2772                         if (rdev && bio == IO_MADE_GOOD) {
2773                                 rdev_clear_badblocks(
2774                                         rdev,
2775                                         r10_bio->devs[m].addr,
2776                                         r10_bio->sectors, 0);
2777                                 rdev_dec_pending(rdev, conf->mddev);
2778                         }
2779                 }
2780                 if (test_bit(R10BIO_WriteError,
2781                              &r10_bio->state))
2782                         close_write(r10_bio);
2783                 raid_end_bio_io(r10_bio);
2784         }
2785 }
2786
2787 static void raid10d(struct md_thread *thread)
2788 {
2789         struct mddev *mddev = thread->mddev;
2790         struct r10bio *r10_bio;
2791         unsigned long flags;
2792         struct r10conf *conf = mddev->private;
2793         struct list_head *head = &conf->retry_list;
2794         struct blk_plug plug;
2795
2796         md_check_recovery(mddev);
2797
2798         blk_start_plug(&plug);
2799         for (;;) {
2800
2801                 flush_pending_writes(conf);
2802
2803                 spin_lock_irqsave(&conf->device_lock, flags);
2804                 if (list_empty(head)) {
2805                         spin_unlock_irqrestore(&conf->device_lock, flags);
2806                         break;
2807                 }
2808                 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2809                 list_del(head->prev);
2810                 conf->nr_queued--;
2811                 spin_unlock_irqrestore(&conf->device_lock, flags);
2812
2813                 mddev = r10_bio->mddev;
2814                 conf = mddev->private;
2815                 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2816                     test_bit(R10BIO_WriteError, &r10_bio->state))
2817                         handle_write_completed(conf, r10_bio);
2818                 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2819                         reshape_request_write(mddev, r10_bio);
2820                 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2821                         sync_request_write(mddev, r10_bio);
2822                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2823                         recovery_request_write(mddev, r10_bio);
2824                 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2825                         handle_read_error(mddev, r10_bio);
2826                 else {
2827                         /* just a partial read to be scheduled from a
2828                          * separate context
2829                          */
2830                         int slot = r10_bio->read_slot;
2831                         generic_make_request(r10_bio->devs[slot].bio);
2832                 }
2833
2834                 cond_resched();
2835                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2836                         md_check_recovery(mddev);
2837         }
2838         blk_finish_plug(&plug);
2839 }
2840
2841 static int init_resync(struct r10conf *conf)
2842 {
2843         int buffs;
2844         int i;
2845
2846         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2847         BUG_ON(conf->r10buf_pool);
2848         conf->have_replacement = 0;
2849         for (i = 0; i < conf->geo.raid_disks; i++)
2850                 if (conf->mirrors[i].replacement)
2851                         conf->have_replacement = 1;
2852         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2853         if (!conf->r10buf_pool)
2854                 return -ENOMEM;
2855         conf->next_resync = 0;
2856         return 0;
2857 }
2858
2859 /*
2860  * perform a "sync" on one "block"
2861  *
2862  * We need to make sure that no normal I/O request - particularly write
2863  * requests - conflict with active sync requests.
2864  *
2865  * This is achieved by tracking pending requests and a 'barrier' concept
2866  * that can be installed to exclude normal IO requests.
2867  *
2868  * Resync and recovery are handled very differently.
2869  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2870  *
2871  * For resync, we iterate over virtual addresses, read all copies,
2872  * and update if there are differences.  If only one copy is live,
2873  * skip it.
2874  * For recovery, we iterate over physical addresses, read a good
2875  * value for each non-in_sync drive, and over-write.
2876  *
2877  * So, for recovery we may have several outstanding complex requests for a
2878  * given address, one for each out-of-sync device.  We model this by allocating
2879  * a number of r10_bio structures, one for each out-of-sync device.
2880  * As we setup these structures, we collect all bio's together into a list
2881  * which we then process collectively to add pages, and then process again
2882  * to pass to generic_make_request.
2883  *
2884  * The r10_bio structures are linked using a borrowed master_bio pointer.
2885  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2886  * has its remaining count decremented to 0, the whole complex operation
2887  * is complete.
2888  *
2889  */
2890
2891 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2892                              int *skipped)
2893 {
2894         struct r10conf *conf = mddev->private;
2895         struct r10bio *r10_bio;
2896         struct bio *biolist = NULL, *bio;
2897         sector_t max_sector, nr_sectors;
2898         int i;
2899         int max_sync;
2900         sector_t sync_blocks;
2901         sector_t sectors_skipped = 0;
2902         int chunks_skipped = 0;
2903         sector_t chunk_mask = conf->geo.chunk_mask;
2904
2905         if (!conf->r10buf_pool)
2906                 if (init_resync(conf))
2907                         return 0;
2908
2909         /*
2910          * Allow skipping a full rebuild for incremental assembly
2911          * of a clean array, like RAID1 does.
2912          */
2913         if (mddev->bitmap == NULL &&
2914             mddev->recovery_cp == MaxSector &&
2915             mddev->reshape_position == MaxSector &&
2916             !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2917             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2918             !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2919             conf->fullsync == 0) {
2920                 *skipped = 1;
2921                 return mddev->dev_sectors - sector_nr;
2922         }
2923
2924  skipped:
2925         max_sector = mddev->dev_sectors;
2926         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2927             test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2928                 max_sector = mddev->resync_max_sectors;
2929         if (sector_nr >= max_sector) {
2930                 /* If we aborted, we need to abort the
2931                  * sync on the 'current' bitmap chucks (there can
2932                  * be several when recovering multiple devices).
2933                  * as we may have started syncing it but not finished.
2934                  * We can find the current address in
2935                  * mddev->curr_resync, but for recovery,
2936                  * we need to convert that to several
2937                  * virtual addresses.
2938                  */
2939                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2940                         end_reshape(conf);
2941                         close_sync(conf);
2942                         return 0;
2943                 }
2944
2945                 if (mddev->curr_resync < max_sector) { /* aborted */
2946                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2947                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2948                                                 &sync_blocks, 1);
2949                         else for (i = 0; i < conf->geo.raid_disks; i++) {
2950                                 sector_t sect =
2951                                         raid10_find_virt(conf, mddev->curr_resync, i);
2952                                 bitmap_end_sync(mddev->bitmap, sect,
2953                                                 &sync_blocks, 1);
2954                         }
2955                 } else {
2956                         /* completed sync */
2957                         if ((!mddev->bitmap || conf->fullsync)
2958                             && conf->have_replacement
2959                             && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2960                                 /* Completed a full sync so the replacements
2961                                  * are now fully recovered.
2962                                  */
2963                                 for (i = 0; i < conf->geo.raid_disks; i++)
2964                                         if (conf->mirrors[i].replacement)
2965                                                 conf->mirrors[i].replacement
2966                                                         ->recovery_offset
2967                                                         = MaxSector;
2968                         }
2969                         conf->fullsync = 0;
2970                 }
2971                 bitmap_close_sync(mddev->bitmap);
2972                 close_sync(conf);
2973                 *skipped = 1;
2974                 return sectors_skipped;
2975         }
2976
2977         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2978                 return reshape_request(mddev, sector_nr, skipped);
2979
2980         if (chunks_skipped >= conf->geo.raid_disks) {
2981                 /* if there has been nothing to do on any drive,
2982                  * then there is nothing to do at all..
2983                  */
2984                 *skipped = 1;
2985                 return (max_sector - sector_nr) + sectors_skipped;
2986         }
2987
2988         if (max_sector > mddev->resync_max)
2989                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2990
2991         /* make sure whole request will fit in a chunk - if chunks
2992          * are meaningful
2993          */
2994         if (conf->geo.near_copies < conf->geo.raid_disks &&
2995             max_sector > (sector_nr | chunk_mask))
2996                 max_sector = (sector_nr | chunk_mask) + 1;
2997
2998         /* Again, very different code for resync and recovery.
2999          * Both must result in an r10bio with a list of bios that
3000          * have bi_end_io, bi_sector, bi_bdev set,
3001          * and bi_private set to the r10bio.
3002          * For recovery, we may actually create several r10bios
3003          * with 2 bios in each, that correspond to the bios in the main one.
3004          * In this case, the subordinate r10bios link back through a
3005          * borrowed master_bio pointer, and the counter in the master
3006          * includes a ref from each subordinate.
3007          */
3008         /* First, we decide what to do and set ->bi_end_io
3009          * To end_sync_read if we want to read, and
3010          * end_sync_write if we will want to write.
