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/slab.h>
  22 #include <linux/delay.h>
  23 #include <linux/blkdev.h>
  24 #include <linux/seq_file.h>
  25 #include "md.h"
  26 #include "raid10.h"
  27 #include "raid0.h"
  28 #include "bitmap.h"
  29
  30 /*
  31  * RAID10 provides a combination of RAID0 and RAID1 functionality.
  32  * The layout of data is defined by
  33  *    chunk_size
  34  *    raid_disks
  35  *    near_copies (stored in low byte of layout)
  36  *    far_copies (stored in second byte of layout)
  37  *    far_offset (stored in bit 16 of layout )
  38  *
  39  * The data to be stored is divided into chunks using chunksize.
  40  * Each device is divided into far_copies sections.
  41  * In each section, chunks are laid out in a style similar to raid0, but
  42  * near_copies copies of each chunk is stored (each on a different drive).
  43  * The starting device for each section is offset near_copies from the starting
  44  * device of the previous section.
  45  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
  46  * drive.
  47  * near_copies and far_copies must be at least one, and their product is at most
  48  * raid_disks.
  49  *
  50  * If far_offset is true, then the far_copies are handled a bit differently.
  51  * The copies are still in different stripes, but instead of be very far apart
  52  * on disk, there are adjacent stripes.
  53  */
  54
  55 /*
  56  * Number of guaranteed r10bios in case of extreme VM load:
  57  */
  58 #define NR_RAID10_BIOS 256
  59
  60 static void unplug_slaves(mddev_t *mddev);
  61
  62 static void allow_barrier(conf_t *conf);
  63 static void lower_barrier(conf_t *conf);
  64
  65 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
  66 {
  67         conf_t *conf = data;
  68         r10bio_t *r10_bio;
  69         int size = offsetof(struct r10bio_s, devs[conf->copies]);
  70
  71         /* allocate a r10bio with room for raid_disks entries in the bios array */
  72         r10_bio = kzalloc(size, gfp_flags);
  73         if (!r10_bio && conf->mddev)
  74                 unplug_slaves(conf->mddev);
  75
  76         return r10_bio;
  77 }
  78
  79 static void r10bio_pool_free(void *r10_bio, void *data)
  80 {
  81         kfree(r10_bio);
  82 }
  83
  84 /* Maximum size of each resync request */
  85 #define RESYNC_BLOCK_SIZE (64*1024)
  86 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  87 /* amount of memory to reserve for resync requests */
  88 #define RESYNC_WINDOW (1024*1024)
  89 /* maximum number of concurrent requests, memory permitting */
  90 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
  91
  92 /*
  93  * When performing a resync, we need to read and compare, so
  94  * we need as many pages are there are copies.
  95  * When performing a recovery, we need 2 bios, one for read,
  96  * one for write (we recover only one drive per r10buf)
  97  *
  98  */
  99 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
 100 {
 101         conf_t *conf = data;
 102         struct page *page;
 103         r10bio_t *r10_bio;
 104         struct bio *bio;
 105         int i, j;
 106         int nalloc;
 107
 108         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 109         if (!r10_bio) {
 110                 unplug_slaves(conf->mddev);
 111                 return NULL;
 112         }
 113
 114         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
 115                 nalloc = conf->copies; /* resync */
 116         else
 117                 nalloc = 2; /* recovery */
 118
 119         /*
 120          * Allocate bios.
 121          */
 122         for (j = nalloc ; j-- ; ) {
 123                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 124                 if (!bio)
 125                         goto out_free_bio;
 126                 r10_bio->devs[j].bio = bio;
 127         }
 128         /*
 129          * Allocate RESYNC_PAGES data pages and attach them
 130          * where needed.
 131          */
 132         for (j = 0 ; j < nalloc; j++) {
 133                 bio = r10_bio->devs[j].bio;
 134                 for (i = 0; i < RESYNC_PAGES; i++) {
 135                         page = alloc_page(gfp_flags);
 136                         if (unlikely(!page))
 137                                 goto out_free_pages;
 138
 139                         bio->bi_io_vec[i].bv_page = page;
 140                 }
 141         }
 142
 143         return r10_bio;
 144
 145 out_free_pages:
 146         for ( ; i > 0 ; i--)
 147                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
 148         while (j--)
 149                 for (i = 0; i < RESYNC_PAGES ; i++)
 150                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 151         j = -1;
 152 out_free_bio:
 153         while ( ++j < nalloc )
 154                 bio_put(r10_bio->devs[j].bio);
 155         r10bio_pool_free(r10_bio, conf);
 156         return NULL;
 157 }
 158
 159 static void r10buf_pool_free(void *__r10_bio, void *data)
 160 {
 161         int i;
 162         conf_t *conf = data;
 163         r10bio_t *r10bio = __r10_bio;
 164         int j;
 165
 166         for (j=0; j < conf->copies; j++) {
 167                 struct bio *bio = r10bio->devs[j].bio;
 168                 if (bio) {
 169                         for (i = 0; i < RESYNC_PAGES; i++) {
 170                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 171                                 bio->bi_io_vec[i].bv_page = NULL;
 172                         }
 173                         bio_put(bio);
 174                 }
 175         }
 176         r10bio_pool_free(r10bio, conf);
 177 }
 178
 179 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
 180 {
 181         int i;
 182
 183         for (i = 0; i < conf->copies; i++) {
 184                 struct bio **bio = & r10_bio->devs[i].bio;
 185                 if (*bio && *bio != IO_BLOCKED)
 186                         bio_put(*bio);
 187                 *bio = NULL;
 188         }
 189 }
 190
 191 static void free_r10bio(r10bio_t *r10_bio)
 192 {
 193         conf_t *conf = r10_bio->mddev->private;
 194
 195         /*
 196          * Wake up any possible resync thread that waits for the device
 197          * to go idle.
 198          */
 199         allow_barrier(conf);
 200
 201         put_all_bios(conf, r10_bio);
 202         mempool_free(r10_bio, conf->r10bio_pool);
 203 }
 204
 205 static void put_buf(r10bio_t *r10_bio)
 206 {
 207         conf_t *conf = r10_bio->mddev->private;
 208
 209         mempool_free(r10_bio, conf->r10buf_pool);
 210
 211         lower_barrier(conf);
 212 }
 213
 214 static void reschedule_retry(r10bio_t *r10_bio)
 215 {
 216         unsigned long flags;
 217         mddev_t *mddev = r10_bio->mddev;
 218         conf_t *conf = mddev->private;
 219
 220         spin_lock_irqsave(&conf->device_lock, flags);
 221         list_add(&r10_bio->retry_list, &conf->retry_list);
 222         conf->nr_queued ++;
 223         spin_unlock_irqrestore(&conf->device_lock, flags);
 224
 225         /* wake up frozen array... */
 226         wake_up(&conf->wait_barrier);
 227
 228         md_wakeup_thread(mddev->thread);
 229 }
 230
 231 /*
 232  * raid_end_bio_io() is called when we have finished servicing a mirrored
 233  * operation and are ready to return a success/failure code to the buffer
 234  * cache layer.
 235  */
 236 static void raid_end_bio_io(r10bio_t *r10_bio)
 237 {
 238         struct bio *bio = r10_bio->master_bio;
 239
 240         bio_endio(bio,
 241                 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
 242         free_r10bio(r10_bio);
 243 }
 244
 245 /*
 246  * Update disk head position estimator based on IRQ completion info.
 247  */
 248 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
 249 {
 250         conf_t *conf = r10_bio->mddev->private;
 251
 252         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 253                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 254 }
 255
 256 static void raid10_end_read_request(struct bio *bio, int error)
 257 {
 258         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 259         r10bio_t *r10_bio = bio->bi_private;
 260         int slot, dev;
 261         conf_t *conf = r10_bio->mddev->private;
 262
 263
 264         slot = r10_bio->read_slot;
 265         dev = r10_bio->devs[slot].devnum;
 266         /*
 267          * this branch is our 'one mirror IO has finished' event handler:
 268          */
 269         update_head_pos(slot, r10_bio);
 270
 271         if (uptodate) {
 272                 /*
 273                  * Set R10BIO_Uptodate in our master bio, so that
 274                  * we will return a good error code to the higher
 275                  * levels even if IO on some other mirrored buffer fails.
 276                  *
 277                  * The 'master' represents the composite IO operation to
 278                  * user-side. So if something waits for IO, then it will
 279                  * wait for the 'master' bio.
 280                  */
 281                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 282                 raid_end_bio_io(r10_bio);
 283         } else {
 284                 /*
 285                  * oops, read error:
 286                  */
 287                 char b[BDEVNAME_SIZE];
 288                 if (printk_ratelimit())
 289                         printk(KERN_ERR "md/raid10:%s: %s: rescheduling sector %llu\n",
 290                                mdname(conf->mddev),
 291                                bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
 292                 reschedule_retry(r10_bio);
 293         }
 294
 295         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 296 }
 297
 298 static void raid10_end_write_request(struct bio *bio, int error)
 299 {
 300         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 301         r10bio_t *r10_bio = bio->bi_private;
 302         int slot, dev;
 303         conf_t *conf = r10_bio->mddev->private;
 304
 305         for (slot = 0; slot < conf->copies; slot++)
 306                 if (r10_bio->devs[slot].bio == bio)
 307                         break;
 308         dev = r10_bio->devs[slot].devnum;
 309
 310         /*
 311          * this branch is our 'one mirror IO has finished' event handler:
 312          */
 313         if (!uptodate) {
 314                 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
 315                 /* an I/O failed, we can't clear the bitmap */
 316                 set_bit(R10BIO_Degraded, &r10_bio->state);
 317         } else
 318                 /*
 319                  * Set R10BIO_Uptodate in our master bio, so that
 320                  * we will return a good error code for to the higher
 321                  * levels even if IO on some other mirrored buffer fails.
