drivers/md/raid10.c

   1 /*
   2  * raid10.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 2000-2004 Neil Brown
   5  *
   6  * RAID-10 support for md.
   7  *
   8  * Base on code in raid1.c.  See raid1.c for futher copyright information.
   9  *
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2, or (at your option)
  14  * any later version.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * (for example /usr/src/linux/COPYING); if not, write to the Free
  18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 #include "dm-bio-list.h"
  22 #include <linux/raid/raid10.h>
  23 #include <linux/raid/bitmap.h>
  24
  25 /*
  26  * RAID10 provides a combination of RAID0 and RAID1 functionality.
  27  * The layout of data is defined by
  28  *    chunk_size
  29  *    raid_disks
  30  *    near_copies (stored in low byte of layout)
  31  *    far_copies (stored in second byte of layout)
  32  *    far_offset (stored in bit 16 of layout )
  33  *
  34  * The data to be stored is divided into chunks using chunksize.
  35  * Each device is divided into far_copies sections.
  36  * In each section, chunks are laid out in a style similar to raid0, but
  37  * near_copies copies of each chunk is stored (each on a different drive).
  38  * The starting device for each section is offset near_copies from the starting
  39  * device of the previous section.
  40  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
  41  * drive.
  42  * near_copies and far_copies must be at least one, and their product is at most
  43  * raid_disks.
  44  *
  45  * If far_offset is true, then the far_copies are handled a bit differently.
  46  * The copies are still in different stripes, but instead of be very far apart
  47  * on disk, there are adjacent stripes.
  48  */
  49
  50 /*
  51  * Number of guaranteed r10bios in case of extreme VM load:
  52  */
  53 #define NR_RAID10_BIOS 256
  54
  55 static void unplug_slaves(mddev_t *mddev);
  56
  57 static void allow_barrier(conf_t *conf);
  58 static void lower_barrier(conf_t *conf);
  59
  60 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
  61 {
  62         conf_t *conf = data;
  63         r10bio_t *r10_bio;
  64         int size = offsetof(struct r10bio_s, devs[conf->copies]);
  65
  66         /* allocate a r10bio with room for raid_disks entries in the bios array */
  67         r10_bio = kzalloc(size, gfp_flags);
  68         if (!r10_bio)
  69                 unplug_slaves(conf->mddev);
  70
  71         return r10_bio;
  72 }
  73
  74 static void r10bio_pool_free(void *r10_bio, void *data)
  75 {
  76         kfree(r10_bio);
  77 }
  78
  79 #define RESYNC_BLOCK_SIZE (64*1024)
  80 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  81 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  82 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  83 #define RESYNC_WINDOW (2048*1024)
  84
  85 /*
  86  * When performing a resync, we need to read and compare, so
  87  * we need as many pages are there are copies.
  88  * When performing a recovery, we need 2 bios, one for read,
  89  * one for write (we recover only one drive per r10buf)
  90  *
  91  */
  92 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
  93 {
  94         conf_t *conf = data;
  95         struct page *page;
  96         r10bio_t *r10_bio;
  97         struct bio *bio;
  98         int i, j;
  99         int nalloc;
 100
 101         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 102         if (!r10_bio) {
 103                 unplug_slaves(conf->mddev);
 104                 return NULL;
 105         }
 106
 107         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
 108                 nalloc = conf->copies; /* resync */
 109         else
 110                 nalloc = 2; /* recovery */
 111
 112         /*
 113          * Allocate bios.
 114          */
 115         for (j = nalloc ; j-- ; ) {
 116                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 117                 if (!bio)
 118                         goto out_free_bio;
 119                 r10_bio->devs[j].bio = bio;
 120         }
 121         /*
 122          * Allocate RESYNC_PAGES data pages and attach them
 123          * where needed.
 124          */
 125         for (j = 0 ; j < nalloc; j++) {
 126                 bio = r10_bio->devs[j].bio;
 127                 for (i = 0; i < RESYNC_PAGES; i++) {
 128                         page = alloc_page(gfp_flags);
 129                         if (unlikely(!page))
 130                                 goto out_free_pages;
 131
 132                         bio->bi_io_vec[i].bv_page = page;
 133                 }
 134         }
 135
 136         return r10_bio;
 137
 138 out_free_pages:
 139         for ( ; i > 0 ; i--)
 140                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
 141         while (j--)
 142                 for (i = 0; i < RESYNC_PAGES ; i++)
 143                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 144         j = -1;
 145 out_free_bio:
 146         while ( ++j < nalloc )
 147                 bio_put(r10_bio->devs[j].bio);
 148         r10bio_pool_free(r10_bio, conf);
 149         return NULL;
 150 }
 151
 152 static void r10buf_pool_free(void *__r10_bio, void *data)
 153 {
 154         int i;
 155         conf_t *conf = data;
 156         r10bio_t *r10bio = __r10_bio;
 157         int j;
 158
 159         for (j=0; j < conf->copies; j++) {
 160                 struct bio *bio = r10bio->devs[j].bio;
 161                 if (bio) {
 162                         for (i = 0; i < RESYNC_PAGES; i++) {
 163                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 164                                 bio->bi_io_vec[i].bv_page = NULL;
 165                         }
 166                         bio_put(bio);
 167                 }
 168         }
 169         r10bio_pool_free(r10bio, conf);
 170 }
 171
 172 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
 173 {
 174         int i;
 175
 176         for (i = 0; i < conf->copies; i++) {
 177                 struct bio **bio = & r10_bio->devs[i].bio;
 178                 if (*bio && *bio != IO_BLOCKED)
 179                         bio_put(*bio);
 180                 *bio = NULL;
 181         }
 182 }
 183
 184 static void free_r10bio(r10bio_t *r10_bio)
 185 {
 186         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 187
 188         /*
 189          * Wake up any possible resync thread that waits for the device
 190          * to go idle.
 191          */
 192         allow_barrier(conf);
 193
 194         put_all_bios(conf, r10_bio);
 195         mempool_free(r10_bio, conf->r10bio_pool);
 196 }
 197
 198 static void put_buf(r10bio_t *r10_bio)
 199 {
 200         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 201
 202         mempool_free(r10_bio, conf->r10buf_pool);
 203
 204         lower_barrier(conf);
 205 }
 206
 207 static void reschedule_retry(r10bio_t *r10_bio)
 208 {
 209         unsigned long flags;
 210         mddev_t *mddev = r10_bio->mddev;
 211         conf_t *conf = mddev_to_conf(mddev);
 212
 213         spin_lock_irqsave(&conf->device_lock, flags);
 214         list_add(&r10_bio->retry_list, &conf->retry_list);
 215         conf->nr_queued ++;
 216         spin_unlock_irqrestore(&conf->device_lock, flags);
 217
 218         /* wake up frozen array... */
 219         wake_up(&conf->wait_barrier);
 220
 221         md_wakeup_thread(mddev->thread);
 222 }
 223
 224 /*
 225  * raid_end_bio_io() is called when we have finished servicing a mirrored
 226  * operation and are ready to return a success/failure code to the buffer
 227  * cache layer.
 228  */
 229 static void raid_end_bio_io(r10bio_t *r10_bio)
 230 {
 231         struct bio *bio = r10_bio->master_bio;
 232
 233         bio_endio(bio,
 234                 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
 235         free_r10bio(r10_bio);
 236 }
 237
 238 /*
 239  * Update disk head position estimator based on IRQ completion info.
 240  */
 241 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
 242 {
 243         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 244
 245         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 246                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 247 }
 248
 249 static void raid10_end_read_request(struct bio *bio, int error)
 250 {
 251         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 252         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
 253         int slot, dev;
 254         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 255
 256
 257         slot = r10_bio->read_slot;
 258         dev = r10_bio->devs[slot].devnum;
 259         /*
 260          * this branch is our 'one mirror IO has finished' event handler:
 261          */
 262         update_head_pos(slot, r10_bio);
 263
 264         if (uptodate) {
 265                 /*
 266                  * Set R10BIO_Uptodate in our master bio, so that
 267                  * we will return a good error code to the higher
 268                  * levels even if IO on some other mirrored buffer fails.
 269                  *
 270                  * The 'master' represents the composite IO operation to
 271                  * user-side. So if something waits for IO, then it will
 272                  * wait for the 'master' bio.
 273                  */
 274                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 275                 raid_end_bio_io(r10_bio);
 276         } else {
 277                 /*
 278                  * oops, read error:
 279                  */
 280                 char b[BDEVNAME_SIZE];
 281                 if (printk_ratelimit())
 282                         printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
 283                                bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
 284                 reschedule_retry(r10_bio);
 285         }
 286
 287         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 288 }
 289
 290 static void raid10_end_write_request(struct bio *bio, int error)
 291 {
 292         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 293         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
 294         int slot, dev;
 295         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 296
 297         for (slot = 0; slot < conf->copies; slot++)
 298                 if (r10_bio->devs[slot].bio == bio)
 299                         break;
 300         dev = r10_bio->devs[slot].devnum;
 301
 302         /*
 303          * this branch is our 'one mirror IO has finished' event handler:
 304          */
 305         if (!uptodate) {
 306                 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
 307                 /* an I/O failed, we can't clear the bitmap */
 308                 set_bit(R10BIO_Degraded, &r10_bio->state);
 309         } else
 310                 /*
 311                  * Set R10BIO_Uptodate in our master bio, so that
 312                  * we will return a good error code for to the higher
 313                  * levels even if IO on some other mirrored buffer fails.
