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  *
  33  * The data to be stored is divided into chunks using chunksize.
  34  * Each device is divided into far_copies sections.
  35  * In each section, chunks are laid out in a style similar to raid0, but
  36  * near_copies copies of each chunk is stored (each on a different drive).
  37  * The starting device for each section is offset near_copies from the starting
  38  * device of the previous section.
  39  * Thus there are (near_copies*far_copies) of each chunk, and each is on a different
  40  * drive.
  41  * near_copies and far_copies must be at least one, and their product is at most
  42  * raid_disks.
  43  */
  44
  45 /*
  46  * Number of guaranteed r10bios in case of extreme VM load:
  47  */
  48 #define NR_RAID10_BIOS 256
  49
  50 static void unplug_slaves(mddev_t *mddev);
  51
  52 static void allow_barrier(conf_t *conf);
  53 static void lower_barrier(conf_t *conf);
  54
  55 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
  56 {
  57         conf_t *conf = data;
  58         r10bio_t *r10_bio;
  59         int size = offsetof(struct r10bio_s, devs[conf->copies]);
  60
  61         /* allocate a r10bio with room for raid_disks entries in the bios array */
  62         r10_bio = kzalloc(size, gfp_flags);
  63         if (!r10_bio)
  64                 unplug_slaves(conf->mddev);
  65
  66         return r10_bio;
  67 }
  68
  69 static void r10bio_pool_free(void *r10_bio, void *data)
  70 {
  71         kfree(r10_bio);
  72 }
  73
  74 #define RESYNC_BLOCK_SIZE (64*1024)
  75 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  76 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  77 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  78 #define RESYNC_WINDOW (2048*1024)
  79
  80 /*
  81  * When performing a resync, we need to read and compare, so
  82  * we need as many pages are there are copies.
  83  * When performing a recovery, we need 2 bios, one for read,
  84  * one for write (we recover only one drive per r10buf)
  85  *
  86  */
  87 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
  88 {
  89         conf_t *conf = data;
  90         struct page *page;
  91         r10bio_t *r10_bio;
  92         struct bio *bio;
  93         int i, j;
  94         int nalloc;
  95
  96         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
  97         if (!r10_bio) {
  98                 unplug_slaves(conf->mddev);
  99                 return NULL;
 100         }
 101
 102         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
 103                 nalloc = conf->copies; /* resync */
 104         else
 105                 nalloc = 2; /* recovery */
 106
 107         /*
 108          * Allocate bios.
 109          */
 110         for (j = nalloc ; j-- ; ) {
 111                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 112                 if (!bio)
 113                         goto out_free_bio;
 114                 r10_bio->devs[j].bio = bio;
 115         }
 116         /*
 117          * Allocate RESYNC_PAGES data pages and attach them
 118          * where needed.
 119          */
 120         for (j = 0 ; j < nalloc; j++) {
 121                 bio = r10_bio->devs[j].bio;
 122                 for (i = 0; i < RESYNC_PAGES; i++) {
 123                         page = alloc_page(gfp_flags);
 124                         if (unlikely(!page))
 125                                 goto out_free_pages;
 126
 127                         bio->bi_io_vec[i].bv_page = page;
 128                 }
 129         }
 130
 131         return r10_bio;
 132
 133 out_free_pages:
 134         for ( ; i > 0 ; i--)
 135                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
 136         while (j--)
 137                 for (i = 0; i < RESYNC_PAGES ; i++)
 138                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 139         j = -1;
 140 out_free_bio:
 141         while ( ++j < nalloc )
 142                 bio_put(r10_bio->devs[j].bio);
 143         r10bio_pool_free(r10_bio, conf);
 144         return NULL;
 145 }
 146
 147 static void r10buf_pool_free(void *__r10_bio, void *data)
 148 {
 149         int i;
 150         conf_t *conf = data;
 151         r10bio_t *r10bio = __r10_bio;
 152         int j;
 153
 154         for (j=0; j < conf->copies; j++) {
 155                 struct bio *bio = r10bio->devs[j].bio;
 156                 if (bio) {
 157                         for (i = 0; i < RESYNC_PAGES; i++) {
 158                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 159                                 bio->bi_io_vec[i].bv_page = NULL;
 160                         }
 161                         bio_put(bio);
 162                 }
 163         }
 164         r10bio_pool_free(r10bio, conf);
 165 }
 166
 167 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
 168 {
 169         int i;
 170
 171         for (i = 0; i < conf->copies; i++) {
 172                 struct bio **bio = & r10_bio->devs[i].bio;
 173                 if (*bio && *bio != IO_BLOCKED)
 174                         bio_put(*bio);
 175                 *bio = NULL;
 176         }
 177 }
 178
 179 static void free_r10bio(r10bio_t *r10_bio)
 180 {
 181         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 182
 183         /*
 184          * Wake up any possible resync thread that waits for the device
 185          * to go idle.
 186          */
 187         allow_barrier(conf);
 188
 189         put_all_bios(conf, r10_bio);
 190         mempool_free(r10_bio, conf->r10bio_pool);
 191 }
 192
 193 static void put_buf(r10bio_t *r10_bio)
 194 {
 195         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 196
 197         mempool_free(r10_bio, conf->r10buf_pool);
 198
 199         lower_barrier(conf);
 200 }
 201
 202 static void reschedule_retry(r10bio_t *r10_bio)
 203 {
 204         unsigned long flags;
 205         mddev_t *mddev = r10_bio->mddev;
 206         conf_t *conf = mddev_to_conf(mddev);
 207
 208         spin_lock_irqsave(&conf->device_lock, flags);
 209         list_add(&r10_bio->retry_list, &conf->retry_list);
 210         conf->nr_queued ++;
 211         spin_unlock_irqrestore(&conf->device_lock, flags);
 212
 213         md_wakeup_thread(mddev->thread);
 214 }
 215
 216 /*
 217  * raid_end_bio_io() is called when we have finished servicing a mirrored
 218  * operation and are ready to return a success/failure code to the buffer
 219  * cache layer.
 220  */
 221 static void raid_end_bio_io(r10bio_t *r10_bio)
 222 {
 223         struct bio *bio = r10_bio->master_bio;
 224
 225         bio_endio(bio, bio->bi_size,
 226                 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
 227         free_r10bio(r10_bio);
 228 }
 229
 230 /*
 231  * Update disk head position estimator based on IRQ completion info.
 232  */
 233 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
 234 {
 235         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 236
 237         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 238                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 239 }
 240
 241 static int raid10_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
 242 {
 243         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 244         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
 245         int slot, dev;
 246         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 247
 248         if (bio->bi_size)
 249                 return 1;
 250
 251         slot = r10_bio->read_slot;
 252         dev = r10_bio->devs[slot].devnum;
 253         /*
 254          * this branch is our 'one mirror IO has finished' event handler:
 255          */
 256         update_head_pos(slot, r10_bio);
 257
 258         if (uptodate) {
 259                 /*
 260                  * Set R10BIO_Uptodate in our master bio, so that
 261                  * we will return a good error code to the higher
 262                  * levels even if IO on some other mirrored buffer fails.
 263                  *
 264                  * The 'master' represents the composite IO operation to
 265                  * user-side. So if something waits for IO, then it will
 266                  * wait for the 'master' bio.
 267                  */
 268                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 269                 raid_end_bio_io(r10_bio);
 270         } else {
 271                 /*
 272                  * oops, read error:
 273                  */
 274                 char b[BDEVNAME_SIZE];
 275                 if (printk_ratelimit())
 276                         printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
 277                                bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
 278                 reschedule_retry(r10_bio);
 279         }
 280
 281         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 282         return 0;
 283 }
 284
 285 static int raid10_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
 286 {
 287         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 288         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
 289         int slot, dev;
 290         conf_t *conf = mddev_to_conf(r10_bio->mddev);
 291
 292         if (bio->bi_size)
 293                 return 1;
 294
 295         for (slot = 0; slot < conf->copies; slot++)
 296                 if (r10_bio->devs[slot].bio == bio)
 297                         break;
 298         dev = r10_bio->devs[slot].devnum;
 299
 300         /*
 301          * this branch is our 'one mirror IO has finished' event handler:
 302          */
 303         if (!uptodate) {
 304                 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
 305                 /* an I/O failed, we can't clear the bitmap */
 306                 set_bit(R10BIO_Degraded, &r10_bio->state);
 307         } else
 308                 /*
 309                  * Set R10BIO_Uptodate in our master bio, so that
 310                  * we will return a good error code for to the higher
 311                  * levels even if IO on some other mirrored buffer fails.
 312                  *
 313                  * The 'master' represents the composite IO operation to
 314                  * user-side. So if something waits for IO, then it will
 315                  * wait for the 'master' bio.
 316                  */
 317                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 318
 319         update_head_pos(slot, r10_bio);
 320
 321         /*
 322          *
 323          * Let's see if all mirrored write operations have finished
 324          * already.
