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 further 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 <linux/slab.h>
  22 #include <linux/delay.h>
  23 #include <linux/blkdev.h>
  24 #include <linux/seq_file.h>
  25 #include "md.h"
  26 #include "raid10.h"
  27 #include "raid0.h"
  28 #include "bitmap.h"
  29
  30 /*
  31  * RAID10 provides a combination of RAID0 and RAID1 functionality.
  32  * The layout of data is defined by
  33  *    chunk_size
  34  *    raid_disks
  35  *    near_copies (stored in low byte of layout)
  36  *    far_copies (stored in second byte of layout)
  37  *    far_offset (stored in bit 16 of layout )
  38  *
  39  * The data to be stored is divided into chunks using chunksize.
  40  * Each device is divided into far_copies sections.
  41  * In each section, chunks are laid out in a style similar to raid0, but
  42  * near_copies copies of each chunk is stored (each on a different drive).
  43  * The starting device for each section is offset near_copies from the starting
  44  * device of the previous section.
  45  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
  46  * drive.
  47  * near_copies and far_copies must be at least one, and their product is at most
  48  * raid_disks.
  49  *
  50  * If far_offset is true, then the far_copies are handled a bit differently.
  51  * The copies are still in different stripes, but instead of be very far apart
  52  * on disk, there are adjacent stripes.
  53  */
  54
  55 /*
  56  * Number of guaranteed r10bios in case of extreme VM load:
  57  */
  58 #define NR_RAID10_BIOS 256
  59
  60 static void allow_barrier(conf_t *conf);
  61 static void lower_barrier(conf_t *conf);
  62
  63 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
  64 {
  65         conf_t *conf = data;
  66         int size = offsetof(struct r10bio_s, devs[conf->copies]);
  67
  68         /* allocate a r10bio with room for raid_disks entries in the bios array */
  69         return kzalloc(size, gfp_flags);
  70 }
  71
  72 static void r10bio_pool_free(void *r10_bio, void *data)
  73 {
  74         kfree(r10_bio);
  75 }
  76
  77 /* Maximum size of each resync request */
  78 #define RESYNC_BLOCK_SIZE (64*1024)
  79 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  80 /* amount of memory to reserve for resync requests */
  81 #define RESYNC_WINDOW (1024*1024)
  82 /* maximum number of concurrent requests, memory permitting */
  83 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
  84
  85 /*
  86  * When performing a resync, we need to read and compare, so
  87  * we need as many pages are there are copies.
  88  * When performing a recovery, we need 2 bios, one for read,
  89  * one for write (we recover only one drive per r10buf)
  90  *
  91  */
  92 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
  93 {
  94         conf_t *conf = data;
  95         struct page *page;
  96         r10bio_t *r10_bio;
  97         struct bio *bio;
  98         int i, j;
  99         int nalloc;
 100
 101         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 102         if (!r10_bio)
 103                 return NULL;
 104
 105         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
 106                 nalloc = conf->copies; /* resync */
 107         else
 108                 nalloc = 2; /* recovery */
 109
 110         /*
 111          * Allocate bios.
 112          */
 113         for (j = nalloc ; j-- ; ) {
 114                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 115                 if (!bio)
 116                         goto out_free_bio;
 117                 r10_bio->devs[j].bio = bio;
 118         }
 119         /*
 120          * Allocate RESYNC_PAGES data pages and attach them
 121          * where needed.
 122          */
 123         for (j = 0 ; j < nalloc; j++) {
 124                 bio = r10_bio->devs[j].bio;
 125                 for (i = 0; i < RESYNC_PAGES; i++) {
 126                         page = alloc_page(gfp_flags);
 127                         if (unlikely(!page))
 128                                 goto out_free_pages;
 129
 130                         bio->bi_io_vec[i].bv_page = page;
 131                 }
 132         }
 133
 134         return r10_bio;
 135
 136 out_free_pages:
 137         for ( ; i > 0 ; i--)
 138                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
 139         while (j--)
 140                 for (i = 0; i < RESYNC_PAGES ; i++)
 141                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 142         j = -1;
 143 out_free_bio:
 144         while ( ++j < nalloc )
 145                 bio_put(r10_bio->devs[j].bio);
 146         r10bio_pool_free(r10_bio, conf);
 147         return NULL;
 148 }
 149
 150 static void r10buf_pool_free(void *__r10_bio, void *data)
 151 {
 152         int i;
 153         conf_t *conf = data;
 154         r10bio_t *r10bio = __r10_bio;
 155         int j;
 156
 157         for (j=0; j < conf->copies; j++) {
 158                 struct bio *bio = r10bio->devs[j].bio;
 159                 if (bio) {
 160                         for (i = 0; i < RESYNC_PAGES; i++) {
 161                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 162                                 bio->bi_io_vec[i].bv_page = NULL;
 163                         }
 164                         bio_put(bio);
 165                 }
 166         }
 167         r10bio_pool_free(r10bio, conf);
 168 }
 169
 170 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
 171 {
 172         int i;
 173
 174         for (i = 0; i < conf->copies; i++) {
 175                 struct bio **bio = & r10_bio->devs[i].bio;
 176                 if (*bio && *bio != IO_BLOCKED)
 177                         bio_put(*bio);
 178                 *bio = NULL;
 179         }
 180 }
 181
 182 static void free_r10bio(r10bio_t *r10_bio)
 183 {
 184         conf_t *conf = r10_bio->mddev->private;
 185
 186         /*
 187          * Wake up any possible resync thread that waits for the device
 188          * to go idle.
 189          */
 190         allow_barrier(conf);
 191
 192         put_all_bios(conf, r10_bio);
 193         mempool_free(r10_bio, conf->r10bio_pool);
 194 }
 195
 196 static void put_buf(r10bio_t *r10_bio)
 197 {
 198         conf_t *conf = r10_bio->mddev->private;
 199
 200         mempool_free(r10_bio, conf->r10buf_pool);
 201
 202         lower_barrier(conf);
 203 }
 204
 205 static void reschedule_retry(r10bio_t *r10_bio)
 206 {
 207         unsigned long flags;
 208         mddev_t *mddev = r10_bio->mddev;
 209         conf_t *conf = mddev->private;
 210
 211         spin_lock_irqsave(&conf->device_lock, flags);
 212         list_add(&r10_bio->retry_list, &conf->retry_list);
 213         conf->nr_queued ++;
 214         spin_unlock_irqrestore(&conf->device_lock, flags);
 215
 216         /* wake up frozen array... */
 217         wake_up(&conf->wait_barrier);
 218
 219         md_wakeup_thread(mddev->thread);
 220 }
 221
 222 /*
 223  * raid_end_bio_io() is called when we have finished servicing a mirrored
 224  * operation and are ready to return a success/failure code to the buffer
 225  * cache layer.
 226  */
 227 static void raid_end_bio_io(r10bio_t *r10_bio)
 228 {
 229         struct bio *bio = r10_bio->master_bio;
 230
 231         bio_endio(bio,
 232                 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
 233         free_r10bio(r10_bio);
 234 }
 235
 236 /*
 237  * Update disk head position estimator based on IRQ completion info.
 238  */
 239 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
 240 {
 241         conf_t *conf = r10_bio->mddev->private;
 242
 243         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 244                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 245 }
 246
 247 static void raid10_end_read_request(struct bio *bio, int error)
 248 {
 249         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 250         r10bio_t *r10_bio = bio->bi_private;
 251         int slot, dev;
 252         conf_t *conf = r10_bio->mddev->private;
 253
 254
 255         slot = r10_bio->read_slot;
 256         dev = r10_bio->devs[slot].devnum;
 257         /*
 258          * this branch is our 'one mirror IO has finished' event handler:
 259          */
 260         update_head_pos(slot, r10_bio);
 261
 262         if (uptodate) {
 263                 /*
 264                  * Set R10BIO_Uptodate in our master bio, so that
 265                  * we will return a good error code to the higher
 266                  * levels even if IO on some other mirrored buffer fails.
 267                  *
 268                  * The 'master' represents the composite IO operation to
 269                  * user-side. So if something waits for IO, then it will
 270                  * wait for the 'master' bio.
 271                  */
 272                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 273                 raid_end_bio_io(r10_bio);
 274                 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 275         } else {
 276                 /*
 277                  * oops, read error - keep the refcount on the rdev
 278                  */
 279                 char b[BDEVNAME_SIZE];
 280                 if (printk_ratelimit())
 281                         printk(KERN_ERR "md/raid10:%s: %s: rescheduling sector %llu\n",
 282                                mdname(conf->mddev),
 283                                bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
 284                 reschedule_retry(r10_bio);
 285         }
 286 }
 287
 288 static void raid10_end_write_request(struct bio *bio, int error)
 289 {
 290         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 291         r10bio_t *r10_bio = bio->bi_private;
 292         int slot, dev;
 293         conf_t *conf = r10_bio->mddev->private;
 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 }
 338
 339
 340 /*
 341  * RAID10 layout manager
 342  * As well as the chunksize and raid_disks count, there are two
 343  * parameters: near_copies and far_copies.
 344  * near_copies * far_copies must be <= raid_disks.
 345  * Normally one of these will be 1.
 346  * If both are 1, we get raid0.
 347  * If near_copies == raid_disks, we get raid1.
 348  *
 349  * Chunks are laid out in raid0 style with near_copies copies of the
 350  * first chunk, followed by near_copies copies of the next chunk and
 351  * so on.
 352  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 353  * as described above, we start again with a device offset of near_copies.
 354  * So we effectively have another copy of the whole array further down all
 355  * the drives, but with blocks on different drives.
 356  * With this layout, and block is never stored twice on the one device.
 357  *
 358  * raid10_find_phys finds the sector offset of a given virtual sector
 359  * on each device that it is on.