 322                  *
 323                  * The 'master' represents the composite IO operation to
 324                  * user-side. So if something waits for IO, then it will
 325                  * wait for the 'master' bio.
 326                  */
 327                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 328
 329         update_head_pos(slot, r10_bio);
 330
 331         /*
 332          *
 333          * Let's see if all mirrored write operations have finished
 334          * already.
 335          */
 336         if (atomic_dec_and_test(&r10_bio->remaining)) {
 337                 /* clear the bitmap if all writes complete successfully */
 338                 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 339                                 r10_bio->sectors,
 340                                 !test_bit(R10BIO_Degraded, &r10_bio->state),
 341                                 0);
 342                 md_write_end(r10_bio->mddev);
 343                 raid_end_bio_io(r10_bio);
 344         }
 345
 346         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 347 }
 348
 349
 350 /*
 351  * RAID10 layout manager
 352  * Aswell as the chunksize and raid_disks count, there are two
 353  * parameters: near_copies and far_copies.
 354  * near_copies * far_copies must be <= raid_disks.
 355  * Normally one of these will be 1.
 356  * If both are 1, we get raid0.
 357  * If near_copies == raid_disks, we get raid1.
 358  *
 359  * Chunks are layed out in raid0 style with near_copies copies of the
 360  * first chunk, followed by near_copies copies of the next chunk and
 361  * so on.
 362  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 363  * as described above, we start again with a device offset of near_copies.
 364  * So we effectively have another copy of the whole array further down all
 365  * the drives, but with blocks on different drives.
 366  * With this layout, and block is never stored twice on the one device.
 367  *
 368  * raid10_find_phys finds the sector offset of a given virtual sector
 369  * on each device that it is on.
 370  *
 371  * raid10_find_virt does the reverse mapping, from a device and a
 372  * sector offset to a virtual address
 373  */
 374
 375 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
 376 {
 377         int n,f;
 378         sector_t sector;
 379         sector_t chunk;
 380         sector_t stripe;
 381         int dev;
 382
 383         int slot = 0;
 384
 385         /* now calculate first sector/dev */
 386         chunk = r10bio->sector >> conf->chunk_shift;
 387         sector = r10bio->sector & conf->chunk_mask;
 388
 389         chunk *= conf->near_copies;
 390         stripe = chunk;
 391         dev = sector_div(stripe, conf->raid_disks);
 392         if (conf->far_offset)
 393                 stripe *= conf->far_copies;
 394
 395         sector += stripe << conf->chunk_shift;
 396
 397         /* and calculate all the others */
 398         for (n=0; n < conf->near_copies; n++) {
 399                 int d = dev;
 400                 sector_t s = sector;
 401                 r10bio->devs[slot].addr = sector;
 402                 r10bio->devs[slot].devnum = d;
 403                 slot++;
 404
 405                 for (f = 1; f < conf->far_copies; f++) {
 406                         d += conf->near_copies;
 407                         if (d >= conf->raid_disks)
 408                                 d -= conf->raid_disks;
 409                         s += conf->stride;
 410                         r10bio->devs[slot].devnum = d;
 411                         r10bio->devs[slot].addr = s;
 412                         slot++;
 413                 }
 414                 dev++;
 415                 if (dev >= conf->raid_disks) {
 416                         dev = 0;
 417                         sector += (conf->chunk_mask + 1);
 418                 }
 419         }
 420         BUG_ON(slot != conf->copies);
 421 }
 422
 423 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
 424 {
 425         sector_t offset, chunk, vchunk;
 426
 427         offset = sector & conf->chunk_mask;
 428         if (conf->far_offset) {
 429                 int fc;
 430                 chunk = sector >> conf->chunk_shift;
 431                 fc = sector_div(chunk, conf->far_copies);
 432                 dev -= fc * conf->near_copies;
 433                 if (dev < 0)
 434                         dev += conf->raid_disks;
 435         } else {
 436                 while (sector >= conf->stride) {
 437                         sector -= conf->stride;
 438                         if (dev < conf->near_copies)
 439                                 dev += conf->raid_disks - conf->near_copies;
 440                         else
 441                                 dev -= conf->near_copies;
 442                 }
 443                 chunk = sector >> conf->chunk_shift;
 444         }
 445         vchunk = chunk * conf->raid_disks + dev;
 446         sector_div(vchunk, conf->near_copies);
 447         return (vchunk << conf->chunk_shift) + offset;
 448 }
 449
 450 /**
 451  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 452  *      @q: request queue
 453  *      @bvm: properties of new bio
 454  *      @biovec: the request that could be merged to it.
 455  *
 456  *      Return amount of bytes we can accept at this offset
 457  *      If near_copies == raid_disk, there are no striping issues,
 458  *      but in that case, the function isn't called at all.
 459  */
 460 static int raid10_mergeable_bvec(struct request_queue *q,
 461                                  struct bvec_merge_data *bvm,
 462                                  struct bio_vec *biovec)
 463 {
 464         mddev_t *mddev = q->queuedata;
 465         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 466         int max;
 467         unsigned int chunk_sectors = mddev->chunk_sectors;
 468         unsigned int bio_sectors = bvm->bi_size >> 9;
 469
 470         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 471         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 472         if (max <= biovec->bv_len && bio_sectors == 0)
 473                 return biovec->bv_len;
 474         else
 475                 return max;
 476 }
 477
 478 /*
 479  * This routine returns the disk from which the requested read should
 480  * be done. There is a per-array 'next expected sequential IO' sector
 481  * number - if this matches on the next IO then we use the last disk.
 482  * There is also a per-disk 'last know head position' sector that is
 483  * maintained from IRQ contexts, both the normal and the resync IO
 484  * completion handlers update this position correctly. If there is no
 485  * perfect sequential match then we pick the disk whose head is closest.
 486  *
 487  * If there are 2 mirrors in the same 2 devices, performance degrades
 488  * because position is mirror, not device based.
 489  *
 490  * The rdev for the device selected will have nr_pending incremented.
 491  */
 492
 493 /*
 494  * FIXME: possibly should rethink readbalancing and do it differently
 495  * depending on near_copies / far_copies geometry.
 496  */
 497 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
 498 {
 499         const sector_t this_sector = r10_bio->sector;
 500         int disk, slot, nslot;
 501         const int sectors = r10_bio->sectors;
 502         sector_t new_distance, current_distance;
 503         mdk_rdev_t *rdev;
 504
 505         raid10_find_phys(conf, r10_bio);
 506         rcu_read_lock();
 507         /*
 508          * Check if we can balance. We can balance on the whole
 509          * device if no resync is going on (recovery is ok), or below
 510          * the resync window. We take the first readable disk when
 511          * above the resync window.
 512          */
 513         if (conf->mddev->recovery_cp < MaxSector
 514             && (this_sector + sectors >= conf->next_resync)) {
 515                 /* make sure that disk is operational */
 516                 slot = 0;
 517                 disk = r10_bio->devs[slot].devnum;
 518
 519                 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 520                        r10_bio->devs[slot].bio == IO_BLOCKED ||
 521                        !test_bit(In_sync, &rdev->flags)) {
 522                         slot++;
 523                         if (slot == conf->copies) {
 524                                 slot = 0;
 525                                 disk = -1;
 526                                 break;
 527                         }
 528                         disk = r10_bio->devs[slot].devnum;
 529                 }
 530                 goto rb_out;
 531         }
 532
 533
 534         /* make sure the disk is operational */
 535         slot = 0;
 536         disk = r10_bio->devs[slot].devnum;
 537         while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 538                r10_bio->devs[slot].bio == IO_BLOCKED ||
 539                !test_bit(In_sync, &rdev->flags)) {
 540                 slot ++;
 541                 if (slot == conf->copies) {
 542                         disk = -1;
 543                         goto rb_out;
 544                 }
 545                 disk = r10_bio->devs[slot].devnum;
 546         }
 547
 548
 549         current_distance = abs(r10_bio->devs[slot].addr -
 550                                conf->mirrors[disk].head_position);
 551
 552         /* Find the disk whose head is closest,
 553          * or - for far > 1 - find the closest to partition beginning */
 554
 555         for (nslot = slot; nslot < conf->copies; nslot++) {
 556                 int ndisk = r10_bio->devs[nslot].devnum;
 557
 558
 559                 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
 560                     r10_bio->devs[nslot].bio == IO_BLOCKED ||
 561                     !test_bit(In_sync, &rdev->flags))
 562                         continue;
 563
 564                 /* This optimisation is debatable, and completely destroys
 565                  * sequential read speed for 'far copies' arrays.  So only
 566                  * keep it for 'near' arrays, and review those later.
