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 "md.h"
  28 #include "raid10.h"
  29 #include "raid0.h"
  30 #include "bitmap.h"
  31
  32 /*
  33  * RAID10 provides a combination of RAID0 and RAID1 functionality.
  34  * The layout of data is defined by
  35  *    chunk_size
  36  *    raid_disks
  37  *    near_copies (stored in low byte of layout)
  38  *    far_copies (stored in second byte of layout)
  39  *    far_offset (stored in bit 16 of layout )
  40  *
  41  * The data to be stored is divided into chunks using chunksize.
  42  * Each device is divided into far_copies sections.
  43  * In each section, chunks are laid out in a style similar to raid0, but
  44  * near_copies copies of each chunk is stored (each on a different drive).
  45  * The starting device for each section is offset near_copies from the starting
  46  * device of the previous section.
  47  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
  48  * drive.
  49  * near_copies and far_copies must be at least one, and their product is at most
  50  * raid_disks.
  51  *
  52  * If far_offset is true, then the far_copies are handled a bit differently.
  53  * The copies are still in different stripes, but instead of be very far apart
  54  * on disk, there are adjacent stripes.
  55  */
  56
  57 /*
  58  * Number of guaranteed r10bios in case of extreme VM load:
  59  */
  60 #define NR_RAID10_BIOS 256
  61
  62 /* When there are this many requests queue to be written by
  63  * the raid10 thread, we become 'congested' to provide back-pressure
  64  * for writeback.
  65  */
  66 static int max_queued_requests = 1024;
  67
  68 static void allow_barrier(struct r10conf *conf);
  69 static void lower_barrier(struct r10conf *conf);
  70
  71 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
  72 {
  73         struct r10conf *conf = data;
  74         int size = offsetof(struct r10bio, devs[conf->copies]);
  75
  76         /* allocate a r10bio with room for raid_disks entries in the
  77          * bios array */
  78         return kzalloc(size, gfp_flags);
  79 }
  80
  81 static void r10bio_pool_free(void *r10_bio, void *data)
  82 {
  83         kfree(r10_bio);
  84 }
  85
  86 /* Maximum size of each resync request */
  87 #define RESYNC_BLOCK_SIZE (64*1024)
  88 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  89 /* amount of memory to reserve for resync requests */
  90 #define RESYNC_WINDOW (1024*1024)
  91 /* maximum number of concurrent requests, memory permitting */
  92 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
  93
  94 /*
  95  * When performing a resync, we need to read and compare, so
  96  * we need as many pages are there are copies.
  97  * When performing a recovery, we need 2 bios, one for read,
  98  * one for write (we recover only one drive per r10buf)
  99  *
 100  */
 101 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
 102 {
 103         struct r10conf *conf = data;
 104         struct page *page;
 105         struct r10bio *r10_bio;
 106         struct bio *bio;
 107         int i, j;
 108         int nalloc;
 109
 110         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 111         if (!r10_bio)
 112                 return NULL;
 113
 114         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
 115                 nalloc = conf->copies; /* resync */
 116         else
 117                 nalloc = 2; /* recovery */
 118
 119         /*
 120          * Allocate bios.
 121          */
 122         for (j = nalloc ; j-- ; ) {
 123                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 124                 if (!bio)
 125                         goto out_free_bio;
 126                 r10_bio->devs[j].bio = bio;
 127                 if (!conf->have_replacement)
 128                         continue;
 129                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 130                 if (!bio)
 131                         goto out_free_bio;
 132                 r10_bio->devs[j].repl_bio = bio;
 133         }
 134         /*
 135          * Allocate RESYNC_PAGES data pages and attach them
 136          * where needed.
 137          */
 138         for (j = 0 ; j < nalloc; j++) {
 139                 struct bio *rbio = r10_bio->devs[j].repl_bio;
 140                 bio = r10_bio->devs[j].bio;
 141                 for (i = 0; i < RESYNC_PAGES; i++) {
 142                         if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
 143                                                 &conf->mddev->recovery)) {
 144                                 /* we can share bv_page's during recovery */
 145                                 struct bio *rbio = r10_bio->devs[0].bio;
 146                                 page = rbio->bi_io_vec[i].bv_page;
 147                                 get_page(page);
 148                         } else
 149                                 page = alloc_page(gfp_flags);
 150                         if (unlikely(!page))
 151                                 goto out_free_pages;
 152
 153                         bio->bi_io_vec[i].bv_page = page;
 154                         if (rbio)
 155                                 rbio->bi_io_vec[i].bv_page = page;
 156                 }
 157         }
 158
 159         return r10_bio;
 160
 161 out_free_pages:
 162         for ( ; i > 0 ; i--)
 163                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
 164         while (j--)
 165                 for (i = 0; i < RESYNC_PAGES ; i++)
 166                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 167         j = -1;
 168 out_free_bio:
 169         while (++j < nalloc) {
 170                 bio_put(r10_bio->devs[j].bio);
 171                 if (r10_bio->devs[j].repl_bio)
 172                         bio_put(r10_bio->devs[j].repl_bio);
 173         }
 174         r10bio_pool_free(r10_bio, conf);
 175         return NULL;
 176 }
 177
 178 static void r10buf_pool_free(void *__r10_bio, void *data)
 179 {
 180         int i;
 181         struct r10conf *conf = data;
 182         struct r10bio *r10bio = __r10_bio;
 183         int j;
 184
 185         for (j=0; j < conf->copies; j++) {
 186                 struct bio *bio = r10bio->devs[j].bio;
 187                 if (bio) {
 188                         for (i = 0; i < RESYNC_PAGES; i++) {
 189                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 190                                 bio->bi_io_vec[i].bv_page = NULL;
 191                         }
 192                         bio_put(bio);
 193                 }
 194                 bio = r10bio->devs[j].repl_bio;
 195                 if (bio)
 196                         bio_put(bio);
 197         }
 198         r10bio_pool_free(r10bio, conf);
 199 }
 200
 201 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
 202 {
 203         int i;
 204
 205         for (i = 0; i < conf->copies; i++) {
 206                 struct bio **bio = & r10_bio->devs[i].bio;
 207                 if (!BIO_SPECIAL(*bio))
 208                         bio_put(*bio);
 209                 *bio = NULL;
 210                 bio = &r10_bio->devs[i].repl_bio;
 211                 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
 212                         bio_put(*bio);
 213                 *bio = NULL;
 214         }
 215 }
 216
 217 static void free_r10bio(struct r10bio *r10_bio)
 218 {
 219         struct r10conf *conf = r10_bio->mddev->private;
 220
 221         put_all_bios(conf, r10_bio);
 222         mempool_free(r10_bio, conf->r10bio_pool);
 223 }
 224
 225 static void put_buf(struct r10bio *r10_bio)
 226 {
 227         struct r10conf *conf = r10_bio->mddev->private;
 228
 229         mempool_free(r10_bio, conf->r10buf_pool);
 230
 231         lower_barrier(conf);
 232 }
 233
 234 static void reschedule_retry(struct r10bio *r10_bio)
 235 {
 236         unsigned long flags;
 237         struct mddev *mddev = r10_bio->mddev;
 238         struct r10conf *conf = mddev->private;
 239
 240         spin_lock_irqsave(&conf->device_lock, flags);
 241         list_add(&r10_bio->retry_list, &conf->retry_list);
 242         conf->nr_queued ++;
 243         spin_unlock_irqrestore(&conf->device_lock, flags);
 244
 245         /* wake up frozen array... */
 246         wake_up(&conf->wait_barrier);
 247
 248         md_wakeup_thread(mddev->thread);
 249 }
 250
 251 /*
 252  * raid_end_bio_io() is called when we have finished servicing a mirrored
 253  * operation and are ready to return a success/failure code to the buffer
 254  * cache layer.
 255  */
 256 static void raid_end_bio_io(struct r10bio *r10_bio)
 257 {
 258         struct bio *bio = r10_bio->master_bio;
 259         int done;
 260         struct r10conf *conf = r10_bio->mddev->private;
 261
 262         if (bio->bi_phys_segments) {
 263                 unsigned long flags;
 264                 spin_lock_irqsave(&conf->device_lock, flags);
 265                 bio->bi_phys_segments--;
 266                 done = (bio->bi_phys_segments == 0);
 267                 spin_unlock_irqrestore(&conf->device_lock, flags);
 268         } else
 269                 done = 1;
 270         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
 271                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 272         if (done) {
 273                 bio_endio(bio, 0);
 274                 /*
 275                  * Wake up any possible resync thread that waits for the device
 276                  * to go idle.
 277                  */
 278                 allow_barrier(conf);
 279         }
 280         free_r10bio(r10_bio);
 281 }
 282
 283 /*
 284  * Update disk head position estimator based on IRQ completion info.
 285  */
 286 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
 287 {
 288         struct r10conf *conf = r10_bio->mddev->private;
 289
 290         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 291                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 292 }
 293
 294 /*
 295  * Find the disk number which triggered given bio
 296  */
 297 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
 298                          struct bio *bio, int *slotp, int *replp)
 299 {
 300         int slot;
 301         int repl = 0;
 302
 303         for (slot = 0; slot < conf->copies; slot++) {
 304                 if (r10_bio->devs[slot].bio == bio)
 305                         break;
 306                 if (r10_bio->devs[slot].repl_bio == bio) {
 307                         repl = 1;
 308                         break;
 309                 }
 310         }
 311
 312         BUG_ON(slot == conf->copies);
 313         update_head_pos(slot, r10_bio);
 314
 315         if (slotp)
 316                 *slotp = slot;
 317         if (replp)
 318                 *replp = repl;
 319         return r10_bio->devs[slot].devnum;
 320 }
 321
 322 static void raid10_end_read_request(struct bio *bio, int error)
 323 {
 324         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 325         struct r10bio *r10_bio = bio->bi_private;
 326         int slot, dev;
 327         struct md_rdev *rdev;
 328         struct r10conf *conf = r10_bio->mddev->private;
 329
 330
 331         slot = r10_bio->read_slot;
 332         dev = r10_bio->devs[slot].devnum;
 333         rdev = r10_bio->devs[slot].rdev;
 334         /*
 335          * this branch is our 'one mirror IO has finished' event handler:
 336          */
 337         update_head_pos(slot, r10_bio);
 338
 339         if (uptodate) {
 340                 /*
 341                  * Set R10BIO_Uptodate in our master bio, so that
 342                  * we will return a good error code to the higher
 343                  * levels even if IO on some other mirrored buffer fails.
 344                  *
 345                  * The 'master' represents the composite IO operation to
 346                  * user-side. So if something waits for IO, then it will
 347                  * wait for the 'master' bio.
 348                  */
 349                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 350                 raid_end_bio_io(r10_bio);
 351                 rdev_dec_pending(rdev, conf->mddev);
 352         } else {
 353                 /*
 354                  * oops, read error - keep the refcount on the rdev
 355                  */
 356                 char b[BDEVNAME_SIZE];
 357                 printk_ratelimited(KERN_ERR
 358                                    "md/raid10:%s: %s: rescheduling sector %llu\n",
 359                                    mdname(conf->mddev),
 360                                    bdevname(rdev->bdev, b),
 361                                    (unsigned long long)r10_bio->sector);
 362                 set_bit(R10BIO_ReadError, &r10_bio->state);
 363                 reschedule_retry(r10_bio);
 364         }
 365 }
 366
 367 static void close_write(struct r10bio *r10_bio)
 368 {
 369         /* clear the bitmap if all writes complete successfully */
 370         bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 371                         r10_bio->sectors,
 372                         !test_bit(R10BIO_Degraded, &r10_bio->state),
 373                         0);
 374         md_write_end(r10_bio->mddev);
 375 }
 376
 377 static void one_write_done(struct r10bio *r10_bio)
 378 {
 379         if (atomic_dec_and_test(&r10_bio->remaining)) {
 380                 if (test_bit(R10BIO_WriteError, &r10_bio->state))
 381                         reschedule_retry(r10_bio);
 382                 else {
 383                         close_write(r10_bio);
 384                         if (test_bit(R10BIO_MadeGood, &r10_bio->state))
 385                                 reschedule_retry(r10_bio);
 386                         else
 387                                 raid_end_bio_io(r10_bio);
 388                 }
 389         }
 390 }
 391
 392 static void raid10_end_write_request(struct bio *bio, int error)
 393 {
 394         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 395         struct r10bio *r10_bio = bio->bi_private;
 396         int dev;
 397         int dec_rdev = 1;
 398         struct r10conf *conf = r10_bio->mddev->private;
 399         int slot, repl;
 400         struct md_rdev *rdev;
 401
 402         dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
 403
 404         if (repl)
 405                 rdev = conf->mirrors[dev].replacement;
 406         else
 407                 rdev = conf->mirrors[dev].rdev;
 408         /*
 409          * this branch is our 'one mirror IO has finished' event handler:
 410          */
 411         if (!uptodate) {
 412                 if (repl)
 413                         /* Never record new bad blocks to replacement,
 414                          * just fail it.
 415                          */
 416                         md_error(rdev->mddev, rdev);
 417                 else {
 418                         set_bit(WriteErrorSeen, &rdev->flags);
 419                         set_bit(R10BIO_WriteError, &r10_bio->state);
 420                         dec_rdev = 0;
 421                 }
 422         } else {
 423                 /*
 424                  * Set R10BIO_Uptodate in our master bio, so that
 425                  * we will return a good error code for to the higher
 426                  * levels even if IO on some other mirrored buffer fails.
 427                  *
 428                  * The 'master' represents the composite IO operation to
 429                  * user-side. So if something waits for IO, then it will
 430                  * wait for the 'master' bio.
 431                  */
 432                 sector_t first_bad;
 433                 int bad_sectors;
 434
 435                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 436
 437                 /* Maybe we can clear some bad blocks. */
 438                 if (is_badblock(rdev,
 439                                 r10_bio->devs[slot].addr,
 440                                 r10_bio->sectors,
 441                                 &first_bad, &bad_sectors)) {
 442                         bio_put(bio);
 443                         if (repl)
 444                                 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
 445                         else
 446                                 r10_bio->devs[slot].bio = IO_MADE_GOOD;
 447                         dec_rdev = 0;
 448                         set_bit(R10BIO_MadeGood, &r10_bio->state);
 449                 }
 450         }
 451
 452         /*
 453          *
 454          * Let's see if all mirrored write operations have finished
 455          * already.
 456          */
 457         one_write_done(r10_bio);
 458         if (dec_rdev)
 459                 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 460 }
 461
 462 /*
 463  * RAID10 layout manager
 464  * As well as the chunksize and raid_disks count, there are two
 465  * parameters: near_copies and far_copies.
