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(Blocked, &rdev->flags);
1025         set_bit(Faulty, &rdev->flags);
1026         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1027         printk(KERN_ALERT
1028                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1029                "md/raid10:%s: Operation continuing on %d devices.\n",
1030                mdname(mddev), bdevname(rdev->bdev, b),
1031                mdname(mddev), conf->raid_disks - mddev->degraded);
1032 }
1033
1034 static void print_conf(conf_t *conf)
1035 {
1036         int i;
1037         mirror_info_t *tmp;
1038
1039         printk(KERN_DEBUG "RAID10 conf printout:\n");
1040         if (!conf) {
1041                 printk(KERN_DEBUG "(!conf)\n");
1042                 return;
1043         }
1044         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1045                 conf->raid_disks);
1046
1047         for (i = 0; i < conf->raid_disks; i++) {
1048                 char b[BDEVNAME_SIZE];
1049                 tmp = conf->mirrors + i;
1050                 if (tmp->rdev)
1051                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1052                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1053                                 !test_bit(Faulty, &tmp->rdev->flags),
1054                                 bdevname(tmp->rdev->bdev,b));
1055         }
1056 }
1057
1058 static void close_sync(conf_t *conf)
1059 {
1060         wait_barrier(conf);
1061         allow_barrier(conf);
1062
1063         mempool_destroy(conf->r10buf_pool);
1064         conf->r10buf_pool = NULL;
1065 }
1066
1067 static int raid10_spare_active(mddev_t *mddev)
1068 {
1069         int i;
1070         conf_t *conf = mddev->private;
1071         mirror_info_t *tmp;
1072         int count = 0;
1073         unsigned long flags;
1074
1075         /*
1076          * Find all non-in_sync disks within the RAID10 configuration
1077          * and mark them in_sync
1078          */
1079         for (i = 0; i < conf->raid_disks; i++) {
1080                 tmp = conf->mirrors + i;
1081                 if (tmp->rdev
1082                     && !test_bit(Faulty, &tmp->rdev->flags)
1083                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1084                         count++;
1085                         sysfs_notify_dirent(tmp->rdev->sysfs_state);
1086                 }
1087         }
1088         spin_lock_irqsave(&conf->device_lock, flags);
1089         mddev->degraded -= count;
1090         spin_unlock_irqrestore(&conf->device_lock, flags);
1091
1092         print_conf(conf);
1093         return count;
1094 }
1095
1096
1097 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1098 {
1099         conf_t *conf = mddev->private;
1100         int err = -EEXIST;
1101         int mirror;
1102         int first = 0;
1103         int last = conf->raid_disks - 1;
1104
1105         if (rdev->badblocks.count)
1106                 return -EINVAL;
1107
1108         if (mddev->recovery_cp < MaxSector)
1109                 /* only hot-add to in-sync arrays, as recovery is
1110                  * very different from resync
1111                  */
1112                 return -EBUSY;
1113         if (!enough(conf, -1))
1114                 return -EINVAL;
1115
1116         if (rdev->raid_disk >= 0)
1117                 first = last = rdev->raid_disk;
1118
1119         if (rdev->saved_raid_disk >= first &&
1120             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1121                 mirror = rdev->saved_raid_disk;
1122         else
1123                 mirror = first;
1124         for ( ; mirror <= last ; mirror++) {
1125                 mirror_info_t *p = &conf->mirrors[mirror];
1126                 if (p->recovery_disabled == mddev->recovery_disabled)
1127                         continue;
1128                 if (!p->rdev)
1129                         continue;
1130
1131                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1132                                   rdev->data_offset << 9);
1133                 /* as we don't honour merge_bvec_fn, we must
1134                  * never risk violating it, so limit
1135                  * ->max_segments to one lying with a single
1136                  * page, as a one page request is never in
1137                  * violation.
1138                  */
1139                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1140                         blk_queue_max_segments(mddev->queue, 1);
1141                         blk_queue_segment_boundary(mddev->queue,
1142                                                    PAGE_CACHE_SIZE - 1);
1143                 }
1144
1145                 p->head_position = 0;
1146                 rdev->raid_disk = mirror;
1147                 err = 0;
1148                 if (rdev->saved_raid_disk != mirror)
1149                         conf->fullsync = 1;
1150                 rcu_assign_pointer(p->rdev, rdev);
1151                 break;
1152         }
1153
1154         md_integrity_add_rdev(rdev, mddev);
1155         print_conf(conf);
1156         return err;
1157 }
1158
1159 static int raid10_remove_disk(mddev_t *mddev, int number)
1160 {
1161         conf_t *conf = mddev->private;
1162         int err = 0;
1163         mdk_rdev_t *rdev;
1164         mirror_info_t *p = conf->mirrors+ number;
1165
1166         print_conf(conf);
1167         rdev = p->rdev;
1168         if (rdev) {
1169                 if (test_bit(In_sync, &rdev->flags) ||
1170                     atomic_read(&rdev->nr_pending)) {
1171                         err = -EBUSY;
1172                         goto abort;
1173                 }
1174                 /* Only remove faulty devices in recovery
1175                  * is not possible.
1176                  */
1177                 if (!test_bit(Faulty, &rdev->flags) &&
1178                     mddev->recovery_disabled != p->recovery_disabled &&
1179                     enough(conf, -1)) {
1180                         err = -EBUSY;
1181                         goto abort;
1182                 }
1183                 p->rdev = NULL;
1184                 synchronize_rcu();
1185                 if (atomic_read(&rdev->nr_pending)) {
1186                         /* lost the race, try later */
1187                         err = -EBUSY;
1188                         p->rdev = rdev;
1189                         goto abort;
1190                 }
1191                 err = md_integrity_register(mddev);
1192         }
1193 abort:
1194
1195         print_conf(conf);
1196         return err;
1197 }
1198
1199
1200 static void end_sync_read(struct bio *bio, int error)
1201 {
1202         r10bio_t *r10_bio = bio->bi_private;
1203         conf_t *conf = r10_bio->mddev->private;
1204         int d;
1205
1206         d = find_bio_disk(conf, r10_bio, bio);
1207
1208         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1209                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1210         else {
1211                 atomic_add(r10_bio->sectors,
1212                            &conf->mirrors[d].rdev->corrected_errors);
1213                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1214                         md_error(r10_bio->mddev,
1215                                  conf->mirrors[d].rdev);
1216         }
1217
1218         /* for reconstruct, we always reschedule after a read.
