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(unsigned int __nocast 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(unsigned int __nocast 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
 500         raid10_find_phys(conf, r10_bio);
 501         rcu_read_lock();
 502         /*
 503          * Check if we can balance. We can balance on the whole
 504          * device if no resync is going on, or below the resync window.
 505          * We take the first readable disk when above the resync window.
 506          */
 507         if (conf->mddev->recovery_cp < MaxSector
 508             && (this_sector + sectors >= conf->next_resync)) {
 509                 /* make sure that disk is operational */
 510                 slot = 0;
 511                 disk = r10_bio->devs[slot].devnum;
 512
 513                 while (!conf->mirrors[disk].rdev ||
 514                        !conf->mirrors[disk].rdev->in_sync) {
 515                         slot++;
 516                         if (slot == conf->copies) {
 517                                 slot = 0;
 518                                 disk = -1;
 519                                 break;
 520                         }
 521                         disk = r10_bio->devs[slot].devnum;
 522                 }
 523                 goto rb_out;
 524         }
 525
 526
 527         /* make sure the disk is operational */
 528         slot = 0;
 529         disk = r10_bio->devs[slot].devnum;
 530         while (!conf->mirrors[disk].rdev ||
 531                !conf->mirrors[disk].rdev->in_sync) {
 532                 slot ++;
 533                 if (slot == conf->copies) {
 534                         disk = -1;
 535                         goto rb_out;
 536                 }
 537                 disk = r10_bio->devs[slot].devnum;
 538         }
 539
 540
 541         current_distance = abs(r10_bio->devs[slot].addr -
 542                                conf->mirrors[disk].head_position);
 543
 544         /* Find the disk whose head is closest */
 545
 546         for (nslot = slot; nslot < conf->copies; nslot++) {
 547                 int ndisk = r10_bio->devs[nslot].devnum;
 548
 549
 550                 if (!conf->mirrors[ndisk].rdev ||
 551                     !conf->mirrors[ndisk].rdev->in_sync)
 552                         continue;
 553
 554                 if (!atomic_read(&conf->mirrors[ndisk].rdev->nr_pending)) {
 555                         disk = ndisk;
 556                         slot = nslot;
 557                         break;
 558                 }
 559                 new_distance = abs(r10_bio->devs[nslot].addr -
 560                                    conf->mirrors[ndisk].head_position);
 561                 if (new_distance < current_distance) {
 562                         current_distance = new_distance;
 563                         disk = ndisk;
 564                         slot = nslot;
 565                 }
 566         }
 567
 568 rb_out:
 569         r10_bio->read_slot = slot;
 570 /*      conf->next_seq_sect = this_sector + sectors;*/
 571
 572         if (disk >= 0 && conf->mirrors[disk].rdev)
 573                 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
 574         rcu_read_unlock();
 575
 576         return disk;
 577 }
 578
 579 static void unplug_slaves(mddev_t *mddev)
 580 {
 581         conf_t *conf = mddev_to_conf(mddev);
 582         int i;
 583
 584         rcu_read_lock();
 585         for (i=0; i<mddev->raid_disks; i++) {
 586                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
 587                 if (rdev && !rdev->faulty && atomic_read(&rdev->nr_pending)) {
 588                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 589
 590                         atomic_inc(&rdev->nr_pending);
 591                         rcu_read_unlock();
 592
 593                         if (r_queue->unplug_fn)
 594                                 r_queue->unplug_fn(r_queue);
 595
 596                         rdev_dec_pending(rdev, mddev);
 597                         rcu_read_lock();
 598                 }
 599         }
 600         rcu_read_unlock();
 601 }
 602
 603 static void raid10_unplug(request_queue_t *q)
 604 {
 605         unplug_slaves(q->queuedata);
 606 }
 607
 608 static int raid10_issue_flush(request_queue_t *q, struct gendisk *disk,
 609                              sector_t *error_sector)
 610 {
 611         mddev_t *mddev = q->queuedata;
 612         conf_t *conf = mddev_to_conf(mddev);
 613         int i, ret = 0;
 614
 615         rcu_read_lock();
 616         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
 617                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
 618                 if (rdev && !rdev->faulty) {
 619                         struct block_device *bdev = rdev->bdev;
 620                         request_queue_t *r_queue = bdev_get_queue(bdev);
 621
 622                         if (!r_queue->issue_flush_fn)
 623                                 ret = -EOPNOTSUPP;
 624                         else {
 625                                 atomic_inc(&rdev->nr_pending);
 626                                 rcu_read_unlock();
 627                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
 628                                                               error_sector);
 629                                 rdev_dec_pending(rdev, mddev);
 630                                 rcu_read_lock();
 631                         }
 632                 }
 633         }
 634         rcu_read_unlock();
 635         return ret;
 636 }
 637
 638 /*
 639  * Throttle resync depth, so that we can both get proper overlapping of
 640  * requests, but are still able to handle normal requests quickly.
 641  */
 642 #define RESYNC_DEPTH 32
 643
 644 static void device_barrier(conf_t *conf, sector_t sect)
 645 {
 646         spin_lock_irq(&conf->resync_lock);
 647         wait_event_lock_irq(conf->wait_idle, !waitqueue_active(&conf->wait_resume),
 648                             conf->resync_lock, unplug_slaves(conf->mddev));
 649
 650         if (!conf->barrier++) {
 651                 wait_event_lock_irq(conf->wait_idle, !conf->nr_pending,
 652                                     conf->resync_lock, unplug_slaves(conf->mddev));
 653                 if (conf->nr_pending)
 654                         BUG();
 655         }
 656         wait_event_lock_irq(conf->wait_resume, conf->barrier < RESYNC_DEPTH,
 657                             conf->resync_lock, unplug_slaves(conf->mddev));
 658         conf->next_resync = sect;
 659         spin_unlock_irq(&conf->resync_lock);
 660 }
 661
 662 static int make_request(request_queue_t *q, struct bio * bio)
 663 {
 664         mddev_t *mddev = q->queuedata;
 665         conf_t *conf = mddev_to_conf(mddev);
 666         mirror_info_t *mirror;
 667         r10bio_t *r10_bio;
 668         struct bio *read_bio;
 669         int i;
 670         int chunk_sects = conf->chunk_mask + 1;
 671
 672         if (unlikely(bio_barrier(bio))) {
 673                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
 674                 return 0;
 675         }
 676
 677         /* If this request crosses a chunk boundary, we need to
 678          * split it.  This will only happen for 1 PAGE (or less) requests.
