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  *
  42  * The data to be stored is divided into chunks using chunksize.
  43  * Each device is divided into far_copies sections.
  44  * In each section, chunks are laid out in a style similar to raid0, but
  45  * near_copies copies of each chunk is stored (each on a different drive).
  46  * The starting device for each section is offset near_copies from the starting
  47  * device of the previous section.
  48  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
  49  * drive.
  50  * near_copies and far_copies must be at least one, and their product is at most
  51  * raid_disks.
  52  *
  53  * If far_offset is true, then the far_copies are handled a bit differently.
  54  * The copies are still in different stripes, but instead of be very far apart
  55  * on disk, there are adjacent stripes.
  56  */
  57
  58 /*
  59  * Number of guaranteed r10bios in case of extreme VM load:
  60  */
  61 #define NR_RAID10_BIOS 256
  62
  63 /* When there are this many requests queue to be written by
  64  * the raid10 thread, we become 'congested' to provide back-pressure
  65  * for writeback.
  66  */
  67 static int max_queued_requests = 1024;
  68
  69 static void allow_barrier(struct r10conf *conf);
  70 static void lower_barrier(struct r10conf *conf);
  71 static int enough(struct r10conf *conf, int ignore);
  72 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
  73                                 int *skipped);
  74 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
  75 static void end_reshape_write(struct bio *bio, int error);
  76 static void end_reshape(struct r10conf *conf);
  77
  78 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
  79 {
  80         struct r10conf *conf = data;
  81         int size = offsetof(struct r10bio, devs[conf->copies]);
  82
  83         /* allocate a r10bio with room for raid_disks entries in the
  84          * bios array */
  85         return kzalloc(size, gfp_flags);
  86 }
  87
  88 static void r10bio_pool_free(void *r10_bio, void *data)
  89 {
  90         kfree(r10_bio);
  91 }
  92
  93 /* Maximum size of each resync request */
  94 #define RESYNC_BLOCK_SIZE (64*1024)
  95 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  96 /* amount of memory to reserve for resync requests */
  97 #define RESYNC_WINDOW (1024*1024)
  98 /* maximum number of concurrent requests, memory permitting */
  99 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
 100
 101 /*
 102  * When performing a resync, we need to read and compare, so
 103  * we need as many pages are there are copies.
 104  * When performing a recovery, we need 2 bios, one for read,
 105  * one for write (we recover only one drive per r10buf)
 106  *
 107  */
 108 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
 109 {
 110         struct r10conf *conf = data;
 111         struct page *page;
 112         struct r10bio *r10_bio;
 113         struct bio *bio;
 114         int i, j;
 115         int nalloc;
 116
 117         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 118         if (!r10_bio)
 119                 return NULL;
 120
 121         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
 122             test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
 123                 nalloc = conf->copies; /* resync */
 124         else
 125                 nalloc = 2; /* recovery */
 126
 127         /*
 128          * Allocate bios.
 129          */
 130         for (j = nalloc ; j-- ; ) {
 131                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 132                 if (!bio)
 133                         goto out_free_bio;
 134                 r10_bio->devs[j].bio = bio;
 135                 if (!conf->have_replacement)
 136                         continue;
 137                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 138                 if (!bio)
 139                         goto out_free_bio;
 140                 r10_bio->devs[j].repl_bio = bio;
 141         }
 142         /*
 143          * Allocate RESYNC_PAGES data pages and attach them
 144          * where needed.
 145          */
 146         for (j = 0 ; j < nalloc; j++) {
 147                 struct bio *rbio = r10_bio->devs[j].repl_bio;
 148                 bio = r10_bio->devs[j].bio;
 149                 for (i = 0; i < RESYNC_PAGES; i++) {
 150                         if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
 151                                                &conf->mddev->recovery)) {
 152                                 /* we can share bv_page's during recovery
 153                                  * and reshape */
 154                                 struct bio *rbio = r10_bio->devs[0].bio;
 155                                 page = rbio->bi_io_vec[i].bv_page;
 156                                 get_page(page);
 157                         } else
 158                                 page = alloc_page(gfp_flags);
 159                         if (unlikely(!page))
 160                                 goto out_free_pages;
 161
 162                         bio->bi_io_vec[i].bv_page = page;
 163                         if (rbio)
 164                                 rbio->bi_io_vec[i].bv_page = page;
 165                 }
 166         }
 167
 168         return r10_bio;
 169
 170 out_free_pages:
 171         for ( ; i > 0 ; i--)
 172                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
 173         while (j--)
 174                 for (i = 0; i < RESYNC_PAGES ; i++)
 175                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 176         j = 0;
 177 out_free_bio:
 178         for ( ; j < nalloc; j++) {
 179                 if (r10_bio->devs[j].bio)
 180                         bio_put(r10_bio->devs[j].bio);
 181                 if (r10_bio->devs[j].repl_bio)
 182                         bio_put(r10_bio->devs[j].repl_bio);
 183         }
 184         r10bio_pool_free(r10_bio, conf);
 185         return NULL;
 186 }
 187
 188 static void r10buf_pool_free(void *__r10_bio, void *data)
 189 {
 190         int i;
 191         struct r10conf *conf = data;
 192         struct r10bio *r10bio = __r10_bio;
 193         int j;
 194
 195         for (j=0; j < conf->copies; j++) {
 196                 struct bio *bio = r10bio->devs[j].bio;
 197                 if (bio) {
 198                         for (i = 0; i < RESYNC_PAGES; i++) {
 199                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 200                                 bio->bi_io_vec[i].bv_page = NULL;
 201                         }
 202                         bio_put(bio);
 203                 }
 204                 bio = r10bio->devs[j].repl_bio;
 205                 if (bio)
 206                         bio_put(bio);
 207         }
 208         r10bio_pool_free(r10bio, conf);
 209 }
 210
 211 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
 212 {
 213         int i;
 214
 215         for (i = 0; i < conf->copies; i++) {
 216                 struct bio **bio = & r10_bio->devs[i].bio;
 217                 if (!BIO_SPECIAL(*bio))
 218                         bio_put(*bio);
 219                 *bio = NULL;
 220                 bio = &r10_bio->devs[i].repl_bio;
 221                 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
 222                         bio_put(*bio);
 223                 *bio = NULL;
 224         }
 225 }
 226
 227 static void free_r10bio(struct r10bio *r10_bio)
 228 {
 229         struct r10conf *conf = r10_bio->mddev->private;
 230
 231         put_all_bios(conf, r10_bio);
 232         mempool_free(r10_bio, conf->r10bio_pool);
 233 }
 234
 235 static void put_buf(struct r10bio *r10_bio)
 236 {
 237         struct r10conf *conf = r10_bio->mddev->private;
 238
 239         mempool_free(r10_bio, conf->r10buf_pool);
 240
 241         lower_barrier(conf);
 242 }
 243
 244 static void reschedule_retry(struct r10bio *r10_bio)
 245 {
 246         unsigned long flags;
 247         struct mddev *mddev = r10_bio->mddev;
 248         struct r10conf *conf = mddev->private;
 249
 250         spin_lock_irqsave(&conf->device_lock, flags);
 251         list_add(&r10_bio->retry_list, &conf->retry_list);
 252         conf->nr_queued ++;
 253         spin_unlock_irqrestore(&conf->device_lock, flags);
 254
 255         /* wake up frozen array... */
 256         wake_up(&conf->wait_barrier);
 257
 258         md_wakeup_thread(mddev->thread);
 259 }
 260
 261 /*
 262  * raid_end_bio_io() is called when we have finished servicing a mirrored
 263  * operation and are ready to return a success/failure code to the buffer
 264  * cache layer.
 265  */
 266 static void raid_end_bio_io(struct r10bio *r10_bio)
 267 {
 268         struct bio *bio = r10_bio->master_bio;
 269         int done;
 270         struct r10conf *conf = r10_bio->mddev->private;
 271
 272         if (bio->bi_phys_segments) {
 273                 unsigned long flags;
 274                 spin_lock_irqsave(&conf->device_lock, flags);
 275                 bio->bi_phys_segments--;
 276                 done = (bio->bi_phys_segments == 0);
 277                 spin_unlock_irqrestore(&conf->device_lock, flags);
 278         } else
 279                 done = 1;
 280         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
 281                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 282         if (done) {
 283                 bio_endio(bio, 0);
 284                 /*
 285                  * Wake up any possible resync thread that waits for the device
 286                  * to go idle.
 287                  */
 288                 allow_barrier(conf);
 289         }
 290         free_r10bio(r10_bio);
 291 }
 292
 293 /*
 294  * Update disk head position estimator based on IRQ completion info.
 295  */
 296 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
 297 {
 298         struct r10conf *conf = r10_bio->mddev->private;
 299
 300         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 301                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 302 }
 303
 304 /*
 305  * Find the disk number which triggered given bio
 306  */
 307 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
 308                          struct bio *bio, int *slotp, int *replp)
 309 {
 310         int slot;
 311         int repl = 0;
 312
 313         for (slot = 0; slot < conf->copies; slot++) {
 314                 if (r10_bio->devs[slot].bio == bio)
 315                         break;
 316                 if (r10_bio->devs[slot].repl_bio == bio) {
 317                         repl = 1;
 318                         break;
 319                 }
 320         }
 321
 322         BUG_ON(slot == conf->copies);
 323         update_head_pos(slot, r10_bio);
 324
 325         if (slotp)
 326                 *slotp = slot;
 327         if (replp)
 328                 *replp = repl;
 329         return r10_bio->devs[slot].devnum;
 330 }
 331
 332 static void raid10_end_read_request(struct bio *bio, int error)
 333 {
 334         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 335         struct r10bio *r10_bio = bio->bi_private;
 336         int slot, dev;
 337         struct md_rdev *rdev;
 338         struct r10conf *conf = r10_bio->mddev->private;
 339
 340
 341         slot = r10_bio->read_slot;
 342         dev = r10_bio->devs[slot].devnum;
 343         rdev = r10_bio->devs[slot].rdev;
 344         /*
 345          * this branch is our 'one mirror IO has finished' event handler:
 346          */
 347         update_head_pos(slot, r10_bio);
 348
 349         if (uptodate) {
 350                 /*
 351                  * Set R10BIO_Uptodate in our master bio, so that
 352                  * we will return a good error code to the higher
 353                  * levels even if IO on some other mirrored buffer fails.
 354                  *
 355                  * The 'master' represents the composite IO operation to
 356                  * user-side. So if something waits for IO, then it will
 357                  * wait for the 'master' bio.
 358                  */
 359                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 360         } else {
 361                 /* If all other devices that store this block have
 362                  * failed, we want to return the error upwards rather
 363                  * than fail the last device.  Here we redefine
 364                  * "uptodate" to mean "Don't want to retry"
 365                  */
 366                 unsigned long flags;
 367                 spin_lock_irqsave(&conf->device_lock, flags);
 368                 if (!enough(conf, rdev->raid_disk))
 369                         uptodate = 1;
 370                 spin_unlock_irqrestore(&conf->device_lock, flags);
 371         }
 372         if (uptodate) {
 373                 raid_end_bio_io(r10_bio);
 374                 rdev_dec_pending(rdev, conf->mddev);
 375         } else {
 376                 /*
 377                  * oops, read error - keep the refcount on the rdev
 378                  */
 379                 char b[BDEVNAME_SIZE];
 380                 printk_ratelimited(KERN_ERR
 381                                    "md/raid10:%s: %s: rescheduling sector %llu\n",
 382                                    mdname(conf->mddev),
 383                                    bdevname(rdev->bdev, b),
 384                                    (unsigned long long)r10_bio->sector);
 385                 set_bit(R10BIO_ReadError, &r10_bio->state);
 386                 reschedule_retry(r10_bio);
 387         }
 388 }
 389
 390 static void close_write(struct r10bio *r10_bio)
 391 {
 392         /* clear the bitmap if all writes complete successfully */
 393         bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 394                         r10_bio->sectors,
 395                         !test_bit(R10BIO_Degraded, &r10_bio->state),
 396                         0);
 397         md_write_end(r10_bio->mddev);
 398 }
 399
 400 static void one_write_done(struct r10bio *r10_bio)
 401 {
 402         if (atomic_dec_and_test(&r10_bio->remaining)) {
 403                 if (test_bit(R10BIO_WriteError, &r10_bio->state))
 404                         reschedule_retry(r10_bio);
 405                 else {
 406                         close_write(r10_bio);
 407                         if (test_bit(R10BIO_MadeGood, &r10_bio->state))
 408                                 reschedule_retry(r10_bio);
 409                         else
 410                                 raid_end_bio_io(r10_bio);
 411                 }
 412         }
 413 }
 414
 415 static void raid10_end_write_request(struct bio *bio, int error)
 416 {
 417         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 418         struct r10bio *r10_bio = bio->bi_private;
 419         int dev;
 420         int dec_rdev = 1;
 421         struct r10conf *conf = r10_bio->mddev->private;
 422         int slot, repl;
 423         struct md_rdev *rdev = NULL;
 424
 425         dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
 426
 427         if (repl)
 428                 rdev = conf->mirrors[dev].replacement;
 429         if (!rdev) {
 430                 smp_rmb();
 431                 repl = 0;
 432                 rdev = conf->mirrors[dev].rdev;
 433         }
 434         /*
 435          * this branch is our 'one mirror IO has finished' event handler:
 436          */
 437         if (!uptodate) {
 438                 if (repl)
 439                         /* Never record new bad blocks to replacement,
 440                          * just fail it.
 441                          */
 442                         md_error(rdev->mddev, rdev);
 443                 else {
 444                         set_bit(WriteErrorSeen, &rdev->flags);
 445                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
 446                                 set_bit(MD_RECOVERY_NEEDED,
 447                                         &rdev->mddev->recovery);
 448                         set_bit(R10BIO_WriteError, &r10_bio->state);
 449                         dec_rdev = 0;
 450                 }
 451         } else {
 452                 /*
 453                  * Set R10BIO_Uptodate in our master bio, so that
 454                  * we will return a good error code for to the higher
 455                  * levels even if IO on some other mirrored buffer fails.
 456                  *
 457                  * The 'master' represents the composite IO operation to
 458                  * user-side. So if something waits for IO, then it will
 459                  * wait for the 'master' bio.
 460                  */
 461                 sector_t first_bad;
 462                 int bad_sectors;
 463
 464                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 465
 466                 /* Maybe we can clear some bad blocks. */
 467                 if (is_badblock(rdev,
 468                                 r10_bio->devs[slot].addr,
 469                                 r10_bio->sectors,
 470                                 &first_bad, &bad_sectors)) {
 471                         bio_put(bio);
 472                         if (repl)
 473                                 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
 474                         else
 475                                 r10_bio->devs[slot].bio = IO_MADE_GOOD;
 476                         dec_rdev = 0;
 477                         set_bit(R10BIO_MadeGood, &r10_bio->state);
 478                 }
 479         }
 480
 481         /*
 482          *
 483          * Let's see if all mirrored write operations have finished
 484          * already.
 485          */
 486         one_write_done(r10_bio);
 487         if (dec_rdev)
 488                 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 489 }
 490
 491 /*
 492  * RAID10 layout manager
 493  * As well as the chunksize and raid_disks count, there are two
 494  * parameters: near_copies and far_copies.
 495  * near_copies * far_copies must be <= raid_disks.
 496  * Normally one of these will be 1.
 497  * If both are 1, we get raid0.
 498  * If near_copies == raid_disks, we get raid1.
 499  *
 500  * Chunks are laid out in raid0 style with near_copies copies of the
 501  * first chunk, followed by near_copies copies of the next chunk and
 502  * so on.
 503  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 504  * as described above, we start again with a device offset of near_copies.
 505  * So we effectively have another copy of the whole array further down all
 506  * the drives, but with blocks on different drives.
 507  * With this layout, and block is never stored twice on the one device.
 508  *
 509  * raid10_find_phys finds the sector offset of a given virtual sector
 510  * on each device that it is on.
 511  *
 512  * raid10_find_virt does the reverse mapping, from a device and a
 513  * sector offset to a virtual address
 514  */
 515
 516 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
 517 {
 518         int n,f;
 519         sector_t sector;
 520         sector_t chunk;
 521         sector_t stripe;
 522         int dev;
 523         int slot = 0;
 524
 525         /* now calculate first sector/dev */
 526         chunk = r10bio->sector >> geo->chunk_shift;
 527         sector = r10bio->sector & geo->chunk_mask;
 528
 529         chunk *= geo->near_copies;
 530         stripe = chunk;
 531         dev = sector_div(stripe, geo->raid_disks);
 532         if (geo->far_offset)
 533                 stripe *= geo->far_copies;
 534
 535         sector += stripe << geo->chunk_shift;
 536
 537         /* and calculate all the others */
 538         for (n = 0; n < geo->near_copies; n++) {
 539                 int d = dev;
 540                 sector_t s = sector;
 541                 r10bio->devs[slot].addr = sector;
 542                 r10bio->devs[slot].devnum = d;
 543                 slot++;
 544
 545                 for (f = 1; f < geo->far_copies; f++) {
 546                         d += geo->near_copies;
 547                         if (d >= geo->raid_disks)
 548                                 d -= geo->raid_disks;
 549                         s += geo->stride;
 550                         r10bio->devs[slot].devnum = d;
 551                         r10bio->devs[slot].addr = s;
 552                         slot++;
 553                 }
 554                 dev++;
 555                 if (dev >= geo->raid_disks) {
 556                         dev = 0;
 557                         sector += (geo->chunk_mask + 1);
 558                 }
 559         }
 560 }
 561
 562 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
 563 {
 564         struct geom *geo = &conf->geo;
 565
 566         if (conf->reshape_progress != MaxSector &&
 567             ((r10bio->sector >= conf->reshape_progress) !=
 568              conf->mddev->reshape_backwards)) {
 569                 set_bit(R10BIO_Previous, &r10bio->state);
 570                 geo = &conf->prev;
 571         } else
 572                 clear_bit(R10BIO_Previous, &r10bio->state);
 573
 574         __raid10_find_phys(geo, r10bio);
 575 }
 576
 577 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
 578 {
 579         sector_t offset, chunk, vchunk;
 580         /* Never use conf->prev as this is only called during resync
 581          * or recovery, so reshape isn't happening
 582          */
 583         struct geom *geo = &conf->geo;
 584
 585         offset = sector & geo->chunk_mask;
 586         if (geo->far_offset) {
 587                 int fc;
 588                 chunk = sector >> geo->chunk_shift;
 589                 fc = sector_div(chunk, geo->far_copies);
 590                 dev -= fc * geo->near_copies;
 591                 if (dev < 0)
 592                         dev += geo->raid_disks;
 593         } else {
 594                 while (sector >= geo->stride) {
 595                         sector -= geo->stride;
 596                         if (dev < geo->near_copies)
 597                                 dev += geo->raid_disks - geo->near_copies;
 598                         else
 599                                 dev -= geo->near_copies;
 600                 }
 601                 chunk = sector >> geo->chunk_shift;
 602         }
 603         vchunk = chunk * geo->raid_disks + dev;
 604         sector_div(vchunk, geo->near_copies);
 605         return (vchunk << geo->chunk_shift) + offset;
 606 }
 607
 608 /**
 609  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 610  *      @q: request queue
 611  *      @bvm: properties of new bio
 612  *      @biovec: the request that could be merged to it.
 613  *
 614  *      Return amount of bytes we can accept at this offset
 615  *      This requires checking for end-of-chunk if near_copies != raid_disks,
 616  *      and for subordinate merge_bvec_fns if merge_check_needed.
