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 <linux/ratelimit.h>
  26 #include "md.h"
  27 #include "raid10.h"
  28 #include "raid0.h"
  29 #include "bitmap.h"
  30
  31 /*
  32  * RAID10 provides a combination of RAID0 and RAID1 functionality.
  33  * The layout of data is defined by
  34  *    chunk_size
  35  *    raid_disks
  36  *    near_copies (stored in low byte of layout)
  37  *    far_copies (stored in second byte of layout)
  38  *    far_offset (stored in bit 16 of layout )
  39  *
  40  * The data to be stored is divided into chunks using chunksize.
  41  * Each device is divided into far_copies sections.
  42  * In each section, chunks are laid out in a style similar to raid0, but
  43  * near_copies copies of each chunk is stored (each on a different drive).
  44  * The starting device for each section is offset near_copies from the starting
  45  * device of the previous section.
  46  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
  47  * drive.
  48  * near_copies and far_copies must be at least one, and their product is at most
  49  * raid_disks.
  50  *
  51  * If far_offset is true, then the far_copies are handled a bit differently.
  52  * The copies are still in different stripes, but instead of be very far apart
  53  * on disk, there are adjacent stripes.
  54  */
  55
  56 /*
  57  * Number of guaranteed r10bios in case of extreme VM load:
  58  */
  59 #define NR_RAID10_BIOS 256
  60
  61 static void allow_barrier(conf_t *conf);
  62 static void lower_barrier(conf_t *conf);
  63
  64 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
  65 {
  66         conf_t *conf = data;
  67         int size = offsetof(struct r10bio_s, devs[conf->copies]);
  68
  69         /* allocate a r10bio with room for raid_disks entries in the bios array */
  70         return kzalloc(size, gfp_flags);
  71 }
  72
  73 static void r10bio_pool_free(void *r10_bio, void *data)
  74 {
  75         kfree(r10_bio);
  76 }
  77
  78 /* Maximum size of each resync request */
  79 #define RESYNC_BLOCK_SIZE (64*1024)
  80 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  81 /* amount of memory to reserve for resync requests */
  82 #define RESYNC_WINDOW (1024*1024)
  83 /* maximum number of concurrent requests, memory permitting */
  84 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
  85
  86 /*
  87  * When performing a resync, we need to read and compare, so
  88  * we need as many pages are there are copies.
  89  * When performing a recovery, we need 2 bios, one for read,
  90  * one for write (we recover only one drive per r10buf)
  91  *
  92  */
  93 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
  94 {
  95         conf_t *conf = data;
  96         struct page *page;
  97         r10bio_t *r10_bio;
  98         struct bio *bio;
  99         int i, j;
 100         int nalloc;
 101
 102         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 103         if (!r10_bio)
 104                 return NULL;
 105
 106         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
 107                 nalloc = conf->copies; /* resync */
 108         else
 109                 nalloc = 2; /* recovery */
 110
 111         /*
 112          * Allocate bios.
 113          */
 114         for (j = nalloc ; j-- ; ) {
 115                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 116                 if (!bio)
 117                         goto out_free_bio;
 118                 r10_bio->devs[j].bio = bio;
 119         }
 120         /*
 121          * Allocate RESYNC_PAGES data pages and attach them
 122          * where needed.
 123          */
 124         for (j = 0 ; j < nalloc; j++) {
 125                 bio = r10_bio->devs[j].bio;
 126                 for (i = 0; i < RESYNC_PAGES; i++) {
 127                         if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
 128                                                 &conf->mddev->recovery)) {
 129                                 /* we can share bv_page's during recovery */
 130                                 struct bio *rbio = r10_bio->devs[0].bio;
 131                                 page = rbio->bi_io_vec[i].bv_page;
 132                                 get_page(page);
 133                         } else
 134                                 page = alloc_page(gfp_flags);
 135                         if (unlikely(!page))
 136                                 goto out_free_pages;
 137
 138                         bio->bi_io_vec[i].bv_page = page;
 139                 }
 140         }
 141
 142         return r10_bio;
 143
 144 out_free_pages:
 145         for ( ; i > 0 ; i--)
 146                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
 147         while (j--)
 148                 for (i = 0; i < RESYNC_PAGES ; i++)
 149                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 150         j = -1;
 151 out_free_bio:
 152         while ( ++j < nalloc )
 153                 bio_put(r10_bio->devs[j].bio);
 154         r10bio_pool_free(r10_bio, conf);
 155         return NULL;
 156 }
 157
 158 static void r10buf_pool_free(void *__r10_bio, void *data)
 159 {
 160         int i;
 161         conf_t *conf = data;
 162         r10bio_t *r10bio = __r10_bio;
 163         int j;
 164
 165         for (j=0; j < conf->copies; j++) {
 166                 struct bio *bio = r10bio->devs[j].bio;
 167                 if (bio) {
 168                         for (i = 0; i < RESYNC_PAGES; i++) {
 169                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 170                                 bio->bi_io_vec[i].bv_page = NULL;
 171                         }
 172                         bio_put(bio);
 173                 }
 174         }
 175         r10bio_pool_free(r10bio, conf);
 176 }
 177
 178 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
 179 {
 180         int i;
 181
 182         for (i = 0; i < conf->copies; i++) {
 183                 struct bio **bio = & r10_bio->devs[i].bio;
 184                 if (*bio && *bio != IO_BLOCKED)
 185                         bio_put(*bio);
 186                 *bio = NULL;
 187         }
 188 }
 189
 190 static void free_r10bio(r10bio_t *r10_bio)
 191 {
 192         conf_t *conf = r10_bio->mddev->private;
 193
 194         /*
 195          * Wake up any possible resync thread that waits for the device
 196          * to go idle.
 197          */
 198         allow_barrier(conf);
 199
 200         put_all_bios(conf, r10_bio);
 201         mempool_free(r10_bio, conf->r10bio_pool);
 202 }
 203
 204 static void put_buf(r10bio_t *r10_bio)
 205 {
 206         conf_t *conf = r10_bio->mddev->private;
 207
 208         mempool_free(r10_bio, conf->r10buf_pool);
 209
 210         lower_barrier(conf);
 211 }
 212
 213 static void reschedule_retry(r10bio_t *r10_bio)
 214 {
 215         unsigned long flags;
 216         mddev_t *mddev = r10_bio->mddev;
 217         conf_t *conf = mddev->private;
 218
 219         spin_lock_irqsave(&conf->device_lock, flags);
 220         list_add(&r10_bio->retry_list, &conf->retry_list);
 221         conf->nr_queued ++;
 222         spin_unlock_irqrestore(&conf->device_lock, flags);
 223
 224         /* wake up frozen array... */
 225         wake_up(&conf->wait_barrier);
 226
 227         md_wakeup_thread(mddev->thread);
 228 }
 229
 230 /*
 231  * raid_end_bio_io() is called when we have finished servicing a mirrored
 232  * operation and are ready to return a success/failure code to the buffer
 233  * cache layer.
 234  */
 235 static void raid_end_bio_io(r10bio_t *r10_bio)
 236 {
 237         struct bio *bio = r10_bio->master_bio;
 238
 239         bio_endio(bio,
 240                 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
 241         free_r10bio(r10_bio);
 242 }
 243
 244 /*
 245  * Update disk head position estimator based on IRQ completion info.
 246  */
 247 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
 248 {
 249         conf_t *conf = r10_bio->mddev->private;
 250
 251         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 252                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 253 }
 254
 255 /*
 256  * Find the disk number which triggered given bio
 257  */
 258 static int find_bio_disk(conf_t *conf, r10bio_t *r10_bio, struct bio *bio)
 259 {
 260         int slot;
 261
 262         for (slot = 0; slot < conf->copies; slot++)
 263                 if (r10_bio->devs[slot].bio == bio)
 264                         break;
 265
 266         BUG_ON(slot == conf->copies);
 267         update_head_pos(slot, r10_bio);
 268
 269         return r10_bio->devs[slot].devnum;
 270 }
 271
 272 static void raid10_end_read_request(struct bio *bio, int error)
 273 {
 274         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 275         r10bio_t *r10_bio = bio->bi_private;
 276         int slot, dev;
 277         conf_t *conf = r10_bio->mddev->private;
 278
 279
 280         slot = r10_bio->read_slot;
 281         dev = r10_bio->devs[slot].devnum;
 282         /*
 283          * this branch is our 'one mirror IO has finished' event handler:
 284          */
 285         update_head_pos(slot, r10_bio);
 286
 287         if (uptodate) {
 288                 /*
 289                  * Set R10BIO_Uptodate in our master bio, so that
 290                  * we will return a good error code to the higher
 291                  * levels even if IO on some other mirrored buffer fails.
 292                  *
 293                  * The 'master' represents the composite IO operation to
 294                  * user-side. So if something waits for IO, then it will
 295                  * wait for the 'master' bio.
 296                  */
 297                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 298                 raid_end_bio_io(r10_bio);
 299                 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 300         } else {
 301                 /*
 302                  * oops, read error - keep the refcount on the rdev
 303                  */
 304                 char b[BDEVNAME_SIZE];
 305                 printk_ratelimited(KERN_ERR
 306                                    "md/raid10:%s: %s: rescheduling sector %llu\n",
 307                                    mdname(conf->mddev),
 308                                    bdevname(conf->mirrors[dev].rdev->bdev, b),
 309                                    (unsigned long long)r10_bio->sector);
 310                 reschedule_retry(r10_bio);
 311         }
 312 }
 313
 314 static void raid10_end_write_request(struct bio *bio, int error)
 315 {
 316         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 317         r10bio_t *r10_bio = bio->bi_private;
 318         int dev;
 319         conf_t *conf = r10_bio->mddev->private;
 320
 321         dev = find_bio_disk(conf, r10_bio, bio);
 322
 323         /*
 324          * this branch is our 'one mirror IO has finished' event handler:
 325          */
 326         if (!uptodate) {
 327                 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
 328                 /* an I/O failed, we can't clear the bitmap */
 329                 set_bit(R10BIO_Degraded, &r10_bio->state);
 330         } else
 331                 /*
 332                  * Set R10BIO_Uptodate in our master bio, so that
 333                  * we will return a good error code for to the higher
 334                  * levels even if IO on some other mirrored buffer fails.
 335                  *
 336                  * The 'master' represents the composite IO operation to
 337                  * user-side. So if something waits for IO, then it will
 338                  * wait for the 'master' bio.
 339                  */
 340                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 341
 342         /*
 343          *
 344          * Let's see if all mirrored write operations have finished
 345          * already.
