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 futher 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 "dm-bio-list.h"
  22 #include <linux/raid/raid10.h>
  23 #include <linux/raid/bitmap.h>
  24
  25 /*
  26  * RAID10 provides a combination of RAID0 and RAID1 functionality.
  27  * The layout of data is defined by
  28  *    chunk_size
  29  *    raid_disks
  30  *    near_copies (stored in low byte of layout)
  31  *    far_copies (stored in second byte of layout)
  32  *
  33  * The data to be stored is divided into chunks using chunksize.
  34  * Each device is divided into far_copies sections.
  35  * In each section, chunks are laid out in a style similar to raid0, but
  36  * near_copies copies of each chunk is stored (each on a different drive).
  37  * The starting device for each section is offset near_copies from the starting
  38  * device of the previous section.
  39  * Thus there are (near_copies*far_copies) of each chunk, and each is on a different
  40  * drive.
  41  * near_copies and far_copies must be at least one, and their product is at most
  42  * raid_disks.
  43  */
  44
  45 /*
  46  * Number of guaranteed r10bios in case of extreme VM load:
  47  */
  48 #define NR_RAID10_BIOS 256
  49
  50 static void unplug_slaves(mddev_t *mddev);
  51
  52 static void allow_barrier(conf_t *conf);
  53 static void lower_barrier(conf_t *conf);
  54
  55 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
  56 {
  57         conf_t *conf = data;
  58         r10bio_t *r10_bio;
  59         int size = offsetof(struct r10bio_s, devs[conf->copies]);
  60
  61         /* allocate a r10bio with room for raid_disks entries in the bios array */
  62         r10_bio = kmalloc(size, gfp_flags);
  63         if (r10_bio)
  64                 memset(r10_bio, 0, size);
  65         else
  66                 unplug_slaves(conf->mddev);
  67
  68         return r10_bio;
  69 }
  70
  71 static void r10bio_pool_free(void *r10_bio, void *data)
  72 {
  73         kfree(r10_bio);
  74 }
  75
  76 #define RESYNC_BLOCK_SIZE (64*1024)
  77 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  78 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  79 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  80 #define RESYNC_WINDOW (2048*1024)
  81
  82 /*
  83  * When performing a resync, we need to read and compare, so
  84  * we need as many pages are there are copies.
  85  * When performing a recovery, we need 2 bios, one for read,
  86  * one for write (we recover only one drive per r10buf)
  87  *
  88  */
  89 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
  90 {
  91         conf_t *conf = data;
  92         struct page *page;
  93         r10bio_t *r10_bio;
  94         struct bio *bio;
  95         int i, j;
  96         int nalloc;
  97
  98         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
  99         if (!r10_bio) {
 100                 unplug_slaves(conf->mddev);
 101                 return NULL;
 102         }
 103
 104         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
 105                 nalloc = conf->copies; /* resync */
 106         else
 107                 nalloc = 2; /* recovery */
 108
 109         /*
 110          * Allocate bios.
 111          */
 112         for (j = nalloc ; j-- ; ) {
 113                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 114                 if (!bio)
 115                         goto out_free_bio;
 116                 r10_bio->devs[j].bio = bio;
 117         }
 118         /*
 119          * Allocate RESYNC_PAGES data pages and attach them
 120          * where needed.
 121          */
 122         for (j = 0 ; j < nalloc; j++) {
 123                 bio = r10_bio->devs[j].bio;
 124                 for (i = 0; i < RESYNC_PAGES; i++) {
 125                         page = alloc_page(gfp_flags);
 126                         if (unlikely(!page))
 127                                 goto out_free_pages;
 128
 129                         bio->bi_io_vec[i].bv_page = page;
 130                 }
 131         }
 132
 133         return r10_bio;
 134
 135 out_free_pages:
 136         for ( ; i > 0 ; i--)
 137                 __free_page(bio->bi_io_vec[i-1].bv_page);
 138         while (j--)
 139                 for (i = 0; i < RESYNC_PAGES ; i++)
 140                         __free_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 141         j = -1;
 142 out_free_bio:
 143         while ( ++j < nalloc )
 144                 bio_put(r10_bio->devs[j].bio);
 145         r10bio_pool_free(r10_bio, conf);
 146         return NULL;
 147 }
 148
 149 static void r10buf_pool_free(void *__r10_bio, void *data)
 150 {
 151         int i;
 152         conf_t *conf = data;
 153         r10bio_t *r10bio = __r10_bio;
 154         int j;
 155
 156         for (j=0; j < conf->copies; j++) {
 157                 struct bio *bio = r10bio->devs[j].bio;
 158                 if (bio) {
 159                         for (i = 0; i < RESYNC_PAGES; i++) {
 160                                 __free_page(bio->bi_io_vec[i].bv_page);
 161                                 bio->bi_io_vec[i].bv_page = NULL;
 162                         }
 163                         bio_put(bio);
 164                 }
 165         }
 166         r10bio_pool_free(r10bio, conf);
 167 }
 168
 169 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
 170 {
 171         int i;
 172
 173         for (i = 0; i < conf->copies; i++) {
 174                 struct bio **bio = & r10_bio->devs[i].bio;
 175                 if (*bio && *bio != IO_BLOCKED)
 176                         bio_put(*bio);
 177                 *bio = NULL;
 178         }
 179 }
 180
 181 static inline void free_r10bio(r10bio_t *r10_bio)
 182 {
 183         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 184
 185         /*
 186          * Wake up any possible resync thread that waits for the device
 187          * to go idle.
 188          */
 189         allow_barrier(conf);
 190
 191         put_all_bios(conf, r10_bio);
 192         mempool_free(r10_bio, conf->r10bio_pool);
 193 }
 194
 195 static inline void put_buf(r10bio_t *r10_bio)
 196 {
 197         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 198
 199         mempool_free(r10_bio, conf->r10buf_pool);
 200
 201         lower_barrier(conf);
 202 }
 203
 204 static void reschedule_retry(r10bio_t *r10_bio)
 205 {
 206         unsigned long flags;
 207         mddev_t *mddev = r10_bio->mddev;
 208         conf_t *conf = mddev_to_conf(mddev);
 209
 210         spin_lock_irqsave(&conf->device_lock, flags);
 211         list_add(&r10_bio->retry_list, &conf->retry_list);
 212         conf->nr_queued ++;
 213         spin_unlock_irqrestore(&conf->device_lock, flags);
 214
 215         md_wakeup_thread(mddev->thread);
 216 }
 217
 218 /*
 219  * raid_end_bio_io() is called when we have finished servicing a mirrored
 220  * operation and are ready to return a success/failure code to the buffer
 221  * cache layer.
 222  */
 223 static void raid_end_bio_io(r10bio_t *r10_bio)
 224 {
 225         struct bio *bio = r10_bio->master_bio;
 226
 227         bio_endio(bio, bio->bi_size,
 228                 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
 229         free_r10bio(r10_bio);
 230 }
 231
 232 /*
 233  * Update disk head position estimator based on IRQ completion info.
 234  */
 235 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
 236 {
 237         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 238
 239         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 240                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 241 }
 242
 243 static int raid10_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
 244 {
 245         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 246         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
 247         int slot, dev;
 248         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 249
 250         if (bio->bi_size)
 251                 return 1;
 252
 253         slot = r10_bio->read_slot;
 254         dev = r10_bio->devs[slot].devnum;
 255         /*
 256          * this branch is our 'one mirror IO has finished' event handler:
 257          */
 258         update_head_pos(slot, r10_bio);
 259
 260         if (uptodate) {
 261                 /*
 262                  * Set R10BIO_Uptodate in our master bio, so that
 263                  * we will return a good error code to the higher
 264                  * levels even if IO on some other mirrored buffer fails.
 265                  *
 266                  * The 'master' represents the composite IO operation to
 267                  * user-side. So if something waits for IO, then it will
 268                  * wait for the 'master' bio.
 269                  */
 270                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 271                 raid_end_bio_io(r10_bio);
 272         } else {
 273                 /*
 274                  * oops, read error:
 275                  */
 276                 char b[BDEVNAME_SIZE];
 277                 if (printk_ratelimit())
 278                         printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
 279                                bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
 280                 reschedule_retry(r10_bio);
 281         }
 282
 283         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 284         return 0;
 285 }
 286
 287 static int raid10_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
 288 {
 289         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 290         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
 291         int slot, dev;
 292         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 293
 294         if (bio->bi_size)
 295                 return 1;
 296
 297         for (slot = 0; slot < conf->copies; slot++)
 298                 if (r10_bio->devs[slot].bio == bio)
 299                         break;
 300         dev = r10_bio->devs[slot].devnum;
 301
 302         /*
 303          * this branch is our 'one mirror IO has finished' event handler:
 304          */
 305         if (!uptodate) {
 306                 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
 307                 /* an I/O failed, we can't clear the bitmap */
 308                 set_bit(R10BIO_Degraded, &r10_bio->state);
 309         } else
 310                 /*
 311                  * Set R10BIO_Uptodate in our master bio, so that
 312                  * we will return a good error code for to the higher
 313                  * levels even if IO on some other mirrored buffer fails.
