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)
  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 never risk
1159                          * violating it, so limit ->max_sector to one PAGE, as
1160                          * a one page request is never in violation.
1161                          */
1162                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1163                             queue_max_sectors(mddev->queue) > (PAGE_SIZE>>9))
1164                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1165
1166                         p->head_position = 0;
1167                         rdev->raid_disk = mirror;
1168                         err = 0;
1169                         if (rdev->saved_raid_disk != mirror)
1170                                 conf->fullsync = 1;
1171                         rcu_assign_pointer(p->rdev, rdev);
1172                         break;
1173                 }
1174
1175         md_integrity_add_rdev(rdev, mddev);
1176         print_conf(conf);
1177         return err;
1178 }
1179
1180 static int raid10_remove_disk(mddev_t *mddev, int number)
1181 {
1182         conf_t *conf = mddev->private;
1183         int err = 0;
1184         mdk_rdev_t *rdev;
1185         mirror_info_t *p = conf->mirrors+ number;
1186
1187         print_conf(conf);
1188         rdev = p->rdev;
1189         if (rdev) {
1190                 if (test_bit(In_sync, &rdev->flags) ||
1191                     atomic_read(&rdev->nr_pending)) {
1192                         err = -EBUSY;
1193                         goto abort;
1194                 }
1195                 /* Only remove faulty devices in recovery
1196                  * is not possible.
1197                  */
1198                 if (!test_bit(Faulty, &rdev->flags) &&
1199                     enough(conf)) {
1200                         err = -EBUSY;
1201                         goto abort;
1202                 }
1203                 p->rdev = NULL;
1204                 synchronize_rcu();
1205                 if (atomic_read(&rdev->nr_pending)) {
1206                         /* lost the race, try later */
1207                         err = -EBUSY;
1208                         p->rdev = rdev;
1209                         goto abort;
1210                 }
1211                 md_integrity_register(mddev);
1212         }
1213 abort:
1214
1215         print_conf(conf);
1216         return err;
1217 }
1218
1219
1220 static void end_sync_read(struct bio *bio, int error)
1221 {
1222         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1223         conf_t *conf = r10_bio->mddev->private;
1224         int i,d;
1225
1226         for (i=0; i<conf->copies; i++)
1227                 if (r10_bio->devs[i].bio == bio)
1228                         break;
1229         BUG_ON(i == conf->copies);
1230         update_head_pos(i, r10_bio);
1231         d = r10_bio->devs[i].devnum;
1232
1233         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1234                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1235         else {
1236                 atomic_add(r10_bio->sectors,
1237                            &conf->mirrors[d].rdev->corrected_errors);
1238                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1239                         md_error(r10_bio->mddev,
1240                                  conf->mirrors[d].rdev);
1241         }
1242
1243         /* for reconstruct, we always reschedule after a read.
1244          * for resync, only after all reads
1245          */
1246         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1247         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1248             atomic_dec_and_test(&r10_bio->remaining)) {
1249                 /* we have read all the blocks,
1250                  * do the comparison in process context in raid10d
1251                  */
1252                 reschedule_retry(r10_bio);
1253         }
1254 }
1255
1256 static void end_sync_write(struct bio *bio, int error)
1257 {
1258         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1259         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1260         mddev_t *mddev = r10_bio->mddev;
1261         conf_t *conf = mddev->private;
1262         int i,d;
1263
1264         for (i = 0; i < conf->copies; i++)
1265                 if (r10_bio->devs[i].bio == bio)
1266                         break;
1267         d = r10_bio->devs[i].devnum;
1268
1269         if (!uptodate)
1270                 md_error(mddev, conf->mirrors[d].rdev);
1271
1272         update_head_pos(i, r10_bio);
1273
1274         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1275         while (atomic_dec_and_test(&r10_bio->remaining)) {
1276                 if (r10_bio->master_bio == NULL) {
1277                         /* the primary of several recovery bios */
1278                         sector_t s = r10_bio->sectors;
1279                         put_buf(r10_bio);
1280                         md_done_sync(mddev, s, 1);
1281                         break;
1282                 } else {
1283                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1284                         put_buf(r10_bio);
1285                         r10_bio = r10_bio2;
1286                 }
1287         }
1288 }
1289
1290 /*
1291  * Note: sync and recover and handled very differently for raid10
1292  * This code is for resync.
1293  * For resync, we read through virtual addresses and read all blocks.
1294  * If there is any error, we schedule a write.  The lowest numbered
1295  * drive is authoritative.
1296  * However requests come for physical address, so we need to map.
1297  * For every physical address there are raid_disks/copies virtual addresses,
1298  * which is always are least one, but is not necessarly an integer.
1299  * This means that a physical address can span multiple chunks, so we may
1300  * have to submit multiple io requests for a single sync request.
