git.oblomov.eu 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, bio->bi_size,
 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 int raid10_end_read_request(struct bio *bio, unsigned int bytes_done, 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         if (bio->bi_size)
 254                 return 1;
 255
 256         slot = r10_bio->read_slot;
 257         dev = r10_bio->devs[slot].devnum;
 258         /*
 259          * this branch is our 'one mirror IO has finished' event handler:
 260          */
 261         update_head_pos(slot, r10_bio);
 262
 263         if (uptodate) {
 264                 /*
 265                  * Set R10BIO_Uptodate in our master bio, so that
 266                  * we will return a good error code to the higher
 267                  * levels even if IO on some other mirrored buffer fails.
 268                  *
 269                  * The 'master' represents the composite IO operation to
 270                  * user-side. So if something waits for IO, then it will
 271                  * wait for the 'master' bio.
 272                  */
 273                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 274                 raid_end_bio_io(r10_bio);
 275         } else {
 276                 /*
 277                  * oops, read error:
 278                  */
 279                 char b[BDEVNAME_SIZE];
 280                 if (printk_ratelimit())
 281                         printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
 282                                bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
 283                 reschedule_retry(r10_bio);
 284         }
 285
 286         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 287         return 0;
 288 }
 289
 290 static int raid10_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
 291 {
 292         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 293         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
 294         int slot, dev;
 295         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 296
 297         if (bio->bi_size)
 298                 return 1;
 299
 300         for (slot = 0; slot < conf->copies; slot++)
 301                 if (r10_bio->devs[slot].bio == bio)
 302                         break;
 303         dev = r10_bio->devs[slot].devnum;
 304
 305         /*
 306          * this branch is our 'one mirror IO has finished' event handler:
 307          */
 308         if (!uptodate) {
 309                 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
 310                 /* an I/O failed, we can't clear the bitmap */
 311                 set_bit(R10BIO_Degraded, &r10_bio->state);
 312         } else
 313                 /*
 314                  * Set R10BIO_Uptodate in our master bio, so that
 315                  * we will return a good error code for to the higher
 316                  * levels even if IO on some other mirrored buffer fails.
 317                  *
 318                  * The 'master' represents the composite IO operation to
 319                  * user-side. So if something waits for IO, then it will
 320                  * wait for the 'master' bio.
 321                  */
 322                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 323
 324         update_head_pos(slot, r10_bio);
 325
 326         /*
 327          *
 328          * Let's see if all mirrored write operations have finished
 329          * already.
 330          */
 331         if (atomic_dec_and_test(&r10_bio->remaining)) {
 332                 /* clear the bitmap if all writes complete successfully */
 333                 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 334                                 r10_bio->sectors,
 335                                 !test_bit(R10BIO_Degraded, &r10_bio->state),
 336                                 0);
 337                 md_write_end(r10_bio->mddev);
 338                 raid_end_bio_io(r10_bio);
 339         }
 340
 341         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 342         return 0;
 343 }
 344
 345
 346 /*
 347  * RAID10 layout manager
 348  * Aswell as the chunksize and raid_disks count, there are two
 349  * parameters: near_copies and far_copies.
 350  * near_copies * far_copies must be <= raid_disks.
 351  * Normally one of these will be 1.
 352  * If both are 1, we get raid0.
 353  * If near_copies == raid_disks, we get raid1.
 354  *
 355  * Chunks are layed out in raid0 style with near_copies copies of the
 356  * first chunk, followed by near_copies copies of the next chunk and
 357  * so on.
 358  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 359  * as described above, we start again with a device offset of near_copies.
 360  * So we effectively have another copy of the whole array further down all
 361  * the drives, but with blocks on different drives.
 362  * With this layout, and block is never stored twice on the one device.
 363  *
 364  * raid10_find_phys finds the sector offset of a given virtual sector
 365  * on each device that it is on.
 366  *
 367  * raid10_find_virt does the reverse mapping, from a device and a
 368  * sector offset to a virtual address
 369  */
 370
 371 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
 372 {
 373         int n,f;
 374         sector_t sector;
 375         sector_t chunk;
 376         sector_t stripe;
 377         int dev;
 378
 379         int slot = 0;
 380
 381         /* now calculate first sector/dev */
 382         chunk = r10bio->sector >> conf->chunk_shift;
 383         sector = r10bio->sector & conf->chunk_mask;
 384
 385         chunk *= conf->near_copies;
 386         stripe = chunk;
 387         dev = sector_div(stripe, conf->raid_disks);
 388         if (conf->far_offset)
 389                 stripe *= conf->far_copies;
 390
 391         sector += stripe << conf->chunk_shift;
 392
 393         /* and calculate all the others */
 394         for (n=0; n < conf->near_copies; n++) {
 395                 int d = dev;
 396                 sector_t s = sector;
 397                 r10bio->devs[slot].addr = sector;
 398                 r10bio->devs[slot].devnum = d;
 399                 slot++;
 400
 401                 for (f = 1; f < conf->far_copies; f++) {
 402                         d += conf->near_copies;
 403                         if (d >= conf->raid_disks)
 404                                 d -= conf->raid_disks;
 405                         s += conf->stride;
 406                         r10bio->devs[slot].devnum = d;
 407                         r10bio->devs[slot].addr = s;
 408                         slot++;
 409                 }
 410                 dev++;
 411                 if (dev >= conf->raid_disks) {
 412                         dev = 0;
 413                         sector += (conf->chunk_mask + 1);
 414                 }
 415         }
 416         BUG_ON(slot != conf->copies);
 417 }
 418
 419 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
 420 {
 421         sector_t offset, chunk, vchunk;
 422
 423         offset = sector & conf->chunk_mask;
 424         if (conf->far_offset) {
 425                 int fc;
 426                 chunk = sector >> conf->chunk_shift;
 427                 fc = sector_div(chunk, conf->far_copies);
 428                 dev -= fc * conf->near_copies;
 429                 if (dev < 0)
 430                         dev += conf->raid_disks;
 431         } else {
 432                 while (sector > conf->stride) {
 433                         sector -= conf->stride;
 434                         if (dev < conf->near_copies)
 435                                 dev += conf->raid_disks - conf->near_copies;
 436                         else
 437                                 dev -= conf->near_copies;
 438                 }
 439                 chunk = sector >> conf->chunk_shift;
 440         }
 441         vchunk = chunk * conf->raid_disks + dev;
 442         sector_div(vchunk, conf->near_copies);
 443         return (vchunk << conf->chunk_shift) + offset;
 444 }
 445
 446 /**
 447  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 448  *      @q: request queue
 449  *      @bio: the buffer head that's been built up so far
 450  *      @biovec: the request that could be merged to it.
 451  *
 452  *      Return amount of bytes we can accept at this offset
 453  *      If near_copies == raid_disk, there are no striping issues,
 454  *      but in that case, the function isn't called at all.
 455  */
 456 static int raid10_mergeable_bvec(request_queue_t *q, struct bio *bio,
 457                                 struct bio_vec *bio_vec)
 458 {
 459         mddev_t *mddev = q->queuedata;
 460         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
 461         int max;
 462         unsigned int chunk_sectors = mddev->chunk_size >> 9;
 463         unsigned int bio_sectors = bio->bi_size >> 9;
 464
 465         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 466         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 467         if (max <= bio_vec->bv_len && bio_sectors == 0)
 468                 return bio_vec->bv_len;
 469         else
 470                 return max;
 471 }
 472
 473 /*
 474  * This routine returns the disk from which the requested read should
 475  * be done. There is a per-array 'next expected sequential IO' sector
 476  * number - if this matches on the next IO then we use the last disk.
 477  * There is also a per-disk 'last know head position' sector that is
 478  * maintained from IRQ contexts, both the normal and the resync IO
 479  * completion handlers update this position correctly. If there is no
 480  * perfect sequential match then we pick the disk whose head is closest.
 481  *
 482  * If there are 2 mirrors in the same 2 devices, performance degrades
 483  * because position is mirror, not device based.
 484  *
 485  * The rdev for the device selected will have nr_pending incremented.
 486  */
 487
 488 /*
 489  * FIXME: possibly should rethink readbalancing and do it differently
 490  * depending on near_copies / far_copies geometry.