 679          */
 680         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 681                       > chunk_sects &&
 682                     conf->near_copies < conf->raid_disks)) {
 683                 struct bio_pair *bp;
 684                 /* Sanity check -- queue functions should prevent this happening */
 685                 if (bio->bi_vcnt != 1 ||
 686                     bio->bi_idx != 0)
 687                         goto bad_map;
 688                 /* This is a one page bio that upper layers
 689                  * refuse to split for us, so we need to split it.
 690                  */
 691                 bp = bio_split(bio, bio_split_pool,
 692                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 693                 if (make_request(q, &bp->bio1))
 694                         generic_make_request(&bp->bio1);
 695                 if (make_request(q, &bp->bio2))
 696                         generic_make_request(&bp->bio2);
 697
 698                 bio_pair_release(bp);
 699                 return 0;
 700         bad_map:
 701                 printk("raid10_make_request bug: can't convert block across chunks"
 702                        " or bigger than %dk %llu %d\n", chunk_sects/2,
 703                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 704
 705                 bio_io_error(bio, bio->bi_size);
 706                 return 0;
 707         }
 708
 709         md_write_start(mddev, bio);
 710
 711         /*
 712          * Register the new request and wait if the reconstruction
 713          * thread has put up a bar for new requests.
 714          * Continue immediately if no resync is active currently.
 715          */
 716         spin_lock_irq(&conf->resync_lock);
 717         wait_event_lock_irq(conf->wait_resume, !conf->barrier, conf->resync_lock, );
 718         conf->nr_pending++;
 719         spin_unlock_irq(&conf->resync_lock);
 720
 721         if (bio_data_dir(bio)==WRITE) {
 722                 disk_stat_inc(mddev->gendisk, writes);
 723                 disk_stat_add(mddev->gendisk, write_sectors, bio_sectors(bio));
 724         } else {
 725                 disk_stat_inc(mddev->gendisk, reads);
 726                 disk_stat_add(mddev->gendisk, read_sectors, bio_sectors(bio));
 727         }
 728
 729         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 730
 731         r10_bio->master_bio = bio;
 732         r10_bio->sectors = bio->bi_size >> 9;
 733
 734         r10_bio->mddev = mddev;
 735         r10_bio->sector = bio->bi_sector;
 736
 737         if (bio_data_dir(bio) == READ) {
 738                 /*
 739                  * read balancing logic:
 740                  */
 741                 int disk = read_balance(conf, r10_bio);
 742                 int slot = r10_bio->read_slot;
 743                 if (disk < 0) {
 744                         raid_end_bio_io(r10_bio);
 745                         return 0;
 746                 }
 747                 mirror = conf->mirrors + disk;
 748
 749                 read_bio = bio_clone(bio, GFP_NOIO);
 750
 751                 r10_bio->devs[slot].bio = read_bio;
 752
 753                 read_bio->bi_sector = r10_bio->devs[slot].addr +
 754                         mirror->rdev->data_offset;
 755                 read_bio->bi_bdev = mirror->rdev->bdev;
 756                 read_bio->bi_end_io = raid10_end_read_request;
 757                 read_bio->bi_rw = READ;
 758                 read_bio->bi_private = r10_bio;
 759
 760                 generic_make_request(read_bio);
 761                 return 0;
 762         }
 763
 764         /*
 765          * WRITE:
 766          */
 767         /* first select target devices under spinlock and
 768          * inc refcount on their rdev.  Record them by setting
 769          * bios[x] to bio
 770          */
 771         raid10_find_phys(conf, r10_bio);
 772         rcu_read_lock();
 773         for (i = 0;  i < conf->copies; i++) {
 774                 int d = r10_bio->devs[i].devnum;
 775                 if (conf->mirrors[d].rdev &&
 776                     !conf->mirrors[d].rdev->faulty) {
 777                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
 778                         r10_bio->devs[i].bio = bio;
 779                 } else
 780                         r10_bio->devs[i].bio = NULL;
 781         }
 782         rcu_read_unlock();
 783
 784         atomic_set(&r10_bio->remaining, 1);
 785
 786         for (i = 0; i < conf->copies; i++) {
 787                 struct bio *mbio;
 788                 int d = r10_bio->devs[i].devnum;
 789                 if (!r10_bio->devs[i].bio)
 790                         continue;
 791
 792                 mbio = bio_clone(bio, GFP_NOIO);
 793                 r10_bio->devs[i].bio = mbio;
 794
 795                 mbio->bi_sector = r10_bio->devs[i].addr+
 796                         conf->mirrors[d].rdev->data_offset;
 797                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
 798                 mbio->bi_end_io = raid10_end_write_request;
 799                 mbio->bi_rw = WRITE;
 800                 mbio->bi_private = r10_bio;
 801
 802                 atomic_inc(&r10_bio->remaining);
 803                 generic_make_request(mbio);
 804         }
 805
 806         if (atomic_dec_and_test(&r10_bio->remaining)) {
 807                 md_write_end(mddev);
 808                 raid_end_bio_io(r10_bio);
 809         }
 810
 811         return 0;
 812 }
 813
 814 static void status(struct seq_file *seq, mddev_t *mddev)
 815 {
 816         conf_t *conf = mddev_to_conf(mddev);
 817         int i;
 818
 819         if (conf->near_copies < conf->raid_disks)
 820                 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
 821         if (conf->near_copies > 1)
 822                 seq_printf(seq, " %d near-copies", conf->near_copies);
 823         if (conf->far_copies > 1)
 824                 seq_printf(seq, " %d far-copies", conf->far_copies);
 825
 826         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 827                                                 conf->working_disks);
 828         for (i = 0; i < conf->raid_disks; i++)
 829                 seq_printf(seq, "%s",
 830                               conf->mirrors[i].rdev &&
 831                               conf->mirrors[i].rdev->in_sync ? "U" : "_");
 832         seq_printf(seq, "]");
 833 }
 834
 835 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 836 {
 837         char b[BDEVNAME_SIZE];
 838         conf_t *conf = mddev_to_conf(mddev);
 839
 840         /*
 841          * If it is not operational, then we have already marked it as dead
 842          * else if it is the last working disks, ignore the error, let the
 843          * next level up know.