1301  */
1302 /*
1303  * We check if all blocks are in-sync and only write to blocks that
1304  * aren't in sync
1305  */
1306 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1307 {
1308         conf_t *conf = mddev->private;
1309         int i, first;
1310         struct bio *tbio, *fbio;
1311
1312         atomic_set(&r10_bio->remaining, 1);
1313
1314         /* find the first device with a block */
1315         for (i=0; i<conf->copies; i++)
1316                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1317                         break;
1318
1319         if (i == conf->copies)
1320                 goto done;
1321
1322         first = i;
1323         fbio = r10_bio->devs[i].bio;
1324
1325         /* now find blocks with errors */
1326         for (i=0 ; i < conf->copies ; i++) {
1327                 int  j, d;
1328                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1329
1330                 tbio = r10_bio->devs[i].bio;
1331
1332                 if (tbio->bi_end_io != end_sync_read)
1333                         continue;
1334                 if (i == first)
1335                         continue;
1336                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1337                         /* We know that the bi_io_vec layout is the same for
1338                          * both 'first' and 'i', so we just compare them.
1339                          * All vec entries are PAGE_SIZE;
1340                          */
1341                         for (j = 0; j < vcnt; j++)
1342                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1343                                            page_address(tbio->bi_io_vec[j].bv_page),
1344                                            PAGE_SIZE))
1345                                         break;
1346                         if (j == vcnt)
1347                                 continue;
1348                         mddev->resync_mismatches += r10_bio->sectors;
1349                 }
1350                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1351                         /* Don't fix anything. */
1352                         continue;
1353                 /* Ok, we need to write this bio
1354                  * First we need to fixup bv_offset, bv_len and
1355                  * bi_vecs, as the read request might have corrupted these
1356                  */
1357                 tbio->bi_vcnt = vcnt;
1358                 tbio->bi_size = r10_bio->sectors << 9;
1359                 tbio->bi_idx = 0;
1360                 tbio->bi_phys_segments = 0;
1361                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1362                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1363                 tbio->bi_next = NULL;
1364                 tbio->bi_rw = WRITE;
1365                 tbio->bi_private = r10_bio;
1366                 tbio->bi_sector = r10_bio->devs[i].addr;
1367
1368                 for (j=0; j < vcnt ; j++) {
1369                         tbio->bi_io_vec[j].bv_offset = 0;
1370                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1371
1372                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1373                                page_address(fbio->bi_io_vec[j].bv_page),
1374                                PAGE_SIZE);
1375                 }
1376                 tbio->bi_end_io = end_sync_write;
1377
1378                 d = r10_bio->devs[i].devnum;
1379                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1380                 atomic_inc(&r10_bio->remaining);
1381                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1382
1383                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1384                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1385                 generic_make_request(tbio);
1386         }
1387
1388 done:
1389         if (atomic_dec_and_test(&r10_bio->remaining)) {
1390                 md_done_sync(mddev, r10_bio->sectors, 1);
1391                 put_buf(r10_bio);
1392         }
1393 }
1394
1395 /*
1396  * Now for the recovery code.
1397  * Recovery happens across physical sectors.
1398  * We recover all non-is_sync drives by finding the virtual address of
1399  * each, and then choose a working drive that also has that virt address.
1400  * There is a separate r10_bio for each non-in_sync drive.
1401  * Only the first two slots are in use. The first for reading,
1402  * The second for writing.
1403  *
1404  */
1405
1406 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1407 {
1408         conf_t *conf = mddev->private;
1409         int i, d;
1410         struct bio *bio, *wbio;
1411
1412
1413         /* move the pages across to the second bio
1414          * and submit the write request
1415          */
1416         bio = r10_bio->devs[0].bio;
1417         wbio = r10_bio->devs[1].bio;
1418         for (i=0; i < wbio->bi_vcnt; i++) {
1419                 struct page *p = bio->bi_io_vec[i].bv_page;
1420                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1421                 wbio->bi_io_vec[i].bv_page = p;
1422         }
1423         d = r10_bio->devs[1].devnum;
1424
1425         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1426         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1427         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1428                 generic_make_request(wbio);
1429         else
1430                 bio_endio(wbio, -EIO);
1431 }
1432
1433
1434 /*
1435  * This is a kernel thread which:
1436  *
1437  *      1.      Retries failed read operations on working mirrors.
1438  *      2.      Updates the raid superblock when problems encounter.
1439  *      3.      Performs writes following reads for array synchronising.
