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