 567                  */
 568                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
 569                         disk = ndisk;
 570                         slot = nslot;
 571                         break;
 572                 }
 573
 574                 /* for far > 1 always use the lowest address */
 575                 if (conf->far_copies > 1)
 576                         new_distance = r10_bio->devs[nslot].addr;
 577                 else
 578                         new_distance = abs(r10_bio->devs[nslot].addr -
 579                                            conf->mirrors[ndisk].head_position);
 580                 if (new_distance < current_distance) {
 581                         current_distance = new_distance;
 582                         disk = ndisk;
 583                         slot = nslot;
 584                 }
 585         }
 586
 587 rb_out:
 588         r10_bio->read_slot = slot;
 589 /*      conf->next_seq_sect = this_sector + sectors;*/
 590
 591         if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
 592                 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
 593         else
 594                 disk = -1;
 595         rcu_read_unlock();
 596
 597         return disk;
 598 }
 599
 600 static void unplug_slaves(mddev_t *mddev)
 601 {
 602         conf_t *conf = mddev->private;
 603         int i;
 604
 605         rcu_read_lock();
 606         for (i=0; i < conf->raid_disks; i++) {
 607                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 608                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 609                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 610
 611                         atomic_inc(&rdev->nr_pending);
 612                         rcu_read_unlock();
 613
 614                         blk_unplug(r_queue);
 615
 616                         rdev_dec_pending(rdev, mddev);
 617                         rcu_read_lock();
 618                 }
 619         }
 620         rcu_read_unlock();
 621 }
 622
 623 static void raid10_unplug(struct request_queue *q)
 624 {
 625         mddev_t *mddev = q->queuedata;
 626
 627         unplug_slaves(q->queuedata);
 628         md_wakeup_thread(mddev->thread);
 629 }
 630
 631 static int raid10_congested(void *data, int bits)
 632 {
 633         mddev_t *mddev = data;
 634         conf_t *conf = mddev->private;
 635         int i, ret = 0;
 636
 637         if (mddev_congested(mddev, bits))
 638                 return 1;
 639         rcu_read_lock();
 640         for (i = 0; i < conf->raid_disks && ret == 0; i++) {
 641                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 642                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 643                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 644
 645                         ret |= bdi_congested(&q->backing_dev_info, bits);
 646                 }
 647         }
 648         rcu_read_unlock();
 649         return ret;
 650 }
 651
 652 static int flush_pending_writes(conf_t *conf)
 653 {
 654         /* Any writes that have been queued but are awaiting
 655          * bitmap updates get flushed here.
 656          * We return 1 if any requests were actually submitted.
 657          */
 658         int rv = 0;
 659
 660         spin_lock_irq(&conf->device_lock);
 661
 662         if (conf->pending_bio_list.head) {
 663                 struct bio *bio;
 664                 bio = bio_list_get(&conf->pending_bio_list);
 665                 blk_remove_plug(conf->mddev->queue);
 666                 spin_unlock_irq(&conf->device_lock);
 667                 /* flush any pending bitmap writes to disk
 668                  * before proceeding w/ I/O */
 669                 bitmap_unplug(conf->mddev->bitmap);
 670
 671                 while (bio) { /* submit pending writes */
 672                         struct bio *next = bio->bi_next;
 673                         bio->bi_next = NULL;
 674                         generic_make_request(bio);
 675                         bio = next;
 676                 }
 677                 rv = 1;
 678         } else
 679                 spin_unlock_irq(&conf->device_lock);
 680         return rv;
 681 }
 682 /* Barriers....
 683  * Sometimes we need to suspend IO while we do something else,
 684  * either some resync/recovery, or reconfigure the array.
 685  * To do this we raise a 'barrier'.
 686  * The 'barrier' is a counter that can be raised multiple times
 687  * to count how many activities are happening which preclude
 688  * normal IO.
 689  * We can only raise the barrier if there is no pending IO.
 690  * i.e. if nr_pending == 0.
 691  * We choose only to raise the barrier if no-one is waiting for the
 692  * barrier to go down.  This means that as soon as an IO request
 693  * is ready, no other operations which require a barrier will start
 694  * until the IO request has had a chance.
 695  *
 696  * So: regular IO calls 'wait_barrier'.  When that returns there
 697  *    is no backgroup IO happening,  It must arrange to call
 698  *    allow_barrier when it has finished its IO.
 699  * backgroup IO calls must call raise_barrier.  Once that returns
 700  *    there is no normal IO happeing.  It must arrange to call
 701  *    lower_barrier when the particular background IO completes.
 702  */
 703
 704 static void raise_barrier(conf_t *conf, int force)
 705 {
 706         BUG_ON(force && !conf->barrier);
 707         spin_lock_irq(&conf->resync_lock);
 708
 709         /* Wait until no block IO is waiting (unless 'force') */
 710         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 711                             conf->resync_lock,
 712                             raid10_unplug(conf->mddev->queue));
 713
 714         /* block any new IO from starting */
 715         conf->barrier++;
 716
 717         /* No wait for all pending IO to complete */
 718         wait_event_lock_irq(conf->wait_barrier,
 719                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 720                             conf->resync_lock,
 721                             raid10_unplug(conf->mddev->queue));
 722
 723         spin_unlock_irq(&conf->resync_lock);
 724 }
 725
 726 static void lower_barrier(conf_t *conf)
 727 {
 728         unsigned long flags;
 729         spin_lock_irqsave(&conf->resync_lock, flags);
 730         conf->barrier--;
 731         spin_unlock_irqrestore(&conf->resync_lock, flags);
 732         wake_up(&conf->wait_barrier);
 733 }
 734
 735 static void wait_barrier(conf_t *conf)
 736 {
 737         spin_lock_irq(&conf->resync_lock);
 738         if (conf->barrier) {
 739                 conf->nr_waiting++;
 740                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 741                                     conf->resync_lock,
 742                                     raid10_unplug(conf->mddev->queue));
 743                 conf->nr_waiting--;
 744         }
 745         conf->nr_pending++;
 746         spin_unlock_irq(&conf->resync_lock);
 747 }
 748
 749 static void allow_barrier(conf_t *conf)
 750 {
 751         unsigned long flags;
 752         spin_lock_irqsave(&conf->resync_lock, flags);
 753         conf->nr_pending--;
 754         spin_unlock_irqrestore(&conf->resync_lock, flags);
 755         wake_up(&conf->wait_barrier);
 756 }
 757
 758 static void freeze_array(conf_t *conf)
 759 {
 760         /* stop syncio and normal IO and wait for everything to
 761          * go quiet.
 762          * We increment barrier and nr_waiting, and then
 763          * wait until nr_pending match nr_queued+1
 764          * This is called in the context of one normal IO request
 765          * that has failed. Thus any sync request that might be pending
 766          * will be blocked by nr_pending, and we need to wait for
 767          * pending IO requests to complete or be queued for re-try.
 768          * Thus the number queued (nr_queued) plus this request (1)
 769          * must match the number of pending IOs (nr_pending) before
 770          * we continue.
 771          */
 772         spin_lock_irq(&conf->resync_lock);
 773         conf->barrier++;
 774         conf->nr_waiting++;
 775         wait_event_lock_irq(conf->wait_barrier,
 776                             conf->nr_pending == conf->nr_queued+1,
 777                             conf->resync_lock,
 778                             ({ flush_pending_writes(conf);
 779                                raid10_unplug(conf->mddev->queue); }));
 780         spin_unlock_irq(&conf->resync_lock);
 781 }
 782
 783 static void unfreeze_array(conf_t *conf)
 784 {
 785         /* reverse the effect of the freeze */
 786         spin_lock_irq(&conf->resync_lock);
 787         conf->barrier--;
 788         conf->nr_waiting--;
 789         wake_up(&conf->wait_barrier);
 790         spin_unlock_irq(&conf->resync_lock);
 791 }
 792
 793 static int make_request(mddev_t *mddev, struct bio * bio)
 794 {
 795         conf_t *conf = mddev->private;
 796         mirror_info_t *mirror;
 797         r10bio_t *r10_bio;
 798         struct bio *read_bio;
 799         int i;
 800         int chunk_sects = conf->chunk_mask + 1;
 801         const int rw = bio_data_dir(bio);
 802         const bool do_sync = bio_rw_flagged(bio, BIO_RW_SYNCIO);
 803         struct bio_list bl;
 804         unsigned long flags;
 805         mdk_rdev_t *blocked_rdev;
 806
 807         if (unlikely(bio_rw_flagged(bio, BIO_RW_BARRIER))) {
 808                 md_barrier_request(mddev, bio);
 809                 return 0;
 810         }
 811
 812         /* If this request crosses a chunk boundary, we need to
 813          * split it.  This will only happen for 1 PAGE (or less) requests.
 814          */
 815         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 816                       > chunk_sects &&
 817                     conf->near_copies < conf->raid_disks)) {
 818                 struct bio_pair *bp;
 819                 /* Sanity check -- queue functions should prevent this happening */
 820                 if (bio->bi_vcnt != 1 ||
 821                     bio->bi_idx != 0)
 822                         goto bad_map;
 823                 /* This is a one page bio that upper layers
 824                  * refuse to split for us, so we need to split it.
 825                  */
 826                 bp = bio_split(bio,
 827                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 828                 if (make_request(mddev, &bp->bio1))
 829                         generic_make_request(&bp->bio1);
 830                 if (make_request(mddev, &bp->bio2))
 831                         generic_make_request(&bp->bio2);
 832
 833                 bio_pair_release(bp);
 834                 return 0;
 835         bad_map:
 836                 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
 837                        " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
 838                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 839
 840                 bio_io_error(bio);
 841                 return 0;
 842         }
 843
 844         md_write_start(mddev, bio);
 845
 846         /*
 847          * Register the new request and wait if the reconstruction
 848          * thread has put up a bar for new requests.
 849          * Continue immediately if no resync is active currently.
