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 /* Barriers....
 638  * Sometimes we need to suspend IO while we do something else,
 639  * either some resync/recovery, or reconfigure the array.
 640  * To do this we raise a 'barrier'.
 641  * The 'barrier' is a counter that can be raised multiple times
 642  * to count how many activities are happening which preclude
 643  * normal IO.
 644  * We can only raise the barrier if there is no pending IO.
 645  * i.e. if nr_pending == 0.
 646  * We choose only to raise the barrier if no-one is waiting for the
 647  * barrier to go down.  This means that as soon as an IO request
 648  * is ready, no other operations which require a barrier will start
 649  * until the IO request has had a chance.
 650  *
 651  * So: regular IO calls 'wait_barrier'.  When that returns there
 652  *    is no backgroup IO happening,  It must arrange to call
 653  *    allow_barrier when it has finished its IO.
 654  * backgroup IO calls must call raise_barrier.  Once that returns
 655  *    there is no normal IO happeing.  It must arrange to call
 656  *    lower_barrier when the particular background IO completes.
 657  */
 658
 659 static void raise_barrier(conf_t *conf, int force)
 660 {
 661         BUG_ON(force && !conf->barrier);
 662         spin_lock_irq(&conf->resync_lock);
 663
 664         /* Wait until no block IO is waiting (unless 'force') */
 665         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 666                             conf->resync_lock, );
 667
 668         /* block any new IO from starting */
 669         conf->barrier++;
 670
 671         /* Now wait for all pending IO to complete */
 672         wait_event_lock_irq(conf->wait_barrier,
 673                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 674                             conf->resync_lock, );
 675
 676         spin_unlock_irq(&conf->resync_lock);
 677 }
 678
 679 static void lower_barrier(conf_t *conf)
 680 {
 681         unsigned long flags;
 682         spin_lock_irqsave(&conf->resync_lock, flags);
 683         conf->barrier--;
 684         spin_unlock_irqrestore(&conf->resync_lock, flags);
 685         wake_up(&conf->wait_barrier);
 686 }
 687
 688 static void wait_barrier(conf_t *conf)
 689 {
 690         spin_lock_irq(&conf->resync_lock);
 691         if (conf->barrier) {
 692                 conf->nr_waiting++;
 693                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 694                                     conf->resync_lock,
 695                                     );
 696                 conf->nr_waiting--;
 697         }
 698         conf->nr_pending++;
 699         spin_unlock_irq(&conf->resync_lock);
 700 }
 701
 702 static void allow_barrier(conf_t *conf)
 703 {
 704         unsigned long flags;
 705         spin_lock_irqsave(&conf->resync_lock, flags);
 706         conf->nr_pending--;
 707         spin_unlock_irqrestore(&conf->resync_lock, flags);
 708         wake_up(&conf->wait_barrier);
 709 }
 710
 711 static void freeze_array(conf_t *conf)
 712 {
 713         /* stop syncio and normal IO and wait for everything to
 714          * go quiet.
 715          * We increment barrier and nr_waiting, and then
 716          * wait until nr_pending match nr_queued+1
 717          * This is called in the context of one normal IO request
 718          * that has failed. Thus any sync request that might be pending
 719          * will be blocked by nr_pending, and we need to wait for
 720          * pending IO requests to complete or be queued for re-try.
 721          * Thus the number queued (nr_queued) plus this request (1)
 722          * must match the number of pending IOs (nr_pending) before
 723          * we continue.
 724          */
 725         spin_lock_irq(&conf->resync_lock);
 726         conf->barrier++;
 727         conf->nr_waiting++;
 728         wait_event_lock_irq(conf->wait_barrier,
 729                             conf->nr_pending == conf->nr_queued+1,
 730                             conf->resync_lock,
 731                             flush_pending_writes(conf));
 732
 733         spin_unlock_irq(&conf->resync_lock);
 734 }
 735
 736 static void unfreeze_array(conf_t *conf)
 737 {
 738         /* reverse the effect of the freeze */
 739         spin_lock_irq(&conf->resync_lock);
 740         conf->barrier--;
 741         conf->nr_waiting--;
 742         wake_up(&conf->wait_barrier);
 743         spin_unlock_irq(&conf->resync_lock);
 744 }
 745
 746 static int make_request(mddev_t *mddev, struct bio * bio)
 747 {
 748         conf_t *conf = mddev->private;
 749         mirror_info_t *mirror;
 750         r10bio_t *r10_bio;
 751         struct bio *read_bio;
 752         int i;
 753         int chunk_sects = conf->chunk_mask + 1;
 754         const int rw = bio_data_dir(bio);
 755         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 756         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
 757         unsigned long flags;
 758         mdk_rdev_t *blocked_rdev;
 759         int plugged;
 760
 761         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
 762                 md_flush_request(mddev, bio);
 763                 return 0;
 764         }
 765
 766         /* If this request crosses a chunk boundary, we need to
 767          * split it.  This will only happen for 1 PAGE (or less) requests.
 768          */
 769         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 770                       > chunk_sects &&
 771                     conf->near_copies < conf->raid_disks)) {
 772                 struct bio_pair *bp;
 773                 /* Sanity check -- queue functions should prevent this happening */
 774                 if (bio->bi_vcnt != 1 ||
 775                     bio->bi_idx != 0)
 776                         goto bad_map;
 777                 /* This is a one page bio that upper layers
 778                  * refuse to split for us, so we need to split it.
 779                  */
 780                 bp = bio_split(bio,
 781                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 782
 783                 /* Each of these 'make_request' calls will call 'wait_barrier'.
 784                  * If the first succeeds but the second blocks due to the resync
 785                  * thread raising the barrier, we will deadlock because the
 786                  * IO to the underlying device will be queued in generic_make_request
 787                  * and will never complete, so will never reduce nr_pending.
 788                  * So increment nr_waiting here so no new raise_barriers will
 789                  * succeed, and so the second wait_barrier cannot block.
 790                  */
 791                 spin_lock_irq(&conf->resync_lock);
 792                 conf->nr_waiting++;
 793                 spin_unlock_irq(&conf->resync_lock);
 794
 795                 if (make_request(mddev, &bp->bio1))
 796                         generic_make_request(&bp->bio1);
 797                 if (make_request(mddev, &bp->bio2))
 798                         generic_make_request(&bp->bio2);
 799
 800                 spin_lock_irq(&conf->resync_lock);
 801                 conf->nr_waiting--;
 802                 wake_up(&conf->wait_barrier);
 803                 spin_unlock_irq(&conf->resync_lock);
 804
 805                 bio_pair_release(bp);
 806                 return 0;
 807         bad_map:
 808                 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
 809                        " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
 810                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 811
 812                 bio_io_error(bio);
 813                 return 0;
 814         }
 815
 816         md_write_start(mddev, bio);
 817
 818         /*
 819          * Register the new request and wait if the reconstruction
 820          * thread has put up a bar for new requests.
 821          * Continue immediately if no resync is active currently.