 491  */
 492 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
 493 {
 494         const unsigned long this_sector = r10_bio->sector;
 495         int disk, slot, nslot;
 496         const int sectors = r10_bio->sectors;
 497         sector_t new_distance, current_distance;
 498         mdk_rdev_t *rdev;
 499
 500         raid10_find_phys(conf, r10_bio);
 501         rcu_read_lock();
 502         /*
 503          * Check if we can balance. We can balance on the whole
 504          * device if no resync is going on (recovery is ok), or below
 505          * the resync window. We take the first readable disk when
 506          * above the resync window.
 507          */
 508         if (conf->mddev->recovery_cp < MaxSector
 509             && (this_sector + sectors >= conf->next_resync)) {
 510                 /* make sure that disk is operational */
 511                 slot = 0;
 512                 disk = r10_bio->devs[slot].devnum;
 513
 514                 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 515                        r10_bio->devs[slot].bio == IO_BLOCKED ||
 516                        !test_bit(In_sync, &rdev->flags)) {
 517                         slot++;
 518                         if (slot == conf->copies) {
 519                                 slot = 0;
 520                                 disk = -1;
 521                                 break;
 522                         }
 523                         disk = r10_bio->devs[slot].devnum;
 524                 }
 525                 goto rb_out;
 526         }
 527
 528
 529         /* make sure the disk is operational */
 530         slot = 0;
 531         disk = r10_bio->devs[slot].devnum;
 532         while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 533                r10_bio->devs[slot].bio == IO_BLOCKED ||
 534                !test_bit(In_sync, &rdev->flags)) {
 535                 slot ++;
 536                 if (slot == conf->copies) {
 537                         disk = -1;
 538                         goto rb_out;
 539                 }
 540                 disk = r10_bio->devs[slot].devnum;
 541         }
 542
 543
 544         current_distance = abs(r10_bio->devs[slot].addr -
 545                                conf->mirrors[disk].head_position);
 546
 547         /* Find the disk whose head is closest */
 548
 549         for (nslot = slot; nslot < conf->copies; nslot++) {
 550                 int ndisk = r10_bio->devs[nslot].devnum;
 551
 552
 553                 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
 554                     r10_bio->devs[nslot].bio == IO_BLOCKED ||
 555                     !test_bit(In_sync, &rdev->flags))
 556                         continue;
 557
 558                 /* This optimisation is debatable, and completely destroys
 559                  * sequential read speed for 'far copies' arrays.  So only
 560                  * keep it for 'near' arrays, and review those later.
 561                  */
 562                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
 563                         disk = ndisk;
 564                         slot = nslot;
 565                         break;
 566                 }
 567                 new_distance = abs(r10_bio->devs[nslot].addr -
 568                                    conf->mirrors[ndisk].head_position);
 569                 if (new_distance < current_distance) {
 570                         current_distance = new_distance;
 571                         disk = ndisk;
 572                         slot = nslot;
 573                 }
 574         }
 575
 576 rb_out:
 577         r10_bio->read_slot = slot;
 578 /*      conf->next_seq_sect = this_sector + sectors;*/
 579
 580         if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
 581                 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
 582         else
 583                 disk = -1;
 584         rcu_read_unlock();
 585
 586         return disk;
 587 }
 588
 589 static void unplug_slaves(mddev_t *mddev)
 590 {
 591         conf_t *conf = mddev_to_conf(mddev);
 592         int i;
 593
 594         rcu_read_lock();
 595         for (i=0; i<mddev->raid_disks; i++) {
 596                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 597                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 598                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 599
 600                         atomic_inc(&rdev->nr_pending);
 601                         rcu_read_unlock();
 602
 603                         if (r_queue->unplug_fn)
 604                                 r_queue->unplug_fn(r_queue);
 605
 606                         rdev_dec_pending(rdev, mddev);
 607                         rcu_read_lock();
 608                 }
 609         }
 610         rcu_read_unlock();
 611 }
 612
 613 static void raid10_unplug(request_queue_t *q)
 614 {
 615         mddev_t *mddev = q->queuedata;
 616
 617         unplug_slaves(q->queuedata);
 618         md_wakeup_thread(mddev->thread);
 619 }
 620
 621 static int raid10_issue_flush(request_queue_t *q, struct gendisk *disk,
 622                              sector_t *error_sector)
 623 {
 624         mddev_t *mddev = q->queuedata;
 625         conf_t *conf = mddev_to_conf(mddev);
 626         int i, ret = 0;
 627
 628         rcu_read_lock();
 629         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
 630                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 631                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 632                         struct block_device *bdev = rdev->bdev;
 633                         request_queue_t *r_queue = bdev_get_queue(bdev);
 634
 635                         if (!r_queue->issue_flush_fn)
 636                                 ret = -EOPNOTSUPP;
 637                         else {
 638                                 atomic_inc(&rdev->nr_pending);
 639                                 rcu_read_unlock();
 640                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
 641                                                               error_sector);
 642                                 rdev_dec_pending(rdev, mddev);
 643                                 rcu_read_lock();
 644                         }
 645                 }
 646         }
 647         rcu_read_unlock();
 648         return ret;
 649 }
 650
 651 /* Barriers....
 652  * Sometimes we need to suspend IO while we do something else,
 653  * either some resync/recovery, or reconfigure the array.
 654  * To do this we raise a 'barrier'.
 655  * The 'barrier' is a counter that can be raised multiple times
 656  * to count how many activities are happening which preclude
 657  * normal IO.
 658  * We can only raise the barrier if there is no pending IO.
 659  * i.e. if nr_pending == 0.
 660  * We choose only to raise the barrier if no-one is waiting for the
 661  * barrier to go down.  This means that as soon as an IO request
 662  * is ready, no other operations which require a barrier will start
 663  * until the IO request has had a chance.
 664  *
 665  * So: regular IO calls 'wait_barrier'.  When that returns there
 666  *    is no backgroup IO happening,  It must arrange to call
 667  *    allow_barrier when it has finished its IO.
 668  * backgroup IO calls must call raise_barrier.  Once that returns
 669  *    there is no normal IO happeing.  It must arrange to call
 670  *    lower_barrier when the particular background IO completes.
 671  */
 672 #define RESYNC_DEPTH 32
 673
 674 static void raise_barrier(conf_t *conf, int force)
 675 {
 676         BUG_ON(force && !conf->barrier);
 677         spin_lock_irq(&conf->resync_lock);
 678
 679         /* Wait until no block IO is waiting (unless 'force') */
 680         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 681                             conf->resync_lock,
 682                             raid10_unplug(conf->mddev->queue));
 683
 684         /* block any new IO from starting */
 685         conf->barrier++;
 686
 687         /* No wait for all pending IO to complete */
 688         wait_event_lock_irq(conf->wait_barrier,
 689                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 690                             conf->resync_lock,
 691                             raid10_unplug(conf->mddev->queue));
 692
 693         spin_unlock_irq(&conf->resync_lock);
 694 }
 695
 696 static void lower_barrier(conf_t *conf)
 697 {
 698         unsigned long flags;
 699         spin_lock_irqsave(&conf->resync_lock, flags);
 700         conf->barrier--;
 701         spin_unlock_irqrestore(&conf->resync_lock, flags);
 702         wake_up(&conf->wait_barrier);
 703 }
 704
 705 static void wait_barrier(conf_t *conf)
 706 {
 707         spin_lock_irq(&conf->resync_lock);
 708         if (conf->barrier) {
 709                 conf->nr_waiting++;
 710                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 711                                     conf->resync_lock,
 712                                     raid10_unplug(conf->mddev->queue));
 713                 conf->nr_waiting--;
 714         }
 715         conf->nr_pending++;
 716         spin_unlock_irq(&conf->resync_lock);
 717 }
 718
 719 static void allow_barrier(conf_t *conf)
 720 {
 721         unsigned long flags;
 722         spin_lock_irqsave(&conf->resync_lock, flags);
 723         conf->nr_pending--;
 724         spin_unlock_irqrestore(&conf->resync_lock, flags);
 725         wake_up(&conf->wait_barrier);
 726 }
 727
 728 static void freeze_array(conf_t *conf)
 729 {
 730         /* stop syncio and normal IO and wait for everything to
 731          * go quiet.