1440  */
1441
1442 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1443 {
1444         int sect = 0; /* Offset from r10_bio->sector */
1445         int sectors = r10_bio->sectors;
1446         mdk_rdev_t*rdev;
1447         while(sectors) {
1448                 int s = sectors;
1449                 int sl = r10_bio->read_slot;
1450                 int success = 0;
1451                 int start;
1452
1453                 if (s > (PAGE_SIZE>>9))
1454                         s = PAGE_SIZE >> 9;
1455
1456                 rcu_read_lock();
1457                 do {
1458                         int d = r10_bio->devs[sl].devnum;
1459                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1460                         if (rdev &&
1461                             test_bit(In_sync, &rdev->flags)) {
1462                                 atomic_inc(&rdev->nr_pending);
1463                                 rcu_read_unlock();
1464                                 success = sync_page_io(rdev->bdev,
1465                                                        r10_bio->devs[sl].addr +
1466                                                        sect + rdev->data_offset,
1467                                                        s<<9,
1468                                                        conf->tmppage, READ);
1469                                 rdev_dec_pending(rdev, mddev);
1470                                 rcu_read_lock();
1471                                 if (success)
1472                                         break;
1473                         }
1474                         sl++;
1475                         if (sl == conf->copies)
1476                                 sl = 0;
1477                 } while (!success && sl != r10_bio->read_slot);
1478                 rcu_read_unlock();
1479
1480                 if (!success) {
1481                         /* Cannot read from anywhere -- bye bye array */
1482                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1483                         md_error(mddev, conf->mirrors[dn].rdev);
1484                         break;
1485                 }
1486
1487                 start = sl;
1488                 /* write it back and re-read */
1489                 rcu_read_lock();
1490                 while (sl != r10_bio->read_slot) {
1491                         int d;
1492                         if (sl==0)
1493                                 sl = conf->copies;
1494                         sl--;
1495                         d = r10_bio->devs[sl].devnum;
1496                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1497                         if (rdev &&
1498                             test_bit(In_sync, &rdev->flags)) {
1499                                 atomic_inc(&rdev->nr_pending);
1500                                 rcu_read_unlock();
1501                                 atomic_add(s, &rdev->corrected_errors);
1502                                 if (sync_page_io(rdev->bdev,
1503                                                  r10_bio->devs[sl].addr +
1504                                                  sect + rdev->data_offset,
1505                                                  s<<9, conf->tmppage, WRITE)
1506                                     == 0)
1507                                         /* Well, this device is dead */
1508                                         md_error(mddev, rdev);
1509                                 rdev_dec_pending(rdev, mddev);
1510                                 rcu_read_lock();
1511                         }
1512                 }
1513                 sl = start;
1514                 while (sl != r10_bio->read_slot) {
1515                         int d;
1516                         if (sl==0)
1517                                 sl = conf->copies;
1518                         sl--;
1519                         d = r10_bio->devs[sl].devnum;
1520                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1521                         if (rdev &&
1522                             test_bit(In_sync, &rdev->flags)) {
1523                                 char b[BDEVNAME_SIZE];
1524                                 atomic_inc(&rdev->nr_pending);
1525                                 rcu_read_unlock();
1526                                 if (sync_page_io(rdev->bdev,
1527                                                  r10_bio->devs[sl].addr +
1528                                                  sect + rdev->data_offset,
1529                                                  s<<9, conf->tmppage, READ) == 0)
1530                                         /* Well, this device is dead */
1531                                         md_error(mddev, rdev);
1532                                 else
1533                                         printk(KERN_INFO
1534                                                "raid10:%s: read error corrected"
1535                                                " (%d sectors at %llu on %s)\n",
1536                                                mdname(mddev), s,
1537                                                (unsigned long long)(sect+
1538                                                     rdev->data_offset),
1539                                                bdevname(rdev->bdev, b));
1540
1541                                 rdev_dec_pending(rdev, mddev);
1542                                 rcu_read_lock();
1543                         }
1544                 }
1545                 rcu_read_unlock();
1546
1547                 sectors -= s;
1548                 sect += s;
1549         }
1550 }
1551
1552 static void raid10d(mddev_t *mddev)
1553 {
1554         r10bio_t *r10_bio;
1555         struct bio *bio;
1556         unsigned long flags;
1557         conf_t *conf = mddev->private;
1558         struct list_head *head = &conf->retry_list;
1559         int unplug=0;
1560         mdk_rdev_t *rdev;
1561
1562         md_check_recovery(mddev);
1563
1564         for (;;) {
1565                 char b[BDEVNAME_SIZE];
1566
1567                 unplug += flush_pending_writes(conf);
1568
1569                 spin_lock_irqsave(&conf->device_lock, flags);
1570                 if (list_empty(head)) {
1571                         spin_unlock_irqrestore(&conf->device_lock, flags);
1572                         break;
1573                 }
1574                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1575                 list_del(head->prev);
1576                 conf->nr_queued--;
1577                 spin_unlock_irqrestore(&conf->device_lock, flags);
1578
1579                 mddev = r10_bio->mddev;
1580                 conf = mddev->private;
1581                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1582                         sync_request_write(mddev, r10_bio);
1583                         unplug = 1;
1584                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1585                         recovery_request_write(mddev, r10_bio);
1586                         unplug = 1;
1587                 } else {
1588                         int mirror;
1589                         /* we got a read error. Maybe the drive is bad.  Maybe just
1590                          * the block and we can fix it.
1591                          * We freeze all other IO, and try reading the block from
1592                          * other devices.  When we find one, we re-write
1593                          * and check it that fixes the read error.
1594                          * This is all done synchronously while the array is
1595                          * frozen.