 844          * else mark the drive as failed
 845          */
 846         if (rdev->in_sync
 847             && conf->working_disks == 1)
 848                 /*
 849                  * Don't fail the drive, just return an IO error.
 850                  * The test should really be more sophisticated than
 851                  * "working_disks == 1", but it isn't critical, and
 852                  * can wait until we do more sophisticated "is the drive
 853                  * really dead" tests...
 854                  */
 855                 return;
 856         if (rdev->in_sync) {
 857                 mddev->degraded++;
 858                 conf->working_disks--;
 859                 /*
 860                  * if recovery is running, make sure it aborts.
 861                  */
 862                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 863         }
 864         rdev->in_sync = 0;
 865         rdev->faulty = 1;
 866         mddev->sb_dirty = 1;
 867         printk(KERN_ALERT "raid10: Disk failure on %s, disabling device. \n"
 868                 "       Operation continuing on %d devices\n",
 869                 bdevname(rdev->bdev,b), conf->working_disks);
 870 }
 871
 872 static void print_conf(conf_t *conf)
 873 {
 874         int i;
 875         mirror_info_t *tmp;
 876
 877         printk("RAID10 conf printout:\n");
 878         if (!conf) {
 879                 printk("(!conf)\n");
 880                 return;
 881         }
 882         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 883                 conf->raid_disks);
 884
 885         for (i = 0; i < conf->raid_disks; i++) {
 886                 char b[BDEVNAME_SIZE];
 887                 tmp = conf->mirrors + i;
 888                 if (tmp->rdev)
 889                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
 890                                 i, !tmp->rdev->in_sync, !tmp->rdev->faulty,
 891                                 bdevname(tmp->rdev->bdev,b));
 892         }
 893 }
 894
 895 static void close_sync(conf_t *conf)
 896 {
 897         spin_lock_irq(&conf->resync_lock);
 898         wait_event_lock_irq(conf->wait_resume, !conf->barrier,
 899                             conf->resync_lock,  unplug_slaves(conf->mddev));
 900         spin_unlock_irq(&conf->resync_lock);
 901
 902         if (conf->barrier) BUG();
 903         if (waitqueue_active(&conf->wait_idle)) BUG();
 904
 905         mempool_destroy(conf->r10buf_pool);
 906         conf->r10buf_pool = NULL;
 907 }
 908
 909 /* check if there are enough drives for
 910  * every block to appear on atleast one
 911  */
 912 static int enough(conf_t *conf)
 913 {
 914         int first = 0;
 915
 916         do {
 917                 int n = conf->copies;
 918                 int cnt = 0;
 919                 while (n--) {
 920                         if (conf->mirrors[first].rdev)
 921                                 cnt++;
 922                         first = (first+1) % conf->raid_disks;
 923                 }
 924                 if (cnt == 0)
 925                         return 0;
 926         } while (first != 0);
 927         return 1;
 928 }
 929
 930 static int raid10_spare_active(mddev_t *mddev)
 931 {
 932         int i;
 933         conf_t *conf = mddev->private;
 934         mirror_info_t *tmp;
 935
 936         /*
 937          * Find all non-in_sync disks within the RAID10 configuration
 938          * and mark them in_sync
 939          */
 940         for (i = 0; i < conf->raid_disks; i++) {
 941                 tmp = conf->mirrors + i;
 942                 if (tmp->rdev
 943                     && !tmp->rdev->faulty
 944                     && !tmp->rdev->in_sync) {
 945                         conf->working_disks++;
 946                         mddev->degraded--;
 947                         tmp->rdev->in_sync = 1;
 948                 }
 949         }
 950
 951         print_conf(conf);
 952         return 0;
 953 }
 954
 955
 956 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 957 {
 958         conf_t *conf = mddev->private;
 959         int found = 0;
 960         int mirror;
 961         mirror_info_t *p;
 962
 963         if (mddev->recovery_cp < MaxSector)
 964                 /* only hot-add to in-sync arrays, as recovery is
 965                  * very different from resync
 966                  */
 967                 return 0;
 968         if (!enough(conf))
 969                 return 0;
 970
 971         for (mirror=0; mirror < mddev->raid_disks; mirror++)
 972                 if ( !(p=conf->mirrors+mirror)->rdev) {
 973
 974                         blk_queue_stack_limits(mddev->queue,
 975                                                rdev->bdev->bd_disk->queue);
 976                         /* as we don't honour merge_bvec_fn, we must never risk
 977                          * violating it, so limit ->max_sector to one PAGE, as
 978                          * a one page request is never in violation.
