drivers/md/raid10.c

   1 /*
   2  * raid10.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 2000-2004 Neil Brown
   5  *
   6  * RAID-10 support for md.
   7  *
   8  * Base on code in raid1.c.  See raid1.c for further copyright information.
   9  *
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2, or (at your option)
  14  * any later version.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * (for example /usr/src/linux/COPYING); if not, write to the Free
  18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 #include <linux/slab.h>
  22 #include <linux/delay.h>
  23 #include <linux/blkdev.h>
  24 #include <linux/module.h>
  25 #include <linux/seq_file.h>
  26 #include <linux/ratelimit.h>
  27 #include "md.h"
  28 #include "raid10.h"
  29 #include "raid0.h"
  30 #include "bitmap.h"
  31
  32 /*
  33  * RAID10 provides a combination of RAID0 and RAID1 functionality.
  34  * The layout of data is defined by
  35  *    chunk_size
  36  *    raid_disks
  37  *    near_copies (stored in low byte of layout)
  38  *    far_copies (stored in second byte of layout)
  39  *    far_offset (stored in bit 16 of layout )
  40  *
  41  * The data to be stored is divided into chunks using chunksize.
  42  * Each device is divided into far_copies sections.
  43  * In each section, chunks are laid out in a style similar to raid0, but
  44  * near_copies copies of each chunk is stored (each on a different drive).
  45  * The starting device for each section is offset near_copies from the starting
  46  * device of the previous section.
  47  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
  48  * drive.
  49  * near_copies and far_copies must be at least one, and their product is at most
  50  * raid_disks.
  51  *
  52  * If far_offset is true, then the far_copies are handled a bit differently.
  53  * The copies are still in different stripes, but instead of be very far apart
  54  * on disk, there are adjacent stripes.
  55  */
  56
  57 /*
  58  * Number of guaranteed r10bios in case of extreme VM load:
  59  */
  60 #define NR_RAID10_BIOS 256
  61
  62 /* When there are this many requests queue to be written by
  63  * the raid10 thread, we become 'congested' to provide back-pressure
  64  * for writeback.
  65  */
  66 static int max_queued_requests = 1024;
  67
  68 static void allow_barrier(struct r10conf *conf);
  69 static void lower_barrier(struct r10conf *conf);
  70
  71 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
  72 {
  73         struct r10conf *conf = data;
  74         int size = offsetof(struct r10bio, devs[conf->copies]);
  75
  76         /* allocate a r10bio with room for raid_disks entries in the
  77          * bios array */
  78         return kzalloc(size, gfp_flags);
  79 }
  80
  81 static void r10bio_pool_free(void *r10_bio, void *data)
  82 {
  83         kfree(r10_bio);
  84 }
  85
  86 /* Maximum size of each resync request */
  87 #define RESYNC_BLOCK_SIZE (64*1024)
  88 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  89 /* amount of memory to reserve for resync requests */
  90 #define RESYNC_WINDOW (1024*1024)
  91 /* maximum number of concurrent requests, memory permitting */
  92 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
  93
  94 /*
  95  * When performing a resync, we need to read and compare, so
  96  * we need as many pages are there are copies.
  97  * When performing a recovery, we need 2 bios, one for read,
  98  * one for write (we recover only one drive per r10buf)
  99  *
 100  */
 101 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
 102 {
 103         struct r10conf *conf = data;
 104         struct page *page;
 105         struct r10bio *r10_bio;
 106         struct bio *bio;
 107         int i, j;
 108         int nalloc;
 109
 110         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 111         if (!r10_bio)
 112                 return NULL;
 113
 114         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
 115                 nalloc = conf->copies; /* resync */
 116         else
 117                 nalloc = 2; /* recovery */
 118
 119         /*
 120          * Allocate bios.
 121          */
 122         for (j = nalloc ; j-- ; ) {
 123                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 124                 if (!bio)
 125                         goto out_free_bio;
 126                 r10_bio->devs[j].bio = bio;
 127                 if (!conf->have_replacement)
 128                         continue;
 129                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 130                 if (!bio)
 131                         goto out_free_bio;
 132                 r10_bio->devs[j].repl_bio = bio;
 133         }
 134         /*
 135          * Allocate RESYNC_PAGES data pages and attach them
 136          * where needed.
 137          */
 138         for (j = 0 ; j < nalloc; j++) {
 139                 struct bio *rbio = r10_bio->devs[j].repl_bio;
 140                 bio = r10_bio->devs[j].bio;
 141                 for (i = 0; i < RESYNC_PAGES; i++) {
 142                         if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
 143                                                 &conf->mddev->recovery)) {
 144                                 /* we can share bv_page's during recovery */
 145                                 struct bio *rbio = r10_bio->devs[0].bio;
 146                                 page = rbio->bi_io_vec[i].bv_page;
 147                                 get_page(page);
 148                         } else
 149                                 page = alloc_page(gfp_flags);
 150                         if (unlikely(!page))
 151                                 goto out_free_pages;
 152
 153                         bio->bi_io_vec[i].bv_page = page;
 154                         if (rbio)
 155                                 rbio->bi_io_vec[i].bv_page = page;
 156                 }
 157         }
 158
 159         return r10_bio;
 160
 161 out_free_pages:
 162         for ( ; i > 0 ; i--)
 163                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
 164         while (j--)
 165                 for (i = 0; i < RESYNC_PAGES ; i++)
 166                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 167         j = -1;
 168 out_free_bio:
 169         while (++j < nalloc) {
 170                 bio_put(r10_bio->devs[j].bio);
 171                 if (r10_bio->devs[j].repl_bio)
 172                         bio_put(r10_bio->devs[j].repl_bio);
 173         }
 174         r10bio_pool_free(r10_bio, conf);
 175         return NULL;
 176 }
 177
 178 static void r10buf_pool_free(void *__r10_bio, void *data)
 179 {
 180         int i;
 181         struct r10conf *conf = data;
 182         struct r10bio *r10bio = __r10_bio;
 183         int j;
 184
 185         for (j=0; j < conf->copies; j++) {
 186                 struct bio *bio = r10bio->devs[j].bio;
 187                 if (bio) {
 188                         for (i = 0; i < RESYNC_PAGES; i++) {
 189                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 190                                 bio->bi_io_vec[i].bv_page = NULL;
 191                         }
 192                         bio_put(bio);
 193                 }
 194                 bio = r10bio->devs[j].repl_bio;
 195                 if (bio)
 196                         bio_put(bio);
 197         }
 198         r10bio_pool_free(r10bio, conf);
 199 }
 200
 201 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
 202 {
 203         int i;
 204
 205         for (i = 0; i < conf->copies; i++) {
 206                 struct bio **bio = & r10_bio->devs[i].bio;
 207                 if (!BIO_SPECIAL(*bio))
 208                         bio_put(*bio);
 209                 *bio = NULL;
 210                 bio = &r10_bio->devs[i].repl_bio;
 211                 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
 212                         bio_put(*bio);
 213                 *bio = NULL;
 214         }
 215 }
 216
 217 static void free_r10bio(struct r10bio *r10_bio)
 218 {
 219         struct r10conf *conf = r10_bio->mddev->private;
 220
 221         put_all_bios(conf, r10_bio);
 222         mempool_free(r10_bio, conf->r10bio_pool);
 223 }
 224
 225 static void put_buf(struct r10bio *r10_bio)
 226 {
 227         struct r10conf *conf = r10_bio->mddev->private;
 228
 229         mempool_free(r10_bio, conf->r10buf_pool);
 230
 231         lower_barrier(conf);
 232 }
 233
 234 static void reschedule_retry(struct r10bio *r10_bio)
 235 {
 236         unsigned long flags;
 237         struct mddev *mddev = r10_bio->mddev;
 238         struct r10conf *conf = mddev->private;
 239
 240         spin_lock_irqsave(&conf->device_lock, flags);
 241         list_add(&r10_bio->retry_list, &conf->retry_list);
 242         conf->nr_queued ++;
 243         spin_unlock_irqrestore(&conf->device_lock, flags);
 244
 245         /* wake up frozen array... */
 246         wake_up(&conf->wait_barrier);
 247
 248         md_wakeup_thread(mddev->thread);
 249 }
 250
 251 /*
 252  * raid_end_bio_io() is called when we have finished servicing a mirrored
 253  * operation and are ready to return a success/failure code to the buffer
 254  * cache layer.
 255  */
 256 static void raid_end_bio_io(struct r10bio *r10_bio)
 257 {
 258         struct bio *bio = r10_bio->master_bio;
 259         int done;
 260         struct r10conf *conf = r10_bio->mddev->private;
 261
 262         if (bio->bi_phys_segments) {
 263                 unsigned long flags;
 264                 spin_lock_irqsave(&conf->device_lock, flags);
 265                 bio->bi_phys_segments--;
 266                 done = (bio->bi_phys_segments == 0);
 267                 spin_unlock_irqrestore(&conf->device_lock, flags);
 268         } else
 269                 done = 1;
 270         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
 271                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 272         if (done) {
 273                 bio_endio(bio, 0);
 274                 /*
 275                  * Wake up any possible resync thread that waits for the device
 276                  * to go idle.
 277                  */
 278                 allow_barrier(conf);
 279         }
 280         free_r10bio(r10_bio);
 281 }
 282
 283 /*
 284  * Update disk head position estimator based on IRQ completion info.
 285  */
 286 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
 287 {
 288         struct r10conf *conf = r10_bio->mddev->private;
 289
 290         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 291                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 292 }
 293
 294 /*
 295  * Find the disk number which triggered given bio
 296  */
 297 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
 298                          struct bio *bio, int *slotp, int *replp)
 299 {
 300         int slot;
 301         int repl = 0;
 302
 303         for (slot = 0; slot < conf->copies; slot++) {
 304                 if (r10_bio->devs[slot].bio == bio)
 305                         break;
 306                 if (r10_bio->devs[slot].repl_bio == bio) {
 307                         repl = 1;
 308                         break;
 309                 }
 310         }
 311
 312         BUG_ON(slot == conf->copies);
 313         update_head_pos(slot, r10_bio);
 314
 315         if (slotp)
 316                 *slotp = slot;
 317         if (replp)
 318                 *replp = repl;
 319         return r10_bio->devs[slot].devnum;
 320 }
 321
 322 static void raid10_end_read_request(struct bio *bio, int error)
 323 {
 324         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 325         struct r10bio *r10_bio = bio->bi_private;
 326         int slot, dev;
 327         struct md_rdev *rdev;
 328         struct r10conf *conf = r10_bio->mddev->private;
 329
 330
 331         slot = r10_bio->read_slot;
 332         dev = r10_bio->devs[slot].devnum;
 333         rdev = r10_bio->devs[slot].rdev;
 334         /*
 335          * this branch is our 'one mirror IO has finished' event handler:
 336          */
 337         update_head_pos(slot, r10_bio);
 338
 339         if (uptodate) {
 340                 /*
 341                  * Set R10BIO_Uptodate in our master bio, so that
 342                  * we will return a good error code to the higher
 343                  * levels even if IO on some other mirrored buffer fails.
 344                  *
 345                  * The 'master' represents the composite IO operation to
 346                  * user-side. So if something waits for IO, then it will
 347                  * wait for the 'master' bio.
 348                  */
 349                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 350                 raid_end_bio_io(r10_bio);
 351                 rdev_dec_pending(rdev, conf->mddev);
 352         } else {
 353                 /*
 354                  * oops, read error - keep the refcount on the rdev
 355                  */
 356                 char b[BDEVNAME_SIZE];
 357                 printk_ratelimited(KERN_ERR
 358                                    "md/raid10:%s: %s: rescheduling sector %llu\n",
 359                                    mdname(conf->mddev),
 360                                    bdevname(rdev->bdev, b),
 361                                    (unsigned long long)r10_bio->sector);
 362                 set_bit(R10BIO_ReadError, &r10_bio->state);
 363                 reschedule_retry(r10_bio);
 364         }
 365 }
 366
 367 static void close_write(struct r10bio *r10_bio)
 368 {
 369         /* clear the bitmap if all writes complete successfully */
 370         bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 371                         r10_bio->sectors,
 372                         !test_bit(R10BIO_Degraded, &r10_bio->state),
 373                         0);
 374         md_write_end(r10_bio->mddev);
 375 }
 376
 377 static void one_write_done(struct r10bio *r10_bio)
 378 {
 379         if (atomic_dec_and_test(&r10_bio->remaining)) {
 380                 if (test_bit(R10BIO_WriteError, &r10_bio->state))
 381                         reschedule_retry(r10_bio);
 382                 else {
 383                         close_write(r10_bio);
 384                         if (test_bit(R10BIO_MadeGood, &r10_bio->state))
 385                                 reschedule_retry(r10_bio);
 386                         else
 387                                 raid_end_bio_io(r10_bio);
 388                 }
 389         }
 390 }
 391
 392 static void raid10_end_write_request(struct bio *bio, int error)
 393 {
 394         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 395         struct r10bio *r10_bio = bio->bi_private;
 396         int dev;
 397         int dec_rdev = 1;
 398         struct r10conf *conf = r10_bio->mddev->private;
 399         int slot, repl;
 400         struct md_rdev *rdev = NULL;
 401
 402         dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
 403
 404         if (repl)
 405                 rdev = conf->mirrors[dev].replacement;
 406         if (!rdev) {
 407                 smp_rmb();
 408                 repl = 0;
 409                 rdev = conf->mirrors[dev].rdev;
 410         }
 411         /*
 412          * this branch is our 'one mirror IO has finished' event handler:
 413          */
 414         if (!uptodate) {
 415                 if (repl)
 416                         /* Never record new bad blocks to replacement,
 417                          * just fail it.
 418                          */
 419                         md_error(rdev->mddev, rdev);
 420                 else {
 421                         set_bit(WriteErrorSeen, &rdev->flags);
 422                         set_bit(R10BIO_WriteError, &r10_bio->state);
 423                         dec_rdev = 0;
 424                 }
 425         } else {
 426                 /*
 427                  * Set R10BIO_Uptodate in our master bio, so that
 428                  * we will return a good error code for to the higher
 429                  * levels even if IO on some other mirrored buffer fails.
 430                  *
 431                  * The 'master' represents the composite IO operation to
 432                  * user-side. So if something waits for IO, then it will
 433                  * wait for the 'master' bio.
 434                  */
 435                 sector_t first_bad;
 436                 int bad_sectors;
 437
 438                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 439
 440                 /* Maybe we can clear some bad blocks. */
 441                 if (is_badblock(rdev,
 442                                 r10_bio->devs[slot].addr,
 443                                 r10_bio->sectors,
 444                                 &first_bad, &bad_sectors)) {
 445                         bio_put(bio);
 446                         if (repl)
 447                                 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
 448                         else
 449                                 r10_bio->devs[slot].bio = IO_MADE_GOOD;
 450                         dec_rdev = 0;
 451                         set_bit(R10BIO_MadeGood, &r10_bio->state);
 452                 }
 453         }
 454
 455         /*
 456          *
 457          * Let's see if all mirrored write operations have finished
 458          * already.
