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         md_wakeup_thread(mddev->thread);
 219 }
 220
 221 /*
 222  * raid_end_bio_io() is called when we have finished servicing a mirrored
 223  * operation and are ready to return a success/failure code to the buffer
 224  * cache layer.
 225  */
 226 static void raid_end_bio_io(r10bio_t *r10_bio)
 227 {
 228         struct bio *bio = r10_bio->master_bio;
 229
 230         bio_endio(bio,
 231                 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
 232         free_r10bio(r10_bio);
 233 }
 234
 235 /*
 236  * Update disk head position estimator based on IRQ completion info.
 237  */
 238 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
 239 {
 240         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 241
 242         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 243                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 244 }
 245
 246 static void raid10_end_read_request(struct bio *bio, int error)
 247 {
 248         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 249         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
 250         int slot, dev;
 251         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 252
 253
 254         slot = r10_bio->read_slot;
 255         dev = r10_bio->devs[slot].devnum;
 256         /*
 257          * this branch is our 'one mirror IO has finished' event handler:
 258          */
 259         update_head_pos(slot, r10_bio);
 260
 261         if (uptodate) {
 262                 /*
 263                  * Set R10BIO_Uptodate in our master bio, so that
 264                  * we will return a good error code to the higher
 265                  * levels even if IO on some other mirrored buffer fails.
 266                  *
 267                  * The 'master' represents the composite IO operation to
 268                  * user-side. So if something waits for IO, then it will
 269                  * wait for the 'master' bio.
 270                  */
 271                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 272                 raid_end_bio_io(r10_bio);
 273         } else {
 274                 /*
 275                  * oops, read error:
 276                  */
 277                 char b[BDEVNAME_SIZE];
 278                 if (printk_ratelimit())
 279                         printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
 280                                bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
 281                 reschedule_retry(r10_bio);
 282         }
 283
 284         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 285 }
 286
 287 static void raid10_end_write_request(struct bio *bio, int error)
 288 {
 289         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 290         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
 291         int slot, dev;
 292         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 293
 294         for (slot = 0; slot < conf->copies; slot++)
 295                 if (r10_bio->devs[slot].bio == bio)
 296                         break;
 297         dev = r10_bio->devs[slot].devnum;
 298
 299         /*
 300          * this branch is our 'one mirror IO has finished' event handler:
 301          */
 302         if (!uptodate) {
 303                 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
 304                 /* an I/O failed, we can't clear the bitmap */
 305                 set_bit(R10BIO_Degraded, &r10_bio->state);
 306         } else
 307                 /*
 308                  * Set R10BIO_Uptodate in our master bio, so that
 309                  * we will return a good error code for to the higher
 310                  * levels even if IO on some other mirrored buffer fails.
 311                  *
 312                  * The 'master' represents the composite IO operation to
 313                  * user-side. So if something waits for IO, then it will
 314                  * wait for the 'master' bio.
 315                  */
 316                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 317
 318         update_head_pos(slot, r10_bio);
 319
 320         /*
 321          *
 322          * Let's see if all mirrored write operations have finished
 323          * already.
 324          */
 325         if (atomic_dec_and_test(&r10_bio->remaining)) {
 326                 /* clear the bitmap if all writes complete successfully */
 327                 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 328                                 r10_bio->sectors,
 329                                 !test_bit(R10BIO_Degraded, &r10_bio->state),
 330                                 0);
 331                 md_write_end(r10_bio->mddev);
 332                 raid_end_bio_io(r10_bio);
 333         }
 334
 335         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 336 }
 337
 338
 339 /*
 340  * RAID10 layout manager
 341  * Aswell as the chunksize and raid_disks count, there are two
 342  * parameters: near_copies and far_copies.
 343  * near_copies * far_copies must be <= raid_disks.
 344  * Normally one of these will be 1.
 345  * If both are 1, we get raid0.
 346  * If near_copies == raid_disks, we get raid1.
 347  *
 348  * Chunks are layed out in raid0 style with near_copies copies of the
 349  * first chunk, followed by near_copies copies of the next chunk and
 350  * so on.
 351  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 352  * as described above, we start again with a device offset of near_copies.
 353  * So we effectively have another copy of the whole array further down all
 354  * the drives, but with blocks on different drives.
 355  * With this layout, and block is never stored twice on the one device.
 356  *
 357  * raid10_find_phys finds the sector offset of a given virtual sector
 358  * on each device that it is on.
 359  *
 360  * raid10_find_virt does the reverse mapping, from a device and a
 361  * sector offset to a virtual address
 362  */
 363
 364 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
 365 {
 366         int n,f;
 367         sector_t sector;
 368         sector_t chunk;
 369         sector_t stripe;
 370         int dev;
 371
 372         int slot = 0;
 373
 374         /* now calculate first sector/dev */
 375         chunk = r10bio->sector >> conf->chunk_shift;
 376         sector = r10bio->sector & conf->chunk_mask;
 377
 378         chunk *= conf->near_copies;
 379         stripe = chunk;
 380         dev = sector_div(stripe, conf->raid_disks);
 381         if (conf->far_offset)
 382                 stripe *= conf->far_copies;
 383
 384         sector += stripe << conf->chunk_shift;
 385
 386         /* and calculate all the others */
 387         for (n=0; n < conf->near_copies; n++) {
 388                 int d = dev;
 389                 sector_t s = sector;
 390                 r10bio->devs[slot].addr = sector;
 391                 r10bio->devs[slot].devnum = d;
 392                 slot++;
 393
 394                 for (f = 1; f < conf->far_copies; f++) {
 395                         d += conf->near_copies;
 396                         if (d >= conf->raid_disks)
 397                                 d -= conf->raid_disks;
 398                         s += conf->stride;
 399                         r10bio->devs[slot].devnum = d;
 400                         r10bio->devs[slot].addr = s;
 401                         slot++;
 402                 }
 403                 dev++;
 404                 if (dev >= conf->raid_disks) {
 405                         dev = 0;
 406                         sector += (conf->chunk_mask + 1);
 407                 }
 408         }
 409         BUG_ON(slot != conf->copies);
 410 }
 411
 412 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
 413 {
 414         sector_t offset, chunk, vchunk;
 415
 416         offset = sector & conf->chunk_mask;
 417         if (conf->far_offset) {
 418                 int fc;
 419                 chunk = sector >> conf->chunk_shift;
 420                 fc = sector_div(chunk, conf->far_copies);
 421                 dev -= fc * conf->near_copies;
 422                 if (dev < 0)
 423                         dev += conf->raid_disks;
 424         } else {
 425                 while (sector >= conf->stride) {
 426                         sector -= conf->stride;
 427                         if (dev < conf->near_copies)
 428                                 dev += conf->raid_disks - conf->near_copies;
 429                         else
 430                                 dev -= conf->near_copies;
 431                 }
 432                 chunk = sector >> conf->chunk_shift;
 433         }
 434         vchunk = chunk * conf->raid_disks + dev;
 435         sector_div(vchunk, conf->near_copies);
 436         return (vchunk << conf->chunk_shift) + offset;
 437 }
 438
 439 /**
 440  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 441  *      @q: request queue
 442  *      @bvm: properties of new bio
 443  *      @biovec: the request that could be merged to it.
 444  *
 445  *      Return amount of bytes we can accept at this offset
 446  *      If near_copies == raid_disk, there are no striping issues,
 447  *      but in that case, the function isn't called at all.
 448  */
 449 static int raid10_mergeable_bvec(struct request_queue *q,
 450                                  struct bvec_merge_data *bvm,
 451                                  struct bio_vec *biovec)
 452 {
 453         mddev_t *mddev = q->queuedata;
 454         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 455         int max;
 456         unsigned int chunk_sectors = mddev->chunk_size >> 9;
 457         unsigned int bio_sectors = bvm->bi_size >> 9;
 458
 459         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 460         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 461         if (max <= biovec->bv_len && bio_sectors == 0)
 462                 return biovec->bv_len;
 463         else
 464                 return max;
 465 }
 466
 467 /*
 468  * This routine returns the disk from which the requested read should
 469  * be done. There is a per-array 'next expected sequential IO' sector
 470  * number - if this matches on the next IO then we use the last disk.
 471  * There is also a per-disk 'last know head position' sector that is
 472  * maintained from IRQ contexts, both the normal and the resync IO
 473  * completion handlers update this position correctly. If there is no
 474  * perfect sequential match then we pick the disk whose head is closest.
 475  *
 476  * If there are 2 mirrors in the same 2 devices, performance degrades
 477  * because position is mirror, not device based.
 478  *
 479  * The rdev for the device selected will have nr_pending incremented.
 480  */
 481
 482 /*
 483  * FIXME: possibly should rethink readbalancing and do it differently
 484  * depending on near_copies / far_copies geometry.