 346          */
 347         if (atomic_dec_and_test(&r10_bio->remaining)) {
 348                 /* clear the bitmap if all writes complete successfully */
 349                 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 350                                 r10_bio->sectors,
 351                                 !test_bit(R10BIO_Degraded, &r10_bio->state),
 352                                 0);
 353                 md_write_end(r10_bio->mddev);
 354                 raid_end_bio_io(r10_bio);
 355         }
 356
 357         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 358 }
 359
 360
 361 /*
 362  * RAID10 layout manager
 363  * As well as the chunksize and raid_disks count, there are two
 364  * parameters: near_copies and far_copies.
 365  * near_copies * far_copies must be <= raid_disks.
 366  * Normally one of these will be 1.
 367  * If both are 1, we get raid0.
 368  * If near_copies == raid_disks, we get raid1.
 369  *
 370  * Chunks are laid out in raid0 style with near_copies copies of the
 371  * first chunk, followed by near_copies copies of the next chunk and
 372  * so on.
 373  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 374  * as described above, we start again with a device offset of near_copies.
 375  * So we effectively have another copy of the whole array further down all
 376  * the drives, but with blocks on different drives.
 377  * With this layout, and block is never stored twice on the one device.
 378  *
 379  * raid10_find_phys finds the sector offset of a given virtual sector
 380  * on each device that it is on.
 381  *
 382  * raid10_find_virt does the reverse mapping, from a device and a
 383  * sector offset to a virtual address
 384  */
 385
 386 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
 387 {
 388         int n,f;
 389         sector_t sector;
 390         sector_t chunk;
 391         sector_t stripe;
 392         int dev;
 393
 394         int slot = 0;
 395
 396         /* now calculate first sector/dev */
 397         chunk = r10bio->sector >> conf->chunk_shift;
 398         sector = r10bio->sector & conf->chunk_mask;
 399
 400         chunk *= conf->near_copies;
 401         stripe = chunk;
 402         dev = sector_div(stripe, conf->raid_disks);
 403         if (conf->far_offset)
 404                 stripe *= conf->far_copies;
 405
 406         sector += stripe << conf->chunk_shift;
 407
 408         /* and calculate all the others */
 409         for (n=0; n < conf->near_copies; n++) {
 410                 int d = dev;
 411                 sector_t s = sector;
 412                 r10bio->devs[slot].addr = sector;
 413                 r10bio->devs[slot].devnum = d;
 414                 slot++;
 415
 416                 for (f = 1; f < conf->far_copies; f++) {
 417                         d += conf->near_copies;
 418                         if (d >= conf->raid_disks)
 419                                 d -= conf->raid_disks;
 420                         s += conf->stride;
 421                         r10bio->devs[slot].devnum = d;
 422                         r10bio->devs[slot].addr = s;
 423                         slot++;
 424                 }
 425                 dev++;
 426                 if (dev >= conf->raid_disks) {
 427                         dev = 0;
 428                         sector += (conf->chunk_mask + 1);
 429                 }
 430         }
 431         BUG_ON(slot != conf->copies);
 432 }
 433
 434 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
 435 {
 436         sector_t offset, chunk, vchunk;
 437
 438         offset = sector & conf->chunk_mask;
 439         if (conf->far_offset) {
 440                 int fc;
 441                 chunk = sector >> conf->chunk_shift;
 442                 fc = sector_div(chunk, conf->far_copies);
 443                 dev -= fc * conf->near_copies;
 444                 if (dev < 0)
 445                         dev += conf->raid_disks;
 446         } else {
 447                 while (sector >= conf->stride) {
 448                         sector -= conf->stride;
 449                         if (dev < conf->near_copies)
 450                                 dev += conf->raid_disks - conf->near_copies;
 451                         else
 452                                 dev -= conf->near_copies;
 453                 }
 454                 chunk = sector >> conf->chunk_shift;
 455         }
 456         vchunk = chunk * conf->raid_disks + dev;
 457         sector_div(vchunk, conf->near_copies);
 458         return (vchunk << conf->chunk_shift) + offset;
 459 }
 460
 461 /**
 462  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 463  *      @q: request queue
 464  *      @bvm: properties of new bio
 465  *      @biovec: the request that could be merged to it.
 466  *
 467  *      Return amount of bytes we can accept at this offset
 468  *      If near_copies == raid_disk, there are no striping issues,
 469  *      but in that case, the function isn't called at all.
 470  */
 471 static int raid10_mergeable_bvec(struct request_queue *q,
 472                                  struct bvec_merge_data *bvm,
 473                                  struct bio_vec *biovec)
 474 {
 475         mddev_t *mddev = q->queuedata;
 476         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 477         int max;
 478         unsigned int chunk_sectors = mddev->chunk_sectors;
 479         unsigned int bio_sectors = bvm->bi_size >> 9;
 480
 481         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 482         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 483         if (max <= biovec->bv_len && bio_sectors == 0)
 484                 return biovec->bv_len;
 485         else
 486                 return max;
 487 }
 488
 489 /*
 490  * This routine returns the disk from which the requested read should
 491  * be done. There is a per-array 'next expected sequential IO' sector
 492  * number - if this matches on the next IO then we use the last disk.
 493  * There is also a per-disk 'last know head position' sector that is
 494  * maintained from IRQ contexts, both the normal and the resync IO
 495  * completion handlers update this position correctly. If there is no
 496  * perfect sequential match then we pick the disk whose head is closest.
 497  *
 498  * If there are 2 mirrors in the same 2 devices, performance degrades
 499  * because position is mirror, not device based.
 500  *
 501  * The rdev for the device selected will have nr_pending incremented.
 502  */
 503
 504 /*
 505  * FIXME: possibly should rethink readbalancing and do it differently
 506  * depending on near_copies / far_copies geometry.
 507  */
 508 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
 509 {
 510         const sector_t this_sector = r10_bio->sector;
 511         int disk, slot;
 512         const int sectors = r10_bio->sectors;
 513         sector_t new_distance, best_dist;
 514         mdk_rdev_t *rdev;
 515         int do_balance;
 516         int best_slot;
 517
 518         raid10_find_phys(conf, r10_bio);
 519         rcu_read_lock();
 520 retry:
 521         best_slot = -1;
 522         best_dist = MaxSector;
 523         do_balance = 1;
 524         /*
 525          * Check if we can balance. We can balance on the whole
 526          * device if no resync is going on (recovery is ok), or below
 527          * the resync window. We take the first readable disk when
 528          * above the resync window.
 529          */
 530         if (conf->mddev->recovery_cp < MaxSector
 531             && (this_sector + sectors >= conf->next_resync))
 532                 do_balance = 0;
 533
 534         for (slot = 0; slot < conf->copies ; slot++) {
 535                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
 536                         continue;
 537                 disk = r10_bio->devs[slot].devnum;
 538                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 539                 if (rdev == NULL)
 540                         continue;
 541                 if (!test_bit(In_sync, &rdev->flags))
 542                         continue;
 543
 544                 if (!do_balance)
 545                         break;
 546
 547                 /* This optimisation is debatable, and completely destroys
 548                  * sequential read speed for 'far copies' arrays.  So only
 549                  * keep it for 'near' arrays, and review those later.
 550                  */
 551                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
 552                         break;
 553
 554                 /* for far > 1 always use the lowest address */
 555                 if (conf->far_copies > 1)
 556                         new_distance = r10_bio->devs[slot].addr;
 557                 else
 558                         new_distance = abs(r10_bio->devs[slot].addr -
 559                                            conf->mirrors[disk].head_position);
 560                 if (new_distance < best_dist) {
 561                         best_dist = new_distance;
 562                         best_slot = slot;
 563                 }
 564         }
 565         if (slot == conf->copies)
 566                 slot = best_slot;
 567
 568         if (slot >= 0) {
 569                 disk = r10_bio->devs[slot].devnum;
 570                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 571                 if (!rdev)
 572                         goto retry;
 573                 atomic_inc(&rdev->nr_pending);
 574                 if (test_bit(Faulty, &rdev->flags)) {
 575                         /* Cannot risk returning a device that failed
 576                          * before we inc'ed nr_pending
 577                          */
 578                         rdev_dec_pending(rdev, conf->mddev);
 579                         goto retry;
 580                 }
 581                 r10_bio->read_slot = slot;
 582         } else
 583                 disk = -1;
 584         rcu_read_unlock();
 585
 586         return disk;
 587 }
 588
 589 static int raid10_congested(void *data, int bits)
 590 {
 591         mddev_t *mddev = data;
 592         conf_t *conf = mddev->private;
 593         int i, ret = 0;
 594
 595         if (mddev_congested(mddev, bits))
 596                 return 1;
 597         rcu_read_lock();
 598         for (i = 0; i < conf->raid_disks && ret == 0; i++) {
 599                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 600                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 601                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 602
 603                         ret |= bdi_congested(&q->backing_dev_info, bits);
 604                 }
 605         }
 606         rcu_read_unlock();
 607         return ret;
 608 }
 609
 610 static void flush_pending_writes(conf_t *conf)
 611 {
 612         /* Any writes that have been queued but are awaiting
 613          * bitmap updates get flushed here.
 614          */
 615         spin_lock_irq(&conf->device_lock);
 616
 617         if (conf->pending_bio_list.head) {
 618                 struct bio *bio;
 619                 bio = bio_list_get(&conf->pending_bio_list);
 620                 spin_unlock_irq(&conf->device_lock);
 621                 /* flush any pending bitmap writes to disk
 622                  * before proceeding w/ I/O */
 623                 bitmap_unplug(conf->mddev->bitmap);
 624
 625                 while (bio) { /* submit pending writes */
 626                         struct bio *next = bio->bi_next;
 627                         bio->bi_next = NULL;
 628                         generic_make_request(bio);
 629                         bio = next;
 630                 }
 631         } else
 632                 spin_unlock_irq(&conf->device_lock);
 633 }
 634
 635 /* Barriers....
 636  * Sometimes we need to suspend IO while we do something else,
 637  * either some resync/recovery, or reconfigure the array.
 638  * To do this we raise a 'barrier'.
 639  * The 'barrier' is a counter that can be raised multiple times
 640  * to count how many activities are happening which preclude
 641  * normal IO.
 642  * We can only raise the barrier if there is no pending IO.
 643  * i.e. if nr_pending == 0.
 644  * We choose only to raise the barrier if no-one is waiting for the
 645  * barrier to go down.  This means that as soon as an IO request
 646  * is ready, no other operations which require a barrier will start
 647  * until the IO request has had a chance.