 314                  *
 315                  * The 'master' represents the composite IO operation to
 316                  * user-side. So if something waits for IO, then it will
 317                  * wait for the 'master' bio.
 318                  */
 319                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 320
 321         update_head_pos(slot, r10_bio);
 322
 323         /*
 324          *
 325          * Let's see if all mirrored write operations have finished
 326          * already.
 327          */
 328         if (atomic_dec_and_test(&r10_bio->remaining)) {
 329                 /* clear the bitmap if all writes complete successfully */
 330                 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 331                                 r10_bio->sectors,
 332                                 !test_bit(R10BIO_Degraded, &r10_bio->state),
 333                                 0);
 334                 md_write_end(r10_bio->mddev);
 335                 raid_end_bio_io(r10_bio);
 336         }
 337
 338         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 339         return 0;
 340 }
 341
 342
 343 /*
 344  * RAID10 layout manager
 345  * Aswell as the chunksize and raid_disks count, there are two
 346  * parameters: near_copies and far_copies.
 347  * near_copies * far_copies must be <= raid_disks.
 348  * Normally one of these will be 1.
 349  * If both are 1, we get raid0.
 350  * If near_copies == raid_disks, we get raid1.
 351  *
 352  * Chunks are layed out in raid0 style with near_copies copies of the
 353  * first chunk, followed by near_copies copies of the next chunk and
 354  * so on.
 355  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 356  * as described above, we start again with a device offset of near_copies.
 357  * So we effectively have another copy of the whole array further down all
 358  * the drives, but with blocks on different drives.
 359  * With this layout, and block is never stored twice on the one device.
 360  *
 361  * raid10_find_phys finds the sector offset of a given virtual sector
 362  * on each device that it is on. If a block isn't on a device,
 363  * that entry in the array is set to MaxSector.
 364  *
 365  * raid10_find_virt does the reverse mapping, from a device and a
 366  * sector offset to a virtual address
 367  */
 368
 369 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
 370 {
 371         int n,f;
 372         sector_t sector;
 373         sector_t chunk;
 374         sector_t stripe;
 375         int dev;
 376
 377         int slot = 0;
 378
 379         /* now calculate first sector/dev */
 380         chunk = r10bio->sector >> conf->chunk_shift;
 381         sector = r10bio->sector & conf->chunk_mask;
 382
 383         chunk *= conf->near_copies;
 384         stripe = chunk;
 385         dev = sector_div(stripe, conf->raid_disks);
 386
 387         sector += stripe << conf->chunk_shift;
 388
 389         /* and calculate all the others */
 390         for (n=0; n < conf->near_copies; n++) {
 391                 int d = dev;
 392                 sector_t s = sector;
 393                 r10bio->devs[slot].addr = sector;
 394                 r10bio->devs[slot].devnum = d;
 395                 slot++;
 396
 397                 for (f = 1; f < conf->far_copies; f++) {
 398                         d += conf->near_copies;
 399                         if (d >= conf->raid_disks)
 400                                 d -= conf->raid_disks;
 401                         s += conf->stride;
 402                         r10bio->devs[slot].devnum = d;
 403                         r10bio->devs[slot].addr = s;
 404                         slot++;
 405                 }
 406                 dev++;
 407                 if (dev >= conf->raid_disks) {
 408                         dev = 0;
 409                         sector += (conf->chunk_mask + 1);
 410                 }
 411         }
 412         BUG_ON(slot != conf->copies);
 413 }
 414
 415 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
 416 {
 417         sector_t offset, chunk, vchunk;
 418
 419         while (sector > conf->stride) {
 420                 sector -= conf->stride;
 421                 if (dev < conf->near_copies)
 422                         dev += conf->raid_disks - conf->near_copies;
 423                 else
 424                         dev -= conf->near_copies;
 425         }
 426
 427         offset = sector & conf->chunk_mask;
 428         chunk = sector >> conf->chunk_shift;
 429         vchunk = chunk * conf->raid_disks + dev;
 430         sector_div(vchunk, conf->near_copies);
 431         return (vchunk << conf->chunk_shift) + offset;
 432 }
 433
 434 /**
 435  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 436  *      @q: request queue
 437  *      @bio: the buffer head that's been built up so far
 438  *      @biovec: the request that could be merged to it.
 439  *
 440  *      Return amount of bytes we can accept at this offset
 441  *      If near_copies == raid_disk, there are no striping issues,
 442  *      but in that case, the function isn't called at all.
 443  */
 444 static int raid10_mergeable_bvec(request_queue_t *q, struct bio *bio,
 445                                 struct bio_vec *bio_vec)
 446 {
 447         mddev_t *mddev = q->queuedata;
 448         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
 449         int max;
 450         unsigned int chunk_sectors = mddev->chunk_size >> 9;
 451         unsigned int bio_sectors = bio->bi_size >> 9;
 452
 453         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 454         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 455         if (max <= bio_vec->bv_len && bio_sectors == 0)
 456                 return bio_vec->bv_len;
 457         else
 458                 return max;
 459 }
 460
 461 /*
 462  * This routine returns the disk from which the requested read should
 463  * be done. There is a per-array 'next expected sequential IO' sector
 464  * number - if this matches on the next IO then we use the last disk.
 465  * There is also a per-disk 'last know head position' sector that is
 466  * maintained from IRQ contexts, both the normal and the resync IO
 467  * completion handlers update this position correctly. If there is no
 468  * perfect sequential match then we pick the disk whose head is closest.
 469  *
 470  * If there are 2 mirrors in the same 2 devices, performance degrades
 471  * because position is mirror, not device based.
 472  *
 473  * The rdev for the device selected will have nr_pending incremented.
 474  */
 475
 476 /*
 477  * FIXME: possibly should rethink readbalancing and do it differently
 478  * depending on near_copies / far_copies geometry.
 479  */
 480 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
 481 {
 482         const unsigned long this_sector = r10_bio->sector;
 483         int disk, slot, nslot;
 484         const int sectors = r10_bio->sectors;
 485         sector_t new_distance, current_distance;
 486         mdk_rdev_t *rdev;
 487
 488         raid10_find_phys(conf, r10_bio);
 489         rcu_read_lock();
 490         /*
 491          * Check if we can balance. We can balance on the whole
 492          * device if no resync is going on (recovery is ok), or below
 493          * the resync window. We take the first readable disk when
 494          * above the resync window.
 495          */
 496         if (conf->mddev->recovery_cp < MaxSector
 497             && (this_sector + sectors >= conf->next_resync)) {
 498                 /* make sure that disk is operational */
 499                 slot = 0;
 500                 disk = r10_bio->devs[slot].devnum;
 501
 502                 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 503                        r10_bio->devs[slot].bio == IO_BLOCKED ||
 504                        !test_bit(In_sync, &rdev->flags)) {
 505                         slot++;
 506                         if (slot == conf->copies) {
 507                                 slot = 0;
 508                                 disk = -1;
 509                                 break;
 510                         }
 511                         disk = r10_bio->devs[slot].devnum;
 512                 }
 513                 goto rb_out;
 514         }
 515
 516
 517         /* make sure the disk is operational */
 518         slot = 0;
 519         disk = r10_bio->devs[slot].devnum;
 520         while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 521                r10_bio->devs[slot].bio == IO_BLOCKED ||
 522                !test_bit(In_sync, &rdev->flags)) {
 523                 slot ++;
 524                 if (slot == conf->copies) {
 525                         disk = -1;
 526                         goto rb_out;
 527                 }
 528                 disk = r10_bio->devs[slot].devnum;
 529         }
 530
 531
 532         current_distance = abs(r10_bio->devs[slot].addr -
 533                                conf->mirrors[disk].head_position);
 534
 535         /* Find the disk whose head is closest */
 536
 537         for (nslot = slot; nslot < conf->copies; nslot++) {
 538                 int ndisk = r10_bio->devs[nslot].devnum;
 539
 540
 541                 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
 542                     r10_bio->devs[nslot].bio == IO_BLOCKED ||
 543                     !test_bit(In_sync, &rdev->flags))
 544                         continue;
 545
 546                 /* This optimisation is debatable, and completely destroys
 547                  * sequential read speed for 'far copies' arrays.  So only
 548                  * keep it for 'near' arrays, and review those later.