3011          */
3012
3013         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3014         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3015                 /* recovery... the complicated one */
3016                 int j;
3017                 r10_bio = NULL;
3018
3019                 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3020                         int still_degraded;
3021                         struct r10bio *rb2;
3022                         sector_t sect;
3023                         int must_sync;
3024                         int any_working;
3025                         struct raid10_info *mirror = &conf->mirrors[i];
3026
3027                         if ((mirror->rdev == NULL ||
3028                              test_bit(In_sync, &mirror->rdev->flags))
3029                             &&
3030                             (mirror->replacement == NULL ||
3031                              test_bit(Faulty,
3032                                       &mirror->replacement->flags)))
3033                                 continue;
3034
3035                         still_degraded = 0;
3036                         /* want to reconstruct this device */
3037                         rb2 = r10_bio;
3038                         sect = raid10_find_virt(conf, sector_nr, i);
3039                         if (sect >= mddev->resync_max_sectors) {
3040                                 /* last stripe is not complete - don't
3041                                  * try to recover this sector.
3042                                  */
3043                                 continue;
3044                         }
3045                         /* Unless we are doing a full sync, or a replacement
3046                          * we only need to recover the block if it is set in
3047                          * the bitmap
3048                          */
3049                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
3050                                                       &sync_blocks, 1);
3051                         if (sync_blocks < max_sync)
3052                                 max_sync = sync_blocks;
3053                         if (!must_sync &&
3054                             mirror->replacement == NULL &&
3055                             !conf->fullsync) {
3056                                 /* yep, skip the sync_blocks here, but don't assume
3057                                  * that there will never be anything to do here
3058                                  */
3059                                 chunks_skipped = -1;
3060                                 continue;
3061                         }
3062
3063                         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3064                         r10_bio->state = 0;
3065                         raise_barrier(conf, rb2 != NULL);
3066                         atomic_set(&r10_bio->remaining, 0);
3067
3068                         r10_bio->master_bio = (struct bio*)rb2;
3069                         if (rb2)
3070                                 atomic_inc(&rb2->remaining);
3071                         r10_bio->mddev = mddev;
3072                         set_bit(R10BIO_IsRecover, &r10_bio->state);
3073                         r10_bio->sector = sect;
3074
3075                         raid10_find_phys(conf, r10_bio);
3076
3077                         /* Need to check if the array will still be
3078                          * degraded
3079                          */
3080                         for (j = 0; j < conf->geo.raid_disks; j++)
3081                                 if (conf->mirrors[j].rdev == NULL ||
3082                                     test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
3083                                         still_degraded = 1;
3084                                         break;
3085                                 }
3086
3087                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
3088                                                       &sync_blocks, still_degraded);
3089
3090                         any_working = 0;
3091                         for (j=0; j<conf->copies;j++) {
3092                                 int k;
3093                                 int d = r10_bio->devs[j].devnum;
3094                                 sector_t from_addr, to_addr;
3095                                 struct md_rdev *rdev;
3096                                 sector_t sector, first_bad;
3097                                 int bad_sectors;
3098                                 if (!conf->mirrors[d].rdev ||
3099                                     !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3100                                         continue;
3101                                 /* This is where we read from */
3102                                 any_working = 1;
3103                                 rdev = conf->mirrors[d].rdev;
3104                                 sector = r10_bio->devs[j].addr;
3105
3106                                 if (is_badblock(rdev, sector, max_sync,
3107                                                 &first_bad, &bad_sectors)) {
3108                                         if (first_bad > sector)
3109                                                 max_sync = first_bad - sector;
3110                                         else {
3111                                                 bad_sectors -= (sector
3112                                                                 - first_bad);
3113                                                 if (max_sync > bad_sectors)
3114                                                         max_sync = bad_sectors;
3115                                                 continue;
3116                                         }
3117                                 }
3118                                 bio = r10_bio->devs[0].bio;
3119                                 bio_reset(bio);
3120                                 bio->bi_next = biolist;
3121                                 biolist = bio;
3122                                 bio->bi_private = r10_bio;
3123                                 bio->bi_end_io = end_sync_read;
3124                                 bio->bi_rw = READ;
3125                                 from_addr = r10_bio->devs[j].addr;
3126                                 bio->bi_iter.bi_sector = from_addr +
3127                                         rdev->data_offset;
3128                                 bio->bi_bdev = rdev->bdev;
3129                                 atomic_inc(&rdev->nr_pending);
3130                                 /* and we write to 'i' (if not in_sync) */
3131
3132                                 for (k=0; k<conf->copies; k++)
3133                                         if (r10_bio->devs[k].devnum == i)
3134                                                 break;
3135                                 BUG_ON(k == conf->copies);
3136                                 to_addr = r10_bio->devs[k].addr;
3137                                 r10_bio->devs[0].devnum = d;
3138                                 r10_bio->devs[0].addr = from_addr;
3139                                 r10_bio->devs[1].devnum = i;
3140                                 r10_bio->devs[1].addr = to_addr;
3141
3142                                 rdev = mirror->rdev;
3143                                 if (!test_bit(In_sync, &rdev->flags)) {
3144                                         bio = r10_bio->devs[1].bio;
3145                                         bio_reset(bio);
3146                                         bio->bi_next = biolist;
3147                                         biolist = bio;
3148                                         bio->bi_private = r10_bio;
3149                                         bio->bi_end_io = end_sync_write;
3150                                         bio->bi_rw = WRITE;
3151                                         bio->bi_iter.bi_sector = to_addr
3152                                                 + rdev->data_offset;
3153                                         bio->bi_bdev = rdev->bdev;
3154                                         atomic_inc(&r10_bio->remaining);
3155                                 } else
3156                                         r10_bio->devs[1].bio->bi_end_io = NULL;
3157
3158                                 /* and maybe write to replacement */
3159                                 bio = r10_bio->devs[1].repl_bio;
3160                                 if (bio)
3161                                         bio->bi_end_io = NULL;
3162                                 rdev = mirror->replacement;
3163                                 /* Note: if rdev != NULL, then bio
3164                                  * cannot be NULL as r10buf_pool_alloc will
3165                                  * have allocated it.
3166                                  * So the second test here is pointless.
3167                                  * But it keeps semantic-checkers happy, and
3168                                  * this comment keeps human reviewers
3169                                  * happy.
3170                                  */
3171                                 if (rdev == NULL || bio == NULL ||
3172                                     test_bit(Faulty, &rdev->flags))
3173                                         break;
3174                                 bio_reset(bio);
3175                                 bio->bi_next = biolist;
3176                                 biolist = bio;
3177                                 bio->bi_private = r10_bio;
3178                                 bio->bi_end_io = end_sync_write;
3179                                 bio->bi_rw = WRITE;
3180                                 bio->bi_iter.bi_sector = to_addr +
3181                                         rdev->data_offset;
3182                                 bio->bi_bdev = rdev->bdev;
3183                                 atomic_inc(&r10_bio->remaining);
3184                                 break;
3185                         }
3186                         if (j == conf->copies) {
3187                                 /* Cannot recover, so abort the recovery or
3188                                  * record a bad block */
3189                                 if (any_working) {
3190                                         /* problem is that there are bad blocks
3191                                          * on other device(s)
3192                                          */
3193                                         int k;
3194                                         for (k = 0; k < conf->copies; k++)
3195                                                 if (r10_bio->devs[k].devnum == i)
3196                                                         break;
3197                                         if (!test_bit(In_sync,
3198                                                       &mirror->rdev->flags)
3199                                             && !rdev_set_badblocks(
3200                                                     mirror->rdev,
3201                                                     r10_bio->devs[k].addr,
3202                                                     max_sync, 0))
3203                                                 any_working = 0;
3204                                         if (mirror->replacement &&
3205                                             !rdev_set_badblocks(
3206                                                     mirror->replacement,
3207                                                     r10_bio->devs[k].addr,
3208                                                     max_sync, 0))