 822          */
 823         wait_barrier(conf);
 824
 825         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 826
 827         r10_bio->master_bio = bio;
 828         r10_bio->sectors = bio->bi_size >> 9;
 829
 830         r10_bio->mddev = mddev;
 831         r10_bio->sector = bio->bi_sector;
 832         r10_bio->state = 0;
 833
 834         if (rw == READ) {
 835                 /*
 836                  * read balancing logic:
 837                  */
 838                 int disk = read_balance(conf, r10_bio);
 839                 int slot = r10_bio->read_slot;
 840                 if (disk < 0) {
 841                         raid_end_bio_io(r10_bio);
 842                         return 0;
 843                 }
 844                 mirror = conf->mirrors + disk;
 845
 846                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 847
 848                 r10_bio->devs[slot].bio = read_bio;
 849
 850                 read_bio->bi_sector = r10_bio->devs[slot].addr +
 851                         mirror->rdev->data_offset;
 852                 read_bio->bi_bdev = mirror->rdev->bdev;
 853                 read_bio->bi_end_io = raid10_end_read_request;
 854                 read_bio->bi_rw = READ | do_sync;
 855                 read_bio->bi_private = r10_bio;
 856
 857                 generic_make_request(read_bio);
 858                 return 0;
 859         }
 860
 861         /*
 862          * WRITE:
 863          */
 864         /* first select target devices under rcu_lock and
 865          * inc refcount on their rdev.  Record them by setting
 866          * bios[x] to bio
 867          */
 868         plugged = mddev_check_plugged(mddev);
 869
 870         raid10_find_phys(conf, r10_bio);
 871  retry_write:
 872         blocked_rdev = NULL;
 873         rcu_read_lock();
 874         for (i = 0;  i < conf->copies; i++) {
 875                 int d = r10_bio->devs[i].devnum;
 876                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
 877                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 878                         atomic_inc(&rdev->nr_pending);
 879                         blocked_rdev = rdev;
 880                         break;
 881                 }
 882                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 883                         atomic_inc(&rdev->nr_pending);
 884                         r10_bio->devs[i].bio = bio;
 885                 } else {
 886                         r10_bio->devs[i].bio = NULL;
 887                         set_bit(R10BIO_Degraded, &r10_bio->state);
 888                 }
 889         }
 890         rcu_read_unlock();
 891
 892         if (unlikely(blocked_rdev)) {
 893                 /* Have to wait for this device to get unblocked, then retry */
 894                 int j;
 895                 int d;
 896
 897                 for (j = 0; j < i; j++)
 898                         if (r10_bio->devs[j].bio) {
 899                                 d = r10_bio->devs[j].devnum;
 900                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
 901                         }
 902                 allow_barrier(conf);
 903                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 904                 wait_barrier(conf);
 905                 goto retry_write;
 906         }
 907
 908         atomic_set(&r10_bio->remaining, 1);
 909         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
 910
 911         for (i = 0; i < conf->copies; i++) {
 912                 struct bio *mbio;
 913                 int d = r10_bio->devs[i].devnum;
 914                 if (!r10_bio->devs[i].bio)
 915                         continue;
 916
 917                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 918                 r10_bio->devs[i].bio = mbio;
 919
 920                 mbio->bi_sector = r10_bio->devs[i].addr+
 921                         conf->mirrors[d].rdev->data_offset;
 922                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
 923                 mbio->bi_end_io = raid10_end_write_request;
 924                 mbio->bi_rw = WRITE | do_sync | do_fua;
 925                 mbio->bi_private = r10_bio;
 926
 927                 atomic_inc(&r10_bio->remaining);
 928                 spin_lock_irqsave(&conf->device_lock, flags);
 929                 bio_list_add(&conf->pending_bio_list, mbio);
 930                 spin_unlock_irqrestore(&conf->device_lock, flags);
 931         }
 932
 933         if (atomic_dec_and_test(&r10_bio->remaining)) {
 934                 /* This matches the end of raid10_end_write_request() */
 935                 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 936                                 r10_bio->sectors,
 937                                 !test_bit(R10BIO_Degraded, &r10_bio->state),
 938                                 0);
 939                 md_write_end(mddev);
 940                 raid_end_bio_io(r10_bio);
 941         }
 942
 943         /* In case raid10d snuck in to freeze_array */
 944         wake_up(&conf->wait_barrier);
 945
 946         if (do_sync || !mddev->bitmap || !plugged)
 947                 md_wakeup_thread(mddev->thread);
 948         return 0;
 949 }
 950
 951 static void status(struct seq_file *seq, mddev_t *mddev)
 952 {
 953         conf_t *conf = mddev->private;
 954         int i;
 955
 956         if (conf->near_copies < conf->raid_disks)
 957                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
 958         if (conf->near_copies > 1)
 959                 seq_printf(seq, " %d near-copies", conf->near_copies);
 960         if (conf->far_copies > 1) {
 961                 if (conf->far_offset)
 962                         seq_printf(seq, " %d offset-copies", conf->far_copies);
 963                 else
 964                         seq_printf(seq, " %d far-copies", conf->far_copies);
 965         }
 966         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 967                                         conf->raid_disks - mddev->degraded);
 968         for (i = 0; i < conf->raid_disks; i++)
 969                 seq_printf(seq, "%s",
 970                               conf->mirrors[i].rdev &&
 971                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
 972         seq_printf(seq, "]");
 973 }
 974
 975 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 976 {
 977         char b[BDEVNAME_SIZE];
 978         conf_t *conf = mddev->private;
 979
 980         /*
 981          * If it is not operational, then we have already marked it as dead
 982          * else if it is the last working disks, ignore the error, let the
 983          * next level up know.
 984          * else mark the drive as failed
 985          */
 986         if (test_bit(In_sync, &rdev->flags)
 987             && conf->raid_disks-mddev->degraded == 1)
 988                 /*
 989                  * Don't fail the drive, just return an IO error.
 990                  * The test should really be more sophisticated than
 991                  * "working_disks == 1", but it isn't critical, and
 992                  * can wait until we do more sophisticated "is the drive
 993                  * really dead" tests...
 994                  */
 995                 return;
 996         if (test_and_clear_bit(In_sync, &rdev->flags)) {
 997                 unsigned long flags;
 998                 spin_lock_irqsave(&conf->device_lock, flags);
 999                 mddev->degraded++;
1000                 spin_unlock_irqrestore(&conf->device_lock, flags);
1001                 /*
1002                  * if recovery is running, make sure it aborts.
1003                  */
1004                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1005         }
1006         set_bit(Faulty, &rdev->flags);
1007         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1008         printk(KERN_ALERT
1009                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1010                "md/raid10:%s: Operation continuing on %d devices.\n",
1011                mdname(mddev), bdevname(rdev->bdev, b),
1012                mdname(mddev), conf->raid_disks - mddev->degraded);
1013 }
1014
1015 static void print_conf(conf_t *conf)
1016 {
1017         int i;
1018         mirror_info_t *tmp;
1019
1020         printk(KERN_DEBUG "RAID10 conf printout:\n");
1021         if (!conf) {
1022                 printk(KERN_DEBUG "(!conf)\n");
1023                 return;
1024         }
1025         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1026                 conf->raid_disks);
1027
1028         for (i = 0; i < conf->raid_disks; i++) {
1029                 char b[BDEVNAME_SIZE];
1030                 tmp = conf->mirrors + i;
1031                 if (tmp->rdev)
1032                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1033                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1034                                 !test_bit(Faulty, &tmp->rdev->flags),
1035                                 bdevname(tmp->rdev->bdev,b));
1036         }
1037 }
1038
1039 static void close_sync(conf_t *conf)
1040 {
1041         wait_barrier(conf);
1042         allow_barrier(conf);
1043
1044         mempool_destroy(conf->r10buf_pool);
1045         conf->r10buf_pool = NULL;
1046 }
1047
1048 /* check if there are enough drives for
1049  * every block to appear on atleast one
1050  */
1051 static int enough(conf_t *conf)
1052 {
1053         int first = 0;
1054
1055         do {
1056                 int n = conf->copies;
1057                 int cnt = 0;
1058                 while (n--) {
1059                         if (conf->mirrors[first].rdev)
1060                                 cnt++;
1061                         first = (first+1) % conf->raid_disks;
1062                 }
1063                 if (cnt == 0)
1064                         return 0;
1065         } while (first != 0);
1066         return 1;
1067 }
1068
1069 static int raid10_spare_active(mddev_t *mddev)
1070 {
1071         int i;
1072         conf_t *conf = mddev->private;
1073         mirror_info_t *tmp;
1074         int count = 0;
1075         unsigned long flags;
1076
1077         /*
1078          * Find all non-in_sync disks within the RAID10 configuration
1079          * and mark them in_sync
1080          */
1081         for (i = 0; i < conf->raid_disks; i++) {
1082                 tmp = conf->mirrors + i;
1083                 if (tmp->rdev
1084                     && !test_bit(Faulty, &tmp->rdev->flags)
1085                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1086                         count++;
1087                         sysfs_notify_dirent(tmp->rdev->sysfs_state);
1088                 }
1089         }
1090         spin_lock_irqsave(&conf->device_lock, flags);
1091         mddev->degraded -= count;
1092         spin_unlock_irqrestore(&conf->device_lock, flags);
1093
1094         print_conf(conf);
1095         return count;
1096 }
1097
1098
1099 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1100 {
1101         conf_t *conf = mddev->private;
1102         int err = -EEXIST;
1103         int mirror;
1104         mirror_info_t *p;
1105         int first = 0;
1106         int last = conf->raid_disks - 1;
1107
1108         if (mddev->recovery_cp < MaxSector)
1109                 /* only hot-add to in-sync arrays, as recovery is
1110                  * very different from resync
1111                  */
1112                 return -EBUSY;
1113         if (!enough(conf))
1114                 return -EINVAL;
1115
1116         if (rdev->raid_disk >= 0)
1117                 first = last = rdev->raid_disk;
1118
1119         if (rdev->saved_raid_disk >= 0 &&
1120             rdev->saved_raid_disk >= first &&
1121             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1122                 mirror = rdev->saved_raid_disk;
1123         else
1124                 mirror = first;
1125         for ( ; mirror <= last ; mirror++)
1126                 if ( !(p=conf->mirrors+mirror)->rdev) {
1127
1128                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1129                                           rdev->data_offset << 9);
1130                         /* as we don't honour merge_bvec_fn, we must
1131                          * never risk violating it, so limit
1132                          * ->max_segments to one lying with a single
1133                          * page, as a one page request is never in
1134                          * violation.