 732          * We increment barrier and nr_waiting, and then
 733          * wait until barrier+nr_pending match nr_queued+2
 734          */
 735         spin_lock_irq(&conf->resync_lock);
 736         conf->barrier++;
 737         conf->nr_waiting++;
 738         wait_event_lock_irq(conf->wait_barrier,
 739                             conf->barrier+conf->nr_pending == conf->nr_queued+2,
 740                             conf->resync_lock,
 741                             raid10_unplug(conf->mddev->queue));
 742         spin_unlock_irq(&conf->resync_lock);
 743 }
 744
 745 static void unfreeze_array(conf_t *conf)
 746 {
 747         /* reverse the effect of the freeze */
 748         spin_lock_irq(&conf->resync_lock);
 749         conf->barrier--;
 750         conf->nr_waiting--;
 751         wake_up(&conf->wait_barrier);
 752         spin_unlock_irq(&conf->resync_lock);
 753 }
 754
 755 static int make_request(request_queue_t *q, struct bio * bio)
 756 {
 757         mddev_t *mddev = q->queuedata;
 758         conf_t *conf = mddev_to_conf(mddev);
 759         mirror_info_t *mirror;
 760         r10bio_t *r10_bio;
 761         struct bio *read_bio;
 762         int i;
 763         int chunk_sects = conf->chunk_mask + 1;
 764         const int rw = bio_data_dir(bio);
 765         struct bio_list bl;
 766         unsigned long flags;
 767
 768         if (unlikely(bio_barrier(bio))) {
 769                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
 770                 return 0;
 771         }
 772
 773         /* If this request crosses a chunk boundary, we need to
 774          * split it.  This will only happen for 1 PAGE (or less) requests.
 775          */
 776         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 777                       > chunk_sects &&
 778                     conf->near_copies < conf->raid_disks)) {
 779                 struct bio_pair *bp;
 780                 /* Sanity check -- queue functions should prevent this happening */
 781                 if (bio->bi_vcnt != 1 ||
 782                     bio->bi_idx != 0)
 783                         goto bad_map;
 784                 /* This is a one page bio that upper layers
 785                  * refuse to split for us, so we need to split it.
 786                  */
 787                 bp = bio_split(bio, bio_split_pool,
 788                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 789                 if (make_request(q, &bp->bio1))
 790                         generic_make_request(&bp->bio1);
 791                 if (make_request(q, &bp->bio2))
 792                         generic_make_request(&bp->bio2);
 793
 794                 bio_pair_release(bp);
 795                 return 0;
 796         bad_map:
 797                 printk("raid10_make_request bug: can't convert block across chunks"
 798                        " or bigger than %dk %llu %d\n", chunk_sects/2,
 799                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 800
 801                 bio_io_error(bio, bio->bi_size);
 802                 return 0;
 803         }
 804
 805         md_write_start(mddev, bio);
 806
 807         /*
 808          * Register the new request and wait if the reconstruction
 809          * thread has put up a bar for new requests.
 810          * Continue immediately if no resync is active currently.
 811          */
 812         wait_barrier(conf);
 813
 814         disk_stat_inc(mddev->gendisk, ios[rw]);
 815         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
 816
 817         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 818
 819         r10_bio->master_bio = bio;
 820         r10_bio->sectors = bio->bi_size >> 9;
 821
 822         r10_bio->mddev = mddev;
 823         r10_bio->sector = bio->bi_sector;
 824         r10_bio->state = 0;
 825
 826         if (rw == READ) {
 827                 /*
 828                  * read balancing logic:
 829                  */
 830                 int disk = read_balance(conf, r10_bio);
 831                 int slot = r10_bio->read_slot;
 832                 if (disk < 0) {
 833                         raid_end_bio_io(r10_bio);
 834                         return 0;
 835                 }
 836                 mirror = conf->mirrors + disk;
 837
 838                 read_bio = bio_clone(bio, GFP_NOIO);
 839
 840                 r10_bio->devs[slot].bio = read_bio;
 841
 842                 read_bio->bi_sector = r10_bio->devs[slot].addr +
 843                         mirror->rdev->data_offset;
 844                 read_bio->bi_bdev = mirror->rdev->bdev;
 845                 read_bio->bi_end_io = raid10_end_read_request;
 846                 read_bio->bi_rw = READ;
 847                 read_bio->bi_private = r10_bio;
 848
 849                 generic_make_request(read_bio);
 850                 return 0;
 851         }
 852
 853         /*
 854          * WRITE:
 855          */
 856         /* first select target devices under spinlock and
 857          * inc refcount on their rdev.  Record them by setting
 858          * bios[x] to bio
 859          */
 860         raid10_find_phys(conf, r10_bio);
 861         rcu_read_lock();
 862         for (i = 0;  i < conf->copies; i++) {
 863                 int d = r10_bio->devs[i].devnum;
 864                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
 865                 if (rdev &&
 866                     !test_bit(Faulty, &rdev->flags)) {
 867                         atomic_inc(&rdev->nr_pending);
 868                         r10_bio->devs[i].bio = bio;
 869                 } else {
 870                         r10_bio->devs[i].bio = NULL;
 871                         set_bit(R10BIO_Degraded, &r10_bio->state);
 872                 }
 873         }
 874         rcu_read_unlock();
 875
 876         atomic_set(&r10_bio->remaining, 0);
 877
 878         bio_list_init(&bl);
 879         for (i = 0; i < conf->copies; i++) {
 880                 struct bio *mbio;
 881                 int d = r10_bio->devs[i].devnum;
 882                 if (!r10_bio->devs[i].bio)
 883                         continue;
 884
 885                 mbio = bio_clone(bio, GFP_NOIO);
 886                 r10_bio->devs[i].bio = mbio;
 887
 888                 mbio->bi_sector = r10_bio->devs[i].addr+
 889                         conf->mirrors[d].rdev->data_offset;
 890                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
 891                 mbio->bi_end_io = raid10_end_write_request;
 892                 mbio->bi_rw = WRITE;
 893                 mbio->bi_private = r10_bio;
 894
 895                 atomic_inc(&r10_bio->remaining);
 896                 bio_list_add(&bl, mbio);
 897         }
 898
 899         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
 900         spin_lock_irqsave(&conf->device_lock, flags);
 901         bio_list_merge(&conf->pending_bio_list, &bl);
 902         blk_plug_device(mddev->queue);
 903         spin_unlock_irqrestore(&conf->device_lock, flags);
 904
 905         return 0;
 906 }
 907
 908 static void status(struct seq_file *seq, mddev_t *mddev)
 909 {
 910         conf_t *conf = mddev_to_conf(mddev);
 911         int i;
 912
 913         if (conf->near_copies < conf->raid_disks)
 914                 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
 915         if (conf->near_copies > 1)
 916                 seq_printf(seq, " %d near-copies", conf->near_copies);
 917         if (conf->far_copies > 1) {
 918                 if (conf->far_offset)
 919                         seq_printf(seq, " %d offset-copies", conf->far_copies);
 920                 else
 921                         seq_printf(seq, " %d far-copies", conf->far_copies);
 922         }
 923         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 924                                                 conf->working_disks);
 925         for (i = 0; i < conf->raid_disks; i++)
 926                 seq_printf(seq, "%s",
 927                               conf->mirrors[i].rdev &&
 928                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
 929         seq_printf(seq, "]");
 930 }
 931
 932 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 933 {
 934         char b[BDEVNAME_SIZE];
 935         conf_t *conf = mddev_to_conf(mddev);
 936
 937         /*
 938          * If it is not operational, then we have already marked it as dead
 939          * else if it is the last working disks, ignore the error, let the
 940          * next level up know.
 941          * else mark the drive as failed
 942          */
 943         if (test_bit(In_sync, &rdev->flags)
 944             && conf->working_disks == 1)
 945                 /*
 946                  * Don't fail the drive, just return an IO error.
 947                  * The test should really be more sophisticated than
 948                  * "working_disks == 1", but it isn't critical, and
 949                  * can wait until we do more sophisticated "is the drive
 950                  * really dead" tests...
 951                  */
 952                 return;
 953         if (test_bit(In_sync, &rdev->flags)) {
 954                 mddev->degraded++;
 955                 conf->working_disks--;
 956                 /*
 957                  * if recovery is running, make sure it aborts.