 549                  */
 550                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
 551                         disk = ndisk;
 552                         slot = nslot;
 553                         break;
 554                 }
 555                 new_distance = abs(r10_bio->devs[nslot].addr -
 556                                    conf->mirrors[ndisk].head_position);
 557                 if (new_distance < current_distance) {
 558                         current_distance = new_distance;
 559                         disk = ndisk;
 560                         slot = nslot;
 561                 }
 562         }
 563
 564 rb_out:
 565         r10_bio->read_slot = slot;
 566 /*      conf->next_seq_sect = this_sector + sectors;*/
 567
 568         if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
 569                 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
 570         rcu_read_unlock();
 571
 572         return disk;
 573 }
 574
 575 static void unplug_slaves(mddev_t *mddev)
 576 {
 577         conf_t *conf = mddev_to_conf(mddev);
 578         int i;
 579
 580         rcu_read_lock();
 581         for (i=0; i<mddev->raid_disks; i++) {
 582                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 583                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 584                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 585
 586                         atomic_inc(&rdev->nr_pending);
 587                         rcu_read_unlock();
 588
 589                         if (r_queue->unplug_fn)
 590                                 r_queue->unplug_fn(r_queue);
 591
 592                         rdev_dec_pending(rdev, mddev);
 593                         rcu_read_lock();
 594                 }
 595         }
 596         rcu_read_unlock();
 597 }
 598
 599 static void raid10_unplug(request_queue_t *q)
 600 {
 601         mddev_t *mddev = q->queuedata;
 602
 603         unplug_slaves(q->queuedata);
 604         md_wakeup_thread(mddev->thread);
 605 }
 606
 607 static int raid10_issue_flush(request_queue_t *q, struct gendisk *disk,
 608                              sector_t *error_sector)
 609 {
 610         mddev_t *mddev = q->queuedata;
 611         conf_t *conf = mddev_to_conf(mddev);
 612         int i, ret = 0;
 613
 614         rcu_read_lock();
 615         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
 616                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 617                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 618                         struct block_device *bdev = rdev->bdev;
 619                         request_queue_t *r_queue = bdev_get_queue(bdev);
 620
 621                         if (!r_queue->issue_flush_fn)
 622                                 ret = -EOPNOTSUPP;
 623                         else {
 624                                 atomic_inc(&rdev->nr_pending);
 625                                 rcu_read_unlock();
 626                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
 627                                                               error_sector);
 628                                 rdev_dec_pending(rdev, mddev);
 629                                 rcu_read_lock();
 630                         }
 631                 }
 632         }
 633         rcu_read_unlock();
 634         return ret;
 635 }
 636
 637 /* Barriers....
 638  * Sometimes we need to suspend IO while we do something else,
 639  * either some resync/recovery, or reconfigure the array.
 640  * To do this we raise a 'barrier'.
 641  * The 'barrier' is a counter that can be raised multiple times
 642  * to count how many activities are happening which preclude
 643  * normal IO.
 644  * We can only raise the barrier if there is no pending IO.
 645  * i.e. if nr_pending == 0.
 646  * We choose only to raise the barrier if no-one is waiting for the
 647  * barrier to go down.  This means that as soon as an IO request
 648  * is ready, no other operations which require a barrier will start
 649  * until the IO request has had a chance.
 650  *
 651  * So: regular IO calls 'wait_barrier'.  When that returns there
 652  *    is no backgroup IO happening,  It must arrange to call
 653  *    allow_barrier when it has finished its IO.
 654  * backgroup IO calls must call raise_barrier.  Once that returns
 655  *    there is no normal IO happeing.  It must arrange to call
 656  *    lower_barrier when the particular background IO completes.
 657  */
 658 #define RESYNC_DEPTH 32
 659
 660 static void raise_barrier(conf_t *conf, int force)
 661 {
 662         BUG_ON(force && !conf->barrier);
 663         spin_lock_irq(&conf->resync_lock);
 664
 665         /* Wait until no block IO is waiting (unless 'force') */
 666         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 667                             conf->resync_lock,
 668                             raid10_unplug(conf->mddev->queue));
 669
 670         /* block any new IO from starting */
 671         conf->barrier++;
 672
 673         /* No wait for all pending IO to complete */
 674         wait_event_lock_irq(conf->wait_barrier,
 675                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 676                             conf->resync_lock,
 677                             raid10_unplug(conf->mddev->queue));
 678
 679         spin_unlock_irq(&conf->resync_lock);
 680 }
 681
 682 static void lower_barrier(conf_t *conf)
 683 {
 684         unsigned long flags;
 685         spin_lock_irqsave(&conf->resync_lock, flags);
 686         conf->barrier--;
 687         spin_unlock_irqrestore(&conf->resync_lock, flags);
 688         wake_up(&conf->wait_barrier);
 689 }
 690
 691 static void wait_barrier(conf_t *conf)
 692 {
 693         spin_lock_irq(&conf->resync_lock);
 694         if (conf->barrier) {
 695                 conf->nr_waiting++;
 696                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 697                                     conf->resync_lock,
 698                                     raid10_unplug(conf->mddev->queue));
 699                 conf->nr_waiting--;
 700         }
 701         conf->nr_pending++;
 702         spin_unlock_irq(&conf->resync_lock);
 703 }
 704
 705 static void allow_barrier(conf_t *conf)
 706 {
 707         unsigned long flags;
 708         spin_lock_irqsave(&conf->resync_lock, flags);
 709         conf->nr_pending--;
 710         spin_unlock_irqrestore(&conf->resync_lock, flags);
 711         wake_up(&conf->wait_barrier);
 712 }
 713
 714 static void freeze_array(conf_t *conf)
 715 {
 716         /* stop syncio and normal IO and wait for everything to
 717          * go quite.
 718          * We increment barrier and nr_waiting, and then
 719          * wait until barrier+nr_pending match nr_queued+2
 720          */
 721         spin_lock_irq(&conf->resync_lock);
 722         conf->barrier++;
 723         conf->nr_waiting++;
 724         wait_event_lock_irq(conf->wait_barrier,
 725                             conf->barrier+conf->nr_pending == conf->nr_queued+2,
 726                             conf->resync_lock,
 727                             raid10_unplug(conf->mddev->queue));
 728         spin_unlock_irq(&conf->resync_lock);
 729 }
 730
 731 static void unfreeze_array(conf_t *conf)
 732 {
 733         /* reverse the effect of the freeze */
 734         spin_lock_irq(&conf->resync_lock);
 735         conf->barrier--;
 736         conf->nr_waiting--;
 737         wake_up(&conf->wait_barrier);
 738         spin_unlock_irq(&conf->resync_lock);
 739 }
 740
 741 static int make_request(request_queue_t *q, struct bio * bio)
 742 {
 743         mddev_t *mddev = q->queuedata;
 744         conf_t *conf = mddev_to_conf(mddev);
 745         mirror_info_t *mirror;
 746         r10bio_t *r10_bio;
 747         struct bio *read_bio;
 748         int i;
 749         int chunk_sects = conf->chunk_mask + 1;
 750         const int rw = bio_data_dir(bio);
 751         struct bio_list bl;
 752         unsigned long flags;
 753
 754         if (unlikely(bio_barrier(bio))) {
 755                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
 756                 return 0;
 757         }
 758
 759         /* If this request crosses a chunk boundary, we need to
 760          * split it.  This will only happen for 1 PAGE (or less) requests.
 761          */
 762         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 763                       > chunk_sects &&
 764                     conf->near_copies < conf->raid_disks)) {
 765                 struct bio_pair *bp;
 766                 /* Sanity check -- queue functions should prevent this happening */
 767                 if (bio->bi_vcnt != 1 ||
 768                     bio->bi_idx != 0)
 769                         goto bad_map;
 770                 /* This is a one page bio that upper layers
 771                  * refuse to split for us, so we need to split it.
 772                  */
 773                 bp = bio_split(bio, bio_split_pool,
 774                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 775                 if (make_request(q, &bp->bio1))
 776                         generic_make_request(&bp->bio1);
 777                 if (make_request(q, &bp->bio2))
 778                         generic_make_request(&bp->bio2);
 779
 780                 bio_pair_release(bp);
 781                 return 0;
 782         bad_map:
 783                 printk("raid10_make_request bug: can't convert block across chunks"
 784                        " or bigger than %dk %llu %d\n", chunk_sects/2,
 785                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 786
 787                 bio_io_error(bio, bio->bi_size);
 788                 return 0;
 789         }
 790
 791         md_write_start(mddev, bio);
 792
 793         /*
 794          * Register the new request and wait if the reconstruction
 795          * thread has put up a bar for new requests.
 796          * Continue immediately if no resync is active currently.