1135                          */
1136                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1137                                 blk_queue_max_segments(mddev->queue, 1);
1138                                 blk_queue_segment_boundary(mddev->queue,
1139                                                            PAGE_CACHE_SIZE - 1);
1140                         }
1141
1142                         p->head_position = 0;
1143                         rdev->raid_disk = mirror;
1144                         err = 0;
1145                         if (rdev->saved_raid_disk != mirror)
1146                                 conf->fullsync = 1;
1147                         rcu_assign_pointer(p->rdev, rdev);
1148                         break;
1149                 }
1150
1151         md_integrity_add_rdev(rdev, mddev);
1152         print_conf(conf);
1153         return err;
1154 }
1155
1156 static int raid10_remove_disk(mddev_t *mddev, int number)
1157 {
1158         conf_t *conf = mddev->private;
1159         int err = 0;
1160         mdk_rdev_t *rdev;
1161         mirror_info_t *p = conf->mirrors+ number;
1162
1163         print_conf(conf);
1164         rdev = p->rdev;
1165         if (rdev) {
1166                 if (test_bit(In_sync, &rdev->flags) ||
1167                     atomic_read(&rdev->nr_pending)) {
1168                         err = -EBUSY;
1169                         goto abort;
1170                 }
1171                 /* Only remove faulty devices in recovery
1172                  * is not possible.
1173                  */
1174                 if (!test_bit(Faulty, &rdev->flags) &&
1175                     enough(conf)) {
1176                         err = -EBUSY;
1177                         goto abort;
1178                 }
1179                 p->rdev = NULL;
1180                 synchronize_rcu();
1181                 if (atomic_read(&rdev->nr_pending)) {
1182                         /* lost the race, try later */
1183                         err = -EBUSY;
1184                         p->rdev = rdev;
1185                         goto abort;
1186                 }
1187                 err = md_integrity_register(mddev);
1188         }
1189 abort:
1190
1191         print_conf(conf);
1192         return err;
1193 }
1194
1195
1196 static void end_sync_read(struct bio *bio, int error)
1197 {
1198         r10bio_t *r10_bio = bio->bi_private;
1199         conf_t *conf = r10_bio->mddev->private;
1200         int i,d;
1201
1202         for (i=0; i<conf->copies; i++)
1203                 if (r10_bio->devs[i].bio == bio)
1204                         break;
1205         BUG_ON(i == conf->copies);
1206         update_head_pos(i, r10_bio);
1207         d = r10_bio->devs[i].devnum;
1208
1209         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1210                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1211         else {
1212                 atomic_add(r10_bio->sectors,
1213                            &conf->mirrors[d].rdev->corrected_errors);
1214                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1215                         md_error(r10_bio->mddev,
1216                                  conf->mirrors[d].rdev);
1217         }
1218
1219         /* for reconstruct, we always reschedule after a read.
1220          * for resync, only after all reads
1221          */
1222         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1223         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1224             atomic_dec_and_test(&r10_bio->remaining)) {
1225                 /* we have read all the blocks,
1226                  * do the comparison in process context in raid10d
1227                  */
1228                 reschedule_retry(r10_bio);
1229         }
1230 }
1231
1232 static void end_sync_write(struct bio *bio, int error)
1233 {
1234         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1235         r10bio_t *r10_bio = bio->bi_private;
1236         mddev_t *mddev = r10_bio->mddev;
1237         conf_t *conf = mddev->private;
1238         int i,d;
1239
1240         for (i = 0; i < conf->copies; i++)
1241                 if (r10_bio->devs[i].bio == bio)
1242                         break;
1243         d = r10_bio->devs[i].devnum;
1244
1245         if (!uptodate)
1246                 md_error(mddev, conf->mirrors[d].rdev);
1247
1248         update_head_pos(i, r10_bio);
1249
1250         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1251         while (atomic_dec_and_test(&r10_bio->remaining)) {
1252                 if (r10_bio->master_bio == NULL) {
1253                         /* the primary of several recovery bios */
1254                         sector_t s = r10_bio->sectors;
1255                         put_buf(r10_bio);
1256                         md_done_sync(mddev, s, 1);
1257                         break;
1258                 } else {
1259                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1260                         put_buf(r10_bio);
1261                         r10_bio = r10_bio2;
1262                 }
1263         }
1264 }
1265
1266 /*
1267  * Note: sync and recover and handled very differently for raid10
1268  * This code is for resync.
1269  * For resync, we read through virtual addresses and read all blocks.
1270  * If there is any error, we schedule a write.  The lowest numbered
1271  * drive is authoritative.
1272  * However requests come for physical address, so we need to map.
1273  * For every physical address there are raid_disks/copies virtual addresses,
1274  * which is always are least one, but is not necessarly an integer.
1275  * This means that a physical address can span multiple chunks, so we may
1276  * have to submit multiple io requests for a single sync request.
1277  */
1278 /*
1279  * We check if all blocks are in-sync and only write to blocks that
1280  * aren't in sync
1281  */
1282 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1283 {
1284         conf_t *conf = mddev->private;
1285         int i, first;
1286         struct bio *tbio, *fbio;
1287
1288         atomic_set(&r10_bio->remaining, 1);
1289
1290         /* find the first device with a block */
1291         for (i=0; i<conf->copies; i++)
1292                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1293                         break;
1294
1295         if (i == conf->copies)
1296                 goto done;
1297
1298         first = i;
1299         fbio = r10_bio->devs[i].bio;
1300
1301         /* now find blocks with errors */
1302         for (i=0 ; i < conf->copies ; i++) {
1303                 int  j, d;
1304                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1305
1306                 tbio = r10_bio->devs[i].bio;
1307
1308                 if (tbio->bi_end_io != end_sync_read)
1309                         continue;
1310                 if (i == first)
1311                         continue;
1312                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1313                         /* We know that the bi_io_vec layout is the same for
1314                          * both 'first' and 'i', so we just compare them.
