drivers/md/raid10.c

   1 /*
   2  * raid10.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 2000-2004 Neil Brown
   5  *
   6  * RAID-10 support for md.
   7  *
   8  * Base on code in raid1.c.  See raid1.c for futher copyright information.
   9  *
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2, or (at your option)
  14  * any later version.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * (for example /usr/src/linux/COPYING); if not, write to the Free
  18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 #include <linux/slab.h>
  22 #include <linux/delay.h>
  23 #include <linux/blkdev.h>
  24 #include <linux/seq_file.h>
  25 #include "md.h"
  26 #include "raid10.h"
  27 #include "bitmap.h"
  28
  29 /*
  30  * RAID10 provides a combination of RAID0 and RAID1 functionality.
  31  * The layout of data is defined by
  32  *    chunk_size
  33  *    raid_disks
  34  *    near_copies (stored in low byte of layout)
  35  *    far_copies (stored in second byte of layout)
  36  *    far_offset (stored in bit 16 of layout )
  37  *
  38  * The data to be stored is divided into chunks using chunksize.
  39  * Each device is divided into far_copies sections.
  40  * In each section, chunks are laid out in a style similar to raid0, but
  41  * near_copies copies of each chunk is stored (each on a different drive).
  42  * The starting device for each section is offset near_copies from the starting
  43  * device of the previous section.
  44  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
  45  * drive.
  46  * near_copies and far_copies must be at least one, and their product is at most
  47  * raid_disks.
  48  *
  49  * If far_offset is true, then the far_copies are handled a bit differently.
  50  * The copies are still in different stripes, but instead of be very far apart
  51  * on disk, there are adjacent stripes.
  52  */
  53
  54 /*
  55  * Number of guaranteed r10bios in case of extreme VM load:
  56  */
  57 #define NR_RAID10_BIOS 256
  58
  59 static void unplug_slaves(mddev_t *mddev);
  60
  61 static void allow_barrier(conf_t *conf);
  62 static void lower_barrier(conf_t *conf);
  63
  64 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
  65 {
  66         conf_t *conf = data;
  67         r10bio_t *r10_bio;
  68         int size = offsetof(struct r10bio_s, devs[conf->copies]);
  69
  70         /* allocate a r10bio with room for raid_disks entries in the bios array */
  71         r10_bio = kzalloc(size, gfp_flags);
  72         if (!r10_bio && conf->mddev)
  73                 unplug_slaves(conf->mddev);
  74
  75         return r10_bio;
  76 }
  77
  78 static void r10bio_pool_free(void *r10_bio, void *data)
  79 {
  80         kfree(r10_bio);
  81 }
  82
  83 /* Maximum size of each resync request */
  84 #define RESYNC_BLOCK_SIZE (64*1024)
  85 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  86 /* amount of memory to reserve for resync requests */
  87 #define RESYNC_WINDOW (1024*1024)
  88 /* maximum number of concurrent requests, memory permitting */
  89 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
  90
  91 /*
  92  * When performing a resync, we need to read and compare, so
  93  * we need as many pages are there are copies.
  94  * When performing a recovery, we need 2 bios, one for read,
  95  * one for write (we recover only one drive per r10buf)
  96  *
  97  */
  98 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
  99 {
 100         conf_t *conf = data;
 101         struct page *page;
 102         r10bio_t *r10_bio;
 103         struct bio *bio;
 104         int i, j;
 105         int nalloc;
 106
 107         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 108         if (!r10_bio) {
 109                 unplug_slaves(conf->mddev);
 110                 return NULL;
 111         }
 112
 113         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
 114                 nalloc = conf->copies; /* resync */
 115         else
 116                 nalloc = 2; /* recovery */
 117
 118         /*
 119          * Allocate bios.
 120          */
 121         for (j = nalloc ; j-- ; ) {
 122                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 123                 if (!bio)
 124                         goto out_free_bio;
 125                 r10_bio->devs[j].bio = bio;
 126         }
 127         /*
 128          * Allocate RESYNC_PAGES data pages and attach them
 129          * where needed.
 130          */
 131         for (j = 0 ; j < nalloc; j++) {
 132                 bio = r10_bio->devs[j].bio;
 133                 for (i = 0; i < RESYNC_PAGES; i++) {
 134                         page = alloc_page(gfp_flags);
 135                         if (unlikely(!page))
 136                                 goto out_free_pages;
 137
 138                         bio->bi_io_vec[i].bv_page = page;
 139                 }
 140         }
 141
 142         return r10_bio;
 143
 144 out_free_pages:
 145         for ( ; i > 0 ; i--)
 146                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
 147         while (j--)
 148                 for (i = 0; i < RESYNC_PAGES ; i++)
 149                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 150         j = -1;
 151 out_free_bio:
 152         while ( ++j < nalloc )
 153                 bio_put(r10_bio->devs[j].bio);
 154         r10bio_pool_free(r10_bio, conf);
 155         return NULL;
 156 }
 157
 158 static void r10buf_pool_free(void *__r10_bio, void *data)
 159 {
 160         int i;
 161         conf_t *conf = data;
 162         r10bio_t *r10bio = __r10_bio;
 163         int j;
 164
 165         for (j=0; j < conf->copies; j++) {
 166                 struct bio *bio = r10bio->devs[j].bio;
 167                 if (bio) {
 168                         for (i = 0; i < RESYNC_PAGES; i++) {
 169                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 170                                 bio->bi_io_vec[i].bv_page = NULL;
 171                         }
 172                         bio_put(bio);
 173                 }
 174         }
 175         r10bio_pool_free(r10bio, conf);
 176 }
 177
 178 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
 179 {
 180         int i;
 181
 182         for (i = 0; i < conf->copies; i++) {
 183                 struct bio **bio = & r10_bio->devs[i].bio;
 184                 if (*bio && *bio != IO_BLOCKED)
 185                         bio_put(*bio);
 186                 *bio = NULL;
 187         }
 188 }
 189
 190 static void free_r10bio(r10bio_t *r10_bio)
 191 {
 192         conf_t *conf = r10_bio->mddev->private;
 193
 194         /*
 195          * Wake up any possible resync thread that waits for the device
 196          * to go idle.
 197          */
 198         allow_barrier(conf);
 199
 200         put_all_bios(conf, r10_bio);
 201         mempool_free(r10_bio, conf->r10bio_pool);
 202 }
 203
 204 static void put_buf(r10bio_t *r10_bio)
 205 {
 206         conf_t *conf = r10_bio->mddev->private;
 207
 208         mempool_free(r10_bio, conf->r10buf_pool);
 209
 210         lower_barrier(conf);
 211 }
 212
 213 static void reschedule_retry(r10bio_t *r10_bio)
 214 {
 215         unsigned long flags;
 216         mddev_t *mddev = r10_bio->mddev;
 217         conf_t *conf = mddev->private;
 218
 219         spin_lock_irqsave(&conf->device_lock, flags);
 220         list_add(&r10_bio->retry_list, &conf->retry_list);
 221         conf->nr_queued ++;
 222         spin_unlock_irqrestore(&conf->device_lock, flags);
 223
 224         /* wake up frozen array... */
 225         wake_up(&conf->wait_barrier);
 226
 227         md_wakeup_thread(mddev->thread);
 228 }
 229
 230 /*
 231  * raid_end_bio_io() is called when we have finished servicing a mirrored
 232  * operation and are ready to return a success/failure code to the buffer
 233  * cache layer.
 234  */
 235 static void raid_end_bio_io(r10bio_t *r10_bio)
 236 {
 237         struct bio *bio = r10_bio->master_bio;
 238
 239         bio_endio(bio,
 240                 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
 241         free_r10bio(r10_bio);
 242 }
 243
 244 /*
 245  * Update disk head position estimator based on IRQ completion info.
 246  */
 247 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
 248 {
 249         conf_t *conf = r10_bio->mddev->private;
 250
 251         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 252                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 253 }
 254
 255 static void raid10_end_read_request(struct bio *bio, int error)
 256 {
 257         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 258         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
 259         int slot, dev;
 260         conf_t *conf = r10_bio->mddev->private;
 261
 262
 263         slot = r10_bio->read_slot;
 264         dev = r10_bio->devs[slot].devnum;
 265         /*
 266          * this branch is our 'one mirror IO has finished' event handler:
 267          */
 268         update_head_pos(slot, r10_bio);
 269
 270         if (uptodate) {
 271                 /*
 272                  * Set R10BIO_Uptodate in our master bio, so that
 273                  * we will return a good error code to the higher
 274                  * levels even if IO on some other mirrored buffer fails.
 275                  *
 276                  * The 'master' represents the composite IO operation to
 277                  * user-side. So if something waits for IO, then it will
 278                  * wait for the 'master' bio.
 279                  */
 280                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 281                 raid_end_bio_io(r10_bio);
 282         } else {
 283                 /*
 284                  * oops, read error:
 285                  */
 286                 char b[BDEVNAME_SIZE];
 287                 if (printk_ratelimit())
 288                         printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
 289                                bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
 290                 reschedule_retry(r10_bio);
 291         }
 292
 293         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 294 }
 295
 296 static void raid10_end_write_request(struct bio *bio, int error)
 297 {
 298         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 299         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
 300         int slot, dev;
 301         conf_t *conf = r10_bio->mddev->private;
 302
 303         for (slot = 0; slot < conf->copies; slot++)
 304                 if (r10_bio->devs[slot].bio == bio)
 305                         break;
 306         dev = r10_bio->devs[slot].devnum;
 307
 308         /*
 309          * this branch is our 'one mirror IO has finished' event handler:
 310          */
 311         if (!uptodate) {
 312                 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
 313                 /* an I/O failed, we can't clear the bitmap */
 314                 set_bit(R10BIO_Degraded, &r10_bio->state);
 315         } else
 316                 /*
 317                  * Set R10BIO_Uptodate in our master bio, so that
 318                  * we will return a good error code for to the higher
 319                  * levels even if IO on some other mirrored buffer fails.
 320                  *
 321                  * The 'master' represents the composite IO operation to
 322                  * user-side. So if something waits for IO, then it will
 323                  * wait for the 'master' bio.
 324                  */
 325                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 326
 327         update_head_pos(slot, r10_bio);
 328
 329         /*
 330          *
 331          * Let's see if all mirrored write operations have finished
 332          * already.