 797          */
 798         wait_barrier(conf);
 799
 800         disk_stat_inc(mddev->gendisk, ios[rw]);
 801         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
 802
 803         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 804
 805         r10_bio->master_bio = bio;
 806         r10_bio->sectors = bio->bi_size >> 9;
 807
 808         r10_bio->mddev = mddev;
 809         r10_bio->sector = bio->bi_sector;
 810         r10_bio->state = 0;
 811
 812         if (rw == READ) {
 813                 /*
 814                  * read balancing logic:
 815                  */
 816                 int disk = read_balance(conf, r10_bio);
 817                 int slot = r10_bio->read_slot;
 818                 if (disk < 0) {
 819                         raid_end_bio_io(r10_bio);
 820                         return 0;
 821                 }
 822                 mirror = conf->mirrors + disk;
 823
 824                 read_bio = bio_clone(bio, GFP_NOIO);
 825
 826                 r10_bio->devs[slot].bio = read_bio;
 827
 828                 read_bio->bi_sector = r10_bio->devs[slot].addr +
 829                         mirror->rdev->data_offset;
 830                 read_bio->bi_bdev = mirror->rdev->bdev;
 831                 read_bio->bi_end_io = raid10_end_read_request;
 832                 read_bio->bi_rw = READ;
 833                 read_bio->bi_private = r10_bio;
 834
 835                 generic_make_request(read_bio);
 836                 return 0;
 837         }
 838
 839         /*
 840          * WRITE:
 841          */
 842         /* first select target devices under spinlock and
 843          * inc refcount on their rdev.  Record them by setting
 844          * bios[x] to bio
 845          */
 846         raid10_find_phys(conf, r10_bio);
 847         rcu_read_lock();
 848         for (i = 0;  i < conf->copies; i++) {
 849                 int d = r10_bio->devs[i].devnum;
 850                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
 851                 if (rdev &&
 852                     !test_bit(Faulty, &rdev->flags)) {
 853                         atomic_inc(&rdev->nr_pending);
 854                         r10_bio->devs[i].bio = bio;
 855                 } else {
 856                         r10_bio->devs[i].bio = NULL;
 857                         set_bit(R10BIO_Degraded, &r10_bio->state);
 858                 }
 859         }
 860         rcu_read_unlock();
 861
 862         atomic_set(&r10_bio->remaining, 0);
 863
 864         bio_list_init(&bl);
 865         for (i = 0; i < conf->copies; i++) {
 866                 struct bio *mbio;
 867                 int d = r10_bio->devs[i].devnum;
 868                 if (!r10_bio->devs[i].bio)
 869                         continue;
 870
 871                 mbio = bio_clone(bio, GFP_NOIO);
 872                 r10_bio->devs[i].bio = mbio;
 873
 874                 mbio->bi_sector = r10_bio->devs[i].addr+
 875                         conf->mirrors[d].rdev->data_offset;
 876                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
 877                 mbio->bi_end_io = raid10_end_write_request;
 878                 mbio->bi_rw = WRITE;
 879                 mbio->bi_private = r10_bio;
 880
 881                 atomic_inc(&r10_bio->remaining);
 882                 bio_list_add(&bl, mbio);
 883         }
 884
 885         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
 886         spin_lock_irqsave(&conf->device_lock, flags);
 887         bio_list_merge(&conf->pending_bio_list, &bl);
 888         blk_plug_device(mddev->queue);
 889         spin_unlock_irqrestore(&conf->device_lock, flags);
 890
 891         return 0;
 892 }
 893
 894 static void status(struct seq_file *seq, mddev_t *mddev)
 895 {
 896         conf_t *conf = mddev_to_conf(mddev);
 897         int i;
 898
 899         if (conf->near_copies < conf->raid_disks)
 900                 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
 901         if (conf->near_copies > 1)
 902                 seq_printf(seq, " %d near-copies", conf->near_copies);
 903         if (conf->far_copies > 1)
 904                 seq_printf(seq, " %d far-copies", conf->far_copies);
 905
 906         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 907                                                 conf->working_disks);
 908         for (i = 0; i < conf->raid_disks; i++)
 909                 seq_printf(seq, "%s",
 910                               conf->mirrors[i].rdev &&
 911                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
 912         seq_printf(seq, "]");
 913 }
 914
 915 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 916 {
 917         char b[BDEVNAME_SIZE];
 918         conf_t *conf = mddev_to_conf(mddev);
 919
 920         /*
 921          * If it is not operational, then we have already marked it as dead
 922          * else if it is the last working disks, ignore the error, let the
 923          * next level up know.
 924          * else mark the drive as failed
 925          */
 926         if (test_bit(In_sync, &rdev->flags)
 927             && conf->working_disks == 1)
 928                 /*
 929                  * Don't fail the drive, just return an IO error.
 930                  * The test should really be more sophisticated than
 931                  * "working_disks == 1", but it isn't critical, and
 932                  * can wait until we do more sophisticated "is the drive
 933                  * really dead" tests...
 934                  */
 935                 return;
 936         if (test_bit(In_sync, &rdev->flags)) {
 937                 mddev->degraded++;
 938                 conf->working_disks--;
 939                 /*
 940                  * if recovery is running, make sure it aborts.
 941                  */
 942                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 943         }
 944         clear_bit(In_sync, &rdev->flags);
 945         set_bit(Faulty, &rdev->flags);
 946         mddev->sb_dirty = 1;
 947         printk(KERN_ALERT "raid10: Disk failure on %s, disabling device. \n"
 948                 "       Operation continuing on %d devices\n",
 949                 bdevname(rdev->bdev,b), conf->working_disks);
 950 }
 951
 952 static void print_conf(conf_t *conf)
 953 {
 954         int i;
 955         mirror_info_t *tmp;
 956
 957         printk("RAID10 conf printout:\n");
 958         if (!conf) {
 959                 printk("(!conf)\n");
 960                 return;
 961         }
 962         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 963                 conf->raid_disks);
 964
 965         for (i = 0; i < conf->raid_disks; i++) {
 966                 char b[BDEVNAME_SIZE];
 967                 tmp = conf->mirrors + i;
 968                 if (tmp->rdev)
 969                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
 970                                 i, !test_bit(In_sync, &tmp->rdev->flags),
 971                                 !test_bit(Faulty, &tmp->rdev->flags),
 972                                 bdevname(tmp->rdev->bdev,b));
 973         }
 974 }
 975
 976 static void close_sync(conf_t *conf)
 977 {
 978         wait_barrier(conf);
 979         allow_barrier(conf);
 980
 981         mempool_destroy(conf->r10buf_pool);
 982         conf->r10buf_pool = NULL;
 983 }
 984
 985 /* check if there are enough drives for
 986  * every block to appear on atleast one
 987  */
 988 static int enough(conf_t *conf)
 989 {
 990         int first = 0;
 991
 992         do {
 993                 int n = conf->copies;
 994                 int cnt = 0;
 995                 while (n--) {
 996                         if (conf->mirrors[first].rdev)
 997                                 cnt++;
 998                         first = (first+1) % conf->raid_disks;
 999                 }
1000                 if (cnt == 0)
1001                         return 0;
1002         } while (first != 0);
1003         return 1;
1004 }
1005
1006 static int raid10_spare_active(mddev_t *mddev)
1007 {
1008         int i;
1009         conf_t *conf = mddev->private;
1010         mirror_info_t *tmp;
1011
1012         /*
1013          * Find all non-in_sync disks within the RAID10 configuration
1014          * and mark them in_sync
1015          */
1016         for (i = 0; i < conf->raid_disks; i++) {
1017                 tmp = conf->mirrors + i;
1018                 if (tmp->rdev
1019                     && !test_bit(Faulty, &tmp->rdev->flags)
1020                     && !test_bit(In_sync, &tmp->rdev->flags)) {
1021                         conf->working_disks++;
1022                         mddev->degraded--;
1023                         set_bit(In_sync, &tmp->rdev->flags);
1024                 }
1025         }
1026
1027         print_conf(conf);
1028         return 0;
1029 }
1030
1031
1032 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1033 {
1034         conf_t *conf = mddev->private;
1035         int found = 0;
1036         int mirror;
1037         mirror_info_t *p;
1038
1039         if (mddev->recovery_cp < MaxSector)
1040                 /* only hot-add to in-sync arrays, as recovery is
1041                  * very different from resync
1042                  */
1043                 return 0;
1044         if (!enough(conf))
1045                 return 0;
1046
1047         if (rdev->saved_raid_disk >= 0 &&
1048             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1049                 mirror = rdev->saved_raid_disk;
1050         else
1051                 mirror = 0;
1052         for ( ; mirror < mddev->raid_disks; mirror++)
1053                 if ( !(p=conf->mirrors+mirror)->rdev) {
1054
1055                         blk_queue_stack_limits(mddev->queue,
1056                                                rdev->bdev->bd_disk->queue);
1057                         /* as we don't honour merge_bvec_fn, we must never risk
1058                          * violating it, so limit ->max_sector to one PAGE, as
1059                          * a one page request is never in violation.
