linux/drivers/md/raid10.c
<<
>>
Prefs
   1/*
   2 * raid10.c : Multiple Devices driver for Linux
   3 *
   4 * Copyright (C) 2000-2004 Neil Brown
   5 *
   6 * RAID-10 support for md.
   7 *
   8 * Base on code in raid1.c.  See raid1.c for further copyright information.
   9 *
  10 *
  11 * This program is free software; you can redistribute it and/or modify
  12 * it under the terms of the GNU General Public License as published by
  13 * the Free Software Foundation; either version 2, or (at your option)
  14 * any later version.
  15 *
  16 * You should have received a copy of the GNU General Public License
  17 * (for example /usr/src/linux/COPYING); if not, write to the Free
  18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19 */
  20
  21#include <linux/slab.h>
  22#include <linux/delay.h>
  23#include <linux/blkdev.h>
  24#include <linux/module.h>
  25#include <linux/seq_file.h>
  26#include <linux/ratelimit.h>
  27#include <linux/kthread.h>
  28#include "md.h"
  29#include "raid10.h"
  30#include "raid0.h"
  31#include "bitmap.h"
  32
  33/*
  34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
  35 * The layout of data is defined by
  36 *    chunk_size
  37 *    raid_disks
  38 *    near_copies (stored in low byte of layout)
  39 *    far_copies (stored in second byte of layout)
  40 *    far_offset (stored in bit 16 of layout )
  41 *    use_far_sets (stored in bit 17 of layout )
  42 *
  43 * The data to be stored is divided into chunks using chunksize.  Each device
  44 * is divided into far_copies sections.   In each section, chunks are laid out
  45 * in a style similar to raid0, but near_copies copies of each chunk is stored
  46 * (each on a different drive).  The starting device for each section is offset
  47 * near_copies from the starting device of the previous section.  Thus there
  48 * are (near_copies * far_copies) of each chunk, and each is on a different
  49 * drive.  near_copies and far_copies must be at least one, and their product
  50 * is at most raid_disks.
  51 *
  52 * If far_offset is true, then the far_copies are handled a bit differently.
  53 * The copies are still in different stripes, but instead of being very far
  54 * apart on disk, there are adjacent stripes.
  55 *
  56 * The far and offset algorithms are handled slightly differently if
  57 * 'use_far_sets' is true.  In this case, the array's devices are grouped into
  58 * sets that are (near_copies * far_copies) in size.  The far copied stripes
  59 * are still shifted by 'near_copies' devices, but this shifting stays confined
  60 * to the set rather than the entire array.  This is done to improve the number
  61 * of device combinations that can fail without causing the array to fail.
  62 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
  63 * on a device):
  64 *    A B C D    A B C D E
  65 *      ...         ...
  66 *    D A B C    E A B C D
  67 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
  68 *    [A B] [C D]    [A B] [C D E]
  69 *    |...| |...|    |...| | ... |
  70 *    [B A] [D C]    [B A] [E C D]
  71 */
  72
  73/*
  74 * Number of guaranteed r10bios in case of extreme VM load:
  75 */
  76#define NR_RAID10_BIOS 256
  77
  78/* when we get a read error on a read-only array, we redirect to another
  79 * device without failing the first device, or trying to over-write to
  80 * correct the read error.  To keep track of bad blocks on a per-bio
  81 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
  82 */
  83#define IO_BLOCKED ((struct bio *)1)
  84/* When we successfully write to a known bad-block, we need to remove the
  85 * bad-block marking which must be done from process context.  So we record
  86 * the success by setting devs[n].bio to IO_MADE_GOOD
  87 */
  88#define IO_MADE_GOOD ((struct bio *)2)
  89
  90#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
  91
  92/* When there are this many requests queued to be written by
  93 * the raid10 thread, we become 'congested' to provide back-pressure
  94 * for writeback.
  95 */
  96static int max_queued_requests = 1024;
  97
  98static void allow_barrier(struct r10conf *conf);
  99static void lower_barrier(struct r10conf *conf);
 100static int _enough(struct r10conf *conf, int previous, int ignore);
 101static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
 102                                int *skipped);
 103static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
 104static void end_reshape_write(struct bio *bio, int error);
 105static void end_reshape(struct r10conf *conf);
 106
 107static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
 108{
 109        struct r10conf *conf = data;
 110        int size = offsetof(struct r10bio, devs[conf->copies]);
 111
 112        /* allocate a r10bio with room for raid_disks entries in the
 113         * bios array */
 114        return kzalloc(size, gfp_flags);
 115}
 116
 117static void r10bio_pool_free(void *r10_bio, void *data)
 118{
 119        kfree(r10_bio);
 120}
 121
 122/* Maximum size of each resync request */
 123#define RESYNC_BLOCK_SIZE (64*1024)
 124#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
 125/* amount of memory to reserve for resync requests */
 126#define RESYNC_WINDOW (1024*1024)
 127/* maximum number of concurrent requests, memory permitting */
 128#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
 129
 130/*
 131 * When performing a resync, we need to read and compare, so
 132 * we need as many pages are there are copies.
 133 * When performing a recovery, we need 2 bios, one for read,
 134 * one for write (we recover only one drive per r10buf)
 135 *
 136 */
 137static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
 138{
 139        struct r10conf *conf = data;
 140        struct page *page;
 141        struct r10bio *r10_bio;
 142        struct bio *bio;
 143        int i, j;
 144        int nalloc;
 145
 146        r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 147        if (!r10_bio)
 148                return NULL;
 149
 150        if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
 151            test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
 152                nalloc = conf->copies; /* resync */
 153        else
 154                nalloc = 2; /* recovery */
 155
 156        /*
 157         * Allocate bios.
 158         */
 159        for (j = nalloc ; j-- ; ) {
 160                bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 161                if (!bio)
 162                        goto out_free_bio;
 163                r10_bio->devs[j].bio = bio;
 164                if (!conf->have_replacement)
 165                        continue;
 166                bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 167                if (!bio)
 168                        goto out_free_bio;
 169                r10_bio->devs[j].repl_bio = bio;
 170        }
 171        /*
 172         * Allocate RESYNC_PAGES data pages and attach them
 173         * where needed.
 174         */
 175        for (j = 0 ; j < nalloc; j++) {
 176                struct bio *rbio = r10_bio->devs[j].repl_bio;
 177                bio = r10_bio->devs[j].bio;
 178                for (i = 0; i < RESYNC_PAGES; i++) {
 179                        if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
 180                                               &conf->mddev->recovery)) {
 181                                /* we can share bv_page's during recovery
 182                                 * and reshape */
 183                                struct bio *rbio = r10_bio->devs[0].bio;
 184                                page = rbio->bi_io_vec[i].bv_page;
 185                                get_page(page);
 186                        } else
 187                                page = alloc_page(gfp_flags);
 188                        if (unlikely(!page))
 189                                goto out_free_pages;
 190
 191                        bio->bi_io_vec[i].bv_page = page;
 192                        if (rbio)
 193                                rbio->bi_io_vec[i].bv_page = page;
 194                }
 195        }
 196
 197        return r10_bio;
 198
 199out_free_pages:
 200        for ( ; i > 0 ; i--)
 201                safe_put_page(bio->bi_io_vec[i-1].bv_page);
 202        while (j--)
 203                for (i = 0; i < RESYNC_PAGES ; i++)
 204                        safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 205        j = 0;
 206out_free_bio:
 207        for ( ; j < nalloc; j++) {
 208                if (r10_bio->devs[j].bio)
 209                        bio_put(r10_bio->devs[j].bio);
 210                if (r10_bio->devs[j].repl_bio)
 211                        bio_put(r10_bio->devs[j].repl_bio);
 212        }
 213        r10bio_pool_free(r10_bio, conf);
 214        return NULL;
 215}
 216
 217static void r10buf_pool_free(void *__r10_bio, void *data)
 218{
 219        int i;
 220        struct r10conf *conf = data;
 221        struct r10bio *r10bio = __r10_bio;
 222        int j;
 223
 224        for (j=0; j < conf->copies; j++) {
 225                struct bio *bio = r10bio->devs[j].bio;
 226                if (bio) {
 227                        for (i = 0; i < RESYNC_PAGES; i++) {
 228                                safe_put_page(bio->bi_io_vec[i].bv_page);
 229                                bio->bi_io_vec[i].bv_page = NULL;
 230                        }
 231                        bio_put(bio);
 232                }
 233                bio = r10bio->devs[j].repl_bio;
 234                if (bio)
 235                        bio_put(bio);
 236        }
 237        r10bio_pool_free(r10bio, conf);
 238}
 239
 240static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
 241{
 242        int i;
 243
 244        for (i = 0; i < conf->copies; i++) {
 245                struct bio **bio = & r10_bio->devs[i].bio;
 246                if (!BIO_SPECIAL(*bio))
 247                        bio_put(*bio);
 248                *bio = NULL;
 249                bio = &r10_bio->devs[i].repl_bio;
 250                if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
 251                        bio_put(*bio);
 252                *bio = NULL;
 253        }
 254}
 255
 256static void free_r10bio(struct r10bio *r10_bio)
 257{
 258        struct r10conf *conf = r10_bio->mddev->private;
 259
 260        put_all_bios(conf, r10_bio);
 261        mempool_free(r10_bio, conf->r10bio_pool);
 262}
 263
 264static void put_buf(struct r10bio *r10_bio)
 265{
 266        struct r10conf *conf = r10_bio->mddev->private;
 267
 268        mempool_free(r10_bio, conf->r10buf_pool);
 269
 270        lower_barrier(conf);
 271}
 272
 273static void reschedule_retry(struct r10bio *r10_bio)
 274{
 275        unsigned long flags;
 276        struct mddev *mddev = r10_bio->mddev;
 277        struct r10conf *conf = mddev->private;
 278
 279        spin_lock_irqsave(&conf->device_lock, flags);
 280        list_add(&r10_bio->retry_list, &conf->retry_list);
 281        conf->nr_queued ++;
 282        spin_unlock_irqrestore(&conf->device_lock, flags);
 283
 284        /* wake up frozen array... */
 285        wake_up(&conf->wait_barrier);
 286
 287        md_wakeup_thread(mddev->thread);
 288}
 289
 290/*
 291 * raid_end_bio_io() is called when we have finished servicing a mirrored
 292 * operation and are ready to return a success/failure code to the buffer
 293 * cache layer.
 294 */
 295static void raid_end_bio_io(struct r10bio *r10_bio)
 296{
 297        struct bio *bio = r10_bio->master_bio;
 298        int done;
 299        struct r10conf *conf = r10_bio->mddev->private;
 300
 301        if (bio->bi_phys_segments) {
 302                unsigned long flags;
 303                spin_lock_irqsave(&conf->device_lock, flags);
 304                bio->bi_phys_segments--;
 305                done = (bio->bi_phys_segments == 0);
 306                spin_unlock_irqrestore(&conf->device_lock, flags);
 307        } else
 308                done = 1;
 309        if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
 310                clear_bit(BIO_UPTODATE, &bio->bi_flags);
 311        if (done) {
 312                bio_endio(bio, 0);
 313                /*
 314                 * Wake up any possible resync thread that waits for the device
 315                 * to go idle.
 316                 */
 317                allow_barrier(conf);
 318        }
 319        free_r10bio(r10_bio);
 320}
 321
 322/*
 323 * Update disk head position estimator based on IRQ completion info.
 324 */
 325static inline void update_head_pos(int slot, struct r10bio *r10_bio)
 326{
 327        struct r10conf *conf = r10_bio->mddev->private;
 328
 329        conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 330                r10_bio->devs[slot].addr + (r10_bio->sectors);
 331}
 332
 333/*
 334 * Find the disk number which triggered given bio
 335 */
 336static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
 337                         struct bio *bio, int *slotp, int *replp)
 338{
 339        int slot;
 340        int repl = 0;
 341
 342        for (slot = 0; slot < conf->copies; slot++) {
 343                if (r10_bio->devs[slot].bio == bio)
 344                        break;
 345                if (r10_bio->devs[slot].repl_bio == bio) {
 346                        repl = 1;
 347                        break;
 348                }
 349        }
 350
 351        BUG_ON(slot == conf->copies);
 352        update_head_pos(slot, r10_bio);
 353
 354        if (slotp)
 355                *slotp = slot;
 356        if (replp)
 357                *replp = repl;
 358        return r10_bio->devs[slot].devnum;
 359}
 360
 361static void raid10_end_read_request(struct bio *bio, int error)
 362{
 363        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 364        struct r10bio *r10_bio = bio->bi_private;
 365        int slot, dev;
 366        struct md_rdev *rdev;
 367        struct r10conf *conf = r10_bio->mddev->private;
 368
 369
 370        slot = r10_bio->read_slot;
 371        dev = r10_bio->devs[slot].devnum;
 372        rdev = r10_bio->devs[slot].rdev;
 373        /*
 374         * this branch is our 'one mirror IO has finished' event handler:
 375         */
 376        update_head_pos(slot, r10_bio);
 377
 378        if (uptodate) {
 379                /*
 380                 * Set R10BIO_Uptodate in our master bio, so that
 381                 * we will return a good error code to the higher
 382                 * levels even if IO on some other mirrored buffer fails.
 383                 *
 384                 * The 'master' represents the composite IO operation to
 385                 * user-side. So if something waits for IO, then it will
 386                 * wait for the 'master' bio.
 387                 */
 388                set_bit(R10BIO_Uptodate, &r10_bio->state);
 389        } else {
 390                /* If all other devices that store this block have
 391                 * failed, we want to return the error upwards rather
 392                 * than fail the last device.  Here we redefine
 393                 * "uptodate" to mean "Don't want to retry"
 394                 */
 395                if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
 396                             rdev->raid_disk))
 397                        uptodate = 1;
 398        }
 399        if (uptodate) {
 400                raid_end_bio_io(r10_bio);
 401                rdev_dec_pending(rdev, conf->mddev);
 402        } else {
 403                /*
 404                 * oops, read error - keep the refcount on the rdev
 405                 */
 406                char b[BDEVNAME_SIZE];
 407                printk_ratelimited(KERN_ERR
 408                                   "md/raid10:%s: %s: rescheduling sector %llu\n",
 409                                   mdname(conf->mddev),
 410                                   bdevname(rdev->bdev, b),
 411                                   (unsigned long long)r10_bio->sector);
 412                set_bit(R10BIO_ReadError, &r10_bio->state);
 413                reschedule_retry(r10_bio);
 414        }
 415}
 416
 417static void close_write(struct r10bio *r10_bio)
 418{
 419        /* clear the bitmap if all writes complete successfully */
 420        bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 421                        r10_bio->sectors,
 422                        !test_bit(R10BIO_Degraded, &r10_bio->state),
 423                        0);
 424        md_write_end(r10_bio->mddev);
 425}
 426
 427static void one_write_done(struct r10bio *r10_bio)
 428{
 429        if (atomic_dec_and_test(&r10_bio->remaining)) {
 430                if (test_bit(R10BIO_WriteError, &r10_bio->state))
 431                        reschedule_retry(r10_bio);
 432                else {
 433                        close_write(r10_bio);
 434                        if (test_bit(R10BIO_MadeGood, &r10_bio->state))
 435                                reschedule_retry(r10_bio);
 436                        else
 437                                raid_end_bio_io(r10_bio);
 438                }
 439        }
 440}
 441
 442static void raid10_end_write_request(struct bio *bio, int error)
 443{
 444        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 445        struct r10bio *r10_bio = bio->bi_private;
 446        int dev;
 447        int dec_rdev = 1;
 448        struct r10conf *conf = r10_bio->mddev->private;
 449        int slot, repl;
 450        struct md_rdev *rdev = NULL;
 451
 452        dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
 453
 454        if (repl)
 455                rdev = conf->mirrors[dev].replacement;
 456        if (!rdev) {
 457                smp_rmb();
 458                repl = 0;
 459                rdev = conf->mirrors[dev].rdev;
 460        }
 461        /*
 462         * this branch is our 'one mirror IO has finished' event handler:
 463         */
 464        if (!uptodate) {
 465                if (repl)
 466                        /* Never record new bad blocks to replacement,
 467                         * just fail it.