 325          */
 326         if (atomic_dec_and_test(&r10_bio->remaining)) {
 327                 /* clear the bitmap if all writes complete successfully */
 328                 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 329                                 r10_bio->sectors,
 330                                 !test_bit(R10BIO_Degraded, &r10_bio->state),
 331                                 0);
 332                 md_write_end(r10_bio->mddev);
 333                 raid_end_bio_io(r10_bio);
 334         }
 335
 336         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 337         return 0;
 338 }
 339
 340
 341 /*
 342  * RAID10 layout manager
 343  * Aswell as the chunksize and raid_disks count, there are two
 344  * parameters: near_copies and far_copies.
 345  * near_copies * far_copies must be <= raid_disks.
 346  * Normally one of these will be 1.
 347  * If both are 1, we get raid0.
 348  * If near_copies == raid_disks, we get raid1.
 349  *
 350  * Chunks are layed out in raid0 style with near_copies copies of the
 351  * first chunk, followed by near_copies copies of the next chunk and
 352  * so on.
 353  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 354  * as described above, we start again with a device offset of near_copies.
 355  * So we effectively have another copy of the whole array further down all
 356  * the drives, but with blocks on different drives.
 357  * With this layout, and block is never stored twice on the one device.
 358  *
 359  * raid10_find_phys finds the sector offset of a given virtual sector
 360  * on each device that it is on. If a block isn't on a device,
 361  * that entry in the array is set to MaxSector.
 362  *
 363  * raid10_find_virt does the reverse mapping, from a device and a
 364  * sector offset to a virtual address
 365  */
 366
 367 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
 368 {
 369         int n,f;
 370         sector_t sector;
 371         sector_t chunk;
 372         sector_t stripe;
 373         int dev;
 374
 375         int slot = 0;
 376
 377         /* now calculate first sector/dev */
 378         chunk = r10bio->sector >> conf->chunk_shift;
 379         sector = r10bio->sector & conf->chunk_mask;
 380
 381         chunk *= conf->near_copies;
 382         stripe = chunk;
 383         dev = sector_div(stripe, conf->raid_disks);
 384
 385         sector += stripe << conf->chunk_shift;
 386
 387         /* and calculate all the others */
 388         for (n=0; n < conf->near_copies; n++) {
 389                 int d = dev;
 390                 sector_t s = sector;
 391                 r10bio->devs[slot].addr = sector;
 392                 r10bio->devs[slot].devnum = d;
 393                 slot++;
 394
 395                 for (f = 1; f < conf->far_copies; f++) {
 396                         d += conf->near_copies;
 397                         if (d >= conf->raid_disks)
 398                                 d -= conf->raid_disks;
 399                         s += conf->stride;
 400                         r10bio->devs[slot].devnum = d;
 401                         r10bio->devs[slot].addr = s;
 402                         slot++;
 403                 }
 404                 dev++;
 405                 if (dev >= conf->raid_disks) {
 406                         dev = 0;
 407                         sector += (conf->chunk_mask + 1);
 408                 }
 409         }
 410         BUG_ON(slot != conf->copies);
 411 }
 412
 413 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
 414 {
 415         sector_t offset, chunk, vchunk;
 416
 417         while (sector > conf->stride) {
 418                 sector -= conf->stride;
 419                 if (dev < conf->near_copies)
 420                         dev += conf->raid_disks - conf->near_copies;
 421                 else
 422                         dev -= conf->near_copies;
 423         }
 424
 425         offset = sector & conf->chunk_mask;
 426         chunk = sector >> conf->chunk_shift;
 427         vchunk = chunk * conf->raid_disks + dev;
 428         sector_div(vchunk, conf->near_copies);
 429         return (vchunk << conf->chunk_shift) + offset;
 430 }
 431
 432 /**
 433  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 434  *      @q: request queue
 435  *      @bio: the buffer head that's been built up so far
 436  *      @biovec: the request that could be merged to it.
 437  *
 438  *      Return amount of bytes we can accept at this offset
 439  *      If near_copies == raid_disk, there are no striping issues,
 440  *      but in that case, the function isn't called at all.
 441  */
 442 static int raid10_mergeable_bvec(request_queue_t *q, struct bio *bio,
 443                                 struct bio_vec *bio_vec)
 444 {
 445         mddev_t *mddev = q->queuedata;
 446         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
 447         int max;
 448         unsigned int chunk_sectors = mddev->chunk_size >> 9;
 449         unsigned int bio_sectors = bio->bi_size >> 9;
 450
 451         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 452         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 453         if (max <= bio_vec->bv_len && bio_sectors == 0)
 454                 return bio_vec->bv_len;
 455         else
 456                 return max;
 457 }
 458
 459 /*
 460  * This routine returns the disk from which the requested read should
 461  * be done. There is a per-array 'next expected sequential IO' sector
 462  * number - if this matches on the next IO then we use the last disk.
 463  * There is also a per-disk 'last know head position' sector that is
 464  * maintained from IRQ contexts, both the normal and the resync IO
 465  * completion handlers update this position correctly. If there is no
 466  * perfect sequential match then we pick the disk whose head is closest.
 467  *
 468  * If there are 2 mirrors in the same 2 devices, performance degrades
 469  * because position is mirror, not device based.
 470  *
 471  * The rdev for the device selected will have nr_pending incremented.
 472  */
 473
 474 /*
 475  * FIXME: possibly should rethink readbalancing and do it differently
 476  * depending on near_copies / far_copies geometry.
 477  */
 478 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
 479 {
 480         const unsigned long this_sector = r10_bio->sector;
 481         int disk, slot, nslot;
 482         const int sectors = r10_bio->sectors;
 483         sector_t new_distance, current_distance;
 484         mdk_rdev_t *rdev;
 485
 486         raid10_find_phys(conf, r10_bio);
 487         rcu_read_lock();
 488         /*
 489          * Check if we can balance. We can balance on the whole
 490          * device if no resync is going on (recovery is ok), or below
 491          * the resync window. We take the first readable disk when
 492          * above the resync window.
 493          */
 494         if (conf->mddev->recovery_cp < MaxSector
 495             && (this_sector + sectors >= conf->next_resync)) {
 496                 /* make sure that disk is operational */
 497                 slot = 0;
 498                 disk = r10_bio->devs[slot].devnum;
 499
 500                 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 501                        r10_bio->devs[slot].bio == IO_BLOCKED ||
 502                        !test_bit(In_sync, &rdev->flags)) {
 503                         slot++;
 504                         if (slot == conf->copies) {
 505                                 slot = 0;
 506                                 disk = -1;
 507                                 break;
 508                         }
 509                         disk = r10_bio->devs[slot].devnum;
 510                 }
 511                 goto rb_out;
 512         }
 513
 514
 515         /* make sure the disk is operational */
 516         slot = 0;
 517         disk = r10_bio->devs[slot].devnum;
 518         while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 519                r10_bio->devs[slot].bio == IO_BLOCKED ||
 520                !test_bit(In_sync, &rdev->flags)) {
 521                 slot ++;
 522                 if (slot == conf->copies) {
 523                         disk = -1;
 524                         goto rb_out;
 525                 }
 526                 disk = r10_bio->devs[slot].devnum;
 527         }
 528
 529
 530         current_distance = abs(r10_bio->devs[slot].addr -
 531                                conf->mirrors[disk].head_position);
 532
 533         /* Find the disk whose head is closest */
 534
 535         for (nslot = slot; nslot < conf->copies; nslot++) {
 536                 int ndisk = r10_bio->devs[nslot].devnum;
 537
 538
 539                 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
 540                     r10_bio->devs[nslot].bio == IO_BLOCKED ||
 541                     !test_bit(In_sync, &rdev->flags))
 542                         continue;
 543
 544                 /* This optimisation is debatable, and completely destroys
 545                  * sequential read speed for 'far copies' arrays.  So only
 546                  * keep it for 'near' arrays, and review those later.