 617  */
 618 static int raid10_mergeable_bvec(struct request_queue *q,
 619                                  struct bvec_merge_data *bvm,
 620                                  struct bio_vec *biovec)
 621 {
 622         struct mddev *mddev = q->queuedata;
 623         struct r10conf *conf = mddev->private;
 624         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 625         int max;
 626         unsigned int chunk_sectors;
 627         unsigned int bio_sectors = bvm->bi_size >> 9;
 628         struct geom *geo = &conf->geo;
 629
 630         chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
 631         if (conf->reshape_progress != MaxSector &&
 632             ((sector >= conf->reshape_progress) !=
 633              conf->mddev->reshape_backwards))
 634                 geo = &conf->prev;
 635
 636         if (geo->near_copies < geo->raid_disks) {
 637                 max = (chunk_sectors - ((sector & (chunk_sectors - 1))
 638                                         + bio_sectors)) << 9;
 639                 if (max < 0)
 640                         /* bio_add cannot handle a negative return */
 641                         max = 0;
 642                 if (max <= biovec->bv_len && bio_sectors == 0)
 643                         return biovec->bv_len;
 644         } else
 645                 max = biovec->bv_len;
 646
 647         if (mddev->merge_check_needed) {
 648                 struct r10bio r10_bio;
 649                 int s;
 650                 if (conf->reshape_progress != MaxSector) {
 651                         /* Cannot give any guidance during reshape */
 652                         if (max <= biovec->bv_len && bio_sectors == 0)
 653                                 return biovec->bv_len;
 654                         return 0;
 655                 }
 656                 r10_bio.sector = sector;
 657                 raid10_find_phys(conf, &r10_bio);
 658                 rcu_read_lock();
 659                 for (s = 0; s < conf->copies; s++) {
 660                         int disk = r10_bio.devs[s].devnum;
 661                         struct md_rdev *rdev = rcu_dereference(
 662                                 conf->mirrors[disk].rdev);
 663                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
 664                                 struct request_queue *q =
 665                                         bdev_get_queue(rdev->bdev);
 666                                 if (q->merge_bvec_fn) {
 667                                         bvm->bi_sector = r10_bio.devs[s].addr
 668                                                 + rdev->data_offset;
 669                                         bvm->bi_bdev = rdev->bdev;
 670                                         max = min(max, q->merge_bvec_fn(
 671                                                           q, bvm, biovec));
 672                                 }
 673                         }
 674                         rdev = rcu_dereference(conf->mirrors[disk].replacement);
 675                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
 676                                 struct request_queue *q =
 677                                         bdev_get_queue(rdev->bdev);
 678                                 if (q->merge_bvec_fn) {
 679                                         bvm->bi_sector = r10_bio.devs[s].addr
 680                                                 + rdev->data_offset;
 681                                         bvm->bi_bdev = rdev->bdev;
 682                                         max = min(max, q->merge_bvec_fn(
 683                                                           q, bvm, biovec));
 684                                 }
 685                         }
 686                 }
 687                 rcu_read_unlock();
 688         }
 689         return max;
 690 }
 691
 692 /*
 693  * This routine returns the disk from which the requested read should
 694  * be done. There is a per-array 'next expected sequential IO' sector
 695  * number - if this matches on the next IO then we use the last disk.
 696  * There is also a per-disk 'last know head position' sector that is
 697  * maintained from IRQ contexts, both the normal and the resync IO
 698  * completion handlers update this position correctly. If there is no
 699  * perfect sequential match then we pick the disk whose head is closest.
 700  *
 701  * If there are 2 mirrors in the same 2 devices, performance degrades
 702  * because position is mirror, not device based.
 703  *
 704  * The rdev for the device selected will have nr_pending incremented.
 705  */
 706
 707 /*
 708  * FIXME: possibly should rethink readbalancing and do it differently
 709  * depending on near_copies / far_copies geometry.
 710  */
 711 static struct md_rdev *read_balance(struct r10conf *conf,
 712                                     struct r10bio *r10_bio,
 713                                     int *max_sectors)
 714 {
 715         const sector_t this_sector = r10_bio->sector;
 716         int disk, slot;
 717         int sectors = r10_bio->sectors;
 718         int best_good_sectors;
 719         sector_t new_distance, best_dist;
 720         struct md_rdev *rdev, *best_rdev;
 721         int do_balance;
 722         int best_slot;
 723         struct geom *geo = &conf->geo;
 724
 725         raid10_find_phys(conf, r10_bio);
 726         rcu_read_lock();
 727 retry:
 728         sectors = r10_bio->sectors;
 729         best_slot = -1;
 730         best_rdev = NULL;
 731         best_dist = MaxSector;
 732         best_good_sectors = 0;
 733         do_balance = 1;
 734         /*
 735          * Check if we can balance. We can balance on the whole
 736          * device if no resync is going on (recovery is ok), or below
 737          * the resync window. We take the first readable disk when
 738          * above the resync window.
 739          */
 740         if (conf->mddev->recovery_cp < MaxSector
 741             && (this_sector + sectors >= conf->next_resync))
 742                 do_balance = 0;
 743
 744         for (slot = 0; slot < conf->copies ; slot++) {
 745                 sector_t first_bad;
 746                 int bad_sectors;
 747                 sector_t dev_sector;
 748
 749                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
 750                         continue;
 751                 disk = r10_bio->devs[slot].devnum;
 752                 rdev = rcu_dereference(conf->mirrors[disk].replacement);
 753                 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
 754                     test_bit(Unmerged, &rdev->flags) ||
 755                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 756                         rdev = rcu_dereference(conf->mirrors[disk].rdev);
 757                 if (rdev == NULL ||
 758                     test_bit(Faulty, &rdev->flags) ||
 759                     test_bit(Unmerged, &rdev->flags))
 760                         continue;
 761                 if (!test_bit(In_sync, &rdev->flags) &&
 762                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 763                         continue;
 764
 765                 dev_sector = r10_bio->devs[slot].addr;
 766                 if (is_badblock(rdev, dev_sector, sectors,
 767                                 &first_bad, &bad_sectors)) {
 768                         if (best_dist < MaxSector)
 769                                 /* Already have a better slot */
 770                                 continue;
 771                         if (first_bad <= dev_sector) {
 772                                 /* Cannot read here.  If this is the
 773                                  * 'primary' device, then we must not read
 774                                  * beyond 'bad_sectors' from another device.
 775                                  */
 776                                 bad_sectors -= (dev_sector - first_bad);
 777                                 if (!do_balance && sectors > bad_sectors)
 778                                         sectors = bad_sectors;
 779                                 if (best_good_sectors > sectors)
 780                                         best_good_sectors = sectors;
 781                         } else {
 782                                 sector_t good_sectors =
 783                                         first_bad - dev_sector;
 784                                 if (good_sectors > best_good_sectors) {
 785                                         best_good_sectors = good_sectors;
 786                                         best_slot = slot;
 787                                         best_rdev = rdev;
 788                                 }
 789                                 if (!do_balance)
 790                                         /* Must read from here */
 791                                         break;
 792                         }
 793                         continue;
 794                 } else
 795                         best_good_sectors = sectors;
 796
 797                 if (!do_balance)
 798                         break;
 799
 800                 /* This optimisation is debatable, and completely destroys
 801                  * sequential read speed for 'far copies' arrays.  So only
 802                  * keep it for 'near' arrays, and review those later.
 803                  */
 804                 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
 805                         break;
 806
 807                 /* for far > 1 always use the lowest address */
 808                 if (geo->far_copies > 1)
 809                         new_distance = r10_bio->devs[slot].addr;
 810                 else
 811                         new_distance = abs(r10_bio->devs[slot].addr -
 812                                            conf->mirrors[disk].head_position);
 813                 if (new_distance < best_dist) {
 814                         best_dist = new_distance;
 815                         best_slot = slot;
 816                         best_rdev = rdev;
 817                 }
 818         }
 819         if (slot >= conf->copies) {
 820                 slot = best_slot;
 821                 rdev = best_rdev;
 822         }
 823
 824         if (slot >= 0) {
 825                 atomic_inc(&rdev->nr_pending);
 826                 if (test_bit(Faulty, &rdev->flags)) {
 827                         /* Cannot risk returning a device that failed
 828                          * before we inc'ed nr_pending
 829                          */
 830                         rdev_dec_pending(rdev, conf->mddev);
 831                         goto retry;
 832                 }
 833                 r10_bio->read_slot = slot;
 834         } else
 835                 rdev = NULL;
 836         rcu_read_unlock();
 837         *max_sectors = best_good_sectors;
 838
 839         return rdev;
 840 }
 841
 842 static int raid10_congested(void *data, int bits)
 843 {
 844         struct mddev *mddev = data;
 845         struct r10conf *conf = mddev->private;
 846         int i, ret = 0;
 847
 848         if ((bits & (1 << BDI_async_congested)) &&
 849             conf->pending_count >= max_queued_requests)
 850                 return 1;
 851
 852         if (mddev_congested(mddev, bits))
 853                 return 1;
 854         rcu_read_lock();
 855         for (i = 0;
 856              (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
 857                      && ret == 0;
 858              i++) {
 859                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 860                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 861                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 862
 863                         ret |= bdi_congested(&q->backing_dev_info, bits);
 864                 }
 865         }
 866         rcu_read_unlock();
 867         return ret;
 868 }
 869
 870 static void flush_pending_writes(struct r10conf *conf)
 871 {
 872         /* Any writes that have been queued but are awaiting
 873          * bitmap updates get flushed here.
 874          */
 875         spin_lock_irq(&conf->device_lock);
 876
 877         if (conf->pending_bio_list.head) {
 878                 struct bio *bio;
 879                 bio = bio_list_get(&conf->pending_bio_list);
 880                 conf->pending_count = 0;
 881                 spin_unlock_irq(&conf->device_lock);
 882                 /* flush any pending bitmap writes to disk
 883                  * before proceeding w/ I/O */
 884                 bitmap_unplug(conf->mddev->bitmap);
 885                 wake_up(&conf->wait_barrier);
 886
 887                 while (bio) { /* submit pending writes */
 888                         struct bio *next = bio->bi_next;
 889                         bio->bi_next = NULL;
 890                         generic_make_request(bio);
 891                         bio = next;
 892                 }
 893         } else
 894                 spin_unlock_irq(&conf->device_lock);
 895 }
 896
 897 /* Barriers....
 898  * Sometimes we need to suspend IO while we do something else,
 899  * either some resync/recovery, or reconfigure the array.
 900  * To do this we raise a 'barrier'.
 901  * The 'barrier' is a counter that can be raised multiple times
 902  * to count how many activities are happening which preclude
 903  * normal IO.
 904  * We can only raise the barrier if there is no pending IO.
 905  * i.e. if nr_pending == 0.
 906  * We choose only to raise the barrier if no-one is waiting for the
 907  * barrier to go down.  This means that as soon as an IO request
 908  * is ready, no other operations which require a barrier will start
 909  * until the IO request has had a chance.
 910  *
 911  * So: regular IO calls 'wait_barrier'.  When that returns there
 912  *    is no backgroup IO happening,  It must arrange to call
 913  *    allow_barrier when it has finished its IO.
 914  * backgroup IO calls must call raise_barrier.  Once that returns
 915  *    there is no normal IO happeing.  It must arrange to call
 916  *    lower_barrier when the particular background IO completes.
 917  */
 918
 919 static void raise_barrier(struct r10conf *conf, int force)
 920 {
 921         BUG_ON(force && !conf->barrier);
 922         spin_lock_irq(&conf->resync_lock);
 923
 924         /* Wait until no block IO is waiting (unless 'force') */
 925         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 926                             conf->resync_lock, );
 927
 928         /* block any new IO from starting */
 929         conf->barrier++;
 930
 931         /* Now wait for all pending IO to complete */
 932         wait_event_lock_irq(conf->wait_barrier,
 933                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 934                             conf->resync_lock, );
 935
 936         spin_unlock_irq(&conf->resync_lock);
 937 }
 938
 939 static void lower_barrier(struct r10conf *conf)
 940 {
 941         unsigned long flags;
 942         spin_lock_irqsave(&conf->resync_lock, flags);
 943         conf->barrier--;
 944         spin_unlock_irqrestore(&conf->resync_lock, flags);
 945         wake_up(&conf->wait_barrier);
 946 }
 947
 948 static void wait_barrier(struct r10conf *conf)
 949 {
 950         spin_lock_irq(&conf->resync_lock);
 951         if (conf->barrier) {
 952                 conf->nr_waiting++;
 953                 /* Wait for the barrier to drop.
 954                  * However if there are already pending
 955                  * requests (preventing the barrier from
 956                  * rising completely), and the
 957                  * pre-process bio queue isn't empty,
 958                  * then don't wait, as we need to empty
 959                  * that queue to get the nr_pending
 960                  * count down.
 961                  */
 962                 wait_event_lock_irq(conf->wait_barrier,
 963                                     !conf->barrier ||
 964                                     (conf->nr_pending &&
 965                                      current->bio_list &&
 966                                      !bio_list_empty(current->bio_list)),
 967                                     conf->resync_lock,
 968                         );
 969                 conf->nr_waiting--;
 970         }
 971         conf->nr_pending++;
 972         spin_unlock_irq(&conf->resync_lock);
 973 }
 974
 975 static void allow_barrier(struct r10conf *conf)
 976 {
 977         unsigned long flags;
 978         spin_lock_irqsave(&conf->resync_lock, flags);
 979         conf->nr_pending--;
 980         spin_unlock_irqrestore(&conf->resync_lock, flags);
 981         wake_up(&conf->wait_barrier);
 982 }
 983
 984 static void freeze_array(struct r10conf *conf)
 985 {
 986         /* stop syncio and normal IO and wait for everything to
 987          * go quiet.
 988          * We increment barrier and nr_waiting, and then
 989          * wait until nr_pending match nr_queued+1
 990          * This is called in the context of one normal IO request
 991          * that has failed. Thus any sync request that might be pending
 992          * will be blocked by nr_pending, and we need to wait for
 993          * pending IO requests to complete or be queued for re-try.
 994          * Thus the number queued (nr_queued) plus this request (1)
 995          * must match the number of pending IOs (nr_pending) before
 996          * we continue.
 997          */
 998         spin_lock_irq(&conf->resync_lock);
 999         conf->barrier++;
1000         conf->nr_waiting++;
1001         wait_event_lock_irq(conf->wait_barrier,
1002                             conf->nr_pending == conf->nr_queued+1,
1003                             conf->resync_lock,
1004                             flush_pending_writes(conf));
1005
1006         spin_unlock_irq(&conf->resync_lock);
1007 }
1008
1009 static void unfreeze_array(struct r10conf *conf)
1010 {
1011         /* reverse the effect of the freeze */
1012         spin_lock_irq(&conf->resync_lock);
1013         conf->barrier--;
1014         conf->nr_waiting--;
1015         wake_up(&conf->wait_barrier);
1016         spin_unlock_irq(&conf->resync_lock);
1017 }
1018
1019 static sector_t choose_data_offset(struct r10bio *r10_bio,
1020                                    struct md_rdev *rdev)
1021 {
1022         if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1023             test_bit(R10BIO_Previous, &r10_bio->state))
1024                 return rdev->data_offset;
1025         else
1026                 return rdev->new_data_offset;
1027 }
1028
1029 static void make_request(struct mddev *mddev, struct bio * bio)
1030 {
1031         struct r10conf *conf = mddev->private;
1032         struct r10bio *r10_bio;
1033         struct bio *read_bio;
1034         int i;
1035         sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1036         int chunk_sects = chunk_mask + 1;
1037         const int rw = bio_data_dir(bio);
1038         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1039         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1040         unsigned long flags;
1041         struct md_rdev *blocked_rdev;
1042         int plugged;
1043         int sectors_handled;
1044         int max_sectors;
1045         int sectors;
1046
1047         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1048                 md_flush_request(mddev, bio);
1049                 return;
1050         }
1051
1052         /* If this request crosses a chunk boundary, we need to
1053          * split it.  This will only happen for 1 PAGE (or less) requests.
1054          */
1055         if (unlikely((bio->bi_sector & chunk_mask) + (bio->bi_size >> 9)
1056                      > chunk_sects
1057                      && (conf->geo.near_copies < conf->geo.raid_disks
1058                          || conf->prev.near_copies < conf->prev.raid_disks))) {
1059                 struct bio_pair *bp;
1060                 /* Sanity check -- queue functions should prevent this happening */
1061                 if (bio->bi_vcnt != 1 ||
1062                     bio->bi_idx != 0)
1063                         goto bad_map;
1064                 /* This is a one page bio that upper layers
1065                  * refuse to split for us, so we need to split it.
1066                  */
1067                 bp = bio_split(bio,
1068                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
1069
1070                 /* Each of these 'make_request' calls will call 'wait_barrier'.
1071                  * If the first succeeds but the second blocks due to the resync
1072                  * thread raising the barrier, we will deadlock because the
1073                  * IO to the underlying device will be queued in generic_make_request
1074                  * and will never complete, so will never reduce nr_pending.
1075                  * So increment nr_waiting here so no new raise_barriers will
1076                  * succeed, and so the second wait_barrier cannot block.
1077                  */
1078                 spin_lock_irq(&conf->resync_lock);
1079                 conf->nr_waiting++;
1080                 spin_unlock_irq(&conf->resync_lock);
1081
1082                 make_request(mddev, &bp->bio1);
1083                 make_request(mddev, &bp->bio2);
1084
1085                 spin_lock_irq(&conf->resync_lock);
1086                 conf->nr_waiting--;
1087                 wake_up(&conf->wait_barrier);
1088                 spin_unlock_irq(&conf->resync_lock);
1089
1090                 bio_pair_release(bp);
1091                 return;
1092         bad_map:
1093                 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1094                        " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
1095                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
1096
1097                 bio_io_error(bio);
1098                 return;
1099         }
1100
1101         md_write_start(mddev, bio);
1102
1103         /*
1104          * Register the new request and wait if the reconstruction
1105          * thread has put up a bar for new requests.
1106          * Continue immediately if no resync is active currently.
1107          */
1108         wait_barrier(conf);
1109
1110         sectors = bio->bi_size >> 9;
1111         while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1112             bio->bi_sector < conf->reshape_progress &&
1113             bio->bi_sector + sectors > conf->reshape_progress) {
1114                 /* IO spans the reshape position.  Need to wait for
1115                  * reshape to pass
1116                  */
1117                 allow_barrier(conf);
1118                 wait_event(conf->wait_barrier,
1119                            conf->reshape_progress <= bio->bi_sector ||
1120                            conf->reshape_progress >= bio->bi_sector + sectors);
1121                 wait_barrier(conf);
1122         }
1123         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1124             bio_data_dir(bio) == WRITE &&
1125             (mddev->reshape_backwards
1126              ? (bio->bi_sector < conf->reshape_safe &&
1127                 bio->bi_sector + sectors > conf->reshape_progress)
1128              : (bio->bi_sector + sectors > conf->reshape_safe &&
1129                 bio->bi_sector < conf->reshape_progress))) {
1130                 /* Need to update reshape_position in metadata */
1131                 mddev->reshape_position = conf->reshape_progress;
1132                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1133                 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1134                 md_wakeup_thread(mddev->thread);
1135                 wait_event(mddev->sb_wait,
1136                            !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1137
1138                 conf->reshape_safe = mddev->reshape_position;
1139         }
1140
1141         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1142
1143         r10_bio->master_bio = bio;
1144         r10_bio->sectors = sectors;
1145
1146         r10_bio->mddev = mddev;
1147         r10_bio->sector = bio->bi_sector;
1148         r10_bio->state = 0;
1149
1150         /* We might need to issue multiple reads to different
1151          * devices if there are bad blocks around, so we keep
1152          * track of the number of reads in bio->bi_phys_segments.
1153          * If this is 0, there is only one r10_bio and no locking
1154          * will be needed when the request completes.  If it is
1155          * non-zero, then it is the number of not-completed requests.