 360  *
 361  * raid10_find_virt does the reverse mapping, from a device and a
 362  * sector offset to a virtual address
 363  */
 364
 365 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
 366 {
 367         int n,f;
 368         sector_t sector;
 369         sector_t chunk;
 370         sector_t stripe;
 371         int dev;
 372
 373         int slot = 0;
 374
 375         /* now calculate first sector/dev */
 376         chunk = r10bio->sector >> conf->chunk_shift;
 377         sector = r10bio->sector & conf->chunk_mask;
 378
 379         chunk *= conf->near_copies;
 380         stripe = chunk;
 381         dev = sector_div(stripe, conf->raid_disks);
 382         if (conf->far_offset)
 383                 stripe *= conf->far_copies;
 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         offset = sector & conf->chunk_mask;
 418         if (conf->far_offset) {
 419                 int fc;
 420                 chunk = sector >> conf->chunk_shift;
 421                 fc = sector_div(chunk, conf->far_copies);
 422                 dev -= fc * conf->near_copies;
 423                 if (dev < 0)
 424                         dev += conf->raid_disks;
 425         } else {
 426                 while (sector >= conf->stride) {
 427                         sector -= conf->stride;
 428                         if (dev < conf->near_copies)
 429                                 dev += conf->raid_disks - conf->near_copies;
 430                         else
 431                                 dev -= conf->near_copies;
 432                 }
 433                 chunk = sector >> conf->chunk_shift;
 434         }
 435         vchunk = chunk * conf->raid_disks + dev;
 436         sector_div(vchunk, conf->near_copies);
 437         return (vchunk << conf->chunk_shift) + offset;
 438 }
 439
 440 /**
 441  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 442  *      @q: request queue
 443  *      @bvm: properties of new bio
 444  *      @biovec: the request that could be merged to it.
 445  *
 446  *      Return amount of bytes we can accept at this offset
 447  *      If near_copies == raid_disk, there are no striping issues,
 448  *      but in that case, the function isn't called at all.
 449  */
 450 static int raid10_mergeable_bvec(struct request_queue *q,
 451                                  struct bvec_merge_data *bvm,
 452                                  struct bio_vec *biovec)
 453 {
 454         mddev_t *mddev = q->queuedata;
 455         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 456         int max;
 457         unsigned int chunk_sectors = mddev->chunk_sectors;
 458         unsigned int bio_sectors = bvm->bi_size >> 9;
 459
 460         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 461         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 462         if (max <= biovec->bv_len && bio_sectors == 0)
 463                 return biovec->bv_len;
 464         else
 465                 return max;
 466 }
 467
 468 /*
 469  * This routine returns the disk from which the requested read should
 470  * be done. There is a per-array 'next expected sequential IO' sector
 471  * number - if this matches on the next IO then we use the last disk.
 472  * There is also a per-disk 'last know head position' sector that is
 473  * maintained from IRQ contexts, both the normal and the resync IO
 474  * completion handlers update this position correctly. If there is no
 475  * perfect sequential match then we pick the disk whose head is closest.
 476  *
 477  * If there are 2 mirrors in the same 2 devices, performance degrades
 478  * because position is mirror, not device based.
 479  *
 480  * The rdev for the device selected will have nr_pending incremented.
 481  */
 482
 483 /*
 484  * FIXME: possibly should rethink readbalancing and do it differently
 485  * depending on near_copies / far_copies geometry.
 486  */
 487 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
 488 {
 489         const sector_t this_sector = r10_bio->sector;
 490         int disk, slot;
 491         const int sectors = r10_bio->sectors;
 492         sector_t new_distance, best_dist;
 493         mdk_rdev_t *rdev;
 494         int do_balance;
 495         int best_slot;
 496
 497         raid10_find_phys(conf, r10_bio);
 498         rcu_read_lock();
 499 retry:
 500         best_slot = -1;
 501         best_dist = MaxSector;
 502         do_balance = 1;
 503         /*
 504          * Check if we can balance. We can balance on the whole
 505          * device if no resync is going on (recovery is ok), or below
 506          * the resync window. We take the first readable disk when
 507          * above the resync window.
 508          */
 509         if (conf->mddev->recovery_cp < MaxSector
 510             && (this_sector + sectors >= conf->next_resync))
 511                 do_balance = 0;
 512
 513         for (slot = 0; slot < conf->copies ; slot++) {
 514                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
 515                         continue;
 516                 disk = r10_bio->devs[slot].devnum;
 517                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 518                 if (rdev == NULL)
 519                         continue;
 520                 if (!test_bit(In_sync, &rdev->flags))
 521                         continue;
 522
 523                 if (!do_balance)
 524                         break;
 525
 526                 /* This optimisation is debatable, and completely destroys
 527                  * sequential read speed for 'far copies' arrays.  So only
 528                  * keep it for 'near' arrays, and review those later.
 529                  */
 530                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
 531                         break;
 532
 533                 /* for far > 1 always use the lowest address */
 534                 if (conf->far_copies > 1)
 535                         new_distance = r10_bio->devs[slot].addr;
 536                 else
 537                         new_distance = abs(r10_bio->devs[slot].addr -
 538                                            conf->mirrors[disk].head_position);
 539                 if (new_distance < best_dist) {
 540                         best_dist = new_distance;
 541                         best_slot = slot;
 542                 }
 543         }
 544         if (slot == conf->copies)
 545                 slot = best_slot;
 546
 547         if (slot >= 0) {
 548                 disk = r10_bio->devs[slot].devnum;
 549                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 550                 if (!rdev)
 551                         goto retry;
 552                 atomic_inc(&rdev->nr_pending);
 553                 if (test_bit(Faulty, &rdev->flags)) {
 554                         /* Cannot risk returning a device that failed
 555                          * before we inc'ed nr_pending
 556                          */
 557                         rdev_dec_pending(rdev, conf->mddev);
 558                         goto retry;
 559                 }
 560                 r10_bio->read_slot = slot;
 561         } else
 562                 disk = -1;
 563         rcu_read_unlock();
 564
 565         return disk;
 566 }
 567
 568 static int raid10_congested(void *data, int bits)
 569 {
 570         mddev_t *mddev = data;
 571         conf_t *conf = mddev->private;
 572         int i, ret = 0;
 573
 574         if (mddev_congested(mddev, bits))
 575                 return 1;
 576         rcu_read_lock();
 577         for (i = 0; i < conf->raid_disks && ret == 0; i++) {
 578                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 579                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 580                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 581
 582                         ret |= bdi_congested(&q->backing_dev_info, bits);
 583                 }
 584         }
 585         rcu_read_unlock();
 586         return ret;
 587 }
 588
 589 static void flush_pending_writes(conf_t *conf)
 590 {
 591         /* Any writes that have been queued but are awaiting
 592          * bitmap updates get flushed here.
 593          */
 594         spin_lock_irq(&conf->device_lock);
 595
 596         if (conf->pending_bio_list.head) {
 597                 struct bio *bio;
 598                 bio = bio_list_get(&conf->pending_bio_list);
 599                 spin_unlock_irq(&conf->device_lock);
 600                 /* flush any pending bitmap writes to disk
 601                  * before proceeding w/ I/O */
 602                 bitmap_unplug(conf->mddev->bitmap);
 603
 604                 while (bio) { /* submit pending writes */
 605                         struct bio *next = bio->bi_next;
 606                         bio->bi_next = NULL;
 607                         generic_make_request(bio);
 608                         bio = next;
 609                 }
 610         } else
 611                 spin_unlock_irq(&conf->device_lock);
 612 }
 613
 614 /* Barriers....
 615  * Sometimes we need to suspend IO while we do something else,
 616  * either some resync/recovery, or reconfigure the array.
 617  * To do this we raise a 'barrier'.
 618  * The 'barrier' is a counter that can be raised multiple times
 619  * to count how many activities are happening which preclude
 620  * normal IO.
 621  * We can only raise the barrier if there is no pending IO.
 622  * i.e. if nr_pending == 0.
 623  * We choose only to raise the barrier if no-one is waiting for the
 624  * barrier to go down.  This means that as soon as an IO request
 625  * is ready, no other operations which require a barrier will start
 626  * until the IO request has had a chance.
 627  *
 628  * So: regular IO calls 'wait_barrier'.  When that returns there
 629  *    is no backgroup IO happening,  It must arrange to call
 630  *    allow_barrier when it has finished its IO.
 631  * backgroup IO calls must call raise_barrier.  Once that returns
 632  *    there is no normal IO happeing.  It must arrange to call
 633  *    lower_barrier when the particular background IO completes.
 634  */
 635
 636 static void raise_barrier(conf_t *conf, int force)
 637 {
 638         BUG_ON(force && !conf->barrier);
 639         spin_lock_irq(&conf->resync_lock);
 640
 641         /* Wait until no block IO is waiting (unless 'force') */
 642         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 643                             conf->resync_lock, );
 644
 645         /* block any new IO from starting */
 646         conf->barrier++;
 647
 648         /* Now wait for all pending IO to complete */
 649         wait_event_lock_irq(conf->wait_barrier,
 650                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 651                             conf->resync_lock, );
 652
 653         spin_unlock_irq(&conf->resync_lock);
 654 }
 655
 656 static void lower_barrier(conf_t *conf)
 657 {
 658         unsigned long flags;
 659         spin_lock_irqsave(&conf->resync_lock, flags);
 660         conf->barrier--;
 661         spin_unlock_irqrestore(&conf->resync_lock, flags);
 662         wake_up(&conf->wait_barrier);
 663 }
 664
 665 static void wait_barrier(conf_t *conf)
 666 {
 667         spin_lock_irq(&conf->resync_lock);
 668         if (conf->barrier) {
 669                 conf->nr_waiting++;
 670                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 671                                     conf->resync_lock,
 672                                     );
 673                 conf->nr_waiting--;
 674         }
 675         conf->nr_pending++;
 676         spin_unlock_irq(&conf->resync_lock);
 677 }
 678
 679 static void allow_barrier(conf_t *conf)
 680 {
 681         unsigned long flags;
 682         spin_lock_irqsave(&conf->resync_lock, flags);
 683         conf->nr_pending--;
 684         spin_unlock_irqrestore(&conf->resync_lock, flags);
 685         wake_up(&conf->wait_barrier);
 686 }
 687
 688 static void freeze_array(conf_t *conf)
 689 {
 690         /* stop syncio and normal IO and wait for everything to
 691          * go quiet.
 692          * We increment barrier and nr_waiting, and then
 693          * wait until nr_pending match nr_queued+1
 694          * This is called in the context of one normal IO request
 695          * that has failed. Thus any sync request that might be pending
 696          * will be blocked by nr_pending, and we need to wait for
 697          * pending IO requests to complete or be queued for re-try.