 958                  */
 959                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 960         }
 961         clear_bit(In_sync, &rdev->flags);
 962         set_bit(Faulty, &rdev->flags);
 963         mddev->sb_dirty = 1;
 964         printk(KERN_ALERT "raid10: Disk failure on %s, disabling device. \n"
 965                 "       Operation continuing on %d devices\n",
 966                 bdevname(rdev->bdev,b), conf->working_disks);
 967 }
 968
 969 static void print_conf(conf_t *conf)
 970 {
 971         int i;
 972         mirror_info_t *tmp;
 973
 974         printk("RAID10 conf printout:\n");
 975         if (!conf) {
 976                 printk("(!conf)\n");
 977                 return;
 978         }
 979         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 980                 conf->raid_disks);
 981
 982         for (i = 0; i < conf->raid_disks; i++) {
 983                 char b[BDEVNAME_SIZE];
 984                 tmp = conf->mirrors + i;
 985                 if (tmp->rdev)
 986                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
 987                                 i, !test_bit(In_sync, &tmp->rdev->flags),
 988                                 !test_bit(Faulty, &tmp->rdev->flags),
 989                                 bdevname(tmp->rdev->bdev,b));
 990         }
 991 }
 992
 993 static void close_sync(conf_t *conf)
 994 {
 995         wait_barrier(conf);
 996         allow_barrier(conf);
 997
 998         mempool_destroy(conf->r10buf_pool);
 999         conf->r10buf_pool = NULL;
1000 }
1001
1002 /* check if there are enough drives for
1003  * every block to appear on atleast one
1004  */
1005 static int enough(conf_t *conf)
1006 {
1007         int first = 0;
1008
1009         do {
1010                 int n = conf->copies;
1011                 int cnt = 0;
1012                 while (n--) {
1013                         if (conf->mirrors[first].rdev)
1014                                 cnt++;
1015                         first = (first+1) % conf->raid_disks;
1016                 }
1017                 if (cnt == 0)
1018                         return 0;
1019         } while (first != 0);
1020         return 1;
1021 }
1022
1023 static int raid10_spare_active(mddev_t *mddev)
1024 {
1025         int i;
1026         conf_t *conf = mddev->private;
1027         mirror_info_t *tmp;
1028
1029         /*
1030          * Find all non-in_sync disks within the RAID10 configuration
1031          * and mark them in_sync
1032          */
1033         for (i = 0; i < conf->raid_disks; i++) {
1034                 tmp = conf->mirrors + i;
1035                 if (tmp->rdev
1036                     && !test_bit(Faulty, &tmp->rdev->flags)
1037                     && !test_bit(In_sync, &tmp->rdev->flags)) {
1038                         conf->working_disks++;
1039                         mddev->degraded--;
1040                         set_bit(In_sync, &tmp->rdev->flags);
1041                 }
1042         }
1043
1044         print_conf(conf);
1045         return 0;
1046 }
1047
1048
1049 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1050 {
1051         conf_t *conf = mddev->private;
1052         int found = 0;
1053         int mirror;
1054         mirror_info_t *p;
1055
1056         if (mddev->recovery_cp < MaxSector)
1057                 /* only hot-add to in-sync arrays, as recovery is
1058                  * very different from resync
1059                  */
1060                 return 0;
1061         if (!enough(conf))
1062                 return 0;
1063
1064         if (rdev->saved_raid_disk >= 0 &&
1065             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1066                 mirror = rdev->saved_raid_disk;
1067         else
1068                 mirror = 0;
1069         for ( ; mirror < mddev->raid_disks; mirror++)
1070                 if ( !(p=conf->mirrors+mirror)->rdev) {
1071
1072                         blk_queue_stack_limits(mddev->queue,
1073                                                rdev->bdev->bd_disk->queue);
1074                         /* as we don't honour merge_bvec_fn, we must never risk
1075                          * violating it, so limit ->max_sector to one PAGE, as
1076                          * a one page request is never in violation.
1077                          */
1078                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1079                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1080                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1081
1082                         p->head_position = 0;
1083                         rdev->raid_disk = mirror;
1084                         found = 1;
1085                         if (rdev->saved_raid_disk != mirror)
1086                                 conf->fullsync = 1;
1087                         rcu_assign_pointer(p->rdev, rdev);
1088                         break;
1089                 }
1090
1091         print_conf(conf);
1092         return found;
1093 }
1094
1095 static int raid10_remove_disk(mddev_t *mddev, int number)
1096 {
1097         conf_t *conf = mddev->private;
1098         int err = 0;
1099         mdk_rdev_t *rdev;
1100         mirror_info_t *p = conf->mirrors+ number;
1101
1102         print_conf(conf);
1103         rdev = p->rdev;
1104         if (rdev) {
1105                 if (test_bit(In_sync, &rdev->flags) ||
1106                     atomic_read(&rdev->nr_pending)) {
1107                         err = -EBUSY;
1108                         goto abort;
1109                 }
1110                 p->rdev = NULL;
1111                 synchronize_rcu();
1112                 if (atomic_read(&rdev->nr_pending)) {
1113                         /* lost the race, try later */
1114                         err = -EBUSY;
1115                         p->rdev = rdev;
1116                 }
1117         }
1118 abort:
1119
1120         print_conf(conf);
1121         return err;
1122 }
1123
1124
1125 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1126 {
1127         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1128         conf_t *conf = mddev_to_conf(r10_bio->mddev);
1129         int i,d;
1130
1131         if (bio->bi_size)
1132                 return 1;
1133
1134         for (i=0; i<conf->copies; i++)
1135                 if (r10_bio->devs[i].bio == bio)
1136                         break;
1137         BUG_ON(i == conf->copies);
1138         update_head_pos(i, r10_bio);
1139         d = r10_bio->devs[i].devnum;
1140
1141         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1142                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1143         else {
1144                 atomic_add(r10_bio->sectors,
1145                            &conf->mirrors[d].rdev->corrected_errors);
1146                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1147                         md_error(r10_bio->mddev,
1148                                  conf->mirrors[d].rdev);
1149         }
1150
1151         /* for reconstruct, we always reschedule after a read.
1152          * for resync, only after all reads
1153          */
1154         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1155             atomic_dec_and_test(&r10_bio->remaining)) {
1156                 /* we have read all the blocks,
1157                  * do the comparison in process context in raid10d
1158                  */
1159                 reschedule_retry(r10_bio);
1160         }
1161         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1162         return 0;
1163 }
1164
1165 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1166 {
1167         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1168         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1169         mddev_t *mddev = r10_bio->mddev;
1170         conf_t *conf = mddev_to_conf(mddev);
1171         int i,d;
1172
1173         if (bio->bi_size)
1174                 return 1;
1175
1176         for (i = 0; i < conf->copies; i++)
1177                 if (r10_bio->devs[i].bio == bio)
1178                         break;
1179         d = r10_bio->devs[i].devnum;
1180
1181         if (!uptodate)
1182                 md_error(mddev, conf->mirrors[d].rdev);
1183         update_head_pos(i, r10_bio);
1184
1185         while (atomic_dec_and_test(&r10_bio->remaining)) {
1186                 if (r10_bio->master_bio == NULL) {
1187                         /* the primary of several recovery bios */
1188                         md_done_sync(mddev, r10_bio->sectors, 1);
1189                         put_buf(r10_bio);
1190                         break;
1191                 } else {
1192                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1193                         put_buf(r10_bio);
1194                         r10_bio = r10_bio2;
1195                 }
1196         }
1197         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1198         return 0;
1199 }
1200
1201 /*
1202  * Note: sync and recover and handled very differently for raid10
1203  * This code is for resync.
1204  * For resync, we read through virtual addresses and read all blocks.
1205  * If there is any error, we schedule a write.  The lowest numbered
1206  * drive is authoritative.
1207  * However requests come for physical address, so we need to map.
1208  * For every physical address there are raid_disks/copies virtual addresses,
1209  * which is always are least one, but is not necessarly an integer.
1210  * This means that a physical address can span multiple chunks, so we may
1211  * have to submit multiple io requests for a single sync request.