 648  *
 649  * So: regular IO calls 'wait_barrier'.  When that returns there
 650  *    is no backgroup IO happening,  It must arrange to call
 651  *    allow_barrier when it has finished its IO.
 652  * backgroup IO calls must call raise_barrier.  Once that returns
 653  *    there is no normal IO happeing.  It must arrange to call
 654  *    lower_barrier when the particular background IO completes.
 655  */
 656
 657 static void raise_barrier(conf_t *conf, int force)
 658 {
 659         BUG_ON(force && !conf->barrier);
 660         spin_lock_irq(&conf->resync_lock);
 661
 662         /* Wait until no block IO is waiting (unless 'force') */
 663         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 664                             conf->resync_lock, );
 665
 666         /* block any new IO from starting */
 667         conf->barrier++;
 668
 669         /* Now wait for all pending IO to complete */
 670         wait_event_lock_irq(conf->wait_barrier,
 671                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 672                             conf->resync_lock, );
 673
 674         spin_unlock_irq(&conf->resync_lock);
 675 }
 676
 677 static void lower_barrier(conf_t *conf)
 678 {
 679         unsigned long flags;
 680         spin_lock_irqsave(&conf->resync_lock, flags);
 681         conf->barrier--;
 682         spin_unlock_irqrestore(&conf->resync_lock, flags);
 683         wake_up(&conf->wait_barrier);
 684 }
 685
 686 static void wait_barrier(conf_t *conf)
 687 {
 688         spin_lock_irq(&conf->resync_lock);
 689         if (conf->barrier) {
 690                 conf->nr_waiting++;
 691                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 692                                     conf->resync_lock,
 693                                     );
 694                 conf->nr_waiting--;
 695         }
 696         conf->nr_pending++;
 697         spin_unlock_irq(&conf->resync_lock);
 698 }
 699
 700 static void allow_barrier(conf_t *conf)
 701 {
 702         unsigned long flags;
 703         spin_lock_irqsave(&conf->resync_lock, flags);
 704         conf->nr_pending--;
 705         spin_unlock_irqrestore(&conf->resync_lock, flags);
 706         wake_up(&conf->wait_barrier);
 707 }
 708
 709 static void freeze_array(conf_t *conf)
 710 {
 711         /* stop syncio and normal IO and wait for everything to
 712          * go quiet.
 713          * We increment barrier and nr_waiting, and then
 714          * wait until nr_pending match nr_queued+1
 715          * This is called in the context of one normal IO request
 716          * that has failed. Thus any sync request that might be pending
 717          * will be blocked by nr_pending, and we need to wait for
 718          * pending IO requests to complete or be queued for re-try.
 719          * Thus the number queued (nr_queued) plus this request (1)
 720          * must match the number of pending IOs (nr_pending) before
 721          * we continue.
 722          */
 723         spin_lock_irq(&conf->resync_lock);
 724         conf->barrier++;
 725         conf->nr_waiting++;
 726         wait_event_lock_irq(conf->wait_barrier,
 727                             conf->nr_pending == conf->nr_queued+1,
 728                             conf->resync_lock,
 729                             flush_pending_writes(conf));
 730
 731         spin_unlock_irq(&conf->resync_lock);
 732 }
 733
 734 static void unfreeze_array(conf_t *conf)
 735 {
 736         /* reverse the effect of the freeze */
 737         spin_lock_irq(&conf->resync_lock);
 738         conf->barrier--;
 739         conf->nr_waiting--;
 740         wake_up(&conf->wait_barrier);
 741         spin_unlock_irq(&conf->resync_lock);
 742 }
 743
 744 static int make_request(mddev_t *mddev, struct bio * bio)
 745 {
 746         conf_t *conf = mddev->private;
 747         mirror_info_t *mirror;
 748         r10bio_t *r10_bio;
 749         struct bio *read_bio;
 750         int i;
 751         int chunk_sects = conf->chunk_mask + 1;
 752         const int rw = bio_data_dir(bio);
 753         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 754         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
 755         unsigned long flags;
 756         mdk_rdev_t *blocked_rdev;
 757         int plugged;
 758
 759         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
 760                 md_flush_request(mddev, bio);
 761                 return 0;
 762         }
 763
 764         /* If this request crosses a chunk boundary, we need to
 765          * split it.  This will only happen for 1 PAGE (or less) requests.
 766          */
 767         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 768                       > chunk_sects &&
 769                     conf->near_copies < conf->raid_disks)) {
 770                 struct bio_pair *bp;
 771                 /* Sanity check -- queue functions should prevent this happening */
 772                 if (bio->bi_vcnt != 1 ||
 773                     bio->bi_idx != 0)
 774                         goto bad_map;
 775                 /* This is a one page bio that upper layers
 776                  * refuse to split for us, so we need to split it.
 777                  */
 778                 bp = bio_split(bio,
 779                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 780
 781                 /* Each of these 'make_request' calls will call 'wait_barrier'.
 782                  * If the first succeeds but the second blocks due to the resync
 783                  * thread raising the barrier, we will deadlock because the
 784                  * IO to the underlying device will be queued in generic_make_request
 785                  * and will never complete, so will never reduce nr_pending.
 786                  * So increment nr_waiting here so no new raise_barriers will
 787                  * succeed, and so the second wait_barrier cannot block.
 788                  */
 789                 spin_lock_irq(&conf->resync_lock);
 790                 conf->nr_waiting++;
 791                 spin_unlock_irq(&conf->resync_lock);
 792
 793                 if (make_request(mddev, &bp->bio1))
 794                         generic_make_request(&bp->bio1);
 795                 if (make_request(mddev, &bp->bio2))
 796                         generic_make_request(&bp->bio2);
 797
 798                 spin_lock_irq(&conf->resync_lock);
 799                 conf->nr_waiting--;
 800                 wake_up(&conf->wait_barrier);
 801                 spin_unlock_irq(&conf->resync_lock);
 802
 803                 bio_pair_release(bp);
 804                 return 0;
 805         bad_map:
 806                 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
 807                        " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
 808                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 809
 810                 bio_io_error(bio);
 811                 return 0;
 812         }
 813
 814         md_write_start(mddev, bio);
 815
 816         /*
 817          * Register the new request and wait if the reconstruction
 818          * thread has put up a bar for new requests.
 819          * Continue immediately if no resync is active currently.
 820          */
 821         wait_barrier(conf);
 822
 823         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 824
 825         r10_bio->master_bio = bio;
 826         r10_bio->sectors = bio->bi_size >> 9;
 827
 828         r10_bio->mddev = mddev;
 829         r10_bio->sector = bio->bi_sector;
 830         r10_bio->state = 0;
 831
 832         if (rw == READ) {
 833                 /*
 834                  * read balancing logic:
 835                  */
 836                 int disk = read_balance(conf, r10_bio);
 837                 int slot = r10_bio->read_slot;
 838                 if (disk < 0) {
 839                         raid_end_bio_io(r10_bio);
 840                         return 0;
 841                 }
 842                 mirror = conf->mirrors + disk;
 843
 844                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 845
 846                 r10_bio->devs[slot].bio = read_bio;
 847
 848                 read_bio->bi_sector = r10_bio->devs[slot].addr +
 849                         mirror->rdev->data_offset;
 850                 read_bio->bi_bdev = mirror->rdev->bdev;
 851                 read_bio->bi_end_io = raid10_end_read_request;
 852                 read_bio->bi_rw = READ | do_sync;
 853                 read_bio->bi_private = r10_bio;
 854
 855                 generic_make_request(read_bio);
 856                 return 0;
 857         }
 858
 859         /*
 860          * WRITE:
 861          */
 862         /* first select target devices under rcu_lock and
 863          * inc refcount on their rdev.  Record them by setting
 864          * bios[x] to bio
 865          */
 866         plugged = mddev_check_plugged(mddev);
 867
 868         raid10_find_phys(conf, r10_bio);
 869  retry_write:
 870         blocked_rdev = NULL;
 871         rcu_read_lock();
 872         for (i = 0;  i < conf->copies; i++) {
 873                 int d = r10_bio->devs[i].devnum;
 874                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
 875                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 876                         atomic_inc(&rdev->nr_pending);
 877                         blocked_rdev = rdev;
 878                         break;
 879                 }
 880                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 881                         atomic_inc(&rdev->nr_pending);
 882                         r10_bio->devs[i].bio = bio;
 883                 } else {
 884                         r10_bio->devs[i].bio = NULL;
 885                         set_bit(R10BIO_Degraded, &r10_bio->state);
 886                 }
 887         }
 888         rcu_read_unlock();
 889
 890         if (unlikely(blocked_rdev)) {
 891                 /* Have to wait for this device to get unblocked, then retry */
 892                 int j;
 893                 int d;
 894
 895                 for (j = 0; j < i; j++)
 896                         if (r10_bio->devs[j].bio) {
 897                                 d = r10_bio->devs[j].devnum;
 898                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
 899                         }
 900                 allow_barrier(conf);
 901                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 902                 wait_barrier(conf);
 903                 goto retry_write;
 904         }
 905
 906         atomic_set(&r10_bio->remaining, 1);
 907         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
 908
 909         for (i = 0; i < conf->copies; i++) {
 910                 struct bio *mbio;
 911                 int d = r10_bio->devs[i].devnum;
 912                 if (!r10_bio->devs[i].bio)
 913                         continue;
 914
 915                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 916                 r10_bio->devs[i].bio = mbio;
 917
 918                 mbio->bi_sector = r10_bio->devs[i].addr+
 919                         conf->mirrors[d].rdev->data_offset;
 920                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
 921                 mbio->bi_end_io = raid10_end_write_request;
 922                 mbio->bi_rw = WRITE | do_sync | do_fua;
 923                 mbio->bi_private = r10_bio;
 924
 925                 atomic_inc(&r10_bio->remaining);
 926                 spin_lock_irqsave(&conf->device_lock, flags);
 927                 bio_list_add(&conf->pending_bio_list, mbio);
 928                 spin_unlock_irqrestore(&conf->device_lock, flags);
 929         }
 930
 931         if (atomic_dec_and_test(&r10_bio->remaining)) {
 932                 /* This matches the end of raid10_end_write_request() */
 933                 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 934                                 r10_bio->sectors,
 935                                 !test_bit(R10BIO_Degraded, &r10_bio->state),
 936                                 0);
 937                 md_write_end(mddev);
 938                 raid_end_bio_io(r10_bio);
 939         }
 940
 941         /* In case raid10d snuck in to freeze_array */
 942         wake_up(&conf->wait_barrier);
 943
 944         if (do_sync || !mddev->bitmap || !plugged)
 945                 md_wakeup_thread(mddev->thread);
 946         return 0;
 947 }
 948
 949 static void status(struct seq_file *seq, mddev_t *mddev)
 950 {
 951         conf_t *conf = mddev->private;
 952         int i;
 953
 954         if (conf->near_copies < conf->raid_disks)
 955                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
 956         if (conf->near_copies > 1)
 957                 seq_printf(seq, " %d near-copies", conf->near_copies);
 958         if (conf->far_copies > 1) {
 959                 if (conf->far_offset)
 960                         seq_printf(seq, " %d offset-copies", conf->far_copies);
 961                 else
 962                         seq_printf(seq, " %d far-copies", conf->far_copies);
 963         }
 964         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 965                                         conf->raid_disks - mddev->degraded);
 966         for (i = 0; i < conf->raid_disks; i++)
 967                 seq_printf(seq, "%s",
 968                               conf->mirrors[i].rdev &&
 969                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
 970         seq_printf(seq, "]");
 971 }
 972
 973 /* check if there are enough drives for
 974  * every block to appear on atleast one.