 314                  *
 315                  * The 'master' represents the composite IO operation to
 316                  * user-side. So if something waits for IO, then it will
 317                  * wait for the 'master' bio.
 318                  */
 319                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 320
 321         update_head_pos(slot, r10_bio);
 322
 323         /*
 324          *
 325          * Let's see if all mirrored write operations have finished
 326          * already.
 327          */
 328         if (atomic_dec_and_test(&r10_bio->remaining)) {
 329                 /* clear the bitmap if all writes complete successfully */
 330                 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 331                                 r10_bio->sectors,
 332                                 !test_bit(R10BIO_Degraded, &r10_bio->state),
 333                                 0);
 334                 md_write_end(r10_bio->mddev);
 335                 raid_end_bio_io(r10_bio);
 336         }
 337
 338         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 339 }
 340
 341
 342 /*
 343  * RAID10 layout manager
 344  * Aswell as the chunksize and raid_disks count, there are two
 345  * parameters: near_copies and far_copies.
 346  * near_copies * far_copies must be <= raid_disks.
 347  * Normally one of these will be 1.
 348  * If both are 1, we get raid0.
 349  * If near_copies == raid_disks, we get raid1.
 350  *
 351  * Chunks are layed out in raid0 style with near_copies copies of the
 352  * first chunk, followed by near_copies copies of the next chunk and
 353  * so on.
 354  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 355  * as described above, we start again with a device offset of near_copies.
 356  * So we effectively have another copy of the whole array further down all
 357  * the drives, but with blocks on different drives.
 358  * With this layout, and block is never stored twice on the one device.
 359  *
 360  * raid10_find_phys finds the sector offset of a given virtual sector
 361  * on each device that it is on.
 362  *
 363  * raid10_find_virt does the reverse mapping, from a device and a
 364  * sector offset to a virtual address
 365  */
 366
 367 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
 368 {
 369         int n,f;
 370         sector_t sector;
 371         sector_t chunk;
 372         sector_t stripe;
 373         int dev;
 374
 375         int slot = 0;
 376
 377         /* now calculate first sector/dev */
 378         chunk = r10bio->sector >> conf->chunk_shift;
 379         sector = r10bio->sector & conf->chunk_mask;
 380
 381         chunk *= conf->near_copies;
 382         stripe = chunk;
 383         dev = sector_div(stripe, conf->raid_disks);
 384         if (conf->far_offset)
 385                 stripe *= conf->far_copies;
 386
 387         sector += stripe << conf->chunk_shift;
 388
 389         /* and calculate all the others */
 390         for (n=0; n < conf->near_copies; n++) {
 391                 int d = dev;
 392                 sector_t s = sector;
 393                 r10bio->devs[slot].addr = sector;
 394                 r10bio->devs[slot].devnum = d;
 395                 slot++;
 396
 397                 for (f = 1; f < conf->far_copies; f++) {
 398                         d += conf->near_copies;
 399                         if (d >= conf->raid_disks)
 400                                 d -= conf->raid_disks;
 401                         s += conf->stride;
 402                         r10bio->devs[slot].devnum = d;
 403                         r10bio->devs[slot].addr = s;
 404                         slot++;
 405                 }
 406                 dev++;
 407                 if (dev >= conf->raid_disks) {
 408                         dev = 0;
 409                         sector += (conf->chunk_mask + 1);
 410                 }
 411         }
 412         BUG_ON(slot != conf->copies);
 413 }
 414
 415 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
 416 {
 417         sector_t offset, chunk, vchunk;
 418
 419         offset = sector & conf->chunk_mask;
 420         if (conf->far_offset) {
 421                 int fc;
 422                 chunk = sector >> conf->chunk_shift;
 423                 fc = sector_div(chunk, conf->far_copies);
 424                 dev -= fc * conf->near_copies;
 425                 if (dev < 0)
 426                         dev += conf->raid_disks;
 427         } else {
 428                 while (sector >= conf->stride) {
 429                         sector -= conf->stride;
 430                         if (dev < conf->near_copies)
 431                                 dev += conf->raid_disks - conf->near_copies;
 432                         else
 433                                 dev -= conf->near_copies;
 434                 }
 435                 chunk = sector >> conf->chunk_shift;
 436         }
 437         vchunk = chunk * conf->raid_disks + dev;
 438         sector_div(vchunk, conf->near_copies);
 439         return (vchunk << conf->chunk_shift) + offset;
 440 }
 441
 442 /**
 443  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 444  *      @q: request queue
 445  *      @bvm: properties of new bio
 446  *      @biovec: the request that could be merged to it.
 447  *
 448  *      Return amount of bytes we can accept at this offset
 449  *      If near_copies == raid_disk, there are no striping issues,
 450  *      but in that case, the function isn't called at all.
 451  */
 452 static int raid10_mergeable_bvec(struct request_queue *q,
 453                                  struct bvec_merge_data *bvm,
 454                                  struct bio_vec *biovec)
 455 {
 456         mddev_t *mddev = q->queuedata;
 457         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 458         int max;
 459         unsigned int chunk_sectors = mddev->chunk_size >> 9;
 460         unsigned int bio_sectors = bvm->bi_size >> 9;
 461
 462         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 463         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 464         if (max <= biovec->bv_len && bio_sectors == 0)
 465                 return biovec->bv_len;
 466         else
 467                 return max;
 468 }
 469
 470 /*
 471  * This routine returns the disk from which the requested read should
 472  * be done. There is a per-array 'next expected sequential IO' sector
 473  * number - if this matches on the next IO then we use the last disk.
 474  * There is also a per-disk 'last know head position' sector that is
 475  * maintained from IRQ contexts, both the normal and the resync IO
 476  * completion handlers update this position correctly. If there is no
 477  * perfect sequential match then we pick the disk whose head is closest.
 478  *
 479  * If there are 2 mirrors in the same 2 devices, performance degrades
 480  * because position is mirror, not device based.
 481  *
 482  * The rdev for the device selected will have nr_pending incremented.
 483  */
 484
 485 /*
 486  * FIXME: possibly should rethink readbalancing and do it differently
 487  * depending on near_copies / far_copies geometry.
 488  */
 489 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
 490 {
 491         const unsigned long this_sector = r10_bio->sector;
 492         int disk, slot, nslot;
 493         const int sectors = r10_bio->sectors;
 494         sector_t new_distance, current_distance;
 495         mdk_rdev_t *rdev;
 496
 497         raid10_find_phys(conf, r10_bio);
 498         rcu_read_lock();
 499         /*
 500          * Check if we can balance. We can balance on the whole
 501          * device if no resync is going on (recovery is ok), or below
 502          * the resync window. We take the first readable disk when
 503          * above the resync window.
 504          */
 505         if (conf->mddev->recovery_cp < MaxSector
 506             && (this_sector + sectors >= conf->next_resync)) {
 507                 /* make sure that disk is operational */
 508                 slot = 0;
 509                 disk = r10_bio->devs[slot].devnum;
 510
 511                 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 512                        r10_bio->devs[slot].bio == IO_BLOCKED ||
 513                        !test_bit(In_sync, &rdev->flags)) {
 514                         slot++;
 515                         if (slot == conf->copies) {
 516                                 slot = 0;
 517                                 disk = -1;
 518                                 break;
 519                         }
 520                         disk = r10_bio->devs[slot].devnum;
 521                 }
 522                 goto rb_out;
 523         }
 524
 525
 526         /* make sure the disk is operational */
 527         slot = 0;
 528         disk = r10_bio->devs[slot].devnum;
 529         while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 530                r10_bio->devs[slot].bio == IO_BLOCKED ||
 531                !test_bit(In_sync, &rdev->flags)) {
 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          * or - for far > 1 - find the closest to partition beginning */
 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                     r10_bio->devs[nslot].bio == IO_BLOCKED ||
 553                     !test_bit(In_sync, &rdev->flags))
 554                         continue;
 555
 556                 /* This optimisation is debatable, and completely destroys
 557                  * sequential read speed for 'far copies' arrays.  So only
 558                  * keep it for 'near' arrays, and review those later.
 559                  */
 560                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
 561                         disk = ndisk;
 562                         slot = nslot;
 563                         break;
 564                 }
 565
 566                 /* for far > 1 always use the lowest address */
 567                 if (conf->far_copies > 1)
 568                         new_distance = r10_bio->devs[nslot].addr;
 569                 else
 570                         new_distance = abs(r10_bio->devs[nslot].addr -
 571                                            conf->mirrors[ndisk].head_position);
 572                 if (new_distance < current_distance) {
 573                         current_distance = new_distance;
 574                         disk = ndisk;
 575                         slot = nslot;
 576                 }
 577         }
 578
 579 rb_out:
 580         r10_bio->read_slot = slot;
 581 /*      conf->next_seq_sect = this_sector + sectors;*/
 582
 583         if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
 584                 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
 585         else
 586                 disk = -1;
 587         rcu_read_unlock();
 588
 589         return disk;
 590 }
 591
 592 static void unplug_slaves(mddev_t *mddev)
 593 {
 594         conf_t *conf = mddev_to_conf(mddev);
 595         int i;
 596
 597         rcu_read_lock();
 598         for (i=0; i<mddev->raid_disks; i++) {
 599                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 600                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 601                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 602
 603                         atomic_inc(&rdev->nr_pending);
 604                         rcu_read_unlock();
 605
 606                         blk_unplug(r_queue);
 607
 608                         rdev_dec_pending(rdev, mddev);
 609                         rcu_read_lock();
 610                 }
 611         }
 612         rcu_read_unlock();
 613 }
 614
 615 static void raid10_unplug(struct request_queue *q)
 616 {
 617         mddev_t *mddev = q->queuedata;
 618
 619         unplug_slaves(q->queuedata);
 620         md_wakeup_thread(mddev->thread);
 621 }
 622
 623 static int raid10_congested(void *data, int bits)
 624 {
 625         mddev_t *mddev = data;
 626         conf_t *conf = mddev_to_conf(mddev);
 627         int i, ret = 0;
 628
 629         rcu_read_lock();
 630         for (i = 0; i < mddev->raid_disks && ret == 0; i++) {
 631                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 632                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 633                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 634
 635                         ret |= bdi_congested(&q->backing_dev_info, bits);
 636                 }
 637         }
 638         rcu_read_unlock();
 639         return ret;
 640 }
 641
 642 static int flush_pending_writes(conf_t *conf)
 643 {
 644         /* Any writes that have been queued but are awaiting
 645          * bitmap updates get flushed here.