 485  */
 486 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
 487 {
 488         const unsigned long this_sector = r10_bio->sector;
 489         int disk, slot, nslot;
 490         const int sectors = r10_bio->sectors;
 491         sector_t new_distance, current_distance;
 492         mdk_rdev_t *rdev;
 493
 494         raid10_find_phys(conf, r10_bio);
 495         rcu_read_lock();
 496         /*
 497          * Check if we can balance. We can balance on the whole
 498          * device if no resync is going on (recovery is ok), or below
 499          * the resync window. We take the first readable disk when
 500          * above the resync window.
 501          */
 502         if (conf->mddev->recovery_cp < MaxSector
 503             && (this_sector + sectors >= conf->next_resync)) {
 504                 /* make sure that disk is operational */
 505                 slot = 0;
 506                 disk = r10_bio->devs[slot].devnum;
 507
 508                 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 509                        r10_bio->devs[slot].bio == IO_BLOCKED ||
 510                        !test_bit(In_sync, &rdev->flags)) {
 511                         slot++;
 512                         if (slot == conf->copies) {
 513                                 slot = 0;
 514                                 disk = -1;
 515                                 break;
 516                         }
 517                         disk = r10_bio->devs[slot].devnum;
 518                 }
 519                 goto rb_out;
 520         }
 521
 522
 523         /* make sure the disk is operational */
 524         slot = 0;
 525         disk = r10_bio->devs[slot].devnum;
 526         while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 527                r10_bio->devs[slot].bio == IO_BLOCKED ||
 528                !test_bit(In_sync, &rdev->flags)) {
 529                 slot ++;
 530                 if (slot == conf->copies) {
 531                         disk = -1;
 532                         goto rb_out;
 533                 }
 534                 disk = r10_bio->devs[slot].devnum;
 535         }
 536
 537
 538         current_distance = abs(r10_bio->devs[slot].addr -
 539                                conf->mirrors[disk].head_position);
 540
 541         /* Find the disk whose head is closest,
 542          * or - for far > 1 - find the closest to partition beginning */
 543
 544         for (nslot = slot; nslot < conf->copies; nslot++) {
 545                 int ndisk = r10_bio->devs[nslot].devnum;
 546
 547
 548                 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
 549                     r10_bio->devs[nslot].bio == IO_BLOCKED ||
 550                     !test_bit(In_sync, &rdev->flags))
 551                         continue;
 552
 553                 /* This optimisation is debatable, and completely destroys
 554                  * sequential read speed for 'far copies' arrays.  So only
 555                  * keep it for 'near' arrays, and review those later.
 556                  */
 557                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
 558                         disk = ndisk;
 559                         slot = nslot;
 560                         break;
 561                 }
 562
 563                 /* for far > 1 always use the lowest address */
 564                 if (conf->far_copies > 1)
 565                         new_distance = r10_bio->devs[nslot].addr;
 566                 else
 567                         new_distance = abs(r10_bio->devs[nslot].addr -
 568                                            conf->mirrors[ndisk].head_position);
 569                 if (new_distance < current_distance) {
 570                         current_distance = new_distance;
 571                         disk = ndisk;
 572                         slot = nslot;
 573                 }
 574         }
 575
 576 rb_out:
 577         r10_bio->read_slot = slot;
 578 /*      conf->next_seq_sect = this_sector + sectors;*/
 579
 580         if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
 581                 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
 582         else
 583                 disk = -1;
 584         rcu_read_unlock();
 585
 586         return disk;
 587 }
 588
 589 static void unplug_slaves(mddev_t *mddev)
 590 {
 591         conf_t *conf = mddev_to_conf(mddev);
 592         int i;
 593
 594         rcu_read_lock();
 595         for (i=0; i<mddev->raid_disks; i++) {
 596                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 597                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 598                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 599
 600                         atomic_inc(&rdev->nr_pending);
 601                         rcu_read_unlock();
 602
 603                         blk_unplug(r_queue);
 604
 605                         rdev_dec_pending(rdev, mddev);
 606                         rcu_read_lock();
 607                 }
 608         }
 609         rcu_read_unlock();
 610 }
 611
 612 static void raid10_unplug(struct request_queue *q)
 613 {
 614         mddev_t *mddev = q->queuedata;
 615
 616         unplug_slaves(q->queuedata);
 617         md_wakeup_thread(mddev->thread);
 618 }
 619
 620 static int raid10_congested(void *data, int bits)
 621 {
 622         mddev_t *mddev = data;
 623         conf_t *conf = mddev_to_conf(mddev);
 624         int i, ret = 0;
 625
 626         rcu_read_lock();
 627         for (i = 0; i < mddev->raid_disks && ret == 0; i++) {
 628                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 629                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 630                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 631
 632                         ret |= bdi_congested(&q->backing_dev_info, bits);
 633                 }
 634         }
 635         rcu_read_unlock();
 636         return ret;
 637 }
 638
 639 static int flush_pending_writes(conf_t *conf)
 640 {
 641         /* Any writes that have been queued but are awaiting
 642          * bitmap updates get flushed here.
 643          * We return 1 if any requests were actually submitted.
 644          */
 645         int rv = 0;
 646
 647         spin_lock_irq(&conf->device_lock);
 648
 649         if (conf->pending_bio_list.head) {
 650                 struct bio *bio;
 651                 bio = bio_list_get(&conf->pending_bio_list);
 652                 blk_remove_plug(conf->mddev->queue);
 653                 spin_unlock_irq(&conf->device_lock);
 654                 /* flush any pending bitmap writes to disk
 655                  * before proceeding w/ I/O */
 656                 bitmap_unplug(conf->mddev->bitmap);
 657
 658                 while (bio) { /* submit pending writes */
 659                         struct bio *next = bio->bi_next;
 660                         bio->bi_next = NULL;
 661                         generic_make_request(bio);
 662                         bio = next;
 663                 }
 664                 rv = 1;
 665         } else
 666                 spin_unlock_irq(&conf->device_lock);
 667         return rv;
 668 }
 669 /* Barriers....
 670  * Sometimes we need to suspend IO while we do something else,
 671  * either some resync/recovery, or reconfigure the array.
 672  * To do this we raise a 'barrier'.
 673  * The 'barrier' is a counter that can be raised multiple times
 674  * to count how many activities are happening which preclude
 675  * normal IO.
 676  * We can only raise the barrier if there is no pending IO.
 677  * i.e. if nr_pending == 0.
 678  * We choose only to raise the barrier if no-one is waiting for the
 679  * barrier to go down.  This means that as soon as an IO request
 680  * is ready, no other operations which require a barrier will start
 681  * until the IO request has had a chance.
 682  *
 683  * So: regular IO calls 'wait_barrier'.  When that returns there
 684  *    is no backgroup IO happening,  It must arrange to call
 685  *    allow_barrier when it has finished its IO.
 686  * backgroup IO calls must call raise_barrier.  Once that returns
 687  *    there is no normal IO happeing.  It must arrange to call
 688  *    lower_barrier when the particular background IO completes.
 689  */
 690 #define RESYNC_DEPTH 32
 691
 692 static void raise_barrier(conf_t *conf, int force)
 693 {
 694         BUG_ON(force && !conf->barrier);
 695         spin_lock_irq(&conf->resync_lock);
 696
 697         /* Wait until no block IO is waiting (unless 'force') */
 698         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 699                             conf->resync_lock,
 700                             raid10_unplug(conf->mddev->queue));
 701
 702         /* block any new IO from starting */
 703         conf->barrier++;
 704
 705         /* No wait for all pending IO to complete */
 706         wait_event_lock_irq(conf->wait_barrier,
 707                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 708                             conf->resync_lock,
 709                             raid10_unplug(conf->mddev->queue));
 710
 711         spin_unlock_irq(&conf->resync_lock);
 712 }
 713
 714 static void lower_barrier(conf_t *conf)
 715 {
 716         unsigned long flags;
 717         spin_lock_irqsave(&conf->resync_lock, flags);
 718         conf->barrier--;
 719         spin_unlock_irqrestore(&conf->resync_lock, flags);
 720         wake_up(&conf->wait_barrier);
 721 }
 722
 723 static void wait_barrier(conf_t *conf)
 724 {
 725         spin_lock_irq(&conf->resync_lock);
 726         if (conf->barrier) {
 727                 conf->nr_waiting++;
 728                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 729                                     conf->resync_lock,
 730                                     raid10_unplug(conf->mddev->queue));
 731                 conf->nr_waiting--;
 732         }
 733         conf->nr_pending++;
 734         spin_unlock_irq(&conf->resync_lock);
 735 }
 736
 737 static void allow_barrier(conf_t *conf)
 738 {
 739         unsigned long flags;
 740         spin_lock_irqsave(&conf->resync_lock, flags);
 741         conf->nr_pending--;
 742         spin_unlock_irqrestore(&conf->resync_lock, flags);
 743         wake_up(&conf->wait_barrier);
 744 }
 745
 746 static void freeze_array(conf_t *conf)
 747 {
 748         /* stop syncio and normal IO and wait for everything to
 749          * go quiet.