 459          */
 460         one_write_done(r10_bio);
 461         if (dec_rdev)
 462                 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 463 }
 464
 465 /*
 466  * RAID10 layout manager
 467  * As well as the chunksize and raid_disks count, there are two
 468  * parameters: near_copies and far_copies.
 469  * near_copies * far_copies must be <= raid_disks.
 470  * Normally one of these will be 1.
 471  * If both are 1, we get raid0.
 472  * If near_copies == raid_disks, we get raid1.
 473  *
 474  * Chunks are laid out in raid0 style with near_copies copies of the
 475  * first chunk, followed by near_copies copies of the next chunk and
 476  * so on.
 477  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 478  * as described above, we start again with a device offset of near_copies.
 479  * So we effectively have another copy of the whole array further down all
 480  * the drives, but with blocks on different drives.
 481  * With this layout, and block is never stored twice on the one device.
 482  *
 483  * raid10_find_phys finds the sector offset of a given virtual sector
 484  * on each device that it is on.
 485  *
 486  * raid10_find_virt does the reverse mapping, from a device and a
 487  * sector offset to a virtual address
 488  */
 489
 490 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
 491 {
 492         int n,f;
 493         sector_t sector;
 494         sector_t chunk;
 495         sector_t stripe;
 496         int dev;
 497
 498         int slot = 0;
 499
 500         /* now calculate first sector/dev */
 501         chunk = r10bio->sector >> conf->chunk_shift;
 502         sector = r10bio->sector & conf->chunk_mask;
 503
 504         chunk *= conf->near_copies;
 505         stripe = chunk;
 506         dev = sector_div(stripe, conf->raid_disks);
 507         if (conf->far_offset)
 508                 stripe *= conf->far_copies;
 509
 510         sector += stripe << conf->chunk_shift;
 511
 512         /* and calculate all the others */
 513         for (n=0; n < conf->near_copies; n++) {
 514                 int d = dev;
 515                 sector_t s = sector;
 516                 r10bio->devs[slot].addr = sector;
 517                 r10bio->devs[slot].devnum = d;
 518                 slot++;
 519
 520                 for (f = 1; f < conf->far_copies; f++) {
 521                         d += conf->near_copies;
 522                         if (d >= conf->raid_disks)
 523                                 d -= conf->raid_disks;
 524                         s += conf->stride;
 525                         r10bio->devs[slot].devnum = d;
 526                         r10bio->devs[slot].addr = s;
 527                         slot++;
 528                 }
 529                 dev++;
 530                 if (dev >= conf->raid_disks) {
 531                         dev = 0;
 532                         sector += (conf->chunk_mask + 1);
 533                 }
 534         }
 535         BUG_ON(slot != conf->copies);
 536 }
 537
 538 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
 539 {
 540         sector_t offset, chunk, vchunk;
 541
 542         offset = sector & conf->chunk_mask;
 543         if (conf->far_offset) {
 544                 int fc;
 545                 chunk = sector >> conf->chunk_shift;
 546                 fc = sector_div(chunk, conf->far_copies);
 547                 dev -= fc * conf->near_copies;
 548                 if (dev < 0)
 549                         dev += conf->raid_disks;
 550         } else {
 551                 while (sector >= conf->stride) {
 552                         sector -= conf->stride;
 553                         if (dev < conf->near_copies)
 554                                 dev += conf->raid_disks - conf->near_copies;
 555                         else
 556                                 dev -= conf->near_copies;
 557                 }
 558                 chunk = sector >> conf->chunk_shift;
 559         }
 560         vchunk = chunk * conf->raid_disks + dev;
 561         sector_div(vchunk, conf->near_copies);
 562         return (vchunk << conf->chunk_shift) + offset;
 563 }
 564
 565 /**
 566  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 567  *      @q: request queue
 568  *      @bvm: properties of new bio
 569  *      @biovec: the request that could be merged to it.
 570  *
 571  *      Return amount of bytes we can accept at this offset
 572  *      If near_copies == raid_disk, there are no striping issues,
 573  *      but in that case, the function isn't called at all.
 574  */
 575 static int raid10_mergeable_bvec(struct request_queue *q,
 576                                  struct bvec_merge_data *bvm,
 577                                  struct bio_vec *biovec)
 578 {
 579         struct mddev *mddev = q->queuedata;
 580         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 581         int max;
 582         unsigned int chunk_sectors = mddev->chunk_sectors;
 583         unsigned int bio_sectors = bvm->bi_size >> 9;
 584
 585         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 586         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 587         if (max <= biovec->bv_len && bio_sectors == 0)
 588                 return biovec->bv_len;
 589         else
 590                 return max;
 591 }
 592
 593 /*
 594  * This routine returns the disk from which the requested read should
 595  * be done. There is a per-array 'next expected sequential IO' sector
 596  * number - if this matches on the next IO then we use the last disk.
 597  * There is also a per-disk 'last know head position' sector that is
 598  * maintained from IRQ contexts, both the normal and the resync IO
 599  * completion handlers update this position correctly. If there is no
 600  * perfect sequential match then we pick the disk whose head is closest.
 601  *
 602  * If there are 2 mirrors in the same 2 devices, performance degrades
 603  * because position is mirror, not device based.
 604  *
 605  * The rdev for the device selected will have nr_pending incremented.
 606  */
 607
 608 /*
 609  * FIXME: possibly should rethink readbalancing and do it differently
 610  * depending on near_copies / far_copies geometry.
 611  */
 612 static struct md_rdev *read_balance(struct r10conf *conf,
 613                                     struct r10bio *r10_bio,
 614                                     int *max_sectors)
 615 {
 616         const sector_t this_sector = r10_bio->sector;
 617         int disk, slot;
 618         int sectors = r10_bio->sectors;
 619         int best_good_sectors;
 620         sector_t new_distance, best_dist;
 621         struct md_rdev *rdev, *best_rdev;
 622         int do_balance;
 623         int best_slot;
 624
 625         raid10_find_phys(conf, r10_bio);
 626         rcu_read_lock();
 627 retry:
 628         sectors = r10_bio->sectors;
 629         best_slot = -1;
 630         best_rdev = NULL;
 631         best_dist = MaxSector;
 632         best_good_sectors = 0;
 633         do_balance = 1;
 634         /*
 635          * Check if we can balance. We can balance on the whole
 636          * device if no resync is going on (recovery is ok), or below
 637          * the resync window. We take the first readable disk when
 638          * above the resync window.
 639          */
 640         if (conf->mddev->recovery_cp < MaxSector
 641             && (this_sector + sectors >= conf->next_resync))
 642                 do_balance = 0;
 643
 644         for (slot = 0; slot < conf->copies ; slot++) {
 645                 sector_t first_bad;
 646                 int bad_sectors;
 647                 sector_t dev_sector;
 648
 649                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
 650                         continue;
 651                 disk = r10_bio->devs[slot].devnum;
 652                 rdev = rcu_dereference(conf->mirrors[disk].replacement);
 653                 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
 654                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 655                         rdev = rcu_dereference(conf->mirrors[disk].rdev);
 656                 if (rdev == NULL)
 657                         continue;
 658                 if (test_bit(Faulty, &rdev->flags))
 659                         continue;
 660                 if (!test_bit(In_sync, &rdev->flags) &&
 661                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 662                         continue;
 663
 664                 dev_sector = r10_bio->devs[slot].addr;
 665                 if (is_badblock(rdev, dev_sector, sectors,
 666                                 &first_bad, &bad_sectors)) {
 667                         if (best_dist < MaxSector)
 668                                 /* Already have a better slot */
 669                                 continue;
 670                         if (first_bad <= dev_sector) {
 671                                 /* Cannot read here.  If this is the
 672                                  * 'primary' device, then we must not read
 673                                  * beyond 'bad_sectors' from another device.
 674                                  */
 675                                 bad_sectors -= (dev_sector - first_bad);
 676                                 if (!do_balance && sectors > bad_sectors)
 677                                         sectors = bad_sectors;
 678                                 if (best_good_sectors > sectors)
 679                                         best_good_sectors = sectors;
 680                         } else {
 681                                 sector_t good_sectors =
 682                                         first_bad - dev_sector;
 683                                 if (good_sectors > best_good_sectors) {
 684                                         best_good_sectors = good_sectors;
 685                                         best_slot = slot;
 686                                         best_rdev = rdev;
 687                                 }
 688                                 if (!do_balance)
 689                                         /* Must read from here */
 690                                         break;
 691                         }
 692                         continue;
 693                 } else
 694                         best_good_sectors = sectors;
 695
 696                 if (!do_balance)
 697                         break;
 698
 699                 /* This optimisation is debatable, and completely destroys
 700                  * sequential read speed for 'far copies' arrays.  So only
 701                  * keep it for 'near' arrays, and review those later.
 702                  */
 703                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
 704                         break;
 705
 706                 /* for far > 1 always use the lowest address */
 707                 if (conf->far_copies > 1)
 708                         new_distance = r10_bio->devs[slot].addr;
 709                 else
 710                         new_distance = abs(r10_bio->devs[slot].addr -
 711                                            conf->mirrors[disk].head_position);
 712                 if (new_distance < best_dist) {
 713                         best_dist = new_distance;
 714                         best_slot = slot;
 715                         best_rdev = rdev;
 716                 }
 717         }
 718         if (slot >= conf->copies) {
 719                 slot = best_slot;
 720                 rdev = best_rdev;
 721         }
 722
 723         if (slot >= 0) {
 724                 atomic_inc(&rdev->nr_pending);
 725                 if (test_bit(Faulty, &rdev->flags)) {
 726                         /* Cannot risk returning a device that failed
 727                          * before we inc'ed nr_pending
 728                          */
 729                         rdev_dec_pending(rdev, conf->mddev);
 730                         goto retry;
 731                 }
 732                 r10_bio->read_slot = slot;
 733         } else
 734                 rdev = NULL;
 735         rcu_read_unlock();
 736         *max_sectors = best_good_sectors;
 737
 738         return rdev;
 739 }
 740
 741 static int raid10_congested(void *data, int bits)
 742 {
 743         struct mddev *mddev = data;
 744         struct r10conf *conf = mddev->private;
 745         int i, ret = 0;
 746
 747         if ((bits & (1 << BDI_async_congested)) &&
 748             conf->pending_count >= max_queued_requests)
 749                 return 1;
 750
 751         if (mddev_congested(mddev, bits))
 752                 return 1;
 753         rcu_read_lock();
 754         for (i = 0; i < conf->raid_disks && ret == 0; i++) {
 755                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 756                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 757                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 758
 759                         ret |= bdi_congested(&q->backing_dev_info, bits);
 760                 }
 761         }
 762         rcu_read_unlock();
 763         return ret;
 764 }
 765
 766 static void flush_pending_writes(struct r10conf *conf)
 767 {
 768         /* Any writes that have been queued but are awaiting
 769          * bitmap updates get flushed here.
 770          */
 771         spin_lock_irq(&conf->device_lock);
 772
 773         if (conf->pending_bio_list.head) {
 774                 struct bio *bio;
 775                 bio = bio_list_get(&conf->pending_bio_list);
 776                 conf->pending_count = 0;
 777                 spin_unlock_irq(&conf->device_lock);
 778                 /* flush any pending bitmap writes to disk
 779                  * before proceeding w/ I/O */
 780                 bitmap_unplug(conf->mddev->bitmap);
 781                 wake_up(&conf->wait_barrier);
 782
 783                 while (bio) { /* submit pending writes */
 784                         struct bio *next = bio->bi_next;
 785                         bio->bi_next = NULL;
 786                         generic_make_request(bio);
 787                         bio = next;
 788                 }
 789         } else
 790                 spin_unlock_irq(&conf->device_lock);
 791 }
 792
 793 /* Barriers....
 794  * Sometimes we need to suspend IO while we do something else,
 795  * either some resync/recovery, or reconfigure the array.
 796  * To do this we raise a 'barrier'.
 797  * The 'barrier' is a counter that can be raised multiple times
 798  * to count how many activities are happening which preclude
 799  * normal IO.
 800  * We can only raise the barrier if there is no pending IO.
 801  * i.e. if nr_pending == 0.
 802  * We choose only to raise the barrier if no-one is waiting for the
 803  * barrier to go down.  This means that as soon as an IO request
 804  * is ready, no other operations which require a barrier will start
 805  * until the IO request has had a chance.
 806  *
 807  * So: regular IO calls 'wait_barrier'.  When that returns there
 808  *    is no backgroup IO happening,  It must arrange to call
 809  *    allow_barrier when it has finished its IO.
 810  * backgroup IO calls must call raise_barrier.  Once that returns
 811  *    there is no normal IO happeing.  It must arrange to call
 812  *    lower_barrier when the particular background IO completes.
 813  */
 814
 815 static void raise_barrier(struct r10conf *conf, int force)
 816 {
 817         BUG_ON(force && !conf->barrier);
 818         spin_lock_irq(&conf->resync_lock);
 819
 820         /* Wait until no block IO is waiting (unless 'force') */
 821         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 822                             conf->resync_lock, );
 823
 824         /* block any new IO from starting */
 825         conf->barrier++;
 826
 827         /* Now wait for all pending IO to complete */
 828         wait_event_lock_irq(conf->wait_barrier,
 829                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 830                             conf->resync_lock, );
 831
 832         spin_unlock_irq(&conf->resync_lock);
 833 }
 834
 835 static void lower_barrier(struct r10conf *conf)
 836 {
 837         unsigned long flags;
 838         spin_lock_irqsave(&conf->resync_lock, flags);
 839         conf->barrier--;
 840         spin_unlock_irqrestore(&conf->resync_lock, flags);
 841         wake_up(&conf->wait_barrier);
 842 }
 843
 844 static void wait_barrier(struct r10conf *conf)
 845 {
 846         spin_lock_irq(&conf->resync_lock);
 847         if (conf->barrier) {
 848                 conf->nr_waiting++;
 849                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 850                                     conf->resync_lock,
 851                                     );
 852                 conf->nr_waiting--;
 853         }
 854         conf->nr_pending++;
 855         spin_unlock_irq(&conf->resync_lock);
 856 }
 857
 858 static void allow_barrier(struct r10conf *conf)
 859 {
 860         unsigned long flags;
 861         spin_lock_irqsave(&conf->resync_lock, flags);
 862         conf->nr_pending--;
 863         spin_unlock_irqrestore(&conf->resync_lock, flags);
 864         wake_up(&conf->wait_barrier);
 865 }
 866
 867 static void freeze_array(struct r10conf *conf)
 868 {
 869         /* stop syncio and normal IO and wait for everything to
 870          * go quiet.
 871          * We increment barrier and nr_waiting, and then
 872          * wait until nr_pending match nr_queued+1
 873          * This is called in the context of one normal IO request
 874          * that has failed. Thus any sync request that might be pending
 875          * will be blocked by nr_pending, and we need to wait for
 876          * pending IO requests to complete or be queued for re-try.