 547                  */
 548                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
 549                         disk = ndisk;
 550                         slot = nslot;
 551                         break;
 552                 }
 553                 new_distance = abs(r10_bio->devs[nslot].addr -
 554                                    conf->mirrors[ndisk].head_position);
 555                 if (new_distance < current_distance) {
 556                         current_distance = new_distance;
 557                         disk = ndisk;
 558                         slot = nslot;
 559                 }
 560         }
 561
 562 rb_out:
 563         r10_bio->read_slot = slot;
 564 /*      conf->next_seq_sect = this_sector + sectors;*/
 565
 566         if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
 567                 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
 568         else
 569                 disk = -1;
 570         rcu_read_unlock();
 571
 572         return disk;
 573 }
 574
 575 static void unplug_slaves(mddev_t *mddev)
 576 {
 577         conf_t *conf = mddev_to_conf(mddev);
 578         int i;
 579
 580         rcu_read_lock();
 581         for (i=0; i<mddev->raid_disks; i++) {
 582                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 583                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 584                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 585
 586                         atomic_inc(&rdev->nr_pending);
 587                         rcu_read_unlock();
 588
 589                         if (r_queue->unplug_fn)
 590                                 r_queue->unplug_fn(r_queue);
 591
 592                         rdev_dec_pending(rdev, mddev);
 593                         rcu_read_lock();
 594                 }
 595         }
 596         rcu_read_unlock();
 597 }
 598
 599 static void raid10_unplug(request_queue_t *q)
 600 {
 601         mddev_t *mddev = q->queuedata;
 602
 603         unplug_slaves(q->queuedata);
 604         md_wakeup_thread(mddev->thread);
 605 }
 606
 607 static int raid10_issue_flush(request_queue_t *q, struct gendisk *disk,
 608                              sector_t *error_sector)
 609 {
 610         mddev_t *mddev = q->queuedata;
 611         conf_t *conf = mddev_to_conf(mddev);
 612         int i, ret = 0;
 613
 614         rcu_read_lock();
 615         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
 616                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 617                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 618                         struct block_device *bdev = rdev->bdev;
 619                         request_queue_t *r_queue = bdev_get_queue(bdev);
 620
 621                         if (!r_queue->issue_flush_fn)
 622                                 ret = -EOPNOTSUPP;
 623                         else {
 624                                 atomic_inc(&rdev->nr_pending);
 625                                 rcu_read_unlock();
 626                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
 627                                                               error_sector);
 628                                 rdev_dec_pending(rdev, mddev);
 629                                 rcu_read_lock();
 630                         }
 631                 }
 632         }
 633         rcu_read_unlock();
 634         return ret;
 635 }
 636
 637 /* Barriers....
 638  * Sometimes we need to suspend IO while we do something else,
 639  * either some resync/recovery, or reconfigure the array.
 640  * To do this we raise a 'barrier'.
 641  * The 'barrier' is a counter that can be raised multiple times
 642  * to count how many activities are happening which preclude
 643  * normal IO.
 644  * We can only raise the barrier if there is no pending IO.
 645  * i.e. if nr_pending == 0.
 646  * We choose only to raise the barrier if no-one is waiting for the
 647  * barrier to go down.  This means that as soon as an IO request
 648  * is ready, no other operations which require a barrier will start
 649  * until the IO request has had a chance.
 650  *
 651  * So: regular IO calls 'wait_barrier'.  When that returns there
 652  *    is no backgroup IO happening,  It must arrange to call
 653  *    allow_barrier when it has finished its IO.
 654  * backgroup IO calls must call raise_barrier.  Once that returns
 655  *    there is no normal IO happeing.  It must arrange to call
 656  *    lower_barrier when the particular background IO completes.
 657  */
 658 #define RESYNC_DEPTH 32
 659
 660 static void raise_barrier(conf_t *conf, int force)
 661 {
 662         BUG_ON(force && !conf->barrier);
 663         spin_lock_irq(&conf->resync_lock);
 664
 665         /* Wait until no block IO is waiting (unless 'force') */
 666         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 667                             conf->resync_lock,
 668                             raid10_unplug(conf->mddev->queue));
 669
 670         /* block any new IO from starting */
 671         conf->barrier++;
 672
 673         /* No wait for all pending IO to complete */
 674         wait_event_lock_irq(conf->wait_barrier,
 675                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 676                             conf->resync_lock,
 677                             raid10_unplug(conf->mddev->queue));
 678
 679         spin_unlock_irq(&conf->resync_lock);
 680 }
 681
 682 static void lower_barrier(conf_t *conf)
 683 {
 684         unsigned long flags;
 685         spin_lock_irqsave(&conf->resync_lock, flags);
 686         conf->barrier--;
 687         spin_unlock_irqrestore(&conf->resync_lock, flags);
 688         wake_up(&conf->wait_barrier);
 689 }
 690
 691 static void wait_barrier(conf_t *conf)
 692 {
 693         spin_lock_irq(&conf->resync_lock);
 694         if (conf->barrier) {
 695                 conf->nr_waiting++;
 696                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 697                                     conf->resync_lock,
 698                                     raid10_unplug(conf->mddev->queue));
 699                 conf->nr_waiting--;
 700         }
 701         conf->nr_pending++;
 702         spin_unlock_irq(&conf->resync_lock);
 703 }
 704
 705 static void allow_barrier(conf_t *conf)
 706 {
 707         unsigned long flags;
 708         spin_lock_irqsave(&conf->resync_lock, flags);
 709         conf->nr_pending--;
 710         spin_unlock_irqrestore(&conf->resync_lock, flags);
 711         wake_up(&conf->wait_barrier);
 712 }
 713
 714 static void freeze_array(conf_t *conf)
 715 {
 716         /* stop syncio and normal IO and wait for everything to
 717          * go quiet.
 718          * We increment barrier and nr_waiting, and then
 719          * wait until barrier+nr_pending match nr_queued+2
 720          */
 721         spin_lock_irq(&conf->resync_lock);
 722         conf->barrier++;
 723         conf->nr_waiting++;
 724         wait_event_lock_irq(conf->wait_barrier,
 725                             conf->barrier+conf->nr_pending == conf->nr_queued+2,
 726                             conf->resync_lock,
 727                             raid10_unplug(conf->mddev->queue));
 728         spin_unlock_irq(&conf->resync_lock);
 729 }
 730
 731 static void unfreeze_array(conf_t *conf)
 732 {
 733         /* reverse the effect of the freeze */
 734         spin_lock_irq(&conf->resync_lock);
 735         conf->barrier--;
 736         conf->nr_waiting--;
 737         wake_up(&conf->wait_barrier);
 738         spin_unlock_irq(&conf->resync_lock);
 739 }
 740
 741 static int make_request(request_queue_t *q, struct bio * bio)
 742 {
 743         mddev_t *mddev = q->queuedata;
 744         conf_t *conf = mddev_to_conf(mddev);
 745         mirror_info_t *mirror;
 746         r10bio_t *r10_bio;
 747         struct bio *read_bio;
 748         int i;
 749         int chunk_sects = conf->chunk_mask + 1;
 750         const int rw = bio_data_dir(bio);
 751         const int do_sync = bio_sync(bio);
 752         struct bio_list bl;
 753         unsigned long flags;
 754
 755         if (unlikely(bio_barrier(bio))) {
 756                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
 757                 return 0;
 758         }
 759
 760         /* If this request crosses a chunk boundary, we need to
 761          * split it.  This will only happen for 1 PAGE (or less) requests.
 762          */
 763         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 764                       > chunk_sects &&
 765                     conf->near_copies < conf->raid_disks)) {
 766                 struct bio_pair *bp;
 767                 /* Sanity check -- queue functions should prevent this happening */
 768                 if (bio->bi_vcnt != 1 ||
 769                     bio->bi_idx != 0)
 770                         goto bad_map;
 771                 /* This is a one page bio that upper layers
 772                  * refuse to split for us, so we need to split it.
 773                  */
 774                 bp = bio_split(bio, bio_split_pool,
 775                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 776                 if (make_request(q, &bp->bio1))
 777                         generic_make_request(&bp->bio1);
 778                 if (make_request(q, &bp->bio2))
 779                         generic_make_request(&bp->bio2);
 780
 781                 bio_pair_release(bp);
 782                 return 0;
 783         bad_map:
 784                 printk("raid10_make_request bug: can't convert block across chunks"
 785                        " or bigger than %dk %llu %d\n", chunk_sects/2,
 786                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 787
 788                 bio_io_error(bio, bio->bi_size);
 789                 return 0;
 790         }
 791
 792         md_write_start(mddev, bio);
 793
 794         /*
 795          * Register the new request and wait if the reconstruction
 796          * thread has put up a bar for new requests.
 797          * Continue immediately if no resync is active currently.