3209                                                 any_working = 0;
3210                                 }
3211                                 if (!any_working)  {
3212                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
3213                                                               &mddev->recovery))
3214                                                 printk(KERN_INFO "md/raid10:%s: insufficient "
3215                                                        "working devices for recovery.\n",
3216                                                        mdname(mddev));
3217                                         mirror->recovery_disabled
3218                                                 = mddev->recovery_disabled;
3219                                 }
3220                                 put_buf(r10_bio);
3221                                 if (rb2)
3222                                         atomic_dec(&rb2->remaining);
3223                                 r10_bio = rb2;
3224                                 break;
3225                         }
3226                 }
3227                 if (biolist == NULL) {
3228                         while (r10_bio) {
3229                                 struct r10bio *rb2 = r10_bio;
3230                                 r10_bio = (struct r10bio*) rb2->master_bio;
3231                                 rb2->master_bio = NULL;
3232                                 put_buf(rb2);
3233                         }
3234                         goto giveup;
3235                 }
3236         } else {
3237                 /* resync. Schedule a read for every block at this virt offset */
3238                 int count = 0;
3239
3240                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3241
3242                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3243                                        &sync_blocks, mddev->degraded) &&
3244                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3245                                                  &mddev->recovery)) {
3246                         /* We can skip this block */
3247                         *skipped = 1;
3248                         return sync_blocks + sectors_skipped;
3249                 }
3250                 if (sync_blocks < max_sync)
3251                         max_sync = sync_blocks;
3252                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3253                 r10_bio->state = 0;
3254
3255                 r10_bio->mddev = mddev;
3256                 atomic_set(&r10_bio->remaining, 0);
3257                 raise_barrier(conf, 0);
3258                 conf->next_resync = sector_nr;
3259
3260                 r10_bio->master_bio = NULL;
3261                 r10_bio->sector = sector_nr;
3262                 set_bit(R10BIO_IsSync, &r10_bio->state);
3263                 raid10_find_phys(conf, r10_bio);
3264                 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3265
3266                 for (i = 0; i < conf->copies; i++) {
3267                         int d = r10_bio->devs[i].devnum;
3268                         sector_t first_bad, sector;
3269                         int bad_sectors;
3270
3271                         if (r10_bio->devs[i].repl_bio)
3272                                 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3273
3274                         bio = r10_bio->devs[i].bio;
3275                         bio_reset(bio);
3276                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
3277                         if (conf->mirrors[d].rdev == NULL ||
3278                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3279                                 continue;
3280                         sector = r10_bio->devs[i].addr;
3281                         if (is_badblock(conf->mirrors[d].rdev,
3282                                         sector, max_sync,
3283                                         &first_bad, &bad_sectors)) {
3284                                 if (first_bad > sector)
3285                                         max_sync = first_bad - sector;
3286                                 else {
3287                                         bad_sectors -= (sector - first_bad);
3288                                         if (max_sync > bad_sectors)
3289                                                 max_sync = bad_sectors;
3290                                         continue;
3291                                 }
3292                         }
3293                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3294                         atomic_inc(&r10_bio->remaining);
3295                         bio->bi_next = biolist;
3296                         biolist = bio;
3297                         bio->bi_private = r10_bio;
3298                         bio->bi_end_io = end_sync_read;
3299                         bio->bi_rw = READ;
3300                         bio->bi_iter.bi_sector = sector +
3301                                 conf->mirrors[d].rdev->data_offset;
3302                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3303                         count++;
3304
3305                         if (conf->mirrors[d].replacement == NULL ||
3306                             test_bit(Faulty,
3307                                      &conf->mirrors[d].replacement->flags))
3308                                 continue;
3309
3310                         /* Need to set up for writing to the replacement */
3311                         bio = r10_bio->devs[i].repl_bio;
3312                         bio_reset(bio);
3313                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
3314
3315                         sector = r10_bio->devs[i].addr;
3316                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3317                         bio->bi_next = biolist;
3318                         biolist = bio;
3319                         bio->bi_private = r10_bio;
3320                         bio->bi_end_io = end_sync_write;
3321                         bio->bi_rw = WRITE;
3322                         bio->bi_iter.bi_sector = sector +
3323                                 conf->mirrors[d].replacement->data_offset;
3324                         bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3325                         count++;
3326                 }
3327
3328                 if (count < 2) {
3329                         for (i=0; i<conf->copies; i++) {
3330                                 int d = r10_bio->devs[i].devnum;
3331                                 if (r10_bio->devs[i].bio->bi_end_io)
3332                                         rdev_dec_pending(conf->mirrors[d].rdev,
3333                                                          mddev);
3334                                 if (r10_bio->devs[i].repl_bio &&
3335                                     r10_bio->devs[i].repl_bio->bi_end_io)
3336                                         rdev_dec_pending(
3337                                                 conf->mirrors[d].replacement,
3338                                                 mddev);
3339                         }
3340                         put_buf(r10_bio);
3341                         biolist = NULL;
3342                         goto giveup;
3343                 }
3344         }
3345
3346         nr_sectors = 0;
3347         if (sector_nr + max_sync < max_sector)
3348                 max_sector = sector_nr + max_sync;
3349         do {
3350                 struct page *page;
3351                 int len = PAGE_SIZE;
3352                 if (sector_nr + (len>>9) > max_sector)
3353                         len = (max_sector - sector_nr) << 9;
3354                 if (len == 0)
3355                         break;
3356                 for (bio= biolist ; bio ; bio=bio->bi_next) {
3357                         struct bio *bio2;
3358                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3359                         if (bio_add_page(bio, page, len, 0))
3360                                 continue;
3361
3362                         /* stop here */
3363                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3364                         for (bio2 = biolist;
3365                              bio2 && bio2 != bio;
3366                              bio2 = bio2->bi_next) {
3367                                 /* remove last page from this bio */
3368                                 bio2->bi_vcnt--;
3369                                 bio2->bi_iter.bi_size -= len;
3370                                 __clear_bit(BIO_SEG_VALID, &bio2->bi_flags);
3371                         }
3372                         goto bio_full;
3373                 }
3374                 nr_sectors += len>>9;
3375                 sector_nr += len>>9;
3376         } while (biolist->bi_vcnt < RESYNC_PAGES);
3377  bio_full:
3378         r10_bio->sectors = nr_sectors;
3379
3380         while (biolist) {
3381                 bio = biolist;
3382                 biolist = biolist->bi_next;
3383
3384                 bio->bi_next = NULL;
3385                 r10_bio = bio->bi_private;
3386                 r10_bio->sectors = nr_sectors;
3387
3388                 if (bio->bi_end_io == end_sync_read) {
3389                         md_sync_acct(bio->bi_bdev, nr_sectors);
3390                         set_bit(BIO_UPTODATE, &bio->bi_flags);
3391                         generic_make_request(bio);
3392                 }
3393         }
3394
3395         if (sectors_skipped)
3396                 /* pretend they weren't skipped, it makes
3397                  * no important difference in this case
3398                  */
3399                 md_done_sync(mddev, sectors_skipped, 1);
3400
3401         return sectors_skipped + nr_sectors;
3402  giveup:
3403         /* There is nowhere to write, so all non-sync
3404          * drives must be failed or in resync, all drives
3405          * have a bad block, so try the next chunk...
3406          */
3407         if (sector_nr + max_sync < max_sector)
3408                 max_sector = sector_nr + max_sync;
3409
3410         sectors_skipped += (max_sector - sector_nr);
3411         chunks_skipped ++;
3412         sector_nr = max_sector;
3413         goto skipped;
3414 }
3415
3416 static sector_t
3417 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3418 {
3419         sector_t size;
3420         struct r10conf *conf = mddev->private;
3421
3422         if (!raid_disks)
3423                 raid_disks = min(conf->geo.raid_disks,
3424                                  conf->prev.raid_disks);
3425         if (!sectors)
3426                 sectors = conf->dev_sectors;
3427
3428         size = sectors >> conf->geo.chunk_shift;
3429         sector_div(size, conf->geo.far_copies);
3430         size = size * raid_disks;
3431         sector_div(size, conf->geo.near_copies);
3432
3433         return size << conf->geo.chunk_shift;
3434 }
3435
3436 static void calc_sectors(struct r10conf *conf, sector_t size)
3437 {
3438         /* Calculate the number of sectors-per-device that will
3439          * actually be used, and set conf->dev_sectors and
3440          * conf->stride
3441          */
3442
3443         size = size >> conf->geo.chunk_shift;
3444         sector_div(size, conf->geo.far_copies);
3445         size = size * conf->geo.raid_disks;
3446         sector_div(size, conf->geo.near_copies);
3447         /* 'size' is now the number of chunks in the array */
3448         /* calculate "used chunks per device" */
3449         size = size * conf->copies;
3450
3451         /* We need to round up when dividing by raid_disks to
3452          * get the stride size.