 877          * Thus the number queued (nr_queued) plus this request (1)
 878          * must match the number of pending IOs (nr_pending) before
 879          * we continue.
 880          */
 881         spin_lock_irq(&conf->resync_lock);
 882         conf->barrier++;
 883         conf->nr_waiting++;
 884         wait_event_lock_irq(conf->wait_barrier,
 885                             conf->nr_pending == conf->nr_queued+1,
 886                             conf->resync_lock,
 887                             flush_pending_writes(conf));
 888
 889         spin_unlock_irq(&conf->resync_lock);
 890 }
 891
 892 static void unfreeze_array(struct r10conf *conf)
 893 {
 894         /* reverse the effect of the freeze */
 895         spin_lock_irq(&conf->resync_lock);
 896         conf->barrier--;
 897         conf->nr_waiting--;
 898         wake_up(&conf->wait_barrier);
 899         spin_unlock_irq(&conf->resync_lock);
 900 }
 901
 902 static void make_request(struct mddev *mddev, struct bio * bio)
 903 {
 904         struct r10conf *conf = mddev->private;
 905         struct r10bio *r10_bio;
 906         struct bio *read_bio;
 907         int i;
 908         int chunk_sects = conf->chunk_mask + 1;
 909         const int rw = bio_data_dir(bio);
 910         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 911         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
 912         unsigned long flags;
 913         struct md_rdev *blocked_rdev;
 914         int plugged;
 915         int sectors_handled;
 916         int max_sectors;
 917
 918         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
 919                 md_flush_request(mddev, bio);
 920                 return;
 921         }
 922
 923         /* If this request crosses a chunk boundary, we need to
 924          * split it.  This will only happen for 1 PAGE (or less) requests.
 925          */
 926         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 927                       > chunk_sects &&
 928                     conf->near_copies < conf->raid_disks)) {
 929                 struct bio_pair *bp;
 930                 /* Sanity check -- queue functions should prevent this happening */
 931                 if (bio->bi_vcnt != 1 ||
 932                     bio->bi_idx != 0)
 933                         goto bad_map;
 934                 /* This is a one page bio that upper layers
 935                  * refuse to split for us, so we need to split it.
 936                  */
 937                 bp = bio_split(bio,
 938                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 939
 940                 /* Each of these 'make_request' calls will call 'wait_barrier'.
 941                  * If the first succeeds but the second blocks due to the resync
 942                  * thread raising the barrier, we will deadlock because the
 943                  * IO to the underlying device will be queued in generic_make_request
 944                  * and will never complete, so will never reduce nr_pending.
 945                  * So increment nr_waiting here so no new raise_barriers will
 946                  * succeed, and so the second wait_barrier cannot block.
 947                  */
 948                 spin_lock_irq(&conf->resync_lock);
 949                 conf->nr_waiting++;
 950                 spin_unlock_irq(&conf->resync_lock);
 951
 952                 make_request(mddev, &bp->bio1);
 953                 make_request(mddev, &bp->bio2);
 954
 955                 spin_lock_irq(&conf->resync_lock);
 956                 conf->nr_waiting--;
 957                 wake_up(&conf->wait_barrier);
 958                 spin_unlock_irq(&conf->resync_lock);
 959
 960                 bio_pair_release(bp);
 961                 return;
 962         bad_map:
 963                 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
 964                        " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
 965                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 966
 967                 bio_io_error(bio);
 968                 return;
 969         }
 970
 971         md_write_start(mddev, bio);
 972
 973         /*
 974          * Register the new request and wait if the reconstruction
 975          * thread has put up a bar for new requests.
 976          * Continue immediately if no resync is active currently.
 977          */
 978         wait_barrier(conf);
 979
 980         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 981
 982         r10_bio->master_bio = bio;
 983         r10_bio->sectors = bio->bi_size >> 9;
 984
 985         r10_bio->mddev = mddev;
 986         r10_bio->sector = bio->bi_sector;
 987         r10_bio->state = 0;
 988
 989         /* We might need to issue multiple reads to different
 990          * devices if there are bad blocks around, so we keep
 991          * track of the number of reads in bio->bi_phys_segments.
 992          * If this is 0, there is only one r10_bio and no locking
 993          * will be needed when the request completes.  If it is
 994          * non-zero, then it is the number of not-completed requests.
 995          */
 996         bio->bi_phys_segments = 0;
 997         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
 998
 999         if (rw == READ) {
1000                 /*
1001                  * read balancing logic:
1002                  */
1003                 struct md_rdev *rdev;
1004                 int slot;
1005
1006 read_again:
1007                 rdev = read_balance(conf, r10_bio, &max_sectors);
1008                 if (!rdev) {
1009                         raid_end_bio_io(r10_bio);
1010                         return;
1011                 }
1012                 slot = r10_bio->read_slot;
1013
1014                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1015                 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1016                             max_sectors);
1017
1018                 r10_bio->devs[slot].bio = read_bio;
1019                 r10_bio->devs[slot].rdev = rdev;
1020
1021                 read_bio->bi_sector = r10_bio->devs[slot].addr +
1022                         rdev->data_offset;
1023                 read_bio->bi_bdev = rdev->bdev;
1024                 read_bio->bi_end_io = raid10_end_read_request;
1025                 read_bio->bi_rw = READ | do_sync;
1026                 read_bio->bi_private = r10_bio;
1027
1028                 if (max_sectors < r10_bio->sectors) {
1029                         /* Could not read all from this device, so we will
1030                          * need another r10_bio.
1031                          */
1032                         sectors_handled = (r10_bio->sectors + max_sectors
1033                                            - bio->bi_sector);
1034                         r10_bio->sectors = max_sectors;
1035                         spin_lock_irq(&conf->device_lock);
1036                         if (bio->bi_phys_segments == 0)
1037                                 bio->bi_phys_segments = 2;
1038                         else
1039                                 bio->bi_phys_segments++;
1040                         spin_unlock(&conf->device_lock);
1041                         /* Cannot call generic_make_request directly
1042                          * as that will be queued in __generic_make_request
1043                          * and subsequent mempool_alloc might block
1044                          * waiting for it.  so hand bio over to raid10d.
1045                          */
1046                         reschedule_retry(r10_bio);
1047
1048                         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1049
1050                         r10_bio->master_bio = bio;
1051                         r10_bio->sectors = ((bio->bi_size >> 9)
1052                                             - sectors_handled);
1053                         r10_bio->state = 0;
1054                         r10_bio->mddev = mddev;
1055                         r10_bio->sector = bio->bi_sector + sectors_handled;
1056                         goto read_again;
1057                 } else
1058                         generic_make_request(read_bio);
1059                 return;
1060         }
1061
1062         /*
1063          * WRITE:
1064          */
1065         if (conf->pending_count >= max_queued_requests) {
1066                 md_wakeup_thread(mddev->thread);
1067                 wait_event(conf->wait_barrier,
1068                            conf->pending_count < max_queued_requests);
1069         }
1070         /* first select target devices under rcu_lock and
1071          * inc refcount on their rdev.  Record them by setting
1072          * bios[x] to bio
1073          * If there are known/acknowledged bad blocks on any device
1074          * on which we have seen a write error, we want to avoid
1075          * writing to those blocks.  This potentially requires several
1076          * writes to write around the bad blocks.  Each set of writes
1077          * gets its own r10_bio with a set of bios attached.  The number
1078          * of r10_bios is recored in bio->bi_phys_segments just as with
1079          * the read case.
1080          */
1081         plugged = mddev_check_plugged(mddev);
1082
1083         r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1084         raid10_find_phys(conf, r10_bio);
1085 retry_write:
1086         blocked_rdev = NULL;
1087         rcu_read_lock();
1088         max_sectors = r10_bio->sectors;
1089
1090         for (i = 0;  i < conf->copies; i++) {
1091                 int d = r10_bio->devs[i].devnum;
1092                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1093                 struct md_rdev *rrdev = rcu_dereference(
1094                         conf->mirrors[d].replacement);
1095                 if (rdev == rrdev)
1096                         rrdev = NULL;
1097                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1098                         atomic_inc(&rdev->nr_pending);
1099                         blocked_rdev = rdev;
1100                         break;
1101                 }
1102                 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1103                         atomic_inc(&rrdev->nr_pending);
1104                         blocked_rdev = rrdev;
1105                         break;
1106                 }
1107                 if (rrdev && test_bit(Faulty, &rrdev->flags))
1108                         rrdev = NULL;
1109
1110                 r10_bio->devs[i].bio = NULL;
1111                 r10_bio->devs[i].repl_bio = NULL;
1112                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1113                         set_bit(R10BIO_Degraded, &r10_bio->state);
1114                         continue;
1115                 }
1116                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1117                         sector_t first_bad;
1118                         sector_t dev_sector = r10_bio->devs[i].addr;
1119                         int bad_sectors;
1120                         int is_bad;
1121
1122                         is_bad = is_badblock(rdev, dev_sector,
1123                                              max_sectors,
1124                                              &first_bad, &bad_sectors);
1125                         if (is_bad < 0) {
1126                                 /* Mustn't write here until the bad block
1127                                  * is acknowledged
1128                                  */
1129                                 atomic_inc(&rdev->nr_pending);
1130                                 set_bit(BlockedBadBlocks, &rdev->flags);
1131                                 blocked_rdev = rdev;
1132                                 break;
1133                         }
1134                         if (is_bad && first_bad <= dev_sector) {
1135                                 /* Cannot write here at all */
1136                                 bad_sectors -= (dev_sector - first_bad);
1137                                 if (bad_sectors < max_sectors)
1138                                         /* Mustn't write more than bad_sectors
1139                                          * to other devices yet
1140                                          */
1141                                         max_sectors = bad_sectors;
1142                                 /* We don't set R10BIO_Degraded as that
1143                                  * only applies if the disk is missing,
1144                                  * so it might be re-added, and we want to
1145                                  * know to recover this chunk.
1146                                  * In this case the device is here, and the
1147                                  * fact that this chunk is not in-sync is
1148                                  * recorded in the bad block log.
1149                                  */
1150                                 continue;
1151                         }
1152                         if (is_bad) {
1153                                 int good_sectors = first_bad - dev_sector;
1154                                 if (good_sectors < max_sectors)
1155                                         max_sectors = good_sectors;
1156                         }
1157                 }
1158                 r10_bio->devs[i].bio = bio;
1159                 atomic_inc(&rdev->nr_pending);
1160                 if (rrdev) {
1161                         r10_bio->devs[i].repl_bio = bio;
1162                         atomic_inc(&rrdev->nr_pending);
1163                 }
1164         }
1165         rcu_read_unlock();
1166
1167         if (unlikely(blocked_rdev)) {
1168                 /* Have to wait for this device to get unblocked, then retry */
1169                 int j;
1170                 int d;
1171
1172                 for (j = 0; j < i; j++) {
1173                         if (r10_bio->devs[j].bio) {
1174                                 d = r10_bio->devs[j].devnum;
1175                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1176                         }
1177                         if (r10_bio->devs[j].repl_bio) {
1178                                 struct md_rdev *rdev;
1179                                 d = r10_bio->devs[j].devnum;
1180                                 rdev = conf->mirrors[d].replacement;
1181                                 if (!rdev) {
1182                                         /* Race with remove_disk */
1183                                         smp_mb();
1184                                         rdev = conf->mirrors[d].rdev;
1185                                 }
1186                                 rdev_dec_pending(rdev, mddev);
1187                         }
1188                 }
1189                 allow_barrier(conf);
1190                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1191                 wait_barrier(conf);
1192                 goto retry_write;
1193         }
1194
1195         if (max_sectors < r10_bio->sectors) {
1196                 /* We are splitting this into multiple parts, so
1197                  * we need to prepare for allocating another r10_bio.
1198                  */
1199                 r10_bio->sectors = max_sectors;
1200                 spin_lock_irq(&conf->device_lock);
1201                 if (bio->bi_phys_segments == 0)
1202                         bio->bi_phys_segments = 2;
1203                 else
1204                         bio->bi_phys_segments++;
1205                 spin_unlock_irq(&conf->device_lock);
1206         }
1207         sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1208
1209         atomic_set(&r10_bio->remaining, 1);
1210         bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1211
1212         for (i = 0; i < conf->copies; i++) {
1213                 struct bio *mbio;
1214                 int d = r10_bio->devs[i].devnum;
1215                 if (!r10_bio->devs[i].bio)
1216                         continue;
1217
1218                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1219                 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1220                             max_sectors);
1221                 r10_bio->devs[i].bio = mbio;
1222
1223                 mbio->bi_sector = (r10_bio->devs[i].addr+
1224                                    conf->mirrors[d].rdev->data_offset);
1225                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1226                 mbio->bi_end_io = raid10_end_write_request;
1227                 mbio->bi_rw = WRITE | do_sync | do_fua;
1228                 mbio->bi_private = r10_bio;
1229
1230                 atomic_inc(&r10_bio->remaining);
1231                 spin_lock_irqsave(&conf->device_lock, flags);
1232                 bio_list_add(&conf->pending_bio_list, mbio);
1233                 conf->pending_count++;
1234                 spin_unlock_irqrestore(&conf->device_lock, flags);
1235
1236                 if (!r10_bio->devs[i].repl_bio)
1237                         continue;
1238
1239                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1240                 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1241                             max_sectors);
1242                 r10_bio->devs[i].repl_bio = mbio;
1243
1244                 /* We are actively writing to the original device
1245                  * so it cannot disappear, so the replacement cannot
1246                  * become NULL here
1247                  */
1248                 mbio->bi_sector = (r10_bio->devs[i].addr+
1249                                    conf->mirrors[d].replacement->data_offset);
1250                 mbio->bi_bdev = conf->mirrors[d].replacement->bdev;
1251                 mbio->bi_end_io = raid10_end_write_request;
1252                 mbio->bi_rw = WRITE | do_sync | do_fua;
1253                 mbio->bi_private = r10_bio;
1254
1255                 atomic_inc(&r10_bio->remaining);
1256                 spin_lock_irqsave(&conf->device_lock, flags);
1257                 bio_list_add(&conf->pending_bio_list, mbio);
1258                 conf->pending_count++;
1259                 spin_unlock_irqrestore(&conf->device_lock, flags);
1260         }
1261
1262         /* Don't remove the bias on 'remaining' (one_write_done) until
1263          * after checking if we need to go around again.
1264          */
1265
1266         if (sectors_handled < (bio->bi_size >> 9)) {
1267                 one_write_done(r10_bio);
1268                 /* We need another r10_bio.  It has already been counted
1269                  * in bio->bi_phys_segments.