1596                          */
1597                         if (mddev->ro == 0) {
1598                                 freeze_array(conf);
1599                                 fix_read_error(conf, mddev, r10_bio);
1600                                 unfreeze_array(conf);
1601                         }
1602
1603                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1604                         r10_bio->devs[r10_bio->read_slot].bio =
1605                                 mddev->ro ? IO_BLOCKED : NULL;
1606                         mirror = read_balance(conf, r10_bio);
1607                         if (mirror == -1) {
1608                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1609                                        " read error for block %llu\n",
1610                                        bdevname(bio->bi_bdev,b),
1611                                        (unsigned long long)r10_bio->sector);
1612                                 raid_end_bio_io(r10_bio);
1613                                 bio_put(bio);
1614                         } else {
1615                                 const bool do_sync = bio_rw_flagged(r10_bio->master_bio, BIO_RW_SYNCIO);
1616                                 bio_put(bio);
1617                                 rdev = conf->mirrors[mirror].rdev;
1618                                 if (printk_ratelimit())
1619                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1620                                                " another mirror\n",
1621                                                bdevname(rdev->bdev,b),
1622                                                (unsigned long long)r10_bio->sector);
1623                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1624                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1625                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1626                                         + rdev->data_offset;
1627                                 bio->bi_bdev = rdev->bdev;
1628                                 bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
1629                                 bio->bi_private = r10_bio;
1630                                 bio->bi_end_io = raid10_end_read_request;
1631                                 unplug = 1;
1632                                 generic_make_request(bio);
1633                         }
1634                 }
1635         }
1636         if (unplug)
1637                 unplug_slaves(mddev);
1638 }
1639
1640
1641 static int init_resync(conf_t *conf)
1642 {
1643         int buffs;
1644
1645         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1646         BUG_ON(conf->r10buf_pool);
1647         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1648         if (!conf->r10buf_pool)
1649                 return -ENOMEM;
1650         conf->next_resync = 0;
1651         return 0;
1652 }
1653
1654 /*
1655  * perform a "sync" on one "block"
1656  *
1657  * We need to make sure that no normal I/O request - particularly write
1658  * requests - conflict with active sync requests.
1659  *
1660  * This is achieved by tracking pending requests and a 'barrier' concept
1661  * that can be installed to exclude normal IO requests.
1662  *
1663  * Resync and recovery are handled very differently.
1664  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1665  *
1666  * For resync, we iterate over virtual addresses, read all copies,
1667  * and update if there are differences.  If only one copy is live,
1668  * skip it.
1669  * For recovery, we iterate over physical addresses, read a good
1670  * value for each non-in_sync drive, and over-write.
1671  *
1672  * So, for recovery we may have several outstanding complex requests for a
1673  * given address, one for each out-of-sync device.  We model this by allocating
1674  * a number of r10_bio structures, one for each out-of-sync device.
1675  * As we setup these structures, we collect all bio's together into a list
1676  * which we then process collectively to add pages, and then process again
1677  * to pass to generic_make_request.
1678  *
1679  * The r10_bio structures are linked using a borrowed master_bio pointer.
1680  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1681  * has its remaining count decremented to 0, the whole complex operation
1682  * is complete.
1683  *
1684  */
1685
1686 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1687 {
1688         conf_t *conf = mddev->private;
1689         r10bio_t *r10_bio;
1690         struct bio *biolist = NULL, *bio;
1691         sector_t max_sector, nr_sectors;
1692         int disk;
1693         int i;
1694         int max_sync;
1695         int sync_blocks;
1696
1697         sector_t sectors_skipped = 0;
1698         int chunks_skipped = 0;
1699
1700         if (!conf->r10buf_pool)
1701                 if (init_resync(conf))
1702                         return 0;
1703
1704  skipped:
1705         max_sector = mddev->dev_sectors;
1706         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1707                 max_sector = mddev->resync_max_sectors;
1708         if (sector_nr >= max_sector) {
1709                 /* If we aborted, we need to abort the
1710                  * sync on the 'current' bitmap chucks (there can
1711                  * be several when recovering multiple devices).
1712                  * as we may have started syncing it but not finished.
1713                  * We can find the current address in
1714                  * mddev->curr_resync, but for recovery,
1715                  * we need to convert that to several
1716                  * virtual addresses.
1717                  */
1718                 if (mddev->curr_resync < max_sector) { /* aborted */
1719                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1720                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1721                                                 &sync_blocks, 1);
1722                         else for (i=0; i<conf->raid_disks; i++) {
1723                                 sector_t sect =
1724                                         raid10_find_virt(conf, mddev->curr_resync, i);
1725                                 bitmap_end_sync(mddev->bitmap, sect,
1726                                                 &sync_blocks, 1);
1727                         }
1728                 } else /* completed sync */
1729                         conf->fullsync = 0;
1730
1731                 bitmap_close_sync(mddev->bitmap);
1732                 close_sync(conf);
1733                 *skipped = 1;
1734                 return sectors_skipped;
1735         }
1736         if (chunks_skipped >= conf->raid_disks) {
1737                 /* if there has been nothing to do on any drive,
1738                  * then there is nothing to do at all..
