err.no Git - linux-2.6/blob - 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  *      @bio: the buffer head that's been built up so far
 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, struct bio *bio,
 450                                 struct bio_vec *bio_vec)
 451 {
 452         mddev_t *mddev = q->queuedata;
 453         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
 454         int max;
 455         unsigned int chunk_sectors = mddev->chunk_size >> 9;
 456         unsigned int bio_sectors = bio->bi_size >> 9;
 457
 458         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 459         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 460         if (max <= bio_vec->bv_len && bio_sectors == 0)
 461                 return bio_vec->bv_len;
 462         else
 463                 return max;
 464 }
 465
 466 /*
 467  * This routine returns the disk from which the requested read should
 468  * be done. There is a per-array 'next expected sequential IO' sector
 469  * number - if this matches on the next IO then we use the last disk.
 470  * There is also a per-disk 'last know head position' sector that is
 471  * maintained from IRQ contexts, both the normal and the resync IO
 472  * completion handlers update this position correctly. If there is no
 473  * perfect sequential match then we pick the disk whose head is closest.
 474  *
 475  * If there are 2 mirrors in the same 2 devices, performance degrades
 476  * because position is mirror, not device based.
 477  *
 478  * The rdev for the device selected will have nr_pending incremented.
 479  */
 480
 481 /*
 482  * FIXME: possibly should rethink readbalancing and do it differently
 483  * depending on near_copies / far_copies geometry.
 484  */
 485 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
 486 {
 487         const unsigned long this_sector = r10_bio->sector;
 488         int disk, slot, nslot;
 489         const int sectors = r10_bio->sectors;
 490         sector_t new_distance, current_distance;
 491         mdk_rdev_t *rdev;
 492
 493         raid10_find_phys(conf, r10_bio);
 494         rcu_read_lock();
 495         /*
 496          * Check if we can balance. We can balance on the whole
 497          * device if no resync is going on (recovery is ok), or below
 498          * the resync window. We take the first readable disk when
 499          * above the resync window.
 500          */
 501         if (conf->mddev->recovery_cp < MaxSector
 502             && (this_sector + sectors >= conf->next_resync)) {
 503                 /* make sure that disk is operational */
 504                 slot = 0;
 505                 disk = r10_bio->devs[slot].devnum;
 506
 507                 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 508                        r10_bio->devs[slot].bio == IO_BLOCKED ||
 509                        !test_bit(In_sync, &rdev->flags)) {
 510                         slot++;
 511                         if (slot == conf->copies) {
 512                                 slot = 0;
 513                                 disk = -1;
 514                                 break;
 515                         }
 516                         disk = r10_bio->devs[slot].devnum;
 517                 }
 518                 goto rb_out;
 519         }
 520
 521
 522         /* make sure the disk is operational */
 523         slot = 0;
 524         disk = r10_bio->devs[slot].devnum;
 525         while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 526                r10_bio->devs[slot].bio == IO_BLOCKED ||
 527                !test_bit(In_sync, &rdev->flags)) {
 528                 slot ++;
 529                 if (slot == conf->copies) {
 530                         disk = -1;
 531                         goto rb_out;
 532                 }
 533                 disk = r10_bio->devs[slot].devnum;
 534         }
 535
 536
 537         current_distance = abs(r10_bio->devs[slot].addr -
 538                                conf->mirrors[disk].head_position);
 539
 540         /* Find the disk whose head is closest */
 541
 542         for (nslot = slot; nslot < conf->copies; nslot++) {
 543                 int ndisk = r10_bio->devs[nslot].devnum;
 544
 545
 546                 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
 547                     r10_bio->devs[nslot].bio == IO_BLOCKED ||
 548                     !test_bit(In_sync, &rdev->flags))
 549                         continue;
 550
 551                 /* This optimisation is debatable, and completely destroys
 552                  * sequential read speed for 'far copies' arrays.  So only
 553                  * keep it for 'near' arrays, and review those later.
 554                  */
 555                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
 556                         disk = ndisk;
 557                         slot = nslot;
 558                         break;
 559                 }
 560                 new_distance = abs(r10_bio->devs[nslot].addr -
 561                                    conf->mirrors[ndisk].head_position);
 562                 if (new_distance < current_distance) {
 563                         current_distance = new_distance;
 564                         disk = ndisk;
 565                         slot = nslot;
 566                 }
 567         }
 568
 569 rb_out:
 570         r10_bio->read_slot = slot;
 571 /*      conf->next_seq_sect = this_sector + sectors;*/
 572
 573         if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
 574                 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
 575         else
 576                 disk = -1;
 577         rcu_read_unlock();
 578
 579         return disk;
 580 }
 581
 582 static void unplug_slaves(mddev_t *mddev)
 583 {
 584         conf_t *conf = mddev_to_conf(mddev);
 585         int i;
 586
 587         rcu_read_lock();
 588         for (i=0; i<mddev->raid_disks; i++) {
 589                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 590                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 591                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 592
 593                         atomic_inc(&rdev->nr_pending);
 594                         rcu_read_unlock();
 595
 596                         blk_unplug(r_queue);
 597
 598                         rdev_dec_pending(rdev, mddev);
 599                         rcu_read_lock();
 600                 }
 601         }
 602         rcu_read_unlock();
 603 }
 604
 605 static void raid10_unplug(struct request_queue *q)
 606 {
 607         mddev_t *mddev = q->queuedata;
 608
 609         unplug_slaves(q->queuedata);
 610         md_wakeup_thread(mddev->thread);
 611 }
 612
 613 static int raid10_congested(void *data, int bits)
 614 {
 615         mddev_t *mddev = data;
 616         conf_t *conf = mddev_to_conf(mddev);
 617         int i, ret = 0;
 618
 619         rcu_read_lock();
 620         for (i = 0; i < mddev->raid_disks && ret == 0; i++) {
 621                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 622                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 623                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 624
 625                         ret |= bdi_congested(&q->backing_dev_info, bits);
 626                 }
 627         }
 628         rcu_read_unlock();
 629         return ret;
 630 }
 631
 632
 633 /* Barriers....
 634  * Sometimes we need to suspend IO while we do something else,
 635  * either some resync/recovery, or reconfigure the array.
 636  * To do this we raise a 'barrier'.
 637  * The 'barrier' is a counter that can be raised multiple times
 638  * to count how many activities are happening which preclude
 639  * normal IO.
 640  * We can only raise the barrier if there is no pending IO.
 641  * i.e. if nr_pending == 0.
 642  * We choose only to raise the barrier if no-one is waiting for the
 643  * barrier to go down.  This means that as soon as an IO request
 644  * is ready, no other operations which require a barrier will start
 645  * until the IO request has had a chance.
 646  *
 647  * So: regular IO calls 'wait_barrier'.  When that returns there
 648  *    is no backgroup IO happening,  It must arrange to call
 649  *    allow_barrier when it has finished its IO.
 650  * backgroup IO calls must call raise_barrier.  Once that returns
 651  *    there is no normal IO happeing.  It must arrange to call
 652  *    lower_barrier when the particular background IO completes.