 798          */
 799         wait_barrier(conf);
 800
 801         disk_stat_inc(mddev->gendisk, ios[rw]);
 802         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
 803
 804         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 805
 806         r10_bio->master_bio = bio;
 807         r10_bio->sectors = bio->bi_size >> 9;
 808
 809         r10_bio->mddev = mddev;
 810         r10_bio->sector = bio->bi_sector;
 811         r10_bio->state = 0;
 812
 813         if (rw == READ) {
 814                 /*
 815                  * read balancing logic:
 816                  */
 817                 int disk = read_balance(conf, r10_bio);
 818                 int slot = r10_bio->read_slot;
 819                 if (disk < 0) {
 820                         raid_end_bio_io(r10_bio);
 821                         return 0;
 822                 }
 823                 mirror = conf->mirrors + disk;
 824
 825                 read_bio = bio_clone(bio, GFP_NOIO);
 826
 827                 r10_bio->devs[slot].bio = read_bio;
 828
 829                 read_bio->bi_sector = r10_bio->devs[slot].addr +
 830                         mirror->rdev->data_offset;
 831                 read_bio->bi_bdev = mirror->rdev->bdev;
 832                 read_bio->bi_end_io = raid10_end_read_request;
 833                 read_bio->bi_rw = READ | do_sync;
 834                 read_bio->bi_private = r10_bio;
 835
 836                 generic_make_request(read_bio);
 837                 return 0;
 838         }
 839
 840         /*
 841          * WRITE:
 842          */
 843         /* first select target devices under spinlock and
 844          * inc refcount on their rdev.  Record them by setting
 845          * bios[x] to bio
 846          */
 847         raid10_find_phys(conf, r10_bio);
 848         rcu_read_lock();
 849         for (i = 0;  i < conf->copies; i++) {
 850                 int d = r10_bio->devs[i].devnum;
 851                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
 852                 if (rdev &&
 853                     !test_bit(Faulty, &rdev->flags)) {
 854                         atomic_inc(&rdev->nr_pending);
 855                         r10_bio->devs[i].bio = bio;
 856                 } else {
 857                         r10_bio->devs[i].bio = NULL;
 858                         set_bit(R10BIO_Degraded, &r10_bio->state);
 859                 }
 860         }
 861         rcu_read_unlock();
 862
 863         atomic_set(&r10_bio->remaining, 0);
 864
 865         bio_list_init(&bl);
 866         for (i = 0; i < conf->copies; i++) {
 867                 struct bio *mbio;
 868                 int d = r10_bio->devs[i].devnum;
 869                 if (!r10_bio->devs[i].bio)
 870                         continue;
 871
 872                 mbio = bio_clone(bio, GFP_NOIO);
 873                 r10_bio->devs[i].bio = mbio;
 874
 875                 mbio->bi_sector = r10_bio->devs[i].addr+
 876                         conf->mirrors[d].rdev->data_offset;
 877                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
 878                 mbio->bi_end_io = raid10_end_write_request;
 879                 mbio->bi_rw = WRITE | do_sync;
 880                 mbio->bi_private = r10_bio;
 881
 882                 atomic_inc(&r10_bio->remaining);
 883                 bio_list_add(&bl, mbio);
 884         }
 885
 886         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
 887         spin_lock_irqsave(&conf->device_lock, flags);
 888         bio_list_merge(&conf->pending_bio_list, &bl);
 889         blk_plug_device(mddev->queue);
 890         spin_unlock_irqrestore(&conf->device_lock, flags);
 891
 892         if (do_sync)
 893                 md_wakeup_thread(mddev->thread);
 894
 895         return 0;
 896 }
 897
 898 static void status(struct seq_file *seq, mddev_t *mddev)
 899 {
 900         conf_t *conf = mddev_to_conf(mddev);
 901         int i;
 902
 903         if (conf->near_copies < conf->raid_disks)
 904                 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
 905         if (conf->near_copies > 1)
 906                 seq_printf(seq, " %d near-copies", conf->near_copies);
 907         if (conf->far_copies > 1)
 908                 seq_printf(seq, " %d far-copies", conf->far_copies);
 909
 910         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 911                                                 conf->working_disks);
 912         for (i = 0; i < conf->raid_disks; i++)
 913                 seq_printf(seq, "%s",
 914                               conf->mirrors[i].rdev &&
 915                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
 916         seq_printf(seq, "]");
 917 }
 918
 919 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 920 {
 921         char b[BDEVNAME_SIZE];
 922         conf_t *conf = mddev_to_conf(mddev);
 923
 924         /*
 925          * If it is not operational, then we have already marked it as dead
 926          * else if it is the last working disks, ignore the error, let the
 927          * next level up know.
 928          * else mark the drive as failed
 929          */
 930         if (test_bit(In_sync, &rdev->flags)
 931             && conf->working_disks == 1)
 932                 /*
 933                  * Don't fail the drive, just return an IO error.
 934                  * The test should really be more sophisticated than
 935                  * "working_disks == 1", but it isn't critical, and
 936                  * can wait until we do more sophisticated "is the drive
 937                  * really dead" tests...
 938                  */
 939                 return;
 940         if (test_bit(In_sync, &rdev->flags)) {
 941                 mddev->degraded++;
 942                 conf->working_disks--;
 943                 /*
 944                  * if recovery is running, make sure it aborts.
 945                  */
 946                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 947         }
 948         clear_bit(In_sync, &rdev->flags);
 949         set_bit(Faulty, &rdev->flags);
 950         mddev->sb_dirty = 1;
 951         printk(KERN_ALERT "raid10: Disk failure on %s, disabling device. \n"
 952                 "       Operation continuing on %d devices\n",
 953                 bdevname(rdev->bdev,b), conf->working_disks);
 954 }
 955
 956 static void print_conf(conf_t *conf)
 957 {
 958         int i;
 959         mirror_info_t *tmp;
 960
 961         printk("RAID10 conf printout:\n");
 962         if (!conf) {
 963                 printk("(!conf)\n");
 964                 return;
 965         }
 966         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 967                 conf->raid_disks);
 968
 969         for (i = 0; i < conf->raid_disks; i++) {
 970                 char b[BDEVNAME_SIZE];
 971                 tmp = conf->mirrors + i;
 972                 if (tmp->rdev)
 973                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
 974                                 i, !test_bit(In_sync, &tmp->rdev->flags),
 975                                 !test_bit(Faulty, &tmp->rdev->flags),
 976                                 bdevname(tmp->rdev->bdev,b));
 977         }
 978 }
 979
 980 static void close_sync(conf_t *conf)
 981 {
 982         wait_barrier(conf);
 983         allow_barrier(conf);
 984
 985         mempool_destroy(conf->r10buf_pool);
 986         conf->r10buf_pool = NULL;
 987 }
 988
 989 /* check if there are enough drives for
 990  * every block to appear on atleast one
 991  */
 992 static int enough(conf_t *conf)
 993 {
 994         int first = 0;
 995
 996         do {
 997                 int n = conf->copies;
 998                 int cnt = 0;
 999                 while (n--) {
1000                         if (conf->mirrors[first].rdev)
1001                                 cnt++;
1002                         first = (first+1) % conf->raid_disks;
1003                 }
1004                 if (cnt == 0)
1005                         return 0;
1006         } while (first != 0);
1007         return 1;
1008 }
1009
1010 static int raid10_spare_active(mddev_t *mddev)
1011 {
1012         int i;
1013         conf_t *conf = mddev->private;
1014         mirror_info_t *tmp;
1015
1016         /*
1017          * Find all non-in_sync disks within the RAID10 configuration
1018          * and mark them in_sync
1019          */
1020         for (i = 0; i < conf->raid_disks; i++) {
1021                 tmp = conf->mirrors + i;
1022                 if (tmp->rdev
1023                     && !test_bit(Faulty, &tmp->rdev->flags)
1024                     && !test_bit(In_sync, &tmp->rdev->flags)) {
1025                         conf->working_disks++;
1026                         mddev->degraded--;
1027                         set_bit(In_sync, &tmp->rdev->flags);
1028                 }
1029         }
1030
1031         print_conf(conf);
1032         return 0;
1033 }
1034
1035
1036 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1037 {
1038         conf_t *conf = mddev->private;
1039         int found = 0;
1040         int mirror;
1041         mirror_info_t *p;
1042
1043         if (mddev->recovery_cp < MaxSector)
1044                 /* only hot-add to in-sync arrays, as recovery is
1045                  * very different from resync
1046                  */
1047                 return 0;
1048         if (!enough(conf))
1049                 return 0;
1050
1051         if (rdev->saved_raid_disk >= 0 &&
1052             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1053                 mirror = rdev->saved_raid_disk;
1054         else
1055                 mirror = 0;
1056         for ( ; mirror < mddev->raid_disks; mirror++)
1057                 if ( !(p=conf->mirrors+mirror)->rdev) {
1058
1059                         blk_queue_stack_limits(mddev->queue,
1060                                                rdev->bdev->bd_disk->queue);
1061                         /* as we don't honour merge_bvec_fn, we must never risk
1062                          * violating it, so limit ->max_sector to one PAGE, as
1063                          * a one page request is never in violation.