1219          * for resync, only after all reads
1220          */
1221         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1222         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1223             atomic_dec_and_test(&r10_bio->remaining)) {
1224                 /* we have read all the blocks,
1225                  * do the comparison in process context in raid10d
1226                  */
1227                 reschedule_retry(r10_bio);
1228         }
1229 }
1230
1231 static void end_sync_write(struct bio *bio, int error)
1232 {
1233         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1234         r10bio_t *r10_bio = bio->bi_private;
1235         mddev_t *mddev = r10_bio->mddev;
1236         conf_t *conf = mddev->private;
1237         int d;
1238
1239         d = find_bio_disk(conf, r10_bio, bio);
1240
1241         if (!uptodate)
1242                 md_error(mddev, conf->mirrors[d].rdev);
1243
1244         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1245         while (atomic_dec_and_test(&r10_bio->remaining)) {
1246                 if (r10_bio->master_bio == NULL) {
1247                         /* the primary of several recovery bios */
1248                         sector_t s = r10_bio->sectors;
1249                         put_buf(r10_bio);
1250                         md_done_sync(mddev, s, 1);
1251                         break;
1252                 } else {
1253                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1254                         put_buf(r10_bio);
1255                         r10_bio = r10_bio2;
1256                 }
1257         }
1258 }
1259
1260 /*
1261  * Note: sync and recover and handled very differently for raid10
1262  * This code is for resync.
1263  * For resync, we read through virtual addresses and read all blocks.
1264  * If there is any error, we schedule a write.  The lowest numbered
1265  * drive is authoritative.
1266  * However requests come for physical address, so we need to map.
1267  * For every physical address there are raid_disks/copies virtual addresses,
1268  * which is always are least one, but is not necessarly an integer.
1269  * This means that a physical address can span multiple chunks, so we may
1270  * have to submit multiple io requests for a single sync request.
1271  */
1272 /*
1273  * We check if all blocks are in-sync and only write to blocks that
1274  * aren't in sync
1275  */
1276 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1277 {
1278         conf_t *conf = mddev->private;
1279         int i, first;
1280         struct bio *tbio, *fbio;
1281
1282         atomic_set(&r10_bio->remaining, 1);
1283
1284         /* find the first device with a block */
1285         for (i=0; i<conf->copies; i++)
1286                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1287                         break;
1288
1289         if (i == conf->copies)
1290                 goto done;
1291
1292         first = i;
1293         fbio = r10_bio->devs[i].bio;
1294
1295         /* now find blocks with errors */
1296         for (i=0 ; i < conf->copies ; i++) {
1297                 int  j, d;
1298                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1299
1300                 tbio = r10_bio->devs[i].bio;
1301
1302                 if (tbio->bi_end_io != end_sync_read)
1303                         continue;
1304                 if (i == first)
1305                         continue;
1306                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1307                         /* We know that the bi_io_vec layout is the same for
1308                          * both 'first' and 'i', so we just compare them.
1309                          * All vec entries are PAGE_SIZE;
1310                          */
1311                         for (j = 0; j < vcnt; j++)
1312                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1313                                            page_address(tbio->bi_io_vec[j].bv_page),
1314                                            PAGE_SIZE))
1315                                         break;
1316                         if (j == vcnt)
1317                                 continue;
1318                         mddev->resync_mismatches += r10_bio->sectors;
1319                 }
1320                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1321                         /* Don't fix anything. */
1322                         continue;
1323                 /* Ok, we need to write this bio
1324                  * First we need to fixup bv_offset, bv_len and
1325                  * bi_vecs, as the read request might have corrupted these
1326                  */
1327                 tbio->bi_vcnt = vcnt;
1328                 tbio->bi_size = r10_bio->sectors << 9;
1329                 tbio->bi_idx = 0;
1330                 tbio->bi_phys_segments = 0;
1331                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1332                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1333                 tbio->bi_next = NULL;
1334                 tbio->bi_rw = WRITE;
1335                 tbio->bi_private = r10_bio;
1336                 tbio->bi_sector = r10_bio->devs[i].addr;
1337
1338                 for (j=0; j < vcnt ; j++) {
1339                         tbio->bi_io_vec[j].bv_offset = 0;
1340                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1341
1342                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1343                                page_address(fbio->bi_io_vec[j].bv_page),
1344                                PAGE_SIZE);
1345                 }
1346                 tbio->bi_end_io = end_sync_write;
1347
1348                 d = r10_bio->devs[i].devnum;
1349                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1350                 atomic_inc(&r10_bio->remaining);
1351                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1352
1353                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1354                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1355                 generic_make_request(tbio);
1356         }
1357
1358 done:
1359         if (atomic_dec_and_test(&r10_bio->remaining)) {
1360                 md_done_sync(mddev, r10_bio->sectors, 1);
1361                 put_buf(r10_bio);
1362         }
1363 }
1364
1365 /*
1366  * Now for the recovery code.
1367  * Recovery happens across physical sectors.
1368  * We recover all non-is_sync drives by finding the virtual address of
1369  * each, and then choose a working drive that also has that virt address.
1370  * There is a separate r10_bio for each non-in_sync drive.
1371  * Only the first two slots are in use. The first for reading,
1372  * The second for writing.
1373  *
1374  */
1375
1376 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1377 {
1378         conf_t *conf = mddev->private;
1379         int d;
1380         struct bio *wbio;
1381
1382         /*
1383          * share the pages with the first bio
1384          * and submit the write request
1385          */
1386         wbio = r10_bio->devs[1].bio;
1387         d = r10_bio->devs[1].devnum;
1388
1389         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1390         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1391         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1392                 generic_make_request(wbio);
1393         else {
1394                 printk(KERN_NOTICE
1395                        "md/raid10:%s: recovery aborted due to read error\n",
1396                        mdname(mddev));
1397                 conf->mirrors[d].recovery_disabled = mddev->recovery_disabled;
1398                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1399                 bio_endio(wbio, 0);
1400         }
1401 }
1402
1403
1404 /*
1405  * Used by fix_read_error() to decay the per rdev read_errors.
1406  * We halve the read error count for every hour that has elapsed
1407  * since the last recorded read error.