1156          */
1157         bio->bi_phys_segments = 0;
1158         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1159
1160         if (rw == READ) {
1161                 /*
1162                  * read balancing logic:
1163                  */
1164                 struct md_rdev *rdev;
1165                 int slot;
1166
1167 read_again:
1168                 rdev = read_balance(conf, r10_bio, &max_sectors);
1169                 if (!rdev) {
1170                         raid_end_bio_io(r10_bio);
1171                         return;
1172                 }
1173                 slot = r10_bio->read_slot;
1174
1175                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1176                 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1177                             max_sectors);
1178
1179                 r10_bio->devs[slot].bio = read_bio;
1180                 r10_bio->devs[slot].rdev = rdev;
1181
1182                 read_bio->bi_sector = r10_bio->devs[slot].addr +
1183                         choose_data_offset(r10_bio, rdev);
1184                 read_bio->bi_bdev = rdev->bdev;
1185                 read_bio->bi_end_io = raid10_end_read_request;
1186                 read_bio->bi_rw = READ | do_sync;
1187                 read_bio->bi_private = r10_bio;
1188
1189                 if (max_sectors < r10_bio->sectors) {
1190                         /* Could not read all from this device, so we will
1191                          * need another r10_bio.
1192                          */
1193                         sectors_handled = (r10_bio->sectors + max_sectors
1194                                            - bio->bi_sector);
1195                         r10_bio->sectors = max_sectors;
1196                         spin_lock_irq(&conf->device_lock);
1197                         if (bio->bi_phys_segments == 0)
1198                                 bio->bi_phys_segments = 2;
1199                         else
1200                                 bio->bi_phys_segments++;
1201                         spin_unlock(&conf->device_lock);
1202                         /* Cannot call generic_make_request directly
1203                          * as that will be queued in __generic_make_request
1204                          * and subsequent mempool_alloc might block
1205                          * waiting for it.  so hand bio over to raid10d.
1206                          */
1207                         reschedule_retry(r10_bio);
1208
1209                         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1210
1211                         r10_bio->master_bio = bio;
1212                         r10_bio->sectors = ((bio->bi_size >> 9)
1213                                             - sectors_handled);
1214                         r10_bio->state = 0;
1215                         r10_bio->mddev = mddev;
1216                         r10_bio->sector = bio->bi_sector + sectors_handled;
1217                         goto read_again;
1218                 } else
1219                         generic_make_request(read_bio);
1220                 return;
1221         }
1222
1223         /*
1224          * WRITE:
1225          */
1226         if (conf->pending_count >= max_queued_requests) {
1227                 md_wakeup_thread(mddev->thread);
1228                 wait_event(conf->wait_barrier,
1229                            conf->pending_count < max_queued_requests);
1230         }
1231         /* first select target devices under rcu_lock and
1232          * inc refcount on their rdev.  Record them by setting
1233          * bios[x] to bio
1234          * If there are known/acknowledged bad blocks on any device
1235          * on which we have seen a write error, we want to avoid
1236          * writing to those blocks.  This potentially requires several
1237          * writes to write around the bad blocks.  Each set of writes
1238          * gets its own r10_bio with a set of bios attached.  The number
1239          * of r10_bios is recored in bio->bi_phys_segments just as with
1240          * the read case.
1241          */
1242         plugged = mddev_check_plugged(mddev);
1243
1244         r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1245         raid10_find_phys(conf, r10_bio);
1246 retry_write:
1247         blocked_rdev = NULL;
1248         rcu_read_lock();
1249         max_sectors = r10_bio->sectors;
1250
1251         for (i = 0;  i < conf->copies; i++) {
1252                 int d = r10_bio->devs[i].devnum;
1253                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1254                 struct md_rdev *rrdev = rcu_dereference(
1255                         conf->mirrors[d].replacement);
1256                 if (rdev == rrdev)
1257                         rrdev = NULL;
1258                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1259                         atomic_inc(&rdev->nr_pending);
1260                         blocked_rdev = rdev;
1261                         break;
1262                 }
1263                 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1264                         atomic_inc(&rrdev->nr_pending);
1265                         blocked_rdev = rrdev;
1266                         break;
1267                 }
1268                 if (rrdev && (test_bit(Faulty, &rrdev->flags)
1269                               || test_bit(Unmerged, &rrdev->flags)))
1270                         rrdev = NULL;
1271
1272                 r10_bio->devs[i].bio = NULL;
1273                 r10_bio->devs[i].repl_bio = NULL;
1274                 if (!rdev || test_bit(Faulty, &rdev->flags) ||
1275                     test_bit(Unmerged, &rdev->flags)) {
1276                         set_bit(R10BIO_Degraded, &r10_bio->state);
1277                         continue;
1278                 }
1279                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1280                         sector_t first_bad;
1281                         sector_t dev_sector = r10_bio->devs[i].addr;
1282                         int bad_sectors;
1283                         int is_bad;
1284
1285                         is_bad = is_badblock(rdev, dev_sector,
1286                                              max_sectors,
1287                                              &first_bad, &bad_sectors);
1288                         if (is_bad < 0) {
1289                                 /* Mustn't write here until the bad block
1290                                  * is acknowledged
1291                                  */
1292                                 atomic_inc(&rdev->nr_pending);
1293                                 set_bit(BlockedBadBlocks, &rdev->flags);
1294                                 blocked_rdev = rdev;
1295                                 break;
1296                         }
1297                         if (is_bad && first_bad <= dev_sector) {
1298                                 /* Cannot write here at all */
1299                                 bad_sectors -= (dev_sector - first_bad);
1300                                 if (bad_sectors < max_sectors)
1301                                         /* Mustn't write more than bad_sectors
1302                                          * to other devices yet
1303                                          */
1304                                         max_sectors = bad_sectors;
1305                                 /* We don't set R10BIO_Degraded as that
1306                                  * only applies if the disk is missing,
1307                                  * so it might be re-added, and we want to
1308                                  * know to recover this chunk.
1309                                  * In this case the device is here, and the
1310                                  * fact that this chunk is not in-sync is
1311                                  * recorded in the bad block log.
1312                                  */
1313                                 continue;
1314                         }
1315                         if (is_bad) {
1316                                 int good_sectors = first_bad - dev_sector;
1317                                 if (good_sectors < max_sectors)
1318                                         max_sectors = good_sectors;
1319                         }
1320                 }
1321                 r10_bio->devs[i].bio = bio;
1322                 atomic_inc(&rdev->nr_pending);
1323                 if (rrdev) {
1324                         r10_bio->devs[i].repl_bio = bio;
1325                         atomic_inc(&rrdev->nr_pending);
1326                 }
1327         }
1328         rcu_read_unlock();
1329
1330         if (unlikely(blocked_rdev)) {
1331                 /* Have to wait for this device to get unblocked, then retry */
1332                 int j;
1333                 int d;
1334
1335                 for (j = 0; j < i; j++) {
1336                         if (r10_bio->devs[j].bio) {
1337                                 d = r10_bio->devs[j].devnum;
1338                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1339                         }
1340                         if (r10_bio->devs[j].repl_bio) {
1341                                 struct md_rdev *rdev;
1342                                 d = r10_bio->devs[j].devnum;
1343                                 rdev = conf->mirrors[d].replacement;
1344                                 if (!rdev) {
1345                                         /* Race with remove_disk */
1346                                         smp_mb();
1347                                         rdev = conf->mirrors[d].rdev;
1348                                 }
1349                                 rdev_dec_pending(rdev, mddev);
1350                         }
1351                 }
1352                 allow_barrier(conf);
1353                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1354                 wait_barrier(conf);
1355                 goto retry_write;
1356         }
1357
1358         if (max_sectors < r10_bio->sectors) {
1359                 /* We are splitting this into multiple parts, so
1360                  * we need to prepare for allocating another r10_bio.
1361                  */
1362                 r10_bio->sectors = max_sectors;
1363                 spin_lock_irq(&conf->device_lock);
1364                 if (bio->bi_phys_segments == 0)
1365                         bio->bi_phys_segments = 2;
1366                 else
1367                         bio->bi_phys_segments++;
1368                 spin_unlock_irq(&conf->device_lock);
1369         }
1370         sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1371
1372         atomic_set(&r10_bio->remaining, 1);
1373         bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1374
1375         for (i = 0; i < conf->copies; i++) {
1376                 struct bio *mbio;
1377                 int d = r10_bio->devs[i].devnum;
1378                 if (!r10_bio->devs[i].bio)
1379                         continue;
1380
1381                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1382                 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1383                             max_sectors);
1384                 r10_bio->devs[i].bio = mbio;
1385
1386                 mbio->bi_sector = (r10_bio->devs[i].addr+
1387                                    choose_data_offset(r10_bio,
1388                                                       conf->mirrors[d].rdev));
1389                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1390                 mbio->bi_end_io = raid10_end_write_request;
1391                 mbio->bi_rw = WRITE | do_sync | do_fua;
1392                 mbio->bi_private = r10_bio;
1393
1394                 atomic_inc(&r10_bio->remaining);
1395                 spin_lock_irqsave(&conf->device_lock, flags);
1396                 bio_list_add(&conf->pending_bio_list, mbio);
1397                 conf->pending_count++;
1398                 spin_unlock_irqrestore(&conf->device_lock, flags);
1399
1400                 if (!r10_bio->devs[i].repl_bio)
1401                         continue;
1402
1403                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1404                 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1405                             max_sectors);
1406                 r10_bio->devs[i].repl_bio = mbio;
1407
1408                 /* We are actively writing to the original device
1409                  * so it cannot disappear, so the replacement cannot
1410                  * become NULL here
1411                  */
1412                 mbio->bi_sector = (r10_bio->devs[i].addr +
1413                                    choose_data_offset(
1414                                            r10_bio,
1415                                            conf->mirrors[d].replacement));
1416                 mbio->bi_bdev = conf->mirrors[d].replacement->bdev;
1417                 mbio->bi_end_io = raid10_end_write_request;
1418                 mbio->bi_rw = WRITE | do_sync | do_fua;
1419                 mbio->bi_private = r10_bio;
1420
1421                 atomic_inc(&r10_bio->remaining);
1422                 spin_lock_irqsave(&conf->device_lock, flags);
1423                 bio_list_add(&conf->pending_bio_list, mbio);
1424                 conf->pending_count++;
1425                 spin_unlock_irqrestore(&conf->device_lock, flags);
1426         }
1427
1428         /* Don't remove the bias on 'remaining' (one_write_done) until
1429          * after checking if we need to go around again.
1430          */
1431
1432         if (sectors_handled < (bio->bi_size >> 9)) {
1433                 one_write_done(r10_bio);
1434                 /* We need another r10_bio.  It has already been counted
1435                  * in bio->bi_phys_segments.
1436                  */
1437                 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1438
1439                 r10_bio->master_bio = bio;
1440                 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1441
1442                 r10_bio->mddev = mddev;
1443                 r10_bio->sector = bio->bi_sector + sectors_handled;
1444                 r10_bio->state = 0;
1445                 goto retry_write;
1446         }
1447         one_write_done(r10_bio);
1448
1449         /* In case raid10d snuck in to freeze_array */
1450         wake_up(&conf->wait_barrier);
1451
1452         if (do_sync || !mddev->bitmap || !plugged)
1453                 md_wakeup_thread(mddev->thread);
1454 }
1455
1456 static void status(struct seq_file *seq, struct mddev *mddev)
1457 {
1458         struct r10conf *conf = mddev->private;
1459         int i;
1460
1461         if (conf->geo.near_copies < conf->geo.raid_disks)
1462                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1463         if (conf->geo.near_copies > 1)
1464                 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1465         if (conf->geo.far_copies > 1) {
1466                 if (conf->geo.far_offset)
1467                         seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1468                 else
1469                         seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1470         }
1471         seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1472                                         conf->geo.raid_disks - mddev->degraded);
1473         for (i = 0; i < conf->geo.raid_disks; i++)
1474                 seq_printf(seq, "%s",
1475                               conf->mirrors[i].rdev &&
1476                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1477         seq_printf(seq, "]");
1478 }
1479
1480 /* check if there are enough drives for
1481  * every block to appear on atleast one.
1482  * Don't consider the device numbered 'ignore'
1483  * as we might be about to remove it.
1484  */
1485 static int _enough(struct r10conf *conf, struct geom *geo, int ignore)
1486 {
1487         int first = 0;
1488
1489         do {
1490                 int n = conf->copies;
1491                 int cnt = 0;
1492                 while (n--) {
1493                         if (conf->mirrors[first].rdev &&
1494                             first != ignore)
1495                                 cnt++;
1496                         first = (first+1) % geo->raid_disks;
1497                 }
1498                 if (cnt == 0)
1499                         return 0;
1500         } while (first != 0);
1501         return 1;
1502 }
1503
1504 static int enough(struct r10conf *conf, int ignore)
1505 {
1506         return _enough(conf, &conf->geo, ignore) &&
1507                 _enough(conf, &conf->prev, ignore);
1508 }
1509
1510 static void error(struct mddev *mddev, struct md_rdev *rdev)
1511 {
1512         char b[BDEVNAME_SIZE];
1513         struct r10conf *conf = mddev->private;
1514
1515         /*
1516          * If it is not operational, then we have already marked it as dead
1517          * else if it is the last working disks, ignore the error, let the
1518          * next level up know.
1519          * else mark the drive as failed
1520          */
1521         if (test_bit(In_sync, &rdev->flags)
1522             && !enough(conf, rdev->raid_disk))
1523                 /*
1524                  * Don't fail the drive, just return an IO error.
1525                  */
1526                 return;
1527         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1528                 unsigned long flags;
1529                 spin_lock_irqsave(&conf->device_lock, flags);
1530                 mddev->degraded++;
1531                 spin_unlock_irqrestore(&conf->device_lock, flags);
1532                 /*
1533                  * if recovery is running, make sure it aborts.
1534                  */
1535                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1536         }
1537         set_bit(Blocked, &rdev->flags);
1538         set_bit(Faulty, &rdev->flags);
1539         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1540         printk(KERN_ALERT
1541                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1542                "md/raid10:%s: Operation continuing on %d devices.\n",
1543                mdname(mddev), bdevname(rdev->bdev, b),
1544                mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1545 }
1546
1547 static void print_conf(struct r10conf *conf)
1548 {
1549         int i;
1550         struct mirror_info *tmp;
1551
1552         printk(KERN_DEBUG "RAID10 conf printout:\n");
1553         if (!conf) {
1554                 printk(KERN_DEBUG "(!conf)\n");
1555                 return;
1556         }
1557         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1558                 conf->geo.raid_disks);
1559
1560         for (i = 0; i < conf->geo.raid_disks; i++) {
1561                 char b[BDEVNAME_SIZE];
1562                 tmp = conf->mirrors + i;
1563                 if (tmp->rdev)
1564                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1565                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1566                                 !test_bit(Faulty, &tmp->rdev->flags),
1567                                 bdevname(tmp->rdev->bdev,b));
1568         }
1569 }
1570
1571 static void close_sync(struct r10conf *conf)
1572 {
1573         wait_barrier(conf);
1574         allow_barrier(conf);
1575
1576         mempool_destroy(conf->r10buf_pool);
1577         conf->r10buf_pool = NULL;
1578 }
1579
1580 static int raid10_spare_active(struct mddev *mddev)
1581 {
1582         int i;
1583         struct r10conf *conf = mddev->private;
1584         struct mirror_info *tmp;
1585         int count = 0;
1586         unsigned long flags;
1587
1588         /*
1589          * Find all non-in_sync disks within the RAID10 configuration
1590          * and mark them in_sync
1591          */
1592         for (i = 0; i < conf->geo.raid_disks; i++) {
1593                 tmp = conf->mirrors + i;
1594                 if (tmp->replacement
1595                     && tmp->replacement->recovery_offset == MaxSector
1596                     && !test_bit(Faulty, &tmp->replacement->flags)
1597                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1598                         /* Replacement has just become active */
1599                         if (!tmp->rdev
1600                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1601                                 count++;
1602                         if (tmp->rdev) {
1603                                 /* Replaced device not technically faulty,
1604                                  * but we need to be sure it gets removed
1605                                  * and never re-added.
1606                                  */
1607                                 set_bit(Faulty, &tmp->rdev->flags);
1608                                 sysfs_notify_dirent_safe(
1609                                         tmp->rdev->sysfs_state);
1610                         }
1611                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1612                 } else if (tmp->rdev
1613                            && !test_bit(Faulty, &tmp->rdev->flags)
1614                            && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1615                         count++;
1616                         sysfs_notify_dirent(tmp->rdev->sysfs_state);
1617                 }
1618         }
1619         spin_lock_irqsave(&conf->device_lock, flags);
1620         mddev->degraded -= count;
1621         spin_unlock_irqrestore(&conf->device_lock, flags);
1622
1623         print_conf(conf);
1624         return count;
1625 }
1626
1627
1628 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1629 {
1630         struct r10conf *conf = mddev->private;
1631         int err = -EEXIST;
1632         int mirror;
1633         int first = 0;
1634         int last = conf->geo.raid_disks - 1;
1635         struct request_queue *q = bdev_get_queue(rdev->bdev);
1636
1637         if (mddev->recovery_cp < MaxSector)
1638                 /* only hot-add to in-sync arrays, as recovery is
1639                  * very different from resync
1640                  */
1641                 return -EBUSY;
1642         if (rdev->saved_raid_disk < 0 && !_enough(conf, &conf->prev, -1))
1643                 return -EINVAL;
1644
1645         if (rdev->raid_disk >= 0)
1646                 first = last = rdev->raid_disk;
1647
1648         if (q->merge_bvec_fn) {
1649                 set_bit(Unmerged, &rdev->flags);
1650                 mddev->merge_check_needed = 1;
1651         }
1652
1653         if (rdev->saved_raid_disk >= first &&
1654             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1655                 mirror = rdev->saved_raid_disk;
1656         else
1657                 mirror = first;
1658         for ( ; mirror <= last ; mirror++) {
1659                 struct mirror_info *p = &conf->mirrors[mirror];
1660                 if (p->recovery_disabled == mddev->recovery_disabled)
1661                         continue;
1662                 if (p->rdev) {
1663                         if (!test_bit(WantReplacement, &p->rdev->flags) ||
1664                             p->replacement != NULL)
1665                                 continue;
1666                         clear_bit(In_sync, &rdev->flags);
1667                         set_bit(Replacement, &rdev->flags);
1668                         rdev->raid_disk = mirror;
1669                         err = 0;
1670                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1671                                           rdev->data_offset << 9);
1672                         conf->fullsync = 1;
1673                         rcu_assign_pointer(p->replacement, rdev);
1674                         break;
1675                 }
1676
1677                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1678                                   rdev->data_offset << 9);
1679
1680                 p->head_position = 0;
1681                 p->recovery_disabled = mddev->recovery_disabled - 1;
1682                 rdev->raid_disk = mirror;
1683                 err = 0;
1684                 if (rdev->saved_raid_disk != mirror)
1685                         conf->fullsync = 1;
1686                 rcu_assign_pointer(p->rdev, rdev);
1687                 break;
1688         }
1689         if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1690                 /* Some requests might not have seen this new
1691                  * merge_bvec_fn.  We must wait for them to complete
1692                  * before merging the device fully.
1693                  * First we make sure any code which has tested
1694                  * our function has submitted the request, then
1695                  * we wait for all outstanding requests to complete.
1696                  */
1697                 synchronize_sched();
1698                 raise_barrier(conf, 0);
1699                 lower_barrier(conf);
1700                 clear_bit(Unmerged, &rdev->flags);
1701         }
1702         md_integrity_add_rdev(rdev, mddev);
1703         print_conf(conf);
1704         return err;
1705 }
1706
1707 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1708 {
1709         struct r10conf *conf = mddev->private;
1710         int err = 0;
1711         int number = rdev->raid_disk;
1712         struct md_rdev **rdevp;
1713         struct mirror_info *p = conf->mirrors + number;
1714
1715         print_conf(conf);
1716         if (rdev == p->rdev)
1717                 rdevp = &p->rdev;
1718         else if (rdev == p->replacement)
1719                 rdevp = &p->replacement;
1720         else
1721                 return 0;
1722
1723         if (test_bit(In_sync, &rdev->flags) ||
1724             atomic_read(&rdev->nr_pending)) {
1725                 err = -EBUSY;
1726                 goto abort;
1727         }
1728         /* Only remove faulty devices if recovery
1729          * is not possible.