 653  */
 654 #define RESYNC_DEPTH 32
 655
 656 static void raise_barrier(conf_t *conf, int force)
 657 {
 658         BUG_ON(force && !conf->barrier);
 659         spin_lock_irq(&conf->resync_lock);
 660
 661         /* Wait until no block IO is waiting (unless 'force') */
 662         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 663                             conf->resync_lock,
 664                             raid10_unplug(conf->mddev->queue));
 665
 666         /* block any new IO from starting */
 667         conf->barrier++;
 668
 669         /* No wait for all pending IO to complete */
 670         wait_event_lock_irq(conf->wait_barrier,
 671                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 672                             conf->resync_lock,
 673                             raid10_unplug(conf->mddev->queue));
 674
 675         spin_unlock_irq(&conf->resync_lock);
 676 }
 677
 678 static void lower_barrier(conf_t *conf)
 679 {
 680         unsigned long flags;
 681         spin_lock_irqsave(&conf->resync_lock, flags);
 682         conf->barrier--;
 683         spin_unlock_irqrestore(&conf->resync_lock, flags);
 684         wake_up(&conf->wait_barrier);
 685 }
 686
 687 static void wait_barrier(conf_t *conf)
 688 {
 689         spin_lock_irq(&conf->resync_lock);
 690         if (conf->barrier) {
 691                 conf->nr_waiting++;
 692                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 693                                     conf->resync_lock,
 694                                     raid10_unplug(conf->mddev->queue));
 695                 conf->nr_waiting--;
 696         }
 697         conf->nr_pending++;
 698         spin_unlock_irq(&conf->resync_lock);
 699 }
 700
 701 static void allow_barrier(conf_t *conf)
 702 {
 703         unsigned long flags;
 704         spin_lock_irqsave(&conf->resync_lock, flags);
 705         conf->nr_pending--;
 706         spin_unlock_irqrestore(&conf->resync_lock, flags);
 707         wake_up(&conf->wait_barrier);
 708 }
 709
 710 static void freeze_array(conf_t *conf)
 711 {
 712         /* stop syncio and normal IO and wait for everything to
 713          * go quiet.
 714          * We increment barrier and nr_waiting, and then
 715          * wait until barrier+nr_pending match nr_queued+2
 716          */
 717         spin_lock_irq(&conf->resync_lock);
 718         conf->barrier++;
 719         conf->nr_waiting++;
 720         wait_event_lock_irq(conf->wait_barrier,
 721                             conf->barrier+conf->nr_pending == conf->nr_queued+2,
 722                             conf->resync_lock,
 723                             raid10_unplug(conf->mddev->queue));
 724         spin_unlock_irq(&conf->resync_lock);
 725 }
 726
 727 static void unfreeze_array(conf_t *conf)
 728 {
 729         /* reverse the effect of the freeze */
 730         spin_lock_irq(&conf->resync_lock);
 731         conf->barrier--;
 732         conf->nr_waiting--;
 733         wake_up(&conf->wait_barrier);
 734         spin_unlock_irq(&conf->resync_lock);
 735 }
 736
 737 static int make_request(struct request_queue *q, struct bio * bio)
 738 {
 739         mddev_t *mddev = q->queuedata;
 740         conf_t *conf = mddev_to_conf(mddev);
 741         mirror_info_t *mirror;
 742         r10bio_t *r10_bio;
 743         struct bio *read_bio;
 744         int i;
 745         int chunk_sects = conf->chunk_mask + 1;
 746         const int rw = bio_data_dir(bio);
 747         const int do_sync = bio_sync(bio);
 748         struct bio_list bl;
 749         unsigned long flags;
 750
 751         if (unlikely(bio_barrier(bio))) {
 752                 bio_endio(bio, -EOPNOTSUPP);
 753                 return 0;
 754         }
 755
 756         /* If this request crosses a chunk boundary, we need to
 757          * split it.  This will only happen for 1 PAGE (or less) requests.
 758          */
 759         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 760                       > chunk_sects &&
 761                     conf->near_copies < conf->raid_disks)) {
 762                 struct bio_pair *bp;
 763                 /* Sanity check -- queue functions should prevent this happening */
 764                 if (bio->bi_vcnt != 1 ||
 765                     bio->bi_idx != 0)
 766                         goto bad_map;
 767                 /* This is a one page bio that upper layers
 768                  * refuse to split for us, so we need to split it.
 769                  */
 770                 bp = bio_split(bio, bio_split_pool,
 771                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 772                 if (make_request(q, &bp->bio1))
 773                         generic_make_request(&bp->bio1);
 774                 if (make_request(q, &bp->bio2))
 775                         generic_make_request(&bp->bio2);
 776
 777                 bio_pair_release(bp);
 778                 return 0;
 779         bad_map:
 780                 printk("raid10_make_request bug: can't convert block across chunks"
 781                        " or bigger than %dk %llu %d\n", chunk_sects/2,
 782                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 783
 784                 bio_io_error(bio);
 785                 return 0;
 786         }
 787
 788         md_write_start(mddev, bio);
 789
 790         /*
 791          * Register the new request and wait if the reconstruction
 792          * thread has put up a bar for new requests.
 793          * Continue immediately if no resync is active currently.
 794          */
 795         wait_barrier(conf);
 796
 797         disk_stat_inc(mddev->gendisk, ios[rw]);
 798         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
 799
 800         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 801
 802         r10_bio->master_bio = bio;
 803         r10_bio->sectors = bio->bi_size >> 9;
 804
 805         r10_bio->mddev = mddev;
 806         r10_bio->sector = bio->bi_sector;
 807         r10_bio->state = 0;
 808
 809         if (rw == READ) {
 810                 /*
 811                  * read balancing logic:
 812                  */
 813                 int disk = read_balance(conf, r10_bio);
 814                 int slot = r10_bio->read_slot;
 815                 if (disk < 0) {
 816                         raid_end_bio_io(r10_bio);
 817                         return 0;
 818                 }
 819                 mirror = conf->mirrors + disk;
 820
 821                 read_bio = bio_clone(bio, GFP_NOIO);
 822
 823                 r10_bio->devs[slot].bio = read_bio;
 824
 825                 read_bio->bi_sector = r10_bio->devs[slot].addr +
 826                         mirror->rdev->data_offset;
 827                 read_bio->bi_bdev = mirror->rdev->bdev;
 828                 read_bio->bi_end_io = raid10_end_read_request;
 829                 read_bio->bi_rw = READ | do_sync;
 830                 read_bio->bi_private = r10_bio;
 831
 832                 generic_make_request(read_bio);
 833                 return 0;
 834         }
 835
 836         /*
 837          * WRITE:
 838          */
 839         /* first select target devices under spinlock and
 840          * inc refcount on their rdev.  Record them by setting
 841          * bios[x] to bio
 842          */
 843         raid10_find_phys(conf, r10_bio);
 844         rcu_read_lock();
 845         for (i = 0;  i < conf->copies; i++) {
 846                 int d = r10_bio->devs[i].devnum;
 847                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
 848                 if (rdev &&
 849                     !test_bit(Faulty, &rdev->flags)) {
 850                         atomic_inc(&rdev->nr_pending);
 851                         r10_bio->devs[i].bio = bio;
 852                 } else {
 853                         r10_bio->devs[i].bio = NULL;
 854                         set_bit(R10BIO_Degraded, &r10_bio->state);
 855                 }
 856         }
 857         rcu_read_unlock();
 858
 859         atomic_set(&r10_bio->remaining, 0);
 860
 861         bio_list_init(&bl);
 862         for (i = 0; i < conf->copies; i++) {
 863                 struct bio *mbio;
 864                 int d = r10_bio->devs[i].devnum;
 865                 if (!r10_bio->devs[i].bio)
 866                         continue;
 867
 868                 mbio = bio_clone(bio, GFP_NOIO);
 869                 r10_bio->devs[i].bio = mbio;
 870
 871                 mbio->bi_sector = r10_bio->devs[i].addr+
 872                         conf->mirrors[d].rdev->data_offset;
 873                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
 874                 mbio->bi_end_io = raid10_end_write_request;
 875                 mbio->bi_rw = WRITE | do_sync;
 876                 mbio->bi_private = r10_bio;
 877
 878                 atomic_inc(&r10_bio->remaining);
 879                 bio_list_add(&bl, mbio);
 880         }
 881
 882         if (unlikely(!atomic_read(&r10_bio->remaining))) {
 883                 /* the array is dead */
 884                 md_write_end(mddev);
 885                 raid_end_bio_io(r10_bio);
 886                 return 0;
 887         }
 888
 889         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
 890         spin_lock_irqsave(&conf->device_lock, flags);
 891         bio_list_merge(&conf->pending_bio_list, &bl);
 892         blk_plug_device(mddev->queue);
 893         spin_unlock_irqrestore(&conf->device_lock, flags);
 894
 895         if (do_sync)
 896                 md_wakeup_thread(mddev->thread);
 897
 898         return 0;
 899 }
 900
 901 static void status(struct seq_file *seq, mddev_t *mddev)
 902 {
 903         conf_t *conf = mddev_to_conf(mddev);
 904         int i;
 905
 906         if (conf->near_copies < conf->raid_disks)
 907                 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
 908         if (conf->near_copies > 1)
 909                 seq_printf(seq, " %d near-copies", conf->near_copies);
 910         if (conf->far_copies > 1) {
 911                 if (conf->far_offset)
 912                         seq_printf(seq, " %d offset-copies", conf->far_copies);
 913                 else
 914                         seq_printf(seq, " %d far-copies", conf->far_copies);
 915         }
 916         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 917                                         conf->raid_disks - mddev->degraded);
 918         for (i = 0; i < conf->raid_disks; i++)
 919                 seq_printf(seq, "%s",
 920                               conf->mirrors[i].rdev &&
 921                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
 922         seq_printf(seq, "]");
 923 }
 924
 925 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 926 {
 927         char b[BDEVNAME_SIZE];
 928         conf_t *conf = mddev_to_conf(mddev);
 929
 930         /*
 931          * If it is not operational, then we have already marked it as dead
 932          * else if it is the last working disks, ignore the error, let the
 933          * next level up know.