1270                  */
1271                 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1272
1273                 r10_bio->master_bio = bio;
1274                 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1275
1276                 r10_bio->mddev = mddev;
1277                 r10_bio->sector = bio->bi_sector + sectors_handled;
1278                 r10_bio->state = 0;
1279                 goto retry_write;
1280         }
1281         one_write_done(r10_bio);
1282
1283         /* In case raid10d snuck in to freeze_array */
1284         wake_up(&conf->wait_barrier);
1285
1286         if (do_sync || !mddev->bitmap || !plugged)
1287                 md_wakeup_thread(mddev->thread);
1288 }
1289
1290 static void status(struct seq_file *seq, struct mddev *mddev)
1291 {
1292         struct r10conf *conf = mddev->private;
1293         int i;
1294
1295         if (conf->near_copies < conf->raid_disks)
1296                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1297         if (conf->near_copies > 1)
1298                 seq_printf(seq, " %d near-copies", conf->near_copies);
1299         if (conf->far_copies > 1) {
1300                 if (conf->far_offset)
1301                         seq_printf(seq, " %d offset-copies", conf->far_copies);
1302                 else
1303                         seq_printf(seq, " %d far-copies", conf->far_copies);
1304         }
1305         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1306                                         conf->raid_disks - mddev->degraded);
1307         for (i = 0; i < conf->raid_disks; i++)
1308                 seq_printf(seq, "%s",
1309                               conf->mirrors[i].rdev &&
1310                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1311         seq_printf(seq, "]");
1312 }
1313
1314 /* check if there are enough drives for
1315  * every block to appear on atleast one.
1316  * Don't consider the device numbered 'ignore'
1317  * as we might be about to remove it.
1318  */
1319 static int enough(struct r10conf *conf, int ignore)
1320 {
1321         int first = 0;
1322
1323         do {
1324                 int n = conf->copies;
1325                 int cnt = 0;
1326                 while (n--) {
1327                         if (conf->mirrors[first].rdev &&
1328                             first != ignore)
1329                                 cnt++;
1330                         first = (first+1) % conf->raid_disks;
1331                 }
1332                 if (cnt == 0)
1333                         return 0;
1334         } while (first != 0);
1335         return 1;
1336 }
1337
1338 static void error(struct mddev *mddev, struct md_rdev *rdev)
1339 {
1340         char b[BDEVNAME_SIZE];
1341         struct r10conf *conf = mddev->private;
1342
1343         /*
1344          * If it is not operational, then we have already marked it as dead
1345          * else if it is the last working disks, ignore the error, let the
1346          * next level up know.
1347          * else mark the drive as failed
1348          */
1349         if (test_bit(In_sync, &rdev->flags)
1350             && !enough(conf, rdev->raid_disk))
1351                 /*
1352                  * Don't fail the drive, just return an IO error.
1353                  */
1354                 return;
1355         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1356                 unsigned long flags;
1357                 spin_lock_irqsave(&conf->device_lock, flags);
1358                 mddev->degraded++;
1359                 spin_unlock_irqrestore(&conf->device_lock, flags);
1360                 /*
1361                  * if recovery is running, make sure it aborts.
1362                  */
1363                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1364         }
1365         set_bit(Blocked, &rdev->flags);
1366         set_bit(Faulty, &rdev->flags);
1367         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1368         printk(KERN_ALERT
1369                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1370                "md/raid10:%s: Operation continuing on %d devices.\n",
1371                mdname(mddev), bdevname(rdev->bdev, b),
1372                mdname(mddev), conf->raid_disks - mddev->degraded);
1373 }
1374
1375 static void print_conf(struct r10conf *conf)
1376 {
1377         int i;
1378         struct mirror_info *tmp;
1379
1380         printk(KERN_DEBUG "RAID10 conf printout:\n");
1381         if (!conf) {
1382                 printk(KERN_DEBUG "(!conf)\n");
1383                 return;
1384         }
1385         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1386                 conf->raid_disks);
1387
1388         for (i = 0; i < conf->raid_disks; i++) {
1389                 char b[BDEVNAME_SIZE];
1390                 tmp = conf->mirrors + i;
1391                 if (tmp->rdev)
1392                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1393                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1394                                 !test_bit(Faulty, &tmp->rdev->flags),
1395                                 bdevname(tmp->rdev->bdev,b));
1396         }
1397 }
1398
1399 static void close_sync(struct r10conf *conf)
1400 {
1401         wait_barrier(conf);
1402         allow_barrier(conf);
1403
1404         mempool_destroy(conf->r10buf_pool);
1405         conf->r10buf_pool = NULL;
1406 }
1407
1408 static int raid10_spare_active(struct mddev *mddev)
1409 {
1410         int i;
1411         struct r10conf *conf = mddev->private;
1412         struct mirror_info *tmp;
1413         int count = 0;
1414         unsigned long flags;
1415
1416         /*
1417          * Find all non-in_sync disks within the RAID10 configuration
1418          * and mark them in_sync
1419          */
1420         for (i = 0; i < conf->raid_disks; i++) {
1421                 tmp = conf->mirrors + i;
1422                 if (tmp->replacement
1423                     && tmp->replacement->recovery_offset == MaxSector
1424                     && !test_bit(Faulty, &tmp->replacement->flags)
1425                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1426                         /* Replacement has just become active */
1427                         if (!tmp->rdev
1428                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1429                                 count++;
1430                         if (tmp->rdev) {
1431                                 /* Replaced device not technically faulty,
1432                                  * but we need to be sure it gets removed
1433                                  * and never re-added.
1434                                  */
1435                                 set_bit(Faulty, &tmp->rdev->flags);
1436                                 sysfs_notify_dirent_safe(
1437                                         tmp->rdev->sysfs_state);
1438                         }
1439                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1440                 } else if (tmp->rdev
1441                            && !test_bit(Faulty, &tmp->rdev->flags)
1442                            && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1443                         count++;
1444                         sysfs_notify_dirent(tmp->rdev->sysfs_state);
1445                 }
1446         }
1447         spin_lock_irqsave(&conf->device_lock, flags);
1448         mddev->degraded -= count;
1449         spin_unlock_irqrestore(&conf->device_lock, flags);
1450
1451         print_conf(conf);
1452         return count;
1453 }
1454
1455
1456 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1457 {
1458         struct r10conf *conf = mddev->private;
1459         int err = -EEXIST;
1460         int mirror;
1461         int first = 0;
1462         int last = conf->raid_disks - 1;
1463
1464         if (mddev->recovery_cp < MaxSector)
1465                 /* only hot-add to in-sync arrays, as recovery is
1466                  * very different from resync
1467                  */
1468                 return -EBUSY;
1469         if (!enough(conf, -1))
1470                 return -EINVAL;
1471
1472         if (rdev->raid_disk >= 0)
1473                 first = last = rdev->raid_disk;
1474
1475         if (rdev->saved_raid_disk >= first &&
1476             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1477                 mirror = rdev->saved_raid_disk;
1478         else
1479                 mirror = first;
1480         for ( ; mirror <= last ; mirror++) {
1481                 struct mirror_info *p = &conf->mirrors[mirror];
1482                 if (p->recovery_disabled == mddev->recovery_disabled)
1483                         continue;
1484                 if (p->rdev)
1485                         continue;
1486
1487                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1488                                   rdev->data_offset << 9);
1489                 /* as we don't honour merge_bvec_fn, we must
1490                  * never risk violating it, so limit
1491                  * ->max_segments to one lying with a single
1492                  * page, as a one page request is never in
1493                  * violation.
1494                  */
1495                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1496                         blk_queue_max_segments(mddev->queue, 1);
1497                         blk_queue_segment_boundary(mddev->queue,
1498                                                    PAGE_CACHE_SIZE - 1);
1499                 }
1500
1501                 p->head_position = 0;
1502                 p->recovery_disabled = mddev->recovery_disabled - 1;
1503                 rdev->raid_disk = mirror;
1504                 err = 0;
1505                 if (rdev->saved_raid_disk != mirror)
1506                         conf->fullsync = 1;
1507                 rcu_assign_pointer(p->rdev, rdev);
1508                 break;
1509         }
1510
1511         md_integrity_add_rdev(rdev, mddev);
1512         print_conf(conf);
1513         return err;
1514 }
1515
1516 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1517 {
1518         struct r10conf *conf = mddev->private;
1519         int err = 0;
1520         int number = rdev->raid_disk;
1521         struct md_rdev **rdevp;
1522         struct mirror_info *p = conf->mirrors + number;
1523
1524         print_conf(conf);
1525         if (rdev == p->rdev)
1526                 rdevp = &p->rdev;
1527         else if (rdev == p->replacement)
1528                 rdevp = &p->replacement;
1529         else
1530                 return 0;
1531
1532         if (test_bit(In_sync, &rdev->flags) ||
1533             atomic_read(&rdev->nr_pending)) {
1534                 err = -EBUSY;
1535                 goto abort;
1536         }
1537         /* Only remove faulty devices if recovery
1538          * is not possible.
1539          */
1540         if (!test_bit(Faulty, &rdev->flags) &&
1541             mddev->recovery_disabled != p->recovery_disabled &&
1542             (!p->replacement || p->replacement == rdev) &&
1543             enough(conf, -1)) {
1544                 err = -EBUSY;
1545                 goto abort;
1546         }
1547         *rdevp = NULL;
1548         synchronize_rcu();
1549         if (atomic_read(&rdev->nr_pending)) {
1550                 /* lost the race, try later */
1551                 err = -EBUSY;
1552                 *rdevp = rdev;
1553                 goto abort;
1554         } else if (p->replacement) {
1555                 /* We must have just cleared 'rdev' */
1556                 p->rdev = p->replacement;
1557                 clear_bit(Replacement, &p->replacement->flags);
1558                 smp_mb(); /* Make sure other CPUs may see both as identical
1559                            * but will never see neither -- if they are careful.
1560                            */
1561                 p->replacement = NULL;
1562                 clear_bit(WantReplacement, &rdev->flags);
1563         } else
1564                 /* We might have just remove the Replacement as faulty
1565                  * Clear the flag just in case
1566                  */
1567                 clear_bit(WantReplacement, &rdev->flags);
1568
1569         err = md_integrity_register(mddev);
1570
1571 abort:
1572
1573         print_conf(conf);
1574         return err;
1575 }
1576
1577
1578 static void end_sync_read(struct bio *bio, int error)
1579 {
1580         struct r10bio *r10_bio = bio->bi_private;
1581         struct r10conf *conf = r10_bio->mddev->private;
1582         int d;
1583
1584         d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1585
1586         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1587                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1588         else
1589                 /* The write handler will notice the lack of
1590                  * R10BIO_Uptodate and record any errors etc
1591                  */
1592                 atomic_add(r10_bio->sectors,
1593                            &conf->mirrors[d].rdev->corrected_errors);
1594
1595         /* for reconstruct, we always reschedule after a read.
1596          * for resync, only after all reads
1597          */
1598         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1599         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1600             atomic_dec_and_test(&r10_bio->remaining)) {
1601                 /* we have read all the blocks,
1602                  * do the comparison in process context in raid10d
1603                  */
1604                 reschedule_retry(r10_bio);
1605         }
1606 }
1607
1608 static void end_sync_request(struct r10bio *r10_bio)
1609 {
1610         struct mddev *mddev = r10_bio->mddev;
1611
1612         while (atomic_dec_and_test(&r10_bio->remaining)) {
1613                 if (r10_bio->master_bio == NULL) {
1614                         /* the primary of several recovery bios */
1615                         sector_t s = r10_bio->sectors;
1616                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1617                             test_bit(R10BIO_WriteError, &r10_bio->state))
1618                                 reschedule_retry(r10_bio);
1619                         else
1620                                 put_buf(r10_bio);
1621                         md_done_sync(mddev, s, 1);
1622                         break;
1623                 } else {
1624                         struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1625                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1626                             test_bit(R10BIO_WriteError, &r10_bio->state))
1627                                 reschedule_retry(r10_bio);
1628                         else
1629                                 put_buf(r10_bio);
1630                         r10_bio = r10_bio2;
1631                 }
1632         }
1633 }
1634
1635 static void end_sync_write(struct bio *bio, int error)
1636 {
1637         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1638         struct r10bio *r10_bio = bio->bi_private;
1639         struct mddev *mddev = r10_bio->mddev;
1640         struct r10conf *conf = mddev->private;
1641         int d;
1642         sector_t first_bad;
1643         int bad_sectors;
1644         int slot;
1645         int repl;
1646         struct md_rdev *rdev = NULL;
1647
1648         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1649         if (repl)
1650                 rdev = conf->mirrors[d].replacement;
1651         if (!rdev) {
1652                 smp_mb();
1653                 rdev = conf->mirrors[d].rdev;
1654         }
1655
1656         if (!uptodate) {
1657                 if (repl)
1658                         md_error(mddev, rdev);
1659                 else {
1660                         set_bit(WriteErrorSeen, &rdev->flags);
1661                         set_bit(R10BIO_WriteError, &r10_bio->state);
1662                 }
1663         } else if (is_badblock(rdev,
1664                              r10_bio->devs[slot].addr,
1665                              r10_bio->sectors,
1666                              &first_bad, &bad_sectors))
1667                 set_bit(R10BIO_MadeGood, &r10_bio->state);
1668
1669         rdev_dec_pending(rdev, mddev);
1670
1671         end_sync_request(r10_bio);
1672 }
1673
1674 /*
1675  * Note: sync and recover and handled very differently for raid10
1676  * This code is for resync.
1677  * For resync, we read through virtual addresses and read all blocks.
1678  * If there is any error, we schedule a write.  The lowest numbered
1679  * drive is authoritative.
1680  * However requests come for physical address, so we need to map.
1681  * For every physical address there are raid_disks/copies virtual addresses,
1682  * which is always are least one, but is not necessarly an integer.
1683  * This means that a physical address can span multiple chunks, so we may
1684  * have to submit multiple io requests for a single sync request.
1685  */
1686 /*
1687  * We check if all blocks are in-sync and only write to blocks that
1688  * aren't in sync
1689  */
1690 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1691 {
1692         struct r10conf *conf = mddev->private;
1693         int i, first;
1694         struct bio *tbio, *fbio;
1695
1696         atomic_set(&r10_bio->remaining, 1);
1697
1698         /* find the first device with a block */
1699         for (i=0; i<conf->copies; i++)
1700                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1701                         break;
1702
1703         if (i == conf->copies)
1704                 goto done;
1705
1706         first = i;
1707         fbio = r10_bio->devs[i].bio;
1708
1709         /* now find blocks with errors */
1710         for (i=0 ; i < conf->copies ; i++) {
1711                 int  j, d;
1712                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1713
1714                 tbio = r10_bio->devs[i].bio;
1715
1716                 if (tbio->bi_end_io != end_sync_read)
1717                         continue;
1718                 if (i == first)
1719                         continue;
1720                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1721                         /* We know that the bi_io_vec layout is the same for
1722                          * both 'first' and 'i', so we just compare them.