 750          * We increment barrier and nr_waiting, and then
 751          * wait until nr_pending match nr_queued+1
 752          * This is called in the context of one normal IO request
 753          * that has failed. Thus any sync request that might be pending
 754          * will be blocked by nr_pending, and we need to wait for
 755          * pending IO requests to complete or be queued for re-try.
 756          * Thus the number queued (nr_queued) plus this request (1)
 757          * must match the number of pending IOs (nr_pending) before
 758          * we continue.
 759          */
 760         spin_lock_irq(&conf->resync_lock);
 761         conf->barrier++;
 762         conf->nr_waiting++;
 763         wait_event_lock_irq(conf->wait_barrier,
 764                             conf->nr_pending == conf->nr_queued+1,
 765                             conf->resync_lock,
 766                             ({ flush_pending_writes(conf);
 767                                raid10_unplug(conf->mddev->queue); }));
 768         spin_unlock_irq(&conf->resync_lock);
 769 }
 770
 771 static void unfreeze_array(conf_t *conf)
 772 {
 773         /* reverse the effect of the freeze */
 774         spin_lock_irq(&conf->resync_lock);
 775         conf->barrier--;
 776         conf->nr_waiting--;
 777         wake_up(&conf->wait_barrier);
 778         spin_unlock_irq(&conf->resync_lock);
 779 }
 780
 781 static int make_request(struct request_queue *q, struct bio * bio)
 782 {
 783         mddev_t *mddev = q->queuedata;
 784         conf_t *conf = mddev_to_conf(mddev);
 785         mirror_info_t *mirror;
 786         r10bio_t *r10_bio;
 787         struct bio *read_bio;
 788         int i;
 789         int chunk_sects = conf->chunk_mask + 1;
 790         const int rw = bio_data_dir(bio);
 791         const int do_sync = bio_sync(bio);
 792         struct bio_list bl;
 793         unsigned long flags;
 794         mdk_rdev_t *blocked_rdev;
 795
 796         if (unlikely(bio_barrier(bio))) {
 797                 bio_endio(bio, -EOPNOTSUPP);
 798                 return 0;
 799         }
 800
 801         /* If this request crosses a chunk boundary, we need to
 802          * split it.  This will only happen for 1 PAGE (or less) requests.
 803          */
 804         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 805                       > chunk_sects &&
 806                     conf->near_copies < conf->raid_disks)) {
 807                 struct bio_pair *bp;
 808                 /* Sanity check -- queue functions should prevent this happening */
 809                 if (bio->bi_vcnt != 1 ||
 810                     bio->bi_idx != 0)
 811                         goto bad_map;
 812                 /* This is a one page bio that upper layers
 813                  * refuse to split for us, so we need to split it.
 814                  */
 815                 bp = bio_split(bio, bio_split_pool,
 816                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 817                 if (make_request(q, &bp->bio1))
 818                         generic_make_request(&bp->bio1);
 819                 if (make_request(q, &bp->bio2))
 820                         generic_make_request(&bp->bio2);
 821
 822                 bio_pair_release(bp);
 823                 return 0;
 824         bad_map:
 825                 printk("raid10_make_request bug: can't convert block across chunks"
 826                        " or bigger than %dk %llu %d\n", chunk_sects/2,
 827                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 828
 829                 bio_io_error(bio);
 830                 return 0;
 831         }
 832
 833         md_write_start(mddev, bio);
 834
 835         /*
 836          * Register the new request and wait if the reconstruction
 837          * thread has put up a bar for new requests.
 838          * Continue immediately if no resync is active currently.
 839          */
 840         wait_barrier(conf);
 841
 842         disk_stat_inc(mddev->gendisk, ios[rw]);
 843         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
 844
 845         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 846
 847         r10_bio->master_bio = bio;
 848         r10_bio->sectors = bio->bi_size >> 9;
 849
 850         r10_bio->mddev = mddev;
 851         r10_bio->sector = bio->bi_sector;
 852         r10_bio->state = 0;
 853
 854         if (rw == READ) {
 855                 /*
 856                  * read balancing logic:
 857                  */
 858                 int disk = read_balance(conf, r10_bio);
 859                 int slot = r10_bio->read_slot;
 860                 if (disk < 0) {
 861                         raid_end_bio_io(r10_bio);
 862                         return 0;
 863                 }
 864                 mirror = conf->mirrors + disk;
 865
 866                 read_bio = bio_clone(bio, GFP_NOIO);
 867
 868                 r10_bio->devs[slot].bio = read_bio;
 869
 870                 read_bio->bi_sector = r10_bio->devs[slot].addr +
 871                         mirror->rdev->data_offset;
 872                 read_bio->bi_bdev = mirror->rdev->bdev;
 873                 read_bio->bi_end_io = raid10_end_read_request;
 874                 read_bio->bi_rw = READ | do_sync;
 875                 read_bio->bi_private = r10_bio;
 876
 877                 generic_make_request(read_bio);
 878                 return 0;
 879         }
 880
 881         /*
 882          * WRITE:
 883          */
 884         /* first select target devices under rcu_lock and
 885          * inc refcount on their rdev.  Record them by setting
 886          * bios[x] to bio
 887          */
 888         raid10_find_phys(conf, r10_bio);
 889  retry_write:
 890         blocked_rdev = NULL;
 891         rcu_read_lock();
 892         for (i = 0;  i < conf->copies; i++) {
 893                 int d = r10_bio->devs[i].devnum;
 894                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
 895                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 896                         atomic_inc(&rdev->nr_pending);
 897                         blocked_rdev = rdev;
 898                         break;
 899                 }
 900                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 901                         atomic_inc(&rdev->nr_pending);
 902                         r10_bio->devs[i].bio = bio;
 903                 } else {
 904                         r10_bio->devs[i].bio = NULL;
 905                         set_bit(R10BIO_Degraded, &r10_bio->state);
 906                 }
 907         }
 908         rcu_read_unlock();
 909
 910         if (unlikely(blocked_rdev)) {
 911                 /* Have to wait for this device to get unblocked, then retry */
 912                 int j;
 913                 int d;
 914
 915                 for (j = 0; j < i; j++)
 916                         if (r10_bio->devs[j].bio) {
 917                                 d = r10_bio->devs[j].devnum;
 918                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
 919                         }
 920                 allow_barrier(conf);
 921                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 922                 wait_barrier(conf);
 923                 goto retry_write;
 924         }
 925
 926         atomic_set(&r10_bio->remaining, 0);
 927
 928         bio_list_init(&bl);
 929         for (i = 0; i < conf->copies; i++) {
 930                 struct bio *mbio;
 931                 int d = r10_bio->devs[i].devnum;
 932                 if (!r10_bio->devs[i].bio)
 933                         continue;
 934
 935                 mbio = bio_clone(bio, GFP_NOIO);
 936                 r10_bio->devs[i].bio = mbio;
 937
 938                 mbio->bi_sector = r10_bio->devs[i].addr+
 939                         conf->mirrors[d].rdev->data_offset;
 940                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
 941                 mbio->bi_end_io = raid10_end_write_request;
 942                 mbio->bi_rw = WRITE | do_sync;
 943                 mbio->bi_private = r10_bio;
 944
 945                 atomic_inc(&r10_bio->remaining);
 946                 bio_list_add(&bl, mbio);
 947         }
 948
 949         if (unlikely(!atomic_read(&r10_bio->remaining))) {
 950                 /* the array is dead */
 951                 md_write_end(mddev);
 952                 raid_end_bio_io(r10_bio);
 953                 return 0;
 954         }
 955
 956         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
 957         spin_lock_irqsave(&conf->device_lock, flags);
 958         bio_list_merge(&conf->pending_bio_list, &bl);
 959         blk_plug_device(mddev->queue);
 960         spin_unlock_irqrestore(&conf->device_lock, flags);
 961
 962         /* In case raid10d snuck in to freeze_array */
 963         wake_up(&conf->wait_barrier);
 964
 965         if (do_sync)
 966                 md_wakeup_thread(mddev->thread);
 967
 968         return 0;
 969 }
 970
 971 static void status(struct seq_file *seq, mddev_t *mddev)
 972 {
 973         conf_t *conf = mddev_to_conf(mddev);
 974         int i;
 975
 976         if (conf->near_copies < conf->raid_disks)
 977                 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
 978         if (conf->near_copies > 1)
 979                 seq_printf(seq, " %d near-copies", conf->near_copies);
 980         if (conf->far_copies > 1) {
 981                 if (conf->far_offset)
 982                         seq_printf(seq, " %d offset-copies", conf->far_copies);
 983                 else
 984                         seq_printf(seq, " %d far-copies", conf->far_copies);
 985         }
 986         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 987                                         conf->raid_disks - mddev->degraded);
 988         for (i = 0; i < conf->raid_disks; i++)
 989                 seq_printf(seq, "%s",
 990                               conf->mirrors[i].rdev &&
 991                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
 992         seq_printf(seq, "]");
 993 }
 994
 995 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 996 {
 997         char b[BDEVNAME_SIZE];
 998         conf_t *conf = mddev_to_conf(mddev);
 999
1000         /*
1001          * If it is not operational, then we have already marked it as dead
1002          * else if it is the last working disks, ignore the error, let the
1003          * next level up know.