1212  */
1213 /*
1214  * We check if all blocks are in-sync and only write to blocks that
1215  * aren't in sync
1216  */
1217 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1218 {
1219         conf_t *conf = mddev_to_conf(mddev);
1220         int i, first;
1221         struct bio *tbio, *fbio;
1222
1223         atomic_set(&r10_bio->remaining, 1);
1224
1225         /* find the first device with a block */
1226         for (i=0; i<conf->copies; i++)
1227                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1228                         break;
1229
1230         if (i == conf->copies)
1231                 goto done;
1232
1233         first = i;
1234         fbio = r10_bio->devs[i].bio;
1235
1236         /* now find blocks with errors */
1237         for (i=0 ; i < conf->copies ; i++) {
1238                 int  j, d;
1239                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1240
1241                 tbio = r10_bio->devs[i].bio;
1242
1243                 if (tbio->bi_end_io != end_sync_read)
1244                         continue;
1245                 if (i == first)
1246                         continue;
1247                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1248                         /* We know that the bi_io_vec layout is the same for
1249                          * both 'first' and 'i', so we just compare them.
1250                          * All vec entries are PAGE_SIZE;
1251                          */
1252                         for (j = 0; j < vcnt; j++)
1253                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1254                                            page_address(tbio->bi_io_vec[j].bv_page),
1255                                            PAGE_SIZE))
1256                                         break;
1257                         if (j == vcnt)
1258                                 continue;
1259                         mddev->resync_mismatches += r10_bio->sectors;
1260                 }
1261                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1262                         /* Don't fix anything. */
1263                         continue;
1264                 /* Ok, we need to write this bio
1265                  * First we need to fixup bv_offset, bv_len and
1266                  * bi_vecs, as the read request might have corrupted these
1267                  */
1268                 tbio->bi_vcnt = vcnt;
1269                 tbio->bi_size = r10_bio->sectors << 9;
1270                 tbio->bi_idx = 0;
1271                 tbio->bi_phys_segments = 0;
1272                 tbio->bi_hw_segments = 0;
1273                 tbio->bi_hw_front_size = 0;
1274                 tbio->bi_hw_back_size = 0;
1275                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1276                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1277                 tbio->bi_next = NULL;
1278                 tbio->bi_rw = WRITE;
1279                 tbio->bi_private = r10_bio;
1280                 tbio->bi_sector = r10_bio->devs[i].addr;
1281
1282                 for (j=0; j < vcnt ; j++) {
1283                         tbio->bi_io_vec[j].bv_offset = 0;
1284                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1285
1286                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1287                                page_address(fbio->bi_io_vec[j].bv_page),
1288                                PAGE_SIZE);
1289                 }
1290                 tbio->bi_end_io = end_sync_write;
1291
1292                 d = r10_bio->devs[i].devnum;
1293                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1294                 atomic_inc(&r10_bio->remaining);
1295                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1296
1297                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1298                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1299                 generic_make_request(tbio);
1300         }
1301
1302 done:
1303         if (atomic_dec_and_test(&r10_bio->remaining)) {
1304                 md_done_sync(mddev, r10_bio->sectors, 1);
1305                 put_buf(r10_bio);
1306         }
1307 }
1308
1309 /*
1310  * Now for the recovery code.
1311  * Recovery happens across physical sectors.
1312  * We recover all non-is_sync drives by finding the virtual address of
1313  * each, and then choose a working drive that also has that virt address.
1314  * There is a separate r10_bio for each non-in_sync drive.
1315  * Only the first two slots are in use. The first for reading,
1316  * The second for writing.
1317  *
1318  */
1319
1320 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1321 {
1322         conf_t *conf = mddev_to_conf(mddev);
1323         int i, d;
1324         struct bio *bio, *wbio;
1325
1326
1327         /* move the pages across to the second bio
1328          * and submit the write request
1329          */
1330         bio = r10_bio->devs[0].bio;
1331         wbio = r10_bio->devs[1].bio;
1332         for (i=0; i < wbio->bi_vcnt; i++) {
1333                 struct page *p = bio->bi_io_vec[i].bv_page;
1334                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1335                 wbio->bi_io_vec[i].bv_page = p;
1336         }
1337         d = r10_bio->devs[1].devnum;
1338
1339         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1340         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1341         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1342                 generic_make_request(wbio);
1343         else
1344                 bio_endio(wbio, wbio->bi_size, -EIO);
1345 }
1346
1347
1348 /*
1349  * This is a kernel thread which:
1350  *
1351  *      1.      Retries failed read operations on working mirrors.
1352  *      2.      Updates the raid superblock when problems encounter.
1353  *      3.      Performs writes following reads for array syncronising.
1354  */
1355
1356 static void raid10d(mddev_t *mddev)
1357 {
1358         r10bio_t *r10_bio;
1359         struct bio *bio;
1360         unsigned long flags;
1361         conf_t *conf = mddev_to_conf(mddev);
1362         struct list_head *head = &conf->retry_list;
1363         int unplug=0;
1364         mdk_rdev_t *rdev;
1365
1366         md_check_recovery(mddev);
1367
1368         for (;;) {
1369                 char b[BDEVNAME_SIZE];
1370                 spin_lock_irqsave(&conf->device_lock, flags);
1371
1372                 if (conf->pending_bio_list.head) {
1373                         bio = bio_list_get(&conf->pending_bio_list);
1374                         blk_remove_plug(mddev->queue);
1375                         spin_unlock_irqrestore(&conf->device_lock, flags);
1376                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1377                         if (bitmap_unplug(mddev->bitmap) != 0)
1378                                 printk("%s: bitmap file write failed!\n", mdname(mddev));
1379
1380                         while (bio) { /* submit pending writes */
1381                                 struct bio *next = bio->bi_next;
1382                                 bio->bi_next = NULL;
1383                                 generic_make_request(bio);
1384                                 bio = next;
1385                         }
1386                         unplug = 1;
1387
1388                         continue;
1389                 }
1390
1391                 if (list_empty(head))
1392                         break;
1393                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1394                 list_del(head->prev);
1395                 conf->nr_queued--;
1396                 spin_unlock_irqrestore(&conf->device_lock, flags);
1397
1398                 mddev = r10_bio->mddev;
1399                 conf = mddev_to_conf(mddev);
1400                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1401                         sync_request_write(mddev, r10_bio);
1402                         unplug = 1;
1403                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1404                         recovery_request_write(mddev, r10_bio);
1405                         unplug = 1;
1406                 } else {
1407                         int mirror;
1408                         /* we got a read error. Maybe the drive is bad.  Maybe just
1409                          * the block and we can fix it.
1410                          * We freeze all other IO, and try reading the block from
1411                          * other devices.  When we find one, we re-write
1412                          * and check it that fixes the read error.
1413                          * This is all done synchronously while the array is
1414                          * frozen.