1723                          * All vec entries are PAGE_SIZE;
1724                          */
1725                         for (j = 0; j < vcnt; j++)
1726                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1727                                            page_address(tbio->bi_io_vec[j].bv_page),
1728                                            PAGE_SIZE))
1729                                         break;
1730                         if (j == vcnt)
1731                                 continue;
1732                         mddev->resync_mismatches += r10_bio->sectors;
1733                         if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1734                                 /* Don't fix anything. */
1735                                 continue;
1736                 }
1737                 /* Ok, we need to write this bio, either to correct an
1738                  * inconsistency or to correct an unreadable block.
1739                  * First we need to fixup bv_offset, bv_len and
1740                  * bi_vecs, as the read request might have corrupted these
1741                  */
1742                 tbio->bi_vcnt = vcnt;
1743                 tbio->bi_size = r10_bio->sectors << 9;
1744                 tbio->bi_idx = 0;
1745                 tbio->bi_phys_segments = 0;
1746                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1747                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1748                 tbio->bi_next = NULL;
1749                 tbio->bi_rw = WRITE;
1750                 tbio->bi_private = r10_bio;
1751                 tbio->bi_sector = r10_bio->devs[i].addr;
1752
1753                 for (j=0; j < vcnt ; j++) {
1754                         tbio->bi_io_vec[j].bv_offset = 0;
1755                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1756
1757                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1758                                page_address(fbio->bi_io_vec[j].bv_page),
1759                                PAGE_SIZE);
1760                 }
1761                 tbio->bi_end_io = end_sync_write;
1762
1763                 d = r10_bio->devs[i].devnum;
1764                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1765                 atomic_inc(&r10_bio->remaining);
1766                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1767
1768                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1769                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1770                 generic_make_request(tbio);
1771         }
1772
1773         /* Now write out to any replacement devices
1774          * that are active
1775          */
1776         for (i = 0; i < conf->copies; i++) {
1777                 int j, d;
1778                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1779
1780                 tbio = r10_bio->devs[i].repl_bio;
1781                 if (!tbio || !tbio->bi_end_io)
1782                         continue;
1783                 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
1784                     && r10_bio->devs[i].bio != fbio)
1785                         for (j = 0; j < vcnt; j++)
1786                                 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1787                                        page_address(fbio->bi_io_vec[j].bv_page),
1788                                        PAGE_SIZE);
1789                 d = r10_bio->devs[i].devnum;
1790                 atomic_inc(&r10_bio->remaining);
1791                 md_sync_acct(conf->mirrors[d].replacement->bdev,
1792                              tbio->bi_size >> 9);
1793                 generic_make_request(tbio);
1794         }
1795
1796 done:
1797         if (atomic_dec_and_test(&r10_bio->remaining)) {
1798                 md_done_sync(mddev, r10_bio->sectors, 1);
1799                 put_buf(r10_bio);
1800         }
1801 }
1802
1803 /*
1804  * Now for the recovery code.
1805  * Recovery happens across physical sectors.
1806  * We recover all non-is_sync drives by finding the virtual address of
1807  * each, and then choose a working drive that also has that virt address.
1808  * There is a separate r10_bio for each non-in_sync drive.
1809  * Only the first two slots are in use. The first for reading,
1810  * The second for writing.
1811  *
1812  */
1813 static void fix_recovery_read_error(struct r10bio *r10_bio)
1814 {
1815         /* We got a read error during recovery.
1816          * We repeat the read in smaller page-sized sections.
1817          * If a read succeeds, write it to the new device or record
1818          * a bad block if we cannot.
1819          * If a read fails, record a bad block on both old and
1820          * new devices.
1821          */
1822         struct mddev *mddev = r10_bio->mddev;
1823         struct r10conf *conf = mddev->private;
1824         struct bio *bio = r10_bio->devs[0].bio;
1825         sector_t sect = 0;
1826         int sectors = r10_bio->sectors;
1827         int idx = 0;
1828         int dr = r10_bio->devs[0].devnum;
1829         int dw = r10_bio->devs[1].devnum;
1830
1831         while (sectors) {
1832                 int s = sectors;
1833                 struct md_rdev *rdev;
1834                 sector_t addr;
1835                 int ok;
1836
1837                 if (s > (PAGE_SIZE>>9))
1838                         s = PAGE_SIZE >> 9;
1839
1840                 rdev = conf->mirrors[dr].rdev;
1841                 addr = r10_bio->devs[0].addr + sect,
1842                 ok = sync_page_io(rdev,
1843                                   addr,
1844                                   s << 9,
1845                                   bio->bi_io_vec[idx].bv_page,
1846                                   READ, false);
1847                 if (ok) {
1848                         rdev = conf->mirrors[dw].rdev;
1849                         addr = r10_bio->devs[1].addr + sect;
1850                         ok = sync_page_io(rdev,
1851                                           addr,
1852                                           s << 9,
1853                                           bio->bi_io_vec[idx].bv_page,
1854                                           WRITE, false);
1855                         if (!ok)
1856                                 set_bit(WriteErrorSeen, &rdev->flags);
1857                 }
1858                 if (!ok) {
1859                         /* We don't worry if we cannot set a bad block -
1860                          * it really is bad so there is no loss in not
1861                          * recording it yet
1862                          */
1863                         rdev_set_badblocks(rdev, addr, s, 0);
1864
1865                         if (rdev != conf->mirrors[dw].rdev) {
1866                                 /* need bad block on destination too */
1867                                 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
1868                                 addr = r10_bio->devs[1].addr + sect;
1869                                 ok = rdev_set_badblocks(rdev2, addr, s, 0);
1870                                 if (!ok) {
1871                                         /* just abort the recovery */
1872                                         printk(KERN_NOTICE
1873                                                "md/raid10:%s: recovery aborted"
1874                                                " due to read error\n",
1875                                                mdname(mddev));
1876
1877                                         conf->mirrors[dw].recovery_disabled
1878                                                 = mddev->recovery_disabled;
1879                                         set_bit(MD_RECOVERY_INTR,
1880                                                 &mddev->recovery);
1881                                         break;
1882                                 }
1883                         }
1884                 }
1885
1886                 sectors -= s;
1887                 sect += s;
1888                 idx++;
1889         }
1890 }
1891
1892 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1893 {
1894         struct r10conf *conf = mddev->private;
1895         int d;
1896         struct bio *wbio, *wbio2;
1897
1898         if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
1899                 fix_recovery_read_error(r10_bio);
1900                 end_sync_request(r10_bio);
1901                 return;
1902         }
1903
1904         /*
1905          * share the pages with the first bio
1906          * and submit the write request
1907          */
1908         d = r10_bio->devs[1].devnum;
1909         wbio = r10_bio->devs[1].bio;
1910         wbio2 = r10_bio->devs[1].repl_bio;
1911         if (wbio->bi_end_io) {
1912                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1913                 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1914                 generic_make_request(wbio);
1915         }
1916         if (wbio2 && wbio2->bi_end_io) {
1917                 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
1918                 md_sync_acct(conf->mirrors[d].replacement->bdev,
1919                              wbio2->bi_size >> 9);
1920                 generic_make_request(wbio2);
1921         }
1922 }
1923
1924
1925 /*
1926  * Used by fix_read_error() to decay the per rdev read_errors.
1927  * We halve the read error count for every hour that has elapsed
1928  * since the last recorded read error.
1929  *
1930  */
1931 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
1932 {
1933         struct timespec cur_time_mon;
1934         unsigned long hours_since_last;
1935         unsigned int read_errors = atomic_read(&rdev->read_errors);
1936
1937         ktime_get_ts(&cur_time_mon);
1938
1939         if (rdev->last_read_error.tv_sec == 0 &&
1940             rdev->last_read_error.tv_nsec == 0) {
1941                 /* first time we've seen a read error */
1942                 rdev->last_read_error = cur_time_mon;
1943                 return;
1944         }
1945
1946         hours_since_last = (cur_time_mon.tv_sec -
1947                             rdev->last_read_error.tv_sec) / 3600;
1948
1949         rdev->last_read_error = cur_time_mon;
1950
1951         /*
1952          * if hours_since_last is > the number of bits in read_errors
1953          * just set read errors to 0. We do this to avoid
1954          * overflowing the shift of read_errors by hours_since_last.
1955          */
1956         if (hours_since_last >= 8 * sizeof(read_errors))
1957                 atomic_set(&rdev->read_errors, 0);
1958         else
1959                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1960 }
1961
1962 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
1963                             int sectors, struct page *page, int rw)
1964 {
1965         sector_t first_bad;
1966         int bad_sectors;
1967
1968         if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
1969             && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
1970                 return -1;
1971         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1972                 /* success */
1973                 return 1;
1974         if (rw == WRITE)
1975                 set_bit(WriteErrorSeen, &rdev->flags);
1976         /* need to record an error - either for the block or the device */
1977         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1978                 md_error(rdev->mddev, rdev);
1979         return 0;
1980 }
1981
1982 /*
1983  * This is a kernel thread which:
1984  *
1985  *      1.      Retries failed read operations on working mirrors.
1986  *      2.      Updates the raid superblock when problems encounter.
1987  *      3.      Performs writes following reads for array synchronising.
1988  */
1989
1990 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
1991 {
1992         int sect = 0; /* Offset from r10_bio->sector */
1993         int sectors = r10_bio->sectors;
1994         struct md_rdev*rdev;
1995         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1996         int d = r10_bio->devs[r10_bio->read_slot].devnum;
1997
1998         /* still own a reference to this rdev, so it cannot
1999          * have been cleared recently.
2000          */
2001         rdev = conf->mirrors[d].rdev;
2002
2003         if (test_bit(Faulty, &rdev->flags))
2004                 /* drive has already been failed, just ignore any
2005                    more fix_read_error() attempts */
2006                 return;
2007
2008         check_decay_read_errors(mddev, rdev);
2009         atomic_inc(&rdev->read_errors);
2010         if (atomic_read(&rdev->read_errors) > max_read_errors) {
2011                 char b[BDEVNAME_SIZE];
2012                 bdevname(rdev->bdev, b);
2013
2014                 printk(KERN_NOTICE
2015                        "md/raid10:%s: %s: Raid device exceeded "
2016                        "read_error threshold [cur %d:max %d]\n",
2017                        mdname(mddev), b,
2018                        atomic_read(&rdev->read_errors), max_read_errors);
2019                 printk(KERN_NOTICE
2020                        "md/raid10:%s: %s: Failing raid device\n",
2021                        mdname(mddev), b);
2022                 md_error(mddev, conf->mirrors[d].rdev);
2023                 return;
2024         }
2025
2026         while(sectors) {
2027                 int s = sectors;
2028                 int sl = r10_bio->read_slot;
2029                 int success = 0;
2030                 int start;
2031
2032                 if (s > (PAGE_SIZE>>9))
2033                         s = PAGE_SIZE >> 9;
2034
2035                 rcu_read_lock();
2036                 do {
2037                         sector_t first_bad;
2038                         int bad_sectors;
2039
2040                         d = r10_bio->devs[sl].devnum;
2041                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2042                         if (rdev &&
2043                             test_bit(In_sync, &rdev->flags) &&
2044                             is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2045                                         &first_bad, &bad_sectors) == 0) {
2046                                 atomic_inc(&rdev->nr_pending);
2047                                 rcu_read_unlock();
2048                                 success = sync_page_io(rdev,
2049                                                        r10_bio->devs[sl].addr +
2050                                                        sect,
2051                                                        s<<9,
2052                                                        conf->tmppage, READ, false);
2053                                 rdev_dec_pending(rdev, mddev);
2054                                 rcu_read_lock();
2055                                 if (success)
2056                                         break;
2057                         }
2058                         sl++;
2059                         if (sl == conf->copies)
2060                                 sl = 0;
2061                 } while (!success && sl != r10_bio->read_slot);
2062                 rcu_read_unlock();
2063
2064                 if (!success) {
2065                         /* Cannot read from anywhere, just mark the block
2066                          * as bad on the first device to discourage future
2067                          * reads.
2068                          */
2069                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2070                         rdev = conf->mirrors[dn].rdev;
2071
2072                         if (!rdev_set_badblocks(
2073                                     rdev,
2074                                     r10_bio->devs[r10_bio->read_slot].addr
2075                                     + sect,
2076                                     s, 0))
2077                                 md_error(mddev, rdev);
2078                         break;
2079                 }
2080
2081                 start = sl;
2082                 /* write it back and re-read */
2083                 rcu_read_lock();
2084                 while (sl != r10_bio->read_slot) {
2085                         char b[BDEVNAME_SIZE];
2086
2087                         if (sl==0)
2088                                 sl = conf->copies;
2089                         sl--;
2090                         d = r10_bio->devs[sl].devnum;
2091                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2092                         if (!rdev ||
2093                             !test_bit(In_sync, &rdev->flags))
2094                                 continue;
2095
2096                         atomic_inc(&rdev->nr_pending);
2097                         rcu_read_unlock();
2098                         if (r10_sync_page_io(rdev,
2099                                              r10_bio->devs[sl].addr +
2100                                              sect,
2101                                              s<<9, conf->tmppage, WRITE)
2102                             == 0) {
2103                                 /* Well, this device is dead */
2104                                 printk(KERN_NOTICE
2105                                        "md/raid10:%s: read correction "
2106                                        "write failed"
2107                                        " (%d sectors at %llu on %s)\n",
2108                                        mdname(mddev), s,
2109                                        (unsigned long long)(
2110                                                sect + rdev->data_offset),
2111                                        bdevname(rdev->bdev, b));
2112                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2113                                        "drive\n",
2114                                        mdname(mddev),
2115                                        bdevname(rdev->bdev, b));
2116                         }
2117                         rdev_dec_pending(rdev, mddev);
2118                         rcu_read_lock();
2119                 }
2120                 sl = start;
2121                 while (sl != r10_bio->read_slot) {
2122                         char b[BDEVNAME_SIZE];
2123
2124                         if (sl==0)
2125                                 sl = conf->copies;
2126                         sl--;
2127                         d = r10_bio->devs[sl].devnum;
2128                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2129                         if (!rdev ||
2130                             !test_bit(In_sync, &rdev->flags))
2131                                 continue;
2132
2133                         atomic_inc(&rdev->nr_pending);
2134                         rcu_read_unlock();
2135                         switch (r10_sync_page_io(rdev,
2136                                              r10_bio->devs[sl].addr +
2137                                              sect,
2138                                              s<<9, conf->tmppage,
2139                                                  READ)) {
2140                         case 0:
2141                                 /* Well, this device is dead */
2142                                 printk(KERN_NOTICE
2143                                        "md/raid10:%s: unable to read back "
2144                                        "corrected sectors"
2145                                        " (%d sectors at %llu on %s)\n",
2146                                        mdname(mddev), s,
2147                                        (unsigned long long)(
2148                                                sect + rdev->data_offset),
2149                                        bdevname(rdev->bdev, b));
2150                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2151                                        "drive\n",
2152                                        mdname(mddev),
2153                                        bdevname(rdev->bdev, b));
2154                                 break;
2155                         case 1:
2156                                 printk(KERN_INFO
2157                                        "md/raid10:%s: read error corrected"
2158                                        " (%d sectors at %llu on %s)\n",
2159                                        mdname(mddev), s,
2160                                        (unsigned long long)(
2161                                                sect + rdev->data_offset),
2162                                        bdevname(rdev->bdev, b));
2163                                 atomic_add(s, &rdev->corrected_errors);
2164                         }
2165
2166                         rdev_dec_pending(rdev, mddev);
2167                         rcu_read_lock();
2168                 }
2169                 rcu_read_unlock();
2170
2171                 sectors -= s;
2172                 sect += s;
2173         }
2174 }
2175
2176 static void bi_complete(struct bio *bio, int error)
2177 {
2178         complete((struct completion *)bio->bi_private);
2179 }
2180
2181 static int submit_bio_wait(int rw, struct bio *bio)
2182 {
2183         struct completion event;
2184         rw |= REQ_SYNC;
2185
2186         init_completion(&event);
2187         bio->bi_private = &event;
2188         bio->bi_end_io = bi_complete;
2189         submit_bio(rw, bio);
2190         wait_for_completion(&event);
2191
2192         return test_bit(BIO_UPTODATE, &bio->bi_flags);
2193 }
2194
2195 static int narrow_write_error(struct r10bio *r10_bio, int i)
2196 {
2197         struct bio *bio = r10_bio->master_bio;
2198         struct mddev *mddev = r10_bio->mddev;
2199         struct r10conf *conf = mddev->private;
2200         struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2201         /* bio has the data to be written to slot 'i' where
2202          * we just recently had a write error.