1060                          */
1061                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1062                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1063                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1064
1065                         p->head_position = 0;
1066                         rdev->raid_disk = mirror;
1067                         found = 1;
1068                         if (rdev->saved_raid_disk != mirror)
1069                                 conf->fullsync = 1;
1070                         rcu_assign_pointer(p->rdev, rdev);
1071                         break;
1072                 }
1073
1074         print_conf(conf);
1075         return found;
1076 }
1077
1078 static int raid10_remove_disk(mddev_t *mddev, int number)
1079 {
1080         conf_t *conf = mddev->private;
1081         int err = 0;
1082         mdk_rdev_t *rdev;
1083         mirror_info_t *p = conf->mirrors+ number;
1084
1085         print_conf(conf);
1086         rdev = p->rdev;
1087         if (rdev) {
1088                 if (test_bit(In_sync, &rdev->flags) ||
1089                     atomic_read(&rdev->nr_pending)) {
1090                         err = -EBUSY;
1091                         goto abort;
1092                 }
1093                 p->rdev = NULL;
1094                 synchronize_rcu();
1095                 if (atomic_read(&rdev->nr_pending)) {
1096                         /* lost the race, try later */
1097                         err = -EBUSY;
1098                         p->rdev = rdev;
1099                 }
1100         }
1101 abort:
1102
1103         print_conf(conf);
1104         return err;
1105 }
1106
1107
1108 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1109 {
1110         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1111         conf_t *conf = mddev_to_conf(r10_bio->mddev);
1112         int i,d;
1113
1114         if (bio->bi_size)
1115                 return 1;
1116
1117         for (i=0; i<conf->copies; i++)
1118                 if (r10_bio->devs[i].bio == bio)
1119                         break;
1120         if (i == conf->copies)
1121                 BUG();
1122         update_head_pos(i, r10_bio);
1123         d = r10_bio->devs[i].devnum;
1124
1125         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1126                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1127         else if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1128                 md_error(r10_bio->mddev,
1129                          conf->mirrors[d].rdev);
1130
1131         /* for reconstruct, we always reschedule after a read.
1132          * for resync, only after all reads
1133          */
1134         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1135             atomic_dec_and_test(&r10_bio->remaining)) {
1136                 /* we have read all the blocks,
1137                  * do the comparison in process context in raid10d
1138                  */
1139                 reschedule_retry(r10_bio);
1140         }
1141         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1142         return 0;
1143 }
1144
1145 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1146 {
1147         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1148         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1149         mddev_t *mddev = r10_bio->mddev;
1150         conf_t *conf = mddev_to_conf(mddev);
1151         int i,d;
1152
1153         if (bio->bi_size)
1154                 return 1;
1155
1156         for (i = 0; i < conf->copies; i++)
1157                 if (r10_bio->devs[i].bio == bio)
1158                         break;
1159         d = r10_bio->devs[i].devnum;
1160
1161         if (!uptodate)
1162                 md_error(mddev, conf->mirrors[d].rdev);
1163         update_head_pos(i, r10_bio);
1164
1165         while (atomic_dec_and_test(&r10_bio->remaining)) {
1166                 if (r10_bio->master_bio == NULL) {
1167                         /* the primary of several recovery bios */
1168                         md_done_sync(mddev, r10_bio->sectors, 1);
1169                         put_buf(r10_bio);
1170                         break;
1171                 } else {
1172                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1173                         put_buf(r10_bio);
1174                         r10_bio = r10_bio2;
1175                 }
1176         }
1177         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1178         return 0;
1179 }
1180
1181 /*
1182  * Note: sync and recover and handled very differently for raid10
1183  * This code is for resync.
1184  * For resync, we read through virtual addresses and read all blocks.
1185  * If there is any error, we schedule a write.  The lowest numbered
1186  * drive is authoritative.
1187  * However requests come for physical address, so we need to map.
1188  * For every physical address there are raid_disks/copies virtual addresses,
1189  * which is always are least one, but is not necessarly an integer.
1190  * This means that a physical address can span multiple chunks, so we may
1191  * have to submit multiple io requests for a single sync request.
1192  */
1193 /*
1194  * We check if all blocks are in-sync and only write to blocks that
1195  * aren't in sync
1196  */
1197 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1198 {
1199         conf_t *conf = mddev_to_conf(mddev);
1200         int i, first;
1201         struct bio *tbio, *fbio;
1202
1203         atomic_set(&r10_bio->remaining, 1);
1204
1205         /* find the first device with a block */
1206         for (i=0; i<conf->copies; i++)
1207                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1208                         break;
1209
1210         if (i == conf->copies)
1211                 goto done;
1212
1213         first = i;
1214         fbio = r10_bio->devs[i].bio;
1215
1216         /* now find blocks with errors */
1217         for (i=0 ; i < conf->copies ; i++) {
1218                 int  j, d;
1219                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1220
1221                 tbio = r10_bio->devs[i].bio;
1222
1223                 if (tbio->bi_end_io != end_sync_read)
1224                         continue;
1225                 if (i == first)
1226                         continue;
1227                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1228                         /* We know that the bi_io_vec layout is the same for
1229                          * both 'first' and 'i', so we just compare them.
1230                          * All vec entries are PAGE_SIZE;
1231                          */
1232                         for (j = 0; j < vcnt; j++)
1233                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1234                                            page_address(tbio->bi_io_vec[j].bv_page),
1235                                            PAGE_SIZE))
1236                                         break;
1237                         if (j == vcnt)
1238                                 continue;
1239                         mddev->resync_mismatches += r10_bio->sectors;
1240                 }
1241                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1242                         /* Don't fix anything. */
1243                         continue;
1244                 /* Ok, we need to write this bio
1245                  * First we need to fixup bv_offset, bv_len and
1246                  * bi_vecs, as the read request might have corrupted these
1247                  */
1248                 tbio->bi_vcnt = vcnt;
1249                 tbio->bi_size = r10_bio->sectors << 9;
1250                 tbio->bi_idx = 0;
1251                 tbio->bi_phys_segments = 0;
1252                 tbio->bi_hw_segments = 0;
1253                 tbio->bi_hw_front_size = 0;
1254                 tbio->bi_hw_back_size = 0;
1255                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1256                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1257                 tbio->bi_next = NULL;
1258                 tbio->bi_rw = WRITE;
1259                 tbio->bi_private = r10_bio;
1260                 tbio->bi_sector = r10_bio->devs[i].addr;
1261
1262                 for (j=0; j < vcnt ; j++) {
1263                         tbio->bi_io_vec[j].bv_offset = 0;
1264                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1265
1266                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1267                                page_address(fbio->bi_io_vec[j].bv_page),
1268                                PAGE_SIZE);
1269                 }
1270                 tbio->bi_end_io = end_sync_write;
1271
1272                 d = r10_bio->devs[i].devnum;
1273                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1274                 atomic_inc(&r10_bio->remaining);
1275                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1276
1277                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1278                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1279                 generic_make_request(tbio);
1280         }
1281
1282 done:
1283         if (atomic_dec_and_test(&r10_bio->remaining)) {
1284                 md_done_sync(mddev, r10_bio->sectors, 1);
1285                 put_buf(r10_bio);
1286         }
1287 }
1288
1289 /*
1290  * Now for the recovery code.
1291  * Recovery happens across physical sectors.
1292  * We recover all non-is_sync drives by finding the virtual address of
1293  * each, and then choose a working drive that also has that virt address.
1294  * There is a separate r10_bio for each non-in_sync drive.
1295  * Only the first two slots are in use. The first for reading,
1296  * The second for writing.
1297  *
1298  */
1299
1300 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1301 {
1302         conf_t *conf = mddev_to_conf(mddev);
1303         int i, d;
1304         struct bio *bio, *wbio;
1305
1306
1307         /* move the pages across to the second bio
1308          * and submit the write request
1309          */
1310         bio = r10_bio->devs[0].bio;
1311         wbio = r10_bio->devs[1].bio;
1312         for (i=0; i < wbio->bi_vcnt; i++) {
1313                 struct page *p = bio->bi_io_vec[i].bv_page;
1314                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1315                 wbio->bi_io_vec[i].bv_page = p;
1316         }
1317         d = r10_bio->devs[1].devnum;
1318
1319         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1320         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1321         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1322                 generic_make_request(wbio);
1323         else
1324                 bio_endio(wbio, wbio->bi_size, -EIO);
1325 }
1326
1327
1328 /*
1329  * This is a kernel thread which:
1330  *
1331  *      1.      Retries failed read operations on working mirrors.
1332  *      2.      Updates the raid superblock when problems encounter.
1333  *      3.      Performs writes following reads for array syncronising.