1408  *
1409  */
1410 static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1411 {
1412         struct timespec cur_time_mon;
1413         unsigned long hours_since_last;
1414         unsigned int read_errors = atomic_read(&rdev->read_errors);
1415
1416         ktime_get_ts(&cur_time_mon);
1417
1418         if (rdev->last_read_error.tv_sec == 0 &&
1419             rdev->last_read_error.tv_nsec == 0) {
1420                 /* first time we've seen a read error */
1421                 rdev->last_read_error = cur_time_mon;
1422                 return;
1423         }
1424
1425         hours_since_last = (cur_time_mon.tv_sec -
1426                             rdev->last_read_error.tv_sec) / 3600;
1427
1428         rdev->last_read_error = cur_time_mon;
1429
1430         /*
1431          * if hours_since_last is > the number of bits in read_errors
1432          * just set read errors to 0. We do this to avoid
1433          * overflowing the shift of read_errors by hours_since_last.
1434          */
1435         if (hours_since_last >= 8 * sizeof(read_errors))
1436                 atomic_set(&rdev->read_errors, 0);
1437         else
1438                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1439 }
1440
1441 /*
1442  * This is a kernel thread which:
1443  *
1444  *      1.      Retries failed read operations on working mirrors.
1445  *      2.      Updates the raid superblock when problems encounter.
1446  *      3.      Performs writes following reads for array synchronising.
1447  */
1448
1449 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1450 {
1451         int sect = 0; /* Offset from r10_bio->sector */
1452         int sectors = r10_bio->sectors;
1453         mdk_rdev_t*rdev;
1454         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1455         int d = r10_bio->devs[r10_bio->read_slot].devnum;
1456
1457         /* still own a reference to this rdev, so it cannot
1458          * have been cleared recently.
1459          */
1460         rdev = conf->mirrors[d].rdev;
1461
1462         if (test_bit(Faulty, &rdev->flags))
1463                 /* drive has already been failed, just ignore any
1464                    more fix_read_error() attempts */
1465                 return;
1466
1467         check_decay_read_errors(mddev, rdev);
1468         atomic_inc(&rdev->read_errors);
1469         if (atomic_read(&rdev->read_errors) > max_read_errors) {
1470                 char b[BDEVNAME_SIZE];
1471                 bdevname(rdev->bdev, b);
1472
1473                 printk(KERN_NOTICE
1474                        "md/raid10:%s: %s: Raid device exceeded "
1475                        "read_error threshold [cur %d:max %d]\n",
1476                        mdname(mddev), b,
1477                        atomic_read(&rdev->read_errors), max_read_errors);
1478                 printk(KERN_NOTICE
1479                        "md/raid10:%s: %s: Failing raid device\n",
1480                        mdname(mddev), b);
1481                 md_error(mddev, conf->mirrors[d].rdev);
1482                 return;
1483         }
1484
1485         while(sectors) {
1486                 int s = sectors;
1487                 int sl = r10_bio->read_slot;
1488                 int success = 0;
1489                 int start;
1490
1491                 if (s > (PAGE_SIZE>>9))
1492                         s = PAGE_SIZE >> 9;
1493
1494                 rcu_read_lock();
1495                 do {
1496                         d = r10_bio->devs[sl].devnum;
1497                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1498                         if (rdev &&
1499                             test_bit(In_sync, &rdev->flags)) {
1500                                 atomic_inc(&rdev->nr_pending);
1501                                 rcu_read_unlock();
1502                                 success = sync_page_io(rdev,
1503                                                        r10_bio->devs[sl].addr +
1504                                                        sect,
1505                                                        s<<9,
1506                                                        conf->tmppage, READ, false);
1507                                 rdev_dec_pending(rdev, mddev);
1508                                 rcu_read_lock();
1509                                 if (success)
1510                                         break;
1511                         }
1512                         sl++;
1513                         if (sl == conf->copies)
1514                                 sl = 0;
1515                 } while (!success && sl != r10_bio->read_slot);
1516                 rcu_read_unlock();
1517
1518                 if (!success) {
1519                         /* Cannot read from anywhere -- bye bye array */
1520                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1521                         md_error(mddev, conf->mirrors[dn].rdev);
1522                         break;
1523                 }
1524
1525                 start = sl;
1526                 /* write it back and re-read */
1527                 rcu_read_lock();
1528                 while (sl != r10_bio->read_slot) {
1529                         char b[BDEVNAME_SIZE];
1530
1531                         if (sl==0)
1532                                 sl = conf->copies;
1533                         sl--;
1534                         d = r10_bio->devs[sl].devnum;
1535                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1536                         if (rdev &&
1537                             test_bit(In_sync, &rdev->flags)) {
1538                                 atomic_inc(&rdev->nr_pending);
1539                                 rcu_read_unlock();
1540                                 if (sync_page_io(rdev,
1541                                                  r10_bio->devs[sl].addr +
1542                                                  sect,
1543                                                  s<<9, conf->tmppage, WRITE, false)
1544                                     == 0) {
1545                                         /* Well, this device is dead */
1546                                         printk(KERN_NOTICE
1547                                                "md/raid10:%s: read correction "
1548                                                "write failed"
1549                                                " (%d sectors at %llu on %s)\n",
1550                                                mdname(mddev), s,
1551                                                (unsigned long long)(
1552                                                        sect + rdev->data_offset),
1553                                                bdevname(rdev->bdev, b));
1554                                         printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1555                                                "drive\n",
1556                                                mdname(mddev),
1557                                                bdevname(rdev->bdev, b));
1558                                         md_error(mddev, rdev);
1559                                 }
1560                                 rdev_dec_pending(rdev, mddev);
1561                                 rcu_read_lock();
1562                         }
1563                 }
1564                 sl = start;
1565                 while (sl != r10_bio->read_slot) {
1566
1567                         if (sl==0)
1568                                 sl = conf->copies;
1569                         sl--;
1570                         d = r10_bio->devs[sl].devnum;
1571                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1572                         if (rdev &&
1573                             test_bit(In_sync, &rdev->flags)) {
1574                                 char b[BDEVNAME_SIZE];
1575                                 atomic_inc(&rdev->nr_pending);
1576                                 rcu_read_unlock();
1577                                 if (sync_page_io(rdev,
1578                                                  r10_bio->devs[sl].addr +
1579                                                  sect,
1580                                                  s<<9, conf->tmppage,
1581                                                  READ, false) == 0) {
1582                                         /* Well, this device is dead */
1583                                         printk(KERN_NOTICE