 698          * Thus the number queued (nr_queued) plus this request (1)
 699          * must match the number of pending IOs (nr_pending) before
 700          * we continue.
 701          */
 702         spin_lock_irq(&conf->resync_lock);
 703         conf->barrier++;
 704         conf->nr_waiting++;
 705         wait_event_lock_irq(conf->wait_barrier,
 706                             conf->nr_pending == conf->nr_queued+1,
 707                             conf->resync_lock,
 708                             flush_pending_writes(conf));
 709
 710         spin_unlock_irq(&conf->resync_lock);
 711 }
 712
 713 static void unfreeze_array(conf_t *conf)
 714 {
 715         /* reverse the effect of the freeze */
 716         spin_lock_irq(&conf->resync_lock);
 717         conf->barrier--;
 718         conf->nr_waiting--;
 719         wake_up(&conf->wait_barrier);
 720         spin_unlock_irq(&conf->resync_lock);
 721 }
 722
 723 static int make_request(mddev_t *mddev, struct bio * bio)
 724 {
 725         conf_t *conf = mddev->private;
 726         mirror_info_t *mirror;
 727         r10bio_t *r10_bio;
 728         struct bio *read_bio;
 729         int i;
 730         int chunk_sects = conf->chunk_mask + 1;
 731         const int rw = bio_data_dir(bio);
 732         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 733         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
 734         unsigned long flags;
 735         mdk_rdev_t *blocked_rdev;
 736         int plugged;
 737
 738         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
 739                 md_flush_request(mddev, bio);
 740                 return 0;
 741         }
 742
 743         /* If this request crosses a chunk boundary, we need to
 744          * split it.  This will only happen for 1 PAGE (or less) requests.
 745          */
 746         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 747                       > chunk_sects &&
 748                     conf->near_copies < conf->raid_disks)) {
 749                 struct bio_pair *bp;
 750                 /* Sanity check -- queue functions should prevent this happening */
 751                 if (bio->bi_vcnt != 1 ||
 752                     bio->bi_idx != 0)
 753                         goto bad_map;
 754                 /* This is a one page bio that upper layers
 755                  * refuse to split for us, so we need to split it.
 756                  */
 757                 bp = bio_split(bio,
 758                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 759
 760                 /* Each of these 'make_request' calls will call 'wait_barrier'.
 761                  * If the first succeeds but the second blocks due to the resync
 762                  * thread raising the barrier, we will deadlock because the
 763                  * IO to the underlying device will be queued in generic_make_request
 764                  * and will never complete, so will never reduce nr_pending.
 765                  * So increment nr_waiting here so no new raise_barriers will
 766                  * succeed, and so the second wait_barrier cannot block.
 767                  */
 768                 spin_lock_irq(&conf->resync_lock);
 769                 conf->nr_waiting++;
 770                 spin_unlock_irq(&conf->resync_lock);
 771
 772                 if (make_request(mddev, &bp->bio1))
 773                         generic_make_request(&bp->bio1);
 774                 if (make_request(mddev, &bp->bio2))
 775                         generic_make_request(&bp->bio2);
 776
 777                 spin_lock_irq(&conf->resync_lock);
 778                 conf->nr_waiting--;
 779                 wake_up(&conf->wait_barrier);
 780                 spin_unlock_irq(&conf->resync_lock);
 781
 782                 bio_pair_release(bp);
 783                 return 0;
 784         bad_map:
 785                 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
 786                        " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
 787                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 788
 789                 bio_io_error(bio);
 790                 return 0;
 791         }
 792
 793         md_write_start(mddev, bio);
 794
 795         /*
 796          * Register the new request and wait if the reconstruction
 797          * thread has put up a bar for new requests.
 798          * Continue immediately if no resync is active currently.
 799          */
 800         wait_barrier(conf);
 801
 802         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 803
 804         r10_bio->master_bio = bio;
 805         r10_bio->sectors = bio->bi_size >> 9;
 806
 807         r10_bio->mddev = mddev;
 808         r10_bio->sector = bio->bi_sector;
 809         r10_bio->state = 0;
 810
 811         if (rw == READ) {
 812                 /*
 813                  * read balancing logic:
 814                  */
 815                 int disk = read_balance(conf, r10_bio);
 816                 int slot = r10_bio->read_slot;
 817                 if (disk < 0) {
 818                         raid_end_bio_io(r10_bio);
 819                         return 0;
 820                 }
 821                 mirror = conf->mirrors + disk;
 822
 823                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 824
 825                 r10_bio->devs[slot].bio = read_bio;
 826
 827                 read_bio->bi_sector = r10_bio->devs[slot].addr +
 828                         mirror->rdev->data_offset;
 829                 read_bio->bi_bdev = mirror->rdev->bdev;
 830                 read_bio->bi_end_io = raid10_end_read_request;
 831                 read_bio->bi_rw = READ | do_sync;
 832                 read_bio->bi_private = r10_bio;
 833
 834                 generic_make_request(read_bio);
 835                 return 0;
 836         }
 837
 838         /*
 839          * WRITE:
 840          */
 841         /* first select target devices under rcu_lock and
 842          * inc refcount on their rdev.  Record them by setting
 843          * bios[x] to bio
 844          */
 845         plugged = mddev_check_plugged(mddev);
 846
 847         raid10_find_phys(conf, r10_bio);
 848  retry_write:
 849         blocked_rdev = NULL;
 850         rcu_read_lock();
 851         for (i = 0;  i < conf->copies; i++) {
 852                 int d = r10_bio->devs[i].devnum;
 853                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
 854                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 855                         atomic_inc(&rdev->nr_pending);
 856                         blocked_rdev = rdev;
 857                         break;
 858                 }
 859                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 860                         atomic_inc(&rdev->nr_pending);
 861                         r10_bio->devs[i].bio = bio;
 862                 } else {
 863                         r10_bio->devs[i].bio = NULL;
 864                         set_bit(R10BIO_Degraded, &r10_bio->state);
 865                 }
 866         }
 867         rcu_read_unlock();
 868
 869         if (unlikely(blocked_rdev)) {
 870                 /* Have to wait for this device to get unblocked, then retry */
 871                 int j;
 872                 int d;
 873
 874                 for (j = 0; j < i; j++)
 875                         if (r10_bio->devs[j].bio) {
 876                                 d = r10_bio->devs[j].devnum;
 877                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
 878                         }
 879                 allow_barrier(conf);
 880                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 881                 wait_barrier(conf);
 882                 goto retry_write;
 883         }
 884
 885         atomic_set(&r10_bio->remaining, 1);
 886         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
 887
 888         for (i = 0; i < conf->copies; i++) {
 889                 struct bio *mbio;
 890                 int d = r10_bio->devs[i].devnum;
 891                 if (!r10_bio->devs[i].bio)
 892                         continue;
 893
 894                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 895                 r10_bio->devs[i].bio = mbio;
 896
 897                 mbio->bi_sector = r10_bio->devs[i].addr+
 898                         conf->mirrors[d].rdev->data_offset;
 899                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
 900                 mbio->bi_end_io = raid10_end_write_request;
 901                 mbio->bi_rw = WRITE | do_sync | do_fua;
 902                 mbio->bi_private = r10_bio;
 903
 904                 atomic_inc(&r10_bio->remaining);
 905                 spin_lock_irqsave(&conf->device_lock, flags);
 906                 bio_list_add(&conf->pending_bio_list, mbio);
 907                 spin_unlock_irqrestore(&conf->device_lock, flags);
 908         }
 909
 910         if (atomic_dec_and_test(&r10_bio->remaining)) {
 911                 /* This matches the end of raid10_end_write_request() */
 912                 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 913                                 r10_bio->sectors,
 914                                 !test_bit(R10BIO_Degraded, &r10_bio->state),
 915                                 0);
 916                 md_write_end(mddev);
 917                 raid_end_bio_io(r10_bio);
 918         }
 919
 920         /* In case raid10d snuck in to freeze_array */
 921         wake_up(&conf->wait_barrier);
 922
 923         if (do_sync || !mddev->bitmap || !plugged)
 924                 md_wakeup_thread(mddev->thread);
 925         return 0;
 926 }
 927
 928 static void status(struct seq_file *seq, mddev_t *mddev)
 929 {
 930         conf_t *conf = mddev->private;
 931         int i;
 932
 933         if (conf->near_copies < conf->raid_disks)
 934                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
 935         if (conf->near_copies > 1)
 936                 seq_printf(seq, " %d near-copies", conf->near_copies);
 937         if (conf->far_copies > 1) {
 938                 if (conf->far_offset)
 939                         seq_printf(seq, " %d offset-copies", conf->far_copies);
 940                 else
 941                         seq_printf(seq, " %d far-copies", conf->far_copies);
 942         }
 943         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 944                                         conf->raid_disks - mddev->degraded);
 945         for (i = 0; i < conf->raid_disks; i++)
 946                 seq_printf(seq, "%s",
 947                               conf->mirrors[i].rdev &&
 948                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
 949         seq_printf(seq, "]");
 950 }
 951
 952 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 953 {
 954         char b[BDEVNAME_SIZE];
 955         conf_t *conf = mddev->private;
 956
 957         /*
 958          * If it is not operational, then we have already marked it as dead
 959          * else if it is the last working disks, ignore the error, let the
 960          * next level up know.
 961          * else mark the drive as failed
 962          */
 963         if (test_bit(In_sync, &rdev->flags)
 964             && conf->raid_disks-mddev->degraded == 1)
 965                 /*
 966                  * Don't fail the drive, just return an IO error.
 967                  * The test should really be more sophisticated than
 968                  * "working_disks == 1", but it isn't critical, and
 969                  * can wait until we do more sophisticated "is the drive
 970                  * really dead" tests...
 971                  */
 972                 return;
 973         if (test_and_clear_bit(In_sync, &rdev->flags)) {
 974                 unsigned long flags;
 975                 spin_lock_irqsave(&conf->device_lock, flags);
 976                 mddev->degraded++;
 977                 spin_unlock_irqrestore(&conf->device_lock, flags);
 978                 /*
 979                  * if recovery is running, make sure it aborts.