 466  * near_copies * far_copies must be <= raid_disks.
 467  * Normally one of these will be 1.
 468  * If both are 1, we get raid0.
 469  * If near_copies == raid_disks, we get raid1.
 470  *
 471  * Chunks are laid out in raid0 style with near_copies copies of the
 472  * first chunk, followed by near_copies copies of the next chunk and
 473  * so on.
 474  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 475  * as described above, we start again with a device offset of near_copies.
 476  * So we effectively have another copy of the whole array further down all
 477  * the drives, but with blocks on different drives.
 478  * With this layout, and block is never stored twice on the one device.
 479  *
 480  * raid10_find_phys finds the sector offset of a given virtual sector
 481  * on each device that it is on.
 482  *
 483  * raid10_find_virt does the reverse mapping, from a device and a
 484  * sector offset to a virtual address
 485  */
 486
 487 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
 488 {
 489         int n,f;
 490         sector_t sector;
 491         sector_t chunk;
 492         sector_t stripe;
 493         int dev;
 494
 495         int slot = 0;
 496
 497         /* now calculate first sector/dev */
 498         chunk = r10bio->sector >> conf->chunk_shift;
 499         sector = r10bio->sector & conf->chunk_mask;
 500
 501         chunk *= conf->near_copies;
 502         stripe = chunk;
 503         dev = sector_div(stripe, conf->raid_disks);
 504         if (conf->far_offset)
 505                 stripe *= conf->far_copies;
 506
 507         sector += stripe << conf->chunk_shift;
 508
 509         /* and calculate all the others */
 510         for (n=0; n < conf->near_copies; n++) {
 511                 int d = dev;
 512                 sector_t s = sector;
 513                 r10bio->devs[slot].addr = sector;
 514                 r10bio->devs[slot].devnum = d;
 515                 slot++;
 516
 517                 for (f = 1; f < conf->far_copies; f++) {
 518                         d += conf->near_copies;
 519                         if (d >= conf->raid_disks)
 520                                 d -= conf->raid_disks;
 521                         s += conf->stride;
 522                         r10bio->devs[slot].devnum = d;
 523                         r10bio->devs[slot].addr = s;
 524                         slot++;
 525                 }
 526                 dev++;
 527                 if (dev >= conf->raid_disks) {
 528                         dev = 0;
 529                         sector += (conf->chunk_mask + 1);
 530                 }
 531         }
 532         BUG_ON(slot != conf->copies);
 533 }
 534
 535 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
 536 {
 537         sector_t offset, chunk, vchunk;
 538
 539         offset = sector & conf->chunk_mask;
 540         if (conf->far_offset) {
 541                 int fc;
 542                 chunk = sector >> conf->chunk_shift;
 543                 fc = sector_div(chunk, conf->far_copies);
 544                 dev -= fc * conf->near_copies;
 545                 if (dev < 0)
 546                         dev += conf->raid_disks;
 547         } else {
 548                 while (sector >= conf->stride) {
 549                         sector -= conf->stride;
 550                         if (dev < conf->near_copies)
 551                                 dev += conf->raid_disks - conf->near_copies;
 552                         else
 553                                 dev -= conf->near_copies;
 554                 }
 555                 chunk = sector >> conf->chunk_shift;
 556         }
 557         vchunk = chunk * conf->raid_disks + dev;
 558         sector_div(vchunk, conf->near_copies);
 559         return (vchunk << conf->chunk_shift) + offset;
 560 }
 561
 562 /**
 563  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 564  *      @q: request queue
 565  *      @bvm: properties of new bio
 566  *      @biovec: the request that could be merged to it.
 567  *
 568  *      Return amount of bytes we can accept at this offset
 569  *      If near_copies == raid_disk, there are no striping issues,
 570  *      but in that case, the function isn't called at all.
 571  */
 572 static int raid10_mergeable_bvec(struct request_queue *q,
 573                                  struct bvec_merge_data *bvm,
 574                                  struct bio_vec *biovec)
 575 {
 576         struct mddev *mddev = q->queuedata;
 577         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 578         int max;
 579         unsigned int chunk_sectors = mddev->chunk_sectors;
 580         unsigned int bio_sectors = bvm->bi_size >> 9;
 581
 582         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 583         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 584         if (max <= biovec->bv_len && bio_sectors == 0)
 585                 return biovec->bv_len;
 586         else
 587                 return max;
 588 }
 589
 590 /*
 591  * This routine returns the disk from which the requested read should
 592  * be done. There is a per-array 'next expected sequential IO' sector
 593  * number - if this matches on the next IO then we use the last disk.
 594  * There is also a per-disk 'last know head position' sector that is
 595  * maintained from IRQ contexts, both the normal and the resync IO
 596  * completion handlers update this position correctly. If there is no
 597  * perfect sequential match then we pick the disk whose head is closest.
 598  *
 599  * If there are 2 mirrors in the same 2 devices, performance degrades
 600  * because position is mirror, not device based.
 601  *
 602  * The rdev for the device selected will have nr_pending incremented.
 603  */
 604
 605 /*
 606  * FIXME: possibly should rethink readbalancing and do it differently
 607  * depending on near_copies / far_copies geometry.
 608  */
 609 static struct md_rdev *read_balance(struct r10conf *conf,
 610                                     struct r10bio *r10_bio,
 611                                     int *max_sectors)
 612 {
 613         const sector_t this_sector = r10_bio->sector;
 614         int disk, slot;
 615         int sectors = r10_bio->sectors;
 616         int best_good_sectors;
 617         sector_t new_distance, best_dist;
 618         struct md_rdev *rdev, *best_rdev;
 619         int do_balance;
 620         int best_slot;
 621
 622         raid10_find_phys(conf, r10_bio);
 623         rcu_read_lock();
 624 retry:
 625         sectors = r10_bio->sectors;
 626         best_slot = -1;
 627         best_rdev = NULL;
 628         best_dist = MaxSector;
 629         best_good_sectors = 0;
 630         do_balance = 1;
 631         /*
 632          * Check if we can balance. We can balance on the whole
 633          * device if no resync is going on (recovery is ok), or below
 634          * the resync window. We take the first readable disk when
 635          * above the resync window.
 636          */
 637         if (conf->mddev->recovery_cp < MaxSector
 638             && (this_sector + sectors >= conf->next_resync))
 639                 do_balance = 0;
 640
 641         for (slot = 0; slot < conf->copies ; slot++) {
 642                 sector_t first_bad;
 643                 int bad_sectors;
 644                 sector_t dev_sector;
 645
 646                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
 647                         continue;
 648                 disk = r10_bio->devs[slot].devnum;
 649                 rdev = rcu_dereference(conf->mirrors[disk].replacement);
 650                 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
 651                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 652                         rdev = rcu_dereference(conf->mirrors[disk].rdev);
 653                 if (rdev == NULL)
 654                         continue;
 655                 if (test_bit(Faulty, &rdev->flags))
 656                         continue;
 657                 if (!test_bit(In_sync, &rdev->flags) &&
 658                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 659                         continue;
 660
 661                 dev_sector = r10_bio->devs[slot].addr;
 662                 if (is_badblock(rdev, dev_sector, sectors,
 663                                 &first_bad, &bad_sectors)) {
 664                         if (best_dist < MaxSector)
 665                                 /* Already have a better slot */
 666                                 continue;
 667                         if (first_bad <= dev_sector) {
 668                                 /* Cannot read here.  If this is the
 669                                  * 'primary' device, then we must not read
 670                                  * beyond 'bad_sectors' from another device.
 671                                  */
 672                                 bad_sectors -= (dev_sector - first_bad);
 673                                 if (!do_balance && sectors > bad_sectors)
 674                                         sectors = bad_sectors;
 675                                 if (best_good_sectors > sectors)
 676                                         best_good_sectors = sectors;
 677                         } else {
 678                                 sector_t good_sectors =
 679                                         first_bad - dev_sector;
 680                                 if (good_sectors > best_good_sectors) {
 681                                         best_good_sectors = good_sectors;
 682                                         best_slot = slot;
 683                                         best_rdev = rdev;
 684                                 }
 685                                 if (!do_balance)
 686                                         /* Must read from here */
 687                                         break;
 688                         }
 689                         continue;
 690                 } else
 691                         best_good_sectors = sectors;
 692
 693                 if (!do_balance)
 694                         break;
 695
 696                 /* This optimisation is debatable, and completely destroys
 697                  * sequential read speed for 'far copies' arrays.  So only
 698                  * keep it for 'near' arrays, and review those later.
 699                  */
 700                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
 701                         break;
 702
 703                 /* for far > 1 always use the lowest address */
 704                 if (conf->far_copies > 1)
 705                         new_distance = r10_bio->devs[slot].addr;
 706                 else
 707                         new_distance = abs(r10_bio->devs[slot].addr -
 708                                            conf->mirrors[disk].head_position);
 709                 if (new_distance < best_dist) {
 710                         best_dist = new_distance;
 711                         best_slot = slot;
 712                         best_rdev = rdev;
 713                 }
 714         }
 715         if (slot >= conf->copies) {
 716                 slot = best_slot;
 717                 rdev = best_rdev;
 718         }
 719
 720         if (slot >= 0) {
 721                 atomic_inc(&rdev->nr_pending);
 722                 if (test_bit(Faulty, &rdev->flags)) {
 723                         /* Cannot risk returning a device that failed
 724                          * before we inc'ed nr_pending
 725                          */
 726                         rdev_dec_pending(rdev, conf->mddev);
 727                         goto retry;
 728                 }
 729                 r10_bio->read_slot = slot;
 730         } else
 731                 rdev = NULL;
 732         rcu_read_unlock();
 733         *max_sectors = best_good_sectors;
 734
 735         return rdev;
 736 }
 737
 738 static int raid10_congested(void *data, int bits)
 739 {
 740         struct mddev *mddev = data;
 741         struct r10conf *conf = mddev->private;
 742         int i, ret = 0;
 743
 744         if ((bits & (1 << BDI_async_congested)) &&
 745             conf->pending_count >= max_queued_requests)
 746                 return 1;
 747
 748         if (mddev_congested(mddev, bits))
 749                 return 1;
 750         rcu_read_lock();
 751         for (i = 0; i < conf->raid_disks && ret == 0; i++) {
 752                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 753                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 754                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 755
 756                         ret |= bdi_congested(&q->backing_dev_info, bits);
 757                 }
 758         }
 759         rcu_read_unlock();
 760         return ret;
 761 }
 762
 763 static void flush_pending_writes(struct r10conf *conf)
 764 {
 765         /* Any writes that have been queued but are awaiting
 766          * bitmap updates get flushed here.
 767          */
 768         spin_lock_irq(&conf->device_lock);
 769
 770         if (conf->pending_bio_list.head) {
 771                 struct bio *bio;
 772                 bio = bio_list_get(&conf->pending_bio_list);
 773                 conf->pending_count = 0;
 774                 spin_unlock_irq(&conf->device_lock);
 775                 /* flush any pending bitmap writes to disk
 776                  * before proceeding w/ I/O */
 777                 bitmap_unplug(conf->mddev->bitmap);
 778                 wake_up(&conf->wait_barrier);
 779
 780                 while (bio) { /* submit pending writes */
 781                         struct bio *next = bio->bi_next;
 782                         bio->bi_next = NULL;
 783                         generic_make_request(bio);
 784                         bio = next;
 785                 }
 786         } else
 787                 spin_unlock_irq(&conf->device_lock);
 788 }
 789
 790 /* Barriers....
 791  * Sometimes we need to suspend IO while we do something else,
 792  * either some resync/recovery, or reconfigure the array.
 793  * To do this we raise a 'barrier'.
 794  * The 'barrier' is a counter that can be raised multiple times
 795  * to count how many activities are happening which preclude
 796  * normal IO.
 797  * We can only raise the barrier if there is no pending IO.
 798  * i.e. if nr_pending == 0.
 799  * We choose only to raise the barrier if no-one is waiting for the
 800  * barrier to go down.  This means that as soon as an IO request
 801  * is ready, no other operations which require a barrier will start
 802  * until the IO request has had a chance.
 803  *
 804  * So: regular IO calls 'wait_barrier'.  When that returns there
 805  *    is no backgroup IO happening,  It must arrange to call
 806  *    allow_barrier when it has finished its IO.
 807  * backgroup IO calls must call raise_barrier.  Once that returns
 808  *    there is no normal IO happeing.  It must arrange to call
 809  *    lower_barrier when the particular background IO completes.
 810  */
 811
 812 static void raise_barrier(struct r10conf *conf, int force)
 813 {
 814         BUG_ON(force && !conf->barrier);
 815         spin_lock_irq(&conf->resync_lock);
 816
 817         /* Wait until no block IO is waiting (unless 'force') */
 818         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 819                             conf->resync_lock, );
 820
 821         /* block any new IO from starting */
 822         conf->barrier++;
 823
 824         /* Now wait for all pending IO to complete */
 825         wait_event_lock_irq(conf->wait_barrier,
 826                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 827                             conf->resync_lock, );
 828
 829         spin_unlock_irq(&conf->resync_lock);
 830 }
 831
 832 static void lower_barrier(struct r10conf *conf)
 833 {
 834         unsigned long flags;
 835         spin_lock_irqsave(&conf->resync_lock, flags);
 836         conf->barrier--;
 837         spin_unlock_irqrestore(&conf->resync_lock, flags);
 838         wake_up(&conf->wait_barrier);
 839 }
 840
 841 static void wait_barrier(struct r10conf *conf)
 842 {
 843         spin_lock_irq(&conf->resync_lock);
 844         if (conf->barrier) {
 845                 conf->nr_waiting++;
 846                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 847                                     conf->resync_lock,
 848                                     );
 849                 conf->nr_waiting--;
 850         }
 851         conf->nr_pending++;
 852         spin_unlock_irq(&conf->resync_lock);
 853 }
 854
 855 static void allow_barrier(struct r10conf *conf)
 856 {
 857         unsigned long flags;
 858         spin_lock_irqsave(&conf->resync_lock, flags);
 859         conf->nr_pending--;
 860         spin_unlock_irqrestore(&conf->resync_lock, flags);
 861         wake_up(&conf->wait_barrier);
 862 }
 863
 864 static void freeze_array(struct r10conf *conf)
 865 {
 866         /* stop syncio and normal IO and wait for everything to
 867          * go quiet.
 868          * We increment barrier and nr_waiting, and then
 869          * wait until nr_pending match nr_queued+1
 870          * This is called in the context of one normal IO request
 871          * that has failed. Thus any sync request that might be pending
 872          * will be blocked by nr_pending, and we need to wait for
 873          * pending IO requests to complete or be queued for re-try.