2203          * We repeatedly clone the bio and trim down to one block,
2204          * then try the write.  Where the write fails we record
2205          * a bad block.
2206          * It is conceivable that the bio doesn't exactly align with
2207          * blocks.  We must handle this.
2208          *
2209          * We currently own a reference to the rdev.
2210          */
2211
2212         int block_sectors;
2213         sector_t sector;
2214         int sectors;
2215         int sect_to_write = r10_bio->sectors;
2216         int ok = 1;
2217
2218         if (rdev->badblocks.shift < 0)
2219                 return 0;
2220
2221         block_sectors = 1 << rdev->badblocks.shift;
2222         sector = r10_bio->sector;
2223         sectors = ((r10_bio->sector + block_sectors)
2224                    & ~(sector_t)(block_sectors - 1))
2225                 - sector;
2226
2227         while (sect_to_write) {
2228                 struct bio *wbio;
2229                 if (sectors > sect_to_write)
2230                         sectors = sect_to_write;
2231                 /* Write at 'sector' for 'sectors' */
2232                 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2233                 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2234                 wbio->bi_sector = (r10_bio->devs[i].addr+
2235                                    rdev->data_offset+
2236                                    (sector - r10_bio->sector));
2237                 wbio->bi_bdev = rdev->bdev;
2238                 if (submit_bio_wait(WRITE, wbio) == 0)
2239                         /* Failure! */
2240                         ok = rdev_set_badblocks(rdev, sector,
2241                                                 sectors, 0)
2242                                 && ok;
2243
2244                 bio_put(wbio);
2245                 sect_to_write -= sectors;
2246                 sector += sectors;
2247                 sectors = block_sectors;
2248         }
2249         return ok;
2250 }
2251
2252 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2253 {
2254         int slot = r10_bio->read_slot;
2255         struct bio *bio;
2256         struct r10conf *conf = mddev->private;
2257         struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2258         char b[BDEVNAME_SIZE];
2259         unsigned long do_sync;
2260         int max_sectors;
2261
2262         /* we got a read error. Maybe the drive is bad.  Maybe just
2263          * the block and we can fix it.
2264          * We freeze all other IO, and try reading the block from
2265          * other devices.  When we find one, we re-write
2266          * and check it that fixes the read error.
2267          * This is all done synchronously while the array is
2268          * frozen.
2269          */
2270         if (mddev->ro == 0) {
2271                 freeze_array(conf);
2272                 fix_read_error(conf, mddev, r10_bio);
2273                 unfreeze_array(conf);
2274         }
2275         rdev_dec_pending(rdev, mddev);
2276
2277         bio = r10_bio->devs[slot].bio;
2278         bdevname(bio->bi_bdev, b);
2279         r10_bio->devs[slot].bio =
2280                 mddev->ro ? IO_BLOCKED : NULL;
2281 read_more:
2282         rdev = read_balance(conf, r10_bio, &max_sectors);
2283         if (rdev == NULL) {
2284                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2285                        " read error for block %llu\n",
2286                        mdname(mddev), b,
2287                        (unsigned long long)r10_bio->sector);
2288                 raid_end_bio_io(r10_bio);
2289                 bio_put(bio);
2290                 return;
2291         }
2292
2293         do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2294         if (bio)
2295                 bio_put(bio);
2296         slot = r10_bio->read_slot;
2297         printk_ratelimited(
2298                 KERN_ERR
2299                 "md/raid10:%s: %s: redirecting"
2300                 "sector %llu to another mirror\n",
2301                 mdname(mddev),
2302                 bdevname(rdev->bdev, b),
2303                 (unsigned long long)r10_bio->sector);
2304         bio = bio_clone_mddev(r10_bio->master_bio,
2305                               GFP_NOIO, mddev);
2306         md_trim_bio(bio,
2307                     r10_bio->sector - bio->bi_sector,
2308                     max_sectors);
2309         r10_bio->devs[slot].bio = bio;
2310         r10_bio->devs[slot].rdev = rdev;
2311         bio->bi_sector = r10_bio->devs[slot].addr
2312                 + rdev->data_offset;
2313         bio->bi_bdev = rdev->bdev;
2314         bio->bi_rw = READ | do_sync;
2315         bio->bi_private = r10_bio;
2316         bio->bi_end_io = raid10_end_read_request;
2317         if (max_sectors < r10_bio->sectors) {
2318                 /* Drat - have to split this up more */
2319                 struct bio *mbio = r10_bio->master_bio;
2320                 int sectors_handled =
2321                         r10_bio->sector + max_sectors
2322                         - mbio->bi_sector;
2323                 r10_bio->sectors = max_sectors;
2324                 spin_lock_irq(&conf->device_lock);
2325                 if (mbio->bi_phys_segments == 0)
2326                         mbio->bi_phys_segments = 2;
2327                 else
2328                         mbio->bi_phys_segments++;
2329                 spin_unlock_irq(&conf->device_lock);
2330                 generic_make_request(bio);
2331                 bio = NULL;
2332
2333                 r10_bio = mempool_alloc(conf->r10bio_pool,
2334                                         GFP_NOIO);
2335                 r10_bio->master_bio = mbio;
2336                 r10_bio->sectors = (mbio->bi_size >> 9)
2337                         - sectors_handled;
2338                 r10_bio->state = 0;
2339                 set_bit(R10BIO_ReadError,
2340                         &r10_bio->state);
2341                 r10_bio->mddev = mddev;
2342                 r10_bio->sector = mbio->bi_sector
2343                         + sectors_handled;
2344
2345                 goto read_more;
2346         } else
2347                 generic_make_request(bio);
2348 }
2349
2350 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2351 {
2352         /* Some sort of write request has finished and it
2353          * succeeded in writing where we thought there was a
2354          * bad block.  So forget the bad block.
2355          * Or possibly if failed and we need to record
2356          * a bad block.
2357          */
2358         int m;
2359         struct md_rdev *rdev;
2360
2361         if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2362             test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2363                 for (m = 0; m < conf->copies; m++) {
2364                         int dev = r10_bio->devs[m].devnum;
2365                         rdev = conf->mirrors[dev].rdev;
2366                         if (r10_bio->devs[m].bio == NULL)
2367                                 continue;
2368                         if (test_bit(BIO_UPTODATE,
2369                                      &r10_bio->devs[m].bio->bi_flags)) {
2370                                 rdev_clear_badblocks(
2371                                         rdev,
2372                                         r10_bio->devs[m].addr,
2373                                         r10_bio->sectors);
2374                         } else {
2375                                 if (!rdev_set_badblocks(
2376                                             rdev,
2377                                             r10_bio->devs[m].addr,
2378                                             r10_bio->sectors, 0))
2379                                         md_error(conf->mddev, rdev);
2380                         }
2381                         rdev = conf->mirrors[dev].replacement;
2382                         if (r10_bio->devs[m].repl_bio == NULL)
2383                                 continue;
2384                         if (test_bit(BIO_UPTODATE,
2385                                      &r10_bio->devs[m].repl_bio->bi_flags)) {
2386                                 rdev_clear_badblocks(
2387                                         rdev,
2388                                         r10_bio->devs[m].addr,
2389                                         r10_bio->sectors);
2390                         } else {
2391                                 if (!rdev_set_badblocks(
2392                                             rdev,
2393                                             r10_bio->devs[m].addr,
2394                                             r10_bio->sectors, 0))
2395                                         md_error(conf->mddev, rdev);
2396                         }
2397                 }
2398                 put_buf(r10_bio);
2399         } else {
2400                 for (m = 0; m < conf->copies; m++) {
2401                         int dev = r10_bio->devs[m].devnum;
2402                         struct bio *bio = r10_bio->devs[m].bio;
2403                         rdev = conf->mirrors[dev].rdev;
2404                         if (bio == IO_MADE_GOOD) {
2405                                 rdev_clear_badblocks(
2406                                         rdev,
2407                                         r10_bio->devs[m].addr,
2408                                         r10_bio->sectors);
2409                                 rdev_dec_pending(rdev, conf->mddev);
2410                         } else if (bio != NULL &&
2411                                    !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2412                                 if (!narrow_write_error(r10_bio, m)) {
2413                                         md_error(conf->mddev, rdev);
2414                                         set_bit(R10BIO_Degraded,
2415                                                 &r10_bio->state);
2416                                 }
2417                                 rdev_dec_pending(rdev, conf->mddev);
2418                         }
2419                         bio = r10_bio->devs[m].repl_bio;
2420                         rdev = conf->mirrors[dev].replacement;
2421                         if (rdev && bio == IO_MADE_GOOD) {
2422                                 rdev_clear_badblocks(
2423                                         rdev,
2424                                         r10_bio->devs[m].addr,
2425                                         r10_bio->sectors);
2426                                 rdev_dec_pending(rdev, conf->mddev);
2427                         }
2428                 }
2429                 if (test_bit(R10BIO_WriteError,
2430                              &r10_bio->state))
2431                         close_write(r10_bio);
2432                 raid_end_bio_io(r10_bio);
2433         }
2434 }
2435
2436 static void raid10d(struct mddev *mddev)
2437 {
2438         struct r10bio *r10_bio;
2439         unsigned long flags;
2440         struct r10conf *conf = mddev->private;
2441         struct list_head *head = &conf->retry_list;
2442         struct blk_plug plug;
2443
2444         md_check_recovery(mddev);
2445
2446         blk_start_plug(&plug);
2447         for (;;) {
2448
2449                 flush_pending_writes(conf);
2450
2451                 spin_lock_irqsave(&conf->device_lock, flags);
2452                 if (list_empty(head)) {
2453                         spin_unlock_irqrestore(&conf->device_lock, flags);
2454                         break;
2455                 }
2456                 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2457                 list_del(head->prev);
2458                 conf->nr_queued--;
2459                 spin_unlock_irqrestore(&conf->device_lock, flags);
2460
2461                 mddev = r10_bio->mddev;
2462                 conf = mddev->private;
2463                 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2464                     test_bit(R10BIO_WriteError, &r10_bio->state))
2465                         handle_write_completed(conf, r10_bio);
2466                 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2467                         sync_request_write(mddev, r10_bio);
2468                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2469                         recovery_request_write(mddev, r10_bio);
2470                 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2471                         handle_read_error(mddev, r10_bio);
2472                 else {
2473                         /* just a partial read to be scheduled from a
2474                          * separate context
2475                          */
2476                         int slot = r10_bio->read_slot;
2477                         generic_make_request(r10_bio->devs[slot].bio);
2478                 }
2479
2480                 cond_resched();
2481                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2482                         md_check_recovery(mddev);
2483         }
2484         blk_finish_plug(&plug);
2485 }
2486
2487
2488 static int init_resync(struct r10conf *conf)
2489 {
2490         int buffs;
2491         int i;
2492
2493         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2494         BUG_ON(conf->r10buf_pool);
2495         conf->have_replacement = 0;
2496         for (i = 0; i < conf->raid_disks; i++)
2497                 if (conf->mirrors[i].replacement)
2498                         conf->have_replacement = 1;
2499         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2500         if (!conf->r10buf_pool)
2501                 return -ENOMEM;
2502         conf->next_resync = 0;
2503         return 0;
2504 }
2505
2506 /*
2507  * perform a "sync" on one "block"
2508  *
2509  * We need to make sure that no normal I/O request - particularly write
2510  * requests - conflict with active sync requests.
2511  *
2512  * This is achieved by tracking pending requests and a 'barrier' concept
2513  * that can be installed to exclude normal IO requests.
2514  *
2515  * Resync and recovery are handled very differently.
2516  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2517  *
2518  * For resync, we iterate over virtual addresses, read all copies,
2519  * and update if there are differences.  If only one copy is live,
2520  * skip it.
2521  * For recovery, we iterate over physical addresses, read a good
2522  * value for each non-in_sync drive, and over-write.
2523  *
2524  * So, for recovery we may have several outstanding complex requests for a
2525  * given address, one for each out-of-sync device.  We model this by allocating
2526  * a number of r10_bio structures, one for each out-of-sync device.
2527  * As we setup these structures, we collect all bio's together into a list
2528  * which we then process collectively to add pages, and then process again
2529  * to pass to generic_make_request.
2530  *
2531  * The r10_bio structures are linked using a borrowed master_bio pointer.
2532  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2533  * has its remaining count decremented to 0, the whole complex operation
2534  * is complete.
2535  *
2536  */
2537
2538 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2539                              int *skipped, int go_faster)
2540 {
2541         struct r10conf *conf = mddev->private;
2542         struct r10bio *r10_bio;
2543         struct bio *biolist = NULL, *bio;
2544         sector_t max_sector, nr_sectors;
2545         int i;
2546         int max_sync;
2547         sector_t sync_blocks;
2548         sector_t sectors_skipped = 0;
2549         int chunks_skipped = 0;
2550
2551         if (!conf->r10buf_pool)
2552                 if (init_resync(conf))
2553                         return 0;
2554
2555  skipped:
2556         max_sector = mddev->dev_sectors;
2557         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2558                 max_sector = mddev->resync_max_sectors;
2559         if (sector_nr >= max_sector) {
2560                 /* If we aborted, we need to abort the
2561                  * sync on the 'current' bitmap chucks (there can
2562                  * be several when recovering multiple devices).
2563                  * as we may have started syncing it but not finished.
2564                  * We can find the current address in
2565                  * mddev->curr_resync, but for recovery,
2566                  * we need to convert that to several
2567                  * virtual addresses.