1739                  */
1740                 *skipped = 1;
1741                 return (max_sector - sector_nr) + sectors_skipped;
1742         }
1743
1744         if (max_sector > mddev->resync_max)
1745                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1746
1747         /* make sure whole request will fit in a chunk - if chunks
1748          * are meaningful
1749          */
1750         if (conf->near_copies < conf->raid_disks &&
1751             max_sector > (sector_nr | conf->chunk_mask))
1752                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1753         /*
1754          * If there is non-resync activity waiting for us then
1755          * put in a delay to throttle resync.
1756          */
1757         if (!go_faster && conf->nr_waiting)
1758                 msleep_interruptible(1000);
1759
1760         /* Again, very different code for resync and recovery.
1761          * Both must result in an r10bio with a list of bios that
1762          * have bi_end_io, bi_sector, bi_bdev set,
1763          * and bi_private set to the r10bio.
1764          * For recovery, we may actually create several r10bios
1765          * with 2 bios in each, that correspond to the bios in the main one.
1766          * In this case, the subordinate r10bios link back through a
1767          * borrowed master_bio pointer, and the counter in the master
1768          * includes a ref from each subordinate.
1769          */
1770         /* First, we decide what to do and set ->bi_end_io
1771          * To end_sync_read if we want to read, and
1772          * end_sync_write if we will want to write.
1773          */
1774
1775         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1776         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1777                 /* recovery... the complicated one */
1778                 int j, k;
1779                 r10_bio = NULL;
1780
1781                 for (i=0 ; i<conf->raid_disks; i++)
1782                         if (conf->mirrors[i].rdev &&
1783                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1784                                 int still_degraded = 0;
1785                                 /* want to reconstruct this device */
1786                                 r10bio_t *rb2 = r10_bio;
1787                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1788                                 int must_sync;
1789                                 /* Unless we are doing a full sync, we only need
1790                                  * to recover the block if it is set in the bitmap
1791                                  */
1792                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1793                                                               &sync_blocks, 1);
1794                                 if (sync_blocks < max_sync)
1795                                         max_sync = sync_blocks;
1796                                 if (!must_sync &&
1797                                     !conf->fullsync) {
1798                                         /* yep, skip the sync_blocks here, but don't assume
1799                                          * that there will never be anything to do here
1800                                          */
1801                                         chunks_skipped = -1;
1802                                         continue;
1803                                 }
1804
1805                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1806                                 raise_barrier(conf, rb2 != NULL);
1807                                 atomic_set(&r10_bio->remaining, 0);
1808
1809                                 r10_bio->master_bio = (struct bio*)rb2;
1810                                 if (rb2)
1811                                         atomic_inc(&rb2->remaining);
1812                                 r10_bio->mddev = mddev;
1813                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1814                                 r10_bio->sector = sect;
1815
1816                                 raid10_find_phys(conf, r10_bio);
1817
1818                                 /* Need to check if the array will still be
1819                                  * degraded
1820                                  */
1821                                 for (j=0; j<conf->raid_disks; j++)
1822                                         if (conf->mirrors[j].rdev == NULL ||
1823                                             test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
1824                                                 still_degraded = 1;
1825                                                 break;
1826                                         }
1827
1828                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1829                                                               &sync_blocks, still_degraded);
1830
1831                                 for (j=0; j<conf->copies;j++) {
1832                                         int d = r10_bio->devs[j].devnum;
1833                                         if (conf->mirrors[d].rdev &&
1834                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1835                                                 /* This is where we read from */
1836                                                 bio = r10_bio->devs[0].bio;
1837                                                 bio->bi_next = biolist;
1838                                                 biolist = bio;
1839                                                 bio->bi_private = r10_bio;
1840                                                 bio->bi_end_io = end_sync_read;
1841                                                 bio->bi_rw = READ;
1842                                                 bio->bi_sector = r10_bio->devs[j].addr +
1843                                                         conf->mirrors[d].rdev->data_offset;
1844                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1845                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1846                                                 atomic_inc(&r10_bio->remaining);
1847                                                 /* and we write to 'i' */
1848
1849                                                 for (k=0; k<conf->copies; k++)
1850                                                         if (r10_bio->devs[k].devnum == i)
1851                                                                 break;
1852                                                 BUG_ON(k == conf->copies);
1853                                                 bio = r10_bio->devs[1].bio;
1854                                                 bio->bi_next = biolist;
1855                                                 biolist = bio;
1856                                                 bio->bi_private = r10_bio;
1857                                                 bio->bi_end_io = end_sync_write;
1858                                                 bio->bi_rw = WRITE;
1859                                                 bio->bi_sector = r10_bio->devs[k].addr +
1860                                                         conf->mirrors[i].rdev->data_offset;
1861                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1862
1863                                                 r10_bio->devs[0].devnum = d;
