drivers/md/raid10.c

   1 /*
   2  * raid10.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 2000-2004 Neil Brown
   5  *
   6  * RAID-10 support for md.
   7  *
   8  * Base on code in raid1.c.  See raid1.c for further copyright information.
   9  *
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2, or (at your option)
  14  * any later version.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * (for example /usr/src/linux/COPYING); if not, write to the Free
  18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 #include <linux/slab.h>
  22 #include <linux/delay.h>
  23 #include <linux/blkdev.h>
  24 #include <linux/seq_file.h>
  25 #include "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 allow_barrier(conf_t *conf);
  61 static void lower_barrier(conf_t *conf);
  62
  63 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
  64 {
  65         conf_t *conf = data;
  66         int size = offsetof(struct r10bio_s, devs[conf->copies]);
  67
  68         /* allocate a r10bio with room for raid_disks entries in the bios array */
  69         return kzalloc(size, gfp_flags);
  70 }
  71
  72 static void r10bio_pool_free(void *r10_bio, void *data)
  73 {
  74         kfree(r10_bio);
  75 }
  76
  77 /* Maximum size of each resync request */
  78 #define RESYNC_BLOCK_SIZE (64*1024)
  79 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  80 /* amount of memory to reserve for resync requests */
  81 #define RESYNC_WINDOW (1024*1024)
  82 /* maximum number of concurrent requests, memory permitting */
  83 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
  84
  85 /*
  86  * When performing a resync, we need to read and compare, so
  87  * we need as many pages are there are copies.
  88  * When performing a recovery, we need 2 bios, one for read,
  89  * one for write (we recover only one drive per r10buf)
  90  *
  91  */
  92 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
  93 {
  94         conf_t *conf = data;
  95         struct page *page;
  96         r10bio_t *r10_bio;
  97         struct bio *bio;
  98         int i, j;
  99         int nalloc;
 100
 101         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 102         if (!r10_bio)
 103                 return NULL;
 104
 105         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
 106                 nalloc = conf->copies; /* resync */
 107         else
 108                 nalloc = 2; /* recovery */
 109
 110         /*
 111          * Allocate bios.
 112          */
 113         for (j = nalloc ; j-- ; ) {
 114                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 115                 if (!bio)
 116                         goto out_free_bio;
 117                 r10_bio->devs[j].bio = bio;
 118         }
 119         /*
 120          * Allocate RESYNC_PAGES data pages and attach them
 121          * where needed.
 122          */
 123         for (j = 0 ; j < nalloc; j++) {
 124                 bio = r10_bio->devs[j].bio;
 125                 for (i = 0; i < RESYNC_PAGES; i++) {
 126                         page = alloc_page(gfp_flags);
 127                         if (unlikely(!page))
 128                                 goto out_free_pages;
 129
 130                         bio->bi_io_vec[i].bv_page = page;
 131                 }
 132         }
 133
 134         return r10_bio;
 135
 136 out_free_pages:
 137         for ( ; i > 0 ; i--)
 138                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
 139         while (j--)
 140                 for (i = 0; i < RESYNC_PAGES ; i++)
 141                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 142         j = -1;
 143 out_free_bio:
 144         while ( ++j < nalloc )
 145                 bio_put(r10_bio->devs[j].bio);
 146         r10bio_pool_free(r10_bio, conf);
 147         return NULL;
 148 }
 149
 150 static void r10buf_pool_free(void *__r10_bio, void *data)
 151 {
 152         int i;
 153         conf_t *conf = data;
 154         r10bio_t *r10bio = __r10_bio;
 155         int j;
 156
 157         for (j=0; j < conf->copies; j++) {
 158                 struct bio *bio = r10bio->devs[j].bio;
 159                 if (bio) {
 160                         for (i = 0; i < RESYNC_PAGES; i++) {
 161                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 162                                 bio->bi_io_vec[i].bv_page = NULL;
 163                         }
 164                         bio_put(bio);
 165                 }
 166         }
 167         r10bio_pool_free(r10bio, conf);
 168 }
 169
 170 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
 171 {
 172         int i;
 173
 174         for (i = 0; i < conf->copies; i++) {
 175                 struct bio **bio = & r10_bio->devs[i].bio;
 176                 if (*bio && *bio != IO_BLOCKED)
 177                         bio_put(*bio);
 178                 *bio = NULL;
 179         }
 180 }
 181
 182 static void free_r10bio(r10bio_t *r10_bio)
 183 {
 184         conf_t *conf = r10_bio->mddev->private;
 185
 186         /*
 187          * Wake up any possible resync thread that waits for the device
 188          * to go idle.
 189          */
 190         allow_barrier(conf);
 191
 192         put_all_bios(conf, r10_bio);
 193         mempool_free(r10_bio, conf->r10bio_pool);
 194 }
 195
 196 static void put_buf(r10bio_t *r10_bio)
 197 {
 198         conf_t *conf = r10_bio->mddev->private;
 199
 200         mempool_free(r10_bio, conf->r10buf_pool);
 201
 202         lower_barrier(conf);
 203 }
 204
 205 static void reschedule_retry(r10bio_t *r10_bio)
 206 {
 207         unsigned long flags;
 208         mddev_t *mddev = r10_bio->mddev;
 209         conf_t *conf = mddev->private;
 210
 211         spin_lock_irqsave(&conf->device_lock, flags);
 212         list_add(&r10_bio->retry_list, &conf->retry_list);
 213         conf->nr_queued ++;
 214         spin_unlock_irqrestore(&conf->device_lock, flags);
 215
 216         /* wake up frozen array... */
 217         wake_up(&conf->wait_barrier);
 218
 219         md_wakeup_thread(mddev->thread);
 220 }
 221
 222 /*
 223  * raid_end_bio_io() is called when we have finished servicing a mirrored
 224  * operation and are ready to return a success/failure code to the buffer
 225  * cache layer.
 226  */
 227 static void raid_end_bio_io(r10bio_t *r10_bio)
 228 {
 229         struct bio *bio = r10_bio->master_bio;
 230
 231         bio_endio(bio,
 232                 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
 233         free_r10bio(r10_bio);
 234 }
 235
 236 /*
 237  * Update disk head position estimator based on IRQ completion info.
 238  */
 239 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
 240 {
 241         conf_t *conf = r10_bio->mddev->private;
 242
 243         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 244                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 245 }
 246
 247 static void raid10_end_read_request(struct bio *bio, int error)
 248 {
 249         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 250         r10bio_t *r10_bio = bio->bi_private;
 251         int slot, dev;
 252         conf_t *conf = r10_bio->mddev->private;
 253
 254
 255         slot = r10_bio->read_slot;
 256         dev = r10_bio->devs[slot].devnum;
 257         /*
 258          * this branch is our 'one mirror IO has finished' event handler:
 259          */
 260         update_head_pos(slot, r10_bio);
 261
 262         if (uptodate) {
 263                 /*
 264                  * Set R10BIO_Uptodate in our master bio, so that
 265                  * we will return a good error code to the higher
 266                  * levels even if IO on some other mirrored buffer fails.
 267                  *
 268                  * The 'master' represents the composite IO operation to
 269                  * user-side. So if something waits for IO, then it will
 270                  * wait for the 'master' bio.
 271                  */
 272                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 273                 raid_end_bio_io(r10_bio);
 274         } else {
 275                 /*
 276                  * oops, read error:
 277                  */
 278                 char b[BDEVNAME_SIZE];
 279                 if (printk_ratelimit())
 280                         printk(KERN_ERR "md/raid10:%s: %s: rescheduling sector %llu\n",
 281                                mdname(conf->mddev),
 282                                bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
 283                 reschedule_retry(r10_bio);
 284         }
 285
 286         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 287 }
 288
 289 static void raid10_end_write_request(struct bio *bio, int error)
 290 {
 291         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 292         r10bio_t *r10_bio = bio->bi_private;
 293         int slot, dev;
 294         conf_t *conf = r10_bio->mddev->private;
 295
 296         for (slot = 0; slot < conf->copies; slot++)
 297                 if (r10_bio->devs[slot].bio == bio)
 298                         break;
 299         dev = r10_bio->devs[slot].devnum;
 300
 301         /*
 302          * this branch is our 'one mirror IO has finished' event handler:
 303          */
 304         if (!uptodate) {
 305                 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
 306                 /* an I/O failed, we can't clear the bitmap */
 307                 set_bit(R10BIO_Degraded, &r10_bio->state);
 308         } else
 309                 /*
 310                  * Set R10BIO_Uptodate in our master bio, so that
 311                  * we will return a good error code for to the higher
 312                  * levels even if IO on some other mirrored buffer fails.
 313                  *
 314                  * The 'master' represents the composite IO operation to
 315                  * user-side. So if something waits for IO, then it will
 316                  * wait for the 'master' bio.
 317                  */
 318                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 319
 320         update_head_pos(slot, r10_bio);
 321
 322         /*
 323          *
 324          * Let's see if all mirrored write operations have finished
 325          * already.
 326          */
 327         if (atomic_dec_and_test(&r10_bio->remaining)) {
 328                 /* clear the bitmap if all writes complete successfully */
 329                 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 330                                 r10_bio->sectors,
 331                                 !test_bit(R10BIO_Degraded, &r10_bio->state),
 332                                 0);
 333                 md_write_end(r10_bio->mddev);
 334                 raid_end_bio_io(r10_bio);
 335         }
 336
 337         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 338 }
 339
 340
 341 /*
 342  * RAID10 layout manager
 343  * As well as the chunksize and raid_disks count, there are two
 344  * parameters: near_copies and far_copies.
 345  * near_copies * far_copies must be <= raid_disks.
 346  * Normally one of these will be 1.
 347  * If both are 1, we get raid0.
 348  * If near_copies == raid_disks, we get raid1.
 349  *
 350  * Chunks are laid out in raid0 style with near_copies copies of the
 351  * first chunk, followed by near_copies copies of the next chunk and
 352  * so on.
 353  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 354  * as described above, we start again with a device offset of near_copies.
 355  * So we effectively have another copy of the whole array further down all
 356  * the drives, but with blocks on different drives.
 357  * With this layout, and block is never stored twice on the one device.
 358  *
 359  * raid10_find_phys finds the sector offset of a given virtual sector
 360  * on each device that it is on.