 333          */
 334         if (atomic_dec_and_test(&r10_bio->remaining)) {
 335                 /* clear the bitmap if all writes complete successfully */
 336                 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 337                                 r10_bio->sectors,
 338                                 !test_bit(R10BIO_Degraded, &r10_bio->state),
 339                                 0);
 340                 md_write_end(r10_bio->mddev);
 341                 raid_end_bio_io(r10_bio);
 342         }
 343
 344         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 345 }
 346
 347
 348 /*
 349  * RAID10 layout manager
 350  * Aswell as the chunksize and raid_disks count, there are two
 351  * parameters: near_copies and far_copies.
 352  * near_copies * far_copies must be <= raid_disks.
 353  * Normally one of these will be 1.
 354  * If both are 1, we get raid0.
 355  * If near_copies == raid_disks, we get raid1.
 356  *
 357  * Chunks are layed out in raid0 style with near_copies copies of the
 358  * first chunk, followed by near_copies copies of the next chunk and
 359  * so on.
 360  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 361  * as described above, we start again with a device offset of near_copies.
 362  * So we effectively have another copy of the whole array further down all
 363  * the drives, but with blocks on different drives.
 364  * With this layout, and block is never stored twice on the one device.
 365  *
 366  * raid10_find_phys finds the sector offset of a given virtual sector
 367  * on each device that it is on.
 368  *
 369  * raid10_find_virt does the reverse mapping, from a device and a
 370  * sector offset to a virtual address
 371  */
 372
 373 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
 374 {
 375         int n,f;
 376         sector_t sector;
 377         sector_t chunk;
 378         sector_t stripe;
 379         int dev;
 380
 381         int slot = 0;
 382
 383         /* now calculate first sector/dev */
 384         chunk = r10bio->sector >> conf->chunk_shift;
 385         sector = r10bio->sector & conf->chunk_mask;
 386
 387         chunk *= conf->near_copies;
 388         stripe = chunk;
 389         dev = sector_div(stripe, conf->raid_disks);
 390         if (conf->far_offset)
 391                 stripe *= conf->far_copies;
 392
 393         sector += stripe << conf->chunk_shift;
 394
 395         /* and calculate all the others */
 396         for (n=0; n < conf->near_copies; n++) {
 397                 int d = dev;
 398                 sector_t s = sector;
 399                 r10bio->devs[slot].addr = sector;
 400                 r10bio->devs[slot].devnum = d;
 401                 slot++;
 402
 403                 for (f = 1; f < conf->far_copies; f++) {
 404                         d += conf->near_copies;
 405                         if (d >= conf->raid_disks)
 406                                 d -= conf->raid_disks;
 407                         s += conf->stride;
 408                         r10bio->devs[slot].devnum = d;
 409                         r10bio->devs[slot].addr = s;
 410                         slot++;
 411                 }
 412                 dev++;
 413                 if (dev >= conf->raid_disks) {
 414                         dev = 0;
 415                         sector += (conf->chunk_mask + 1);
 416                 }
 417         }
 418         BUG_ON(slot != conf->copies);
 419 }
 420
 421 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
 422 {
 423         sector_t offset, chunk, vchunk;
 424
 425         offset = sector & conf->chunk_mask;
 426         if (conf->far_offset) {
 427                 int fc;
 428                 chunk = sector >> conf->chunk_shift;
 429                 fc = sector_div(chunk, conf->far_copies);
 430                 dev -= fc * conf->near_copies;
 431                 if (dev < 0)
 432                         dev += conf->raid_disks;
 433         } else {
 434                 while (sector >= conf->stride) {
 435                         sector -= conf->stride;
 436                         if (dev < conf->near_copies)
 437                                 dev += conf->raid_disks - conf->near_copies;
 438                         else
 439                                 dev -= conf->near_copies;
 440                 }
 441                 chunk = sector >> conf->chunk_shift;
 442         }
 443         vchunk = chunk * conf->raid_disks + dev;
 444         sector_div(vchunk, conf->near_copies);
 445         return (vchunk << conf->chunk_shift) + offset;
 446 }
 447
 448 /**
 449  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 450  *      @q: request queue
 451  *      @bvm: properties of new bio
 452  *      @biovec: the request that could be merged to it.
 453  *
 454  *      Return amount of bytes we can accept at this offset
 455  *      If near_copies == raid_disk, there are no striping issues,
 456  *      but in that case, the function isn't called at all.
 457  */
 458 static int raid10_mergeable_bvec(struct request_queue *q,
 459                                  struct bvec_merge_data *bvm,
 460                                  struct bio_vec *biovec)
 461 {
 462         mddev_t *mddev = q->queuedata;
 463         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 464         int max;
 465         unsigned int chunk_sectors = mddev->chunk_sectors;
 466         unsigned int bio_sectors = bvm->bi_size >> 9;
 467
 468         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 469         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 470         if (max <= biovec->bv_len && bio_sectors == 0)
 471                 return biovec->bv_len;
 472         else
 473                 return max;
 474 }
 475
 476 /*
 477  * This routine returns the disk from which the requested read should
 478  * be done. There is a per-array 'next expected sequential IO' sector
 479  * number - if this matches on the next IO then we use the last disk.
 480  * There is also a per-disk 'last know head position' sector that is
 481  * maintained from IRQ contexts, both the normal and the resync IO
 482  * completion handlers update this position correctly. If there is no
 483  * perfect sequential match then we pick the disk whose head is closest.
 484  *
 485  * If there are 2 mirrors in the same 2 devices, performance degrades
 486  * because position is mirror, not device based.
 487  *
 488  * The rdev for the device selected will have nr_pending incremented.
 489  */
 490
 491 /*
 492  * FIXME: possibly should rethink readbalancing and do it differently
 493  * depending on near_copies / far_copies geometry.
 494  */
 495 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
 496 {
 497         const unsigned long this_sector = r10_bio->sector;
 498         int disk, slot, nslot;
 499         const int sectors = r10_bio->sectors;
 500         sector_t new_distance, current_distance;
 501         mdk_rdev_t *rdev;
 502
 503         raid10_find_phys(conf, r10_bio);
 504         rcu_read_lock();
 505         /*
 506          * Check if we can balance. We can balance on the whole
 507          * device if no resync is going on (recovery is ok), or below
 508          * the resync window. We take the first readable disk when
 509          * above the resync window.
 510          */
 511         if (conf->mddev->recovery_cp < MaxSector
 512             && (this_sector + sectors >= conf->next_resync)) {
 513                 /* make sure that disk is operational */
 514                 slot = 0;
 515                 disk = r10_bio->devs[slot].devnum;
 516
 517                 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 518                        r10_bio->devs[slot].bio == IO_BLOCKED ||
 519                        !test_bit(In_sync, &rdev->flags)) {
 520                         slot++;
 521                         if (slot == conf->copies) {
 522                                 slot = 0;
 523                                 disk = -1;
 524                                 break;
 525                         }
 526                         disk = r10_bio->devs[slot].devnum;
 527                 }
 528                 goto rb_out;
 529         }
 530
 531
 532         /* make sure the disk is operational */
 533         slot = 0;
 534         disk = r10_bio->devs[slot].devnum;
 535         while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 536                r10_bio->devs[slot].bio == IO_BLOCKED ||
 537                !test_bit(In_sync, &rdev->flags)) {
 538                 slot ++;
 539                 if (slot == conf->copies) {
 540                         disk = -1;
 541                         goto rb_out;
 542                 }
 543                 disk = r10_bio->devs[slot].devnum;
 544         }
 545
 546
 547         current_distance = abs(r10_bio->devs[slot].addr -
 548                                conf->mirrors[disk].head_position);
 549
 550         /* Find the disk whose head is closest,
 551          * or - for far > 1 - find the closest to partition beginning */
 552
 553         for (nslot = slot; nslot < conf->copies; nslot++) {
 554                 int ndisk = r10_bio->devs[nslot].devnum;
 555
 556
 557                 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
 558                     r10_bio->devs[nslot].bio == IO_BLOCKED ||
 559                     !test_bit(In_sync, &rdev->flags))
 560                         continue;
 561
 562                 /* This optimisation is debatable, and completely destroys
 563                  * sequential read speed for 'far copies' arrays.  So only
 564                  * keep it for 'near' arrays, and review those later.
 565                  */
 566                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
 567                         disk = ndisk;
 568                         slot = nslot;
 569                         break;
 570                 }
 571
 572                 /* for far > 1 always use the lowest address */
 573                 if (conf->far_copies > 1)
 574                         new_distance = r10_bio->devs[nslot].addr;
 575                 else
 576                         new_distance = abs(r10_bio->devs[nslot].addr -
 577                                            conf->mirrors[ndisk].head_position);
 578                 if (new_distance < current_distance) {
 579                         current_distance = new_distance;
 580                         disk = ndisk;
 581                         slot = nslot;
 582                 }
 583         }
 584
 585 rb_out:
 586         r10_bio->read_slot = slot;
 587 /*      conf->next_seq_sect = this_sector + sectors;*/
 588
 589         if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
 590                 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
 591         else
 592                 disk = -1;
 593         rcu_read_unlock();
 594
 595         return disk;
 596 }
 597
 598 static void unplug_slaves(mddev_t *mddev)
 599 {
 600         conf_t *conf = mddev->private;
 601         int i;
 602
 603         rcu_read_lock();
 604         for (i=0; i<mddev->raid_disks; i++) {
 605                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 606                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 607                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 608
 609                         atomic_inc(&rdev->nr_pending);
 610                         rcu_read_unlock();
 611
 612                         blk_unplug(r_queue);
 613
 614                         rdev_dec_pending(rdev, mddev);
 615                         rcu_read_lock();
 616                 }
 617         }
 618         rcu_read_unlock();
 619 }
 620
 621 static void raid10_unplug(struct request_queue *q)
 622 {
 623         mddev_t *mddev = q->queuedata;
 624
 625         unplug_slaves(q->queuedata);
 626         md_wakeup_thread(mddev->thread);
 627 }
 628
 629 static int raid10_congested(void *data, int bits)
 630 {
 631         mddev_t *mddev = data;
 632         conf_t *conf = mddev->private;
 633         int i, ret = 0;
 634
 635         if (mddev_congested(mddev, bits))
 636                 return 1;
 637         rcu_read_lock();
 638         for (i = 0; i < mddev->raid_disks && ret == 0; i++) {
 639                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 640                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 641                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 642
 643                         ret |= bdi_congested(&q->backing_dev_info, bits);
 644                 }
 645         }
 646         rcu_read_unlock();
 647         return ret;
 648 }
 649
 650 static int flush_pending_writes(conf_t *conf)
 651 {
 652         /* Any writes that have been queued but are awaiting
 653          * bitmap updates get flushed here.