2568                  */
2569                 if (mddev->curr_resync < max_sector) { /* aborted */
2570                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2571                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2572                                                 &sync_blocks, 1);
2573                         else for (i=0; i<conf->raid_disks; i++) {
2574                                 sector_t sect =
2575                                         raid10_find_virt(conf, mddev->curr_resync, i);
2576                                 bitmap_end_sync(mddev->bitmap, sect,
2577                                                 &sync_blocks, 1);
2578                         }
2579                 } else {
2580                         /* completed sync */
2581                         if ((!mddev->bitmap || conf->fullsync)
2582                             && conf->have_replacement
2583                             && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2584                                 /* Completed a full sync so the replacements
2585                                  * are now fully recovered.
2586                                  */
2587                                 for (i = 0; i < conf->raid_disks; i++)
2588                                         if (conf->mirrors[i].replacement)
2589                                                 conf->mirrors[i].replacement
2590                                                         ->recovery_offset
2591                                                         = MaxSector;
2592                         }
2593                         conf->fullsync = 0;
2594                 }
2595                 bitmap_close_sync(mddev->bitmap);
2596                 close_sync(conf);
2597                 *skipped = 1;
2598                 return sectors_skipped;
2599         }
2600         if (chunks_skipped >= conf->raid_disks) {
2601                 /* if there has been nothing to do on any drive,
2602                  * then there is nothing to do at all..
2603                  */
2604                 *skipped = 1;
2605                 return (max_sector - sector_nr) + sectors_skipped;
2606         }
2607
2608         if (max_sector > mddev->resync_max)
2609                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2610
2611         /* make sure whole request will fit in a chunk - if chunks
2612          * are meaningful
2613          */
2614         if (conf->near_copies < conf->raid_disks &&
2615             max_sector > (sector_nr | conf->chunk_mask))
2616                 max_sector = (sector_nr | conf->chunk_mask) + 1;
2617         /*
2618          * If there is non-resync activity waiting for us then
2619          * put in a delay to throttle resync.
2620          */
2621         if (!go_faster && conf->nr_waiting)
2622                 msleep_interruptible(1000);
2623
2624         /* Again, very different code for resync and recovery.
2625          * Both must result in an r10bio with a list of bios that
2626          * have bi_end_io, bi_sector, bi_bdev set,
2627          * and bi_private set to the r10bio.
2628          * For recovery, we may actually create several r10bios
2629          * with 2 bios in each, that correspond to the bios in the main one.
2630          * In this case, the subordinate r10bios link back through a
2631          * borrowed master_bio pointer, and the counter in the master
2632          * includes a ref from each subordinate.
2633          */
2634         /* First, we decide what to do and set ->bi_end_io
2635          * To end_sync_read if we want to read, and
2636          * end_sync_write if we will want to write.
2637          */
2638
2639         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2640         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2641                 /* recovery... the complicated one */
2642                 int j;
2643                 r10_bio = NULL;
2644
2645                 for (i=0 ; i<conf->raid_disks; i++) {
2646                         int still_degraded;
2647                         struct r10bio *rb2;
2648                         sector_t sect;
2649                         int must_sync;
2650                         int any_working;
2651                         struct mirror_info *mirror = &conf->mirrors[i];
2652
2653                         if ((mirror->rdev == NULL ||
2654                              test_bit(In_sync, &mirror->rdev->flags))
2655                             &&
2656                             (mirror->replacement == NULL ||
2657                              test_bit(Faulty,
2658                                       &mirror->replacement->flags)))
2659                                 continue;
2660
2661                         still_degraded = 0;
2662                         /* want to reconstruct this device */
2663                         rb2 = r10_bio;
2664                         sect = raid10_find_virt(conf, sector_nr, i);
2665                         /* Unless we are doing a full sync, or a replacement
2666                          * we only need to recover the block if it is set in
2667                          * the bitmap
2668                          */
2669                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2670                                                       &sync_blocks, 1);
2671                         if (sync_blocks < max_sync)
2672                                 max_sync = sync_blocks;
2673                         if (!must_sync &&
2674                             mirror->replacement == NULL &&
2675                             !conf->fullsync) {
2676                                 /* yep, skip the sync_blocks here, but don't assume
2677                                  * that there will never be anything to do here
2678                                  */
2679                                 chunks_skipped = -1;
2680                                 continue;
2681                         }
2682
2683                         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2684                         raise_barrier(conf, rb2 != NULL);
2685                         atomic_set(&r10_bio->remaining, 0);
2686
2687                         r10_bio->master_bio = (struct bio*)rb2;
2688                         if (rb2)
2689                                 atomic_inc(&rb2->remaining);
2690                         r10_bio->mddev = mddev;
2691                         set_bit(R10BIO_IsRecover, &r10_bio->state);
2692                         r10_bio->sector = sect;
2693
2694                         raid10_find_phys(conf, r10_bio);
2695
2696                         /* Need to check if the array will still be
2697                          * degraded
2698                          */
2699                         for (j=0; j<conf->raid_disks; j++)
2700                                 if (conf->mirrors[j].rdev == NULL ||
2701                                     test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2702                                         still_degraded = 1;
2703                                         break;
2704                                 }
2705
2706                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2707                                                       &sync_blocks, still_degraded);
2708
2709                         any_working = 0;
2710                         for (j=0; j<conf->copies;j++) {
2711                                 int k;
2712                                 int d = r10_bio->devs[j].devnum;
2713                                 sector_t from_addr, to_addr;
2714                                 struct md_rdev *rdev;
2715                                 sector_t sector, first_bad;
2716                                 int bad_sectors;
2717                                 if (!conf->mirrors[d].rdev ||
2718                                     !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2719                                         continue;
2720                                 /* This is where we read from */
2721                                 any_working = 1;
2722                                 rdev = conf->mirrors[d].rdev;
2723                                 sector = r10_bio->devs[j].addr;
2724
2725                                 if (is_badblock(rdev, sector, max_sync,
2726                                                 &first_bad, &bad_sectors)) {
2727                                         if (first_bad > sector)
2728                                                 max_sync = first_bad - sector;
2729                                         else {
2730                                                 bad_sectors -= (sector
2731                                                                 - first_bad);
2732                                                 if (max_sync > bad_sectors)
2733                                                         max_sync = bad_sectors;
2734                                                 continue;
2735                                         }
2736                                 }
2737                                 bio = r10_bio->devs[0].bio;
2738                                 bio->bi_next = biolist;
2739                                 biolist = bio;
2740                                 bio->bi_private = r10_bio;
2741                                 bio->bi_end_io = end_sync_read;
2742                                 bio->bi_rw = READ;
2743                                 from_addr = r10_bio->devs[j].addr;
2744                                 bio->bi_sector = from_addr + rdev->data_offset;
2745                                 bio->bi_bdev = rdev->bdev;
2746                                 atomic_inc(&rdev->nr_pending);
2747                                 /* and we write to 'i' (if not in_sync) */
2748
2749                                 for (k=0; k<conf->copies; k++)
2750                                         if (r10_bio->devs[k].devnum == i)
2751                                                 break;
2752                                 BUG_ON(k == conf->copies);
2753                                 to_addr = r10_bio->devs[k].addr;
2754                                 r10_bio->devs[0].devnum = d;
2755                                 r10_bio->devs[0].addr = from_addr;
2756                                 r10_bio->devs[1].devnum = i;
2757                                 r10_bio->devs[1].addr = to_addr;
2758
2759                                 rdev = mirror->rdev;
2760                                 if (!test_bit(In_sync, &rdev->flags)) {
2761                                         bio = r10_bio->devs[1].bio;
2762                                         bio->bi_next = biolist;
2763                                         biolist = bio;
2764                                         bio->bi_private = r10_bio;
2765                                         bio->bi_end_io = end_sync_write;
2766                                         bio->bi_rw = WRITE;
2767                                         bio->bi_sector = to_addr
2768                                                 + rdev->data_offset;
2769                                         bio->bi_bdev = rdev->bdev;
2770                                         atomic_inc(&r10_bio->remaining);
2771                                 } else
2772                                         r10_bio->devs[1].bio->bi_end_io = NULL;
2773
2774                                 /* and maybe write to replacement */
2775                                 bio = r10_bio->devs[1].repl_bio;
2776                                 if (bio)
2777                                         bio->bi_end_io = NULL;
2778                                 rdev = mirror->replacement;
2779                                 /* Note: if rdev != NULL, then bio
2780                                  * cannot be NULL as r10buf_pool_alloc will
2781                                  * have allocated it.
2782                                  * So the second test here is pointless.
2783                                  * But it keeps semantic-checkers happy, and
2784                                  * this comment keeps human reviewers
2785                                  * happy.
2786                                  */
2787                                 if (rdev == NULL || bio == NULL ||
2788                                     test_bit(Faulty, &rdev->flags))
2789                                         break;
2790                                 bio->bi_next = biolist;
2791                                 biolist = bio;
2792                                 bio->bi_private = r10_bio;
2793                                 bio->bi_end_io = end_sync_write;
2794                                 bio->bi_rw = WRITE;
2795                                 bio->bi_sector = to_addr + rdev->data_offset;
2796                                 bio->bi_bdev = rdev->bdev;
2797                                 atomic_inc(&r10_bio->remaining);
2798                                 break;
2799                         }
2800                         if (j == conf->copies) {
2801                                 /* Cannot recover, so abort the recovery or
2802                                  * record a bad block */
2803                                 put_buf(r10_bio);
2804                                 if (rb2)
2805                                         atomic_dec(&rb2->remaining);
2806                                 r10_bio = rb2;
2807                                 if (any_working) {
2808                                         /* problem is that there are bad blocks
2809                                          * on other device(s)
2810                                          */
2811                                         int k;
2812                                         for (k = 0; k < conf->copies; k++)
2813                                                 if (r10_bio->devs[k].devnum == i)
2814                                                         break;
2815                                         if (!test_bit(In_sync,
2816                                                       &mirror->rdev->flags)
2817                                             && !rdev_set_badblocks(
2818                                                     mirror->rdev,
2819                                                     r10_bio->devs[k].addr,
2820                                                     max_sync, 0))
2821                                                 any_working = 0;
2822                                         if (mirror->replacement &&
2823                                             !rdev_set_badblocks(
2824                                                     mirror->replacement,
2825                                                     r10_bio->devs[k].addr,
2826                                                     max_sync, 0))
2827                                                 any_working = 0;
2828                                 }
2829                                 if (!any_working)  {
2830                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
2831                                                               &mddev->recovery))
2832                                                 printk(KERN_INFO "md/raid10:%s: insufficient "
2833                                                        "working devices for recovery.\n",
2834                                                        mdname(mddev));
2835                                         mirror->recovery_disabled
2836                                                 = mddev->recovery_disabled;
2837                                 }
2838                                 break;
2839                         }
2840                 }
2841                 if (biolist == NULL) {
2842                         while (r10_bio) {
2843                                 struct r10bio *rb2 = r10_bio;
2844                                 r10_bio = (struct r10bio*) rb2->master_bio;
2845                                 rb2->master_bio = NULL;
2846                                 put_buf(rb2);
2847                         }
2848                         goto giveup;
2849                 }
2850         } else {
2851                 /* resync. Schedule a read for every block at this virt offset */
2852                 int count = 0;
2853
2854                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2855
2856                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2857                                        &sync_blocks, mddev->degraded) &&
2858                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
2859                                                  &mddev->recovery)) {
2860                         /* We can skip this block */
2861                         *skipped = 1;
2862                         return sync_blocks + sectors_skipped;
2863                 }
2864                 if (sync_blocks < max_sync)
2865                         max_sync = sync_blocks;
2866                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2867
2868                 r10_bio->mddev = mddev;
2869                 atomic_set(&r10_bio->remaining, 0);
2870                 raise_barrier(conf, 0);
2871                 conf->next_resync = sector_nr;
2872
2873                 r10_bio->master_bio = NULL;
2874                 r10_bio->sector = sector_nr;
2875                 set_bit(R10BIO_IsSync, &r10_bio->state);
2876                 raid10_find_phys(conf, r10_bio);
2877                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2878
2879                 for (i=0; i<conf->copies; i++) {
2880                         int d = r10_bio->devs[i].devnum;
2881                         sector_t first_bad, sector;
2882                         int bad_sectors;
2883
2884                         if (r10_bio->devs[i].repl_bio)
2885                                 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
2886
2887                         bio = r10_bio->devs[i].bio;
2888                         bio->bi_end_io = NULL;
2889                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
2890                         if (conf->mirrors[d].rdev == NULL ||
2891                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2892                                 continue;
2893                         sector = r10_bio->devs[i].addr;
2894                         if (is_badblock(conf->mirrors[d].rdev,
2895                                         sector, max_sync,
2896                                         &first_bad, &bad_sectors)) {
2897                                 if (first_bad > sector)
2898                                         max_sync = first_bad - sector;
2899                                 else {
2900                                         bad_sectors -= (sector - first_bad);
2901                                         if (max_sync > bad_sectors)
2902                                                 max_sync = max_sync;
2903                                         continue;
2904                                 }
2905                         }
2906                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2907                         atomic_inc(&r10_bio->remaining);
2908                         bio->bi_next = biolist;
2909                         biolist = bio;
2910                         bio->bi_private = r10_bio;
2911                         bio->bi_end_io = end_sync_read;
2912                         bio->bi_rw = READ;
2913                         bio->bi_sector = sector +
2914                                 conf->mirrors[d].rdev->data_offset;
2915                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2916                         count++;
2917
2918                         if (conf->mirrors[d].replacement == NULL ||
2919                             test_bit(Faulty,
2920                                      &conf->mirrors[d].replacement->flags))
2921                                 continue;
2922
2923                         /* Need to set up for writing to the replacement */
2924                         bio = r10_bio->devs[i].repl_bio;
2925                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
2926
2927                         sector = r10_bio->devs[i].addr;
2928                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2929                         bio->bi_next = biolist;
2930                         biolist = bio;
2931                         bio->bi_private = r10_bio;
2932                         bio->bi_end_io = end_sync_write;
2933                         bio->bi_rw = WRITE;
2934                         bio->bi_sector = sector +
2935                                 conf->mirrors[d].replacement->data_offset;
2936                         bio->bi_bdev = conf->mirrors[d].replacement->bdev;
2937                         count++;
2938                 }
2939
2940                 if (count < 2) {
2941                         for (i=0; i<conf->copies; i++) {
2942                                 int d = r10_bio->devs[i].devnum;
2943                                 if (r10_bio->devs[i].bio->bi_end_io)
2944                                         rdev_dec_pending(conf->mirrors[d].rdev,
2945                                                          mddev);
2946                                 if (r10_bio->devs[i].repl_bio &&
2947                                     r10_bio->devs[i].repl_bio->bi_end_io)
2948                                         rdev_dec_pending(
2949                                                 conf->mirrors[d].replacement,
2950                                                 mddev);
2951                         }
2952                         put_buf(r10_bio);
2953                         biolist = NULL;
2954                         goto giveup;
2955                 }
2956         }
2957
2958         for (bio = biolist; bio ; bio=bio->bi_next) {
2959
2960                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2961                 if (bio->bi_end_io)
2962                         bio->bi_flags |= 1 << BIO_UPTODATE;
2963                 bio->bi_vcnt = 0;
2964                 bio->bi_idx = 0;
2965                 bio->bi_phys_segments = 0;
2966                 bio->bi_size = 0;
2967         }
2968
2969         nr_sectors = 0;
2970         if (sector_nr + max_sync < max_sector)
2971                 max_sector = sector_nr + max_sync;
2972         do {
2973                 struct page *page;
2974                 int len = PAGE_SIZE;
2975                 if (sector_nr + (len>>9) > max_sector)
2976                         len = (max_sector - sector_nr) << 9;
2977                 if (len == 0)
2978                         break;
2979                 for (bio= biolist ; bio ; bio=bio->bi_next) {
2980                         struct bio *bio2;
2981                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2982                         if (bio_add_page(bio, page, len, 0))
2983                                 continue;
2984
2985                         /* stop here */
2986                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2987                         for (bio2 = biolist;
2988                              bio2 && bio2 != bio;
2989                              bio2 = bio2->bi_next) {
2990                                 /* remove last page from this bio */
2991                                 bio2->bi_vcnt--;
2992                                 bio2->bi_size -= len;
2993                                 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2994                         }
2995                         goto bio_full;
2996                 }
2997                 nr_sectors += len>>9;
2998                 sector_nr += len>>9;
2999         } while (biolist->bi_vcnt < RESYNC_PAGES);
3000  bio_full:
3001         r10_bio->sectors = nr_sectors;
3002
3003         while (biolist) {
3004                 bio = biolist;
3005                 biolist = biolist->bi_next;
3006
3007                 bio->bi_next = NULL;
3008                 r10_bio = bio->bi_private;
3009                 r10_bio->sectors = nr_sectors;
3010
3011                 if (bio->bi_end_io == end_sync_read) {
3012                         md_sync_acct(bio->bi_bdev, nr_sectors);
3013                         generic_make_request(bio);
3014                 }
3015         }
3016
3017         if (sectors_skipped)
3018                 /* pretend they weren't skipped, it makes
3019                  * no important difference in this case
3020                  */
3021                 md_done_sync(mddev, sectors_skipped, 1);
3022
3023         return sectors_skipped + nr_sectors;
3024  giveup:
3025         /* There is nowhere to write, so all non-sync
3026          * drives must be failed or in resync, all drives
3027          * have a bad block, so try the next chunk...
