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