 361  *
 362  * raid10_find_virt does the reverse mapping, from a device and a
 363  * sector offset to a virtual address
 364  */
 365
 366 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
 367 {
 368         int n,f;
 369         sector_t sector;
 370         sector_t chunk;
 371         sector_t stripe;
 372         int dev;
 373
 374         int slot = 0;
 375
 376         /* now calculate first sector/dev */
 377         chunk = r10bio->sector >> conf->chunk_shift;
 378         sector = r10bio->sector & conf->chunk_mask;
 379
 380         chunk *= conf->near_copies;
 381         stripe = chunk;
 382         dev = sector_div(stripe, conf->raid_disks);
 383         if (conf->far_offset)
 384                 stripe *= conf->far_copies;
 385
 386         sector += stripe << conf->chunk_shift;
 387
 388         /* and calculate all the others */
 389         for (n=0; n < conf->near_copies; n++) {
 390                 int d = dev;
 391                 sector_t s = sector;
 392                 r10bio->devs[slot].addr = sector;
 393                 r10bio->devs[slot].devnum = d;
 394                 slot++;
 395
 396                 for (f = 1; f < conf->far_copies; f++) {
 397                         d += conf->near_copies;
 398                         if (d >= conf->raid_disks)
 399                                 d -= conf->raid_disks;
 400                         s += conf->stride;
 401                         r10bio->devs[slot].devnum = d;
 402                         r10bio->devs[slot].addr = s;
 403                         slot++;
 404                 }
 405                 dev++;
 406                 if (dev >= conf->raid_disks) {
 407                         dev = 0;
 408                         sector += (conf->chunk_mask + 1);
 409                 }
 410         }
 411         BUG_ON(slot != conf->copies);
 412 }
 413
 414 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
 415 {
 416         sector_t offset, chunk, vchunk;
 417
 418         offset = sector & conf->chunk_mask;
 419         if (conf->far_offset) {
 420                 int fc;
 421                 chunk = sector >> conf->chunk_shift;
 422                 fc = sector_div(chunk, conf->far_copies);
 423                 dev -= fc * conf->near_copies;
 424                 if (dev < 0)
 425                         dev += conf->raid_disks;
 426         } else {
 427                 while (sector >= conf->stride) {
 428                         sector -= conf->stride;
 429                         if (dev < conf->near_copies)
 430                                 dev += conf->raid_disks - conf->near_copies;
 431                         else
 432                                 dev -= conf->near_copies;
 433                 }
 434                 chunk = sector >> conf->chunk_shift;
 435         }
 436         vchunk = chunk * conf->raid_disks + dev;
 437         sector_div(vchunk, conf->near_copies);
 438         return (vchunk << conf->chunk_shift) + offset;
 439 }
 440
 441 /**
 442  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 443  *      @q: request queue
 444  *      @bvm: properties of new bio
 445  *      @biovec: the request that could be merged to it.
 446  *
 447  *      Return amount of bytes we can accept at this offset
 448  *      If near_copies == raid_disk, there are no striping issues,
 449  *      but in that case, the function isn't called at all.
 450  */
 451 static int raid10_mergeable_bvec(struct request_queue *q,
 452                                  struct bvec_merge_data *bvm,
 453                                  struct bio_vec *biovec)
 454 {
 455         mddev_t *mddev = q->queuedata;
 456         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 457         int max;
 458         unsigned int chunk_sectors = mddev->chunk_sectors;
 459         unsigned int bio_sectors = bvm->bi_size >> 9;
 460
 461         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 462         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 463         if (max <= biovec->bv_len && bio_sectors == 0)
 464                 return biovec->bv_len;
 465         else
 466                 return max;
 467 }
 468
 469 /*
 470  * This routine returns the disk from which the requested read should
 471  * be done. There is a per-array 'next expected sequential IO' sector
 472  * number - if this matches on the next IO then we use the last disk.
 473  * There is also a per-disk 'last know head position' sector that is
 474  * maintained from IRQ contexts, both the normal and the resync IO
 475  * completion handlers update this position correctly. If there is no
 476  * perfect sequential match then we pick the disk whose head is closest.
 477  *
 478  * If there are 2 mirrors in the same 2 devices, performance degrades
 479  * because position is mirror, not device based.
 480  *
 481  * The rdev for the device selected will have nr_pending incremented.
 482  */
 483
 484 /*
 485  * FIXME: possibly should rethink readbalancing and do it differently
 486  * depending on near_copies / far_copies geometry.
 487  */
 488 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
 489 {
 490         const sector_t this_sector = r10_bio->sector;
 491         int disk, slot, nslot;
 492         const int sectors = r10_bio->sectors;
 493         sector_t new_distance, current_distance;
 494         mdk_rdev_t *rdev;
 495
 496         raid10_find_phys(conf, r10_bio);
 497         rcu_read_lock();
 498         /*
 499          * Check if we can balance. We can balance on the whole
 500          * device if no resync is going on (recovery is ok), or below
 501          * the resync window. We take the first readable disk when
 502          * above the resync window.
 503          */
 504         if (conf->mddev->recovery_cp < MaxSector
 505             && (this_sector + sectors >= conf->next_resync)) {
 506                 /* make sure that disk is operational */
 507                 slot = 0;
 508                 disk = r10_bio->devs[slot].devnum;
 509
 510                 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 511                        r10_bio->devs[slot].bio == IO_BLOCKED ||
 512                        !test_bit(In_sync, &rdev->flags)) {
 513                         slot++;
 514                         if (slot == conf->copies) {
 515                                 slot = 0;
 516                                 disk = -1;
 517                                 break;
 518                         }
 519                         disk = r10_bio->devs[slot].devnum;
 520                 }
 521                 goto rb_out;
 522         }
 523
 524
 525         /* make sure the disk is operational */
 526         slot = 0;
 527         disk = r10_bio->devs[slot].devnum;
 528         while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 529                r10_bio->devs[slot].bio == IO_BLOCKED ||
 530                !test_bit(In_sync, &rdev->flags)) {
 531                 slot ++;
 532                 if (slot == conf->copies) {
 533                         disk = -1;
 534                         goto rb_out;
 535                 }
 536                 disk = r10_bio->devs[slot].devnum;
 537         }
 538
 539
 540         current_distance = abs(r10_bio->devs[slot].addr -
 541                                conf->mirrors[disk].head_position);
 542
 543         /* Find the disk whose head is closest,
 544          * or - for far > 1 - find the closest to partition beginning */
 545
 546         for (nslot = slot; nslot < conf->copies; nslot++) {
 547                 int ndisk = r10_bio->devs[nslot].devnum;
 548
 549
 550                 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
 551                     r10_bio->devs[nslot].bio == IO_BLOCKED ||
 552                     !test_bit(In_sync, &rdev->flags))
 553                         continue;
 554
 555                 /* This optimisation is debatable, and completely destroys
 556                  * sequential read speed for 'far copies' arrays.  So only
 557                  * keep it for 'near' arrays, and review those later.
 558                  */
 559                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
 560                         disk = ndisk;
 561                         slot = nslot;
 562                         break;
 563                 }
 564
 565                 /* for far > 1 always use the lowest address */
 566                 if (conf->far_copies > 1)
 567                         new_distance = r10_bio->devs[nslot].addr;
 568                 else
 569                         new_distance = abs(r10_bio->devs[nslot].addr -
 570                                            conf->mirrors[ndisk].head_position);
 571                 if (new_distance < current_distance) {
 572                         current_distance = new_distance;
 573                         disk = ndisk;
 574                         slot = nslot;
 575                 }
 576         }
 577
 578 rb_out:
 579         r10_bio->read_slot = slot;
 580 /*      conf->next_seq_sect = this_sector + sectors;*/
 581
 582         if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
 583                 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
 584         else
 585                 disk = -1;
 586         rcu_read_unlock();
 587
 588         return disk;
 589 }
 590
 591 static int raid10_congested(void *data, int bits)
 592 {
 593         mddev_t *mddev = data;
 594         conf_t *conf = mddev->private;
 595         int i, ret = 0;
 596
 597         if (mddev_congested(mddev, bits))
 598                 return 1;
 599         rcu_read_lock();
 600         for (i = 0; i < conf->raid_disks && ret == 0; i++) {
 601                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 602                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 603                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 604
 605                         ret |= bdi_congested(&q->backing_dev_info, bits);
 606                 }
 607         }
 608         rcu_read_unlock();
 609         return ret;
 610 }
 611
 612 static void flush_pending_writes(conf_t *conf)
 613 {
 614         /* Any writes that have been queued but are awaiting
 615          * bitmap updates get flushed here.
 616          */
 617         spin_lock_irq(&conf->device_lock);
 618
 619         if (conf->pending_bio_list.head) {
 620                 struct bio *bio;
 621                 bio = bio_list_get(&conf->pending_bio_list);
 622                 spin_unlock_irq(&conf->device_lock);
 623                 /* flush any pending bitmap writes to disk
 624                  * before proceeding w/ I/O */
 625                 bitmap_unplug(conf->mddev->bitmap);
 626
 627                 while (bio) { /* submit pending writes */
 628                         struct bio *next = bio->bi_next;
 629                         bio->bi_next = NULL;
 630                         generic_make_request(bio);
 631                         bio = next;
 632                 }
 633         } else
 634                 spin_unlock_irq(&conf->device_lock);
 635 }
 636
 637 static void md_kick_device(mddev_t *mddev)
 638 {
 639         blk_flush_plug(current);
 640         md_wakeup_thread(mddev->thread);
 641 }
 642
 643 /* Barriers....
 644  * Sometimes we need to suspend IO while we do something else,
 645  * either some resync/recovery, or reconfigure the array.
 646  * To do this we raise a 'barrier'.
 647  * The 'barrier' is a counter that can be raised multiple times
 648  * to count how many activities are happening which preclude
 649  * normal IO.
 650  * We can only raise the barrier if there is no pending IO.
 651  * i.e. if nr_pending == 0.
 652  * We choose only to raise the barrier if no-one is waiting for the
 653  * barrier to go down.  This means that as soon as an IO request
 654  * is ready, no other operations which require a barrier will start
 655  * until the IO request has had a chance.
 656  *
 657  * So: regular IO calls 'wait_barrier'.  When that returns there
 658  *    is no backgroup IO happening,  It must arrange to call
 659  *    allow_barrier when it has finished its IO.
 660  * backgroup IO calls must call raise_barrier.  Once that returns
 661  *    there is no normal IO happeing.  It must arrange to call
 662  *    lower_barrier when the particular background IO completes.