1315                          * All vec entries are PAGE_SIZE;
1316                          */
1317                         for (j = 0; j < vcnt; j++)
1318                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1319                                            page_address(tbio->bi_io_vec[j].bv_page),
1320                                            PAGE_SIZE))
1321                                         break;
1322                         if (j == vcnt)
1323                                 continue;
1324                         mddev->resync_mismatches += r10_bio->sectors;
1325                 }
1326                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1327                         /* Don't fix anything. */
1328                         continue;
1329                 /* Ok, we need to write this bio
1330                  * First we need to fixup bv_offset, bv_len and
1331                  * bi_vecs, as the read request might have corrupted these
1332                  */
1333                 tbio->bi_vcnt = vcnt;
1334                 tbio->bi_size = r10_bio->sectors << 9;
1335                 tbio->bi_idx = 0;
1336                 tbio->bi_phys_segments = 0;
1337                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1338                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1339                 tbio->bi_next = NULL;
1340                 tbio->bi_rw = WRITE;
1341                 tbio->bi_private = r10_bio;
1342                 tbio->bi_sector = r10_bio->devs[i].addr;
1343
1344                 for (j=0; j < vcnt ; j++) {
1345                         tbio->bi_io_vec[j].bv_offset = 0;
1346                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1347
1348                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1349                                page_address(fbio->bi_io_vec[j].bv_page),
1350                                PAGE_SIZE);
1351                 }
1352                 tbio->bi_end_io = end_sync_write;
1353
1354                 d = r10_bio->devs[i].devnum;
1355                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1356                 atomic_inc(&r10_bio->remaining);
1357                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1358
1359                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1360                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1361                 generic_make_request(tbio);
1362         }
1363
1364 done:
1365         if (atomic_dec_and_test(&r10_bio->remaining)) {
1366                 md_done_sync(mddev, r10_bio->sectors, 1);
1367                 put_buf(r10_bio);
1368         }
1369 }
1370
1371 /*
1372  * Now for the recovery code.
1373  * Recovery happens across physical sectors.
1374  * We recover all non-is_sync drives by finding the virtual address of
1375  * each, and then choose a working drive that also has that virt address.
1376  * There is a separate r10_bio for each non-in_sync drive.
1377  * Only the first two slots are in use. The first for reading,
1378  * The second for writing.
1379  *
1380  */
1381
1382 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1383 {
1384         conf_t *conf = mddev->private;
1385         int i, d;
1386         struct bio *bio, *wbio;
1387
1388
1389         /* move the pages across to the second bio
1390          * and submit the write request
1391          */
1392         bio = r10_bio->devs[0].bio;
1393         wbio = r10_bio->devs[1].bio;
1394         for (i=0; i < wbio->bi_vcnt; i++) {
1395                 struct page *p = bio->bi_io_vec[i].bv_page;
1396                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1397                 wbio->bi_io_vec[i].bv_page = p;
1398         }
1399         d = r10_bio->devs[1].devnum;
1400
1401         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1402         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1403         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1404                 generic_make_request(wbio);
1405         else
1406                 bio_endio(wbio, -EIO);
1407 }
1408
1409
1410 /*
1411  * Used by fix_read_error() to decay the per rdev read_errors.
1412  * We halve the read error count for every hour that has elapsed
1413  * since the last recorded read error.
1414  *
1415  */
1416 static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1417 {
1418         struct timespec cur_time_mon;
1419         unsigned long hours_since_last;
1420         unsigned int read_errors = atomic_read(&rdev->read_errors);
1421
1422         ktime_get_ts(&cur_time_mon);
1423
1424         if (rdev->last_read_error.tv_sec == 0 &&
1425             rdev->last_read_error.tv_nsec == 0) {
1426                 /* first time we've seen a read error */
1427                 rdev->last_read_error = cur_time_mon;
1428                 return;
1429         }
1430
1431         hours_since_last = (cur_time_mon.tv_sec -
1432                             rdev->last_read_error.tv_sec) / 3600;
1433
1434         rdev->last_read_error = cur_time_mon;
1435
1436         /*
1437          * if hours_since_last is > the number of bits in read_errors
1438          * just set read errors to 0. We do this to avoid
1439          * overflowing the shift of read_errors by hours_since_last.
1440          */
1441         if (hours_since_last >= 8 * sizeof(read_errors))
1442                 atomic_set(&rdev->read_errors, 0);
1443         else
1444                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1445 }
1446
1447 /*
1448  * This is a kernel thread which:
1449  *
1450  *      1.      Retries failed read operations on working mirrors.
1451  *      2.      Updates the raid superblock when problems encounter.
1452  *      3.      Performs writes following reads for array synchronising.
1453  */
1454
1455 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1456 {
1457         int sect = 0; /* Offset from r10_bio->sector */
1458         int sectors = r10_bio->sectors;
1459         mdk_rdev_t*rdev;
1460         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1461         int d = r10_bio->devs[r10_bio->read_slot].devnum;
1462
1463         rcu_read_lock();
1464         rdev = rcu_dereference(conf->mirrors[d].rdev);
1465         if (rdev) { /* If rdev is not NULL */
1466                 char b[BDEVNAME_SIZE];
1467                 int cur_read_error_count = 0;
1468
1469                 bdevname(rdev->bdev, b);
1470
1471                 if (test_bit(Faulty, &rdev->flags)) {
1472                         rcu_read_unlock();
1473                         /* drive has already been failed, just ignore any
1474                            more fix_read_error() attempts */
1475                         return;
1476                 }
1477
1478                 check_decay_read_errors(mddev, rdev);
1479                 atomic_inc(&rdev->read_errors);
1480                 cur_read_error_count = atomic_read(&rdev->read_errors);
1481                 if (cur_read_error_count > max_read_errors) {
1482                         rcu_read_unlock();
1483                         printk(KERN_NOTICE
1484                                "md/raid10:%s: %s: Raid device exceeded "
1485                                "read_error threshold "
1486                                "[cur %d:max %d]\n",
1487                                mdname(mddev),
1488                                b, cur_read_error_count, max_read_errors);
1489                         printk(KERN_NOTICE
1490                                "md/raid10:%s: %s: Failing raid "
1491                                "device\n", mdname(mddev), b);
1492                         md_error(mddev, conf->mirrors[d].rdev);
1493                         return;
1494                 }
1495         }
1496         rcu_read_unlock();
1497
1498         while(sectors) {
1499                 int s = sectors;
1500                 int sl = r10_bio->read_slot;
1501                 int success = 0;
1502                 int start;
1503
1504                 if (s > (PAGE_SIZE>>9))
1505                         s = PAGE_SIZE >> 9;
1506
1507                 rcu_read_lock();
1508                 do {
1509                         d = r10_bio->devs[sl].devnum;
1510                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1511                         if (rdev &&
1512                             test_bit(In_sync, &rdev->flags)) {
1513                                 atomic_inc(&rdev->nr_pending);
1514                                 rcu_read_unlock();
1515                                 success = sync_page_io(rdev,
1516                                                        r10_bio->devs[sl].addr +
1517                                                        sect,
1518                                                        s<<9,
1519                                                        conf->tmppage, READ, false);
1520                                 rdev_dec_pending(rdev, mddev);
1521                                 rcu_read_lock();
1522                                 if (success)
1523                                         break;
1524                         }
1525                         sl++;
1526                         if (sl == conf->copies)
1527                                 sl = 0;
1528                 } while (!success && sl != r10_bio->read_slot);
1529                 rcu_read_unlock();
1530
1531                 if (!success) {
1532                         /* Cannot read from anywhere -- bye bye array */
1533                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1534                         md_error(mddev, conf->mirrors[dn].rdev);
1535                         break;
1536                 }
1537
1538                 start = sl;
1539                 /* write it back and re-read */
1540                 rcu_read_lock();
1541                 while (sl != r10_bio->read_slot) {
1542                         char b[BDEVNAME_SIZE];
1543
1544                         if (sl==0)
1545                                 sl = conf->copies;
1546                         sl--;
1547                         d = r10_bio->devs[sl].devnum;
1548                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1549                         if (rdev &&
1550                             test_bit(In_sync, &rdev->flags)) {
1551                                 atomic_inc(&rdev->nr_pending);
1552                                 rcu_read_unlock();
1553                                 atomic_add(s, &rdev->corrected_errors);
1554                                 if (sync_page_io(rdev,
1555                                                  r10_bio->devs[sl].addr +
1556                                                  sect,
1557                                                  s<<9, conf->tmppage, WRITE, false)
1558                                     == 0) {
1559                                         /* Well, this device is dead */
1560                                         printk(KERN_NOTICE
1561                                                "md/raid10:%s: read correction "
1562                                                "write failed"
1563                                                " (%d sectors at %llu on %s)\n",
1564                                                mdname(mddev), s,