 850          */
 851         wait_barrier(conf);
 852
 853         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 854
 855         r10_bio->master_bio = bio;
 856         r10_bio->sectors = bio->bi_size >> 9;
 857
 858         r10_bio->mddev = mddev;
 859         r10_bio->sector = bio->bi_sector;
 860         r10_bio->state = 0;
 861
 862         if (rw == READ) {
 863                 /*
 864                  * read balancing logic:
 865                  */
 866                 int disk = read_balance(conf, r10_bio);
 867                 int slot = r10_bio->read_slot;
 868                 if (disk < 0) {
 869                         raid_end_bio_io(r10_bio);
 870                         return 0;
 871                 }
 872                 mirror = conf->mirrors + disk;
 873
 874                 read_bio = bio_clone(bio, GFP_NOIO);
 875
 876                 r10_bio->devs[slot].bio = read_bio;
 877
 878                 read_bio->bi_sector = r10_bio->devs[slot].addr +
 879                         mirror->rdev->data_offset;
 880                 read_bio->bi_bdev = mirror->rdev->bdev;
 881                 read_bio->bi_end_io = raid10_end_read_request;
 882                 read_bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
 883                 read_bio->bi_private = r10_bio;
 884
 885                 generic_make_request(read_bio);
 886                 return 0;
 887         }
 888
 889         /*
 890          * WRITE:
 891          */
 892         /* first select target devices under rcu_lock and
 893          * inc refcount on their rdev.  Record them by setting
 894          * bios[x] to bio
 895          */
 896         raid10_find_phys(conf, r10_bio);
 897  retry_write:
 898         blocked_rdev = NULL;
 899         rcu_read_lock();
 900         for (i = 0;  i < conf->copies; i++) {
 901                 int d = r10_bio->devs[i].devnum;
 902                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
 903                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 904                         atomic_inc(&rdev->nr_pending);
 905                         blocked_rdev = rdev;
 906                         break;
 907                 }
 908                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 909                         atomic_inc(&rdev->nr_pending);
 910                         r10_bio->devs[i].bio = bio;
 911                 } else {
 912                         r10_bio->devs[i].bio = NULL;
 913                         set_bit(R10BIO_Degraded, &r10_bio->state);
 914                 }
 915         }
 916         rcu_read_unlock();
 917
 918         if (unlikely(blocked_rdev)) {
 919                 /* Have to wait for this device to get unblocked, then retry */
 920                 int j;
 921                 int d;
 922
 923                 for (j = 0; j < i; j++)
 924                         if (r10_bio->devs[j].bio) {
 925                                 d = r10_bio->devs[j].devnum;
 926                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
 927                         }
 928                 allow_barrier(conf);
 929                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 930                 wait_barrier(conf);
 931                 goto retry_write;
 932         }
 933
 934         atomic_set(&r10_bio->remaining, 0);
 935
 936         bio_list_init(&bl);
 937         for (i = 0; i < conf->copies; i++) {
 938                 struct bio *mbio;
 939                 int d = r10_bio->devs[i].devnum;
 940                 if (!r10_bio->devs[i].bio)
 941                         continue;
 942
 943                 mbio = bio_clone(bio, GFP_NOIO);
 944                 r10_bio->devs[i].bio = mbio;
 945
 946                 mbio->bi_sector = r10_bio->devs[i].addr+
 947                         conf->mirrors[d].rdev->data_offset;
 948                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
 949                 mbio->bi_end_io = raid10_end_write_request;
 950                 mbio->bi_rw = WRITE | (do_sync << BIO_RW_SYNCIO);
 951                 mbio->bi_private = r10_bio;
 952
 953                 atomic_inc(&r10_bio->remaining);
 954                 bio_list_add(&bl, mbio);
 955         }
 956
 957         if (unlikely(!atomic_read(&r10_bio->remaining))) {
 958                 /* the array is dead */
 959                 md_write_end(mddev);
 960                 raid_end_bio_io(r10_bio);
 961                 return 0;
 962         }
 963
 964         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
 965         spin_lock_irqsave(&conf->device_lock, flags);
 966         bio_list_merge(&conf->pending_bio_list, &bl);
 967         blk_plug_device(mddev->queue);
 968         spin_unlock_irqrestore(&conf->device_lock, flags);
 969
 970         /* In case raid10d snuck in to freeze_array */
 971         wake_up(&conf->wait_barrier);
 972
 973         if (do_sync)
 974                 md_wakeup_thread(mddev->thread);
 975
 976         return 0;
 977 }
 978
 979 static void status(struct seq_file *seq, mddev_t *mddev)
 980 {
 981         conf_t *conf = mddev->private;
 982         int i;
 983
 984         if (conf->near_copies < conf->raid_disks)
 985                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
 986         if (conf->near_copies > 1)
 987                 seq_printf(seq, " %d near-copies", conf->near_copies);
 988         if (conf->far_copies > 1) {
 989                 if (conf->far_offset)
 990                         seq_printf(seq, " %d offset-copies", conf->far_copies);
 991                 else
 992                         seq_printf(seq, " %d far-copies", conf->far_copies);
 993         }
 994         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 995                                         conf->raid_disks - mddev->degraded);
 996         for (i = 0; i < conf->raid_disks; i++)
 997                 seq_printf(seq, "%s",
 998                               conf->mirrors[i].rdev &&
 999                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1000         seq_printf(seq, "]");
1001 }
1002
1003 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1004 {
1005         char b[BDEVNAME_SIZE];
1006         conf_t *conf = mddev->private;
1007
1008         /*
1009          * If it is not operational, then we have already marked it as dead
1010          * else if it is the last working disks, ignore the error, let the
1011          * next level up know.
1012          * else mark the drive as failed
1013          */
1014         if (test_bit(In_sync, &rdev->flags)
1015             && conf->raid_disks-mddev->degraded == 1)
1016                 /*
1017                  * Don't fail the drive, just return an IO error.
1018                  * The test should really be more sophisticated than
1019                  * "working_disks == 1", but it isn't critical, and
1020                  * can wait until we do more sophisticated "is the drive
1021                  * really dead" tests...
1022                  */
1023                 return;
1024         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1025                 unsigned long flags;
1026                 spin_lock_irqsave(&conf->device_lock, flags);
1027                 mddev->degraded++;
1028                 spin_unlock_irqrestore(&conf->device_lock, flags);
1029                 /*
1030                  * if recovery is running, make sure it aborts.
1031                  */
1032                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1033         }
1034         set_bit(Faulty, &rdev->flags);
1035         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1036         printk(KERN_ALERT "md/raid10:%s: Disk failure on %s, disabling device.\n"
1037                KERN_ALERT "md/raid10:%s: Operation continuing on %d devices.\n",
1038                mdname(mddev), bdevname(rdev->bdev, b),
1039                mdname(mddev), conf->raid_disks - mddev->degraded);
1040 }
1041
1042 static void print_conf(conf_t *conf)
1043 {
1044         int i;
1045         mirror_info_t *tmp;
1046
1047         printk(KERN_DEBUG "RAID10 conf printout:\n");
1048         if (!conf) {
1049                 printk(KERN_DEBUG "(!conf)\n");
1050                 return;
1051         }
1052         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1053                 conf->raid_disks);
1054
1055         for (i = 0; i < conf->raid_disks; i++) {
1056                 char b[BDEVNAME_SIZE];
1057                 tmp = conf->mirrors + i;
1058                 if (tmp->rdev)
1059                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1060                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1061                                 !test_bit(Faulty, &tmp->rdev->flags),
1062                                 bdevname(tmp->rdev->bdev,b));
1063         }
1064 }
1065
1066 static void close_sync(conf_t *conf)
1067 {
1068         wait_barrier(conf);
1069         allow_barrier(conf);
1070
1071         mempool_destroy(conf->r10buf_pool);
1072         conf->r10buf_pool = NULL;
1073 }
1074
1075 /* check if there are enough drives for
1076  * every block to appear on atleast one
1077  */
1078 static int enough(conf_t *conf)
1079 {
1080         int first = 0;
1081
1082         do {
1083                 int n = conf->copies;
1084                 int cnt = 0;
1085                 while (n--) {
1086                         if (conf->mirrors[first].rdev)
1087                                 cnt++;
1088                         first = (first+1) % conf->raid_disks;
1089                 }
1090                 if (cnt == 0)
1091                         return 0;
1092         } while (first != 0);
1093         return 1;
1094 }
1095
1096 static int raid10_spare_active(mddev_t *mddev)
1097 {
1098         int i;
1099         conf_t *conf = mddev->private;
1100         mirror_info_t *tmp;
1101
1102         /*
1103          * Find all non-in_sync disks within the RAID10 configuration
1104          * and mark them in_sync
1105          */
1106         for (i = 0; i < conf->raid_disks; i++) {
1107                 tmp = conf->mirrors + i;
1108                 if (tmp->rdev
1109                     && !test_bit(Faulty, &tmp->rdev->flags)
1110                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1111                         unsigned long flags;
1112                         spin_lock_irqsave(&conf->device_lock, flags);
1113                         mddev->degraded--;
1114                         spin_unlock_irqrestore(&conf->device_lock, flags);
1115                 }
1116         }
1117
1118         print_conf(conf);
1119         return 0;
1120 }
1121
1122
1123 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1124 {
1125         conf_t *conf = mddev->private;
1126         int err = -EEXIST;
1127         int mirror;
1128         mirror_info_t *p;
1129         int first = 0;
1130         int last = conf->raid_disks - 1;
1131
1132         if (mddev->recovery_cp < MaxSector)
1133                 /* only hot-add to in-sync arrays, as recovery is
1134                  * very different from resync
1135                  */
1136                 return -EBUSY;
1137         if (!enough(conf))
1138                 return -EINVAL;
1139
1140         if (rdev->raid_disk >= 0)
1141                 first = last = rdev->raid_disk;
1142
1143         if (rdev->saved_raid_disk >= 0 &&
1144             rdev->saved_raid_disk >= first &&
1145             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1146                 mirror = rdev->saved_raid_disk;
1147         else
1148                 mirror = first;
1149         for ( ; mirror <= last ; mirror++)
1150                 if ( !(p=conf->mirrors+mirror)->rdev) {
1151
1152                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1153                                           rdev->data_offset << 9);
1154                         /* as we don't honour merge_bvec_fn, we must
1155                          * never risk violating it, so limit
1156                          * ->max_segments to one lying with a single
1157                          * page, as a one page request is never in
1158                          * violation.