 468                         */
 469                        md_error(rdev->mddev, rdev);
 470                else {
 471                        set_bit(WriteErrorSeen, &rdev->flags);
 472                        if (!test_and_set_bit(WantReplacement, &rdev->flags))
 473                                set_bit(MD_RECOVERY_NEEDED,
 474                                        &rdev->mddev->recovery);
 475                        set_bit(R10BIO_WriteError, &r10_bio->state);
 476                        dec_rdev = 0;
 477                }
 478        } else {
 479                /*
 480                 * Set R10BIO_Uptodate in our master bio, so that
 481                 * we will return a good error code for to the higher
 482                 * levels even if IO on some other mirrored buffer fails.
 483                 *
 484                 * The 'master' represents the composite IO operation to
 485                 * user-side. So if something waits for IO, then it will
 486                 * wait for the 'master' bio.
 487                 */
 488                sector_t first_bad;
 489                int bad_sectors;
 490
 491                /*
 492                 * Do not set R10BIO_Uptodate if the current device is
 493                 * rebuilding or Faulty. This is because we cannot use
 494                 * such device for properly reading the data back (we could
 495                 * potentially use it, if the current write would have felt
 496                 * before rdev->recovery_offset, but for simplicity we don't
 497                 * check this here.
 498                 */
 499                if (test_bit(In_sync, &rdev->flags) &&
 500                    !test_bit(Faulty, &rdev->flags))
 501                        set_bit(R10BIO_Uptodate, &r10_bio->state);
 502
 503                /* Maybe we can clear some bad blocks. */
 504                if (is_badblock(rdev,
 505                                r10_bio->devs[slot].addr,
 506                                r10_bio->sectors,
 507                                &first_bad, &bad_sectors)) {
 508                        bio_put(bio);
 509                        if (repl)
 510                                r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
 511                        else
 512                                r10_bio->devs[slot].bio = IO_MADE_GOOD;
 513                        dec_rdev = 0;
 514                        set_bit(R10BIO_MadeGood, &r10_bio->state);
 515                }
 516        }
 517
 518        /*
 519         *
 520         * Let's see if all mirrored write operations have finished
 521         * already.
 522         */
 523        one_write_done(r10_bio);
 524        if (dec_rdev)
 525                rdev_dec_pending(rdev, conf->mddev);
 526}
 527
 528/*
 529 * RAID10 layout manager
 530 * As well as the chunksize and raid_disks count, there are two
 531 * parameters: near_copies and far_copies.
 532 * near_copies * far_copies must be <= raid_disks.
 533 * Normally one of these will be 1.
 534 * If both are 1, we get raid0.
 535 * If near_copies == raid_disks, we get raid1.
 536 *
 537 * Chunks are laid out in raid0 style with near_copies copies of the
 538 * first chunk, followed by near_copies copies of the next chunk and
 539 * so on.
 540 * If far_copies > 1, then after 1/far_copies of the array has been assigned
 541 * as described above, we start again with a device offset of near_copies.
 542 * So we effectively have another copy of the whole array further down all
 543 * the drives, but with blocks on different drives.
 544 * With this layout, and block is never stored twice on the one device.
 545 *
 546 * raid10_find_phys finds the sector offset of a given virtual sector
 547 * on each device that it is on.
 548 *
 549 * raid10_find_virt does the reverse mapping, from a device and a
 550 * sector offset to a virtual address
 551 */
 552
 553static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
 554{
 555        int n,f;
 556        sector_t sector;
 557        sector_t chunk;
 558        sector_t stripe;
 559        int dev;
 560        int slot = 0;
 561        int last_far_set_start, last_far_set_size;
 562
 563        last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
 564        last_far_set_start *= geo->far_set_size;
 565
 566        last_far_set_size = geo->far_set_size;
 567        last_far_set_size += (geo->raid_disks % geo->far_set_size);
 568
 569        /* now calculate first sector/dev */
 570        chunk = r10bio->sector >> geo->chunk_shift;
 571        sector = r10bio->sector & geo->chunk_mask;
 572
 573        chunk *= geo->near_copies;
 574        stripe = chunk;
 575        dev = sector_div(stripe, geo->raid_disks);
 576        if (geo->far_offset)
 577                stripe *= geo->far_copies;
 578
 579        sector += stripe << geo->chunk_shift;
 580
 581        /* and calculate all the others */
 582        for (n = 0; n < geo->near_copies; n++) {
 583                int d = dev;
 584                int set;
 585                sector_t s = sector;
 586                r10bio->devs[slot].devnum = d;
 587                r10bio->devs[slot].addr = s;
 588                slot++;
 589
 590                for (f = 1; f < geo->far_copies; f++) {
 591                        set = d / geo->far_set_size;
 592                        d += geo->near_copies;
 593
 594                        if ((geo->raid_disks % geo->far_set_size) &&
 595                            (d > last_far_set_start)) {
 596                                d -= last_far_set_start;
 597                                d %= last_far_set_size;
 598                                d += last_far_set_start;
 599                        } else {
 600                                d %= geo->far_set_size;
 601                                d += geo->far_set_size * set;
 602                        }
 603                        s += geo->stride;
 604                        r10bio->devs[slot].devnum = d;
 605                        r10bio->devs[slot].addr = s;
 606                        slot++;
 607                }
 608                dev++;
 609                if (dev >= geo->raid_disks) {
 610                        dev = 0;
 611                        sector += (geo->chunk_mask + 1);
 612                }
 613        }
 614}
 615
 616static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
 617{
 618        struct geom *geo = &conf->geo;
 619
 620        if (conf->reshape_progress != MaxSector &&
 621            ((r10bio->sector >= conf->reshape_progress) !=
 622             conf->mddev->reshape_backwards)) {
 623                set_bit(R10BIO_Previous, &r10bio->state);
 624                geo = &conf->prev;
 625        } else
 626                clear_bit(R10BIO_Previous, &r10bio->state);
 627
 628        __raid10_find_phys(geo, r10bio);
 629}
 630
 631static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
 632{
 633        sector_t offset, chunk, vchunk;
 634        /* Never use conf->prev as this is only called during resync
 635         * or recovery, so reshape isn't happening
 636         */
 637        struct geom *geo = &conf->geo;
 638        int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
 639        int far_set_size = geo->far_set_size;
 640        int last_far_set_start;
 641
 642        if (geo->raid_disks % geo->far_set_size) {
 643                last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
 644                last_far_set_start *= geo->far_set_size;
 645
 646                if (dev >= last_far_set_start) {
 647                        far_set_size = geo->far_set_size;
 648                        far_set_size += (geo->raid_disks % geo->far_set_size);
 649                        far_set_start = last_far_set_start;
 650                }
 651        }
 652
 653        offset = sector & geo->chunk_mask;
 654        if (geo->far_offset) {
 655                int fc;
 656                chunk = sector >> geo->chunk_shift;
 657                fc = sector_div(chunk, geo->far_copies);
 658                dev -= fc * geo->near_copies;
 659                if (dev < far_set_start)
 660                        dev += far_set_size;
 661        } else {
 662                while (sector >= geo->stride) {
 663                        sector -= geo->stride;
 664                        if (dev < (geo->near_copies + far_set_start))
 665                                dev += far_set_size - geo->near_copies;
 666                        else
 667                                dev -= geo->near_copies;
 668                }
 669                chunk = sector >> geo->chunk_shift;
 670        }
 671        vchunk = chunk * geo->raid_disks + dev;
 672        sector_div(vchunk, geo->near_copies);
 673        return (vchunk << geo->chunk_shift) + offset;
 674}
 675
 676/**
 677 *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 678 *      @q: request queue
 679 *      @bvm: properties of new bio
 680 *      @biovec: the request that could be merged to it.
 681 *
 682 *      Return amount of bytes we can accept at this offset
 683 *      This requires checking for end-of-chunk if near_copies != raid_disks,
 684 *      and for subordinate merge_bvec_fns if merge_check_needed.
 685 */
 686static int raid10_mergeable_bvec(struct request_queue *q,
 687                                 struct bvec_merge_data *bvm,
 688                                 struct bio_vec *biovec)
 689{
 690        struct mddev *mddev = q->queuedata;
 691        struct r10conf *conf = mddev->private;
 692        sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 693        int max;
 694        unsigned int chunk_sectors;
 695        unsigned int bio_sectors = bvm->bi_size >> 9;
 696        struct geom *geo = &conf->geo;
 697
 698        chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
 699        if (conf->reshape_progress != MaxSector &&
 700            ((sector >= conf->reshape_progress) !=
 701             conf->mddev->reshape_backwards))
 702                geo = &conf->prev;
 703
 704        if (geo->near_copies < geo->raid_disks) {
 705                max = (chunk_sectors - ((sector & (chunk_sectors - 1))
 706                                        + bio_sectors)) << 9;
 707                if (max < 0)
 708                        /* bio_add cannot handle a negative return */
 709                        max = 0;
 710                if (max <= biovec->bv_len && bio_sectors == 0)
 711                        return biovec->bv_len;
 712        } else
 713                max = biovec->bv_len;
 714
 715        if (mddev->merge_check_needed) {
 716                struct {
 717                        struct r10bio r10_bio;
 718                        struct r10dev devs[conf->copies];
 719                } on_stack;
 720                struct r10bio *r10_bio = &on_stack.r10_bio;
 721                int s;
 722                if (conf->reshape_progress != MaxSector) {
 723                        /* Cannot give any guidance during reshape */
 724                        if (max <= biovec->bv_len && bio_sectors == 0)
 725                                return biovec->bv_len;
 726                        return 0;
 727                }
 728                r10_bio->sector = sector;
 729                raid10_find_phys(conf, r10_bio);
 730                rcu_read_lock();
 731                for (s = 0; s < conf->copies; s++) {
 732                        int disk = r10_bio->devs[s].devnum;
 733                        struct md_rdev *rdev = rcu_dereference(
 734                                conf->mirrors[disk].rdev);
 735                        if (rdev && !test_bit(Faulty, &rdev->flags)) {
 736                                struct request_queue *q =
 737                                        bdev_get_queue(rdev->bdev);
 738                                if (q->merge_bvec_fn) {
 739                                        bvm->bi_sector = r10_bio->devs[s].addr
 740                                                + rdev->data_offset;
 741                                        bvm->bi_bdev = rdev->bdev;
 742                                        max = min(max, q->merge_bvec_fn(
 743                                                          q, bvm, biovec));
 744                                }
 745                        }
 746                        rdev = rcu_dereference(conf->mirrors[disk].replacement);
 747                        if (rdev && !test_bit(Faulty, &rdev->flags)) {
 748                                struct request_queue *q =
 749                                        bdev_get_queue(rdev->bdev);
 750                                if (q->merge_bvec_fn) {
 751                                        bvm->bi_sector = r10_bio->devs[s].addr
 752                                                + rdev->data_offset;
 753                                        bvm->bi_bdev = rdev->bdev;
 754                                        max = min(max, q->merge_bvec_fn(
 755                                                          q, bvm, biovec));
 756                                }
 757                        }
 758                }
 759                rcu_read_unlock();
 760        }
 761        return max;
 762}
 763
 764/*
 765 * This routine returns the disk from which the requested read should
 766 * be done. There is a per-array 'next expected sequential IO' sector
 767 * number - if this matches on the next IO then we use the last disk.
 768 * There is also a per-disk 'last know head position' sector that is
 769 * maintained from IRQ contexts, both the normal and the resync IO
 770 * completion handlers update this position correctly. If there is no
 771 * perfect sequential match then we pick the disk whose head is closest.
 772 *
 773 * If there are 2 mirrors in the same 2 devices, performance degrades
 774 * because position is mirror, not device based.
 775 *
 776 * The rdev for the device selected will have nr_pending incremented.
 777 */
 778
 779/*
 780 * FIXME: possibly should rethink readbalancing and do it differently
 781 * depending on near_copies / far_copies geometry.
 782 */
 783static struct md_rdev *read_balance(struct r10conf *conf,
 784                                    struct r10bio *r10_bio,
 785                                    int *max_sectors)
 786{
 787        const sector_t this_sector = r10_bio->sector;
 788        int disk, slot;
 789        int sectors = r10_bio->sectors;
 790        int best_good_sectors;
 791        sector_t new_distance, best_dist;
 792        struct md_rdev *best_rdev, *rdev = NULL;
 793        int do_balance;
 794        int best_slot;
 795        struct geom *geo = &conf->geo;
 796
 797        raid10_find_phys(conf, r10_bio);
 798        rcu_read_lock();
 799retry:
 800        sectors = r10_bio->sectors;
 801        best_slot = -1;
 802        best_rdev = NULL;
 803        best_dist = MaxSector;
 804        best_good_sectors = 0;
 805        do_balance = 1;
 806        /*
 807         * Check if we can balance. We can balance on the whole
 808         * device if no resync is going on (recovery is ok), or below
 809         * the resync window. We take the first readable disk when
 810         * above the resync window.
 811         */
 812        if (conf->mddev->recovery_cp < MaxSector
 813            && (this_sector + sectors >= conf->next_resync))
 814                do_balance = 0;
 815
 816        for (slot = 0; slot < conf->copies ; slot++) {
 817                sector_t first_bad;
 818                int bad_sectors;
 819                sector_t dev_sector;
 820
 821                if (r10_bio->devs[slot].bio == IO_BLOCKED)
 822                        continue;
 823                disk = r10_bio->devs[slot].devnum;
 824                rdev = rcu_dereference(conf->mirrors[disk].replacement);
 825                if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
 826                    test_bit(Unmerged, &rdev->flags) ||
 827                    r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 828                        rdev = rcu_dereference(conf->mirrors[disk].rdev);
 829                if (rdev == NULL ||
 830                    test_bit(Faulty, &rdev->flags) ||
 831                    test_bit(Unmerged, &rdev->flags))
 832                        continue;
 833                if (!test_bit(In_sync, &rdev->flags) &&
 834                    r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 835                        continue;
 836
 837                dev_sector = r10_bio->devs[slot].addr;
 838                if (is_badblock(rdev, dev_sector, sectors,
 839                                &first_bad, &bad_sectors)) {
 840                        if (best_dist < MaxSector)
 841                                /* Already have a better slot */
 842                                continue;
 843                        if (first_bad <= dev_sector) {
 844                                /* Cannot read here.  If this is the
 845                                 * 'primary' device, then we must not read
 846                                 * beyond 'bad_sectors' from another device.
 847                                 */
 848                                bad_sectors -= (dev_sector - first_bad);
 849                                if (!do_balance && sectors > bad_sectors)
 850                                        sectors = bad_sectors;
 851                                if (best_good_sectors > sectors)
 852                                        best_good_sectors = sectors;
 853                        } else {
 854                                sector_t good_sectors =
 855                                        first_bad - dev_sector;
 856                                if (good_sectors > best_good_sectors) {
 857                                        best_good_sectors = good_sectors;
 858                                        best_slot = slot;
 859                                        best_rdev = rdev;
 860                                }
 861                                if (!do_balance)
 862                                        /* Must read from here */
 863                                        break;
 864                        }
 865                        continue;
 866                } else
 867                        best_good_sectors = sectors;
 868
 869                if (!do_balance)
 870                        break;
 871
 872                /* This optimisation is debatable, and completely destroys
 873                 * sequential read speed for 'far copies' arrays.  So only
 874                 * keep it for 'near' arrays, and review those later.