 975  * Don't consider the device numbered 'ignore'
 976  * as we might be about to remove it.
 977  */
 978 static int enough(conf_t *conf, int ignore)
 979 {
 980         int first = 0;
 981
 982         do {
 983                 int n = conf->copies;
 984                 int cnt = 0;
 985                 while (n--) {
 986                         if (conf->mirrors[first].rdev &&
 987                             first != ignore)
 988                                 cnt++;
 989                         first = (first+1) % conf->raid_disks;
 990                 }
 991                 if (cnt == 0)
 992                         return 0;
 993         } while (first != 0);
 994         return 1;
 995 }
 996
 997 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 998 {
 999         char b[BDEVNAME_SIZE];
1000         conf_t *conf = mddev->private;
1001
1002         /*
1003          * If it is not operational, then we have already marked it as dead
1004          * else if it is the last working disks, ignore the error, let the
1005          * next level up know.
1006          * else mark the drive as failed
1007          */
1008         if (test_bit(In_sync, &rdev->flags)
1009             && !enough(conf, rdev->raid_disk))
1010                 /*
1011                  * Don't fail the drive, just return an IO error.
1012                  */
1013                 return;
1014         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1015                 unsigned long flags;
1016                 spin_lock_irqsave(&conf->device_lock, flags);
1017                 mddev->degraded++;
1018                 spin_unlock_irqrestore(&conf->device_lock, flags);
1019                 /*
1020                  * if recovery is running, make sure it aborts.
1021                  */
1022                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1023         }
1024         set_bit(Faulty, &rdev->flags);
1025         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1026         printk(KERN_ALERT
1027                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1028                "md/raid10:%s: Operation continuing on %d devices.\n",
1029                mdname(mddev), bdevname(rdev->bdev, b),
1030                mdname(mddev), conf->raid_disks - mddev->degraded);
1031 }
1032
1033 static void print_conf(conf_t *conf)
1034 {
1035         int i;
1036         mirror_info_t *tmp;
1037
1038         printk(KERN_DEBUG "RAID10 conf printout:\n");
1039         if (!conf) {
1040                 printk(KERN_DEBUG "(!conf)\n");
1041                 return;
1042         }
1043         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1044                 conf->raid_disks);
1045
1046         for (i = 0; i < conf->raid_disks; i++) {
1047                 char b[BDEVNAME_SIZE];
1048                 tmp = conf->mirrors + i;
1049                 if (tmp->rdev)
1050                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1051                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1052                                 !test_bit(Faulty, &tmp->rdev->flags),
1053                                 bdevname(tmp->rdev->bdev,b));
1054         }
1055 }
1056
1057 static void close_sync(conf_t *conf)
1058 {
1059         wait_barrier(conf);
1060         allow_barrier(conf);
1061
1062         mempool_destroy(conf->r10buf_pool);
1063         conf->r10buf_pool = NULL;
1064 }
1065
1066 static int raid10_spare_active(mddev_t *mddev)
1067 {
1068         int i;
1069         conf_t *conf = mddev->private;
1070         mirror_info_t *tmp;
1071         int count = 0;
1072         unsigned long flags;
1073
1074         /*
1075          * Find all non-in_sync disks within the RAID10 configuration
1076          * and mark them in_sync
1077          */
1078         for (i = 0; i < conf->raid_disks; i++) {
1079                 tmp = conf->mirrors + i;
1080                 if (tmp->rdev
1081                     && !test_bit(Faulty, &tmp->rdev->flags)
1082                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1083                         count++;
1084                         sysfs_notify_dirent(tmp->rdev->sysfs_state);
1085                 }
1086         }
1087         spin_lock_irqsave(&conf->device_lock, flags);
1088         mddev->degraded -= count;
1089         spin_unlock_irqrestore(&conf->device_lock, flags);
1090
1091         print_conf(conf);
1092         return count;
1093 }
1094
1095
1096 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1097 {
1098         conf_t *conf = mddev->private;
1099         int err = -EEXIST;
1100         int mirror;
1101         int first = 0;
1102         int last = conf->raid_disks - 1;
1103
1104         if (mddev->recovery_cp < MaxSector)
1105                 /* only hot-add to in-sync arrays, as recovery is
1106                  * very different from resync
1107                  */
1108                 return -EBUSY;
1109         if (!enough(conf, -1))
1110                 return -EINVAL;
1111
1112         if (rdev->raid_disk >= 0)
1113                 first = last = rdev->raid_disk;
1114
1115         if (rdev->saved_raid_disk >= first &&
1116             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1117                 mirror = rdev->saved_raid_disk;
1118         else
1119                 mirror = first;
1120         for ( ; mirror <= last ; mirror++) {
1121                 mirror_info_t *p = &conf->mirrors[mirror];
1122                 if (p->recovery_disabled == mddev->recovery_disabled)
1123                         continue;
1124                 if (!p->rdev)
1125                         continue;
1126
1127                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1128                                   rdev->data_offset << 9);
1129                 /* as we don't honour merge_bvec_fn, we must
1130                  * never risk violating it, so limit
1131                  * ->max_segments to one lying with a single
1132                  * page, as a one page request is never in
1133                  * violation.
1134                  */
1135                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1136                         blk_queue_max_segments(mddev->queue, 1);
1137                         blk_queue_segment_boundary(mddev->queue,
1138                                                    PAGE_CACHE_SIZE - 1);
1139                 }
1140
1141                 p->head_position = 0;
1142                 rdev->raid_disk = mirror;
1143                 err = 0;
1144                 if (rdev->saved_raid_disk != mirror)
1145                         conf->fullsync = 1;
1146                 rcu_assign_pointer(p->rdev, rdev);
1147                 break;
1148         }
1149
1150         md_integrity_add_rdev(rdev, mddev);
1151         print_conf(conf);
1152         return err;
1153 }
1154
1155 static int raid10_remove_disk(mddev_t *mddev, int number)
1156 {
1157         conf_t *conf = mddev->private;
1158         int err = 0;
1159         mdk_rdev_t *rdev;
1160         mirror_info_t *p = conf->mirrors+ number;
1161
1162         print_conf(conf);
1163         rdev = p->rdev;
1164         if (rdev) {
1165                 if (test_bit(In_sync, &rdev->flags) ||
1166                     atomic_read(&rdev->nr_pending)) {
1167                         err = -EBUSY;
1168                         goto abort;
1169                 }
1170                 /* Only remove faulty devices in recovery
1171                  * is not possible.
1172                  */
1173                 if (!test_bit(Faulty, &rdev->flags) &&
1174                     mddev->recovery_disabled != p->recovery_disabled &&
1175                     enough(conf, -1)) {
1176                         err = -EBUSY;
1177                         goto abort;
1178                 }
1179                 p->rdev = NULL;
1180                 synchronize_rcu();
1181                 if (atomic_read(&rdev->nr_pending)) {
1182                         /* lost the race, try later */
1183                         err = -EBUSY;
1184                         p->rdev = rdev;
1185                         goto abort;
1186                 }
1187                 err = md_integrity_register(mddev);
1188         }
1189 abort:
1190
1191         print_conf(conf);
1192         return err;
1193 }
1194
1195
1196 static void end_sync_read(struct bio *bio, int error)
1197 {
1198         r10bio_t *r10_bio = bio->bi_private;
1199         conf_t *conf = r10_bio->mddev->private;
1200         int d;
1201
1202         d = find_bio_disk(conf, r10_bio, bio);
1203
1204         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1205                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1206         else {
1207                 atomic_add(r10_bio->sectors,
1208                            &conf->mirrors[d].rdev->corrected_errors);
1209                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1210                         md_error(r10_bio->mddev,
1211                                  conf->mirrors[d].rdev);
1212         }
1213
1214         /* for reconstruct, we always reschedule after a read.
1215          * for resync, only after all reads
1216          */
1217         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1218         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1219             atomic_dec_and_test(&r10_bio->remaining)) {
1220                 /* we have read all the blocks,
1221                  * do the comparison in process context in raid10d
1222                  */
1223                 reschedule_retry(r10_bio);
1224         }
1225 }
1226
1227 static void end_sync_write(struct bio *bio, int error)
1228 {
1229         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1230         r10bio_t *r10_bio = bio->bi_private;
1231         mddev_t *mddev = r10_bio->mddev;
1232         conf_t *conf = mddev->private;
1233         int d;
1234
1235         d = find_bio_disk(conf, r10_bio, bio);
1236
1237         if (!uptodate)
1238                 md_error(mddev, conf->mirrors[d].rdev);
1239
1240         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1241         while (atomic_dec_and_test(&r10_bio->remaining)) {
1242                 if (r10_bio->master_bio == NULL) {
1243                         /* the primary of several recovery bios */
1244                         sector_t s = r10_bio->sectors;
1245                         put_buf(r10_bio);
1246                         md_done_sync(mddev, s, 1);
1247                         break;
1248                 } else {
1249                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1250                         put_buf(r10_bio);
1251                         r10_bio = r10_bio2;
1252                 }
1253         }
1254 }
1255
1256 /*
1257  * Note: sync and recover and handled very differently for raid10
1258  * This code is for resync.
1259  * For resync, we read through virtual addresses and read all blocks.
1260  * If there is any error, we schedule a write.  The lowest numbered
1261  * drive is authoritative.
1262  * However requests come for physical address, so we need to map.
1263  * For every physical address there are raid_disks/copies virtual addresses,
1264  * which is always are least one, but is not necessarly an integer.