1730          */
1731         if (!test_bit(Faulty, &rdev->flags) &&
1732             mddev->recovery_disabled != p->recovery_disabled &&
1733             (!p->replacement || p->replacement == rdev) &&
1734             number < conf->geo.raid_disks &&
1735             enough(conf, -1)) {
1736                 err = -EBUSY;
1737                 goto abort;
1738         }
1739         *rdevp = NULL;
1740         synchronize_rcu();
1741         if (atomic_read(&rdev->nr_pending)) {
1742                 /* lost the race, try later */
1743                 err = -EBUSY;
1744                 *rdevp = rdev;
1745                 goto abort;
1746         } else if (p->replacement) {
1747                 /* We must have just cleared 'rdev' */
1748                 p->rdev = p->replacement;
1749                 clear_bit(Replacement, &p->replacement->flags);
1750                 smp_mb(); /* Make sure other CPUs may see both as identical
1751                            * but will never see neither -- if they are careful.
1752                            */
1753                 p->replacement = NULL;
1754                 clear_bit(WantReplacement, &rdev->flags);
1755         } else
1756                 /* We might have just remove the Replacement as faulty
1757                  * Clear the flag just in case
1758                  */
1759                 clear_bit(WantReplacement, &rdev->flags);
1760
1761         err = md_integrity_register(mddev);
1762
1763 abort:
1764
1765         print_conf(conf);
1766         return err;
1767 }
1768
1769
1770 static void end_sync_read(struct bio *bio, int error)
1771 {
1772         struct r10bio *r10_bio = bio->bi_private;
1773         struct r10conf *conf = r10_bio->mddev->private;
1774         int d;
1775
1776         if (bio == r10_bio->master_bio) {
1777                 /* this is a reshape read */
1778                 d = r10_bio->read_slot; /* really the read dev */
1779         } else
1780                 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1781
1782         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1783                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1784         else
1785                 /* The write handler will notice the lack of
1786                  * R10BIO_Uptodate and record any errors etc
1787                  */
1788                 atomic_add(r10_bio->sectors,
1789                            &conf->mirrors[d].rdev->corrected_errors);
1790
1791         /* for reconstruct, we always reschedule after a read.
1792          * for resync, only after all reads
1793          */
1794         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1795         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1796             atomic_dec_and_test(&r10_bio->remaining)) {
1797                 /* we have read all the blocks,
1798                  * do the comparison in process context in raid10d
1799                  */
1800                 reschedule_retry(r10_bio);
1801         }
1802 }
1803
1804 static void end_sync_request(struct r10bio *r10_bio)
1805 {
1806         struct mddev *mddev = r10_bio->mddev;
1807
1808         while (atomic_dec_and_test(&r10_bio->remaining)) {
1809                 if (r10_bio->master_bio == NULL) {
1810                         /* the primary of several recovery bios */
1811                         sector_t s = r10_bio->sectors;
1812                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1813                             test_bit(R10BIO_WriteError, &r10_bio->state))
1814                                 reschedule_retry(r10_bio);
1815                         else
1816                                 put_buf(r10_bio);
1817                         md_done_sync(mddev, s, 1);
1818                         break;
1819                 } else {
1820                         struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1821                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1822                             test_bit(R10BIO_WriteError, &r10_bio->state))
1823                                 reschedule_retry(r10_bio);
1824                         else
1825                                 put_buf(r10_bio);
1826                         r10_bio = r10_bio2;
1827                 }
1828         }
1829 }
1830
1831 static void end_sync_write(struct bio *bio, int error)
1832 {
1833         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1834         struct r10bio *r10_bio = bio->bi_private;
1835         struct mddev *mddev = r10_bio->mddev;
1836         struct r10conf *conf = mddev->private;
1837         int d;
1838         sector_t first_bad;
1839         int bad_sectors;
1840         int slot;
1841         int repl;
1842         struct md_rdev *rdev = NULL;
1843
1844         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1845         if (repl)
1846                 rdev = conf->mirrors[d].replacement;
1847         else
1848                 rdev = conf->mirrors[d].rdev;
1849
1850         if (!uptodate) {
1851                 if (repl)
1852                         md_error(mddev, rdev);
1853                 else {
1854                         set_bit(WriteErrorSeen, &rdev->flags);
1855                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
1856                                 set_bit(MD_RECOVERY_NEEDED,
1857                                         &rdev->mddev->recovery);
1858                         set_bit(R10BIO_WriteError, &r10_bio->state);
1859                 }
1860         } else if (is_badblock(rdev,
1861                              r10_bio->devs[slot].addr,
1862                              r10_bio->sectors,
1863                              &first_bad, &bad_sectors))
1864                 set_bit(R10BIO_MadeGood, &r10_bio->state);
1865
1866         rdev_dec_pending(rdev, mddev);
1867
1868         end_sync_request(r10_bio);
1869 }
1870
1871 /*
1872  * Note: sync and recover and handled very differently for raid10
1873  * This code is for resync.
1874  * For resync, we read through virtual addresses and read all blocks.
1875  * If there is any error, we schedule a write.  The lowest numbered
1876  * drive is authoritative.
1877  * However requests come for physical address, so we need to map.
1878  * For every physical address there are raid_disks/copies virtual addresses,
1879  * which is always are least one, but is not necessarly an integer.
1880  * This means that a physical address can span multiple chunks, so we may
1881  * have to submit multiple io requests for a single sync request.
1882  */
1883 /*
1884  * We check if all blocks are in-sync and only write to blocks that
1885  * aren't in sync
1886  */
1887 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1888 {
1889         struct r10conf *conf = mddev->private;
1890         int i, first;
1891         struct bio *tbio, *fbio;
1892         int vcnt;
1893
1894         atomic_set(&r10_bio->remaining, 1);
1895
1896         /* find the first device with a block */
1897         for (i=0; i<conf->copies; i++)
1898                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1899                         break;
1900
1901         if (i == conf->copies)
1902                 goto done;
1903
1904         first = i;
1905         fbio = r10_bio->devs[i].bio;
1906
1907         vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
1908         /* now find blocks with errors */
1909         for (i=0 ; i < conf->copies ; i++) {
1910                 int  j, d;
1911
1912                 tbio = r10_bio->devs[i].bio;
1913
1914                 if (tbio->bi_end_io != end_sync_read)
1915                         continue;
1916                 if (i == first)
1917                         continue;
1918                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1919                         /* We know that the bi_io_vec layout is the same for
1920                          * both 'first' and 'i', so we just compare them.
1921                          * All vec entries are PAGE_SIZE;
1922                          */
1923                         for (j = 0; j < vcnt; j++)
1924                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1925                                            page_address(tbio->bi_io_vec[j].bv_page),
1926                                            fbio->bi_io_vec[j].bv_len))
1927                                         break;
1928                         if (j == vcnt)
1929                                 continue;
1930                         mddev->resync_mismatches += r10_bio->sectors;
1931                         if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1932                                 /* Don't fix anything. */
1933                                 continue;
1934                 }
1935                 /* Ok, we need to write this bio, either to correct an
1936                  * inconsistency or to correct an unreadable block.
1937                  * First we need to fixup bv_offset, bv_len and
1938                  * bi_vecs, as the read request might have corrupted these
1939                  */
1940                 tbio->bi_vcnt = vcnt;
1941                 tbio->bi_size = r10_bio->sectors << 9;
1942                 tbio->bi_idx = 0;
1943                 tbio->bi_phys_segments = 0;
1944                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1945                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1946                 tbio->bi_next = NULL;
1947                 tbio->bi_rw = WRITE;
1948                 tbio->bi_private = r10_bio;
1949                 tbio->bi_sector = r10_bio->devs[i].addr;
1950
1951                 for (j=0; j < vcnt ; j++) {
1952                         tbio->bi_io_vec[j].bv_offset = 0;
1953                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1954
1955                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1956                                page_address(fbio->bi_io_vec[j].bv_page),
1957                                PAGE_SIZE);
1958                 }
1959                 tbio->bi_end_io = end_sync_write;
1960
1961                 d = r10_bio->devs[i].devnum;
1962                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1963                 atomic_inc(&r10_bio->remaining);
1964                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1965
1966                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1967                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1968                 generic_make_request(tbio);
1969         }
1970
1971         /* Now write out to any replacement devices
1972          * that are active
1973          */
1974         for (i = 0; i < conf->copies; i++) {
1975                 int j, d;
1976
1977                 tbio = r10_bio->devs[i].repl_bio;
1978                 if (!tbio || !tbio->bi_end_io)
1979                         continue;
1980                 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
1981                     && r10_bio->devs[i].bio != fbio)
1982                         for (j = 0; j < vcnt; j++)
1983                                 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1984                                        page_address(fbio->bi_io_vec[j].bv_page),
1985                                        PAGE_SIZE);
1986                 d = r10_bio->devs[i].devnum;
1987                 atomic_inc(&r10_bio->remaining);
1988                 md_sync_acct(conf->mirrors[d].replacement->bdev,
1989                              tbio->bi_size >> 9);
1990                 generic_make_request(tbio);
1991         }
1992
1993 done:
1994         if (atomic_dec_and_test(&r10_bio->remaining)) {
1995                 md_done_sync(mddev, r10_bio->sectors, 1);
1996                 put_buf(r10_bio);
1997         }
1998 }
1999
2000 /*
2001  * Now for the recovery code.
2002  * Recovery happens across physical sectors.
2003  * We recover all non-is_sync drives by finding the virtual address of
2004  * each, and then choose a working drive that also has that virt address.
2005  * There is a separate r10_bio for each non-in_sync drive.
2006  * Only the first two slots are in use. The first for reading,
2007  * The second for writing.
2008  *
2009  */
2010 static void fix_recovery_read_error(struct r10bio *r10_bio)
2011 {
2012         /* We got a read error during recovery.
2013          * We repeat the read in smaller page-sized sections.
2014          * If a read succeeds, write it to the new device or record
2015          * a bad block if we cannot.
2016          * If a read fails, record a bad block on both old and
2017          * new devices.
2018          */
2019         struct mddev *mddev = r10_bio->mddev;
2020         struct r10conf *conf = mddev->private;
2021         struct bio *bio = r10_bio->devs[0].bio;
2022         sector_t sect = 0;
2023         int sectors = r10_bio->sectors;
2024         int idx = 0;
2025         int dr = r10_bio->devs[0].devnum;
2026         int dw = r10_bio->devs[1].devnum;
2027
2028         while (sectors) {
2029                 int s = sectors;
2030                 struct md_rdev *rdev;
2031                 sector_t addr;
2032                 int ok;
2033
2034                 if (s > (PAGE_SIZE>>9))
2035                         s = PAGE_SIZE >> 9;
2036
2037                 rdev = conf->mirrors[dr].rdev;
2038                 addr = r10_bio->devs[0].addr + sect,
2039                 ok = sync_page_io(rdev,
2040                                   addr,
2041                                   s << 9,
2042                                   bio->bi_io_vec[idx].bv_page,
2043                                   READ, false);
2044                 if (ok) {
2045                         rdev = conf->mirrors[dw].rdev;
2046                         addr = r10_bio->devs[1].addr + sect;
2047                         ok = sync_page_io(rdev,
2048                                           addr,
2049                                           s << 9,
2050                                           bio->bi_io_vec[idx].bv_page,
2051                                           WRITE, false);
2052                         if (!ok) {
2053                                 set_bit(WriteErrorSeen, &rdev->flags);
2054                                 if (!test_and_set_bit(WantReplacement,
2055                                                       &rdev->flags))
2056                                         set_bit(MD_RECOVERY_NEEDED,
2057                                                 &rdev->mddev->recovery);
2058                         }
2059                 }
2060                 if (!ok) {
2061                         /* We don't worry if we cannot set a bad block -
2062                          * it really is bad so there is no loss in not
2063                          * recording it yet
2064                          */
2065                         rdev_set_badblocks(rdev, addr, s, 0);
2066
2067                         if (rdev != conf->mirrors[dw].rdev) {
2068                                 /* need bad block on destination too */
2069                                 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2070                                 addr = r10_bio->devs[1].addr + sect;
2071                                 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2072                                 if (!ok) {
2073                                         /* just abort the recovery */
2074                                         printk(KERN_NOTICE
2075                                                "md/raid10:%s: recovery aborted"
2076                                                " due to read error\n",
2077                                                mdname(mddev));
2078
2079                                         conf->mirrors[dw].recovery_disabled
2080                                                 = mddev->recovery_disabled;
2081                                         set_bit(MD_RECOVERY_INTR,
2082                                                 &mddev->recovery);
2083                                         break;
2084                                 }
2085                         }
2086                 }
2087
2088                 sectors -= s;
2089                 sect += s;
2090                 idx++;
2091         }
2092 }
2093
2094 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2095 {
2096         struct r10conf *conf = mddev->private;
2097         int d;
2098         struct bio *wbio, *wbio2;
2099
2100         if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2101                 fix_recovery_read_error(r10_bio);
2102                 end_sync_request(r10_bio);
2103                 return;
2104         }
2105
2106         /*
2107          * share the pages with the first bio
2108          * and submit the write request
2109          */
2110         d = r10_bio->devs[1].devnum;
2111         wbio = r10_bio->devs[1].bio;
2112         wbio2 = r10_bio->devs[1].repl_bio;
2113         if (wbio->bi_end_io) {
2114                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2115                 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
2116                 generic_make_request(wbio);
2117         }
2118         if (wbio2 && wbio2->bi_end_io) {
2119                 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2120                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2121                              wbio2->bi_size >> 9);
2122                 generic_make_request(wbio2);
2123         }
2124 }
2125
2126
2127 /*
2128  * Used by fix_read_error() to decay the per rdev read_errors.
2129  * We halve the read error count for every hour that has elapsed
2130  * since the last recorded read error.
2131  *
2132  */
2133 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2134 {
2135         struct timespec cur_time_mon;
2136         unsigned long hours_since_last;
2137         unsigned int read_errors = atomic_read(&rdev->read_errors);
2138
2139         ktime_get_ts(&cur_time_mon);
2140
2141         if (rdev->last_read_error.tv_sec == 0 &&
2142             rdev->last_read_error.tv_nsec == 0) {
2143                 /* first time we've seen a read error */
2144                 rdev->last_read_error = cur_time_mon;
2145                 return;
2146         }
2147
2148         hours_since_last = (cur_time_mon.tv_sec -
2149                             rdev->last_read_error.tv_sec) / 3600;
2150
2151         rdev->last_read_error = cur_time_mon;
2152
2153         /*
2154          * if hours_since_last is > the number of bits in read_errors
2155          * just set read errors to 0. We do this to avoid
2156          * overflowing the shift of read_errors by hours_since_last.
2157          */
2158         if (hours_since_last >= 8 * sizeof(read_errors))
2159                 atomic_set(&rdev->read_errors, 0);
2160         else
2161                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2162 }
2163
2164 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2165                             int sectors, struct page *page, int rw)
2166 {
2167         sector_t first_bad;
2168         int bad_sectors;
2169
2170         if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2171             && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2172                 return -1;
2173         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2174                 /* success */
2175                 return 1;
2176         if (rw == WRITE) {
2177                 set_bit(WriteErrorSeen, &rdev->flags);
2178                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2179                         set_bit(MD_RECOVERY_NEEDED,
2180                                 &rdev->mddev->recovery);
2181         }
2182         /* need to record an error - either for the block or the device */
2183         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2184                 md_error(rdev->mddev, rdev);
2185         return 0;
2186 }
2187
2188 /*
2189  * This is a kernel thread which:
2190  *
2191  *      1.      Retries failed read operations on working mirrors.
2192  *      2.      Updates the raid superblock when problems encounter.
2193  *      3.      Performs writes following reads for array synchronising.
2194  */
2195
2196 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2197 {
2198         int sect = 0; /* Offset from r10_bio->sector */
2199         int sectors = r10_bio->sectors;
2200         struct md_rdev*rdev;
2201         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2202         int d = r10_bio->devs[r10_bio->read_slot].devnum;
2203
2204         /* still own a reference to this rdev, so it cannot
2205          * have been cleared recently.
2206          */
2207         rdev = conf->mirrors[d].rdev;
2208
2209         if (test_bit(Faulty, &rdev->flags))
2210                 /* drive has already been failed, just ignore any
2211                    more fix_read_error() attempts */
2212                 return;
2213
2214         check_decay_read_errors(mddev, rdev);
2215         atomic_inc(&rdev->read_errors);
2216         if (atomic_read(&rdev->read_errors) > max_read_errors) {
2217                 char b[BDEVNAME_SIZE];
2218                 bdevname(rdev->bdev, b);
2219
2220                 printk(KERN_NOTICE
2221                        "md/raid10:%s: %s: Raid device exceeded "
2222                        "read_error threshold [cur %d:max %d]\n",
2223                        mdname(mddev), b,
2224                        atomic_read(&rdev->read_errors), max_read_errors);
2225                 printk(KERN_NOTICE
2226                        "md/raid10:%s: %s: Failing raid device\n",
2227                        mdname(mddev), b);
2228                 md_error(mddev, conf->mirrors[d].rdev);
2229                 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2230                 return;
2231         }
2232
2233         while(sectors) {
2234                 int s = sectors;
2235                 int sl = r10_bio->read_slot;
2236                 int success = 0;
2237                 int start;
2238
2239                 if (s > (PAGE_SIZE>>9))
2240                         s = PAGE_SIZE >> 9;
2241
2242                 rcu_read_lock();
2243                 do {
2244                         sector_t first_bad;
2245                         int bad_sectors;
2246
2247                         d = r10_bio->devs[sl].devnum;
2248                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2249                         if (rdev &&
2250                             !test_bit(Unmerged, &rdev->flags) &&
2251                             test_bit(In_sync, &rdev->flags) &&
2252                             is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2253                                         &first_bad, &bad_sectors) == 0) {
2254                                 atomic_inc(&rdev->nr_pending);
2255                                 rcu_read_unlock();
2256                                 success = sync_page_io(rdev,
2257                                                        r10_bio->devs[sl].addr +
2258                                                        sect,
2259                                                        s<<9,
2260                                                        conf->tmppage, READ, false);
2261                                 rdev_dec_pending(rdev, mddev);
2262                                 rcu_read_lock();
2263                                 if (success)
2264                                         break;
2265                         }
2266                         sl++;
2267                         if (sl == conf->copies)
2268                                 sl = 0;
2269                 } while (!success && sl != r10_bio->read_slot);
2270                 rcu_read_unlock();
2271
2272                 if (!success) {
2273                         /* Cannot read from anywhere, just mark the block
2274                          * as bad on the first device to discourage future
2275                          * reads.