1584                                                "md/raid10:%s: unable to read back "
1585                                                "corrected sectors"
1586                                                " (%d sectors at %llu on %s)\n",
1587                                                mdname(mddev), s,
1588                                                (unsigned long long)(
1589                                                        sect + rdev->data_offset),
1590                                                bdevname(rdev->bdev, b));
1591                                         printk(KERN_NOTICE "md/raid10:%s: %s: failing drive\n",
1592                                                mdname(mddev),
1593                                                bdevname(rdev->bdev, b));
1594
1595                                         md_error(mddev, rdev);
1596                                 } else {
1597                                         printk(KERN_INFO
1598                                                "md/raid10:%s: read error corrected"
1599                                                " (%d sectors at %llu on %s)\n",
1600                                                mdname(mddev), s,
1601                                                (unsigned long long)(
1602                                                        sect + rdev->data_offset),
1603                                                bdevname(rdev->bdev, b));
1604                                         atomic_add(s, &rdev->corrected_errors);
1605                                 }
1606
1607                                 rdev_dec_pending(rdev, mddev);
1608                                 rcu_read_lock();
1609                         }
1610                 }
1611                 rcu_read_unlock();
1612
1613                 sectors -= s;
1614                 sect += s;
1615         }
1616 }
1617
1618 static void raid10d(mddev_t *mddev)
1619 {
1620         r10bio_t *r10_bio;
1621         struct bio *bio;
1622         unsigned long flags;
1623         conf_t *conf = mddev->private;
1624         struct list_head *head = &conf->retry_list;
1625         mdk_rdev_t *rdev;
1626         struct blk_plug plug;
1627
1628         md_check_recovery(mddev);
1629
1630         blk_start_plug(&plug);
1631         for (;;) {
1632                 char b[BDEVNAME_SIZE];
1633
1634                 flush_pending_writes(conf);
1635
1636                 spin_lock_irqsave(&conf->device_lock, flags);
1637                 if (list_empty(head)) {
1638                         spin_unlock_irqrestore(&conf->device_lock, flags);
1639                         break;
1640                 }
1641                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1642                 list_del(head->prev);
1643                 conf->nr_queued--;
1644                 spin_unlock_irqrestore(&conf->device_lock, flags);
1645
1646                 mddev = r10_bio->mddev;
1647                 conf = mddev->private;
1648                 if (test_bit(R10BIO_IsSync, &r10_bio->state))
1649                         sync_request_write(mddev, r10_bio);
1650                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
1651                         recovery_request_write(mddev, r10_bio);
1652                 else {
1653                         int slot = r10_bio->read_slot;
1654                         int mirror = r10_bio->devs[slot].devnum;
1655                         /* we got a read error. Maybe the drive is bad.  Maybe just
1656                          * the block and we can fix it.
1657                          * We freeze all other IO, and try reading the block from
1658                          * other devices.  When we find one, we re-write
1659                          * and check it that fixes the read error.
1660                          * This is all done synchronously while the array is
1661                          * frozen.
1662                          */
1663                         if (mddev->ro == 0) {
1664                                 freeze_array(conf);
1665                                 fix_read_error(conf, mddev, r10_bio);
1666                                 unfreeze_array(conf);
1667                         }
1668                         rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
1669
1670                         bio = r10_bio->devs[slot].bio;
1671                         r10_bio->devs[slot].bio =
1672                                 mddev->ro ? IO_BLOCKED : NULL;
1673                         mirror = read_balance(conf, r10_bio);
1674                         if (mirror == -1) {
1675                                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
1676                                        " read error for block %llu\n",
1677                                        mdname(mddev),
1678                                        bdevname(bio->bi_bdev,b),
1679                                        (unsigned long long)r10_bio->sector);
1680                                 raid_end_bio_io(r10_bio);
1681                                 bio_put(bio);
1682                         } else {
1683                                 const unsigned long do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
1684                                 bio_put(bio);
1685                                 slot = r10_bio->read_slot;
1686                                 rdev = conf->mirrors[mirror].rdev;
1687                                 printk_ratelimited(
1688                                         KERN_ERR
1689                                         "md/raid10:%s: %s: redirecting"
1690                                         "sector %llu to another mirror\n",
1691                                         mdname(mddev),
1692                                         bdevname(rdev->bdev, b),
1693                                         (unsigned long long)r10_bio->sector);
1694                                 bio = bio_clone_mddev(r10_bio->master_bio,
1695                                                       GFP_NOIO, mddev);
1696                                 r10_bio->devs[slot].bio = bio;
1697                                 bio->bi_sector = r10_bio->devs[slot].addr
1698                                         + rdev->data_offset;
1699                                 bio->bi_bdev = rdev->bdev;
1700                                 bio->bi_rw = READ | do_sync;
1701                                 bio->bi_private = r10_bio;
1702                                 bio->bi_end_io = raid10_end_read_request;
1703                                 generic_make_request(bio);
1704                         }
1705                 }
1706                 cond_resched();
1707                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
1708                         md_check_recovery(mddev);
1709         }
1710         blk_finish_plug(&plug);
1711 }
1712
1713
1714 static int init_resync(conf_t *conf)
1715 {
1716         int buffs;
1717
1718         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1719         BUG_ON(conf->r10buf_pool);
1720         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1721         if (!conf->r10buf_pool)
1722                 return -ENOMEM;
1723         conf->next_resync = 0;
1724         return 0;
1725 }
1726
1727 /*
1728  * perform a "sync" on one "block"
1729  *
1730  * We need to make sure that no normal I/O request - particularly write
1731  * requests - conflict with active sync requests.
1732  *
1733  * This is achieved by tracking pending requests and a 'barrier' concept
1734  * that can be installed to exclude normal IO requests.
1735  *
1736  * Resync and recovery are handled very differently.
1737  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1738  *
1739  * For resync, we iterate over virtual addresses, read all copies,
1740  * and update if there are differences.  If only one copy is live,
1741  * skip it.
1742  * For recovery, we iterate over physical addresses, read a good
1743  * value for each non-in_sync drive, and over-write.
1744  *
1745  * So, for recovery we may have several outstanding complex requests for a
1746  * given address, one for each out-of-sync device.  We model this by allocating
1747  * a number of r10_bio structures, one for each out-of-sync device.
1748  * As we setup these structures, we collect all bio's together into a list
1749  * which we then process collectively to add pages, and then process again
1750  * to pass to generic_make_request.