1064                          */
1065                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1066                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1067                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1068
1069                         p->head_position = 0;
1070                         rdev->raid_disk = mirror;
1071                         found = 1;
1072                         if (rdev->saved_raid_disk != mirror)
1073                                 conf->fullsync = 1;
1074                         rcu_assign_pointer(p->rdev, rdev);
1075                         break;
1076                 }
1077
1078         print_conf(conf);
1079         return found;
1080 }
1081
1082 static int raid10_remove_disk(mddev_t *mddev, int number)
1083 {
1084         conf_t *conf = mddev->private;
1085         int err = 0;
1086         mdk_rdev_t *rdev;
1087         mirror_info_t *p = conf->mirrors+ number;
1088
1089         print_conf(conf);
1090         rdev = p->rdev;
1091         if (rdev) {
1092                 if (test_bit(In_sync, &rdev->flags) ||
1093                     atomic_read(&rdev->nr_pending)) {
1094                         err = -EBUSY;
1095                         goto abort;
1096                 }
1097                 p->rdev = NULL;
1098                 synchronize_rcu();
1099                 if (atomic_read(&rdev->nr_pending)) {
1100                         /* lost the race, try later */
1101                         err = -EBUSY;
1102                         p->rdev = rdev;
1103                 }
1104         }
1105 abort:
1106
1107         print_conf(conf);
1108         return err;
1109 }
1110
1111
1112 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1113 {
1114         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1115         conf_t *conf = mddev_to_conf(r10_bio->mddev);
1116         int i,d;
1117
1118         if (bio->bi_size)
1119                 return 1;
1120
1121         for (i=0; i<conf->copies; i++)
1122                 if (r10_bio->devs[i].bio == bio)
1123                         break;
1124         if (i == conf->copies)
1125                 BUG();
1126         update_head_pos(i, r10_bio);
1127         d = r10_bio->devs[i].devnum;
1128
1129         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1130                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1131         else {
1132                 atomic_add(r10_bio->sectors,
1133                            &conf->mirrors[d].rdev->corrected_errors);
1134                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1135                         md_error(r10_bio->mddev,
1136                                  conf->mirrors[d].rdev);
1137         }
1138
1139         /* for reconstruct, we always reschedule after a read.
1140          * for resync, only after all reads
1141          */
1142         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1143             atomic_dec_and_test(&r10_bio->remaining)) {
1144                 /* we have read all the blocks,
1145                  * do the comparison in process context in raid10d
1146                  */
1147                 reschedule_retry(r10_bio);
1148         }
1149         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1150         return 0;
1151 }
1152
1153 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1154 {
1155         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1156         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1157         mddev_t *mddev = r10_bio->mddev;
1158         conf_t *conf = mddev_to_conf(mddev);
1159         int i,d;
1160
1161         if (bio->bi_size)
1162                 return 1;
1163
1164         for (i = 0; i < conf->copies; i++)
1165                 if (r10_bio->devs[i].bio == bio)
1166                         break;
1167         d = r10_bio->devs[i].devnum;
1168
1169         if (!uptodate)
1170                 md_error(mddev, conf->mirrors[d].rdev);
1171         update_head_pos(i, r10_bio);
1172
1173         while (atomic_dec_and_test(&r10_bio->remaining)) {
1174                 if (r10_bio->master_bio == NULL) {
1175                         /* the primary of several recovery bios */
1176                         md_done_sync(mddev, r10_bio->sectors, 1);
1177                         put_buf(r10_bio);
1178                         break;
1179                 } else {
1180                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1181                         put_buf(r10_bio);
1182                         r10_bio = r10_bio2;
1183                 }
1184         }
1185         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1186         return 0;
1187 }
1188
1189 /*
1190  * Note: sync and recover and handled very differently for raid10
1191  * This code is for resync.
1192  * For resync, we read through virtual addresses and read all blocks.
1193  * If there is any error, we schedule a write.  The lowest numbered
1194  * drive is authoritative.
1195  * However requests come for physical address, so we need to map.
1196  * For every physical address there are raid_disks/copies virtual addresses,
1197  * which is always are least one, but is not necessarly an integer.
1198  * This means that a physical address can span multiple chunks, so we may
1199  * have to submit multiple io requests for a single sync request.
1200  */
1201 /*
1202  * We check if all blocks are in-sync and only write to blocks that
1203  * aren't in sync
1204  */
1205 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1206 {
1207         conf_t *conf = mddev_to_conf(mddev);
1208         int i, first;
1209         struct bio *tbio, *fbio;
1210
1211         atomic_set(&r10_bio->remaining, 1);
1212
1213         /* find the first device with a block */
1214         for (i=0; i<conf->copies; i++)
1215                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1216                         break;
1217
1218         if (i == conf->copies)
1219                 goto done;
1220
1221         first = i;
1222         fbio = r10_bio->devs[i].bio;
1223
1224         /* now find blocks with errors */
1225         for (i=0 ; i < conf->copies ; i++) {
1226                 int  j, d;
1227                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1228
1229                 tbio = r10_bio->devs[i].bio;
1230
1231                 if (tbio->bi_end_io != end_sync_read)
1232                         continue;
1233                 if (i == first)
1234                         continue;
1235                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1236                         /* We know that the bi_io_vec layout is the same for
1237                          * both 'first' and 'i', so we just compare them.
1238                          * All vec entries are PAGE_SIZE;
1239                          */
1240                         for (j = 0; j < vcnt; j++)
1241                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1242                                            page_address(tbio->bi_io_vec[j].bv_page),
1243                                            PAGE_SIZE))
1244                                         break;
1245                         if (j == vcnt)
1246                                 continue;
1247                         mddev->resync_mismatches += r10_bio->sectors;
1248                 }
1249                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1250                         /* Don't fix anything. */
1251                         continue;
1252                 /* Ok, we need to write this bio
1253                  * First we need to fixup bv_offset, bv_len and
1254                  * bi_vecs, as the read request might have corrupted these
1255                  */
1256                 tbio->bi_vcnt = vcnt;
1257                 tbio->bi_size = r10_bio->sectors << 9;
1258                 tbio->bi_idx = 0;
1259                 tbio->bi_phys_segments = 0;
1260                 tbio->bi_hw_segments = 0;
1261                 tbio->bi_hw_front_size = 0;
1262                 tbio->bi_hw_back_size = 0;
1263                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1264                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1265                 tbio->bi_next = NULL;
1266                 tbio->bi_rw = WRITE;
1267                 tbio->bi_private = r10_bio;
1268                 tbio->bi_sector = r10_bio->devs[i].addr;
1269
1270                 for (j=0; j < vcnt ; j++) {
1271                         tbio->bi_io_vec[j].bv_offset = 0;
1272                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1273
1274                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1275                                page_address(fbio->bi_io_vec[j].bv_page),
1276                                PAGE_SIZE);
1277                 }
1278                 tbio->bi_end_io = end_sync_write;
1279
1280                 d = r10_bio->devs[i].devnum;
1281                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1282                 atomic_inc(&r10_bio->remaining);
1283                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1284
1285                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1286                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1287                 generic_make_request(tbio);
1288         }
1289
1290 done:
1291         if (atomic_dec_and_test(&r10_bio->remaining)) {
1292                 md_done_sync(mddev, r10_bio->sectors, 1);
1293                 put_buf(r10_bio);
1294         }
1295 }
1296
1297 /*
1298  * Now for the recovery code.
1299  * Recovery happens across physical sectors.
1300  * We recover all non-is_sync drives by finding the virtual address of
1301  * each, and then choose a working drive that also has that virt address.
1302  * There is a separate r10_bio for each non-in_sync drive.
1303  * Only the first two slots are in use. The first for reading,
1304  * The second for writing.
1305  *
1306  */
1307
1308 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1309 {
1310         conf_t *conf = mddev_to_conf(mddev);
1311         int i, d;
1312         struct bio *bio, *wbio;
1313
1314
1315         /* move the pages across to the second bio
1316          * and submit the write request
1317          */
1318         bio = r10_bio->devs[0].bio;
1319         wbio = r10_bio->devs[1].bio;
1320         for (i=0; i < wbio->bi_vcnt; i++) {
1321                 struct page *p = bio->bi_io_vec[i].bv_page;
1322                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1323                 wbio->bi_io_vec[i].bv_page = p;
1324         }
1325         d = r10_bio->devs[1].devnum;
1326
1327         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1328         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1329         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1330                 generic_make_request(wbio);
1331         else
1332                 bio_endio(wbio, wbio->bi_size, -EIO);
1333 }
1334
1335
1336 /*
1337  * This is a kernel thread which:
1338  *
1339  *      1.      Retries failed read operations on working mirrors.
1340  *      2.      Updates the raid superblock when problems encounter.
1341  *      3.      Performs writes following reads for array syncronising.
1342  */
1343
1344 static void raid10d(mddev_t *mddev)
1345 {
1346         r10bio_t *r10_bio;
1347         struct bio *bio;
1348         unsigned long flags;
1349         conf_t *conf = mddev_to_conf(mddev);
1350         struct list_head *head = &conf->retry_list;
1351         int unplug=0;
1352         mdk_rdev_t *rdev;
1353
1354         md_check_recovery(mddev);
1355
1356         for (;;) {
1357                 char b[BDEVNAME_SIZE];
1358                 spin_lock_irqsave(&conf->device_lock, flags);
1359
1360                 if (conf->pending_bio_list.head) {
1361                         bio = bio_list_get(&conf->pending_bio_list);
1362                         blk_remove_plug(mddev->queue);
1363                         spin_unlock_irqrestore(&conf->device_lock, flags);
1364                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1365                         if (bitmap_unplug(mddev->bitmap) != 0)
1366                                 printk("%s: bitmap file write failed!\n", mdname(mddev));
1367
1368                         while (bio) { /* submit pending writes */
1369                                 struct bio *next = bio->bi_next;
1370                                 bio->bi_next = NULL;
1371                                 generic_make_request(bio);
1372                                 bio = next;
1373                         }
1374                         unplug = 1;
1375
1376                         continue;
1377                 }
1378
1379                 if (list_empty(head))
1380                         break;
1381                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1382                 list_del(head->prev);
1383                 conf->nr_queued--;
1384                 spin_unlock_irqrestore(&conf->device_lock, flags);
1385
1386                 mddev = r10_bio->mddev;
1387                 conf = mddev_to_conf(mddev);
1388                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1389                         sync_request_write(mddev, r10_bio);
1390                         unplug = 1;
1391                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1392                         recovery_request_write(mddev, r10_bio);
1393                         unplug = 1;
1394                 } else {
1395                         int mirror;
1396                         /* we got a read error. Maybe the drive is bad.  Maybe just
1397                          * the block and we can fix it.
