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