 654          * We return 1 if any requests were actually submitted.
 655          */
 656         int rv = 0;
 657
 658         spin_lock_irq(&conf->device_lock);
 659
 660         if (conf->pending_bio_list.head) {
 661                 struct bio *bio;
 662                 bio = bio_list_get(&conf->pending_bio_list);
 663                 blk_remove_plug(conf->mddev->queue);
 664                 spin_unlock_irq(&conf->device_lock);
 665                 /* flush any pending bitmap writes to disk
 666                  * before proceeding w/ I/O */
 667                 bitmap_unplug(conf->mddev->bitmap);
 668
 669                 while (bio) { /* submit pending writes */
 670                         struct bio *next = bio->bi_next;
 671                         bio->bi_next = NULL;
 672                         generic_make_request(bio);
 673                         bio = next;
 674                 }
 675                 rv = 1;
 676         } else
 677                 spin_unlock_irq(&conf->device_lock);
 678         return rv;
 679 }
 680 /* Barriers....
 681  * Sometimes we need to suspend IO while we do something else,
 682  * either some resync/recovery, or reconfigure the array.
 683  * To do this we raise a 'barrier'.
 684  * The 'barrier' is a counter that can be raised multiple times
 685  * to count how many activities are happening which preclude
 686  * normal IO.
 687  * We can only raise the barrier if there is no pending IO.
 688  * i.e. if nr_pending == 0.
 689  * We choose only to raise the barrier if no-one is waiting for the
 690  * barrier to go down.  This means that as soon as an IO request
 691  * is ready, no other operations which require a barrier will start
 692  * until the IO request has had a chance.
 693  *
 694  * So: regular IO calls 'wait_barrier'.  When that returns there
 695  *    is no backgroup IO happening,  It must arrange to call
 696  *    allow_barrier when it has finished its IO.
 697  * backgroup IO calls must call raise_barrier.  Once that returns
 698  *    there is no normal IO happeing.  It must arrange to call
 699  *    lower_barrier when the particular background IO completes.
 700  */
 701
 702 static void raise_barrier(conf_t *conf, int force)
 703 {
 704         BUG_ON(force && !conf->barrier);
 705         spin_lock_irq(&conf->resync_lock);
 706
 707         /* Wait until no block IO is waiting (unless 'force') */
 708         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 709                             conf->resync_lock,
 710                             raid10_unplug(conf->mddev->queue));
 711
 712         /* block any new IO from starting */
 713         conf->barrier++;
 714
 715         /* No wait for all pending IO to complete */
 716         wait_event_lock_irq(conf->wait_barrier,
 717                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 718                             conf->resync_lock,
 719                             raid10_unplug(conf->mddev->queue));
 720
 721         spin_unlock_irq(&conf->resync_lock);
 722 }
 723
 724 static void lower_barrier(conf_t *conf)
 725 {
 726         unsigned long flags;
 727         spin_lock_irqsave(&conf->resync_lock, flags);
 728         conf->barrier--;
 729         spin_unlock_irqrestore(&conf->resync_lock, flags);
 730         wake_up(&conf->wait_barrier);
 731 }
 732
 733 static void wait_barrier(conf_t *conf)
 734 {
 735         spin_lock_irq(&conf->resync_lock);
 736         if (conf->barrier) {
 737                 conf->nr_waiting++;
 738                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 739                                     conf->resync_lock,
 740                                     raid10_unplug(conf->mddev->queue));
 741                 conf->nr_waiting--;
 742         }
 743         conf->nr_pending++;
 744         spin_unlock_irq(&conf->resync_lock);
 745 }
 746
 747 static void allow_barrier(conf_t *conf)
 748 {
 749         unsigned long flags;
 750         spin_lock_irqsave(&conf->resync_lock, flags);
 751         conf->nr_pending--;
 752         spin_unlock_irqrestore(&conf->resync_lock, flags);
 753         wake_up(&conf->wait_barrier);
 754 }
 755
 756 static void freeze_array(conf_t *conf)
 757 {
 758         /* stop syncio and normal IO and wait for everything to
 759          * go quiet.
 760          * We increment barrier and nr_waiting, and then
 761          * wait until nr_pending match nr_queued+1
 762          * This is called in the context of one normal IO request
 763          * that has failed. Thus any sync request that might be pending
 764          * will be blocked by nr_pending, and we need to wait for
 765          * pending IO requests to complete or be queued for re-try.
 766          * Thus the number queued (nr_queued) plus this request (1)
 767          * must match the number of pending IOs (nr_pending) before
 768          * we continue.
 769          */
 770         spin_lock_irq(&conf->resync_lock);
 771         conf->barrier++;
 772         conf->nr_waiting++;
 773         wait_event_lock_irq(conf->wait_barrier,
 774                             conf->nr_pending == conf->nr_queued+1,
 775                             conf->resync_lock,
 776                             ({ flush_pending_writes(conf);
 777                                raid10_unplug(conf->mddev->queue); }));
 778         spin_unlock_irq(&conf->resync_lock);
 779 }
 780
 781 static void unfreeze_array(conf_t *conf)
 782 {
 783         /* reverse the effect of the freeze */
 784         spin_lock_irq(&conf->resync_lock);
 785         conf->barrier--;
 786         conf->nr_waiting--;
 787         wake_up(&conf->wait_barrier);
 788         spin_unlock_irq(&conf->resync_lock);
 789 }
 790
 791 static int make_request(struct request_queue *q, struct bio * bio)
 792 {
 793         mddev_t *mddev = q->queuedata;
 794         conf_t *conf = mddev->private;
 795         mirror_info_t *mirror;
 796         r10bio_t *r10_bio;
 797         struct bio *read_bio;
 798         int cpu;
 799         int i;
 800         int chunk_sects = conf->chunk_mask + 1;
 801         const int rw = bio_data_dir(bio);
 802         const bool do_sync = bio_rw_flagged(bio, BIO_RW_SYNCIO);
 803         struct bio_list bl;
 804         unsigned long flags;
 805         mdk_rdev_t *blocked_rdev;
 806
 807         if (unlikely(bio_rw_flagged(bio, BIO_RW_BARRIER))) {
 808                 md_barrier_request(mddev, bio);
 809                 return 0;
 810         }
 811
 812         /* If this request crosses a chunk boundary, we need to
 813          * split it.  This will only happen for 1 PAGE (or less) requests.
 814          */
 815         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 816                       > chunk_sects &&
 817                     conf->near_copies < conf->raid_disks)) {
 818                 struct bio_pair *bp;
 819                 /* Sanity check -- queue functions should prevent this happening */
 820                 if (bio->bi_vcnt != 1 ||
 821                     bio->bi_idx != 0)
 822                         goto bad_map;
 823                 /* This is a one page bio that upper layers
 824                  * refuse to split for us, so we need to split it.
 825                  */
 826                 bp = bio_split(bio,
 827                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 828                 if (make_request(q, &bp->bio1))
 829                         generic_make_request(&bp->bio1);
 830                 if (make_request(q, &bp->bio2))
 831                         generic_make_request(&bp->bio2);
 832
 833                 bio_pair_release(bp);
 834                 return 0;
 835         bad_map:
 836                 printk("raid10_make_request bug: can't convert block across chunks"
 837                        " or bigger than %dk %llu %d\n", chunk_sects/2,
 838                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 839
 840                 bio_io_error(bio);
 841                 return 0;
 842         }
 843
 844         md_write_start(mddev, bio);
 845
 846         /*
 847          * Register the new request and wait if the reconstruction
 848          * thread has put up a bar for new requests.
 849          * Continue immediately if no resync is active currently.