 646          * We return 1 if any requests were actually submitted.
 647          */
 648         int rv = 0;
 649
 650         spin_lock_irq(&conf->device_lock);
 651
 652         if (conf->pending_bio_list.head) {
 653                 struct bio *bio;
 654                 bio = bio_list_get(&conf->pending_bio_list);
 655                 blk_remove_plug(conf->mddev->queue);
 656                 spin_unlock_irq(&conf->device_lock);
 657                 /* flush any pending bitmap writes to disk
 658                  * before proceeding w/ I/O */
 659                 bitmap_unplug(conf->mddev->bitmap);
 660
 661                 while (bio) { /* submit pending writes */
 662                         struct bio *next = bio->bi_next;
 663                         bio->bi_next = NULL;
 664                         generic_make_request(bio);
 665                         bio = next;
 666                 }
 667                 rv = 1;
 668         } else
 669                 spin_unlock_irq(&conf->device_lock);
 670         return rv;
 671 }
 672 /* Barriers....
 673  * Sometimes we need to suspend IO while we do something else,
 674  * either some resync/recovery, or reconfigure the array.
 675  * To do this we raise a 'barrier'.
 676  * The 'barrier' is a counter that can be raised multiple times
 677  * to count how many activities are happening which preclude
 678  * normal IO.
 679  * We can only raise the barrier if there is no pending IO.
 680  * i.e. if nr_pending == 0.
 681  * We choose only to raise the barrier if no-one is waiting for the
 682  * barrier to go down.  This means that as soon as an IO request
 683  * is ready, no other operations which require a barrier will start
 684  * until the IO request has had a chance.
 685  *
 686  * So: regular IO calls 'wait_barrier'.  When that returns there
 687  *    is no backgroup IO happening,  It must arrange to call
 688  *    allow_barrier when it has finished its IO.
 689  * backgroup IO calls must call raise_barrier.  Once that returns
 690  *    there is no normal IO happeing.  It must arrange to call
 691  *    lower_barrier when the particular background IO completes.
 692  */
 693 #define RESYNC_DEPTH 32
 694
 695 static void raise_barrier(conf_t *conf, int force)
 696 {
 697         BUG_ON(force && !conf->barrier);
 698         spin_lock_irq(&conf->resync_lock);
 699
 700         /* Wait until no block IO is waiting (unless 'force') */
 701         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 702                             conf->resync_lock,
 703                             raid10_unplug(conf->mddev->queue));
 704
 705         /* block any new IO from starting */
 706         conf->barrier++;
 707
 708         /* No wait for all pending IO to complete */
 709         wait_event_lock_irq(conf->wait_barrier,
 710                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 711                             conf->resync_lock,
 712                             raid10_unplug(conf->mddev->queue));
 713
 714         spin_unlock_irq(&conf->resync_lock);
 715 }
 716
 717 static void lower_barrier(conf_t *conf)
 718 {
 719         unsigned long flags;
 720         spin_lock_irqsave(&conf->resync_lock, flags);
 721         conf->barrier--;
 722         spin_unlock_irqrestore(&conf->resync_lock, flags);
 723         wake_up(&conf->wait_barrier);
 724 }
 725
 726 static void wait_barrier(conf_t *conf)
 727 {
 728         spin_lock_irq(&conf->resync_lock);
 729         if (conf->barrier) {
 730                 conf->nr_waiting++;
 731                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 732                                     conf->resync_lock,
 733                                     raid10_unplug(conf->mddev->queue));
 734                 conf->nr_waiting--;
 735         }
 736         conf->nr_pending++;
 737         spin_unlock_irq(&conf->resync_lock);
 738 }
 739
 740 static void allow_barrier(conf_t *conf)
 741 {
 742         unsigned long flags;
 743         spin_lock_irqsave(&conf->resync_lock, flags);
 744         conf->nr_pending--;
 745         spin_unlock_irqrestore(&conf->resync_lock, flags);
 746         wake_up(&conf->wait_barrier);
 747 }
 748
 749 static void freeze_array(conf_t *conf)
 750 {
 751         /* stop syncio and normal IO and wait for everything to
 752          * go quiet.
 753          * We increment barrier and nr_waiting, and then
 754          * wait until nr_pending match nr_queued+1
 755          * This is called in the context of one normal IO request
 756          * that has failed. Thus any sync request that might be pending
 757          * will be blocked by nr_pending, and we need to wait for
 758          * pending IO requests to complete or be queued for re-try.
 759          * Thus the number queued (nr_queued) plus this request (1)
 760          * must match the number of pending IOs (nr_pending) before
 761          * we continue.
 762          */
 763         spin_lock_irq(&conf->resync_lock);
 764         conf->barrier++;
 765         conf->nr_waiting++;
 766         wait_event_lock_irq(conf->wait_barrier,
 767                             conf->nr_pending == conf->nr_queued+1,
 768                             conf->resync_lock,
 769                             ({ flush_pending_writes(conf);
 770                                raid10_unplug(conf->mddev->queue); }));
 771         spin_unlock_irq(&conf->resync_lock);
 772 }
 773
 774 static void unfreeze_array(conf_t *conf)
 775 {
 776         /* reverse the effect of the freeze */
 777         spin_lock_irq(&conf->resync_lock);
 778         conf->barrier--;
 779         conf->nr_waiting--;
 780         wake_up(&conf->wait_barrier);
 781         spin_unlock_irq(&conf->resync_lock);
 782 }
 783
 784 static int make_request(struct request_queue *q, struct bio * bio)
 785 {
 786         mddev_t *mddev = q->queuedata;
 787         conf_t *conf = mddev_to_conf(mddev);
 788         mirror_info_t *mirror;
 789         r10bio_t *r10_bio;
 790         struct bio *read_bio;
 791         int i;
 792         int chunk_sects = conf->chunk_mask + 1;
 793         const int rw = bio_data_dir(bio);
 794         const int do_sync = bio_sync(bio);
 795         struct bio_list bl;
 796         unsigned long flags;
 797         mdk_rdev_t *blocked_rdev;
 798
 799         if (unlikely(bio_barrier(bio))) {
 800                 bio_endio(bio, -EOPNOTSUPP);
 801                 return 0;
 802         }
 803
 804         /* If this request crosses a chunk boundary, we need to
 805          * split it.  This will only happen for 1 PAGE (or less) requests.
 806          */
 807         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 808                       > chunk_sects &&
 809                     conf->near_copies < conf->raid_disks)) {
 810                 struct bio_pair *bp;
 811                 /* Sanity check -- queue functions should prevent this happening */
 812                 if (bio->bi_vcnt != 1 ||
 813                     bio->bi_idx != 0)
 814                         goto bad_map;
 815                 /* This is a one page bio that upper layers
 816                  * refuse to split for us, so we need to split it.
 817                  */
 818                 bp = bio_split(bio, bio_split_pool,
 819                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 820                 if (make_request(q, &bp->bio1))
 821                         generic_make_request(&bp->bio1);
 822                 if (make_request(q, &bp->bio2))
 823                         generic_make_request(&bp->bio2);
 824
 825                 bio_pair_release(bp);
 826                 return 0;
 827         bad_map:
 828                 printk("raid10_make_request bug: can't convert block across chunks"
 829                        " or bigger than %dk %llu %d\n", chunk_sects/2,
 830                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 831
 832                 bio_io_error(bio);
 833                 return 0;
 834         }
 835
 836         md_write_start(mddev, bio);
 837
 838         /*
 839          * Register the new request and wait if the reconstruction
 840          * thread has put up a bar for new requests.
 841          * Continue immediately if no resync is active currently.