1159                          */
1160                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1161                                 blk_queue_max_segments(mddev->queue, 1);
1162                                 blk_queue_segment_boundary(mddev->queue,
1163                                                            PAGE_CACHE_SIZE - 1);
1164                         }
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 = 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 = 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  * Used by fix_read_error() to decay the per rdev read_errors.
1436  * We halve the read error count for every hour that has elapsed
1437  * since the last recorded read error.
1438  *
1439  */
1440 static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1441 {
1442         struct timespec cur_time_mon;
1443         unsigned long hours_since_last;
1444         unsigned int read_errors = atomic_read(&rdev->read_errors);
1445
1446         ktime_get_ts(&cur_time_mon);
1447
1448         if (rdev->last_read_error.tv_sec == 0 &&
1449             rdev->last_read_error.tv_nsec == 0) {
1450                 /* first time we've seen a read error */
1451                 rdev->last_read_error = cur_time_mon;
1452                 return;
1453         }
1454
1455         hours_since_last = (cur_time_mon.tv_sec -
1456                             rdev->last_read_error.tv_sec) / 3600;
1457
1458         rdev->last_read_error = cur_time_mon;
1459
1460         /*
1461          * if hours_since_last is > the number of bits in read_errors
1462          * just set read errors to 0. We do this to avoid
1463          * overflowing the shift of read_errors by hours_since_last.
1464          */
1465         if (hours_since_last >= 8 * sizeof(read_errors))
1466                 atomic_set(&rdev->read_errors, 0);
1467         else
1468                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1469 }
1470
1471 /*
1472  * This is a kernel thread which:
1473  *
1474  *      1.      Retries failed read operations on working mirrors.
1475  *      2.      Updates the raid superblock when problems encounter.
1476  *      3.      Performs writes following reads for array synchronising.
1477  */
1478
1479 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1480 {
1481         int sect = 0; /* Offset from r10_bio->sector */
1482         int sectors = r10_bio->sectors;
1483         mdk_rdev_t*rdev;
1484         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1485
1486         rcu_read_lock();
1487         {
1488                 int d = r10_bio->devs[r10_bio->read_slot].devnum;
1489                 char b[BDEVNAME_SIZE];
1490                 int cur_read_error_count = 0;
1491
1492                 rdev = rcu_dereference(conf->mirrors[d].rdev);
1493                 bdevname(rdev->bdev, b);
1494
1495                 if (test_bit(Faulty, &rdev->flags)) {
1496                         rcu_read_unlock();
1497                         /* drive has already been failed, just ignore any
1498                            more fix_read_error() attempts */
1499                         return;
1500                 }
1501
1502                 check_decay_read_errors(mddev, rdev);
1503                 atomic_inc(&rdev->read_errors);
1504                 cur_read_error_count = atomic_read(&rdev->read_errors);
1505                 if (cur_read_error_count > max_read_errors) {
1506                         rcu_read_unlock();
1507                         printk(KERN_NOTICE
1508                                "md/raid10:%s: %s: Raid device exceeded "
1509                                "read_error threshold "
1510                                "[cur %d:max %d]\n",
1511                                mdname(mddev),
1512                                b, cur_read_error_count, max_read_errors);
1513                         printk(KERN_NOTICE
1514                                "md/raid10:%s: %s: Failing raid "
1515                                "device\n", mdname(mddev), b);
1516                         md_error(mddev, conf->mirrors[d].rdev);
1517                         return;
1518                 }
1519         }
1520         rcu_read_unlock();
1521
1522         while(sectors) {
1523                 int s = sectors;
1524                 int sl = r10_bio->read_slot;
1525                 int success = 0;
1526                 int start;
1527
1528                 if (s > (PAGE_SIZE>>9))
1529                         s = PAGE_SIZE >> 9;
1530
1531                 rcu_read_lock();
1532                 do {
1533                         int d = r10_bio->devs[sl].devnum;
1534                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1535                         if (rdev &&
1536                             test_bit(In_sync, &rdev->flags)) {
1537                                 atomic_inc(&rdev->nr_pending);
1538                                 rcu_read_unlock();
1539                                 success = sync_page_io(rdev->bdev,
1540                                                        r10_bio->devs[sl].addr +
1541                                                        sect + rdev->data_offset,
1542                                                        s<<9,
1543                                                        conf->tmppage, READ);
1544                                 rdev_dec_pending(rdev, mddev);
1545                                 rcu_read_lock();
1546                                 if (success)
1547                                         break;
1548                         }
1549                         sl++;
1550                         if (sl == conf->copies)
1551                                 sl = 0;
1552                 } while (!success && sl != r10_bio->read_slot);
1553                 rcu_read_unlock();
1554
1555                 if (!success) {
1556                         /* Cannot read from anywhere -- bye bye array */
1557                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1558                         md_error(mddev, conf->mirrors[dn].rdev);
1559                         break;
1560                 }
1561
1562                 start = sl;
1563                 /* write it back and re-read */
1564                 rcu_read_lock();
1565                 while (sl != r10_bio->read_slot) {
1566                         char b[BDEVNAME_SIZE];
1567                         int d;
1568                         if (sl==0)
1569                                 sl = conf->copies;
1570                         sl--;
1571                         d = r10_bio->devs[sl].devnum;
1572                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1573                         if (rdev &&
1574                             test_bit(In_sync, &rdev->flags)) {
1575                                 atomic_inc(&rdev->nr_pending);
1576                                 rcu_read_unlock();
1577                                 atomic_add(s, &rdev->corrected_errors);
1578                                 if (sync_page_io(rdev->bdev,
1579                                                  r10_bio->devs[sl].addr +
1580                                                  sect + rdev->data_offset,
1581                                                  s<<9, conf->tmppage, WRITE)
1582                                     == 0) {
1583                                         /* Well, this device is dead */
1584                                         printk(KERN_NOTICE
1585                                                "md/raid10:%s: read correction "
1586                                                "write failed"
1587                                                " (%d sectors at %llu on %s)\n",
1588                                                mdname(mddev), s,
1589                                                (unsigned long long)(sect+
1590                                                rdev->data_offset),
1591                                                bdevname(rdev->bdev, b));
1592                                         printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1593                                                "drive\n",
1594                                                mdname(mddev),
1595                                                bdevname(rdev->bdev, b));
1596                                         md_error(mddev, rdev);
1597                                 }
1598                                 rdev_dec_pending(rdev, mddev);
1599                                 rcu_read_lock();
1600                         }
1601                 }
1602                 sl = start;
1603                 while (sl != r10_bio->read_slot) {
1604                         int d;
1605                         if (sl==0)
1606                                 sl = conf->copies;
1607                         sl--;
1608                         d = r10_bio->devs[sl].devnum;
1609                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1610                         if (rdev &&
1611                             test_bit(In_sync, &rdev->flags)) {
1612                                 char b[BDEVNAME_SIZE];
1613                                 atomic_inc(&rdev->nr_pending);
1614                                 rcu_read_unlock();
1615                                 if (sync_page_io(rdev->bdev,
1616                                                  r10_bio->devs[sl].addr +
1617                                                  sect + rdev->data_offset,
1618                                                  s<<9, conf->tmppage,
1619                                                  READ) == 0) {
1620                                         /* Well, this device is dead */
1621                                         printk(KERN_NOTICE
1622                                                "md/raid10:%s: unable to read back "
1623                                                "corrected sectors"
1624                                                " (%d sectors at %llu on %s)\n",
1625                                                mdname(mddev), s,
1626                                                (unsigned long long)(sect+
1627                                                     rdev->data_offset),
1628                                                bdevname(rdev->bdev, b));
1629                                         printk(KERN_NOTICE "md/raid10:%s: %s: failing drive\n",
1630                                                mdname(mddev),
1631                                                bdevname(rdev->bdev, b));
1632
1633                                         md_error(mddev, rdev);
1634                                 } else {
1635                                         printk(KERN_INFO
1636                                                "md/raid10:%s: read error corrected"
1637                                                " (%d sectors at %llu on %s)\n",
1638                                                mdname(mddev), s,
1639                                                (unsigned long long)(sect+
1640                                                     rdev->data_offset),
1641                                                bdevname(rdev->bdev, b));
1642                                 }
1643
1644                                 rdev_dec_pending(rdev, mddev);
1645                                 rcu_read_lock();
1646                         }
1647                 }
1648                 rcu_read_unlock();
1649
1650                 sectors -= s;
1651                 sect += s;
1652         }
1653 }
1654
1655 static void raid10d(mddev_t *mddev)
1656 {
1657         r10bio_t *r10_bio;
1658         struct bio *bio;
1659         unsigned long flags;
1660         conf_t *conf = mddev->private;
1661         struct list_head *head = &conf->retry_list;
1662         int unplug=0;
1663         mdk_rdev_t *rdev;
1664
1665         md_check_recovery(mddev);
1666
1667         for (;;) {
1668                 char b[BDEVNAME_SIZE];
1669
1670                 unplug += flush_pending_writes(conf);
1671
1672                 spin_lock_irqsave(&conf->device_lock, flags);
1673                 if (list_empty(head)) {
1674                         spin_unlock_irqrestore(&conf->device_lock, flags);
1675                         break;
1676                 }
1677                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1678                 list_del(head->prev);
1679                 conf->nr_queued--;
1680                 spin_unlock_irqrestore(&conf->device_lock, flags);
1681
1682                 mddev = r10_bio->mddev;
1683                 conf = mddev->private;
1684                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1685                         sync_request_write(mddev, r10_bio);
1686                         unplug = 1;
1687                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1688                         recovery_request_write(mddev, r10_bio);
1689                         unplug = 1;
1690                 } else {
1691                         int mirror;
1692                         /* we got a read error. Maybe the drive is bad.  Maybe just
1693                          * the block and we can fix it.