1415                          */
1416                         int sect = 0; /* Offset from r10_bio->sector */
1417                         int sectors = r10_bio->sectors;
1418                         freeze_array(conf);
1419                         if (mddev->ro == 0) while(sectors) {
1420                                 int s = sectors;
1421                                 int sl = r10_bio->read_slot;
1422                                 int success = 0;
1423
1424                                 if (s > (PAGE_SIZE>>9))
1425                                         s = PAGE_SIZE >> 9;
1426
1427                                 rcu_read_lock();
1428                                 do {
1429                                         int d = r10_bio->devs[sl].devnum;
1430                                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1431                                         if (rdev &&
1432                                             test_bit(In_sync, &rdev->flags)) {
1433                                                 atomic_inc(&rdev->nr_pending);
1434                                                 rcu_read_unlock();
1435                                                 success = sync_page_io(rdev->bdev,
1436                                                                        r10_bio->devs[sl].addr +
1437                                                                        sect + rdev->data_offset,
1438                                                                        s<<9,
1439                                                                        conf->tmppage, READ);
1440                                                 rdev_dec_pending(rdev, mddev);
1441                                                 rcu_read_lock();
1442                                                 if (success)
1443                                                         break;
1444                                         }
1445                                         sl++;
1446                                         if (sl == conf->copies)
1447                                                 sl = 0;
1448                                 } while (!success && sl != r10_bio->read_slot);
1449                                 rcu_read_unlock();
1450
1451                                 if (success) {
1452                                         int start = sl;
1453                                         /* write it back and re-read */
1454                                         rcu_read_lock();
1455                                         while (sl != r10_bio->read_slot) {
1456                                                 int d;
1457                                                 if (sl==0)
1458                                                         sl = conf->copies;
1459                                                 sl--;
1460                                                 d = r10_bio->devs[sl].devnum;
1461                                                 rdev = rcu_dereference(conf->mirrors[d].rdev);
1462                                                 if (rdev &&
1463                                                     test_bit(In_sync, &rdev->flags)) {
1464                                                         atomic_inc(&rdev->nr_pending);
1465                                                         rcu_read_unlock();
1466                                                         atomic_add(s, &rdev->corrected_errors);
1467                                                         if (sync_page_io(rdev->bdev,
1468                                                                          r10_bio->devs[sl].addr +
1469                                                                          sect + rdev->data_offset,
1470                                                                          s<<9, conf->tmppage, WRITE) == 0)
1471                                                                 /* Well, this device is dead */
1472                                                                 md_error(mddev, rdev);
1473                                                         rdev_dec_pending(rdev, mddev);
1474                                                         rcu_read_lock();
1475                                                 }
1476                                         }
1477                                         sl = start;
1478                                         while (sl != r10_bio->read_slot) {
1479                                                 int d;
1480                                                 if (sl==0)
1481                                                         sl = conf->copies;
1482                                                 sl--;
1483                                                 d = r10_bio->devs[sl].devnum;
1484                                                 rdev = rcu_dereference(conf->mirrors[d].rdev);
1485                                                 if (rdev &&
1486                                                     test_bit(In_sync, &rdev->flags)) {
1487                                                         atomic_inc(&rdev->nr_pending);
1488                                                         rcu_read_unlock();
1489                                                         if (sync_page_io(rdev->bdev,
1490                                                                          r10_bio->devs[sl].addr +
1491                                                                          sect + rdev->data_offset,
1492                                                                          s<<9, conf->tmppage, READ) == 0)
1493                                                                 /* Well, this device is dead */
1494                                                                 md_error(mddev, rdev);
1495                                                         rdev_dec_pending(rdev, mddev);
1496                                                         rcu_read_lock();
1497                                                 }
1498                                         }
1499                                         rcu_read_unlock();
1500                                 } else {
1501                                         /* Cannot read from anywhere -- bye bye array */
1502                                         md_error(mddev, conf->mirrors[r10_bio->devs[r10_bio->read_slot].devnum].rdev);
1503                                         break;
1504                                 }
1505                                 sectors -= s;
1506                                 sect += s;
1507                         }
1508
1509                         unfreeze_array(conf);
1510
1511                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1512                         r10_bio->devs[r10_bio->read_slot].bio =
1513                                 mddev->ro ? IO_BLOCKED : NULL;
1514                         bio_put(bio);
1515                         mirror = read_balance(conf, r10_bio);
1516                         if (mirror == -1) {
1517                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1518                                        " read error for block %llu\n",
1519                                        bdevname(bio->bi_bdev,b),
1520                                        (unsigned long long)r10_bio->sector);
1521                                 raid_end_bio_io(r10_bio);
1522                         } else {
1523                                 rdev = conf->mirrors[mirror].rdev;
1524                                 if (printk_ratelimit())
1525                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1526                                                " another mirror\n",
1527                                                bdevname(rdev->bdev,b),
1528                                                (unsigned long long)r10_bio->sector);
1529                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1530                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1531                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1532                                         + rdev->data_offset;
1533                                 bio->bi_bdev = rdev->bdev;
1534                                 bio->bi_rw = READ;
1535                                 bio->bi_private = r10_bio;
1536                                 bio->bi_end_io = raid10_end_read_request;
1537                                 unplug = 1;
1538                                 generic_make_request(bio);
1539                         }
1540                 }
1541         }
1542         spin_unlock_irqrestore(&conf->device_lock, flags);
1543         if (unplug)
1544                 unplug_slaves(mddev);
1545 }
1546
1547
1548 static int init_resync(conf_t *conf)
1549 {
1550         int buffs;
1551
1552         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1553         BUG_ON(conf->r10buf_pool);
1554         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1555         if (!conf->r10buf_pool)
1556                 return -ENOMEM;
1557         conf->next_resync = 0;
1558         return 0;
1559 }
1560
1561 /*
1562  * perform a "sync" on one "block"
1563  *
1564  * We need to make sure that no normal I/O request - particularly write
1565  * requests - conflict with active sync requests.
1566  *
1567  * This is achieved by tracking pending requests and a 'barrier' concept
1568  * that can be installed to exclude normal IO requests.
1569  *
1570  * Resync and recovery are handled very differently.
1571  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1572  *
1573  * For resync, we iterate over virtual addresses, read all copies,
1574  * and update if there are differences.  If only one copy is live,
1575  * skip it.
1576  * For recovery, we iterate over physical addresses, read a good
1577  * value for each non-in_sync drive, and over-write.
1578  *
1579  * So, for recovery we may have several outstanding complex requests for a
1580  * given address, one for each out-of-sync device.  We model this by allocating
1581  * a number of r10_bio structures, one for each out-of-sync device.
1582  * As we setup these structures, we collect all bio's together into a list
1583  * which we then process collectively to add pages, and then process again
1584  * to pass to generic_make_request.
1585  *
1586  * The r10_bio structures are linked using a borrowed master_bio pointer.
1587  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1588  * has its remaining count decremented to 0, the whole complex operation
1589  * is complete.
1590  *
1591  */
1592
1593 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1594 {
1595         conf_t *conf = mddev_to_conf(mddev);
1596         r10bio_t *r10_bio;
1597         struct bio *biolist = NULL, *bio;
1598         sector_t max_sector, nr_sectors;
1599         int disk;
1600         int i;
1601         int max_sync;
1602         int sync_blocks;
1603
1604         sector_t sectors_skipped = 0;
1605         int chunks_skipped = 0;
1606
1607         if (!conf->r10buf_pool)
1608                 if (init_resync(conf))
1609                         return 0;
1610
1611  skipped:
1612         max_sector = mddev->size << 1;
1613         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1614                 max_sector = mddev->resync_max_sectors;
1615         if (sector_nr >= max_sector) {
1616                 /* If we aborted, we need to abort the
1617                  * sync on the 'current' bitmap chucks (there can
1618                  * be several when recovering multiple devices).
1619                  * as we may have started syncing it but not finished.
1620                  * We can find the current address in
1621                  * mddev->curr_resync, but for recovery,
1622                  * we need to convert that to several
1623                  * virtual addresses.
1624                  */
1625                 if (mddev->curr_resync < max_sector) { /* aborted */
1626                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1627                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1628                                                 &sync_blocks, 1);
1629                         else for (i=0; i<conf->raid_disks; i++) {
1630                                 sector_t sect =
1631                                         raid10_find_virt(conf, mddev->curr_resync, i);
1632                                 bitmap_end_sync(mddev->bitmap, sect,
1633                                                 &sync_blocks, 1);
1634                         }
1635                 } else /* completed sync */
1636                         conf->fullsync = 0;
1637
1638                 bitmap_close_sync(mddev->bitmap);
1639                 close_sync(conf);
1640                 *skipped = 1;
1641                 return sectors_skipped;
1642         }
1643         if (chunks_skipped >= conf->raid_disks) {
1644                 /* if there has been nothing to do on any drive,
1645                  * then there is nothing to do at all..
1646                  */
1647                 *skipped = 1;
1648                 return (max_sector - sector_nr) + sectors_skipped;
1649         }
1650
1651         /* make sure whole request will fit in a chunk - if chunks
1652          * are meaningful
1653          */
1654         if (conf->near_copies < conf->raid_disks &&
1655             max_sector > (sector_nr | conf->chunk_mask))
1656                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1657         /*
1658          * If there is non-resync activity waiting for us then
1659          * put in a delay to throttle resync.
1660          */
1661         if (!go_faster && conf->nr_waiting)
1662                 msleep_interruptible(1000);
1663
1664         /* Again, very different code for resync and recovery.
1665          * Both must result in an r10bio with a list of bios that
1666          * have bi_end_io, bi_sector, bi_bdev set,
1667          * and bi_private set to the r10bio.
1668          * For recovery, we may actually create several r10bios
1669          * with 2 bios in each, that correspond to the bios in the main one.
1670          * In this case, the subordinate r10bios link back through a
1671          * borrowed master_bio pointer, and the counter in the master
1672          * includes a ref from each subordinate.
1673          */
1674         /* First, we decide what to do and set ->bi_end_io
1675          * To end_sync_read if we want to read, and
1676          * end_sync_write if we will want to write.