 842          */
 843         wait_barrier(conf);
 844
 845         disk_stat_inc(mddev->gendisk, ios[rw]);
 846         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
 847
 848         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 849
 850         r10_bio->master_bio = bio;
 851         r10_bio->sectors = bio->bi_size >> 9;
 852
 853         r10_bio->mddev = mddev;
 854         r10_bio->sector = bio->bi_sector;
 855         r10_bio->state = 0;
 856
 857         if (rw == READ) {
 858                 /*
 859                  * read balancing logic:
 860                  */
 861                 int disk = read_balance(conf, r10_bio);
 862                 int slot = r10_bio->read_slot;
 863                 if (disk < 0) {
 864                         raid_end_bio_io(r10_bio);
 865                         return 0;
 866                 }
 867                 mirror = conf->mirrors + disk;
 868
 869                 read_bio = bio_clone(bio, GFP_NOIO);
 870
 871                 r10_bio->devs[slot].bio = read_bio;
 872
 873                 read_bio->bi_sector = r10_bio->devs[slot].addr +
 874                         mirror->rdev->data_offset;
 875                 read_bio->bi_bdev = mirror->rdev->bdev;
 876                 read_bio->bi_end_io = raid10_end_read_request;
 877                 read_bio->bi_rw = READ | do_sync;
 878                 read_bio->bi_private = r10_bio;
 879
 880                 generic_make_request(read_bio);
 881                 return 0;
 882         }
 883
 884         /*
 885          * WRITE:
 886          */
 887         /* first select target devices under rcu_lock and
 888          * inc refcount on their rdev.  Record them by setting
 889          * bios[x] to bio
 890          */
 891         raid10_find_phys(conf, r10_bio);
 892  retry_write:
 893         blocked_rdev = NULL;
 894         rcu_read_lock();
 895         for (i = 0;  i < conf->copies; i++) {
 896                 int d = r10_bio->devs[i].devnum;
 897                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
 898                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 899                         atomic_inc(&rdev->nr_pending);
 900                         blocked_rdev = rdev;
 901                         break;
 902                 }
 903                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 904                         atomic_inc(&rdev->nr_pending);
 905                         r10_bio->devs[i].bio = bio;
 906                 } else {
 907                         r10_bio->devs[i].bio = NULL;
 908                         set_bit(R10BIO_Degraded, &r10_bio->state);
 909                 }
 910         }
 911         rcu_read_unlock();
 912
 913         if (unlikely(blocked_rdev)) {
 914                 /* Have to wait for this device to get unblocked, then retry */
 915                 int j;
 916                 int d;
 917
 918                 for (j = 0; j < i; j++)
 919                         if (r10_bio->devs[j].bio) {
 920                                 d = r10_bio->devs[j].devnum;
 921                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
 922                         }
 923                 allow_barrier(conf);
 924                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 925                 wait_barrier(conf);
 926                 goto retry_write;
 927         }
 928
 929         atomic_set(&r10_bio->remaining, 0);
 930
 931         bio_list_init(&bl);
 932         for (i = 0; i < conf->copies; i++) {
 933                 struct bio *mbio;
 934                 int d = r10_bio->devs[i].devnum;
 935                 if (!r10_bio->devs[i].bio)
 936                         continue;
 937
 938                 mbio = bio_clone(bio, GFP_NOIO);
 939                 r10_bio->devs[i].bio = mbio;
 940
 941                 mbio->bi_sector = r10_bio->devs[i].addr+
 942                         conf->mirrors[d].rdev->data_offset;
 943                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
 944                 mbio->bi_end_io = raid10_end_write_request;
 945                 mbio->bi_rw = WRITE | do_sync;
 946                 mbio->bi_private = r10_bio;
 947
 948                 atomic_inc(&r10_bio->remaining);
 949                 bio_list_add(&bl, mbio);
 950         }
 951
 952         if (unlikely(!atomic_read(&r10_bio->remaining))) {
 953                 /* the array is dead */
 954                 md_write_end(mddev);
 955                 raid_end_bio_io(r10_bio);
 956                 return 0;
 957         }
 958
 959         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
 960         spin_lock_irqsave(&conf->device_lock, flags);
 961         bio_list_merge(&conf->pending_bio_list, &bl);
 962         blk_plug_device(mddev->queue);
 963         spin_unlock_irqrestore(&conf->device_lock, flags);
 964
 965         /* In case raid10d snuck in to freeze_array */
 966         wake_up(&conf->wait_barrier);
 967
 968         if (do_sync)
 969                 md_wakeup_thread(mddev->thread);
 970
 971         return 0;
 972 }
 973
 974 static void status(struct seq_file *seq, mddev_t *mddev)
 975 {
 976         conf_t *conf = mddev_to_conf(mddev);
 977         int i;
 978
 979         if (conf->near_copies < conf->raid_disks)
 980                 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
 981         if (conf->near_copies > 1)
 982                 seq_printf(seq, " %d near-copies", conf->near_copies);
 983         if (conf->far_copies > 1) {
 984                 if (conf->far_offset)
 985                         seq_printf(seq, " %d offset-copies", conf->far_copies);
 986                 else
 987                         seq_printf(seq, " %d far-copies", conf->far_copies);
 988         }
 989         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 990                                         conf->raid_disks - mddev->degraded);
 991         for (i = 0; i < conf->raid_disks; i++)
 992                 seq_printf(seq, "%s",
 993                               conf->mirrors[i].rdev &&
 994                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
 995         seq_printf(seq, "]");
 996 }
 997
 998 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 999 {
1000         char b[BDEVNAME_SIZE];
1001         conf_t *conf = mddev_to_conf(mddev);
1002
1003         /*
1004          * If it is not operational, then we have already marked it as dead
1005          * else if it is the last working disks, ignore the error, let the
1006          * next level up know.
1007          * else mark the drive as failed
1008          */
1009         if (test_bit(In_sync, &rdev->flags)
1010             && conf->raid_disks-mddev->degraded == 1)
1011                 /*
1012                  * Don't fail the drive, just return an IO error.
1013                  * The test should really be more sophisticated than
1014                  * "working_disks == 1", but it isn't critical, and
1015                  * can wait until we do more sophisticated "is the drive
1016                  * really dead" tests...
1017                  */
1018                 return;
1019         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1020                 unsigned long flags;
1021                 spin_lock_irqsave(&conf->device_lock, flags);
1022                 mddev->degraded++;
1023                 spin_unlock_irqrestore(&conf->device_lock, flags);
1024                 /*
1025                  * if recovery is running, make sure it aborts.
1026                  */
1027                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1028         }
1029         set_bit(Faulty, &rdev->flags);
1030         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1031         printk(KERN_ALERT "raid10: Disk failure on %s, disabling device.\n"
1032                 "raid10: Operation continuing on %d devices.\n",
1033                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1034 }
1035
1036 static void print_conf(conf_t *conf)
1037 {
1038         int i;
1039         mirror_info_t *tmp;
1040
1041         printk("RAID10 conf printout:\n");
1042         if (!conf) {
1043                 printk("(!conf)\n");
1044                 return;
1045         }
1046         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1047                 conf->raid_disks);
1048
1049         for (i = 0; i < conf->raid_disks; i++) {
1050                 char b[BDEVNAME_SIZE];
1051                 tmp = conf->mirrors + i;
1052                 if (tmp->rdev)
1053                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1054                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1055                                 !test_bit(Faulty, &tmp->rdev->flags),
1056                                 bdevname(tmp->rdev->bdev,b));
1057         }
1058 }
1059
1060 static void close_sync(conf_t *conf)
1061 {
1062         wait_barrier(conf);
1063         allow_barrier(conf);
1064
1065         mempool_destroy(conf->r10buf_pool);
1066         conf->r10buf_pool = NULL;
1067 }
1068
1069 /* check if there are enough drives for
1070  * every block to appear on atleast one
1071  */
1072 static int enough(conf_t *conf)
1073 {
1074         int first = 0;
1075
1076         do {
1077                 int n = conf->copies;
1078                 int cnt = 0;
1079                 while (n--) {
1080                         if (conf->mirrors[first].rdev)
1081                                 cnt++;
1082                         first = (first+1) % conf->raid_disks;
1083                 }
1084                 if (cnt == 0)
1085                         return 0;
1086         } while (first != 0);
1087         return 1;
1088 }
1089
1090 static int raid10_spare_active(mddev_t *mddev)
1091 {
1092         int i;
1093         conf_t *conf = mddev->private;
1094         mirror_info_t *tmp;
1095
1096         /*
1097          * Find all non-in_sync disks within the RAID10 configuration
1098          * and mark them in_sync
1099          */
1100         for (i = 0; i < conf->raid_disks; i++) {
1101                 tmp = conf->mirrors + i;
1102                 if (tmp->rdev
1103                     && !test_bit(Faulty, &tmp->rdev->flags)
1104                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1105                         unsigned long flags;
1106                         spin_lock_irqsave(&conf->device_lock, flags);
1107                         mddev->degraded--;
1108                         spin_unlock_irqrestore(&conf->device_lock, flags);
1109                 }
1110         }
1111
1112         print_conf(conf);
1113         return 0;
1114 }
1115
1116
1117 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1118 {
1119         conf_t *conf = mddev->private;
1120         int err = -EEXIST;
1121         int mirror;
1122         mirror_info_t *p;
1123         int first = 0;
1124         int last = mddev->raid_disks - 1;
1125
1126         if (mddev->recovery_cp < MaxSector)
1127                 /* only hot-add to in-sync arrays, as recovery is
1128                  * very different from resync
1129                  */
1130                 return -EBUSY;
1131         if (!enough(conf))
1132                 return -EINVAL;
1133
1134         if (rdev->raid_disk)
1135                 first = last = rdev->raid_disk;
1136
1137         if (rdev->saved_raid_disk >= 0 &&
1138             rdev->saved_raid_disk >= first &&
1139             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1140                 mirror = rdev->saved_raid_disk;
1141         else
1142                 mirror = first;
1143         for ( ; mirror <= last ; mirror++)
1144                 if ( !(p=conf->mirrors+mirror)->rdev) {
1145
1146                         blk_queue_stack_limits(mddev->queue,
1147                                                rdev->bdev->bd_disk->queue);
1148                         /* as we don't honour merge_bvec_fn, we must never risk
1149                          * violating it, so limit ->max_sector to one PAGE, as
1150                          * a one page request is never in violation.