3453          */
3454         size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3455
3456         conf->dev_sectors = size << conf->geo.chunk_shift;
3457
3458         if (conf->geo.far_offset)
3459                 conf->geo.stride = 1 << conf->geo.chunk_shift;
3460         else {
3461                 sector_div(size, conf->geo.far_copies);
3462                 conf->geo.stride = size << conf->geo.chunk_shift;
3463         }
3464 }
3465
3466 enum geo_type {geo_new, geo_old, geo_start};
3467 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3468 {
3469         int nc, fc, fo;
3470         int layout, chunk, disks;
3471         switch (new) {
3472         case geo_old:
3473                 layout = mddev->layout;
3474                 chunk = mddev->chunk_sectors;
3475                 disks = mddev->raid_disks - mddev->delta_disks;
3476                 break;
3477         case geo_new:
3478                 layout = mddev->new_layout;
3479                 chunk = mddev->new_chunk_sectors;
3480                 disks = mddev->raid_disks;
3481                 break;
3482         default: /* avoid 'may be unused' warnings */
3483         case geo_start: /* new when starting reshape - raid_disks not
3484                          * updated yet. */
3485                 layout = mddev->new_layout;
3486                 chunk = mddev->new_chunk_sectors;
3487                 disks = mddev->raid_disks + mddev->delta_disks;
3488                 break;
3489         }
3490         if (layout >> 18)
3491                 return -1;
3492         if (chunk < (PAGE_SIZE >> 9) ||
3493             !is_power_of_2(chunk))
3494                 return -2;
3495         nc = layout & 255;
3496         fc = (layout >> 8) & 255;
3497         fo = layout & (1<<16);
3498         geo->raid_disks = disks;
3499         geo->near_copies = nc;
3500         geo->far_copies = fc;
3501         geo->far_offset = fo;
3502         geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks;
3503         geo->chunk_mask = chunk - 1;
3504         geo->chunk_shift = ffz(~chunk);
3505         return nc*fc;
3506 }
3507
3508 static struct r10conf *setup_conf(struct mddev *mddev)
3509 {
3510         struct r10conf *conf = NULL;
3511         int err = -EINVAL;
3512         struct geom geo;
3513         int copies;
3514
3515         copies = setup_geo(&geo, mddev, geo_new);
3516
3517         if (copies == -2) {
3518                 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3519                        "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3520                        mdname(mddev), PAGE_SIZE);
3521                 goto out;
3522         }
3523
3524         if (copies < 2 || copies > mddev->raid_disks) {
3525                 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3526                        mdname(mddev), mddev->new_layout);
3527                 goto out;
3528         }
3529
3530         err = -ENOMEM;
3531         conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3532         if (!conf)
3533                 goto out;
3534
3535         /* FIXME calc properly */
3536         conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3537                                                             max(0,-mddev->delta_disks)),
3538                                 GFP_KERNEL);
3539         if (!conf->mirrors)
3540                 goto out;
3541
3542         conf->tmppage = alloc_page(GFP_KERNEL);
3543         if (!conf->tmppage)
3544                 goto out;
3545
3546         conf->geo = geo;
3547         conf->copies = copies;
3548         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3549                                            r10bio_pool_free, conf);
3550         if (!conf->r10bio_pool)
3551                 goto out;
3552
3553         calc_sectors(conf, mddev->dev_sectors);
3554         if (mddev->reshape_position == MaxSector) {
3555                 conf->prev = conf->geo;
3556                 conf->reshape_progress = MaxSector;
3557         } else {
3558                 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3559                         err = -EINVAL;
3560                         goto out;
3561                 }
3562                 conf->reshape_progress = mddev->reshape_position;
3563                 if (conf->prev.far_offset)
3564                         conf->prev.stride = 1 << conf->prev.chunk_shift;
3565                 else
3566                         /* far_copies must be 1 */
3567                         conf->prev.stride = conf->dev_sectors;
3568         }
3569         conf->reshape_safe = conf->reshape_progress;
3570         spin_lock_init(&conf->device_lock);
3571         INIT_LIST_HEAD(&conf->retry_list);
3572
3573         spin_lock_init(&conf->resync_lock);
3574         init_waitqueue_head(&conf->wait_barrier);
3575
3576         conf->thread = md_register_thread(raid10d, mddev, "raid10");
3577         if (!conf->thread)
3578                 goto out;
3579
3580         conf->mddev = mddev;
3581         return conf;
3582
3583  out:
3584         if (err == -ENOMEM)
3585                 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3586                        mdname(mddev));
3587         if (conf) {
3588                 if (conf->r10bio_pool)
3589                         mempool_destroy(conf->r10bio_pool);
3590                 kfree(conf->mirrors);
3591                 safe_put_page(conf->tmppage);
3592                 kfree(conf);
3593         }
3594         return ERR_PTR(err);
3595 }
3596
3597 static int run(struct mddev *mddev)
3598 {
3599         struct r10conf *conf;
3600         int i, disk_idx, chunk_size;
3601         struct raid10_info *disk;
3602         struct md_rdev *rdev;
3603         sector_t size;
3604         sector_t min_offset_diff = 0;
3605         int first = 1;
3606         bool discard_supported = false;
3607
3608         if (mddev->private == NULL) {
3609                 conf = setup_conf(mddev);
3610                 if (IS_ERR(conf))
3611                         return PTR_ERR(conf);
3612                 mddev->private = conf;
3613         }
3614         conf = mddev->private;
3615         if (!conf)
3616                 goto out;
3617
3618         mddev->thread = conf->thread;
3619         conf->thread = NULL;
3620
3621         chunk_size = mddev->chunk_sectors << 9;
3622         if (mddev->queue) {
3623                 blk_queue_max_discard_sectors(mddev->queue,
3624                                               mddev->chunk_sectors);
3625                 blk_queue_max_write_same_sectors(mddev->queue, 0);
3626                 blk_queue_io_min(mddev->queue, chunk_size);
3627                 if (conf->geo.raid_disks % conf->geo.near_copies)
3628                         blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3629                 else
3630                         blk_queue_io_opt(mddev->queue, chunk_size *
3631                                          (conf->geo.raid_disks / conf->geo.near_copies));
3632         }
3633
3634         rdev_for_each(rdev, mddev) {
3635                 long long diff;
3636                 struct request_queue *q;
3637
3638                 disk_idx = rdev->raid_disk;
3639                 if (disk_idx < 0)
3640                         continue;
3641                 if (disk_idx >= conf->geo.raid_disks &&
3642                     disk_idx >= conf->prev.raid_disks)
3643                         continue;
3644                 disk = conf->mirrors + disk_idx;
3645
3646                 if (test_bit(Replacement, &rdev->flags)) {
3647                         if (disk->replacement)
3648                                 goto out_free_conf;
3649                         disk->replacement = rdev;
3650                 } else {
3651                         if (disk->rdev)
3652                                 goto out_free_conf;
3653                         disk->rdev = rdev;
3654                 }
3655                 q = bdev_get_queue(rdev->bdev);
3656                 if (q->merge_bvec_fn)
3657                         mddev->merge_check_needed = 1;
3658                 diff = (rdev->new_data_offset - rdev->data_offset);
3659                 if (!mddev->reshape_backwards)
3660                         diff = -diff;
3661                 if (diff < 0)
3662                         diff = 0;
3663                 if (first || diff < min_offset_diff)
3664                         min_offset_diff = diff;
3665
3666                 if (mddev->gendisk)
3667                         disk_stack_limits(mddev->gendisk, rdev->bdev,
3668                                           rdev->data_offset << 9);
3669
3670                 disk->head_position = 0;
3671
3672                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3673                         discard_supported = true;
3674         }
3675
3676         if (mddev->queue) {
3677                 if (discard_supported)
3678                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3679                                                 mddev->queue);
3680                 else
3681                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3682                                                   mddev->queue);
3683         }
3684         /* need to check that every block has at least one working mirror */
3685         if (!enough(conf, -1)) {
3686                 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3687                        mdname(mddev));
3688                 goto out_free_conf;
3689         }
3690
3691         if (conf->reshape_progress != MaxSector) {
3692                 /* must ensure that shape change is supported */
3693                 if (conf->geo.far_copies != 1 &&
3694                     conf->geo.far_offset == 0)
3695                         goto out_free_conf;
3696                 if (conf->prev.far_copies != 1 &&
3697                     conf->prev.far_offset == 0)
3698                         goto out_free_conf;
3699         }
3700
3701         mddev->degraded = 0;
3702         for (i = 0;
3703              i < conf->geo.raid_disks
3704                      || i < conf->prev.raid_disks;
3705              i++) {
3706
3707                 disk = conf->mirrors + i;
3708
3709                 if (!disk->rdev && disk->replacement) {
3710                         /* The replacement is all we have - use it */
3711                         disk->rdev = disk->replacement;
3712                         disk->replacement = NULL;
3713                         clear_bit(Replacement, &disk->rdev->flags);
3714                 }
3715
3716                 if (!disk->rdev ||
3717                     !test_bit(In_sync, &disk->rdev->flags)) {
3718                         disk->head_position = 0;
3719                         mddev->degraded++;
3720                         if (disk->rdev &&
3721                             disk->rdev->saved_raid_disk < 0)
3722                                 conf->fullsync = 1;
3723                 }
3724                 disk->recovery_disabled = mddev->recovery_disabled - 1;
3725         }
3726
3727         if (mddev->recovery_cp != MaxSector)
3728                 printk(KERN_NOTICE "md/raid10:%s: not clean"
3729                        " -- starting background reconstruction\n",
3730                        mdname(mddev));
3731         printk(KERN_INFO
3732                 "md/raid10:%s: active with %d out of %d devices\n",
3733                 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3734                 conf->geo.raid_disks);
3735         /*
3736          * Ok, everything is just fine now
3737          */
3738         mddev->dev_sectors = conf->dev_sectors;
3739         size = raid10_size(mddev, 0, 0);
3740         md_set_array_sectors(mddev, size);
3741         mddev->resync_max_sectors = size;
3742
3743         if (mddev->queue) {
3744                 int stripe = conf->geo.raid_disks *
3745                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3746
3747                 /* Calculate max read-ahead size.