1694                          * We freeze all other IO, and try reading the block from
1695                          * other devices.  When we find one, we re-write
1696                          * and check it that fixes the read error.
1697                          * This is all done synchronously while the array is
1698                          * frozen.
1699                          */
1700                         if (mddev->ro == 0) {
1701                                 freeze_array(conf);
1702                                 fix_read_error(conf, mddev, r10_bio);
1703                                 unfreeze_array(conf);
1704                         }
1705
1706                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1707                         r10_bio->devs[r10_bio->read_slot].bio =
1708                                 mddev->ro ? IO_BLOCKED : NULL;
1709                         mirror = read_balance(conf, r10_bio);
1710                         if (mirror == -1) {
1711                                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
1712                                        " read error for block %llu\n",
1713                                        mdname(mddev),
1714                                        bdevname(bio->bi_bdev,b),
1715                                        (unsigned long long)r10_bio->sector);
1716                                 raid_end_bio_io(r10_bio);
1717                                 bio_put(bio);
1718                         } else {
1719                                 const bool do_sync = bio_rw_flagged(r10_bio->master_bio, BIO_RW_SYNCIO);
1720                                 bio_put(bio);
1721                                 rdev = conf->mirrors[mirror].rdev;
1722                                 if (printk_ratelimit())
1723                                         printk(KERN_ERR "md/raid10:%s: %s: redirecting sector %llu to"
1724                                                " another mirror\n",
1725                                                mdname(mddev),
1726                                                bdevname(rdev->bdev,b),
1727                                                (unsigned long long)r10_bio->sector);
1728                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1729                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1730                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1731                                         + rdev->data_offset;
1732                                 bio->bi_bdev = rdev->bdev;
1733                                 bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
1734                                 bio->bi_private = r10_bio;
1735                                 bio->bi_end_io = raid10_end_read_request;
1736                                 unplug = 1;
1737                                 generic_make_request(bio);
1738                         }
1739                 }
1740                 cond_resched();
1741         }
1742         if (unplug)
1743                 unplug_slaves(mddev);
1744 }
1745
1746
1747 static int init_resync(conf_t *conf)
1748 {
1749         int buffs;
1750
1751         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1752         BUG_ON(conf->r10buf_pool);
1753         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1754         if (!conf->r10buf_pool)
1755                 return -ENOMEM;
1756         conf->next_resync = 0;
1757         return 0;
1758 }
1759
1760 /*
1761  * perform a "sync" on one "block"
1762  *
1763  * We need to make sure that no normal I/O request - particularly write
1764  * requests - conflict with active sync requests.
1765  *
1766  * This is achieved by tracking pending requests and a 'barrier' concept
1767  * that can be installed to exclude normal IO requests.
1768  *
1769  * Resync and recovery are handled very differently.
1770  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1771  *
1772  * For resync, we iterate over virtual addresses, read all copies,
1773  * and update if there are differences.  If only one copy is live,
1774  * skip it.
1775  * For recovery, we iterate over physical addresses, read a good
1776  * value for each non-in_sync drive, and over-write.
1777  *
1778  * So, for recovery we may have several outstanding complex requests for a
1779  * given address, one for each out-of-sync device.  We model this by allocating
1780  * a number of r10_bio structures, one for each out-of-sync device.
1781  * As we setup these structures, we collect all bio's together into a list
1782  * which we then process collectively to add pages, and then process again
1783  * to pass to generic_make_request.
1784  *
1785  * The r10_bio structures are linked using a borrowed master_bio pointer.
1786  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1787  * has its remaining count decremented to 0, the whole complex operation
1788  * is complete.
1789  *
1790  */
1791
1792 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1793 {
1794         conf_t *conf = mddev->private;
1795         r10bio_t *r10_bio;
1796         struct bio *biolist = NULL, *bio;
1797         sector_t max_sector, nr_sectors;
1798         int disk;
1799         int i;
1800         int max_sync;
1801         int sync_blocks;
1802
1803         sector_t sectors_skipped = 0;
1804         int chunks_skipped = 0;
1805
1806         if (!conf->r10buf_pool)
1807                 if (init_resync(conf))
1808                         return 0;
1809
1810  skipped:
1811         max_sector = mddev->dev_sectors;
1812         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1813                 max_sector = mddev->resync_max_sectors;
1814         if (sector_nr >= max_sector) {
1815                 /* If we aborted, we need to abort the
1816                  * sync on the 'current' bitmap chucks (there can
1817                  * be several when recovering multiple devices).
1818                  * as we may have started syncing it but not finished.
1819                  * We can find the current address in
1820                  * mddev->curr_resync, but for recovery,
1821                  * we need to convert that to several
1822                  * virtual addresses.
1823                  */
1824                 if (mddev->curr_resync < max_sector) { /* aborted */
1825                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1826                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1827                                                 &sync_blocks, 1);
1828                         else for (i=0; i<conf->raid_disks; i++) {
1829                                 sector_t sect =
1830                                         raid10_find_virt(conf, mddev->curr_resync, i);
1831                                 bitmap_end_sync(mddev->bitmap, sect,
1832                                                 &sync_blocks, 1);
1833                         }
1834                 } else /* completed sync */
1835                         conf->fullsync = 0;
1836
1837                 bitmap_close_sync(mddev->bitmap);
1838                 close_sync(conf);
1839                 *skipped = 1;
1840                 return sectors_skipped;
1841         }
1842         if (chunks_skipped >= conf->raid_disks) {
1843                 /* if there has been nothing to do on any drive,
1844                  * then there is nothing to do at all..
1845                  */
1846                 *skipped = 1;
1847                 return (max_sector - sector_nr) + sectors_skipped;
1848         }
1849
1850         if (max_sector > mddev->resync_max)
1851                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1852
1853         /* make sure whole request will fit in a chunk - if chunks
1854          * are meaningful
1855          */
1856         if (conf->near_copies < conf->raid_disks &&
1857             max_sector > (sector_nr | conf->chunk_mask))
1858                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1859         /*
1860          * If there is non-resync activity waiting for us then
1861          * put in a delay to throttle resync.
1862          */
1863         if (!go_faster && conf->nr_waiting)
1864                 msleep_interruptible(1000);
1865
1866         /* Again, very different code for resync and recovery.
1867          * Both must result in an r10bio with a list of bios that
1868          * have bi_end_io, bi_sector, bi_bdev set,
1869          * and bi_private set to the r10bio.
1870          * For recovery, we may actually create several r10bios
1871          * with 2 bios in each, that correspond to the bios in the main one.
1872          * In this case, the subordinate r10bios link back through a
1873          * borrowed master_bio pointer, and the counter in the master
1874          * includes a ref from each subordinate.
1875          */
1876         /* First, we decide what to do and set ->bi_end_io
1877          * To end_sync_read if we want to read, and
1878          * end_sync_write if we will want to write.