1751  *
1752  * The r10_bio structures are linked using a borrowed master_bio pointer.
1753  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1754  * has its remaining count decremented to 0, the whole complex operation
1755  * is complete.
1756  *
1757  */
1758
1759 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr,
1760                              int *skipped, int go_faster)
1761 {
1762         conf_t *conf = mddev->private;
1763         r10bio_t *r10_bio;
1764         struct bio *biolist = NULL, *bio;
1765         sector_t max_sector, nr_sectors;
1766         int i;
1767         int max_sync;
1768         sector_t sync_blocks;
1769
1770         sector_t sectors_skipped = 0;
1771         int chunks_skipped = 0;
1772
1773         if (!conf->r10buf_pool)
1774                 if (init_resync(conf))
1775                         return 0;
1776
1777  skipped:
1778         max_sector = mddev->dev_sectors;
1779         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1780                 max_sector = mddev->resync_max_sectors;
1781         if (sector_nr >= max_sector) {
1782                 /* If we aborted, we need to abort the
1783                  * sync on the 'current' bitmap chucks (there can
1784                  * be several when recovering multiple devices).
1785                  * as we may have started syncing it but not finished.
1786                  * We can find the current address in
1787                  * mddev->curr_resync, but for recovery,
1788                  * we need to convert that to several
1789                  * virtual addresses.
1790                  */
1791                 if (mddev->curr_resync < max_sector) { /* aborted */
1792                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1793                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1794                                                 &sync_blocks, 1);
1795                         else for (i=0; i<conf->raid_disks; i++) {
1796                                 sector_t sect =
1797                                         raid10_find_virt(conf, mddev->curr_resync, i);
1798                                 bitmap_end_sync(mddev->bitmap, sect,
1799                                                 &sync_blocks, 1);
1800                         }
1801                 } else /* completed sync */
1802                         conf->fullsync = 0;
1803
1804                 bitmap_close_sync(mddev->bitmap);
1805                 close_sync(conf);
1806                 *skipped = 1;
1807                 return sectors_skipped;
1808         }
1809         if (chunks_skipped >= conf->raid_disks) {
1810                 /* if there has been nothing to do on any drive,
1811                  * then there is nothing to do at all..
1812                  */
1813                 *skipped = 1;
1814                 return (max_sector - sector_nr) + sectors_skipped;
1815         }
1816
1817         if (max_sector > mddev->resync_max)
1818                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1819
1820         /* make sure whole request will fit in a chunk - if chunks
1821          * are meaningful
1822          */
1823         if (conf->near_copies < conf->raid_disks &&
1824             max_sector > (sector_nr | conf->chunk_mask))
1825                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1826         /*
1827          * If there is non-resync activity waiting for us then
1828          * put in a delay to throttle resync.
1829          */
1830         if (!go_faster && conf->nr_waiting)
1831                 msleep_interruptible(1000);
1832
1833         /* Again, very different code for resync and recovery.
1834          * Both must result in an r10bio with a list of bios that
1835          * have bi_end_io, bi_sector, bi_bdev set,
1836          * and bi_private set to the r10bio.
1837          * For recovery, we may actually create several r10bios
1838          * with 2 bios in each, that correspond to the bios in the main one.
1839          * In this case, the subordinate r10bios link back through a
1840          * borrowed master_bio pointer, and the counter in the master
1841          * includes a ref from each subordinate.
1842          */
1843         /* First, we decide what to do and set ->bi_end_io
1844          * To end_sync_read if we want to read, and
1845          * end_sync_write if we will want to write.
1846          */
1847
1848         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1849         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1850                 /* recovery... the complicated one */
1851                 int j, k;
1852                 r10_bio = NULL;
1853
1854                 for (i=0 ; i<conf->raid_disks; i++) {
1855                         int still_degraded;
1856                         r10bio_t *rb2;
1857                         sector_t sect;
1858                         int must_sync;
1859
1860                         if (conf->mirrors[i].rdev == NULL ||
1861                             test_bit(In_sync, &conf->mirrors[i].rdev->flags))
1862                                 continue;
1863
1864                         still_degraded = 0;
1865                         /* want to reconstruct this device */
1866                         rb2 = r10_bio;
1867                         sect = raid10_find_virt(conf, sector_nr, i);
1868                         /* Unless we are doing a full sync, we only need
1869                          * to recover the block if it is set in the bitmap
1870                          */
1871                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
1872                                                       &sync_blocks, 1);
1873                         if (sync_blocks < max_sync)
1874                                 max_sync = sync_blocks;
1875                         if (!must_sync &&
1876                             !conf->fullsync) {
1877                                 /* yep, skip the sync_blocks here, but don't assume
1878                                  * that there will never be anything to do here
1879                                  */
1880                                 chunks_skipped = -1;
1881                                 continue;
1882                         }
1883
1884                         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1885                         raise_barrier(conf, rb2 != NULL);
1886                         atomic_set(&r10_bio->remaining, 0);
1887
1888                         r10_bio->master_bio = (struct bio*)rb2;
1889                         if (rb2)
1890                                 atomic_inc(&rb2->remaining);
1891                         r10_bio->mddev = mddev;
1892                         set_bit(R10BIO_IsRecover, &r10_bio->state);
1893                         r10_bio->sector = sect;
1894
1895                         raid10_find_phys(conf, r10_bio);
1896
1897                         /* Need to check if the array will still be
1898                          * degraded
1899                          */
1900                         for (j=0; j<conf->raid_disks; j++)
1901                                 if (conf->mirrors[j].rdev == NULL ||
1902                                     test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
1903                                         still_degraded = 1;
1904                                         break;
1905                                 }
1906
1907                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
1908                                                       &sync_blocks, still_degraded);
1909
1910                         for (j=0; j<conf->copies;j++) {
1911                                 int d = r10_bio->devs[j].devnum;
1912                                 if (!conf->mirrors[d].rdev ||
1913                                     !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
1914                                         continue;
1915                                 /* This is where we read from */
1916                                 bio = r10_bio->devs[0].bio;
1917                                 bio->bi_next = biolist;
1918                                 biolist = bio;
1919                                 bio->bi_private = r10_bio;
1920                                 bio->bi_end_io = end_sync_read;
1921                                 bio->bi_rw = READ;
1922                                 bio->bi_sector = r10_bio->devs[j].addr +
1923                                         conf->mirrors[d].rdev->data_offset;
