drivers/md/raid10.c

   1 /*
   2  * raid10.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 2000-2004 Neil Brown
   5  *
   6  * RAID-10 support for md.
   7  *
   8  * Base on code in raid1.c.  See raid1.c for futher copyright information.
   9  *
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2, or (at your option)
  14  * any later version.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * (for example /usr/src/linux/COPYING); if not, write to the Free
  18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 #include <linux/delay.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/seq_file.h>
  24 #include "md.h"
  25 #include "raid10.h"
  26 #include "bitmap.h"
  27
  28 /*
  29  * RAID10 provides a combination of RAID0 and RAID1 functionality.
  30  * The layout of data is defined by
  31  *    chunk_size
  32  *    raid_disks
  33  *    near_copies (stored in low byte of layout)
  34  *    far_copies (stored in second byte of layout)
  35  *    far_offset (stored in bit 16 of layout )
  36  *
  37  * The data to be stored is divided into chunks using chunksize.
  38  * Each device is divided into far_copies sections.
  39  * In each section, chunks are laid out in a style similar to raid0, but
  40  * near_copies copies of each chunk is stored (each on a different drive).
  41  * The starting device for each section is offset near_copies from the starting
  42  * device of the previous section.
  43  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
  44  * drive.
  45  * near_copies and far_copies must be at least one, and their product is at most
  46  * raid_disks.
  47  *
  48  * If far_offset is true, then the far_copies are handled a bit differently.
  49  * The copies are still in different stripes, but instead of be very far apart
  50  * on disk, there are adjacent stripes.
  51  */
  52
  53 /*
  54  * Number of guaranteed r10bios in case of extreme VM load:
  55  */
  56 #define NR_RAID10_BIOS 256
  57
  58 static void unplug_slaves(mddev_t *mddev);
  59
  60 static void allow_barrier(conf_t *conf);
  61 static void lower_barrier(conf_t *conf);
  62
  63 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
  64 {
  65         conf_t *conf = data;
  66         r10bio_t *r10_bio;
  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         r10_bio = kzalloc(size, gfp_flags);
  71         if (!r10_bio && conf->mddev)
  72                 unplug_slaves(conf->mddev);
  73
  74         return r10_bio;
  75 }
  76
  77 static void r10bio_pool_free(void *r10_bio, void *data)
  78 {
  79         kfree(r10_bio);
  80 }
  81
  82 /* Maximum size of each resync request */
  83 #define RESYNC_BLOCK_SIZE (64*1024)
  84 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  85 /* amount of memory to reserve for resync requests */
  86 #define RESYNC_WINDOW (1024*1024)
  87 /* maximum number of concurrent requests, memory permitting */
  88 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
  89
  90 /*
  91  * When performing a resync, we need to read and compare, so
  92  * we need as many pages are there are copies.
  93  * When performing a recovery, we need 2 bios, one for read,
  94  * one for write (we recover only one drive per r10buf)
  95  *
  96  */
  97 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
  98 {
  99         conf_t *conf = data;
 100         struct page *page;
 101         r10bio_t *r10_bio;
 102         struct bio *bio;
 103         int i, j;
 104         int nalloc;
 105
 106         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 107         if (!r10_bio) {
 108                 unplug_slaves(conf->mddev);
 109                 return NULL;
 110         }
 111
 112         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
 113                 nalloc = conf->copies; /* resync */
 114         else
 115                 nalloc = 2; /* recovery */
 116
 117         /*
 118          * Allocate bios.
 119          */
 120         for (j = nalloc ; j-- ; ) {
 121                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 122                 if (!bio)
 123                         goto out_free_bio;
 124                 r10_bio->devs[j].bio = bio;
 125         }
 126         /*
 127          * Allocate RESYNC_PAGES data pages and attach them
 128          * where needed.
 129          */
 130         for (j = 0 ; j < nalloc; j++) {
 131                 bio = r10_bio->devs[j].bio;
 132                 for (i = 0; i < RESYNC_PAGES; i++) {
 133                         page = alloc_page(gfp_flags);
 134                         if (unlikely(!page))
 135                                 goto out_free_pages;
 136
 137                         bio->bi_io_vec[i].bv_page = page;
 138                 }
 139         }
 140
 141         return r10_bio;
 142
 143 out_free_pages:
 144         for ( ; i > 0 ; i--)
 145                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
 146         while (j--)
 147                 for (i = 0; i < RESYNC_PAGES ; i++)
 148                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 149         j = -1;
 150 out_free_bio:
 151         while ( ++j < nalloc )
 152                 bio_put(r10_bio->devs[j].bio);
 153         r10bio_pool_free(r10_bio, conf);
 154         return NULL;
 155 }
 156
 157 static void r10buf_pool_free(void *__r10_bio, void *data)
 158 {
 159         int i;
 160         conf_t *conf = data;
 161         r10bio_t *r10bio = __r10_bio;
 162         int j;
 163
 164         for (j=0; j < conf->copies; j++) {
 165                 struct bio *bio = r10bio->devs[j].bio;
 166                 if (bio) {
 167                         for (i = 0; i < RESYNC_PAGES; i++) {
 168                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 169                                 bio->bi_io_vec[i].bv_page = NULL;
 170                         }
 171                         bio_put(bio);
 172                 }
 173         }
 174         r10bio_pool_free(r10bio, conf);
 175 }
 176
 177 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
 178 {
 179         int i;
 180
 181         for (i = 0; i < conf->copies; i++) {
 182                 struct bio **bio = & r10_bio->devs[i].bio;
 183                 if (*bio && *bio != IO_BLOCKED)
 184                         bio_put(*bio);
 185                 *bio = NULL;
 186         }
 187 }
 188
 189 static void free_r10bio(r10bio_t *r10_bio)
 190 {
 191         conf_t *conf = r10_bio->mddev->private;
 192
 193         /*
 194          * Wake up any possible resync thread that waits for the device
 195          * to go idle.
 196          */
 197         allow_barrier(conf);
 198
 199         put_all_bios(conf, r10_bio);
 200         mempool_free(r10_bio, conf->r10bio_pool);
 201 }
 202
 203 static void put_buf(r10bio_t *r10_bio)
 204 {
 205         conf_t *conf = r10_bio->mddev->private;
 206
 207         mempool_free(r10_bio, conf->r10buf_pool);
 208
 209         lower_barrier(conf);
 210 }
 211
 212 static void reschedule_retry(r10bio_t *r10_bio)
 213 {
 214         unsigned long flags;
 215         mddev_t *mddev = r10_bio->mddev;
 216         conf_t *conf = mddev->private;
 217
 218         spin_lock_irqsave(&conf->device_lock, flags);
 219         list_add(&r10_bio->retry_list, &conf->retry_list);
 220         conf->nr_queued ++;
 221         spin_unlock_irqrestore(&conf->device_lock, flags);
 222
 223         /* wake up frozen array... */
 224         wake_up(&conf->wait_barrier);
 225
 226         md_wakeup_thread(mddev->thread);
 227 }
 228
 229 /*
 230  * raid_end_bio_io() is called when we have finished servicing a mirrored
 231  * operation and are ready to return a success/failure code to the buffer
 232  * cache layer.
 233  */
 234 static void raid_end_bio_io(r10bio_t *r10_bio)
 235 {
 236         struct bio *bio = r10_bio->master_bio;
 237
 238         bio_endio(bio,
 239                 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
 240         free_r10bio(r10_bio);
 241 }
 242
 243 /*
 244  * Update disk head position estimator based on IRQ completion info.
 245  */
 246 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
 247 {
 248         conf_t *conf = r10_bio->mddev->private;
 249
 250         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 251                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 252 }
 253
 254 static void raid10_end_read_request(struct bio *bio, int error)
 255 {
 256         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 257         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
 258         int slot, dev;
 259         conf_t *conf = r10_bio->mddev->private;
 260
 261
 262         slot = r10_bio->read_slot;
 263         dev = r10_bio->devs[slot].devnum;
 264         /*
 265          * this branch is our 'one mirror IO has finished' event handler:
 266          */
 267         update_head_pos(slot, r10_bio);
 268
 269         if (uptodate) {
 270                 /*
 271                  * Set R10BIO_Uptodate in our master bio, so that
 272                  * we will return a good error code to the higher
 273                  * levels even if IO on some other mirrored buffer fails.
 274                  *
 275                  * The 'master' represents the composite IO operation to
 276                  * user-side. So if something waits for IO, then it will
 277                  * wait for the 'master' bio.
 278                  */
 279                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 280                 raid_end_bio_io(r10_bio);
 281         } else {
 282                 /*
 283                  * oops, read error:
 284                  */
 285                 char b[BDEVNAME_SIZE];
 286                 if (printk_ratelimit())
 287                         printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
 288                                bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
 289                 reschedule_retry(r10_bio);
 290         }
 291
 292         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 293 }
 294
 295 static void raid10_end_write_request(struct bio *bio, int error)
 296 {
 297         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 298         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
 299         int slot, dev;
 300         conf_t *conf = r10_bio->mddev->private;
 301
 302         for (slot = 0; slot < conf->copies; slot++)
 303                 if (r10_bio->devs[slot].bio == bio)
 304                         break;
 305         dev = r10_bio->devs[slot].devnum;
 306
 307         /*
 308          * this branch is our 'one mirror IO has finished' event handler:
 309          */
 310         if (!uptodate) {
 311                 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
 312                 /* an I/O failed, we can't clear the bitmap */
 313                 set_bit(R10BIO_Degraded, &r10_bio->state);
 314         } else
 315                 /*
 316                  * Set R10BIO_Uptodate in our master bio, so that
 317                  * we will return a good error code for to the higher
 318                  * levels even if IO on some other mirrored buffer fails.
 319                  *
 320                  * The 'master' represents the composite IO operation to
 321                  * user-side. So if something waits for IO, then it will
 322                  * wait for the 'master' bio.