 875                 */
 876                if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
 877                        break;
 878
 879                /* for far > 1 always use the lowest address */
 880                if (geo->far_copies > 1)
 881                        new_distance = r10_bio->devs[slot].addr;
 882                else
 883                        new_distance = abs(r10_bio->devs[slot].addr -
 884                                           conf->mirrors[disk].head_position);
 885                if (new_distance < best_dist) {
 886                        best_dist = new_distance;
 887                        best_slot = slot;
 888                        best_rdev = rdev;
 889                }
 890        }
 891        if (slot >= conf->copies) {
 892                slot = best_slot;
 893                rdev = best_rdev;
 894        }
 895
 896        if (slot >= 0) {
 897                atomic_inc(&rdev->nr_pending);
 898                if (test_bit(Faulty, &rdev->flags)) {
 899                        /* Cannot risk returning a device that failed
 900                         * before we inc'ed nr_pending
 901                         */
 902                        rdev_dec_pending(rdev, conf->mddev);
 903                        goto retry;
 904                }
 905                r10_bio->read_slot = slot;
 906        } else
 907                rdev = NULL;
 908        rcu_read_unlock();
 909        *max_sectors = best_good_sectors;
 910
 911        return rdev;
 912}
 913
 914int md_raid10_congested(struct mddev *mddev, int bits)
 915{
 916        struct r10conf *conf = mddev->private;
 917        int i, ret = 0;
 918
 919        if ((bits & (1 << BDI_async_congested)) &&
 920            conf->pending_count >= max_queued_requests)
 921                return 1;
 922
 923        rcu_read_lock();
 924        for (i = 0;
 925             (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
 926                     && ret == 0;
 927             i++) {
 928                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 929                if (rdev && !test_bit(Faulty, &rdev->flags)) {
 930                        struct request_queue *q = bdev_get_queue(rdev->bdev);
 931
 932                        ret |= bdi_congested(&q->backing_dev_info, bits);
 933                }
 934        }
 935        rcu_read_unlock();
 936        return ret;
 937}
 938EXPORT_SYMBOL_GPL(md_raid10_congested);
 939
 940static int raid10_congested(void *data, int bits)
 941{
 942        struct mddev *mddev = data;
 943
 944        return mddev_congested(mddev, bits) ||
 945                md_raid10_congested(mddev, bits);
 946}
 947
 948static void flush_pending_writes(struct r10conf *conf)
 949{
 950        /* Any writes that have been queued but are awaiting
 951         * bitmap updates get flushed here.
 952         */
 953        spin_lock_irq(&conf->device_lock);
 954
 955        if (conf->pending_bio_list.head) {
 956                struct bio *bio;
 957                bio = bio_list_get(&conf->pending_bio_list);
 958                conf->pending_count = 0;
 959                spin_unlock_irq(&conf->device_lock);
 960                /* flush any pending bitmap writes to disk
 961                 * before proceeding w/ I/O */
 962                bitmap_unplug(conf->mddev->bitmap);
 963                wake_up(&conf->wait_barrier);
 964
 965                while (bio) { /* submit pending writes */
 966                        struct bio *next = bio->bi_next;
 967                        bio->bi_next = NULL;
 968                        if (unlikely((bio->bi_rw & REQ_DISCARD) &&
 969                            !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 970                                /* Just ignore it */
 971                                bio_endio(bio, 0);
 972                        else
 973                                generic_make_request(bio);
 974                        bio = next;
 975                }
 976        } else
 977                spin_unlock_irq(&conf->device_lock);
 978}
 979
 980/* Barriers....
 981 * Sometimes we need to suspend IO while we do something else,
 982 * either some resync/recovery, or reconfigure the array.
 983 * To do this we raise a 'barrier'.
 984 * The 'barrier' is a counter that can be raised multiple times
 985 * to count how many activities are happening which preclude
 986 * normal IO.
 987 * We can only raise the barrier if there is no pending IO.
 988 * i.e. if nr_pending == 0.
 989 * We choose only to raise the barrier if no-one is waiting for the
 990 * barrier to go down.  This means that as soon as an IO request
 991 * is ready, no other operations which require a barrier will start
 992 * until the IO request has had a chance.
 993 *
 994 * So: regular IO calls 'wait_barrier'.  When that returns there
 995 *    is no backgroup IO happening,  It must arrange to call
 996 *    allow_barrier when it has finished its IO.
 997 * backgroup IO calls must call raise_barrier.  Once that returns
 998 *    there is no normal IO happeing.  It must arrange to call
 999 *    lower_barrier when the particular background IO completes.
1000 */
1001
1002static void raise_barrier(struct r10conf *conf, int force)
1003{
1004        BUG_ON(force && !conf->barrier);
1005        spin_lock_irq(&conf->resync_lock);
1006
1007        /* Wait until no block IO is waiting (unless 'force') */
1008        wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
1009                            conf->resync_lock);
1010
1011        /* block any new IO from starting */
1012        conf->barrier++;
1013
1014        /* Now wait for all pending IO to complete */
1015        wait_event_lock_irq(conf->wait_barrier,
1016                            !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
1017                            conf->resync_lock);
1018
1019        spin_unlock_irq(&conf->resync_lock);
1020}
1021
1022static void lower_barrier(struct r10conf *conf)
1023{
1024        unsigned long flags;
1025        spin_lock_irqsave(&conf->resync_lock, flags);
1026        conf->barrier--;
1027        spin_unlock_irqrestore(&conf->resync_lock, flags);
1028        wake_up(&conf->wait_barrier);
1029}
1030
1031static void wait_barrier(struct r10conf *conf)
1032{
1033        spin_lock_irq(&conf->resync_lock);
1034        if (conf->barrier) {
1035                conf->nr_waiting++;
1036                /* Wait for the barrier to drop.
1037                 * However if there are already pending
1038                 * requests (preventing the barrier from
1039                 * rising completely), and the
1040                 * pre-process bio queue isn't empty,
1041                 * then don't wait, as we need to empty
1042                 * that queue to get the nr_pending
1043                 * count down.
1044                 */
1045                wait_event_lock_irq(conf->wait_barrier,
1046                                    !conf->barrier ||
1047                                    (conf->nr_pending &&
1048                                     current->bio_list &&
1049                                     !bio_list_empty(current->bio_list)),
1050                                    conf->resync_lock);
1051                conf->nr_waiting--;
1052        }
1053        conf->nr_pending++;
1054        spin_unlock_irq(&conf->resync_lock);
1055}
1056
1057static void allow_barrier(struct r10conf *conf)
1058{
1059        unsigned long flags;
1060        spin_lock_irqsave(&conf->resync_lock, flags);
1061        conf->nr_pending--;
1062        spin_unlock_irqrestore(&conf->resync_lock, flags);
1063        wake_up(&conf->wait_barrier);
1064}
1065
1066static void freeze_array(struct r10conf *conf, int extra)
1067{
1068        /* stop syncio and normal IO and wait for everything to
1069         * go quiet.
1070         * We increment barrier and nr_waiting, and then
1071         * wait until nr_pending match nr_queued+extra
1072         * This is called in the context of one normal IO request
1073         * that has failed. Thus any sync request that might be pending
1074         * will be blocked by nr_pending, and we need to wait for
1075         * pending IO requests to complete or be queued for re-try.
1076         * Thus the number queued (nr_queued) plus this request (extra)
1077         * must match the number of pending IOs (nr_pending) before
1078         * we continue.
1079         */
1080        spin_lock_irq(&conf->resync_lock);
1081        conf->barrier++;
1082        conf->nr_waiting++;
1083        wait_event_lock_irq_cmd(conf->wait_barrier,
1084                                conf->nr_pending == conf->nr_queued+extra,
1085                                conf->resync_lock,
1086                                flush_pending_writes(conf));
1087
1088        spin_unlock_irq(&conf->resync_lock);
1089}
1090
1091static void unfreeze_array(struct r10conf *conf)
1092{
1093        /* reverse the effect of the freeze */
1094        spin_lock_irq(&conf->resync_lock);
1095        conf->barrier--;
1096        conf->nr_waiting--;
1097        wake_up(&conf->wait_barrier);
1098        spin_unlock_irq(&conf->resync_lock);
1099}
1100
1101static sector_t choose_data_offset(struct r10bio *r10_bio,
1102                                   struct md_rdev *rdev)
1103{
1104        if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1105            test_bit(R10BIO_Previous, &r10_bio->state))
1106                return rdev->data_offset;
1107        else
1108                return rdev->new_data_offset;
1109}
1110
1111struct raid10_plug_cb {
1112        struct blk_plug_cb      cb;
1113        struct bio_list         pending;
1114        int                     pending_cnt;
1115};
1116
1117static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1118{
1119        struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1120                                                   cb);
1121        struct mddev *mddev = plug->cb.data;
1122        struct r10conf *conf = mddev->private;
1123        struct bio *bio;
1124
1125        if (from_schedule || current->bio_list) {
1126                spin_lock_irq(&conf->device_lock);
1127                bio_list_merge(&conf->pending_bio_list, &plug->pending);
1128                conf->pending_count += plug->pending_cnt;
1129                spin_unlock_irq(&conf->device_lock);
1130                wake_up(&conf->wait_barrier);
1131                md_wakeup_thread(mddev->thread);
1132                kfree(plug);
1133                return;
1134        }
1135
1136        /* we aren't scheduling, so we can do the write-out directly. */
1137        bio = bio_list_get(&plug->pending);
1138        bitmap_unplug(mddev->bitmap);
1139        wake_up(&conf->wait_barrier);
1140
1141        while (bio) { /* submit pending writes */
1142                struct bio *next = bio->bi_next;
1143                bio->bi_next = NULL;
1144                if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1145                    !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1146                        /* Just ignore it */
1147                        bio_endio(bio, 0);
1148                else
1149                        generic_make_request(bio);
1150                bio = next;
1151        }
1152        kfree(plug);
1153}
1154
1155static void make_request(struct mddev *mddev, struct bio * bio)
1156{
1157        struct r10conf *conf = mddev->private;
1158        struct r10bio *r10_bio;
1159        struct bio *read_bio;
1160        int i;
1161        sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1162        int chunk_sects = chunk_mask + 1;
1163        const int rw = bio_data_dir(bio);
1164        const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1165        const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1166        const unsigned long do_discard = (bio->bi_rw
1167                                          & (REQ_DISCARD | REQ_SECURE));
1168        const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1169        unsigned long flags;
1170        struct md_rdev *blocked_rdev;
1171        struct blk_plug_cb *cb;
1172        struct raid10_plug_cb *plug = NULL;
1173        int sectors_handled;
1174        int max_sectors;
1175        int sectors;
1176
1177        if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1178                md_flush_request(mddev, bio);
1179                return;
1180        }
1181
1182        /* If this request crosses a chunk boundary, we need to
1183         * split it.  This will only happen for 1 PAGE (or less) requests.
1184         */
1185        if (unlikely((bio->bi_sector & chunk_mask) + bio_sectors(bio)
1186                     > chunk_sects
1187                     && (conf->geo.near_copies < conf->geo.raid_disks
1188                         || conf->prev.near_copies < conf->prev.raid_disks))) {
1189                struct bio_pair *bp;
1190                /* Sanity check -- queue functions should prevent this happening */
1191                if (bio_segments(bio) > 1)
1192                        goto bad_map;
1193                /* This is a one page bio that upper layers
1194                 * refuse to split for us, so we need to split it.
1195                 */
1196                bp = bio_split(bio,
1197                               chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
1198
1199                /* Each of these 'make_request' calls will call 'wait_barrier'.
1200                 * If the first succeeds but the second blocks due to the resync
1201                 * thread raising the barrier, we will deadlock because the
1202                 * IO to the underlying device will be queued in generic_make_request
1203                 * and will never complete, so will never reduce nr_pending.
1204                 * So increment nr_waiting here so no new raise_barriers will
1205                 * succeed, and so the second wait_barrier cannot block.
1206                 */
1207                spin_lock_irq(&conf->resync_lock);
1208                conf->nr_waiting++;
1209                spin_unlock_irq(&conf->resync_lock);
1210
1211                make_request(mddev, &bp->bio1);
1212                make_request(mddev, &bp->bio2);
1213
1214                spin_lock_irq(&conf->resync_lock);
1215                conf->nr_waiting--;
1216                wake_up(&conf->wait_barrier);
1217                spin_unlock_irq(&conf->resync_lock);
1218
1219                bio_pair_release(bp);
1220                return;
1221        bad_map:
1222                printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1223                       " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
1224                       (unsigned long long)bio->bi_sector, bio_sectors(bio) / 2);
1225
1226                bio_io_error(bio);
1227                return;
1228        }
1229
1230        md_write_start(mddev, bio);
1231
1232        /*
1233         * Register the new request and wait if the reconstruction
1234         * thread has put up a bar for new requests.
1235         * Continue immediately if no resync is active currently.
1236         */
1237        wait_barrier(conf);
1238
1239        sectors = bio_sectors(bio);
1240        while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1241            bio->bi_sector < conf->reshape_progress &&
1242            bio->bi_sector + sectors > conf->reshape_progress) {
1243                /* IO spans the reshape position.  Need to wait for
1244                 * reshape to pass
1245                 */
1246                allow_barrier(conf);
1247                wait_event(conf->wait_barrier,
1248                           conf->reshape_progress <= bio->bi_sector ||
1249                           conf->reshape_progress >= bio->bi_sector + sectors);
1250                wait_barrier(conf);
1251        }
1252        if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1253            bio_data_dir(bio) == WRITE &&
1254            (mddev->reshape_backwards
1255             ? (bio->bi_sector < conf->reshape_safe &&
1256                bio->bi_sector + sectors > conf->reshape_progress)
1257             : (bio->bi_sector + sectors > conf->reshape_safe &&
1258                bio->bi_sector < conf->reshape_progress))) {
1259                /* Need to update reshape_position in metadata */
1260                mddev->reshape_position = conf->reshape_progress;
1261                set_bit(MD_CHANGE_DEVS, &mddev->flags);
1262                set_bit(MD_CHANGE_PENDING, &mddev->flags);
1263                md_wakeup_thread(mddev->thread);
1264                wait_event(mddev->sb_wait,
1265                           !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1266
1267                conf->reshape_safe = mddev->reshape_position;
1268        }
1269
1270        r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1271
1272        r10_bio->master_bio = bio;
1273        r10_bio->sectors = sectors;
1274
1275        r10_bio->mddev = mddev;
1276        r10_bio->sector = bio->bi_sector;
1277        r10_bio->state = 0;
1278
1279        /* We might need to issue multiple reads to different
1280         * devices if there are bad blocks around, so we keep
1281         * track of the number of reads in bio->bi_phys_segments.
1282         * If this is 0, there is only one r10_bio and no locking
1283         * will be needed when the request completes.  If it is
1284         * non-zero, then it is the number of not-completed requests.
1285         */
1286        bio->bi_phys_segments = 0;
1287        clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1288
1289        if (rw == READ) {
1290                /*
1291                 * read balancing logic:
1292                 */
1293                struct md_rdev *rdev;
1294                int slot;
1295
1296read_again:
1297                rdev = read_balance(conf, r10_bio, &max_sectors);
1298                if (!rdev) {
1299                        raid_end_bio_io(r10_bio);
1300                        return;
1301                }
1302                slot = r10_bio->read_slot;
1303
1304                read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1305                md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1306                            max_sectors);
1307
1308                r10_bio->devs[slot].bio = read_bio;
1309                r10_bio->devs[slot].rdev = rdev;
1310
1311                read_bio->bi_sector = r10_bio->devs[slot].addr +
1312                        choose_data_offset(r10_bio, rdev);
1313                read_bio->bi_bdev = rdev->bdev;
1314                read_bio->bi_end_io = raid10_end_read_request;
1315                read_bio->bi_rw = READ | do_sync;
1316                read_bio->bi_private = r10_bio;
1317
1318                if (max_sectors < r10_bio->sectors) {
1319                        /* Could not read all from this device, so we will
1320                         * need another r10_bio.