1924                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1925                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1926                                 atomic_inc(&r10_bio->remaining);
1927                                 /* and we write to 'i' */
1928
1929                                 for (k=0; k<conf->copies; k++)
1930                                         if (r10_bio->devs[k].devnum == i)
1931                                                 break;
1932                                 BUG_ON(k == conf->copies);
1933                                 bio = r10_bio->devs[1].bio;
1934                                 bio->bi_next = biolist;
1935                                 biolist = bio;
1936                                 bio->bi_private = r10_bio;
1937                                 bio->bi_end_io = end_sync_write;
1938                                 bio->bi_rw = WRITE;
1939                                 bio->bi_sector = r10_bio->devs[k].addr +
1940                                         conf->mirrors[i].rdev->data_offset;
1941                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1942
1943                                 r10_bio->devs[0].devnum = d;
1944                                 r10_bio->devs[1].devnum = i;
1945
1946                                 break;
1947                         }
1948                         if (j == conf->copies) {
1949                                 /* Cannot recover, so abort the recovery */
1950                                 put_buf(r10_bio);
1951                                 if (rb2)
1952                                         atomic_dec(&rb2->remaining);
1953                                 r10_bio = rb2;
1954                                 if (!test_and_set_bit(MD_RECOVERY_INTR,
1955                                                       &mddev->recovery))
1956                                         printk(KERN_INFO "md/raid10:%s: insufficient "
1957                                                "working devices for recovery.\n",
1958                                                mdname(mddev));
1959                                 break;
1960                         }
1961                 }
1962                 if (biolist == NULL) {
1963                         while (r10_bio) {
1964                                 r10bio_t *rb2 = r10_bio;
1965                                 r10_bio = (r10bio_t*) rb2->master_bio;
1966                                 rb2->master_bio = NULL;
1967                                 put_buf(rb2);
1968                         }
1969                         goto giveup;
1970                 }
1971         } else {
1972                 /* resync. Schedule a read for every block at this virt offset */
1973                 int count = 0;
1974
1975                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1976
1977                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1978                                        &sync_blocks, mddev->degraded) &&
1979                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
1980                                                  &mddev->recovery)) {
1981                         /* We can skip this block */
1982                         *skipped = 1;
1983                         return sync_blocks + sectors_skipped;
1984                 }
1985                 if (sync_blocks < max_sync)
1986                         max_sync = sync_blocks;
1987                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1988
1989                 r10_bio->mddev = mddev;
1990                 atomic_set(&r10_bio->remaining, 0);
1991                 raise_barrier(conf, 0);
1992                 conf->next_resync = sector_nr;
1993
1994                 r10_bio->master_bio = NULL;
1995                 r10_bio->sector = sector_nr;
1996                 set_bit(R10BIO_IsSync, &r10_bio->state);
1997                 raid10_find_phys(conf, r10_bio);
1998                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1999
2000                 for (i=0; i<conf->copies; i++) {
2001                         int d = r10_bio->devs[i].devnum;
2002                         bio = r10_bio->devs[i].bio;
2003                         bio->bi_end_io = NULL;
2004                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
2005                         if (conf->mirrors[d].rdev == NULL ||
2006                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2007                                 continue;
2008                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2009                         atomic_inc(&r10_bio->remaining);
2010                         bio->bi_next = biolist;
2011                         biolist = bio;
2012                         bio->bi_private = r10_bio;
2013                         bio->bi_end_io = end_sync_read;
2014                         bio->bi_rw = READ;
2015                         bio->bi_sector = r10_bio->devs[i].addr +
2016                                 conf->mirrors[d].rdev->data_offset;
2017                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2018                         count++;
2019                 }
2020
2021                 if (count < 2) {
2022                         for (i=0; i<conf->copies; i++) {
2023                                 int d = r10_bio->devs[i].devnum;
2024                                 if (r10_bio->devs[i].bio->bi_end_io)
2025                                         rdev_dec_pending(conf->mirrors[d].rdev,
2026                                                          mddev);
2027                         }
2028                         put_buf(r10_bio);
2029                         biolist = NULL;
2030                         goto giveup;
2031                 }
2032         }
2033
2034         for (bio = biolist; bio ; bio=bio->bi_next) {
2035
2036                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2037                 if (bio->bi_end_io)
2038                         bio->bi_flags |= 1 << BIO_UPTODATE;
2039                 bio->bi_vcnt = 0;
2040                 bio->bi_idx = 0;
2041                 bio->bi_phys_segments = 0;
2042                 bio->bi_size = 0;
2043         }
2044
2045         nr_sectors = 0;
2046         if (sector_nr + max_sync < max_sector)
2047                 max_sector = sector_nr + max_sync;
2048         do {
2049                 struct page *page;
2050                 int len = PAGE_SIZE;
2051                 if (sector_nr + (len>>9) > max_sector)
2052                         len = (max_sector - sector_nr) << 9;
2053                 if (len == 0)
2054                         break;
2055                 for (bio= biolist ; bio ; bio=bio->bi_next) {
2056                         struct bio *bio2;
2057                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2058                         if (bio_add_page(bio, page, len, 0))
2059                                 continue;
2060
2061                         /* stop here */
2062                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2063                         for (bio2 = biolist;
2064                              bio2 && bio2 != bio;
2065                              bio2 = bio2->bi_next) {
2066                                 /* remove last page from this bio */
2067                                 bio2->bi_vcnt--;
2068                                 bio2->bi_size -= len;
2069                                 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2070                         }
2071                         goto bio_full;
2072                 }
2073                 nr_sectors += len>>9;
2074                 sector_nr += len>>9;
2075         } while (biolist->bi_vcnt < RESYNC_PAGES);
2076  bio_full:
2077         r10_bio->sectors = nr_sectors;
2078
2079         while (biolist) {
2080                 bio = biolist;
2081                 biolist = biolist->bi_next;
2082
2083                 bio->bi_next = NULL;
2084                 r10_bio = bio->bi_private;
2085                 r10_bio->sectors = nr_sectors;
2086
2087                 if (bio->bi_end_io == end_sync_read) {
2088                         md_sync_acct(bio->bi_bdev, nr_sectors);
2089                         generic_make_request(bio);
2090                 }
2091         }
2092
2093         if (sectors_skipped)
2094                 /* pretend they weren't skipped, it makes
2095                  * no important difference in this case
2096                  */
2097                 md_done_sync(mddev, sectors_skipped, 1);
2098
2099         return sectors_skipped + nr_sectors;
2100  giveup:
2101         /* There is nowhere to write, so all non-sync
2102          * drives must be failed, so try the next chunk...