2276                          */
2277                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2278                         rdev = conf->mirrors[dn].rdev;
2279
2280                         if (!rdev_set_badblocks(
2281                                     rdev,
2282                                     r10_bio->devs[r10_bio->read_slot].addr
2283                                     + sect,
2284                                     s, 0)) {
2285                                 md_error(mddev, rdev);
2286                                 r10_bio->devs[r10_bio->read_slot].bio
2287                                         = IO_BLOCKED;
2288                         }
2289                         break;
2290                 }
2291
2292                 start = sl;
2293                 /* write it back and re-read */
2294                 rcu_read_lock();
2295                 while (sl != r10_bio->read_slot) {
2296                         char b[BDEVNAME_SIZE];
2297
2298                         if (sl==0)
2299                                 sl = conf->copies;
2300                         sl--;
2301                         d = r10_bio->devs[sl].devnum;
2302                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2303                         if (!rdev ||
2304                             test_bit(Unmerged, &rdev->flags) ||
2305                             !test_bit(In_sync, &rdev->flags))
2306                                 continue;
2307
2308                         atomic_inc(&rdev->nr_pending);
2309                         rcu_read_unlock();
2310                         if (r10_sync_page_io(rdev,
2311                                              r10_bio->devs[sl].addr +
2312                                              sect,
2313                                              s<<9, conf->tmppage, WRITE)
2314                             == 0) {
2315                                 /* Well, this device is dead */
2316                                 printk(KERN_NOTICE
2317                                        "md/raid10:%s: read correction "
2318                                        "write failed"
2319                                        " (%d sectors at %llu on %s)\n",
2320                                        mdname(mddev), s,
2321                                        (unsigned long long)(
2322                                                sect +
2323                                                choose_data_offset(r10_bio,
2324                                                                   rdev)),
2325                                        bdevname(rdev->bdev, b));
2326                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2327                                        "drive\n",
2328                                        mdname(mddev),
2329                                        bdevname(rdev->bdev, b));
2330                         }
2331                         rdev_dec_pending(rdev, mddev);
2332                         rcu_read_lock();
2333                 }
2334                 sl = start;
2335                 while (sl != r10_bio->read_slot) {
2336                         char b[BDEVNAME_SIZE];
2337
2338                         if (sl==0)
2339                                 sl = conf->copies;
2340                         sl--;
2341                         d = r10_bio->devs[sl].devnum;
2342                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2343                         if (!rdev ||
2344                             !test_bit(In_sync, &rdev->flags))
2345                                 continue;
2346
2347                         atomic_inc(&rdev->nr_pending);
2348                         rcu_read_unlock();
2349                         switch (r10_sync_page_io(rdev,
2350                                              r10_bio->devs[sl].addr +
2351                                              sect,
2352                                              s<<9, conf->tmppage,
2353                                                  READ)) {
2354                         case 0:
2355                                 /* Well, this device is dead */
2356                                 printk(KERN_NOTICE
2357                                        "md/raid10:%s: unable to read back "
2358                                        "corrected sectors"
2359                                        " (%d sectors at %llu on %s)\n",
2360                                        mdname(mddev), s,
2361                                        (unsigned long long)(
2362                                                sect +
2363                                                choose_data_offset(r10_bio, rdev)),
2364                                        bdevname(rdev->bdev, b));
2365                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2366                                        "drive\n",
2367                                        mdname(mddev),
2368                                        bdevname(rdev->bdev, b));
2369                                 break;
2370                         case 1:
2371                                 printk(KERN_INFO
2372                                        "md/raid10:%s: read error corrected"
2373                                        " (%d sectors at %llu on %s)\n",
2374                                        mdname(mddev), s,
2375                                        (unsigned long long)(
2376                                                sect +
2377                                                choose_data_offset(r10_bio, rdev)),
2378                                        bdevname(rdev->bdev, b));
2379                                 atomic_add(s, &rdev->corrected_errors);
2380                         }
2381
2382                         rdev_dec_pending(rdev, mddev);
2383                         rcu_read_lock();
2384                 }
2385                 rcu_read_unlock();
2386
2387                 sectors -= s;
2388                 sect += s;
2389         }
2390 }
2391
2392 static void bi_complete(struct bio *bio, int error)
2393 {
2394         complete((struct completion *)bio->bi_private);
2395 }
2396
2397 static int submit_bio_wait(int rw, struct bio *bio)
2398 {
2399         struct completion event;
2400         rw |= REQ_SYNC;
2401
2402         init_completion(&event);
2403         bio->bi_private = &event;
2404         bio->bi_end_io = bi_complete;
2405         submit_bio(rw, bio);
2406         wait_for_completion(&event);
2407
2408         return test_bit(BIO_UPTODATE, &bio->bi_flags);
2409 }
2410
2411 static int narrow_write_error(struct r10bio *r10_bio, int i)
2412 {
2413         struct bio *bio = r10_bio->master_bio;
2414         struct mddev *mddev = r10_bio->mddev;
2415         struct r10conf *conf = mddev->private;
2416         struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2417         /* bio has the data to be written to slot 'i' where
2418          * we just recently had a write error.
2419          * We repeatedly clone the bio and trim down to one block,
2420          * then try the write.  Where the write fails we record
2421          * a bad block.
2422          * It is conceivable that the bio doesn't exactly align with
2423          * blocks.  We must handle this.
2424          *
2425          * We currently own a reference to the rdev.
2426          */
2427
2428         int block_sectors;
2429         sector_t sector;
2430         int sectors;
2431         int sect_to_write = r10_bio->sectors;
2432         int ok = 1;
2433
2434         if (rdev->badblocks.shift < 0)
2435                 return 0;
2436
2437         block_sectors = 1 << rdev->badblocks.shift;
2438         sector = r10_bio->sector;
2439         sectors = ((r10_bio->sector + block_sectors)
2440                    & ~(sector_t)(block_sectors - 1))
2441                 - sector;
2442
2443         while (sect_to_write) {
2444                 struct bio *wbio;
2445                 if (sectors > sect_to_write)
2446                         sectors = sect_to_write;
2447                 /* Write at 'sector' for 'sectors' */
2448                 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2449                 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2450                 wbio->bi_sector = (r10_bio->devs[i].addr+
2451                                    choose_data_offset(r10_bio, rdev) +
2452                                    (sector - r10_bio->sector));
2453                 wbio->bi_bdev = rdev->bdev;
2454                 if (submit_bio_wait(WRITE, wbio) == 0)
2455                         /* Failure! */
2456                         ok = rdev_set_badblocks(rdev, sector,
2457                                                 sectors, 0)
2458                                 && ok;
2459
2460                 bio_put(wbio);
2461                 sect_to_write -= sectors;
2462                 sector += sectors;
2463                 sectors = block_sectors;
2464         }
2465         return ok;
2466 }
2467
2468 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2469 {
2470         int slot = r10_bio->read_slot;
2471         struct bio *bio;
2472         struct r10conf *conf = mddev->private;
2473         struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2474         char b[BDEVNAME_SIZE];
2475         unsigned long do_sync;
2476         int max_sectors;
2477
2478         /* we got a read error. Maybe the drive is bad.  Maybe just
2479          * the block and we can fix it.
2480          * We freeze all other IO, and try reading the block from
2481          * other devices.  When we find one, we re-write
2482          * and check it that fixes the read error.
2483          * This is all done synchronously while the array is
2484          * frozen.
2485          */
2486         bio = r10_bio->devs[slot].bio;
2487         bdevname(bio->bi_bdev, b);
2488         bio_put(bio);
2489         r10_bio->devs[slot].bio = NULL;
2490
2491         if (mddev->ro == 0) {
2492                 freeze_array(conf);
2493                 fix_read_error(conf, mddev, r10_bio);
2494                 unfreeze_array(conf);
2495         } else
2496                 r10_bio->devs[slot].bio = IO_BLOCKED;
2497
2498         rdev_dec_pending(rdev, mddev);
2499
2500 read_more:
2501         rdev = read_balance(conf, r10_bio, &max_sectors);
2502         if (rdev == NULL) {
2503                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2504                        " read error for block %llu\n",
2505                        mdname(mddev), b,
2506                        (unsigned long long)r10_bio->sector);
2507                 raid_end_bio_io(r10_bio);
2508                 return;
2509         }
2510
2511         do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2512         slot = r10_bio->read_slot;
2513         printk_ratelimited(
2514                 KERN_ERR
2515                 "md/raid10:%s: %s: redirecting"
2516                 "sector %llu to another mirror\n",
2517                 mdname(mddev),
2518                 bdevname(rdev->bdev, b),
2519                 (unsigned long long)r10_bio->sector);
2520         bio = bio_clone_mddev(r10_bio->master_bio,
2521                               GFP_NOIO, mddev);
2522         md_trim_bio(bio,
2523                     r10_bio->sector - bio->bi_sector,
2524                     max_sectors);
2525         r10_bio->devs[slot].bio = bio;
2526         r10_bio->devs[slot].rdev = rdev;
2527         bio->bi_sector = r10_bio->devs[slot].addr
2528                 + choose_data_offset(r10_bio, rdev);
2529         bio->bi_bdev = rdev->bdev;
2530         bio->bi_rw = READ | do_sync;
2531         bio->bi_private = r10_bio;
2532         bio->bi_end_io = raid10_end_read_request;
2533         if (max_sectors < r10_bio->sectors) {
2534                 /* Drat - have to split this up more */
2535                 struct bio *mbio = r10_bio->master_bio;
2536                 int sectors_handled =
2537                         r10_bio->sector + max_sectors
2538                         - mbio->bi_sector;
2539                 r10_bio->sectors = max_sectors;
2540                 spin_lock_irq(&conf->device_lock);
2541                 if (mbio->bi_phys_segments == 0)
2542                         mbio->bi_phys_segments = 2;
2543                 else
2544                         mbio->bi_phys_segments++;
2545                 spin_unlock_irq(&conf->device_lock);
2546                 generic_make_request(bio);
2547
2548                 r10_bio = mempool_alloc(conf->r10bio_pool,
2549                                         GFP_NOIO);
2550                 r10_bio->master_bio = mbio;
2551                 r10_bio->sectors = (mbio->bi_size >> 9)
2552                         - sectors_handled;
2553                 r10_bio->state = 0;
2554                 set_bit(R10BIO_ReadError,
2555                         &r10_bio->state);
2556                 r10_bio->mddev = mddev;
2557                 r10_bio->sector = mbio->bi_sector
2558                         + sectors_handled;
2559
2560                 goto read_more;
2561         } else
2562                 generic_make_request(bio);
2563 }
2564
2565 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2566 {
2567         /* Some sort of write request has finished and it
2568          * succeeded in writing where we thought there was a
2569          * bad block.  So forget the bad block.
2570          * Or possibly if failed and we need to record
2571          * a bad block.
2572          */
2573         int m;
2574         struct md_rdev *rdev;
2575
2576         if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2577             test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2578                 for (m = 0; m < conf->copies; m++) {
2579                         int dev = r10_bio->devs[m].devnum;
2580                         rdev = conf->mirrors[dev].rdev;
2581                         if (r10_bio->devs[m].bio == NULL)
2582                                 continue;
2583                         if (test_bit(BIO_UPTODATE,
2584                                      &r10_bio->devs[m].bio->bi_flags)) {
2585                                 rdev_clear_badblocks(
2586                                         rdev,
2587                                         r10_bio->devs[m].addr,
2588                                         r10_bio->sectors, 0);
2589                         } else {
2590                                 if (!rdev_set_badblocks(
2591                                             rdev,
2592                                             r10_bio->devs[m].addr,
2593                                             r10_bio->sectors, 0))
2594                                         md_error(conf->mddev, rdev);
2595                         }
2596                         rdev = conf->mirrors[dev].replacement;
2597                         if (r10_bio->devs[m].repl_bio == NULL)
2598                                 continue;
2599                         if (test_bit(BIO_UPTODATE,
2600                                      &r10_bio->devs[m].repl_bio->bi_flags)) {
2601                                 rdev_clear_badblocks(
2602                                         rdev,
2603                                         r10_bio->devs[m].addr,
2604                                         r10_bio->sectors, 0);
2605                         } else {
2606                                 if (!rdev_set_badblocks(
2607                                             rdev,
2608                                             r10_bio->devs[m].addr,
2609                                             r10_bio->sectors, 0))
2610                                         md_error(conf->mddev, rdev);
2611                         }
2612                 }
2613                 put_buf(r10_bio);
2614         } else {
2615                 for (m = 0; m < conf->copies; m++) {
2616                         int dev = r10_bio->devs[m].devnum;
2617                         struct bio *bio = r10_bio->devs[m].bio;
2618                         rdev = conf->mirrors[dev].rdev;
2619                         if (bio == IO_MADE_GOOD) {
2620                                 rdev_clear_badblocks(
2621                                         rdev,
2622                                         r10_bio->devs[m].addr,
2623                                         r10_bio->sectors, 0);
2624                                 rdev_dec_pending(rdev, conf->mddev);
2625                         } else if (bio != NULL &&
2626                                    !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2627                                 if (!narrow_write_error(r10_bio, m)) {
2628                                         md_error(conf->mddev, rdev);
2629                                         set_bit(R10BIO_Degraded,
2630                                                 &r10_bio->state);
2631                                 }
2632                                 rdev_dec_pending(rdev, conf->mddev);
2633                         }
2634                         bio = r10_bio->devs[m].repl_bio;
2635                         rdev = conf->mirrors[dev].replacement;
2636                         if (rdev && bio == IO_MADE_GOOD) {
2637                                 rdev_clear_badblocks(
2638                                         rdev,
2639                                         r10_bio->devs[m].addr,
2640                                         r10_bio->sectors, 0);
2641                                 rdev_dec_pending(rdev, conf->mddev);
2642                         }
2643                 }
2644                 if (test_bit(R10BIO_WriteError,
2645                              &r10_bio->state))
2646                         close_write(r10_bio);
2647                 raid_end_bio_io(r10_bio);
2648         }
2649 }
2650
2651 static void raid10d(struct mddev *mddev)
2652 {
2653         struct r10bio *r10_bio;
2654         unsigned long flags;
2655         struct r10conf *conf = mddev->private;
2656         struct list_head *head = &conf->retry_list;
2657         struct blk_plug plug;
2658
2659         md_check_recovery(mddev);
2660
2661         blk_start_plug(&plug);
2662         for (;;) {
2663
2664                 flush_pending_writes(conf);
2665
2666                 spin_lock_irqsave(&conf->device_lock, flags);
2667                 if (list_empty(head)) {
2668                         spin_unlock_irqrestore(&conf->device_lock, flags);
2669                         break;
2670                 }
2671                 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2672                 list_del(head->prev);
2673                 conf->nr_queued--;
2674                 spin_unlock_irqrestore(&conf->device_lock, flags);
2675
2676                 mddev = r10_bio->mddev;
2677                 conf = mddev->private;
2678                 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2679                     test_bit(R10BIO_WriteError, &r10_bio->state))
2680                         handle_write_completed(conf, r10_bio);
2681                 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2682                         reshape_request_write(mddev, r10_bio);
2683                 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2684                         sync_request_write(mddev, r10_bio);
2685                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2686                         recovery_request_write(mddev, r10_bio);
2687                 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2688                         handle_read_error(mddev, r10_bio);
2689                 else {
2690                         /* just a partial read to be scheduled from a
2691                          * separate context
2692                          */
2693                         int slot = r10_bio->read_slot;
2694                         generic_make_request(r10_bio->devs[slot].bio);
2695                 }
2696
2697                 cond_resched();
2698                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2699                         md_check_recovery(mddev);
2700         }
2701         blk_finish_plug(&plug);
2702 }
2703
2704
2705 static int init_resync(struct r10conf *conf)
2706 {
2707         int buffs;
2708         int i;
2709
2710         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2711         BUG_ON(conf->r10buf_pool);
2712         conf->have_replacement = 0;
2713         for (i = 0; i < conf->geo.raid_disks; i++)
2714                 if (conf->mirrors[i].replacement)
2715                         conf->have_replacement = 1;
2716         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2717         if (!conf->r10buf_pool)
2718                 return -ENOMEM;
2719         conf->next_resync = 0;
2720         return 0;
2721 }
2722
2723 /*
2724  * perform a "sync" on one "block"
2725  *
2726  * We need to make sure that no normal I/O request - particularly write
2727  * requests - conflict with active sync requests.
2728  *
2729  * This is achieved by tracking pending requests and a 'barrier' concept
2730  * that can be installed to exclude normal IO requests.
2731  *
2732  * Resync and recovery are handled very differently.
2733  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2734  *
2735  * For resync, we iterate over virtual addresses, read all copies,
2736  * and update if there are differences.  If only one copy is live,
2737  * skip it.
2738  * For recovery, we iterate over physical addresses, read a good
2739  * value for each non-in_sync drive, and over-write.
2740  *
2741  * So, for recovery we may have several outstanding complex requests for a
2742  * given address, one for each out-of-sync device.  We model this by allocating
2743  * a number of r10_bio structures, one for each out-of-sync device.
2744  * As we setup these structures, we collect all bio's together into a list
2745  * which we then process collectively to add pages, and then process again
2746  * to pass to generic_make_request.
2747  *
2748  * The r10_bio structures are linked using a borrowed master_bio pointer.
2749  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2750  * has its remaining count decremented to 0, the whole complex operation
2751  * is complete.
2752  *
2753  */
2754
2755 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2756                              int *skipped, int go_faster)
2757 {
2758         struct r10conf *conf = mddev->private;
2759         struct r10bio *r10_bio;
2760         struct bio *biolist = NULL, *bio;
2761         sector_t max_sector, nr_sectors;
2762         int i;
2763         int max_sync;
2764         sector_t sync_blocks;
2765         sector_t sectors_skipped = 0;
2766         int chunks_skipped = 0;
2767         sector_t chunk_mask = conf->geo.chunk_mask;
2768
2769         if (!conf->r10buf_pool)
2770                 if (init_resync(conf))
2771                         return 0;
2772
2773  skipped:
2774         max_sector = mddev->dev_sectors;
2775         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2776             test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2777                 max_sector = mddev->resync_max_sectors;
2778         if (sector_nr >= max_sector) {
2779                 /* If we aborted, we need to abort the
2780                  * sync on the 'current' bitmap chucks (there can
2781                  * be several when recovering multiple devices).
2782                  * as we may have started syncing it but not finished.
2783                  * We can find the current address in
2784                  * mddev->curr_resync, but for recovery,
2785                  * we need to convert that to several
2786                  * virtual addresses.
2787                  */
2788                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2789                         end_reshape(conf);
2790                         return 0;
2791                 }
2792
2793                 if (mddev->curr_resync < max_sector) { /* aborted */
2794                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2795                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2796                                                 &sync_blocks, 1);
2797                         else for (i = 0; i < conf->geo.raid_disks; i++) {
2798                                 sector_t sect =
2799                                         raid10_find_virt(conf, mddev->curr_resync, i);
2800                                 bitmap_end_sync(mddev->bitmap, sect,
2801                                                 &sync_blocks, 1);
2802                         }
2803                 } else {
2804                         /* completed sync */
2805                         if ((!mddev->bitmap || conf->fullsync)
2806                             && conf->have_replacement
2807                             && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2808                                 /* Completed a full sync so the replacements
2809                                  * are now fully recovered.
2810                                  */
2811                                 for (i = 0; i < conf->geo.raid_disks; i++)
2812                                         if (conf->mirrors[i].replacement)
2813                                                 conf->mirrors[i].replacement
2814                                                         ->recovery_offset
2815                                                         = MaxSector;
2816                         }
2817                         conf->fullsync = 0;
2818                 }
2819                 bitmap_close_sync(mddev->bitmap);
2820                 close_sync(conf);
2821                 *skipped = 1;
2822                 return sectors_skipped;
2823         }
2824
2825         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2826                 return reshape_request(mddev, sector_nr, skipped);
2827
2828         if (chunks_skipped >= conf->geo.raid_disks) {
2829                 /* if there has been nothing to do on any drive,
2830                  * then there is nothing to do at all..
2831                  */
2832                 *skipped = 1;
2833                 return (max_sector - sector_nr) + sectors_skipped;
2834         }
2835
2836         if (max_sector > mddev->resync_max)
2837                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2838
2839         /* make sure whole request will fit in a chunk - if chunks
2840          * are meaningful
2841          */
2842         if (conf->geo.near_copies < conf->geo.raid_disks &&
2843             max_sector > (sector_nr | chunk_mask))
2844                 max_sector = (sector_nr | chunk_mask) + 1;
2845         /*
2846          * If there is non-resync activity waiting for us then
2847          * put in a delay to throttle resync.
2848          */
2849         if (!go_faster && conf->nr_waiting)
2850                 msleep_interruptible(1000);
2851
2852         /* Again, very different code for resync and recovery.
2853          * Both must result in an r10bio with a list of bios that
2854          * have bi_end_io, bi_sector, bi_bdev set,
2855          * and bi_private set to the r10bio.
2856          * For recovery, we may actually create several r10bios
2857          * with 2 bios in each, that correspond to the bios in the main one.
2858          * In this case, the subordinate r10bios link back through a
2859          * borrowed master_bio pointer, and the counter in the master
2860          * includes a ref from each subordinate.