 980                  */
 981                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
 982         }
 983         set_bit(Faulty, &rdev->flags);
 984         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 985         printk(KERN_ALERT
 986                "md/raid10:%s: Disk failure on %s, disabling device.\n"
 987                "md/raid10:%s: Operation continuing on %d devices.\n",
 988                mdname(mddev), bdevname(rdev->bdev, b),
 989                mdname(mddev), conf->raid_disks - mddev->degraded);
 990 }
 991
 992 static void print_conf(conf_t *conf)
 993 {
 994         int i;
 995         mirror_info_t *tmp;
 996
 997         printk(KERN_DEBUG "RAID10 conf printout:\n");
 998         if (!conf) {
 999                 printk(KERN_DEBUG "(!conf)\n");
1000                 return;
1001         }
1002         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1003                 conf->raid_disks);
1004
1005         for (i = 0; i < conf->raid_disks; i++) {
1006                 char b[BDEVNAME_SIZE];
1007                 tmp = conf->mirrors + i;
1008                 if (tmp->rdev)
1009                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1010                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1011                                 !test_bit(Faulty, &tmp->rdev->flags),
1012                                 bdevname(tmp->rdev->bdev,b));
1013         }
1014 }
1015
1016 static void close_sync(conf_t *conf)
1017 {
1018         wait_barrier(conf);
1019         allow_barrier(conf);
1020
1021         mempool_destroy(conf->r10buf_pool);
1022         conf->r10buf_pool = NULL;
1023 }
1024
1025 /* check if there are enough drives for
1026  * every block to appear on atleast one
1027  */
1028 static int enough(conf_t *conf)
1029 {
1030         int first = 0;
1031
1032         do {
1033                 int n = conf->copies;
1034                 int cnt = 0;
1035                 while (n--) {
1036                         if (conf->mirrors[first].rdev)
1037                                 cnt++;
1038                         first = (first+1) % conf->raid_disks;
1039                 }
1040                 if (cnt == 0)
1041                         return 0;
1042         } while (first != 0);
1043         return 1;
1044 }
1045
1046 static int raid10_spare_active(mddev_t *mddev)
1047 {
1048         int i;
1049         conf_t *conf = mddev->private;
1050         mirror_info_t *tmp;
1051         int count = 0;
1052         unsigned long flags;
1053
1054         /*
1055          * Find all non-in_sync disks within the RAID10 configuration
1056          * and mark them in_sync
1057          */
1058         for (i = 0; i < conf->raid_disks; i++) {
1059                 tmp = conf->mirrors + i;
1060                 if (tmp->rdev
1061                     && !test_bit(Faulty, &tmp->rdev->flags)
1062                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1063                         count++;
1064                         sysfs_notify_dirent(tmp->rdev->sysfs_state);
1065                 }
1066         }
1067         spin_lock_irqsave(&conf->device_lock, flags);
1068         mddev->degraded -= count;
1069         spin_unlock_irqrestore(&conf->device_lock, flags);
1070
1071         print_conf(conf);
1072         return count;
1073 }
1074
1075
1076 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1077 {
1078         conf_t *conf = mddev->private;
1079         int err = -EEXIST;
1080         int mirror;
1081         mirror_info_t *p;
1082         int first = 0;
1083         int last = conf->raid_disks - 1;
1084
1085         if (mddev->recovery_cp < MaxSector)
1086                 /* only hot-add to in-sync arrays, as recovery is
1087                  * very different from resync
1088                  */
1089                 return -EBUSY;
1090         if (!enough(conf))
1091                 return -EINVAL;
1092
1093         if (rdev->raid_disk >= 0)
1094                 first = last = rdev->raid_disk;
1095
1096         if (rdev->saved_raid_disk >= 0 &&
1097             rdev->saved_raid_disk >= first &&
1098             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1099                 mirror = rdev->saved_raid_disk;
1100         else
1101                 mirror = first;
1102         for ( ; mirror <= last ; mirror++)
1103                 if ( !(p=conf->mirrors+mirror)->rdev) {
1104
1105                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1106                                           rdev->data_offset << 9);
1107                         /* as we don't honour merge_bvec_fn, we must
1108                          * never risk violating it, so limit
1109                          * ->max_segments to one lying with a single
1110                          * page, as a one page request is never in
1111                          * violation.
1112                          */
1113                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1114                                 blk_queue_max_segments(mddev->queue, 1);
1115                                 blk_queue_segment_boundary(mddev->queue,
1116                                                            PAGE_CACHE_SIZE - 1);
1117                         }
1118
1119                         p->head_position = 0;
1120                         rdev->raid_disk = mirror;
1121                         err = 0;
1122                         if (rdev->saved_raid_disk != mirror)
1123                                 conf->fullsync = 1;
1124                         rcu_assign_pointer(p->rdev, rdev);
1125                         break;
1126                 }
1127
1128         md_integrity_add_rdev(rdev, mddev);
1129         print_conf(conf);
1130         return err;
1131 }
1132
1133 static int raid10_remove_disk(mddev_t *mddev, int number)
1134 {
1135         conf_t *conf = mddev->private;
1136         int err = 0;
1137         mdk_rdev_t *rdev;
1138         mirror_info_t *p = conf->mirrors+ number;
1139
1140         print_conf(conf);
1141         rdev = p->rdev;
1142         if (rdev) {
1143                 if (test_bit(In_sync, &rdev->flags) ||
1144                     atomic_read(&rdev->nr_pending)) {
1145                         err = -EBUSY;
1146                         goto abort;
1147                 }
1148                 /* Only remove faulty devices in recovery
1149                  * is not possible.
1150                  */
1151                 if (!test_bit(Faulty, &rdev->flags) &&
1152                     enough(conf)) {
1153                         err = -EBUSY;
1154                         goto abort;
1155                 }
1156                 p->rdev = NULL;
1157                 synchronize_rcu();
1158                 if (atomic_read(&rdev->nr_pending)) {
1159                         /* lost the race, try later */
1160                         err = -EBUSY;
1161                         p->rdev = rdev;
1162                         goto abort;
1163                 }
1164                 err = md_integrity_register(mddev);
1165         }
1166 abort:
1167
1168         print_conf(conf);
1169         return err;
1170 }
1171
1172
1173 static void end_sync_read(struct bio *bio, int error)
1174 {
1175         r10bio_t *r10_bio = bio->bi_private;
1176         conf_t *conf = r10_bio->mddev->private;
1177         int i,d;
1178
1179         for (i=0; i<conf->copies; i++)
1180                 if (r10_bio->devs[i].bio == bio)
1181                         break;
1182         BUG_ON(i == conf->copies);
1183         update_head_pos(i, r10_bio);
1184         d = r10_bio->devs[i].devnum;
1185
1186         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1187                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1188         else {
1189                 atomic_add(r10_bio->sectors,
1190                            &conf->mirrors[d].rdev->corrected_errors);
1191                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1192                         md_error(r10_bio->mddev,
1193                                  conf->mirrors[d].rdev);
1194         }
1195
1196         /* for reconstruct, we always reschedule after a read.
1197          * for resync, only after all reads
1198          */
1199         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1200         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1201             atomic_dec_and_test(&r10_bio->remaining)) {
1202                 /* we have read all the blocks,
1203                  * do the comparison in process context in raid10d
1204                  */
1205                 reschedule_retry(r10_bio);
1206         }
1207 }
1208
1209 static void end_sync_write(struct bio *bio, int error)
1210 {
1211         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1212         r10bio_t *r10_bio = bio->bi_private;
1213         mddev_t *mddev = r10_bio->mddev;
1214         conf_t *conf = mddev->private;
1215         int i,d;
1216
1217         for (i = 0; i < conf->copies; i++)
1218                 if (r10_bio->devs[i].bio == bio)
1219                         break;
1220         d = r10_bio->devs[i].devnum;
1221
1222         if (!uptodate)
1223                 md_error(mddev, conf->mirrors[d].rdev);
1224
1225         update_head_pos(i, r10_bio);
1226
1227         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1228         while (atomic_dec_and_test(&r10_bio->remaining)) {
1229                 if (r10_bio->master_bio == NULL) {
1230                         /* the primary of several recovery bios */
1231                         sector_t s = r10_bio->sectors;
1232                         put_buf(r10_bio);
1233                         md_done_sync(mddev, s, 1);
1234                         break;
1235                 } else {
1236                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1237                         put_buf(r10_bio);
1238                         r10_bio = r10_bio2;
1239                 }
1240         }
1241 }
1242
1243 /*
1244  * Note: sync and recover and handled very differently for raid10
1245  * This code is for resync.
1246  * For resync, we read through virtual addresses and read all blocks.
1247  * If there is any error, we schedule a write.  The lowest numbered
1248  * drive is authoritative.
1249  * However requests come for physical address, so we need to map.
1250  * For every physical address there are raid_disks/copies virtual addresses,
1251  * which is always are least one, but is not necessarly an integer.
1252  * This means that a physical address can span multiple chunks, so we may
1253  * have to submit multiple io requests for a single sync request.
1254  */
1255 /*
1256  * We check if all blocks are in-sync and only write to blocks that
1257  * aren't in sync
1258  */
1259 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1260 {
1261         conf_t *conf = mddev->private;
1262         int i, first;
1263         struct bio *tbio, *fbio;
1264
1265         atomic_set(&r10_bio->remaining, 1);
1266
1267         /* find the first device with a block */
1268         for (i=0; i<conf->copies; i++)
1269                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1270                         break;
1271
1272         if (i == conf->copies)
1273                 goto done;
1274
1275         first = i;
1276         fbio = r10_bio->devs[i].bio;
1277
1278         /* now find blocks with errors */
1279         for (i=0 ; i < conf->copies ; i++) {
1280                 int  j, d;
1281                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1282
1283                 tbio = r10_bio->devs[i].bio;
1284
1285                 if (tbio->bi_end_io != end_sync_read)
1286                         continue;
1287                 if (i == first)
1288                         continue;
1289                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1290                         /* We know that the bi_io_vec layout is the same for
1291                          * both 'first' and 'i', so we just compare them.