1265  * This means that a physical address can span multiple chunks, so we may
1266  * have to submit multiple io requests for a single sync request.
1267  */
1268 /*
1269  * We check if all blocks are in-sync and only write to blocks that
1270  * aren't in sync
1271  */
1272 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1273 {
1274         conf_t *conf = mddev->private;
1275         int i, first;
1276         struct bio *tbio, *fbio;
1277
1278         atomic_set(&r10_bio->remaining, 1);
1279
1280         /* find the first device with a block */
1281         for (i=0; i<conf->copies; i++)
1282                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1283                         break;
1284
1285         if (i == conf->copies)
1286                 goto done;
1287
1288         first = i;
1289         fbio = r10_bio->devs[i].bio;
1290
1291         /* now find blocks with errors */
1292         for (i=0 ; i < conf->copies ; i++) {
1293                 int  j, d;
1294                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1295
1296                 tbio = r10_bio->devs[i].bio;
1297
1298                 if (tbio->bi_end_io != end_sync_read)
1299                         continue;
1300                 if (i == first)
1301                         continue;
1302                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1303                         /* We know that the bi_io_vec layout is the same for
1304                          * both 'first' and 'i', so we just compare them.
1305                          * All vec entries are PAGE_SIZE;
1306                          */
1307                         for (j = 0; j < vcnt; j++)
1308                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1309                                            page_address(tbio->bi_io_vec[j].bv_page),
1310                                            PAGE_SIZE))
1311                                         break;
1312                         if (j == vcnt)
1313                                 continue;
1314                         mddev->resync_mismatches += r10_bio->sectors;
1315                 }
1316                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1317                         /* Don't fix anything. */
1318                         continue;
1319                 /* Ok, we need to write this bio
1320                  * First we need to fixup bv_offset, bv_len and
1321                  * bi_vecs, as the read request might have corrupted these
1322                  */
1323                 tbio->bi_vcnt = vcnt;
1324                 tbio->bi_size = r10_bio->sectors << 9;
1325                 tbio->bi_idx = 0;
1326                 tbio->bi_phys_segments = 0;
1327                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1328                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1329                 tbio->bi_next = NULL;
1330                 tbio->bi_rw = WRITE;
1331                 tbio->bi_private = r10_bio;
1332                 tbio->bi_sector = r10_bio->devs[i].addr;
1333
1334                 for (j=0; j < vcnt ; j++) {
1335                         tbio->bi_io_vec[j].bv_offset = 0;
1336                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1337
1338                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1339                                page_address(fbio->bi_io_vec[j].bv_page),
1340                                PAGE_SIZE);
1341                 }
1342                 tbio->bi_end_io = end_sync_write;
1343
1344                 d = r10_bio->devs[i].devnum;
1345                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1346                 atomic_inc(&r10_bio->remaining);
1347                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1348
1349                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1350                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1351                 generic_make_request(tbio);
1352         }
1353
1354 done:
1355         if (atomic_dec_and_test(&r10_bio->remaining)) {
1356                 md_done_sync(mddev, r10_bio->sectors, 1);
1357                 put_buf(r10_bio);
1358         }
1359 }
1360
1361 /*
1362  * Now for the recovery code.
1363  * Recovery happens across physical sectors.
1364  * We recover all non-is_sync drives by finding the virtual address of
1365  * each, and then choose a working drive that also has that virt address.
1366  * There is a separate r10_bio for each non-in_sync drive.
1367  * Only the first two slots are in use. The first for reading,
1368  * The second for writing.
1369  *
1370  */
1371
1372 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1373 {
1374         conf_t *conf = mddev->private;
1375         int d;
1376         struct bio *wbio;
1377
1378         /*
1379          * share the pages with the first bio
1380          * and submit the write request
1381          */
1382         wbio = r10_bio->devs[1].bio;
1383         d = r10_bio->devs[1].devnum;
1384
1385         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1386         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1387         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1388                 generic_make_request(wbio);
1389         else {
1390                 printk(KERN_NOTICE
1391                        "md/raid10:%s: recovery aborted due to read error\n",
1392                        mdname(mddev));
1393                 conf->mirrors[d].recovery_disabled = mddev->recovery_disabled;
1394                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1395                 bio_endio(wbio, 0);
1396         }
1397 }
1398
1399
1400 /*
1401  * Used by fix_read_error() to decay the per rdev read_errors.
1402  * We halve the read error count for every hour that has elapsed
1403  * since the last recorded read error.
1404  *
1405  */
1406 static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1407 {
1408         struct timespec cur_time_mon;
1409         unsigned long hours_since_last;
1410         unsigned int read_errors = atomic_read(&rdev->read_errors);
1411
1412         ktime_get_ts(&cur_time_mon);
1413
1414         if (rdev->last_read_error.tv_sec == 0 &&
1415             rdev->last_read_error.tv_nsec == 0) {
1416                 /* first time we've seen a read error */
1417                 rdev->last_read_error = cur_time_mon;
1418                 return;
1419         }
1420
1421         hours_since_last = (cur_time_mon.tv_sec -
1422                             rdev->last_read_error.tv_sec) / 3600;
1423
1424         rdev->last_read_error = cur_time_mon;
1425
1426         /*
1427          * if hours_since_last is > the number of bits in read_errors
1428          * just set read errors to 0. We do this to avoid
1429          * overflowing the shift of read_errors by hours_since_last.
1430          */
1431         if (hours_since_last >= 8 * sizeof(read_errors))
1432                 atomic_set(&rdev->read_errors, 0);
1433         else
1434                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1435 }
1436
1437 /*
1438  * This is a kernel thread which:
1439  *
1440  *      1.      Retries failed read operations on working mirrors.
1441  *      2.      Updates the raid superblock when problems encounter.
1442  *      3.      Performs writes following reads for array synchronising.
1443  */
1444
1445 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1446 {
1447         int sect = 0; /* Offset from r10_bio->sector */
1448         int sectors = r10_bio->sectors;
1449         mdk_rdev_t*rdev;
1450         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1451         int d = r10_bio->devs[r10_bio->read_slot].devnum;
1452
1453         /* still own a reference to this rdev, so it cannot
1454          * have been cleared recently.
1455          */
1456         rdev = conf->mirrors[d].rdev;
1457
1458         if (test_bit(Faulty, &rdev->flags))
1459                 /* drive has already been failed, just ignore any
1460                    more fix_read_error() attempts */
1461                 return;
1462
1463         check_decay_read_errors(mddev, rdev);
1464         atomic_inc(&rdev->read_errors);
1465         if (atomic_read(&rdev->read_errors) > max_read_errors) {
1466                 char b[BDEVNAME_SIZE];
1467                 bdevname(rdev->bdev, b);
1468
1469                 printk(KERN_NOTICE
1470                        "md/raid10:%s: %s: Raid device exceeded "
1471                        "read_error threshold [cur %d:max %d]\n",
1472                        mdname(mddev), b,
1473                        atomic_read(&rdev->read_errors), max_read_errors);
1474                 printk(KERN_NOTICE
1475                        "md/raid10:%s: %s: Failing raid device\n",
1476                        mdname(mddev), b);
1477                 md_error(mddev, conf->mirrors[d].rdev);
1478                 return;
1479         }
1480
1481         while(sectors) {
1482                 int s = sectors;
1483                 int sl = r10_bio->read_slot;
1484                 int success = 0;
1485                 int start;
1486
1487                 if (s > (PAGE_SIZE>>9))
1488                         s = PAGE_SIZE >> 9;
1489
1490                 rcu_read_lock();
1491                 do {
1492                         d = r10_bio->devs[sl].devnum;
1493                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1494                         if (rdev &&
1495                             test_bit(In_sync, &rdev->flags)) {
1496                                 atomic_inc(&rdev->nr_pending);
1497                                 rcu_read_unlock();
1498                                 success = sync_page_io(rdev,
1499                                                        r10_bio->devs[sl].addr +
1500                                                        sect,
1501                                                        s<<9,
1502                                                        conf->tmppage, READ, false);
1503                                 rdev_dec_pending(rdev, mddev);
1504                                 rcu_read_lock();
1505                                 if (success)
1506                                         break;
1507                         }
1508                         sl++;
1509                         if (sl == conf->copies)
1510                                 sl = 0;
1511                 } while (!success && sl != r10_bio->read_slot);
1512                 rcu_read_unlock();
1513
1514                 if (!success) {
1515                         /* Cannot read from anywhere -- bye bye array */
1516                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1517                         md_error(mddev, conf->mirrors[dn].rdev);
1518                         break;
1519                 }
1520
1521                 start = sl;
1522                 /* write it back and re-read */
1523                 rcu_read_lock();
1524                 while (sl != r10_bio->read_slot) {
1525                         char b[BDEVNAME_SIZE];
1526
1527                         if (sl==0)
1528                                 sl = conf->copies;
1529                         sl--;
1530                         d = r10_bio->devs[sl].devnum;
1531                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1532                         if (rdev &&
1533                             test_bit(In_sync, &rdev->flags)) {
1534                                 atomic_inc(&rdev->nr_pending);
1535                                 rcu_read_unlock();
1536                                 if (sync_page_io(rdev,
1537                                                  r10_bio->devs[sl].addr +
1538                                                  sect,
1539                                                  s<<9, conf->tmppage, WRITE, false)
1540                                     == 0) {
1541                                         /* Well, this device is dead */
1542                                         printk(KERN_NOTICE
1543                                                "md/raid10:%s: read correction "
1544                                                "write failed"
1545                                                " (%d sectors at %llu on %s)\n",
1546                                                mdname(mddev), s,
1547                                                (unsigned long long)(
1548                                                        sect + rdev->data_offset),
1549                                                bdevname(rdev->bdev, b));
1550                                         printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1551                                                "drive\n",
1552                                                mdname(mddev),
1553                                                bdevname(rdev->bdev, b));
1554                                         md_error(mddev, rdev);
1555                                 }
1556                                 rdev_dec_pending(rdev, mddev);
1557                                 rcu_read_lock();
1558                         }
1559                 }
1560                 sl = start;
1561                 while (sl != r10_bio->read_slot) {
1562
1563                         if (sl==0)
1564                                 sl = conf->copies;
1565                         sl--;
1566                         d = r10_bio->devs[sl].devnum;
1567                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1568                         if (rdev &&
1569                             test_bit(In_sync, &rdev->flags)) {
1570                                 char b[BDEVNAME_SIZE];
1571                                 atomic_inc(&rdev->nr_pending);
1572                                 rcu_read_unlock();
1573                                 if (sync_page_io(rdev,
1574                                                  r10_bio->devs[sl].addr +
1575                                                  sect,
1576                                                  s<<9, conf->tmppage,
1577                                                  READ, false) == 0) {
1578                                         /* Well, this device is dead */
1579                                         printk(KERN_NOTICE
1580                                                "md/raid10:%s: unable to read back "
1581                                                "corrected sectors"
1582                                                " (%d sectors at %llu on %s)\n",
1583                                                mdname(mddev), s,
1584                                                (unsigned long long)(
1585                                                        sect + rdev->data_offset),
1586                                                bdevname(rdev->bdev, b));
1587                                         printk(KERN_NOTICE "md/raid10:%s: %s: failing drive\n",
1588                                                mdname(mddev),
1589                                                bdevname(rdev->bdev, b));
1590
1591                                         md_error(mddev, rdev);
1592                                 } else {
1593                                         printk(KERN_INFO
1594                                                "md/raid10:%s: read error corrected"
1595                                                " (%d sectors at %llu on %s)\n",
1596                                                mdname(mddev), s,
1597                                                (unsigned long long)(
1598                                                        sect + rdev->data_offset),
1599                                                bdevname(rdev->bdev, b));
1600                                         atomic_add(s, &rdev->corrected_errors);
1601                                 }
1602
1603                                 rdev_dec_pending(rdev, mddev);
1604                                 rcu_read_lock();
1605                         }
1606                 }
1607                 rcu_read_unlock();
1608
1609                 sectors -= s;
1610                 sect += s;
1611         }
1612 }
1613
1614 static void raid10d(mddev_t *mddev)
1615 {
1616         r10bio_t *r10_bio;
1617         struct bio *bio;
1618         unsigned long flags;
1619         conf_t *conf = mddev->private;
1620         struct list_head *head = &conf->retry_list;
1621         mdk_rdev_t *rdev;
1622         struct blk_plug plug;
1623
1624         md_check_recovery(mddev);
1625
1626         blk_start_plug(&plug);
1627         for (;;) {
1628                 char b[BDEVNAME_SIZE];
1629
1630                 flush_pending_writes(conf);
1631
1632                 spin_lock_irqsave(&conf->device_lock, flags);
1633                 if (list_empty(head)) {
1634                         spin_unlock_irqrestore(&conf->device_lock, flags);
1635                         break;
1636                 }
1637                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1638                 list_del(head->prev);
1639                 conf->nr_queued--;
1640                 spin_unlock_irqrestore(&conf->device_lock, flags);
1641
1642                 mddev = r10_bio->mddev;
1643                 conf = mddev->private;
1644                 if (test_bit(R10BIO_IsSync, &r10_bio->state))
1645                         sync_request_write(mddev, r10_bio);
1646                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
1647                         recovery_request_write(mddev, r10_bio);
1648                 else {
1649                         int slot = r10_bio->read_slot;
1650                         int mirror = r10_bio->devs[slot].devnum;
1651                         /* we got a read error. Maybe the drive is bad.  Maybe just
1652                          * the block and we can fix it.