1004          * else mark the drive as failed
1005          */
1006         if (test_bit(In_sync, &rdev->flags)
1007             && conf->raid_disks-mddev->degraded == 1)
1008                 /*
1009                  * Don't fail the drive, just return an IO error.
1010                  * The test should really be more sophisticated than
1011                  * "working_disks == 1", but it isn't critical, and
1012                  * can wait until we do more sophisticated "is the drive
1013                  * really dead" tests...
1014                  */
1015                 return;
1016         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1017                 unsigned long flags;
1018                 spin_lock_irqsave(&conf->device_lock, flags);
1019                 mddev->degraded++;
1020                 spin_unlock_irqrestore(&conf->device_lock, flags);
1021                 /*
1022                  * if recovery is running, make sure it aborts.
1023                  */
1024                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1025         }
1026         set_bit(Faulty, &rdev->flags);
1027         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1028         printk(KERN_ALERT "raid10: Disk failure on %s, disabling device.\n"
1029                 "raid10: Operation continuing on %d devices.\n",
1030                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1031 }
1032
1033 static void print_conf(conf_t *conf)
1034 {
1035         int i;
1036         mirror_info_t *tmp;
1037
1038         printk("RAID10 conf printout:\n");
1039         if (!conf) {
1040                 printk("(!conf)\n");
1041                 return;
1042         }
1043         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1044                 conf->raid_disks);
1045
1046         for (i = 0; i < conf->raid_disks; i++) {
1047                 char b[BDEVNAME_SIZE];
1048                 tmp = conf->mirrors + i;
1049                 if (tmp->rdev)
1050                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1051                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1052                                 !test_bit(Faulty, &tmp->rdev->flags),
1053                                 bdevname(tmp->rdev->bdev,b));
1054         }
1055 }
1056
1057 static void close_sync(conf_t *conf)
1058 {
1059         wait_barrier(conf);
1060         allow_barrier(conf);
1061
1062         mempool_destroy(conf->r10buf_pool);
1063         conf->r10buf_pool = NULL;
1064 }
1065
1066 /* check if there are enough drives for
1067  * every block to appear on atleast one
1068  */
1069 static int enough(conf_t *conf)
1070 {
1071         int first = 0;
1072
1073         do {
1074                 int n = conf->copies;
1075                 int cnt = 0;
1076                 while (n--) {
1077                         if (conf->mirrors[first].rdev)
1078                                 cnt++;
1079                         first = (first+1) % conf->raid_disks;
1080                 }
1081                 if (cnt == 0)
1082                         return 0;
1083         } while (first != 0);
1084         return 1;
1085 }
1086
1087 static int raid10_spare_active(mddev_t *mddev)
1088 {
1089         int i;
1090         conf_t *conf = mddev->private;
1091         mirror_info_t *tmp;
1092
1093         /*
1094          * Find all non-in_sync disks within the RAID10 configuration
1095          * and mark them in_sync
1096          */
1097         for (i = 0; i < conf->raid_disks; i++) {
1098                 tmp = conf->mirrors + i;
1099                 if (tmp->rdev
1100                     && !test_bit(Faulty, &tmp->rdev->flags)
1101                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1102                         unsigned long flags;
1103                         spin_lock_irqsave(&conf->device_lock, flags);
1104                         mddev->degraded--;
1105                         spin_unlock_irqrestore(&conf->device_lock, flags);
1106                 }
1107         }
1108
1109         print_conf(conf);
1110         return 0;
1111 }
1112
1113
1114 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1115 {
1116         conf_t *conf = mddev->private;
1117         int err = -EEXIST;
1118         int mirror;
1119         mirror_info_t *p;
1120         int first = 0;
1121         int last = mddev->raid_disks - 1;
1122
1123         if (mddev->recovery_cp < MaxSector)
1124                 /* only hot-add to in-sync arrays, as recovery is
1125                  * very different from resync
1126                  */
1127                 return -EBUSY;
1128         if (!enough(conf))
1129                 return -EINVAL;
1130
1131         if (rdev->raid_disk)
1132                 first = last = rdev->raid_disk;
1133
1134         if (rdev->saved_raid_disk >= 0 &&
1135             rdev->saved_raid_disk >= first &&
1136             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1137                 mirror = rdev->saved_raid_disk;
1138         else
1139                 mirror = first;
1140         for ( ; mirror <= last ; mirror++)
1141                 if ( !(p=conf->mirrors+mirror)->rdev) {
1142
1143                         blk_queue_stack_limits(mddev->queue,
1144                                                rdev->bdev->bd_disk->queue);
1145                         /* as we don't honour merge_bvec_fn, we must never risk
1146                          * violating it, so limit ->max_sector to one PAGE, as
1147                          * a one page request is never in violation.
1148                          */
1149                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1150                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1151                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1152
1153                         p->head_position = 0;
1154                         rdev->raid_disk = mirror;
1155                         err = 0;
1156                         if (rdev->saved_raid_disk != mirror)
1157                                 conf->fullsync = 1;
1158                         rcu_assign_pointer(p->rdev, rdev);
1159                         break;
1160                 }
1161
1162         print_conf(conf);
1163         return err;
1164 }
1165
1166 static int raid10_remove_disk(mddev_t *mddev, int number)
1167 {
1168         conf_t *conf = mddev->private;
1169         int err = 0;
1170         mdk_rdev_t *rdev;
1171         mirror_info_t *p = conf->mirrors+ number;
1172
1173         print_conf(conf);
1174         rdev = p->rdev;
1175         if (rdev) {
1176                 if (test_bit(In_sync, &rdev->flags) ||
1177                     atomic_read(&rdev->nr_pending)) {
1178                         err = -EBUSY;
1179                         goto abort;
1180                 }
1181                 /* Only remove faulty devices in recovery
1182                  * is not possible.
1183                  */
1184                 if (!test_bit(Faulty, &rdev->flags) &&
1185                     enough(conf)) {
1186                         err = -EBUSY;
1187                         goto abort;
1188                 }
1189                 p->rdev = NULL;
1190                 synchronize_rcu();
1191                 if (atomic_read(&rdev->nr_pending)) {
1192                         /* lost the race, try later */
1193                         err = -EBUSY;
1194                         p->rdev = rdev;
1195                 }
1196         }
1197 abort:
1198
1199         print_conf(conf);
1200         return err;
1201 }
1202
1203
1204 static void end_sync_read(struct bio *bio, int error)
1205 {
1206         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1207         conf_t *conf = mddev_to_conf(r10_bio->mddev);
1208         int i,d;
1209
1210         for (i=0; i<conf->copies; i++)
1211                 if (r10_bio->devs[i].bio == bio)
1212                         break;
1213         BUG_ON(i == conf->copies);
1214         update_head_pos(i, r10_bio);
1215         d = r10_bio->devs[i].devnum;
1216
1217         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1218                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1219         else {
1220                 atomic_add(r10_bio->sectors,
1221                            &conf->mirrors[d].rdev->corrected_errors);
1222                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1223                         md_error(r10_bio->mddev,
1224                                  conf->mirrors[d].rdev);
1225         }
1226
1227         /* for reconstruct, we always reschedule after a read.
1228          * for resync, only after all reads
1229          */
1230         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1231             atomic_dec_and_test(&r10_bio->remaining)) {
1232                 /* we have read all the blocks,
1233                  * do the comparison in process context in raid10d
1234                  */
1235                 reschedule_retry(r10_bio);
1236         }
1237         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1238 }
1239
1240 static void end_sync_write(struct bio *bio, int error)
1241 {
1242         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1243         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1244         mddev_t *mddev = r10_bio->mddev;
1245         conf_t *conf = mddev_to_conf(mddev);
1246         int i,d;
1247
1248         for (i = 0; i < conf->copies; i++)
1249                 if (r10_bio->devs[i].bio == bio)
1250                         break;
1251         d = r10_bio->devs[i].devnum;
1252
1253         if (!uptodate)
1254                 md_error(mddev, conf->mirrors[d].rdev);
1255
1256         update_head_pos(i, r10_bio);
1257
1258         while (atomic_dec_and_test(&r10_bio->remaining)) {
1259                 if (r10_bio->master_bio == NULL) {
1260                         /* the primary of several recovery bios */
1261                         md_done_sync(mddev, r10_bio->sectors, 1);
1262                         put_buf(r10_bio);
1263                         break;
1264                 } else {
1265                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1266                         put_buf(r10_bio);
1267                         r10_bio = r10_bio2;
1268                 }
1269         }
1270         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1271 }
1272
1273 /*
1274  * Note: sync and recover and handled very differently for raid10
1275  * This code is for resync.