 323                  */
 324                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 325
 326         update_head_pos(slot, r10_bio);
 327
 328         /*
 329          *
 330          * Let's see if all mirrored write operations have finished
 331          * already.
 332          */
 333         if (atomic_dec_and_test(&r10_bio->remaining)) {
 334                 /* clear the bitmap if all writes complete successfully */
 335                 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 336                                 r10_bio->sectors,
 337                                 !test_bit(R10BIO_Degraded, &r10_bio->state),
 338                                 0);
 339                 md_write_end(r10_bio->mddev);
 340                 raid_end_bio_io(r10_bio);
 341         }
 342
 343         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 344 }
 345
 346
 347 /*
 348  * RAID10 layout manager
 349  * Aswell as the chunksize and raid_disks count, there are two
 350  * parameters: near_copies and far_copies.
 351  * near_copies * far_copies must be <= raid_disks.
 352  * Normally one of these will be 1.
 353  * If both are 1, we get raid0.
 354  * If near_copies == raid_disks, we get raid1.
 355  *
 356  * Chunks are layed out in raid0 style with near_copies copies of the
 357  * first chunk, followed by near_copies copies of the next chunk and
 358  * so on.
 359  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 360  * as described above, we start again with a device offset of near_copies.
 361  * So we effectively have another copy of the whole array further down all
 362  * the drives, but with blocks on different drives.
 363  * With this layout, and block is never stored twice on the one device.
 364  *
 365  * raid10_find_phys finds the sector offset of a given virtual sector
 366  * on each device that it is on.
 367  *
 368  * raid10_find_virt does the reverse mapping, from a device and a
 369  * sector offset to a virtual address
 370  */
 371
 372 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
 373 {
 374         int n,f;
 375         sector_t sector;
 376         sector_t chunk;
 377         sector_t stripe;
 378         int dev;
 379
 380         int slot = 0;
 381
 382         /* now calculate first sector/dev */
 383         chunk = r10bio->sector >> conf->chunk_shift;
 384         sector = r10bio->sector & conf->chunk_mask;
 385
 386         chunk *= conf->near_copies;
 387         stripe = chunk;
 388         dev = sector_div(stripe, conf->raid_disks);
 389         if (conf->far_offset)
 390                 stripe *= conf->far_copies;
 391
 392         sector += stripe << conf->chunk_shift;
 393
 394         /* and calculate all the others */
 395         for (n=0; n < conf->near_copies; n++) {
 396                 int d = dev;
 397                 sector_t s = sector;
 398                 r10bio->devs[slot].addr = sector;
 399                 r10bio->devs[slot].devnum = d;
 400                 slot++;
 401
 402                 for (f = 1; f < conf->far_copies; f++) {
 403                         d += conf->near_copies;
 404                         if (d >= conf->raid_disks)
 405                                 d -= conf->raid_disks;
 406                         s += conf->stride;
 407                         r10bio->devs[slot].devnum = d;
 408                         r10bio->devs[slot].addr = s;
 409                         slot++;
 410                 }
 411                 dev++;
 412                 if (dev >= conf->raid_disks) {
 413                         dev = 0;
 414                         sector += (conf->chunk_mask + 1);
 415                 }
 416         }
 417         BUG_ON(slot != conf->copies);
 418 }
 419
 420 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
 421 {
 422         sector_t offset, chunk, vchunk;
 423
 424         offset = sector & conf->chunk_mask;
 425         if (conf->far_offset) {
 426                 int fc;
 427                 chunk = sector >> conf->chunk_shift;
 428                 fc = sector_div(chunk, conf->far_copies);
 429                 dev -= fc * conf->near_copies;
 430                 if (dev < 0)
 431                         dev += conf->raid_disks;
 432         } else {
 433                 while (sector >= conf->stride) {
 434                         sector -= conf->stride;
 435                         if (dev < conf->near_copies)
 436                                 dev += conf->raid_disks - conf->near_copies;
 437                         else
 438                                 dev -= conf->near_copies;
 439                 }
 440                 chunk = sector >> conf->chunk_shift;
 441         }
 442         vchunk = chunk * conf->raid_disks + dev;
 443         sector_div(vchunk, conf->near_copies);
 444         return (vchunk << conf->chunk_shift) + offset;
 445 }
 446
 447 /**
 448  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 449  *      @q: request queue
 450  *      @bvm: properties of new bio
 451  *      @biovec: the request that could be merged to it.
 452  *
 453  *      Return amount of bytes we can accept at this offset
 454  *      If near_copies == raid_disk, there are no striping issues,
 455  *      but in that case, the function isn't called at all.
 456  */
 457 static int raid10_mergeable_bvec(struct request_queue *q,
 458                                  struct bvec_merge_data *bvm,
 459                                  struct bio_vec *biovec)
 460 {
 461         mddev_t *mddev = q->queuedata;
 462         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 463         int max;
 464         unsigned int chunk_sectors = mddev->chunk_sectors;
 465         unsigned int bio_sectors = bvm->bi_size >> 9;
 466
 467         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 468         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 469         if (max <= biovec->bv_len && bio_sectors == 0)
 470                 return biovec->bv_len;
 471         else
 472                 return max;
 473 }
 474
 475 /*
 476  * This routine returns the disk from which the requested read should
 477  * be done. There is a per-array 'next expected sequential IO' sector
 478  * number - if this matches on the next IO then we use the last disk.
 479  * There is also a per-disk 'last know head position' sector that is
 480  * maintained from IRQ contexts, both the normal and the resync IO
 481  * completion handlers update this position correctly. If there is no
 482  * perfect sequential match then we pick the disk whose head is closest.
 483  *
 484  * If there are 2 mirrors in the same 2 devices, performance degrades
 485  * because position is mirror, not device based.
 486  *
 487  * The rdev for the device selected will have nr_pending incremented.
 488  */
 489
 490 /*
 491  * FIXME: possibly should rethink readbalancing and do it differently
 492  * depending on near_copies / far_copies geometry.
 493  */
 494 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
 495 {
 496         const unsigned long this_sector = r10_bio->sector;
 497         int disk, slot, nslot;
 498         const int sectors = r10_bio->sectors;
 499         sector_t new_distance, current_distance;
 500         mdk_rdev_t *rdev;
 501
 502         raid10_find_phys(conf, r10_bio);
 503         rcu_read_lock();
 504         /*
 505          * Check if we can balance. We can balance on the whole
 506          * device if no resync is going on (recovery is ok), or below
 507          * the resync window. We take the first readable disk when
 508          * above the resync window.
 509          */
 510         if (conf->mddev->recovery_cp < MaxSector
 511             && (this_sector + sectors >= conf->next_resync)) {
 512                 /* make sure that disk is operational */
 513                 slot = 0;
 514                 disk = r10_bio->devs[slot].devnum;
 515
 516                 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 517                        r10_bio->devs[slot].bio == IO_BLOCKED ||
 518                        !test_bit(In_sync, &rdev->flags)) {
 519                         slot++;
 520                         if (slot == conf->copies) {
 521                                 slot = 0;
 522                                 disk = -1;
 523                                 break;
 524                         }
 525                         disk = r10_bio->devs[slot].devnum;
 526                 }
 527                 goto rb_out;
 528         }
 529
 530
 531         /* make sure the disk is operational */
 532         slot = 0;
 533         disk = r10_bio->devs[slot].devnum;
 534         while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 535                r10_bio->devs[slot].bio == IO_BLOCKED ||
 536                !test_bit(In_sync, &rdev->flags)) {
 537                 slot ++;
 538                 if (slot == conf->copies) {
 539                         disk = -1;
 540                         goto rb_out;
 541                 }
 542                 disk = r10_bio->devs[slot].devnum;
 543         }
 544
 545
 546         current_distance = abs(r10_bio->devs[slot].addr -
 547                                conf->mirrors[disk].head_position);
 548
 549         /* Find the disk whose head is closest,
 550          * or - for far > 1 - find the closest to partition beginning */
 551
 552         for (nslot = slot; nslot < conf->copies; nslot++) {
 553                 int ndisk = r10_bio->devs[nslot].devnum;
 554
 555
 556                 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
 557                     r10_bio->devs[nslot].bio == IO_BLOCKED ||
 558                     !test_bit(In_sync, &rdev->flags))
 559                         continue;
 560
 561                 /* This optimisation is debatable, and completely destroys
 562                  * sequential read speed for 'far copies' arrays.  So only
 563                  * keep it for 'near' arrays, and review those later.
 564                  */
 565                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
 566                         disk = ndisk;
 567                         slot = nslot;
 568                         break;
 569                 }
 570
 571                 /* for far > 1 always use the lowest address */
 572                 if (conf->far_copies > 1)
 573                         new_distance = r10_bio->devs[nslot].addr;
 574                 else
 575                         new_distance = abs(r10_bio->devs[nslot].addr -
 576                                            conf->mirrors[ndisk].head_position);
 577                 if (new_distance < current_distance) {
 578                         current_distance = new_distance;
 579                         disk = ndisk;
 580                         slot = nslot;
 581                 }
 582         }
 583
 584 rb_out:
 585         r10_bio->read_slot = slot;
 586 /*      conf->next_seq_sect = this_sector + sectors;*/
 587
 588         if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
 589                 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
 590         else
 591                 disk = -1;
 592         rcu_read_unlock();
 593
 594         return disk;
 595 }
 596
 597 static void unplug_slaves(mddev_t *mddev)
 598 {
 599         conf_t *conf = mddev->private;
 600         int i;
 601
 602         rcu_read_lock();
 603         for (i=0; i<mddev->raid_disks; i++) {
 604                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 605                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 606                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 607
 608                         atomic_inc(&rdev->nr_pending);
 609                         rcu_read_unlock();
 610
 611                         blk_unplug(r_queue);
 612
 613                         rdev_dec_pending(rdev, mddev);
 614                         rcu_read_lock();
 615                 }
 616         }
 617         rcu_read_unlock();
 618 }
 619
 620 static void raid10_unplug(struct request_queue *q)
 621 {
 622         mddev_t *mddev = q->queuedata;
 623
 624         unplug_slaves(q->queuedata);
 625         md_wakeup_thread(mddev->thread);
 626 }
 627
 628 static int raid10_congested(void *data, int bits)
 629 {
 630         mddev_t *mddev = data;
 631         conf_t *conf = mddev->private;
 632         int i, ret = 0;
 633
 634         if (mddev_congested(mddev, bits))
 635                 return 1;
 636         rcu_read_lock();
 637         for (i = 0; i < mddev->raid_disks && ret == 0; i++) {
 638                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 639                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 640                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 641
 642                         ret |= bdi_congested(&q->backing_dev_info, bits);
 643                 }
 644         }
 645         rcu_read_unlock();
 646         return ret;
 647 }
 648
 649 static int flush_pending_writes(conf_t *conf)
 650 {
 651         /* Any writes that have been queued but are awaiting
 652          * bitmap updates get flushed here.