1653                          * We freeze all other IO, and try reading the block from
1654                          * other devices.  When we find one, we re-write
1655                          * and check it that fixes the read error.
1656                          * This is all done synchronously while the array is
1657                          * frozen.
1658                          */
1659                         if (mddev->ro == 0) {
1660                                 freeze_array(conf);
1661                                 fix_read_error(conf, mddev, r10_bio);
1662                                 unfreeze_array(conf);
1663                         }
1664                         rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
1665
1666                         bio = r10_bio->devs[slot].bio;
1667                         r10_bio->devs[slot].bio =
1668                                 mddev->ro ? IO_BLOCKED : NULL;
1669                         mirror = read_balance(conf, r10_bio);
1670                         if (mirror == -1) {
1671                                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
1672                                        " read error for block %llu\n",
1673                                        mdname(mddev),
1674                                        bdevname(bio->bi_bdev,b),
1675                                        (unsigned long long)r10_bio->sector);
1676                                 raid_end_bio_io(r10_bio);
1677                                 bio_put(bio);
1678                         } else {
1679                                 const unsigned long do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
1680                                 bio_put(bio);
1681                                 slot = r10_bio->read_slot;
1682                                 rdev = conf->mirrors[mirror].rdev;
1683                                 printk_ratelimited(
1684                                         KERN_ERR
1685                                         "md/raid10:%s: %s: redirecting"
1686                                         "sector %llu to another mirror\n",
1687                                         mdname(mddev),
1688                                         bdevname(rdev->bdev, b),
1689                                         (unsigned long long)r10_bio->sector);
1690                                 bio = bio_clone_mddev(r10_bio->master_bio,
1691                                                       GFP_NOIO, mddev);
1692                                 r10_bio->devs[slot].bio = bio;
1693                                 bio->bi_sector = r10_bio->devs[slot].addr
1694                                         + rdev->data_offset;
1695                                 bio->bi_bdev = rdev->bdev;
1696                                 bio->bi_rw = READ | do_sync;
1697                                 bio->bi_private = r10_bio;
1698                                 bio->bi_end_io = raid10_end_read_request;
1699                                 generic_make_request(bio);
1700                         }
1701                 }
1702                 cond_resched();
1703         }
1704         blk_finish_plug(&plug);
1705 }
1706
1707
1708 static int init_resync(conf_t *conf)
1709 {
1710         int buffs;
1711
1712         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1713         BUG_ON(conf->r10buf_pool);
1714         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1715         if (!conf->r10buf_pool)
1716                 return -ENOMEM;
1717         conf->next_resync = 0;
1718         return 0;
1719 }
1720
1721 /*
1722  * perform a "sync" on one "block"
1723  *
1724  * We need to make sure that no normal I/O request - particularly write
1725  * requests - conflict with active sync requests.
1726  *
1727  * This is achieved by tracking pending requests and a 'barrier' concept
1728  * that can be installed to exclude normal IO requests.
1729  *
1730  * Resync and recovery are handled very differently.
1731  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1732  *
1733  * For resync, we iterate over virtual addresses, read all copies,
1734  * and update if there are differences.  If only one copy is live,
1735  * skip it.
1736  * For recovery, we iterate over physical addresses, read a good
1737  * value for each non-in_sync drive, and over-write.
1738  *
1739  * So, for recovery we may have several outstanding complex requests for a
1740  * given address, one for each out-of-sync device.  We model this by allocating
1741  * a number of r10_bio structures, one for each out-of-sync device.
1742  * As we setup these structures, we collect all bio's together into a list
1743  * which we then process collectively to add pages, and then process again
1744  * to pass to generic_make_request.
1745  *
1746  * The r10_bio structures are linked using a borrowed master_bio pointer.
1747  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1748  * has its remaining count decremented to 0, the whole complex operation
1749  * is complete.
1750  *
1751  */
1752
1753 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr,
1754                              int *skipped, int go_faster)
1755 {
1756         conf_t *conf = mddev->private;
1757         r10bio_t *r10_bio;
1758         struct bio *biolist = NULL, *bio;
1759         sector_t max_sector, nr_sectors;
1760         int i;
1761         int max_sync;
1762         sector_t sync_blocks;
1763
1764         sector_t sectors_skipped = 0;
1765         int chunks_skipped = 0;
1766
1767         if (!conf->r10buf_pool)
1768                 if (init_resync(conf))
1769                         return 0;
1770
1771  skipped:
1772         max_sector = mddev->dev_sectors;
1773         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1774                 max_sector = mddev->resync_max_sectors;
1775         if (sector_nr >= max_sector) {
1776                 /* If we aborted, we need to abort the
1777                  * sync on the 'current' bitmap chucks (there can
1778                  * be several when recovering multiple devices).
1779                  * as we may have started syncing it but not finished.
1780                  * We can find the current address in
1781                  * mddev->curr_resync, but for recovery,
1782                  * we need to convert that to several
1783                  * virtual addresses.
1784                  */
1785                 if (mddev->curr_resync < max_sector) { /* aborted */
1786                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1787                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1788                                                 &sync_blocks, 1);
1789                         else for (i=0; i<conf->raid_disks; i++) {
1790                                 sector_t sect =
1791                                         raid10_find_virt(conf, mddev->curr_resync, i);
1792                                 bitmap_end_sync(mddev->bitmap, sect,
1793                                                 &sync_blocks, 1);
1794                         }
1795                 } else /* completed sync */
1796                         conf->fullsync = 0;
1797
1798                 bitmap_close_sync(mddev->bitmap);
1799                 close_sync(conf);
1800                 *skipped = 1;
1801                 return sectors_skipped;
1802         }
1803         if (chunks_skipped >= conf->raid_disks) {
1804                 /* if there has been nothing to do on any drive,
1805                  * then there is nothing to do at all..
1806                  */
1807                 *skipped = 1;
1808                 return (max_sector - sector_nr) + sectors_skipped;
1809         }
1810
1811         if (max_sector > mddev->resync_max)
1812                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1813
1814         /* make sure whole request will fit in a chunk - if chunks
1815          * are meaningful
1816          */
1817         if (conf->near_copies < conf->raid_disks &&
1818             max_sector > (sector_nr | conf->chunk_mask))
1819                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1820         /*
1821          * If there is non-resync activity waiting for us then
1822          * put in a delay to throttle resync.
1823          */
1824         if (!go_faster && conf->nr_waiting)
1825                 msleep_interruptible(1000);
1826
1827         /* Again, very different code for resync and recovery.
1828          * Both must result in an r10bio with a list of bios that
1829          * have bi_end_io, bi_sector, bi_bdev set,
1830          * and bi_private set to the r10bio.
1831          * For recovery, we may actually create several r10bios
1832          * with 2 bios in each, that correspond to the bios in the main one.
1833          * In this case, the subordinate r10bios link back through a
1834          * borrowed master_bio pointer, and the counter in the master
1835          * includes a ref from each subordinate.
1836          */
1837         /* First, we decide what to do and set ->bi_end_io
1838          * To end_sync_read if we want to read, and
1839          * end_sync_write if we will want to write.