1276  * For resync, we read through virtual addresses and read all blocks.
1277  * If there is any error, we schedule a write.  The lowest numbered
1278  * drive is authoritative.
1279  * However requests come for physical address, so we need to map.
1280  * For every physical address there are raid_disks/copies virtual addresses,
1281  * which is always are least one, but is not necessarly an integer.
1282  * This means that a physical address can span multiple chunks, so we may
1283  * have to submit multiple io requests for a single sync request.
1284  */
1285 /*
1286  * We check if all blocks are in-sync and only write to blocks that
1287  * aren't in sync
1288  */
1289 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1290 {
1291         conf_t *conf = mddev_to_conf(mddev);
1292         int i, first;
1293         struct bio *tbio, *fbio;
1294
1295         atomic_set(&r10_bio->remaining, 1);
1296
1297         /* find the first device with a block */
1298         for (i=0; i<conf->copies; i++)
1299                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1300                         break;
1301
1302         if (i == conf->copies)
1303                 goto done;
1304
1305         first = i;
1306         fbio = r10_bio->devs[i].bio;
1307
1308         /* now find blocks with errors */
1309         for (i=0 ; i < conf->copies ; i++) {
1310                 int  j, d;
1311                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1312
1313                 tbio = r10_bio->devs[i].bio;
1314
1315                 if (tbio->bi_end_io != end_sync_read)
1316                         continue;
1317                 if (i == first)
1318                         continue;
1319                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1320                         /* We know that the bi_io_vec layout is the same for
1321                          * both 'first' and 'i', so we just compare them.
1322                          * All vec entries are PAGE_SIZE;
1323                          */
1324                         for (j = 0; j < vcnt; j++)
1325                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1326                                            page_address(tbio->bi_io_vec[j].bv_page),
1327                                            PAGE_SIZE))
1328                                         break;
1329                         if (j == vcnt)
1330                                 continue;
1331                         mddev->resync_mismatches += r10_bio->sectors;
1332                 }
1333                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1334                         /* Don't fix anything. */
1335                         continue;
1336                 /* Ok, we need to write this bio
1337                  * First we need to fixup bv_offset, bv_len and
1338                  * bi_vecs, as the read request might have corrupted these
1339                  */
1340                 tbio->bi_vcnt = vcnt;
1341                 tbio->bi_size = r10_bio->sectors << 9;
1342                 tbio->bi_idx = 0;
1343                 tbio->bi_phys_segments = 0;
1344                 tbio->bi_hw_segments = 0;
1345                 tbio->bi_hw_front_size = 0;
1346                 tbio->bi_hw_back_size = 0;
1347                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1348                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1349                 tbio->bi_next = NULL;
1350                 tbio->bi_rw = WRITE;
1351                 tbio->bi_private = r10_bio;
1352                 tbio->bi_sector = r10_bio->devs[i].addr;
1353
1354                 for (j=0; j < vcnt ; j++) {
1355                         tbio->bi_io_vec[j].bv_offset = 0;
1356                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1357
1358                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1359                                page_address(fbio->bi_io_vec[j].bv_page),
1360                                PAGE_SIZE);
1361                 }
1362                 tbio->bi_end_io = end_sync_write;
1363
1364                 d = r10_bio->devs[i].devnum;
1365                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1366                 atomic_inc(&r10_bio->remaining);
1367                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1368
1369                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1370                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1371                 generic_make_request(tbio);
1372         }
1373
1374 done:
1375         if (atomic_dec_and_test(&r10_bio->remaining)) {
1376                 md_done_sync(mddev, r10_bio->sectors, 1);
1377                 put_buf(r10_bio);
1378         }
1379 }
1380
1381 /*
1382  * Now for the recovery code.
1383  * Recovery happens across physical sectors.
1384  * We recover all non-is_sync drives by finding the virtual address of
1385  * each, and then choose a working drive that also has that virt address.
1386  * There is a separate r10_bio for each non-in_sync drive.
1387  * Only the first two slots are in use. The first for reading,
1388  * The second for writing.
1389  *
1390  */
1391
1392 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1393 {
1394         conf_t *conf = mddev_to_conf(mddev);
1395         int i, d;
1396         struct bio *bio, *wbio;
1397
1398
1399         /* move the pages across to the second bio
1400          * and submit the write request
1401          */
1402         bio = r10_bio->devs[0].bio;
1403         wbio = r10_bio->devs[1].bio;
1404         for (i=0; i < wbio->bi_vcnt; i++) {
1405                 struct page *p = bio->bi_io_vec[i].bv_page;
1406                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1407                 wbio->bi_io_vec[i].bv_page = p;
1408         }
1409         d = r10_bio->devs[1].devnum;
1410
1411         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1412         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1413         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1414                 generic_make_request(wbio);
1415         else
1416                 bio_endio(wbio, -EIO);
1417 }
1418
1419
1420 /*
1421  * This is a kernel thread which:
1422  *
1423  *      1.      Retries failed read operations on working mirrors.
1424  *      2.      Updates the raid superblock when problems encounter.
1425  *      3.      Performs writes following reads for array synchronising.
1426  */
1427
1428 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1429 {
1430         int sect = 0; /* Offset from r10_bio->sector */
1431         int sectors = r10_bio->sectors;
1432         mdk_rdev_t*rdev;
1433         while(sectors) {
1434                 int s = sectors;
1435                 int sl = r10_bio->read_slot;
1436                 int success = 0;
1437                 int start;
1438
1439                 if (s > (PAGE_SIZE>>9))
1440                         s = PAGE_SIZE >> 9;
1441
1442                 rcu_read_lock();
1443                 do {
1444                         int d = r10_bio->devs[sl].devnum;
1445                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1446                         if (rdev &&
1447                             test_bit(In_sync, &rdev->flags)) {
1448                                 atomic_inc(&rdev->nr_pending);
1449                                 rcu_read_unlock();
1450                                 success = sync_page_io(rdev->bdev,
1451                                                        r10_bio->devs[sl].addr +
1452                                                        sect + rdev->data_offset,
1453                                                        s<<9,
1454                                                        conf->tmppage, READ);
1455                                 rdev_dec_pending(rdev, mddev);
1456                                 rcu_read_lock();
1457                                 if (success)
1458                                         break;
1459                         }
1460                         sl++;
1461                         if (sl == conf->copies)
1462                                 sl = 0;
1463                 } while (!success && sl != r10_bio->read_slot);
1464                 rcu_read_unlock();
1465
1466                 if (!success) {
1467                         /* Cannot read from anywhere -- bye bye array */
1468                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1469                         md_error(mddev, conf->mirrors[dn].rdev);
1470                         break;
1471                 }
1472
1473                 start = sl;
1474                 /* write it back and re-read */
1475                 rcu_read_lock();
1476                 while (sl != r10_bio->read_slot) {
1477                         int d;
1478                         if (sl==0)
1479                                 sl = conf->copies;
1480                         sl--;
1481                         d = r10_bio->devs[sl].devnum;
1482                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1483                         if (rdev &&
1484                             test_bit(In_sync, &rdev->flags)) {
1485                                 atomic_inc(&rdev->nr_pending);
1486                                 rcu_read_unlock();
1487                                 atomic_add(s, &rdev->corrected_errors);
1488                                 if (sync_page_io(rdev->bdev,
1489                                                  r10_bio->devs[sl].addr +
1490                                                  sect + rdev->data_offset,
1491                                                  s<<9, conf->tmppage, WRITE)
1492                                     == 0)
1493                                         /* Well, this device is dead */
1494                                         md_error(mddev, rdev);
1495                                 rdev_dec_pending(rdev, mddev);
1496                                 rcu_read_lock();
1497                         }
1498                 }
1499                 sl = start;
1500                 while (sl != r10_bio->read_slot) {
1501                         int d;
1502                         if (sl==0)
1503                                 sl = conf->copies;
1504                         sl--;
1505                         d = r10_bio->devs[sl].devnum;
1506                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1507                         if (rdev &&
1508                             test_bit(In_sync, &rdev->flags)) {
1509                                 char b[BDEVNAME_SIZE];
1510                                 atomic_inc(&rdev->nr_pending);
1511                                 rcu_read_unlock();
1512                                 if (sync_page_io(rdev->bdev,
1513                                                  r10_bio->devs[sl].addr +
1514                                                  sect + rdev->data_offset,
1515                                                  s<<9, conf->tmppage, READ) == 0)
1516                                         /* Well, this device is dead */
1517                                         md_error(mddev, rdev);
1518                                 else
1519                                         printk(KERN_INFO
1520                                                "raid10:%s: read error corrected"
1521                                                " (%d sectors at %llu on %s)\n",
1522                                                mdname(mddev), s,
1523                                                (unsigned long long)(sect+
1524                                                     rdev->data_offset),
1525                                                bdevname(rdev->bdev, b));
1526
1527                                 rdev_dec_pending(rdev, mddev);
1528                                 rcu_read_lock();
1529                         }
1530                 }
1531                 rcu_read_unlock();
1532
1533                 sectors -= s;
1534                 sect += s;
1535         }
1536 }
1537
1538 static void raid10d(mddev_t *mddev)
1539 {
1540         r10bio_t *r10_bio;
1541         struct bio *bio;
1542         unsigned long flags;
1543         conf_t *conf = mddev_to_conf(mddev);
1544         struct list_head *head = &conf->retry_list;
1545         int unplug=0;
1546         mdk_rdev_t *rdev;
1547
1548         md_check_recovery(mddev);
1549
1550         for (;;) {
1551                 char b[BDEVNAME_SIZE];
1552
1553                 unplug += flush_pending_writes(conf);
1554
1555                 spin_lock_irqsave(&conf->device_lock, flags);
1556                 if (list_empty(head)) {
1557                         spin_unlock_irqrestore(&conf->device_lock, flags);
1558                         break;
1559                 }
1560                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1561                 list_del(head->prev);
1562                 conf->nr_queued--;
1563                 spin_unlock_irqrestore(&conf->device_lock, flags);
1564
1565                 mddev = r10_bio->mddev;
1566                 conf = mddev_to_conf(mddev);
1567                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1568                         sync_request_write(mddev, r10_bio);
1569                         unplug = 1;
1570                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1571                         recovery_request_write(mddev, r10_bio);
1572                         unplug = 1;
1573                 } else {
1574                         int mirror;
1575                         /* we got a read error. Maybe the drive is bad.  Maybe just
1576                          * the block and we can fix it.