 850          */
 851         wait_barrier(conf);
 852
 853         cpu = part_stat_lock();
 854         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
 855         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
 856                       bio_sectors(bio));
 857         part_stat_unlock();
 858
 859         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 860
 861         r10_bio->master_bio = bio;
 862         r10_bio->sectors = bio->bi_size >> 9;
 863
 864         r10_bio->mddev = mddev;
 865         r10_bio->sector = bio->bi_sector;
 866         r10_bio->state = 0;
 867
 868         if (rw == READ) {
 869                 /*
 870                  * read balancing logic:
 871                  */
 872                 int disk = read_balance(conf, r10_bio);
 873                 int slot = r10_bio->read_slot;
 874                 if (disk < 0) {
 875                         raid_end_bio_io(r10_bio);
 876                         return 0;
 877                 }
 878                 mirror = conf->mirrors + disk;
 879
 880                 read_bio = bio_clone(bio, GFP_NOIO);
 881
 882                 r10_bio->devs[slot].bio = read_bio;
 883
 884                 read_bio->bi_sector = r10_bio->devs[slot].addr +
 885                         mirror->rdev->data_offset;
 886                 read_bio->bi_bdev = mirror->rdev->bdev;
 887                 read_bio->bi_end_io = raid10_end_read_request;
 888                 read_bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
 889                 read_bio->bi_private = r10_bio;
 890
 891                 generic_make_request(read_bio);
 892                 return 0;
 893         }
 894
 895         /*
 896          * WRITE:
 897          */
 898         /* first select target devices under rcu_lock and
 899          * inc refcount on their rdev.  Record them by setting
 900          * bios[x] to bio
 901          */
 902         raid10_find_phys(conf, r10_bio);
 903  retry_write:
 904         blocked_rdev = NULL;
 905         rcu_read_lock();
 906         for (i = 0;  i < conf->copies; i++) {
 907                 int d = r10_bio->devs[i].devnum;
 908                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
 909                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 910                         atomic_inc(&rdev->nr_pending);
 911                         blocked_rdev = rdev;
 912                         break;
 913                 }
 914                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 915                         atomic_inc(&rdev->nr_pending);
 916                         r10_bio->devs[i].bio = bio;
 917                 } else {
 918                         r10_bio->devs[i].bio = NULL;
 919                         set_bit(R10BIO_Degraded, &r10_bio->state);
 920                 }
 921         }
 922         rcu_read_unlock();
 923
 924         if (unlikely(blocked_rdev)) {
 925                 /* Have to wait for this device to get unblocked, then retry */
 926                 int j;
 927                 int d;
 928
 929                 for (j = 0; j < i; j++)
 930                         if (r10_bio->devs[j].bio) {
 931                                 d = r10_bio->devs[j].devnum;
 932                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
 933                         }
 934                 allow_barrier(conf);
 935                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 936                 wait_barrier(conf);
 937                 goto retry_write;
 938         }
 939
 940         atomic_set(&r10_bio->remaining, 0);
 941
 942         bio_list_init(&bl);
 943         for (i = 0; i < conf->copies; i++) {
 944                 struct bio *mbio;
 945                 int d = r10_bio->devs[i].devnum;
 946                 if (!r10_bio->devs[i].bio)
 947                         continue;
 948
 949                 mbio = bio_clone(bio, GFP_NOIO);
 950                 r10_bio->devs[i].bio = mbio;
 951
 952                 mbio->bi_sector = r10_bio->devs[i].addr+
 953                         conf->mirrors[d].rdev->data_offset;
 954                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
 955                 mbio->bi_end_io = raid10_end_write_request;
 956                 mbio->bi_rw = WRITE | (do_sync << BIO_RW_SYNCIO);
 957                 mbio->bi_private = r10_bio;
 958
 959                 atomic_inc(&r10_bio->remaining);
 960                 bio_list_add(&bl, mbio);
 961         }
 962
 963         if (unlikely(!atomic_read(&r10_bio->remaining))) {
 964                 /* the array is dead */
 965                 md_write_end(mddev);
 966                 raid_end_bio_io(r10_bio);
 967                 return 0;
 968         }
 969
 970         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
 971         spin_lock_irqsave(&conf->device_lock, flags);
 972         bio_list_merge(&conf->pending_bio_list, &bl);
 973         blk_plug_device(mddev->queue);
 974         spin_unlock_irqrestore(&conf->device_lock, flags);
 975
 976         /* In case raid10d snuck in to freeze_array */
 977         wake_up(&conf->wait_barrier);
 978
 979         if (do_sync)
 980                 md_wakeup_thread(mddev->thread);
 981
 982         return 0;
 983 }
 984
 985 static void status(struct seq_file *seq, mddev_t *mddev)
 986 {
 987         conf_t *conf = mddev->private;
 988         int i;
 989
 990         if (conf->near_copies < conf->raid_disks)
 991                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
 992         if (conf->near_copies > 1)
 993                 seq_printf(seq, " %d near-copies", conf->near_copies);
 994         if (conf->far_copies > 1) {
 995                 if (conf->far_offset)
 996                         seq_printf(seq, " %d offset-copies", conf->far_copies);
 997                 else
 998                         seq_printf(seq, " %d far-copies", conf->far_copies);
 999         }
1000         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1001                                         conf->raid_disks - mddev->degraded);
1002         for (i = 0; i < conf->raid_disks; i++)
1003                 seq_printf(seq, "%s",
1004                               conf->mirrors[i].rdev &&
1005                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1006         seq_printf(seq, "]");
1007 }
1008
1009 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1010 {
1011         char b[BDEVNAME_SIZE];
1012         conf_t *conf = mddev->private;
1013
1014         /*
1015          * If it is not operational, then we have already marked it as dead
1016          * else if it is the last working disks, ignore the error, let the
1017          * next level up know.
1018          * else mark the drive as failed
1019          */
1020         if (test_bit(In_sync, &rdev->flags)
1021             && conf->raid_disks-mddev->degraded == 1)
1022                 /*
1023                  * Don't fail the drive, just return an IO error.
1024                  * The test should really be more sophisticated than
1025                  * "working_disks == 1", but it isn't critical, and
1026                  * can wait until we do more sophisticated "is the drive
1027                  * really dead" tests...
1028                  */
1029                 return;
1030         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1031                 unsigned long flags;
1032                 spin_lock_irqsave(&conf->device_lock, flags);
1033                 mddev->degraded++;
1034                 spin_unlock_irqrestore(&conf->device_lock, flags);
1035                 /*
1036                  * if recovery is running, make sure it aborts.
1037                  */
1038                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1039         }
1040         set_bit(Faulty, &rdev->flags);
1041         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1042         printk(KERN_ALERT "raid10: Disk failure on %s, disabling device.\n"
1043                 "raid10: Operation continuing on %d devices.\n",
1044                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1045 }
1046
1047 static void print_conf(conf_t *conf)
1048 {
1049         int i;
1050         mirror_info_t *tmp;
1051
1052         printk("RAID10 conf printout:\n");
1053         if (!conf) {
1054                 printk("(!conf)\n");
1055                 return;
1056         }
1057         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1058                 conf->raid_disks);
1059
1060         for (i = 0; i < conf->raid_disks; i++) {
1061                 char b[BDEVNAME_SIZE];
1062                 tmp = conf->mirrors + i;
1063                 if (tmp->rdev)
1064                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1065                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1066                                 !test_bit(Faulty, &tmp->rdev->flags),
1067                                 bdevname(tmp->rdev->bdev,b));
1068         }
1069 }
1070
1071 static void close_sync(conf_t *conf)
1072 {
1073         wait_barrier(conf);
1074         allow_barrier(conf);
1075
1076         mempool_destroy(conf->r10buf_pool);
1077         conf->r10buf_pool = NULL;
1078 }
1079
1080 /* check if there are enough drives for
1081  * every block to appear on atleast one
1082  */
1083 static int enough(conf_t *conf)
1084 {
1085         int first = 0;
1086
1087         do {
1088                 int n = conf->copies;
1089                 int cnt = 0;
1090                 while (n--) {
1091                         if (conf->mirrors[first].rdev)
1092                                 cnt++;
1093                         first = (first+1) % conf->raid_disks;
1094                 }
1095                 if (cnt == 0)
1096                         return 0;
1097         } while (first != 0);
1098         return 1;
1099 }
1100
1101 static int raid10_spare_active(mddev_t *mddev)
1102 {
1103         int i;
1104         conf_t *conf = mddev->private;
1105         mirror_info_t *tmp;
1106
1107         /*
1108          * Find all non-in_sync disks within the RAID10 configuration
1109          * and mark them in_sync
1110          */
1111         for (i = 0; i < conf->raid_disks; i++) {
1112                 tmp = conf->mirrors + i;
1113                 if (tmp->rdev
1114                     && !test_bit(Faulty, &tmp->rdev->flags)
1115                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1116                         unsigned long flags;
1117                         spin_lock_irqsave(&conf->device_lock, flags);
1118                         mddev->degraded--;
1119                         spin_unlock_irqrestore(&conf->device_lock, flags);
1120                 }
1121         }
1122
1123         print_conf(conf);
1124         return 0;
1125 }
1126
1127
1128 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1129 {
1130         conf_t *conf = mddev->private;
1131         int err = -EEXIST;
1132         int mirror;
1133         mirror_info_t *p;
1134         int first = 0;
1135         int last = mddev->raid_disks - 1;
1136
1137         if (mddev->recovery_cp < MaxSector)
1138                 /* only hot-add to in-sync arrays, as recovery is
1139                  * very different from resync
1140                  */
1141                 return -EBUSY;
1142         if (!enough(conf))
1143                 return -EINVAL;
1144
1145         if (rdev->raid_disk >= 0)
1146                 first = last = rdev->raid_disk;
1147
1148         if (rdev->saved_raid_disk >= 0 &&
1149             rdev->saved_raid_disk >= first &&
1150             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1151                 mirror = rdev->saved_raid_disk;
1152         else
1153                 mirror = first;
1154         for ( ; mirror <= last ; mirror++)
1155                 if ( !(p=conf->mirrors+mirror)->rdev) {
1156
1157                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1158                                           rdev->data_offset << 9);
1159                         /* as we don't honour merge_bvec_fn, we must
1160                          * never risk violating it, so limit
1161                          * ->max_segments to one lying with a single
1162                          * page, as a one page request is never in
1163                          * violation.
1164                          */
1165                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1166                                 blk_queue_max_segments(mddev->queue, 1);
1167                                 blk_queue_segment_boundary(mddev->queue,
1168                                                            PAGE_CACHE_SIZE - 1);
1169                         }
1170
1171                         p->head_position = 0;
1172                         rdev->raid_disk = mirror;
1173                         err = 0;
1174                         if (rdev->saved_raid_disk != mirror)
1175                                 conf->fullsync = 1;
1176                         rcu_assign_pointer(p->rdev, rdev);
1177                         break;
1178                 }
1179
1180         md_integrity_add_rdev(rdev, mddev);
1181         print_conf(conf);
1182         return err;
1183 }
1184
1185 static int raid10_remove_disk(mddev_t *mddev, int number)
1186 {
1187         conf_t *conf = mddev->private;
1188         int err = 0;
1189         mdk_rdev_t *rdev;
1190         mirror_info_t *p = conf->mirrors+ number;
1191
1192         print_conf(conf);
1193         rdev = p->rdev;
1194         if (rdev) {
1195                 if (test_bit(In_sync, &rdev->flags) ||
1196                     atomic_read(&rdev->nr_pending)) {
1197                         err = -EBUSY;
1198                         goto abort;
1199                 }
1200                 /* Only remove faulty devices in recovery
1201                  * is not possible.
1202                  */
1203                 if (!test_bit(Faulty, &rdev->flags) &&
1204                     enough(conf)) {
1205                         err = -EBUSY;
1206                         goto abort;
1207                 }
1208                 p->rdev = NULL;
1209                 synchronize_rcu();
1210                 if (atomic_read(&rdev->nr_pending)) {
1211                         /* lost the race, try later */
1212                         err = -EBUSY;
1213                         p->rdev = rdev;
1214                         goto abort;
1215                 }
1216                 md_integrity_register(mddev);
1217         }
1218 abort:
1219
1220         print_conf(conf);
1221         return err;
1222 }
1223
1224
1225 static void end_sync_read(struct bio *bio, int error)
1226 {
1227         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1228         conf_t *conf = r10_bio->mddev->private;
1229         int i,d;
1230
1231         for (i=0; i<conf->copies; i++)
1232                 if (r10_bio->devs[i].bio == bio)
1233                         break;
1234         BUG_ON(i == conf->copies);
1235         update_head_pos(i, r10_bio);
1236         d = r10_bio->devs[i].devnum;
1237
1238         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1239                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1240         else {
1241                 atomic_add(r10_bio->sectors,
1242                            &conf->mirrors[d].rdev->corrected_errors);
1243                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1244                         md_error(r10_bio->mddev,
1245                                  conf->mirrors[d].rdev);
1246         }
1247
1248         /* for reconstruct, we always reschedule after a read.