1864                                                 r10_bio->devs[1].devnum = i;
1865
1866                                                 break;
1867                                         }
1868                                 }
1869                                 if (j == conf->copies) {
1870                                         /* Cannot recover, so abort the recovery */
1871                                         put_buf(r10_bio);
1872                                         if (rb2)
1873                                                 atomic_dec(&rb2->remaining);
1874                                         r10_bio = rb2;
1875                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
1876                                                               &mddev->recovery))
1877                                                 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1878                                                        mdname(mddev));
1879                                         break;
1880                                 }
1881                         }
1882                 if (biolist == NULL) {
1883                         while (r10_bio) {
1884                                 r10bio_t *rb2 = r10_bio;
1885                                 r10_bio = (r10bio_t*) rb2->master_bio;
1886                                 rb2->master_bio = NULL;
1887                                 put_buf(rb2);
1888                         }
1889                         goto giveup;
1890                 }
1891         } else {
1892                 /* resync. Schedule a read for every block at this virt offset */
1893                 int count = 0;
1894
1895                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1896
1897                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1898                                        &sync_blocks, mddev->degraded) &&
1899                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1900                         /* We can skip this block */
1901                         *skipped = 1;
1902                         return sync_blocks + sectors_skipped;
1903                 }
1904                 if (sync_blocks < max_sync)
1905                         max_sync = sync_blocks;
1906                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1907
1908                 r10_bio->mddev = mddev;
1909                 atomic_set(&r10_bio->remaining, 0);
1910                 raise_barrier(conf, 0);
1911                 conf->next_resync = sector_nr;
1912
1913                 r10_bio->master_bio = NULL;
1914                 r10_bio->sector = sector_nr;
1915                 set_bit(R10BIO_IsSync, &r10_bio->state);
1916                 raid10_find_phys(conf, r10_bio);
1917                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1918
1919                 for (i=0; i<conf->copies; i++) {
1920                         int d = r10_bio->devs[i].devnum;
1921                         bio = r10_bio->devs[i].bio;
1922                         bio->bi_end_io = NULL;
1923                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
1924                         if (conf->mirrors[d].rdev == NULL ||
1925                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1926                                 continue;
1927                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1928                         atomic_inc(&r10_bio->remaining);
1929                         bio->bi_next = biolist;
1930                         biolist = bio;
1931                         bio->bi_private = r10_bio;
1932                         bio->bi_end_io = end_sync_read;
1933                         bio->bi_rw = READ;
1934                         bio->bi_sector = r10_bio->devs[i].addr +
1935                                 conf->mirrors[d].rdev->data_offset;
1936                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1937                         count++;
1938                 }
1939
1940                 if (count < 2) {
1941                         for (i=0; i<conf->copies; i++) {
1942                                 int d = r10_bio->devs[i].devnum;
1943                                 if (r10_bio->devs[i].bio->bi_end_io)
1944                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1945                         }
1946                         put_buf(r10_bio);
1947                         biolist = NULL;
1948                         goto giveup;
1949                 }
1950         }
1951
1952         for (bio = biolist; bio ; bio=bio->bi_next) {
1953
1954                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1955                 if (bio->bi_end_io)
1956                         bio->bi_flags |= 1 << BIO_UPTODATE;
1957                 bio->bi_vcnt = 0;
1958                 bio->bi_idx = 0;
1959                 bio->bi_phys_segments = 0;
1960                 bio->bi_size = 0;
1961         }
1962
1963         nr_sectors = 0;
1964         if (sector_nr + max_sync < max_sector)
1965                 max_sector = sector_nr + max_sync;
1966         do {
1967                 struct page *page;
1968                 int len = PAGE_SIZE;
1969                 disk = 0;
1970                 if (sector_nr + (len>>9) > max_sector)
1971                         len = (max_sector - sector_nr) << 9;
1972                 if (len == 0)
1973                         break;
1974                 for (bio= biolist ; bio ; bio=bio->bi_next) {
1975                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1976                         if (bio_add_page(bio, page, len, 0) == 0) {
1977                                 /* stop here */
1978                                 struct bio *bio2;
1979                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1980                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1981                                         /* remove last page from this bio */
1982                                         bio2->bi_vcnt--;
1983                                         bio2->bi_size -= len;
1984                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1985                                 }
1986                                 goto bio_full;
1987                         }
1988                         disk = i;
1989                 }
1990                 nr_sectors += len>>9;
1991                 sector_nr += len>>9;
1992         } while (biolist->bi_vcnt < RESYNC_PAGES);
1993  bio_full:
1994         r10_bio->sectors = nr_sectors;
1995
1996         while (biolist) {
1997                 bio = biolist;
1998                 biolist = biolist->bi_next;
1999
2000                 bio->bi_next = NULL;
2001                 r10_bio = bio->bi_private;
2002                 r10_bio->sectors = nr_sectors;
2003
2004                 if (bio->bi_end_io == end_sync_read) {
2005                         md_sync_acct(bio->bi_bdev, nr_sectors);
2006                         generic_make_request(bio);
2007                 }
2008         }
2009
2010         if (sectors_skipped)
2011                 /* pretend they weren't skipped, it makes
2012                  * no important difference in this case
2013                  */
2014                 md_done_sync(mddev, sectors_skipped, 1);
2015
2016         return sectors_skipped + nr_sectors;
2017  giveup:
2018         /* There is nowhere to write, so all non-sync
2019          * drives must be failed, so try the next chunk...