1321                         */
1322                        sectors_handled = (r10_bio->sectors + max_sectors
1323                                           - bio->bi_sector);
1324                        r10_bio->sectors = max_sectors;
1325                        spin_lock_irq(&conf->device_lock);
1326                        if (bio->bi_phys_segments == 0)
1327                                bio->bi_phys_segments = 2;
1328                        else
1329                                bio->bi_phys_segments++;
1330                        spin_unlock(&conf->device_lock);
1331                        /* Cannot call generic_make_request directly
1332                         * as that will be queued in __generic_make_request
1333                         * and subsequent mempool_alloc might block
1334                         * waiting for it.  so hand bio over to raid10d.
1335                         */
1336                        reschedule_retry(r10_bio);
1337
1338                        r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1339
1340                        r10_bio->master_bio = bio;
1341                        r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1342                        r10_bio->state = 0;
1343                        r10_bio->mddev = mddev;
1344                        r10_bio->sector = bio->bi_sector + sectors_handled;
1345                        goto read_again;
1346                } else
1347                        generic_make_request(read_bio);
1348                return;
1349        }
1350
1351        /*
1352         * WRITE:
1353         */
1354        if (conf->pending_count >= max_queued_requests) {
1355                md_wakeup_thread(mddev->thread);
1356                wait_event(conf->wait_barrier,
1357                           conf->pending_count < max_queued_requests);
1358        }
1359        /* first select target devices under rcu_lock and
1360         * inc refcount on their rdev.  Record them by setting
1361         * bios[x] to bio
1362         * If there are known/acknowledged bad blocks on any device
1363         * on which we have seen a write error, we want to avoid
1364         * writing to those blocks.  This potentially requires several
1365         * writes to write around the bad blocks.  Each set of writes
1366         * gets its own r10_bio with a set of bios attached.  The number
1367         * of r10_bios is recored in bio->bi_phys_segments just as with
1368         * the read case.
1369         */
1370
1371        r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1372        raid10_find_phys(conf, r10_bio);
1373retry_write:
1374        blocked_rdev = NULL;
1375        rcu_read_lock();
1376        max_sectors = r10_bio->sectors;
1377
1378        for (i = 0;  i < conf->copies; i++) {
1379                int d = r10_bio->devs[i].devnum;
1380                struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1381                struct md_rdev *rrdev = rcu_dereference(
1382                        conf->mirrors[d].replacement);
1383                if (rdev == rrdev)
1384                        rrdev = NULL;
1385                if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1386                        atomic_inc(&rdev->nr_pending);
1387                        blocked_rdev = rdev;
1388                        break;
1389                }
1390                if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1391                        atomic_inc(&rrdev->nr_pending);
1392                        blocked_rdev = rrdev;
1393                        break;
1394                }
1395                if (rdev && (test_bit(Faulty, &rdev->flags)
1396                             || test_bit(Unmerged, &rdev->flags)))
1397                        rdev = NULL;
1398                if (rrdev && (test_bit(Faulty, &rrdev->flags)
1399                              || test_bit(Unmerged, &rrdev->flags)))
1400                        rrdev = NULL;
1401
1402                r10_bio->devs[i].bio = NULL;
1403                r10_bio->devs[i].repl_bio = NULL;
1404
1405                if (!rdev && !rrdev) {
1406                        set_bit(R10BIO_Degraded, &r10_bio->state);
1407                        continue;
1408                }
1409                if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1410                        sector_t first_bad;
1411                        sector_t dev_sector = r10_bio->devs[i].addr;
1412                        int bad_sectors;
1413                        int is_bad;
1414
1415                        is_bad = is_badblock(rdev, dev_sector,
1416                                             max_sectors,
1417                                             &first_bad, &bad_sectors);
1418                        if (is_bad < 0) {
1419                                /* Mustn't write here until the bad block
1420                                 * is acknowledged
1421                                 */
1422                                atomic_inc(&rdev->nr_pending);
1423                                set_bit(BlockedBadBlocks, &rdev->flags);
1424                                blocked_rdev = rdev;
1425                                break;
1426                        }
1427                        if (is_bad && first_bad <= dev_sector) {
1428                                /* Cannot write here at all */
1429                                bad_sectors -= (dev_sector - first_bad);
1430                                if (bad_sectors < max_sectors)
1431                                        /* Mustn't write more than bad_sectors
1432                                         * to other devices yet
1433                                         */
1434                                        max_sectors = bad_sectors;
1435                                /* We don't set R10BIO_Degraded as that
1436                                 * only applies if the disk is missing,
1437                                 * so it might be re-added, and we want to
1438                                 * know to recover this chunk.
1439                                 * In this case the device is here, and the
1440                                 * fact that this chunk is not in-sync is
1441                                 * recorded in the bad block log.
1442                                 */
1443                                continue;
1444                        }
1445                        if (is_bad) {
1446                                int good_sectors = first_bad - dev_sector;
1447                                if (good_sectors < max_sectors)
1448                                        max_sectors = good_sectors;
1449                        }
1450                }
1451                if (rdev) {
1452                        r10_bio->devs[i].bio = bio;
1453                        atomic_inc(&rdev->nr_pending);
1454                }
1455                if (rrdev) {
1456                        r10_bio->devs[i].repl_bio = bio;
1457                        atomic_inc(&rrdev->nr_pending);
1458                }
1459        }
1460        rcu_read_unlock();
1461
1462        if (unlikely(blocked_rdev)) {
1463                /* Have to wait for this device to get unblocked, then retry */
1464                int j;
1465                int d;
1466
1467                for (j = 0; j < i; j++) {
1468                        if (r10_bio->devs[j].bio) {
1469                                d = r10_bio->devs[j].devnum;
1470                                rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1471                        }
1472                        if (r10_bio->devs[j].repl_bio) {
1473                                struct md_rdev *rdev;
1474                                d = r10_bio->devs[j].devnum;
1475                                rdev = conf->mirrors[d].replacement;
1476                                if (!rdev) {
1477                                        /* Race with remove_disk */
1478                                        smp_mb();
1479                                        rdev = conf->mirrors[d].rdev;
1480                                }
1481                                rdev_dec_pending(rdev, mddev);
1482                        }
1483                }
1484                allow_barrier(conf);
1485                md_wait_for_blocked_rdev(blocked_rdev, mddev);
1486                wait_barrier(conf);
1487                goto retry_write;
1488        }
1489
1490        if (max_sectors < r10_bio->sectors) {
1491                /* We are splitting this into multiple parts, so
1492                 * we need to prepare for allocating another r10_bio.
1493                 */
1494                r10_bio->sectors = max_sectors;
1495                spin_lock_irq(&conf->device_lock);
1496                if (bio->bi_phys_segments == 0)
1497                        bio->bi_phys_segments = 2;
1498                else
1499                        bio->bi_phys_segments++;
1500                spin_unlock_irq(&conf->device_lock);
1501        }
1502        sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1503
1504        atomic_set(&r10_bio->remaining, 1);
1505        bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1506
1507        for (i = 0; i < conf->copies; i++) {
1508                struct bio *mbio;
1509                int d = r10_bio->devs[i].devnum;
1510                if (r10_bio->devs[i].bio) {
1511                        struct md_rdev *rdev = conf->mirrors[d].rdev;
1512                        mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1513                        md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1514                                    max_sectors);
1515                        r10_bio->devs[i].bio = mbio;
1516
1517                        mbio->bi_sector = (r10_bio->devs[i].addr+
1518                                           choose_data_offset(r10_bio,
1519                                                              rdev));
1520                        mbio->bi_bdev = rdev->bdev;
1521                        mbio->bi_end_io = raid10_end_write_request;
1522                        mbio->bi_rw =
1523                                WRITE | do_sync | do_fua | do_discard | do_same;
1524                        mbio->bi_private = r10_bio;
1525
1526                        atomic_inc(&r10_bio->remaining);
1527
1528                        cb = blk_check_plugged(raid10_unplug, mddev,
1529                                               sizeof(*plug));
1530                        if (cb)
1531                                plug = container_of(cb, struct raid10_plug_cb,
1532                                                    cb);
1533                        else
1534                                plug = NULL;
1535                        spin_lock_irqsave(&conf->device_lock, flags);
1536                        if (plug) {
1537                                bio_list_add(&plug->pending, mbio);
1538                                plug->pending_cnt++;
1539                        } else {
1540                                bio_list_add(&conf->pending_bio_list, mbio);
1541                                conf->pending_count++;
1542                        }
1543                        spin_unlock_irqrestore(&conf->device_lock, flags);
1544                        if (!plug)
1545                                md_wakeup_thread(mddev->thread);
1546                }
1547
1548                if (r10_bio->devs[i].repl_bio) {
1549                        struct md_rdev *rdev = conf->mirrors[d].replacement;
1550                        if (rdev == NULL) {
1551                                /* Replacement just got moved to main 'rdev' */
1552                                smp_mb();
1553                                rdev = conf->mirrors[d].rdev;
1554                        }
1555                        mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1556                        md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1557                                    max_sectors);
1558                        r10_bio->devs[i].repl_bio = mbio;
1559
1560                        mbio->bi_sector = (r10_bio->devs[i].addr +
1561                                           choose_data_offset(
1562                                                   r10_bio, rdev));
1563                        mbio->bi_bdev = rdev->bdev;
1564                        mbio->bi_end_io = raid10_end_write_request;
1565                        mbio->bi_rw =
1566                                WRITE | do_sync | do_fua | do_discard | do_same;
1567                        mbio->bi_private = r10_bio;
1568
1569                        atomic_inc(&r10_bio->remaining);
1570                        spin_lock_irqsave(&conf->device_lock, flags);
1571                        bio_list_add(&conf->pending_bio_list, mbio);
1572                        conf->pending_count++;
1573                        spin_unlock_irqrestore(&conf->device_lock, flags);
1574                        if (!mddev_check_plugged(mddev))
1575                                md_wakeup_thread(mddev->thread);
1576                }
1577        }
1578
1579        /* Don't remove the bias on 'remaining' (one_write_done) until
1580         * after checking if we need to go around again.
1581         */
1582
1583        if (sectors_handled < bio_sectors(bio)) {
1584                one_write_done(r10_bio);
1585                /* We need another r10_bio.  It has already been counted
1586                 * in bio->bi_phys_segments.
1587                 */
1588                r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1589
1590                r10_bio->master_bio = bio;
1591                r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1592
1593                r10_bio->mddev = mddev;
1594                r10_bio->sector = bio->bi_sector + sectors_handled;
1595                r10_bio->state = 0;
1596                goto retry_write;
1597        }
1598        one_write_done(r10_bio);
1599
1600        /* In case raid10d snuck in to freeze_array */
1601        wake_up(&conf->wait_barrier);
1602}
1603
1604static void status(struct seq_file *seq, struct mddev *mddev)
1605{
1606        struct r10conf *conf = mddev->private;
1607        int i;
1608
1609        if (conf->geo.near_copies < conf->geo.raid_disks)
1610                seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1611        if (conf->geo.near_copies > 1)
1612                seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1613        if (conf->geo.far_copies > 1) {
1614                if (conf->geo.far_offset)
1615                        seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1616                else
1617                        seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1618        }
1619        seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1620                                        conf->geo.raid_disks - mddev->degraded);
1621        for (i = 0; i < conf->geo.raid_disks; i++)
1622                seq_printf(seq, "%s",
1623                              conf->mirrors[i].rdev &&
1624                              test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1625        seq_printf(seq, "]");
1626}
1627
1628/* check if there are enough drives for
1629 * every block to appear on atleast one.
1630 * Don't consider the device numbered 'ignore'
1631 * as we might be about to remove it.
1632 */
1633static int _enough(struct r10conf *conf, int previous, int ignore)
1634{
1635        int first = 0;
1636        int has_enough = 0;
1637        int disks, ncopies;
1638        if (previous) {
1639                disks = conf->prev.raid_disks;
1640                ncopies = conf->prev.near_copies;
1641        } else {
1642                disks = conf->geo.raid_disks;
1643                ncopies = conf->geo.near_copies;
1644        }
1645
1646        rcu_read_lock();
1647        do {
1648                int n = conf->copies;
1649                int cnt = 0;
1650                int this = first;
1651                while (n--) {
1652                        struct md_rdev *rdev;
1653                        if (this != ignore &&
1654                            (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1655                            test_bit(In_sync, &rdev->flags))
1656                                cnt++;
1657                        this = (this+1) % disks;
1658                }
1659                if (cnt == 0)
1660                        goto out;
1661                first = (first + ncopies) % disks;
1662        } while (first != 0);
1663        has_enough = 1;
1664out:
1665        rcu_read_unlock();
1666        return has_enough;
1667}
1668
1669static int enough(struct r10conf *conf, int ignore)
1670{
1671        /* when calling 'enough', both 'prev' and 'geo' must
1672         * be stable.
1673         * This is ensured if ->reconfig_mutex or ->device_lock
1674         * is held.
1675         */
1676        return _enough(conf, 0, ignore) &&
1677                _enough(conf, 1, ignore);
1678}
1679
1680static void error(struct mddev *mddev, struct md_rdev *rdev)
1681{
1682        char b[BDEVNAME_SIZE];
1683        struct r10conf *conf = mddev->private;
1684        unsigned long flags;
1685
1686        /*
1687         * If it is not operational, then we have already marked it as dead
1688         * else if it is the last working disks, ignore the error, let the
1689         * next level up know.
1690         * else mark the drive as failed
1691         */
1692        spin_lock_irqsave(&conf->device_lock, flags);
1693        if (test_bit(In_sync, &rdev->flags)
1694            && !enough(conf, rdev->raid_disk)) {
1695                /*
1696                 * Don't fail the drive, just return an IO error.
1697                 */
1698                spin_unlock_irqrestore(&conf->device_lock, flags);
1699                return;
1700        }
1701        if (test_and_clear_bit(In_sync, &rdev->flags)) {
1702                mddev->degraded++;
1703                        /*
1704                 * if recovery is running, make sure it aborts.
1705                 */
1706                set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1707        }
1708        set_bit(Blocked, &rdev->flags);
1709        set_bit(Faulty, &rdev->flags);
1710        set_bit(MD_CHANGE_DEVS, &mddev->flags);
1711        spin_unlock_irqrestore(&conf->device_lock, flags);
1712        printk(KERN_ALERT
1713               "md/raid10:%s: Disk failure on %s, disabling device.\n"
1714               "md/raid10:%s: Operation continuing on %d devices.\n",
1715               mdname(mddev), bdevname(rdev->bdev, b),
1716               mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1717}
1718
1719static void print_conf(struct r10conf *conf)
1720{
1721        int i;
1722        struct raid10_info *tmp;
1723
1724        printk(KERN_DEBUG "RAID10 conf printout:\n");
1725        if (!conf) {
1726                printk(KERN_DEBUG "(!conf)\n");
1727                return;
1728        }
1729        printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1730                conf->geo.raid_disks);
1731
1732        for (i = 0; i < conf->geo.raid_disks; i++) {
1733                char b[BDEVNAME_SIZE];
1734                tmp = conf->mirrors + i;
1735                if (tmp->rdev)
1736                        printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1737                                i, !test_bit(In_sync, &tmp->rdev->flags),
1738                                !test_bit(Faulty, &tmp->rdev->flags),
1739                                bdevname(tmp->rdev->bdev,b));
1740        }
1741}
1742
1743static void close_sync(struct r10conf *conf)
1744{
1745        wait_barrier(conf);
1746        allow_barrier(conf);
1747
1748        mempool_destroy(conf->r10buf_pool);
1749        conf->r10buf_pool = NULL;
1750}
1751
1752static int raid10_spare_active(struct mddev *mddev)
1753{
1754        int i;
1755        struct r10conf *conf = mddev->private;
1756        struct raid10_info *tmp;
1757        int count = 0;
1758        unsigned long flags;
1759
1760        /*
1761         * Find all non-in_sync disks within the RAID10 configuration
1762         * and mark them in_sync
1763         */
1764        for (i = 0; i < conf->geo.raid_disks; i++) {
1765                tmp = conf->mirrors + i;
1766                if (tmp->replacement
1767                    && tmp->replacement->recovery_offset == MaxSector
1768                    && !test_bit(Faulty, &tmp->replacement->flags)
1769                    && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1770                        /* Replacement has just become active */
1771                        if (!tmp->rdev
1772                            || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1773                                count++;
1774                        if (tmp->rdev) {
1775                                /* Replaced device not technically faulty,
1776                                 * but we need to be sure it gets removed
1777                                 * and never re-added.