1249          * for resync, only after all reads
1250          */
1251         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1252         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1253             atomic_dec_and_test(&r10_bio->remaining)) {
1254                 /* we have read all the blocks,
1255                  * do the comparison in process context in raid10d
1256                  */
1257                 reschedule_retry(r10_bio);
1258         }
1259 }
1260
1261 static void end_sync_write(struct bio *bio, int error)
1262 {
1263         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1264         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1265         mddev_t *mddev = r10_bio->mddev;
1266         conf_t *conf = mddev->private;
1267         int i,d;
1268
1269         for (i = 0; i < conf->copies; i++)
1270                 if (r10_bio->devs[i].bio == bio)
1271                         break;
1272         d = r10_bio->devs[i].devnum;
1273
1274         if (!uptodate)
1275                 md_error(mddev, conf->mirrors[d].rdev);
1276
1277         update_head_pos(i, r10_bio);
1278
1279         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1280         while (atomic_dec_and_test(&r10_bio->remaining)) {
1281                 if (r10_bio->master_bio == NULL) {
1282                         /* the primary of several recovery bios */
1283                         sector_t s = r10_bio->sectors;
1284                         put_buf(r10_bio);
1285                         md_done_sync(mddev, s, 1);
1286                         break;
1287                 } else {
1288                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1289                         put_buf(r10_bio);
1290                         r10_bio = r10_bio2;
1291                 }
1292         }
1293 }
1294
1295 /*
1296  * Note: sync and recover and handled very differently for raid10
1297  * This code is for resync.
1298  * For resync, we read through virtual addresses and read all blocks.
1299  * If there is any error, we schedule a write.  The lowest numbered
1300  * drive is authoritative.
1301  * However requests come for physical address, so we need to map.
1302  * For every physical address there are raid_disks/copies virtual addresses,
1303  * which is always are least one, but is not necessarly an integer.
1304  * This means that a physical address can span multiple chunks, so we may
1305  * have to submit multiple io requests for a single sync request.
1306  */
1307 /*
1308  * We check if all blocks are in-sync and only write to blocks that
1309  * aren't in sync
1310  */
1311 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1312 {
1313         conf_t *conf = mddev->private;
1314         int i, first;
1315         struct bio *tbio, *fbio;
1316
1317         atomic_set(&r10_bio->remaining, 1);
1318
1319         /* find the first device with a block */
1320         for (i=0; i<conf->copies; i++)
1321                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1322                         break;
1323
1324         if (i == conf->copies)
1325                 goto done;
1326
1327         first = i;
1328         fbio = r10_bio->devs[i].bio;
1329
1330         /* now find blocks with errors */
1331         for (i=0 ; i < conf->copies ; i++) {
1332                 int  j, d;
1333                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1334
1335                 tbio = r10_bio->devs[i].bio;
1336
1337                 if (tbio->bi_end_io != end_sync_read)
1338                         continue;
1339                 if (i == first)
1340                         continue;
1341                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1342                         /* We know that the bi_io_vec layout is the same for
1343                          * both 'first' and 'i', so we just compare them.
1344                          * All vec entries are PAGE_SIZE;
1345                          */
1346                         for (j = 0; j < vcnt; j++)
1347                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1348                                            page_address(tbio->bi_io_vec[j].bv_page),
1349                                            PAGE_SIZE))
1350                                         break;
1351                         if (j == vcnt)
1352                                 continue;
1353                         mddev->resync_mismatches += r10_bio->sectors;
1354                 }
1355                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1356                         /* Don't fix anything. */
1357                         continue;
1358                 /* Ok, we need to write this bio
1359                  * First we need to fixup bv_offset, bv_len and
1360                  * bi_vecs, as the read request might have corrupted these
1361                  */
1362                 tbio->bi_vcnt = vcnt;
1363                 tbio->bi_size = r10_bio->sectors << 9;
1364                 tbio->bi_idx = 0;
1365                 tbio->bi_phys_segments = 0;
1366                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1367                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1368                 tbio->bi_next = NULL;
1369                 tbio->bi_rw = WRITE;
1370                 tbio->bi_private = r10_bio;
1371                 tbio->bi_sector = r10_bio->devs[i].addr;
1372
1373                 for (j=0; j < vcnt ; j++) {
1374                         tbio->bi_io_vec[j].bv_offset = 0;
1375                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1376
1377                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1378                                page_address(fbio->bi_io_vec[j].bv_page),
1379                                PAGE_SIZE);
1380                 }
1381                 tbio->bi_end_io = end_sync_write;
1382
1383                 d = r10_bio->devs[i].devnum;
1384                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1385                 atomic_inc(&r10_bio->remaining);
1386                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1387
1388                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1389                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1390                 generic_make_request(tbio);
1391         }
1392
1393 done:
1394         if (atomic_dec_and_test(&r10_bio->remaining)) {
1395                 md_done_sync(mddev, r10_bio->sectors, 1);
1396                 put_buf(r10_bio);
1397         }
1398 }
1399
1400 /*
1401  * Now for the recovery code.
1402  * Recovery happens across physical sectors.
1403  * We recover all non-is_sync drives by finding the virtual address of
1404  * each, and then choose a working drive that also has that virt address.
1405  * There is a separate r10_bio for each non-in_sync drive.
1406  * Only the first two slots are in use. The first for reading,
1407  * The second for writing.
1408  *
1409  */
1410
1411 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1412 {
1413         conf_t *conf = mddev->private;
1414         int i, d;
1415         struct bio *bio, *wbio;
1416
1417
1418         /* move the pages across to the second bio
1419          * and submit the write request
1420          */
1421         bio = r10_bio->devs[0].bio;
1422         wbio = r10_bio->devs[1].bio;
1423         for (i=0; i < wbio->bi_vcnt; i++) {
1424                 struct page *p = bio->bi_io_vec[i].bv_page;
1425                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1426                 wbio->bi_io_vec[i].bv_page = p;
1427         }
1428         d = r10_bio->devs[1].devnum;
1429
1430         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1431         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1432         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1433                 generic_make_request(wbio);
1434         else
1435                 bio_endio(wbio, -EIO);
1436 }
1437
1438
1439 /*
1440  * Used by fix_read_error() to decay the per rdev read_errors.
1441  * We halve the read error count for every hour that has elapsed
1442  * since the last recorded read error.
1443  *
1444  */
1445 static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1446 {
1447         struct timespec cur_time_mon;
1448         unsigned long hours_since_last;
1449         unsigned int read_errors = atomic_read(&rdev->read_errors);
1450
1451         ktime_get_ts(&cur_time_mon);
1452
1453         if (rdev->last_read_error.tv_sec == 0 &&
1454             rdev->last_read_error.tv_nsec == 0) {
1455                 /* first time we've seen a read error */
1456                 rdev->last_read_error = cur_time_mon;
1457                 return;
1458         }
1459
1460         hours_since_last = (cur_time_mon.tv_sec -
1461                             rdev->last_read_error.tv_sec) / 3600;
1462
1463         rdev->last_read_error = cur_time_mon;
1464
1465         /*
1466          * if hours_since_last is > the number of bits in read_errors
1467          * just set read errors to 0. We do this to avoid
1468          * overflowing the shift of read_errors by hours_since_last.
1469          */
1470         if (hours_since_last >= 8 * sizeof(read_errors))
1471                 atomic_set(&rdev->read_errors, 0);
1472         else
1473                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1474 }
1475
1476 /*
1477  * This is a kernel thread which:
1478  *
1479  *      1.      Retries failed read operations on working mirrors.
1480  *      2.      Updates the raid superblock when problems encounter.
1481  *      3.      Performs writes following reads for array synchronising.