 934          * else mark the drive as failed
 935          */
 936         if (test_bit(In_sync, &rdev->flags)
 937             && conf->raid_disks-mddev->degraded == 1)
 938                 /*
 939                  * Don't fail the drive, just return an IO error.
 940                  * The test should really be more sophisticated than
 941                  * "working_disks == 1", but it isn't critical, and
 942                  * can wait until we do more sophisticated "is the drive
 943                  * really dead" tests...
 944                  */
 945                 return;
 946         if (test_and_clear_bit(In_sync, &rdev->flags)) {
 947                 unsigned long flags;
 948                 spin_lock_irqsave(&conf->device_lock, flags);
 949                 mddev->degraded++;
 950                 spin_unlock_irqrestore(&conf->device_lock, flags);
 951                 /*
 952                  * if recovery is running, make sure it aborts.
 953                  */
 954                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 955         }
 956         set_bit(Faulty, &rdev->flags);
 957         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 958         printk(KERN_ALERT "raid10: Disk failure on %s, disabling device. \n"
 959                 "       Operation continuing on %d devices\n",
 960                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
 961 }
 962
 963 static void print_conf(conf_t *conf)
 964 {
 965         int i;
 966         mirror_info_t *tmp;
 967
 968         printk("RAID10 conf printout:\n");
 969         if (!conf) {
 970                 printk("(!conf)\n");
 971                 return;
 972         }
 973         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
 974                 conf->raid_disks);
 975
 976         for (i = 0; i < conf->raid_disks; i++) {
 977                 char b[BDEVNAME_SIZE];
 978                 tmp = conf->mirrors + i;
 979                 if (tmp->rdev)
 980                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
 981                                 i, !test_bit(In_sync, &tmp->rdev->flags),
 982                                 !test_bit(Faulty, &tmp->rdev->flags),
 983                                 bdevname(tmp->rdev->bdev,b));
 984         }
 985 }
 986
 987 static void close_sync(conf_t *conf)
 988 {
 989         wait_barrier(conf);
 990         allow_barrier(conf);
 991
 992         mempool_destroy(conf->r10buf_pool);
 993         conf->r10buf_pool = NULL;
 994 }
 995
 996 /* check if there are enough drives for
 997  * every block to appear on atleast one
 998  */
 999 static int enough(conf_t *conf)
1000 {
1001         int first = 0;
1002
1003         do {
1004                 int n = conf->copies;
1005                 int cnt = 0;
1006                 while (n--) {
1007                         if (conf->mirrors[first].rdev)
1008                                 cnt++;
1009                         first = (first+1) % conf->raid_disks;
1010                 }
1011                 if (cnt == 0)
1012                         return 0;
1013         } while (first != 0);
1014         return 1;
1015 }
1016
1017 static int raid10_spare_active(mddev_t *mddev)
1018 {
1019         int i;
1020         conf_t *conf = mddev->private;
1021         mirror_info_t *tmp;
1022
1023         /*
1024          * Find all non-in_sync disks within the RAID10 configuration
1025          * and mark them in_sync
1026          */
1027         for (i = 0; i < conf->raid_disks; i++) {
1028                 tmp = conf->mirrors + i;
1029                 if (tmp->rdev
1030                     && !test_bit(Faulty, &tmp->rdev->flags)
1031                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1032                         unsigned long flags;
1033                         spin_lock_irqsave(&conf->device_lock, flags);
1034                         mddev->degraded--;
1035                         spin_unlock_irqrestore(&conf->device_lock, flags);
1036                 }
1037         }
1038
1039         print_conf(conf);
1040         return 0;
1041 }
1042
1043
1044 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1045 {
1046         conf_t *conf = mddev->private;
1047         int found = 0;
1048         int mirror;
1049         mirror_info_t *p;
1050
1051         if (mddev->recovery_cp < MaxSector)
1052                 /* only hot-add to in-sync arrays, as recovery is
1053                  * very different from resync
1054                  */
1055                 return 0;
1056         if (!enough(conf))
1057                 return 0;
1058
1059         if (rdev->saved_raid_disk >= 0 &&
1060             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1061                 mirror = rdev->saved_raid_disk;
1062         else
1063                 mirror = 0;
1064         for ( ; mirror < mddev->raid_disks; mirror++)
1065                 if ( !(p=conf->mirrors+mirror)->rdev) {
1066
1067                         blk_queue_stack_limits(mddev->queue,
1068                                                rdev->bdev->bd_disk->queue);
1069                         /* as we don't honour merge_bvec_fn, we must never risk
1070                          * violating it, so limit ->max_sector to one PAGE, as
1071                          * a one page request is never in violation.