1778                                 */
1779                                set_bit(Faulty, &tmp->rdev->flags);
1780                                sysfs_notify_dirent_safe(
1781                                        tmp->rdev->sysfs_state);
1782                        }
1783                        sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1784                } else if (tmp->rdev
1785                           && tmp->rdev->recovery_offset == MaxSector
1786                           && !test_bit(Faulty, &tmp->rdev->flags)
1787                           && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1788                        count++;
1789                        sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1790                }
1791        }
1792        spin_lock_irqsave(&conf->device_lock, flags);
1793        mddev->degraded -= count;
1794        spin_unlock_irqrestore(&conf->device_lock, flags);
1795
1796        print_conf(conf);
1797        return count;
1798}
1799
1800
1801static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1802{
1803        struct r10conf *conf = mddev->private;
1804        int err = -EEXIST;
1805        int mirror;
1806        int first = 0;
1807        int last = conf->geo.raid_disks - 1;
1808        struct request_queue *q = bdev_get_queue(rdev->bdev);
1809
1810        if (mddev->recovery_cp < MaxSector)
1811                /* only hot-add to in-sync arrays, as recovery is
1812                 * very different from resync
1813                 */
1814                return -EBUSY;
1815        if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1816                return -EINVAL;
1817
1818        if (rdev->raid_disk >= 0)
1819                first = last = rdev->raid_disk;
1820
1821        if (q->merge_bvec_fn) {
1822                set_bit(Unmerged, &rdev->flags);
1823                mddev->merge_check_needed = 1;
1824        }
1825
1826        if (rdev->saved_raid_disk >= first &&
1827            conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1828                mirror = rdev->saved_raid_disk;
1829        else
1830                mirror = first;
1831        for ( ; mirror <= last ; mirror++) {
1832                struct raid10_info *p = &conf->mirrors[mirror];
1833                if (p->recovery_disabled == mddev->recovery_disabled)
1834                        continue;
1835                if (p->rdev) {
1836                        if (!test_bit(WantReplacement, &p->rdev->flags) ||
1837                            p->replacement != NULL)
1838                                continue;
1839                        clear_bit(In_sync, &rdev->flags);
1840                        set_bit(Replacement, &rdev->flags);
1841                        rdev->raid_disk = mirror;
1842                        err = 0;
1843                        if (mddev->gendisk)
1844                                disk_stack_limits(mddev->gendisk, rdev->bdev,
1845                                                  rdev->data_offset << 9);
1846                        conf->fullsync = 1;
1847                        rcu_assign_pointer(p->replacement, rdev);
1848                        break;
1849                }
1850
1851                if (mddev->gendisk)
1852                        disk_stack_limits(mddev->gendisk, rdev->bdev,
1853                                          rdev->data_offset << 9);
1854
1855                p->head_position = 0;
1856                p->recovery_disabled = mddev->recovery_disabled - 1;
1857                rdev->raid_disk = mirror;
1858                err = 0;
1859                if (rdev->saved_raid_disk != mirror)
1860                        conf->fullsync = 1;
1861                rcu_assign_pointer(p->rdev, rdev);
1862                break;
1863        }
1864        if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1865                /* Some requests might not have seen this new
1866                 * merge_bvec_fn.  We must wait for them to complete
1867                 * before merging the device fully.
1868                 * First we make sure any code which has tested
1869                 * our function has submitted the request, then
1870                 * we wait for all outstanding requests to complete.
1871                 */
1872                synchronize_sched();
1873                freeze_array(conf, 0);
1874                unfreeze_array(conf);
1875                clear_bit(Unmerged, &rdev->flags);
1876        }
1877        md_integrity_add_rdev(rdev, mddev);
1878        if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1879                queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1880
1881        print_conf(conf);
1882        return err;
1883}
1884
1885static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1886{
1887        struct r10conf *conf = mddev->private;
1888        int err = 0;
1889        int number = rdev->raid_disk;
1890        struct md_rdev **rdevp;
1891        struct raid10_info *p = conf->mirrors + number;
1892
1893        print_conf(conf);
1894        if (rdev == p->rdev)
1895                rdevp = &p->rdev;
1896        else if (rdev == p->replacement)
1897                rdevp = &p->replacement;
1898        else
1899                return 0;
1900
1901        if (test_bit(In_sync, &rdev->flags) ||
1902            atomic_read(&rdev->nr_pending)) {
1903                err = -EBUSY;
1904                goto abort;
1905        }
1906        /* Only remove faulty devices if recovery
1907         * is not possible.
1908         */
1909        if (!test_bit(Faulty, &rdev->flags) &&
1910            mddev->recovery_disabled != p->recovery_disabled &&
1911            (!p->replacement || p->replacement == rdev) &&
1912            number < conf->geo.raid_disks &&
1913            enough(conf, -1)) {
1914                err = -EBUSY;
1915                goto abort;
1916        }
1917        *rdevp = NULL;
1918        synchronize_rcu();
1919        if (atomic_read(&rdev->nr_pending)) {
1920                /* lost the race, try later */
1921                err = -EBUSY;
1922                *rdevp = rdev;
1923                goto abort;
1924        } else if (p->replacement) {
1925                /* We must have just cleared 'rdev' */
1926                p->rdev = p->replacement;
1927                clear_bit(Replacement, &p->replacement->flags);
1928                smp_mb(); /* Make sure other CPUs may see both as identical
1929                           * but will never see neither -- if they are careful.
1930                           */
1931                p->replacement = NULL;
1932                clear_bit(WantReplacement, &rdev->flags);
1933        } else
1934                /* We might have just remove the Replacement as faulty
1935                 * Clear the flag just in case
1936                 */
1937                clear_bit(WantReplacement, &rdev->flags);
1938
1939        err = md_integrity_register(mddev);
1940
1941abort:
1942
1943        print_conf(conf);
1944        return err;
1945}
1946
1947
1948static void end_sync_read(struct bio *bio, int error)
1949{
1950        struct r10bio *r10_bio = bio->bi_private;
1951        struct r10conf *conf = r10_bio->mddev->private;
1952        int d;
1953
1954        if (bio == r10_bio->master_bio) {
1955                /* this is a reshape read */
1956                d = r10_bio->read_slot; /* really the read dev */
1957        } else
1958                d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1959
1960        if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1961                set_bit(R10BIO_Uptodate, &r10_bio->state);
1962        else
1963                /* The write handler will notice the lack of
1964                 * R10BIO_Uptodate and record any errors etc
1965                 */
1966                atomic_add(r10_bio->sectors,
1967                           &conf->mirrors[d].rdev->corrected_errors);
1968
1969        /* for reconstruct, we always reschedule after a read.
1970         * for resync, only after all reads
1971         */
1972        rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1973        if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1974            atomic_dec_and_test(&r10_bio->remaining)) {
1975                /* we have read all the blocks,
1976                 * do the comparison in process context in raid10d
1977                 */
1978                reschedule_retry(r10_bio);
1979        }
1980}
1981
1982static void end_sync_request(struct r10bio *r10_bio)
1983{
1984        struct mddev *mddev = r10_bio->mddev;
1985
1986        while (atomic_dec_and_test(&r10_bio->remaining)) {
1987                if (r10_bio->master_bio == NULL) {
1988                        /* the primary of several recovery bios */
1989                        sector_t s = r10_bio->sectors;
1990                        if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1991                            test_bit(R10BIO_WriteError, &r10_bio->state))
1992                                reschedule_retry(r10_bio);
1993                        else
1994                                put_buf(r10_bio);
1995                        md_done_sync(mddev, s, 1);
1996                        break;
1997                } else {
1998                        struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1999                        if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2000                            test_bit(R10BIO_WriteError, &r10_bio->state))
2001                                reschedule_retry(r10_bio);
2002                        else
2003                                put_buf(r10_bio);
2004                        r10_bio = r10_bio2;
2005                }
2006        }
2007}
2008
2009static void end_sync_write(struct bio *bio, int error)
2010{
2011        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2012        struct r10bio *r10_bio = bio->bi_private;
2013        struct mddev *mddev = r10_bio->mddev;
2014        struct r10conf *conf = mddev->private;
2015        int d;
2016        sector_t first_bad;
2017        int bad_sectors;
2018        int slot;
2019        int repl;
2020        struct md_rdev *rdev = NULL;
2021
2022        d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2023        if (repl)
2024                rdev = conf->mirrors[d].replacement;
2025        else
2026                rdev = conf->mirrors[d].rdev;
2027
2028        if (!uptodate) {
2029                if (repl)
2030                        md_error(mddev, rdev);
2031                else {
2032                        set_bit(WriteErrorSeen, &rdev->flags);
2033                        if (!test_and_set_bit(WantReplacement, &rdev->flags))
2034                                set_bit(MD_RECOVERY_NEEDED,
2035                                        &rdev->mddev->recovery);
2036                        set_bit(R10BIO_WriteError, &r10_bio->state);
2037                }
2038        } else if (is_badblock(rdev,
2039                             r10_bio->devs[slot].addr,
2040                             r10_bio->sectors,
2041                             &first_bad, &bad_sectors))
2042                set_bit(R10BIO_MadeGood, &r10_bio->state);
2043
2044        rdev_dec_pending(rdev, mddev);
2045
2046        end_sync_request(r10_bio);
2047}
2048
2049/*
2050 * Note: sync and recover and handled very differently for raid10
2051 * This code is for resync.
2052 * For resync, we read through virtual addresses and read all blocks.
2053 * If there is any error, we schedule a write.  The lowest numbered
2054 * drive is authoritative.
2055 * However requests come for physical address, so we need to map.
2056 * For every physical address there are raid_disks/copies virtual addresses,
2057 * which is always are least one, but is not necessarly an integer.
2058 * This means that a physical address can span multiple chunks, so we may
2059 * have to submit multiple io requests for a single sync request.
2060 */
2061/*
2062 * We check if all blocks are in-sync and only write to blocks that
2063 * aren't in sync
2064 */
2065static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2066{
2067        struct r10conf *conf = mddev->private;
2068        int i, first;
2069        struct bio *tbio, *fbio;
2070        int vcnt;
2071
2072        atomic_set(&r10_bio->remaining, 1);
2073
2074        /* find the first device with a block */
2075        for (i=0; i<conf->copies; i++)
2076                if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
2077                        break;
2078
2079        if (i == conf->copies)
2080                goto done;
2081
2082        first = i;
2083        fbio = r10_bio->devs[i].bio;
2084
2085        vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2086        /* now find blocks with errors */
2087        for (i=0 ; i < conf->copies ; i++) {
2088                int  j, d;
2089
2090                tbio = r10_bio->devs[i].bio;
2091
2092                if (tbio->bi_end_io != end_sync_read)
2093                        continue;
2094                if (i == first)
2095                        continue;
2096                if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
2097                        /* We know that the bi_io_vec layout is the same for
2098                         * both 'first' and 'i', so we just compare them.
2099                         * All vec entries are PAGE_SIZE;
2100                         */
2101                        int sectors = r10_bio->sectors;
2102                        for (j = 0; j < vcnt; j++) {
2103                                int len = PAGE_SIZE;
2104                                if (sectors < (len / 512))
2105                                        len = sectors * 512;
2106                                if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2107                                           page_address(tbio->bi_io_vec[j].bv_page),
2108                                           len))
2109                                        break;
2110                                sectors -= len/512;
2111                        }
2112                        if (j == vcnt)
2113                                continue;
2114                        atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2115                        if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2116                                /* Don't fix anything. */
2117                                continue;
2118                }
2119                /* Ok, we need to write this bio, either to correct an
2120                 * inconsistency or to correct an unreadable block.
2121                 * First we need to fixup bv_offset, bv_len and
2122                 * bi_vecs, as the read request might have corrupted these
2123                 */
2124                bio_reset(tbio);
2125
2126                tbio->bi_vcnt = vcnt;
2127                tbio->bi_size = r10_bio->sectors << 9;
2128                tbio->bi_rw = WRITE;
2129                tbio->bi_private = r10_bio;
2130                tbio->bi_sector = r10_bio->devs[i].addr;
2131
2132                for (j=0; j < vcnt ; j++) {
2133                        tbio->bi_io_vec[j].bv_offset = 0;
2134                        tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
2135
2136                        memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2137                               page_address(fbio->bi_io_vec[j].bv_page),
2138                               PAGE_SIZE);
2139                }
2140                tbio->bi_end_io = end_sync_write;
2141
2142                d = r10_bio->devs[i].devnum;
2143                atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2144                atomic_inc(&r10_bio->remaining);
2145                md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2146
2147                tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
2148                tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2149                generic_make_request(tbio);
2150        }
2151
2152        /* Now write out to any replacement devices
2153         * that are active
2154         */
2155        for (i = 0; i < conf->copies; i++) {
2156                int j, d;
2157
2158                tbio = r10_bio->devs[i].repl_bio;
2159                if (!tbio || !tbio->bi_end_io)
2160                        continue;
2161                if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2162                    && r10_bio->devs[i].bio != fbio)
2163                        for (j = 0; j < vcnt; j++)
2164                                memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2165                                       page_address(fbio->bi_io_vec[j].bv_page),
2166                                       PAGE_SIZE);
2167                d = r10_bio->devs[i].devnum;
2168                atomic_inc(&r10_bio->remaining);
2169                md_sync_acct(conf->mirrors[d].replacement->bdev,
2170                             bio_sectors(tbio));
2171                generic_make_request(tbio);
2172        }
2173
2174done:
2175        if (atomic_dec_and_test(&r10_bio->remaining)) {
2176                md_done_sync(mddev, r10_bio->sectors, 1);
2177                put_buf(r10_bio);
2178        }
2179}
2180
2181/*
2182 * Now for the recovery code.
2183 * Recovery happens across physical sectors.
2184 * We recover all non-is_sync drives by finding the virtual address of
2185 * each, and then choose a working drive that also has that virt address.
2186 * There is a separate r10_bio for each non-in_sync drive.
2187 * Only the first two slots are in use. The first for reading,
2188 * The second for writing.
2189 *
2190 */
2191static void fix_recovery_read_error(struct r10bio *r10_bio)
2192{
2193        /* We got a read error during recovery.
2194         * We repeat the read in smaller page-sized sections.
2195         * If a read succeeds, write it to the new device or record
2196         * a bad block if we cannot.
2197         * If a read fails, record a bad block on both old and
2198         * new devices.