1151                          */
1152                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1153                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1154                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1155
1156                         p->head_position = 0;
1157                         rdev->raid_disk = mirror;
1158                         err = 0;
1159                         if (rdev->saved_raid_disk != mirror)
1160                                 conf->fullsync = 1;
1161                         rcu_assign_pointer(p->rdev, rdev);
1162                         break;
1163                 }
1164
1165         print_conf(conf);
1166         return err;
1167 }
1168
1169 static int raid10_remove_disk(mddev_t *mddev, int number)
1170 {
1171         conf_t *conf = mddev->private;
1172         int err = 0;
1173         mdk_rdev_t *rdev;
1174         mirror_info_t *p = conf->mirrors+ number;
1175
1176         print_conf(conf);
1177         rdev = p->rdev;
1178         if (rdev) {
1179                 if (test_bit(In_sync, &rdev->flags) ||
1180                     atomic_read(&rdev->nr_pending)) {
1181                         err = -EBUSY;
1182                         goto abort;
1183                 }
1184                 /* Only remove faulty devices in recovery
1185                  * is not possible.
1186                  */
1187                 if (!test_bit(Faulty, &rdev->flags) &&
1188                     enough(conf)) {
1189                         err = -EBUSY;
1190                         goto abort;
1191                 }
1192                 p->rdev = NULL;
1193                 synchronize_rcu();
1194                 if (atomic_read(&rdev->nr_pending)) {
1195                         /* lost the race, try later */
1196                         err = -EBUSY;
1197                         p->rdev = rdev;
1198                 }
1199         }
1200 abort:
1201
1202         print_conf(conf);
1203         return err;
1204 }
1205
1206
1207 static void end_sync_read(struct bio *bio, int error)
1208 {
1209         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1210         conf_t *conf = mddev_to_conf(r10_bio->mddev);
1211         int i,d;
1212
1213         for (i=0; i<conf->copies; i++)
1214                 if (r10_bio->devs[i].bio == bio)
1215                         break;
1216         BUG_ON(i == conf->copies);
1217         update_head_pos(i, r10_bio);
1218         d = r10_bio->devs[i].devnum;
1219
1220         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1221                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1222         else {
1223                 atomic_add(r10_bio->sectors,
1224                            &conf->mirrors[d].rdev->corrected_errors);
1225                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1226                         md_error(r10_bio->mddev,
1227                                  conf->mirrors[d].rdev);
1228         }
1229
1230         /* for reconstruct, we always reschedule after a read.
1231          * for resync, only after all reads
1232          */
1233         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1234             atomic_dec_and_test(&r10_bio->remaining)) {
1235                 /* we have read all the blocks,
1236                  * do the comparison in process context in raid10d
1237                  */
1238                 reschedule_retry(r10_bio);
1239         }
1240         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1241 }
1242
1243 static void end_sync_write(struct bio *bio, int error)
1244 {
1245         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1246         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1247         mddev_t *mddev = r10_bio->mddev;
1248         conf_t *conf = mddev_to_conf(mddev);
1249         int i,d;
1250
1251         for (i = 0; i < conf->copies; i++)
1252                 if (r10_bio->devs[i].bio == bio)
1253                         break;
1254         d = r10_bio->devs[i].devnum;
1255
1256         if (!uptodate)
1257                 md_error(mddev, conf->mirrors[d].rdev);
1258
1259         update_head_pos(i, r10_bio);
1260
1261         while (atomic_dec_and_test(&r10_bio->remaining)) {
1262                 if (r10_bio->master_bio == NULL) {
1263                         /* the primary of several recovery bios */
1264                         md_done_sync(mddev, r10_bio->sectors, 1);
1265                         put_buf(r10_bio);
1266                         break;
1267                 } else {
1268                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1269                         put_buf(r10_bio);
1270                         r10_bio = r10_bio2;
1271                 }
1272         }
1273         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1274 }
1275
1276 /*
1277  * Note: sync and recover and handled very differently for raid10
1278  * This code is for resync.
1279  * For resync, we read through virtual addresses and read all blocks.
1280  * If there is any error, we schedule a write.  The lowest numbered
1281  * drive is authoritative.
1282  * However requests come for physical address, so we need to map.
1283  * For every physical address there are raid_disks/copies virtual addresses,
1284  * which is always are least one, but is not necessarly an integer.
1285  * This means that a physical address can span multiple chunks, so we may
1286  * have to submit multiple io requests for a single sync request.
1287  */
1288 /*
1289  * We check if all blocks are in-sync and only write to blocks that
1290  * aren't in sync
1291  */
1292 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1293 {
1294         conf_t *conf = mddev_to_conf(mddev);
1295         int i, first;
1296         struct bio *tbio, *fbio;
1297
1298         atomic_set(&r10_bio->remaining, 1);
1299
1300         /* find the first device with a block */
1301         for (i=0; i<conf->copies; i++)
1302                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1303                         break;
1304
1305         if (i == conf->copies)
1306                 goto done;
1307
1308         first = i;
1309         fbio = r10_bio->devs[i].bio;
1310
1311         /* now find blocks with errors */
1312         for (i=0 ; i < conf->copies ; i++) {
1313                 int  j, d;
1314                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1315
1316                 tbio = r10_bio->devs[i].bio;
1317
1318                 if (tbio->bi_end_io != end_sync_read)
1319                         continue;
1320                 if (i == first)
1321                         continue;
1322                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1323                         /* We know that the bi_io_vec layout is the same for
1324                          * both 'first' and 'i', so we just compare them.
1325                          * All vec entries are PAGE_SIZE;
1326                          */
1327                         for (j = 0; j < vcnt; j++)
1328                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1329                                            page_address(tbio->bi_io_vec[j].bv_page),
1330                                            PAGE_SIZE))
1331                                         break;
1332                         if (j == vcnt)
1333                                 continue;
1334                         mddev->resync_mismatches += r10_bio->sectors;
1335                 }
1336                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1337                         /* Don't fix anything. */
1338                         continue;
1339                 /* Ok, we need to write this bio
1340                  * First we need to fixup bv_offset, bv_len and
1341                  * bi_vecs, as the read request might have corrupted these
1342                  */
1343                 tbio->bi_vcnt = vcnt;
1344                 tbio->bi_size = r10_bio->sectors << 9;
1345                 tbio->bi_idx = 0;
1346                 tbio->bi_phys_segments = 0;
1347                 tbio->bi_hw_segments = 0;
1348                 tbio->bi_hw_front_size = 0;
1349                 tbio->bi_hw_back_size = 0;
1350                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1351                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1352                 tbio->bi_next = NULL;
1353                 tbio->bi_rw = WRITE;
1354                 tbio->bi_private = r10_bio;
1355                 tbio->bi_sector = r10_bio->devs[i].addr;
1356
1357                 for (j=0; j < vcnt ; j++) {
1358                         tbio->bi_io_vec[j].bv_offset = 0;
1359                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1360
1361                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1362                                page_address(fbio->bi_io_vec[j].bv_page),
1363                                PAGE_SIZE);
1364                 }
1365                 tbio->bi_end_io = end_sync_write;
1366
1367                 d = r10_bio->devs[i].devnum;
1368                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1369                 atomic_inc(&r10_bio->remaining);
1370                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1371
1372                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1373                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1374                 generic_make_request(tbio);
1375         }
1376
1377 done:
1378         if (atomic_dec_and_test(&r10_bio->remaining)) {
1379                 md_done_sync(mddev, r10_bio->sectors, 1);
1380                 put_buf(r10_bio);
1381         }
1382 }
1383
1384 /*
1385  * Now for the recovery code.
1386  * Recovery happens across physical sectors.
1387  * We recover all non-is_sync drives by finding the virtual address of
1388  * each, and then choose a working drive that also has that virt address.
1389  * There is a separate r10_bio for each non-in_sync drive.
1390  * Only the first two slots are in use. The first for reading,
1391  * The second for writing.
1392  *
1393  */
1394
1395 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1396 {
1397         conf_t *conf = mddev_to_conf(mddev);
1398         int i, d;
1399         struct bio *bio, *wbio;
1400
1401
1402         /* move the pages across to the second bio
1403          * and submit the write request
1404          */
1405         bio = r10_bio->devs[0].bio;
1406         wbio = r10_bio->devs[1].bio;
1407         for (i=0; i < wbio->bi_vcnt; i++) {
1408                 struct page *p = bio->bi_io_vec[i].bv_page;
1409                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1410                 wbio->bi_io_vec[i].bv_page = p;
1411         }
1412         d = r10_bio->devs[1].devnum;
1413
1414         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1415         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1416         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1417                 generic_make_request(wbio);
1418         else
1419                 bio_endio(wbio, -EIO);
1420 }
1421
1422
1423 /*
1424  * This is a kernel thread which:
1425  *
1426  *      1.      Retries failed read operations on working mirrors.
1427  *      2.      Updates the raid superblock when problems encounter.
1428  *      3.      Performs writes following reads for array synchronising.