1577                          * We freeze all other IO, and try reading the block from
1578                          * other devices.  When we find one, we re-write
1579                          * and check it that fixes the read error.
1580                          * This is all done synchronously while the array is
1581                          * frozen.
1582                          */
1583                         if (mddev->ro == 0) {
1584                                 freeze_array(conf);
1585                                 fix_read_error(conf, mddev, r10_bio);
1586                                 unfreeze_array(conf);
1587                         }
1588
1589                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1590                         r10_bio->devs[r10_bio->read_slot].bio =
1591                                 mddev->ro ? IO_BLOCKED : NULL;
1592                         mirror = read_balance(conf, r10_bio);
1593                         if (mirror == -1) {
1594                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1595                                        " read error for block %llu\n",
1596                                        bdevname(bio->bi_bdev,b),
1597                                        (unsigned long long)r10_bio->sector);
1598                                 raid_end_bio_io(r10_bio);
1599                                 bio_put(bio);
1600                         } else {
1601                                 const int do_sync = bio_sync(r10_bio->master_bio);
1602                                 bio_put(bio);
1603                                 rdev = conf->mirrors[mirror].rdev;
1604                                 if (printk_ratelimit())
1605                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1606                                                " another mirror\n",
1607                                                bdevname(rdev->bdev,b),
1608                                                (unsigned long long)r10_bio->sector);
1609                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1610                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1611                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1612                                         + rdev->data_offset;
1613                                 bio->bi_bdev = rdev->bdev;
1614                                 bio->bi_rw = READ | do_sync;
1615                                 bio->bi_private = r10_bio;
1616                                 bio->bi_end_io = raid10_end_read_request;
1617                                 unplug = 1;
1618                                 generic_make_request(bio);
1619                         }
1620                 }
1621         }
1622         if (unplug)
1623                 unplug_slaves(mddev);
1624 }
1625
1626
1627 static int init_resync(conf_t *conf)
1628 {
1629         int buffs;
1630
1631         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1632         BUG_ON(conf->r10buf_pool);
1633         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1634         if (!conf->r10buf_pool)
1635                 return -ENOMEM;
1636         conf->next_resync = 0;
1637         return 0;
1638 }
1639
1640 /*
1641  * perform a "sync" on one "block"
1642  *
1643  * We need to make sure that no normal I/O request - particularly write
1644  * requests - conflict with active sync requests.
1645  *
1646  * This is achieved by tracking pending requests and a 'barrier' concept
1647  * that can be installed to exclude normal IO requests.
1648  *
1649  * Resync and recovery are handled very differently.
1650  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1651  *
1652  * For resync, we iterate over virtual addresses, read all copies,
1653  * and update if there are differences.  If only one copy is live,
1654  * skip it.
1655  * For recovery, we iterate over physical addresses, read a good
1656  * value for each non-in_sync drive, and over-write.
1657  *
1658  * So, for recovery we may have several outstanding complex requests for a
1659  * given address, one for each out-of-sync device.  We model this by allocating
1660  * a number of r10_bio structures, one for each out-of-sync device.
1661  * As we setup these structures, we collect all bio's together into a list
1662  * which we then process collectively to add pages, and then process again
1663  * to pass to generic_make_request.
1664  *
1665  * The r10_bio structures are linked using a borrowed master_bio pointer.
1666  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1667  * has its remaining count decremented to 0, the whole complex operation
1668  * is complete.
1669  *
1670  */
1671
1672 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1673 {
1674         conf_t *conf = mddev_to_conf(mddev);
1675         r10bio_t *r10_bio;
1676         struct bio *biolist = NULL, *bio;
1677         sector_t max_sector, nr_sectors;
1678         int disk;
1679         int i;
1680         int max_sync;
1681         int sync_blocks;
1682
1683         sector_t sectors_skipped = 0;
1684         int chunks_skipped = 0;
1685
1686         if (!conf->r10buf_pool)
1687                 if (init_resync(conf))
1688                         return 0;
1689
1690  skipped:
1691         max_sector = mddev->size << 1;
1692         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1693                 max_sector = mddev->resync_max_sectors;
1694         if (sector_nr >= max_sector) {
1695                 /* If we aborted, we need to abort the
1696                  * sync on the 'current' bitmap chucks (there can
1697                  * be several when recovering multiple devices).
1698                  * as we may have started syncing it but not finished.
1699                  * We can find the current address in
1700                  * mddev->curr_resync, but for recovery,
1701                  * we need to convert that to several
1702                  * virtual addresses.
1703                  */
1704                 if (mddev->curr_resync < max_sector) { /* aborted */
1705                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1706                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1707                                                 &sync_blocks, 1);
1708                         else for (i=0; i<conf->raid_disks; i++) {
1709                                 sector_t sect =
1710                                         raid10_find_virt(conf, mddev->curr_resync, i);
1711                                 bitmap_end_sync(mddev->bitmap, sect,
1712                                                 &sync_blocks, 1);
1713                         }
1714                 } else /* completed sync */
1715                         conf->fullsync = 0;
1716
1717                 bitmap_close_sync(mddev->bitmap);
1718                 close_sync(conf);
1719                 *skipped = 1;
1720                 return sectors_skipped;
1721         }
1722         if (chunks_skipped >= conf->raid_disks) {
1723                 /* if there has been nothing to do on any drive,
1724                  * then there is nothing to do at all..
1725                  */
1726                 *skipped = 1;
1727                 return (max_sector - sector_nr) + sectors_skipped;
1728         }
1729
1730         if (max_sector > mddev->resync_max)
1731                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1732
1733         /* make sure whole request will fit in a chunk - if chunks
1734          * are meaningful
1735          */
1736         if (conf->near_copies < conf->raid_disks &&
1737             max_sector > (sector_nr | conf->chunk_mask))
1738                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1739         /*
1740          * If there is non-resync activity waiting for us then
1741          * put in a delay to throttle resync.
1742          */
1743         if (!go_faster && conf->nr_waiting)
1744                 msleep_interruptible(1000);
1745
1746         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1747
1748         /* Again, very different code for resync and recovery.
1749          * Both must result in an r10bio with a list of bios that
1750          * have bi_end_io, bi_sector, bi_bdev set,
1751          * and bi_private set to the r10bio.
1752          * For recovery, we may actually create several r10bios
1753          * with 2 bios in each, that correspond to the bios in the main one.