1565                                                (unsigned long long)(sect+
1566                                                rdev->data_offset),
1567                                                bdevname(rdev->bdev, b));
1568                                         printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1569                                                "drive\n",
1570                                                mdname(mddev),
1571                                                bdevname(rdev->bdev, b));
1572                                         md_error(mddev, rdev);
1573                                 }
1574                                 rdev_dec_pending(rdev, mddev);
1575                                 rcu_read_lock();
1576                         }
1577                 }
1578                 sl = start;
1579                 while (sl != r10_bio->read_slot) {
1580
1581                         if (sl==0)
1582                                 sl = conf->copies;
1583                         sl--;
1584                         d = r10_bio->devs[sl].devnum;
1585                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1586                         if (rdev &&
1587                             test_bit(In_sync, &rdev->flags)) {
1588                                 char b[BDEVNAME_SIZE];
1589                                 atomic_inc(&rdev->nr_pending);
1590                                 rcu_read_unlock();
1591                                 if (sync_page_io(rdev,
1592                                                  r10_bio->devs[sl].addr +
1593                                                  sect,
1594                                                  s<<9, conf->tmppage,
1595                                                  READ, false) == 0) {
1596                                         /* Well, this device is dead */
1597                                         printk(KERN_NOTICE
1598                                                "md/raid10:%s: unable to read back "
1599                                                "corrected sectors"
1600                                                " (%d sectors at %llu on %s)\n",
1601                                                mdname(mddev), s,
1602                                                (unsigned long long)(sect+
1603                                                     rdev->data_offset),
1604                                                bdevname(rdev->bdev, b));
1605                                         printk(KERN_NOTICE "md/raid10:%s: %s: failing drive\n",
1606                                                mdname(mddev),
1607                                                bdevname(rdev->bdev, b));
1608
1609                                         md_error(mddev, rdev);
1610                                 } else {
1611                                         printk(KERN_INFO
1612                                                "md/raid10:%s: read error corrected"
1613                                                " (%d sectors at %llu on %s)\n",
1614                                                mdname(mddev), s,
1615                                                (unsigned long long)(sect+
1616                                                     rdev->data_offset),
1617                                                bdevname(rdev->bdev, b));
1618                                 }
1619
1620                                 rdev_dec_pending(rdev, mddev);
1621                                 rcu_read_lock();
1622                         }
1623                 }
1624                 rcu_read_unlock();
1625
1626                 sectors -= s;
1627                 sect += s;
1628         }
1629 }
1630
1631 static void raid10d(mddev_t *mddev)
1632 {
1633         r10bio_t *r10_bio;
1634         struct bio *bio;
1635         unsigned long flags;
1636         conf_t *conf = mddev->private;
1637         struct list_head *head = &conf->retry_list;
1638         mdk_rdev_t *rdev;
1639         struct blk_plug plug;
1640
1641         md_check_recovery(mddev);
1642
1643         blk_start_plug(&plug);
1644         for (;;) {
1645                 char b[BDEVNAME_SIZE];
1646
1647                 flush_pending_writes(conf);
1648
1649                 spin_lock_irqsave(&conf->device_lock, flags);
1650                 if (list_empty(head)) {
1651                         spin_unlock_irqrestore(&conf->device_lock, flags);
1652                         break;
1653                 }
1654                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1655                 list_del(head->prev);
1656                 conf->nr_queued--;
1657                 spin_unlock_irqrestore(&conf->device_lock, flags);
1658
1659                 mddev = r10_bio->mddev;
1660                 conf = mddev->private;
1661                 if (test_bit(R10BIO_IsSync, &r10_bio->state))
1662                         sync_request_write(mddev, r10_bio);
1663                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
1664                         recovery_request_write(mddev, r10_bio);
1665                 else {
1666                         int mirror;
1667                         /* we got a read error. Maybe the drive is bad.  Maybe just
1668                          * the block and we can fix it.
1669                          * We freeze all other IO, and try reading the block from
1670                          * other devices.  When we find one, we re-write
1671                          * and check it that fixes the read error.
1672                          * This is all done synchronously while the array is
1673                          * frozen.
1674                          */
1675                         if (mddev->ro == 0) {
1676                                 freeze_array(conf);
1677                                 fix_read_error(conf, mddev, r10_bio);
1678                                 unfreeze_array(conf);
1679                         }
1680
1681                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1682                         r10_bio->devs[r10_bio->read_slot].bio =
1683                                 mddev->ro ? IO_BLOCKED : NULL;
1684                         mirror = read_balance(conf, r10_bio);
1685                         if (mirror == -1) {
1686                                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
1687                                        " read error for block %llu\n",
1688                                        mdname(mddev),
1689                                        bdevname(bio->bi_bdev,b),
1690                                        (unsigned long long)r10_bio->sector);
1691                                 raid_end_bio_io(r10_bio);
1692                                 bio_put(bio);
1693                         } else {
1694                                 const unsigned long do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
1695                                 bio_put(bio);
1696                                 rdev = conf->mirrors[mirror].rdev;
1697                                 if (printk_ratelimit())
1698                                         printk(KERN_ERR "md/raid10:%s: %s: redirecting sector %llu to"
1699                                                " another mirror\n",
1700                                                mdname(mddev),
1701                                                bdevname(rdev->bdev,b),
1702                                                (unsigned long long)r10_bio->sector);
1703                                 bio = bio_clone_mddev(r10_bio->master_bio,
1704                                                       GFP_NOIO, mddev);
1705                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1706                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1707                                         + rdev->data_offset;
1708                                 bio->bi_bdev = rdev->bdev;
1709                                 bio->bi_rw = READ | do_sync;
1710                                 bio->bi_private = r10_bio;
1711                                 bio->bi_end_io = raid10_end_read_request;
1712                                 generic_make_request(bio);
1713                         }
1714                 }
1715                 cond_resched();
1716         }
1717         blk_finish_plug(&plug);
1718 }
1719
1720
1721 static int init_resync(conf_t *conf)
1722 {
1723         int buffs;
1724
1725         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1726         BUG_ON(conf->r10buf_pool);
1727         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1728         if (!conf->r10buf_pool)
1729                 return -ENOMEM;
1730         conf->next_resync = 0;
1731         return 0;
1732 }
1733
1734 /*
1735  * perform a "sync" on one "block"
1736  *
1737  * We need to make sure that no normal I/O request - particularly write
1738  * requests - conflict with active sync requests.
1739  *
1740  * This is achieved by tracking pending requests and a 'barrier' concept
1741  * that can be installed to exclude normal IO requests.
1742  *
1743  * Resync and recovery are handled very differently.
1744  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1745  *
1746  * For resync, we iterate over virtual addresses, read all copies,
1747  * and update if there are differences.  If only one copy is live,
1748  * skip it.
1749  * For recovery, we iterate over physical addresses, read a good
1750  * value for each non-in_sync drive, and over-write.
1751  *
1752  * So, for recovery we may have several outstanding complex requests for a
1753  * given address, one for each out-of-sync device.  We model this by allocating
1754  * a number of r10_bio structures, one for each out-of-sync device.
1755  * As we setup these structures, we collect all bio's together into a list
1756  * which we then process collectively to add pages, and then process again
1757  * to pass to generic_make_request.
1758  *
1759  * The r10_bio structures are linked using a borrowed master_bio pointer.
1760  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1761  * has its remaining count decremented to 0, the whole complex operation
1762  * is complete.
1763  *
1764  */
1765
1766 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1767 {
1768         conf_t *conf = mddev->private;
1769         r10bio_t *r10_bio;
1770         struct bio *biolist = NULL, *bio;
1771         sector_t max_sector, nr_sectors;
1772         int disk;
1773         int i;
1774         int max_sync;
1775         sector_t sync_blocks;
1776
1777         sector_t sectors_skipped = 0;
1778         int chunks_skipped = 0;
1779
1780         if (!conf->r10buf_pool)
1781                 if (init_resync(conf))
1782                         return 0;
1783
1784  skipped:
1785         max_sector = mddev->dev_sectors;
1786         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1787                 max_sector = mddev->resync_max_sectors;
1788         if (sector_nr >= max_sector) {
1789                 /* If we aborted, we need to abort the
1790                  * sync on the 'current' bitmap chucks (there can
1791                  * be several when recovering multiple devices).