 874          * Thus the number queued (nr_queued) plus this request (1)
 875          * must match the number of pending IOs (nr_pending) before
 876          * we continue.
 877          */
 878         spin_lock_irq(&conf->resync_lock);
 879         conf->barrier++;
 880         conf->nr_waiting++;
 881         wait_event_lock_irq(conf->wait_barrier,
 882                             conf->nr_pending == conf->nr_queued+1,
 883                             conf->resync_lock,
 884                             flush_pending_writes(conf));
 885
 886         spin_unlock_irq(&conf->resync_lock);
 887 }
 888
 889 static void unfreeze_array(struct r10conf *conf)
 890 {
 891         /* reverse the effect of the freeze */
 892         spin_lock_irq(&conf->resync_lock);
 893         conf->barrier--;
 894         conf->nr_waiting--;
 895         wake_up(&conf->wait_barrier);
 896         spin_unlock_irq(&conf->resync_lock);
 897 }
 898
 899 static void make_request(struct mddev *mddev, struct bio * bio)
 900 {
 901         struct r10conf *conf = mddev->private;
 902         struct r10bio *r10_bio;
 903         struct bio *read_bio;
 904         int i;
 905         int chunk_sects = conf->chunk_mask + 1;
 906         const int rw = bio_data_dir(bio);
 907         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 908         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
 909         unsigned long flags;
 910         struct md_rdev *blocked_rdev;
 911         int plugged;
 912         int sectors_handled;
 913         int max_sectors;
 914
 915         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
 916                 md_flush_request(mddev, bio);
 917                 return;
 918         }
 919
 920         /* If this request crosses a chunk boundary, we need to
 921          * split it.  This will only happen for 1 PAGE (or less) requests.
 922          */
 923         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 924                       > chunk_sects &&
 925                     conf->near_copies < conf->raid_disks)) {
 926                 struct bio_pair *bp;
 927                 /* Sanity check -- queue functions should prevent this happening */
 928                 if (bio->bi_vcnt != 1 ||
 929                     bio->bi_idx != 0)
 930                         goto bad_map;
 931                 /* This is a one page bio that upper layers
 932                  * refuse to split for us, so we need to split it.
 933                  */
 934                 bp = bio_split(bio,
 935                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 936
 937                 /* Each of these 'make_request' calls will call 'wait_barrier'.
 938                  * If the first succeeds but the second blocks due to the resync
 939                  * thread raising the barrier, we will deadlock because the
 940                  * IO to the underlying device will be queued in generic_make_request
 941                  * and will never complete, so will never reduce nr_pending.
 942                  * So increment nr_waiting here so no new raise_barriers will
 943                  * succeed, and so the second wait_barrier cannot block.
 944                  */
 945                 spin_lock_irq(&conf->resync_lock);
 946                 conf->nr_waiting++;
 947                 spin_unlock_irq(&conf->resync_lock);
 948
 949                 make_request(mddev, &bp->bio1);
 950                 make_request(mddev, &bp->bio2);
 951
 952                 spin_lock_irq(&conf->resync_lock);
 953                 conf->nr_waiting--;
 954                 wake_up(&conf->wait_barrier);
 955                 spin_unlock_irq(&conf->resync_lock);
 956
 957                 bio_pair_release(bp);
 958                 return;
 959         bad_map:
 960                 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
 961                        " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
 962                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 963
 964                 bio_io_error(bio);
 965                 return;
 966         }
 967
 968         md_write_start(mddev, bio);
 969
 970         /*
 971          * Register the new request and wait if the reconstruction
 972          * thread has put up a bar for new requests.
 973          * Continue immediately if no resync is active currently.
 974          */
 975         wait_barrier(conf);
 976
 977         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 978
 979         r10_bio->master_bio = bio;
 980         r10_bio->sectors = bio->bi_size >> 9;
 981
 982         r10_bio->mddev = mddev;
 983         r10_bio->sector = bio->bi_sector;
 984         r10_bio->state = 0;
 985
 986         /* We might need to issue multiple reads to different
 987          * devices if there are bad blocks around, so we keep
 988          * track of the number of reads in bio->bi_phys_segments.
 989          * If this is 0, there is only one r10_bio and no locking
 990          * will be needed when the request completes.  If it is
 991          * non-zero, then it is the number of not-completed requests.
 992          */
 993         bio->bi_phys_segments = 0;
 994         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
 995
 996         if (rw == READ) {
 997                 /*
 998                  * read balancing logic:
 999                  */
1000                 struct md_rdev *rdev;
1001                 int slot;
1002
1003 read_again:
1004                 rdev = read_balance(conf, r10_bio, &max_sectors);
1005                 if (!rdev) {
1006                         raid_end_bio_io(r10_bio);
1007                         return;
1008                 }
1009                 slot = r10_bio->read_slot;
1010
1011                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1012                 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1013                             max_sectors);
1014
1015                 r10_bio->devs[slot].bio = read_bio;
1016                 r10_bio->devs[slot].rdev = rdev;
1017
1018                 read_bio->bi_sector = r10_bio->devs[slot].addr +
1019                         rdev->data_offset;
1020                 read_bio->bi_bdev = rdev->bdev;
1021                 read_bio->bi_end_io = raid10_end_read_request;
1022                 read_bio->bi_rw = READ | do_sync;
1023                 read_bio->bi_private = r10_bio;
1024
1025                 if (max_sectors < r10_bio->sectors) {
1026                         /* Could not read all from this device, so we will
1027                          * need another r10_bio.
1028                          */
1029                         sectors_handled = (r10_bio->sectors + max_sectors
1030                                            - bio->bi_sector);
1031                         r10_bio->sectors = max_sectors;
1032                         spin_lock_irq(&conf->device_lock);
1033                         if (bio->bi_phys_segments == 0)
1034                                 bio->bi_phys_segments = 2;
1035                         else
1036                                 bio->bi_phys_segments++;
1037                         spin_unlock(&conf->device_lock);
1038                         /* Cannot call generic_make_request directly
1039                          * as that will be queued in __generic_make_request
1040                          * and subsequent mempool_alloc might block
1041                          * waiting for it.  so hand bio over to raid10d.
1042                          */
1043                         reschedule_retry(r10_bio);
1044
1045                         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1046
1047                         r10_bio->master_bio = bio;
1048                         r10_bio->sectors = ((bio->bi_size >> 9)
1049                                             - sectors_handled);
1050                         r10_bio->state = 0;
1051                         r10_bio->mddev = mddev;
1052                         r10_bio->sector = bio->bi_sector + sectors_handled;
1053                         goto read_again;
1054                 } else
1055                         generic_make_request(read_bio);
1056                 return;
1057         }
1058
1059         /*
1060          * WRITE:
1061          */
1062         if (conf->pending_count >= max_queued_requests) {
1063                 md_wakeup_thread(mddev->thread);
1064                 wait_event(conf->wait_barrier,
1065                            conf->pending_count < max_queued_requests);
1066         }
1067         /* first select target devices under rcu_lock and
1068          * inc refcount on their rdev.  Record them by setting
1069          * bios[x] to bio
1070          * If there are known/acknowledged bad blocks on any device
1071          * on which we have seen a write error, we want to avoid
1072          * writing to those blocks.  This potentially requires several
1073          * writes to write around the bad blocks.  Each set of writes
1074          * gets its own r10_bio with a set of bios attached.  The number
1075          * of r10_bios is recored in bio->bi_phys_segments just as with
1076          * the read case.
1077          */
1078         plugged = mddev_check_plugged(mddev);
1079
1080         r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1081         raid10_find_phys(conf, r10_bio);
1082 retry_write:
1083         blocked_rdev = NULL;
1084         rcu_read_lock();
1085         max_sectors = r10_bio->sectors;
1086
1087         for (i = 0;  i < conf->copies; i++) {
1088                 int d = r10_bio->devs[i].devnum;
1089                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1090                 struct md_rdev *rrdev = rcu_dereference(
1091                         conf->mirrors[d].replacement);
1092                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1093                         atomic_inc(&rdev->nr_pending);
1094                         blocked_rdev = rdev;
1095                         break;
1096                 }
1097                 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1098                         atomic_inc(&rrdev->nr_pending);
1099                         blocked_rdev = rrdev;
1100                         break;
1101                 }
1102                 if (rrdev && test_bit(Faulty, &rrdev->flags))
1103                         rrdev = NULL;
1104
1105                 r10_bio->devs[i].bio = NULL;
1106                 r10_bio->devs[i].repl_bio = NULL;
1107                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1108                         set_bit(R10BIO_Degraded, &r10_bio->state);
1109                         continue;
1110                 }
1111                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1112                         sector_t first_bad;
1113                         sector_t dev_sector = r10_bio->devs[i].addr;
1114                         int bad_sectors;
1115                         int is_bad;
1116
1117                         is_bad = is_badblock(rdev, dev_sector,
1118                                              max_sectors,
1119                                              &first_bad, &bad_sectors);
1120                         if (is_bad < 0) {
1121                                 /* Mustn't write here until the bad block
1122                                  * is acknowledged
1123                                  */
1124                                 atomic_inc(&rdev->nr_pending);
1125                                 set_bit(BlockedBadBlocks, &rdev->flags);
1126                                 blocked_rdev = rdev;
1127                                 break;
1128                         }
1129                         if (is_bad && first_bad <= dev_sector) {
1130                                 /* Cannot write here at all */
1131                                 bad_sectors -= (dev_sector - first_bad);
1132                                 if (bad_sectors < max_sectors)
1133                                         /* Mustn't write more than bad_sectors
1134                                          * to other devices yet
1135                                          */
1136                                         max_sectors = bad_sectors;
1137                                 /* We don't set R10BIO_Degraded as that
1138                                  * only applies if the disk is missing,
1139                                  * so it might be re-added, and we want to
1140                                  * know to recover this chunk.
1141                                  * In this case the device is here, and the
1142                                  * fact that this chunk is not in-sync is
1143                                  * recorded in the bad block log.
1144                                  */
1145                                 continue;
1146                         }
1147                         if (is_bad) {
1148                                 int good_sectors = first_bad - dev_sector;
1149                                 if (good_sectors < max_sectors)
1150                                         max_sectors = good_sectors;
1151                         }
1152                 }
1153                 r10_bio->devs[i].bio = bio;
1154                 atomic_inc(&rdev->nr_pending);
1155                 if (rrdev) {
1156                         r10_bio->devs[i].repl_bio = bio;
1157                         atomic_inc(&rrdev->nr_pending);
1158                 }
1159         }
1160         rcu_read_unlock();
1161
1162         if (unlikely(blocked_rdev)) {
1163                 /* Have to wait for this device to get unblocked, then retry */
1164                 int j;
1165                 int d;
1166
1167                 for (j = 0; j < i; j++) {
1168                         if (r10_bio->devs[j].bio) {
1169                                 d = r10_bio->devs[j].devnum;
1170                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1171                         }
1172                         if (r10_bio->devs[j].repl_bio) {
1173                                 d = r10_bio->devs[j].devnum;
1174                                 rdev_dec_pending(
1175                                         conf->mirrors[d].replacement, mddev);
1176                         }
1177                 }
1178                 allow_barrier(conf);
1179                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1180                 wait_barrier(conf);
1181                 goto retry_write;
1182         }
1183
1184         if (max_sectors < r10_bio->sectors) {
1185                 /* We are splitting this into multiple parts, so
1186                  * we need to prepare for allocating another r10_bio.
1187                  */
1188                 r10_bio->sectors = max_sectors;
1189                 spin_lock_irq(&conf->device_lock);
1190                 if (bio->bi_phys_segments == 0)
1191                         bio->bi_phys_segments = 2;
1192                 else
1193                         bio->bi_phys_segments++;
1194                 spin_unlock_irq(&conf->device_lock);
1195         }
1196         sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1197
1198         atomic_set(&r10_bio->remaining, 1);
1199         bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1200
1201         for (i = 0; i < conf->copies; i++) {
1202                 struct bio *mbio;
1203                 int d = r10_bio->devs[i].devnum;
1204                 if (!r10_bio->devs[i].bio)
1205                         continue;
1206
1207                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1208                 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1209                             max_sectors);
1210                 r10_bio->devs[i].bio = mbio;
1211
1212                 mbio->bi_sector = (r10_bio->devs[i].addr+
1213                                    conf->mirrors[d].rdev->data_offset);
1214                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1215                 mbio->bi_end_io = raid10_end_write_request;
1216                 mbio->bi_rw = WRITE | do_sync | do_fua;
1217                 mbio->bi_private = r10_bio;
1218
1219                 atomic_inc(&r10_bio->remaining);
1220                 spin_lock_irqsave(&conf->device_lock, flags);
1221                 bio_list_add(&conf->pending_bio_list, mbio);
1222                 conf->pending_count++;
1223                 spin_unlock_irqrestore(&conf->device_lock, flags);
1224
1225                 if (!r10_bio->devs[i].repl_bio)
1226                         continue;
1227
1228                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1229                 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1230                             max_sectors);
1231                 r10_bio->devs[i].repl_bio = mbio;
1232
1233                 mbio->bi_sector = (r10_bio->devs[i].addr+
1234                                    conf->mirrors[d].replacement->data_offset);
1235                 mbio->bi_bdev = conf->mirrors[d].replacement->bdev;
1236                 mbio->bi_end_io = raid10_end_write_request;
1237                 mbio->bi_rw = WRITE | do_sync | do_fua;
1238                 mbio->bi_private = r10_bio;
1239
1240                 atomic_inc(&r10_bio->remaining);
1241                 spin_lock_irqsave(&conf->device_lock, flags);
1242                 bio_list_add(&conf->pending_bio_list, mbio);
1243                 conf->pending_count++;
1244                 spin_unlock_irqrestore(&conf->device_lock, flags);
1245         }
1246
1247         /* Don't remove the bias on 'remaining' (one_write_done) until
1248          * after checking if we need to go around again.
1249          */
1250
1251         if (sectors_handled < (bio->bi_size >> 9)) {
1252                 one_write_done(r10_bio);
1253                 /* We need another r10_bio.  It has already been counted
1254                  * in bio->bi_phys_segments.