1334  */
1335
1336 static void raid10d(mddev_t *mddev)
1337 {
1338         r10bio_t *r10_bio;
1339         struct bio *bio;
1340         unsigned long flags;
1341         conf_t *conf = mddev_to_conf(mddev);
1342         struct list_head *head = &conf->retry_list;
1343         int unplug=0;
1344         mdk_rdev_t *rdev;
1345
1346         md_check_recovery(mddev);
1347
1348         for (;;) {
1349                 char b[BDEVNAME_SIZE];
1350                 spin_lock_irqsave(&conf->device_lock, flags);
1351
1352                 if (conf->pending_bio_list.head) {
1353                         bio = bio_list_get(&conf->pending_bio_list);
1354                         blk_remove_plug(mddev->queue);
1355                         spin_unlock_irqrestore(&conf->device_lock, flags);
1356                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1357                         if (bitmap_unplug(mddev->bitmap) != 0)
1358                                 printk("%s: bitmap file write failed!\n", mdname(mddev));
1359
1360                         while (bio) { /* submit pending writes */
1361                                 struct bio *next = bio->bi_next;
1362                                 bio->bi_next = NULL;
1363                                 generic_make_request(bio);
1364                                 bio = next;
1365                         }
1366                         unplug = 1;
1367
1368                         continue;
1369                 }
1370
1371                 if (list_empty(head))
1372                         break;
1373                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1374                 list_del(head->prev);
1375                 conf->nr_queued--;
1376                 spin_unlock_irqrestore(&conf->device_lock, flags);
1377
1378                 mddev = r10_bio->mddev;
1379                 conf = mddev_to_conf(mddev);
1380                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1381                         sync_request_write(mddev, r10_bio);
1382                         unplug = 1;
1383                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1384                         recovery_request_write(mddev, r10_bio);
1385                         unplug = 1;
1386                 } else {
1387                         int mirror;
1388                         /* we got a read error. Maybe the drive is bad.  Maybe just
1389                          * the block and we can fix it.
1390                          * We freeze all other IO, and try reading the block from
1391                          * other devices.  When we find one, we re-write
1392                          * and check it that fixes the read error.
1393                          * This is all done synchronously while the array is
1394                          * frozen.
1395                          */
1396                         int sect = 0; /* Offset from r10_bio->sector */
1397                         int sectors = r10_bio->sectors;
1398                         freeze_array(conf);
1399                         if (mddev->ro == 0) while(sectors) {
1400                                 int s = sectors;
1401                                 int sl = r10_bio->read_slot;
1402                                 int success = 0;
1403
1404                                 if (s > (PAGE_SIZE>>9))
1405                                         s = PAGE_SIZE >> 9;
1406
1407                                 do {
1408                                         int d = r10_bio->devs[sl].devnum;
1409                                         rdev = conf->mirrors[d].rdev;
1410                                         if (rdev &&
1411                                             test_bit(In_sync, &rdev->flags) &&
1412                                             sync_page_io(rdev->bdev,
1413                                                          r10_bio->devs[sl].addr +
1414                                                          sect + rdev->data_offset,
1415                                                          s<<9,
1416                                                          conf->tmppage, READ))
1417                                                 success = 1;
1418                                         else {
1419                                                 sl++;
1420                                                 if (sl == conf->copies)
1421                                                         sl = 0;
1422                                         }
1423                                 } while (!success && sl != r10_bio->read_slot);
1424
1425                                 if (success) {
1426                                         /* write it back and re-read */
1427                                         while (sl != r10_bio->read_slot) {
1428                                                 int d;
1429                                                 if (sl==0)
1430                                                         sl = conf->copies;
1431                                                 sl--;
1432                                                 d = r10_bio->devs[sl].devnum;
1433                                                 rdev = conf->mirrors[d].rdev;
1434                                                 if (rdev &&
1435                                                     test_bit(In_sync, &rdev->flags)) {
1436                                                         if (sync_page_io(rdev->bdev,
1437                                                                          r10_bio->devs[sl].addr +
1438                                                                          sect + rdev->data_offset,
1439                                                                          s<<9, conf->tmppage, WRITE) == 0 ||
1440                                                             sync_page_io(rdev->bdev,
1441                                                                          r10_bio->devs[sl].addr +
1442                                                                          sect + rdev->data_offset,
1443                                                                          s<<9, conf->tmppage, READ) == 0) {
1444                                                                 /* Well, this device is dead */
1445                                                                 md_error(mddev, rdev);
1446                                                         }
1447                                                 }
1448                                         }
1449                                 } else {
1450                                         /* Cannot read from anywhere -- bye bye array */
1451                                         md_error(mddev, conf->mirrors[r10_bio->devs[r10_bio->read_slot].devnum].rdev);
1452                                         break;
1453                                 }
1454                                 sectors -= s;
1455                                 sect += s;
1456                         }
1457
1458                         unfreeze_array(conf);
1459
1460                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1461                         r10_bio->devs[r10_bio->read_slot].bio =
1462                                 mddev->ro ? IO_BLOCKED : NULL;
1463                         bio_put(bio);
1464                         mirror = read_balance(conf, r10_bio);
1465                         if (mirror == -1) {
1466                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1467                                        " read error for block %llu\n",
1468                                        bdevname(bio->bi_bdev,b),
1469                                        (unsigned long long)r10_bio->sector);
1470                                 raid_end_bio_io(r10_bio);
1471                         } else {
1472                                 rdev = conf->mirrors[mirror].rdev;
1473                                 if (printk_ratelimit())
1474                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1475                                                " another mirror\n",
1476                                                bdevname(rdev->bdev,b),
1477                                                (unsigned long long)r10_bio->sector);
1478                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1479                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1480                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1481                                         + rdev->data_offset;
1482                                 bio->bi_bdev = rdev->bdev;
1483                                 bio->bi_rw = READ;
1484                                 bio->bi_private = r10_bio;
1485                                 bio->bi_end_io = raid10_end_read_request;
1486                                 unplug = 1;
1487                                 generic_make_request(bio);
1488                         }
1489                 }
1490         }
1491         spin_unlock_irqrestore(&conf->device_lock, flags);
1492         if (unplug)
1493                 unplug_slaves(mddev);
1494 }
1495
1496
1497 static int init_resync(conf_t *conf)
1498 {
1499         int buffs;
1500
1501         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1502         if (conf->r10buf_pool)
1503                 BUG();
1504         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1505         if (!conf->r10buf_pool)
1506                 return -ENOMEM;
1507         conf->next_resync = 0;
1508         return 0;
1509 }
1510
1511 /*
1512  * perform a "sync" on one "block"
1513  *
1514  * We need to make sure that no normal I/O request - particularly write
1515  * requests - conflict with active sync requests.
1516  *
1517  * This is achieved by tracking pending requests and a 'barrier' concept
1518  * that can be installed to exclude normal IO requests.
1519  *
1520  * Resync and recovery are handled very differently.
1521  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1522  *
1523  * For resync, we iterate over virtual addresses, read all copies,
1524  * and update if there are differences.  If only one copy is live,
1525  * skip it.
1526  * For recovery, we iterate over physical addresses, read a good
1527  * value for each non-in_sync drive, and over-write.
1528  *
1529  * So, for recovery we may have several outstanding complex requests for a
1530  * given address, one for each out-of-sync device.  We model this by allocating
1531  * a number of r10_bio structures, one for each out-of-sync device.
1532  * As we setup these structures, we collect all bio's together into a list
1533  * which we then process collectively to add pages, and then process again
1534  * to pass to generic_make_request.
1535  *
1536  * The r10_bio structures are linked using a borrowed master_bio pointer.
1537  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1538  * has its remaining count decremented to 0, the whole complex operation
1539  * is complete.
1540  *
1541  */
1542
1543 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1544 {
1545         conf_t *conf = mddev_to_conf(mddev);
1546         r10bio_t *r10_bio;
1547         struct bio *biolist = NULL, *bio;
1548         sector_t max_sector, nr_sectors;
1549         int disk;
1550         int i;
1551         int max_sync;
1552         int sync_blocks;
1553
1554         sector_t sectors_skipped = 0;
1555         int chunks_skipped = 0;
1556
1557         if (!conf->r10buf_pool)
1558                 if (init_resync(conf))
1559                         return 0;
1560
1561  skipped:
1562         max_sector = mddev->size << 1;
1563         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1564                 max_sector = mddev->resync_max_sectors;
1565         if (sector_nr >= max_sector) {
1566                 /* If we aborted, we need to abort the
1567                  * sync on the 'current' bitmap chucks (there can
1568                  * be several when recovering multiple devices).
1569                  * as we may have started syncing it but not finished.
1570                  * We can find the current address in
1571                  * mddev->curr_resync, but for recovery,
1572                  * we need to convert that to several
1573                  * virtual addresses.
1574                  */
1575                 if (mddev->curr_resync < max_sector) { /* aborted */
1576                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1577                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1578                                                 &sync_blocks, 1);
1579                         else for (i=0; i<conf->raid_disks; i++) {
1580                                 sector_t sect =
1581                                         raid10_find_virt(conf, mddev->curr_resync, i);
1582                                 bitmap_end_sync(mddev->bitmap, sect,
1583                                                 &sync_blocks, 1);
1584                         }
1585                 } else /* completed sync */
1586                         conf->fullsync = 0;
1587
1588                 bitmap_close_sync(mddev->bitmap);
1589                 close_sync(conf);
1590                 *skipped = 1;
1591                 return sectors_skipped;
1592         }
1593         if (chunks_skipped >= conf->raid_disks) {
1594                 /* if there has been nothing to do on any drive,
1595                  * then there is nothing to do at all..