1072                          */
1073                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1074                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1075                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1076
1077                         p->head_position = 0;
1078                         rdev->raid_disk = mirror;
1079                         found = 1;
1080                         if (rdev->saved_raid_disk != mirror)
1081                                 conf->fullsync = 1;
1082                         rcu_assign_pointer(p->rdev, rdev);
1083                         break;
1084                 }
1085
1086         print_conf(conf);
1087         return found;
1088 }
1089
1090 static int raid10_remove_disk(mddev_t *mddev, int number)
1091 {
1092         conf_t *conf = mddev->private;
1093         int err = 0;
1094         mdk_rdev_t *rdev;
1095         mirror_info_t *p = conf->mirrors+ number;
1096
1097         print_conf(conf);
1098         rdev = p->rdev;
1099         if (rdev) {
1100                 if (test_bit(In_sync, &rdev->flags) ||
1101                     atomic_read(&rdev->nr_pending)) {
1102                         err = -EBUSY;
1103                         goto abort;
1104                 }
1105                 p->rdev = NULL;
1106                 synchronize_rcu();
1107                 if (atomic_read(&rdev->nr_pending)) {
1108                         /* lost the race, try later */
1109                         err = -EBUSY;
1110                         p->rdev = rdev;
1111                 }
1112         }
1113 abort:
1114
1115         print_conf(conf);
1116         return err;
1117 }
1118
1119
1120 static void end_sync_read(struct bio *bio, int error)
1121 {
1122         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1123         conf_t *conf = mddev_to_conf(r10_bio->mddev);
1124         int i,d;
1125
1126         for (i=0; i<conf->copies; i++)
1127                 if (r10_bio->devs[i].bio == bio)
1128                         break;
1129         BUG_ON(i == conf->copies);
1130         update_head_pos(i, r10_bio);
1131         d = r10_bio->devs[i].devnum;
1132
1133         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1134                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1135         else {
1136                 atomic_add(r10_bio->sectors,
1137                            &conf->mirrors[d].rdev->corrected_errors);
1138                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1139                         md_error(r10_bio->mddev,
1140                                  conf->mirrors[d].rdev);
1141         }
1142
1143         /* for reconstruct, we always reschedule after a read.
1144          * for resync, only after all reads
1145          */
1146         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1147             atomic_dec_and_test(&r10_bio->remaining)) {
1148                 /* we have read all the blocks,
1149                  * do the comparison in process context in raid10d
1150                  */
1151                 reschedule_retry(r10_bio);
1152         }
1153         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1154 }
1155
1156 static void end_sync_write(struct bio *bio, int error)
1157 {
1158         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1159         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1160         mddev_t *mddev = r10_bio->mddev;
1161         conf_t *conf = mddev_to_conf(mddev);
1162         int i,d;
1163
1164         for (i = 0; i < conf->copies; i++)
1165                 if (r10_bio->devs[i].bio == bio)
1166                         break;
1167         d = r10_bio->devs[i].devnum;
1168
1169         if (!uptodate)
1170                 md_error(mddev, conf->mirrors[d].rdev);
1171         update_head_pos(i, r10_bio);
1172
1173         while (atomic_dec_and_test(&r10_bio->remaining)) {
1174                 if (r10_bio->master_bio == NULL) {
1175                         /* the primary of several recovery bios */
1176                         md_done_sync(mddev, r10_bio->sectors, 1);
1177                         put_buf(r10_bio);
1178                         break;
1179                 } else {
1180                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1181                         put_buf(r10_bio);
1182                         r10_bio = r10_bio2;
1183                 }
1184         }
1185         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1186 }
1187
1188 /*
1189  * Note: sync and recover and handled very differently for raid10
1190  * This code is for resync.
1191  * For resync, we read through virtual addresses and read all blocks.
1192  * If there is any error, we schedule a write.  The lowest numbered
1193  * drive is authoritative.
1194  * However requests come for physical address, so we need to map.
1195  * For every physical address there are raid_disks/copies virtual addresses,
1196  * which is always are least one, but is not necessarly an integer.
1197  * This means that a physical address can span multiple chunks, so we may
1198  * have to submit multiple io requests for a single sync request.
1199  */
1200 /*
1201  * We check if all blocks are in-sync and only write to blocks that
1202  * aren't in sync
1203  */
1204 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1205 {
1206         conf_t *conf = mddev_to_conf(mddev);
1207         int i, first;
1208         struct bio *tbio, *fbio;
1209
1210         atomic_set(&r10_bio->remaining, 1);
1211
1212         /* find the first device with a block */
1213         for (i=0; i<conf->copies; i++)
1214                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1215                         break;
1216
1217         if (i == conf->copies)
1218                 goto done;
1219
1220         first = i;
1221         fbio = r10_bio->devs[i].bio;
1222
1223         /* now find blocks with errors */
1224         for (i=0 ; i < conf->copies ; i++) {
1225                 int  j, d;
1226                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1227
1228                 tbio = r10_bio->devs[i].bio;
1229
1230                 if (tbio->bi_end_io != end_sync_read)
1231                         continue;
1232                 if (i == first)
1233                         continue;
1234                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1235                         /* We know that the bi_io_vec layout is the same for
1236                          * both 'first' and 'i', so we just compare them.
1237                          * All vec entries are PAGE_SIZE;
1238                          */
1239                         for (j = 0; j < vcnt; j++)
1240                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1241                                            page_address(tbio->bi_io_vec[j].bv_page),
1242                                            PAGE_SIZE))
1243                                         break;
1244                         if (j == vcnt)
1245                                 continue;
1246                         mddev->resync_mismatches += r10_bio->sectors;
1247                 }
1248                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1249                         /* Don't fix anything. */
1250                         continue;
1251                 /* Ok, we need to write this bio
1252                  * First we need to fixup bv_offset, bv_len and
1253                  * bi_vecs, as the read request might have corrupted these
1254                  */
1255                 tbio->bi_vcnt = vcnt;
1256                 tbio->bi_size = r10_bio->sectors << 9;
1257                 tbio->bi_idx = 0;
1258                 tbio->bi_phys_segments = 0;
1259                 tbio->bi_hw_segments = 0;
1260                 tbio->bi_hw_front_size = 0;
1261                 tbio->bi_hw_back_size = 0;
1262                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1263                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1264                 tbio->bi_next = NULL;
1265                 tbio->bi_rw = WRITE;
1266                 tbio->bi_private = r10_bio;
1267                 tbio->bi_sector = r10_bio->devs[i].addr;
1268
1269                 for (j=0; j < vcnt ; j++) {
1270                         tbio->bi_io_vec[j].bv_offset = 0;
1271                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1272
1273                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1274                                page_address(fbio->bi_io_vec[j].bv_page),
1275                                PAGE_SIZE);
1276                 }
1277                 tbio->bi_end_io = end_sync_write;
1278
1279                 d = r10_bio->devs[i].devnum;
1280                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1281                 atomic_inc(&r10_bio->remaining);
1282                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1283
1284                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1285                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1286                 generic_make_request(tbio);
1287         }
1288
1289 done:
1290         if (atomic_dec_and_test(&r10_bio->remaining)) {
1291                 md_done_sync(mddev, r10_bio->sectors, 1);
1292                 put_buf(r10_bio);
1293         }
1294 }
1295
1296 /*
1297  * Now for the recovery code.
1298  * Recovery happens across physical sectors.
1299  * We recover all non-is_sync drives by finding the virtual address of
1300  * each, and then choose a working drive that also has that virt address.
1301  * There is a separate r10_bio for each non-in_sync drive.
1302  * Only the first two slots are in use. The first for reading,
1303  * The second for writing.
1304  *
1305  */
1306
1307 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1308 {
1309         conf_t *conf = mddev_to_conf(mddev);
1310         int i, d;
1311         struct bio *bio, *wbio;
1312
1313
1314         /* move the pages across to the second bio
1315          * and submit the write request
1316          */
1317         bio = r10_bio->devs[0].bio;
1318         wbio = r10_bio->devs[1].bio;
1319         for (i=0; i < wbio->bi_vcnt; i++) {
1320                 struct page *p = bio->bi_io_vec[i].bv_page;
1321                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1322                 wbio->bi_io_vec[i].bv_page = p;
1323         }
1324         d = r10_bio->devs[1].devnum;
1325
1326         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1327         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1328         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1329                 generic_make_request(wbio);
1330         else
1331                 bio_endio(wbio, -EIO);
1332 }
1333
1334
1335 /*
1336  * This is a kernel thread which:
1337  *
1338  *      1.      Retries failed read operations on working mirrors.