 979                          */
 980                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 981                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
 982                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
 983
 984                         p->head_position = 0;
 985                         rdev->raid_disk = mirror;
 986                         found = 1;
 987                         p->rdev = rdev;
 988                         break;
 989                 }
 990
 991         print_conf(conf);
 992         return found;
 993 }
 994
 995 static int raid10_remove_disk(mddev_t *mddev, int number)
 996 {
 997         conf_t *conf = mddev->private;
 998         int err = 0;
 999         mdk_rdev_t *rdev;
1000         mirror_info_t *p = conf->mirrors+ number;
1001
1002         print_conf(conf);
1003         rdev = p->rdev;
1004         if (rdev) {
1005                 if (rdev->in_sync ||
1006                     atomic_read(&rdev->nr_pending)) {
1007                         err = -EBUSY;
1008                         goto abort;
1009                 }
1010                 p->rdev = NULL;
1011                 synchronize_rcu();
1012                 if (atomic_read(&rdev->nr_pending)) {
1013                         /* lost the race, try later */
1014                         err = -EBUSY;
1015                         p->rdev = rdev;
1016                 }
1017         }
1018 abort:
1019
1020         print_conf(conf);
1021         return err;
1022 }
1023
1024
1025 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1026 {
1027         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1028         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1029         conf_t *conf = mddev_to_conf(r10_bio->mddev);
1030         int i,d;
1031
1032         if (bio->bi_size)
1033                 return 1;
1034
1035         for (i=0; i<conf->copies; i++)
1036                 if (r10_bio->devs[i].bio == bio)
1037                         break;
1038         if (i == conf->copies)
1039                 BUG();
1040         update_head_pos(i, r10_bio);
1041         d = r10_bio->devs[i].devnum;
1042         if (!uptodate)
1043                 md_error(r10_bio->mddev,
1044                          conf->mirrors[d].rdev);
1045
1046         /* for reconstruct, we always reschedule after a read.
1047          * for resync, only after all reads
1048          */
1049         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1050             atomic_dec_and_test(&r10_bio->remaining)) {
1051                 /* we have read all the blocks,
1052                  * do the comparison in process context in raid10d
1053                  */
1054                 reschedule_retry(r10_bio);
1055         }
1056         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1057         return 0;
1058 }
1059
1060 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1061 {
1062         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1063         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1064         mddev_t *mddev = r10_bio->mddev;
1065         conf_t *conf = mddev_to_conf(mddev);
1066         int i,d;
1067
1068         if (bio->bi_size)
1069                 return 1;
1070
1071         for (i = 0; i < conf->copies; i++)
1072                 if (r10_bio->devs[i].bio == bio)
1073                         break;
1074         d = r10_bio->devs[i].devnum;
1075
1076         if (!uptodate)
1077                 md_error(mddev, conf->mirrors[d].rdev);
1078         update_head_pos(i, r10_bio);
1079
1080         while (atomic_dec_and_test(&r10_bio->remaining)) {
1081                 if (r10_bio->master_bio == NULL) {
1082                         /* the primary of several recovery bios */
1083                         md_done_sync(mddev, r10_bio->sectors, 1);
1084                         put_buf(r10_bio);
1085                         break;
1086                 } else {
1087                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1088                         put_buf(r10_bio);
1089                         r10_bio = r10_bio2;
1090                 }
1091         }
1092         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1093         return 0;
1094 }
1095
1096 /*
1097  * Note: sync and recover and handled very differently for raid10
1098  * This code is for resync.
1099  * For resync, we read through virtual addresses and read all blocks.
1100  * If there is any error, we schedule a write.  The lowest numbered
1101  * drive is authoritative.
1102  * However requests come for physical address, so we need to map.
1103  * For every physical address there are raid_disks/copies virtual addresses,
1104  * which is always are least one, but is not necessarly an integer.
1105  * This means that a physical address can span multiple chunks, so we may
1106  * have to submit multiple io requests for a single sync request.
1107  */
1108 /*
1109  * We check if all blocks are in-sync and only write to blocks that
1110  * aren't in sync
1111  */
1112 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1113 {
1114         conf_t *conf = mddev_to_conf(mddev);
1115         int i, first;
1116         struct bio *tbio, *fbio;
1117
1118         atomic_set(&r10_bio->remaining, 1);
1119
1120         /* find the first device with a block */
1121         for (i=0; i<conf->copies; i++)
1122                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1123                         break;
1124
1125         if (i == conf->copies)
1126                 goto done;
1127
1128         first = i;
1129         fbio = r10_bio->devs[i].bio;
1130
1131         /* now find blocks with errors */
1132         for (i=first+1 ; i < conf->copies ; i++) {
1133                 int vcnt, j, d;
1134
1135                 if (!test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1136                         continue;
1137                 /* We know that the bi_io_vec layout is the same for
1138                  * both 'first' and 'i', so we just compare them.
1139                  * All vec entries are PAGE_SIZE;
1140                  */
1141                 tbio = r10_bio->devs[i].bio;
1142                 vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1143                 for (j = 0; j < vcnt; j++)
1144                         if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1145                                    page_address(tbio->bi_io_vec[j].bv_page),
1146                                    PAGE_SIZE))
1147                                 break;
1148                 if (j == vcnt)
1149                         continue;
1150                 /* Ok, we need to write this bio
1151                  * First we need to fixup bv_offset, bv_len and
1152                  * bi_vecs, as the read request might have corrupted these
1153                  */
1154                 tbio->bi_vcnt = vcnt;
1155                 tbio->bi_size = r10_bio->sectors << 9;
1156                 tbio->bi_idx = 0;
1157                 tbio->bi_phys_segments = 0;
1158                 tbio->bi_hw_segments = 0;
1159                 tbio->bi_hw_front_size = 0;
1160                 tbio->bi_hw_back_size = 0;
1161                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1162                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1163                 tbio->bi_next = NULL;
1164                 tbio->bi_rw = WRITE;
1165                 tbio->bi_private = r10_bio;
1166                 tbio->bi_sector = r10_bio->devs[i].addr;
1167
1168                 for (j=0; j < vcnt ; j++) {
1169                         tbio->bi_io_vec[j].bv_offset = 0;
1170                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1171
1172                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1173                                page_address(fbio->bi_io_vec[j].bv_page),
1174                                PAGE_SIZE);
1175                 }
1176                 tbio->bi_end_io = end_sync_write;
1177
1178                 d = r10_bio->devs[i].devnum;
1179                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1180                 atomic_inc(&r10_bio->remaining);
1181                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1182
1183                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1184                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1185                 generic_make_request(tbio);
1186         }
1187
1188 done:
1189         if (atomic_dec_and_test(&r10_bio->remaining)) {
1190                 md_done_sync(mddev, r10_bio->sectors, 1);
1191                 put_buf(r10_bio);
1192         }
1193 }
1194
1195 /*
1196  * Now for the recovery code.