1482  */
1483
1484 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1485 {
1486         int sect = 0; /* Offset from r10_bio->sector */
1487         int sectors = r10_bio->sectors;
1488         mdk_rdev_t*rdev;
1489         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1490
1491         rcu_read_lock();
1492         {
1493                 int d = r10_bio->devs[r10_bio->read_slot].devnum;
1494                 char b[BDEVNAME_SIZE];
1495                 int cur_read_error_count = 0;
1496
1497                 rdev = rcu_dereference(conf->mirrors[d].rdev);
1498                 bdevname(rdev->bdev, b);
1499
1500                 if (test_bit(Faulty, &rdev->flags)) {
1501                         rcu_read_unlock();
1502                         /* drive has already been failed, just ignore any
1503                            more fix_read_error() attempts */
1504                         return;
1505                 }
1506
1507                 check_decay_read_errors(mddev, rdev);
1508                 atomic_inc(&rdev->read_errors);
1509                 cur_read_error_count = atomic_read(&rdev->read_errors);
1510                 if (cur_read_error_count > max_read_errors) {
1511                         rcu_read_unlock();
1512                         printk(KERN_NOTICE
1513                                "raid10: %s: Raid device exceeded "
1514                                "read_error threshold "
1515                                "[cur %d:max %d]\n",
1516                                b, cur_read_error_count, max_read_errors);
1517                         printk(KERN_NOTICE
1518                                "raid10: %s: Failing raid "
1519                                "device\n", b);
1520                         md_error(mddev, conf->mirrors[d].rdev);
1521                         return;
1522                 }
1523         }
1524         rcu_read_unlock();
1525
1526         while(sectors) {
1527                 int s = sectors;
1528                 int sl = r10_bio->read_slot;
1529                 int success = 0;
1530                 int start;
1531
1532                 if (s > (PAGE_SIZE>>9))
1533                         s = PAGE_SIZE >> 9;
1534
1535                 rcu_read_lock();
1536                 do {
1537                         int d = r10_bio->devs[sl].devnum;
1538                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1539                         if (rdev &&
1540                             test_bit(In_sync, &rdev->flags)) {
1541                                 atomic_inc(&rdev->nr_pending);
1542                                 rcu_read_unlock();
1543                                 success = sync_page_io(rdev->bdev,
1544                                                        r10_bio->devs[sl].addr +
1545                                                        sect + rdev->data_offset,
1546                                                        s<<9,
1547                                                        conf->tmppage, READ);
1548                                 rdev_dec_pending(rdev, mddev);
1549                                 rcu_read_lock();
1550                                 if (success)
1551                                         break;
1552                         }
1553                         sl++;
1554                         if (sl == conf->copies)
1555                                 sl = 0;
1556                 } while (!success && sl != r10_bio->read_slot);
1557                 rcu_read_unlock();
1558
1559                 if (!success) {
1560                         /* Cannot read from anywhere -- bye bye array */
1561                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1562                         md_error(mddev, conf->mirrors[dn].rdev);
1563                         break;
1564                 }
1565
1566                 start = sl;
1567                 /* write it back and re-read */
1568                 rcu_read_lock();
1569                 while (sl != r10_bio->read_slot) {
1570                         char b[BDEVNAME_SIZE];
1571                         int d;
1572                         if (sl==0)
1573                                 sl = conf->copies;
1574                         sl--;
1575                         d = r10_bio->devs[sl].devnum;
1576                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1577                         if (rdev &&
1578                             test_bit(In_sync, &rdev->flags)) {
1579                                 atomic_inc(&rdev->nr_pending);
1580                                 rcu_read_unlock();
1581                                 atomic_add(s, &rdev->corrected_errors);
1582                                 if (sync_page_io(rdev->bdev,
1583                                                  r10_bio->devs[sl].addr +
1584                                                  sect + rdev->data_offset,
1585                                                  s<<9, conf->tmppage, WRITE)
1586                                     == 0) {
1587                                         /* Well, this device is dead */
1588                                         printk(KERN_NOTICE
1589                                                "raid10:%s: read correction "
1590                                                "write failed"
1591                                                " (%d sectors at %llu on %s)\n",
1592                                                mdname(mddev), s,
1593                                                (unsigned long long)(sect+
1594                                                rdev->data_offset),
1595                                                bdevname(rdev->bdev, b));
1596                                         printk(KERN_NOTICE "raid10:%s: failing "
1597                                                "drive\n",
1598                                                bdevname(rdev->bdev, b));
1599                                         md_error(mddev, rdev);
1600                                 }
1601                                 rdev_dec_pending(rdev, mddev);
1602                                 rcu_read_lock();
1603                         }
1604                 }
1605                 sl = start;
1606                 while (sl != r10_bio->read_slot) {
1607                         int d;
1608                         if (sl==0)
1609                                 sl = conf->copies;
1610                         sl--;
1611                         d = r10_bio->devs[sl].devnum;
1612                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1613                         if (rdev &&
1614                             test_bit(In_sync, &rdev->flags)) {
1615                                 char b[BDEVNAME_SIZE];
1616                                 atomic_inc(&rdev->nr_pending);
1617                                 rcu_read_unlock();
1618                                 if (sync_page_io(rdev->bdev,
1619                                                  r10_bio->devs[sl].addr +
1620                                                  sect + rdev->data_offset,
1621                                                  s<<9, conf->tmppage,
1622                                                  READ) == 0) {
1623                                         /* Well, this device is dead */
1624                                         printk(KERN_NOTICE
1625                                                "raid10:%s: unable to read back "
1626                                                "corrected sectors"
1627                                                " (%d sectors at %llu on %s)\n",
1628                                                mdname(mddev), s,
1629                                                (unsigned long long)(sect+
1630                                                     rdev->data_offset),
1631                                                bdevname(rdev->bdev, b));
1632                                         printk(KERN_NOTICE "raid10:%s: failing drive\n",
1633                                                bdevname(rdev->bdev, b));
1634
1635                                         md_error(mddev, rdev);
1636                                 } else {
1637                                         printk(KERN_INFO
1638                                                "raid10:%s: read error corrected"
1639                                                " (%d sectors at %llu on %s)\n",
1640                                                mdname(mddev), s,
1641                                                (unsigned long long)(sect+
1642                                                     rdev->data_offset),
1643                                                bdevname(rdev->bdev, b));
1644                                 }
1645
1646                                 rdev_dec_pending(rdev, mddev);
1647                                 rcu_read_lock();
1648                         }
1649                 }
1650                 rcu_read_unlock();
1651
1652                 sectors -= s;
1653                 sect += s;
1654         }
1655 }
1656
1657 static void raid10d(mddev_t *mddev)
1658 {
1659         r10bio_t *r10_bio;
1660         struct bio *bio;
1661         unsigned long flags;
1662         conf_t *conf = mddev->private;
1663         struct list_head *head = &conf->retry_list;
1664         int unplug=0;
1665         mdk_rdev_t *rdev;
1666
1667         md_check_recovery(mddev);
1668
1669         for (;;) {
1670                 char b[BDEVNAME_SIZE];
1671
1672                 unplug += flush_pending_writes(conf);
1673
1674                 spin_lock_irqsave(&conf->device_lock, flags);
1675                 if (list_empty(head)) {
1676                         spin_unlock_irqrestore(&conf->device_lock, flags);
1677                         break;
1678                 }
1679                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1680                 list_del(head->prev);
1681                 conf->nr_queued--;
1682                 spin_unlock_irqrestore(&conf->device_lock, flags);
1683
1684                 mddev = r10_bio->mddev;
1685                 conf = mddev->private;
1686                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1687                         sync_request_write(mddev, r10_bio);
1688                         unplug = 1;
1689                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1690                         recovery_request_write(mddev, r10_bio);
1691                         unplug = 1;
1692                 } else {
1693                         int mirror;
1694                         /* we got a read error. Maybe the drive is bad.  Maybe just
1695                          * the block and we can fix it.
1696                          * We freeze all other IO, and try reading the block from
1697                          * other devices.  When we find one, we re-write
1698                          * and check it that fixes the read error.
1699                          * This is all done synchronously while the array is
1700                          * frozen.
1701                          */
1702                         if (mddev->ro == 0) {
1703                                 freeze_array(conf);
1704                                 fix_read_error(conf, mddev, r10_bio);
1705                                 unfreeze_array(conf);
1706                         }
1707
1708                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1709                         r10_bio->devs[r10_bio->read_slot].bio =
1710                                 mddev->ro ? IO_BLOCKED : NULL;
1711                         mirror = read_balance(conf, r10_bio);
1712                         if (mirror == -1) {
1713                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1714                                        " read error for block %llu\n",
1715                                        bdevname(bio->bi_bdev,b),
1716                                        (unsigned long long)r10_bio->sector);
1717                                 raid_end_bio_io(r10_bio);
1718                                 bio_put(bio);
1719                         } else {
1720                                 const bool do_sync = bio_rw_flagged(r10_bio->master_bio, BIO_RW_SYNCIO);
1721                                 bio_put(bio);
1722                                 rdev = conf->mirrors[mirror].rdev;
1723                                 if (printk_ratelimit())
1724                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1725                                                " another mirror\n",
1726                                                bdevname(rdev->bdev,b),
1727                                                (unsigned long long)r10_bio->sector);
1728                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1729                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1730                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1731                                         + rdev->data_offset;
1732                                 bio->bi_bdev = rdev->bdev;
1733                                 bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
1734                                 bio->bi_private = r10_bio;
1735                                 bio->bi_end_io = raid10_end_read_request;
1736                                 unplug = 1;
1737                                 generic_make_request(bio);
1738                         }
1739                 }
1740                 cond_resched();
1741         }
1742         if (unplug)
1743                 unplug_slaves(mddev);
1744 }
1745
1746
1747 static int init_resync(conf_t *conf)
1748 {
1749         int buffs;
1750
1751         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1752         BUG_ON(conf->r10buf_pool);
1753         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1754         if (!conf->r10buf_pool)
1755                 return -ENOMEM;
1756         conf->next_resync = 0;
1757         return 0;
1758 }
1759
1760 /*
1761  * perform a "sync" on one "block"
1762  *
1763  * We need to make sure that no normal I/O request - particularly write
1764  * requests - conflict with active sync requests.
1765  *
1766  * This is achieved by tracking pending requests and a 'barrier' concept
1767  * that can be installed to exclude normal IO requests.
1768  *
1769  * Resync and recovery are handled very differently.
1770  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1771  *
1772  * For resync, we iterate over virtual addresses, read all copies,
1773  * and update if there are differences.  If only one copy is live,
1774  * skip it.
1775  * For recovery, we iterate over physical addresses, read a good
1776  * value for each non-in_sync drive, and over-write.
1777  *
1778  * So, for recovery we may have several outstanding complex requests for a
1779  * given address, one for each out-of-sync device.  We model this by allocating
1780  * a number of r10_bio structures, one for each out-of-sync device.
1781  * As we setup these structures, we collect all bio's together into a list
1782  * which we then process collectively to add pages, and then process again
1783  * to pass to generic_make_request.
1784  *
1785  * The r10_bio structures are linked using a borrowed master_bio pointer.
1786  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1787  * has its remaining count decremented to 0, the whole complex operation
1788  * is complete.
1789  *
1790  */
1791
1792 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1793 {
1794         conf_t *conf = mddev->private;
1795         r10bio_t *r10_bio;
1796         struct bio *biolist = NULL, *bio;
1797         sector_t max_sector, nr_sectors;
1798         int disk;
1799         int i;
1800         int max_sync;
1801         int sync_blocks;
1802
1803         sector_t sectors_skipped = 0;
1804         int chunks_skipped = 0;
1805
1806         if (!conf->r10buf_pool)
1807                 if (init_resync(conf))
1808                         return 0;
1809
1810  skipped:
1811         max_sector = mddev->dev_sectors;
1812         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1813                 max_sector = mddev->resync_max_sectors;
1814         if (sector_nr >= max_sector) {
1815                 /* If we aborted, we need to abort the
1816                  * sync on the 'current' bitmap chucks (there can
1817                  * be several when recovering multiple devices).
1818                  * as we may have started syncing it but not finished.
1819                  * We can find the current address in
1820                  * mddev->curr_resync, but for recovery,
1821                  * we need to convert that to several
1822                  * virtual addresses.