 663  */
 664
 665 static void raise_barrier(conf_t *conf, int force)
 666 {
 667         BUG_ON(force && !conf->barrier);
 668         spin_lock_irq(&conf->resync_lock);
 669
 670         /* Wait until no block IO is waiting (unless 'force') */
 671         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 672                             conf->resync_lock, md_kick_device(conf->mddev));
 673
 674         /* block any new IO from starting */
 675         conf->barrier++;
 676
 677         /* No wait for all pending IO to complete */
 678         wait_event_lock_irq(conf->wait_barrier,
 679                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 680                             conf->resync_lock, md_kick_device(conf->mddev));
 681
 682         spin_unlock_irq(&conf->resync_lock);
 683 }
 684
 685 static void lower_barrier(conf_t *conf)
 686 {
 687         unsigned long flags;
 688         spin_lock_irqsave(&conf->resync_lock, flags);
 689         conf->barrier--;
 690         spin_unlock_irqrestore(&conf->resync_lock, flags);
 691         wake_up(&conf->wait_barrier);
 692 }
 693
 694 static void wait_barrier(conf_t *conf)
 695 {
 696         spin_lock_irq(&conf->resync_lock);
 697         if (conf->barrier) {
 698                 conf->nr_waiting++;
 699                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 700                                     conf->resync_lock,
 701                                     md_kick_device(conf->mddev));
 702                 conf->nr_waiting--;
 703         }
 704         conf->nr_pending++;
 705         spin_unlock_irq(&conf->resync_lock);
 706 }
 707
 708 static void allow_barrier(conf_t *conf)
 709 {
 710         unsigned long flags;
 711         spin_lock_irqsave(&conf->resync_lock, flags);
 712         conf->nr_pending--;
 713         spin_unlock_irqrestore(&conf->resync_lock, flags);
 714         wake_up(&conf->wait_barrier);
 715 }
 716
 717 static void freeze_array(conf_t *conf)
 718 {
 719         /* stop syncio and normal IO and wait for everything to
 720          * go quiet.
 721          * We increment barrier and nr_waiting, and then
 722          * wait until nr_pending match nr_queued+1
 723          * This is called in the context of one normal IO request
 724          * that has failed. Thus any sync request that might be pending
 725          * will be blocked by nr_pending, and we need to wait for
 726          * pending IO requests to complete or be queued for re-try.
 727          * Thus the number queued (nr_queued) plus this request (1)
 728          * must match the number of pending IOs (nr_pending) before
 729          * we continue.
 730          */
 731         spin_lock_irq(&conf->resync_lock);
 732         conf->barrier++;
 733         conf->nr_waiting++;
 734         wait_event_lock_irq(conf->wait_barrier,
 735                             conf->nr_pending == conf->nr_queued+1,
 736                             conf->resync_lock,
 737                             ({ flush_pending_writes(conf);
 738                                md_kick_device(conf->mddev); }));
 739         spin_unlock_irq(&conf->resync_lock);
 740 }
 741
 742 static void unfreeze_array(conf_t *conf)
 743 {
 744         /* reverse the effect of the freeze */
 745         spin_lock_irq(&conf->resync_lock);
 746         conf->barrier--;
 747         conf->nr_waiting--;
 748         wake_up(&conf->wait_barrier);
 749         spin_unlock_irq(&conf->resync_lock);
 750 }
 751
 752 static int make_request(mddev_t *mddev, struct bio * bio)
 753 {
 754         conf_t *conf = mddev->private;
 755         mirror_info_t *mirror;
 756         r10bio_t *r10_bio;
 757         struct bio *read_bio;
 758         int i;
 759         int chunk_sects = conf->chunk_mask + 1;
 760         const int rw = bio_data_dir(bio);
 761         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 762         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
 763         unsigned long flags;
 764         mdk_rdev_t *blocked_rdev;
 765
 766         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
 767                 md_flush_request(mddev, bio);
 768                 return 0;
 769         }
 770
 771         /* If this request crosses a chunk boundary, we need to
 772          * split it.  This will only happen for 1 PAGE (or less) requests.
 773          */
 774         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 775                       > chunk_sects &&
 776                     conf->near_copies < conf->raid_disks)) {
 777                 struct bio_pair *bp;
 778                 /* Sanity check -- queue functions should prevent this happening */
 779                 if (bio->bi_vcnt != 1 ||
 780                     bio->bi_idx != 0)
 781                         goto bad_map;
 782                 /* This is a one page bio that upper layers
 783                  * refuse to split for us, so we need to split it.
 784                  */
 785                 bp = bio_split(bio,
 786                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 787
 788                 /* Each of these 'make_request' calls will call 'wait_barrier'.
 789                  * If the first succeeds but the second blocks due to the resync
 790                  * thread raising the barrier, we will deadlock because the
 791                  * IO to the underlying device will be queued in generic_make_request
 792                  * and will never complete, so will never reduce nr_pending.
 793                  * So increment nr_waiting here so no new raise_barriers will
 794                  * succeed, and so the second wait_barrier cannot block.
 795                  */
 796                 spin_lock_irq(&conf->resync_lock);
 797                 conf->nr_waiting++;
 798                 spin_unlock_irq(&conf->resync_lock);
 799
 800                 if (make_request(mddev, &bp->bio1))
 801                         generic_make_request(&bp->bio1);
 802                 if (make_request(mddev, &bp->bio2))
 803                         generic_make_request(&bp->bio2);
 804
 805                 spin_lock_irq(&conf->resync_lock);
 806                 conf->nr_waiting--;
 807                 wake_up(&conf->wait_barrier);
 808                 spin_unlock_irq(&conf->resync_lock);
 809
 810                 bio_pair_release(bp);
 811                 return 0;
 812         bad_map:
 813                 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
 814                        " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
 815                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 816
 817                 bio_io_error(bio);
 818                 return 0;
 819         }
 820
 821         md_write_start(mddev, bio);
 822
 823         /*
 824          * Register the new request and wait if the reconstruction
 825          * thread has put up a bar for new requests.
 826          * Continue immediately if no resync is active currently.
 827          */
 828         wait_barrier(conf);
 829
 830         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 831
 832         r10_bio->master_bio = bio;
 833         r10_bio->sectors = bio->bi_size >> 9;
 834
 835         r10_bio->mddev = mddev;
 836         r10_bio->sector = bio->bi_sector;
 837         r10_bio->state = 0;
 838
 839         if (rw == READ) {
 840                 /*
 841                  * read balancing logic:
 842                  */
 843                 int disk = read_balance(conf, r10_bio);
 844                 int slot = r10_bio->read_slot;
 845                 if (disk < 0) {
 846                         raid_end_bio_io(r10_bio);
 847                         return 0;
 848                 }
 849                 mirror = conf->mirrors + disk;
 850
 851                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 852
 853                 r10_bio->devs[slot].bio = read_bio;
 854
 855                 read_bio->bi_sector = r10_bio->devs[slot].addr +
 856                         mirror->rdev->data_offset;
 857                 read_bio->bi_bdev = mirror->rdev->bdev;
 858                 read_bio->bi_end_io = raid10_end_read_request;
 859                 read_bio->bi_rw = READ | do_sync;
 860                 read_bio->bi_private = r10_bio;
 861
 862                 generic_make_request(read_bio);
 863                 return 0;
 864         }
 865
 866         /*
 867          * WRITE:
 868          */
 869         /* first select target devices under rcu_lock and
 870          * inc refcount on their rdev.  Record them by setting
 871          * bios[x] to bio
 872          */
 873         raid10_find_phys(conf, r10_bio);
 874  retry_write:
 875         blocked_rdev = NULL;
 876         rcu_read_lock();
 877         for (i = 0;  i < conf->copies; i++) {
 878                 int d = r10_bio->devs[i].devnum;
 879                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
 880                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 881                         atomic_inc(&rdev->nr_pending);
 882                         blocked_rdev = rdev;
 883                         break;
 884                 }
 885                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 886                         atomic_inc(&rdev->nr_pending);
 887                         r10_bio->devs[i].bio = bio;
 888                 } else {
 889                         r10_bio->devs[i].bio = NULL;
 890                         set_bit(R10BIO_Degraded, &r10_bio->state);
 891                 }
 892         }
 893         rcu_read_unlock();
 894
 895         if (unlikely(blocked_rdev)) {
 896                 /* Have to wait for this device to get unblocked, then retry */
 897                 int j;
 898                 int d;
 899
 900                 for (j = 0; j < i; j++)
 901                         if (r10_bio->devs[j].bio) {
 902                                 d = r10_bio->devs[j].devnum;
 903                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
 904                         }
 905                 allow_barrier(conf);
 906                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 907                 wait_barrier(conf);
 908                 goto retry_write;
 909         }
 910
 911         atomic_set(&r10_bio->remaining, 1);
 912         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
 913
 914         for (i = 0; i < conf->copies; i++) {
 915                 struct bio *mbio;
 916                 int d = r10_bio->devs[i].devnum;
 917                 if (!r10_bio->devs[i].bio)
 918                         continue;
 919
 920                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 921                 r10_bio->devs[i].bio = mbio;
 922
 923                 mbio->bi_sector = r10_bio->devs[i].addr+
 924                         conf->mirrors[d].rdev->data_offset;
 925                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
 926                 mbio->bi_end_io = raid10_end_write_request;
 927                 mbio->bi_rw = WRITE | do_sync | do_fua;
 928                 mbio->bi_private = r10_bio;
 929
 930                 atomic_inc(&r10_bio->remaining);
 931                 spin_lock_irqsave(&conf->device_lock, flags);
 932                 bio_list_add(&conf->pending_bio_list, mbio);
 933                 spin_unlock_irqrestore(&conf->device_lock, flags);
 934         }
 935
 936         if (atomic_dec_and_test(&r10_bio->remaining)) {
 937                 /* This matches the end of raid10_end_write_request() */
 938                 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 939                                 r10_bio->sectors,
 940                                 !test_bit(R10BIO_Degraded, &r10_bio->state),
 941                                 0);
 942                 md_write_end(mddev);
 943                 raid_end_bio_io(r10_bio);
 944         }
 945
 946         /* In case raid10d snuck in to freeze_array */
 947         wake_up(&conf->wait_barrier);
 948
 949         if (do_sync || !mddev->bitmap)
 950                 md_wakeup_thread(mddev->thread);
 951
 952         return 0;
 953 }
 954
 955 static void status(struct seq_file *seq, mddev_t *mddev)
 956 {
 957         conf_t *conf = mddev->private;
 958         int i;
 959
 960         if (conf->near_copies < conf->raid_disks)
 961                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
 962         if (conf->near_copies > 1)
 963                 seq_printf(seq, " %d near-copies", conf->near_copies);
 964         if (conf->far_copies > 1) {
 965                 if (conf->far_offset)
 966                         seq_printf(seq, " %d offset-copies", conf->far_copies);
 967                 else
 968                         seq_printf(seq, " %d far-copies", conf->far_copies);
 969         }
 970         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 971                                         conf->raid_disks - mddev->degraded);
 972         for (i = 0; i < conf->raid_disks; i++)
 973                 seq_printf(seq, "%s",
 974                               conf->mirrors[i].rdev &&
 975                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
 976         seq_printf(seq, "]");
 977 }
 978
 979 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 980 {
 981         char b[BDEVNAME_SIZE];
 982         conf_t *conf = mddev->private;
 983
 984         /*
 985          * If it is not operational, then we have already marked it as dead
 986          * else if it is the last working disks, ignore the error, let the
 987          * next level up know.