2861          */
2862         /* First, we decide what to do and set ->bi_end_io
2863          * To end_sync_read if we want to read, and
2864          * end_sync_write if we will want to write.
2865          */
2866
2867         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2868         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2869                 /* recovery... the complicated one */
2870                 int j;
2871                 r10_bio = NULL;
2872
2873                 for (i = 0 ; i < conf->geo.raid_disks; i++) {
2874                         int still_degraded;
2875                         struct r10bio *rb2;
2876                         sector_t sect;
2877                         int must_sync;
2878                         int any_working;
2879                         struct mirror_info *mirror = &conf->mirrors[i];
2880
2881                         if ((mirror->rdev == NULL ||
2882                              test_bit(In_sync, &mirror->rdev->flags))
2883                             &&
2884                             (mirror->replacement == NULL ||
2885                              test_bit(Faulty,
2886                                       &mirror->replacement->flags)))
2887                                 continue;
2888
2889                         still_degraded = 0;
2890                         /* want to reconstruct this device */
2891                         rb2 = r10_bio;
2892                         sect = raid10_find_virt(conf, sector_nr, i);
2893                         /* Unless we are doing a full sync, or a replacement
2894                          * we only need to recover the block if it is set in
2895                          * the bitmap
2896                          */
2897                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2898                                                       &sync_blocks, 1);
2899                         if (sync_blocks < max_sync)
2900                                 max_sync = sync_blocks;
2901                         if (!must_sync &&
2902                             mirror->replacement == NULL &&
2903                             !conf->fullsync) {
2904                                 /* yep, skip the sync_blocks here, but don't assume
2905                                  * that there will never be anything to do here
2906                                  */
2907                                 chunks_skipped = -1;
2908                                 continue;
2909                         }
2910
2911                         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2912                         raise_barrier(conf, rb2 != NULL);
2913                         atomic_set(&r10_bio->remaining, 0);
2914
2915                         r10_bio->master_bio = (struct bio*)rb2;
2916                         if (rb2)
2917                                 atomic_inc(&rb2->remaining);
2918                         r10_bio->mddev = mddev;
2919                         set_bit(R10BIO_IsRecover, &r10_bio->state);
2920                         r10_bio->sector = sect;
2921
2922                         raid10_find_phys(conf, r10_bio);
2923
2924                         /* Need to check if the array will still be
2925                          * degraded
2926                          */
2927                         for (j = 0; j < conf->geo.raid_disks; j++)
2928                                 if (conf->mirrors[j].rdev == NULL ||
2929                                     test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2930                                         still_degraded = 1;
2931                                         break;
2932                                 }
2933
2934                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2935                                                       &sync_blocks, still_degraded);
2936
2937                         any_working = 0;
2938                         for (j=0; j<conf->copies;j++) {
2939                                 int k;
2940                                 int d = r10_bio->devs[j].devnum;
2941                                 sector_t from_addr, to_addr;
2942                                 struct md_rdev *rdev;
2943                                 sector_t sector, first_bad;
2944                                 int bad_sectors;
2945                                 if (!conf->mirrors[d].rdev ||
2946                                     !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2947                                         continue;
2948                                 /* This is where we read from */
2949                                 any_working = 1;
2950                                 rdev = conf->mirrors[d].rdev;
2951                                 sector = r10_bio->devs[j].addr;
2952
2953                                 if (is_badblock(rdev, sector, max_sync,
2954                                                 &first_bad, &bad_sectors)) {
2955                                         if (first_bad > sector)
2956                                                 max_sync = first_bad - sector;
2957                                         else {
2958                                                 bad_sectors -= (sector
2959                                                                 - first_bad);
2960                                                 if (max_sync > bad_sectors)
2961                                                         max_sync = bad_sectors;
2962                                                 continue;
2963                                         }
2964                                 }
2965                                 bio = r10_bio->devs[0].bio;
2966                                 bio->bi_next = biolist;
2967                                 biolist = bio;
2968                                 bio->bi_private = r10_bio;
2969                                 bio->bi_end_io = end_sync_read;
2970                                 bio->bi_rw = READ;
2971                                 from_addr = r10_bio->devs[j].addr;
2972                                 bio->bi_sector = from_addr + rdev->data_offset;
2973                                 bio->bi_bdev = rdev->bdev;
2974                                 atomic_inc(&rdev->nr_pending);
2975                                 /* and we write to 'i' (if not in_sync) */
2976
2977                                 for (k=0; k<conf->copies; k++)
2978                                         if (r10_bio->devs[k].devnum == i)
2979                                                 break;
2980                                 BUG_ON(k == conf->copies);
2981                                 to_addr = r10_bio->devs[k].addr;
2982                                 r10_bio->devs[0].devnum = d;
2983                                 r10_bio->devs[0].addr = from_addr;
2984                                 r10_bio->devs[1].devnum = i;
2985                                 r10_bio->devs[1].addr = to_addr;
2986
2987                                 rdev = mirror->rdev;
2988                                 if (!test_bit(In_sync, &rdev->flags)) {
2989                                         bio = r10_bio->devs[1].bio;
2990                                         bio->bi_next = biolist;
2991                                         biolist = bio;
2992                                         bio->bi_private = r10_bio;
2993                                         bio->bi_end_io = end_sync_write;
2994                                         bio->bi_rw = WRITE;
2995                                         bio->bi_sector = to_addr
2996                                                 + rdev->data_offset;
2997                                         bio->bi_bdev = rdev->bdev;
2998                                         atomic_inc(&r10_bio->remaining);
2999                                 } else
3000                                         r10_bio->devs[1].bio->bi_end_io = NULL;
3001
3002                                 /* and maybe write to replacement */
3003                                 bio = r10_bio->devs[1].repl_bio;
3004                                 if (bio)
3005                                         bio->bi_end_io = NULL;
3006                                 rdev = mirror->replacement;
3007                                 /* Note: if rdev != NULL, then bio
3008                                  * cannot be NULL as r10buf_pool_alloc will
3009                                  * have allocated it.
3010                                  * So the second test here is pointless.
3011                                  * But it keeps semantic-checkers happy, and
3012                                  * this comment keeps human reviewers
3013                                  * happy.
3014                                  */
3015                                 if (rdev == NULL || bio == NULL ||
3016                                     test_bit(Faulty, &rdev->flags))
3017                                         break;
3018                                 bio->bi_next = biolist;
3019                                 biolist = bio;
3020                                 bio->bi_private = r10_bio;
3021                                 bio->bi_end_io = end_sync_write;
3022                                 bio->bi_rw = WRITE;
3023                                 bio->bi_sector = to_addr + rdev->data_offset;
3024                                 bio->bi_bdev = rdev->bdev;
3025                                 atomic_inc(&r10_bio->remaining);
3026                                 break;
3027                         }
3028                         if (j == conf->copies) {
3029                                 /* Cannot recover, so abort the recovery or
3030                                  * record a bad block */
3031                                 put_buf(r10_bio);
3032                                 if (rb2)
3033                                         atomic_dec(&rb2->remaining);
3034                                 r10_bio = rb2;
3035                                 if (any_working) {
3036                                         /* problem is that there are bad blocks
3037                                          * on other device(s)
3038                                          */
3039                                         int k;
3040                                         for (k = 0; k < conf->copies; k++)
3041                                                 if (r10_bio->devs[k].devnum == i)
3042                                                         break;
3043                                         if (!test_bit(In_sync,
3044                                                       &mirror->rdev->flags)
3045                                             && !rdev_set_badblocks(
3046                                                     mirror->rdev,
3047                                                     r10_bio->devs[k].addr,
3048                                                     max_sync, 0))
3049                                                 any_working = 0;
3050                                         if (mirror->replacement &&
3051                                             !rdev_set_badblocks(
3052                                                     mirror->replacement,
3053                                                     r10_bio->devs[k].addr,
3054                                                     max_sync, 0))
3055                                                 any_working = 0;
3056                                 }
3057                                 if (!any_working)  {
3058                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
3059                                                               &mddev->recovery))
3060                                                 printk(KERN_INFO "md/raid10:%s: insufficient "
3061                                                        "working devices for recovery.\n",
3062                                                        mdname(mddev));
3063                                         mirror->recovery_disabled
3064                                                 = mddev->recovery_disabled;
3065                                 }
3066                                 break;
3067                         }
3068                 }
3069                 if (biolist == NULL) {
3070                         while (r10_bio) {
3071                                 struct r10bio *rb2 = r10_bio;
3072                                 r10_bio = (struct r10bio*) rb2->master_bio;
3073                                 rb2->master_bio = NULL;
3074                                 put_buf(rb2);
3075                         }
3076                         goto giveup;
3077                 }
3078         } else {
3079                 /* resync. Schedule a read for every block at this virt offset */
3080                 int count = 0;
3081
3082                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3083
3084                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3085                                        &sync_blocks, mddev->degraded) &&
3086                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3087                                                  &mddev->recovery)) {
3088                         /* We can skip this block */
3089                         *skipped = 1;
3090                         return sync_blocks + sectors_skipped;
3091                 }
3092                 if (sync_blocks < max_sync)
3093                         max_sync = sync_blocks;
3094                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3095
3096                 r10_bio->mddev = mddev;
3097                 atomic_set(&r10_bio->remaining, 0);
3098                 raise_barrier(conf, 0);
3099                 conf->next_resync = sector_nr;
3100
3101                 r10_bio->master_bio = NULL;
3102                 r10_bio->sector = sector_nr;
3103                 set_bit(R10BIO_IsSync, &r10_bio->state);
3104                 raid10_find_phys(conf, r10_bio);
3105                 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3106
3107                 for (i = 0; i < conf->copies; i++) {
3108                         int d = r10_bio->devs[i].devnum;
3109                         sector_t first_bad, sector;
3110                         int bad_sectors;
3111
3112                         if (r10_bio->devs[i].repl_bio)
3113                                 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3114
3115                         bio = r10_bio->devs[i].bio;
3116                         bio->bi_end_io = NULL;
3117                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
3118                         if (conf->mirrors[d].rdev == NULL ||
3119                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3120                                 continue;
3121                         sector = r10_bio->devs[i].addr;
3122                         if (is_badblock(conf->mirrors[d].rdev,
3123                                         sector, max_sync,
3124                                         &first_bad, &bad_sectors)) {
3125                                 if (first_bad > sector)
3126                                         max_sync = first_bad - sector;
3127                                 else {
3128                                         bad_sectors -= (sector - first_bad);
3129                                         if (max_sync > bad_sectors)
3130                                                 max_sync = max_sync;
3131                                         continue;
3132                                 }
3133                         }
3134                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3135                         atomic_inc(&r10_bio->remaining);
3136                         bio->bi_next = biolist;
3137                         biolist = bio;
3138                         bio->bi_private = r10_bio;
3139                         bio->bi_end_io = end_sync_read;
3140                         bio->bi_rw = READ;
3141                         bio->bi_sector = sector +
3142                                 conf->mirrors[d].rdev->data_offset;
3143                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3144                         count++;
3145
3146                         if (conf->mirrors[d].replacement == NULL ||
3147                             test_bit(Faulty,
3148                                      &conf->mirrors[d].replacement->flags))
3149                                 continue;
3150
3151                         /* Need to set up for writing to the replacement */
3152                         bio = r10_bio->devs[i].repl_bio;
3153                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
3154
3155                         sector = r10_bio->devs[i].addr;
3156                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3157                         bio->bi_next = biolist;
3158                         biolist = bio;
3159                         bio->bi_private = r10_bio;
3160                         bio->bi_end_io = end_sync_write;
3161                         bio->bi_rw = WRITE;
3162                         bio->bi_sector = sector +
3163                                 conf->mirrors[d].replacement->data_offset;
3164                         bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3165                         count++;
3166                 }
3167
3168                 if (count < 2) {
3169                         for (i=0; i<conf->copies; i++) {
3170                                 int d = r10_bio->devs[i].devnum;
3171                                 if (r10_bio->devs[i].bio->bi_end_io)
3172                                         rdev_dec_pending(conf->mirrors[d].rdev,
3173                                                          mddev);
3174                                 if (r10_bio->devs[i].repl_bio &&
3175                                     r10_bio->devs[i].repl_bio->bi_end_io)
3176                                         rdev_dec_pending(
3177                                                 conf->mirrors[d].replacement,
3178                                                 mddev);
3179                         }
3180                         put_buf(r10_bio);
3181                         biolist = NULL;
3182                         goto giveup;
3183                 }
3184         }
3185
3186         for (bio = biolist; bio ; bio=bio->bi_next) {
3187
3188                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
3189                 if (bio->bi_end_io)
3190                         bio->bi_flags |= 1 << BIO_UPTODATE;
3191                 bio->bi_vcnt = 0;
3192                 bio->bi_idx = 0;
3193                 bio->bi_phys_segments = 0;
3194                 bio->bi_size = 0;
3195         }
3196
3197         nr_sectors = 0;
3198         if (sector_nr + max_sync < max_sector)
3199                 max_sector = sector_nr + max_sync;
3200         do {
3201                 struct page *page;
3202                 int len = PAGE_SIZE;
3203                 if (sector_nr + (len>>9) > max_sector)
3204                         len = (max_sector - sector_nr) << 9;
3205                 if (len == 0)
3206                         break;
3207                 for (bio= biolist ; bio ; bio=bio->bi_next) {
3208                         struct bio *bio2;
3209                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3210                         if (bio_add_page(bio, page, len, 0))
3211                                 continue;
3212
3213                         /* stop here */
3214                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3215                         for (bio2 = biolist;
3216                              bio2 && bio2 != bio;
3217                              bio2 = bio2->bi_next) {
3218                                 /* remove last page from this bio */
3219                                 bio2->bi_vcnt--;
3220                                 bio2->bi_size -= len;
3221                                 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3222                         }
3223                         goto bio_full;
3224                 }
3225                 nr_sectors += len>>9;
3226                 sector_nr += len>>9;
3227         } while (biolist->bi_vcnt < RESYNC_PAGES);
3228  bio_full:
3229         r10_bio->sectors = nr_sectors;
3230
3231         while (biolist) {
3232                 bio = biolist;
3233                 biolist = biolist->bi_next;
3234
3235                 bio->bi_next = NULL;
3236                 r10_bio = bio->bi_private;
3237                 r10_bio->sectors = nr_sectors;
3238
3239                 if (bio->bi_end_io == end_sync_read) {
3240                         md_sync_acct(bio->bi_bdev, nr_sectors);
3241                         generic_make_request(bio);
3242                 }
3243         }
3244
3245         if (sectors_skipped)
3246                 /* pretend they weren't skipped, it makes
3247                  * no important difference in this case
3248                  */
3249                 md_done_sync(mddev, sectors_skipped, 1);
3250
3251         return sectors_skipped + nr_sectors;
3252  giveup:
3253         /* There is nowhere to write, so all non-sync
3254          * drives must be failed or in resync, all drives
3255          * have a bad block, so try the next chunk...
3256          */
3257         if (sector_nr + max_sync < max_sector)
3258                 max_sector = sector_nr + max_sync;
3259
3260         sectors_skipped += (max_sector - sector_nr);
3261         chunks_skipped ++;
3262         sector_nr = max_sector;
3263         goto skipped;
3264 }
3265
3266 static sector_t
3267 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3268 {
3269         sector_t size;
3270         struct r10conf *conf = mddev->private;
3271
3272         if (!raid_disks)
3273                 raid_disks = min(conf->geo.raid_disks,
3274                                  conf->prev.raid_disks);
3275         if (!sectors)
3276                 sectors = conf->dev_sectors;
3277
3278         size = sectors >> conf->geo.chunk_shift;
3279         sector_div(size, conf->geo.far_copies);
3280         size = size * raid_disks;
3281         sector_div(size, conf->geo.near_copies);
3282
3283         return size << conf->geo.chunk_shift;
3284 }
3285
3286 static void calc_sectors(struct r10conf *conf, sector_t size)
3287 {
3288         /* Calculate the number of sectors-per-device that will
3289          * actually be used, and set conf->dev_sectors and
3290          * conf->stride
3291          */
3292
3293         size = size >> conf->geo.chunk_shift;
3294         sector_div(size, conf->geo.far_copies);
3295         size = size * conf->geo.raid_disks;
3296         sector_div(size, conf->geo.near_copies);
3297         /* 'size' is now the number of chunks in the array */
3298         /* calculate "used chunks per device" */
3299         size = size * conf->copies;
3300
3301         /* We need to round up when dividing by raid_disks to
3302          * get the stride size.