1197  * Recovery happens across physical sectors.
1198  * We recover all non-is_sync drives by finding the virtual address of
1199  * each, and then choose a working drive that also has that virt address.
1200  * There is a separate r10_bio for each non-in_sync drive.
1201  * Only the first two slots are in use. The first for reading,
1202  * The second for writing.
1203  *
1204  */
1205
1206 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1207 {
1208         conf_t *conf = mddev_to_conf(mddev);
1209         int i, d;
1210         struct bio *bio, *wbio;
1211
1212
1213         /* move the pages across to the second bio
1214          * and submit the write request
1215          */
1216         bio = r10_bio->devs[0].bio;
1217         wbio = r10_bio->devs[1].bio;
1218         for (i=0; i < wbio->bi_vcnt; i++) {
1219                 struct page *p = bio->bi_io_vec[i].bv_page;
1220                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1221                 wbio->bi_io_vec[i].bv_page = p;
1222         }
1223         d = r10_bio->devs[1].devnum;
1224
1225         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1226         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1227         generic_make_request(wbio);
1228 }
1229
1230
1231 /*
1232  * This is a kernel thread which:
1233  *
1234  *      1.      Retries failed read operations on working mirrors.
1235  *      2.      Updates the raid superblock when problems encounter.
1236  *      3.      Performs writes following reads for array syncronising.
1237  */
1238
1239 static void raid10d(mddev_t *mddev)
1240 {
1241         r10bio_t *r10_bio;
1242         struct bio *bio;
1243         unsigned long flags;
1244         conf_t *conf = mddev_to_conf(mddev);
1245         struct list_head *head = &conf->retry_list;
1246         int unplug=0;
1247         mdk_rdev_t *rdev;
1248
1249         md_check_recovery(mddev);
1250
1251         for (;;) {
1252                 char b[BDEVNAME_SIZE];
1253                 spin_lock_irqsave(&conf->device_lock, flags);
1254                 if (list_empty(head))
1255                         break;
1256                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1257                 list_del(head->prev);
1258                 spin_unlock_irqrestore(&conf->device_lock, flags);
1259
1260                 mddev = r10_bio->mddev;
1261                 conf = mddev_to_conf(mddev);
1262                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1263                         sync_request_write(mddev, r10_bio);
1264                         unplug = 1;
1265                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1266                         recovery_request_write(mddev, r10_bio);
1267                         unplug = 1;
1268                 } else {
1269                         int mirror;
1270                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1271                         r10_bio->devs[r10_bio->read_slot].bio = NULL;
1272                         bio_put(bio);
1273                         mirror = read_balance(conf, r10_bio);
1274                         if (mirror == -1) {
1275                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1276                                        " read error for block %llu\n",
1277                                        bdevname(bio->bi_bdev,b),
1278                                        (unsigned long long)r10_bio->sector);
1279                                 raid_end_bio_io(r10_bio);
1280                         } else {
1281                                 rdev = conf->mirrors[mirror].rdev;
1282                                 if (printk_ratelimit())
1283                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1284                                                " another mirror\n",
1285                                                bdevname(rdev->bdev,b),
1286                                                (unsigned long long)r10_bio->sector);
1287                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1288                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1289                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1290                                         + rdev->data_offset;
1291                                 bio->bi_bdev = rdev->bdev;
1292                                 bio->bi_rw = READ;
1293                                 bio->bi_private = r10_bio;
1294                                 bio->bi_end_io = raid10_end_read_request;
1295                                 unplug = 1;
1296                                 generic_make_request(bio);
1297                         }
1298                 }
1299         }
1300         spin_unlock_irqrestore(&conf->device_lock, flags);
1301         if (unplug)
1302                 unplug_slaves(mddev);
1303 }
1304
1305
1306 static int init_resync(conf_t *conf)
1307 {
1308         int buffs;
1309
1310         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1311         if (conf->r10buf_pool)
1312                 BUG();
1313         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1314         if (!conf->r10buf_pool)
1315                 return -ENOMEM;
1316         conf->next_resync = 0;
1317         return 0;
1318 }
1319
1320 /*
1321  * perform a "sync" on one "block"
1322  *
1323  * We need to make sure that no normal I/O request - particularly write
1324  * requests - conflict with active sync requests.
1325  *
1326  * This is achieved by tracking pending requests and a 'barrier' concept
1327  * that can be installed to exclude normal IO requests.
1328  *
1329  * Resync and recovery are handled very differently.
1330  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1331  *
1332  * For resync, we iterate over virtual addresses, read all copies,
1333  * and update if there are differences.  If only one copy is live,
1334  * skip it.
1335  * For recovery, we iterate over physical addresses, read a good
1336  * value for each non-in_sync drive, and over-write.