 988          * else mark the drive as failed
 989          */
 990         if (test_bit(In_sync, &rdev->flags)
 991             && conf->raid_disks-mddev->degraded == 1)
 992                 /*
 993                  * Don't fail the drive, just return an IO error.
 994                  * The test should really be more sophisticated than
 995                  * "working_disks == 1", but it isn't critical, and
 996                  * can wait until we do more sophisticated "is the drive
 997                  * really dead" tests...
 998                  */
 999                 return;
1000         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1001                 unsigned long flags;
1002                 spin_lock_irqsave(&conf->device_lock, flags);
1003                 mddev->degraded++;
1004                 spin_unlock_irqrestore(&conf->device_lock, flags);
1005                 /*
1006                  * if recovery is running, make sure it aborts.
1007                  */
1008                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1009         }
1010         set_bit(Faulty, &rdev->flags);
1011         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1012         printk(KERN_ALERT
1013                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1014                "md/raid10:%s: Operation continuing on %d devices.\n",
1015                mdname(mddev), bdevname(rdev->bdev, b),
1016                mdname(mddev), conf->raid_disks - mddev->degraded);
1017 }
1018
1019 static void print_conf(conf_t *conf)
1020 {
1021         int i;
1022         mirror_info_t *tmp;
1023
1024         printk(KERN_DEBUG "RAID10 conf printout:\n");
1025         if (!conf) {
1026                 printk(KERN_DEBUG "(!conf)\n");
1027                 return;
1028         }
1029         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1030                 conf->raid_disks);
1031
1032         for (i = 0; i < conf->raid_disks; i++) {
1033                 char b[BDEVNAME_SIZE];
1034                 tmp = conf->mirrors + i;
1035                 if (tmp->rdev)
1036                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1037                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1038                                 !test_bit(Faulty, &tmp->rdev->flags),
1039                                 bdevname(tmp->rdev->bdev,b));
1040         }
1041 }
1042
1043 static void close_sync(conf_t *conf)
1044 {
1045         wait_barrier(conf);
1046         allow_barrier(conf);
1047
1048         mempool_destroy(conf->r10buf_pool);
1049         conf->r10buf_pool = NULL;
1050 }
1051
1052 /* check if there are enough drives for
1053  * every block to appear on atleast one
1054  */
1055 static int enough(conf_t *conf)
1056 {
1057         int first = 0;
1058
1059         do {
1060                 int n = conf->copies;
1061                 int cnt = 0;
1062                 while (n--) {
1063                         if (conf->mirrors[first].rdev)
1064                                 cnt++;
1065                         first = (first+1) % conf->raid_disks;
1066                 }
1067                 if (cnt == 0)
1068                         return 0;
1069         } while (first != 0);
1070         return 1;
1071 }
1072
1073 static int raid10_spare_active(mddev_t *mddev)
1074 {
1075         int i;
1076         conf_t *conf = mddev->private;
1077         mirror_info_t *tmp;
1078         int count = 0;
1079         unsigned long flags;
1080
1081         /*
1082          * Find all non-in_sync disks within the RAID10 configuration
1083          * and mark them in_sync
1084          */
1085         for (i = 0; i < conf->raid_disks; i++) {
1086                 tmp = conf->mirrors + i;
1087                 if (tmp->rdev
1088                     && !test_bit(Faulty, &tmp->rdev->flags)
1089                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1090                         count++;
1091                         sysfs_notify_dirent(tmp->rdev->sysfs_state);
1092                 }
1093         }
1094         spin_lock_irqsave(&conf->device_lock, flags);
1095         mddev->degraded -= count;
1096         spin_unlock_irqrestore(&conf->device_lock, flags);
1097
1098         print_conf(conf);
1099         return count;
1100 }
1101
1102
1103 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1104 {
1105         conf_t *conf = mddev->private;
1106         int err = -EEXIST;
1107         int mirror;
1108         mirror_info_t *p;
1109         int first = 0;
1110         int last = conf->raid_disks - 1;
1111
1112         if (mddev->recovery_cp < MaxSector)
1113                 /* only hot-add to in-sync arrays, as recovery is
1114                  * very different from resync
1115                  */
1116                 return -EBUSY;
1117         if (!enough(conf))
1118                 return -EINVAL;
1119
1120         if (rdev->raid_disk >= 0)
1121                 first = last = rdev->raid_disk;
1122
1123         if (rdev->saved_raid_disk >= 0 &&
1124             rdev->saved_raid_disk >= first &&
1125             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1126                 mirror = rdev->saved_raid_disk;
1127         else
1128                 mirror = first;
1129         for ( ; mirror <= last ; mirror++)
1130                 if ( !(p=conf->mirrors+mirror)->rdev) {
1131
1132                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1133                                           rdev->data_offset << 9);
1134                         /* as we don't honour merge_bvec_fn, we must
1135                          * never risk violating it, so limit
1136                          * ->max_segments to one lying with a single
1137                          * page, as a one page request is never in
1138                          * violation.
1139                          */
1140                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1141                                 blk_queue_max_segments(mddev->queue, 1);
1142                                 blk_queue_segment_boundary(mddev->queue,
1143                                                            PAGE_CACHE_SIZE - 1);
1144                         }
1145
1146                         p->head_position = 0;
1147                         rdev->raid_disk = mirror;
1148                         err = 0;
1149                         if (rdev->saved_raid_disk != mirror)
1150                                 conf->fullsync = 1;
1151                         rcu_assign_pointer(p->rdev, rdev);
1152                         break;
1153                 }
1154
1155         md_integrity_add_rdev(rdev, mddev);
1156         print_conf(conf);
1157         return err;
1158 }
1159
1160 static int raid10_remove_disk(mddev_t *mddev, int number)
1161 {
1162         conf_t *conf = mddev->private;
1163         int err = 0;
1164         mdk_rdev_t *rdev;
1165         mirror_info_t *p = conf->mirrors+ number;
1166
1167         print_conf(conf);
1168         rdev = p->rdev;
1169         if (rdev) {
1170                 if (test_bit(In_sync, &rdev->flags) ||
1171                     atomic_read(&rdev->nr_pending)) {
1172                         err = -EBUSY;
1173                         goto abort;
1174                 }
1175                 /* Only remove faulty devices in recovery
1176                  * is not possible.
1177                  */
1178                 if (!test_bit(Faulty, &rdev->flags) &&
1179                     enough(conf)) {
1180                         err = -EBUSY;
1181                         goto abort;
1182                 }
1183                 p->rdev = NULL;
1184                 synchronize_rcu();
1185                 if (atomic_read(&rdev->nr_pending)) {
1186                         /* lost the race, try later */
1187                         err = -EBUSY;
1188                         p->rdev = rdev;
1189                         goto abort;
1190                 }
1191                 err = md_integrity_register(mddev);
1192         }
1193 abort:
1194
1195         print_conf(conf);
1196         return err;
1197 }
1198
1199
1200 static void end_sync_read(struct bio *bio, int error)
1201 {
1202         r10bio_t *r10_bio = bio->bi_private;
1203         conf_t *conf = r10_bio->mddev->private;
1204         int i,d;
1205
1206         for (i=0; i<conf->copies; i++)
1207                 if (r10_bio->devs[i].bio == bio)
1208                         break;
1209         BUG_ON(i == conf->copies);
1210         update_head_pos(i, r10_bio);
1211         d = r10_bio->devs[i].devnum;
1212
1213         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1214                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1215         else {
1216                 atomic_add(r10_bio->sectors,
1217                            &conf->mirrors[d].rdev->corrected_errors);
1218                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1219                         md_error(r10_bio->mddev,
1220                                  conf->mirrors[d].rdev);
1221         }
1222
1223         /* for reconstruct, we always reschedule after a read.
1224          * for resync, only after all reads
1225          */
1226         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1227         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1228             atomic_dec_and_test(&r10_bio->remaining)) {
1229                 /* we have read all the blocks,
1230                  * do the comparison in process context in raid10d
1231                  */
1232                 reschedule_retry(r10_bio);
1233         }
1234 }
1235
1236 static void end_sync_write(struct bio *bio, int error)
1237 {
1238         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1239         r10bio_t *r10_bio = bio->bi_private;
1240         mddev_t *mddev = r10_bio->mddev;
1241         conf_t *conf = mddev->private;
1242         int i,d;
1243
1244         for (i = 0; i < conf->copies; i++)
1245                 if (r10_bio->devs[i].bio == bio)
1246                         break;
1247         d = r10_bio->devs[i].devnum;
1248
1249         if (!uptodate)
1250                 md_error(mddev, conf->mirrors[d].rdev);
1251
1252         update_head_pos(i, r10_bio);
1253
1254         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1255         while (atomic_dec_and_test(&r10_bio->remaining)) {
1256                 if (r10_bio->master_bio == NULL) {
1257                         /* the primary of several recovery bios */
1258                         sector_t s = r10_bio->sectors;
1259                         put_buf(r10_bio);
1260                         md_done_sync(mddev, s, 1);
1261                         break;
1262                 } else {
1263                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1264                         put_buf(r10_bio);
1265                         r10_bio = r10_bio2;
1266                 }
1267         }
1268 }
1269
1270 /*
1271  * Note: sync and recover and handled very differently for raid10
1272  * This code is for resync.
1273  * For resync, we read through virtual addresses and read all blocks.
1274  * If there is any error, we schedule a write.  The lowest numbered
1275  * drive is authoritative.