3303          */
3304         size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3305
3306         conf->dev_sectors = size << conf->geo.chunk_shift;
3307
3308         if (conf->geo.far_offset)
3309                 conf->geo.stride = 1 << conf->geo.chunk_shift;
3310         else {
3311                 sector_div(size, conf->geo.far_copies);
3312                 conf->geo.stride = size << conf->geo.chunk_shift;
3313         }
3314 }
3315
3316 enum geo_type {geo_new, geo_old, geo_start};
3317 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3318 {
3319         int nc, fc, fo;
3320         int layout, chunk, disks;
3321         switch (new) {
3322         case geo_old:
3323                 layout = mddev->layout;
3324                 chunk = mddev->chunk_sectors;
3325                 disks = mddev->raid_disks - mddev->delta_disks;
3326                 break;
3327         case geo_new:
3328                 layout = mddev->new_layout;
3329                 chunk = mddev->new_chunk_sectors;
3330                 disks = mddev->raid_disks;
3331                 break;
3332         default: /* avoid 'may be unused' warnings */
3333         case geo_start: /* new when starting reshape - raid_disks not
3334                          * updated yet. */
3335                 layout = mddev->new_layout;
3336                 chunk = mddev->new_chunk_sectors;
3337                 disks = mddev->raid_disks + mddev->delta_disks;
3338                 break;
3339         }
3340         if (layout >> 17)
3341                 return -1;
3342         if (chunk < (PAGE_SIZE >> 9) ||
3343             !is_power_of_2(chunk))
3344                 return -2;
3345         nc = layout & 255;
3346         fc = (layout >> 8) & 255;
3347         fo = layout & (1<<16);
3348         geo->raid_disks = disks;
3349         geo->near_copies = nc;
3350         geo->far_copies = fc;
3351         geo->far_offset = fo;
3352         geo->chunk_mask = chunk - 1;
3353         geo->chunk_shift = ffz(~chunk);
3354         return nc*fc;
3355 }
3356
3357 static struct r10conf *setup_conf(struct mddev *mddev)
3358 {
3359         struct r10conf *conf = NULL;
3360         int err = -EINVAL;
3361         struct geom geo;
3362         int copies;
3363
3364         copies = setup_geo(&geo, mddev, geo_new);
3365
3366         if (copies == -2) {
3367                 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3368                        "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3369                        mdname(mddev), PAGE_SIZE);
3370                 goto out;
3371         }
3372
3373         if (copies < 2 || copies > mddev->raid_disks) {
3374                 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3375                        mdname(mddev), mddev->new_layout);
3376                 goto out;
3377         }
3378
3379         err = -ENOMEM;
3380         conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3381         if (!conf)
3382                 goto out;
3383
3384         /* FIXME calc properly */
3385         conf->mirrors = kzalloc(sizeof(struct mirror_info)*(mddev->raid_disks +
3386                                                             max(0,mddev->delta_disks)),
3387                                 GFP_KERNEL);
3388         if (!conf->mirrors)
3389                 goto out;
3390
3391         conf->tmppage = alloc_page(GFP_KERNEL);
3392         if (!conf->tmppage)
3393                 goto out;
3394
3395         conf->geo = geo;
3396         conf->copies = copies;
3397         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3398                                            r10bio_pool_free, conf);
3399         if (!conf->r10bio_pool)
3400                 goto out;
3401
3402         calc_sectors(conf, mddev->dev_sectors);
3403         if (mddev->reshape_position == MaxSector) {
3404                 conf->prev = conf->geo;
3405                 conf->reshape_progress = MaxSector;
3406         } else {
3407                 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3408                         err = -EINVAL;
3409                         goto out;
3410                 }
3411                 conf->reshape_progress = mddev->reshape_position;
3412                 if (conf->prev.far_offset)
3413                         conf->prev.stride = 1 << conf->prev.chunk_shift;
3414                 else
3415                         /* far_copies must be 1 */
3416                         conf->prev.stride = conf->dev_sectors;
3417         }
3418         spin_lock_init(&conf->device_lock);
3419         INIT_LIST_HEAD(&conf->retry_list);
3420
3421         spin_lock_init(&conf->resync_lock);
3422         init_waitqueue_head(&conf->wait_barrier);
3423
3424         conf->thread = md_register_thread(raid10d, mddev, NULL);
3425         if (!conf->thread)
3426                 goto out;
3427
3428         conf->mddev = mddev;
3429         return conf;
3430
3431  out:
3432         if (err == -ENOMEM)
3433                 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3434                        mdname(mddev));
3435         if (conf) {
3436                 if (conf->r10bio_pool)
3437                         mempool_destroy(conf->r10bio_pool);
3438                 kfree(conf->mirrors);
3439                 safe_put_page(conf->tmppage);
3440                 kfree(conf);
3441         }
3442         return ERR_PTR(err);
3443 }
3444
3445 static int run(struct mddev *mddev)
3446 {
3447         struct r10conf *conf;
3448         int i, disk_idx, chunk_size;
3449         struct mirror_info *disk;
3450         struct md_rdev *rdev;
3451         sector_t size;
3452         sector_t min_offset_diff = 0;
3453         int first = 1;
3454
3455         if (mddev->private == NULL) {
3456                 conf = setup_conf(mddev);
3457                 if (IS_ERR(conf))
3458                         return PTR_ERR(conf);
3459                 mddev->private = conf;
3460         }
3461         conf = mddev->private;
3462         if (!conf)
3463                 goto out;
3464
3465         mddev->thread = conf->thread;
3466         conf->thread = NULL;
3467
3468         chunk_size = mddev->chunk_sectors << 9;
3469         blk_queue_io_min(mddev->queue, chunk_size);
3470         if (conf->geo.raid_disks % conf->geo.near_copies)
3471                 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3472         else
3473                 blk_queue_io_opt(mddev->queue, chunk_size *
3474                                  (conf->geo.raid_disks / conf->geo.near_copies));
3475
3476         rdev_for_each(rdev, mddev) {
3477                 long long diff;
3478
3479                 disk_idx = rdev->raid_disk;
3480                 if (disk_idx < 0)
3481                         continue;
3482                 if (disk_idx >= conf->geo.raid_disks &&
3483                     disk_idx >= conf->prev.raid_disks)
3484                         continue;
3485                 disk = conf->mirrors + disk_idx;
3486
3487                 if (test_bit(Replacement, &rdev->flags)) {
3488                         if (disk->replacement)
3489                                 goto out_free_conf;
3490                         disk->replacement = rdev;
3491                 } else {
3492                         if (disk->rdev)
3493                                 goto out_free_conf;
3494                         disk->rdev = rdev;
3495                 }
3496                 diff = (rdev->new_data_offset - rdev->data_offset);
3497                 if (!mddev->reshape_backwards)
3498                         diff = -diff;
3499                 if (diff < 0)
3500                         diff = 0;
3501                 if (first || diff < min_offset_diff)
3502                         min_offset_diff = diff;
3503
3504                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3505                                   rdev->data_offset << 9);
3506
3507                 disk->head_position = 0;
3508         }
3509
3510         /* need to check that every block has at least one working mirror */
3511         if (!enough(conf, -1)) {
3512                 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3513                        mdname(mddev));
3514                 goto out_free_conf;
3515         }
3516
3517         if (conf->reshape_progress != MaxSector) {
3518                 /* must ensure that shape change is supported */
3519                 if (conf->geo.far_copies != 1 &&
3520                     conf->geo.far_offset == 0)
3521                         goto out_free_conf;
3522                 if (conf->prev.far_copies != 1 &&
3523                     conf->geo.far_offset == 0)
3524                         goto out_free_conf;
3525         }
3526
3527         mddev->degraded = 0;
3528         for (i = 0;
3529              i < conf->geo.raid_disks
3530                      || i < conf->prev.raid_disks;
3531              i++) {
3532
3533                 disk = conf->mirrors + i;
3534
3535                 if (!disk->rdev && disk->replacement) {
3536                         /* The replacement is all we have - use it */
3537                         disk->rdev = disk->replacement;
3538                         disk->replacement = NULL;
3539                         clear_bit(Replacement, &disk->rdev->flags);
3540                 }
3541
3542                 if (!disk->rdev ||
3543                     !test_bit(In_sync, &disk->rdev->flags)) {
3544                         disk->head_position = 0;
3545                         mddev->degraded++;
3546                         if (disk->rdev)
3547                                 conf->fullsync = 1;
3548                 }
3549                 disk->recovery_disabled = mddev->recovery_disabled - 1;
3550         }
3551
3552         if (mddev->recovery_cp != MaxSector)
3553                 printk(KERN_NOTICE "md/raid10:%s: not clean"
3554                        " -- starting background reconstruction\n",
3555                        mdname(mddev));
3556         printk(KERN_INFO
3557                 "md/raid10:%s: active with %d out of %d devices\n",
3558                 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3559                 conf->geo.raid_disks);
3560         /*
3561          * Ok, everything is just fine now
3562          */
3563         mddev->dev_sectors = conf->dev_sectors;
3564         size = raid10_size(mddev, 0, 0);
3565         md_set_array_sectors(mddev, size);
3566         mddev->resync_max_sectors = size;
3567
3568         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3569         mddev->queue->backing_dev_info.congested_data = mddev;
3570
3571         /* Calculate max read-ahead size.
3572          * We need to readahead at least twice a whole stripe....
3573          * maybe...
3574          */
3575         {
3576                 int stripe = conf->geo.raid_disks *
3577                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3578                 stripe /= conf->geo.near_copies;
3579                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3580                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3581         }
3582
3583         blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3584
3585         if (md_integrity_register(mddev))
3586                 goto out_free_conf;
3587
3588         if (conf->reshape_progress != MaxSector) {
3589                 unsigned long before_length, after_length;
3590
3591                 before_length = ((1 << conf->prev.chunk_shift) *
3592                                  conf->prev.far_copies);
3593                 after_length = ((1 << conf->geo.chunk_shift) *
3594                                 conf->geo.far_copies);
3595
3596                 if (max(before_length, after_length) > min_offset_diff) {
3597                         /* This cannot work */
3598                         printk("md/raid10: offset difference not enough to continue reshape\n");
3599                         goto out_free_conf;
3600                 }
3601                 conf->offset_diff = min_offset_diff;
3602
3603                 conf->reshape_safe = conf->reshape_progress;
3604                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3605                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3606                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3607                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3608                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3609                                                         "reshape");
3610         }
3611
3612         return 0;
3613
3614 out_free_conf:
3615         md_unregister_thread(&mddev->thread);
3616         if (conf->r10bio_pool)
3617                 mempool_destroy(conf->r10bio_pool);
3618         safe_put_page(conf->tmppage);
3619         kfree(conf->mirrors);
3620         kfree(conf);
3621         mddev->private = NULL;
3622 out:
3623         return -EIO;
3624 }
3625
3626 static int stop(struct mddev *mddev)
3627 {
3628         struct r10conf *conf = mddev->private;
3629
3630         raise_barrier(conf, 0);
3631         lower_barrier(conf);
3632
3633         md_unregister_thread(&mddev->thread);
3634         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
3635         if (conf->r10bio_pool)
3636                 mempool_destroy(conf->r10bio_pool);
3637         kfree(conf->mirrors);
3638         kfree(conf);
3639         mddev->private = NULL;
3640         return 0;
3641 }
3642
3643 static void raid10_quiesce(struct mddev *mddev, int state)
3644 {
3645         struct r10conf *conf = mddev->private;
3646
3647         switch(state) {
3648         case 1:
3649                 raise_barrier(conf, 0);
3650                 break;
3651         case 0:
3652                 lower_barrier(conf);
3653                 break;
3654         }
3655 }
3656
3657 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3658 {
3659         /* Resize of 'far' arrays is not supported.
3660          * For 'near' and 'offset' arrays we can set the
3661          * number of sectors used to be an appropriate multiple
3662          * of the chunk size.
3663          * For 'offset', this is far_copies*chunksize.
3664          * For 'near' the multiplier is the LCM of
3665          * near_copies and raid_disks.
3666          * So if far_copies > 1 && !far_offset, fail.
3667          * Else find LCM(raid_disks, near_copy)*far_copies and
3668          * multiply by chunk_size.  Then round to this number.
3669          * This is mostly done by raid10_size()
3670          */
3671         struct r10conf *conf = mddev->private;
3672         sector_t oldsize, size;
3673
3674         if (mddev->reshape_position != MaxSector)
3675                 return -EBUSY;
3676
3677         if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3678                 return -EINVAL;
3679
3680         oldsize = raid10_size(mddev, 0, 0);
3681         size = raid10_size(mddev, sectors, 0);
3682         if (mddev->external_size &&
3683             mddev->array_sectors > size)
3684                 return -EINVAL;
3685         if (mddev->bitmap) {
3686                 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3687                 if (ret)
3688                         return ret;
3689         }
3690         md_set_array_sectors(mddev, size);
3691         set_capacity(mddev->gendisk, mddev->array_sectors);
3692         revalidate_disk(mddev->gendisk);
3693         if (sectors > mddev->dev_sectors &&
3694             mddev->recovery_cp > oldsize) {
3695                 mddev->recovery_cp = oldsize;
3696                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3697         }
3698         calc_sectors(conf, sectors);
3699         mddev->dev_sectors = conf->dev_sectors;
3700         mddev->resync_max_sectors = size;
3701         return 0;
3702 }
3703
3704 static void *raid10_takeover_raid0(struct mddev *mddev)
3705 {
3706         struct md_rdev *rdev;
3707         struct r10conf *conf;
3708
3709         if (mddev->degraded > 0) {
3710                 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3711                        mdname(mddev));
3712                 return ERR_PTR(-EINVAL);
3713         }
3714
3715         /* Set new parameters */
3716         mddev->new_level = 10;
3717         /* new layout: far_copies = 1, near_copies = 2 */
3718         mddev->new_layout = (1<<8) + 2;
3719         mddev->new_chunk_sectors = mddev->chunk_sectors;
3720         mddev->delta_disks = mddev->raid_disks;
3721         mddev->raid_disks *= 2;
3722         /* make sure it will be not marked as dirty */
3723         mddev->recovery_cp = MaxSector;
3724
3725         conf = setup_conf(mddev);
3726         if (!IS_ERR(conf)) {
3727                 rdev_for_each(rdev, mddev)
3728                         if (rdev->raid_disk >= 0)
3729                                 rdev->new_raid_disk = rdev->raid_disk * 2;
3730                 conf->barrier = 1;
3731         }
3732
3733         return conf;
3734 }
3735
3736 static void *raid10_takeover(struct mddev *mddev)
3737 {
3738         struct r0conf *raid0_conf;
3739
3740         /* raid10 can take over:
3741          *  raid0 - providing it has only two drives
3742          */
3743         if (mddev->level == 0) {
3744                 /* for raid0 takeover only one zone is supported */
3745                 raid0_conf = mddev->private;
3746                 if (raid0_conf->nr_strip_zones > 1) {
3747                         printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3748                                " with more than one zone.\n",
3749                                mdname(mddev));
3750                         return ERR_PTR(-EINVAL);
3751                 }
3752                 return raid10_takeover_raid0(mddev);
3753         }
3754         return ERR_PTR(-EINVAL);
3755 }
3756
3757 static int raid10_check_reshape(struct mddev *mddev)
3758 {
3759         /* Called when there is a request to change
3760          * - layout (to ->new_layout)
3761          * - chunk size (to ->new_chunk_sectors)
3762          * - raid_disks (by delta_disks)
3763          * or when trying to restart a reshape that was ongoing.
3764          *
3765          * We need to validate the request and possibly allocate
3766          * space if that might be an issue later.
3767          *
3768          * Currently we reject any reshape of a 'far' mode array,
3769          * allow chunk size to change if new is generally acceptable,
3770          * allow raid_disks to increase, and allow
3771          * a switch between 'near' mode and 'offset' mode.
3772          */
3773         struct r10conf *conf = mddev->private;
3774         struct geom geo;
3775
3776         if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3777                 return -EINVAL;
3778
3779         if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3780                 /* mustn't change number of copies */
3781                 return -EINVAL;
3782         if (geo.far_copies > 1 && !geo.far_offset)
3783                 /* Cannot switch to 'far' mode */
3784                 return -EINVAL;
3785
3786         if (mddev->array_sectors & geo.chunk_mask)
3787                         /* not factor of array size */
3788                         return -EINVAL;
3789
3790         if (!enough(conf, -1))
3791                 return -EINVAL;
3792
3793         kfree(conf->mirrors_new);
3794         conf->mirrors_new = NULL;
3795         if (mddev->delta_disks > 0) {
3796                 /* allocate new 'mirrors' list */
3797                 conf->mirrors_new = kzalloc(
3798                         sizeof(struct mirror_info)
3799                         *(mddev->raid_disks +
3800                           mddev->delta_disks),
3801                         GFP_KERNEL);
3802                 if (!conf->mirrors_new)
3803                         return -ENOMEM;
3804         }
3805         return 0;
3806 }
3807
3808 /*
3809  * Need to check if array has failed when deciding whether to:
3810  *  - start an array
3811  *  - remove non-faulty devices
3812  *  - add a spare
3813  *  - allow a reshape
3814  * This determination is simple when no reshape is happening.
3815  * However if there is a reshape, we need to carefully check
3816  * both the before and after sections.
3817  * This is because some failed devices may only affect one
3818  * of the two sections, and some non-in_sync devices may
3819  * be insync in the section most affected by failed devices.
3820  */
3821 static int calc_degraded(struct r10conf *conf)
3822 {
3823         int degraded, degraded2;
3824         int i;
3825
3826         rcu_read_lock();
3827         degraded = 0;
3828         /* 'prev' section first */
3829         for (i = 0; i < conf->prev.raid_disks; i++) {
3830                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3831                 if (!rdev || test_bit(Faulty, &rdev->flags))
3832                         degraded++;
3833                 else if (!test_bit(In_sync, &rdev->flags))
3834                         /* When we can reduce the number of devices in
3835                          * an array, this might not contribute to
3836                          * 'degraded'.  It does now.
3837                          */
3838                         degraded++;
3839         }
3840         rcu_read_unlock();
3841         if (conf->geo.raid_disks == conf->prev.raid_disks)
3842                 return degraded;
3843         rcu_read_lock();
3844         degraded2 = 0;
3845         for (i = 0; i < conf->geo.raid_disks; i++) {
3846                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3847                 if (!rdev || test_bit(Faulty, &rdev->flags))
3848                         degraded2++;
3849                 else if (!test_bit(In_sync, &rdev->flags)) {
3850                         /* If reshape is increasing the number of devices,
3851                          * this section has already been recovered, so
3852                          * it doesn't contribute to degraded.
3853                          * else it does.
3854                          */
3855                         if (conf->geo.raid_disks <= conf->prev.raid_disks)
3856                                 degraded2++;
3857                 }
3858         }
3859         rcu_read_unlock();
3860         if (degraded2 > degraded)
3861                 return degraded2;
3862         return degraded;
3863 }
3864
3865 static int raid10_start_reshape(struct mddev *mddev)
3866 {
3867         /* A 'reshape' has been requested. This commits
3868          * the various 'new' fields and sets MD_RECOVER_RESHAPE
3869          * This also checks if there are enough spares and adds them
3870          * to the array.
3871          * We currently require enough spares to make the final
3872          * array non-degraded.  We also require that the difference
3873          * between old and new data_offset - on each device - is
3874          * enough that we never risk over-writing.
3875          */
3876
3877         unsigned long before_length, after_length;
3878         sector_t min_offset_diff = 0;
3879         int first = 1;
3880         struct geom new;
3881         struct r10conf *conf = mddev->private;
3882         struct md_rdev *rdev;
3883         int spares = 0;
3884         int ret;
3885
3886         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
3887                 return -EBUSY;
3888
3889         if (setup_geo(&new, mddev, geo_start) != conf->copies)
3890                 return -EINVAL;
3891
3892         before_length = ((1 << conf->prev.chunk_shift) *
3893                          conf->prev.far_copies);
3894         after_length = ((1 << conf->geo.chunk_shift) *
3895                         conf->geo.far_copies);
3896
3897         rdev_for_each(rdev, mddev) {
3898                 if (!test_bit(In_sync, &rdev->flags)
3899                     && !test_bit(Faulty, &rdev->flags))
3900                         spares++;
3901                 if (rdev->raid_disk >= 0) {
3902                         long long diff = (rdev->new_data_offset
3903                                           - rdev->data_offset);
3904                         if (!mddev->reshape_backwards)
3905                                 diff = -diff;
3906                         if (diff < 0)
3907                                 diff = 0;
3908                         if (first || diff < min_offset_diff)
3909                                 min_offset_diff = diff;
3910                 }
3911         }
3912
3913         if (max(before_length, after_length) > min_offset_diff)
3914                 return -EINVAL;
3915
3916         if (spares < mddev->delta_disks)
3917                 return -EINVAL;
3918
3919         conf->offset_diff = min_offset_diff;
3920         spin_lock_irq(&conf->device_lock);
3921         if (conf->mirrors_new) {
3922                 memcpy(conf->mirrors_new, conf->mirrors,
3923                        sizeof(struct mirror_info)*conf->prev.raid_disks);
3924                 smp_mb();
3925                 kfree(conf->mirrors_old); /* FIXME and elsewhere */
3926                 conf->mirrors_old = conf->mirrors;
3927                 conf->mirrors = conf->mirrors_new;
3928                 conf->mirrors_new = NULL;
3929         }
3930         setup_geo(&conf->geo, mddev, geo_start);
3931         smp_mb();
3932         if (mddev->reshape_backwards) {
3933                 sector_t size = raid10_size(mddev, 0, 0);
3934                 if (size < mddev->array_sectors) {
3935                         spin_unlock_irq(&conf->device_lock);
3936                         printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
3937                                mdname(mddev));
3938                         return -EINVAL;
3939                 }
3940                 mddev->resync_max_sectors = size;
3941                 conf->reshape_progress = size;
3942         } else
3943                 conf->reshape_progress = 0;
3944         spin_unlock_irq(&conf->device_lock);
3945
3946         if (mddev->delta_disks && mddev->bitmap) {
3947                 ret = bitmap_resize(mddev->bitmap,
3948                                     raid10_size(mddev, 0,
3949                                                 conf->geo.raid_disks),
3950                                     0, 0);
3951                 if (ret)
3952                         goto abort;
3953         }
3954         if (mddev->delta_disks > 0) {
3955                 rdev_for_each(rdev, mddev)
3956                         if (rdev->raid_disk < 0 &&
3957                             !test_bit(Faulty, &rdev->flags)) {
3958                                 if (raid10_add_disk(mddev, rdev) == 0) {
3959                                         if (rdev->raid_disk >=
3960                                             conf->prev.raid_disks)
3961                                                 set_bit(In_sync, &rdev->flags);
3962                                         else
3963                                                 rdev->recovery_offset = 0;
3964
3965                                         if (sysfs_link_rdev(mddev, rdev))
3966                                                 /* Failure here  is OK */;
3967                                 }
3968                         } else if (rdev->raid_disk >= conf->prev.raid_disks
3969                                    && !test_bit(Faulty, &rdev->flags)) {
3970                                 /* This is a spare that was manually added */
3971                                 set_bit(In_sync, &rdev->flags);
3972                         }
3973         }
3974         /* When a reshape changes the number of devices,
3975          * ->degraded is measured against the larger of the
3976          * pre and  post numbers.