2020          */
2021         if (sector_nr + max_sync < max_sector)
2022                 max_sector = sector_nr + max_sync;
2023
2024         sectors_skipped += (max_sector - sector_nr);
2025         chunks_skipped ++;
2026         sector_nr = max_sector;
2027         goto skipped;
2028 }
2029
2030 static sector_t
2031 raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2032 {
2033         sector_t size;
2034         conf_t *conf = mddev->private;
2035
2036         if (!raid_disks)
2037                 raid_disks = mddev->raid_disks;
2038         if (!sectors)
2039                 sectors = mddev->dev_sectors;
2040
2041         size = sectors >> conf->chunk_shift;
2042         sector_div(size, conf->far_copies);
2043         size = size * raid_disks;
2044         sector_div(size, conf->near_copies);
2045
2046         return size << conf->chunk_shift;
2047 }
2048
2049 static int run(mddev_t *mddev)
2050 {
2051         conf_t *conf;
2052         int i, disk_idx, chunk_size;
2053         mirror_info_t *disk;
2054         mdk_rdev_t *rdev;
2055         int nc, fc, fo;
2056         sector_t stride, size;
2057
2058         if (mddev->chunk_sectors < (PAGE_SIZE >> 9) ||
2059             !is_power_of_2(mddev->chunk_sectors)) {
2060                 printk(KERN_ERR "md/raid10: chunk size must be "
2061                        "at least PAGE_SIZE(%ld) and be a power of 2.\n", PAGE_SIZE);
2062                 return -EINVAL;
2063         }
2064
2065         nc = mddev->layout & 255;
2066         fc = (mddev->layout >> 8) & 255;
2067         fo = mddev->layout & (1<<16);
2068         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2069             (mddev->layout >> 17)) {
2070                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
2071                        mdname(mddev), mddev->layout);
2072                 goto out;
2073         }
2074         /*
2075          * copy the already verified devices into our private RAID10
2076          * bookkeeping area. [whatever we allocate in run(),
2077          * should be freed in stop()]
2078          */
2079         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2080         mddev->private = conf;
2081         if (!conf) {
2082                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2083                         mdname(mddev));
2084                 goto out;
2085         }
2086         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2087                                  GFP_KERNEL);
2088         if (!conf->mirrors) {
2089                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2090                        mdname(mddev));
2091                 goto out_free_conf;
2092         }
2093
2094         conf->tmppage = alloc_page(GFP_KERNEL);
2095         if (!conf->tmppage)
2096                 goto out_free_conf;
2097
2098         conf->mddev = mddev;
2099         conf->raid_disks = mddev->raid_disks;
2100         conf->near_copies = nc;
2101         conf->far_copies = fc;
2102         conf->copies = nc*fc;
2103         conf->far_offset = fo;
2104         conf->chunk_mask = mddev->chunk_sectors - 1;
2105         conf->chunk_shift = ffz(~mddev->chunk_sectors);
2106         size = mddev->dev_sectors >> conf->chunk_shift;
2107         sector_div(size, fc);
2108         size = size * conf->raid_disks;
2109         sector_div(size, nc);
2110         /* 'size' is now the number of chunks in the array */
2111         /* calculate "used chunks per device" in 'stride' */
2112         stride = size * conf->copies;
2113
2114         /* We need to round up when dividing by raid_disks to
2115          * get the stride size.
2116          */
2117         stride += conf->raid_disks - 1;
2118         sector_div(stride, conf->raid_disks);
2119         mddev->dev_sectors = stride << conf->chunk_shift;
2120
2121         if (fo)
2122                 stride = 1;
2123         else
2124                 sector_div(stride, fc);
2125         conf->stride = stride << conf->chunk_shift;
2126
2127         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2128                                                 r10bio_pool_free, conf);
2129         if (!conf->r10bio_pool) {
2130                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2131                         mdname(mddev));
2132                 goto out_free_conf;
2133         }
2134
2135         spin_lock_init(&conf->device_lock);
2136         mddev->queue->queue_lock = &conf->device_lock;
2137
2138         chunk_size = mddev->chunk_sectors << 9;
2139         blk_queue_io_min(mddev->queue, chunk_size);
2140         if (conf->raid_disks % conf->near_copies)
2141                 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2142         else
2143                 blk_queue_io_opt(mddev->queue, chunk_size *
2144                                  (conf->raid_disks / conf->near_copies));
2145
2146         list_for_each_entry(rdev, &mddev->disks, same_set) {
2147                 disk_idx = rdev->raid_disk;
2148                 if (disk_idx >= mddev->raid_disks
2149                     || disk_idx < 0)
2150                         continue;
2151                 disk = conf->mirrors + disk_idx;
2152
2153                 disk->rdev = rdev;
2154                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2155                                   rdev->data_offset << 9);
2156                 /* as we don't honour merge_bvec_fn, we must never risk
2157                  * violating it, so limit ->max_sector to one PAGE, as
2158                  * a one page request is never in violation.