1292                          * All vec entries are PAGE_SIZE;
1293                          */
1294                         for (j = 0; j < vcnt; j++)
1295                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1296                                            page_address(tbio->bi_io_vec[j].bv_page),
1297                                            PAGE_SIZE))
1298                                         break;
1299                         if (j == vcnt)
1300                                 continue;
1301                         mddev->resync_mismatches += r10_bio->sectors;
1302                 }
1303                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1304                         /* Don't fix anything. */
1305                         continue;
1306                 /* Ok, we need to write this bio
1307                  * First we need to fixup bv_offset, bv_len and
1308                  * bi_vecs, as the read request might have corrupted these
1309                  */
1310                 tbio->bi_vcnt = vcnt;
1311                 tbio->bi_size = r10_bio->sectors << 9;
1312                 tbio->bi_idx = 0;
1313                 tbio->bi_phys_segments = 0;
1314                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1315                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1316                 tbio->bi_next = NULL;
1317                 tbio->bi_rw = WRITE;
1318                 tbio->bi_private = r10_bio;
1319                 tbio->bi_sector = r10_bio->devs[i].addr;
1320
1321                 for (j=0; j < vcnt ; j++) {
1322                         tbio->bi_io_vec[j].bv_offset = 0;
1323                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1324
1325                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1326                                page_address(fbio->bi_io_vec[j].bv_page),
1327                                PAGE_SIZE);
1328                 }
1329                 tbio->bi_end_io = end_sync_write;
1330
1331                 d = r10_bio->devs[i].devnum;
1332                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1333                 atomic_inc(&r10_bio->remaining);
1334                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1335
1336                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1337                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1338                 generic_make_request(tbio);
1339         }
1340
1341 done:
1342         if (atomic_dec_and_test(&r10_bio->remaining)) {
1343                 md_done_sync(mddev, r10_bio->sectors, 1);
1344                 put_buf(r10_bio);
1345         }
1346 }
1347
1348 /*
1349  * Now for the recovery code.
1350  * Recovery happens across physical sectors.
1351  * We recover all non-is_sync drives by finding the virtual address of
1352  * each, and then choose a working drive that also has that virt address.
1353  * There is a separate r10_bio for each non-in_sync drive.
1354  * Only the first two slots are in use. The first for reading,
1355  * The second for writing.
1356  *
1357  */
1358
1359 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1360 {
1361         conf_t *conf = mddev->private;
1362         int i, d;
1363         struct bio *bio, *wbio;
1364
1365
1366         /* move the pages across to the second bio
1367          * and submit the write request
1368          */
1369         bio = r10_bio->devs[0].bio;
1370         wbio = r10_bio->devs[1].bio;
1371         for (i=0; i < wbio->bi_vcnt; i++) {
1372                 struct page *p = bio->bi_io_vec[i].bv_page;
1373                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1374                 wbio->bi_io_vec[i].bv_page = p;
1375         }
1376         d = r10_bio->devs[1].devnum;
1377
1378         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1379         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1380         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1381                 generic_make_request(wbio);
1382         else
1383                 bio_endio(wbio, -EIO);
1384 }
1385
1386
1387 /*
1388  * Used by fix_read_error() to decay the per rdev read_errors.
1389  * We halve the read error count for every hour that has elapsed
1390  * since the last recorded read error.
1391  *
1392  */
1393 static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1394 {
1395         struct timespec cur_time_mon;
1396         unsigned long hours_since_last;
1397         unsigned int read_errors = atomic_read(&rdev->read_errors);
1398
1399         ktime_get_ts(&cur_time_mon);
1400
1401         if (rdev->last_read_error.tv_sec == 0 &&
1402             rdev->last_read_error.tv_nsec == 0) {
1403                 /* first time we've seen a read error */
1404                 rdev->last_read_error = cur_time_mon;
1405                 return;
1406         }
1407
1408         hours_since_last = (cur_time_mon.tv_sec -
1409                             rdev->last_read_error.tv_sec) / 3600;
1410
1411         rdev->last_read_error = cur_time_mon;
1412
1413         /*
1414          * if hours_since_last is > the number of bits in read_errors
1415          * just set read errors to 0. We do this to avoid
1416          * overflowing the shift of read_errors by hours_since_last.
1417          */
1418         if (hours_since_last >= 8 * sizeof(read_errors))
1419                 atomic_set(&rdev->read_errors, 0);
1420         else
1421                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1422 }
1423
1424 /*
1425  * This is a kernel thread which:
1426  *
1427  *      1.      Retries failed read operations on working mirrors.
1428  *      2.      Updates the raid superblock when problems encounter.
1429  *      3.      Performs writes following reads for array synchronising.
1430  */
1431
1432 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1433 {
1434         int sect = 0; /* Offset from r10_bio->sector */
1435         int sectors = r10_bio->sectors;
1436         mdk_rdev_t*rdev;
1437         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1438         int d = r10_bio->devs[r10_bio->read_slot].devnum;
1439
1440         /* still own a reference to this rdev, so it cannot
1441          * have been cleared recently.
1442          */
1443         rdev = conf->mirrors[d].rdev;
1444
1445         if (test_bit(Faulty, &rdev->flags))
1446                 /* drive has already been failed, just ignore any
1447                    more fix_read_error() attempts */
1448                 return;
1449
1450         check_decay_read_errors(mddev, rdev);
1451         atomic_inc(&rdev->read_errors);
1452         if (atomic_read(&rdev->read_errors) > max_read_errors) {
1453                 char b[BDEVNAME_SIZE];
1454                 bdevname(rdev->bdev, b);
1455
1456                 printk(KERN_NOTICE
1457                        "md/raid10:%s: %s: Raid device exceeded "
1458                        "read_error threshold [cur %d:max %d]\n",
1459                        mdname(mddev), b,
1460                        atomic_read(&rdev->read_errors), max_read_errors);
1461                 printk(KERN_NOTICE
1462                        "md/raid10:%s: %s: Failing raid device\n",
1463                        mdname(mddev), b);
1464                 md_error(mddev, conf->mirrors[d].rdev);
1465                 return;
1466         }
1467
1468         while(sectors) {
1469                 int s = sectors;
1470                 int sl = r10_bio->read_slot;
1471                 int success = 0;
1472                 int start;
1473
1474                 if (s > (PAGE_SIZE>>9))
1475                         s = PAGE_SIZE >> 9;
1476
1477                 rcu_read_lock();
1478                 do {
1479                         d = r10_bio->devs[sl].devnum;
1480                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1481                         if (rdev &&
1482                             test_bit(In_sync, &rdev->flags)) {
1483                                 atomic_inc(&rdev->nr_pending);
1484                                 rcu_read_unlock();
1485                                 success = sync_page_io(rdev,
1486                                                        r10_bio->devs[sl].addr +
1487                                                        sect,
1488                                                        s<<9,
1489                                                        conf->tmppage, READ, false);
1490                                 rdev_dec_pending(rdev, mddev);
1491                                 rcu_read_lock();
1492                                 if (success)
1493                                         break;
1494                         }
1495                         sl++;
1496                         if (sl == conf->copies)
1497                                 sl = 0;
1498                 } while (!success && sl != r10_bio->read_slot);
1499                 rcu_read_unlock();
1500
1501                 if (!success) {
1502                         /* Cannot read from anywhere -- bye bye array */
1503                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1504                         md_error(mddev, conf->mirrors[dn].rdev);
1505                         break;
1506                 }
1507
1508                 start = sl;
1509                 /* write it back and re-read */
1510                 rcu_read_lock();
1511                 while (sl != r10_bio->read_slot) {
1512                         char b[BDEVNAME_SIZE];
1513
1514                         if (sl==0)
1515                                 sl = conf->copies;
1516                         sl--;
1517                         d = r10_bio->devs[sl].devnum;
1518                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1519                         if (rdev &&
1520                             test_bit(In_sync, &rdev->flags)) {
1521                                 atomic_inc(&rdev->nr_pending);
1522                                 rcu_read_unlock();
1523                                 atomic_add(s, &rdev->corrected_errors);
1524                                 if (sync_page_io(rdev,
1525                                                  r10_bio->devs[sl].addr +
1526                                                  sect,
1527                                                  s<<9, conf->tmppage, WRITE, false)
1528                                     == 0) {
1529                                         /* Well, this device is dead */
1530                                         printk(KERN_NOTICE
1531                                                "md/raid10:%s: read correction "
1532                                                "write failed"
1533                                                " (%d sectors at %llu on %s)\n",
1534                                                mdname(mddev), s,
1535                                                (unsigned long long)(
1536                                                        sect + rdev->data_offset),
1537                                                bdevname(rdev->bdev, b));
1538                                         printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1539                                                "drive\n",
1540                                                mdname(mddev),
1541                                                bdevname(rdev->bdev, b));
1542                                         md_error(mddev, rdev);
1543                                 }
1544                                 rdev_dec_pending(rdev, mddev);
1545                                 rcu_read_lock();
1546                         }
1547                 }
1548                 sl = start;
1549                 while (sl != r10_bio->read_slot) {
1550
1551                         if (sl==0)
1552                                 sl = conf->copies;
1553                         sl--;
1554                         d = r10_bio->devs[sl].devnum;
1555                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1556                         if (rdev &&
1557                             test_bit(In_sync, &rdev->flags)) {
1558                                 char b[BDEVNAME_SIZE];
1559                                 atomic_inc(&rdev->nr_pending);
1560                                 rcu_read_unlock();
1561                                 if (sync_page_io(rdev,
1562                                                  r10_bio->devs[sl].addr +
1563                                                  sect,
1564                                                  s<<9, conf->tmppage,
1565                                                  READ, false) == 0) {
1566                                         /* Well, this device is dead */
1567                                         printk(KERN_NOTICE
1568                                                "md/raid10:%s: unable to read back "
1569                                                "corrected sectors"
1570                                                " (%d sectors at %llu on %s)\n",
1571                                                mdname(mddev), s,
1572                                                (unsigned long long)(
1573                                                        sect + rdev->data_offset),
1574                                                bdevname(rdev->bdev, b));
1575                                         printk(KERN_NOTICE "md/raid10:%s: %s: failing drive\n",
1576                                                mdname(mddev),
1577                                                bdevname(rdev->bdev, b));
1578
1579                                         md_error(mddev, rdev);
1580                                 } else {
1581                                         printk(KERN_INFO
1582                                                "md/raid10:%s: read error corrected"
1583                                                " (%d sectors at %llu on %s)\n",
1584                                                mdname(mddev), s,
1585                                                (unsigned long long)(
1586                                                        sect + rdev->data_offset),
1587                                                bdevname(rdev->bdev, b));
1588                                 }
1589
1590                                 rdev_dec_pending(rdev, mddev);
1591                                 rcu_read_lock();
1592                         }
1593                 }
1594                 rcu_read_unlock();
1595
1596                 sectors -= s;
1597                 sect += s;
1598         }
1599 }
1600
1601 static void raid10d(mddev_t *mddev)
1602 {
1603         r10bio_t *r10_bio;
1604         struct bio *bio;
1605         unsigned long flags;
1606         conf_t *conf = mddev->private;
1607         struct list_head *head = &conf->retry_list;
1608         mdk_rdev_t *rdev;
1609         struct blk_plug plug;
1610
1611         md_check_recovery(mddev);
1612
1613         blk_start_plug(&plug);
1614         for (;;) {
1615                 char b[BDEVNAME_SIZE];
1616
1617                 flush_pending_writes(conf);
1618
1619                 spin_lock_irqsave(&conf->device_lock, flags);
1620                 if (list_empty(head)) {
1621                         spin_unlock_irqrestore(&conf->device_lock, flags);
1622                         break;
1623                 }
1624                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1625                 list_del(head->prev);
1626                 conf->nr_queued--;
1627                 spin_unlock_irqrestore(&conf->device_lock, flags);
1628
1629                 mddev = r10_bio->mddev;
1630                 conf = mddev->private;
1631                 if (test_bit(R10BIO_IsSync, &r10_bio->state))
1632                         sync_request_write(mddev, r10_bio);
1633                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
1634                         recovery_request_write(mddev, r10_bio);
1635                 else {
1636                         int slot = r10_bio->read_slot;
1637                         int mirror = r10_bio->devs[slot].devnum;
1638                         /* we got a read error. Maybe the drive is bad.  Maybe just
1639                          * the block and we can fix it.