1596                  */
1597                 *skipped = 1;
1598                 return (max_sector - sector_nr) + sectors_skipped;
1599         }
1600
1601         /* make sure whole request will fit in a chunk - if chunks
1602          * are meaningful
1603          */
1604         if (conf->near_copies < conf->raid_disks &&
1605             max_sector > (sector_nr | conf->chunk_mask))
1606                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1607         /*
1608          * If there is non-resync activity waiting for us then
1609          * put in a delay to throttle resync.
1610          */
1611         if (!go_faster && conf->nr_waiting)
1612                 msleep_interruptible(1000);
1613
1614         /* Again, very different code for resync and recovery.
1615          * Both must result in an r10bio with a list of bios that
1616          * have bi_end_io, bi_sector, bi_bdev set,
1617          * and bi_private set to the r10bio.
1618          * For recovery, we may actually create several r10bios
1619          * with 2 bios in each, that correspond to the bios in the main one.
1620          * In this case, the subordinate r10bios link back through a
1621          * borrowed master_bio pointer, and the counter in the master
1622          * includes a ref from each subordinate.
1623          */
1624         /* First, we decide what to do and set ->bi_end_io
1625          * To end_sync_read if we want to read, and
1626          * end_sync_write if we will want to write.
1627          */
1628
1629         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1630         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1631                 /* recovery... the complicated one */
1632                 int i, j, k;
1633                 r10_bio = NULL;
1634
1635                 for (i=0 ; i<conf->raid_disks; i++)
1636                         if (conf->mirrors[i].rdev &&
1637                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1638                                 int still_degraded = 0;
1639                                 /* want to reconstruct this device */
1640                                 r10bio_t *rb2 = r10_bio;
1641                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1642                                 int must_sync;
1643                                 /* Unless we are doing a full sync, we only need
1644                                  * to recover the block if it is set in the bitmap
1645                                  */
1646                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1647                                                               &sync_blocks, 1);
1648                                 if (sync_blocks < max_sync)
1649                                         max_sync = sync_blocks;
1650                                 if (!must_sync &&
1651                                     !conf->fullsync) {
1652                                         /* yep, skip the sync_blocks here, but don't assume
1653                                          * that there will never be anything to do here
1654                                          */
1655                                         chunks_skipped = -1;
1656                                         continue;
1657                                 }
1658
1659                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1660                                 raise_barrier(conf, rb2 != NULL);
1661                                 atomic_set(&r10_bio->remaining, 0);
1662
1663                                 r10_bio->master_bio = (struct bio*)rb2;
1664                                 if (rb2)
1665                                         atomic_inc(&rb2->remaining);
1666                                 r10_bio->mddev = mddev;
1667                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1668                                 r10_bio->sector = sect;
1669
1670                                 raid10_find_phys(conf, r10_bio);
1671                                 /* Need to check if this section will still be
1672                                  * degraded
1673                                  */
1674                                 for (j=0; j<conf->copies;j++) {
1675                                         int d = r10_bio->devs[j].devnum;
1676                                         if (conf->mirrors[d].rdev == NULL ||
1677                                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1678                                                 still_degraded = 1;
1679                                 }
1680                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1681                                                               &sync_blocks, still_degraded);
1682
1683                                 for (j=0; j<conf->copies;j++) {
1684                                         int d = r10_bio->devs[j].devnum;
1685                                         if (conf->mirrors[d].rdev &&
1686                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1687                                                 /* This is where we read from */
1688                                                 bio = r10_bio->devs[0].bio;
1689                                                 bio->bi_next = biolist;
1690                                                 biolist = bio;
1691                                                 bio->bi_private = r10_bio;
1692                                                 bio->bi_end_io = end_sync_read;
1693                                                 bio->bi_rw = 0;
1694                                                 bio->bi_sector = r10_bio->devs[j].addr +
1695                                                         conf->mirrors[d].rdev->data_offset;
1696                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1697                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1698                                                 atomic_inc(&r10_bio->remaining);
1699                                                 /* and we write to 'i' */
1700
1701                                                 for (k=0; k<conf->copies; k++)
1702                                                         if (r10_bio->devs[k].devnum == i)
1703                                                                 break;
1704                                                 bio = r10_bio->devs[1].bio;
1705                                                 bio->bi_next = biolist;
1706                                                 biolist = bio;
1707                                                 bio->bi_private = r10_bio;
1708                                                 bio->bi_end_io = end_sync_write;
1709                                                 bio->bi_rw = 1;
1710                                                 bio->bi_sector = r10_bio->devs[k].addr +
1711                                                         conf->mirrors[i].rdev->data_offset;
1712                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1713
1714                                                 r10_bio->devs[0].devnum = d;
1715                                                 r10_bio->devs[1].devnum = i;
1716
1717                                                 break;
1718                                         }
1719                                 }
1720                                 if (j == conf->copies) {
1721                                         /* Cannot recover, so abort the recovery */
1722                                         put_buf(r10_bio);
1723                                         r10_bio = rb2;
1724                                         if (!test_and_set_bit(MD_RECOVERY_ERR, &mddev->recovery))
1725                                                 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1726                                                        mdname(mddev));
1727                                         break;
1728                                 }
1729                         }
1730                 if (biolist == NULL) {
1731                         while (r10_bio) {
1732                                 r10bio_t *rb2 = r10_bio;
1733                                 r10_bio = (r10bio_t*) rb2->master_bio;
1734                                 rb2->master_bio = NULL;
1735                                 put_buf(rb2);
1736                         }
1737                         goto giveup;
1738                 }
1739         } else {
1740                 /* resync. Schedule a read for every block at this virt offset */
1741                 int count = 0;
1742
1743                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1744                                        &sync_blocks, mddev->degraded) &&
1745                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1746                         /* We can skip this block */
1747                         *skipped = 1;
1748                         return sync_blocks + sectors_skipped;
1749                 }
1750                 if (sync_blocks < max_sync)
1751                         max_sync = sync_blocks;
1752                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1753
1754                 r10_bio->mddev = mddev;
1755                 atomic_set(&r10_bio->remaining, 0);
1756                 raise_barrier(conf, 0);
1757                 conf->next_resync = sector_nr;
1758
1759                 r10_bio->master_bio = NULL;
1760                 r10_bio->sector = sector_nr;
1761                 set_bit(R10BIO_IsSync, &r10_bio->state);
1762                 raid10_find_phys(conf, r10_bio);
1763                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1764
1765                 for (i=0; i<conf->copies; i++) {
1766                         int d = r10_bio->devs[i].devnum;
1767                         bio = r10_bio->devs[i].bio;
1768                         bio->bi_end_io = NULL;
1769                         if (conf->mirrors[d].rdev == NULL ||
1770                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1771                                 continue;
1772                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1773                         atomic_inc(&r10_bio->remaining);
1774                         bio->bi_next = biolist;
1775                         biolist = bio;
1776                         bio->bi_private = r10_bio;
1777                         bio->bi_end_io = end_sync_read;
1778                         bio->bi_rw = 0;
1779                         bio->bi_sector = r10_bio->devs[i].addr +
1780                                 conf->mirrors[d].rdev->data_offset;
1781                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1782                         count++;
1783                 }
1784
1785                 if (count < 2) {
1786                         for (i=0; i<conf->copies; i++) {
1787                                 int d = r10_bio->devs[i].devnum;
1788                                 if (r10_bio->devs[i].bio->bi_end_io)
1789                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1790                         }
1791                         put_buf(r10_bio);
1792                         biolist = NULL;
1793                         goto giveup;
1794                 }
1795         }
1796
1797         for (bio = biolist; bio ; bio=bio->bi_next) {
1798
1799                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1800                 if (bio->bi_end_io)
1801                         bio->bi_flags |= 1 << BIO_UPTODATE;
1802                 bio->bi_vcnt = 0;
1803                 bio->bi_idx = 0;
1804                 bio->bi_phys_segments = 0;
1805                 bio->bi_hw_segments = 0;
1806                 bio->bi_size = 0;
1807         }
1808
1809         nr_sectors = 0;
1810         if (sector_nr + max_sync < max_sector)
1811                 max_sector = sector_nr + max_sync;
1812         do {
1813                 struct page *page;
1814                 int len = PAGE_SIZE;
1815                 disk = 0;
1816                 if (sector_nr + (len>>9) > max_sector)
1817                         len = (max_sector - sector_nr) << 9;
1818                 if (len == 0)
1819                         break;
1820                 for (bio= biolist ; bio ; bio=bio->bi_next) {
1821                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1822                         if (bio_add_page(bio, page, len, 0) == 0) {
1823                                 /* stop here */
1824                                 struct bio *bio2;
1825                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1826                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1827                                         /* remove last page from this bio */
1828                                         bio2->bi_vcnt--;
1829                                         bio2->bi_size -= len;
1830                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1831                                 }
1832                                 goto bio_full;
1833                         }
1834                         disk = i;
1835                 }
1836                 nr_sectors += len>>9;
1837                 sector_nr += len>>9;
1838         } while (biolist->bi_vcnt < RESYNC_PAGES);
1839  bio_full:
1840         r10_bio->sectors = nr_sectors;
1841
1842         while (biolist) {
1843                 bio = biolist;
1844                 biolist = biolist->bi_next;
1845
1846                 bio->bi_next = NULL;
1847                 r10_bio = bio->bi_private;
1848                 r10_bio->sectors = nr_sectors;
1849
1850                 if (bio->bi_end_io == end_sync_read) {
1851                         md_sync_acct(bio->bi_bdev, nr_sectors);
1852                         generic_make_request(bio);
1853                 }
1854         }
1855
1856         if (sectors_skipped)
1857                 /* pretend they weren't skipped, it makes
1858                  * no important difference in this case
1859                  */
1860                 md_done_sync(mddev, sectors_skipped, 1);
1861
1862         return sectors_skipped + nr_sectors;
1863  giveup:
1864         /* There is nowhere to write, so all non-sync
1865          * drives must be failed, so try the next chunk...