2103          */
2104         if (sector_nr + max_sync < max_sector)
2105                 max_sector = sector_nr + max_sync;
2106
2107         sectors_skipped += (max_sector - sector_nr);
2108         chunks_skipped ++;
2109         sector_nr = max_sector;
2110         goto skipped;
2111 }
2112
2113 static sector_t
2114 raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2115 {
2116         sector_t size;
2117         conf_t *conf = mddev->private;
2118
2119         if (!raid_disks)
2120                 raid_disks = conf->raid_disks;
2121         if (!sectors)
2122                 sectors = conf->dev_sectors;
2123
2124         size = sectors >> conf->chunk_shift;
2125         sector_div(size, conf->far_copies);
2126         size = size * raid_disks;
2127         sector_div(size, conf->near_copies);
2128
2129         return size << conf->chunk_shift;
2130 }
2131
2132
2133 static conf_t *setup_conf(mddev_t *mddev)
2134 {
2135         conf_t *conf = NULL;
2136         int nc, fc, fo;
2137         sector_t stride, size;
2138         int err = -EINVAL;
2139
2140         if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2141             !is_power_of_2(mddev->new_chunk_sectors)) {
2142                 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2143                        "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2144                        mdname(mddev), PAGE_SIZE);
2145                 goto out;
2146         }
2147
2148         nc = mddev->new_layout & 255;
2149         fc = (mddev->new_layout >> 8) & 255;
2150         fo = mddev->new_layout & (1<<16);
2151
2152         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2153             (mddev->new_layout >> 17)) {
2154                 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2155                        mdname(mddev), mddev->new_layout);
2156                 goto out;
2157         }
2158
2159         err = -ENOMEM;
2160         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2161         if (!conf)
2162                 goto out;
2163
2164         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2165                                 GFP_KERNEL);
2166         if (!conf->mirrors)
2167                 goto out;
2168
2169         conf->tmppage = alloc_page(GFP_KERNEL);
2170         if (!conf->tmppage)
2171                 goto out;
2172
2173
2174         conf->raid_disks = mddev->raid_disks;
2175         conf->near_copies = nc;
2176         conf->far_copies = fc;
2177         conf->copies = nc*fc;
2178         conf->far_offset = fo;
2179         conf->chunk_mask = mddev->new_chunk_sectors - 1;
2180         conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2181
2182         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2183                                            r10bio_pool_free, conf);
2184         if (!conf->r10bio_pool)
2185                 goto out;
2186
2187         size = mddev->dev_sectors >> conf->chunk_shift;
2188         sector_div(size, fc);
2189         size = size * conf->raid_disks;
2190         sector_div(size, nc);
2191         /* 'size' is now the number of chunks in the array */
2192         /* calculate "used chunks per device" in 'stride' */
2193         stride = size * conf->copies;
2194
2195         /* We need to round up when dividing by raid_disks to
2196          * get the stride size.
2197          */
2198         stride += conf->raid_disks - 1;
2199         sector_div(stride, conf->raid_disks);
2200
2201         conf->dev_sectors = stride << conf->chunk_shift;
2202
2203         if (fo)
2204                 stride = 1;
2205         else
2206                 sector_div(stride, fc);
2207         conf->stride = stride << conf->chunk_shift;
2208
2209
2210         spin_lock_init(&conf->device_lock);
2211         INIT_LIST_HEAD(&conf->retry_list);
2212
2213         spin_lock_init(&conf->resync_lock);
2214         init_waitqueue_head(&conf->wait_barrier);
2215
2216         conf->thread = md_register_thread(raid10d, mddev, NULL);
2217         if (!conf->thread)
2218                 goto out;
2219
2220         conf->mddev = mddev;
2221         return conf;
2222
2223  out:
2224         printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2225                mdname(mddev));
2226         if (conf) {
2227                 if (conf->r10bio_pool)
2228                         mempool_destroy(conf->r10bio_pool);
2229                 kfree(conf->mirrors);
2230                 safe_put_page(conf->tmppage);
2231                 kfree(conf);
2232         }
2233         return ERR_PTR(err);
2234 }
2235
2236 static int run(mddev_t *mddev)
2237 {
2238         conf_t *conf;
2239         int i, disk_idx, chunk_size;
2240         mirror_info_t *disk;
2241         mdk_rdev_t *rdev;
2242         sector_t size;
2243
2244         /*
2245          * copy the already verified devices into our private RAID10
2246          * bookkeeping area. [whatever we allocate in run(),
2247          * should be freed in stop()]
2248          */
2249
2250         if (mddev->private == NULL) {
2251                 conf = setup_conf(mddev);
2252                 if (IS_ERR(conf))
2253                         return PTR_ERR(conf);
2254                 mddev->private = conf;
2255         }
2256         conf = mddev->private;
2257         if (!conf)
2258                 goto out;
2259
2260         mddev->thread = conf->thread;
2261         conf->thread = NULL;
2262
2263         chunk_size = mddev->chunk_sectors << 9;
2264         blk_queue_io_min(mddev->queue, chunk_size);
2265         if (conf->raid_disks % conf->near_copies)
2266                 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2267         else
2268                 blk_queue_io_opt(mddev->queue, chunk_size *
2269                                  (conf->raid_disks / conf->near_copies));
2270
2271         list_for_each_entry(rdev, &mddev->disks, same_set) {
2272
2273                 if (rdev->badblocks.count) {
2274                         printk(KERN_ERR "md/raid10: cannot handle bad blocks yet\n");
2275                         goto out_free_conf;
2276                 }
2277                 disk_idx = rdev->raid_disk;
2278                 if (disk_idx >= conf->raid_disks
2279                     || disk_idx < 0)
2280                         continue;
2281                 disk = conf->mirrors + disk_idx;
2282
2283                 disk->rdev = rdev;
2284                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2285                                   rdev->data_offset << 9);
2286                 /* as we don't honour merge_bvec_fn, we must never risk
2287                  * violating it, so limit max_segments to 1 lying
2288                  * within a single page.