3748                  * We need to readahead at least twice a whole stripe....
3749                  * maybe...
3750                  */
3751                 stripe /= conf->geo.near_copies;
3752                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3753                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3754         }
3755
3756         if (md_integrity_register(mddev))
3757                 goto out_free_conf;
3758
3759         if (conf->reshape_progress != MaxSector) {
3760                 unsigned long before_length, after_length;
3761
3762                 before_length = ((1 << conf->prev.chunk_shift) *
3763                                  conf->prev.far_copies);
3764                 after_length = ((1 << conf->geo.chunk_shift) *
3765                                 conf->geo.far_copies);
3766
3767                 if (max(before_length, after_length) > min_offset_diff) {
3768                         /* This cannot work */
3769                         printk("md/raid10: offset difference not enough to continue reshape\n");
3770                         goto out_free_conf;
3771                 }
3772                 conf->offset_diff = min_offset_diff;
3773
3774                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3775                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3776                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3777                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3778                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3779                                                         "reshape");
3780         }
3781
3782         return 0;
3783
3784 out_free_conf:
3785         md_unregister_thread(&mddev->thread);
3786         if (conf->r10bio_pool)
3787                 mempool_destroy(conf->r10bio_pool);
3788         safe_put_page(conf->tmppage);
3789         kfree(conf->mirrors);
3790         kfree(conf);
3791         mddev->private = NULL;
3792 out:
3793         return -EIO;
3794 }
3795
3796 static void raid10_free(struct mddev *mddev, void *priv)
3797 {
3798         struct r10conf *conf = priv;
3799
3800         if (conf->r10bio_pool)
3801                 mempool_destroy(conf->r10bio_pool);
3802         safe_put_page(conf->tmppage);
3803         kfree(conf->mirrors);
3804         kfree(conf->mirrors_old);
3805         kfree(conf->mirrors_new);
3806         kfree(conf);
3807 }
3808
3809 static void raid10_quiesce(struct mddev *mddev, int state)
3810 {
3811         struct r10conf *conf = mddev->private;
3812
3813         switch(state) {
3814         case 1:
3815                 raise_barrier(conf, 0);
3816                 break;
3817         case 0:
3818                 lower_barrier(conf);
3819                 break;
3820         }
3821 }
3822
3823 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3824 {
3825         /* Resize of 'far' arrays is not supported.
3826          * For 'near' and 'offset' arrays we can set the
3827          * number of sectors used to be an appropriate multiple
3828          * of the chunk size.
3829          * For 'offset', this is far_copies*chunksize.
3830          * For 'near' the multiplier is the LCM of
3831          * near_copies and raid_disks.
3832          * So if far_copies > 1 && !far_offset, fail.
3833          * Else find LCM(raid_disks, near_copy)*far_copies and
3834          * multiply by chunk_size.  Then round to this number.
3835          * This is mostly done by raid10_size()
3836          */
3837         struct r10conf *conf = mddev->private;
3838         sector_t oldsize, size;
3839
3840         if (mddev->reshape_position != MaxSector)
3841                 return -EBUSY;
3842
3843         if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3844                 return -EINVAL;
3845
3846         oldsize = raid10_size(mddev, 0, 0);
3847         size = raid10_size(mddev, sectors, 0);
3848         if (mddev->external_size &&
3849             mddev->array_sectors > size)
3850                 return -EINVAL;
3851         if (mddev->bitmap) {
3852                 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3853                 if (ret)
3854                         return ret;
3855         }
3856         md_set_array_sectors(mddev, size);
3857         set_capacity(mddev->gendisk, mddev->array_sectors);
3858         revalidate_disk(mddev->gendisk);
3859         if (sectors > mddev->dev_sectors &&
3860             mddev->recovery_cp > oldsize) {
3861                 mddev->recovery_cp = oldsize;
3862                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3863         }
3864         calc_sectors(conf, sectors);
3865         mddev->dev_sectors = conf->dev_sectors;
3866         mddev->resync_max_sectors = size;
3867         return 0;
3868 }
3869
3870 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3871 {
3872         struct md_rdev *rdev;
3873         struct r10conf *conf;
3874
3875         if (mddev->degraded > 0) {
3876                 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3877                        mdname(mddev));
3878                 return ERR_PTR(-EINVAL);
3879         }
3880         sector_div(size, devs);
3881
3882         /* Set new parameters */
3883         mddev->new_level = 10;
3884         /* new layout: far_copies = 1, near_copies = 2 */
3885         mddev->new_layout = (1<<8) + 2;
3886         mddev->new_chunk_sectors = mddev->chunk_sectors;
3887         mddev->delta_disks = mddev->raid_disks;
3888         mddev->raid_disks *= 2;
3889         /* make sure it will be not marked as dirty */
3890         mddev->recovery_cp = MaxSector;
3891         mddev->dev_sectors = size;
3892
3893         conf = setup_conf(mddev);
3894         if (!IS_ERR(conf)) {
3895                 rdev_for_each(rdev, mddev)
3896                         if (rdev->raid_disk >= 0) {
3897                                 rdev->new_raid_disk = rdev->raid_disk * 2;
3898                                 rdev->sectors = size;
3899                         }
3900                 conf->barrier = 1;
3901         }
3902
3903         return conf;
3904 }
3905
3906 static void *raid10_takeover(struct mddev *mddev)
3907 {
3908         struct r0conf *raid0_conf;
3909
3910         /* raid10 can take over:
3911          *  raid0 - providing it has only two drives
3912          */
3913         if (mddev->level == 0) {
3914                 /* for raid0 takeover only one zone is supported */
3915                 raid0_conf = mddev->private;
3916                 if (raid0_conf->nr_strip_zones > 1) {
3917                         printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3918                                " with more than one zone.\n",
3919                                mdname(mddev));
3920                         return ERR_PTR(-EINVAL);
3921                 }
3922                 return raid10_takeover_raid0(mddev,
3923                         raid0_conf->strip_zone->zone_end,
3924                         raid0_conf->strip_zone->nb_dev);
3925         }
3926         return ERR_PTR(-EINVAL);
3927 }
3928
3929 static int raid10_check_reshape(struct mddev *mddev)
3930 {
3931         /* Called when there is a request to change
3932          * - layout (to ->new_layout)
3933          * - chunk size (to ->new_chunk_sectors)
3934          * - raid_disks (by delta_disks)
3935          * or when trying to restart a reshape that was ongoing.
3936          *
3937          * We need to validate the request and possibly allocate
3938          * space if that might be an issue later.
3939          *
3940          * Currently we reject any reshape of a 'far' mode array,
3941          * allow chunk size to change if new is generally acceptable,
3942          * allow raid_disks to increase, and allow
3943          * a switch between 'near' mode and 'offset' mode.
3944          */
3945         struct r10conf *conf = mddev->private;
3946         struct geom geo;
3947
3948         if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3949                 return -EINVAL;
3950
3951         if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3952                 /* mustn't change number of copies */
3953                 return -EINVAL;
3954         if (geo.far_copies > 1 && !geo.far_offset)
3955                 /* Cannot switch to 'far' mode */
3956                 return -EINVAL;
3957
3958         if (mddev->array_sectors & geo.chunk_mask)
3959                         /* not factor of array size */
3960                         return -EINVAL;
3961
3962         if (!enough(conf, -1))
3963                 return -EINVAL;
3964
3965         kfree(conf->mirrors_new);
3966         conf->mirrors_new = NULL;
3967         if (mddev->delta_disks > 0) {
3968                 /* allocate new 'mirrors' list */
3969                 conf->mirrors_new = kzalloc(
3970                         sizeof(struct raid10_info)
3971                         *(mddev->raid_disks +
3972                           mddev->delta_disks),
3973                         GFP_KERNEL);
3974                 if (!conf->mirrors_new)
3975                         return -ENOMEM;
3976         }
3977         return 0;
3978 }
3979
3980 /*
3981  * Need to check if array has failed when deciding whether to:
3982  *  - start an array
3983  *  - remove non-faulty devices
3984  *  - add a spare
3985  *  - allow a reshape
3986  * This determination is simple when no reshape is happening.
3987  * However if there is a reshape, we need to carefully check
3988  * both the before and after sections.
3989  * This is because some failed devices may only affect one
3990  * of the two sections, and some non-in_sync devices may
3991  * be insync in the section most affected by failed devices.