1276  * However requests come for physical address, so we need to map.
1277  * For every physical address there are raid_disks/copies virtual addresses,
1278  * which is always are least one, but is not necessarly an integer.
1279  * This means that a physical address can span multiple chunks, so we may
1280  * have to submit multiple io requests for a single sync request.
1281  */
1282 /*
1283  * We check if all blocks are in-sync and only write to blocks that
1284  * aren't in sync
1285  */
1286 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1287 {
1288         conf_t *conf = mddev->private;
1289         int i, first;
1290         struct bio *tbio, *fbio;
1291
1292         atomic_set(&r10_bio->remaining, 1);
1293
1294         /* find the first device with a block */
1295         for (i=0; i<conf->copies; i++)
1296                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1297                         break;
1298
1299         if (i == conf->copies)
1300                 goto done;
1301
1302         first = i;
1303         fbio = r10_bio->devs[i].bio;
1304
1305         /* now find blocks with errors */
1306         for (i=0 ; i < conf->copies ; i++) {
1307                 int  j, d;
1308                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1309
1310                 tbio = r10_bio->devs[i].bio;
1311
1312                 if (tbio->bi_end_io != end_sync_read)
1313                         continue;
1314                 if (i == first)
1315                         continue;
1316                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1317                         /* We know that the bi_io_vec layout is the same for
1318                          * both 'first' and 'i', so we just compare them.
1319                          * All vec entries are PAGE_SIZE;
1320                          */
1321                         for (j = 0; j < vcnt; j++)
1322                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1323                                            page_address(tbio->bi_io_vec[j].bv_page),
1324                                            PAGE_SIZE))
1325                                         break;
1326                         if (j == vcnt)
1327                                 continue;
1328                         mddev->resync_mismatches += r10_bio->sectors;
1329                 }
1330                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1331                         /* Don't fix anything. */
1332                         continue;
1333                 /* Ok, we need to write this bio
1334                  * First we need to fixup bv_offset, bv_len and
1335                  * bi_vecs, as the read request might have corrupted these
1336                  */
1337                 tbio->bi_vcnt = vcnt;
1338                 tbio->bi_size = r10_bio->sectors << 9;
1339                 tbio->bi_idx = 0;
1340                 tbio->bi_phys_segments = 0;
1341                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1342                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1343                 tbio->bi_next = NULL;
1344                 tbio->bi_rw = WRITE;
1345                 tbio->bi_private = r10_bio;
1346                 tbio->bi_sector = r10_bio->devs[i].addr;
1347
1348                 for (j=0; j < vcnt ; j++) {
1349                         tbio->bi_io_vec[j].bv_offset = 0;
1350                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1351
1352                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1353                                page_address(fbio->bi_io_vec[j].bv_page),
1354                                PAGE_SIZE);
1355                 }
1356                 tbio->bi_end_io = end_sync_write;
1357
1358                 d = r10_bio->devs[i].devnum;
1359                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1360                 atomic_inc(&r10_bio->remaining);
1361                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1362
1363                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1364                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1365                 generic_make_request(tbio);
1366         }
1367
1368 done:
1369         if (atomic_dec_and_test(&r10_bio->remaining)) {
1370                 md_done_sync(mddev, r10_bio->sectors, 1);
1371                 put_buf(r10_bio);
1372         }
1373 }
1374
1375 /*
1376  * Now for the recovery code.
1377  * Recovery happens across physical sectors.
1378  * We recover all non-is_sync drives by finding the virtual address of
1379  * each, and then choose a working drive that also has that virt address.
1380  * There is a separate r10_bio for each non-in_sync drive.
1381  * Only the first two slots are in use. The first for reading,
1382  * The second for writing.
1383  *
1384  */
1385
1386 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1387 {
1388         conf_t *conf = mddev->private;
1389         int i, d;
1390         struct bio *bio, *wbio;
1391
1392
1393         /* move the pages across to the second bio
1394          * and submit the write request
1395          */
1396         bio = r10_bio->devs[0].bio;
1397         wbio = r10_bio->devs[1].bio;
1398         for (i=0; i < wbio->bi_vcnt; i++) {
1399                 struct page *p = bio->bi_io_vec[i].bv_page;
1400                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1401                 wbio->bi_io_vec[i].bv_page = p;
1402         }
1403         d = r10_bio->devs[1].devnum;
1404
1405         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1406         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1407         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1408                 generic_make_request(wbio);
1409         else
1410                 bio_endio(wbio, -EIO);
1411 }
1412
1413
1414 /*
1415  * Used by fix_read_error() to decay the per rdev read_errors.
1416  * We halve the read error count for every hour that has elapsed
1417  * since the last recorded read error.
1418  *
1419  */
1420 static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1421 {
1422         struct timespec cur_time_mon;
1423         unsigned long hours_since_last;
1424         unsigned int read_errors = atomic_read(&rdev->read_errors);
1425
1426         ktime_get_ts(&cur_time_mon);
1427
1428         if (rdev->last_read_error.tv_sec == 0 &&
1429             rdev->last_read_error.tv_nsec == 0) {
1430                 /* first time we've seen a read error */
1431                 rdev->last_read_error = cur_time_mon;
1432                 return;
1433         }
1434
1435         hours_since_last = (cur_time_mon.tv_sec -
1436                             rdev->last_read_error.tv_sec) / 3600;
1437
1438         rdev->last_read_error = cur_time_mon;
1439
1440         /*
1441          * if hours_since_last is > the number of bits in read_errors
1442          * just set read errors to 0. We do this to avoid
1443          * overflowing the shift of read_errors by hours_since_last.
1444          */
1445         if (hours_since_last >= 8 * sizeof(read_errors))
1446                 atomic_set(&rdev->read_errors, 0);
1447         else
1448                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1449 }
1450
1451 /*
1452  * This is a kernel thread which:
1453  *
1454  *      1.      Retries failed read operations on working mirrors.
1455  *      2.      Updates the raid superblock when problems encounter.
1456  *      3.      Performs writes following reads for array synchronising.
1457  */
1458
1459 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1460 {
1461         int sect = 0; /* Offset from r10_bio->sector */
1462         int sectors = r10_bio->sectors;
1463         mdk_rdev_t*rdev;
1464         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1465         int d = r10_bio->devs[r10_bio->read_slot].devnum;
1466
1467         rcu_read_lock();
1468         rdev = rcu_dereference(conf->mirrors[d].rdev);
1469         if (rdev) { /* If rdev is not NULL */
1470                 char b[BDEVNAME_SIZE];
1471                 int cur_read_error_count = 0;
1472
1473                 bdevname(rdev->bdev, b);
1474
1475                 if (test_bit(Faulty, &rdev->flags)) {
1476                         rcu_read_unlock();
1477                         /* drive has already been failed, just ignore any
1478                            more fix_read_error() attempts */
1479                         return;
1480                 }
1481
1482                 check_decay_read_errors(mddev, rdev);
1483                 atomic_inc(&rdev->read_errors);
1484                 cur_read_error_count = atomic_read(&rdev->read_errors);
1485                 if (cur_read_error_count > max_read_errors) {
1486                         rcu_read_unlock();
1487                         printk(KERN_NOTICE
1488                                "md/raid10:%s: %s: Raid device exceeded "
1489                                "read_error threshold "
1490                                "[cur %d:max %d]\n",
1491                                mdname(mddev),
1492                                b, cur_read_error_count, max_read_errors);
1493                         printk(KERN_NOTICE
1494                                "md/raid10:%s: %s: Failing raid "
1495                                "device\n", mdname(mddev), b);
1496                         md_error(mddev, conf->mirrors[d].rdev);
1497                         return;
1498                 }
1499         }
1500         rcu_read_unlock();
1501
1502         while(sectors) {
1503                 int s = sectors;
1504                 int sl = r10_bio->read_slot;
1505                 int success = 0;
1506                 int start;
1507
1508                 if (s > (PAGE_SIZE>>9))
1509                         s = PAGE_SIZE >> 9;
1510
1511                 rcu_read_lock();
1512                 do {
1513                         d = r10_bio->devs[sl].devnum;
1514                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1515                         if (rdev &&
1516                             test_bit(In_sync, &rdev->flags)) {
1517                                 atomic_inc(&rdev->nr_pending);
1518                                 rcu_read_unlock();
1519                                 success = sync_page_io(rdev,
1520                                                        r10_bio->devs[sl].addr +
1521                                                        sect,
1522                                                        s<<9,
1523                                                        conf->tmppage, READ, false);
1524                                 rdev_dec_pending(rdev, mddev);
1525                                 rcu_read_lock();
1526                                 if (success)
1527                                         break;
1528                         }
1529                         sl++;
1530                         if (sl == conf->copies)
1531                                 sl = 0;
1532                 } while (!success && sl != r10_bio->read_slot);
1533                 rcu_read_unlock();
1534
1535                 if (!success) {
1536                         /* Cannot read from anywhere -- bye bye array */
1537                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1538                         md_error(mddev, conf->mirrors[dn].rdev);
1539                         break;
1540                 }
1541
1542                 start = sl;
1543                 /* write it back and re-read */
1544                 rcu_read_lock();
1545                 while (sl != r10_bio->read_slot) {
1546                         char b[BDEVNAME_SIZE];
1547
1548                         if (sl==0)
1549                                 sl = conf->copies;
1550                         sl--;
1551                         d = r10_bio->devs[sl].devnum;
1552                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1553                         if (rdev &&
1554                             test_bit(In_sync, &rdev->flags)) {
1555                                 atomic_inc(&rdev->nr_pending);
1556                                 rcu_read_unlock();
1557                                 atomic_add(s, &rdev->corrected_errors);
1558                                 if (sync_page_io(rdev,
1559                                                  r10_bio->devs[sl].addr +
1560                                                  sect,
1561                                                  s<<9, conf->tmppage, WRITE, false)
1562                                     == 0) {
1563                                         /* Well, this device is dead */
1564                                         printk(KERN_NOTICE
1565                                                "md/raid10:%s: read correction "
1566                                                "write failed"
1567                                                " (%d sectors at %llu on %s)\n",
1568                                                mdname(mddev), s,
1569                                                (unsigned long long)(sect+
1570                                                rdev->data_offset),
1571                                                bdevname(rdev->bdev, b));
1572                                         printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1573                                                "drive\n",
1574                                                mdname(mddev),
1575                                                bdevname(rdev->bdev, b));
1576                                         md_error(mddev, rdev);