1337  *
1338  * So, for recovery we may have several outstanding complex requests for a
1339  * given address, one for each out-of-sync device.  We model this by allocating
1340  * a number of r10_bio structures, one for each out-of-sync device.
1341  * As we setup these structures, we collect all bio's together into a list
1342  * which we then process collectively to add pages, and then process again
1343  * to pass to generic_make_request.
1344  *
1345  * The r10_bio structures are linked using a borrowed master_bio pointer.
1346  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1347  * has its remaining count decremented to 0, the whole complex operation
1348  * is complete.
1349  *
1350  */
1351
1352 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1353 {
1354         conf_t *conf = mddev_to_conf(mddev);
1355         r10bio_t *r10_bio;
1356         struct bio *biolist = NULL, *bio;
1357         sector_t max_sector, nr_sectors;
1358         int disk;
1359         int i;
1360
1361         sector_t sectors_skipped = 0;
1362         int chunks_skipped = 0;
1363
1364         if (!conf->r10buf_pool)
1365                 if (init_resync(conf))
1366                         return 0;
1367
1368  skipped:
1369         max_sector = mddev->size << 1;
1370         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1371                 max_sector = mddev->resync_max_sectors;
1372         if (sector_nr >= max_sector) {
1373                 close_sync(conf);
1374                 *skipped = 1;
1375                 return sectors_skipped;
1376         }
1377         if (chunks_skipped >= conf->raid_disks) {
1378                 /* if there has been nothing to do on any drive,
1379                  * then there is nothing to do at all..
1380                  */
1381                 *skipped = 1;
1382                 return (max_sector - sector_nr) + sectors_skipped;
1383         }
1384
1385         /* make sure whole request will fit in a chunk - if chunks
1386          * are meaningful
1387          */
1388         if (conf->near_copies < conf->raid_disks &&
1389             max_sector > (sector_nr | conf->chunk_mask))
1390                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1391         /*
1392          * If there is non-resync activity waiting for us then
1393          * put in a delay to throttle resync.
1394          */
1395         if (!go_faster && waitqueue_active(&conf->wait_resume))
1396                 msleep_interruptible(1000);
1397         device_barrier(conf, sector_nr + RESYNC_SECTORS);
1398
1399         /* Again, very different code for resync and recovery.
1400          * Both must result in an r10bio with a list of bios that
1401          * have bi_end_io, bi_sector, bi_bdev set,
1402          * and bi_private set to the r10bio.
1403          * For recovery, we may actually create several r10bios
1404          * with 2 bios in each, that correspond to the bios in the main one.
1405          * In this case, the subordinate r10bios link back through a
1406          * borrowed master_bio pointer, and the counter in the master
1407          * includes a ref from each subordinate.
1408          */
1409         /* First, we decide what to do and set ->bi_end_io
1410          * To end_sync_read if we want to read, and
1411          * end_sync_write if we will want to write.
1412          */
1413
1414         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1415                 /* recovery... the complicated one */
1416                 int i, j, k;
1417                 r10_bio = NULL;
1418
1419                 for (i=0 ; i<conf->raid_disks; i++)
1420                         if (conf->mirrors[i].rdev &&
1421                             !conf->mirrors[i].rdev->in_sync) {
1422                                 /* want to reconstruct this device */
1423                                 r10bio_t *rb2 = r10_bio;
1424
1425                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1426                                 spin_lock_irq(&conf->resync_lock);
1427                                 conf->nr_pending++;
1428                                 if (rb2) conf->barrier++;
1429                                 spin_unlock_irq(&conf->resync_lock);
1430                                 atomic_set(&r10_bio->remaining, 0);
1431
1432                                 r10_bio->master_bio = (struct bio*)rb2;
1433                                 if (rb2)
1434                                         atomic_inc(&rb2->remaining);
1435                                 r10_bio->mddev = mddev;
1436                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1437                                 r10_bio->sector = raid10_find_virt(conf, sector_nr, i);
1438                                 raid10_find_phys(conf, r10_bio);
1439                                 for (j=0; j<conf->copies;j++) {
1440                                         int d = r10_bio->devs[j].devnum;
1441                                         if (conf->mirrors[d].rdev &&
1442                                             conf->mirrors[d].rdev->in_sync) {
1443                                                 /* This is where we read from */
1444                                                 bio = r10_bio->devs[0].bio;
1445                                                 bio->bi_next = biolist;
1446                                                 biolist = bio;
1447                                                 bio->bi_private = r10_bio;
1448                                                 bio->bi_end_io = end_sync_read;
1449                                                 bio->bi_rw = 0;
1450                                                 bio->bi_sector = r10_bio->devs[j].addr +
1451                                                         conf->mirrors[d].rdev->data_offset;
1452                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1453                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1454                                                 atomic_inc(&r10_bio->remaining);
1455                                                 /* and we write to 'i' */
1456
1457                                                 for (k=0; k<conf->copies; k++)
1458                                                         if (r10_bio->devs[k].devnum == i)
1459                                                                 break;
1460                                                 bio = r10_bio->devs[1].bio;
1461                                                 bio->bi_next = biolist;
1462                                                 biolist = bio;