1255                  */
1256                 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1257
1258                 r10_bio->master_bio = bio;
1259                 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1260
1261                 r10_bio->mddev = mddev;
1262                 r10_bio->sector = bio->bi_sector + sectors_handled;
1263                 r10_bio->state = 0;
1264                 goto retry_write;
1265         }
1266         one_write_done(r10_bio);
1267
1268         /* In case raid10d snuck in to freeze_array */
1269         wake_up(&conf->wait_barrier);
1270
1271         if (do_sync || !mddev->bitmap || !plugged)
1272                 md_wakeup_thread(mddev->thread);
1273 }
1274
1275 static void status(struct seq_file *seq, struct mddev *mddev)
1276 {
1277         struct r10conf *conf = mddev->private;
1278         int i;
1279
1280         if (conf->near_copies < conf->raid_disks)
1281                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1282         if (conf->near_copies > 1)
1283                 seq_printf(seq, " %d near-copies", conf->near_copies);
1284         if (conf->far_copies > 1) {
1285                 if (conf->far_offset)
1286                         seq_printf(seq, " %d offset-copies", conf->far_copies);
1287                 else
1288                         seq_printf(seq, " %d far-copies", conf->far_copies);
1289         }
1290         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1291                                         conf->raid_disks - mddev->degraded);
1292         for (i = 0; i < conf->raid_disks; i++)
1293                 seq_printf(seq, "%s",
1294                               conf->mirrors[i].rdev &&
1295                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1296         seq_printf(seq, "]");
1297 }
1298
1299 /* check if there are enough drives for
1300  * every block to appear on atleast one.
1301  * Don't consider the device numbered 'ignore'
1302  * as we might be about to remove it.
1303  */
1304 static int enough(struct r10conf *conf, int ignore)
1305 {
1306         int first = 0;
1307
1308         do {
1309                 int n = conf->copies;
1310                 int cnt = 0;
1311                 while (n--) {
1312                         if (conf->mirrors[first].rdev &&
1313                             first != ignore)
1314                                 cnt++;
1315                         first = (first+1) % conf->raid_disks;
1316                 }
1317                 if (cnt == 0)
1318                         return 0;
1319         } while (first != 0);
1320         return 1;
1321 }
1322
1323 static void error(struct mddev *mddev, struct md_rdev *rdev)
1324 {
1325         char b[BDEVNAME_SIZE];
1326         struct r10conf *conf = mddev->private;
1327
1328         /*
1329          * If it is not operational, then we have already marked it as dead
1330          * else if it is the last working disks, ignore the error, let the
1331          * next level up know.
1332          * else mark the drive as failed
1333          */
1334         if (test_bit(In_sync, &rdev->flags)
1335             && !enough(conf, rdev->raid_disk))
1336                 /*
1337                  * Don't fail the drive, just return an IO error.
1338                  */
1339                 return;
1340         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1341                 unsigned long flags;
1342                 spin_lock_irqsave(&conf->device_lock, flags);
1343                 mddev->degraded++;
1344                 spin_unlock_irqrestore(&conf->device_lock, flags);
1345                 /*
1346                  * if recovery is running, make sure it aborts.
1347                  */
1348                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1349         }
1350         set_bit(Blocked, &rdev->flags);
1351         set_bit(Faulty, &rdev->flags);
1352         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1353         printk(KERN_ALERT
1354                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1355                "md/raid10:%s: Operation continuing on %d devices.\n",
1356                mdname(mddev), bdevname(rdev->bdev, b),
1357                mdname(mddev), conf->raid_disks - mddev->degraded);
1358 }
1359
1360 static void print_conf(struct r10conf *conf)
1361 {
1362         int i;
1363         struct mirror_info *tmp;
1364
1365         printk(KERN_DEBUG "RAID10 conf printout:\n");
1366         if (!conf) {
1367                 printk(KERN_DEBUG "(!conf)\n");
1368                 return;
1369         }
1370         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1371                 conf->raid_disks);
1372
1373         for (i = 0; i < conf->raid_disks; i++) {
1374                 char b[BDEVNAME_SIZE];
1375                 tmp = conf->mirrors + i;
1376                 if (tmp->rdev)
1377                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1378                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1379                                 !test_bit(Faulty, &tmp->rdev->flags),
1380                                 bdevname(tmp->rdev->bdev,b));
1381         }
1382 }
1383
1384 static void close_sync(struct r10conf *conf)
1385 {
1386         wait_barrier(conf);
1387         allow_barrier(conf);
1388
1389         mempool_destroy(conf->r10buf_pool);
1390         conf->r10buf_pool = NULL;
1391 }
1392
1393 static int raid10_spare_active(struct mddev *mddev)
1394 {
1395         int i;
1396         struct r10conf *conf = mddev->private;
1397         struct mirror_info *tmp;
1398         int count = 0;
1399         unsigned long flags;
1400
1401         /*
1402          * Find all non-in_sync disks within the RAID10 configuration
1403          * and mark them in_sync
1404          */
1405         for (i = 0; i < conf->raid_disks; i++) {
1406                 tmp = conf->mirrors + i;
1407                 if (tmp->rdev
1408                     && !test_bit(Faulty, &tmp->rdev->flags)
1409                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1410                         count++;
1411                         sysfs_notify_dirent(tmp->rdev->sysfs_state);
1412                 }
1413         }
1414         spin_lock_irqsave(&conf->device_lock, flags);
1415         mddev->degraded -= count;
1416         spin_unlock_irqrestore(&conf->device_lock, flags);
1417
1418         print_conf(conf);
1419         return count;
1420 }
1421
1422
1423 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1424 {
1425         struct r10conf *conf = mddev->private;
1426         int err = -EEXIST;
1427         int mirror;
1428         int first = 0;
1429         int last = conf->raid_disks - 1;
1430
1431         if (mddev->recovery_cp < MaxSector)
1432                 /* only hot-add to in-sync arrays, as recovery is
1433                  * very different from resync
1434                  */
1435                 return -EBUSY;
1436         if (!enough(conf, -1))
1437                 return -EINVAL;
1438
1439         if (rdev->raid_disk >= 0)
1440                 first = last = rdev->raid_disk;
1441
1442         if (rdev->saved_raid_disk >= first &&
1443             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1444                 mirror = rdev->saved_raid_disk;
1445         else
1446                 mirror = first;
1447         for ( ; mirror <= last ; mirror++) {
1448                 struct mirror_info *p = &conf->mirrors[mirror];
1449                 if (p->recovery_disabled == mddev->recovery_disabled)
1450                         continue;
1451                 if (p->rdev)
1452                         continue;
1453
1454                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1455                                   rdev->data_offset << 9);
1456                 /* as we don't honour merge_bvec_fn, we must
1457                  * never risk violating it, so limit
1458                  * ->max_segments to one lying with a single
1459                  * page, as a one page request is never in
1460                  * violation.
1461                  */
1462                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1463                         blk_queue_max_segments(mddev->queue, 1);
1464                         blk_queue_segment_boundary(mddev->queue,
1465                                                    PAGE_CACHE_SIZE - 1);
1466                 }
1467
1468                 p->head_position = 0;
1469                 p->recovery_disabled = mddev->recovery_disabled - 1;
1470                 rdev->raid_disk = mirror;
1471                 err = 0;
1472                 if (rdev->saved_raid_disk != mirror)
1473                         conf->fullsync = 1;
1474                 rcu_assign_pointer(p->rdev, rdev);
1475                 break;
1476         }
1477
1478         md_integrity_add_rdev(rdev, mddev);
1479         print_conf(conf);
1480         return err;
1481 }
1482
1483 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1484 {
1485         struct r10conf *conf = mddev->private;
1486         int err = 0;
1487         int number = rdev->raid_disk;
1488         struct md_rdev **rdevp;
1489         struct mirror_info *p = conf->mirrors + number;
1490
1491         print_conf(conf);
1492         if (rdev == p->rdev)
1493                 rdevp = &p->rdev;
1494         else if (rdev == p->replacement)
1495                 rdevp = &p->replacement;
1496         else
1497                 return 0;
1498
1499         if (test_bit(In_sync, &rdev->flags) ||
1500             atomic_read(&rdev->nr_pending)) {
1501                 err = -EBUSY;
1502                 goto abort;
1503         }
1504         /* Only remove faulty devices if recovery
1505          * is not possible.
1506          */
1507         if (!test_bit(Faulty, &rdev->flags) &&
1508             mddev->recovery_disabled != p->recovery_disabled &&
1509             enough(conf, -1)) {
1510                 err = -EBUSY;
1511                 goto abort;
1512         }
1513         *rdevp = NULL;
1514         synchronize_rcu();
1515         if (atomic_read(&rdev->nr_pending)) {
1516                 /* lost the race, try later */
1517                 err = -EBUSY;
1518                 *rdevp = rdev;
1519                 goto abort;
1520         }
1521         err = md_integrity_register(mddev);
1522
1523 abort:
1524
1525         print_conf(conf);
1526         return err;
1527 }
1528
1529
1530 static void end_sync_read(struct bio *bio, int error)
1531 {
1532         struct r10bio *r10_bio = bio->bi_private;
1533         struct r10conf *conf = r10_bio->mddev->private;
1534         int d;
1535
1536         d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1537
1538         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1539                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1540         else
1541                 /* The write handler will notice the lack of
1542                  * R10BIO_Uptodate and record any errors etc
1543                  */
1544                 atomic_add(r10_bio->sectors,
1545                            &conf->mirrors[d].rdev->corrected_errors);
1546
1547         /* for reconstruct, we always reschedule after a read.
1548          * for resync, only after all reads
1549          */
1550         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1551         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1552             atomic_dec_and_test(&r10_bio->remaining)) {
1553                 /* we have read all the blocks,
1554                  * do the comparison in process context in raid10d
1555                  */
1556                 reschedule_retry(r10_bio);
1557         }
1558 }
1559
1560 static void end_sync_request(struct r10bio *r10_bio)
1561 {
1562         struct mddev *mddev = r10_bio->mddev;
1563
1564         while (atomic_dec_and_test(&r10_bio->remaining)) {
1565                 if (r10_bio->master_bio == NULL) {
1566                         /* the primary of several recovery bios */
1567                         sector_t s = r10_bio->sectors;
1568                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1569                             test_bit(R10BIO_WriteError, &r10_bio->state))
1570                                 reschedule_retry(r10_bio);
1571                         else
1572                                 put_buf(r10_bio);
1573                         md_done_sync(mddev, s, 1);
1574                         break;
1575                 } else {
1576                         struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1577                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1578                             test_bit(R10BIO_WriteError, &r10_bio->state))
1579                                 reschedule_retry(r10_bio);
1580                         else
1581                                 put_buf(r10_bio);
1582                         r10_bio = r10_bio2;
1583                 }
1584         }
1585 }
1586
1587 static void end_sync_write(struct bio *bio, int error)
1588 {
1589         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1590         struct r10bio *r10_bio = bio->bi_private;
1591         struct mddev *mddev = r10_bio->mddev;
1592         struct r10conf *conf = mddev->private;
1593         int d;
1594         sector_t first_bad;
1595         int bad_sectors;
1596         int slot;
1597         int repl;
1598         struct md_rdev *rdev;
1599
1600         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1601         if (repl)
1602                 rdev = conf->mirrors[d].replacement;
1603         else
1604                 rdev = conf->mirrors[d].rdev;
1605
1606         if (!uptodate) {
1607                 if (repl)
1608                         md_error(mddev, rdev);
1609                 else {
1610                         set_bit(WriteErrorSeen, &rdev->flags);
1611                         set_bit(R10BIO_WriteError, &r10_bio->state);
1612                 }
1613         } else if (is_badblock(rdev,
1614                              r10_bio->devs[slot].addr,
1615                              r10_bio->sectors,
1616                              &first_bad, &bad_sectors))
1617                 set_bit(R10BIO_MadeGood, &r10_bio->state);
1618
1619         rdev_dec_pending(rdev, mddev);
1620
1621         end_sync_request(r10_bio);
1622 }
1623
1624 /*
1625  * Note: sync and recover and handled very differently for raid10
1626  * This code is for resync.
1627  * For resync, we read through virtual addresses and read all blocks.
1628  * If there is any error, we schedule a write.  The lowest numbered
1629  * drive is authoritative.
1630  * However requests come for physical address, so we need to map.
1631  * For every physical address there are raid_disks/copies virtual addresses,
1632  * which is always are least one, but is not necessarly an integer.
1633  * This means that a physical address can span multiple chunks, so we may
1634  * have to submit multiple io requests for a single sync request.
1635  */
1636 /*
1637  * We check if all blocks are in-sync and only write to blocks that
1638  * aren't in sync
1639  */
1640 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1641 {
1642         struct r10conf *conf = mddev->private;
1643         int i, first;
1644         struct bio *tbio, *fbio;
1645
1646         atomic_set(&r10_bio->remaining, 1);
1647
1648         /* find the first device with a block */
1649         for (i=0; i<conf->copies; i++)
1650                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1651                         break;
1652
1653         if (i == conf->copies)
1654                 goto done;
1655
1656         first = i;
1657         fbio = r10_bio->devs[i].bio;
1658
1659         /* now find blocks with errors */
1660         for (i=0 ; i < conf->copies ; i++) {
1661                 int  j, d;
1662                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1663
1664                 tbio = r10_bio->devs[i].bio;
1665
1666                 if (tbio->bi_end_io != end_sync_read)
1667                         continue;
1668                 if (i == first)
1669                         continue;
1670                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1671                         /* We know that the bi_io_vec layout is the same for
1672                          * both 'first' and 'i', so we just compare them.
1673                          * All vec entries are PAGE_SIZE;
1674                          */
1675                         for (j = 0; j < vcnt; j++)
1676                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1677                                            page_address(tbio->bi_io_vec[j].bv_page),
1678                                            PAGE_SIZE))
1679                                         break;
1680                         if (j == vcnt)
1681                                 continue;
1682                         mddev->resync_mismatches += r10_bio->sectors;
1683                         if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1684                                 /* Don't fix anything. */
1685                                 continue;
1686                 }
1687                 /* Ok, we need to write this bio, either to correct an
1688                  * inconsistency or to correct an unreadable block.