3028          */
3029         if (sector_nr + max_sync < max_sector)
3030                 max_sector = sector_nr + max_sync;
3031
3032         sectors_skipped += (max_sector - sector_nr);
3033         chunks_skipped ++;
3034         sector_nr = max_sector;
3035         goto skipped;
3036 }
3037
3038 static sector_t
3039 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3040 {
3041         sector_t size;
3042         struct r10conf *conf = mddev->private;
3043
3044         if (!raid_disks)
3045                 raid_disks = conf->raid_disks;
3046         if (!sectors)
3047                 sectors = conf->dev_sectors;
3048
3049         size = sectors >> conf->chunk_shift;
3050         sector_div(size, conf->far_copies);
3051         size = size * raid_disks;
3052         sector_div(size, conf->near_copies);
3053
3054         return size << conf->chunk_shift;
3055 }
3056
3057
3058 static struct r10conf *setup_conf(struct mddev *mddev)
3059 {
3060         struct r10conf *conf = NULL;
3061         int nc, fc, fo;
3062         sector_t stride, size;
3063         int err = -EINVAL;
3064
3065         if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
3066             !is_power_of_2(mddev->new_chunk_sectors)) {
3067                 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3068                        "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3069                        mdname(mddev), PAGE_SIZE);
3070                 goto out;
3071         }
3072
3073         nc = mddev->new_layout & 255;
3074         fc = (mddev->new_layout >> 8) & 255;
3075         fo = mddev->new_layout & (1<<16);
3076
3077         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
3078             (mddev->new_layout >> 17)) {
3079                 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3080                        mdname(mddev), mddev->new_layout);
3081                 goto out;
3082         }
3083
3084         err = -ENOMEM;
3085         conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3086         if (!conf)
3087                 goto out;
3088
3089         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
3090                                 GFP_KERNEL);
3091         if (!conf->mirrors)
3092                 goto out;
3093
3094         conf->tmppage = alloc_page(GFP_KERNEL);
3095         if (!conf->tmppage)
3096                 goto out;
3097
3098
3099         conf->raid_disks = mddev->raid_disks;
3100         conf->near_copies = nc;
3101         conf->far_copies = fc;
3102         conf->copies = nc*fc;
3103         conf->far_offset = fo;
3104         conf->chunk_mask = mddev->new_chunk_sectors - 1;
3105         conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
3106
3107         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3108                                            r10bio_pool_free, conf);
3109         if (!conf->r10bio_pool)
3110                 goto out;
3111
3112         size = mddev->dev_sectors >> conf->chunk_shift;
3113         sector_div(size, fc);
3114         size = size * conf->raid_disks;
3115         sector_div(size, nc);
3116         /* 'size' is now the number of chunks in the array */
3117         /* calculate "used chunks per device" in 'stride' */
3118         stride = size * conf->copies;
3119
3120         /* We need to round up when dividing by raid_disks to
3121          * get the stride size.
3122          */
3123         stride += conf->raid_disks - 1;
3124         sector_div(stride, conf->raid_disks);
3125
3126         conf->dev_sectors = stride << conf->chunk_shift;
3127
3128         if (fo)
3129                 stride = 1;
3130         else
3131                 sector_div(stride, fc);
3132         conf->stride = stride << conf->chunk_shift;
3133
3134
3135         spin_lock_init(&conf->device_lock);
3136         INIT_LIST_HEAD(&conf->retry_list);
3137
3138         spin_lock_init(&conf->resync_lock);
3139         init_waitqueue_head(&conf->wait_barrier);
3140
3141         conf->thread = md_register_thread(raid10d, mddev, NULL);
3142         if (!conf->thread)
3143                 goto out;
3144
3145         conf->mddev = mddev;
3146         return conf;
3147
3148  out:
3149         printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3150                mdname(mddev));
3151         if (conf) {
3152                 if (conf->r10bio_pool)
3153                         mempool_destroy(conf->r10bio_pool);
3154                 kfree(conf->mirrors);
3155                 safe_put_page(conf->tmppage);
3156                 kfree(conf);
3157         }
3158         return ERR_PTR(err);
3159 }
3160
3161 static int run(struct mddev *mddev)
3162 {
3163         struct r10conf *conf;
3164         int i, disk_idx, chunk_size;
3165         struct mirror_info *disk;
3166         struct md_rdev *rdev;
3167         sector_t size;
3168
3169         /*
3170          * copy the already verified devices into our private RAID10
3171          * bookkeeping area. [whatever we allocate in run(),
3172          * should be freed in stop()]
3173          */
3174
3175         if (mddev->private == NULL) {
3176                 conf = setup_conf(mddev);
3177                 if (IS_ERR(conf))
3178                         return PTR_ERR(conf);
3179                 mddev->private = conf;
3180         }
3181         conf = mddev->private;
3182         if (!conf)
3183                 goto out;
3184
3185         mddev->thread = conf->thread;
3186         conf->thread = NULL;
3187
3188         chunk_size = mddev->chunk_sectors << 9;
3189         blk_queue_io_min(mddev->queue, chunk_size);
3190         if (conf->raid_disks % conf->near_copies)
3191                 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
3192         else
3193                 blk_queue_io_opt(mddev->queue, chunk_size *
3194                                  (conf->raid_disks / conf->near_copies));
3195
3196         list_for_each_entry(rdev, &mddev->disks, same_set) {
3197
3198                 disk_idx = rdev->raid_disk;
3199                 if (disk_idx >= conf->raid_disks
3200                     || disk_idx < 0)
3201                         continue;
3202                 disk = conf->mirrors + disk_idx;
3203
3204                 disk->rdev = rdev;
3205                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3206                                   rdev->data_offset << 9);
3207                 /* as we don't honour merge_bvec_fn, we must never risk
3208                  * violating it, so limit max_segments to 1 lying
3209                  * within a single page.
3210                  */
3211                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
3212                         blk_queue_max_segments(mddev->queue, 1);
3213                         blk_queue_segment_boundary(mddev->queue,
3214                                                    PAGE_CACHE_SIZE - 1);
3215                 }
3216
3217                 disk->head_position = 0;
3218         }
3219         /* need to check that every block has at least one working mirror */
3220         if (!enough(conf, -1)) {
3221                 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3222                        mdname(mddev));
3223                 goto out_free_conf;
3224         }
3225
3226         mddev->degraded = 0;
3227         for (i = 0; i < conf->raid_disks; i++) {
3228
3229                 disk = conf->mirrors + i;
3230
3231                 if (!disk->rdev ||
3232                     !test_bit(In_sync, &disk->rdev->flags)) {
3233                         disk->head_position = 0;
3234                         mddev->degraded++;
3235                         if (disk->rdev)
3236                                 conf->fullsync = 1;
3237                 }
3238                 disk->recovery_disabled = mddev->recovery_disabled - 1;
3239         }
3240
3241         if (mddev->recovery_cp != MaxSector)
3242                 printk(KERN_NOTICE "md/raid10:%s: not clean"
3243                        " -- starting background reconstruction\n",
3244                        mdname(mddev));
3245         printk(KERN_INFO
3246                 "md/raid10:%s: active with %d out of %d devices\n",
3247                 mdname(mddev), conf->raid_disks - mddev->degraded,
3248                 conf->raid_disks);
3249         /*
3250          * Ok, everything is just fine now
3251          */
3252         mddev->dev_sectors = conf->dev_sectors;
3253         size = raid10_size(mddev, 0, 0);
3254         md_set_array_sectors(mddev, size);
3255         mddev->resync_max_sectors = size;
3256
3257         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3258         mddev->queue->backing_dev_info.congested_data = mddev;
3259
3260         /* Calculate max read-ahead size.
3261          * We need to readahead at least twice a whole stripe....
3262          * maybe...
3263          */
3264         {
3265                 int stripe = conf->raid_disks *
3266                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3267                 stripe /= conf->near_copies;
3268                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
3269                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
3270         }
3271
3272         if (conf->near_copies < conf->raid_disks)
3273                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3274
3275         if (md_integrity_register(mddev))
3276                 goto out_free_conf;
3277
3278         return 0;
3279
3280 out_free_conf:
3281         md_unregister_thread(&mddev->thread);
3282         if (conf->r10bio_pool)
3283                 mempool_destroy(conf->r10bio_pool);
3284         safe_put_page(conf->tmppage);
3285         kfree(conf->mirrors);
3286         kfree(conf);
3287         mddev->private = NULL;
3288 out:
3289         return -EIO;
3290 }
3291
3292 static int stop(struct mddev *mddev)
3293 {
3294         struct r10conf *conf = mddev->private;
3295
3296         raise_barrier(conf, 0);
3297         lower_barrier(conf);
3298
3299         md_unregister_thread(&mddev->thread);
3300         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
3301         if (conf->r10bio_pool)
3302                 mempool_destroy(conf->r10bio_pool);
3303         kfree(conf->mirrors);
3304         kfree(conf);
3305         mddev->private = NULL;
3306         return 0;
3307 }
3308
3309 static void raid10_quiesce(struct mddev *mddev, int state)
3310 {
3311         struct r10conf *conf = mddev->private;
3312
3313         switch(state) {
3314         case 1:
3315                 raise_barrier(conf, 0);
3316                 break;
3317         case 0:
3318                 lower_barrier(conf);
3319                 break;
3320         }
3321 }
3322
3323 static void *raid10_takeover_raid0(struct mddev *mddev)
3324 {
3325         struct md_rdev *rdev;
3326         struct r10conf *conf;
3327
3328         if (mddev->degraded > 0) {
3329                 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3330                        mdname(mddev));
3331                 return ERR_PTR(-EINVAL);
3332         }
3333
3334         /* Set new parameters */
3335         mddev->new_level = 10;
3336         /* new layout: far_copies = 1, near_copies = 2 */
3337         mddev->new_layout = (1<<8) + 2;
3338         mddev->new_chunk_sectors = mddev->chunk_sectors;
3339         mddev->delta_disks = mddev->raid_disks;
3340         mddev->raid_disks *= 2;
3341         /* make sure it will be not marked as dirty */
3342         mddev->recovery_cp = MaxSector;
3343
3344         conf = setup_conf(mddev);
3345         if (!IS_ERR(conf)) {
3346                 list_for_each_entry(rdev, &mddev->disks, same_set)
3347                         if (rdev->raid_disk >= 0)
3348                                 rdev->new_raid_disk = rdev->raid_disk * 2;
3349                 conf->barrier = 1;
3350         }
3351
3352         return conf;
3353 }
3354
3355 static void *raid10_takeover(struct mddev *mddev)
3356 {
3357         struct r0conf *raid0_conf;
3358
3359         /* raid10 can take over:
3360          *  raid0 - providing it has only two drives
3361          */
3362         if (mddev->level == 0) {
3363                 /* for raid0 takeover only one zone is supported */
3364                 raid0_conf = mddev->private;
3365                 if (raid0_conf->nr_strip_zones > 1) {
3366                         printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3367                                " with more than one zone.\n",
3368                                mdname(mddev));
3369                         return ERR_PTR(-EINVAL);
3370                 }
3371                 return raid10_takeover_raid0(mddev);
3372         }
3373         return ERR_PTR(-EINVAL);
3374 }
3375
3376 static struct md_personality raid10_personality =
3377 {
3378         .name           = "raid10",
3379         .level          = 10,
3380         .owner          = THIS_MODULE,
3381         .make_request   = make_request,
3382         .run            = run,
3383         .stop           = stop,
3384         .status         = status,
3385         .error_handler  = error,
3386         .hot_add_disk   = raid10_add_disk,
3387         .hot_remove_disk= raid10_remove_disk,
3388         .spare_active   = raid10_spare_active,
3389         .sync_request   = sync_request,
3390         .quiesce        = raid10_quiesce,
3391         .size           = raid10_size,
3392         .takeover       = raid10_takeover,
3393 };
3394
3395 static int __init raid_init(void)
3396 {
3397         return register_md_personality(&raid10_personality);
3398 }
3399
3400 static void raid_exit(void)
3401 {
3402         unregister_md_personality(&raid10_personality);
3403 }
3404
3405 module_init(raid_init);
3406 module_exit(raid_exit);
3407 MODULE_LICENSE("GPL");
3408 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3409 MODULE_ALIAS("md-personality-9"); /* RAID10 */
3410 MODULE_ALIAS("md-raid10");
3411 MODULE_ALIAS("md-level-10");
3412
3413 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);