 653          * We return 1 if any requests were actually submitted.
 654          */
 655         int rv = 0;
 656
 657         spin_lock_irq(&conf->device_lock);
 658
 659         if (conf->pending_bio_list.head) {
 660                 struct bio *bio;
 661                 bio = bio_list_get(&conf->pending_bio_list);
 662                 blk_remove_plug(conf->mddev->queue);
 663                 spin_unlock_irq(&conf->device_lock);
 664                 /* flush any pending bitmap writes to disk
 665                  * before proceeding w/ I/O */
 666                 bitmap_unplug(conf->mddev->bitmap);
 667
 668                 while (bio) { /* submit pending writes */
 669                         struct bio *next = bio->bi_next;
 670                         bio->bi_next = NULL;
 671                         generic_make_request(bio);
 672                         bio = next;
 673                 }
 674                 rv = 1;
 675         } else
 676                 spin_unlock_irq(&conf->device_lock);
 677         return rv;
 678 }
 679 /* Barriers....
 680  * Sometimes we need to suspend IO while we do something else,
 681  * either some resync/recovery, or reconfigure the array.
 682  * To do this we raise a 'barrier'.
 683  * The 'barrier' is a counter that can be raised multiple times
 684  * to count how many activities are happening which preclude
 685  * normal IO.
 686  * We can only raise the barrier if there is no pending IO.
 687  * i.e. if nr_pending == 0.
 688  * We choose only to raise the barrier if no-one is waiting for the
 689  * barrier to go down.  This means that as soon as an IO request
 690  * is ready, no other operations which require a barrier will start
 691  * until the IO request has had a chance.
 692  *
 693  * So: regular IO calls 'wait_barrier'.  When that returns there
 694  *    is no backgroup IO happening,  It must arrange to call
 695  *    allow_barrier when it has finished its IO.
 696  * backgroup IO calls must call raise_barrier.  Once that returns
 697  *    there is no normal IO happeing.  It must arrange to call
 698  *    lower_barrier when the particular background IO completes.
 699  */
 700
 701 static void raise_barrier(conf_t *conf, int force)
 702 {
 703         BUG_ON(force && !conf->barrier);
 704         spin_lock_irq(&conf->resync_lock);
 705
 706         /* Wait until no block IO is waiting (unless 'force') */
 707         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 708                             conf->resync_lock,
 709                             raid10_unplug(conf->mddev->queue));
 710
 711         /* block any new IO from starting */
 712         conf->barrier++;
 713
 714         /* No wait for all pending IO to complete */
 715         wait_event_lock_irq(conf->wait_barrier,
 716                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 717                             conf->resync_lock,
 718                             raid10_unplug(conf->mddev->queue));
 719
 720         spin_unlock_irq(&conf->resync_lock);
 721 }
 722
 723 static void lower_barrier(conf_t *conf)
 724 {
 725         unsigned long flags;
 726         spin_lock_irqsave(&conf->resync_lock, flags);
 727         conf->barrier--;
 728         spin_unlock_irqrestore(&conf->resync_lock, flags);
 729         wake_up(&conf->wait_barrier);
 730 }
 731
 732 static void wait_barrier(conf_t *conf)
 733 {
 734         spin_lock_irq(&conf->resync_lock);
 735         if (conf->barrier) {
 736                 conf->nr_waiting++;
 737                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 738                                     conf->resync_lock,
 739                                     raid10_unplug(conf->mddev->queue));
 740                 conf->nr_waiting--;
 741         }
 742         conf->nr_pending++;
 743         spin_unlock_irq(&conf->resync_lock);
 744 }
 745
 746 static void allow_barrier(conf_t *conf)
 747 {
 748         unsigned long flags;
 749         spin_lock_irqsave(&conf->resync_lock, flags);
 750         conf->nr_pending--;
 751         spin_unlock_irqrestore(&conf->resync_lock, flags);
 752         wake_up(&conf->wait_barrier);
 753 }
 754
 755 static void freeze_array(conf_t *conf)
 756 {
 757         /* stop syncio and normal IO and wait for everything to
 758          * go quiet.
 759          * We increment barrier and nr_waiting, and then
 760          * wait until nr_pending match nr_queued+1
 761          * This is called in the context of one normal IO request
 762          * that has failed. Thus any sync request that might be pending
 763          * will be blocked by nr_pending, and we need to wait for
 764          * pending IO requests to complete or be queued for re-try.
 765          * Thus the number queued (nr_queued) plus this request (1)
 766          * must match the number of pending IOs (nr_pending) before
 767          * we continue.
 768          */
 769         spin_lock_irq(&conf->resync_lock);
 770         conf->barrier++;
 771         conf->nr_waiting++;
 772         wait_event_lock_irq(conf->wait_barrier,
 773                             conf->nr_pending == conf->nr_queued+1,
 774                             conf->resync_lock,
 775                             ({ flush_pending_writes(conf);
 776                                raid10_unplug(conf->mddev->queue); }));
 777         spin_unlock_irq(&conf->resync_lock);
 778 }
 779
 780 static void unfreeze_array(conf_t *conf)
 781 {
 782         /* reverse the effect of the freeze */
 783         spin_lock_irq(&conf->resync_lock);
 784         conf->barrier--;
 785         conf->nr_waiting--;
 786         wake_up(&conf->wait_barrier);
 787         spin_unlock_irq(&conf->resync_lock);
 788 }
 789
 790 static int make_request(struct request_queue *q, struct bio * bio)
 791 {
 792         mddev_t *mddev = q->queuedata;
 793         conf_t *conf = mddev->private;
 794         mirror_info_t *mirror;
 795         r10bio_t *r10_bio;
 796         struct bio *read_bio;
 797         int cpu;
 798         int i;
 799         int chunk_sects = conf->chunk_mask + 1;
 800         const int rw = bio_data_dir(bio);
 801         const bool do_sync = bio_rw_flagged(bio, BIO_RW_SYNCIO);
 802         struct bio_list bl;
 803         unsigned long flags;
 804         mdk_rdev_t *blocked_rdev;
 805
 806         if (unlikely(bio_rw_flagged(bio, BIO_RW_BARRIER))) {
 807                 bio_endio(bio, -EOPNOTSUPP);
 808                 return 0;
 809         }
 810
 811         /* If this request crosses a chunk boundary, we need to
 812          * split it.  This will only happen for 1 PAGE (or less) requests.
 813          */
 814         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 815                       > chunk_sects &&
 816                     conf->near_copies < conf->raid_disks)) {
 817                 struct bio_pair *bp;
 818                 /* Sanity check -- queue functions should prevent this happening */
 819                 if (bio->bi_vcnt != 1 ||
 820                     bio->bi_idx != 0)
 821                         goto bad_map;
 822                 /* This is a one page bio that upper layers
 823                  * refuse to split for us, so we need to split it.
 824                  */
 825                 bp = bio_split(bio,
 826                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 827                 if (make_request(q, &bp->bio1))
 828                         generic_make_request(&bp->bio1);
 829                 if (make_request(q, &bp->bio2))
 830                         generic_make_request(&bp->bio2);
 831
 832                 bio_pair_release(bp);
 833                 return 0;
 834         bad_map:
 835                 printk("raid10_make_request bug: can't convert block across chunks"
 836                        " or bigger than %dk %llu %d\n", chunk_sects/2,
 837                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 838
 839                 bio_io_error(bio);
 840                 return 0;
 841         }
 842
 843         md_write_start(mddev, bio);
 844
 845         /*
 846          * Register the new request and wait if the reconstruction
 847          * thread has put up a bar for new requests.
 848          * Continue immediately if no resync is active currently.