1792                  * as we may have started syncing it but not finished.
1793                  * We can find the current address in
1794                  * mddev->curr_resync, but for recovery,
1795                  * we need to convert that to several
1796                  * virtual addresses.
1797                  */
1798                 if (mddev->curr_resync < max_sector) { /* aborted */
1799                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1800                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1801                                                 &sync_blocks, 1);
1802                         else for (i=0; i<conf->raid_disks; i++) {
1803                                 sector_t sect =
1804                                         raid10_find_virt(conf, mddev->curr_resync, i);
1805                                 bitmap_end_sync(mddev->bitmap, sect,
1806                                                 &sync_blocks, 1);
1807                         }
1808                 } else /* completed sync */
1809                         conf->fullsync = 0;
1810
1811                 bitmap_close_sync(mddev->bitmap);
1812                 close_sync(conf);
1813                 *skipped = 1;
1814                 return sectors_skipped;
1815         }
1816         if (chunks_skipped >= conf->raid_disks) {
1817                 /* if there has been nothing to do on any drive,
1818                  * then there is nothing to do at all..
1819                  */
1820                 *skipped = 1;
1821                 return (max_sector - sector_nr) + sectors_skipped;
1822         }
1823
1824         if (max_sector > mddev->resync_max)
1825                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1826
1827         /* make sure whole request will fit in a chunk - if chunks
1828          * are meaningful
1829          */
1830         if (conf->near_copies < conf->raid_disks &&
1831             max_sector > (sector_nr | conf->chunk_mask))
1832                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1833         /*
1834          * If there is non-resync activity waiting for us then
1835          * put in a delay to throttle resync.
1836          */
1837         if (!go_faster && conf->nr_waiting)
1838                 msleep_interruptible(1000);
1839
1840         /* Again, very different code for resync and recovery.
1841          * Both must result in an r10bio with a list of bios that
1842          * have bi_end_io, bi_sector, bi_bdev set,
1843          * and bi_private set to the r10bio.
1844          * For recovery, we may actually create several r10bios
1845          * with 2 bios in each, that correspond to the bios in the main one.
1846          * In this case, the subordinate r10bios link back through a
1847          * borrowed master_bio pointer, and the counter in the master
1848          * includes a ref from each subordinate.
1849          */
1850         /* First, we decide what to do and set ->bi_end_io
1851          * To end_sync_read if we want to read, and
1852          * end_sync_write if we will want to write.
1853          */
1854
1855         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1856         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1857                 /* recovery... the complicated one */
1858                 int j, k;
1859                 r10_bio = NULL;
1860
1861                 for (i=0 ; i<conf->raid_disks; i++)
1862                         if (conf->mirrors[i].rdev &&
1863                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1864                                 int still_degraded = 0;
1865                                 /* want to reconstruct this device */
1866                                 r10bio_t *rb2 = r10_bio;
1867                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1868                                 int must_sync;
1869                                 /* Unless we are doing a full sync, we only need
1870                                  * to recover the block if it is set in the bitmap
1871                                  */
1872                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1873                                                               &sync_blocks, 1);
1874                                 if (sync_blocks < max_sync)
1875                                         max_sync = sync_blocks;
1876                                 if (!must_sync &&
1877                                     !conf->fullsync) {
1878                                         /* yep, skip the sync_blocks here, but don't assume
1879                                          * that there will never be anything to do here
1880                                          */
1881                                         chunks_skipped = -1;
1882                                         continue;
1883                                 }
1884
1885                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1886                                 raise_barrier(conf, rb2 != NULL);
1887                                 atomic_set(&r10_bio->remaining, 0);
1888
1889                                 r10_bio->master_bio = (struct bio*)rb2;
1890                                 if (rb2)
1891                                         atomic_inc(&rb2->remaining);
1892                                 r10_bio->mddev = mddev;
1893                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1894                                 r10_bio->sector = sect;
1895
1896                                 raid10_find_phys(conf, r10_bio);
1897
1898                                 /* Need to check if the array will still be
1899                                  * degraded
1900                                  */
1901                                 for (j=0; j<conf->raid_disks; j++)
1902                                         if (conf->mirrors[j].rdev == NULL ||
1903                                             test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
1904                                                 still_degraded = 1;
1905                                                 break;
1906                                         }
1907
1908                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1909                                                               &sync_blocks, still_degraded);
1910
1911                                 for (j=0; j<conf->copies;j++) {
1912                                         int d = r10_bio->devs[j].devnum;
1913                                         if (conf->mirrors[d].rdev &&
1914                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1915                                                 /* This is where we read from */
1916                                                 bio = r10_bio->devs[0].bio;
1917                                                 bio->bi_next = biolist;
1918                                                 biolist = bio;
1919                                                 bio->bi_private = r10_bio;
1920                                                 bio->bi_end_io = end_sync_read;
1921                                                 bio->bi_rw = READ;
1922                                                 bio->bi_sector = r10_bio->devs[j].addr +
1923                                                         conf->mirrors[d].rdev->data_offset;
1924                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1925                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1926                                                 atomic_inc(&r10_bio->remaining);
1927                                                 /* and we write to 'i' */
1928
1929                                                 for (k=0; k<conf->copies; k++)
1930                                                         if (r10_bio->devs[k].devnum == i)
1931                                                                 break;
1932                                                 BUG_ON(k == conf->copies);
1933                                                 bio = r10_bio->devs[1].bio;
1934                                                 bio->bi_next = biolist;
1935                                                 biolist = bio;
1936                                                 bio->bi_private = r10_bio;
1937                                                 bio->bi_end_io = end_sync_write;
1938                                                 bio->bi_rw = WRITE;
1939                                                 bio->bi_sector = r10_bio->devs[k].addr +
1940                                                         conf->mirrors[i].rdev->data_offset;
1941                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1942
1943                                                 r10_bio->devs[0].devnum = d;
1944                                                 r10_bio->devs[1].devnum = i;
1945
1946                                                 break;
1947                                         }
1948                                 }
1949                                 if (j == conf->copies) {
1950                                         /* Cannot recover, so abort the recovery */
1951                                         put_buf(r10_bio);
1952                                         if (rb2)
1953                                                 atomic_dec(&rb2->remaining);
1954                                         r10_bio = rb2;
1955                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
1956                                                               &mddev->recovery))
1957                                                 printk(KERN_INFO "md/raid10:%s: insufficient "
1958                                                        "working devices for recovery.\n",