1577                                 }
1578                                 rdev_dec_pending(rdev, mddev);
1579                                 rcu_read_lock();
1580                         }
1581                 }
1582                 sl = start;
1583                 while (sl != r10_bio->read_slot) {
1584
1585                         if (sl==0)
1586                                 sl = conf->copies;
1587                         sl--;
1588                         d = r10_bio->devs[sl].devnum;
1589                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1590                         if (rdev &&
1591                             test_bit(In_sync, &rdev->flags)) {
1592                                 char b[BDEVNAME_SIZE];
1593                                 atomic_inc(&rdev->nr_pending);
1594                                 rcu_read_unlock();
1595                                 if (sync_page_io(rdev,
1596                                                  r10_bio->devs[sl].addr +
1597                                                  sect,
1598                                                  s<<9, conf->tmppage,
1599                                                  READ, false) == 0) {
1600                                         /* Well, this device is dead */
1601                                         printk(KERN_NOTICE
1602                                                "md/raid10:%s: unable to read back "
1603                                                "corrected sectors"
1604                                                " (%d sectors at %llu on %s)\n",
1605                                                mdname(mddev), s,
1606                                                (unsigned long long)(sect+
1607                                                     rdev->data_offset),
1608                                                bdevname(rdev->bdev, b));
1609                                         printk(KERN_NOTICE "md/raid10:%s: %s: failing drive\n",
1610                                                mdname(mddev),
1611                                                bdevname(rdev->bdev, b));
1612
1613                                         md_error(mddev, rdev);
1614                                 } else {
1615                                         printk(KERN_INFO
1616                                                "md/raid10:%s: read error corrected"
1617                                                " (%d sectors at %llu on %s)\n",
1618                                                mdname(mddev), s,
1619                                                (unsigned long long)(sect+
1620                                                     rdev->data_offset),
1621                                                bdevname(rdev->bdev, b));
1622                                 }
1623
1624                                 rdev_dec_pending(rdev, mddev);
1625                                 rcu_read_lock();
1626                         }
1627                 }
1628                 rcu_read_unlock();
1629
1630                 sectors -= s;
1631                 sect += s;
1632         }
1633 }
1634
1635 static void raid10d(mddev_t *mddev)
1636 {
1637         r10bio_t *r10_bio;
1638         struct bio *bio;
1639         unsigned long flags;
1640         conf_t *conf = mddev->private;
1641         struct list_head *head = &conf->retry_list;
1642         mdk_rdev_t *rdev;
1643
1644         md_check_recovery(mddev);
1645
1646         for (;;) {
1647                 char b[BDEVNAME_SIZE];
1648
1649                 flush_pending_writes(conf);
1650
1651                 spin_lock_irqsave(&conf->device_lock, flags);
1652                 if (list_empty(head)) {
1653                         spin_unlock_irqrestore(&conf->device_lock, flags);
1654                         break;
1655                 }
1656                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1657                 list_del(head->prev);
1658                 conf->nr_queued--;
1659                 spin_unlock_irqrestore(&conf->device_lock, flags);
1660
1661                 mddev = r10_bio->mddev;
1662                 conf = mddev->private;
1663                 if (test_bit(R10BIO_IsSync, &r10_bio->state))
1664                         sync_request_write(mddev, r10_bio);
1665                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
1666                         recovery_request_write(mddev, r10_bio);
1667                 else {
1668                         int mirror;
1669                         /* we got a read error. Maybe the drive is bad.  Maybe just
1670                          * the block and we can fix it.
1671                          * We freeze all other IO, and try reading the block from
1672                          * other devices.  When we find one, we re-write
1673                          * and check it that fixes the read error.
1674                          * This is all done synchronously while the array is
1675                          * frozen.
1676                          */
1677                         if (mddev->ro == 0) {
1678                                 freeze_array(conf);
1679                                 fix_read_error(conf, mddev, r10_bio);
1680                                 unfreeze_array(conf);
1681                         }
1682
1683                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1684                         r10_bio->devs[r10_bio->read_slot].bio =
1685                                 mddev->ro ? IO_BLOCKED : NULL;
1686                         mirror = read_balance(conf, r10_bio);
1687                         if (mirror == -1) {
1688                                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
1689                                        " read error for block %llu\n",
1690                                        mdname(mddev),
1691                                        bdevname(bio->bi_bdev,b),
1692                                        (unsigned long long)r10_bio->sector);
1693                                 raid_end_bio_io(r10_bio);
1694                                 bio_put(bio);
1695                         } else {
1696                                 const unsigned long do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
1697                                 bio_put(bio);
1698                                 rdev = conf->mirrors[mirror].rdev;
1699                                 if (printk_ratelimit())
1700                                         printk(KERN_ERR "md/raid10:%s: %s: redirecting sector %llu to"
1701                                                " another mirror\n",
1702                                                mdname(mddev),
1703                                                bdevname(rdev->bdev,b),
1704                                                (unsigned long long)r10_bio->sector);
1705                                 bio = bio_clone_mddev(r10_bio->master_bio,
1706                                                       GFP_NOIO, mddev);
1707                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1708                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1709                                         + rdev->data_offset;
1710                                 bio->bi_bdev = rdev->bdev;
1711                                 bio->bi_rw = READ | do_sync;
1712                                 bio->bi_private = r10_bio;
1713                                 bio->bi_end_io = raid10_end_read_request;
1714                                 generic_make_request(bio);
1715                         }
1716                 }
1717                 cond_resched();
1718         }
1719 }
1720
1721
1722 static int init_resync(conf_t *conf)
1723 {
1724         int buffs;
1725
1726         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1727         BUG_ON(conf->r10buf_pool);
1728         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1729         if (!conf->r10buf_pool)
1730                 return -ENOMEM;
1731         conf->next_resync = 0;
1732         return 0;
1733 }
1734
1735 /*
1736  * perform a "sync" on one "block"
1737  *
1738  * We need to make sure that no normal I/O request - particularly write
1739  * requests - conflict with active sync requests.
1740  *
1741  * This is achieved by tracking pending requests and a 'barrier' concept
1742  * that can be installed to exclude normal IO requests.
1743  *
1744  * Resync and recovery are handled very differently.
1745  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1746  *
1747  * For resync, we iterate over virtual addresses, read all copies,
1748  * and update if there are differences.  If only one copy is live,
1749  * skip it.
1750  * For recovery, we iterate over physical addresses, read a good
1751  * value for each non-in_sync drive, and over-write.
1752  *
1753  * So, for recovery we may have several outstanding complex requests for a
1754  * given address, one for each out-of-sync device.  We model this by allocating
1755  * a number of r10_bio structures, one for each out-of-sync device.
1756  * As we setup these structures, we collect all bio's together into a list
1757  * which we then process collectively to add pages, and then process again
1758  * to pass to generic_make_request.
1759  *
1760  * The r10_bio structures are linked using a borrowed master_bio pointer.
1761  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1762  * has its remaining count decremented to 0, the whole complex operation
1763  * is complete.
1764  *
1765  */
1766
1767 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1768 {
1769         conf_t *conf = mddev->private;
1770         r10bio_t *r10_bio;
1771         struct bio *biolist = NULL, *bio;
1772         sector_t max_sector, nr_sectors;
1773         int disk;
1774         int i;
1775         int max_sync;
1776         sector_t sync_blocks;
1777
1778         sector_t sectors_skipped = 0;
1779         int chunks_skipped = 0;
1780
1781         if (!conf->r10buf_pool)
1782                 if (init_resync(conf))
1783                         return 0;
1784
1785  skipped:
1786         max_sector = mddev->dev_sectors;
1787         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1788                 max_sector = mddev->resync_max_sectors;
1789         if (sector_nr >= max_sector) {
1790                 /* If we aborted, we need to abort the
1791                  * sync on the 'current' bitmap chucks (there can
1792                  * be several when recovering multiple devices).
1793                  * as we may have started syncing it but not finished.
1794                  * We can find the current address in
1795                  * mddev->curr_resync, but for recovery,
1796                  * we need to convert that to several
1797                  * virtual addresses.
1798                  */
1799                 if (mddev->curr_resync < max_sector) { /* aborted */
1800                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1801                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1802                                                 &sync_blocks, 1);
1803                         else for (i=0; i<conf->raid_disks; i++) {
1804                                 sector_t sect =
1805                                         raid10_find_virt(conf, mddev->curr_resync, i);
1806                                 bitmap_end_sync(mddev->bitmap, sect,
1807                                                 &sync_blocks, 1);
1808                         }
1809                 } else /* completed sync */
1810                         conf->fullsync = 0;
1811
1812                 bitmap_close_sync(mddev->bitmap);
1813                 close_sync(conf);
1814                 *skipped = 1;
1815                 return sectors_skipped;
1816         }
1817         if (chunks_skipped >= conf->raid_disks) {
1818                 /* if there has been nothing to do on any drive,
1819                  * then there is nothing to do at all..
1820                  */
1821                 *skipped = 1;
1822                 return (max_sector - sector_nr) + sectors_skipped;
1823         }
1824
1825         if (max_sector > mddev->resync_max)
1826                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1827
1828         /* make sure whole request will fit in a chunk - if chunks
1829          * are meaningful
1830          */
1831         if (conf->near_copies < conf->raid_disks &&
1832             max_sector > (sector_nr | conf->chunk_mask))
1833                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1834         /*
1835          * If there is non-resync activity waiting for us then
1836          * put in a delay to throttle resync.
1837          */
1838         if (!go_faster && conf->nr_waiting)
1839                 msleep_interruptible(1000);
1840
1841         /* Again, very different code for resync and recovery.
1842          * Both must result in an r10bio with a list of bios that
1843          * have bi_end_io, bi_sector, bi_bdev set,
1844          * and bi_private set to the r10bio.
1845          * For recovery, we may actually create several r10bios
1846          * with 2 bios in each, that correspond to the bios in the main one.
1847          * In this case, the subordinate r10bios link back through a
1848          * borrowed master_bio pointer, and the counter in the master
1849          * includes a ref from each subordinate.
1850          */
1851         /* First, we decide what to do and set ->bi_end_io
1852          * To end_sync_read if we want to read, and
1853          * end_sync_write if we will want to write.