 849          */
 850         wait_barrier(conf);
 851
 852         cpu = part_stat_lock();
 853         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
 854         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
 855                       bio_sectors(bio));
 856         part_stat_unlock();
 857
 858         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 859
 860         r10_bio->master_bio = bio;
 861         r10_bio->sectors = bio->bi_size >> 9;
 862
 863         r10_bio->mddev = mddev;
 864         r10_bio->sector = bio->bi_sector;
 865         r10_bio->state = 0;
 866
 867         if (rw == READ) {
 868                 /*
 869                  * read balancing logic:
 870                  */
 871                 int disk = read_balance(conf, r10_bio);
 872                 int slot = r10_bio->read_slot;
 873                 if (disk < 0) {
 874                         raid_end_bio_io(r10_bio);
 875                         return 0;
 876                 }
 877                 mirror = conf->mirrors + disk;
 878
 879                 read_bio = bio_clone(bio, GFP_NOIO);
 880
 881                 r10_bio->devs[slot].bio = read_bio;
 882
 883                 read_bio->bi_sector = r10_bio->devs[slot].addr +
 884                         mirror->rdev->data_offset;
 885                 read_bio->bi_bdev = mirror->rdev->bdev;
 886                 read_bio->bi_end_io = raid10_end_read_request;
 887                 read_bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
 888                 read_bio->bi_private = r10_bio;
 889
 890                 generic_make_request(read_bio);
 891                 return 0;
 892         }
 893
 894         /*
 895          * WRITE:
 896          */
 897         /* first select target devices under rcu_lock and
 898          * inc refcount on their rdev.  Record them by setting
 899          * bios[x] to bio
 900          */
 901         raid10_find_phys(conf, r10_bio);
 902  retry_write:
 903         blocked_rdev = NULL;
 904         rcu_read_lock();
 905         for (i = 0;  i < conf->copies; i++) {
 906                 int d = r10_bio->devs[i].devnum;
 907                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
 908                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 909                         atomic_inc(&rdev->nr_pending);
 910                         blocked_rdev = rdev;
 911                         break;
 912                 }
 913                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 914                         atomic_inc(&rdev->nr_pending);
 915                         r10_bio->devs[i].bio = bio;
 916                 } else {
 917                         r10_bio->devs[i].bio = NULL;
 918                         set_bit(R10BIO_Degraded, &r10_bio->state);
 919                 }
 920         }
 921         rcu_read_unlock();
 922
 923         if (unlikely(blocked_rdev)) {
 924                 /* Have to wait for this device to get unblocked, then retry */
 925                 int j;
 926                 int d;
 927
 928                 for (j = 0; j < i; j++)
 929                         if (r10_bio->devs[j].bio) {
 930                                 d = r10_bio->devs[j].devnum;
 931                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
 932                         }
 933                 allow_barrier(conf);
 934                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 935                 wait_barrier(conf);
 936                 goto retry_write;
 937         }
 938
 939         atomic_set(&r10_bio->remaining, 0);
 940
 941         bio_list_init(&bl);
 942         for (i = 0; i < conf->copies; i++) {
 943                 struct bio *mbio;
 944                 int d = r10_bio->devs[i].devnum;
 945                 if (!r10_bio->devs[i].bio)
 946                         continue;
 947
 948                 mbio = bio_clone(bio, GFP_NOIO);
 949                 r10_bio->devs[i].bio = mbio;
 950
 951                 mbio->bi_sector = r10_bio->devs[i].addr+
 952                         conf->mirrors[d].rdev->data_offset;
 953                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
 954                 mbio->bi_end_io = raid10_end_write_request;
 955                 mbio->bi_rw = WRITE | (do_sync << BIO_RW_SYNCIO);
 956                 mbio->bi_private = r10_bio;
 957
 958                 atomic_inc(&r10_bio->remaining);
 959                 bio_list_add(&bl, mbio);
 960         }
 961
 962         if (unlikely(!atomic_read(&r10_bio->remaining))) {
 963                 /* the array is dead */
 964                 md_write_end(mddev);
 965                 raid_end_bio_io(r10_bio);
 966                 return 0;
 967         }
 968
 969         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
 970         spin_lock_irqsave(&conf->device_lock, flags);
 971         bio_list_merge(&conf->pending_bio_list, &bl);
 972         blk_plug_device(mddev->queue);
 973         spin_unlock_irqrestore(&conf->device_lock, flags);
 974
 975         /* In case raid10d snuck in to freeze_array */
 976         wake_up(&conf->wait_barrier);
 977
 978         if (do_sync)
 979                 md_wakeup_thread(mddev->thread);
 980
 981         return 0;
 982 }
 983
 984 static void status(struct seq_file *seq, mddev_t *mddev)
 985 {
 986         conf_t *conf = mddev->private;
 987         int i;
 988
 989         if (conf->near_copies < conf->raid_disks)
 990                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
 991         if (conf->near_copies > 1)
 992                 seq_printf(seq, " %d near-copies", conf->near_copies);
 993         if (conf->far_copies > 1) {
 994                 if (conf->far_offset)
 995                         seq_printf(seq, " %d offset-copies", conf->far_copies);
 996                 else
 997                         seq_printf(seq, " %d far-copies", conf->far_copies);
 998         }
 999         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1000                                         conf->raid_disks - mddev->degraded);
1001         for (i = 0; i < conf->raid_disks; i++)
1002                 seq_printf(seq, "%s",
1003                               conf->mirrors[i].rdev &&
1004                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1005         seq_printf(seq, "]");
1006 }
1007
1008 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1009 {
1010         char b[BDEVNAME_SIZE];
1011         conf_t *conf = mddev->private;
1012
1013         /*
1014          * If it is not operational, then we have already marked it as dead
1015          * else if it is the last working disks, ignore the error, let the
1016          * next level up know.
1017          * else mark the drive as failed
1018          */
1019         if (test_bit(In_sync, &rdev->flags)
1020             && conf->raid_disks-mddev->degraded == 1)
1021                 /*
1022                  * Don't fail the drive, just return an IO error.
1023                  * The test should really be more sophisticated than
1024                  * "working_disks == 1", but it isn't critical, and
1025                  * can wait until we do more sophisticated "is the drive
1026                  * really dead" tests...
1027                  */
1028                 return;
1029         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1030                 unsigned long flags;
1031                 spin_lock_irqsave(&conf->device_lock, flags);
1032                 mddev->degraded++;
1033                 spin_unlock_irqrestore(&conf->device_lock, flags);
1034                 /*
1035                  * if recovery is running, make sure it aborts.
1036                  */
1037                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1038         }
1039         set_bit(Faulty, &rdev->flags);
1040         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1041         printk(KERN_ALERT "raid10: Disk failure on %s, disabling device.\n"
1042                 "raid10: Operation continuing on %d devices.\n",
1043                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1044 }
1045
1046 static void print_conf(conf_t *conf)
1047 {
1048         int i;
1049         mirror_info_t *tmp;
1050
1051         printk("RAID10 conf printout:\n");
1052         if (!conf) {
1053                 printk("(!conf)\n");
1054                 return;
1055         }
1056         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1057                 conf->raid_disks);
1058
1059         for (i = 0; i < conf->raid_disks; i++) {
1060                 char b[BDEVNAME_SIZE];
1061                 tmp = conf->mirrors + i;
1062                 if (tmp->rdev)
1063                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1064                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1065                                 !test_bit(Faulty, &tmp->rdev->flags),
1066                                 bdevname(tmp->rdev->bdev,b));
1067         }
1068 }
1069
1070 static void close_sync(conf_t *conf)
1071 {
1072         wait_barrier(conf);
1073         allow_barrier(conf);
1074
1075         mempool_destroy(conf->r10buf_pool);
1076         conf->r10buf_pool = NULL;
1077 }
1078
1079 /* check if there are enough drives for
1080  * every block to appear on atleast one
1081  */
1082 static int enough(conf_t *conf)
1083 {
1084         int first = 0;
1085
1086         do {
1087                 int n = conf->copies;
1088                 int cnt = 0;
1089                 while (n--) {
1090                         if (conf->mirrors[first].rdev)
1091                                 cnt++;
1092                         first = (first+1) % conf->raid_disks;
1093                 }
1094                 if (cnt == 0)
1095                         return 0;
1096         } while (first != 0);
1097         return 1;
1098 }
1099
1100 static int raid10_spare_active(mddev_t *mddev)
1101 {
1102         int i;
1103         conf_t *conf = mddev->private;
1104         mirror_info_t *tmp;
1105
1106         /*
1107          * Find all non-in_sync disks within the RAID10 configuration
1108          * and mark them in_sync
1109          */
1110         for (i = 0; i < conf->raid_disks; i++) {
1111                 tmp = conf->mirrors + i;
1112                 if (tmp->rdev
1113                     && !test_bit(Faulty, &tmp->rdev->flags)
1114                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1115                         unsigned long flags;
1116                         spin_lock_irqsave(&conf->device_lock, flags);
1117                         mddev->degraded--;
1118                         spin_unlock_irqrestore(&conf->device_lock, flags);
1119                 }
1120         }
1121
1122         print_conf(conf);
1123         return 0;
1124 }
1125
1126
1127 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1128 {
1129         conf_t *conf = mddev->private;
1130         int err = -EEXIST;
1131         int mirror;
1132         mirror_info_t *p;
1133         int first = 0;
1134         int last = mddev->raid_disks - 1;
1135
1136         if (mddev->recovery_cp < MaxSector)
1137                 /* only hot-add to in-sync arrays, as recovery is
1138                  * very different from resync
1139                  */
1140                 return -EBUSY;
1141         if (!enough(conf))
1142                 return -EINVAL;
1143
1144         if (rdev->raid_disk >= 0)
1145                 first = last = rdev->raid_disk;
1146
1147         if (rdev->saved_raid_disk >= 0 &&
1148             rdev->saved_raid_disk >= first &&
1149             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1150                 mirror = rdev->saved_raid_disk;
1151         else
1152                 mirror = first;
1153         for ( ; mirror <= last ; mirror++)
1154                 if ( !(p=conf->mirrors+mirror)->rdev) {
1155
1156                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1157                                           rdev->data_offset << 9);
1158                         /* as we don't honour merge_bvec_fn, we must
1159                          * never risk violating it, so limit
1160                          * ->max_phys_segments to one lying with a single
1161                          * page, as a one page request is never in
1162                          * violation.