1339  *      2.      Updates the raid superblock when problems encounter.
1340  *      3.      Performs writes following reads for array synchronising.
1341  */
1342
1343 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1344 {
1345         int sect = 0; /* Offset from r10_bio->sector */
1346         int sectors = r10_bio->sectors;
1347         mdk_rdev_t*rdev;
1348         while(sectors) {
1349                 int s = sectors;
1350                 int sl = r10_bio->read_slot;
1351                 int success = 0;
1352                 int start;
1353
1354                 if (s > (PAGE_SIZE>>9))
1355                         s = PAGE_SIZE >> 9;
1356
1357                 rcu_read_lock();
1358                 do {
1359                         int d = r10_bio->devs[sl].devnum;
1360                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1361                         if (rdev &&
1362                             test_bit(In_sync, &rdev->flags)) {
1363                                 atomic_inc(&rdev->nr_pending);
1364                                 rcu_read_unlock();
1365                                 success = sync_page_io(rdev->bdev,
1366                                                        r10_bio->devs[sl].addr +
1367                                                        sect + rdev->data_offset,
1368                                                        s<<9,
1369                                                        conf->tmppage, READ);
1370                                 rdev_dec_pending(rdev, mddev);
1371                                 rcu_read_lock();
1372                                 if (success)
1373                                         break;
1374                         }
1375                         sl++;
1376                         if (sl == conf->copies)
1377                                 sl = 0;
1378                 } while (!success && sl != r10_bio->read_slot);
1379                 rcu_read_unlock();
1380
1381                 if (!success) {
1382                         /* Cannot read from anywhere -- bye bye array */
1383                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1384                         md_error(mddev, conf->mirrors[dn].rdev);
1385                         break;
1386                 }
1387
1388                 start = sl;
1389                 /* write it back and re-read */
1390                 rcu_read_lock();
1391                 while (sl != r10_bio->read_slot) {
1392                         int d;
1393                         if (sl==0)
1394                                 sl = conf->copies;
1395                         sl--;
1396                         d = r10_bio->devs[sl].devnum;
1397                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1398                         if (rdev &&
1399                             test_bit(In_sync, &rdev->flags)) {
1400                                 atomic_inc(&rdev->nr_pending);
1401                                 rcu_read_unlock();
1402                                 atomic_add(s, &rdev->corrected_errors);
1403                                 if (sync_page_io(rdev->bdev,
1404                                                  r10_bio->devs[sl].addr +
1405                                                  sect + rdev->data_offset,
1406                                                  s<<9, conf->tmppage, WRITE)
1407                                     == 0)
1408                                         /* Well, this device is dead */
1409                                         md_error(mddev, rdev);
1410                                 rdev_dec_pending(rdev, mddev);
1411                                 rcu_read_lock();
1412                         }
1413                 }
1414                 sl = start;
1415                 while (sl != r10_bio->read_slot) {
1416                         int d;
1417                         if (sl==0)
1418                                 sl = conf->copies;
1419                         sl--;
1420                         d = r10_bio->devs[sl].devnum;
1421                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1422                         if (rdev &&
1423                             test_bit(In_sync, &rdev->flags)) {
1424                                 char b[BDEVNAME_SIZE];
1425                                 atomic_inc(&rdev->nr_pending);
1426                                 rcu_read_unlock();
1427                                 if (sync_page_io(rdev->bdev,
1428                                                  r10_bio->devs[sl].addr +
1429                                                  sect + rdev->data_offset,
1430                                                  s<<9, conf->tmppage, READ) == 0)
1431                                         /* Well, this device is dead */
1432                                         md_error(mddev, rdev);
1433                                 else
1434                                         printk(KERN_INFO
1435                                                "raid10:%s: read error corrected"
1436                                                " (%d sectors at %llu on %s)\n",
1437                                                mdname(mddev), s,
1438                                                (unsigned long long)(sect+
1439                                                     rdev->data_offset),
1440                                                bdevname(rdev->bdev, b));
1441
1442                                 rdev_dec_pending(rdev, mddev);
1443                                 rcu_read_lock();
1444                         }
1445                 }
1446                 rcu_read_unlock();
1447
1448                 sectors -= s;
1449                 sect += s;
1450         }
1451 }
1452
1453 static void raid10d(mddev_t *mddev)
1454 {
1455         r10bio_t *r10_bio;
1456         struct bio *bio;
1457         unsigned long flags;
1458         conf_t *conf = mddev_to_conf(mddev);
1459         struct list_head *head = &conf->retry_list;
1460         int unplug=0;
1461         mdk_rdev_t *rdev;
1462
1463         md_check_recovery(mddev);
1464
1465         for (;;) {
1466                 char b[BDEVNAME_SIZE];
1467                 spin_lock_irqsave(&conf->device_lock, flags);
1468
1469                 if (conf->pending_bio_list.head) {
1470                         bio = bio_list_get(&conf->pending_bio_list);
1471                         blk_remove_plug(mddev->queue);
1472                         spin_unlock_irqrestore(&conf->device_lock, flags);
1473                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1474                         bitmap_unplug(mddev->bitmap);
1475
1476                         while (bio) { /* submit pending writes */
1477                                 struct bio *next = bio->bi_next;
1478                                 bio->bi_next = NULL;
1479                                 generic_make_request(bio);
1480                                 bio = next;
1481                         }
1482                         unplug = 1;
1483
1484                         continue;
1485                 }
1486
1487                 if (list_empty(head))
1488                         break;
1489                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1490                 list_del(head->prev);
1491                 conf->nr_queued--;
1492                 spin_unlock_irqrestore(&conf->device_lock, flags);
1493
1494                 mddev = r10_bio->mddev;
1495                 conf = mddev_to_conf(mddev);
1496                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1497                         sync_request_write(mddev, r10_bio);
1498                         unplug = 1;
1499                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1500                         recovery_request_write(mddev, r10_bio);
1501                         unplug = 1;
1502                 } else {
1503                         int mirror;
1504                         /* we got a read error. Maybe the drive is bad.  Maybe just
1505                          * the block and we can fix it.
1506                          * We freeze all other IO, and try reading the block from
1507                          * other devices.  When we find one, we re-write
1508                          * and check it that fixes the read error.
1509                          * This is all done synchronously while the array is
1510                          * frozen.