1840          */
1841
1842         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1843         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1844                 /* recovery... the complicated one */
1845                 int j, k;
1846                 r10_bio = NULL;
1847
1848                 for (i=0 ; i<conf->raid_disks; i++) {
1849                         int still_degraded;
1850                         r10bio_t *rb2;
1851                         sector_t sect;
1852                         int must_sync;
1853
1854                         if (conf->mirrors[i].rdev == NULL ||
1855                             test_bit(In_sync, &conf->mirrors[i].rdev->flags))
1856                                 continue;
1857
1858                         still_degraded = 0;
1859                         /* want to reconstruct this device */
1860                         rb2 = r10_bio;
1861                         sect = raid10_find_virt(conf, sector_nr, i);
1862                         /* Unless we are doing a full sync, we only need
1863                          * to recover the block if it is set in the bitmap
1864                          */
1865                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
1866                                                       &sync_blocks, 1);
1867                         if (sync_blocks < max_sync)
1868                                 max_sync = sync_blocks;
1869                         if (!must_sync &&
1870                             !conf->fullsync) {
1871                                 /* yep, skip the sync_blocks here, but don't assume
1872                                  * that there will never be anything to do here
1873                                  */
1874                                 chunks_skipped = -1;
1875                                 continue;
1876                         }
1877
1878                         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1879                         raise_barrier(conf, rb2 != NULL);
1880                         atomic_set(&r10_bio->remaining, 0);
1881
1882                         r10_bio->master_bio = (struct bio*)rb2;
1883                         if (rb2)
1884                                 atomic_inc(&rb2->remaining);
1885                         r10_bio->mddev = mddev;
1886                         set_bit(R10BIO_IsRecover, &r10_bio->state);
1887                         r10_bio->sector = sect;
1888
1889                         raid10_find_phys(conf, r10_bio);
1890
1891                         /* Need to check if the array will still be
1892                          * degraded
1893                          */
1894                         for (j=0; j<conf->raid_disks; j++)
1895                                 if (conf->mirrors[j].rdev == NULL ||
1896                                     test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
1897                                         still_degraded = 1;
1898                                         break;
1899                                 }
1900
1901                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
1902                                                       &sync_blocks, still_degraded);
1903
1904                         for (j=0; j<conf->copies;j++) {
1905                                 int d = r10_bio->devs[j].devnum;
1906                                 if (!conf->mirrors[d].rdev ||
1907                                     !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
1908                                         continue;
1909                                 /* This is where we read from */
1910                                 bio = r10_bio->devs[0].bio;
1911                                 bio->bi_next = biolist;
1912                                 biolist = bio;
1913                                 bio->bi_private = r10_bio;
1914                                 bio->bi_end_io = end_sync_read;
1915                                 bio->bi_rw = READ;
1916                                 bio->bi_sector = r10_bio->devs[j].addr +
1917                                         conf->mirrors[d].rdev->data_offset;
1918                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1919                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1920                                 atomic_inc(&r10_bio->remaining);
1921                                 /* and we write to 'i' */
1922
1923                                 for (k=0; k<conf->copies; k++)
1924                                         if (r10_bio->devs[k].devnum == i)
1925                                                 break;
1926                                 BUG_ON(k == conf->copies);
1927                                 bio = r10_bio->devs[1].bio;
1928                                 bio->bi_next = biolist;
1929                                 biolist = bio;
1930                                 bio->bi_private = r10_bio;
1931                                 bio->bi_end_io = end_sync_write;
1932                                 bio->bi_rw = WRITE;
1933                                 bio->bi_sector = r10_bio->devs[k].addr +
1934                                         conf->mirrors[i].rdev->data_offset;
1935                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1936
1937                                 r10_bio->devs[0].devnum = d;
1938                                 r10_bio->devs[1].devnum = i;
1939
1940                                 break;
1941                         }
1942                         if (j == conf->copies) {
1943                                 /* Cannot recover, so abort the recovery */
1944                                 put_buf(r10_bio);
1945                                 if (rb2)
1946                                         atomic_dec(&rb2->remaining);
1947                                 r10_bio = rb2;
1948                                 if (!test_and_set_bit(MD_RECOVERY_INTR,
1949                                                       &mddev->recovery))
1950                                         printk(KERN_INFO "md/raid10:%s: insufficient "
1951                                                "working devices for recovery.\n",
1952                                                mdname(mddev));
1953                                 break;
1954                         }
1955                 }
1956                 if (biolist == NULL) {
1957                         while (r10_bio) {
1958                                 r10bio_t *rb2 = r10_bio;
1959                                 r10_bio = (r10bio_t*) rb2->master_bio;
1960                                 rb2->master_bio = NULL;
1961                                 put_buf(rb2);
1962                         }
1963                         goto giveup;
1964                 }
1965         } else {
1966                 /* resync. Schedule a read for every block at this virt offset */
1967                 int count = 0;
1968
1969                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1970
1971                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1972                                        &sync_blocks, mddev->degraded) &&
1973                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
1974                                                  &mddev->recovery)) {
1975                         /* We can skip this block */
1976                         *skipped = 1;
1977                         return sync_blocks + sectors_skipped;
1978                 }
1979                 if (sync_blocks < max_sync)
1980                         max_sync = sync_blocks;
1981                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1982
1983                 r10_bio->mddev = mddev;
1984                 atomic_set(&r10_bio->remaining, 0);
1985                 raise_barrier(conf, 0);
1986                 conf->next_resync = sector_nr;
1987
1988                 r10_bio->master_bio = NULL;
1989                 r10_bio->sector = sector_nr;
1990                 set_bit(R10BIO_IsSync, &r10_bio->state);
1991                 raid10_find_phys(conf, r10_bio);
1992                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1993
1994                 for (i=0; i<conf->copies; i++) {
1995                         int d = r10_bio->devs[i].devnum;
1996                         bio = r10_bio->devs[i].bio;
1997                         bio->bi_end_io = NULL;
1998                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
1999                         if (conf->mirrors[d].rdev == NULL ||
2000                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2001                                 continue;
2002                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2003                         atomic_inc(&r10_bio->remaining);
2004                         bio->bi_next = biolist;
2005                         biolist = bio;
2006                         bio->bi_private = r10_bio;
2007                         bio->bi_end_io = end_sync_read;
2008                         bio->bi_rw = READ;
2009                         bio->bi_sector = r10_bio->devs[i].addr +
2010                                 conf->mirrors[d].rdev->data_offset;
2011                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2012                         count++;
2013                 }
2014
2015                 if (count < 2) {
2016                         for (i=0; i<conf->copies; i++) {
2017                                 int d = r10_bio->devs[i].devnum;
2018                                 if (r10_bio->devs[i].bio->bi_end_io)
2019                                         rdev_dec_pending(conf->mirrors[d].rdev,
2020                                                          mddev);
2021                         }
2022                         put_buf(r10_bio);
2023                         biolist = NULL;
2024                         goto giveup;
2025                 }
2026         }
2027
2028         for (bio = biolist; bio ; bio=bio->bi_next) {
2029
2030                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2031                 if (bio->bi_end_io)
2032                         bio->bi_flags |= 1 << BIO_UPTODATE;
2033                 bio->bi_vcnt = 0;
2034                 bio->bi_idx = 0;
2035                 bio->bi_phys_segments = 0;
2036                 bio->bi_size = 0;
2037         }
2038
2039         nr_sectors = 0;
2040         if (sector_nr + max_sync < max_sector)
2041                 max_sector = sector_nr + max_sync;
2042         do {
2043                 struct page *page;
2044                 int len = PAGE_SIZE;
2045                 if (sector_nr + (len>>9) > max_sector)
2046                         len = (max_sector - sector_nr) << 9;
2047                 if (len == 0)
2048                         break;
2049                 for (bio= biolist ; bio ; bio=bio->bi_next) {
2050                         struct bio *bio2;
2051                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2052                         if (bio_add_page(bio, page, len, 0))
2053                                 continue;
2054
2055                         /* stop here */
2056                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2057                         for (bio2 = biolist;
2058                              bio2 && bio2 != bio;
2059                              bio2 = bio2->bi_next) {
2060                                 /* remove last page from this bio */
2061                                 bio2->bi_vcnt--;
2062                                 bio2->bi_size -= len;
2063                                 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2064                         }
2065                         goto bio_full;
2066                 }
2067                 nr_sectors += len>>9;
2068                 sector_nr += len>>9;
2069         } while (biolist->bi_vcnt < RESYNC_PAGES);
2070  bio_full:
2071         r10_bio->sectors = nr_sectors;
2072
2073         while (biolist) {
2074                 bio = biolist;
2075                 biolist = biolist->bi_next;
2076
2077                 bio->bi_next = NULL;
2078                 r10_bio = bio->bi_private;
2079                 r10_bio->sectors = nr_sectors;
2080
2081                 if (bio->bi_end_io == end_sync_read) {
2082                         md_sync_acct(bio->bi_bdev, nr_sectors);
2083                         generic_make_request(bio);
2084                 }
2085         }
2086
2087         if (sectors_skipped)
2088                 /* pretend they weren't skipped, it makes
2089                  * no important difference in this case
2090                  */
2091                 md_done_sync(mddev, sectors_skipped, 1);
2092
2093         return sectors_skipped + nr_sectors;
2094  giveup:
2095         /* There is nowhere to write, so all non-sync
2096          * drives must be failed, so try the next chunk...
2097          */
2098         if (sector_nr + max_sync < max_sector)
2099                 max_sector = sector_nr + max_sync;
2100
2101         sectors_skipped += (max_sector - sector_nr);
2102         chunks_skipped ++;
2103         sector_nr = max_sector;
2104         goto skipped;
2105 }
2106
2107 static sector_t
2108 raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2109 {
2110         sector_t size;
2111         conf_t *conf = mddev->private;
2112
2113         if (!raid_disks)
2114                 raid_disks = conf->raid_disks;
2115         if (!sectors)
2116                 sectors = conf->dev_sectors;
2117
2118         size = sectors >> conf->chunk_shift;
2119         sector_div(size, conf->far_copies);
2120         size = size * raid_disks;
2121         sector_div(size, conf->near_copies);
2122
2123         return size << conf->chunk_shift;
2124 }
2125
2126
2127 static conf_t *setup_conf(mddev_t *mddev)
2128 {
2129         conf_t *conf = NULL;
2130         int nc, fc, fo;
2131         sector_t stride, size;
2132         int err = -EINVAL;
2133
2134         if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2135             !is_power_of_2(mddev->new_chunk_sectors)) {
2136                 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2137                        "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2138                        mdname(mddev), PAGE_SIZE);
2139                 goto out;
2140         }
2141
2142         nc = mddev->new_layout & 255;
2143         fc = (mddev->new_layout >> 8) & 255;
2144         fo = mddev->new_layout & (1<<16);
2145
2146         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2147             (mddev->new_layout >> 17)) {
2148                 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2149                        mdname(mddev), mddev->new_layout);
2150                 goto out;
2151         }
2152
2153         err = -ENOMEM;
2154         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2155         if (!conf)
2156                 goto out;
2157
2158         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2159                                 GFP_KERNEL);
2160         if (!conf->mirrors)
2161                 goto out;
2162
2163         conf->tmppage = alloc_page(GFP_KERNEL);
2164         if (!conf->tmppage)
2165                 goto out;
2166
2167
2168         conf->raid_disks = mddev->raid_disks;
2169         conf->near_copies = nc;
2170         conf->far_copies = fc;
2171         conf->copies = nc*fc;
2172         conf->far_offset = fo;
2173         conf->chunk_mask = mddev->new_chunk_sectors - 1;
2174         conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2175
2176         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2177                                            r10bio_pool_free, conf);
2178         if (!conf->r10bio_pool)
2179                 goto out;
2180
2181         size = mddev->dev_sectors >> conf->chunk_shift;
2182         sector_div(size, fc);
2183         size = size * conf->raid_disks;
2184         sector_div(size, nc);
2185         /* 'size' is now the number of chunks in the array */
2186         /* calculate "used chunks per device" in 'stride' */
2187         stride = size * conf->copies;
2188
2189         /* We need to round up when dividing by raid_disks to
2190          * get the stride size.