1823                  */
1824                 if (mddev->curr_resync < max_sector) { /* aborted */
1825                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1826                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1827                                                 &sync_blocks, 1);
1828                         else for (i=0; i<conf->raid_disks; i++) {
1829                                 sector_t sect =
1830                                         raid10_find_virt(conf, mddev->curr_resync, i);
1831                                 bitmap_end_sync(mddev->bitmap, sect,
1832                                                 &sync_blocks, 1);
1833                         }
1834                 } else /* completed sync */
1835                         conf->fullsync = 0;
1836
1837                 bitmap_close_sync(mddev->bitmap);
1838                 close_sync(conf);
1839                 *skipped = 1;
1840                 return sectors_skipped;
1841         }
1842         if (chunks_skipped >= conf->raid_disks) {
1843                 /* if there has been nothing to do on any drive,
1844                  * then there is nothing to do at all..
1845                  */
1846                 *skipped = 1;
1847                 return (max_sector - sector_nr) + sectors_skipped;
1848         }
1849
1850         if (max_sector > mddev->resync_max)
1851                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1852
1853         /* make sure whole request will fit in a chunk - if chunks
1854          * are meaningful
1855          */
1856         if (conf->near_copies < conf->raid_disks &&
1857             max_sector > (sector_nr | conf->chunk_mask))
1858                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1859         /*
1860          * If there is non-resync activity waiting for us then
1861          * put in a delay to throttle resync.
1862          */
1863         if (!go_faster && conf->nr_waiting)
1864                 msleep_interruptible(1000);
1865
1866         /* Again, very different code for resync and recovery.
1867          * Both must result in an r10bio with a list of bios that
1868          * have bi_end_io, bi_sector, bi_bdev set,
1869          * and bi_private set to the r10bio.
1870          * For recovery, we may actually create several r10bios
1871          * with 2 bios in each, that correspond to the bios in the main one.
1872          * In this case, the subordinate r10bios link back through a
1873          * borrowed master_bio pointer, and the counter in the master
1874          * includes a ref from each subordinate.
1875          */
1876         /* First, we decide what to do and set ->bi_end_io
1877          * To end_sync_read if we want to read, and
1878          * end_sync_write if we will want to write.
1879          */
1880
1881         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1882         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1883                 /* recovery... the complicated one */
1884                 int j, k;
1885                 r10_bio = NULL;
1886
1887                 for (i=0 ; i<conf->raid_disks; i++)
1888                         if (conf->mirrors[i].rdev &&
1889                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1890                                 int still_degraded = 0;
1891                                 /* want to reconstruct this device */
1892                                 r10bio_t *rb2 = r10_bio;
1893                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1894                                 int must_sync;
1895                                 /* Unless we are doing a full sync, we only need
1896                                  * to recover the block if it is set in the bitmap
1897                                  */
1898                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1899                                                               &sync_blocks, 1);
1900                                 if (sync_blocks < max_sync)
1901                                         max_sync = sync_blocks;
1902                                 if (!must_sync &&
1903                                     !conf->fullsync) {
1904                                         /* yep, skip the sync_blocks here, but don't assume
1905                                          * that there will never be anything to do here
1906                                          */
1907                                         chunks_skipped = -1;
1908                                         continue;
1909                                 }
1910
1911                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1912                                 raise_barrier(conf, rb2 != NULL);
1913                                 atomic_set(&r10_bio->remaining, 0);
1914
1915                                 r10_bio->master_bio = (struct bio*)rb2;
1916                                 if (rb2)
1917                                         atomic_inc(&rb2->remaining);
1918                                 r10_bio->mddev = mddev;
1919                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1920                                 r10_bio->sector = sect;
1921
1922                                 raid10_find_phys(conf, r10_bio);
1923
1924                                 /* Need to check if the array will still be
1925                                  * degraded
1926                                  */
1927                                 for (j=0; j<conf->raid_disks; j++)
1928                                         if (conf->mirrors[j].rdev == NULL ||
1929                                             test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
1930                                                 still_degraded = 1;
1931                                                 break;
1932                                         }
1933
1934                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1935                                                               &sync_blocks, still_degraded);
1936
1937                                 for (j=0; j<conf->copies;j++) {
1938                                         int d = r10_bio->devs[j].devnum;
1939                                         if (conf->mirrors[d].rdev &&
1940                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1941                                                 /* This is where we read from */
1942                                                 bio = r10_bio->devs[0].bio;
1943                                                 bio->bi_next = biolist;
1944                                                 biolist = bio;
1945                                                 bio->bi_private = r10_bio;
1946                                                 bio->bi_end_io = end_sync_read;
1947                                                 bio->bi_rw = READ;
1948                                                 bio->bi_sector = r10_bio->devs[j].addr +
1949                                                         conf->mirrors[d].rdev->data_offset;
1950                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1951                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1952                                                 atomic_inc(&r10_bio->remaining);
1953                                                 /* and we write to 'i' */
1954
1955                                                 for (k=0; k<conf->copies; k++)
1956                                                         if (r10_bio->devs[k].devnum == i)
1957                                                                 break;
1958                                                 BUG_ON(k == conf->copies);
1959                                                 bio = r10_bio->devs[1].bio;
1960                                                 bio->bi_next = biolist;
1961                                                 biolist = bio;
1962                                                 bio->bi_private = r10_bio;
1963                                                 bio->bi_end_io = end_sync_write;
1964                                                 bio->bi_rw = WRITE;
1965                                                 bio->bi_sector = r10_bio->devs[k].addr +
1966                                                         conf->mirrors[i].rdev->data_offset;
1967                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1968
1969                                                 r10_bio->devs[0].devnum = d;
1970                                                 r10_bio->devs[1].devnum = i;
1971
1972                                                 break;
1973                                         }
1974                                 }
1975                                 if (j == conf->copies) {
1976                                         /* Cannot recover, so abort the recovery */
1977                                         put_buf(r10_bio);
1978                                         if (rb2)
1979                                                 atomic_dec(&rb2->remaining);
1980                                         r10_bio = rb2;
1981                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
1982                                                               &mddev->recovery))
1983                                                 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1984                                                        mdname(mddev));
1985                                         break;
1986                                 }
1987                         }
1988                 if (biolist == NULL) {
1989                         while (r10_bio) {
1990                                 r10bio_t *rb2 = r10_bio;
1991                                 r10_bio = (r10bio_t*) rb2->master_bio;
1992                                 rb2->master_bio = NULL;
1993                                 put_buf(rb2);
1994                         }
1995                         goto giveup;
1996                 }
1997         } else {
1998                 /* resync. Schedule a read for every block at this virt offset */
1999                 int count = 0;
2000
2001                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2002
2003                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2004                                        &sync_blocks, mddev->degraded) &&
2005                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2006                         /* We can skip this block */
2007                         *skipped = 1;
2008                         return sync_blocks + sectors_skipped;
2009                 }
2010                 if (sync_blocks < max_sync)
2011                         max_sync = sync_blocks;
2012                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2013
2014                 r10_bio->mddev = mddev;
2015                 atomic_set(&r10_bio->remaining, 0);
2016                 raise_barrier(conf, 0);
2017                 conf->next_resync = sector_nr;
2018
2019                 r10_bio->master_bio = NULL;
2020                 r10_bio->sector = sector_nr;
2021                 set_bit(R10BIO_IsSync, &r10_bio->state);
2022                 raid10_find_phys(conf, r10_bio);
2023                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2024
2025                 for (i=0; i<conf->copies; i++) {
2026                         int d = r10_bio->devs[i].devnum;
2027                         bio = r10_bio->devs[i].bio;
2028                         bio->bi_end_io = NULL;
2029                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
2030                         if (conf->mirrors[d].rdev == NULL ||
2031                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2032                                 continue;
2033                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2034                         atomic_inc(&r10_bio->remaining);
2035                         bio->bi_next = biolist;
2036                         biolist = bio;
2037                         bio->bi_private = r10_bio;
2038                         bio->bi_end_io = end_sync_read;
2039                         bio->bi_rw = READ;
2040                         bio->bi_sector = r10_bio->devs[i].addr +
2041                                 conf->mirrors[d].rdev->data_offset;
2042                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2043                         count++;
2044                 }
2045
2046                 if (count < 2) {
2047                         for (i=0; i<conf->copies; i++) {
2048                                 int d = r10_bio->devs[i].devnum;
2049                                 if (r10_bio->devs[i].bio->bi_end_io)
2050                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
2051                         }
2052                         put_buf(r10_bio);
2053                         biolist = NULL;
2054                         goto giveup;
2055                 }
2056         }
2057
2058         for (bio = biolist; bio ; bio=bio->bi_next) {
2059
2060                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2061                 if (bio->bi_end_io)
2062                         bio->bi_flags |= 1 << BIO_UPTODATE;
2063                 bio->bi_vcnt = 0;
2064                 bio->bi_idx = 0;
2065                 bio->bi_phys_segments = 0;
2066                 bio->bi_size = 0;
2067         }
2068
2069         nr_sectors = 0;
2070         if (sector_nr + max_sync < max_sector)
2071                 max_sector = sector_nr + max_sync;
2072         do {
2073                 struct page *page;
2074                 int len = PAGE_SIZE;
2075                 disk = 0;
2076                 if (sector_nr + (len>>9) > max_sector)
2077                         len = (max_sector - sector_nr) << 9;
2078                 if (len == 0)
2079                         break;
2080                 for (bio= biolist ; bio ; bio=bio->bi_next) {
2081                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2082                         if (bio_add_page(bio, page, len, 0) == 0) {
2083                                 /* stop here */
2084                                 struct bio *bio2;
2085                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2086                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
2087                                         /* remove last page from this bio */
2088                                         bio2->bi_vcnt--;
2089                                         bio2->bi_size -= len;
2090                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2091                                 }
2092                                 goto bio_full;
2093                         }
2094                         disk = i;
2095                 }
2096                 nr_sectors += len>>9;
2097                 sector_nr += len>>9;
2098         } while (biolist->bi_vcnt < RESYNC_PAGES);
2099  bio_full:
2100         r10_bio->sectors = nr_sectors;
2101
2102         while (biolist) {
2103                 bio = biolist;
2104                 biolist = biolist->bi_next;
2105
2106                 bio->bi_next = NULL;
2107                 r10_bio = bio->bi_private;
2108                 r10_bio->sectors = nr_sectors;
2109
2110                 if (bio->bi_end_io == end_sync_read) {
2111                         md_sync_acct(bio->bi_bdev, nr_sectors);
2112                         generic_make_request(bio);
2113                 }
2114         }
2115
2116         if (sectors_skipped)
2117                 /* pretend they weren't skipped, it makes
2118                  * no important difference in this case
2119                  */
2120                 md_done_sync(mddev, sectors_skipped, 1);
2121
2122         return sectors_skipped + nr_sectors;
2123  giveup:
2124         /* There is nowhere to write, so all non-sync
2125          * drives must be failed, so try the next chunk...