1163                          */
1164                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1165                                 blk_queue_max_phys_segments(mddev->queue, 1);
1166                                 blk_queue_segment_boundary(mddev->queue,
1167                                                            PAGE_CACHE_SIZE - 1);
1168                         }
1169
1170                         p->head_position = 0;
1171                         rdev->raid_disk = mirror;
1172                         err = 0;
1173                         if (rdev->saved_raid_disk != mirror)
1174                                 conf->fullsync = 1;
1175                         rcu_assign_pointer(p->rdev, rdev);
1176                         break;
1177                 }
1178
1179         md_integrity_add_rdev(rdev, mddev);
1180         print_conf(conf);
1181         return err;
1182 }
1183
1184 static int raid10_remove_disk(mddev_t *mddev, int number)
1185 {
1186         conf_t *conf = mddev->private;
1187         int err = 0;
1188         mdk_rdev_t *rdev;
1189         mirror_info_t *p = conf->mirrors+ number;
1190
1191         print_conf(conf);
1192         rdev = p->rdev;
1193         if (rdev) {
1194                 if (test_bit(In_sync, &rdev->flags) ||
1195                     atomic_read(&rdev->nr_pending)) {
1196                         err = -EBUSY;
1197                         goto abort;
1198                 }
1199                 /* Only remove faulty devices in recovery
1200                  * is not possible.
1201                  */
1202                 if (!test_bit(Faulty, &rdev->flags) &&
1203                     enough(conf)) {
1204                         err = -EBUSY;
1205                         goto abort;
1206                 }
1207                 p->rdev = NULL;
1208                 synchronize_rcu();
1209                 if (atomic_read(&rdev->nr_pending)) {
1210                         /* lost the race, try later */
1211                         err = -EBUSY;
1212                         p->rdev = rdev;
1213                         goto abort;
1214                 }
1215                 md_integrity_register(mddev);
1216         }
1217 abort:
1218
1219         print_conf(conf);
1220         return err;
1221 }
1222
1223
1224 static void end_sync_read(struct bio *bio, int error)
1225 {
1226         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1227         conf_t *conf = r10_bio->mddev->private;
1228         int i,d;
1229
1230         for (i=0; i<conf->copies; i++)
1231                 if (r10_bio->devs[i].bio == bio)
1232                         break;
1233         BUG_ON(i == conf->copies);
1234         update_head_pos(i, r10_bio);
1235         d = r10_bio->devs[i].devnum;
1236
1237         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1238                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1239         else {
1240                 atomic_add(r10_bio->sectors,
1241                            &conf->mirrors[d].rdev->corrected_errors);
1242                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1243                         md_error(r10_bio->mddev,
1244                                  conf->mirrors[d].rdev);
1245         }
1246
1247         /* for reconstruct, we always reschedule after a read.
1248          * for resync, only after all reads
1249          */
1250         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1251         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1252             atomic_dec_and_test(&r10_bio->remaining)) {
1253                 /* we have read all the blocks,
1254                  * do the comparison in process context in raid10d
1255                  */
1256                 reschedule_retry(r10_bio);
1257         }
1258 }
1259
1260 static void end_sync_write(struct bio *bio, int error)
1261 {
1262         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1263         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1264         mddev_t *mddev = r10_bio->mddev;
1265         conf_t *conf = mddev->private;
1266         int i,d;
1267
1268         for (i = 0; i < conf->copies; i++)
1269                 if (r10_bio->devs[i].bio == bio)
1270                         break;
1271         d = r10_bio->devs[i].devnum;
1272
1273         if (!uptodate)
1274                 md_error(mddev, conf->mirrors[d].rdev);
1275
1276         update_head_pos(i, r10_bio);
1277
1278         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1279         while (atomic_dec_and_test(&r10_bio->remaining)) {
1280                 if (r10_bio->master_bio == NULL) {
1281                         /* the primary of several recovery bios */
1282                         sector_t s = r10_bio->sectors;
1283                         put_buf(r10_bio);
1284                         md_done_sync(mddev, s, 1);
1285                         break;
1286                 } else {
1287                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1288                         put_buf(r10_bio);
1289                         r10_bio = r10_bio2;
1290                 }
1291         }
1292 }
1293
1294 /*
1295  * Note: sync and recover and handled very differently for raid10
1296  * This code is for resync.
1297  * For resync, we read through virtual addresses and read all blocks.
1298  * If there is any error, we schedule a write.  The lowest numbered
1299  * drive is authoritative.
1300  * However requests come for physical address, so we need to map.
1301  * For every physical address there are raid_disks/copies virtual addresses,
1302  * which is always are least one, but is not necessarly an integer.
1303  * This means that a physical address can span multiple chunks, so we may
1304  * have to submit multiple io requests for a single sync request.
1305  */
1306 /*
1307  * We check if all blocks are in-sync and only write to blocks that
1308  * aren't in sync
1309  */
1310 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1311 {
1312         conf_t *conf = mddev->private;
1313         int i, first;
1314         struct bio *tbio, *fbio;
1315
1316         atomic_set(&r10_bio->remaining, 1);
1317
1318         /* find the first device with a block */
1319         for (i=0; i<conf->copies; i++)
1320                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1321                         break;
1322
1323         if (i == conf->copies)
1324                 goto done;
1325
1326         first = i;
1327         fbio = r10_bio->devs[i].bio;
1328
1329         /* now find blocks with errors */
1330         for (i=0 ; i < conf->copies ; i++) {
1331                 int  j, d;
1332                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1333
1334                 tbio = r10_bio->devs[i].bio;
1335
1336                 if (tbio->bi_end_io != end_sync_read)
1337                         continue;
1338                 if (i == first)
1339                         continue;
1340                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1341                         /* We know that the bi_io_vec layout is the same for
1342                          * both 'first' and 'i', so we just compare them.
1343                          * All vec entries are PAGE_SIZE;
1344                          */
1345                         for (j = 0; j < vcnt; j++)
1346                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1347                                            page_address(tbio->bi_io_vec[j].bv_page),
1348                                            PAGE_SIZE))
1349                                         break;
1350                         if (j == vcnt)
1351                                 continue;
1352                         mddev->resync_mismatches += r10_bio->sectors;
1353                 }
1354                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1355                         /* Don't fix anything. */
1356                         continue;
1357                 /* Ok, we need to write this bio
1358                  * First we need to fixup bv_offset, bv_len and
1359                  * bi_vecs, as the read request might have corrupted these
1360                  */
1361                 tbio->bi_vcnt = vcnt;
1362                 tbio->bi_size = r10_bio->sectors << 9;
1363                 tbio->bi_idx = 0;
1364                 tbio->bi_phys_segments = 0;
1365                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1366                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1367                 tbio->bi_next = NULL;
1368                 tbio->bi_rw = WRITE;
1369                 tbio->bi_private = r10_bio;
1370                 tbio->bi_sector = r10_bio->devs[i].addr;
1371
1372                 for (j=0; j < vcnt ; j++) {
1373                         tbio->bi_io_vec[j].bv_offset = 0;
1374                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1375
1376                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1377                                page_address(fbio->bi_io_vec[j].bv_page),
1378                                PAGE_SIZE);
1379                 }
1380                 tbio->bi_end_io = end_sync_write;
1381
1382                 d = r10_bio->devs[i].devnum;
1383                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1384                 atomic_inc(&r10_bio->remaining);
1385                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1386
1387                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1388                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1389                 generic_make_request(tbio);
1390         }
1391
1392 done:
1393         if (atomic_dec_and_test(&r10_bio->remaining)) {
1394                 md_done_sync(mddev, r10_bio->sectors, 1);
1395                 put_buf(r10_bio);
1396         }
1397 }
1398
1399 /*
1400  * Now for the recovery code.
1401  * Recovery happens across physical sectors.
1402  * We recover all non-is_sync drives by finding the virtual address of
1403  * each, and then choose a working drive that also has that virt address.
1404  * There is a separate r10_bio for each non-in_sync drive.
1405  * Only the first two slots are in use. The first for reading,
1406  * The second for writing.
1407  *
1408  */
1409
1410 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1411 {
1412         conf_t *conf = mddev->private;
1413         int i, d;
1414         struct bio *bio, *wbio;
1415
1416
1417         /* move the pages across to the second bio
1418          * and submit the write request
1419          */
1420         bio = r10_bio->devs[0].bio;
1421         wbio = r10_bio->devs[1].bio;
1422         for (i=0; i < wbio->bi_vcnt; i++) {
1423                 struct page *p = bio->bi_io_vec[i].bv_page;
1424                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1425                 wbio->bi_io_vec[i].bv_page = p;
1426         }
1427         d = r10_bio->devs[1].devnum;
1428
1429         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1430         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1431         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1432                 generic_make_request(wbio);
1433         else
1434                 bio_endio(wbio, -EIO);
1435 }
1436
1437
1438 /*
1439  * This is a kernel thread which:
1440  *
1441  *      1.      Retries failed read operations on working mirrors.
1442  *      2.      Updates the raid superblock when problems encounter.
1443  *      3.      Performs writes following reads for array synchronising.