1854          */
1855
1856         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1857         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1858                 /* recovery... the complicated one */
1859                 int j, k;
1860                 r10_bio = NULL;
1861
1862                 for (i=0 ; i<conf->raid_disks; i++)
1863                         if (conf->mirrors[i].rdev &&
1864                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1865                                 int still_degraded = 0;
1866                                 /* want to reconstruct this device */
1867                                 r10bio_t *rb2 = r10_bio;
1868                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1869                                 int must_sync;
1870                                 /* Unless we are doing a full sync, we only need
1871                                  * to recover the block if it is set in the bitmap
1872                                  */
1873                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1874                                                               &sync_blocks, 1);
1875                                 if (sync_blocks < max_sync)
1876                                         max_sync = sync_blocks;
1877                                 if (!must_sync &&
1878                                     !conf->fullsync) {
1879                                         /* yep, skip the sync_blocks here, but don't assume
1880                                          * that there will never be anything to do here
1881                                          */
1882                                         chunks_skipped = -1;
1883                                         continue;
1884                                 }
1885
1886                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1887                                 raise_barrier(conf, rb2 != NULL);
1888                                 atomic_set(&r10_bio->remaining, 0);
1889
1890                                 r10_bio->master_bio = (struct bio*)rb2;
1891                                 if (rb2)
1892                                         atomic_inc(&rb2->remaining);
1893                                 r10_bio->mddev = mddev;
1894                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1895                                 r10_bio->sector = sect;
1896
1897                                 raid10_find_phys(conf, r10_bio);
1898
1899                                 /* Need to check if the array will still be
1900                                  * degraded
1901                                  */
1902                                 for (j=0; j<conf->raid_disks; j++)
1903                                         if (conf->mirrors[j].rdev == NULL ||
1904                                             test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
1905                                                 still_degraded = 1;
1906                                                 break;
1907                                         }
1908
1909                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1910                                                               &sync_blocks, still_degraded);
1911
1912                                 for (j=0; j<conf->copies;j++) {
1913                                         int d = r10_bio->devs[j].devnum;
1914                                         if (conf->mirrors[d].rdev &&
1915                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1916                                                 /* This is where we read from */
1917                                                 bio = r10_bio->devs[0].bio;
1918                                                 bio->bi_next = biolist;
1919                                                 biolist = bio;
1920                                                 bio->bi_private = r10_bio;
1921                                                 bio->bi_end_io = end_sync_read;
1922                                                 bio->bi_rw = READ;
1923                                                 bio->bi_sector = r10_bio->devs[j].addr +
1924                                                         conf->mirrors[d].rdev->data_offset;
1925                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1926                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1927                                                 atomic_inc(&r10_bio->remaining);
1928                                                 /* and we write to 'i' */
1929
1930                                                 for (k=0; k<conf->copies; k++)
1931                                                         if (r10_bio->devs[k].devnum == i)
1932                                                                 break;
1933                                                 BUG_ON(k == conf->copies);
1934                                                 bio = r10_bio->devs[1].bio;
1935                                                 bio->bi_next = biolist;
1936                                                 biolist = bio;
1937                                                 bio->bi_private = r10_bio;
1938                                                 bio->bi_end_io = end_sync_write;
1939                                                 bio->bi_rw = WRITE;
1940                                                 bio->bi_sector = r10_bio->devs[k].addr +
1941                                                         conf->mirrors[i].rdev->data_offset;
1942                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1943
1944                                                 r10_bio->devs[0].devnum = d;
1945                                                 r10_bio->devs[1].devnum = i;
1946
1947                                                 break;
1948                                         }
1949                                 }
1950                                 if (j == conf->copies) {
1951                                         /* Cannot recover, so abort the recovery */
1952                                         put_buf(r10_bio);
1953                                         if (rb2)
1954                                                 atomic_dec(&rb2->remaining);
1955                                         r10_bio = rb2;
1956                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
1957                                                               &mddev->recovery))
1958                                                 printk(KERN_INFO "md/raid10:%s: insufficient "
1959                                                        "working devices for recovery.\n",
1960                                                        mdname(mddev));
1961                                         break;
1962                                 }
1963                         }
1964                 if (biolist == NULL) {
1965                         while (r10_bio) {
1966                                 r10bio_t *rb2 = r10_bio;
1967                                 r10_bio = (r10bio_t*) rb2->master_bio;
1968                                 rb2->master_bio = NULL;
1969                                 put_buf(rb2);
1970                         }
1971                         goto giveup;
1972                 }
1973         } else {
1974                 /* resync. Schedule a read for every block at this virt offset */
1975                 int count = 0;
1976
1977                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1978
1979                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1980                                        &sync_blocks, mddev->degraded) &&
1981                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1982                         /* We can skip this block */
1983                         *skipped = 1;
1984                         return sync_blocks + sectors_skipped;
1985                 }
1986                 if (sync_blocks < max_sync)
1987                         max_sync = sync_blocks;
1988                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1989
1990                 r10_bio->mddev = mddev;
1991                 atomic_set(&r10_bio->remaining, 0);
1992                 raise_barrier(conf, 0);
1993                 conf->next_resync = sector_nr;
1994
1995                 r10_bio->master_bio = NULL;
1996                 r10_bio->sector = sector_nr;
1997                 set_bit(R10BIO_IsSync, &r10_bio->state);
1998                 raid10_find_phys(conf, r10_bio);
1999                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2000
2001                 for (i=0; i<conf->copies; i++) {
2002                         int d = r10_bio->devs[i].devnum;
2003                         bio = r10_bio->devs[i].bio;
2004                         bio->bi_end_io = NULL;
2005                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
2006                         if (conf->mirrors[d].rdev == NULL ||
2007                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2008                                 continue;
2009                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2010                         atomic_inc(&r10_bio->remaining);
2011                         bio->bi_next = biolist;
2012                         biolist = bio;
2013                         bio->bi_private = r10_bio;
2014                         bio->bi_end_io = end_sync_read;
2015                         bio->bi_rw = READ;
2016                         bio->bi_sector = r10_bio->devs[i].addr +
2017                                 conf->mirrors[d].rdev->data_offset;
2018                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2019                         count++;
2020                 }
2021
2022                 if (count < 2) {
2023                         for (i=0; i<conf->copies; i++) {
2024                                 int d = r10_bio->devs[i].devnum;
2025                                 if (r10_bio->devs[i].bio->bi_end_io)
2026                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
2027                         }
2028                         put_buf(r10_bio);
2029                         biolist = NULL;
2030                         goto giveup;
2031                 }
2032         }
2033
2034         for (bio = biolist; bio ; bio=bio->bi_next) {
2035
2036                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2037                 if (bio->bi_end_io)
2038                         bio->bi_flags |= 1 << BIO_UPTODATE;
2039                 bio->bi_vcnt = 0;
2040                 bio->bi_idx = 0;
2041                 bio->bi_phys_segments = 0;
2042                 bio->bi_size = 0;
2043         }
2044
2045         nr_sectors = 0;
2046         if (sector_nr + max_sync < max_sector)
2047                 max_sector = sector_nr + max_sync;
2048         do {
2049                 struct page *page;
2050                 int len = PAGE_SIZE;
2051                 disk = 0;
2052                 if (sector_nr + (len>>9) > max_sector)
2053                         len = (max_sector - sector_nr) << 9;
2054                 if (len == 0)
2055                         break;
2056                 for (bio= biolist ; bio ; bio=bio->bi_next) {
2057                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2058                         if (bio_add_page(bio, page, len, 0) == 0) {
2059                                 /* stop here */
2060                                 struct bio *bio2;
2061                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2062                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
2063                                         /* remove last page from this bio */
2064                                         bio2->bi_vcnt--;
2065                                         bio2->bi_size -= len;
2066                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2067                                 }
2068                                 goto bio_full;
2069                         }
2070                         disk = i;
2071                 }
2072                 nr_sectors += len>>9;
2073                 sector_nr += len>>9;
2074         } while (biolist->bi_vcnt < RESYNC_PAGES);
2075  bio_full:
2076         r10_bio->sectors = nr_sectors;
2077
2078         while (biolist) {
2079                 bio = biolist;
2080                 biolist = biolist->bi_next;
2081
2082                 bio->bi_next = NULL;
2083                 r10_bio = bio->bi_private;
2084                 r10_bio->sectors = nr_sectors;
2085
2086                 if (bio->bi_end_io == end_sync_read) {
2087                         md_sync_acct(bio->bi_bdev, nr_sectors);
2088                         generic_make_request(bio);
2089                 }
2090         }
2091
2092         if (sectors_skipped)
2093                 /* pretend they weren't skipped, it makes
2094                  * no important difference in this case
2095                  */
2096                 md_done_sync(mddev, sectors_skipped, 1);
2097
2098         return sectors_skipped + nr_sectors;
2099  giveup:
2100         /* There is nowhere to write, so all non-sync
2101          * drives must be failed, so try the next chunk...
2102          */
2103         if (sector_nr + max_sync < max_sector)
2104                 max_sector = sector_nr + max_sync;
2105
2106         sectors_skipped += (max_sector - sector_nr);
2107         chunks_skipped ++;
2108         sector_nr = max_sector;
2109         goto skipped;
2110 }
2111
2112 static sector_t
2113 raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2114 {
2115         sector_t size;
2116         conf_t *conf = mddev->private;
2117
2118         if (!raid_disks)
2119                 raid_disks = conf->raid_disks;
2120         if (!sectors)
2121                 sectors = conf->dev_sectors;
2122
2123         size = sectors >> conf->chunk_shift;
2124         sector_div(size, conf->far_copies);
2125         size = size * raid_disks;
2126         sector_div(size, conf->near_copies);
2127
2128         return size << conf->chunk_shift;
2129 }
2130
2131
2132 static conf_t *setup_conf(mddev_t *mddev)
2133 {
2134         conf_t *conf = NULL;
2135         int nc, fc, fo;
2136         sector_t stride, size;
2137         int err = -EINVAL;
2138
2139         if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2140             !is_power_of_2(mddev->new_chunk_sectors)) {
2141                 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2142                        "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2143                        mdname(mddev), PAGE_SIZE);
2144                 goto out;
2145         }
2146
2147         nc = mddev->new_layout & 255;
2148         fc = (mddev->new_layout >> 8) & 255;
2149         fo = mddev->new_layout & (1<<16);
2150
2151         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2152             (mddev->new_layout >> 17)) {
2153                 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2154                        mdname(mddev), mddev->new_layout);
2155                 goto out;
2156         }
2157
2158         err = -ENOMEM;
2159         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2160         if (!conf)
2161                 goto out;
2162
2163         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2164                                 GFP_KERNEL);
2165         if (!conf->mirrors)
2166                 goto out;
2167
2168         conf->tmppage = alloc_page(GFP_KERNEL);
2169         if (!conf->tmppage)
2170                 goto out;
2171
2172
2173         conf->raid_disks = mddev->raid_disks;
2174         conf->near_copies = nc;
2175         conf->far_copies = fc;
2176         conf->copies = nc*fc;
2177         conf->far_offset = fo;
2178         conf->chunk_mask = mddev->new_chunk_sectors - 1;
2179         conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2180
2181         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2182                                            r10bio_pool_free, conf);
2183         if (!conf->r10bio_pool)
2184                 goto out;
2185
2186         size = mddev->dev_sectors >> conf->chunk_shift;
2187         sector_div(size, fc);
2188         size = size * conf->raid_disks;
2189         sector_div(size, nc);
2190         /* 'size' is now the number of chunks in the array */
2191         /* calculate "used chunks per device" in 'stride' */
2192         stride = size * conf->copies;
2193
2194         /* We need to round up when dividing by raid_disks to
2195          * get the stride size.