1866          */
1867         {
1868         sector_t sec = max_sector - sector_nr;
1869         sectors_skipped += sec;
1870         chunks_skipped ++;
1871         sector_nr = max_sector;
1872         goto skipped;
1873         }
1874 }
1875
1876 static int run(mddev_t *mddev)
1877 {
1878         conf_t *conf;
1879         int i, disk_idx;
1880         mirror_info_t *disk;
1881         mdk_rdev_t *rdev;
1882         struct list_head *tmp;
1883         int nc, fc;
1884         sector_t stride, size;
1885
1886         if (mddev->level != 10) {
1887                 printk(KERN_ERR "raid10: %s: raid level not set correctly... (%d)\n",
1888                        mdname(mddev), mddev->level);
1889                 goto out;
1890         }
1891         nc = mddev->layout & 255;
1892         fc = (mddev->layout >> 8) & 255;
1893         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
1894             (mddev->layout >> 16)) {
1895                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
1896                        mdname(mddev), mddev->layout);
1897                 goto out;
1898         }
1899         /*
1900          * copy the already verified devices into our private RAID10
1901          * bookkeeping area. [whatever we allocate in run(),
1902          * should be freed in stop()]
1903          */
1904         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1905         mddev->private = conf;
1906         if (!conf) {
1907                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1908                         mdname(mddev));
1909                 goto out;
1910         }
1911         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1912                                  GFP_KERNEL);
1913         if (!conf->mirrors) {
1914                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1915                        mdname(mddev));
1916                 goto out_free_conf;
1917         }
1918
1919         conf->tmppage = alloc_page(GFP_KERNEL);
1920         if (!conf->tmppage)
1921                 goto out_free_conf;
1922
1923         conf->near_copies = nc;
1924         conf->far_copies = fc;
1925         conf->copies = nc*fc;
1926         conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
1927         conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
1928         stride = mddev->size >> (conf->chunk_shift-1);
1929         sector_div(stride, fc);
1930         conf->stride = stride << conf->chunk_shift;
1931
1932         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
1933                                                 r10bio_pool_free, conf);
1934         if (!conf->r10bio_pool) {
1935                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1936                         mdname(mddev));
1937                 goto out_free_conf;
1938         }
1939
1940         ITERATE_RDEV(mddev, rdev, tmp) {
1941                 disk_idx = rdev->raid_disk;
1942                 if (disk_idx >= mddev->raid_disks
1943                     || disk_idx < 0)
1944                         continue;
1945                 disk = conf->mirrors + disk_idx;
1946
1947                 disk->rdev = rdev;
1948
1949                 blk_queue_stack_limits(mddev->queue,
1950                                        rdev->bdev->bd_disk->queue);
1951                 /* as we don't honour merge_bvec_fn, we must never risk
1952                  * violating it, so limit ->max_sector to one PAGE, as
1953                  * a one page request is never in violation.
1954                  */
1955                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1956                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1957                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
1958
1959                 disk->head_position = 0;
1960                 if (!test_bit(Faulty, &rdev->flags) && test_bit(In_sync, &rdev->flags))
1961                         conf->working_disks++;
1962         }
1963         conf->raid_disks = mddev->raid_disks;
1964         conf->mddev = mddev;
1965         spin_lock_init(&conf->device_lock);
1966         INIT_LIST_HEAD(&conf->retry_list);
1967
1968         spin_lock_init(&conf->resync_lock);
1969         init_waitqueue_head(&conf->wait_barrier);
1970
1971         /* need to check that every block has at least one working mirror */
1972         if (!enough(conf)) {
1973                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
1974                        mdname(mddev));
1975                 goto out_free_conf;
1976         }
1977
1978         mddev->degraded = 0;
1979         for (i = 0; i < conf->raid_disks; i++) {
1980
1981                 disk = conf->mirrors + i;
1982
1983                 if (!disk->rdev) {
1984                         disk->head_position = 0;
1985                         mddev->degraded++;
1986                 }
1987         }
1988
1989
1990         mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
1991         if (!mddev->thread) {
1992                 printk(KERN_ERR
1993                        "raid10: couldn't allocate thread for %s\n",
1994                        mdname(mddev));
1995                 goto out_free_conf;
1996         }
1997
1998         printk(KERN_INFO
1999                 "raid10: raid set %s active with %d out of %d devices\n",
2000                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2001                 mddev->raid_disks);
2002         /*
2003          * Ok, everything is just fine now
2004          */
2005         size = conf->stride * conf->raid_disks;
2006         sector_div(size, conf->near_copies);
2007         mddev->array_size = size/2;
2008         mddev->resync_max_sectors = size;
2009
2010         mddev->queue->unplug_fn = raid10_unplug;
2011         mddev->queue->issue_flush_fn = raid10_issue_flush;
2012
2013         /* Calculate max read-ahead size.
2014          * We need to readahead at least twice a whole stripe....
2015          * maybe...
2016          */
2017         {
2018                 int stripe = conf->raid_disks * mddev->chunk_size / PAGE_CACHE_SIZE;
2019                 stripe /= conf->near_copies;
2020                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2021                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2022         }
2023
2024         if (conf->near_copies < mddev->raid_disks)
2025                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2026         return 0;
2027
2028 out_free_conf:
2029         if (conf->r10bio_pool)
2030                 mempool_destroy(conf->r10bio_pool);
2031         put_page(conf->tmppage);
2032         kfree(conf->mirrors);
2033         kfree(conf);
2034         mddev->private = NULL;
2035 out:
2036         return -EIO;
2037 }
2038
2039 static int stop(mddev_t *mddev)
2040 {
2041         conf_t *conf = mddev_to_conf(mddev);
2042
2043         md_unregister_thread(mddev->thread);
2044         mddev->thread = NULL;
2045         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2046         if (conf->r10bio_pool)
2047                 mempool_destroy(conf->r10bio_pool);
2048         kfree(conf->mirrors);
2049         kfree(conf);
2050         mddev->private = NULL;
2051         return 0;
2052 }
2053
2054 static void raid10_quiesce(mddev_t *mddev, int state)
2055 {
2056         conf_t *conf = mddev_to_conf(mddev);
2057
2058         switch(state) {
2059         case 1:
2060                 raise_barrier(conf, 0);
2061                 break;
2062         case 0:
2063                 lower_barrier(conf);
2064                 break;
2065         }
2066         if (mddev->thread) {
2067                 if (mddev->bitmap)
2068                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2069                 else
2070                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2071                 md_wakeup_thread(mddev->thread);
2072         }
2073 }
2074
2075 static mdk_personality_t raid10_personality =
2076 {
2077         .name           = "raid10",
2078         .owner          = THIS_MODULE,
2079         .make_request   = make_request,
2080         .run            = run,
2081         .stop           = stop,
2082         .status         = status,
2083         .error_handler  = error,
2084         .hot_add_disk   = raid10_add_disk,
2085         .hot_remove_disk= raid10_remove_disk,
2086         .spare_active   = raid10_spare_active,
2087         .sync_request   = sync_request,
2088         .quiesce        = raid10_quiesce,
2089 };
2090
2091 static int __init raid_init(void)
2092 {
2093         return register_md_personality(RAID10, &raid10_personality);
2094 }
2095
2096 static void raid_exit(void)
2097 {
2098         unregister_md_personality(RAID10);
2099 }
2100
2101 module_init(raid_init);
2102 module_exit(raid_exit);
2103 MODULE_LICENSE("GPL");
2104 MODULE_ALIAS("md-personality-9"); /* RAID10 */