1689                  * First we need to fixup bv_offset, bv_len and
1690                  * bi_vecs, as the read request might have corrupted these
1691                  */
1692                 tbio->bi_vcnt = vcnt;
1693                 tbio->bi_size = r10_bio->sectors << 9;
1694                 tbio->bi_idx = 0;
1695                 tbio->bi_phys_segments = 0;
1696                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1697                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1698                 tbio->bi_next = NULL;
1699                 tbio->bi_rw = WRITE;
1700                 tbio->bi_private = r10_bio;
1701                 tbio->bi_sector = r10_bio->devs[i].addr;
1702
1703                 for (j=0; j < vcnt ; j++) {
1704                         tbio->bi_io_vec[j].bv_offset = 0;
1705                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1706
1707                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1708                                page_address(fbio->bi_io_vec[j].bv_page),
1709                                PAGE_SIZE);
1710                 }
1711                 tbio->bi_end_io = end_sync_write;
1712
1713                 d = r10_bio->devs[i].devnum;
1714                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1715                 atomic_inc(&r10_bio->remaining);
1716                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1717
1718                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1719                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1720                 generic_make_request(tbio);
1721         }
1722
1723         /* Now write out to any replacement devices
1724          * that are active
1725          */
1726         for (i = 0; i < conf->copies; i++) {
1727                 int j, d;
1728                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1729
1730                 tbio = r10_bio->devs[i].repl_bio;
1731                 if (!tbio || !tbio->bi_end_io)
1732                         continue;
1733                 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
1734                     && r10_bio->devs[i].bio != fbio)
1735                         for (j = 0; j < vcnt; j++)
1736                                 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1737                                        page_address(fbio->bi_io_vec[j].bv_page),
1738                                        PAGE_SIZE);
1739                 d = r10_bio->devs[i].devnum;
1740                 atomic_inc(&r10_bio->remaining);
1741                 md_sync_acct(conf->mirrors[d].replacement->bdev,
1742                              tbio->bi_size >> 9);
1743                 generic_make_request(tbio);
1744         }
1745
1746 done:
1747         if (atomic_dec_and_test(&r10_bio->remaining)) {
1748                 md_done_sync(mddev, r10_bio->sectors, 1);
1749                 put_buf(r10_bio);
1750         }
1751 }
1752
1753 /*
1754  * Now for the recovery code.
1755  * Recovery happens across physical sectors.
1756  * We recover all non-is_sync drives by finding the virtual address of
1757  * each, and then choose a working drive that also has that virt address.
1758  * There is a separate r10_bio for each non-in_sync drive.
1759  * Only the first two slots are in use. The first for reading,
1760  * The second for writing.
1761  *
1762  */
1763 static void fix_recovery_read_error(struct r10bio *r10_bio)
1764 {
1765         /* We got a read error during recovery.
1766          * We repeat the read in smaller page-sized sections.
1767          * If a read succeeds, write it to the new device or record
1768          * a bad block if we cannot.
1769          * If a read fails, record a bad block on both old and
1770          * new devices.
1771          */
1772         struct mddev *mddev = r10_bio->mddev;
1773         struct r10conf *conf = mddev->private;
1774         struct bio *bio = r10_bio->devs[0].bio;
1775         sector_t sect = 0;
1776         int sectors = r10_bio->sectors;
1777         int idx = 0;
1778         int dr = r10_bio->devs[0].devnum;
1779         int dw = r10_bio->devs[1].devnum;
1780
1781         while (sectors) {
1782                 int s = sectors;
1783                 struct md_rdev *rdev;
1784                 sector_t addr;
1785                 int ok;
1786
1787                 if (s > (PAGE_SIZE>>9))
1788                         s = PAGE_SIZE >> 9;
1789
1790                 rdev = conf->mirrors[dr].rdev;
1791                 addr = r10_bio->devs[0].addr + sect,
1792                 ok = sync_page_io(rdev,
1793                                   addr,
1794                                   s << 9,
1795                                   bio->bi_io_vec[idx].bv_page,
1796                                   READ, false);
1797                 if (ok) {
1798                         rdev = conf->mirrors[dw].rdev;
1799                         addr = r10_bio->devs[1].addr + sect;
1800                         ok = sync_page_io(rdev,
1801                                           addr,
1802                                           s << 9,
1803                                           bio->bi_io_vec[idx].bv_page,
1804                                           WRITE, false);
1805                         if (!ok)
1806                                 set_bit(WriteErrorSeen, &rdev->flags);
1807                 }
1808                 if (!ok) {
1809                         /* We don't worry if we cannot set a bad block -
1810                          * it really is bad so there is no loss in not
1811                          * recording it yet
1812                          */
1813                         rdev_set_badblocks(rdev, addr, s, 0);
1814
1815                         if (rdev != conf->mirrors[dw].rdev) {
1816                                 /* need bad block on destination too */
1817                                 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
1818                                 addr = r10_bio->devs[1].addr + sect;
1819                                 ok = rdev_set_badblocks(rdev2, addr, s, 0);
1820                                 if (!ok) {
1821                                         /* just abort the recovery */
1822                                         printk(KERN_NOTICE
1823                                                "md/raid10:%s: recovery aborted"
1824                                                " due to read error\n",
1825                                                mdname(mddev));
1826
1827                                         conf->mirrors[dw].recovery_disabled
1828                                                 = mddev->recovery_disabled;
1829                                         set_bit(MD_RECOVERY_INTR,
1830                                                 &mddev->recovery);
1831                                         break;
1832                                 }
1833                         }
1834                 }
1835
1836                 sectors -= s;
1837                 sect += s;
1838                 idx++;
1839         }
1840 }
1841
1842 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1843 {
1844         struct r10conf *conf = mddev->private;
1845         int d;
1846         struct bio *wbio, *wbio2;
1847
1848         if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
1849                 fix_recovery_read_error(r10_bio);
1850                 end_sync_request(r10_bio);
1851                 return;
1852         }
1853
1854         /*
1855          * share the pages with the first bio
1856          * and submit the write request
1857          */
1858         d = r10_bio->devs[1].devnum;
1859         wbio = r10_bio->devs[1].bio;
1860         wbio2 = r10_bio->devs[1].repl_bio;
1861         if (wbio->bi_end_io) {
1862                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1863                 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1864                 generic_make_request(wbio);
1865         }
1866         if (wbio2 && wbio2->bi_end_io) {
1867                 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
1868                 md_sync_acct(conf->mirrors[d].replacement->bdev,
1869                              wbio2->bi_size >> 9);
1870                 generic_make_request(wbio2);
1871         }
1872 }
1873
1874
1875 /*
1876  * Used by fix_read_error() to decay the per rdev read_errors.
1877  * We halve the read error count for every hour that has elapsed
1878  * since the last recorded read error.
1879  *
1880  */
1881 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
1882 {
1883         struct timespec cur_time_mon;
1884         unsigned long hours_since_last;
1885         unsigned int read_errors = atomic_read(&rdev->read_errors);
1886
1887         ktime_get_ts(&cur_time_mon);
1888
1889         if (rdev->last_read_error.tv_sec == 0 &&
1890             rdev->last_read_error.tv_nsec == 0) {
1891                 /* first time we've seen a read error */
1892                 rdev->last_read_error = cur_time_mon;
1893                 return;
1894         }
1895
1896         hours_since_last = (cur_time_mon.tv_sec -
1897                             rdev->last_read_error.tv_sec) / 3600;
1898
1899         rdev->last_read_error = cur_time_mon;
1900
1901         /*
1902          * if hours_since_last is > the number of bits in read_errors
1903          * just set read errors to 0. We do this to avoid
1904          * overflowing the shift of read_errors by hours_since_last.
1905          */
1906         if (hours_since_last >= 8 * sizeof(read_errors))
1907                 atomic_set(&rdev->read_errors, 0);
1908         else
1909                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1910 }
1911
1912 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
1913                             int sectors, struct page *page, int rw)
1914 {
1915         sector_t first_bad;
1916         int bad_sectors;
1917
1918         if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
1919             && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
1920                 return -1;
1921         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1922                 /* success */
1923                 return 1;
1924         if (rw == WRITE)
1925                 set_bit(WriteErrorSeen, &rdev->flags);
1926         /* need to record an error - either for the block or the device */
1927         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1928                 md_error(rdev->mddev, rdev);
1929         return 0;
1930 }
1931
1932 /*
1933  * This is a kernel thread which:
1934  *
1935  *      1.      Retries failed read operations on working mirrors.
1936  *      2.      Updates the raid superblock when problems encounter.
1937  *      3.      Performs writes following reads for array synchronising.
1938  */
1939
1940 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
1941 {
1942         int sect = 0; /* Offset from r10_bio->sector */
1943         int sectors = r10_bio->sectors;
1944         struct md_rdev*rdev;
1945         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1946         int d = r10_bio->devs[r10_bio->read_slot].devnum;
1947
1948         /* still own a reference to this rdev, so it cannot
1949          * have been cleared recently.
1950          */
1951         rdev = conf->mirrors[d].rdev;
1952
1953         if (test_bit(Faulty, &rdev->flags))
1954                 /* drive has already been failed, just ignore any
1955                    more fix_read_error() attempts */
1956                 return;
1957
1958         check_decay_read_errors(mddev, rdev);
1959         atomic_inc(&rdev->read_errors);
1960         if (atomic_read(&rdev->read_errors) > max_read_errors) {
1961                 char b[BDEVNAME_SIZE];
1962                 bdevname(rdev->bdev, b);
1963
1964                 printk(KERN_NOTICE
1965                        "md/raid10:%s: %s: Raid device exceeded "
1966                        "read_error threshold [cur %d:max %d]\n",
1967                        mdname(mddev), b,
1968                        atomic_read(&rdev->read_errors), max_read_errors);
1969                 printk(KERN_NOTICE
1970                        "md/raid10:%s: %s: Failing raid device\n",
1971                        mdname(mddev), b);
1972                 md_error(mddev, conf->mirrors[d].rdev);
1973                 return;
1974         }
1975
1976         while(sectors) {
1977                 int s = sectors;
1978                 int sl = r10_bio->read_slot;
1979                 int success = 0;
1980                 int start;
1981
1982                 if (s > (PAGE_SIZE>>9))
1983                         s = PAGE_SIZE >> 9;
1984
1985                 rcu_read_lock();
1986                 do {
1987                         sector_t first_bad;
1988                         int bad_sectors;
1989
1990                         d = r10_bio->devs[sl].devnum;
1991                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1992                         if (rdev &&
1993                             test_bit(In_sync, &rdev->flags) &&
1994                             is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
1995                                         &first_bad, &bad_sectors) == 0) {
1996                                 atomic_inc(&rdev->nr_pending);
1997                                 rcu_read_unlock();
1998                                 success = sync_page_io(rdev,
1999                                                        r10_bio->devs[sl].addr +
2000                                                        sect,
2001                                                        s<<9,
2002                                                        conf->tmppage, READ, false);
2003                                 rdev_dec_pending(rdev, mddev);
2004                                 rcu_read_lock();
2005                                 if (success)
2006                                         break;
2007                         }
2008                         sl++;
2009                         if (sl == conf->copies)
2010                                 sl = 0;
2011                 } while (!success && sl != r10_bio->read_slot);
2012                 rcu_read_unlock();
2013
2014                 if (!success) {
2015                         /* Cannot read from anywhere, just mark the block
2016                          * as bad on the first device to discourage future
2017                          * reads.
2018                          */
2019                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2020                         rdev = conf->mirrors[dn].rdev;
2021
2022                         if (!rdev_set_badblocks(
2023                                     rdev,
2024                                     r10_bio->devs[r10_bio->read_slot].addr
2025                                     + sect,
2026                                     s, 0))
2027                                 md_error(mddev, rdev);
2028                         break;
2029                 }
2030
2031                 start = sl;
2032                 /* write it back and re-read */
2033                 rcu_read_lock();
2034                 while (sl != r10_bio->read_slot) {
2035                         char b[BDEVNAME_SIZE];
2036
2037                         if (sl==0)
2038                                 sl = conf->copies;
2039                         sl--;
2040                         d = r10_bio->devs[sl].devnum;
2041                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2042                         if (!rdev ||
2043                             !test_bit(In_sync, &rdev->flags))
2044                                 continue;
2045
2046                         atomic_inc(&rdev->nr_pending);
2047                         rcu_read_unlock();
2048                         if (r10_sync_page_io(rdev,
2049                                              r10_bio->devs[sl].addr +
2050                                              sect,
2051                                              s<<9, conf->tmppage, WRITE)
2052                             == 0) {
2053                                 /* Well, this device is dead */
2054                                 printk(KERN_NOTICE
2055                                        "md/raid10:%s: read correction "
2056                                        "write failed"
2057                                        " (%d sectors at %llu on %s)\n",
2058                                        mdname(mddev), s,
2059                                        (unsigned long long)(
2060                                                sect + rdev->data_offset),
2061                                        bdevname(rdev->bdev, b));
2062                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2063                                        "drive\n",
2064                                        mdname(mddev),
2065                                        bdevname(rdev->bdev, b));
2066                         }
2067                         rdev_dec_pending(rdev, mddev);
2068                         rcu_read_lock();
2069                 }
2070                 sl = start;
2071                 while (sl != r10_bio->read_slot) {
2072                         char b[BDEVNAME_SIZE];
2073
2074                         if (sl==0)
2075                                 sl = conf->copies;
2076                         sl--;
2077                         d = r10_bio->devs[sl].devnum;
2078                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2079                         if (!rdev ||
2080                             !test_bit(In_sync, &rdev->flags))
2081                                 continue;
2082
2083                         atomic_inc(&rdev->nr_pending);
2084                         rcu_read_unlock();
2085                         switch (r10_sync_page_io(rdev,
2086                                              r10_bio->devs[sl].addr +
2087                                              sect,
2088                                              s<<9, conf->tmppage,
2089                                                  READ)) {
2090                         case 0:
2091                                 /* Well, this device is dead */
2092                                 printk(KERN_NOTICE
2093                                        "md/raid10:%s: unable to read back "
2094                                        "corrected sectors"
2095                                        " (%d sectors at %llu on %s)\n",
2096                                        mdname(mddev), s,
2097                                        (unsigned long long)(
2098                                                sect + rdev->data_offset),
2099                                        bdevname(rdev->bdev, b));
2100                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2101                                        "drive\n",
2102                                        mdname(mddev),
2103                                        bdevname(rdev->bdev, b));
2104                                 break;
2105                         case 1:
2106                                 printk(KERN_INFO
2107                                        "md/raid10:%s: read error corrected"
2108                                        " (%d sectors at %llu on %s)\n",
2109                                        mdname(mddev), s,
2110                                        (unsigned long long)(
2111                                                sect + rdev->data_offset),
2112                                        bdevname(rdev->bdev, b));
2113                                 atomic_add(s, &rdev->corrected_errors);
2114                         }
2115
2116                         rdev_dec_pending(rdev, mddev);
2117                         rcu_read_lock();
2118                 }
2119                 rcu_read_unlock();
2120
2121                 sectors -= s;
2122                 sect += s;
2123         }
2124 }
2125
2126 static void bi_complete(struct bio *bio, int error)
2127 {
2128         complete((struct completion *)bio->bi_private);
2129 }
2130
2131 static int submit_bio_wait(int rw, struct bio *bio)
2132 {
2133         struct completion event;
2134         rw |= REQ_SYNC;
2135
2136         init_completion(&event);
2137         bio->bi_private = &event;
2138         bio->bi_end_io = bi_complete;
2139         submit_bio(rw, bio);
2140         wait_for_completion(&event);
2141
2142         return test_bit(BIO_UPTODATE, &bio->bi_flags);
2143 }
2144
2145 static int narrow_write_error(struct r10bio *r10_bio, int i)
2146 {
2147         struct bio *bio = r10_bio->master_bio;
2148         struct mddev *mddev = r10_bio->mddev;
2149         struct r10conf *conf = mddev->private;
2150         struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2151         /* bio has the data to be written to slot 'i' where
2152          * we just recently had a write error.