2191          */
2192         stride += conf->raid_disks - 1;
2193         sector_div(stride, conf->raid_disks);
2194
2195         conf->dev_sectors = stride << conf->chunk_shift;
2196
2197         if (fo)
2198                 stride = 1;
2199         else
2200                 sector_div(stride, fc);
2201         conf->stride = stride << conf->chunk_shift;
2202
2203
2204         spin_lock_init(&conf->device_lock);
2205         INIT_LIST_HEAD(&conf->retry_list);
2206
2207         spin_lock_init(&conf->resync_lock);
2208         init_waitqueue_head(&conf->wait_barrier);
2209
2210         conf->thread = md_register_thread(raid10d, mddev, NULL);
2211         if (!conf->thread)
2212                 goto out;
2213
2214         conf->mddev = mddev;
2215         return conf;
2216
2217  out:
2218         printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2219                mdname(mddev));
2220         if (conf) {
2221                 if (conf->r10bio_pool)
2222                         mempool_destroy(conf->r10bio_pool);
2223                 kfree(conf->mirrors);
2224                 safe_put_page(conf->tmppage);
2225                 kfree(conf);
2226         }
2227         return ERR_PTR(err);
2228 }
2229
2230 static int run(mddev_t *mddev)
2231 {
2232         conf_t *conf;
2233         int i, disk_idx, chunk_size;
2234         mirror_info_t *disk;
2235         mdk_rdev_t *rdev;
2236         sector_t size;
2237
2238         /*
2239          * copy the already verified devices into our private RAID10
2240          * bookkeeping area. [whatever we allocate in run(),
2241          * should be freed in stop()]
2242          */
2243
2244         if (mddev->private == NULL) {
2245                 conf = setup_conf(mddev);
2246                 if (IS_ERR(conf))
2247                         return PTR_ERR(conf);
2248                 mddev->private = conf;
2249         }
2250         conf = mddev->private;
2251         if (!conf)
2252                 goto out;
2253
2254         mddev->thread = conf->thread;
2255         conf->thread = NULL;
2256
2257         chunk_size = mddev->chunk_sectors << 9;
2258         blk_queue_io_min(mddev->queue, chunk_size);
2259         if (conf->raid_disks % conf->near_copies)
2260                 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2261         else
2262                 blk_queue_io_opt(mddev->queue, chunk_size *
2263                                  (conf->raid_disks / conf->near_copies));
2264
2265         list_for_each_entry(rdev, &mddev->disks, same_set) {
2266                 disk_idx = rdev->raid_disk;
2267                 if (disk_idx >= conf->raid_disks
2268                     || disk_idx < 0)
2269                         continue;
2270                 disk = conf->mirrors + disk_idx;
2271
2272                 disk->rdev = rdev;
2273                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2274                                   rdev->data_offset << 9);
2275                 /* as we don't honour merge_bvec_fn, we must never risk
2276                  * violating it, so limit max_segments to 1 lying
2277                  * within a single page.
2278                  */
2279                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2280                         blk_queue_max_segments(mddev->queue, 1);
2281                         blk_queue_segment_boundary(mddev->queue,
2282                                                    PAGE_CACHE_SIZE - 1);
2283                 }
2284
2285                 disk->head_position = 0;
2286         }
2287         /* need to check that every block has at least one working mirror */
2288         if (!enough(conf, -1)) {
2289                 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2290                        mdname(mddev));
2291                 goto out_free_conf;
2292         }
2293
2294         mddev->degraded = 0;
2295         for (i = 0; i < conf->raid_disks; i++) {
2296
2297                 disk = conf->mirrors + i;
2298
2299                 if (!disk->rdev ||
2300                     !test_bit(In_sync, &disk->rdev->flags)) {
2301                         disk->head_position = 0;
2302                         mddev->degraded++;
2303                         if (disk->rdev)
2304                                 conf->fullsync = 1;
2305                 }
2306         }
2307
2308         if (mddev->recovery_cp != MaxSector)
2309                 printk(KERN_NOTICE "md/raid10:%s: not clean"
2310                        " -- starting background reconstruction\n",
2311                        mdname(mddev));
2312         printk(KERN_INFO
2313                 "md/raid10:%s: active with %d out of %d devices\n",
2314                 mdname(mddev), conf->raid_disks - mddev->degraded,
2315                 conf->raid_disks);
2316         /*
2317          * Ok, everything is just fine now
2318          */
2319         mddev->dev_sectors = conf->dev_sectors;
2320         size = raid10_size(mddev, 0, 0);
2321         md_set_array_sectors(mddev, size);
2322         mddev->resync_max_sectors = size;
2323
2324         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2325         mddev->queue->backing_dev_info.congested_data = mddev;
2326
2327         /* Calculate max read-ahead size.
2328          * We need to readahead at least twice a whole stripe....
2329          * maybe...
2330          */
2331         {
2332                 int stripe = conf->raid_disks *
2333                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2334                 stripe /= conf->near_copies;
2335                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2336                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2337         }
2338
2339         if (conf->near_copies < conf->raid_disks)
2340                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2341
2342         if (md_integrity_register(mddev))
2343                 goto out_free_conf;
2344
2345         return 0;
2346
2347 out_free_conf:
2348         md_unregister_thread(mddev->thread);
2349         if (conf->r10bio_pool)
2350                 mempool_destroy(conf->r10bio_pool);
2351         safe_put_page(conf->tmppage);
2352         kfree(conf->mirrors);
2353         kfree(conf);
2354         mddev->private = NULL;
2355 out:
2356         return -EIO;
2357 }
2358
2359 static int stop(mddev_t *mddev)
2360 {
2361         conf_t *conf = mddev->private;
2362
2363         raise_barrier(conf, 0);
2364         lower_barrier(conf);
2365
2366         md_unregister_thread(mddev->thread);
2367         mddev->thread = NULL;
2368         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2369         if (conf->r10bio_pool)
2370                 mempool_destroy(conf->r10bio_pool);
2371         kfree(conf->mirrors);
2372         kfree(conf);
2373         mddev->private = NULL;
2374         return 0;
2375 }
2376
2377 static void raid10_quiesce(mddev_t *mddev, int state)
2378 {
2379         conf_t *conf = mddev->private;
2380
2381         switch(state) {
2382         case 1:
2383                 raise_barrier(conf, 0);
2384                 break;
2385         case 0:
2386                 lower_barrier(conf);
2387                 break;
2388         }
2389 }
2390
2391 static void *raid10_takeover_raid0(mddev_t *mddev)
2392 {
2393         mdk_rdev_t *rdev;
2394         conf_t *conf;
2395
2396         if (mddev->degraded > 0) {
2397                 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
2398                        mdname(mddev));
2399                 return ERR_PTR(-EINVAL);
2400         }
2401
2402         /* Set new parameters */
2403         mddev->new_level = 10;
2404         /* new layout: far_copies = 1, near_copies = 2 */
2405         mddev->new_layout = (1<<8) + 2;
2406         mddev->new_chunk_sectors = mddev->chunk_sectors;
2407         mddev->delta_disks = mddev->raid_disks;
2408         mddev->raid_disks *= 2;
2409         /* make sure it will be not marked as dirty */
2410         mddev->recovery_cp = MaxSector;
2411
2412         conf = setup_conf(mddev);
2413         if (!IS_ERR(conf)) {
2414                 list_for_each_entry(rdev, &mddev->disks, same_set)
2415                         if (rdev->raid_disk >= 0)
2416                                 rdev->new_raid_disk = rdev->raid_disk * 2;
2417                 conf->barrier = 1;
2418         }
2419
2420         return conf;
2421 }
2422
2423 static void *raid10_takeover(mddev_t *mddev)
2424 {
2425         struct raid0_private_data *raid0_priv;
2426
2427         /* raid10 can take over:
2428          *  raid0 - providing it has only two drives
2429          */
2430         if (mddev->level == 0) {
2431                 /* for raid0 takeover only one zone is supported */
2432                 raid0_priv = mddev->private;
2433                 if (raid0_priv->nr_strip_zones > 1) {
2434                         printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
2435                                " with more than one zone.\n",
2436                                mdname(mddev));
2437                         return ERR_PTR(-EINVAL);
2438                 }
2439                 return raid10_takeover_raid0(mddev);
2440         }
2441         return ERR_PTR(-EINVAL);
2442 }
2443
2444 static struct mdk_personality raid10_personality =
2445 {
2446         .name           = "raid10",
2447         .level          = 10,
2448         .owner          = THIS_MODULE,
2449         .make_request   = make_request,
2450         .run            = run,
2451         .stop           = stop,
2452         .status         = status,
2453         .error_handler  = error,
2454         .hot_add_disk   = raid10_add_disk,
2455         .hot_remove_disk= raid10_remove_disk,
2456         .spare_active   = raid10_spare_active,
2457         .sync_request   = sync_request,
2458         .quiesce        = raid10_quiesce,
2459         .size           = raid10_size,
2460         .takeover       = raid10_takeover,
2461 };
2462
2463 static int __init raid_init(void)
2464 {
2465         return register_md_personality(&raid10_personality);
2466 }
2467
2468 static void raid_exit(void)
2469 {
2470         unregister_md_personality(&raid10_personality);
2471 }
2472
2473 module_init(raid_init);
2474 module_exit(raid_exit);
2475 MODULE_LICENSE("GPL");
2476 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
2477 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2478 MODULE_ALIAS("md-raid10");
2479 MODULE_ALIAS("md-level-10");