1511                          */
1512                         if (mddev->ro == 0) {
1513                                 freeze_array(conf);
1514                                 fix_read_error(conf, mddev, r10_bio);
1515                                 unfreeze_array(conf);
1516                         }
1517
1518                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1519                         r10_bio->devs[r10_bio->read_slot].bio =
1520                                 mddev->ro ? IO_BLOCKED : NULL;
1521                         mirror = read_balance(conf, r10_bio);
1522                         if (mirror == -1) {
1523                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1524                                        " read error for block %llu\n",
1525                                        bdevname(bio->bi_bdev,b),
1526                                        (unsigned long long)r10_bio->sector);
1527                                 raid_end_bio_io(r10_bio);
1528                                 bio_put(bio);
1529                         } else {
1530                                 const int do_sync = bio_sync(r10_bio->master_bio);
1531                                 bio_put(bio);
1532                                 rdev = conf->mirrors[mirror].rdev;
1533                                 if (printk_ratelimit())
1534                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1535                                                " another mirror\n",
1536                                                bdevname(rdev->bdev,b),
1537                                                (unsigned long long)r10_bio->sector);
1538                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1539                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1540                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1541                                         + rdev->data_offset;
1542                                 bio->bi_bdev = rdev->bdev;
1543                                 bio->bi_rw = READ | do_sync;
1544                                 bio->bi_private = r10_bio;
1545                                 bio->bi_end_io = raid10_end_read_request;
1546                                 unplug = 1;
1547                                 generic_make_request(bio);
1548                         }
1549                 }
1550         }
1551         spin_unlock_irqrestore(&conf->device_lock, flags);
1552         if (unplug)
1553                 unplug_slaves(mddev);
1554 }
1555
1556
1557 static int init_resync(conf_t *conf)
1558 {
1559         int buffs;
1560
1561         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1562         BUG_ON(conf->r10buf_pool);
1563         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1564         if (!conf->r10buf_pool)
1565                 return -ENOMEM;
1566         conf->next_resync = 0;
1567         return 0;
1568 }
1569
1570 /*
1571  * perform a "sync" on one "block"
1572  *
1573  * We need to make sure that no normal I/O request - particularly write
1574  * requests - conflict with active sync requests.
1575  *
1576  * This is achieved by tracking pending requests and a 'barrier' concept
1577  * that can be installed to exclude normal IO requests.
1578  *
1579  * Resync and recovery are handled very differently.
1580  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1581  *
1582  * For resync, we iterate over virtual addresses, read all copies,
1583  * and update if there are differences.  If only one copy is live,
1584  * skip it.
1585  * For recovery, we iterate over physical addresses, read a good
1586  * value for each non-in_sync drive, and over-write.
1587  *
1588  * So, for recovery we may have several outstanding complex requests for a
1589  * given address, one for each out-of-sync device.  We model this by allocating
1590  * a number of r10_bio structures, one for each out-of-sync device.
1591  * As we setup these structures, we collect all bio's together into a list
1592  * which we then process collectively to add pages, and then process again
1593  * to pass to generic_make_request.
1594  *
1595  * The r10_bio structures are linked using a borrowed master_bio pointer.
1596  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1597  * has its remaining count decremented to 0, the whole complex operation
1598  * is complete.
1599  *
1600  */
1601
1602 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1603 {
1604         conf_t *conf = mddev_to_conf(mddev);
1605         r10bio_t *r10_bio;
1606         struct bio *biolist = NULL, *bio;
1607         sector_t max_sector, nr_sectors;
1608         int disk;
1609         int i;
1610         int max_sync;
1611         int sync_blocks;
1612
1613         sector_t sectors_skipped = 0;
1614         int chunks_skipped = 0;
1615
1616         if (!conf->r10buf_pool)
1617                 if (init_resync(conf))
1618                         return 0;
1619
1620  skipped:
1621         max_sector = mddev->size << 1;
1622         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1623                 max_sector = mddev->resync_max_sectors;
1624         if (sector_nr >= max_sector) {
1625                 /* If we aborted, we need to abort the
1626                  * sync on the 'current' bitmap chucks (there can
1627                  * be several when recovering multiple devices).
1628                  * as we may have started syncing it but not finished.
1629                  * We can find the current address in
1630                  * mddev->curr_resync, but for recovery,
1631                  * we need to convert that to several
1632                  * virtual addresses.
1633                  */
1634                 if (mddev->curr_resync < max_sector) { /* aborted */
1635                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1636                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1637                                                 &sync_blocks, 1);
1638                         else for (i=0; i<conf->raid_disks; i++) {
1639                                 sector_t sect =
1640                                         raid10_find_virt(conf, mddev->curr_resync, i);
1641                                 bitmap_end_sync(mddev->bitmap, sect,
1642                                                 &sync_blocks, 1);
1643                         }
1644                 } else /* completed sync */
1645                         conf->fullsync = 0;
1646
1647                 bitmap_close_sync(mddev->bitmap);
1648                 close_sync(conf);
1649                 *skipped = 1;
1650                 return sectors_skipped;
1651         }
1652         if (chunks_skipped >= conf->raid_disks) {
1653                 /* if there has been nothing to do on any drive,
1654                  * then there is nothing to do at all..
1655                  */
1656                 *skipped = 1;
1657                 return (max_sector - sector_nr) + sectors_skipped;
1658         }
1659
1660         if (max_sector > mddev->resync_max)
1661                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1662
1663         /* make sure whole request will fit in a chunk - if chunks
1664          * are meaningful
1665          */
1666         if (conf->near_copies < conf->raid_disks &&
1667             max_sector > (sector_nr | conf->chunk_mask))
1668                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1669         /*
1670          * If there is non-resync activity waiting for us then
1671          * put in a delay to throttle resync.
1672          */
1673         if (!go_faster && conf->nr_waiting)
1674                 msleep_interruptible(1000);
1675
1676         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1677
1678         /* Again, very different code for resync and recovery.
1679          * Both must result in an r10bio with a list of bios that
1680          * have bi_end_io, bi_sector, bi_bdev set,
1681          * and bi_private set to the r10bio.
1682          * For recovery, we may actually create several r10bios
1683          * with 2 bios in each, that correspond to the bios in the main one.
1684          * In this case, the subordinate r10bios link back through a
1685          * borrowed master_bio pointer, and the counter in the master
1686          * includes a ref from each subordinate.
1687          */
1688         /* First, we decide what to do and set ->bi_end_io
1689          * To end_sync_read if we want to read, and
1690          * end_sync_write if we will want to write.