2153          * We repeatedly clone the bio and trim down to one block,
2154          * then try the write.  Where the write fails we record
2155          * a bad block.
2156          * It is conceivable that the bio doesn't exactly align with
2157          * blocks.  We must handle this.
2158          *
2159          * We currently own a reference to the rdev.
2160          */
2161
2162         int block_sectors;
2163         sector_t sector;
2164         int sectors;
2165         int sect_to_write = r10_bio->sectors;
2166         int ok = 1;
2167
2168         if (rdev->badblocks.shift < 0)
2169                 return 0;
2170
2171         block_sectors = 1 << rdev->badblocks.shift;
2172         sector = r10_bio->sector;
2173         sectors = ((r10_bio->sector + block_sectors)
2174                    & ~(sector_t)(block_sectors - 1))
2175                 - sector;
2176
2177         while (sect_to_write) {
2178                 struct bio *wbio;
2179                 if (sectors > sect_to_write)
2180                         sectors = sect_to_write;
2181                 /* Write at 'sector' for 'sectors' */
2182                 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2183                 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2184                 wbio->bi_sector = (r10_bio->devs[i].addr+
2185                                    rdev->data_offset+
2186                                    (sector - r10_bio->sector));
2187                 wbio->bi_bdev = rdev->bdev;
2188                 if (submit_bio_wait(WRITE, wbio) == 0)
2189                         /* Failure! */
2190                         ok = rdev_set_badblocks(rdev, sector,
2191                                                 sectors, 0)
2192                                 && ok;
2193
2194                 bio_put(wbio);
2195                 sect_to_write -= sectors;
2196                 sector += sectors;
2197                 sectors = block_sectors;
2198         }
2199         return ok;
2200 }
2201
2202 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2203 {
2204         int slot = r10_bio->read_slot;
2205         struct bio *bio;
2206         struct r10conf *conf = mddev->private;
2207         struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2208         char b[BDEVNAME_SIZE];
2209         unsigned long do_sync;
2210         int max_sectors;
2211
2212         /* we got a read error. Maybe the drive is bad.  Maybe just
2213          * the block and we can fix it.
2214          * We freeze all other IO, and try reading the block from
2215          * other devices.  When we find one, we re-write
2216          * and check it that fixes the read error.
2217          * This is all done synchronously while the array is
2218          * frozen.
2219          */
2220         if (mddev->ro == 0) {
2221                 freeze_array(conf);
2222                 fix_read_error(conf, mddev, r10_bio);
2223                 unfreeze_array(conf);
2224         }
2225         rdev_dec_pending(rdev, mddev);
2226
2227         bio = r10_bio->devs[slot].bio;
2228         bdevname(bio->bi_bdev, b);
2229         r10_bio->devs[slot].bio =
2230                 mddev->ro ? IO_BLOCKED : NULL;
2231 read_more:
2232         rdev = read_balance(conf, r10_bio, &max_sectors);
2233         if (rdev == NULL) {
2234                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2235                        " read error for block %llu\n",
2236                        mdname(mddev), b,
2237                        (unsigned long long)r10_bio->sector);
2238                 raid_end_bio_io(r10_bio);
2239                 bio_put(bio);
2240                 return;
2241         }
2242
2243         do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2244         if (bio)
2245                 bio_put(bio);
2246         slot = r10_bio->read_slot;
2247         printk_ratelimited(
2248                 KERN_ERR
2249                 "md/raid10:%s: %s: redirecting"
2250                 "sector %llu to another mirror\n",
2251                 mdname(mddev),
2252                 bdevname(rdev->bdev, b),
2253                 (unsigned long long)r10_bio->sector);
2254         bio = bio_clone_mddev(r10_bio->master_bio,
2255                               GFP_NOIO, mddev);
2256         md_trim_bio(bio,
2257                     r10_bio->sector - bio->bi_sector,
2258                     max_sectors);
2259         r10_bio->devs[slot].bio = bio;
2260         r10_bio->devs[slot].rdev = rdev;
2261         bio->bi_sector = r10_bio->devs[slot].addr
2262                 + rdev->data_offset;
2263         bio->bi_bdev = rdev->bdev;
2264         bio->bi_rw = READ | do_sync;
2265         bio->bi_private = r10_bio;
2266         bio->bi_end_io = raid10_end_read_request;
2267         if (max_sectors < r10_bio->sectors) {
2268                 /* Drat - have to split this up more */
2269                 struct bio *mbio = r10_bio->master_bio;
2270                 int sectors_handled =
2271                         r10_bio->sector + max_sectors
2272                         - mbio->bi_sector;
2273                 r10_bio->sectors = max_sectors;
2274                 spin_lock_irq(&conf->device_lock);
2275                 if (mbio->bi_phys_segments == 0)
2276                         mbio->bi_phys_segments = 2;
2277                 else
2278                         mbio->bi_phys_segments++;
2279                 spin_unlock_irq(&conf->device_lock);
2280                 generic_make_request(bio);
2281                 bio = NULL;
2282
2283                 r10_bio = mempool_alloc(conf->r10bio_pool,
2284                                         GFP_NOIO);
2285                 r10_bio->master_bio = mbio;
2286                 r10_bio->sectors = (mbio->bi_size >> 9)
2287                         - sectors_handled;
2288                 r10_bio->state = 0;
2289                 set_bit(R10BIO_ReadError,
2290                         &r10_bio->state);
2291                 r10_bio->mddev = mddev;
2292                 r10_bio->sector = mbio->bi_sector
2293                         + sectors_handled;
2294
2295                 goto read_more;
2296         } else
2297                 generic_make_request(bio);
2298 }
2299
2300 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2301 {
2302         /* Some sort of write request has finished and it
2303          * succeeded in writing where we thought there was a
2304          * bad block.  So forget the bad block.
2305          * Or possibly if failed and we need to record
2306          * a bad block.
2307          */
2308         int m;
2309         struct md_rdev *rdev;
2310
2311         if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2312             test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2313                 for (m = 0; m < conf->copies; m++) {
2314                         int dev = r10_bio->devs[m].devnum;
2315                         rdev = conf->mirrors[dev].rdev;
2316                         if (r10_bio->devs[m].bio == NULL)
2317                                 continue;
2318                         if (test_bit(BIO_UPTODATE,
2319                                      &r10_bio->devs[m].bio->bi_flags)) {
2320                                 rdev_clear_badblocks(
2321                                         rdev,
2322                                         r10_bio->devs[m].addr,
2323                                         r10_bio->sectors);
2324                         } else {
2325                                 if (!rdev_set_badblocks(
2326                                             rdev,
2327                                             r10_bio->devs[m].addr,
2328                                             r10_bio->sectors, 0))
2329                                         md_error(conf->mddev, rdev);
2330                         }
2331                         rdev = conf->mirrors[dev].replacement;
2332                         if (r10_bio->devs[m].repl_bio == NULL)
2333                                 continue;
2334                         if (test_bit(BIO_UPTODATE,
2335                                      &r10_bio->devs[m].repl_bio->bi_flags)) {
2336                                 rdev_clear_badblocks(
2337                                         rdev,
2338                                         r10_bio->devs[m].addr,
2339                                         r10_bio->sectors);
2340                         } else {
2341                                 if (!rdev_set_badblocks(
2342                                             rdev,
2343                                             r10_bio->devs[m].addr,
2344                                             r10_bio->sectors, 0))
2345                                         md_error(conf->mddev, rdev);
2346                         }
2347                 }
2348                 put_buf(r10_bio);
2349         } else {
2350                 for (m = 0; m < conf->copies; m++) {
2351                         int dev = r10_bio->devs[m].devnum;
2352                         struct bio *bio = r10_bio->devs[m].bio;
2353                         rdev = conf->mirrors[dev].rdev;
2354                         if (bio == IO_MADE_GOOD) {
2355                                 rdev_clear_badblocks(
2356                                         rdev,
2357                                         r10_bio->devs[m].addr,
2358                                         r10_bio->sectors);
2359                                 rdev_dec_pending(rdev, conf->mddev);
2360                         } else if (bio != NULL &&
2361                                    !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2362                                 if (!narrow_write_error(r10_bio, m)) {
2363                                         md_error(conf->mddev, rdev);
2364                                         set_bit(R10BIO_Degraded,
2365                                                 &r10_bio->state);
2366                                 }
2367                                 rdev_dec_pending(rdev, conf->mddev);
2368                         }
2369                         bio = r10_bio->devs[m].repl_bio;
2370                         rdev = conf->mirrors[dev].replacement;
2371                         if (bio == IO_MADE_GOOD) {
2372                                 rdev_clear_badblocks(
2373                                         rdev,
2374                                         r10_bio->devs[m].addr,
2375                                         r10_bio->sectors);
2376                                 rdev_dec_pending(rdev, conf->mddev);
2377                         }
2378                 }
2379                 if (test_bit(R10BIO_WriteError,
2380                              &r10_bio->state))
2381                         close_write(r10_bio);
2382                 raid_end_bio_io(r10_bio);
2383         }
2384 }
2385
2386 static void raid10d(struct mddev *mddev)
2387 {
2388         struct r10bio *r10_bio;
2389         unsigned long flags;
2390         struct r10conf *conf = mddev->private;
2391         struct list_head *head = &conf->retry_list;
2392         struct blk_plug plug;
2393
2394         md_check_recovery(mddev);
2395
2396         blk_start_plug(&plug);
2397         for (;;) {
2398
2399                 flush_pending_writes(conf);
2400
2401                 spin_lock_irqsave(&conf->device_lock, flags);
2402                 if (list_empty(head)) {
2403                         spin_unlock_irqrestore(&conf->device_lock, flags);
2404                         break;
2405                 }
2406                 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2407                 list_del(head->prev);
2408                 conf->nr_queued--;
2409                 spin_unlock_irqrestore(&conf->device_lock, flags);
2410
2411                 mddev = r10_bio->mddev;
2412                 conf = mddev->private;
2413                 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2414                     test_bit(R10BIO_WriteError, &r10_bio->state))
2415                         handle_write_completed(conf, r10_bio);
2416                 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2417                         sync_request_write(mddev, r10_bio);
2418                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2419                         recovery_request_write(mddev, r10_bio);
2420                 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2421                         handle_read_error(mddev, r10_bio);
2422                 else {
2423                         /* just a partial read to be scheduled from a
2424                          * separate context
2425                          */
2426                         int slot = r10_bio->read_slot;
2427                         generic_make_request(r10_bio->devs[slot].bio);
2428                 }
2429
2430                 cond_resched();
2431                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2432                         md_check_recovery(mddev);
2433         }
2434         blk_finish_plug(&plug);
2435 }
2436
2437
2438 static int init_resync(struct r10conf *conf)
2439 {
2440         int buffs;
2441         int i;
2442
2443         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2444         BUG_ON(conf->r10buf_pool);
2445         conf->have_replacement = 0;
2446         for (i = 0; i < conf->raid_disks; i++)
2447                 if (conf->mirrors[i].replacement)
2448                         conf->have_replacement = 1;
2449         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2450         if (!conf->r10buf_pool)
2451                 return -ENOMEM;
2452         conf->next_resync = 0;
2453         return 0;
2454 }
2455
2456 /*
2457  * perform a "sync" on one "block"
2458  *
2459  * We need to make sure that no normal I/O request - particularly write
2460  * requests - conflict with active sync requests.
2461  *
2462  * This is achieved by tracking pending requests and a 'barrier' concept
2463  * that can be installed to exclude normal IO requests.
2464  *
2465  * Resync and recovery are handled very differently.
2466  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2467  *
2468  * For resync, we iterate over virtual addresses, read all copies,
2469  * and update if there are differences.  If only one copy is live,
2470  * skip it.
2471  * For recovery, we iterate over physical addresses, read a good
2472  * value for each non-in_sync drive, and over-write.
2473  *
2474  * So, for recovery we may have several outstanding complex requests for a
2475  * given address, one for each out-of-sync device.  We model this by allocating
2476  * a number of r10_bio structures, one for each out-of-sync device.
2477  * As we setup these structures, we collect all bio's together into a list
2478  * which we then process collectively to add pages, and then process again
2479  * to pass to generic_make_request.
2480  *
2481  * The r10_bio structures are linked using a borrowed master_bio pointer.
2482  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2483  * has its remaining count decremented to 0, the whole complex operation
2484  * is complete.
2485  *
2486  */
2487
2488 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2489                              int *skipped, int go_faster)
2490 {
2491         struct r10conf *conf = mddev->private;
2492         struct r10bio *r10_bio;
2493         struct bio *biolist = NULL, *bio;
2494         sector_t max_sector, nr_sectors;
2495         int i;
2496         int max_sync;
2497         sector_t sync_blocks;
2498         sector_t sectors_skipped = 0;
2499         int chunks_skipped = 0;
2500
2501         if (!conf->r10buf_pool)
2502                 if (init_resync(conf))
2503                         return 0;
2504
2505  skipped:
2506         max_sector = mddev->dev_sectors;
2507         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2508                 max_sector = mddev->resync_max_sectors;
2509         if (sector_nr >= max_sector) {
2510                 /* If we aborted, we need to abort the
2511                  * sync on the 'current' bitmap chucks (there can
2512                  * be several when recovering multiple devices).
2513                  * as we may have started syncing it but not finished.
2514                  * We can find the current address in
2515                  * mddev->curr_resync, but for recovery,
2516                  * we need to convert that to several
2517                  * virtual addresses.
2518                  */
2519                 if (mddev->curr_resync < max_sector) { /* aborted */
2520                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2521                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2522                                                 &sync_blocks, 1);
2523                         else for (i=0; i<conf->raid_disks; i++) {
2524                                 sector_t sect =
2525                                         raid10_find_virt(conf, mddev->curr_resync, i);
2526                                 bitmap_end_sync(mddev->bitmap, sect,
2527                                                 &sync_blocks, 1);
2528                         }
2529                 } else {
2530                         /* completed sync */
2531                         if ((!mddev->bitmap || conf->fullsync)
2532                             && conf->have_replacement
2533                             && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2534                                 /* Completed a full sync so the replacements
2535                                  * are now fully recovered.