1463                                                 bio->bi_private = r10_bio;
1464                                                 bio->bi_end_io = end_sync_write;
1465                                                 bio->bi_rw = 1;
1466                                                 bio->bi_sector = r10_bio->devs[k].addr +
1467                                                         conf->mirrors[i].rdev->data_offset;
1468                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1469
1470                                                 r10_bio->devs[0].devnum = d;
1471                                                 r10_bio->devs[1].devnum = i;
1472
1473                                                 break;
1474                                         }
1475                                 }
1476                                 if (j == conf->copies) {
1477                                         BUG();
1478                                 }
1479                         }
1480                 if (biolist == NULL) {
1481                         while (r10_bio) {
1482                                 r10bio_t *rb2 = r10_bio;
1483                                 r10_bio = (r10bio_t*) rb2->master_bio;
1484                                 rb2->master_bio = NULL;
1485                                 put_buf(rb2);
1486                         }
1487                         goto giveup;
1488                 }
1489         } else {
1490                 /* resync. Schedule a read for every block at this virt offset */
1491                 int count = 0;
1492                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1493
1494                 spin_lock_irq(&conf->resync_lock);
1495                 conf->nr_pending++;
1496                 spin_unlock_irq(&conf->resync_lock);
1497
1498                 r10_bio->mddev = mddev;
1499                 atomic_set(&r10_bio->remaining, 0);
1500
1501                 r10_bio->master_bio = NULL;
1502                 r10_bio->sector = sector_nr;
1503                 set_bit(R10BIO_IsSync, &r10_bio->state);
1504                 raid10_find_phys(conf, r10_bio);
1505                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1506
1507                 for (i=0; i<conf->copies; i++) {
1508                         int d = r10_bio->devs[i].devnum;
1509                         bio = r10_bio->devs[i].bio;
1510                         bio->bi_end_io = NULL;
1511                         if (conf->mirrors[d].rdev == NULL ||
1512                             conf->mirrors[d].rdev->faulty)
1513                                 continue;
1514                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1515                         atomic_inc(&r10_bio->remaining);
1516                         bio->bi_next = biolist;
1517                         biolist = bio;
1518                         bio->bi_private = r10_bio;
1519                         bio->bi_end_io = end_sync_read;
1520                         bio->bi_rw = 0;
1521                         bio->bi_sector = r10_bio->devs[i].addr +
1522                                 conf->mirrors[d].rdev->data_offset;
1523                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1524                         count++;
1525                 }
1526
1527                 if (count < 2) {
1528                         for (i=0; i<conf->copies; i++) {
1529                                 int d = r10_bio->devs[i].devnum;
1530                                 if (r10_bio->devs[i].bio->bi_end_io)
1531                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1532                         }
1533                         put_buf(r10_bio);
1534                         biolist = NULL;
1535                         goto giveup;
1536                 }
1537         }
1538
1539         for (bio = biolist; bio ; bio=bio->bi_next) {
1540
1541                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1542                 if (bio->bi_end_io)
1543                         bio->bi_flags |= 1 << BIO_UPTODATE;
1544                 bio->bi_vcnt = 0;
1545                 bio->bi_idx = 0;
1546                 bio->bi_phys_segments = 0;
1547                 bio->bi_hw_segments = 0;
1548                 bio->bi_size = 0;
1549         }
1550
1551         nr_sectors = 0;
1552         do {
1553                 struct page *page;
1554                 int len = PAGE_SIZE;
1555                 disk = 0;
1556                 if (sector_nr + (len>>9) > max_sector)
1557                         len = (max_sector - sector_nr) << 9;
1558                 if (len == 0)
1559                         break;
1560                 for (bio= biolist ; bio ; bio=bio->bi_next) {
1561                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1562                         if (bio_add_page(bio, page, len, 0) == 0) {
1563                                 /* stop here */
1564                                 struct bio *bio2;
1565                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1566                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1567                                         /* remove last page from this bio */
1568                                         bio2->bi_vcnt--;
1569                                         bio2->bi_size -= len;
1570                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1571                                 }
1572                                 goto bio_full;
1573                         }
1574                         disk = i;
1575                 }
1576                 nr_sectors += len>>9;
1577                 sector_nr += len>>9;
1578         } while (biolist->bi_vcnt < RESYNC_PAGES);
1579  bio_full:
1580         r10_bio->sectors = nr_sectors;
1581
1582         while (biolist) {
1583                 bio = biolist;
1584                 biolist = biolist->bi_next;
1585
1586                 bio->bi_next = NULL;
1587                 r10_bio = bio->bi_private;
1588                 r10_bio->sectors = nr_sectors;
1589
1590                 if (bio->bi_end_io == end_sync_read) {
1591                         md_sync_acct(bio->bi_bdev, nr_sectors);
1592                         generic_make_request(bio);
1593                 }
1594         }
1595
1596         if (sectors_skipped)
1597                 /* pretend they weren't skipped, it makes
1598                  * no important difference in this case
1599                  */
1600                 md_done_sync(mddev, sectors_skipped, 1);
1601
1602         return sectors_skipped + nr_sectors;
1603  giveup:
1604         /* There is nowhere to write, so all non-sync
1605          * drives must be failed, so try the next chunk...