1398                          * We freeze all other IO, and try reading the block from
1399                          * other devices.  When we find one, we re-write
1400                          * and check it that fixes the read error.
1401                          * This is all done synchronously while the array is
1402                          * frozen.
1403                          */
1404                         int sect = 0; /* Offset from r10_bio->sector */
1405                         int sectors = r10_bio->sectors;
1406                         freeze_array(conf);
1407                         if (mddev->ro == 0) while(sectors) {
1408                                 int s = sectors;
1409                                 int sl = r10_bio->read_slot;
1410                                 int success = 0;
1411
1412                                 if (s > (PAGE_SIZE>>9))
1413                                         s = PAGE_SIZE >> 9;
1414
1415                                 do {
1416                                         int d = r10_bio->devs[sl].devnum;
1417                                         rdev = conf->mirrors[d].rdev;
1418                                         if (rdev &&
1419                                             test_bit(In_sync, &rdev->flags) &&
1420                                             sync_page_io(rdev->bdev,
1421                                                          r10_bio->devs[sl].addr +
1422                                                          sect + rdev->data_offset,
1423                                                          s<<9,
1424                                                          conf->tmppage, READ))
1425                                                 success = 1;
1426                                         else {
1427                                                 sl++;
1428                                                 if (sl == conf->copies)
1429                                                         sl = 0;
1430                                         }
1431                                 } while (!success && sl != r10_bio->read_slot);
1432
1433                                 if (success) {
1434                                         int start = sl;
1435                                         /* write it back and re-read */
1436                                         while (sl != r10_bio->read_slot) {
1437                                                 int d;
1438                                                 if (sl==0)
1439                                                         sl = conf->copies;
1440                                                 sl--;
1441                                                 d = r10_bio->devs[sl].devnum;
1442                                                 rdev = conf->mirrors[d].rdev;
1443                                                 if (rdev &&
1444                                                     test_bit(In_sync, &rdev->flags)) {
1445                                                         atomic_add(s, &rdev->corrected_errors);
1446                                                         if (sync_page_io(rdev->bdev,
1447                                                                          r10_bio->devs[sl].addr +
1448                                                                          sect + rdev->data_offset,
1449                                                                          s<<9, conf->tmppage, WRITE) == 0)
1450                                                                 /* Well, this device is dead */
1451                                                                 md_error(mddev, rdev);
1452                                                 }
1453                                         }
1454                                         sl = start;
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 = conf->mirrors[d].rdev;
1462                                                 if (rdev &&
1463                                                     test_bit(In_sync, &rdev->flags)) {
1464                                                         if (sync_page_io(rdev->bdev,
1465                                                                          r10_bio->devs[sl].addr +
1466                                                                          sect + rdev->data_offset,
1467                                                                          s<<9, conf->tmppage, READ) == 0)
1468                                                                 /* Well, this device is dead */
1469                                                                 md_error(mddev, rdev);
1470                                                 }
1471                                         }
1472                                 } else {
1473                                         /* Cannot read from anywhere -- bye bye array */
1474                                         md_error(mddev, conf->mirrors[r10_bio->devs[r10_bio->read_slot].devnum].rdev);
1475                                         break;
1476                                 }
1477                                 sectors -= s;
1478                                 sect += s;
1479                         }
1480
1481                         unfreeze_array(conf);
1482
1483                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1484                         r10_bio->devs[r10_bio->read_slot].bio =
1485                                 mddev->ro ? IO_BLOCKED : NULL;
1486                         bio_put(bio);
1487                         mirror = read_balance(conf, r10_bio);
1488                         if (mirror == -1) {
1489                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1490                                        " read error for block %llu\n",
1491                                        bdevname(bio->bi_bdev,b),
1492                                        (unsigned long long)r10_bio->sector);
1493                                 raid_end_bio_io(r10_bio);
1494                         } else {
1495                                 const int do_sync = bio_sync(r10_bio->master_bio);
1496                                 rdev = conf->mirrors[mirror].rdev;
1497                                 if (printk_ratelimit())
1498                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1499                                                " another mirror\n",
1500                                                bdevname(rdev->bdev,b),
1501                                                (unsigned long long)r10_bio->sector);
1502                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1503                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1504                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1505                                         + rdev->data_offset;
1506                                 bio->bi_bdev = rdev->bdev;
1507                                 bio->bi_rw = READ | do_sync;
1508                                 bio->bi_private = r10_bio;
1509                                 bio->bi_end_io = raid10_end_read_request;
1510                                 unplug = 1;
1511                                 generic_make_request(bio);
1512                         }
1513                 }
1514         }
1515         spin_unlock_irqrestore(&conf->device_lock, flags);
1516         if (unplug)
1517                 unplug_slaves(mddev);
1518 }
1519
1520
1521 static int init_resync(conf_t *conf)
1522 {
1523         int buffs;
1524
1525         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1526         if (conf->r10buf_pool)
1527                 BUG();
1528         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1529         if (!conf->r10buf_pool)
1530                 return -ENOMEM;
1531         conf->next_resync = 0;
1532         return 0;
1533 }
1534
1535 /*
1536  * perform a "sync" on one "block"
1537  *
1538  * We need to make sure that no normal I/O request - particularly write
1539  * requests - conflict with active sync requests.
1540  *
1541  * This is achieved by tracking pending requests and a 'barrier' concept
1542  * that can be installed to exclude normal IO requests.
1543  *
1544  * Resync and recovery are handled very differently.
1545  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1546  *
1547  * For resync, we iterate over virtual addresses, read all copies,
1548  * and update if there are differences.  If only one copy is live,
1549  * skip it.
1550  * For recovery, we iterate over physical addresses, read a good
1551  * value for each non-in_sync drive, and over-write.
1552  *
1553  * So, for recovery we may have several outstanding complex requests for a
1554  * given address, one for each out-of-sync device.  We model this by allocating
1555  * a number of r10_bio structures, one for each out-of-sync device.
1556  * As we setup these structures, we collect all bio's together into a list
1557  * which we then process collectively to add pages, and then process again
1558  * to pass to generic_make_request.
1559  *
1560  * The r10_bio structures are linked using a borrowed master_bio pointer.
1561  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1562  * has its remaining count decremented to 0, the whole complex operation
1563  * is complete.
1564  *
1565  */
1566
1567 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1568 {
1569         conf_t *conf = mddev_to_conf(mddev);
1570         r10bio_t *r10_bio;
1571         struct bio *biolist = NULL, *bio;
1572         sector_t max_sector, nr_sectors;
1573         int disk;
1574         int i;
1575         int max_sync;
1576         int sync_blocks;
1577
1578         sector_t sectors_skipped = 0;
1579         int chunks_skipped = 0;
1580
1581         if (!conf->r10buf_pool)
1582                 if (init_resync(conf))
1583                         return 0;
1584
1585  skipped:
1586         max_sector = mddev->size << 1;
1587         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1588                 max_sector = mddev->resync_max_sectors;
1589         if (sector_nr >= max_sector) {
1590                 /* If we aborted, we need to abort the
1591                  * sync on the 'current' bitmap chucks (there can
1592                  * be several when recovering multiple devices).
1593                  * as we may have started syncing it but not finished.
1594                  * We can find the current address in
1595                  * mddev->curr_resync, but for recovery,
1596                  * we need to convert that to several
1597                  * virtual addresses.
1598                  */
1599                 if (mddev->curr_resync < max_sector) { /* aborted */
1600                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1601                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1602                                                 &sync_blocks, 1);
1603                         else for (i=0; i<conf->raid_disks; i++) {
1604                                 sector_t sect =
1605                                         raid10_find_virt(conf, mddev->curr_resync, i);
1606                                 bitmap_end_sync(mddev->bitmap, sect,
1607                                                 &sync_blocks, 1);
1608                         }
1609                 } else /* completed sync */
1610                         conf->fullsync = 0;
1611
1612                 bitmap_close_sync(mddev->bitmap);
1613                 close_sync(conf);
1614                 *skipped = 1;
1615                 return sectors_skipped;
1616         }
1617         if (chunks_skipped >= conf->raid_disks) {
1618                 /* if there has been nothing to do on any drive,
1619                  * then there is nothing to do at all..