3977          */
3978         spin_lock_irq(&conf->device_lock);
3979         mddev->degraded = calc_degraded(conf);
3980         spin_unlock_irq(&conf->device_lock);
3981         mddev->raid_disks = conf->geo.raid_disks;
3982         mddev->reshape_position = conf->reshape_progress;
3983         set_bit(MD_CHANGE_DEVS, &mddev->flags);
3984
3985         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3986         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3987         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3988         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3989
3990         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3991                                                 "reshape");
3992         if (!mddev->sync_thread) {
3993                 ret = -EAGAIN;
3994                 goto abort;
3995         }
3996         conf->reshape_checkpoint = jiffies;
3997         md_wakeup_thread(mddev->sync_thread);
3998         md_new_event(mddev);
3999         return 0;
4000
4001 abort:
4002         mddev->recovery = 0;
4003         spin_lock_irq(&conf->device_lock);
4004         conf->geo = conf->prev;
4005         mddev->raid_disks = conf->geo.raid_disks;
4006         rdev_for_each(rdev, mddev)
4007                 rdev->new_data_offset = rdev->data_offset;
4008         smp_wmb();
4009         conf->reshape_progress = MaxSector;
4010         mddev->reshape_position = MaxSector;
4011         spin_unlock_irq(&conf->device_lock);
4012         return ret;
4013 }
4014
4015 /* Calculate the last device-address that could contain
4016  * any block from the chunk that includes the array-address 's'
4017  * and report the next address.
4018  * i.e. the address returned will be chunk-aligned and after
4019  * any data that is in the chunk containing 's'.
4020  */
4021 static sector_t last_dev_address(sector_t s, struct geom *geo)
4022 {
4023         s = (s | geo->chunk_mask) + 1;
4024         s >>= geo->chunk_shift;
4025         s *= geo->near_copies;
4026         s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4027         s *= geo->far_copies;
4028         s <<= geo->chunk_shift;
4029         return s;
4030 }
4031
4032 /* Calculate the first device-address that could contain
4033  * any block from the chunk that includes the array-address 's'.
4034  * This too will be the start of a chunk
4035  */
4036 static sector_t first_dev_address(sector_t s, struct geom *geo)
4037 {
4038         s >>= geo->chunk_shift;
4039         s *= geo->near_copies;
4040         sector_div(s, geo->raid_disks);
4041         s *= geo->far_copies;
4042         s <<= geo->chunk_shift;
4043         return s;
4044 }
4045
4046 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4047                                 int *skipped)
4048 {
4049         /* We simply copy at most one chunk (smallest of old and new)
4050          * at a time, possibly less if that exceeds RESYNC_PAGES,
4051          * or we hit a bad block or something.
4052          * This might mean we pause for normal IO in the middle of
4053          * a chunk, but that is not a problem was mddev->reshape_position
4054          * can record any location.
4055          *
4056          * If we will want to write to a location that isn't
4057          * yet recorded as 'safe' (i.e. in metadata on disk) then
4058          * we need to flush all reshape requests and update the metadata.
4059          *
4060          * When reshaping forwards (e.g. to more devices), we interpret
4061          * 'safe' as the earliest block which might not have been copied
4062          * down yet.  We divide this by previous stripe size and multiply
4063          * by previous stripe length to get lowest device offset that we
4064          * cannot write to yet.
4065          * We interpret 'sector_nr' as an address that we want to write to.
4066          * From this we use last_device_address() to find where we might
4067          * write to, and first_device_address on the  'safe' position.
4068          * If this 'next' write position is after the 'safe' position,
4069          * we must update the metadata to increase the 'safe' position.
4070          *
4071          * When reshaping backwards, we round in the opposite direction
4072          * and perform the reverse test:  next write position must not be
4073          * less than current safe position.
4074          *
4075          * In all this the minimum difference in data offsets
4076          * (conf->offset_diff - always positive) allows a bit of slack,
4077          * so next can be after 'safe', but not by more than offset_disk
4078          *
4079          * We need to prepare all the bios here before we start any IO
4080          * to ensure the size we choose is acceptable to all devices.
4081          * The means one for each copy for write-out and an extra one for
4082          * read-in.
4083          * We store the read-in bio in ->master_bio and the others in
4084          * ->devs[x].bio and ->devs[x].repl_bio.
4085          */
4086         struct r10conf *conf = mddev->private;
4087         struct r10bio *r10_bio;
4088         sector_t next, safe, last;
4089         int max_sectors;
4090         int nr_sectors;
4091         int s;
4092         struct md_rdev *rdev;
4093         int need_flush = 0;
4094         struct bio *blist;
4095         struct bio *bio, *read_bio;
4096         int sectors_done = 0;
4097
4098         if (sector_nr == 0) {
4099                 /* If restarting in the middle, skip the initial sectors */
4100                 if (mddev->reshape_backwards &&
4101                     conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4102                         sector_nr = (raid10_size(mddev, 0, 0)
4103                                      - conf->reshape_progress);
4104                 } else if (!mddev->reshape_backwards &&
4105                            conf->reshape_progress > 0)
4106                         sector_nr = conf->reshape_progress;
4107                 if (sector_nr) {
4108                         mddev->curr_resync_completed = sector_nr;
4109                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4110                         *skipped = 1;
4111                         return sector_nr;
4112                 }
4113         }
4114
4115         /* We don't use sector_nr to track where we are up to
4116          * as that doesn't work well for ->reshape_backwards.
4117          * So just use ->reshape_progress.
4118          */
4119         if (mddev->reshape_backwards) {
4120                 /* 'next' is the earliest device address that we might
4121                  * write to for this chunk in the new layout
4122                  */
4123                 next = first_dev_address(conf->reshape_progress - 1,
4124                                          &conf->geo);
4125
4126                 /* 'safe' is the last device address that we might read from
4127                  * in the old layout after a restart
4128                  */
4129                 safe = last_dev_address(conf->reshape_safe - 1,
4130                                         &conf->prev);
4131
4132                 if (next + conf->offset_diff < safe)
4133                         need_flush = 1;
4134
4135                 last = conf->reshape_progress - 1;
4136                 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4137                                                & conf->prev.chunk_mask);
4138                 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4139                         sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4140         } else {
4141                 /* 'next' is after the last device address that we
4142                  * might write to for this chunk in the new layout
4143                  */
4144                 next = last_dev_address(conf->reshape_progress, &conf->geo);
4145
4146                 /* 'safe' is the earliest device address that we might
4147                  * read from in the old layout after a restart
4148                  */
4149                 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4150
4151                 /* Need to update metadata if 'next' might be beyond 'safe'
4152                  * as that would possibly corrupt data
4153                  */
4154                 if (next > safe + conf->offset_diff)
4155                         need_flush = 1;
4156
4157                 sector_nr = conf->reshape_progress;
4158                 last  = sector_nr | (conf->geo.chunk_mask
4159                                      & conf->prev.chunk_mask);
4160
4161                 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4162                         last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4163         }
4164
4165         if (need_flush ||
4166             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4167                 /* Need to update reshape_position in metadata */
4168                 wait_barrier(conf);
4169                 mddev->reshape_position = conf->reshape_progress;
4170                 if (mddev->reshape_backwards)
4171                         mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4172                                 - conf->reshape_progress;
4173                 else
4174                         mddev->curr_resync_completed = conf->reshape_progress;
4175                 conf->reshape_checkpoint = jiffies;
4176                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4177                 md_wakeup_thread(mddev->thread);
4178                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4179                            kthread_should_stop());
4180                 conf->reshape_safe = mddev->reshape_position;
4181                 allow_barrier(conf);
4182         }
4183
4184 read_more:
4185         /* Now schedule reads for blocks from sector_nr to last */
4186         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4187         raise_barrier(conf, sectors_done != 0);
4188         atomic_set(&r10_bio->remaining, 0);
4189         r10_bio->mddev = mddev;
4190         r10_bio->sector = sector_nr;
4191         set_bit(R10BIO_IsReshape, &r10_bio->state);
4192         r10_bio->sectors = last - sector_nr + 1;
4193         rdev = read_balance(conf, r10_bio, &max_sectors);
4194         BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4195
4196         if (!rdev) {
4197                 /* Cannot read from here, so need to record bad blocks
4198                  * on all the target devices.
4199                  */
4200                 // FIXME
4201                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4202                 return sectors_done;
4203         }
4204
4205         read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4206
4207         read_bio->bi_bdev = rdev->bdev;
4208         read_bio->bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4209                                + rdev->data_offset);
4210         read_bio->bi_private = r10_bio;
4211         read_bio->bi_end_io = end_sync_read;
4212         read_bio->bi_rw = READ;
4213         read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
4214         read_bio->bi_flags |= 1 << BIO_UPTODATE;
4215         read_bio->bi_vcnt = 0;
4216         read_bio->bi_idx = 0;
4217         read_bio->bi_size = 0;
4218         r10_bio->master_bio = read_bio;
4219         r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4220
4221         /* Now find the locations in the new layout */
4222         __raid10_find_phys(&conf->geo, r10_bio);
4223
4224         blist = read_bio;
4225         read_bio->bi_next = NULL;
4226
4227         for (s = 0; s < conf->copies*2; s++) {
4228                 struct bio *b;
4229                 int d = r10_bio->devs[s/2].devnum;
4230                 struct md_rdev *rdev2;
4231                 if (s&1) {
4232                         rdev2 = conf->mirrors[d].replacement;
4233                         b = r10_bio->devs[s/2].repl_bio;
4234                 } else {
4235                         rdev2 = conf->mirrors[d].rdev;
4236                         b = r10_bio->devs[s/2].bio;
4237                 }
4238                 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4239                         continue;
4240                 b->bi_bdev = rdev2->bdev;
4241                 b->bi_sector = r10_bio->devs[s/2].addr + rdev2->new_data_offset;
4242                 b->bi_private = r10_bio;
4243                 b->bi_end_io = end_reshape_write;
4244                 b->bi_rw = WRITE;
4245                 b->bi_flags &= ~(BIO_POOL_MASK - 1);
4246                 b->bi_flags |= 1 << BIO_UPTODATE;
4247                 b->bi_next = blist;
4248                 b->bi_vcnt = 0;
4249                 b->bi_idx = 0;
4250                 b->bi_size = 0;
4251                 blist = b;
4252         }
4253
4254         /* Now add as many pages as possible to all of these bios. */
4255
4256         nr_sectors = 0;
4257         for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4258                 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4259                 int len = (max_sectors - s) << 9;
4260                 if (len > PAGE_SIZE)
4261                         len = PAGE_SIZE;
4262                 for (bio = blist; bio ; bio = bio->bi_next) {
4263                         struct bio *bio2;
4264                         if (bio_add_page(bio, page, len, 0))
4265                                 continue;
4266
4267                         /* Didn't fit, must stop */
4268                         for (bio2 = blist;
4269                              bio2 && bio2 != bio;
4270                              bio2 = bio2->bi_next) {
4271                                 /* Remove last page from this bio */
4272                                 bio2->bi_vcnt--;
4273                                 bio2->bi_size -= len;
4274                                 bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4275                         }
4276                         goto bio_full;
4277                 }
4278                 sector_nr += len >> 9;
4279                 nr_sectors += len >> 9;
4280         }
4281 bio_full:
4282         r10_bio->sectors = nr_sectors;
4283
4284         /* Now submit the read */
4285         md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4286         atomic_inc(&r10_bio->remaining);
4287         read_bio->bi_next = NULL;
4288         generic_make_request(read_bio);
4289         sector_nr += nr_sectors;
4290         sectors_done += nr_sectors;
4291         if (sector_nr <= last)
4292                 goto read_more;
4293
4294         /* Now that we have done the whole section we can
4295          * update reshape_progress
4296          */
4297         if (mddev->reshape_backwards)
4298                 conf->reshape_progress -= sectors_done;
4299         else
4300                 conf->reshape_progress += sectors_done;
4301
4302         return sectors_done;
4303 }
4304
4305 static void end_reshape_request(struct r10bio *r10_bio);
4306 static int handle_reshape_read_error(struct mddev *mddev,
4307                                      struct r10bio *r10_bio);
4308 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4309 {
4310         /* Reshape read completed.  Hopefully we have a block
4311          * to write out.
4312          * If we got a read error then we do sync 1-page reads from
4313          * elsewhere until we find the data - or give up.
4314          */
4315         struct r10conf *conf = mddev->private;
4316         int s;
4317
4318         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4319                 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4320                         /* Reshape has been aborted */
4321                         md_done_sync(mddev, r10_bio->sectors, 0);
4322                         return;
4323                 }
4324
4325         /* We definitely have the data in the pages, schedule the
4326          * writes.
4327          */
4328         atomic_set(&r10_bio->remaining, 1);
4329         for (s = 0; s < conf->copies*2; s++) {
4330                 struct bio *b;
4331                 int d = r10_bio->devs[s/2].devnum;
4332                 struct md_rdev *rdev;
4333                 if (s&1) {
4334                         rdev = conf->mirrors[d].replacement;
4335                         b = r10_bio->devs[s/2].repl_bio;
4336                 } else {
4337                         rdev = conf->mirrors[d].rdev;
4338                         b = r10_bio->devs[s/2].bio;
4339                 }
4340                 if (!rdev || test_bit(Faulty, &rdev->flags))
4341                         continue;
4342                 atomic_inc(&rdev->nr_pending);
4343                 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4344                 atomic_inc(&r10_bio->remaining);
4345                 b->bi_next = NULL;
4346                 generic_make_request(b);
4347         }
4348         end_reshape_request(r10_bio);
4349 }
4350
4351 static void end_reshape(struct r10conf *conf)
4352 {
4353         if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4354                 return;
4355
4356         spin_lock_irq(&conf->device_lock);
4357         conf->prev = conf->geo;
4358         md_finish_reshape(conf->mddev);
4359         smp_wmb();
4360         conf->reshape_progress = MaxSector;
4361         spin_unlock_irq(&conf->device_lock);
4362
4363         /* read-ahead size must cover two whole stripes, which is
4364          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4365          */
4366         if (conf->mddev->queue) {
4367                 int stripe = conf->geo.raid_disks *
4368                         ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4369                 stripe /= conf->geo.near_copies;
4370                 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4371                         conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4372         }
4373         conf->fullsync = 0;
4374 }
4375
4376
4377 static int handle_reshape_read_error(struct mddev *mddev,
4378                                      struct r10bio *r10_bio)
4379 {
4380         /* Use sync reads to get the blocks from somewhere else */
4381         int sectors = r10_bio->sectors;
4382         struct r10bio r10b;
4383         struct r10conf *conf = mddev->private;
4384         int slot = 0;
4385         int idx = 0;
4386         struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4387
4388         r10b.sector = r10_bio->sector;
4389         __raid10_find_phys(&conf->prev, &r10b);
4390
4391         while (sectors) {
4392                 int s = sectors;
4393                 int success = 0;
4394                 int first_slot = slot;
4395
4396                 if (s > (PAGE_SIZE >> 9))
4397                         s = PAGE_SIZE >> 9;
4398
4399                 while (!success) {
4400                         int d = r10b.devs[slot].devnum;
4401                         struct md_rdev *rdev = conf->mirrors[d].rdev;
4402                         sector_t addr;
4403                         if (rdev == NULL ||
4404                             test_bit(Faulty, &rdev->flags) ||
4405                             !test_bit(In_sync, &rdev->flags))
4406                                 goto failed;
4407
4408                         addr = r10b.devs[slot].addr + idx * PAGE_SIZE;
4409                         success = sync_page_io(rdev,
4410                                                addr,
4411                                                s << 9,
4412                                                bvec[idx].bv_page,
4413                                                READ, false);
4414                         if (success)
4415                                 break;
4416                 failed:
4417                         slot++;
4418                         if (slot >= conf->copies)
4419                                 slot = 0;
4420                         if (slot == first_slot)
4421                                 break;
4422                 }
4423                 if (!success) {
4424                         /* couldn't read this block, must give up */
4425                         set_bit(MD_RECOVERY_INTR,
4426                                 &mddev->recovery);
4427                         return -EIO;
4428                 }
4429                 sectors -= s;
4430                 idx++;
4431         }
4432         return 0;
4433 }
4434
4435 static void end_reshape_write(struct bio *bio, int error)
4436 {
4437         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4438         struct r10bio *r10_bio = bio->bi_private;
4439         struct mddev *mddev = r10_bio->mddev;
4440         struct r10conf *conf = mddev->private;
4441         int d;
4442         int slot;
4443         int repl;
4444         struct md_rdev *rdev = NULL;
4445
4446         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4447         if (repl)
4448                 rdev = conf->mirrors[d].replacement;
4449         if (!rdev) {
4450                 smp_mb();
4451                 rdev = conf->mirrors[d].rdev;
4452         }
4453
4454         if (!uptodate) {
4455                 /* FIXME should record badblock */
4456                 md_error(mddev, rdev);
4457         }
4458
4459         rdev_dec_pending(rdev, mddev);
4460         end_reshape_request(r10_bio);
4461 }
4462
4463 static void end_reshape_request(struct r10bio *r10_bio)
4464 {
4465         if (!atomic_dec_and_test(&r10_bio->remaining))
4466                 return;
4467         md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4468         bio_put(r10_bio->master_bio);
4469         put_buf(r10_bio);
4470 }
4471
4472 static void raid10_finish_reshape(struct mddev *mddev)
4473 {
4474         struct r10conf *conf = mddev->private;
4475
4476         if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4477                 return;
4478
4479         if (mddev->delta_disks > 0) {
4480                 sector_t size = raid10_size(mddev, 0, 0);
4481                 md_set_array_sectors(mddev, size);
4482                 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4483                         mddev->recovery_cp = mddev->resync_max_sectors;
4484                         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4485                 }
4486                 mddev->resync_max_sectors = size;
4487                 set_capacity(mddev->gendisk, mddev->array_sectors);
4488                 revalidate_disk(mddev->gendisk);
4489         } else {
4490                 int d;
4491                 for (d = conf->geo.raid_disks ;
4492                      d < conf->geo.raid_disks - mddev->delta_disks;
4493                      d++) {
4494                         struct md_rdev *rdev = conf->mirrors[d].rdev;
4495                         if (rdev)
4496                                 clear_bit(In_sync, &rdev->flags);
4497                         rdev = conf->mirrors[d].replacement;
4498                         if (rdev)
4499                                 clear_bit(In_sync, &rdev->flags);
4500                 }
4501         }
4502         mddev->layout = mddev->new_layout;
4503         mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4504         mddev->reshape_position = MaxSector;
4505         mddev->delta_disks = 0;
4506         mddev->reshape_backwards = 0;
4507 }
4508
4509 static struct md_personality raid10_personality =
4510 {
4511         .name           = "raid10",
4512         .level          = 10,
4513         .owner          = THIS_MODULE,
4514         .make_request   = make_request,
4515         .run            = run,
4516         .stop           = stop,
4517         .status         = status,
4518         .error_handler  = error,
4519         .hot_add_disk   = raid10_add_disk,
4520         .hot_remove_disk= raid10_remove_disk,
4521         .spare_active   = raid10_spare_active,
4522         .sync_request   = sync_request,
4523         .quiesce        = raid10_quiesce,
4524         .size           = raid10_size,
4525         .resize         = raid10_resize,
4526         .takeover       = raid10_takeover,
4527         .check_reshape  = raid10_check_reshape,
4528         .start_reshape  = raid10_start_reshape,
4529         .finish_reshape = raid10_finish_reshape,
4530 };
4531
4532 static int __init raid_init(void)
4533 {
4534         return register_md_personality(&raid10_personality);
4535 }
4536
4537 static void raid_exit(void)
4538 {
4539         unregister_md_personality(&raid10_personality);
4540 }
4541
4542 module_init(raid_init);
4543 module_exit(raid_exit);
4544 MODULE_LICENSE("GPL");
4545 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4546 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4547 MODULE_ALIAS("md-raid10");
4548 MODULE_ALIAS("md-level-10");
4549
4550 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);