2196          */
2197         stride += conf->raid_disks - 1;
2198         sector_div(stride, conf->raid_disks);
2199
2200         conf->dev_sectors = stride << conf->chunk_shift;
2201
2202         if (fo)
2203                 stride = 1;
2204         else
2205                 sector_div(stride, fc);
2206         conf->stride = stride << conf->chunk_shift;
2207
2208
2209         spin_lock_init(&conf->device_lock);
2210         INIT_LIST_HEAD(&conf->retry_list);
2211
2212         spin_lock_init(&conf->resync_lock);
2213         init_waitqueue_head(&conf->wait_barrier);
2214
2215         conf->thread = md_register_thread(raid10d, mddev, NULL);
2216         if (!conf->thread)
2217                 goto out;
2218
2219         conf->mddev = mddev;
2220         return conf;
2221
2222  out:
2223         printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2224                mdname(mddev));
2225         if (conf) {
2226                 if (conf->r10bio_pool)
2227                         mempool_destroy(conf->r10bio_pool);
2228                 kfree(conf->mirrors);
2229                 safe_put_page(conf->tmppage);
2230                 kfree(conf);
2231         }
2232         return ERR_PTR(err);
2233 }
2234
2235 static int run(mddev_t *mddev)
2236 {
2237         conf_t *conf;
2238         int i, disk_idx, chunk_size;
2239         mirror_info_t *disk;
2240         mdk_rdev_t *rdev;
2241         sector_t size;
2242
2243         /*
2244          * copy the already verified devices into our private RAID10
2245          * bookkeeping area. [whatever we allocate in run(),
2246          * should be freed in stop()]
2247          */
2248
2249         if (mddev->private == NULL) {
2250                 conf = setup_conf(mddev);
2251                 if (IS_ERR(conf))
2252                         return PTR_ERR(conf);
2253                 mddev->private = conf;
2254         }
2255         conf = mddev->private;
2256         if (!conf)
2257                 goto out;
2258
2259         mddev->thread = conf->thread;
2260         conf->thread = NULL;
2261
2262         chunk_size = mddev->chunk_sectors << 9;
2263         blk_queue_io_min(mddev->queue, chunk_size);
2264         if (conf->raid_disks % conf->near_copies)
2265                 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2266         else
2267                 blk_queue_io_opt(mddev->queue, chunk_size *
2268                                  (conf->raid_disks / conf->near_copies));
2269
2270         list_for_each_entry(rdev, &mddev->disks, same_set) {
2271                 disk_idx = rdev->raid_disk;
2272                 if (disk_idx >= conf->raid_disks
2273                     || disk_idx < 0)
2274                         continue;
2275                 disk = conf->mirrors + disk_idx;
2276
2277                 disk->rdev = rdev;
2278                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2279                                   rdev->data_offset << 9);
2280                 /* as we don't honour merge_bvec_fn, we must never risk
2281                  * violating it, so limit max_segments to 1 lying
2282                  * within a single page.
2283                  */
2284                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2285                         blk_queue_max_segments(mddev->queue, 1);
2286                         blk_queue_segment_boundary(mddev->queue,
2287                                                    PAGE_CACHE_SIZE - 1);
2288                 }
2289
2290                 disk->head_position = 0;
2291         }
2292         /* need to check that every block has at least one working mirror */
2293         if (!enough(conf)) {
2294                 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2295                        mdname(mddev));
2296                 goto out_free_conf;
2297         }
2298
2299         mddev->degraded = 0;
2300         for (i = 0; i < conf->raid_disks; i++) {
2301
2302                 disk = conf->mirrors + i;
2303
2304                 if (!disk->rdev ||
2305                     !test_bit(In_sync, &disk->rdev->flags)) {
2306                         disk->head_position = 0;
2307                         mddev->degraded++;
2308                         if (disk->rdev)
2309                                 conf->fullsync = 1;
2310                 }
2311         }
2312
2313         if (mddev->recovery_cp != MaxSector)
2314                 printk(KERN_NOTICE "md/raid10:%s: not clean"
2315                        " -- starting background reconstruction\n",
2316                        mdname(mddev));
2317         printk(KERN_INFO
2318                 "md/raid10:%s: active with %d out of %d devices\n",
2319                 mdname(mddev), conf->raid_disks - mddev->degraded,
2320                 conf->raid_disks);
2321         /*
2322          * Ok, everything is just fine now
2323          */
2324         mddev->dev_sectors = conf->dev_sectors;
2325         size = raid10_size(mddev, 0, 0);
2326         md_set_array_sectors(mddev, size);
2327         mddev->resync_max_sectors = size;
2328
2329         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2330         mddev->queue->backing_dev_info.congested_data = mddev;
2331
2332         /* Calculate max read-ahead size.
2333          * We need to readahead at least twice a whole stripe....
2334          * maybe...
2335          */
2336         {
2337                 int stripe = conf->raid_disks *
2338                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2339                 stripe /= conf->near_copies;
2340                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2341                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2342         }
2343
2344         if (conf->near_copies < conf->raid_disks)
2345                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2346
2347         if (md_integrity_register(mddev))
2348                 goto out_free_conf;
2349
2350         return 0;
2351
2352 out_free_conf:
2353         md_unregister_thread(mddev->thread);
2354         if (conf->r10bio_pool)
2355                 mempool_destroy(conf->r10bio_pool);
2356         safe_put_page(conf->tmppage);
2357         kfree(conf->mirrors);
2358         kfree(conf);
2359         mddev->private = NULL;
2360 out:
2361         return -EIO;
2362 }
2363
2364 static int stop(mddev_t *mddev)
2365 {
2366         conf_t *conf = mddev->private;
2367
2368         raise_barrier(conf, 0);
2369         lower_barrier(conf);
2370
2371         md_unregister_thread(mddev->thread);
2372         mddev->thread = NULL;
2373         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2374         if (conf->r10bio_pool)
2375                 mempool_destroy(conf->r10bio_pool);
2376         kfree(conf->mirrors);
2377         kfree(conf);
2378         mddev->private = NULL;
2379         return 0;
2380 }
2381
2382 static void raid10_quiesce(mddev_t *mddev, int state)
2383 {
2384         conf_t *conf = mddev->private;
2385
2386         switch(state) {
2387         case 1:
2388                 raise_barrier(conf, 0);
2389                 break;
2390         case 0:
2391                 lower_barrier(conf);
2392                 break;
2393         }
2394 }
2395
2396 static void *raid10_takeover_raid0(mddev_t *mddev)
2397 {
2398         mdk_rdev_t *rdev;
2399         conf_t *conf;
2400
2401         if (mddev->degraded > 0) {
2402                 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
2403                        mdname(mddev));
2404                 return ERR_PTR(-EINVAL);
2405         }
2406
2407         /* Set new parameters */
2408         mddev->new_level = 10;
2409         /* new layout: far_copies = 1, near_copies = 2 */
2410         mddev->new_layout = (1<<8) + 2;
2411         mddev->new_chunk_sectors = mddev->chunk_sectors;
2412         mddev->delta_disks = mddev->raid_disks;
2413         mddev->raid_disks *= 2;
2414         /* make sure it will be not marked as dirty */
2415         mddev->recovery_cp = MaxSector;
2416
2417         conf = setup_conf(mddev);
2418         if (!IS_ERR(conf)) {
2419                 list_for_each_entry(rdev, &mddev->disks, same_set)
2420                         if (rdev->raid_disk >= 0)
2421                                 rdev->new_raid_disk = rdev->raid_disk * 2;
2422                 conf->barrier = 1;
2423         }
2424
2425         return conf;
2426 }
2427
2428 static void *raid10_takeover(mddev_t *mddev)
2429 {
2430         struct raid0_private_data *raid0_priv;
2431
2432         /* raid10 can take over:
2433          *  raid0 - providing it has only two drives
2434          */
2435         if (mddev->level == 0) {
2436                 /* for raid0 takeover only one zone is supported */
2437                 raid0_priv = mddev->private;
2438                 if (raid0_priv->nr_strip_zones > 1) {
2439                         printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
2440                                " with more than one zone.\n",
2441                                mdname(mddev));
2442                         return ERR_PTR(-EINVAL);
2443                 }
2444                 return raid10_takeover_raid0(mddev);
2445         }
2446         return ERR_PTR(-EINVAL);
2447 }
2448
2449 static struct mdk_personality raid10_personality =
2450 {
2451         .name           = "raid10",
2452         .level          = 10,
2453         .owner          = THIS_MODULE,
2454         .make_request   = make_request,
2455         .run            = run,
2456         .stop           = stop,
2457         .status         = status,
2458         .error_handler  = error,
2459         .hot_add_disk   = raid10_add_disk,
2460         .hot_remove_disk= raid10_remove_disk,
2461         .spare_active   = raid10_spare_active,
2462         .sync_request   = sync_request,
2463         .quiesce        = raid10_quiesce,
2464         .size           = raid10_size,
2465         .takeover       = raid10_takeover,
2466 };
2467
2468 static int __init raid_init(void)
2469 {
2470         return register_md_personality(&raid10_personality);
2471 }
2472
2473 static void raid_exit(void)
2474 {
2475         unregister_md_personality(&raid10_personality);
2476 }
2477
2478 module_init(raid_init);
2479 module_exit(raid_exit);
2480 MODULE_LICENSE("GPL");
2481 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
2482 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2483 MODULE_ALIAS("md-raid10");
2484 MODULE_ALIAS("md-level-10");