1879          */
1880
1881         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1882         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1883                 /* recovery... the complicated one */
1884                 int j, k;
1885                 r10_bio = NULL;
1886
1887                 for (i=0 ; i<conf->raid_disks; i++)
1888                         if (conf->mirrors[i].rdev &&
1889                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1890                                 int still_degraded = 0;
1891                                 /* want to reconstruct this device */
1892                                 r10bio_t *rb2 = r10_bio;
1893                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1894                                 int must_sync;
1895                                 /* Unless we are doing a full sync, we only need
1896                                  * to recover the block if it is set in the bitmap
1897                                  */
1898                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1899                                                               &sync_blocks, 1);
1900                                 if (sync_blocks < max_sync)
1901                                         max_sync = sync_blocks;
1902                                 if (!must_sync &&
1903                                     !conf->fullsync) {
1904                                         /* yep, skip the sync_blocks here, but don't assume
1905                                          * that there will never be anything to do here
1906                                          */
1907                                         chunks_skipped = -1;
1908                                         continue;
1909                                 }
1910
1911                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1912                                 raise_barrier(conf, rb2 != NULL);
1913                                 atomic_set(&r10_bio->remaining, 0);
1914
1915                                 r10_bio->master_bio = (struct bio*)rb2;
1916                                 if (rb2)
1917                                         atomic_inc(&rb2->remaining);
1918                                 r10_bio->mddev = mddev;
1919                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1920                                 r10_bio->sector = sect;
1921
1922                                 raid10_find_phys(conf, r10_bio);
1923
1924                                 /* Need to check if the array will still be
1925                                  * degraded
1926                                  */
1927                                 for (j=0; j<conf->raid_disks; j++)
1928                                         if (conf->mirrors[j].rdev == NULL ||
1929                                             test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
1930                                                 still_degraded = 1;
1931                                                 break;
1932                                         }
1933
1934                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1935                                                               &sync_blocks, still_degraded);
1936
1937                                 for (j=0; j<conf->copies;j++) {
1938                                         int d = r10_bio->devs[j].devnum;
1939                                         if (conf->mirrors[d].rdev &&
1940                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1941                                                 /* This is where we read from */
1942                                                 bio = r10_bio->devs[0].bio;
1943                                                 bio->bi_next = biolist;
1944                                                 biolist = bio;
1945                                                 bio->bi_private = r10_bio;
1946                                                 bio->bi_end_io = end_sync_read;
1947                                                 bio->bi_rw = READ;
1948                                                 bio->bi_sector = r10_bio->devs[j].addr +
1949                                                         conf->mirrors[d].rdev->data_offset;
1950                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1951                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1952                                                 atomic_inc(&r10_bio->remaining);
1953                                                 /* and we write to 'i' */
1954
1955                                                 for (k=0; k<conf->copies; k++)
1956                                                         if (r10_bio->devs[k].devnum == i)
1957                                                                 break;
1958                                                 BUG_ON(k == conf->copies);
1959                                                 bio = r10_bio->devs[1].bio;
1960                                                 bio->bi_next = biolist;
1961                                                 biolist = bio;
1962                                                 bio->bi_private = r10_bio;
1963                                                 bio->bi_end_io = end_sync_write;
1964                                                 bio->bi_rw = WRITE;
1965                                                 bio->bi_sector = r10_bio->devs[k].addr +
1966                                                         conf->mirrors[i].rdev->data_offset;
1967                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1968
1969                                                 r10_bio->devs[0].devnum = d;
1970                                                 r10_bio->devs[1].devnum = i;
1971
1972                                                 break;
1973                                         }
1974                                 }
1975                                 if (j == conf->copies) {
1976                                         /* Cannot recover, so abort the recovery */
1977                                         put_buf(r10_bio);
1978                                         if (rb2)
1979                                                 atomic_dec(&rb2->remaining);
1980                                         r10_bio = rb2;
1981                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
1982                                                               &mddev->recovery))
1983                                                 printk(KERN_INFO "md/raid10:%s: insufficient "
1984                                                        "working devices for recovery.\n",
1985                                                        mdname(mddev));
1986                                         break;
1987                                 }
1988                         }
1989                 if (biolist == NULL) {
1990                         while (r10_bio) {
1991                                 r10bio_t *rb2 = r10_bio;
1992                                 r10_bio = (r10bio_t*) rb2->master_bio;
1993                                 rb2->master_bio = NULL;
1994                                 put_buf(rb2);
1995                         }
1996                         goto giveup;
1997                 }
1998         } else {
1999                 /* resync. Schedule a read for every block at this virt offset */
2000                 int count = 0;
2001
2002                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2003
2004                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2005                                        &sync_blocks, mddev->degraded) &&
2006                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2007                         /* We can skip this block */
2008                         *skipped = 1;
2009                         return sync_blocks + sectors_skipped;
2010                 }
2011                 if (sync_blocks < max_sync)
2012                         max_sync = sync_blocks;
2013                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2014
2015                 r10_bio->mddev = mddev;
2016                 atomic_set(&r10_bio->remaining, 0);
2017                 raise_barrier(conf, 0);
2018                 conf->next_resync = sector_nr;
2019
2020                 r10_bio->master_bio = NULL;
2021                 r10_bio->sector = sector_nr;
2022                 set_bit(R10BIO_IsSync, &r10_bio->state);
2023                 raid10_find_phys(conf, r10_bio);
2024                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2025
2026                 for (i=0; i<conf->copies; i++) {
2027                         int d = r10_bio->devs[i].devnum;
2028                         bio = r10_bio->devs[i].bio;
2029                         bio->bi_end_io = NULL;
2030                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
2031                         if (conf->mirrors[d].rdev == NULL ||
2032                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2033                                 continue;
2034                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2035                         atomic_inc(&r10_bio->remaining);
2036                         bio->bi_next = biolist;
2037                         biolist = bio;
2038                         bio->bi_private = r10_bio;
2039                         bio->bi_end_io = end_sync_read;
2040                         bio->bi_rw = READ;
2041                         bio->bi_sector = r10_bio->devs[i].addr +
2042                                 conf->mirrors[d].rdev->data_offset;
2043                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2044                         count++;
2045                 }
2046
2047                 if (count < 2) {
2048                         for (i=0; i<conf->copies; i++) {
2049                                 int d = r10_bio->devs[i].devnum;
2050                                 if (r10_bio->devs[i].bio->bi_end_io)
2051                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
2052                         }
2053                         put_buf(r10_bio);
2054                         biolist = NULL;
2055                         goto giveup;
2056                 }
2057         }
2058
2059         for (bio = biolist; bio ; bio=bio->bi_next) {
2060
2061                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2062                 if (bio->bi_end_io)
2063                         bio->bi_flags |= 1 << BIO_UPTODATE;
2064                 bio->bi_vcnt = 0;
2065                 bio->bi_idx = 0;
2066                 bio->bi_phys_segments = 0;
2067                 bio->bi_size = 0;
2068         }
2069
2070         nr_sectors = 0;
2071         if (sector_nr + max_sync < max_sector)
2072                 max_sector = sector_nr + max_sync;
2073         do {
2074                 struct page *page;
2075                 int len = PAGE_SIZE;
2076                 disk = 0;
2077                 if (sector_nr + (len>>9) > max_sector)
2078                         len = (max_sector - sector_nr) << 9;
2079                 if (len == 0)
2080                         break;
2081                 for (bio= biolist ; bio ; bio=bio->bi_next) {
2082                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2083                         if (bio_add_page(bio, page, len, 0) == 0) {
2084                                 /* stop here */
2085                                 struct bio *bio2;
2086                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2087                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
2088                                         /* remove last page from this bio */
2089                                         bio2->bi_vcnt--;
2090                                         bio2->bi_size -= len;
2091                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2092                                 }
2093                                 goto bio_full;
2094                         }
2095                         disk = i;
2096                 }
2097                 nr_sectors += len>>9;
2098                 sector_nr += len>>9;
2099         } while (biolist->bi_vcnt < RESYNC_PAGES);
2100  bio_full:
2101         r10_bio->sectors = nr_sectors;
2102
2103         while (biolist) {
2104                 bio = biolist;
2105                 biolist = biolist->bi_next;
2106
2107                 bio->bi_next = NULL;
2108                 r10_bio = bio->bi_private;
2109                 r10_bio->sectors = nr_sectors;
2110
2111                 if (bio->bi_end_io == end_sync_read) {
2112                         md_sync_acct(bio->bi_bdev, nr_sectors);
2113                         generic_make_request(bio);
2114                 }
2115         }
2116
2117         if (sectors_skipped)
2118                 /* pretend they weren't skipped, it makes
2119                  * no important difference in this case
2120                  */
2121                 md_done_sync(mddev, sectors_skipped, 1);
2122
2123         return sectors_skipped + nr_sectors;
2124  giveup:
2125         /* There is nowhere to write, so all non-sync
2126          * drives must be failed, so try the next chunk...
2127          */
2128         if (sector_nr + max_sync < max_sector)
2129                 max_sector = sector_nr + max_sync;
2130
2131         sectors_skipped += (max_sector - sector_nr);
2132         chunks_skipped ++;
2133         sector_nr = max_sector;
2134         goto skipped;
2135 }
2136
2137 static sector_t
2138 raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2139 {
2140         sector_t size;
2141         conf_t *conf = mddev->private;
2142
2143         if (!raid_disks)
2144                 raid_disks = conf->raid_disks;
2145         if (!sectors)
2146                 sectors = conf->dev_sectors;
2147
2148         size = sectors >> conf->chunk_shift;
2149         sector_div(size, conf->far_copies);
2150         size = size * raid_disks;
2151         sector_div(size, conf->near_copies);
2152
2153         return size << conf->chunk_shift;
2154 }
2155
2156
2157 static conf_t *setup_conf(mddev_t *mddev)
2158 {
2159         conf_t *conf = NULL;
2160         int nc, fc, fo;
2161         sector_t stride, size;
2162         int err = -EINVAL;
2163
2164         if (mddev->chunk_sectors < (PAGE_SIZE >> 9) ||
2165             !is_power_of_2(mddev->chunk_sectors)) {
2166                 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2167                        "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2168                        mdname(mddev), PAGE_SIZE);
2169                 goto out;
2170         }
2171
2172         nc = mddev->layout & 255;
2173         fc = (mddev->layout >> 8) & 255;
2174         fo = mddev->layout & (1<<16);
2175
2176         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2177             (mddev->layout >> 17)) {
2178                 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2179                        mdname(mddev), mddev->layout);
2180                 goto out;
2181         }
2182
2183         err = -ENOMEM;
2184         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2185         if (!conf)
2186                 goto out;
2187
2188         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2189                                 GFP_KERNEL);
2190         if (!conf->mirrors)
2191                 goto out;
2192
2193         conf->tmppage = alloc_page(GFP_KERNEL);
2194         if (!conf->tmppage)
2195                 goto out;
2196
2197
2198         conf->raid_disks = mddev->raid_disks;
2199         conf->near_copies = nc;
2200         conf->far_copies = fc;
2201         conf->copies = nc*fc;
2202         conf->far_offset = fo;
2203         conf->chunk_mask = mddev->new_chunk_sectors - 1;
2204         conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2205
2206         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2207                                            r10bio_pool_free, conf);
2208         if (!conf->r10bio_pool)
2209                 goto out;
2210
2211         size = mddev->dev_sectors >> conf->chunk_shift;
2212         sector_div(size, fc);
2213         size = size * conf->raid_disks;
2214         sector_div(size, nc);
2215         /* 'size' is now the number of chunks in the array */
2216         /* calculate "used chunks per device" in 'stride' */
2217         stride = size * conf->copies;
2218
2219         /* We need to round up when dividing by raid_disks to
2220          * get the stride size.