1640                          * We freeze all other IO, and try reading the block from
1641                          * other devices.  When we find one, we re-write
1642                          * and check it that fixes the read error.
1643                          * This is all done synchronously while the array is
1644                          * frozen.
1645                          */
1646                         if (mddev->ro == 0) {
1647                                 freeze_array(conf);
1648                                 fix_read_error(conf, mddev, r10_bio);
1649                                 unfreeze_array(conf);
1650                         }
1651                         rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
1652
1653                         bio = r10_bio->devs[slot].bio;
1654                         r10_bio->devs[slot].bio =
1655                                 mddev->ro ? IO_BLOCKED : NULL;
1656                         mirror = read_balance(conf, r10_bio);
1657                         if (mirror == -1) {
1658                                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
1659                                        " read error for block %llu\n",
1660                                        mdname(mddev),
1661                                        bdevname(bio->bi_bdev,b),
1662                                        (unsigned long long)r10_bio->sector);
1663                                 raid_end_bio_io(r10_bio);
1664                                 bio_put(bio);
1665                         } else {
1666                                 const unsigned long do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
1667                                 bio_put(bio);
1668                                 slot = r10_bio->read_slot;
1669                                 rdev = conf->mirrors[mirror].rdev;
1670                                 if (printk_ratelimit())
1671                                         printk(KERN_ERR "md/raid10:%s: %s: redirecting sector %llu to"
1672                                                " another mirror\n",
1673                                                mdname(mddev),
1674                                                bdevname(rdev->bdev,b),
1675                                                (unsigned long long)r10_bio->sector);
1676                                 bio = bio_clone_mddev(r10_bio->master_bio,
1677                                                       GFP_NOIO, mddev);
1678                                 r10_bio->devs[slot].bio = bio;
1679                                 bio->bi_sector = r10_bio->devs[slot].addr
1680                                         + rdev->data_offset;
1681                                 bio->bi_bdev = rdev->bdev;
1682                                 bio->bi_rw = READ | do_sync;
1683                                 bio->bi_private = r10_bio;
1684                                 bio->bi_end_io = raid10_end_read_request;
1685                                 generic_make_request(bio);
1686                         }
1687                 }
1688                 cond_resched();
1689         }
1690         blk_finish_plug(&plug);
1691 }
1692
1693
1694 static int init_resync(conf_t *conf)
1695 {
1696         int buffs;
1697
1698         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1699         BUG_ON(conf->r10buf_pool);
1700         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1701         if (!conf->r10buf_pool)
1702                 return -ENOMEM;
1703         conf->next_resync = 0;
1704         return 0;
1705 }
1706
1707 /*
1708  * perform a "sync" on one "block"
1709  *
1710  * We need to make sure that no normal I/O request - particularly write
1711  * requests - conflict with active sync requests.
1712  *
1713  * This is achieved by tracking pending requests and a 'barrier' concept
1714  * that can be installed to exclude normal IO requests.
1715  *
1716  * Resync and recovery are handled very differently.
1717  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1718  *
1719  * For resync, we iterate over virtual addresses, read all copies,
1720  * and update if there are differences.  If only one copy is live,
1721  * skip it.
1722  * For recovery, we iterate over physical addresses, read a good
1723  * value for each non-in_sync drive, and over-write.
1724  *
1725  * So, for recovery we may have several outstanding complex requests for a
1726  * given address, one for each out-of-sync device.  We model this by allocating
1727  * a number of r10_bio structures, one for each out-of-sync device.
1728  * As we setup these structures, we collect all bio's together into a list
1729  * which we then process collectively to add pages, and then process again
1730  * to pass to generic_make_request.
1731  *
1732  * The r10_bio structures are linked using a borrowed master_bio pointer.
1733  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1734  * has its remaining count decremented to 0, the whole complex operation
1735  * is complete.
1736  *
1737  */
1738
1739 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr,
1740                              int *skipped, int go_faster)
1741 {
1742         conf_t *conf = mddev->private;
1743         r10bio_t *r10_bio;
1744         struct bio *biolist = NULL, *bio;
1745         sector_t max_sector, nr_sectors;
1746         int i;
1747         int max_sync;
1748         sector_t sync_blocks;
1749
1750         sector_t sectors_skipped = 0;
1751         int chunks_skipped = 0;
1752
1753         if (!conf->r10buf_pool)
1754                 if (init_resync(conf))
1755                         return 0;
1756
1757  skipped:
1758         max_sector = mddev->dev_sectors;
1759         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1760                 max_sector = mddev->resync_max_sectors;
1761         if (sector_nr >= max_sector) {
1762                 /* If we aborted, we need to abort the
1763                  * sync on the 'current' bitmap chucks (there can
1764                  * be several when recovering multiple devices).
1765                  * as we may have started syncing it but not finished.
1766                  * We can find the current address in
1767                  * mddev->curr_resync, but for recovery,
1768                  * we need to convert that to several
1769                  * virtual addresses.
1770                  */
1771                 if (mddev->curr_resync < max_sector) { /* aborted */
1772                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1773                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1774                                                 &sync_blocks, 1);
1775                         else for (i=0; i<conf->raid_disks; i++) {
1776                                 sector_t sect =
1777                                         raid10_find_virt(conf, mddev->curr_resync, i);
1778                                 bitmap_end_sync(mddev->bitmap, sect,
1779                                                 &sync_blocks, 1);
1780                         }
1781                 } else /* completed sync */
1782                         conf->fullsync = 0;
1783
1784                 bitmap_close_sync(mddev->bitmap);
1785                 close_sync(conf);
1786                 *skipped = 1;
1787                 return sectors_skipped;
1788         }
1789         if (chunks_skipped >= conf->raid_disks) {
1790                 /* if there has been nothing to do on any drive,
1791                  * then there is nothing to do at all..
1792                  */
1793                 *skipped = 1;
1794                 return (max_sector - sector_nr) + sectors_skipped;
1795         }
1796
1797         if (max_sector > mddev->resync_max)
1798                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1799
1800         /* make sure whole request will fit in a chunk - if chunks
1801          * are meaningful
1802          */
1803         if (conf->near_copies < conf->raid_disks &&
1804             max_sector > (sector_nr | conf->chunk_mask))
1805                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1806         /*
1807          * If there is non-resync activity waiting for us then
1808          * put in a delay to throttle resync.
1809          */
1810         if (!go_faster && conf->nr_waiting)
1811                 msleep_interruptible(1000);
1812
1813         /* Again, very different code for resync and recovery.
1814          * Both must result in an r10bio with a list of bios that
1815          * have bi_end_io, bi_sector, bi_bdev set,
1816          * and bi_private set to the r10bio.
1817          * For recovery, we may actually create several r10bios
1818          * with 2 bios in each, that correspond to the bios in the main one.
1819          * In this case, the subordinate r10bios link back through a
1820          * borrowed master_bio pointer, and the counter in the master
1821          * includes a ref from each subordinate.
1822          */
1823         /* First, we decide what to do and set ->bi_end_io
1824          * To end_sync_read if we want to read, and
1825          * end_sync_write if we will want to write.