2199         */
2200        struct mddev *mddev = r10_bio->mddev;
2201        struct r10conf *conf = mddev->private;
2202        struct bio *bio = r10_bio->devs[0].bio;
2203        sector_t sect = 0;
2204        int sectors = r10_bio->sectors;
2205        int idx = 0;
2206        int dr = r10_bio->devs[0].devnum;
2207        int dw = r10_bio->devs[1].devnum;
2208
2209        while (sectors) {
2210                int s = sectors;
2211                struct md_rdev *rdev;
2212                sector_t addr;
2213                int ok;
2214
2215                if (s > (PAGE_SIZE>>9))
2216                        s = PAGE_SIZE >> 9;
2217
2218                rdev = conf->mirrors[dr].rdev;
2219                addr = r10_bio->devs[0].addr + sect,
2220                ok = sync_page_io(rdev,
2221                                  addr,
2222                                  s << 9,
2223                                  bio->bi_io_vec[idx].bv_page,
2224                                  READ, false);
2225                if (ok) {
2226                        rdev = conf->mirrors[dw].rdev;
2227                        addr = r10_bio->devs[1].addr + sect;
2228                        ok = sync_page_io(rdev,
2229                                          addr,
2230                                          s << 9,
2231                                          bio->bi_io_vec[idx].bv_page,
2232                                          WRITE, false);
2233                        if (!ok) {
2234                                set_bit(WriteErrorSeen, &rdev->flags);
2235                                if (!test_and_set_bit(WantReplacement,
2236                                                      &rdev->flags))
2237                                        set_bit(MD_RECOVERY_NEEDED,
2238                                                &rdev->mddev->recovery);
2239                        }
2240                }
2241                if (!ok) {
2242                        /* We don't worry if we cannot set a bad block -
2243                         * it really is bad so there is no loss in not
2244                         * recording it yet
2245                         */
2246                        rdev_set_badblocks(rdev, addr, s, 0);
2247
2248                        if (rdev != conf->mirrors[dw].rdev) {
2249                                /* need bad block on destination too */
2250                                struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2251                                addr = r10_bio->devs[1].addr + sect;
2252                                ok = rdev_set_badblocks(rdev2, addr, s, 0);
2253                                if (!ok) {
2254                                        /* just abort the recovery */
2255                                        printk(KERN_NOTICE
2256                                               "md/raid10:%s: recovery aborted"
2257                                               " due to read error\n",
2258                                               mdname(mddev));
2259
2260                                        conf->mirrors[dw].recovery_disabled
2261                                                = mddev->recovery_disabled;
2262                                        set_bit(MD_RECOVERY_INTR,
2263                                                &mddev->recovery);
2264                                        break;
2265                                }
2266                        }
2267                }
2268
2269                sectors -= s;
2270                sect += s;
2271                idx++;
2272        }
2273}
2274
2275static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2276{
2277        struct r10conf *conf = mddev->private;
2278        int d;
2279        struct bio *wbio, *wbio2;
2280
2281        if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2282                fix_recovery_read_error(r10_bio);
2283                end_sync_request(r10_bio);
2284                return;
2285        }
2286
2287        /*
2288         * share the pages with the first bio
2289         * and submit the write request
2290         */
2291        d = r10_bio->devs[1].devnum;
2292        wbio = r10_bio->devs[1].bio;
2293        wbio2 = r10_bio->devs[1].repl_bio;
2294        /* Need to test wbio2->bi_end_io before we call
2295         * generic_make_request as if the former is NULL,
2296         * the latter is free to free wbio2.
2297         */
2298        if (wbio2 && !wbio2->bi_end_io)
2299                wbio2 = NULL;
2300        if (wbio->bi_end_io) {
2301                atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2302                md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2303                generic_make_request(wbio);
2304        }
2305        if (wbio2) {
2306                atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2307                md_sync_acct(conf->mirrors[d].replacement->bdev,
2308                             bio_sectors(wbio2));
2309                generic_make_request(wbio2);
2310        }
2311}
2312
2313
2314/*
2315 * Used by fix_read_error() to decay the per rdev read_errors.
2316 * We halve the read error count for every hour that has elapsed
2317 * since the last recorded read error.
2318 *
2319 */
2320static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2321{
2322        struct timespec cur_time_mon;
2323        unsigned long hours_since_last;
2324        unsigned int read_errors = atomic_read(&rdev->read_errors);
2325
2326        ktime_get_ts(&cur_time_mon);
2327
2328        if (rdev->last_read_error.tv_sec == 0 &&
2329            rdev->last_read_error.tv_nsec == 0) {
2330                /* first time we've seen a read error */
2331                rdev->last_read_error = cur_time_mon;
2332                return;
2333        }
2334
2335        hours_since_last = (cur_time_mon.tv_sec -
2336                            rdev->last_read_error.tv_sec) / 3600;
2337
2338        rdev->last_read_error = cur_time_mon;
2339
2340        /*
2341         * if hours_since_last is > the number of bits in read_errors
2342         * just set read errors to 0. We do this to avoid
2343         * overflowing the shift of read_errors by hours_since_last.
2344         */
2345        if (hours_since_last >= 8 * sizeof(read_errors))
2346                atomic_set(&rdev->read_errors, 0);
2347        else
2348                atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2349}
2350
2351static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2352                            int sectors, struct page *page, int rw)
2353{
2354        sector_t first_bad;
2355        int bad_sectors;
2356
2357        if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2358            && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2359                return -1;
2360        if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2361                /* success */
2362                return 1;
2363        if (rw == WRITE) {
2364                set_bit(WriteErrorSeen, &rdev->flags);
2365                if (!test_and_set_bit(WantReplacement, &rdev->flags))
2366                        set_bit(MD_RECOVERY_NEEDED,
2367                                &rdev->mddev->recovery);
2368        }
2369        /* need to record an error - either for the block or the device */
2370        if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2371                md_error(rdev->mddev, rdev);
2372        return 0;
2373}
2374
2375/*
2376 * This is a kernel thread which:
2377 *
2378 *      1.      Retries failed read operations on working mirrors.
2379 *      2.      Updates the raid superblock when problems encounter.
2380 *      3.      Performs writes following reads for array synchronising.
2381 */
2382
2383static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2384{
2385        int sect = 0; /* Offset from r10_bio->sector */
2386        int sectors = r10_bio->sectors;
2387        struct md_rdev*rdev;
2388        int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2389        int d = r10_bio->devs[r10_bio->read_slot].devnum;
2390
2391        /* still own a reference to this rdev, so it cannot
2392         * have been cleared recently.
2393         */
2394        rdev = conf->mirrors[d].rdev;
2395
2396        if (test_bit(Faulty, &rdev->flags))
2397                /* drive has already been failed, just ignore any
2398                   more fix_read_error() attempts */
2399                return;
2400
2401        check_decay_read_errors(mddev, rdev);
2402        atomic_inc(&rdev->read_errors);
2403        if (atomic_read(&rdev->read_errors) > max_read_errors) {
2404                char b[BDEVNAME_SIZE];
2405                bdevname(rdev->bdev, b);
2406
2407                printk(KERN_NOTICE
2408                       "md/raid10:%s: %s: Raid device exceeded "
2409                       "read_error threshold [cur %d:max %d]\n",
2410                       mdname(mddev), b,
2411                       atomic_read(&rdev->read_errors), max_read_errors);
2412                printk(KERN_NOTICE
2413                       "md/raid10:%s: %s: Failing raid device\n",
2414                       mdname(mddev), b);
2415                md_error(mddev, conf->mirrors[d].rdev);
2416                r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2417                return;
2418        }
2419
2420        while(sectors) {
2421                int s = sectors;
2422                int sl = r10_bio->read_slot;
2423                int success = 0;
2424                int start;
2425
2426                if (s > (PAGE_SIZE>>9))
2427                        s = PAGE_SIZE >> 9;
2428
2429                rcu_read_lock();
2430                do {
2431                        sector_t first_bad;
2432                        int bad_sectors;
2433
2434                        d = r10_bio->devs[sl].devnum;
2435                        rdev = rcu_dereference(conf->mirrors[d].rdev);
2436                        if (rdev &&
2437                            !test_bit(Unmerged, &rdev->flags) &&
2438                            test_bit(In_sync, &rdev->flags) &&
2439                            is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2440                                        &first_bad, &bad_sectors) == 0) {
2441                                atomic_inc(&rdev->nr_pending);
2442                                rcu_read_unlock();
2443                                success = sync_page_io(rdev,
2444                                                       r10_bio->devs[sl].addr +
2445                                                       sect,
2446                                                       s<<9,
2447                                                       conf->tmppage, READ, false);
2448                                rdev_dec_pending(rdev, mddev);
2449                                rcu_read_lock();
2450                                if (success)
2451                                        break;
2452                        }
2453                        sl++;
2454                        if (sl == conf->copies)
2455                                sl = 0;
2456                } while (!success && sl != r10_bio->read_slot);
2457                rcu_read_unlock();
2458
2459                if (!success) {
2460                        /* Cannot read from anywhere, just mark the block
2461                         * as bad on the first device to discourage future
2462                         * reads.
2463                         */
2464                        int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2465                        rdev = conf->mirrors[dn].rdev;
2466
2467                        if (!rdev_set_badblocks(
2468                                    rdev,
2469                                    r10_bio->devs[r10_bio->read_slot].addr
2470                                    + sect,
2471                                    s, 0)) {
2472                                md_error(mddev, rdev);
2473                                r10_bio->devs[r10_bio->read_slot].bio
2474                                        = IO_BLOCKED;
2475                        }
2476                        break;
2477                }
2478
2479                start = sl;
2480                /* write it back and re-read */
2481                rcu_read_lock();
2482                while (sl != r10_bio->read_slot) {
2483                        char b[BDEVNAME_SIZE];
2484
2485                        if (sl==0)
2486                                sl = conf->copies;
2487                        sl--;
2488                        d = r10_bio->devs[sl].devnum;
2489                        rdev = rcu_dereference(conf->mirrors[d].rdev);
2490                        if (!rdev ||
2491                            test_bit(Unmerged, &rdev->flags) ||
2492                            !test_bit(In_sync, &rdev->flags))
2493                                continue;
2494
2495                        atomic_inc(&rdev->nr_pending);
2496                        rcu_read_unlock();
2497                        if (r10_sync_page_io(rdev,
2498                                             r10_bio->devs[sl].addr +
2499                                             sect,
2500                                             s, conf->tmppage, WRITE)
2501                            == 0) {
2502                                /* Well, this device is dead */
2503                                printk(KERN_NOTICE
2504                                       "md/raid10:%s: read correction "
2505                                       "write failed"
2506                                       " (%d sectors at %llu on %s)\n",
2507                                       mdname(mddev), s,
2508                                       (unsigned long long)(
2509                                               sect +
2510                                               choose_data_offset(r10_bio,
2511                                                                  rdev)),
2512                                       bdevname(rdev->bdev, b));
2513                                printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2514                                       "drive\n",
2515                                       mdname(mddev),
2516                                       bdevname(rdev->bdev, b));
2517                        }
2518                        rdev_dec_pending(rdev, mddev);
2519                        rcu_read_lock();
2520                }
2521                sl = start;
2522                while (sl != r10_bio->read_slot) {
2523                        char b[BDEVNAME_SIZE];
2524
2525                        if (sl==0)
2526                                sl = conf->copies;
2527                        sl--;
2528                        d = r10_bio->devs[sl].devnum;
2529                        rdev = rcu_dereference(conf->mirrors[d].rdev);
2530                        if (!rdev ||
2531                            !test_bit(In_sync, &rdev->flags))
2532                                continue;
2533
2534                        atomic_inc(&rdev->nr_pending);
2535                        rcu_read_unlock();
2536                        switch (r10_sync_page_io(rdev,
2537                                             r10_bio->devs[sl].addr +
2538                                             sect,
2539                                             s, conf->tmppage,
2540                                                 READ)) {
2541                        case 0:
2542                                /* Well, this device is dead */
2543                                printk(KERN_NOTICE
2544                                       "md/raid10:%s: unable to read back "
2545                                       "corrected sectors"
2546                                       " (%d sectors at %llu on %s)\n",
2547                                       mdname(mddev), s,
2548                                       (unsigned long long)(
2549                                               sect +
2550                                               choose_data_offset(r10_bio, rdev)),
2551                                       bdevname(rdev->bdev, b));
2552                                printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2553                                       "drive\n",
2554                                       mdname(mddev),
2555                                       bdevname(rdev->bdev, b));
2556                                break;
2557                        case 1:
2558                                printk(KERN_INFO
2559                                       "md/raid10:%s: read error corrected"
2560                                       " (%d sectors at %llu on %s)\n",
2561                                       mdname(mddev), s,
2562                                       (unsigned long long)(
2563                                               sect +
2564                                               choose_data_offset(r10_bio, rdev)),
2565                                       bdevname(rdev->bdev, b));
2566                                atomic_add(s, &rdev->corrected_errors);
2567                        }
2568
2569                        rdev_dec_pending(rdev, mddev);
2570                        rcu_read_lock();
2571                }
2572                rcu_read_unlock();
2573
2574                sectors -= s;
2575                sect += s;
2576        }
2577}
2578
2579static int narrow_write_error(struct r10bio *r10_bio, int i)
2580{
2581        struct bio *bio = r10_bio->master_bio;
2582        struct mddev *mddev = r10_bio->mddev;
2583        struct r10conf *conf = mddev->private;
2584        struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2585        /* bio has the data to be written to slot 'i' where
2586         * we just recently had a write error.
2587         * We repeatedly clone the bio and trim down to one block,
2588         * then try the write.  Where the write fails we record
2589         * a bad block.
2590         * It is conceivable that the bio doesn't exactly align with
2591         * blocks.  We must handle this.
2592         *
2593         * We currently own a reference to the rdev.
2594         */
2595
2596        int block_sectors;
2597        sector_t sector;
2598        int sectors;
2599        int sect_to_write = r10_bio->sectors;
2600        int ok = 1;
2601
2602        if (rdev->badblocks.shift < 0)
2603                return 0;
2604
2605        block_sectors = 1 << rdev->badblocks.shift;
2606        sector = r10_bio->sector;
2607        sectors = ((r10_bio->sector + block_sectors)
2608                   & ~(sector_t)(block_sectors - 1))
2609                - sector;
2610
2611        while (sect_to_write) {
2612                struct bio *wbio;
2613                if (sectors > sect_to_write)
2614                        sectors = sect_to_write;
2615                /* Write at 'sector' for 'sectors' */
2616                wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2617                md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2618                wbio->bi_sector = (r10_bio->devs[i].addr+
2619                                   choose_data_offset(r10_bio, rdev) +
2620                                   (sector - r10_bio->sector));
2621                wbio->bi_bdev = rdev->bdev;
2622                if (submit_bio_wait(WRITE, wbio) == 0)
2623                        /* Failure! */
2624                        ok = rdev_set_badblocks(rdev, sector,
2625                                                sectors, 0)
2626                                && ok;
2627
2628                bio_put(wbio);
2629                sect_to_write -= sectors;
2630                sector += sectors;
2631                sectors = block_sectors;
2632        }
2633        return ok;
2634}
2635
2636static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2637{
2638        int slot = r10_bio->read_slot;
2639        struct bio *bio;
2640        struct r10conf *conf = mddev->private;
2641        struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2642        char b[BDEVNAME_SIZE];
2643        unsigned long do_sync;
2644        int max_sectors;
2645
2646        /* we got a read error. Maybe the drive is bad.  Maybe just
2647         * the block and we can fix it.
2648         * We freeze all other IO, and try reading the block from
2649         * other devices.  When we find one, we re-write
2650         * and check it that fixes the read error.
2651         * This is all done synchronously while the array is
2652         * frozen.