2221          */
2222         stride += conf->raid_disks - 1;
2223         sector_div(stride, conf->raid_disks);
2224
2225         conf->dev_sectors = stride << conf->chunk_shift;
2226
2227         if (fo)
2228                 stride = 1;
2229         else
2230                 sector_div(stride, fc);
2231         conf->stride = stride << conf->chunk_shift;
2232
2233
2234         spin_lock_init(&conf->device_lock);
2235         INIT_LIST_HEAD(&conf->retry_list);
2236
2237         spin_lock_init(&conf->resync_lock);
2238         init_waitqueue_head(&conf->wait_barrier);
2239
2240         conf->thread = md_register_thread(raid10d, mddev, NULL);
2241         if (!conf->thread)
2242                 goto out;
2243
2244         conf->scale_disks = 0;
2245         conf->mddev = mddev;
2246         return conf;
2247
2248  out:
2249         printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2250                mdname(mddev));
2251         if (conf) {
2252                 if (conf->r10bio_pool)
2253                         mempool_destroy(conf->r10bio_pool);
2254                 kfree(conf->mirrors);
2255                 safe_put_page(conf->tmppage);
2256                 kfree(conf);
2257         }
2258         return ERR_PTR(err);
2259 }
2260
2261 static int run(mddev_t *mddev)
2262 {
2263         conf_t *conf;
2264         int i, disk_idx, chunk_size;
2265         mirror_info_t *disk;
2266         mdk_rdev_t *rdev;
2267         sector_t size;
2268
2269         /*
2270          * copy the already verified devices into our private RAID10
2271          * bookkeeping area. [whatever we allocate in run(),
2272          * should be freed in stop()]
2273          */
2274
2275         if (mddev->private == NULL) {
2276                 conf = setup_conf(mddev);
2277                 if (IS_ERR(conf))
2278                         return PTR_ERR(conf);
2279                 mddev->private = conf;
2280         }
2281         conf = mddev->private;
2282         if (!conf)
2283                 goto out;
2284
2285         mddev->queue->queue_lock = &conf->device_lock;
2286
2287         mddev->thread = conf->thread;
2288         conf->thread = NULL;
2289
2290         chunk_size = mddev->chunk_sectors << 9;
2291         blk_queue_io_min(mddev->queue, chunk_size);
2292         if (conf->raid_disks % conf->near_copies)
2293                 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2294         else
2295                 blk_queue_io_opt(mddev->queue, chunk_size *
2296                                  (conf->raid_disks / conf->near_copies));
2297
2298         list_for_each_entry(rdev, &mddev->disks, same_set) {
2299                 disk_idx = rdev->raid_disk;
2300                 if (disk_idx >= conf->raid_disks
2301                     || disk_idx < 0)
2302                         continue;
2303                 if (conf->scale_disks) {
2304                         disk_idx *= conf->scale_disks;
2305                         rdev->raid_disk = disk_idx;
2306                         /* MOVE 'rd%d' link !! */
2307                 }
2308                 disk = conf->mirrors + disk_idx;
2309
2310                 disk->rdev = rdev;
2311                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2312                                   rdev->data_offset << 9);
2313                 /* as we don't honour merge_bvec_fn, we must never risk
2314                  * violating it, so limit max_segments to 1 lying
2315                  * within a single page.
2316                  */
2317                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2318                         blk_queue_max_segments(mddev->queue, 1);
2319                         blk_queue_segment_boundary(mddev->queue,
2320                                                    PAGE_CACHE_SIZE - 1);
2321                 }
2322
2323                 disk->head_position = 0;
2324         }
2325         /* need to check that every block has at least one working mirror */
2326         if (!enough(conf)) {
2327                 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2328                        mdname(mddev));
2329                 goto out_free_conf;
2330         }
2331
2332         mddev->degraded = 0;
2333         for (i = 0; i < conf->raid_disks; i++) {
2334
2335                 disk = conf->mirrors + i;
2336
2337                 if (!disk->rdev ||
2338                     !test_bit(In_sync, &disk->rdev->flags)) {
2339                         disk->head_position = 0;
2340                         mddev->degraded++;
2341                         if (disk->rdev)
2342                                 conf->fullsync = 1;
2343                 }
2344         }
2345
2346         if (mddev->recovery_cp != MaxSector)
2347                 printk(KERN_NOTICE "md/raid10:%s: not clean"
2348                        " -- starting background reconstruction\n",
2349                        mdname(mddev));
2350         printk(KERN_INFO
2351                 "md/raid10:%s: active with %d out of %d devices\n",
2352                 mdname(mddev), conf->raid_disks - mddev->degraded,
2353                 conf->raid_disks);
2354         /*
2355          * Ok, everything is just fine now
2356          */
2357         mddev->dev_sectors = conf->dev_sectors;
2358         size = raid10_size(mddev, 0, 0);
2359         md_set_array_sectors(mddev, size);
2360         mddev->resync_max_sectors = size;
2361
2362         mddev->queue->unplug_fn = raid10_unplug;
2363         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2364         mddev->queue->backing_dev_info.congested_data = mddev;
2365
2366         /* Calculate max read-ahead size.
2367          * We need to readahead at least twice a whole stripe....
2368          * maybe...
2369          */
2370         {
2371                 int stripe = conf->raid_disks *
2372                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2373                 stripe /= conf->near_copies;
2374                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2375                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2376         }
2377
2378         if (conf->near_copies < conf->raid_disks)
2379                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2380         md_integrity_register(mddev);
2381         return 0;
2382
2383 out_free_conf:
2384         if (conf->r10bio_pool)
2385                 mempool_destroy(conf->r10bio_pool);
2386         safe_put_page(conf->tmppage);
2387         kfree(conf->mirrors);
2388         kfree(conf);
2389         mddev->private = NULL;
2390         md_unregister_thread(mddev->thread);
2391 out:
2392         return -EIO;
2393 }
2394
2395 static int stop(mddev_t *mddev)
2396 {
2397         conf_t *conf = mddev->private;
2398
2399         raise_barrier(conf, 0);
2400         lower_barrier(conf);
2401
2402         md_unregister_thread(mddev->thread);
2403         mddev->thread = NULL;
2404         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2405         if (conf->r10bio_pool)
2406                 mempool_destroy(conf->r10bio_pool);
2407         kfree(conf->mirrors);
2408         kfree(conf);
2409         mddev->private = NULL;
2410         return 0;
2411 }
2412
2413 static void raid10_quiesce(mddev_t *mddev, int state)
2414 {
2415         conf_t *conf = mddev->private;
2416
2417         switch(state) {
2418         case 1:
2419                 raise_barrier(conf, 0);
2420                 break;
2421         case 0:
2422                 lower_barrier(conf);
2423                 break;
2424         }
2425 }
2426
2427 static void *raid10_takeover_raid0(mddev_t *mddev)
2428 {
2429         mdk_rdev_t *rdev;
2430         conf_t *conf;
2431
2432         if (mddev->degraded > 0) {
2433                 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
2434                        mdname(mddev));
2435                 return ERR_PTR(-EINVAL);
2436         }
2437
2438         /* Update slot numbers to obtain
2439          * degraded raid10 with missing mirrors
2440          */
2441         list_for_each_entry(rdev, &mddev->disks, same_set) {
2442                 rdev->raid_disk *= 2;
2443         }
2444
2445         /* Set new parameters */
2446         mddev->new_level = 10;
2447         /* new layout: far_copies = 1, near_copies = 2 */
2448         mddev->new_layout = (1<<8) + 2;
2449         mddev->new_chunk_sectors = mddev->chunk_sectors;
2450         mddev->delta_disks = mddev->raid_disks;
2451         mddev->degraded = mddev->raid_disks;
2452         mddev->raid_disks *= 2;
2453         /* make sure it will be not marked as dirty */
2454         mddev->recovery_cp = MaxSector;
2455
2456         conf = setup_conf(mddev);
2457         conf->scale_disks = 2;
2458         return conf;
2459 }
2460
2461 static void *raid10_takeover(mddev_t *mddev)
2462 {
2463         struct raid0_private_data *raid0_priv;
2464
2465         /* raid10 can take over:
2466          *  raid0 - providing it has only two drives
2467          */
2468         if (mddev->level == 0) {
2469                 /* for raid0 takeover only one zone is supported */
2470                 raid0_priv = mddev->private;
2471                 if (raid0_priv->nr_strip_zones > 1) {
2472                         printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
2473                                " with more than one zone.\n",
2474                                mdname(mddev));
2475                         return ERR_PTR(-EINVAL);
2476                 }
2477                 return raid10_takeover_raid0(mddev);
2478         }
2479         return ERR_PTR(-EINVAL);
2480 }
2481
2482 static struct mdk_personality raid10_personality =
2483 {
2484         .name           = "raid10",
2485         .level          = 10,
2486         .owner          = THIS_MODULE,
2487         .make_request   = make_request,
2488         .run            = run,
2489         .stop           = stop,
2490         .status         = status,
2491         .error_handler  = error,
2492         .hot_add_disk   = raid10_add_disk,
2493         .hot_remove_disk= raid10_remove_disk,
2494         .spare_active   = raid10_spare_active,
2495         .sync_request   = sync_request,
2496         .quiesce        = raid10_quiesce,
2497         .size           = raid10_size,
2498         .takeover       = raid10_takeover,
2499 };
2500
2501 static int __init raid_init(void)
2502 {
2503         return register_md_personality(&raid10_personality);
2504 }
2505
2506 static void raid_exit(void)
2507 {
2508         unregister_md_personality(&raid10_personality);
2509 }
2510
2511 module_init(raid_init);
2512 module_exit(raid_exit);
2513 MODULE_LICENSE("GPL");
2514 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
2515 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2516 MODULE_ALIAS("md-raid10");
2517 MODULE_ALIAS("md-level-10");