1826          */
1827
1828         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1829         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1830                 /* recovery... the complicated one */
1831                 int j, k;
1832                 r10_bio = NULL;
1833
1834                 for (i=0 ; i<conf->raid_disks; i++) {
1835                         int still_degraded;
1836                         r10bio_t *rb2;
1837                         sector_t sect;
1838                         int must_sync;
1839
1840                         if (conf->mirrors[i].rdev == NULL ||
1841                             test_bit(In_sync, &conf->mirrors[i].rdev->flags))
1842                                 continue;
1843
1844                         still_degraded = 0;
1845                         /* want to reconstruct this device */
1846                         rb2 = r10_bio;
1847                         sect = raid10_find_virt(conf, sector_nr, i);
1848                         /* Unless we are doing a full sync, we only need
1849                          * to recover the block if it is set in the bitmap
1850                          */
1851                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
1852                                                       &sync_blocks, 1);
1853                         if (sync_blocks < max_sync)
1854                                 max_sync = sync_blocks;
1855                         if (!must_sync &&
1856                             !conf->fullsync) {
1857                                 /* yep, skip the sync_blocks here, but don't assume
1858                                  * that there will never be anything to do here
1859                                  */
1860                                 chunks_skipped = -1;
1861                                 continue;
1862                         }
1863
1864                         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1865                         raise_barrier(conf, rb2 != NULL);
1866                         atomic_set(&r10_bio->remaining, 0);
1867
1868                         r10_bio->master_bio = (struct bio*)rb2;
1869                         if (rb2)
1870                                 atomic_inc(&rb2->remaining);
1871                         r10_bio->mddev = mddev;
1872                         set_bit(R10BIO_IsRecover, &r10_bio->state);
1873                         r10_bio->sector = sect;
1874
1875                         raid10_find_phys(conf, r10_bio);
1876
1877                         /* Need to check if the array will still be
1878                          * degraded
1879                          */
1880                         for (j=0; j<conf->raid_disks; j++)
1881                                 if (conf->mirrors[j].rdev == NULL ||
1882                                     test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
1883                                         still_degraded = 1;
1884                                         break;
1885                                 }
1886
1887                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
1888                                                       &sync_blocks, still_degraded);
1889
1890                         for (j=0; j<conf->copies;j++) {
1891                                 int d = r10_bio->devs[j].devnum;
1892                                 if (!conf->mirrors[d].rdev ||
1893                                     !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
1894                                         continue;
1895                                 /* This is where we read from */
1896                                 bio = r10_bio->devs[0].bio;
1897                                 bio->bi_next = biolist;
1898                                 biolist = bio;
1899                                 bio->bi_private = r10_bio;
1900                                 bio->bi_end_io = end_sync_read;
1901                                 bio->bi_rw = READ;
1902                                 bio->bi_sector = r10_bio->devs[j].addr +
1903                                         conf->mirrors[d].rdev->data_offset;
1904                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1905                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1906                                 atomic_inc(&r10_bio->remaining);
1907                                 /* and we write to 'i' */
1908
1909                                 for (k=0; k<conf->copies; k++)
1910                                         if (r10_bio->devs[k].devnum == i)
1911                                                 break;
1912                                 BUG_ON(k == conf->copies);
1913                                 bio = r10_bio->devs[1].bio;
1914                                 bio->bi_next = biolist;
1915                                 biolist = bio;
1916                                 bio->bi_private = r10_bio;
1917                                 bio->bi_end_io = end_sync_write;
1918                                 bio->bi_rw = WRITE;
1919                                 bio->bi_sector = r10_bio->devs[k].addr +
1920                                         conf->mirrors[i].rdev->data_offset;
1921                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1922
1923                                 r10_bio->devs[0].devnum = d;
1924                                 r10_bio->devs[1].devnum = i;
1925
1926                                 break;
1927                         }
1928                         if (j == conf->copies) {
1929                                 /* Cannot recover, so abort the recovery */
1930                                 put_buf(r10_bio);
1931                                 if (rb2)
1932                                         atomic_dec(&rb2->remaining);
1933                                 r10_bio = rb2;
1934                                 if (!test_and_set_bit(MD_RECOVERY_INTR,
1935                                                       &mddev->recovery))
1936                                         printk(KERN_INFO "md/raid10:%s: insufficient "
1937                                                "working devices for recovery.\n",
1938                                                mdname(mddev));
1939                                 break;
1940                         }
1941                 }
1942                 if (biolist == NULL) {
1943                         while (r10_bio) {
1944                                 r10bio_t *rb2 = r10_bio;
1945                                 r10_bio = (r10bio_t*) rb2->master_bio;
1946                                 rb2->master_bio = NULL;
1947                                 put_buf(rb2);
1948                         }
1949                         goto giveup;
1950                 }
1951         } else {
1952                 /* resync. Schedule a read for every block at this virt offset */
1953                 int count = 0;
1954
1955                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1956
1957                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1958                                        &sync_blocks, mddev->degraded) &&
1959                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
1960                                                  &mddev->recovery)) {
1961                         /* We can skip this block */
1962                         *skipped = 1;
1963                         return sync_blocks + sectors_skipped;
1964                 }
1965                 if (sync_blocks < max_sync)
1966                         max_sync = sync_blocks;
1967                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1968
1969                 r10_bio->mddev = mddev;
1970                 atomic_set(&r10_bio->remaining, 0);
1971                 raise_barrier(conf, 0);
1972                 conf->next_resync = sector_nr;
1973
1974                 r10_bio->master_bio = NULL;
1975                 r10_bio->sector = sector_nr;
1976                 set_bit(R10BIO_IsSync, &r10_bio->state);
1977                 raid10_find_phys(conf, r10_bio);
1978                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1979
1980                 for (i=0; i<conf->copies; i++) {
1981                         int d = r10_bio->devs[i].devnum;
1982                         bio = r10_bio->devs[i].bio;
1983                         bio->bi_end_io = NULL;
1984                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
1985                         if (conf->mirrors[d].rdev == NULL ||
1986                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1987                                 continue;
1988                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1989                         atomic_inc(&r10_bio->remaining);
1990                         bio->bi_next = biolist;
1991                         biolist = bio;
1992                         bio->bi_private = r10_bio;
1993                         bio->bi_end_io = end_sync_read;
1994                         bio->bi_rw = READ;
1995                         bio->bi_sector = r10_bio->devs[i].addr +
1996                                 conf->mirrors[d].rdev->data_offset;
1997                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1998                         count++;
1999                 }
2000
2001                 if (count < 2) {
2002                         for (i=0; i<conf->copies; i++) {
2003                                 int d = r10_bio->devs[i].devnum;
2004                                 if (r10_bio->devs[i].bio->bi_end_io)
2005                                         rdev_dec_pending(conf->mirrors[d].rdev,
2006                                                          mddev);
2007                         }
2008                         put_buf(r10_bio);
2009                         biolist = NULL;
2010                         goto giveup;
2011                 }
2012         }
2013
2014         for (bio = biolist; bio ; bio=bio->bi_next) {
2015
2016                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2017                 if (bio->bi_end_io)
2018                         bio->bi_flags |= 1 << BIO_UPTODATE;
2019                 bio->bi_vcnt = 0;
2020                 bio->bi_idx = 0;
2021                 bio->bi_phys_segments = 0;
2022                 bio->bi_size = 0;
2023         }
2024
2025         nr_sectors = 0;
2026         if (sector_nr + max_sync < max_sector)
2027                 max_sector = sector_nr + max_sync;
2028         do {
2029                 struct page *page;
2030                 int len = PAGE_SIZE;
2031                 if (sector_nr + (len>>9) > max_sector)
2032                         len = (max_sector - sector_nr) << 9;
2033                 if (len == 0)
2034                         break;
2035                 for (bio= biolist ; bio ; bio=bio->bi_next) {
2036                         struct bio *bio2;
2037                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2038                         if (bio_add_page(bio, page, len, 0))
2039                                 continue;
2040
2041                         /* stop here */
2042                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2043                         for (bio2 = biolist;
2044                              bio2 && bio2 != bio;
2045                              bio2 = bio2->bi_next) {
2046                                 /* remove last page from this bio */
2047                                 bio2->bi_vcnt--;
2048                                 bio2->bi_size -= len;
2049                                 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2050                         }
2051                         goto bio_full;
2052                 }
2053                 nr_sectors += len>>9;
2054                 sector_nr += len>>9;
2055         } while (biolist->bi_vcnt < RESYNC_PAGES);
2056  bio_full:
2057         r10_bio->sectors = nr_sectors;
2058
2059         while (biolist) {
2060                 bio = biolist;
2061                 biolist = biolist->bi_next;
2062
2063                 bio->bi_next = NULL;
2064                 r10_bio = bio->bi_private;
2065                 r10_bio->sectors = nr_sectors;
2066
2067                 if (bio->bi_end_io == end_sync_read) {
2068                         md_sync_acct(bio->bi_bdev, nr_sectors);
2069                         generic_make_request(bio);
2070                 }
2071         }
2072
2073         if (sectors_skipped)
2074                 /* pretend they weren't skipped, it makes
2075                  * no important difference in this case
2076                  */
2077                 md_done_sync(mddev, sectors_skipped, 1);
2078
2079         return sectors_skipped + nr_sectors;
2080  giveup:
2081         /* There is nowhere to write, so all non-sync
2082          * drives must be failed, so try the next chunk...
2083          */
2084         if (sector_nr + max_sync < max_sector)
2085                 max_sector = sector_nr + max_sync;
2086
2087         sectors_skipped += (max_sector - sector_nr);
2088         chunks_skipped ++;
2089         sector_nr = max_sector;
2090         goto skipped;
2091 }
2092
2093 static sector_t
2094 raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2095 {
2096         sector_t size;
2097         conf_t *conf = mddev->private;
2098
2099         if (!raid_disks)
2100                 raid_disks = conf->raid_disks;
2101         if (!sectors)
2102                 sectors = conf->dev_sectors;
2103
2104         size = sectors >> conf->chunk_shift;
2105         sector_div(size, conf->far_copies);
2106         size = size * raid_disks;
2107         sector_div(size, conf->near_copies);
2108
2109         return size << conf->chunk_shift;
2110 }
2111
2112
2113 static conf_t *setup_conf(mddev_t *mddev)
2114 {
2115         conf_t *conf = NULL;
2116         int nc, fc, fo;
2117         sector_t stride, size;
2118         int err = -EINVAL;
2119
2120         if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2121             !is_power_of_2(mddev->new_chunk_sectors)) {
2122                 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2123                        "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2124                        mdname(mddev), PAGE_SIZE);
2125                 goto out;
2126         }
2127
2128         nc = mddev->new_layout & 255;
2129         fc = (mddev->new_layout >> 8) & 255;
2130         fo = mddev->new_layout & (1<<16);
2131
2132         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2133             (mddev->new_layout >> 17)) {
2134                 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2135                        mdname(mddev), mddev->new_layout);
2136                 goto out;
2137         }
2138
2139         err = -ENOMEM;
2140         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2141         if (!conf)
2142                 goto out;
2143
2144         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2145                                 GFP_KERNEL);
2146         if (!conf->mirrors)
2147                 goto out;
2148
2149         conf->tmppage = alloc_page(GFP_KERNEL);
2150         if (!conf->tmppage)
2151                 goto out;
2152
2153
2154         conf->raid_disks = mddev->raid_disks;
2155         conf->near_copies = nc;
2156         conf->far_copies = fc;
2157         conf->copies = nc*fc;
2158         conf->far_offset = fo;
2159         conf->chunk_mask = mddev->new_chunk_sectors - 1;
2160         conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2161
2162         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2163                                            r10bio_pool_free, conf);
2164         if (!conf->r10bio_pool)
2165                 goto out;
2166
2167         size = mddev->dev_sectors >> conf->chunk_shift;
2168         sector_div(size, fc);
2169         size = size * conf->raid_disks;
2170         sector_div(size, nc);
2171         /* 'size' is now the number of chunks in the array */
2172         /* calculate "used chunks per device" in 'stride' */
2173         stride = size * conf->copies;
2174
2175         /* We need to round up when dividing by raid_disks to
2176          * get the stride size.