2536                                  */
2537                                 for (i = 0; i < conf->raid_disks; i++)
2538                                         if (conf->mirrors[i].replacement)
2539                                                 conf->mirrors[i].replacement
2540                                                         ->recovery_offset
2541                                                         = MaxSector;
2542                         }
2543                         conf->fullsync = 0;
2544                 }
2545                 bitmap_close_sync(mddev->bitmap);
2546                 close_sync(conf);
2547                 *skipped = 1;
2548                 return sectors_skipped;
2549         }
2550         if (chunks_skipped >= conf->raid_disks) {
2551                 /* if there has been nothing to do on any drive,
2552                  * then there is nothing to do at all..
2553                  */
2554                 *skipped = 1;
2555                 return (max_sector - sector_nr) + sectors_skipped;
2556         }
2557
2558         if (max_sector > mddev->resync_max)
2559                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2560
2561         /* make sure whole request will fit in a chunk - if chunks
2562          * are meaningful
2563          */
2564         if (conf->near_copies < conf->raid_disks &&
2565             max_sector > (sector_nr | conf->chunk_mask))
2566                 max_sector = (sector_nr | conf->chunk_mask) + 1;
2567         /*
2568          * If there is non-resync activity waiting for us then
2569          * put in a delay to throttle resync.
2570          */
2571         if (!go_faster && conf->nr_waiting)
2572                 msleep_interruptible(1000);
2573
2574         /* Again, very different code for resync and recovery.
2575          * Both must result in an r10bio with a list of bios that
2576          * have bi_end_io, bi_sector, bi_bdev set,
2577          * and bi_private set to the r10bio.
2578          * For recovery, we may actually create several r10bios
2579          * with 2 bios in each, that correspond to the bios in the main one.
2580          * In this case, the subordinate r10bios link back through a
2581          * borrowed master_bio pointer, and the counter in the master
2582          * includes a ref from each subordinate.
2583          */
2584         /* First, we decide what to do and set ->bi_end_io
2585          * To end_sync_read if we want to read, and
2586          * end_sync_write if we will want to write.
2587          */
2588
2589         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2590         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2591                 /* recovery... the complicated one */
2592                 int j;
2593                 r10_bio = NULL;
2594
2595                 for (i=0 ; i<conf->raid_disks; i++) {
2596                         int still_degraded;
2597                         struct r10bio *rb2;
2598                         sector_t sect;
2599                         int must_sync;
2600                         int any_working;
2601                         struct mirror_info *mirror = &conf->mirrors[i];
2602
2603                         if ((mirror->rdev == NULL ||
2604                              test_bit(In_sync, &mirror->rdev->flags))
2605                             &&
2606                             (mirror->replacement == NULL ||
2607                              test_bit(Faulty,
2608                                       &mirror->replacement->flags)))
2609                                 continue;
2610
2611                         still_degraded = 0;
2612                         /* want to reconstruct this device */
2613                         rb2 = r10_bio;
2614                         sect = raid10_find_virt(conf, sector_nr, i);
2615                         /* Unless we are doing a full sync, or a replacement
2616                          * we only need to recover the block if it is set in
2617                          * the bitmap
2618                          */
2619                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2620                                                       &sync_blocks, 1);
2621                         if (sync_blocks < max_sync)
2622                                 max_sync = sync_blocks;
2623                         if (!must_sync &&
2624                             mirror->replacement == NULL &&
2625                             !conf->fullsync) {
2626                                 /* yep, skip the sync_blocks here, but don't assume
2627                                  * that there will never be anything to do here
2628                                  */
2629                                 chunks_skipped = -1;
2630                                 continue;
2631                         }
2632
2633                         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2634                         raise_barrier(conf, rb2 != NULL);
2635                         atomic_set(&r10_bio->remaining, 0);
2636
2637                         r10_bio->master_bio = (struct bio*)rb2;
2638                         if (rb2)
2639                                 atomic_inc(&rb2->remaining);
2640                         r10_bio->mddev = mddev;
2641                         set_bit(R10BIO_IsRecover, &r10_bio->state);
2642                         r10_bio->sector = sect;
2643
2644                         raid10_find_phys(conf, r10_bio);
2645
2646                         /* Need to check if the array will still be
2647                          * degraded
2648                          */
2649                         for (j=0; j<conf->raid_disks; j++)
2650                                 if (conf->mirrors[j].rdev == NULL ||
2651                                     test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2652                                         still_degraded = 1;
2653                                         break;
2654                                 }
2655
2656                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2657                                                       &sync_blocks, still_degraded);
2658
2659                         any_working = 0;
2660                         for (j=0; j<conf->copies;j++) {
2661                                 int k;
2662                                 int d = r10_bio->devs[j].devnum;
2663                                 sector_t from_addr, to_addr;
2664                                 struct md_rdev *rdev;
2665                                 sector_t sector, first_bad;
2666                                 int bad_sectors;
2667                                 if (!conf->mirrors[d].rdev ||
2668                                     !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2669                                         continue;
2670                                 /* This is where we read from */
2671                                 any_working = 1;
2672                                 rdev = conf->mirrors[d].rdev;
2673                                 sector = r10_bio->devs[j].addr;
2674
2675                                 if (is_badblock(rdev, sector, max_sync,
2676                                                 &first_bad, &bad_sectors)) {
2677                                         if (first_bad > sector)
2678                                                 max_sync = first_bad - sector;
2679                                         else {
2680                                                 bad_sectors -= (sector
2681                                                                 - first_bad);
2682                                                 if (max_sync > bad_sectors)
2683                                                         max_sync = bad_sectors;
2684                                                 continue;
2685                                         }
2686                                 }
2687                                 bio = r10_bio->devs[0].bio;
2688                                 bio->bi_next = biolist;
2689                                 biolist = bio;
2690                                 bio->bi_private = r10_bio;
2691                                 bio->bi_end_io = end_sync_read;
2692                                 bio->bi_rw = READ;
2693                                 from_addr = r10_bio->devs[j].addr;
2694                                 bio->bi_sector = from_addr + rdev->data_offset;
2695                                 bio->bi_bdev = rdev->bdev;
2696                                 atomic_inc(&rdev->nr_pending);
2697                                 /* and we write to 'i' (if not in_sync) */
2698
2699                                 for (k=0; k<conf->copies; k++)
2700                                         if (r10_bio->devs[k].devnum == i)
2701                                                 break;
2702                                 BUG_ON(k == conf->copies);
2703                                 to_addr = r10_bio->devs[k].addr;
2704                                 r10_bio->devs[0].devnum = d;
2705                                 r10_bio->devs[0].addr = from_addr;
2706                                 r10_bio->devs[1].devnum = i;
2707                                 r10_bio->devs[1].addr = to_addr;
2708
2709                                 rdev = mirror->rdev;
2710                                 if (!test_bit(In_sync, &rdev->flags)) {
2711                                         bio = r10_bio->devs[1].bio;
2712                                         bio->bi_next = biolist;
2713                                         biolist = bio;
2714                                         bio->bi_private = r10_bio;
2715                                         bio->bi_end_io = end_sync_write;
2716                                         bio->bi_rw = WRITE;
2717                                         bio->bi_sector = to_addr
2718                                                 + rdev->data_offset;
2719                                         bio->bi_bdev = rdev->bdev;
2720                                         atomic_inc(&r10_bio->remaining);
2721                                 } else
2722                                         r10_bio->devs[1].bio->bi_end_io = NULL;
2723
2724                                 /* and maybe write to replacement */
2725                                 bio = r10_bio->devs[1].repl_bio;
2726                                 if (bio)
2727                                         bio->bi_end_io = NULL;
2728                                 rdev = mirror->replacement;
2729                                 /* Note: if rdev != NULL, then bio
2730                                  * cannot be NULL as r10buf_pool_alloc will
2731                                  * have allocated it.
2732                                  * So the second test here is pointless.
2733                                  * But it keeps semantic-checkers happy, and
2734                                  * this comment keeps human reviewers
2735                                  * happy.
2736                                  */
2737                                 if (rdev == NULL || bio == NULL ||
2738                                     test_bit(Faulty, &rdev->flags))
2739                                         break;
2740                                 bio->bi_next = biolist;
2741                                 biolist = bio;
2742                                 bio->bi_private = r10_bio;
2743                                 bio->bi_end_io = end_sync_write;
2744                                 bio->bi_rw = WRITE;
2745                                 bio->bi_sector = to_addr + rdev->data_offset;
2746                                 bio->bi_bdev = rdev->bdev;
2747                                 atomic_inc(&r10_bio->remaining);
2748                                 break;
2749                         }
2750                         if (j == conf->copies) {
2751                                 /* Cannot recover, so abort the recovery or
2752                                  * record a bad block */
2753                                 put_buf(r10_bio);
2754                                 if (rb2)
2755                                         atomic_dec(&rb2->remaining);
2756                                 r10_bio = rb2;
2757                                 if (any_working) {
2758                                         /* problem is that there are bad blocks
2759                                          * on other device(s)
2760                                          */
2761                                         int k;
2762                                         for (k = 0; k < conf->copies; k++)
2763                                                 if (r10_bio->devs[k].devnum == i)
2764                                                         break;
2765                                         if (!test_bit(In_sync,
2766                                                       &mirror->rdev->flags)
2767                                             && !rdev_set_badblocks(
2768                                                     mirror->rdev,
2769                                                     r10_bio->devs[k].addr,
2770                                                     max_sync, 0))
2771                                                 any_working = 0;
2772                                         if (mirror->replacement &&
2773                                             !rdev_set_badblocks(
2774                                                     mirror->replacement,
2775                                                     r10_bio->devs[k].addr,
2776                                                     max_sync, 0))
2777                                                 any_working = 0;
2778                                 }
2779                                 if (!any_working)  {
2780                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
2781                                                               &mddev->recovery))
2782                                                 printk(KERN_INFO "md/raid10:%s: insufficient "
2783                                                        "working devices for recovery.\n",
2784                                                        mdname(mddev));
2785                                         mirror->recovery_disabled
2786                                                 = mddev->recovery_disabled;
2787                                 }
2788                                 break;
2789                         }
2790                 }
2791                 if (biolist == NULL) {
2792                         while (r10_bio) {
2793                                 struct r10bio *rb2 = r10_bio;
2794                                 r10_bio = (struct r10bio*) rb2->master_bio;
2795                                 rb2->master_bio = NULL;
2796                                 put_buf(rb2);
2797                         }
2798                         goto giveup;
2799                 }
2800         } else {
2801                 /* resync. Schedule a read for every block at this virt offset */
2802                 int count = 0;
2803
2804                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2805
2806                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2807                                        &sync_blocks, mddev->degraded) &&
2808                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
2809                                                  &mddev->recovery)) {
2810                         /* We can skip this block */
2811                         *skipped = 1;
2812                         return sync_blocks + sectors_skipped;
2813                 }
2814                 if (sync_blocks < max_sync)
2815                         max_sync = sync_blocks;
2816                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2817
2818                 r10_bio->mddev = mddev;
2819                 atomic_set(&r10_bio->remaining, 0);
2820                 raise_barrier(conf, 0);
2821                 conf->next_resync = sector_nr;
2822
2823                 r10_bio->master_bio = NULL;
2824                 r10_bio->sector = sector_nr;
2825                 set_bit(R10BIO_IsSync, &r10_bio->state);
2826                 raid10_find_phys(conf, r10_bio);
2827                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2828
2829                 for (i=0; i<conf->copies; i++) {
2830                         int d = r10_bio->devs[i].devnum;
2831                         sector_t first_bad, sector;
2832                         int bad_sectors;
2833
2834                         if (r10_bio->devs[i].repl_bio)
2835                                 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
2836
2837                         bio = r10_bio->devs[i].bio;
2838                         bio->bi_end_io = NULL;
2839                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
2840                         if (conf->mirrors[d].rdev == NULL ||
2841                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2842                                 continue;
2843                         sector = r10_bio->devs[i].addr;
2844                         if (is_badblock(conf->mirrors[d].rdev,
2845                                         sector, max_sync,
2846                                         &first_bad, &bad_sectors)) {
2847                                 if (first_bad > sector)
2848                                         max_sync = first_bad - sector;
2849                                 else {
2850                                         bad_sectors -= (sector - first_bad);
2851                                         if (max_sync > bad_sectors)
2852                                                 max_sync = max_sync;
2853                                         continue;
2854                                 }
2855                         }
2856                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2857                         atomic_inc(&r10_bio->remaining);
2858                         bio->bi_next = biolist;
2859                         biolist = bio;
2860                         bio->bi_private = r10_bio;
2861                         bio->bi_end_io = end_sync_read;
2862                         bio->bi_rw = READ;
2863                         bio->bi_sector = sector +
2864                                 conf->mirrors[d].rdev->data_offset;
2865                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2866                         count++;
2867
2868                         if (conf->mirrors[d].replacement == NULL ||
2869                             test_bit(Faulty,
2870                                      &conf->mirrors[d].replacement->flags))
2871                                 continue;
2872
2873                         /* Need to set up for writing to the replacement */
2874                         bio = r10_bio->devs[i].repl_bio;
2875                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
2876
2877                         sector = r10_bio->devs[i].addr;
2878                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2879                         bio->bi_next = biolist;
2880                         biolist = bio;
2881                         bio->bi_private = r10_bio;
2882                         bio->bi_end_io = end_sync_write;
2883                         bio->bi_rw = WRITE;
2884                         bio->bi_sector = sector +
2885                                 conf->mirrors[d].replacement->data_offset;
2886                         bio->bi_bdev = conf->mirrors[d].replacement->bdev;
2887                         count++;
2888                 }
2889
2890                 if (count < 2) {
2891                         for (i=0; i<conf->copies; i++) {
2892                                 int d = r10_bio->devs[i].devnum;
2893                                 if (r10_bio->devs[i].bio->bi_end_io)
2894                                         rdev_dec_pending(conf->mirrors[d].rdev,
2895                                                          mddev);
2896                                 if (r10_bio->devs[i].repl_bio &&
2897                                     r10_bio->devs[i].repl_bio->bi_end_io)
2898                                         rdev_dec_pending(
2899                                                 conf->mirrors[d].replacement,
2900                                                 mddev);
2901                         }
2902                         put_buf(r10_bio);
2903                         biolist = NULL;
2904                         goto giveup;
2905                 }
2906         }
2907
2908         for (bio = biolist; bio ; bio=bio->bi_next) {
2909
2910                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2911                 if (bio->bi_end_io)
2912                         bio->bi_flags |= 1 << BIO_UPTODATE;
2913                 bio->bi_vcnt = 0;
2914                 bio->bi_idx = 0;
2915                 bio->bi_phys_segments = 0;
2916                 bio->bi_size = 0;
2917         }
2918
2919         nr_sectors = 0;
2920         if (sector_nr + max_sync < max_sector)
2921                 max_sector = sector_nr + max_sync;
2922         do {
2923                 struct page *page;
2924                 int len = PAGE_SIZE;
2925                 if (sector_nr + (len>>9) > max_sector)
2926                         len = (max_sector - sector_nr) << 9;
2927                 if (len == 0)
2928                         break;
2929                 for (bio= biolist ; bio ; bio=bio->bi_next) {
2930                         struct bio *bio2;
2931                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2932                         if (bio_add_page(bio, page, len, 0))
2933                                 continue;
2934
2935                         /* stop here */
2936                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2937                         for (bio2 = biolist;
2938                              bio2 && bio2 != bio;
2939                              bio2 = bio2->bi_next) {
2940                                 /* remove last page from this bio */
2941                                 bio2->bi_vcnt--;
2942                                 bio2->bi_size -= len;
2943                                 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2944                         }
2945                         goto bio_full;
2946                 }
2947                 nr_sectors += len>>9;
2948                 sector_nr += len>>9;
2949         } while (biolist->bi_vcnt < RESYNC_PAGES);
2950  bio_full:
2951         r10_bio->sectors = nr_sectors;
2952
2953         while (biolist) {
2954                 bio = biolist;
2955                 biolist = biolist->bi_next;
2956
2957                 bio->bi_next = NULL;
2958                 r10_bio = bio->bi_private;
2959                 r10_bio->sectors = nr_sectors;
2960
2961                 if (bio->bi_end_io == end_sync_read) {
2962                         md_sync_acct(bio->bi_bdev, nr_sectors);
2963                         generic_make_request(bio);
2964                 }
2965         }
2966
2967         if (sectors_skipped)
2968                 /* pretend they weren't skipped, it makes
2969                  * no important difference in this case
2970                  */
2971                 md_done_sync(mddev, sectors_skipped, 1);
2972
2973         return sectors_skipped + nr_sectors;
2974  giveup:
2975         /* There is nowhere to write, so all non-sync
2976          * drives must be failed or in resync, all drives
2977          * have a bad block, so try the next chunk...