1754          * In this case, the subordinate r10bios link back through a
1755          * borrowed master_bio pointer, and the counter in the master
1756          * includes a ref from each subordinate.
1757          */
1758         /* First, we decide what to do and set ->bi_end_io
1759          * To end_sync_read if we want to read, and
1760          * end_sync_write if we will want to write.
1761          */
1762
1763         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1764         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1765                 /* recovery... the complicated one */
1766                 int i, j, k;
1767                 r10_bio = NULL;
1768
1769                 for (i=0 ; i<conf->raid_disks; i++)
1770                         if (conf->mirrors[i].rdev &&
1771                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1772                                 int still_degraded = 0;
1773                                 /* want to reconstruct this device */
1774                                 r10bio_t *rb2 = r10_bio;
1775                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1776                                 int must_sync;
1777                                 /* Unless we are doing a full sync, we only need
1778                                  * to recover the block if it is set in the bitmap
1779                                  */
1780                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1781                                                               &sync_blocks, 1);
1782                                 if (sync_blocks < max_sync)
1783                                         max_sync = sync_blocks;
1784                                 if (!must_sync &&
1785                                     !conf->fullsync) {
1786                                         /* yep, skip the sync_blocks here, but don't assume
1787                                          * that there will never be anything to do here
1788                                          */
1789                                         chunks_skipped = -1;
1790                                         continue;
1791                                 }
1792
1793                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1794                                 raise_barrier(conf, rb2 != NULL);
1795                                 atomic_set(&r10_bio->remaining, 0);
1796
1797                                 r10_bio->master_bio = (struct bio*)rb2;
1798                                 if (rb2)
1799                                         atomic_inc(&rb2->remaining);
1800                                 r10_bio->mddev = mddev;
1801                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1802                                 r10_bio->sector = sect;
1803
1804                                 raid10_find_phys(conf, r10_bio);
1805                                 /* Need to check if this section will still be
1806                                  * degraded
1807                                  */
1808                                 for (j=0; j<conf->copies;j++) {
1809                                         int d = r10_bio->devs[j].devnum;
1810                                         if (conf->mirrors[d].rdev == NULL ||
1811                                             test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
1812                                                 still_degraded = 1;
1813                                                 break;
1814                                         }
1815                                 }
1816                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1817                                                               &sync_blocks, still_degraded);
1818
1819                                 for (j=0; j<conf->copies;j++) {
1820                                         int d = r10_bio->devs[j].devnum;
1821                                         if (conf->mirrors[d].rdev &&
1822                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1823                                                 /* This is where we read from */
1824                                                 bio = r10_bio->devs[0].bio;
1825                                                 bio->bi_next = biolist;
1826                                                 biolist = bio;
1827                                                 bio->bi_private = r10_bio;
1828                                                 bio->bi_end_io = end_sync_read;
1829                                                 bio->bi_rw = READ;
1830                                                 bio->bi_sector = r10_bio->devs[j].addr +
1831                                                         conf->mirrors[d].rdev->data_offset;
1832                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1833                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1834                                                 atomic_inc(&r10_bio->remaining);
1835                                                 /* and we write to 'i' */
1836
1837                                                 for (k=0; k<conf->copies; k++)
1838                                                         if (r10_bio->devs[k].devnum == i)
1839                                                                 break;
1840                                                 BUG_ON(k == conf->copies);
1841                                                 bio = r10_bio->devs[1].bio;
1842                                                 bio->bi_next = biolist;
1843                                                 biolist = bio;
1844                                                 bio->bi_private = r10_bio;
1845                                                 bio->bi_end_io = end_sync_write;
1846                                                 bio->bi_rw = WRITE;
1847                                                 bio->bi_sector = r10_bio->devs[k].addr +
1848                                                         conf->mirrors[i].rdev->data_offset;
1849                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1850
1851                                                 r10_bio->devs[0].devnum = d;
1852                                                 r10_bio->devs[1].devnum = i;
1853
1854                                                 break;
1855                                         }
1856                                 }
1857                                 if (j == conf->copies) {
1858                                         /* Cannot recover, so abort the recovery */
1859                                         put_buf(r10_bio);
1860                                         if (rb2)
1861                                                 atomic_dec(&rb2->remaining);
1862                                         r10_bio = rb2;
1863                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
1864                                                               &mddev->recovery))
1865                                                 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1866                                                        mdname(mddev));
1867                                         break;
1868                                 }
1869                         }
1870                 if (biolist == NULL) {
1871                         while (r10_bio) {
1872                                 r10bio_t *rb2 = r10_bio;
1873                                 r10_bio = (r10bio_t*) rb2->master_bio;
1874                                 rb2->master_bio = NULL;
1875                                 put_buf(rb2);
1876                         }
1877                         goto giveup;
1878                 }
1879         } else {
1880                 /* resync. Schedule a read for every block at this virt offset */
1881                 int count = 0;
1882
1883                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1884                                        &sync_blocks, mddev->degraded) &&
1885                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1886                         /* We can skip this block */
1887                         *skipped = 1;
1888                         return sync_blocks + sectors_skipped;
1889                 }
1890                 if (sync_blocks < max_sync)
1891                         max_sync = sync_blocks;
1892                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1893
1894                 r10_bio->mddev = mddev;
1895                 atomic_set(&r10_bio->remaining, 0);
1896                 raise_barrier(conf, 0);
1897                 conf->next_resync = sector_nr;
1898
1899                 r10_bio->master_bio = NULL;
1900                 r10_bio->sector = sector_nr;
1901                 set_bit(R10BIO_IsSync, &r10_bio->state);
1902                 raid10_find_phys(conf, r10_bio);
1903                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1904
1905                 for (i=0; i<conf->copies; i++) {
1906                         int d = r10_bio->devs[i].devnum;
1907                         bio = r10_bio->devs[i].bio;
1908                         bio->bi_end_io = NULL;
1909                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
1910                         if (conf->mirrors[d].rdev == NULL ||
1911                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1912                                 continue;
1913                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1914                         atomic_inc(&r10_bio->remaining);
1915                         bio->bi_next = biolist;
1916                         biolist = bio;
1917                         bio->bi_private = r10_bio;
1918                         bio->bi_end_io = end_sync_read;
1919                         bio->bi_rw = READ;
1920                         bio->bi_sector = r10_bio->devs[i].addr +
1921                                 conf->mirrors[d].rdev->data_offset;
1922                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1923                         count++;
1924                 }
1925
1926                 if (count < 2) {
1927                         for (i=0; i<conf->copies; i++) {
1928                                 int d = r10_bio->devs[i].devnum;
1929                                 if (r10_bio->devs[i].bio->bi_end_io)
1930                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1931                         }
1932                         put_buf(r10_bio);
1933                         biolist = NULL;
1934                         goto giveup;
1935                 }
1936         }
1937
1938         for (bio = biolist; bio ; bio=bio->bi_next) {
1939
1940                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1941                 if (bio->bi_end_io)
1942                         bio->bi_flags |= 1 << BIO_UPTODATE;
1943                 bio->bi_vcnt = 0;
1944                 bio->bi_idx = 0;
1945                 bio->bi_phys_segments = 0;
1946                 bio->bi_hw_segments = 0;
1947                 bio->bi_size = 0;
1948         }
1949
1950         nr_sectors = 0;
1951         if (sector_nr + max_sync < max_sector)
1952                 max_sector = sector_nr + max_sync;
1953         do {
1954                 struct page *page;
1955                 int len = PAGE_SIZE;
1956                 disk = 0;
1957                 if (sector_nr + (len>>9) > max_sector)
1958                         len = (max_sector - sector_nr) << 9;
1959                 if (len == 0)
1960                         break;
1961                 for (bio= biolist ; bio ; bio=bio->bi_next) {
1962                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1963                         if (bio_add_page(bio, page, len, 0) == 0) {
1964                                 /* stop here */
1965                                 struct bio *bio2;
1966                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1967                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1968                                         /* remove last page from this bio */
1969                                         bio2->bi_vcnt--;
1970                                         bio2->bi_size -= len;
1971                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1972                                 }
1973                                 goto bio_full;
1974                         }
1975                         disk = i;
1976                 }
1977                 nr_sectors += len>>9;
1978                 sector_nr += len>>9;
1979         } while (biolist->bi_vcnt < RESYNC_PAGES);
1980  bio_full:
1981         r10_bio->sectors = nr_sectors;
1982
1983         while (biolist) {
1984                 bio = biolist;
1985                 biolist = biolist->bi_next;
1986
1987                 bio->bi_next = NULL;
1988                 r10_bio = bio->bi_private;
1989                 r10_bio->sectors = nr_sectors;
1990
1991                 if (bio->bi_end_io == end_sync_read) {
1992                         md_sync_acct(bio->bi_bdev, nr_sectors);
1993                         generic_make_request(bio);
1994                 }
1995         }
1996
1997         if (sectors_skipped)
1998                 /* pretend they weren't skipped, it makes
1999                  * no important difference in this case
2000                  */
2001                 md_done_sync(mddev, sectors_skipped, 1);
2002
2003         return sectors_skipped + nr_sectors;
2004  giveup:
2005         /* There is nowhere to write, so all non-sync
2006          * drives must be failed, so try the next chunk...