3992  */
3993 static int calc_degraded(struct r10conf *conf)
3994 {
3995         int degraded, degraded2;
3996         int i;
3997
3998         rcu_read_lock();
3999         degraded = 0;
4000         /* 'prev' section first */
4001         for (i = 0; i < conf->prev.raid_disks; i++) {
4002                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4003                 if (!rdev || test_bit(Faulty, &rdev->flags))
4004                         degraded++;
4005                 else if (!test_bit(In_sync, &rdev->flags))
4006                         /* When we can reduce the number of devices in
4007                          * an array, this might not contribute to
4008                          * 'degraded'.  It does now.
4009                          */
4010                         degraded++;
4011         }
4012         rcu_read_unlock();
4013         if (conf->geo.raid_disks == conf->prev.raid_disks)
4014                 return degraded;
4015         rcu_read_lock();
4016         degraded2 = 0;
4017         for (i = 0; i < conf->geo.raid_disks; i++) {
4018                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4019                 if (!rdev || test_bit(Faulty, &rdev->flags))
4020                         degraded2++;
4021                 else if (!test_bit(In_sync, &rdev->flags)) {
4022                         /* If reshape is increasing the number of devices,
4023                          * this section has already been recovered, so
4024                          * it doesn't contribute to degraded.
4025                          * else it does.
4026                          */
4027                         if (conf->geo.raid_disks <= conf->prev.raid_disks)
4028                                 degraded2++;
4029                 }
4030         }
4031         rcu_read_unlock();
4032         if (degraded2 > degraded)
4033                 return degraded2;
4034         return degraded;
4035 }
4036
4037 static int raid10_start_reshape(struct mddev *mddev)
4038 {
4039         /* A 'reshape' has been requested. This commits
4040          * the various 'new' fields and sets MD_RECOVER_RESHAPE
4041          * This also checks if there are enough spares and adds them
4042          * to the array.
4043          * We currently require enough spares to make the final
4044          * array non-degraded.  We also require that the difference
4045          * between old and new data_offset - on each device - is
4046          * enough that we never risk over-writing.
4047          */
4048
4049         unsigned long before_length, after_length;
4050         sector_t min_offset_diff = 0;
4051         int first = 1;
4052         struct geom new;
4053         struct r10conf *conf = mddev->private;
4054         struct md_rdev *rdev;
4055         int spares = 0;
4056         int ret;
4057
4058         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4059                 return -EBUSY;
4060
4061         if (setup_geo(&new, mddev, geo_start) != conf->copies)
4062                 return -EINVAL;
4063
4064         before_length = ((1 << conf->prev.chunk_shift) *
4065                          conf->prev.far_copies);
4066         after_length = ((1 << conf->geo.chunk_shift) *
4067                         conf->geo.far_copies);
4068
4069         rdev_for_each(rdev, mddev) {
4070                 if (!test_bit(In_sync, &rdev->flags)
4071                     && !test_bit(Faulty, &rdev->flags))
4072                         spares++;
4073                 if (rdev->raid_disk >= 0) {
4074                         long long diff = (rdev->new_data_offset
4075                                           - rdev->data_offset);
4076                         if (!mddev->reshape_backwards)
4077                                 diff = -diff;
4078                         if (diff < 0)
4079                                 diff = 0;
4080                         if (first || diff < min_offset_diff)
4081                                 min_offset_diff = diff;
4082                 }
4083         }
4084
4085         if (max(before_length, after_length) > min_offset_diff)
4086                 return -EINVAL;
4087
4088         if (spares < mddev->delta_disks)
4089                 return -EINVAL;
4090
4091         conf->offset_diff = min_offset_diff;
4092         spin_lock_irq(&conf->device_lock);
4093         if (conf->mirrors_new) {
4094                 memcpy(conf->mirrors_new, conf->mirrors,
4095                        sizeof(struct raid10_info)*conf->prev.raid_disks);
4096                 smp_mb();
4097                 kfree(conf->mirrors_old);
4098                 conf->mirrors_old = conf->mirrors;
4099                 conf->mirrors = conf->mirrors_new;
4100                 conf->mirrors_new = NULL;
4101         }
4102         setup_geo(&conf->geo, mddev, geo_start);
4103         smp_mb();
4104         if (mddev->reshape_backwards) {
4105                 sector_t size = raid10_size(mddev, 0, 0);
4106                 if (size < mddev->array_sectors) {
4107                         spin_unlock_irq(&conf->device_lock);
4108                         printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4109                                mdname(mddev));
4110                         return -EINVAL;
4111                 }
4112                 mddev->resync_max_sectors = size;
4113                 conf->reshape_progress = size;
4114         } else
4115                 conf->reshape_progress = 0;
4116         conf->reshape_safe = conf->reshape_progress;
4117         spin_unlock_irq(&conf->device_lock);
4118
4119         if (mddev->delta_disks && mddev->bitmap) {
4120                 ret = bitmap_resize(mddev->bitmap,
4121                                     raid10_size(mddev, 0,
4122                                                 conf->geo.raid_disks),
4123                                     0, 0);
4124                 if (ret)
4125                         goto abort;
4126         }
4127         if (mddev->delta_disks > 0) {
4128                 rdev_for_each(rdev, mddev)
4129                         if (rdev->raid_disk < 0 &&
4130                             !test_bit(Faulty, &rdev->flags)) {
4131                                 if (raid10_add_disk(mddev, rdev) == 0) {
4132                                         if (rdev->raid_disk >=
4133                                             conf->prev.raid_disks)
4134                                                 set_bit(In_sync, &rdev->flags);
4135                                         else
4136                                                 rdev->recovery_offset = 0;
4137
4138                                         if (sysfs_link_rdev(mddev, rdev))
4139                                                 /* Failure here  is OK */;
4140                                 }
4141                         } else if (rdev->raid_disk >= conf->prev.raid_disks
4142                                    && !test_bit(Faulty, &rdev->flags)) {
4143                                 /* This is a spare that was manually added */
4144                                 set_bit(In_sync, &rdev->flags);
4145                         }
4146         }
4147         /* When a reshape changes the number of devices,
4148          * ->degraded is measured against the larger of the
4149          * pre and  post numbers.
4150          */
4151         spin_lock_irq(&conf->device_lock);
4152         mddev->degraded = calc_degraded(conf);
4153         spin_unlock_irq(&conf->device_lock);
4154         mddev->raid_disks = conf->geo.raid_disks;
4155         mddev->reshape_position = conf->reshape_progress;
4156         set_bit(MD_CHANGE_DEVS, &mddev->flags);
4157
4158         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4159         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4160         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4161         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4162         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4163
4164         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4165                                                 "reshape");
4166         if (!mddev->sync_thread) {
4167                 ret = -EAGAIN;
4168                 goto abort;
4169         }
4170         conf->reshape_checkpoint = jiffies;
4171         md_wakeup_thread(mddev->sync_thread);
4172         md_new_event(mddev);
4173         return 0;
4174
4175 abort:
4176         mddev->recovery = 0;
4177         spin_lock_irq(&conf->device_lock);
4178         conf->geo = conf->prev;
4179         mddev->raid_disks = conf->geo.raid_disks;
4180         rdev_for_each(rdev, mddev)
4181                 rdev->new_data_offset = rdev->data_offset;
4182         smp_wmb();
4183         conf->reshape_progress = MaxSector;
4184         conf->reshape_safe = MaxSector;
4185         mddev->reshape_position = MaxSector;
4186         spin_unlock_irq(&conf->device_lock);
4187         return ret;
4188 }
4189
4190 /* Calculate the last device-address that could contain
4191  * any block from the chunk that includes the array-address 's'
4192  * and report the next address.
4193  * i.e. the address returned will be chunk-aligned and after
4194  * any data that is in the chunk containing 's'.
4195  */
4196 static sector_t last_dev_address(sector_t s, struct geom *geo)
4197 {
4198         s = (s | geo->chunk_mask) + 1;
4199         s >>= geo->chunk_shift;
4200         s *= geo->near_copies;
4201         s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4202         s *= geo->far_copies;
4203         s <<= geo->chunk_shift;
4204         return s;
4205 }
4206
4207 /* Calculate the first device-address that could contain
4208  * any block from the chunk that includes the array-address 's'.
4209  * This too will be the start of a chunk
4210  */
4211 static sector_t first_dev_address(sector_t s, struct geom *geo)
4212 {
4213         s >>= geo->chunk_shift;
4214         s *= geo->near_copies;
4215         sector_div(s, geo->raid_disks);
4216         s *= geo->far_copies;
4217         s <<= geo->chunk_shift;
4218         return s;
4219 }
4220
4221 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4222                                 int *skipped)
4223 {
4224         /* We simply copy at most one chunk (smallest of old and new)
4225          * at a time, possibly less if that exceeds RESYNC_PAGES,
4226          * or we hit a bad block or something.
4227          * This might mean we pause for normal IO in the middle of
4228          * a chunk, but that is not a problem was mddev->reshape_position
4229          * can record any location.