2126          */
2127         if (sector_nr + max_sync < max_sector)
2128                 max_sector = sector_nr + max_sync;
2129
2130         sectors_skipped += (max_sector - sector_nr);
2131         chunks_skipped ++;
2132         sector_nr = max_sector;
2133         goto skipped;
2134 }
2135
2136 static sector_t
2137 raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2138 {
2139         sector_t size;
2140         conf_t *conf = mddev->private;
2141
2142         if (!raid_disks)
2143                 raid_disks = mddev->raid_disks;
2144         if (!sectors)
2145                 sectors = mddev->dev_sectors;
2146
2147         size = sectors >> conf->chunk_shift;
2148         sector_div(size, conf->far_copies);
2149         size = size * raid_disks;
2150         sector_div(size, conf->near_copies);
2151
2152         return size << conf->chunk_shift;
2153 }
2154
2155 static int run(mddev_t *mddev)
2156 {
2157         conf_t *conf;
2158         int i, disk_idx, chunk_size;
2159         mirror_info_t *disk;
2160         mdk_rdev_t *rdev;
2161         int nc, fc, fo;
2162         sector_t stride, size;
2163
2164         if (mddev->chunk_sectors < (PAGE_SIZE >> 9) ||
2165             !is_power_of_2(mddev->chunk_sectors)) {
2166                 printk(KERN_ERR "md/raid10: chunk size must be "
2167                        "at least PAGE_SIZE(%ld) and be a power of 2.\n", PAGE_SIZE);
2168                 return -EINVAL;
2169         }
2170
2171         nc = mddev->layout & 255;
2172         fc = (mddev->layout >> 8) & 255;
2173         fo = mddev->layout & (1<<16);
2174         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2175             (mddev->layout >> 17)) {
2176                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
2177                        mdname(mddev), mddev->layout);
2178                 goto out;
2179         }
2180         /*
2181          * copy the already verified devices into our private RAID10
2182          * bookkeeping area. [whatever we allocate in run(),
2183          * should be freed in stop()]
2184          */
2185         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2186         mddev->private = conf;
2187         if (!conf) {
2188                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2189                         mdname(mddev));
2190                 goto out;
2191         }
2192         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2193                                  GFP_KERNEL);
2194         if (!conf->mirrors) {
2195                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2196                        mdname(mddev));
2197                 goto out_free_conf;
2198         }
2199
2200         conf->tmppage = alloc_page(GFP_KERNEL);
2201         if (!conf->tmppage)
2202                 goto out_free_conf;
2203
2204         conf->raid_disks = mddev->raid_disks;
2205         conf->near_copies = nc;
2206         conf->far_copies = fc;
2207         conf->copies = nc*fc;
2208         conf->far_offset = fo;
2209         conf->chunk_mask = mddev->chunk_sectors - 1;
2210         conf->chunk_shift = ffz(~mddev->chunk_sectors);
2211         size = mddev->dev_sectors >> conf->chunk_shift;
2212         sector_div(size, fc);
2213         size = size * conf->raid_disks;
2214         sector_div(size, nc);
2215         /* 'size' is now the number of chunks in the array */
2216         /* calculate "used chunks per device" in 'stride' */
2217         stride = size * conf->copies;
2218
2219         /* We need to round up when dividing by raid_disks to
2220          * get the stride size.
2221          */
2222         stride += conf->raid_disks - 1;
2223         sector_div(stride, conf->raid_disks);
2224         mddev->dev_sectors = stride << conf->chunk_shift;
2225
2226         if (fo)
2227                 stride = 1;
2228         else
2229                 sector_div(stride, fc);
2230         conf->stride = stride << conf->chunk_shift;
2231
2232         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2233                                                 r10bio_pool_free, conf);
2234         if (!conf->r10bio_pool) {
2235                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2236                         mdname(mddev));
2237                 goto out_free_conf;
2238         }
2239
2240         conf->mddev = mddev;
2241         spin_lock_init(&conf->device_lock);
2242         mddev->queue->queue_lock = &conf->device_lock;
2243
2244         chunk_size = mddev->chunk_sectors << 9;
2245         blk_queue_io_min(mddev->queue, chunk_size);
2246         if (conf->raid_disks % conf->near_copies)
2247                 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2248         else
2249                 blk_queue_io_opt(mddev->queue, chunk_size *
2250                                  (conf->raid_disks / conf->near_copies));
2251
2252         list_for_each_entry(rdev, &mddev->disks, same_set) {
2253                 disk_idx = rdev->raid_disk;
2254                 if (disk_idx >= mddev->raid_disks
2255                     || disk_idx < 0)
2256                         continue;
2257                 disk = conf->mirrors + disk_idx;
2258
2259                 disk->rdev = rdev;
2260                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2261                                   rdev->data_offset << 9);
2262                 /* as we don't honour merge_bvec_fn, we must never risk
2263                  * violating it, so limit max_segments to 1 lying
2264                  * within a single page.
2265                  */
2266                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2267                         blk_queue_max_segments(mddev->queue, 1);
2268                         blk_queue_segment_boundary(mddev->queue,
2269                                                    PAGE_CACHE_SIZE - 1);
2270                 }
2271
2272                 disk->head_position = 0;
2273         }
2274         INIT_LIST_HEAD(&conf->retry_list);
2275
2276         spin_lock_init(&conf->resync_lock);
2277         init_waitqueue_head(&conf->wait_barrier);
2278
2279         /* need to check that every block has at least one working mirror */
2280         if (!enough(conf)) {
2281                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2282                        mdname(mddev));
2283                 goto out_free_conf;
2284         }
2285
2286         mddev->degraded = 0;
2287         for (i = 0; i < conf->raid_disks; i++) {
2288
2289                 disk = conf->mirrors + i;
2290
2291                 if (!disk->rdev ||
2292                     !test_bit(In_sync, &disk->rdev->flags)) {
2293                         disk->head_position = 0;
2294                         mddev->degraded++;
2295                         if (disk->rdev)
2296                                 conf->fullsync = 1;
2297                 }
2298         }
2299
2300
2301         mddev->thread = md_register_thread(raid10d, mddev, NULL);
2302         if (!mddev->thread) {
2303                 printk(KERN_ERR
2304                        "raid10: couldn't allocate thread for %s\n",
2305                        mdname(mddev));
2306                 goto out_free_conf;
2307         }
2308
2309         if (mddev->recovery_cp != MaxSector)
2310                 printk(KERN_NOTICE "raid10: %s is not clean"
2311                        " -- starting background reconstruction\n",
2312                        mdname(mddev));
2313         printk(KERN_INFO
2314                 "raid10: raid set %s active with %d out of %d devices\n",
2315                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2316                 mddev->raid_disks);
2317         /*
2318          * Ok, everything is just fine now
2319          */
2320         md_set_array_sectors(mddev, raid10_size(mddev, 0, 0));
2321         mddev->resync_max_sectors = raid10_size(mddev, 0, 0);
2322
2323         mddev->queue->unplug_fn = raid10_unplug;
2324         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2325         mddev->queue->backing_dev_info.congested_data = mddev;
2326
2327         /* Calculate max read-ahead size.
2328          * We need to readahead at least twice a whole stripe....
2329          * maybe...
2330          */
2331         {
2332                 int stripe = conf->raid_disks *
2333                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2334                 stripe /= conf->near_copies;
2335                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2336                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2337         }
2338
2339         if (conf->near_copies < mddev->raid_disks)
2340                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2341         md_integrity_register(mddev);
2342         return 0;
2343
2344 out_free_conf:
2345         if (conf->r10bio_pool)
2346                 mempool_destroy(conf->r10bio_pool);
2347         safe_put_page(conf->tmppage);
2348         kfree(conf->mirrors);
2349         kfree(conf);
2350         mddev->private = NULL;
2351 out:
2352         return -EIO;
2353 }
2354
2355 static int stop(mddev_t *mddev)
2356 {
2357         conf_t *conf = mddev->private;
2358
2359         raise_barrier(conf, 0);
2360         lower_barrier(conf);
2361
2362         md_unregister_thread(mddev->thread);
2363         mddev->thread = NULL;
2364         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2365         if (conf->r10bio_pool)
2366                 mempool_destroy(conf->r10bio_pool);
2367         kfree(conf->mirrors);
2368         kfree(conf);
2369         mddev->private = NULL;
2370         return 0;
2371 }
2372
2373 static void raid10_quiesce(mddev_t *mddev, int state)
2374 {
2375         conf_t *conf = mddev->private;
2376
2377         switch(state) {
2378         case 1:
2379                 raise_barrier(conf, 0);
2380                 break;
2381         case 0:
2382                 lower_barrier(conf);
2383                 break;
2384         }
2385 }
2386
2387 static struct mdk_personality raid10_personality =
2388 {
2389         .name           = "raid10",
2390         .level          = 10,
2391         .owner          = THIS_MODULE,
2392         .make_request   = make_request,
2393         .run            = run,
2394         .stop           = stop,
2395         .status         = status,
2396         .error_handler  = error,
2397         .hot_add_disk   = raid10_add_disk,
2398         .hot_remove_disk= raid10_remove_disk,
2399         .spare_active   = raid10_spare_active,
2400         .sync_request   = sync_request,
2401         .quiesce        = raid10_quiesce,
2402         .size           = raid10_size,
2403 };
2404
2405 static int __init raid_init(void)
2406 {
2407         return register_md_personality(&raid10_personality);
2408 }
2409
2410 static void raid_exit(void)
2411 {
2412         unregister_md_personality(&raid10_personality);
2413 }
2414
2415 module_init(raid_init);
2416 module_exit(raid_exit);
2417 MODULE_LICENSE("GPL");
2418 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
2419 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2420 MODULE_ALIAS("md-raid10");
2421 MODULE_ALIAS("md-level-10");