2007          */
2008         {
2009         sector_t sec = max_sector - sector_nr;
2010         sectors_skipped += sec;
2011         chunks_skipped ++;
2012         sector_nr = max_sector;
2013         goto skipped;
2014         }
2015 }
2016
2017 static int run(mddev_t *mddev)
2018 {
2019         conf_t *conf;
2020         int i, disk_idx;
2021         mirror_info_t *disk;
2022         mdk_rdev_t *rdev;
2023         struct list_head *tmp;
2024         int nc, fc, fo;
2025         sector_t stride, size;
2026
2027         if (mddev->chunk_size == 0) {
2028                 printk(KERN_ERR "md/raid10: non-zero chunk size required.\n");
2029                 return -EINVAL;
2030         }
2031
2032         nc = mddev->layout & 255;
2033         fc = (mddev->layout >> 8) & 255;
2034         fo = mddev->layout & (1<<16);
2035         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2036             (mddev->layout >> 17)) {
2037                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
2038                        mdname(mddev), mddev->layout);
2039                 goto out;
2040         }
2041         /*
2042          * copy the already verified devices into our private RAID10
2043          * bookkeeping area. [whatever we allocate in run(),
2044          * should be freed in stop()]
2045          */
2046         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2047         mddev->private = conf;
2048         if (!conf) {
2049                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2050                         mdname(mddev));
2051                 goto out;
2052         }
2053         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2054                                  GFP_KERNEL);
2055         if (!conf->mirrors) {
2056                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2057                        mdname(mddev));
2058                 goto out_free_conf;
2059         }
2060
2061         conf->tmppage = alloc_page(GFP_KERNEL);
2062         if (!conf->tmppage)
2063                 goto out_free_conf;
2064
2065         conf->mddev = mddev;
2066         conf->raid_disks = mddev->raid_disks;
2067         conf->near_copies = nc;
2068         conf->far_copies = fc;
2069         conf->copies = nc*fc;
2070         conf->far_offset = fo;
2071         conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
2072         conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
2073         size = mddev->size >> (conf->chunk_shift-1);
2074         sector_div(size, fc);
2075         size = size * conf->raid_disks;
2076         sector_div(size, nc);
2077         /* 'size' is now the number of chunks in the array */
2078         /* calculate "used chunks per device" in 'stride' */
2079         stride = size * conf->copies;
2080
2081         /* We need to round up when dividing by raid_disks to
2082          * get the stride size.
2083          */
2084         stride += conf->raid_disks - 1;
2085         sector_div(stride, conf->raid_disks);
2086         mddev->size = stride  << (conf->chunk_shift-1);
2087
2088         if (fo)
2089                 stride = 1;
2090         else
2091                 sector_div(stride, fc);
2092         conf->stride = stride << conf->chunk_shift;
2093
2094         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2095                                                 r10bio_pool_free, conf);
2096         if (!conf->r10bio_pool) {
2097                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2098                         mdname(mddev));
2099                 goto out_free_conf;
2100         }
2101
2102         spin_lock_init(&conf->device_lock);
2103         mddev->queue->queue_lock = &conf->device_lock;
2104
2105         rdev_for_each(rdev, tmp, mddev) {
2106                 disk_idx = rdev->raid_disk;
2107                 if (disk_idx >= mddev->raid_disks
2108                     || disk_idx < 0)
2109                         continue;
2110                 disk = conf->mirrors + disk_idx;
2111
2112                 disk->rdev = rdev;
2113
2114                 blk_queue_stack_limits(mddev->queue,
2115                                        rdev->bdev->bd_disk->queue);
2116                 /* as we don't honour merge_bvec_fn, we must never risk
2117                  * violating it, so limit ->max_sector to one PAGE, as
2118                  * a one page request is never in violation.
2119                  */
2120                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2121                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
2122                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
2123
2124                 disk->head_position = 0;
2125         }
2126         INIT_LIST_HEAD(&conf->retry_list);
2127
2128         spin_lock_init(&conf->resync_lock);
2129         init_waitqueue_head(&conf->wait_barrier);
2130
2131         /* need to check that every block has at least one working mirror */
2132         if (!enough(conf)) {
2133                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2134                        mdname(mddev));
2135                 goto out_free_conf;
2136         }
2137
2138         mddev->degraded = 0;
2139         for (i = 0; i < conf->raid_disks; i++) {
2140
2141                 disk = conf->mirrors + i;
2142
2143                 if (!disk->rdev ||
2144                     !test_bit(In_sync, &disk->rdev->flags)) {
2145                         disk->head_position = 0;
2146                         mddev->degraded++;
2147                         if (disk->rdev)
2148                                 conf->fullsync = 1;
2149                 }
2150         }
2151
2152
2153         mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2154         if (!mddev->thread) {
2155                 printk(KERN_ERR
2156                        "raid10: couldn't allocate thread for %s\n",
2157                        mdname(mddev));
2158                 goto out_free_conf;
2159         }
2160
2161         printk(KERN_INFO
2162                 "raid10: raid set %s active with %d out of %d devices\n",
2163                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2164                 mddev->raid_disks);
2165         /*
2166          * Ok, everything is just fine now
2167          */
2168         mddev->array_sectors = size << conf->chunk_shift;
2169         mddev->resync_max_sectors = size << conf->chunk_shift;
2170
2171         mddev->queue->unplug_fn = raid10_unplug;
2172         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2173         mddev->queue->backing_dev_info.congested_data = mddev;
2174
2175         /* Calculate max read-ahead size.
2176          * We need to readahead at least twice a whole stripe....
2177          * maybe...
2178          */
2179         {
2180                 int stripe = conf->raid_disks * (mddev->chunk_size / PAGE_SIZE);
2181                 stripe /= conf->near_copies;
2182                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2183                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2184         }
2185
2186         if (conf->near_copies < mddev->raid_disks)
2187                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2188         return 0;
2189
2190 out_free_conf:
2191         if (conf->r10bio_pool)
2192                 mempool_destroy(conf->r10bio_pool);
2193         safe_put_page(conf->tmppage);
2194         kfree(conf->mirrors);
2195         kfree(conf);
2196         mddev->private = NULL;
2197 out:
2198         return -EIO;
2199 }
2200
2201 static int stop(mddev_t *mddev)
2202 {
2203         conf_t *conf = mddev_to_conf(mddev);
2204
2205         md_unregister_thread(mddev->thread);
2206         mddev->thread = NULL;
2207         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2208         if (conf->r10bio_pool)
2209                 mempool_destroy(conf->r10bio_pool);
2210         kfree(conf->mirrors);
2211         kfree(conf);
2212         mddev->private = NULL;
2213         return 0;
2214 }
2215
2216 static void raid10_quiesce(mddev_t *mddev, int state)
2217 {
2218         conf_t *conf = mddev_to_conf(mddev);
2219
2220         switch(state) {
2221         case 1:
2222                 raise_barrier(conf, 0);
2223                 break;
2224         case 0:
2225                 lower_barrier(conf);
2226                 break;
2227         }
2228         if (mddev->thread) {
2229                 if (mddev->bitmap)
2230                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2231                 else
2232                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2233                 md_wakeup_thread(mddev->thread);
2234         }
2235 }
2236
2237 static struct mdk_personality raid10_personality =
2238 {
2239         .name           = "raid10",
2240         .level          = 10,
2241         .owner          = THIS_MODULE,
2242         .make_request   = make_request,
2243         .run            = run,
2244         .stop           = stop,
2245         .status         = status,
2246         .error_handler  = error,
2247         .hot_add_disk   = raid10_add_disk,
2248         .hot_remove_disk= raid10_remove_disk,
2249         .spare_active   = raid10_spare_active,
2250         .sync_request   = sync_request,
2251         .quiesce        = raid10_quiesce,
2252 };
2253
2254 static int __init raid_init(void)
2255 {
2256         return register_md_personality(&raid10_personality);
2257 }
2258
2259 static void raid_exit(void)
2260 {
2261         unregister_md_personality(&raid10_personality);
2262 }
2263
2264 module_init(raid_init);
2265 module_exit(raid_exit);
2266 MODULE_LICENSE("GPL");
2267 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2268 MODULE_ALIAS("md-raid10");
2269 MODULE_ALIAS("md-level-10");