1677          */
1678
1679         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1680         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1681                 /* recovery... the complicated one */
1682                 int i, j, k;
1683                 r10_bio = NULL;
1684
1685                 for (i=0 ; i<conf->raid_disks; i++)
1686                         if (conf->mirrors[i].rdev &&
1687                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1688                                 int still_degraded = 0;
1689                                 /* want to reconstruct this device */
1690                                 r10bio_t *rb2 = r10_bio;
1691                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1692                                 int must_sync;
1693                                 /* Unless we are doing a full sync, we only need
1694                                  * to recover the block if it is set in the bitmap
1695                                  */
1696                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1697                                                               &sync_blocks, 1);
1698                                 if (sync_blocks < max_sync)
1699                                         max_sync = sync_blocks;
1700                                 if (!must_sync &&
1701                                     !conf->fullsync) {
1702                                         /* yep, skip the sync_blocks here, but don't assume
1703                                          * that there will never be anything to do here
1704                                          */
1705                                         chunks_skipped = -1;
1706                                         continue;
1707                                 }
1708
1709                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1710                                 raise_barrier(conf, rb2 != NULL);
1711                                 atomic_set(&r10_bio->remaining, 0);
1712
1713                                 r10_bio->master_bio = (struct bio*)rb2;
1714                                 if (rb2)
1715                                         atomic_inc(&rb2->remaining);
1716                                 r10_bio->mddev = mddev;
1717                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1718                                 r10_bio->sector = sect;
1719
1720                                 raid10_find_phys(conf, r10_bio);
1721                                 /* Need to check if this section will still be
1722                                  * degraded
1723                                  */
1724                                 for (j=0; j<conf->copies;j++) {
1725                                         int d = r10_bio->devs[j].devnum;
1726                                         if (conf->mirrors[d].rdev == NULL ||
1727                                             test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
1728                                                 still_degraded = 1;
1729                                                 break;
1730                                         }
1731                                 }
1732                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1733                                                               &sync_blocks, still_degraded);
1734
1735                                 for (j=0; j<conf->copies;j++) {
1736                                         int d = r10_bio->devs[j].devnum;
1737                                         if (conf->mirrors[d].rdev &&
1738                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1739                                                 /* This is where we read from */
1740                                                 bio = r10_bio->devs[0].bio;
1741                                                 bio->bi_next = biolist;
1742                                                 biolist = bio;
1743                                                 bio->bi_private = r10_bio;
1744                                                 bio->bi_end_io = end_sync_read;
1745                                                 bio->bi_rw = 0;
1746                                                 bio->bi_sector = r10_bio->devs[j].addr +
1747                                                         conf->mirrors[d].rdev->data_offset;
1748                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1749                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1750                                                 atomic_inc(&r10_bio->remaining);
1751                                                 /* and we write to 'i' */
1752
1753                                                 for (k=0; k<conf->copies; k++)
1754                                                         if (r10_bio->devs[k].devnum == i)
1755                                                                 break;
1756                                                 bio = r10_bio->devs[1].bio;
1757                                                 bio->bi_next = biolist;
1758                                                 biolist = bio;
1759                                                 bio->bi_private = r10_bio;
1760                                                 bio->bi_end_io = end_sync_write;
1761                                                 bio->bi_rw = 1;
1762                                                 bio->bi_sector = r10_bio->devs[k].addr +
1763                                                         conf->mirrors[i].rdev->data_offset;
1764                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1765
1766                                                 r10_bio->devs[0].devnum = d;
1767                                                 r10_bio->devs[1].devnum = i;
1768
1769                                                 break;
1770                                         }
1771                                 }
1772                                 if (j == conf->copies) {
1773                                         /* Cannot recover, so abort the recovery */
1774                                         put_buf(r10_bio);
1775                                         r10_bio = rb2;
1776                                         if (!test_and_set_bit(MD_RECOVERY_ERR, &mddev->recovery))
1777                                                 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1778                                                        mdname(mddev));
1779                                         break;
1780                                 }
1781                         }
1782                 if (biolist == NULL) {
1783                         while (r10_bio) {
1784                                 r10bio_t *rb2 = r10_bio;
1785                                 r10_bio = (r10bio_t*) rb2->master_bio;
1786                                 rb2->master_bio = NULL;
1787                                 put_buf(rb2);
1788                         }
1789                         goto giveup;
1790                 }
1791         } else {
1792                 /* resync. Schedule a read for every block at this virt offset */
1793                 int count = 0;
1794
1795                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1796                                        &sync_blocks, mddev->degraded) &&
1797                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1798                         /* We can skip this block */
1799                         *skipped = 1;
1800                         return sync_blocks + sectors_skipped;
1801                 }
1802                 if (sync_blocks < max_sync)
1803                         max_sync = sync_blocks;
1804                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1805
1806                 r10_bio->mddev = mddev;
1807                 atomic_set(&r10_bio->remaining, 0);
1808                 raise_barrier(conf, 0);
1809                 conf->next_resync = sector_nr;
1810
1811                 r10_bio->master_bio = NULL;
1812                 r10_bio->sector = sector_nr;
1813                 set_bit(R10BIO_IsSync, &r10_bio->state);
1814                 raid10_find_phys(conf, r10_bio);
1815                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1816
1817                 for (i=0; i<conf->copies; i++) {
1818                         int d = r10_bio->devs[i].devnum;
1819                         bio = r10_bio->devs[i].bio;
1820                         bio->bi_end_io = NULL;
1821                         if (conf->mirrors[d].rdev == NULL ||
1822                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1823                                 continue;
1824                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1825                         atomic_inc(&r10_bio->remaining);
1826                         bio->bi_next = biolist;
1827                         biolist = bio;
1828                         bio->bi_private = r10_bio;
1829                         bio->bi_end_io = end_sync_read;
1830                         bio->bi_rw = 0;
1831                         bio->bi_sector = r10_bio->devs[i].addr +
1832                                 conf->mirrors[d].rdev->data_offset;
1833                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1834                         count++;
1835                 }
1836
1837                 if (count < 2) {
1838                         for (i=0; i<conf->copies; i++) {
1839                                 int d = r10_bio->devs[i].devnum;
1840                                 if (r10_bio->devs[i].bio->bi_end_io)
1841                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1842                         }
1843                         put_buf(r10_bio);
1844                         biolist = NULL;
1845                         goto giveup;
1846                 }
1847         }
1848
1849         for (bio = biolist; bio ; bio=bio->bi_next) {
1850
1851                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1852                 if (bio->bi_end_io)
1853                         bio->bi_flags |= 1 << BIO_UPTODATE;
1854                 bio->bi_vcnt = 0;
1855                 bio->bi_idx = 0;
1856                 bio->bi_phys_segments = 0;
1857                 bio->bi_hw_segments = 0;
1858                 bio->bi_size = 0;
1859         }
1860
1861         nr_sectors = 0;
1862         if (sector_nr + max_sync < max_sector)
1863                 max_sector = sector_nr + max_sync;
1864         do {
1865                 struct page *page;
1866                 int len = PAGE_SIZE;
1867                 disk = 0;
1868                 if (sector_nr + (len>>9) > max_sector)
1869                         len = (max_sector - sector_nr) << 9;
1870                 if (len == 0)
1871                         break;
1872                 for (bio= biolist ; bio ; bio=bio->bi_next) {
1873                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1874                         if (bio_add_page(bio, page, len, 0) == 0) {
1875                                 /* stop here */
1876                                 struct bio *bio2;
1877                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1878                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1879                                         /* remove last page from this bio */
1880                                         bio2->bi_vcnt--;
1881                                         bio2->bi_size -= len;
1882                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1883                                 }
1884                                 goto bio_full;
1885                         }
1886                         disk = i;
1887                 }
1888                 nr_sectors += len>>9;
1889                 sector_nr += len>>9;
1890         } while (biolist->bi_vcnt < RESYNC_PAGES);
1891  bio_full:
1892         r10_bio->sectors = nr_sectors;
1893
1894         while (biolist) {
1895                 bio = biolist;
1896                 biolist = biolist->bi_next;
1897
1898                 bio->bi_next = NULL;
1899                 r10_bio = bio->bi_private;
1900                 r10_bio->sectors = nr_sectors;
1901
1902                 if (bio->bi_end_io == end_sync_read) {
1903                         md_sync_acct(bio->bi_bdev, nr_sectors);
1904                         generic_make_request(bio);
1905                 }
1906         }
1907
1908         if (sectors_skipped)
1909                 /* pretend they weren't skipped, it makes
1910                  * no important difference in this case
1911                  */
1912                 md_done_sync(mddev, sectors_skipped, 1);
1913
1914         return sectors_skipped + nr_sectors;
1915  giveup:
1916         /* There is nowhere to write, so all non-sync
1917          * drives must be failed, so try the next chunk...