2177          */
2178         stride += conf->raid_disks - 1;
2179         sector_div(stride, conf->raid_disks);
2180
2181         conf->dev_sectors = stride << conf->chunk_shift;
2182
2183         if (fo)
2184                 stride = 1;
2185         else
2186                 sector_div(stride, fc);
2187         conf->stride = stride << conf->chunk_shift;
2188
2189
2190         spin_lock_init(&conf->device_lock);
2191         INIT_LIST_HEAD(&conf->retry_list);
2192
2193         spin_lock_init(&conf->resync_lock);
2194         init_waitqueue_head(&conf->wait_barrier);
2195
2196         conf->thread = md_register_thread(raid10d, mddev, NULL);
2197         if (!conf->thread)
2198                 goto out;
2199
2200         conf->mddev = mddev;
2201         return conf;
2202
2203  out:
2204         printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2205                mdname(mddev));
2206         if (conf) {
2207                 if (conf->r10bio_pool)
2208                         mempool_destroy(conf->r10bio_pool);
2209                 kfree(conf->mirrors);
2210                 safe_put_page(conf->tmppage);
2211                 kfree(conf);
2212         }
2213         return ERR_PTR(err);
2214 }
2215
2216 static int run(mddev_t *mddev)
2217 {
2218         conf_t *conf;
2219         int i, disk_idx, chunk_size;
2220         mirror_info_t *disk;
2221         mdk_rdev_t *rdev;
2222         sector_t size;
2223
2224         /*
2225          * copy the already verified devices into our private RAID10
2226          * bookkeeping area. [whatever we allocate in run(),
2227          * should be freed in stop()]
2228          */
2229
2230         if (mddev->private == NULL) {
2231                 conf = setup_conf(mddev);
2232                 if (IS_ERR(conf))
2233                         return PTR_ERR(conf);
2234                 mddev->private = conf;
2235         }
2236         conf = mddev->private;
2237         if (!conf)
2238                 goto out;
2239
2240         mddev->thread = conf->thread;
2241         conf->thread = NULL;
2242
2243         chunk_size = mddev->chunk_sectors << 9;
2244         blk_queue_io_min(mddev->queue, chunk_size);
2245         if (conf->raid_disks % conf->near_copies)
2246                 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2247         else
2248                 blk_queue_io_opt(mddev->queue, chunk_size *
2249                                  (conf->raid_disks / conf->near_copies));
2250
2251         list_for_each_entry(rdev, &mddev->disks, same_set) {
2252                 disk_idx = rdev->raid_disk;
2253                 if (disk_idx >= conf->raid_disks
2254                     || disk_idx < 0)
2255                         continue;
2256                 disk = conf->mirrors + disk_idx;
2257
2258                 disk->rdev = rdev;
2259                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2260                                   rdev->data_offset << 9);
2261                 /* as we don't honour merge_bvec_fn, we must never risk
2262                  * violating it, so limit max_segments to 1 lying
2263                  * within a single page.
2264                  */
2265                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2266                         blk_queue_max_segments(mddev->queue, 1);
2267                         blk_queue_segment_boundary(mddev->queue,
2268                                                    PAGE_CACHE_SIZE - 1);
2269                 }
2270
2271                 disk->head_position = 0;
2272         }
2273         /* need to check that every block has at least one working mirror */
2274         if (!enough(conf)) {
2275                 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2276                        mdname(mddev));
2277                 goto out_free_conf;
2278         }
2279
2280         mddev->degraded = 0;
2281         for (i = 0; i < conf->raid_disks; i++) {
2282
2283                 disk = conf->mirrors + i;
2284
2285                 if (!disk->rdev ||
2286                     !test_bit(In_sync, &disk->rdev->flags)) {
2287                         disk->head_position = 0;
2288                         mddev->degraded++;
2289                         if (disk->rdev)
2290                                 conf->fullsync = 1;
2291                 }
2292         }
2293
2294         if (mddev->recovery_cp != MaxSector)
2295                 printk(KERN_NOTICE "md/raid10:%s: not clean"
2296                        " -- starting background reconstruction\n",
2297                        mdname(mddev));
2298         printk(KERN_INFO
2299                 "md/raid10:%s: active with %d out of %d devices\n",
2300                 mdname(mddev), conf->raid_disks - mddev->degraded,
2301                 conf->raid_disks);
2302         /*
2303          * Ok, everything is just fine now
2304          */
2305         mddev->dev_sectors = conf->dev_sectors;
2306         size = raid10_size(mddev, 0, 0);
2307         md_set_array_sectors(mddev, size);
2308         mddev->resync_max_sectors = size;
2309
2310         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2311         mddev->queue->backing_dev_info.congested_data = mddev;
2312
2313         /* Calculate max read-ahead size.
2314          * We need to readahead at least twice a whole stripe....
2315          * maybe...
2316          */
2317         {
2318                 int stripe = conf->raid_disks *
2319                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2320                 stripe /= conf->near_copies;
2321                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2322                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2323         }
2324
2325         if (conf->near_copies < conf->raid_disks)
2326                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2327
2328         if (md_integrity_register(mddev))
2329                 goto out_free_conf;
2330
2331         return 0;
2332
2333 out_free_conf:
2334         md_unregister_thread(mddev->thread);
2335         if (conf->r10bio_pool)
2336                 mempool_destroy(conf->r10bio_pool);
2337         safe_put_page(conf->tmppage);
2338         kfree(conf->mirrors);
2339         kfree(conf);
2340         mddev->private = NULL;
2341 out:
2342         return -EIO;
2343 }
2344
2345 static int stop(mddev_t *mddev)
2346 {
2347         conf_t *conf = mddev->private;
2348
2349         raise_barrier(conf, 0);
2350         lower_barrier(conf);
2351
2352         md_unregister_thread(mddev->thread);
2353         mddev->thread = NULL;
2354         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2355         if (conf->r10bio_pool)
2356                 mempool_destroy(conf->r10bio_pool);
2357         kfree(conf->mirrors);
2358         kfree(conf);
2359         mddev->private = NULL;
2360         return 0;
2361 }
2362
2363 static void raid10_quiesce(mddev_t *mddev, int state)
2364 {
2365         conf_t *conf = mddev->private;
2366
2367         switch(state) {
2368         case 1:
2369                 raise_barrier(conf, 0);
2370                 break;
2371         case 0:
2372                 lower_barrier(conf);
2373                 break;
2374         }
2375 }
2376
2377 static void *raid10_takeover_raid0(mddev_t *mddev)
2378 {
2379         mdk_rdev_t *rdev;
2380         conf_t *conf;
2381
2382         if (mddev->degraded > 0) {
2383                 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
2384                        mdname(mddev));
2385                 return ERR_PTR(-EINVAL);
2386         }
2387
2388         /* Set new parameters */
2389         mddev->new_level = 10;
2390         /* new layout: far_copies = 1, near_copies = 2 */
2391         mddev->new_layout = (1<<8) + 2;
2392         mddev->new_chunk_sectors = mddev->chunk_sectors;
2393         mddev->delta_disks = mddev->raid_disks;
2394         mddev->raid_disks *= 2;
2395         /* make sure it will be not marked as dirty */
2396         mddev->recovery_cp = MaxSector;
2397
2398         conf = setup_conf(mddev);
2399         if (!IS_ERR(conf)) {
2400                 list_for_each_entry(rdev, &mddev->disks, same_set)
2401                         if (rdev->raid_disk >= 0)
2402                                 rdev->new_raid_disk = rdev->raid_disk * 2;
2403                 conf->barrier = 1;
2404         }
2405
2406         return conf;
2407 }
2408
2409 static void *raid10_takeover(mddev_t *mddev)
2410 {
2411         struct raid0_private_data *raid0_priv;
2412
2413         /* raid10 can take over:
2414          *  raid0 - providing it has only two drives
2415          */
2416         if (mddev->level == 0) {
2417                 /* for raid0 takeover only one zone is supported */
2418                 raid0_priv = mddev->private;
2419                 if (raid0_priv->nr_strip_zones > 1) {
2420                         printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
2421                                " with more than one zone.\n",
2422                                mdname(mddev));
2423                         return ERR_PTR(-EINVAL);
2424                 }
2425                 return raid10_takeover_raid0(mddev);
2426         }
2427         return ERR_PTR(-EINVAL);
2428 }
2429
2430 static struct mdk_personality raid10_personality =
2431 {
2432         .name           = "raid10",
2433         .level          = 10,
2434         .owner          = THIS_MODULE,
2435         .make_request   = make_request,
2436         .run            = run,
2437         .stop           = stop,
2438         .status         = status,
2439         .error_handler  = error,
2440         .hot_add_disk   = raid10_add_disk,
2441         .hot_remove_disk= raid10_remove_disk,
2442         .spare_active   = raid10_spare_active,
2443         .sync_request   = sync_request,
2444         .quiesce        = raid10_quiesce,
2445         .size           = raid10_size,
2446         .takeover       = raid10_takeover,
2447 };
2448
2449 static int __init raid_init(void)
2450 {
2451         return register_md_personality(&raid10_personality);
2452 }
2453
2454 static void raid_exit(void)
2455 {
2456         unregister_md_personality(&raid10_personality);
2457 }
2458
2459 module_init(raid_init);
2460 module_exit(raid_exit);
2461 MODULE_LICENSE("GPL");
2462 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
2463 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2464 MODULE_ALIAS("md-raid10");
2465 MODULE_ALIAS("md-level-10");