1444  */
1445
1446 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1447 {
1448         int sect = 0; /* Offset from r10_bio->sector */
1449         int sectors = r10_bio->sectors;
1450         mdk_rdev_t*rdev;
1451         while(sectors) {
1452                 int s = sectors;
1453                 int sl = r10_bio->read_slot;
1454                 int success = 0;
1455                 int start;
1456
1457                 if (s > (PAGE_SIZE>>9))
1458                         s = PAGE_SIZE >> 9;
1459
1460                 rcu_read_lock();
1461                 do {
1462                         int d = r10_bio->devs[sl].devnum;
1463                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1464                         if (rdev &&
1465                             test_bit(In_sync, &rdev->flags)) {
1466                                 atomic_inc(&rdev->nr_pending);
1467                                 rcu_read_unlock();
1468                                 success = sync_page_io(rdev->bdev,
1469                                                        r10_bio->devs[sl].addr +
1470                                                        sect + rdev->data_offset,
1471                                                        s<<9,
1472                                                        conf->tmppage, READ);
1473                                 rdev_dec_pending(rdev, mddev);
1474                                 rcu_read_lock();
1475                                 if (success)
1476                                         break;
1477                         }
1478                         sl++;
1479                         if (sl == conf->copies)
1480                                 sl = 0;
1481                 } while (!success && sl != r10_bio->read_slot);
1482                 rcu_read_unlock();
1483
1484                 if (!success) {
1485                         /* Cannot read from anywhere -- bye bye array */
1486                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1487                         md_error(mddev, conf->mirrors[dn].rdev);
1488                         break;
1489                 }
1490
1491                 start = sl;
1492                 /* write it back and re-read */
1493                 rcu_read_lock();
1494                 while (sl != r10_bio->read_slot) {
1495                         int d;
1496                         if (sl==0)
1497                                 sl = conf->copies;
1498                         sl--;
1499                         d = r10_bio->devs[sl].devnum;
1500                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1501                         if (rdev &&
1502                             test_bit(In_sync, &rdev->flags)) {
1503                                 atomic_inc(&rdev->nr_pending);
1504                                 rcu_read_unlock();
1505                                 atomic_add(s, &rdev->corrected_errors);
1506                                 if (sync_page_io(rdev->bdev,
1507                                                  r10_bio->devs[sl].addr +
1508                                                  sect + rdev->data_offset,
1509                                                  s<<9, conf->tmppage, WRITE)
1510                                     == 0)
1511                                         /* Well, this device is dead */
1512                                         md_error(mddev, rdev);
1513                                 rdev_dec_pending(rdev, mddev);
1514                                 rcu_read_lock();
1515                         }
1516                 }
1517                 sl = start;
1518                 while (sl != r10_bio->read_slot) {
1519                         int d;
1520                         if (sl==0)
1521                                 sl = conf->copies;
1522                         sl--;
1523                         d = r10_bio->devs[sl].devnum;
1524                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1525                         if (rdev &&
1526                             test_bit(In_sync, &rdev->flags)) {
1527                                 char b[BDEVNAME_SIZE];
1528                                 atomic_inc(&rdev->nr_pending);
1529                                 rcu_read_unlock();
1530                                 if (sync_page_io(rdev->bdev,
1531                                                  r10_bio->devs[sl].addr +
1532                                                  sect + rdev->data_offset,
1533                                                  s<<9, conf->tmppage, READ) == 0)
1534                                         /* Well, this device is dead */
1535                                         md_error(mddev, rdev);
1536                                 else
1537                                         printk(KERN_INFO
1538                                                "raid10:%s: read error corrected"
1539                                                " (%d sectors at %llu on %s)\n",
1540                                                mdname(mddev), s,
1541                                                (unsigned long long)(sect+
1542                                                     rdev->data_offset),
1543                                                bdevname(rdev->bdev, b));
1544
1545                                 rdev_dec_pending(rdev, mddev);
1546                                 rcu_read_lock();
1547                         }
1548                 }
1549                 rcu_read_unlock();
1550
1551                 sectors -= s;
1552                 sect += s;
1553         }
1554 }
1555
1556 static void raid10d(mddev_t *mddev)
1557 {
1558         r10bio_t *r10_bio;
1559         struct bio *bio;
1560         unsigned long flags;
1561         conf_t *conf = mddev->private;
1562         struct list_head *head = &conf->retry_list;
1563         int unplug=0;
1564         mdk_rdev_t *rdev;
1565
1566         md_check_recovery(mddev);
1567
1568         for (;;) {
1569                 char b[BDEVNAME_SIZE];
1570
1571                 unplug += flush_pending_writes(conf);
1572
1573                 spin_lock_irqsave(&conf->device_lock, flags);
1574                 if (list_empty(head)) {
1575                         spin_unlock_irqrestore(&conf->device_lock, flags);
1576                         break;
1577                 }
1578                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1579                 list_del(head->prev);
1580                 conf->nr_queued--;
1581                 spin_unlock_irqrestore(&conf->device_lock, flags);
1582
1583                 mddev = r10_bio->mddev;
1584                 conf = mddev->private;
1585                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1586                         sync_request_write(mddev, r10_bio);
1587                         unplug = 1;
1588                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1589                         recovery_request_write(mddev, r10_bio);
1590                         unplug = 1;
1591                 } else {
1592                         int mirror;
1593                         /* we got a read error. Maybe the drive is bad.  Maybe just
1594                          * the block and we can fix it.
1595                          * We freeze all other IO, and try reading the block from
1596                          * other devices.  When we find one, we re-write
1597                          * and check it that fixes the read error.
1598                          * This is all done synchronously while the array is
1599                          * frozen.
1600                          */
1601                         if (mddev->ro == 0) {
1602                                 freeze_array(conf);
1603                                 fix_read_error(conf, mddev, r10_bio);
1604                                 unfreeze_array(conf);
1605                         }
1606
1607                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1608                         r10_bio->devs[r10_bio->read_slot].bio =
1609                                 mddev->ro ? IO_BLOCKED : NULL;
1610                         mirror = read_balance(conf, r10_bio);
1611                         if (mirror == -1) {
1612                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1613                                        " read error for block %llu\n",
1614                                        bdevname(bio->bi_bdev,b),
1615                                        (unsigned long long)r10_bio->sector);
1616                                 raid_end_bio_io(r10_bio);
1617                                 bio_put(bio);
1618                         } else {
1619                                 const bool do_sync = bio_rw_flagged(r10_bio->master_bio, BIO_RW_SYNCIO);
1620                                 bio_put(bio);
1621                                 rdev = conf->mirrors[mirror].rdev;
1622                                 if (printk_ratelimit())
1623                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1624                                                " another mirror\n",
1625                                                bdevname(rdev->bdev,b),
1626                                                (unsigned long long)r10_bio->sector);
1627                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1628                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1629                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1630                                         + rdev->data_offset;
1631                                 bio->bi_bdev = rdev->bdev;
1632                                 bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
1633                                 bio->bi_private = r10_bio;
1634                                 bio->bi_end_io = raid10_end_read_request;
1635                                 unplug = 1;
1636                                 generic_make_request(bio);
1637                         }
1638                 }
1639                 cond_resched();
1640         }
1641         if (unplug)
1642                 unplug_slaves(mddev);
1643 }
1644
1645
1646 static int init_resync(conf_t *conf)
1647 {
1648         int buffs;
1649
1650         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1651         BUG_ON(conf->r10buf_pool);
1652         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1653         if (!conf->r10buf_pool)
1654                 return -ENOMEM;
1655         conf->next_resync = 0;
1656         return 0;
1657 }
1658
1659 /*
1660  * perform a "sync" on one "block"
1661  *
1662  * We need to make sure that no normal I/O request - particularly write
1663  * requests - conflict with active sync requests.
1664  *
1665  * This is achieved by tracking pending requests and a 'barrier' concept
1666  * that can be installed to exclude normal IO requests.
1667  *
1668  * Resync and recovery are handled very differently.
1669  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1670  *
1671  * For resync, we iterate over virtual addresses, read all copies,
1672  * and update if there are differences.  If only one copy is live,
1673  * skip it.
1674  * For recovery, we iterate over physical addresses, read a good
1675  * value for each non-in_sync drive, and over-write.
1676  *
1677  * So, for recovery we may have several outstanding complex requests for a
1678  * given address, one for each out-of-sync device.  We model this by allocating
1679  * a number of r10_bio structures, one for each out-of-sync device.
1680  * As we setup these structures, we collect all bio's together into a list
1681  * which we then process collectively to add pages, and then process again
1682  * to pass to generic_make_request.
1683  *
1684  * The r10_bio structures are linked using a borrowed master_bio pointer.
1685  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1686  * has its remaining count decremented to 0, the whole complex operation
1687  * is complete.
1688  *
1689  */
1690
1691 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1692 {
1693         conf_t *conf = mddev->private;
1694         r10bio_t *r10_bio;
1695         struct bio *biolist = NULL, *bio;
1696         sector_t max_sector, nr_sectors;
1697         int disk;
1698         int i;
1699         int max_sync;
1700         int sync_blocks;
1701
1702         sector_t sectors_skipped = 0;
1703         int chunks_skipped = 0;
1704
1705         if (!conf->r10buf_pool)
1706                 if (init_resync(conf))
1707                         return 0;
1708
1709  skipped:
1710         max_sector = mddev->dev_sectors;
1711         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1712                 max_sector = mddev->resync_max_sectors;
1713         if (sector_nr >= max_sector) {
1714                 /* If we aborted, we need to abort the
1715                  * sync on the 'current' bitmap chucks (there can
1716                  * be several when recovering multiple devices).
1717                  * as we may have started syncing it but not finished.
1718                  * We can find the current address in
1719                  * mddev->curr_resync, but for recovery,
1720                  * we need to convert that to several
1721                  * virtual addresses.
1722                  */
1723                 if (mddev->curr_resync < max_sector) { /* aborted */
1724                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1725                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1726                                                 &sync_blocks, 1);
1727                         else for (i=0; i<conf->raid_disks; i++) {
1728                                 sector_t sect =
1729                                         raid10_find_virt(conf, mddev->curr_resync, i);
1730                                 bitmap_end_sync(mddev->bitmap, sect,
1731                                                 &sync_blocks, 1);
1732                         }
1733                 } else /* completed sync */
1734                         conf->fullsync = 0;
1735
1736                 bitmap_close_sync(mddev->bitmap);
1737                 close_sync(conf);
1738                 *skipped = 1;
1739                 return sectors_skipped;
1740         }
1741         if (chunks_skipped >= conf->raid_disks) {
1742                 /* if there has been nothing to do on any drive,
1743                  * then there is nothing to do at all..
1744                  */
1745                 *skipped = 1;
1746                 return (max_sector - sector_nr) + sectors_skipped;
1747         }
1748
1749         if (max_sector > mddev->resync_max)
1750                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1751
1752         /* make sure whole request will fit in a chunk - if chunks
1753          * are meaningful
1754          */
1755         if (conf->near_copies < conf->raid_disks &&
1756             max_sector > (sector_nr | conf->chunk_mask))
1757                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1758         /*
1759          * If there is non-resync activity waiting for us then
1760          * put in a delay to throttle resync.