1606          */
1607         {
1608         sector_t sec = max_sector - sector_nr;
1609         sectors_skipped += sec;
1610         chunks_skipped ++;
1611         sector_nr = max_sector;
1612         goto skipped;
1613         }
1614 }
1615
1616 static int run(mddev_t *mddev)
1617 {
1618         conf_t *conf;
1619         int i, disk_idx;
1620         mirror_info_t *disk;
1621         mdk_rdev_t *rdev;
1622         struct list_head *tmp;
1623         int nc, fc;
1624         sector_t stride, size;
1625
1626         if (mddev->level != 10) {
1627                 printk(KERN_ERR "raid10: %s: raid level not set correctly... (%d)\n",
1628                        mdname(mddev), mddev->level);
1629                 goto out;
1630         }
1631         nc = mddev->layout & 255;
1632         fc = (mddev->layout >> 8) & 255;
1633         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
1634             (mddev->layout >> 16)) {
1635                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
1636                        mdname(mddev), mddev->layout);
1637                 goto out;
1638         }
1639         /*
1640          * copy the already verified devices into our private RAID10
1641          * bookkeeping area. [whatever we allocate in run(),
1642          * should be freed in stop()]
1643          */
1644         conf = kmalloc(sizeof(conf_t), GFP_KERNEL);
1645         mddev->private = conf;
1646         if (!conf) {
1647                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1648                         mdname(mddev));
1649                 goto out;
1650         }
1651         memset(conf, 0, sizeof(*conf));
1652         conf->mirrors = kmalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1653                                  GFP_KERNEL);
1654         if (!conf->mirrors) {
1655                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1656                        mdname(mddev));
1657                 goto out_free_conf;
1658         }
1659         memset(conf->mirrors, 0, sizeof(struct mirror_info)*mddev->raid_disks);
1660
1661         conf->near_copies = nc;
1662         conf->far_copies = fc;
1663         conf->copies = nc*fc;
1664         conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
1665         conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
1666         stride = mddev->size >> (conf->chunk_shift-1);
1667         sector_div(stride, fc);
1668         conf->stride = stride << conf->chunk_shift;
1669
1670         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
1671                                                 r10bio_pool_free, conf);
1672         if (!conf->r10bio_pool) {
1673                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1674                         mdname(mddev));
1675                 goto out_free_conf;
1676         }
1677
1678         ITERATE_RDEV(mddev, rdev, tmp) {
1679                 disk_idx = rdev->raid_disk;
1680                 if (disk_idx >= mddev->raid_disks
1681                     || disk_idx < 0)
1682                         continue;
1683                 disk = conf->mirrors + disk_idx;
1684
1685                 disk->rdev = rdev;
1686
1687                 blk_queue_stack_limits(mddev->queue,
1688                                        rdev->bdev->bd_disk->queue);
1689                 /* as we don't honour merge_bvec_fn, we must never risk
1690                  * violating it, so limit ->max_sector to one PAGE, as
1691                  * a one page request is never in violation.
1692                  */
1693                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1694                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1695                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
1696
1697                 disk->head_position = 0;
1698                 if (!rdev->faulty && rdev->in_sync)
1699                         conf->working_disks++;
1700         }
1701         conf->raid_disks = mddev->raid_disks;
1702         conf->mddev = mddev;
1703         spin_lock_init(&conf->device_lock);
1704         INIT_LIST_HEAD(&conf->retry_list);
1705
1706         spin_lock_init(&conf->resync_lock);
1707         init_waitqueue_head(&conf->wait_idle);
1708         init_waitqueue_head(&conf->wait_resume);
1709
1710         /* need to check that every block has at least one working mirror */
1711         if (!enough(conf)) {
1712                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
1713                        mdname(mddev));
1714                 goto out_free_conf;
1715         }
1716
1717         mddev->degraded = 0;
1718         for (i = 0; i < conf->raid_disks; i++) {
1719
1720                 disk = conf->mirrors + i;
1721
1722                 if (!disk->rdev) {
1723                         disk->head_position = 0;
1724                         mddev->degraded++;
1725                 }
1726         }
1727
1728
1729         mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
1730         if (!mddev->thread) {
1731                 printk(KERN_ERR
1732                        "raid10: couldn't allocate thread for %s\n",
1733                        mdname(mddev));
1734                 goto out_free_conf;
1735         }
1736
1737         printk(KERN_INFO
1738                 "raid10: raid set %s active with %d out of %d devices\n",
1739                 mdname(mddev), mddev->raid_disks - mddev->degraded,
1740                 mddev->raid_disks);
1741         /*
1742          * Ok, everything is just fine now
1743          */
1744         size = conf->stride * conf->raid_disks;
1745         sector_div(size, conf->near_copies);
1746         mddev->array_size = size/2;
1747         mddev->resync_max_sectors = size;
1748
1749         mddev->queue->unplug_fn = raid10_unplug;
1750         mddev->queue->issue_flush_fn = raid10_issue_flush;
1751
1752         /* Calculate max read-ahead size.
1753          * We need to readahead at least twice a whole stripe....
1754          * maybe...
1755          */
1756         {
1757                 int stripe = conf->raid_disks * mddev->chunk_size / PAGE_CACHE_SIZE;
1758                 stripe /= conf->near_copies;
1759                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
1760                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
1761         }
1762
1763         if (conf->near_copies < mddev->raid_disks)
1764                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
1765         return 0;
1766
1767 out_free_conf:
1768         if (conf->r10bio_pool)
1769                 mempool_destroy(conf->r10bio_pool);
1770         kfree(conf->mirrors);
1771         kfree(conf);
1772         mddev->private = NULL;
1773 out:
1774         return -EIO;
1775 }
1776
1777 static int stop(mddev_t *mddev)
1778 {
1779         conf_t *conf = mddev_to_conf(mddev);
1780
1781         md_unregister_thread(mddev->thread);
1782         mddev->thread = NULL;
1783         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1784         if (conf->r10bio_pool)
1785                 mempool_destroy(conf->r10bio_pool);
1786         kfree(conf->mirrors);
1787         kfree(conf);
1788         mddev->private = NULL;
1789         return 0;
1790 }
1791
1792
1793 static mdk_personality_t raid10_personality =
1794 {
1795         .name           = "raid10",
1796         .owner          = THIS_MODULE,
1797         .make_request   = make_request,
1798         .run            = run,
1799         .stop           = stop,
1800         .status         = status,
1801         .error_handler  = error,
1802         .hot_add_disk   = raid10_add_disk,
1803         .hot_remove_disk= raid10_remove_disk,
1804         .spare_active   = raid10_spare_active,
1805         .sync_request   = sync_request,
1806 };
1807
1808 static int __init raid_init(void)
1809 {
1810         return register_md_personality(RAID10, &raid10_personality);
1811 }
1812
1813 static void raid_exit(void)
1814 {
1815         unregister_md_personality(RAID10);
1816 }
1817
1818 module_init(raid_init);
1819 module_exit(raid_exit);
1820 MODULE_LICENSE("GPL");
1821 MODULE_ALIAS("md-personality-9"); /* RAID10 */