4230          *
4231          * If we will want to write to a location that isn't
4232          * yet recorded as 'safe' (i.e. in metadata on disk) then
4233          * we need to flush all reshape requests and update the metadata.
4234          *
4235          * When reshaping forwards (e.g. to more devices), we interpret
4236          * 'safe' as the earliest block which might not have been copied
4237          * down yet.  We divide this by previous stripe size and multiply
4238          * by previous stripe length to get lowest device offset that we
4239          * cannot write to yet.
4240          * We interpret 'sector_nr' as an address that we want to write to.
4241          * From this we use last_device_address() to find where we might
4242          * write to, and first_device_address on the  'safe' position.
4243          * If this 'next' write position is after the 'safe' position,
4244          * we must update the metadata to increase the 'safe' position.
4245          *
4246          * When reshaping backwards, we round in the opposite direction
4247          * and perform the reverse test:  next write position must not be
4248          * less than current safe position.
4249          *
4250          * In all this the minimum difference in data offsets
4251          * (conf->offset_diff - always positive) allows a bit of slack,
4252          * so next can be after 'safe', but not by more than offset_disk
4253          *
4254          * We need to prepare all the bios here before we start any IO
4255          * to ensure the size we choose is acceptable to all devices.
4256          * The means one for each copy for write-out and an extra one for
4257          * read-in.
4258          * We store the read-in bio in ->master_bio and the others in
4259          * ->devs[x].bio and ->devs[x].repl_bio.
4260          */
4261         struct r10conf *conf = mddev->private;
4262         struct r10bio *r10_bio;
4263         sector_t next, safe, last;
4264         int max_sectors;
4265         int nr_sectors;
4266         int s;
4267         struct md_rdev *rdev;
4268         int need_flush = 0;
4269         struct bio *blist;
4270         struct bio *bio, *read_bio;
4271         int sectors_done = 0;
4272
4273         if (sector_nr == 0) {
4274                 /* If restarting in the middle, skip the initial sectors */
4275                 if (mddev->reshape_backwards &&
4276                     conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4277                         sector_nr = (raid10_size(mddev, 0, 0)
4278                                      - conf->reshape_progress);
4279                 } else if (!mddev->reshape_backwards &&
4280                            conf->reshape_progress > 0)
4281                         sector_nr = conf->reshape_progress;
4282                 if (sector_nr) {
4283                         mddev->curr_resync_completed = sector_nr;
4284                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4285                         *skipped = 1;
4286                         return sector_nr;
4287                 }
4288         }
4289
4290         /* We don't use sector_nr to track where we are up to
4291          * as that doesn't work well for ->reshape_backwards.
4292          * So just use ->reshape_progress.
4293          */
4294         if (mddev->reshape_backwards) {
4295                 /* 'next' is the earliest device address that we might
4296                  * write to for this chunk in the new layout
4297                  */
4298                 next = first_dev_address(conf->reshape_progress - 1,
4299                                          &conf->geo);
4300
4301                 /* 'safe' is the last device address that we might read from
4302                  * in the old layout after a restart
4303                  */
4304                 safe = last_dev_address(conf->reshape_safe - 1,
4305                                         &conf->prev);
4306
4307                 if (next + conf->offset_diff < safe)
4308                         need_flush = 1;
4309
4310                 last = conf->reshape_progress - 1;
4311                 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4312                                                & conf->prev.chunk_mask);
4313                 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4314                         sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4315         } else {
4316                 /* 'next' is after the last device address that we
4317                  * might write to for this chunk in the new layout
4318                  */
4319                 next = last_dev_address(conf->reshape_progress, &conf->geo);
4320
4321                 /* 'safe' is the earliest device address that we might
4322                  * read from in the old layout after a restart
4323                  */
4324                 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4325
4326                 /* Need to update metadata if 'next' might be beyond 'safe'
4327                  * as that would possibly corrupt data
4328                  */
4329                 if (next > safe + conf->offset_diff)
4330                         need_flush = 1;
4331
4332                 sector_nr = conf->reshape_progress;
4333                 last  = sector_nr | (conf->geo.chunk_mask
4334                                      & conf->prev.chunk_mask);
4335
4336                 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4337                         last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4338         }
4339
4340         if (need_flush ||
4341             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4342                 /* Need to update reshape_position in metadata */
4343                 wait_barrier(conf);
4344                 mddev->reshape_position = conf->reshape_progress;
4345                 if (mddev->reshape_backwards)
4346                         mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4347                                 - conf->reshape_progress;
4348                 else
4349                         mddev->curr_resync_completed = conf->reshape_progress;
4350                 conf->reshape_checkpoint = jiffies;
4351                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4352                 md_wakeup_thread(mddev->thread);
4353                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4354                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4355                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4356                         allow_barrier(conf);
4357                         return sectors_done;
4358                 }
4359                 conf->reshape_safe = mddev->reshape_position;
4360                 allow_barrier(conf);
4361         }
4362
4363 read_more:
4364         /* Now schedule reads for blocks from sector_nr to last */
4365         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4366         r10_bio->state = 0;
4367         raise_barrier(conf, sectors_done != 0);
4368         atomic_set(&r10_bio->remaining, 0);
4369         r10_bio->mddev = mddev;
4370         r10_bio->sector = sector_nr;
4371         set_bit(R10BIO_IsReshape, &r10_bio->state);
4372         r10_bio->sectors = last - sector_nr + 1;
4373         rdev = read_balance(conf, r10_bio, &max_sectors);
4374         BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4375
4376         if (!rdev) {
4377                 /* Cannot read from here, so need to record bad blocks
4378                  * on all the target devices.
4379                  */
4380                 // FIXME
4381                 mempool_free(r10_bio, conf->r10buf_pool);
4382                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4383                 return sectors_done;
4384         }
4385
4386         read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4387
4388         read_bio->bi_bdev = rdev->bdev;
4389         read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4390                                + rdev->data_offset);
4391         read_bio->bi_private = r10_bio;
4392         read_bio->bi_end_io = end_sync_read;
4393         read_bio->bi_rw = READ;
4394         read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4395         __set_bit(BIO_UPTODATE, &read_bio->bi_flags);
4396         read_bio->bi_vcnt = 0;
4397         read_bio->bi_iter.bi_size = 0;
4398         r10_bio->master_bio = read_bio;
4399         r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4400
4401         /* Now find the locations in the new layout */
4402         __raid10_find_phys(&conf->geo, r10_bio);
4403
4404         blist = read_bio;
4405         read_bio->bi_next = NULL;
4406
4407         for (s = 0; s < conf->copies*2; s++) {
4408                 struct bio *b;
4409                 int d = r10_bio->devs[s/2].devnum;
4410                 struct md_rdev *rdev2;
4411                 if (s&1) {
4412                         rdev2 = conf->mirrors[d].replacement;
4413                         b = r10_bio->devs[s/2].repl_bio;
4414                 } else {
4415                         rdev2 = conf->mirrors[d].rdev;
4416                         b = r10_bio->devs[s/2].bio;
4417                 }
4418                 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4419                         continue;
4420
4421                 bio_reset(b);
4422                 b->bi_bdev = rdev2->bdev;
4423                 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4424                         rdev2->new_data_offset;
4425                 b->bi_private = r10_bio;
4426                 b->bi_end_io = end_reshape_write;
4427                 b->bi_rw = WRITE;
4428                 b->bi_next = blist;
4429                 blist = b;
4430         }
4431
4432         /* Now add as many pages as possible to all of these bios. */
4433
4434         nr_sectors = 0;
4435         for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4436                 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4437                 int len = (max_sectors - s) << 9;
4438                 if (len > PAGE_SIZE)
4439                         len = PAGE_SIZE;
4440                 for (bio = blist; bio ; bio = bio->bi_next) {
4441                         struct bio *bio2;
4442                         if (bio_add_page(bio, page, len, 0))
4443                                 continue;
4444
4445                         /* Didn't fit, must stop */
4446                         for (bio2 = blist;
4447                              bio2 && bio2 != bio;
4448                              bio2 = bio2->bi_next) {
4449                                 /* Remove last page from this bio */
4450                                 bio2->bi_vcnt--;
4451                                 bio2->bi_iter.bi_size -= len;
4452                                 __clear_bit(BIO_SEG_VALID, &bio2->bi_flags);
4453                         }
4454                         goto bio_full;
4455                 }
4456                 sector_nr += len >> 9;
4457                 nr_sectors += len >> 9;
4458         }
4459 bio_full:
4460         r10_bio->sectors = nr_sectors;
4461
4462         /* Now submit the read */
4463         md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4464         atomic_inc(&r10_bio->remaining);
4465         read_bio->bi_next = NULL;
4466         generic_make_request(read_bio);
4467         sector_nr += nr_sectors;
4468         sectors_done += nr_sectors;
4469         if (sector_nr <= last)
4470                 goto read_more;
4471
4472         /* Now that we have done the whole section we can
4473          * update reshape_progress
4474          */
4475         if (mddev->reshape_backwards)
4476                 conf->reshape_progress -= sectors_done;
4477         else
4478                 conf->reshape_progress += sectors_done;
4479
4480         return sectors_done;
4481 }
4482
4483 static void end_reshape_request(struct r10bio *r10_bio);
4484 static int handle_reshape_read_error(struct mddev *mddev,
4485                                      struct r10bio *r10_bio);
4486 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4487 {
4488         /* Reshape read completed.  Hopefully we have a block
4489          * to write out.