1761          */
1762         if (!go_faster && conf->nr_waiting)
1763                 msleep_interruptible(1000);
1764
1765         /* Again, very different code for resync and recovery.
1766          * Both must result in an r10bio with a list of bios that
1767          * have bi_end_io, bi_sector, bi_bdev set,
1768          * and bi_private set to the r10bio.
1769          * For recovery, we may actually create several r10bios
1770          * with 2 bios in each, that correspond to the bios in the main one.
1771          * In this case, the subordinate r10bios link back through a
1772          * borrowed master_bio pointer, and the counter in the master
1773          * includes a ref from each subordinate.
1774          */
1775         /* First, we decide what to do and set ->bi_end_io
1776          * To end_sync_read if we want to read, and
1777          * end_sync_write if we will want to write.
1778          */
1779
1780         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1781         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1782                 /* recovery... the complicated one */
1783                 int j, k;
1784                 r10_bio = NULL;
1785
1786                 for (i=0 ; i<conf->raid_disks; i++)
1787                         if (conf->mirrors[i].rdev &&
1788                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1789                                 int still_degraded = 0;
1790                                 /* want to reconstruct this device */
1791                                 r10bio_t *rb2 = r10_bio;
1792                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1793                                 int must_sync;
1794                                 /* Unless we are doing a full sync, we only need
1795                                  * to recover the block if it is set in the bitmap
1796                                  */
1797                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1798                                                               &sync_blocks, 1);
1799                                 if (sync_blocks < max_sync)
1800                                         max_sync = sync_blocks;
1801                                 if (!must_sync &&
1802                                     !conf->fullsync) {
1803                                         /* yep, skip the sync_blocks here, but don't assume
1804                                          * that there will never be anything to do here
1805                                          */
1806                                         chunks_skipped = -1;
1807                                         continue;
1808                                 }
1809
1810                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1811                                 raise_barrier(conf, rb2 != NULL);
1812                                 atomic_set(&r10_bio->remaining, 0);
1813
1814                                 r10_bio->master_bio = (struct bio*)rb2;
1815                                 if (rb2)
1816                                         atomic_inc(&rb2->remaining);
1817                                 r10_bio->mddev = mddev;
1818                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1819                                 r10_bio->sector = sect;
1820
1821                                 raid10_find_phys(conf, r10_bio);
1822
1823                                 /* Need to check if the array will still be
1824                                  * degraded
1825                                  */
1826                                 for (j=0; j<conf->raid_disks; j++)
1827                                         if (conf->mirrors[j].rdev == NULL ||
1828                                             test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
1829                                                 still_degraded = 1;
1830                                                 break;
1831                                         }
1832
1833                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1834                                                               &sync_blocks, still_degraded);
1835
1836                                 for (j=0; j<conf->copies;j++) {
1837                                         int d = r10_bio->devs[j].devnum;
1838                                         if (conf->mirrors[d].rdev &&
1839                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1840                                                 /* This is where we read from */
1841                                                 bio = r10_bio->devs[0].bio;
1842                                                 bio->bi_next = biolist;
1843                                                 biolist = bio;
1844                                                 bio->bi_private = r10_bio;
1845                                                 bio->bi_end_io = end_sync_read;
1846                                                 bio->bi_rw = READ;
1847                                                 bio->bi_sector = r10_bio->devs[j].addr +
1848                                                         conf->mirrors[d].rdev->data_offset;
1849                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1850                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1851                                                 atomic_inc(&r10_bio->remaining);
1852                                                 /* and we write to 'i' */
1853
1854                                                 for (k=0; k<conf->copies; k++)
1855                                                         if (r10_bio->devs[k].devnum == i)
1856                                                                 break;
1857                                                 BUG_ON(k == conf->copies);
1858                                                 bio = r10_bio->devs[1].bio;
1859                                                 bio->bi_next = biolist;
1860                                                 biolist = bio;
1861                                                 bio->bi_private = r10_bio;
1862                                                 bio->bi_end_io = end_sync_write;
1863                                                 bio->bi_rw = WRITE;
1864                                                 bio->bi_sector = r10_bio->devs[k].addr +
1865                                                         conf->mirrors[i].rdev->data_offset;
1866                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1867
1868                                                 r10_bio->devs[0].devnum = d;
1869                                                 r10_bio->devs[1].devnum = i;
1870
1871                                                 break;
1872                                         }
1873                                 }
1874                                 if (j == conf->copies) {
1875                                         /* Cannot recover, so abort the recovery */
1876                                         put_buf(r10_bio);
1877                                         if (rb2)
1878                                                 atomic_dec(&rb2->remaining);
1879                                         r10_bio = rb2;
1880                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
1881                                                               &mddev->recovery))
1882                                                 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1883                                                        mdname(mddev));
1884                                         break;
1885                                 }
1886                         }
1887                 if (biolist == NULL) {
1888                         while (r10_bio) {
1889                                 r10bio_t *rb2 = r10_bio;
1890                                 r10_bio = (r10bio_t*) rb2->master_bio;
1891                                 rb2->master_bio = NULL;
1892                                 put_buf(rb2);
1893                         }
1894                         goto giveup;
1895                 }
1896         } else {
1897                 /* resync. Schedule a read for every block at this virt offset */
1898                 int count = 0;
1899
1900                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1901
1902                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1903                                        &sync_blocks, mddev->degraded) &&
1904                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1905                         /* We can skip this block */
1906                         *skipped = 1;
1907                         return sync_blocks + sectors_skipped;
1908                 }
1909                 if (sync_blocks < max_sync)
1910                         max_sync = sync_blocks;
1911                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1912
1913                 r10_bio->mddev = mddev;
1914                 atomic_set(&r10_bio->remaining, 0);
1915                 raise_barrier(conf, 0);
1916                 conf->next_resync = sector_nr;
1917
1918                 r10_bio->master_bio = NULL;
1919                 r10_bio->sector = sector_nr;
1920                 set_bit(R10BIO_IsSync, &r10_bio->state);
1921                 raid10_find_phys(conf, r10_bio);
1922                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1923
1924                 for (i=0; i<conf->copies; i++) {
1925                         int d = r10_bio->devs[i].devnum;
1926                         bio = r10_bio->devs[i].bio;
1927                         bio->bi_end_io = NULL;
1928                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
1929                         if (conf->mirrors[d].rdev == NULL ||
1930                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1931                                 continue;
1932                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1933                         atomic_inc(&r10_bio->remaining);
1934                         bio->bi_next = biolist;
1935                         biolist = bio;
1936                         bio->bi_private = r10_bio;
1937                         bio->bi_end_io = end_sync_read;
1938                         bio->bi_rw = READ;
1939                         bio->bi_sector = r10_bio->devs[i].addr +
1940                                 conf->mirrors[d].rdev->data_offset;
1941                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1942                         count++;
1943                 }
1944
1945                 if (count < 2) {
1946                         for (i=0; i<conf->copies; i++) {
1947                                 int d = r10_bio->devs[i].devnum;
1948                                 if (r10_bio->devs[i].bio->bi_end_io)
1949                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1950                         }
1951                         put_buf(r10_bio);
1952                         biolist = NULL;
1953                         goto giveup;
1954                 }
1955         }
1956
1957         for (bio = biolist; bio ; bio=bio->bi_next) {
1958
1959                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1960                 if (bio->bi_end_io)
1961                         bio->bi_flags |= 1 << BIO_UPTODATE;
1962                 bio->bi_vcnt = 0;
1963                 bio->bi_idx = 0;
1964                 bio->bi_phys_segments = 0;
1965                 bio->bi_size = 0;
1966         }
1967
1968         nr_sectors = 0;
1969         if (sector_nr + max_sync < max_sector)
1970                 max_sector = sector_nr + max_sync;
1971         do {
1972                 struct page *page;
1973                 int len = PAGE_SIZE;
1974                 disk = 0;
1975                 if (sector_nr + (len>>9) > max_sector)
1976                         len = (max_sector - sector_nr) << 9;
1977                 if (len == 0)
1978                         break;
1979                 for (bio= biolist ; bio ; bio=bio->bi_next) {
1980                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1981                         if (bio_add_page(bio, page, len, 0) == 0) {
1982                                 /* stop here */
1983                                 struct bio *bio2;
1984                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1985                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1986                                         /* remove last page from this bio */
1987                                         bio2->bi_vcnt--;
1988                                         bio2->bi_size -= len;
1989                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1990                                 }
1991                                 goto bio_full;
1992                         }
1993                         disk = i;
1994                 }
1995                 nr_sectors += len>>9;
1996                 sector_nr += len>>9;
1997         } while (biolist->bi_vcnt < RESYNC_PAGES);
1998  bio_full:
1999         r10_bio->sectors = nr_sectors;
2000
2001         while (biolist) {
2002                 bio = biolist;
2003                 biolist = biolist->bi_next;
2004
2005                 bio->bi_next = NULL;
2006                 r10_bio = bio->bi_private;
2007                 r10_bio->sectors = nr_sectors;
2008
2009                 if (bio->bi_end_io == end_sync_read) {
2010                         md_sync_acct(bio->bi_bdev, nr_sectors);
2011                         generic_make_request(bio);
2012                 }
2013         }
2014
2015         if (sectors_skipped)
2016                 /* pretend they weren't skipped, it makes
2017                  * no important difference in this case
2018                  */
2019                 md_done_sync(mddev, sectors_skipped, 1);
2020
2021         return sectors_skipped + nr_sectors;
2022  giveup:
2023         /* There is nowhere to write, so all non-sync
2024          * drives must be failed, so try the next chunk...