2978          */
2979         if (sector_nr + max_sync < max_sector)
2980                 max_sector = sector_nr + max_sync;
2981
2982         sectors_skipped += (max_sector - sector_nr);
2983         chunks_skipped ++;
2984         sector_nr = max_sector;
2985         goto skipped;
2986 }
2987
2988 static sector_t
2989 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2990 {
2991         sector_t size;
2992         struct r10conf *conf = mddev->private;
2993
2994         if (!raid_disks)
2995                 raid_disks = conf->raid_disks;
2996         if (!sectors)
2997                 sectors = conf->dev_sectors;
2998
2999         size = sectors >> conf->chunk_shift;
3000         sector_div(size, conf->far_copies);
3001         size = size * raid_disks;
3002         sector_div(size, conf->near_copies);
3003
3004         return size << conf->chunk_shift;
3005 }
3006
3007
3008 static struct r10conf *setup_conf(struct mddev *mddev)
3009 {
3010         struct r10conf *conf = NULL;
3011         int nc, fc, fo;
3012         sector_t stride, size;
3013         int err = -EINVAL;
3014
3015         if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
3016             !is_power_of_2(mddev->new_chunk_sectors)) {
3017                 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3018                        "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3019                        mdname(mddev), PAGE_SIZE);
3020                 goto out;
3021         }
3022
3023         nc = mddev->new_layout & 255;
3024         fc = (mddev->new_layout >> 8) & 255;
3025         fo = mddev->new_layout & (1<<16);
3026
3027         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
3028             (mddev->new_layout >> 17)) {
3029                 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3030                        mdname(mddev), mddev->new_layout);
3031                 goto out;
3032         }
3033
3034         err = -ENOMEM;
3035         conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3036         if (!conf)
3037                 goto out;
3038
3039         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
3040                                 GFP_KERNEL);
3041         if (!conf->mirrors)
3042                 goto out;
3043
3044         conf->tmppage = alloc_page(GFP_KERNEL);
3045         if (!conf->tmppage)
3046                 goto out;
3047
3048
3049         conf->raid_disks = mddev->raid_disks;
3050         conf->near_copies = nc;
3051         conf->far_copies = fc;
3052         conf->copies = nc*fc;
3053         conf->far_offset = fo;
3054         conf->chunk_mask = mddev->new_chunk_sectors - 1;
3055         conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
3056
3057         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3058                                            r10bio_pool_free, conf);
3059         if (!conf->r10bio_pool)
3060                 goto out;
3061
3062         size = mddev->dev_sectors >> conf->chunk_shift;
3063         sector_div(size, fc);
3064         size = size * conf->raid_disks;
3065         sector_div(size, nc);
3066         /* 'size' is now the number of chunks in the array */
3067         /* calculate "used chunks per device" in 'stride' */
3068         stride = size * conf->copies;
3069
3070         /* We need to round up when dividing by raid_disks to
3071          * get the stride size.
3072          */
3073         stride += conf->raid_disks - 1;
3074         sector_div(stride, conf->raid_disks);
3075
3076         conf->dev_sectors = stride << conf->chunk_shift;
3077
3078         if (fo)
3079                 stride = 1;
3080         else
3081                 sector_div(stride, fc);
3082         conf->stride = stride << conf->chunk_shift;
3083
3084
3085         spin_lock_init(&conf->device_lock);
3086         INIT_LIST_HEAD(&conf->retry_list);
3087
3088         spin_lock_init(&conf->resync_lock);
3089         init_waitqueue_head(&conf->wait_barrier);
3090
3091         conf->thread = md_register_thread(raid10d, mddev, NULL);
3092         if (!conf->thread)
3093                 goto out;
3094
3095         conf->mddev = mddev;
3096         return conf;
3097
3098  out:
3099         printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3100                mdname(mddev));
3101         if (conf) {
3102                 if (conf->r10bio_pool)
3103                         mempool_destroy(conf->r10bio_pool);
3104                 kfree(conf->mirrors);
3105                 safe_put_page(conf->tmppage);
3106                 kfree(conf);
3107         }
3108         return ERR_PTR(err);
3109 }
3110
3111 static int run(struct mddev *mddev)
3112 {
3113         struct r10conf *conf;
3114         int i, disk_idx, chunk_size;
3115         struct mirror_info *disk;
3116         struct md_rdev *rdev;
3117         sector_t size;
3118
3119         /*
3120          * copy the already verified devices into our private RAID10
3121          * bookkeeping area. [whatever we allocate in run(),
3122          * should be freed in stop()]
3123          */
3124
3125         if (mddev->private == NULL) {
3126                 conf = setup_conf(mddev);
3127                 if (IS_ERR(conf))
3128                         return PTR_ERR(conf);
3129                 mddev->private = conf;
3130         }
3131         conf = mddev->private;
3132         if (!conf)
3133                 goto out;
3134
3135         mddev->thread = conf->thread;
3136         conf->thread = NULL;
3137
3138         chunk_size = mddev->chunk_sectors << 9;
3139         blk_queue_io_min(mddev->queue, chunk_size);
3140         if (conf->raid_disks % conf->near_copies)
3141                 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
3142         else
3143                 blk_queue_io_opt(mddev->queue, chunk_size *
3144                                  (conf->raid_disks / conf->near_copies));
3145
3146         list_for_each_entry(rdev, &mddev->disks, same_set) {
3147
3148                 disk_idx = rdev->raid_disk;
3149                 if (disk_idx >= conf->raid_disks
3150                     || disk_idx < 0)
3151                         continue;
3152                 disk = conf->mirrors + disk_idx;
3153
3154                 disk->rdev = rdev;
3155                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3156                                   rdev->data_offset << 9);
3157                 /* as we don't honour merge_bvec_fn, we must never risk
3158                  * violating it, so limit max_segments to 1 lying
3159                  * within a single page.
3160                  */
3161                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
3162                         blk_queue_max_segments(mddev->queue, 1);
3163                         blk_queue_segment_boundary(mddev->queue,
3164                                                    PAGE_CACHE_SIZE - 1);
3165                 }
3166
3167                 disk->head_position = 0;
3168         }
3169         /* need to check that every block has at least one working mirror */
3170         if (!enough(conf, -1)) {
3171                 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3172                        mdname(mddev));
3173                 goto out_free_conf;
3174         }
3175
3176         mddev->degraded = 0;
3177         for (i = 0; i < conf->raid_disks; i++) {
3178
3179                 disk = conf->mirrors + i;
3180
3181                 if (!disk->rdev ||
3182                     !test_bit(In_sync, &disk->rdev->flags)) {
3183                         disk->head_position = 0;
3184                         mddev->degraded++;
3185                         if (disk->rdev)
3186                                 conf->fullsync = 1;
3187                 }
3188                 disk->recovery_disabled = mddev->recovery_disabled - 1;
3189         }
3190
3191         if (mddev->recovery_cp != MaxSector)
3192                 printk(KERN_NOTICE "md/raid10:%s: not clean"
3193                        " -- starting background reconstruction\n",
3194                        mdname(mddev));
3195         printk(KERN_INFO
3196                 "md/raid10:%s: active with %d out of %d devices\n",
3197                 mdname(mddev), conf->raid_disks - mddev->degraded,
3198                 conf->raid_disks);
3199         /*
3200          * Ok, everything is just fine now
3201          */
3202         mddev->dev_sectors = conf->dev_sectors;
3203         size = raid10_size(mddev, 0, 0);
3204         md_set_array_sectors(mddev, size);
3205         mddev->resync_max_sectors = size;
3206
3207         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3208         mddev->queue->backing_dev_info.congested_data = mddev;
3209
3210         /* Calculate max read-ahead size.
3211          * We need to readahead at least twice a whole stripe....
3212          * maybe...
3213          */
3214         {
3215                 int stripe = conf->raid_disks *
3216                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3217                 stripe /= conf->near_copies;
3218                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
3219                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
3220         }
3221
3222         if (conf->near_copies < conf->raid_disks)
3223                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3224
3225         if (md_integrity_register(mddev))
3226                 goto out_free_conf;
3227
3228         return 0;
3229
3230 out_free_conf:
3231         md_unregister_thread(&mddev->thread);
3232         if (conf->r10bio_pool)
3233                 mempool_destroy(conf->r10bio_pool);
3234         safe_put_page(conf->tmppage);
3235         kfree(conf->mirrors);
3236         kfree(conf);
3237         mddev->private = NULL;
3238 out:
3239         return -EIO;
3240 }
3241
3242 static int stop(struct mddev *mddev)
3243 {
3244         struct r10conf *conf = mddev->private;
3245
3246         raise_barrier(conf, 0);
3247         lower_barrier(conf);
3248
3249         md_unregister_thread(&mddev->thread);
3250         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
3251         if (conf->r10bio_pool)
3252                 mempool_destroy(conf->r10bio_pool);
3253         kfree(conf->mirrors);
3254         kfree(conf);
3255         mddev->private = NULL;
3256         return 0;
3257 }
3258
3259 static void raid10_quiesce(struct mddev *mddev, int state)
3260 {
3261         struct r10conf *conf = mddev->private;
3262
3263         switch(state) {
3264         case 1:
3265                 raise_barrier(conf, 0);
3266                 break;
3267         case 0:
3268                 lower_barrier(conf);
3269                 break;
3270         }
3271 }
3272
3273 static void *raid10_takeover_raid0(struct mddev *mddev)
3274 {
3275         struct md_rdev *rdev;
3276         struct r10conf *conf;
3277
3278         if (mddev->degraded > 0) {
3279                 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3280                        mdname(mddev));
3281                 return ERR_PTR(-EINVAL);
3282         }
3283
3284         /* Set new parameters */
3285         mddev->new_level = 10;
3286         /* new layout: far_copies = 1, near_copies = 2 */
3287         mddev->new_layout = (1<<8) + 2;
3288         mddev->new_chunk_sectors = mddev->chunk_sectors;
3289         mddev->delta_disks = mddev->raid_disks;
3290         mddev->raid_disks *= 2;
3291         /* make sure it will be not marked as dirty */
3292         mddev->recovery_cp = MaxSector;
3293
3294         conf = setup_conf(mddev);
3295         if (!IS_ERR(conf)) {
3296                 list_for_each_entry(rdev, &mddev->disks, same_set)
3297                         if (rdev->raid_disk >= 0)
3298                                 rdev->new_raid_disk = rdev->raid_disk * 2;
3299                 conf->barrier = 1;
3300         }
3301
3302         return conf;
3303 }
3304
3305 static void *raid10_takeover(struct mddev *mddev)
3306 {
3307         struct r0conf *raid0_conf;
3308
3309         /* raid10 can take over:
3310          *  raid0 - providing it has only two drives
3311          */
3312         if (mddev->level == 0) {
3313                 /* for raid0 takeover only one zone is supported */
3314                 raid0_conf = mddev->private;
3315                 if (raid0_conf->nr_strip_zones > 1) {
3316                         printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3317                                " with more than one zone.\n",
3318                                mdname(mddev));
3319                         return ERR_PTR(-EINVAL);
3320                 }
3321                 return raid10_takeover_raid0(mddev);
3322         }
3323         return ERR_PTR(-EINVAL);
3324 }
3325
3326 static struct md_personality raid10_personality =
3327 {
3328         .name           = "raid10",
3329         .level          = 10,
3330         .owner          = THIS_MODULE,
3331         .make_request   = make_request,
3332         .run            = run,
3333         .stop           = stop,
3334         .status         = status,
3335         .error_handler  = error,
3336         .hot_add_disk   = raid10_add_disk,
3337         .hot_remove_disk= raid10_remove_disk,
3338         .spare_active   = raid10_spare_active,
3339         .sync_request   = sync_request,
3340         .quiesce        = raid10_quiesce,
3341         .size           = raid10_size,
3342         .takeover       = raid10_takeover,
3343 };
3344
3345 static int __init raid_init(void)
3346 {
3347         return register_md_personality(&raid10_personality);
3348 }
3349
3350 static void raid_exit(void)
3351 {
3352         unregister_md_personality(&raid10_personality);
3353 }
3354
3355 module_init(raid_init);
3356 module_exit(raid_exit);
3357 MODULE_LICENSE("GPL");
3358 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3359 MODULE_ALIAS("md-personality-9"); /* RAID10 */
3360 MODULE_ALIAS("md-raid10");
3361 MODULE_ALIAS("md-level-10");
3362
3363 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);