2159                  */
2160                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2161                     queue_max_sectors(mddev->queue) > (PAGE_SIZE>>9))
2162                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
2163
2164                 disk->head_position = 0;
2165         }
2166         INIT_LIST_HEAD(&conf->retry_list);
2167
2168         spin_lock_init(&conf->resync_lock);
2169         init_waitqueue_head(&conf->wait_barrier);
2170
2171         /* need to check that every block has at least one working mirror */
2172         if (!enough(conf)) {
2173                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2174                        mdname(mddev));
2175                 goto out_free_conf;
2176         }
2177
2178         mddev->degraded = 0;
2179         for (i = 0; i < conf->raid_disks; i++) {
2180
2181                 disk = conf->mirrors + i;
2182
2183                 if (!disk->rdev ||
2184                     !test_bit(In_sync, &disk->rdev->flags)) {
2185                         disk->head_position = 0;
2186                         mddev->degraded++;
2187                         if (disk->rdev)
2188                                 conf->fullsync = 1;
2189                 }
2190         }
2191
2192
2193         mddev->thread = md_register_thread(raid10d, mddev, NULL);
2194         if (!mddev->thread) {
2195                 printk(KERN_ERR
2196                        "raid10: couldn't allocate thread for %s\n",
2197                        mdname(mddev));
2198                 goto out_free_conf;
2199         }
2200
2201         if (mddev->recovery_cp != MaxSector)
2202                 printk(KERN_NOTICE "raid10: %s is not clean"
2203                        " -- starting background reconstruction\n",
2204                        mdname(mddev));
2205         printk(KERN_INFO
2206                 "raid10: raid set %s active with %d out of %d devices\n",
2207                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2208                 mddev->raid_disks);
2209         /*
2210          * Ok, everything is just fine now
2211          */
2212         md_set_array_sectors(mddev, raid10_size(mddev, 0, 0));
2213         mddev->resync_max_sectors = raid10_size(mddev, 0, 0);
2214
2215         mddev->queue->unplug_fn = raid10_unplug;
2216         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2217         mddev->queue->backing_dev_info.congested_data = mddev;
2218
2219         /* Calculate max read-ahead size.
2220          * We need to readahead at least twice a whole stripe....
2221          * maybe...
2222          */
2223         {
2224                 int stripe = conf->raid_disks *
2225                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2226                 stripe /= conf->near_copies;
2227                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2228                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2229         }
2230
2231         if (conf->near_copies < mddev->raid_disks)
2232                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2233         md_integrity_register(mddev);
2234         return 0;
2235
2236 out_free_conf:
2237         if (conf->r10bio_pool)
2238                 mempool_destroy(conf->r10bio_pool);
2239         safe_put_page(conf->tmppage);
2240         kfree(conf->mirrors);
2241         kfree(conf);
2242         mddev->private = NULL;
2243 out:
2244         return -EIO;
2245 }
2246
2247 static int stop(mddev_t *mddev)
2248 {
2249         conf_t *conf = mddev->private;
2250
2251         raise_barrier(conf, 0);
2252         lower_barrier(conf);
2253
2254         md_unregister_thread(mddev->thread);
2255         mddev->thread = NULL;
2256         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2257         if (conf->r10bio_pool)
2258                 mempool_destroy(conf->r10bio_pool);
2259         kfree(conf->mirrors);
2260         kfree(conf);
2261         mddev->private = NULL;
2262         return 0;
2263 }
2264
2265 static void raid10_quiesce(mddev_t *mddev, int state)
2266 {
2267         conf_t *conf = mddev->private;
2268
2269         switch(state) {
2270         case 1:
2271                 raise_barrier(conf, 0);
2272                 break;
2273         case 0:
2274                 lower_barrier(conf);
2275                 break;
2276         }
2277         if (mddev->thread) {
2278                 if (mddev->bitmap)
2279                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2280                 else
2281                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2282                 md_wakeup_thread(mddev->thread);
2283         }
2284 }
2285
2286 static struct mdk_personality raid10_personality =
2287 {
2288         .name           = "raid10",
2289         .level          = 10,
2290         .owner          = THIS_MODULE,
2291         .make_request   = make_request,
2292         .run            = run,
2293         .stop           = stop,
2294         .status         = status,
2295         .error_handler  = error,
2296         .hot_add_disk   = raid10_add_disk,
2297         .hot_remove_disk= raid10_remove_disk,
2298         .spare_active   = raid10_spare_active,
2299         .sync_request   = sync_request,
2300         .quiesce        = raid10_quiesce,
2301         .size           = raid10_size,
2302 };
2303
2304 static int __init raid_init(void)
2305 {
2306         return register_md_personality(&raid10_personality);
2307 }
2308
2309 static void raid_exit(void)
2310 {
2311         unregister_md_personality(&raid10_personality);
2312 }
2313
2314 module_init(raid_init);
2315 module_exit(raid_exit);
2316 MODULE_LICENSE("GPL");
2317 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2318 MODULE_ALIAS("md-raid10");
2319 MODULE_ALIAS("md-level-10");