1620                  */
1621                 *skipped = 1;
1622                 return (max_sector - sector_nr) + sectors_skipped;
1623         }
1624
1625         /* make sure whole request will fit in a chunk - if chunks
1626          * are meaningful
1627          */
1628         if (conf->near_copies < conf->raid_disks &&
1629             max_sector > (sector_nr | conf->chunk_mask))
1630                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1631         /*
1632          * If there is non-resync activity waiting for us then
1633          * put in a delay to throttle resync.
1634          */
1635         if (!go_faster && conf->nr_waiting)
1636                 msleep_interruptible(1000);
1637
1638         /* Again, very different code for resync and recovery.
1639          * Both must result in an r10bio with a list of bios that
1640          * have bi_end_io, bi_sector, bi_bdev set,
1641          * and bi_private set to the r10bio.
1642          * For recovery, we may actually create several r10bios
1643          * with 2 bios in each, that correspond to the bios in the main one.
1644          * In this case, the subordinate r10bios link back through a
1645          * borrowed master_bio pointer, and the counter in the master
1646          * includes a ref from each subordinate.
1647          */
1648         /* First, we decide what to do and set ->bi_end_io
1649          * To end_sync_read if we want to read, and
1650          * end_sync_write if we will want to write.
1651          */
1652
1653         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1654         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1655                 /* recovery... the complicated one */
1656                 int i, j, k;
1657                 r10_bio = NULL;
1658
1659                 for (i=0 ; i<conf->raid_disks; i++)
1660                         if (conf->mirrors[i].rdev &&
1661                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1662                                 int still_degraded = 0;
1663                                 /* want to reconstruct this device */
1664                                 r10bio_t *rb2 = r10_bio;
1665                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1666                                 int must_sync;
1667                                 /* Unless we are doing a full sync, we only need
1668                                  * to recover the block if it is set in the bitmap
1669                                  */
1670                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1671                                                               &sync_blocks, 1);
1672                                 if (sync_blocks < max_sync)
1673                                         max_sync = sync_blocks;
1674                                 if (!must_sync &&
1675                                     !conf->fullsync) {
1676                                         /* yep, skip the sync_blocks here, but don't assume
1677                                          * that there will never be anything to do here
1678                                          */
1679                                         chunks_skipped = -1;
1680                                         continue;
1681                                 }
1682
1683                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1684                                 raise_barrier(conf, rb2 != NULL);
1685                                 atomic_set(&r10_bio->remaining, 0);
1686
1687                                 r10_bio->master_bio = (struct bio*)rb2;
1688                                 if (rb2)
1689                                         atomic_inc(&rb2->remaining);
1690                                 r10_bio->mddev = mddev;
1691                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1692                                 r10_bio->sector = sect;
1693
1694                                 raid10_find_phys(conf, r10_bio);
1695                                 /* Need to check if this section will still be
1696                                  * degraded
1697                                  */
1698                                 for (j=0; j<conf->copies;j++) {
1699                                         int d = r10_bio->devs[j].devnum;
1700                                         if (conf->mirrors[d].rdev == NULL ||
1701                                             test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
1702                                                 still_degraded = 1;
1703                                                 break;
1704                                         }
1705                                 }
1706                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1707                                                               &sync_blocks, still_degraded);
1708
1709                                 for (j=0; j<conf->copies;j++) {
1710                                         int d = r10_bio->devs[j].devnum;
1711                                         if (conf->mirrors[d].rdev &&
1712                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1713                                                 /* This is where we read from */
1714                                                 bio = r10_bio->devs[0].bio;
1715                                                 bio->bi_next = biolist;
1716                                                 biolist = bio;
1717                                                 bio->bi_private = r10_bio;
1718                                                 bio->bi_end_io = end_sync_read;
1719                                                 bio->bi_rw = 0;
1720                                                 bio->bi_sector = r10_bio->devs[j].addr +
1721                                                         conf->mirrors[d].rdev->data_offset;
1722                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1723                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1724                                                 atomic_inc(&r10_bio->remaining);
1725                                                 /* and we write to 'i' */
1726
1727                                                 for (k=0; k<conf->copies; k++)
1728                                                         if (r10_bio->devs[k].devnum == i)
1729                                                                 break;
1730                                                 bio = r10_bio->devs[1].bio;
1731                                                 bio->bi_next = biolist;
1732                                                 biolist = bio;
1733                                                 bio->bi_private = r10_bio;
1734                                                 bio->bi_end_io = end_sync_write;
1735                                                 bio->bi_rw = 1;
1736                                                 bio->bi_sector = r10_bio->devs[k].addr +
1737                                                         conf->mirrors[i].rdev->data_offset;
1738                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1739
1740                                                 r10_bio->devs[0].devnum = d;
1741                                                 r10_bio->devs[1].devnum = i;
1742
1743                                                 break;
1744                                         }
1745                                 }
1746                                 if (j == conf->copies) {
1747                                         /* Cannot recover, so abort the recovery */
1748                                         put_buf(r10_bio);
1749                                         r10_bio = rb2;
1750                                         if (!test_and_set_bit(MD_RECOVERY_ERR, &mddev->recovery))
1751                                                 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1752                                                        mdname(mddev));
1753                                         break;
1754                                 }
1755                         }
1756                 if (biolist == NULL) {
1757                         while (r10_bio) {
1758                                 r10bio_t *rb2 = r10_bio;
1759                                 r10_bio = (r10bio_t*) rb2->master_bio;
1760                                 rb2->master_bio = NULL;
1761                                 put_buf(rb2);
1762                         }
1763                         goto giveup;
1764                 }
1765         } else {
1766                 /* resync. Schedule a read for every block at this virt offset */
1767                 int count = 0;
1768
1769                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1770                                        &sync_blocks, mddev->degraded) &&
1771                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1772                         /* We can skip this block */
1773                         *skipped = 1;
1774                         return sync_blocks + sectors_skipped;
1775                 }
1776                 if (sync_blocks < max_sync)
1777                         max_sync = sync_blocks;
1778                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1779
1780                 r10_bio->mddev = mddev;
1781                 atomic_set(&r10_bio->remaining, 0);
1782                 raise_barrier(conf, 0);
1783                 conf->next_resync = sector_nr;
1784
1785                 r10_bio->master_bio = NULL;
1786                 r10_bio->sector = sector_nr;
1787                 set_bit(R10BIO_IsSync, &r10_bio->state);
1788                 raid10_find_phys(conf, r10_bio);
1789                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1790
1791                 for (i=0; i<conf->copies; i++) {
1792                         int d = r10_bio->devs[i].devnum;
1793                         bio = r10_bio->devs[i].bio;
1794                         bio->bi_end_io = NULL;
1795                         if (conf->mirrors[d].rdev == NULL ||
1796                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1797                                 continue;
1798                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1799                         atomic_inc(&r10_bio->remaining);
1800                         bio->bi_next = biolist;
1801                         biolist = bio;
1802                         bio->bi_private = r10_bio;
1803                         bio->bi_end_io = end_sync_read;
1804                         bio->bi_rw = 0;
1805                         bio->bi_sector = r10_bio->devs[i].addr +
1806                                 conf->mirrors[d].rdev->data_offset;
1807                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1808                         count++;
1809                 }
1810
1811                 if (count < 2) {
1812                         for (i=0; i<conf->copies; i++) {
1813                                 int d = r10_bio->devs[i].devnum;
1814                                 if (r10_bio->devs[i].bio->bi_end_io)
1815                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1816                         }
1817                         put_buf(r10_bio);
1818                         biolist = NULL;
1819                         goto giveup;
1820                 }
1821         }
1822
1823         for (bio = biolist; bio ; bio=bio->bi_next) {
1824
1825                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1826                 if (bio->bi_end_io)
1827                         bio->bi_flags |= 1 << BIO_UPTODATE;
1828                 bio->bi_vcnt = 0;
1829                 bio->bi_idx = 0;
1830                 bio->bi_phys_segments = 0;
1831                 bio->bi_hw_segments = 0;
1832                 bio->bi_size = 0;
1833         }
1834
1835         nr_sectors = 0;
1836         if (sector_nr + max_sync < max_sector)
1837                 max_sector = sector_nr + max_sync;
1838         do {
1839                 struct page *page;
1840                 int len = PAGE_SIZE;
1841                 disk = 0;
1842                 if (sector_nr + (len>>9) > max_sector)
1843                         len = (max_sector - sector_nr) << 9;
1844                 if (len == 0)
1845                         break;
1846                 for (bio= biolist ; bio ; bio=bio->bi_next) {
1847                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1848                         if (bio_add_page(bio, page, len, 0) == 0) {
1849                                 /* stop here */
1850                                 struct bio *bio2;
1851                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1852                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1853                                         /* remove last page from this bio */
1854                                         bio2->bi_vcnt--;
1855                                         bio2->bi_size -= len;
1856                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1857                                 }
1858                                 goto bio_full;
1859                         }
1860                         disk = i;
1861                 }
1862                 nr_sectors += len>>9;
1863                 sector_nr += len>>9;
1864         } while (biolist->bi_vcnt < RESYNC_PAGES);
1865  bio_full:
1866         r10_bio->sectors = nr_sectors;
1867
1868         while (biolist) {
1869                 bio = biolist;
1870                 biolist = biolist->bi_next;
1871
1872                 bio->bi_next = NULL;
1873                 r10_bio = bio->bi_private;
1874                 r10_bio->sectors = nr_sectors;
1875
1876                 if (bio->bi_end_io == end_sync_read) {
1877                         md_sync_acct(bio->bi_bdev, nr_sectors);
1878                         generic_make_request(bio);
1879                 }
1880         }
1881
1882         if (sectors_skipped)
1883                 /* pretend they weren't skipped, it makes
1884                  * no important difference in this case
1885                  */
1886                 md_done_sync(mddev, sectors_skipped, 1);
1887
1888         return sectors_skipped + nr_sectors;
1889  giveup:
1890         /* There is nowhere to write, so all non-sync
1891          * drives must be failed, so try the next chunk...