2653         */
2654        bio = r10_bio->devs[slot].bio;
2655        bdevname(bio->bi_bdev, b);
2656        bio_put(bio);
2657        r10_bio->devs[slot].bio = NULL;
2658
2659        if (mddev->ro == 0) {
2660                freeze_array(conf, 1);
2661                fix_read_error(conf, mddev, r10_bio);
2662                unfreeze_array(conf);
2663        } else
2664                r10_bio->devs[slot].bio = IO_BLOCKED;
2665
2666        rdev_dec_pending(rdev, mddev);
2667
2668read_more:
2669        rdev = read_balance(conf, r10_bio, &max_sectors);
2670        if (rdev == NULL) {
2671                printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2672                       " read error for block %llu\n",
2673                       mdname(mddev), b,
2674                       (unsigned long long)r10_bio->sector);
2675                raid_end_bio_io(r10_bio);
2676                return;
2677        }
2678
2679        do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2680        slot = r10_bio->read_slot;
2681        printk_ratelimited(
2682                KERN_ERR
2683                "md/raid10:%s: %s: redirecting "
2684                "sector %llu to another mirror\n",
2685                mdname(mddev),
2686                bdevname(rdev->bdev, b),
2687                (unsigned long long)r10_bio->sector);
2688        bio = bio_clone_mddev(r10_bio->master_bio,
2689                              GFP_NOIO, mddev);
2690        md_trim_bio(bio,
2691                    r10_bio->sector - bio->bi_sector,
2692                    max_sectors);
2693        r10_bio->devs[slot].bio = bio;
2694        r10_bio->devs[slot].rdev = rdev;
2695        bio->bi_sector = r10_bio->devs[slot].addr
2696                + choose_data_offset(r10_bio, rdev);
2697        bio->bi_bdev = rdev->bdev;
2698        bio->bi_rw = READ | do_sync;
2699        bio->bi_private = r10_bio;
2700        bio->bi_end_io = raid10_end_read_request;
2701        if (max_sectors < r10_bio->sectors) {
2702                /* Drat - have to split this up more */
2703                struct bio *mbio = r10_bio->master_bio;
2704                int sectors_handled =
2705                        r10_bio->sector + max_sectors
2706                        - mbio->bi_sector;
2707                r10_bio->sectors = max_sectors;
2708                spin_lock_irq(&conf->device_lock);
2709                if (mbio->bi_phys_segments == 0)
2710                        mbio->bi_phys_segments = 2;
2711                else
2712                        mbio->bi_phys_segments++;
2713                spin_unlock_irq(&conf->device_lock);
2714                generic_make_request(bio);
2715
2716                r10_bio = mempool_alloc(conf->r10bio_pool,
2717                                        GFP_NOIO);
2718                r10_bio->master_bio = mbio;
2719                r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2720                r10_bio->state = 0;
2721                set_bit(R10BIO_ReadError,
2722                        &r10_bio->state);
2723                r10_bio->mddev = mddev;
2724                r10_bio->sector = mbio->bi_sector
2725                        + sectors_handled;
2726
2727                goto read_more;
2728        } else
2729                generic_make_request(bio);
2730}
2731
2732static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2733{
2734        /* Some sort of write request has finished and it
2735         * succeeded in writing where we thought there was a
2736         * bad block.  So forget the bad block.
2737         * Or possibly if failed and we need to record
2738         * a bad block.
2739         */
2740        int m;
2741        struct md_rdev *rdev;
2742
2743        if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2744            test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2745                for (m = 0; m < conf->copies; m++) {
2746                        int dev = r10_bio->devs[m].devnum;
2747                        rdev = conf->mirrors[dev].rdev;
2748                        if (r10_bio->devs[m].bio == NULL)
2749                                continue;
2750                        if (test_bit(BIO_UPTODATE,
2751                                     &r10_bio->devs[m].bio->bi_flags)) {
2752                                rdev_clear_badblocks(
2753                                        rdev,
2754                                        r10_bio->devs[m].addr,
2755                                        r10_bio->sectors, 0);
2756                        } else {
2757                                if (!rdev_set_badblocks(
2758                                            rdev,
2759                                            r10_bio->devs[m].addr,
2760                                            r10_bio->sectors, 0))
2761                                        md_error(conf->mddev, rdev);
2762                        }
2763                        rdev = conf->mirrors[dev].replacement;
2764                        if (r10_bio->devs[m].repl_bio == NULL)
2765                                continue;
2766                        if (test_bit(BIO_UPTODATE,
2767                                     &r10_bio->devs[m].repl_bio->bi_flags)) {
2768                                rdev_clear_badblocks(
2769                                        rdev,
2770                                        r10_bio->devs[m].addr,
2771                                        r10_bio->sectors, 0);
2772                        } else {
2773                                if (!rdev_set_badblocks(
2774                                            rdev,
2775                                            r10_bio->devs[m].addr,
2776                                            r10_bio->sectors, 0))
2777                                        md_error(conf->mddev, rdev);
2778                        }
2779                }
2780                put_buf(r10_bio);
2781        } else {
2782                for (m = 0; m < conf->copies; m++) {
2783                        int dev = r10_bio->devs[m].devnum;
2784                        struct bio *bio = r10_bio->devs[m].bio;
2785                        rdev = conf->mirrors[dev].rdev;
2786                        if (bio == IO_MADE_GOOD) {
2787                                rdev_clear_badblocks(
2788                                        rdev,
2789                                        r10_bio->devs[m].addr,
2790                                        r10_bio->sectors, 0);
2791                                rdev_dec_pending(rdev, conf->mddev);
2792                        } else if (bio != NULL &&
2793                                   !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2794                                if (!narrow_write_error(r10_bio, m)) {
2795                                        md_error(conf->mddev, rdev);
2796                                        set_bit(R10BIO_Degraded,
2797                                                &r10_bio->state);
2798                                }
2799                                rdev_dec_pending(rdev, conf->mddev);
2800                        }
2801                        bio = r10_bio->devs[m].repl_bio;
2802                        rdev = conf->mirrors[dev].replacement;
2803                        if (rdev && bio == IO_MADE_GOOD) {
2804                                rdev_clear_badblocks(
2805                                        rdev,
2806                                        r10_bio->devs[m].addr,
2807                                        r10_bio->sectors, 0);
2808                                rdev_dec_pending(rdev, conf->mddev);
2809                        }
2810                }
2811                if (test_bit(R10BIO_WriteError,
2812                             &r10_bio->state))
2813                        close_write(r10_bio);
2814                raid_end_bio_io(r10_bio);
2815        }
2816}
2817
2818static void raid10d(struct md_thread *thread)
2819{
2820        struct mddev *mddev = thread->mddev;
2821        struct r10bio *r10_bio;
2822        unsigned long flags;
2823        struct r10conf *conf = mddev->private;
2824        struct list_head *head = &conf->retry_list;
2825        struct blk_plug plug;
2826
2827        md_check_recovery(mddev);
2828
2829        blk_start_plug(&plug);
2830        for (;;) {
2831
2832                flush_pending_writes(conf);
2833
2834                spin_lock_irqsave(&conf->device_lock, flags);
2835                if (list_empty(head)) {
2836                        spin_unlock_irqrestore(&conf->device_lock, flags);
2837                        break;
2838                }
2839                r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2840                list_del(head->prev);
2841                conf->nr_queued--;
2842                spin_unlock_irqrestore(&conf->device_lock, flags);
2843
2844                mddev = r10_bio->mddev;
2845                conf = mddev->private;
2846                if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2847                    test_bit(R10BIO_WriteError, &r10_bio->state))
2848                        handle_write_completed(conf, r10_bio);
2849                else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2850                        reshape_request_write(mddev, r10_bio);
2851                else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2852                        sync_request_write(mddev, r10_bio);
2853                else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2854                        recovery_request_write(mddev, r10_bio);
2855                else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2856                        handle_read_error(mddev, r10_bio);
2857                else {
2858                        /* just a partial read to be scheduled from a
2859                         * separate context
2860                         */
2861                        int slot = r10_bio->read_slot;
2862                        generic_make_request(r10_bio->devs[slot].bio);
2863                }
2864
2865                cond_resched();
2866                if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2867                        md_check_recovery(mddev);
2868        }
2869        blk_finish_plug(&plug);
2870}
2871
2872
2873static int init_resync(struct r10conf *conf)
2874{
2875        int buffs;
2876        int i;
2877
2878        buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2879        BUG_ON(conf->r10buf_pool);
2880        conf->have_replacement = 0;
2881        for (i = 0; i < conf->geo.raid_disks; i++)
2882                if (conf->mirrors[i].replacement)
2883                        conf->have_replacement = 1;
2884        conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2885        if (!conf->r10buf_pool)
2886                return -ENOMEM;
2887        conf->next_resync = 0;
2888        return 0;
2889}
2890
2891/*
2892 * perform a "sync" on one "block"
2893 *
2894 * We need to make sure that no normal I/O request - particularly write
2895 * requests - conflict with active sync requests.
2896 *
2897 * This is achieved by tracking pending requests and a 'barrier' concept
2898 * that can be installed to exclude normal IO requests.
2899 *
2900 * Resync and recovery are handled very differently.
2901 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2902 *
2903 * For resync, we iterate over virtual addresses, read all copies,
2904 * and update if there are differences.  If only one copy is live,
2905 * skip it.
2906 * For recovery, we iterate over physical addresses, read a good
2907 * value for each non-in_sync drive, and over-write.
2908 *
2909 * So, for recovery we may have several outstanding complex requests for a
2910 * given address, one for each out-of-sync device.  We model this by allocating
2911 * a number of r10_bio structures, one for each out-of-sync device.
2912 * As we setup these structures, we collect all bio's together into a list
2913 * which we then process collectively to add pages, and then process again
2914 * to pass to generic_make_request.
2915 *
2916 * The r10_bio structures are linked using a borrowed master_bio pointer.
2917 * This link is counted in ->remaining.  When the r10_bio that points to NULL
2918 * has its remaining count decremented to 0, the whole complex operation
2919 * is complete.
2920 *
2921 */
2922
2923static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2924                             int *skipped, int go_faster)
2925{
2926        struct r10conf *conf = mddev->private;
2927        struct r10bio *r10_bio;
2928        struct bio *biolist = NULL, *bio;
2929        sector_t max_sector, nr_sectors;
2930        int i;
2931        int max_sync;
2932        sector_t sync_blocks;
2933        sector_t sectors_skipped = 0;
2934        int chunks_skipped = 0;
2935        sector_t chunk_mask = conf->geo.chunk_mask;
2936
2937        if (!conf->r10buf_pool)
2938                if (init_resync(conf))
2939                        return 0;
2940
2941        /*
2942         * Allow skipping a full rebuild for incremental assembly
2943         * of a clean array, like RAID1 does.
2944         */
2945        if (mddev->bitmap == NULL &&
2946            mddev->recovery_cp == MaxSector &&
2947            mddev->reshape_position == MaxSector &&
2948            !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2949            !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2950            !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2951            conf->fullsync == 0) {
2952                *skipped = 1;
2953                return mddev->dev_sectors - sector_nr;
2954        }
2955
2956 skipped:
2957        max_sector = mddev->dev_sectors;
2958        if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2959            test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2960                max_sector = mddev->resync_max_sectors;
2961        if (sector_nr >= max_sector) {
2962                /* If we aborted, we need to abort the
2963                 * sync on the 'current' bitmap chucks (there can
2964                 * be several when recovering multiple devices).
2965                 * as we may have started syncing it but not finished.
2966                 * We can find the current address in
2967                 * mddev->curr_resync, but for recovery,
2968                 * we need to convert that to several
2969                 * virtual addresses.
2970                 */
2971                if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2972                        end_reshape(conf);
2973                        return 0;
2974                }
2975
2976                if (mddev->curr_resync < max_sector) { /* aborted */
2977                        if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2978                                bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2979                                                &sync_blocks, 1);
2980                        else for (i = 0; i < conf->geo.raid_disks; i++) {
2981                                sector_t sect =
2982                                        raid10_find_virt(conf, mddev->curr_resync, i);
2983                                bitmap_end_sync(mddev->bitmap, sect,
2984                                                &sync_blocks, 1);
2985                        }
2986                } else {
2987                        /* completed sync */
2988                        if ((!mddev->bitmap || conf->fullsync)
2989                            && conf->have_replacement
2990                            && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2991                                /* Completed a full sync so the replacements
2992                                 * are now fully recovered.
2993                                 */
2994                                for (i = 0; i < conf->geo.raid_disks; i++)
2995                                        if (conf->mirrors[i].replacement)
2996                                                conf->mirrors[i].replacement
2997                                                        ->recovery_offset
2998                                                        = MaxSector;
2999                        }
3000                        conf->fullsync = 0;
3001                }
3002                bitmap_close_sync(mddev->bitmap);
3003                close_sync(conf);
3004                *skipped = 1;
3005                return sectors_skipped;
3006        }
3007
3008        if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3009                return reshape_request(mddev, sector_nr, skipped);
3010
3011        if (chunks_skipped >= conf->geo.raid_disks) {
3012                /* if there has been nothing to do on any drive,
3013                 * then there is nothing to do at all..
3014                 */
3015                *skipped = 1;
3016                return (max_sector - sector_nr) + sectors_skipped;
3017        }
3018
3019        if (max_sector > mddev->resync_max)
3020                max_sector = mddev->resync_max; /* Don't do IO beyond here */
3021
3022        /* make sure whole request will fit in a chunk - if chunks
3023         * are meaningful
3024         */
3025        if (conf->geo.near_copies < conf->geo.raid_disks &&
3026            max_sector > (sector_nr | chunk_mask))
3027                max_sector = (sector_nr | chunk_mask) + 1;
3028        /*
3029         * If there is non-resync activity waiting for us then
3030         * put in a delay to throttle resync.
3031         */
3032        if (!go_faster && conf->nr_waiting)
3033                msleep_interruptible(1000);
3034
3035        /* Again, very different code for resync and recovery.
3036         * Both must result in an r10bio with a list of bios that
3037         * have bi_end_io, bi_sector, bi_bdev set,
3038         * and bi_private set to the r10bio.
3039         * For recovery, we may actually create several r10bios
3040         * with 2 bios in each, that correspond to the bios in the main one.
3041         * In this case, the subordinate r10bios link back through a
3042         * borrowed master_bio pointer, and the counter in the master
3043         * includes a ref from each subordinate.
3044         */
3045        /* First, we decide what to do and set ->bi_end_io
3046         * To end_sync_read if we want to read, and
3047         * end_sync_write if we will want to write.
3048         */
3049
3050        max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3051        if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3052                /* recovery... the complicated one */
3053                int j;
3054                r10_bio = NULL;
3055
3056                for (i = 0 ; i < conf->geo.raid_disks; i++) {
3057                        int still_degraded;
3058                        struct r10bio *rb2;
3059                        sector_t sect;
3060                        int must_sync;
3061                        int any_working;
3062                        struct raid10_info *mirror = &conf->mirrors[i];
3063
3064                        if ((mirror->rdev == NULL ||
3065                             test_bit(In_sync, &mirror->rdev->flags))
3066                            &&
3067                            (mirror->replacement == NULL ||
3068                             test_bit(Faulty,
3069                                      &mirror->replacement->flags)))
3070                                continue;
3071
3072                        still_degraded = 0;
3073                        /* want to reconstruct this device */
3074                        rb2 = r10_bio;
3075                        sect = raid10_find_virt(conf, sector_nr, i);
3076                        if (sect >= mddev->resync_max_sectors) {
3077                                /* last stripe is not complete - don't
3078                                 * try to recover this sector.