1429  */
1430
1431 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1432 {
1433         int sect = 0; /* Offset from r10_bio->sector */
1434         int sectors = r10_bio->sectors;
1435         mdk_rdev_t*rdev;
1436         while(sectors) {
1437                 int s = sectors;
1438                 int sl = r10_bio->read_slot;
1439                 int success = 0;
1440                 int start;
1441
1442                 if (s > (PAGE_SIZE>>9))
1443                         s = PAGE_SIZE >> 9;
1444
1445                 rcu_read_lock();
1446                 do {
1447                         int d = r10_bio->devs[sl].devnum;
1448                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1449                         if (rdev &&
1450                             test_bit(In_sync, &rdev->flags)) {
1451                                 atomic_inc(&rdev->nr_pending);
1452                                 rcu_read_unlock();
1453                                 success = sync_page_io(rdev->bdev,
1454                                                        r10_bio->devs[sl].addr +
1455                                                        sect + rdev->data_offset,
1456                                                        s<<9,
1457                                                        conf->tmppage, READ);
1458                                 rdev_dec_pending(rdev, mddev);
1459                                 rcu_read_lock();
1460                                 if (success)
1461                                         break;
1462                         }
1463                         sl++;
1464                         if (sl == conf->copies)
1465                                 sl = 0;
1466                 } while (!success && sl != r10_bio->read_slot);
1467                 rcu_read_unlock();
1468
1469                 if (!success) {
1470                         /* Cannot read from anywhere -- bye bye array */
1471                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1472                         md_error(mddev, conf->mirrors[dn].rdev);
1473                         break;
1474                 }
1475
1476                 start = sl;
1477                 /* write it back and re-read */
1478                 rcu_read_lock();
1479                 while (sl != r10_bio->read_slot) {
1480                         int d;
1481                         if (sl==0)
1482                                 sl = conf->copies;
1483                         sl--;
1484                         d = r10_bio->devs[sl].devnum;
1485                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1486                         if (rdev &&
1487                             test_bit(In_sync, &rdev->flags)) {
1488                                 atomic_inc(&rdev->nr_pending);
1489                                 rcu_read_unlock();
1490                                 atomic_add(s, &rdev->corrected_errors);
1491                                 if (sync_page_io(rdev->bdev,
1492                                                  r10_bio->devs[sl].addr +
1493                                                  sect + rdev->data_offset,
1494                                                  s<<9, conf->tmppage, WRITE)
1495                                     == 0)
1496                                         /* Well, this device is dead */
1497                                         md_error(mddev, rdev);
1498                                 rdev_dec_pending(rdev, mddev);
1499                                 rcu_read_lock();
1500                         }
1501                 }
1502                 sl = start;
1503                 while (sl != r10_bio->read_slot) {
1504                         int d;
1505                         if (sl==0)
1506                                 sl = conf->copies;
1507                         sl--;
1508                         d = r10_bio->devs[sl].devnum;
1509                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1510                         if (rdev &&
1511                             test_bit(In_sync, &rdev->flags)) {
1512                                 char b[BDEVNAME_SIZE];
1513                                 atomic_inc(&rdev->nr_pending);
1514                                 rcu_read_unlock();
1515                                 if (sync_page_io(rdev->bdev,
1516                                                  r10_bio->devs[sl].addr +
1517                                                  sect + rdev->data_offset,
1518                                                  s<<9, conf->tmppage, READ) == 0)
1519                                         /* Well, this device is dead */
1520                                         md_error(mddev, rdev);
1521                                 else
1522                                         printk(KERN_INFO
1523                                                "raid10:%s: read error corrected"
1524                                                " (%d sectors at %llu on %s)\n",
1525                                                mdname(mddev), s,
1526                                                (unsigned long long)(sect+
1527                                                     rdev->data_offset),
1528                                                bdevname(rdev->bdev, b));
1529
1530                                 rdev_dec_pending(rdev, mddev);
1531                                 rcu_read_lock();
1532                         }
1533                 }
1534                 rcu_read_unlock();
1535
1536                 sectors -= s;
1537                 sect += s;
1538         }
1539 }
1540
1541 static void raid10d(mddev_t *mddev)
1542 {
1543         r10bio_t *r10_bio;
1544         struct bio *bio;
1545         unsigned long flags;
1546         conf_t *conf = mddev_to_conf(mddev);
1547         struct list_head *head = &conf->retry_list;
1548         int unplug=0;
1549         mdk_rdev_t *rdev;
1550
1551         md_check_recovery(mddev);
1552
1553         for (;;) {
1554                 char b[BDEVNAME_SIZE];
1555
1556                 unplug += flush_pending_writes(conf);
1557
1558                 spin_lock_irqsave(&conf->device_lock, flags);
1559                 if (list_empty(head)) {
1560                         spin_unlock_irqrestore(&conf->device_lock, flags);
1561                         break;
1562                 }
1563                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1564                 list_del(head->prev);
1565                 conf->nr_queued--;
1566                 spin_unlock_irqrestore(&conf->device_lock, flags);
1567
1568                 mddev = r10_bio->mddev;
1569                 conf = mddev_to_conf(mddev);
1570                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1571                         sync_request_write(mddev, r10_bio);
1572                         unplug = 1;
1573                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1574                         recovery_request_write(mddev, r10_bio);
1575                         unplug = 1;
1576                 } else {
1577                         int mirror;
1578                         /* we got a read error. Maybe the drive is bad.  Maybe just
1579                          * the block and we can fix it.
1580                          * We freeze all other IO, and try reading the block from
1581                          * other devices.  When we find one, we re-write
1582                          * and check it that fixes the read error.
1583                          * This is all done synchronously while the array is
1584                          * frozen.
1585                          */
1586                         if (mddev->ro == 0) {
1587                                 freeze_array(conf);
1588                                 fix_read_error(conf, mddev, r10_bio);
1589                                 unfreeze_array(conf);
1590                         }
1591
1592                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1593                         r10_bio->devs[r10_bio->read_slot].bio =
1594                                 mddev->ro ? IO_BLOCKED : NULL;
1595                         mirror = read_balance(conf, r10_bio);
1596                         if (mirror == -1) {
1597                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1598                                        " read error for block %llu\n",
1599                                        bdevname(bio->bi_bdev,b),
1600                                        (unsigned long long)r10_bio->sector);
1601                                 raid_end_bio_io(r10_bio);
1602                                 bio_put(bio);
1603                         } else {
1604                                 const int do_sync = bio_sync(r10_bio->master_bio);
1605                                 bio_put(bio);
1606                                 rdev = conf->mirrors[mirror].rdev;
1607                                 if (printk_ratelimit())
1608                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1609                                                " another mirror\n",
1610                                                bdevname(rdev->bdev,b),
1611                                                (unsigned long long)r10_bio->sector);
1612                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1613                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1614                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1615                                         + rdev->data_offset;
1616                                 bio->bi_bdev = rdev->bdev;
1617                                 bio->bi_rw = READ | do_sync;
1618                                 bio->bi_private = r10_bio;
1619                                 bio->bi_end_io = raid10_end_read_request;
1620                                 unplug = 1;
1621                                 generic_make_request(bio);
1622                         }
1623                 }
1624         }
1625         if (unplug)
1626                 unplug_slaves(mddev);
1627 }
1628
1629
1630 static int init_resync(conf_t *conf)
1631 {
1632         int buffs;
1633
1634         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1635         BUG_ON(conf->r10buf_pool);
1636         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1637         if (!conf->r10buf_pool)
1638                 return -ENOMEM;
1639         conf->next_resync = 0;
1640         return 0;
1641 }
1642
1643 /*
1644  * perform a "sync" on one "block"
1645  *
1646  * We need to make sure that no normal I/O request - particularly write
1647  * requests - conflict with active sync requests.
1648  *
1649  * This is achieved by tracking pending requests and a 'barrier' concept
1650  * that can be installed to exclude normal IO requests.
1651  *
1652  * Resync and recovery are handled very differently.
1653  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1654  *
1655  * For resync, we iterate over virtual addresses, read all copies,
1656  * and update if there are differences.  If only one copy is live,
1657  * skip it.
1658  * For recovery, we iterate over physical addresses, read a good
1659  * value for each non-in_sync drive, and over-write.
1660  *
1661  * So, for recovery we may have several outstanding complex requests for a
1662  * given address, one for each out-of-sync device.  We model this by allocating
1663  * a number of r10_bio structures, one for each out-of-sync device.
1664  * As we setup these structures, we collect all bio's together into a list
1665  * which we then process collectively to add pages, and then process again
1666  * to pass to generic_make_request.
1667  *
1668  * The r10_bio structures are linked using a borrowed master_bio pointer.
1669  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1670  * has its remaining count decremented to 0, the whole complex operation
1671  * is complete.
1672  *
1673  */
1674
1675 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1676 {
1677         conf_t *conf = mddev_to_conf(mddev);
1678         r10bio_t *r10_bio;
1679         struct bio *biolist = NULL, *bio;
1680         sector_t max_sector, nr_sectors;
1681         int disk;
1682         int i;
1683         int max_sync;
1684         int sync_blocks;
1685
1686         sector_t sectors_skipped = 0;
1687         int chunks_skipped = 0;
1688
1689         if (!conf->r10buf_pool)
1690                 if (init_resync(conf))
1691                         return 0;
1692
1693  skipped:
1694         max_sector = mddev->size << 1;
1695         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1696                 max_sector = mddev->resync_max_sectors;
1697         if (sector_nr >= max_sector) {
1698                 /* If we aborted, we need to abort the
1699                  * sync on the 'current' bitmap chucks (there can
1700                  * be several when recovering multiple devices).
1701                  * as we may have started syncing it but not finished.
1702                  * We can find the current address in
1703                  * mddev->curr_resync, but for recovery,
1704                  * we need to convert that to several
1705                  * virtual addresses.