1959                                                        mdname(mddev));
1960                                         break;
1961                                 }
1962                         }
1963                 if (biolist == NULL) {
1964                         while (r10_bio) {
1965                                 r10bio_t *rb2 = r10_bio;
1966                                 r10_bio = (r10bio_t*) rb2->master_bio;
1967                                 rb2->master_bio = NULL;
1968                                 put_buf(rb2);
1969                         }
1970                         goto giveup;
1971                 }
1972         } else {
1973                 /* resync. Schedule a read for every block at this virt offset */
1974                 int count = 0;
1975
1976                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1977
1978                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1979                                        &sync_blocks, mddev->degraded) &&
1980                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1981                         /* We can skip this block */
1982                         *skipped = 1;
1983                         return sync_blocks + sectors_skipped;
1984                 }
1985                 if (sync_blocks < max_sync)
1986                         max_sync = sync_blocks;
1987                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1988
1989                 r10_bio->mddev = mddev;
1990                 atomic_set(&r10_bio->remaining, 0);
1991                 raise_barrier(conf, 0);
1992                 conf->next_resync = sector_nr;
1993
1994                 r10_bio->master_bio = NULL;
1995                 r10_bio->sector = sector_nr;
1996                 set_bit(R10BIO_IsSync, &r10_bio->state);
1997                 raid10_find_phys(conf, r10_bio);
1998                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1999
2000                 for (i=0; i<conf->copies; i++) {
2001                         int d = r10_bio->devs[i].devnum;
2002                         bio = r10_bio->devs[i].bio;
2003                         bio->bi_end_io = NULL;
2004                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
2005                         if (conf->mirrors[d].rdev == NULL ||
2006                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2007                                 continue;
2008                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2009                         atomic_inc(&r10_bio->remaining);
2010                         bio->bi_next = biolist;
2011                         biolist = bio;
2012                         bio->bi_private = r10_bio;
2013                         bio->bi_end_io = end_sync_read;
2014                         bio->bi_rw = READ;
2015                         bio->bi_sector = r10_bio->devs[i].addr +
2016                                 conf->mirrors[d].rdev->data_offset;
2017                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2018                         count++;
2019                 }
2020
2021                 if (count < 2) {
2022                         for (i=0; i<conf->copies; i++) {
2023                                 int d = r10_bio->devs[i].devnum;
2024                                 if (r10_bio->devs[i].bio->bi_end_io)
2025                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
2026                         }
2027                         put_buf(r10_bio);
2028                         biolist = NULL;
2029                         goto giveup;
2030                 }
2031         }
2032
2033         for (bio = biolist; bio ; bio=bio->bi_next) {
2034
2035                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2036                 if (bio->bi_end_io)
2037                         bio->bi_flags |= 1 << BIO_UPTODATE;
2038                 bio->bi_vcnt = 0;
2039                 bio->bi_idx = 0;
2040                 bio->bi_phys_segments = 0;
2041                 bio->bi_size = 0;
2042         }
2043
2044         nr_sectors = 0;
2045         if (sector_nr + max_sync < max_sector)
2046                 max_sector = sector_nr + max_sync;
2047         do {
2048                 struct page *page;
2049                 int len = PAGE_SIZE;
2050                 disk = 0;
2051                 if (sector_nr + (len>>9) > max_sector)
2052                         len = (max_sector - sector_nr) << 9;
2053                 if (len == 0)
2054                         break;
2055                 for (bio= biolist ; bio ; bio=bio->bi_next) {
2056                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2057                         if (bio_add_page(bio, page, len, 0) == 0) {
2058                                 /* stop here */
2059                                 struct bio *bio2;
2060                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2061                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
2062                                         /* remove last page from this bio */
2063                                         bio2->bi_vcnt--;
2064                                         bio2->bi_size -= len;
2065                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2066                                 }
2067                                 goto bio_full;
2068                         }
2069                         disk = i;
2070                 }
2071                 nr_sectors += len>>9;
2072                 sector_nr += len>>9;
2073         } while (biolist->bi_vcnt < RESYNC_PAGES);
2074  bio_full:
2075         r10_bio->sectors = nr_sectors;
2076
2077         while (biolist) {
2078                 bio = biolist;
2079                 biolist = biolist->bi_next;
2080
2081                 bio->bi_next = NULL;
2082                 r10_bio = bio->bi_private;
2083                 r10_bio->sectors = nr_sectors;
2084
2085                 if (bio->bi_end_io == end_sync_read) {
2086                         md_sync_acct(bio->bi_bdev, nr_sectors);
2087                         generic_make_request(bio);
2088                 }
2089         }
2090
2091         if (sectors_skipped)
2092                 /* pretend they weren't skipped, it makes
2093                  * no important difference in this case
2094                  */
2095                 md_done_sync(mddev, sectors_skipped, 1);
2096
2097         return sectors_skipped + nr_sectors;
2098  giveup:
2099         /* There is nowhere to write, so all non-sync
2100          * drives must be failed, so try the next chunk...
2101          */
2102         if (sector_nr + max_sync < max_sector)
2103                 max_sector = sector_nr + max_sync;
2104
2105         sectors_skipped += (max_sector - sector_nr);
2106         chunks_skipped ++;
2107         sector_nr = max_sector;
2108         goto skipped;
2109 }
2110
2111 static sector_t
2112 raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2113 {
2114         sector_t size;
2115         conf_t *conf = mddev->private;
2116
2117         if (!raid_disks)
2118                 raid_disks = conf->raid_disks;
2119         if (!sectors)
2120                 sectors = conf->dev_sectors;
2121
2122         size = sectors >> conf->chunk_shift;
2123         sector_div(size, conf->far_copies);
2124         size = size * raid_disks;
2125         sector_div(size, conf->near_copies);
2126
2127         return size << conf->chunk_shift;
2128 }
2129
2130
2131 static conf_t *setup_conf(mddev_t *mddev)
2132 {
2133         conf_t *conf = NULL;
2134         int nc, fc, fo;
2135         sector_t stride, size;
2136         int err = -EINVAL;
2137
2138         if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2139             !is_power_of_2(mddev->new_chunk_sectors)) {
2140                 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2141                        "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2142                        mdname(mddev), PAGE_SIZE);
2143                 goto out;
2144         }
2145
2146         nc = mddev->new_layout & 255;
2147         fc = (mddev->new_layout >> 8) & 255;
2148         fo = mddev->new_layout & (1<<16);
2149
2150         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2151             (mddev->new_layout >> 17)) {
2152                 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2153                        mdname(mddev), mddev->new_layout);
2154                 goto out;
2155         }
2156
2157         err = -ENOMEM;
2158         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2159         if (!conf)
2160                 goto out;
2161
2162         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2163                                 GFP_KERNEL);
2164         if (!conf->mirrors)
2165                 goto out;
2166
2167         conf->tmppage = alloc_page(GFP_KERNEL);
2168         if (!conf->tmppage)
2169                 goto out;
2170
2171
2172         conf->raid_disks = mddev->raid_disks;
2173         conf->near_copies = nc;
2174         conf->far_copies = fc;
2175         conf->copies = nc*fc;
2176         conf->far_offset = fo;
2177         conf->chunk_mask = mddev->new_chunk_sectors - 1;
2178         conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2179
2180         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2181                                            r10bio_pool_free, conf);
2182         if (!conf->r10bio_pool)
2183                 goto out;
2184
2185         size = mddev->dev_sectors >> conf->chunk_shift;
2186         sector_div(size, fc);
2187         size = size * conf->raid_disks;
2188         sector_div(size, nc);
2189         /* 'size' is now the number of chunks in the array */
2190         /* calculate "used chunks per device" in 'stride' */
2191         stride = size * conf->copies;
2192
2193         /* We need to round up when dividing by raid_disks to
2194          * get the stride size.