4490          * If we got a read error then we do sync 1-page reads from
4491          * elsewhere until we find the data - or give up.
4492          */
4493         struct r10conf *conf = mddev->private;
4494         int s;
4495
4496         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4497                 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4498                         /* Reshape has been aborted */
4499                         md_done_sync(mddev, r10_bio->sectors, 0);
4500                         return;
4501                 }
4502
4503         /* We definitely have the data in the pages, schedule the
4504          * writes.
4505          */
4506         atomic_set(&r10_bio->remaining, 1);
4507         for (s = 0; s < conf->copies*2; s++) {
4508                 struct bio *b;
4509                 int d = r10_bio->devs[s/2].devnum;
4510                 struct md_rdev *rdev;
4511                 if (s&1) {
4512                         rdev = conf->mirrors[d].replacement;
4513                         b = r10_bio->devs[s/2].repl_bio;
4514                 } else {
4515                         rdev = conf->mirrors[d].rdev;
4516                         b = r10_bio->devs[s/2].bio;
4517                 }
4518                 if (!rdev || test_bit(Faulty, &rdev->flags))
4519                         continue;
4520                 atomic_inc(&rdev->nr_pending);
4521                 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4522                 atomic_inc(&r10_bio->remaining);
4523                 b->bi_next = NULL;
4524                 generic_make_request(b);
4525         }
4526         end_reshape_request(r10_bio);
4527 }
4528
4529 static void end_reshape(struct r10conf *conf)
4530 {
4531         if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4532                 return;
4533
4534         spin_lock_irq(&conf->device_lock);
4535         conf->prev = conf->geo;
4536         md_finish_reshape(conf->mddev);
4537         smp_wmb();
4538         conf->reshape_progress = MaxSector;
4539         conf->reshape_safe = MaxSector;
4540         spin_unlock_irq(&conf->device_lock);
4541
4542         /* read-ahead size must cover two whole stripes, which is
4543          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4544          */
4545         if (conf->mddev->queue) {
4546                 int stripe = conf->geo.raid_disks *
4547                         ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4548                 stripe /= conf->geo.near_copies;
4549                 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4550                         conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4551         }
4552         conf->fullsync = 0;
4553 }
4554
4555 static int handle_reshape_read_error(struct mddev *mddev,
4556                                      struct r10bio *r10_bio)
4557 {
4558         /* Use sync reads to get the blocks from somewhere else */
4559         int sectors = r10_bio->sectors;
4560         struct r10conf *conf = mddev->private;
4561         struct {
4562                 struct r10bio r10_bio;
4563                 struct r10dev devs[conf->copies];
4564         } on_stack;
4565         struct r10bio *r10b = &on_stack.r10_bio;
4566         int slot = 0;
4567         int idx = 0;
4568         struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4569
4570         r10b->sector = r10_bio->sector;
4571         __raid10_find_phys(&conf->prev, r10b);
4572
4573         while (sectors) {
4574                 int s = sectors;
4575                 int success = 0;
4576                 int first_slot = slot;
4577
4578                 if (s > (PAGE_SIZE >> 9))
4579                         s = PAGE_SIZE >> 9;
4580
4581                 while (!success) {
4582                         int d = r10b->devs[slot].devnum;
4583                         struct md_rdev *rdev = conf->mirrors[d].rdev;
4584                         sector_t addr;
4585                         if (rdev == NULL ||
4586                             test_bit(Faulty, &rdev->flags) ||
4587                             !test_bit(In_sync, &rdev->flags))
4588                                 goto failed;
4589
4590                         addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4591                         success = sync_page_io(rdev,
4592                                                addr,
4593                                                s << 9,
4594                                                bvec[idx].bv_page,
4595                                                READ, false);
4596                         if (success)
4597                                 break;
4598                 failed:
4599                         slot++;
4600                         if (slot >= conf->copies)
4601                                 slot = 0;
4602                         if (slot == first_slot)
4603                                 break;
4604                 }
4605                 if (!success) {
4606                         /* couldn't read this block, must give up */
4607                         set_bit(MD_RECOVERY_INTR,
4608                                 &mddev->recovery);
4609                         return -EIO;
4610                 }
4611                 sectors -= s;
4612                 idx++;
4613         }
4614         return 0;
4615 }
4616
4617 static void end_reshape_write(struct bio *bio, int error)
4618 {
4619         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4620         struct r10bio *r10_bio = bio->bi_private;
4621         struct mddev *mddev = r10_bio->mddev;
4622         struct r10conf *conf = mddev->private;
4623         int d;
4624         int slot;
4625         int repl;
4626         struct md_rdev *rdev = NULL;
4627
4628         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4629         if (repl)
4630                 rdev = conf->mirrors[d].replacement;
4631         if (!rdev) {
4632                 smp_mb();
4633                 rdev = conf->mirrors[d].rdev;
4634         }
4635
4636         if (!uptodate) {
4637                 /* FIXME should record badblock */
4638                 md_error(mddev, rdev);
4639         }
4640
4641         rdev_dec_pending(rdev, mddev);
4642         end_reshape_request(r10_bio);
4643 }
4644
4645 static void end_reshape_request(struct r10bio *r10_bio)
4646 {
4647         if (!atomic_dec_and_test(&r10_bio->remaining))
4648                 return;
4649         md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4650         bio_put(r10_bio->master_bio);
4651         put_buf(r10_bio);
4652 }
4653
4654 static void raid10_finish_reshape(struct mddev *mddev)
4655 {
4656         struct r10conf *conf = mddev->private;
4657
4658         if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4659                 return;
4660
4661         if (mddev->delta_disks > 0) {
4662                 sector_t size = raid10_size(mddev, 0, 0);
4663                 md_set_array_sectors(mddev, size);
4664                 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4665                         mddev->recovery_cp = mddev->resync_max_sectors;
4666                         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4667                 }
4668                 mddev->resync_max_sectors = size;
4669                 set_capacity(mddev->gendisk, mddev->array_sectors);
4670                 revalidate_disk(mddev->gendisk);
4671         } else {
4672                 int d;
4673                 for (d = conf->geo.raid_disks ;
4674                      d < conf->geo.raid_disks - mddev->delta_disks;
4675                      d++) {
4676                         struct md_rdev *rdev = conf->mirrors[d].rdev;
4677                         if (rdev)
4678                                 clear_bit(In_sync, &rdev->flags);
4679                         rdev = conf->mirrors[d].replacement;
4680                         if (rdev)
4681                                 clear_bit(In_sync, &rdev->flags);
4682                 }
4683         }
4684         mddev->layout = mddev->new_layout;
4685         mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4686         mddev->reshape_position = MaxSector;
4687         mddev->delta_disks = 0;
4688         mddev->reshape_backwards = 0;
4689 }
4690
4691 static struct md_personality raid10_personality =
4692 {
4693         .name           = "raid10",
4694         .level          = 10,
4695         .owner          = THIS_MODULE,
4696         .make_request   = make_request,
4697         .run            = run,
4698         .free           = raid10_free,
4699         .status         = status,
4700         .error_handler  = error,
4701         .hot_add_disk   = raid10_add_disk,
4702         .hot_remove_disk= raid10_remove_disk,
4703         .spare_active   = raid10_spare_active,
4704         .sync_request   = sync_request,
4705         .quiesce        = raid10_quiesce,
4706         .size           = raid10_size,
4707         .resize         = raid10_resize,
4708         .takeover       = raid10_takeover,
4709         .check_reshape  = raid10_check_reshape,
4710         .start_reshape  = raid10_start_reshape,
4711         .finish_reshape = raid10_finish_reshape,
4712         .congested      = raid10_congested,
4713         .mergeable_bvec = raid10_mergeable_bvec,
4714 };
4715
4716 static int __init raid_init(void)
4717 {
4718         return register_md_personality(&raid10_personality);
4719 }
4720
4721 static void raid_exit(void)
4722 {
4723         unregister_md_personality(&raid10_personality);
4724 }
4725
4726 module_init(raid_init);
4727 module_exit(raid_exit);
4728 MODULE_LICENSE("GPL");
4729 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4730 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4731 MODULE_ALIAS("md-raid10");
4732 MODULE_ALIAS("md-level-10");
4733
4734 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);