1706                  */
1707                 if (mddev->curr_resync < max_sector) { /* aborted */
1708                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1709                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1710                                                 &sync_blocks, 1);
1711                         else for (i=0; i<conf->raid_disks; i++) {
1712                                 sector_t sect =
1713                                         raid10_find_virt(conf, mddev->curr_resync, i);
1714                                 bitmap_end_sync(mddev->bitmap, sect,
1715                                                 &sync_blocks, 1);
1716                         }
1717                 } else /* completed sync */
1718                         conf->fullsync = 0;
1719
1720                 bitmap_close_sync(mddev->bitmap);
1721                 close_sync(conf);
1722                 *skipped = 1;
1723                 return sectors_skipped;
1724         }
1725         if (chunks_skipped >= conf->raid_disks) {
1726                 /* if there has been nothing to do on any drive,
1727                  * then there is nothing to do at all..
1728                  */
1729                 *skipped = 1;
1730                 return (max_sector - sector_nr) + sectors_skipped;
1731         }
1732
1733         if (max_sector > mddev->resync_max)
1734                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1735
1736         /* make sure whole request will fit in a chunk - if chunks
1737          * are meaningful
1738          */
1739         if (conf->near_copies < conf->raid_disks &&
1740             max_sector > (sector_nr | conf->chunk_mask))
1741                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1742         /*
1743          * If there is non-resync activity waiting for us then
1744          * put in a delay to throttle resync.
1745          */
1746         if (!go_faster && conf->nr_waiting)
1747                 msleep_interruptible(1000);
1748
1749         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1750
1751         /* Again, very different code for resync and recovery.
1752          * Both must result in an r10bio with a list of bios that
1753          * have bi_end_io, bi_sector, bi_bdev set,
1754          * and bi_private set to the r10bio.
1755          * For recovery, we may actually create several r10bios
1756          * with 2 bios in each, that correspond to the bios in the main one.
1757          * In this case, the subordinate r10bios link back through a
1758          * borrowed master_bio pointer, and the counter in the master
1759          * includes a ref from each subordinate.
1760          */
1761         /* First, we decide what to do and set ->bi_end_io
1762          * To end_sync_read if we want to read, and
1763          * end_sync_write if we will want to write.
1764          */
1765
1766         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1767         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1768                 /* recovery... the complicated one */
1769                 int i, j, k;
1770                 r10_bio = NULL;
1771
1772                 for (i=0 ; i<conf->raid_disks; i++)
1773                         if (conf->mirrors[i].rdev &&
1774                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1775                                 int still_degraded = 0;
1776                                 /* want to reconstruct this device */
1777                                 r10bio_t *rb2 = r10_bio;
1778                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1779                                 int must_sync;
1780                                 /* Unless we are doing a full sync, we only need
1781                                  * to recover the block if it is set in the bitmap
1782                                  */
1783                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1784                                                               &sync_blocks, 1);
1785                                 if (sync_blocks < max_sync)
1786                                         max_sync = sync_blocks;
1787                                 if (!must_sync &&
1788                                     !conf->fullsync) {
1789                                         /* yep, skip the sync_blocks here, but don't assume
1790                                          * that there will never be anything to do here
1791                                          */
1792                                         chunks_skipped = -1;
1793                                         continue;
1794                                 }
1795
1796                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1797                                 raise_barrier(conf, rb2 != NULL);
1798                                 atomic_set(&r10_bio->remaining, 0);
1799
1800                                 r10_bio->master_bio = (struct bio*)rb2;
1801                                 if (rb2)
1802                                         atomic_inc(&rb2->remaining);
1803                                 r10_bio->mddev = mddev;
1804                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1805                                 r10_bio->sector = sect;
1806
1807                                 raid10_find_phys(conf, r10_bio);
1808                                 /* Need to check if this section will still be
1809                                  * degraded
1810                                  */
1811                                 for (j=0; j<conf->copies;j++) {
1812                                         int d = r10_bio->devs[j].devnum;
1813                                         if (conf->mirrors[d].rdev == NULL ||
1814                                             test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
1815                                                 still_degraded = 1;
1816                                                 break;
1817                                         }
1818                                 }
1819                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1820                                                               &sync_blocks, still_degraded);
1821
1822                                 for (j=0; j<conf->copies;j++) {
1823                                         int d = r10_bio->devs[j].devnum;
1824                                         if (conf->mirrors[d].rdev &&
1825                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1826                                                 /* This is where we read from */
1827                                                 bio = r10_bio->devs[0].bio;
1828                                                 bio->bi_next = biolist;
1829                                                 biolist = bio;
1830                                                 bio->bi_private = r10_bio;
1831                                                 bio->bi_end_io = end_sync_read;
1832                                                 bio->bi_rw = READ;
1833                                                 bio->bi_sector = r10_bio->devs[j].addr +
1834                                                         conf->mirrors[d].rdev->data_offset;
1835                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1836                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1837                                                 atomic_inc(&r10_bio->remaining);
1838                                                 /* and we write to 'i' */
1839
1840                                                 for (k=0; k<conf->copies; k++)
1841                                                         if (r10_bio->devs[k].devnum == i)
1842                                                                 break;
1843                                                 BUG_ON(k == conf->copies);
1844                                                 bio = r10_bio->devs[1].bio;
1845                                                 bio->bi_next = biolist;
1846                                                 biolist = bio;
1847                                                 bio->bi_private = r10_bio;
1848                                                 bio->bi_end_io = end_sync_write;
1849                                                 bio->bi_rw = WRITE;
1850                                                 bio->bi_sector = r10_bio->devs[k].addr +
1851                                                         conf->mirrors[i].rdev->data_offset;
1852                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1853
1854                                                 r10_bio->devs[0].devnum = d;
1855                                                 r10_bio->devs[1].devnum = i;
1856
1857                                                 break;
1858                                         }
1859                                 }
1860                                 if (j == conf->copies) {
1861                                         /* Cannot recover, so abort the recovery */
1862                                         put_buf(r10_bio);
1863                                         if (rb2)
1864                                                 atomic_dec(&rb2->remaining);
1865                                         r10_bio = rb2;
1866                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
1867                                                               &mddev->recovery))
1868                                                 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1869                                                        mdname(mddev));
1870                                         break;
1871                                 }
1872                         }
1873                 if (biolist == NULL) {
1874                         while (r10_bio) {
1875                                 r10bio_t *rb2 = r10_bio;
1876                                 r10_bio = (r10bio_t*) rb2->master_bio;
1877                                 rb2->master_bio = NULL;
1878                                 put_buf(rb2);
1879                         }
1880                         goto giveup;
1881                 }
1882         } else {
1883                 /* resync. Schedule a read for every block at this virt offset */
1884                 int count = 0;
1885
1886                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1887                                        &sync_blocks, mddev->degraded) &&
1888                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1889                         /* We can skip this block */
1890                         *skipped = 1;
1891                         return sync_blocks + sectors_skipped;
1892                 }
1893                 if (sync_blocks < max_sync)
1894                         max_sync = sync_blocks;
1895                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1896
1897                 r10_bio->mddev = mddev;
1898                 atomic_set(&r10_bio->remaining, 0);
1899                 raise_barrier(conf, 0);
1900                 conf->next_resync = sector_nr;
1901
1902                 r10_bio->master_bio = NULL;
1903                 r10_bio->sector = sector_nr;
1904                 set_bit(R10BIO_IsSync, &r10_bio->state);
1905                 raid10_find_phys(conf, r10_bio);
1906                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1907
1908                 for (i=0; i<conf->copies; i++) {
1909                         int d = r10_bio->devs[i].devnum;
1910                         bio = r10_bio->devs[i].bio;
1911                         bio->bi_end_io = NULL;
1912                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
1913                         if (conf->mirrors[d].rdev == NULL ||
1914                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1915                                 continue;
1916                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1917                         atomic_inc(&r10_bio->remaining);
1918                         bio->bi_next = biolist;
1919                         biolist = bio;
1920                         bio->bi_private = r10_bio;
1921                         bio->bi_end_io = end_sync_read;
1922                         bio->bi_rw = READ;
1923                         bio->bi_sector = r10_bio->devs[i].addr +
1924                                 conf->mirrors[d].rdev->data_offset;
1925                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1926                         count++;
1927                 }
1928
1929                 if (count < 2) {
1930                         for (i=0; i<conf->copies; i++) {
1931                                 int d = r10_bio->devs[i].devnum;
1932                                 if (r10_bio->devs[i].bio->bi_end_io)
1933                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1934                         }
1935                         put_buf(r10_bio);
1936                         biolist = NULL;
1937                         goto giveup;
1938                 }
1939         }
1940
1941         for (bio = biolist; bio ; bio=bio->bi_next) {
1942
1943                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1944                 if (bio->bi_end_io)
1945                         bio->bi_flags |= 1 << BIO_UPTODATE;
1946                 bio->bi_vcnt = 0;
1947                 bio->bi_idx = 0;
1948                 bio->bi_phys_segments = 0;
1949                 bio->bi_hw_segments = 0;
1950                 bio->bi_size = 0;
1951         }
1952
1953         nr_sectors = 0;
1954         if (sector_nr + max_sync < max_sector)
1955                 max_sector = sector_nr + max_sync;
1956         do {
1957                 struct page *page;
1958                 int len = PAGE_SIZE;
1959                 disk = 0;
1960                 if (sector_nr + (len>>9) > max_sector)
1961                         len = (max_sector - sector_nr) << 9;
1962                 if (len == 0)
1963                         break;
1964                 for (bio= biolist ; bio ; bio=bio->bi_next) {
1965                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1966                         if (bio_add_page(bio, page, len, 0) == 0) {
1967                                 /* stop here */
1968                                 struct bio *bio2;
1969                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1970                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1971                                         /* remove last page from this bio */
1972                                         bio2->bi_vcnt--;
1973                                         bio2->bi_size -= len;
1974                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1975                                 }
1976                                 goto bio_full;
1977                         }
1978                         disk = i;
1979                 }
1980                 nr_sectors += len>>9;
1981                 sector_nr += len>>9;
1982         } while (biolist->bi_vcnt < RESYNC_PAGES);
1983  bio_full:
1984         r10_bio->sectors = nr_sectors;
1985
1986         while (biolist) {
1987                 bio = biolist;
1988                 biolist = biolist->bi_next;
1989
1990                 bio->bi_next = NULL;
1991                 r10_bio = bio->bi_private;
1992                 r10_bio->sectors = nr_sectors;
1993
1994                 if (bio->bi_end_io == end_sync_read) {
1995                         md_sync_acct(bio->bi_bdev, nr_sectors);
1996                         generic_make_request(bio);
1997                 }
1998         }
1999
2000         if (sectors_skipped)
2001                 /* pretend they weren't skipped, it makes
2002                  * no important difference in this case
2003                  */
2004                 md_done_sync(mddev, sectors_skipped, 1);
2005
2006         return sectors_skipped + nr_sectors;
2007  giveup:
2008         /* There is nowhere to write, so all non-sync
2009          * drives must be failed, so try the next chunk...