2289                  */
2290                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2291                         blk_queue_max_segments(mddev->queue, 1);
2292                         blk_queue_segment_boundary(mddev->queue,
2293                                                    PAGE_CACHE_SIZE - 1);
2294                 }
2295
2296                 disk->head_position = 0;
2297         }
2298         /* need to check that every block has at least one working mirror */
2299         if (!enough(conf, -1)) {
2300                 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2301                        mdname(mddev));
2302                 goto out_free_conf;
2303         }
2304
2305         mddev->degraded = 0;
2306         for (i = 0; i < conf->raid_disks; i++) {
2307
2308                 disk = conf->mirrors + i;
2309
2310                 if (!disk->rdev ||
2311                     !test_bit(In_sync, &disk->rdev->flags)) {
2312                         disk->head_position = 0;
2313                         mddev->degraded++;
2314                         if (disk->rdev)
2315                                 conf->fullsync = 1;
2316                 }
2317         }
2318
2319         if (mddev->recovery_cp != MaxSector)
2320                 printk(KERN_NOTICE "md/raid10:%s: not clean"
2321                        " -- starting background reconstruction\n",
2322                        mdname(mddev));
2323         printk(KERN_INFO
2324                 "md/raid10:%s: active with %d out of %d devices\n",
2325                 mdname(mddev), conf->raid_disks - mddev->degraded,
2326                 conf->raid_disks);
2327         /*
2328          * Ok, everything is just fine now
2329          */
2330         mddev->dev_sectors = conf->dev_sectors;
2331         size = raid10_size(mddev, 0, 0);
2332         md_set_array_sectors(mddev, size);
2333         mddev->resync_max_sectors = size;
2334
2335         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2336         mddev->queue->backing_dev_info.congested_data = mddev;
2337
2338         /* Calculate max read-ahead size.
2339          * We need to readahead at least twice a whole stripe....
2340          * maybe...
2341          */
2342         {
2343                 int stripe = conf->raid_disks *
2344                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2345                 stripe /= conf->near_copies;
2346                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2347                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2348         }
2349
2350         if (conf->near_copies < conf->raid_disks)
2351                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2352
2353         if (md_integrity_register(mddev))
2354                 goto out_free_conf;
2355
2356         return 0;
2357
2358 out_free_conf:
2359         md_unregister_thread(mddev->thread);
2360         if (conf->r10bio_pool)
2361                 mempool_destroy(conf->r10bio_pool);
2362         safe_put_page(conf->tmppage);
2363         kfree(conf->mirrors);
2364         kfree(conf);
2365         mddev->private = NULL;
2366 out:
2367         return -EIO;
2368 }
2369
2370 static int stop(mddev_t *mddev)
2371 {
2372         conf_t *conf = mddev->private;
2373
2374         raise_barrier(conf, 0);
2375         lower_barrier(conf);
2376
2377         md_unregister_thread(mddev->thread);
2378         mddev->thread = NULL;
2379         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2380         if (conf->r10bio_pool)
2381                 mempool_destroy(conf->r10bio_pool);
2382         kfree(conf->mirrors);
2383         kfree(conf);
2384         mddev->private = NULL;
2385         return 0;
2386 }
2387
2388 static void raid10_quiesce(mddev_t *mddev, int state)
2389 {
2390         conf_t *conf = mddev->private;
2391
2392         switch(state) {
2393         case 1:
2394                 raise_barrier(conf, 0);
2395                 break;
2396         case 0:
2397                 lower_barrier(conf);
2398                 break;
2399         }
2400 }
2401
2402 static void *raid10_takeover_raid0(mddev_t *mddev)
2403 {
2404         mdk_rdev_t *rdev;
2405         conf_t *conf;
2406
2407         if (mddev->degraded > 0) {
2408                 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
2409                        mdname(mddev));
2410                 return ERR_PTR(-EINVAL);
2411         }
2412
2413         /* Set new parameters */
2414         mddev->new_level = 10;
2415         /* new layout: far_copies = 1, near_copies = 2 */
2416         mddev->new_layout = (1<<8) + 2;
2417         mddev->new_chunk_sectors = mddev->chunk_sectors;
2418         mddev->delta_disks = mddev->raid_disks;
2419         mddev->raid_disks *= 2;
2420         /* make sure it will be not marked as dirty */
2421         mddev->recovery_cp = MaxSector;
2422
2423         conf = setup_conf(mddev);
2424         if (!IS_ERR(conf)) {
2425                 list_for_each_entry(rdev, &mddev->disks, same_set)
2426                         if (rdev->raid_disk >= 0)
2427                                 rdev->new_raid_disk = rdev->raid_disk * 2;
2428                 conf->barrier = 1;
2429         }
2430
2431         return conf;
2432 }
2433
2434 static void *raid10_takeover(mddev_t *mddev)
2435 {
2436         struct raid0_private_data *raid0_priv;
2437
2438         /* raid10 can take over:
2439          *  raid0 - providing it has only two drives
2440          */
2441         if (mddev->level == 0) {
2442                 /* for raid0 takeover only one zone is supported */
2443                 raid0_priv = mddev->private;
2444                 if (raid0_priv->nr_strip_zones > 1) {
2445                         printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
2446                                " with more than one zone.\n",
2447                                mdname(mddev));
2448                         return ERR_PTR(-EINVAL);
2449                 }
2450                 return raid10_takeover_raid0(mddev);
2451         }
2452         return ERR_PTR(-EINVAL);
2453 }
2454
2455 static struct mdk_personality raid10_personality =
2456 {
2457         .name           = "raid10",
2458         .level          = 10,
2459         .owner          = THIS_MODULE,
2460         .make_request   = make_request,
2461         .run            = run,
2462         .stop           = stop,
2463         .status         = status,
2464         .error_handler  = error,
2465         .hot_add_disk   = raid10_add_disk,
2466         .hot_remove_disk= raid10_remove_disk,
2467         .spare_active   = raid10_spare_active,
2468         .sync_request   = sync_request,
2469         .quiesce        = raid10_quiesce,
2470         .size           = raid10_size,
2471         .takeover       = raid10_takeover,
2472 };
2473
2474 static int __init raid_init(void)
2475 {
2476         return register_md_personality(&raid10_personality);
2477 }
2478
2479 static void raid_exit(void)
2480 {
2481         unregister_md_personality(&raid10_personality);
2482 }
2483
2484 module_init(raid_init);
2485 module_exit(raid_exit);
2486 MODULE_LICENSE("GPL");
2487 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
2488 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2489 MODULE_ALIAS("md-raid10");
2490 MODULE_ALIAS("md-level-10");