1691          */
1692
1693         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1694         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1695                 /* recovery... the complicated one */
1696                 int i, j, k;
1697                 r10_bio = NULL;
1698
1699                 for (i=0 ; i<conf->raid_disks; i++)
1700                         if (conf->mirrors[i].rdev &&
1701                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1702                                 int still_degraded = 0;
1703                                 /* want to reconstruct this device */
1704                                 r10bio_t *rb2 = r10_bio;
1705                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1706                                 int must_sync;
1707                                 /* Unless we are doing a full sync, we only need
1708                                  * to recover the block if it is set in the bitmap
1709                                  */
1710                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1711                                                               &sync_blocks, 1);
1712                                 if (sync_blocks < max_sync)
1713                                         max_sync = sync_blocks;
1714                                 if (!must_sync &&
1715                                     !conf->fullsync) {
1716                                         /* yep, skip the sync_blocks here, but don't assume
1717                                          * that there will never be anything to do here
1718                                          */
1719                                         chunks_skipped = -1;
1720                                         continue;
1721                                 }
1722
1723                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1724                                 raise_barrier(conf, rb2 != NULL);
1725                                 atomic_set(&r10_bio->remaining, 0);
1726
1727                                 r10_bio->master_bio = (struct bio*)rb2;
1728                                 if (rb2)
1729                                         atomic_inc(&rb2->remaining);
1730                                 r10_bio->mddev = mddev;
1731                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1732                                 r10_bio->sector = sect;
1733
1734                                 raid10_find_phys(conf, r10_bio);
1735                                 /* Need to check if this section will still be
1736                                  * degraded
1737                                  */
1738                                 for (j=0; j<conf->copies;j++) {
1739                                         int d = r10_bio->devs[j].devnum;
1740                                         if (conf->mirrors[d].rdev == NULL ||
1741                                             test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
1742                                                 still_degraded = 1;
1743                                                 break;
1744                                         }
1745                                 }
1746                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1747                                                               &sync_blocks, still_degraded);
1748
1749                                 for (j=0; j<conf->copies;j++) {
1750                                         int d = r10_bio->devs[j].devnum;
1751                                         if (conf->mirrors[d].rdev &&
1752                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1753                                                 /* This is where we read from */
1754                                                 bio = r10_bio->devs[0].bio;
1755                                                 bio->bi_next = biolist;
1756                                                 biolist = bio;
1757                                                 bio->bi_private = r10_bio;
1758                                                 bio->bi_end_io = end_sync_read;
1759                                                 bio->bi_rw = READ;
1760                                                 bio->bi_sector = r10_bio->devs[j].addr +
1761                                                         conf->mirrors[d].rdev->data_offset;
1762                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1763                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1764                                                 atomic_inc(&r10_bio->remaining);
1765                                                 /* and we write to 'i' */
1766
1767                                                 for (k=0; k<conf->copies; k++)
1768                                                         if (r10_bio->devs[k].devnum == i)
1769                                                                 break;
1770                                                 BUG_ON(k == conf->copies);
1771                                                 bio = r10_bio->devs[1].bio;
1772                                                 bio->bi_next = biolist;
1773                                                 biolist = bio;
1774                                                 bio->bi_private = r10_bio;
1775                                                 bio->bi_end_io = end_sync_write;
1776                                                 bio->bi_rw = WRITE;
1777                                                 bio->bi_sector = r10_bio->devs[k].addr +
1778                                                         conf->mirrors[i].rdev->data_offset;
1779                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1780
1781                                                 r10_bio->devs[0].devnum = d;
1782                                                 r10_bio->devs[1].devnum = i;
1783
1784                                                 break;
1785                                         }
1786                                 }
1787                                 if (j == conf->copies) {
1788                                         /* Cannot recover, so abort the recovery */
1789                                         put_buf(r10_bio);
1790                                         r10_bio = rb2;
1791                                         if (!test_and_set_bit(MD_RECOVERY_ERR, &mddev->recovery))
1792                                                 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1793                                                        mdname(mddev));
1794                                         break;
1795                                 }
1796                         }
1797                 if (biolist == NULL) {
1798                         while (r10_bio) {
1799                                 r10bio_t *rb2 = r10_bio;
1800                                 r10_bio = (r10bio_t*) rb2->master_bio;
1801                                 rb2->master_bio = NULL;
1802                                 put_buf(rb2);
1803                         }
1804                         goto giveup;
1805                 }
1806         } else {
1807                 /* resync. Schedule a read for every block at this virt offset */
1808                 int count = 0;
1809
1810                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1811                                        &sync_blocks, mddev->degraded) &&
1812                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1813                         /* We can skip this block */
1814                         *skipped = 1;
1815                         return sync_blocks + sectors_skipped;
1816                 }
1817                 if (sync_blocks < max_sync)
1818                         max_sync = sync_blocks;
1819                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1820
1821                 r10_bio->mddev = mddev;
1822                 atomic_set(&r10_bio->remaining, 0);
1823                 raise_barrier(conf, 0);
1824                 conf->next_resync = sector_nr;
1825
1826                 r10_bio->master_bio = NULL;
1827                 r10_bio->sector = sector_nr;
1828                 set_bit(R10BIO_IsSync, &r10_bio->state);
1829                 raid10_find_phys(conf, r10_bio);
1830                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1831
1832                 for (i=0; i<conf->copies; i++) {
1833                         int d = r10_bio->devs[i].devnum;
1834                         bio = r10_bio->devs[i].bio;
1835                         bio->bi_end_io = NULL;
1836                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
1837                         if (conf->mirrors[d].rdev == NULL ||
1838                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1839                                 continue;
1840                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1841                         atomic_inc(&r10_bio->remaining);
1842                         bio->bi_next = biolist;
1843                         biolist = bio;
1844                         bio->bi_private = r10_bio;
1845                         bio->bi_end_io = end_sync_read;
1846                         bio->bi_rw = READ;
1847                         bio->bi_sector = r10_bio->devs[i].addr +
1848                                 conf->mirrors[d].rdev->data_offset;
1849                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1850                         count++;
1851                 }
1852
1853                 if (count < 2) {
1854                         for (i=0; i<conf->copies; i++) {
1855                                 int d = r10_bio->devs[i].devnum;
1856                                 if (r10_bio->devs[i].bio->bi_end_io)
1857                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1858                         }
1859                         put_buf(r10_bio);
1860                         biolist = NULL;
1861                         goto giveup;
1862                 }
1863         }
1864
1865         for (bio = biolist; bio ; bio=bio->bi_next) {
1866
1867                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1868                 if (bio->bi_end_io)
1869                         bio->bi_flags |= 1 << BIO_UPTODATE;
1870                 bio->bi_vcnt = 0;
1871                 bio->bi_idx = 0;
1872                 bio->bi_phys_segments = 0;
1873                 bio->bi_hw_segments = 0;
1874                 bio->bi_size = 0;
1875         }
1876
1877         nr_sectors = 0;
1878         if (sector_nr + max_sync < max_sector)
1879                 max_sector = sector_nr + max_sync;
1880         do {
1881                 struct page *page;
1882                 int len = PAGE_SIZE;
1883                 disk = 0;
1884                 if (sector_nr + (len>>9) > max_sector)
1885                         len = (max_sector - sector_nr) << 9;
1886                 if (len == 0)
1887                         break;
1888                 for (bio= biolist ; bio ; bio=bio->bi_next) {
1889                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1890                         if (bio_add_page(bio, page, len, 0) == 0) {
1891                                 /* stop here */
1892                                 struct bio *bio2;
1893                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1894                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1895                                         /* remove last page from this bio */
1896                                         bio2->bi_vcnt--;
1897                                         bio2->bi_size -= len;
1898                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1899                                 }
1900                                 goto bio_full;
1901                         }
1902                         disk = i;
1903                 }
1904                 nr_sectors += len>>9;
1905                 sector_nr += len>>9;
1906         } while (biolist->bi_vcnt < RESYNC_PAGES);
1907  bio_full:
1908         r10_bio->sectors = nr_sectors;
1909
1910         while (biolist) {
1911                 bio = biolist;
1912                 biolist = biolist->bi_next;
1913
1914                 bio->bi_next = NULL;
1915                 r10_bio = bio->bi_private;
1916                 r10_bio->sectors = nr_sectors;
1917
1918                 if (bio->bi_end_io == end_sync_read) {
1919                         md_sync_acct(bio->bi_bdev, nr_sectors);
1920                         generic_make_request(bio);
1921                 }
1922         }
1923
1924         if (sectors_skipped)
1925                 /* pretend they weren't skipped, it makes
1926                  * no important difference in this case
1927                  */
1928                 md_done_sync(mddev, sectors_skipped, 1);
1929
1930         return sectors_skipped + nr_sectors;
1931  giveup:
1932         /* There is nowhere to write, so all non-sync
1933          * drives must be failed, so try the next chunk...