2025          */
2026         if (sector_nr + max_sync < max_sector)
2027                 max_sector = sector_nr + max_sync;
2028
2029         sectors_skipped += (max_sector - sector_nr);
2030         chunks_skipped ++;
2031         sector_nr = max_sector;
2032         goto skipped;
2033 }
2034
2035 static sector_t
2036 raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2037 {
2038         sector_t size;
2039         conf_t *conf = mddev->private;
2040
2041         if (!raid_disks)
2042                 raid_disks = mddev->raid_disks;
2043         if (!sectors)
2044                 sectors = mddev->dev_sectors;
2045
2046         size = sectors >> conf->chunk_shift;
2047         sector_div(size, conf->far_copies);
2048         size = size * raid_disks;
2049         sector_div(size, conf->near_copies);
2050
2051         return size << conf->chunk_shift;
2052 }
2053
2054 static int run(mddev_t *mddev)
2055 {
2056         conf_t *conf;
2057         int i, disk_idx, chunk_size;
2058         mirror_info_t *disk;
2059         mdk_rdev_t *rdev;
2060         int nc, fc, fo;
2061         sector_t stride, size;
2062
2063         if (mddev->chunk_sectors < (PAGE_SIZE >> 9) ||
2064             !is_power_of_2(mddev->chunk_sectors)) {
2065                 printk(KERN_ERR "md/raid10: chunk size must be "
2066                        "at least PAGE_SIZE(%ld) and be a power of 2.\n", PAGE_SIZE);
2067                 return -EINVAL;
2068         }
2069
2070         nc = mddev->layout & 255;
2071         fc = (mddev->layout >> 8) & 255;
2072         fo = mddev->layout & (1<<16);
2073         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2074             (mddev->layout >> 17)) {
2075                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
2076                        mdname(mddev), mddev->layout);
2077                 goto out;
2078         }
2079         /*
2080          * copy the already verified devices into our private RAID10
2081          * bookkeeping area. [whatever we allocate in run(),
2082          * should be freed in stop()]
2083          */
2084         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2085         mddev->private = conf;
2086         if (!conf) {
2087                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2088                         mdname(mddev));
2089                 goto out;
2090         }
2091         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2092                                  GFP_KERNEL);
2093         if (!conf->mirrors) {
2094                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2095                        mdname(mddev));
2096                 goto out_free_conf;
2097         }
2098
2099         conf->tmppage = alloc_page(GFP_KERNEL);
2100         if (!conf->tmppage)
2101                 goto out_free_conf;
2102
2103         conf->raid_disks = mddev->raid_disks;
2104         conf->near_copies = nc;
2105         conf->far_copies = fc;
2106         conf->copies = nc*fc;
2107         conf->far_offset = fo;
2108         conf->chunk_mask = mddev->chunk_sectors - 1;
2109         conf->chunk_shift = ffz(~mddev->chunk_sectors);
2110         size = mddev->dev_sectors >> conf->chunk_shift;
2111         sector_div(size, fc);
2112         size = size * conf->raid_disks;
2113         sector_div(size, nc);
2114         /* 'size' is now the number of chunks in the array */
2115         /* calculate "used chunks per device" in 'stride' */
2116         stride = size * conf->copies;
2117
2118         /* We need to round up when dividing by raid_disks to
2119          * get the stride size.
2120          */
2121         stride += conf->raid_disks - 1;
2122         sector_div(stride, conf->raid_disks);
2123         mddev->dev_sectors = stride << conf->chunk_shift;
2124
2125         if (fo)
2126                 stride = 1;
2127         else
2128                 sector_div(stride, fc);
2129         conf->stride = stride << conf->chunk_shift;
2130
2131         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2132                                                 r10bio_pool_free, conf);
2133         if (!conf->r10bio_pool) {
2134                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2135                         mdname(mddev));
2136                 goto out_free_conf;
2137         }
2138
2139         conf->mddev = mddev;
2140         spin_lock_init(&conf->device_lock);
2141         mddev->queue->queue_lock = &conf->device_lock;
2142
2143         chunk_size = mddev->chunk_sectors << 9;
2144         blk_queue_io_min(mddev->queue, chunk_size);
2145         if (conf->raid_disks % conf->near_copies)
2146                 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2147         else
2148                 blk_queue_io_opt(mddev->queue, chunk_size *
2149                                  (conf->raid_disks / conf->near_copies));
2150
2151         list_for_each_entry(rdev, &mddev->disks, same_set) {
2152                 disk_idx = rdev->raid_disk;
2153                 if (disk_idx >= mddev->raid_disks
2154                     || disk_idx < 0)
2155                         continue;
2156                 disk = conf->mirrors + disk_idx;
2157
2158                 disk->rdev = rdev;
2159                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2160                                   rdev->data_offset << 9);
2161                 /* as we don't honour merge_bvec_fn, we must never risk
2162                  * violating it, so limit max_phys_segments to 1 lying
2163                  * within a single page.
2164                  */
2165                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2166                         blk_queue_max_phys_segments(mddev->queue, 1);
2167                         blk_queue_segment_boundary(mddev->queue,
2168                                                    PAGE_CACHE_SIZE - 1);
2169                 }
2170
2171                 disk->head_position = 0;
2172         }
2173         INIT_LIST_HEAD(&conf->retry_list);
2174
2175         spin_lock_init(&conf->resync_lock);
2176         init_waitqueue_head(&conf->wait_barrier);
2177
2178         /* need to check that every block has at least one working mirror */
2179         if (!enough(conf)) {
2180                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2181                        mdname(mddev));
2182                 goto out_free_conf;
2183         }
2184
2185         mddev->degraded = 0;
2186         for (i = 0; i < conf->raid_disks; i++) {
2187
2188                 disk = conf->mirrors + i;
2189
2190                 if (!disk->rdev ||
2191                     !test_bit(In_sync, &disk->rdev->flags)) {
2192                         disk->head_position = 0;
2193                         mddev->degraded++;
2194                         if (disk->rdev)
2195                                 conf->fullsync = 1;
2196                 }
2197         }
2198
2199
2200         mddev->thread = md_register_thread(raid10d, mddev, NULL);
2201         if (!mddev->thread) {
2202                 printk(KERN_ERR
2203                        "raid10: couldn't allocate thread for %s\n",
2204                        mdname(mddev));
2205                 goto out_free_conf;
2206         }
2207
2208         if (mddev->recovery_cp != MaxSector)
2209                 printk(KERN_NOTICE "raid10: %s is not clean"
2210                        " -- starting background reconstruction\n",
2211                        mdname(mddev));
2212         printk(KERN_INFO
2213                 "raid10: raid set %s active with %d out of %d devices\n",
2214                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2215                 mddev->raid_disks);
2216         /*
2217          * Ok, everything is just fine now
2218          */
2219         md_set_array_sectors(mddev, raid10_size(mddev, 0, 0));
2220         mddev->resync_max_sectors = raid10_size(mddev, 0, 0);
2221
2222         mddev->queue->unplug_fn = raid10_unplug;
2223         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2224         mddev->queue->backing_dev_info.congested_data = mddev;
2225
2226         /* Calculate max read-ahead size.
2227          * We need to readahead at least twice a whole stripe....
2228          * maybe...
2229          */
2230         {
2231                 int stripe = conf->raid_disks *
2232                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2233                 stripe /= conf->near_copies;
2234                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2235                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2236         }
2237
2238         if (conf->near_copies < mddev->raid_disks)
2239                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2240         md_integrity_register(mddev);
2241         return 0;
2242
2243 out_free_conf:
2244         if (conf->r10bio_pool)
2245                 mempool_destroy(conf->r10bio_pool);
2246         safe_put_page(conf->tmppage);
2247         kfree(conf->mirrors);
2248         kfree(conf);
2249         mddev->private = NULL;
2250 out:
2251         return -EIO;
2252 }
2253
2254 static int stop(mddev_t *mddev)
2255 {
2256         conf_t *conf = mddev->private;
2257
2258         raise_barrier(conf, 0);
2259         lower_barrier(conf);
2260
2261         md_unregister_thread(mddev->thread);
2262         mddev->thread = NULL;
2263         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2264         if (conf->r10bio_pool)
2265                 mempool_destroy(conf->r10bio_pool);
2266         kfree(conf->mirrors);
2267         kfree(conf);
2268         mddev->private = NULL;
2269         return 0;
2270 }
2271
2272 static void raid10_quiesce(mddev_t *mddev, int state)
2273 {
2274         conf_t *conf = mddev->private;
2275
2276         switch(state) {
2277         case 1:
2278                 raise_barrier(conf, 0);
2279                 break;
2280         case 0:
2281                 lower_barrier(conf);
2282                 break;
2283         }
2284         if (mddev->thread) {
2285                 if (mddev->bitmap)
2286                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2287                 else
2288                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2289                 md_wakeup_thread(mddev->thread);
2290         }
2291 }
2292
2293 static struct mdk_personality raid10_personality =
2294 {
2295         .name           = "raid10",
2296         .level          = 10,
2297         .owner          = THIS_MODULE,
2298         .make_request   = make_request,
2299         .run            = run,
2300         .stop           = stop,
2301         .status         = status,
2302         .error_handler  = error,
2303         .hot_add_disk   = raid10_add_disk,
2304         .hot_remove_disk= raid10_remove_disk,
2305         .spare_active   = raid10_spare_active,
2306         .sync_request   = sync_request,
2307         .quiesce        = raid10_quiesce,
2308         .size           = raid10_size,
2309 };
2310
2311 static int __init raid_init(void)
2312 {
2313         return register_md_personality(&raid10_personality);
2314 }
2315
2316 static void raid_exit(void)
2317 {
2318         unregister_md_personality(&raid10_personality);
2319 }
2320
2321 module_init(raid_init);
2322 module_exit(raid_exit);
2323 MODULE_LICENSE("GPL");
2324 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2325 MODULE_ALIAS("md-raid10");
2326 MODULE_ALIAS("md-level-10");