1918          */
1919         {
1920         sector_t sec = max_sector - sector_nr;
1921         sectors_skipped += sec;
1922         chunks_skipped ++;
1923         sector_nr = max_sector;
1924         goto skipped;
1925         }
1926 }
1927
1928 static int run(mddev_t *mddev)
1929 {
1930         conf_t *conf;
1931         int i, disk_idx;
1932         mirror_info_t *disk;
1933         mdk_rdev_t *rdev;
1934         struct list_head *tmp;
1935         int nc, fc, fo;
1936         sector_t stride, size;
1937
1938         if (mddev->chunk_size == 0) {
1939                 printk(KERN_ERR "md/raid10: non-zero chunk size required.\n");
1940                 return -EINVAL;
1941         }
1942
1943         nc = mddev->layout & 255;
1944         fc = (mddev->layout >> 8) & 255;
1945         fo = mddev->layout & (1<<16);
1946         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
1947             (mddev->layout >> 17)) {
1948                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
1949                        mdname(mddev), mddev->layout);
1950                 goto out;
1951         }
1952         /*
1953          * copy the already verified devices into our private RAID10
1954          * bookkeeping area. [whatever we allocate in run(),
1955          * should be freed in stop()]
1956          */
1957         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1958         mddev->private = conf;
1959         if (!conf) {
1960                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1961                         mdname(mddev));
1962                 goto out;
1963         }
1964         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1965                                  GFP_KERNEL);
1966         if (!conf->mirrors) {
1967                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1968                        mdname(mddev));
1969                 goto out_free_conf;
1970         }
1971
1972         conf->tmppage = alloc_page(GFP_KERNEL);
1973         if (!conf->tmppage)
1974                 goto out_free_conf;
1975
1976         conf->near_copies = nc;
1977         conf->far_copies = fc;
1978         conf->copies = nc*fc;
1979         conf->far_offset = fo;
1980         conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
1981         conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
1982         if (fo)
1983                 conf->stride = 1 << conf->chunk_shift;
1984         else {
1985                 stride = mddev->size >> (conf->chunk_shift-1);
1986                 sector_div(stride, fc);
1987                 conf->stride = stride << conf->chunk_shift;
1988         }
1989         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
1990                                                 r10bio_pool_free, conf);
1991         if (!conf->r10bio_pool) {
1992                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1993                         mdname(mddev));
1994                 goto out_free_conf;
1995         }
1996
1997         ITERATE_RDEV(mddev, rdev, tmp) {
1998                 disk_idx = rdev->raid_disk;
1999                 if (disk_idx >= mddev->raid_disks
2000                     || disk_idx < 0)
2001                         continue;
2002                 disk = conf->mirrors + disk_idx;
2003
2004                 disk->rdev = rdev;
2005
2006                 blk_queue_stack_limits(mddev->queue,
2007                                        rdev->bdev->bd_disk->queue);
2008                 /* as we don't honour merge_bvec_fn, we must never risk
2009                  * violating it, so limit ->max_sector to one PAGE, as
2010                  * a one page request is never in violation.
2011                  */
2012                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2013                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
2014                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
2015
2016                 disk->head_position = 0;
2017                 if (!test_bit(Faulty, &rdev->flags) && test_bit(In_sync, &rdev->flags))
2018                         conf->working_disks++;
2019         }
2020         conf->raid_disks = mddev->raid_disks;
2021         conf->mddev = mddev;
2022         spin_lock_init(&conf->device_lock);
2023         INIT_LIST_HEAD(&conf->retry_list);
2024
2025         spin_lock_init(&conf->resync_lock);
2026         init_waitqueue_head(&conf->wait_barrier);
2027
2028         /* need to check that every block has at least one working mirror */
2029         if (!enough(conf)) {
2030                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2031                        mdname(mddev));
2032                 goto out_free_conf;
2033         }
2034
2035         mddev->degraded = 0;
2036         for (i = 0; i < conf->raid_disks; i++) {
2037
2038                 disk = conf->mirrors + i;
2039
2040                 if (!disk->rdev ||
2041                     !test_bit(In_sync, &rdev->flags)) {
2042                         disk->head_position = 0;
2043                         mddev->degraded++;
2044                 }
2045         }
2046
2047
2048         mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2049         if (!mddev->thread) {
2050                 printk(KERN_ERR
2051                        "raid10: couldn't allocate thread for %s\n",
2052                        mdname(mddev));
2053                 goto out_free_conf;
2054         }
2055
2056         printk(KERN_INFO
2057                 "raid10: raid set %s active with %d out of %d devices\n",
2058                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2059                 mddev->raid_disks);
2060         /*
2061          * Ok, everything is just fine now
2062          */
2063         if (conf->far_offset) {
2064                 size = mddev->size >> (conf->chunk_shift-1);
2065                 size *= conf->raid_disks;
2066                 size <<= conf->chunk_shift;
2067                 sector_div(size, conf->far_copies);
2068         } else
2069                 size = conf->stride * conf->raid_disks;
2070         sector_div(size, conf->near_copies);
2071         mddev->array_size = size/2;
2072         mddev->resync_max_sectors = size;
2073
2074         mddev->queue->unplug_fn = raid10_unplug;
2075         mddev->queue->issue_flush_fn = raid10_issue_flush;
2076
2077         /* Calculate max read-ahead size.
2078          * We need to readahead at least twice a whole stripe....
2079          * maybe...
2080          */
2081         {
2082                 int stripe = conf->raid_disks * (mddev->chunk_size / PAGE_SIZE);
2083                 stripe /= conf->near_copies;
2084                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2085                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2086         }
2087
2088         if (conf->near_copies < mddev->raid_disks)
2089                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2090         return 0;
2091
2092 out_free_conf:
2093         if (conf->r10bio_pool)
2094                 mempool_destroy(conf->r10bio_pool);
2095         safe_put_page(conf->tmppage);
2096         kfree(conf->mirrors);
2097         kfree(conf);
2098         mddev->private = NULL;
2099 out:
2100         return -EIO;
2101 }
2102
2103 static int stop(mddev_t *mddev)
2104 {
2105         conf_t *conf = mddev_to_conf(mddev);
2106
2107         md_unregister_thread(mddev->thread);
2108         mddev->thread = NULL;
2109         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2110         if (conf->r10bio_pool)
2111                 mempool_destroy(conf->r10bio_pool);
2112         kfree(conf->mirrors);
2113         kfree(conf);
2114         mddev->private = NULL;
2115         return 0;
2116 }
2117
2118 static void raid10_quiesce(mddev_t *mddev, int state)
2119 {
2120         conf_t *conf = mddev_to_conf(mddev);
2121
2122         switch(state) {
2123         case 1:
2124                 raise_barrier(conf, 0);
2125                 break;
2126         case 0:
2127                 lower_barrier(conf);
2128                 break;
2129         }
2130         if (mddev->thread) {
2131                 if (mddev->bitmap)
2132                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2133                 else
2134                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2135                 md_wakeup_thread(mddev->thread);
2136         }
2137 }
2138
2139 static struct mdk_personality raid10_personality =
2140 {
2141         .name           = "raid10",
2142         .level          = 10,
2143         .owner          = THIS_MODULE,
2144         .make_request   = make_request,
2145         .run            = run,
2146         .stop           = stop,
2147         .status         = status,
2148         .error_handler  = error,
2149         .hot_add_disk   = raid10_add_disk,
2150         .hot_remove_disk= raid10_remove_disk,
2151         .spare_active   = raid10_spare_active,
2152         .sync_request   = sync_request,
2153         .quiesce        = raid10_quiesce,
2154 };
2155
2156 static int __init raid_init(void)
2157 {
2158         return register_md_personality(&raid10_personality);
2159 }
2160
2161 static void raid_exit(void)
2162 {
2163         unregister_md_personality(&raid10_personality);
2164 }
2165
2166 module_init(raid_init);
2167 module_exit(raid_exit);
2168 MODULE_LICENSE("GPL");
2169 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2170 MODULE_ALIAS("md-raid10");
2171 MODULE_ALIAS("md-level-10");