1934          */
1935         {
1936         sector_t sec = max_sector - sector_nr;
1937         sectors_skipped += sec;
1938         chunks_skipped ++;
1939         sector_nr = max_sector;
1940         goto skipped;
1941         }
1942 }
1943
1944 static int run(mddev_t *mddev)
1945 {
1946         conf_t *conf;
1947         int i, disk_idx;
1948         mirror_info_t *disk;
1949         mdk_rdev_t *rdev;
1950         struct list_head *tmp;
1951         int nc, fc, fo;
1952         sector_t stride, size;
1953
1954         if (mddev->chunk_size == 0) {
1955                 printk(KERN_ERR "md/raid10: non-zero chunk size required.\n");
1956                 return -EINVAL;
1957         }
1958
1959         nc = mddev->layout & 255;
1960         fc = (mddev->layout >> 8) & 255;
1961         fo = mddev->layout & (1<<16);
1962         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
1963             (mddev->layout >> 17)) {
1964                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
1965                        mdname(mddev), mddev->layout);
1966                 goto out;
1967         }
1968         /*
1969          * copy the already verified devices into our private RAID10
1970          * bookkeeping area. [whatever we allocate in run(),
1971          * should be freed in stop()]
1972          */
1973         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1974         mddev->private = conf;
1975         if (!conf) {
1976                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1977                         mdname(mddev));
1978                 goto out;
1979         }
1980         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1981                                  GFP_KERNEL);
1982         if (!conf->mirrors) {
1983                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1984                        mdname(mddev));
1985                 goto out_free_conf;
1986         }
1987
1988         conf->tmppage = alloc_page(GFP_KERNEL);
1989         if (!conf->tmppage)
1990                 goto out_free_conf;
1991
1992         conf->mddev = mddev;
1993         conf->raid_disks = mddev->raid_disks;
1994         conf->near_copies = nc;
1995         conf->far_copies = fc;
1996         conf->copies = nc*fc;
1997         conf->far_offset = fo;
1998         conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
1999         conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
2000         size = mddev->size >> (conf->chunk_shift-1);
2001         sector_div(size, fc);
2002         size = size * conf->raid_disks;
2003         sector_div(size, nc);
2004         /* 'size' is now the number of chunks in the array */
2005         /* calculate "used chunks per device" in 'stride' */
2006         stride = size * conf->copies;
2007
2008         /* We need to round up when dividing by raid_disks to
2009          * get the stride size.
2010          */
2011         stride += conf->raid_disks - 1;
2012         sector_div(stride, conf->raid_disks);
2013         mddev->size = stride  << (conf->chunk_shift-1);
2014
2015         if (fo)
2016                 stride = 1;
2017         else
2018                 sector_div(stride, fc);
2019         conf->stride = stride << conf->chunk_shift;
2020
2021         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2022                                                 r10bio_pool_free, conf);
2023         if (!conf->r10bio_pool) {
2024                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2025                         mdname(mddev));
2026                 goto out_free_conf;
2027         }
2028
2029         rdev_for_each(rdev, tmp, mddev) {
2030                 disk_idx = rdev->raid_disk;
2031                 if (disk_idx >= mddev->raid_disks
2032                     || disk_idx < 0)
2033                         continue;
2034                 disk = conf->mirrors + disk_idx;
2035
2036                 disk->rdev = rdev;
2037
2038                 blk_queue_stack_limits(mddev->queue,
2039                                        rdev->bdev->bd_disk->queue);
2040                 /* as we don't honour merge_bvec_fn, we must never risk
2041                  * violating it, so limit ->max_sector to one PAGE, as
2042                  * a one page request is never in violation.
2043                  */
2044                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2045                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
2046                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
2047
2048                 disk->head_position = 0;
2049         }
2050         spin_lock_init(&conf->device_lock);
2051         INIT_LIST_HEAD(&conf->retry_list);
2052
2053         spin_lock_init(&conf->resync_lock);
2054         init_waitqueue_head(&conf->wait_barrier);
2055
2056         /* need to check that every block has at least one working mirror */
2057         if (!enough(conf)) {
2058                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2059                        mdname(mddev));
2060                 goto out_free_conf;
2061         }
2062
2063         mddev->degraded = 0;
2064         for (i = 0; i < conf->raid_disks; i++) {
2065
2066                 disk = conf->mirrors + i;
2067
2068                 if (!disk->rdev ||
2069                     !test_bit(In_sync, &disk->rdev->flags)) {
2070                         disk->head_position = 0;
2071                         mddev->degraded++;
2072                 }
2073         }
2074
2075
2076         mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2077         if (!mddev->thread) {
2078                 printk(KERN_ERR
2079                        "raid10: couldn't allocate thread for %s\n",
2080                        mdname(mddev));
2081                 goto out_free_conf;
2082         }
2083
2084         printk(KERN_INFO
2085                 "raid10: raid set %s active with %d out of %d devices\n",
2086                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2087                 mddev->raid_disks);
2088         /*
2089          * Ok, everything is just fine now
2090          */
2091         mddev->array_size = size << (conf->chunk_shift-1);
2092         mddev->resync_max_sectors = size << conf->chunk_shift;
2093
2094         mddev->queue->unplug_fn = raid10_unplug;
2095         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2096         mddev->queue->backing_dev_info.congested_data = mddev;
2097
2098         /* Calculate max read-ahead size.
2099          * We need to readahead at least twice a whole stripe....
2100          * maybe...
2101          */
2102         {
2103                 int stripe = conf->raid_disks * (mddev->chunk_size / PAGE_SIZE);
2104                 stripe /= conf->near_copies;
2105                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2106                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2107         }
2108
2109         if (conf->near_copies < mddev->raid_disks)
2110                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2111         return 0;
2112
2113 out_free_conf:
2114         if (conf->r10bio_pool)
2115                 mempool_destroy(conf->r10bio_pool);
2116         safe_put_page(conf->tmppage);
2117         kfree(conf->mirrors);
2118         kfree(conf);
2119         mddev->private = NULL;
2120 out:
2121         return -EIO;
2122 }
2123
2124 static int stop(mddev_t *mddev)
2125 {
2126         conf_t *conf = mddev_to_conf(mddev);
2127
2128         md_unregister_thread(mddev->thread);
2129         mddev->thread = NULL;
2130         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2131         if (conf->r10bio_pool)
2132                 mempool_destroy(conf->r10bio_pool);
2133         kfree(conf->mirrors);
2134         kfree(conf);
2135         mddev->private = NULL;
2136         return 0;
2137 }
2138
2139 static void raid10_quiesce(mddev_t *mddev, int state)
2140 {
2141         conf_t *conf = mddev_to_conf(mddev);
2142
2143         switch(state) {
2144         case 1:
2145                 raise_barrier(conf, 0);
2146                 break;
2147         case 0:
2148                 lower_barrier(conf);
2149                 break;
2150         }
2151         if (mddev->thread) {
2152                 if (mddev->bitmap)
2153                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2154                 else
2155                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2156                 md_wakeup_thread(mddev->thread);
2157         }
2158 }
2159
2160 static struct mdk_personality raid10_personality =
2161 {
2162         .name           = "raid10",
2163         .level          = 10,
2164         .owner          = THIS_MODULE,
2165         .make_request   = make_request,
2166         .run            = run,
2167         .stop           = stop,
2168         .status         = status,
2169         .error_handler  = error,
2170         .hot_add_disk   = raid10_add_disk,
2171         .hot_remove_disk= raid10_remove_disk,
2172         .spare_active   = raid10_spare_active,
2173         .sync_request   = sync_request,
2174         .quiesce        = raid10_quiesce,
2175 };
2176
2177 static int __init raid_init(void)
2178 {
2179         return register_md_personality(&raid10_personality);
2180 }
2181
2182 static void raid_exit(void)
2183 {
2184         unregister_md_personality(&raid10_personality);
2185 }
2186
2187 module_init(raid_init);
2188 module_exit(raid_exit);
2189 MODULE_LICENSE("GPL");
2190 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2191 MODULE_ALIAS("md-raid10");
2192 MODULE_ALIAS("md-level-10");