2010          */
2011         {
2012         sector_t sec = max_sector - sector_nr;
2013         sectors_skipped += sec;
2014         chunks_skipped ++;
2015         sector_nr = max_sector;
2016         goto skipped;
2017         }
2018 }
2019
2020 static int run(mddev_t *mddev)
2021 {
2022         conf_t *conf;
2023         int i, disk_idx;
2024         mirror_info_t *disk;
2025         mdk_rdev_t *rdev;
2026         struct list_head *tmp;
2027         int nc, fc, fo;
2028         sector_t stride, size;
2029
2030         if (mddev->chunk_size == 0) {
2031                 printk(KERN_ERR "md/raid10: non-zero chunk size required.\n");
2032                 return -EINVAL;
2033         }
2034
2035         nc = mddev->layout & 255;
2036         fc = (mddev->layout >> 8) & 255;
2037         fo = mddev->layout & (1<<16);
2038         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2039             (mddev->layout >> 17)) {
2040                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
2041                        mdname(mddev), mddev->layout);
2042                 goto out;
2043         }
2044         /*
2045          * copy the already verified devices into our private RAID10
2046          * bookkeeping area. [whatever we allocate in run(),
2047          * should be freed in stop()]
2048          */
2049         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2050         mddev->private = conf;
2051         if (!conf) {
2052                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2053                         mdname(mddev));
2054                 goto out;
2055         }
2056         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2057                                  GFP_KERNEL);
2058         if (!conf->mirrors) {
2059                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2060                        mdname(mddev));
2061                 goto out_free_conf;
2062         }
2063
2064         conf->tmppage = alloc_page(GFP_KERNEL);
2065         if (!conf->tmppage)
2066                 goto out_free_conf;
2067
2068         conf->mddev = mddev;
2069         conf->raid_disks = mddev->raid_disks;
2070         conf->near_copies = nc;
2071         conf->far_copies = fc;
2072         conf->copies = nc*fc;
2073         conf->far_offset = fo;
2074         conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
2075         conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
2076         size = mddev->size >> (conf->chunk_shift-1);
2077         sector_div(size, fc);
2078         size = size * conf->raid_disks;
2079         sector_div(size, nc);
2080         /* 'size' is now the number of chunks in the array */
2081         /* calculate "used chunks per device" in 'stride' */
2082         stride = size * conf->copies;
2083
2084         /* We need to round up when dividing by raid_disks to
2085          * get the stride size.
2086          */
2087         stride += conf->raid_disks - 1;
2088         sector_div(stride, conf->raid_disks);
2089         mddev->size = stride  << (conf->chunk_shift-1);
2090
2091         if (fo)
2092                 stride = 1;
2093         else
2094                 sector_div(stride, fc);
2095         conf->stride = stride << conf->chunk_shift;
2096
2097         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2098                                                 r10bio_pool_free, conf);
2099         if (!conf->r10bio_pool) {
2100                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2101                         mdname(mddev));
2102                 goto out_free_conf;
2103         }
2104
2105         spin_lock_init(&conf->device_lock);
2106         mddev->queue->queue_lock = &conf->device_lock;
2107
2108         rdev_for_each(rdev, tmp, mddev) {
2109                 disk_idx = rdev->raid_disk;
2110                 if (disk_idx >= mddev->raid_disks
2111                     || disk_idx < 0)
2112                         continue;
2113                 disk = conf->mirrors + disk_idx;
2114
2115                 disk->rdev = rdev;
2116
2117                 blk_queue_stack_limits(mddev->queue,
2118                                        rdev->bdev->bd_disk->queue);
2119                 /* as we don't honour merge_bvec_fn, we must never risk
2120                  * violating it, so limit ->max_sector to one PAGE, as
2121                  * a one page request is never in violation.
2122                  */
2123                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2124                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
2125                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
2126
2127                 disk->head_position = 0;
2128         }
2129         INIT_LIST_HEAD(&conf->retry_list);
2130
2131         spin_lock_init(&conf->resync_lock);
2132         init_waitqueue_head(&conf->wait_barrier);
2133
2134         /* need to check that every block has at least one working mirror */
2135         if (!enough(conf)) {
2136                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2137                        mdname(mddev));
2138                 goto out_free_conf;
2139         }
2140
2141         mddev->degraded = 0;
2142         for (i = 0; i < conf->raid_disks; i++) {
2143
2144                 disk = conf->mirrors + i;
2145
2146                 if (!disk->rdev ||
2147                     !test_bit(In_sync, &disk->rdev->flags)) {
2148                         disk->head_position = 0;
2149                         mddev->degraded++;
2150                         if (disk->rdev)
2151                                 conf->fullsync = 1;
2152                 }
2153         }
2154
2155
2156         mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2157         if (!mddev->thread) {
2158                 printk(KERN_ERR
2159                        "raid10: couldn't allocate thread for %s\n",
2160                        mdname(mddev));
2161                 goto out_free_conf;
2162         }
2163
2164         printk(KERN_INFO
2165                 "raid10: raid set %s active with %d out of %d devices\n",
2166                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2167                 mddev->raid_disks);
2168         /*
2169          * Ok, everything is just fine now
2170          */
2171         mddev->array_sectors = size << conf->chunk_shift;
2172         mddev->resync_max_sectors = size << conf->chunk_shift;
2173
2174         mddev->queue->unplug_fn = raid10_unplug;
2175         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2176         mddev->queue->backing_dev_info.congested_data = mddev;
2177
2178         /* Calculate max read-ahead size.
2179          * We need to readahead at least twice a whole stripe....
2180          * maybe...
2181          */
2182         {
2183                 int stripe = conf->raid_disks * (mddev->chunk_size / PAGE_SIZE);
2184                 stripe /= conf->near_copies;
2185                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2186                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2187         }
2188
2189         if (conf->near_copies < mddev->raid_disks)
2190                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2191         return 0;
2192
2193 out_free_conf:
2194         if (conf->r10bio_pool)
2195                 mempool_destroy(conf->r10bio_pool);
2196         safe_put_page(conf->tmppage);
2197         kfree(conf->mirrors);
2198         kfree(conf);
2199         mddev->private = NULL;
2200 out:
2201         return -EIO;
2202 }
2203
2204 static int stop(mddev_t *mddev)
2205 {
2206         conf_t *conf = mddev_to_conf(mddev);
2207
2208         md_unregister_thread(mddev->thread);
2209         mddev->thread = NULL;
2210         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2211         if (conf->r10bio_pool)
2212                 mempool_destroy(conf->r10bio_pool);
2213         kfree(conf->mirrors);
2214         kfree(conf);
2215         mddev->private = NULL;
2216         return 0;
2217 }
2218
2219 static void raid10_quiesce(mddev_t *mddev, int state)
2220 {
2221         conf_t *conf = mddev_to_conf(mddev);
2222
2223         switch(state) {
2224         case 1:
2225                 raise_barrier(conf, 0);
2226                 break;
2227         case 0:
2228                 lower_barrier(conf);
2229                 break;
2230         }
2231         if (mddev->thread) {
2232                 if (mddev->bitmap)
2233                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2234                 else
2235                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2236                 md_wakeup_thread(mddev->thread);
2237         }
2238 }
2239
2240 static struct mdk_personality raid10_personality =
2241 {
2242         .name           = "raid10",
2243         .level          = 10,
2244         .owner          = THIS_MODULE,
2245         .make_request   = make_request,
2246         .run            = run,
2247         .stop           = stop,
2248         .status         = status,
2249         .error_handler  = error,
2250         .hot_add_disk   = raid10_add_disk,
2251         .hot_remove_disk= raid10_remove_disk,
2252         .spare_active   = raid10_spare_active,
2253         .sync_request   = sync_request,
2254         .quiesce        = raid10_quiesce,
2255 };
2256
2257 static int __init raid_init(void)
2258 {
2259         return register_md_personality(&raid10_personality);
2260 }
2261
2262 static void raid_exit(void)
2263 {
2264         unregister_md_personality(&raid10_personality);
2265 }
2266
2267 module_init(raid_init);
2268 module_exit(raid_exit);
2269 MODULE_LICENSE("GPL");
2270 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2271 MODULE_ALIAS("md-raid10");
2272 MODULE_ALIAS("md-level-10");