2195          */
2196         stride += conf->raid_disks - 1;
2197         sector_div(stride, conf->raid_disks);
2198
2199         conf->dev_sectors = stride << conf->chunk_shift;
2200
2201         if (fo)
2202                 stride = 1;
2203         else
2204                 sector_div(stride, fc);
2205         conf->stride = stride << conf->chunk_shift;
2206
2207
2208         spin_lock_init(&conf->device_lock);
2209         INIT_LIST_HEAD(&conf->retry_list);
2210
2211         spin_lock_init(&conf->resync_lock);
2212         init_waitqueue_head(&conf->wait_barrier);
2213
2214         conf->thread = md_register_thread(raid10d, mddev, NULL);
2215         if (!conf->thread)
2216                 goto out;
2217
2218         conf->mddev = mddev;
2219         return conf;
2220
2221  out:
2222         printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2223                mdname(mddev));
2224         if (conf) {
2225                 if (conf->r10bio_pool)
2226                         mempool_destroy(conf->r10bio_pool);
2227                 kfree(conf->mirrors);
2228                 safe_put_page(conf->tmppage);
2229                 kfree(conf);
2230         }
2231         return ERR_PTR(err);
2232 }
2233
2234 static int run(mddev_t *mddev)
2235 {
2236         conf_t *conf;
2237         int i, disk_idx, chunk_size;
2238         mirror_info_t *disk;
2239         mdk_rdev_t *rdev;
2240         sector_t size;
2241
2242         /*
2243          * copy the already verified devices into our private RAID10
2244          * bookkeeping area. [whatever we allocate in run(),
2245          * should be freed in stop()]
2246          */
2247
2248         if (mddev->private == NULL) {
2249                 conf = setup_conf(mddev);
2250                 if (IS_ERR(conf))
2251                         return PTR_ERR(conf);
2252                 mddev->private = conf;
2253         }
2254         conf = mddev->private;
2255         if (!conf)
2256                 goto out;
2257
2258         mddev->thread = conf->thread;
2259         conf->thread = NULL;
2260
2261         chunk_size = mddev->chunk_sectors << 9;
2262         blk_queue_io_min(mddev->queue, chunk_size);
2263         if (conf->raid_disks % conf->near_copies)
2264                 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2265         else
2266                 blk_queue_io_opt(mddev->queue, chunk_size *
2267                                  (conf->raid_disks / conf->near_copies));
2268
2269         list_for_each_entry(rdev, &mddev->disks, same_set) {
2270                 disk_idx = rdev->raid_disk;
2271                 if (disk_idx >= conf->raid_disks
2272                     || disk_idx < 0)
2273                         continue;
2274                 disk = conf->mirrors + disk_idx;
2275
2276                 disk->rdev = rdev;
2277                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2278                                   rdev->data_offset << 9);
2279                 /* as we don't honour merge_bvec_fn, we must never risk
2280                  * violating it, so limit max_segments to 1 lying
2281                  * within a single page.
2282                  */
2283                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2284                         blk_queue_max_segments(mddev->queue, 1);
2285                         blk_queue_segment_boundary(mddev->queue,
2286                                                    PAGE_CACHE_SIZE - 1);
2287                 }
2288
2289                 disk->head_position = 0;
2290         }
2291         /* need to check that every block has at least one working mirror */
2292         if (!enough(conf)) {
2293                 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2294                        mdname(mddev));
2295                 goto out_free_conf;
2296         }
2297
2298         mddev->degraded = 0;
2299         for (i = 0; i < conf->raid_disks; i++) {
2300
2301                 disk = conf->mirrors + i;
2302
2303                 if (!disk->rdev ||
2304                     !test_bit(In_sync, &disk->rdev->flags)) {
2305                         disk->head_position = 0;
2306                         mddev->degraded++;
2307                         if (disk->rdev)
2308                                 conf->fullsync = 1;
2309                 }
2310         }
2311
2312         if (mddev->recovery_cp != MaxSector)
2313                 printk(KERN_NOTICE "md/raid10:%s: not clean"
2314                        " -- starting background reconstruction\n",
2315                        mdname(mddev));
2316         printk(KERN_INFO
2317                 "md/raid10:%s: active with %d out of %d devices\n",
2318                 mdname(mddev), conf->raid_disks - mddev->degraded,
2319                 conf->raid_disks);
2320         /*
2321          * Ok, everything is just fine now
2322          */
2323         mddev->dev_sectors = conf->dev_sectors;
2324         size = raid10_size(mddev, 0, 0);
2325         md_set_array_sectors(mddev, size);
2326         mddev->resync_max_sectors = size;
2327
2328         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2329         mddev->queue->backing_dev_info.congested_data = mddev;
2330
2331         /* Calculate max read-ahead size.
2332          * We need to readahead at least twice a whole stripe....
2333          * maybe...
2334          */
2335         {
2336                 int stripe = conf->raid_disks *
2337                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2338                 stripe /= conf->near_copies;
2339                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2340                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2341         }
2342
2343         if (conf->near_copies < conf->raid_disks)
2344                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2345
2346         if (md_integrity_register(mddev))
2347                 goto out_free_conf;
2348
2349         return 0;
2350
2351 out_free_conf:
2352         md_unregister_thread(mddev->thread);
2353         if (conf->r10bio_pool)
2354                 mempool_destroy(conf->r10bio_pool);
2355         safe_put_page(conf->tmppage);
2356         kfree(conf->mirrors);
2357         kfree(conf);
2358         mddev->private = NULL;
2359 out:
2360         return -EIO;
2361 }
2362
2363 static int stop(mddev_t *mddev)
2364 {
2365         conf_t *conf = mddev->private;
2366
2367         raise_barrier(conf, 0);
2368         lower_barrier(conf);
2369
2370         md_unregister_thread(mddev->thread);
2371         mddev->thread = NULL;
2372         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2373         if (conf->r10bio_pool)
2374                 mempool_destroy(conf->r10bio_pool);
2375         kfree(conf->mirrors);
2376         kfree(conf);
2377         mddev->private = NULL;
2378         return 0;
2379 }
2380
2381 static void raid10_quiesce(mddev_t *mddev, int state)
2382 {
2383         conf_t *conf = mddev->private;
2384
2385         switch(state) {
2386         case 1:
2387                 raise_barrier(conf, 0);
2388                 break;
2389         case 0:
2390                 lower_barrier(conf);
2391                 break;
2392         }
2393 }
2394
2395 static void *raid10_takeover_raid0(mddev_t *mddev)
2396 {
2397         mdk_rdev_t *rdev;
2398         conf_t *conf;
2399
2400         if (mddev->degraded > 0) {
2401                 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
2402                        mdname(mddev));
2403                 return ERR_PTR(-EINVAL);
2404         }
2405
2406         /* Set new parameters */
2407         mddev->new_level = 10;
2408         /* new layout: far_copies = 1, near_copies = 2 */
2409         mddev->new_layout = (1<<8) + 2;
2410         mddev->new_chunk_sectors = mddev->chunk_sectors;
2411         mddev->delta_disks = mddev->raid_disks;
2412         mddev->raid_disks *= 2;
2413         /* make sure it will be not marked as dirty */
2414         mddev->recovery_cp = MaxSector;
2415
2416         conf = setup_conf(mddev);
2417         if (!IS_ERR(conf)) {
2418                 list_for_each_entry(rdev, &mddev->disks, same_set)
2419                         if (rdev->raid_disk >= 0)
2420                                 rdev->new_raid_disk = rdev->raid_disk * 2;
2421                 conf->barrier = 1;
2422         }
2423
2424         return conf;
2425 }
2426
2427 static void *raid10_takeover(mddev_t *mddev)
2428 {
2429         struct raid0_private_data *raid0_priv;
2430
2431         /* raid10 can take over:
2432          *  raid0 - providing it has only two drives
2433          */
2434         if (mddev->level == 0) {
2435                 /* for raid0 takeover only one zone is supported */
2436                 raid0_priv = mddev->private;
2437                 if (raid0_priv->nr_strip_zones > 1) {
2438                         printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
2439                                " with more than one zone.\n",
2440                                mdname(mddev));
2441                         return ERR_PTR(-EINVAL);
2442                 }
2443                 return raid10_takeover_raid0(mddev);
2444         }
2445         return ERR_PTR(-EINVAL);
2446 }
2447
2448 static struct mdk_personality raid10_personality =
2449 {
2450         .name           = "raid10",
2451         .level          = 10,
2452         .owner          = THIS_MODULE,
2453         .make_request   = make_request,
2454         .run            = run,
2455         .stop           = stop,
2456         .status         = status,
2457         .error_handler  = error,
2458         .hot_add_disk   = raid10_add_disk,
2459         .hot_remove_disk= raid10_remove_disk,
2460         .spare_active   = raid10_spare_active,
2461         .sync_request   = sync_request,
2462         .quiesce        = raid10_quiesce,
2463         .size           = raid10_size,
2464         .takeover       = raid10_takeover,
2465 };
2466
2467 static int __init raid_init(void)
2468 {
2469         return register_md_personality(&raid10_personality);
2470 }
2471
2472 static void raid_exit(void)
2473 {
2474         unregister_md_personality(&raid10_personality);
2475 }
2476
2477 module_init(raid_init);
2478 module_exit(raid_exit);
2479 MODULE_LICENSE("GPL");
2480 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
2481 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2482 MODULE_ALIAS("md-raid10");
2483 MODULE_ALIAS("md-level-10");