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