1892          */
1893         {
1894         sector_t sec = max_sector - sector_nr;
1895         sectors_skipped += sec;
1896         chunks_skipped ++;
1897         sector_nr = max_sector;
1898         goto skipped;
1899         }
1900 }
1901
1902 static int run(mddev_t *mddev)
1903 {
1904         conf_t *conf;
1905         int i, disk_idx;
1906         mirror_info_t *disk;
1907         mdk_rdev_t *rdev;
1908         struct list_head *tmp;
1909         int nc, fc;
1910         sector_t stride, size;
1911
1912         if (mddev->chunk_size == 0) {
1913                 printk(KERN_ERR "md/raid10: non-zero chunk size required.\n");
1914                 return -EINVAL;
1915         }
1916
1917         nc = mddev->layout & 255;
1918         fc = (mddev->layout >> 8) & 255;
1919         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
1920             (mddev->layout >> 16)) {
1921                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
1922                        mdname(mddev), mddev->layout);
1923                 goto out;
1924         }
1925         /*
1926          * copy the already verified devices into our private RAID10
1927          * bookkeeping area. [whatever we allocate in run(),
1928          * should be freed in stop()]
1929          */
1930         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1931         mddev->private = conf;
1932         if (!conf) {
1933                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1934                         mdname(mddev));
1935                 goto out;
1936         }
1937         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1938                                  GFP_KERNEL);
1939         if (!conf->mirrors) {
1940                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1941                        mdname(mddev));
1942                 goto out_free_conf;
1943         }
1944
1945         conf->tmppage = alloc_page(GFP_KERNEL);
1946         if (!conf->tmppage)
1947                 goto out_free_conf;
1948
1949         conf->near_copies = nc;
1950         conf->far_copies = fc;
1951         conf->copies = nc*fc;
1952         conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
1953         conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
1954         stride = mddev->size >> (conf->chunk_shift-1);
1955         sector_div(stride, fc);
1956         conf->stride = stride << conf->chunk_shift;
1957
1958         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
1959                                                 r10bio_pool_free, conf);
1960         if (!conf->r10bio_pool) {
1961                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1962                         mdname(mddev));
1963                 goto out_free_conf;
1964         }
1965
1966         ITERATE_RDEV(mddev, rdev, tmp) {
1967                 disk_idx = rdev->raid_disk;
1968                 if (disk_idx >= mddev->raid_disks
1969                     || disk_idx < 0)
1970                         continue;
1971                 disk = conf->mirrors + disk_idx;
1972
1973                 disk->rdev = rdev;
1974
1975                 blk_queue_stack_limits(mddev->queue,
1976                                        rdev->bdev->bd_disk->queue);
1977                 /* as we don't honour merge_bvec_fn, we must never risk
1978                  * violating it, so limit ->max_sector to one PAGE, as
1979                  * a one page request is never in violation.
1980                  */
1981                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1982                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1983                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
1984
1985                 disk->head_position = 0;
1986                 if (!test_bit(Faulty, &rdev->flags) && test_bit(In_sync, &rdev->flags))
1987                         conf->working_disks++;
1988         }
1989         conf->raid_disks = mddev->raid_disks;
1990         conf->mddev = mddev;
1991         spin_lock_init(&conf->device_lock);
1992         INIT_LIST_HEAD(&conf->retry_list);
1993
1994         spin_lock_init(&conf->resync_lock);
1995         init_waitqueue_head(&conf->wait_barrier);
1996
1997         /* need to check that every block has at least one working mirror */
1998         if (!enough(conf)) {
1999                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2000                        mdname(mddev));
2001                 goto out_free_conf;
2002         }
2003
2004         mddev->degraded = 0;
2005         for (i = 0; i < conf->raid_disks; i++) {
2006
2007                 disk = conf->mirrors + i;
2008
2009                 if (!disk->rdev) {
2010                         disk->head_position = 0;
2011                         mddev->degraded++;
2012                 }
2013         }
2014
2015
2016         mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2017         if (!mddev->thread) {
2018                 printk(KERN_ERR
2019                        "raid10: couldn't allocate thread for %s\n",
2020                        mdname(mddev));
2021                 goto out_free_conf;
2022         }
2023
2024         printk(KERN_INFO
2025                 "raid10: raid set %s active with %d out of %d devices\n",
2026                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2027                 mddev->raid_disks);
2028         /*
2029          * Ok, everything is just fine now
2030          */
2031         size = conf->stride * conf->raid_disks;
2032         sector_div(size, conf->near_copies);
2033         mddev->array_size = size/2;
2034         mddev->resync_max_sectors = size;
2035
2036         mddev->queue->unplug_fn = raid10_unplug;
2037         mddev->queue->issue_flush_fn = raid10_issue_flush;
2038
2039         /* Calculate max read-ahead size.
2040          * We need to readahead at least twice a whole stripe....
2041          * maybe...
2042          */
2043         {
2044                 int stripe = conf->raid_disks * mddev->chunk_size / PAGE_SIZE;
2045                 stripe /= conf->near_copies;
2046                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2047                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2048         }
2049
2050         if (conf->near_copies < mddev->raid_disks)
2051                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2052         return 0;
2053
2054 out_free_conf:
2055         if (conf->r10bio_pool)
2056                 mempool_destroy(conf->r10bio_pool);
2057         safe_put_page(conf->tmppage);
2058         kfree(conf->mirrors);
2059         kfree(conf);
2060         mddev->private = NULL;
2061 out:
2062         return -EIO;
2063 }
2064
2065 static int stop(mddev_t *mddev)
2066 {
2067         conf_t *conf = mddev_to_conf(mddev);
2068
2069         md_unregister_thread(mddev->thread);
2070         mddev->thread = NULL;
2071         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2072         if (conf->r10bio_pool)
2073                 mempool_destroy(conf->r10bio_pool);
2074         kfree(conf->mirrors);
2075         kfree(conf);
2076         mddev->private = NULL;
2077         return 0;
2078 }
2079
2080 static void raid10_quiesce(mddev_t *mddev, int state)
2081 {
2082         conf_t *conf = mddev_to_conf(mddev);
2083
2084         switch(state) {
2085         case 1:
2086                 raise_barrier(conf, 0);
2087                 break;
2088         case 0:
2089                 lower_barrier(conf);
2090                 break;
2091         }
2092         if (mddev->thread) {
2093                 if (mddev->bitmap)
2094                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2095                 else
2096                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2097                 md_wakeup_thread(mddev->thread);
2098         }
2099 }
2100
2101 static struct mdk_personality raid10_personality =
2102 {
2103         .name           = "raid10",
2104         .level          = 10,
2105         .owner          = THIS_MODULE,
2106         .make_request   = make_request,
2107         .run            = run,
2108         .stop           = stop,
2109         .status         = status,
2110         .error_handler  = error,
2111         .hot_add_disk   = raid10_add_disk,
2112         .hot_remove_disk= raid10_remove_disk,
2113         .spare_active   = raid10_spare_active,
2114         .sync_request   = sync_request,
2115         .quiesce        = raid10_quiesce,
2116 };
2117
2118 static int __init raid_init(void)
2119 {
2120         return register_md_personality(&raid10_personality);
2121 }
2122
2123 static void raid_exit(void)
2124 {
2125         unregister_md_personality(&raid10_personality);
2126 }
2127
2128 module_init(raid_init);
2129 module_exit(raid_exit);
2130 MODULE_LICENSE("GPL");
2131 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2132 MODULE_ALIAS("md-raid10");
2133 MODULE_ALIAS("md-level-10");