3079                                 */
3080                                continue;
3081                        }
3082                        /* Unless we are doing a full sync, or a replacement
3083                         * we only need to recover the block if it is set in
3084                         * the bitmap
3085                         */
3086                        must_sync = bitmap_start_sync(mddev->bitmap, sect,
3087                                                      &sync_blocks, 1);
3088                        if (sync_blocks < max_sync)
3089                                max_sync = sync_blocks;
3090                        if (!must_sync &&
3091                            mirror->replacement == NULL &&
3092                            !conf->fullsync) {
3093                                /* yep, skip the sync_blocks here, but don't assume
3094                                 * that there will never be anything to do here
3095                                 */
3096                                chunks_skipped = -1;
3097                                continue;
3098                        }
3099
3100                        r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3101                        raise_barrier(conf, rb2 != NULL);
3102                        atomic_set(&r10_bio->remaining, 0);
3103
3104                        r10_bio->master_bio = (struct bio*)rb2;
3105                        if (rb2)
3106                                atomic_inc(&rb2->remaining);
3107                        r10_bio->mddev = mddev;
3108                        set_bit(R10BIO_IsRecover, &r10_bio->state);
3109                        r10_bio->sector = sect;
3110
3111                        raid10_find_phys(conf, r10_bio);
3112
3113                        /* Need to check if the array will still be
3114                         * degraded
3115                         */
3116                        for (j = 0; j < conf->geo.raid_disks; j++)
3117                                if (conf->mirrors[j].rdev == NULL ||
3118                                    test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
3119                                        still_degraded = 1;
3120                                        break;
3121                                }
3122
3123                        must_sync = bitmap_start_sync(mddev->bitmap, sect,
3124                                                      &sync_blocks, still_degraded);
3125
3126                        any_working = 0;
3127                        for (j=0; j<conf->copies;j++) {
3128                                int k;
3129                                int d = r10_bio->devs[j].devnum;
3130                                sector_t from_addr, to_addr;
3131                                struct md_rdev *rdev;
3132                                sector_t sector, first_bad;
3133                                int bad_sectors;
3134                                if (!conf->mirrors[d].rdev ||
3135                                    !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3136                                        continue;
3137                                /* This is where we read from */
3138                                any_working = 1;
3139                                rdev = conf->mirrors[d].rdev;
3140                                sector = r10_bio->devs[j].addr;
3141
3142                                if (is_badblock(rdev, sector, max_sync,
3143                                                &first_bad, &bad_sectors)) {
3144                                        if (first_bad > sector)
3145                                                max_sync = first_bad - sector;
3146                                        else {
3147                                                bad_sectors -= (sector
3148                                                                - first_bad);
3149                                                if (max_sync > bad_sectors)
3150                                                        max_sync = bad_sectors;
3151                                                continue;
3152                                        }
3153                                }
3154                                bio = r10_bio->devs[0].bio;
3155                                bio_reset(bio);
3156                                bio->bi_next = biolist;
3157                                biolist = bio;
3158                                bio->bi_private = r10_bio;
3159                                bio->bi_end_io = end_sync_read;
3160                                bio->bi_rw = READ;
3161                                from_addr = r10_bio->devs[j].addr;
3162                                bio->bi_sector = from_addr + rdev->data_offset;
3163                                bio->bi_bdev = rdev->bdev;
3164                                atomic_inc(&rdev->nr_pending);
3165                                /* and we write to 'i' (if not in_sync) */
3166
3167                                for (k=0; k<conf->copies; k++)
3168                                        if (r10_bio->devs[k].devnum == i)
3169                                                break;
3170                                BUG_ON(k == conf->copies);
3171                                to_addr = r10_bio->devs[k].addr;
3172                                r10_bio->devs[0].devnum = d;
3173                                r10_bio->devs[0].addr = from_addr;
3174                                r10_bio->devs[1].devnum = i;
3175                                r10_bio->devs[1].addr = to_addr;
3176
3177                                rdev = mirror->rdev;
3178                                if (!test_bit(In_sync, &rdev->flags)) {
3179                                        bio = r10_bio->devs[1].bio;
3180                                        bio_reset(bio);
3181                                        bio->bi_next = biolist;
3182                                        biolist = bio;
3183                                        bio->bi_private = r10_bio;
3184                                        bio->bi_end_io = end_sync_write;
3185                                        bio->bi_rw = WRITE;
3186                                        bio->bi_sector = to_addr
3187                                                + rdev->data_offset;
3188                                        bio->bi_bdev = rdev->bdev;
3189                                        atomic_inc(&r10_bio->remaining);
3190                                } else
3191                                        r10_bio->devs[1].bio->bi_end_io = NULL;
3192
3193                                /* and maybe write to replacement */
3194                                bio = r10_bio->devs[1].repl_bio;
3195                                if (bio)
3196                                        bio->bi_end_io = NULL;
3197                                rdev = mirror->replacement;
3198                                /* Note: if rdev != NULL, then bio
3199                                 * cannot be NULL as r10buf_pool_alloc will
3200                                 * have allocated it.
3201                                 * So the second test here is pointless.
3202                                 * But it keeps semantic-checkers happy, and
3203                                 * this comment keeps human reviewers
3204                                 * happy.
3205                                 */
3206                                if (rdev == NULL || bio == NULL ||
3207                                    test_bit(Faulty, &rdev->flags))
3208                                        break;
3209                                bio_reset(bio);
3210                                bio->bi_next = biolist;
3211                                biolist = bio;
3212                                bio->bi_private = r10_bio;
3213                                bio->bi_end_io = end_sync_write;
3214                                bio->bi_rw = WRITE;
3215                                bio->bi_sector = to_addr + rdev->data_offset;
3216                                bio->bi_bdev = rdev->bdev;
3217                                atomic_inc(&r10_bio->remaining);
3218                                break;
3219                        }
3220                        if (j == conf->copies) {
3221                                /* Cannot recover, so abort the recovery or
3222                                 * record a bad block */
3223                                put_buf(r10_bio);
3224                                if (rb2)
3225                                        atomic_dec(&rb2->remaining);
3226                                r10_bio = rb2;
3227                                if (any_working) {
3228                                        /* problem is that there are bad blocks
3229                                         * on other device(s)
3230                                         */
3231                                        int k;
3232                                        for (k = 0; k < conf->copies; k++)
3233                                                if (r10_bio->devs[k].devnum == i)
3234                                                        break;
3235                                        if (!test_bit(In_sync,
3236                                                      &mirror->rdev->flags)
3237                                            && !rdev_set_badblocks(
3238                                                    mirror->rdev,
3239                                                    r10_bio->devs[k].addr,
3240                                                    max_sync, 0))
3241                                                any_working = 0;
3242                                        if (mirror->replacement &&
3243                                            !rdev_set_badblocks(
3244                                                    mirror->replacement,
3245                                                    r10_bio->devs[k].addr,
3246                                                    max_sync, 0))
3247                                                any_working = 0;
3248                                }
3249                                if (!any_working)  {
3250                                        if (!test_and_set_bit(MD_RECOVERY_INTR,
3251                                                              &mddev->recovery))
3252                                                printk(KERN_INFO "md/raid10:%s: insufficient "
3253                                                       "working devices for recovery.\n",
3254                                                       mdname(mddev));
3255                                        mirror->recovery_disabled
3256                                                = mddev->recovery_disabled;
3257                                }
3258                                break;
3259                        }
3260                }
3261                if (biolist == NULL) {
3262                        while (r10_bio) {
3263                                struct r10bio *rb2 = r10_bio;
3264                                r10_bio = (struct r10bio*) rb2->master_bio;
3265                                rb2->master_bio = NULL;
3266                                put_buf(rb2);
3267                        }
3268                        goto giveup;
3269                }
3270        } else {
3271                /* resync. Schedule a read for every block at this virt offset */
3272                int count = 0;
3273
3274                bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3275
3276                if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3277                                       &sync_blocks, mddev->degraded) &&
3278                    !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3279                                                 &mddev->recovery)) {
3280                        /* We can skip this block */
3281                        *skipped = 1;
3282                        return sync_blocks + sectors_skipped;
3283                }
3284                if (sync_blocks < max_sync)
3285                        max_sync = sync_blocks;
3286                r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3287
3288                r10_bio->mddev = mddev;
3289                atomic_set(&r10_bio->remaining, 0);
3290                raise_barrier(conf, 0);
3291                conf->next_resync = sector_nr;
3292
3293                r10_bio->master_bio = NULL;
3294                r10_bio->sector = sector_nr;
3295                set_bit(R10BIO_IsSync, &r10_bio->state);
3296                raid10_find_phys(conf, r10_bio);
3297                r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3298
3299                for (i = 0; i < conf->copies; i++) {
3300                        int d = r10_bio->devs[i].devnum;
3301                        sector_t first_bad, sector;
3302                        int bad_sectors;
3303
3304                        if (r10_bio->devs[i].repl_bio)
3305                                r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3306
3307                        bio = r10_bio->devs[i].bio;
3308                        bio_reset(bio);
3309                        clear_bit(BIO_UPTODATE, &bio->bi_flags);
3310                        if (conf->mirrors[d].rdev == NULL ||
3311                            test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3312                                continue;
3313                        sector = r10_bio->devs[i].addr;
3314                        if (is_badblock(conf->mirrors[d].rdev,
3315                                        sector, max_sync,
3316                                        &first_bad, &bad_sectors)) {
3317                                if (first_bad > sector)
3318                                        max_sync = first_bad - sector;
3319                                else {
3320                                        bad_sectors -= (sector - first_bad);
3321                                        if (max_sync > bad_sectors)
3322                                                max_sync = bad_sectors;
3323                                        continue;
3324                                }
3325                        }
3326                        atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3327                        atomic_inc(&r10_bio->remaining);
3328                        bio->bi_next = biolist;
3329                        biolist = bio;
3330                        bio->bi_private = r10_bio;
3331                        bio->bi_end_io = end_sync_read;
3332                        bio->bi_rw = READ;
3333                        bio->bi_sector = sector +
3334                                conf->mirrors[d].rdev->data_offset;
3335                        bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3336                        count++;
3337
3338                        if (conf->mirrors[d].replacement == NULL ||
3339                            test_bit(Faulty,
3340                                     &conf->mirrors[d].replacement->flags))
3341                                continue;
3342
3343                        /* Need to set up for writing to the replacement */
3344                        bio = r10_bio->devs[i].repl_bio;
3345                        bio_reset(bio);
3346                        clear_bit(BIO_UPTODATE, &bio->bi_flags);
3347
3348                        sector = r10_bio->devs[i].addr;
3349                        atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3350                        bio->bi_next = biolist;
3351                        biolist = bio;
3352                        bio->bi_private = r10_bio;
3353                        bio->bi_end_io = end_sync_write;
3354                        bio->bi_rw = WRITE;
3355                        bio->bi_sector = sector +
3356                                conf->mirrors[d].replacement->data_offset;
3357                        bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3358                        count++;
3359                }
3360
3361                if (count < 2) {
3362                        for (i=0; i<conf->copies; i++) {
3363                                int d = r10_bio->devs[i].devnum;
3364                                if (r10_bio->devs[i].bio->bi_end_io)
3365                                        rdev_dec_pending(conf->mirrors[d].rdev,
3366                                                         mddev);
3367                                if (r10_bio->devs[i].repl_bio &&
3368                                    r10_bio->devs[i].repl_bio->bi_end_io)
3369                                        rdev_dec_pending(
3370                                                conf->mirrors[d].replacement,
3371                                                mddev);
3372                        }
3373                        put_buf(r10_bio);
3374                        biolist = NULL;
3375                        goto giveup;
3376                }
3377        }
3378
3379        nr_sectors = 0;
3380        if (sector_nr + max_sync < max_sector)
3381                max_sector = sector_nr + max_sync;
3382        do {
3383                struct page *page;
3384                int len = PAGE_SIZE;
3385                if (sector_nr + (len>>9) > max_sector)
3386                        len = (max_sector - sector_nr) << 9;
3387                if (len == 0)
3388                        break;
3389                for (bio= biolist ; bio ; bio=bio->bi_next) {
3390                        struct bio *bio2;
3391                        page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3392                        if (bio_add_page(bio, page, len, 0))
3393                                continue;
3394
3395                        /* stop here */
3396                        bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3397                        for (bio2 = biolist;
3398                             bio2 && bio2 != bio;
3399                             bio2 = bio2->bi_next) {
3400                                /* remove last page from this bio */
3401                                bio2->bi_vcnt--;
3402                                bio2->bi_size -= len;
3403                                bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3404                        }
3405                        goto bio_full;
3406                }
3407                nr_sectors += len>>9;
3408                sector_nr += len>>9;
3409        } while (biolist->bi_vcnt < RESYNC_PAGES);
3410 bio_full:
3411        r10_bio->sectors = nr_sectors;
3412
3413        while (biolist) {
3414                bio = biolist;
3415                biolist = biolist->bi_next;
3416
3417                bio->bi_next = NULL;
3418                r10_bio = bio->bi_private;
3419                r10_bio->sectors = nr_sectors;
3420
3421                if (bio->bi_end_io == end_sync_read) {
3422                        md_sync_acct(bio->bi_bdev, nr_sectors);
3423                        set_bit(BIO_UPTODATE, &bio->bi_flags);
3424                        generic_make_request(bio);
3425                }
3426        }
3427
3428        if (sectors_skipped)
3429                /* pretend they weren't skipped, it makes
3430                 * no important difference in this case
3431                 */
3432                md_done_sync(mddev, sectors_skipped, 1);
3433
3434        return sectors_skipped + nr_sectors;
3435 giveup:
3436        /* There is nowhere to write, so all non-sync
3437         * drives must be failed or in resync, all drives
3438         * have a bad block, so try the next chunk...
3439         */
3440        if (sector_nr + max_sync < max_sector)
3441                max_sector = sector_nr + max_sync;
3442
3443        sectors_skipped += (max_sector - sector_nr);
3444        chunks_skipped ++;
3445        sector_nr = max_sector;
3446        goto skipped;
3447}
3448
3449static sector_t
3450raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3451{
3452        sector_t size;
3453        struct r10conf *conf = mddev->private;
3454
3455        if (!raid_disks)
3456                raid_disks = min(conf->geo.raid_disks,
3457                                 conf->prev.raid_disks);
3458        if (!sectors)
3459                sectors = conf->dev_sectors;
3460
3461        size = sectors >> conf->geo.chunk_shift;
3462        sector_div(size, conf->geo.far_copies);
3463        size = size * raid_disks;
3464        sector_div(size, conf->geo.near_copies);
3465
3466        return size << conf->geo.chunk_shift;
3467}
3468
3469static void calc_sectors(struct r10conf *conf, sector_t size)
3470{
3471        /* Calculate the number of sectors-per-device that will
3472         * actually be used, and set conf->dev_sectors and
3473         * conf->stride
3474         */
3475
3476        size = size >> conf->geo.chunk_shift;
3477        sector_div(size, conf->geo.far_copies);
3478        size = size * conf->geo.raid_disks;
3479        sector_div(size, conf->geo.near_copies);
3480        /* 'size' is now the number of chunks in the array */
3481        /* calculate "used chunks per device" */
3482        size = size * conf->copies;
3483
3484        /* We need to round up when dividing by raid_disks to
3485         * get the stride size.
3486         */
3487        size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3488
3489        conf->dev_sectors = size << conf->geo.chunk_shift;
3490
3491        if (conf->geo.far_offset)
3492                conf->geo.stride = 1 << conf->geo.chunk_shift;
3493        else {
3494                sector_div(size, conf->geo.far_copies);
3495                conf->geo.stride = size << conf->geo.chunk_shift;
3496        }
3497}
3498
3499enum geo_type {geo_new, geo_old, geo_start};
3500static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3501{
3502        int nc, fc, fo;
3503        int layout, chunk, disks;
3504        switch (new) {
3505        case geo_old:
3506                layout = mddev->layout;
3507                chunk = mddev->chunk_sectors;
3508                disks = mddev->raid_disks - mddev->delta_disks;
3509                break;
3510        case geo_new:
3511                layout = mddev->new_layout;
3512                chunk = mddev->new_chunk_sectors;
3513                disks = mddev->raid_disks;
3514                break;
3515        default: /* avoid 'may be unused' warnings */
3516        case geo_start: /* new when starting reshape - raid_disks not
3517                         * updated yet. */
3518                layout = mddev->new_layout;
3519                chunk = mddev->new_chunk_sectors;
3520                disks = mddev->raid_disks + mddev->delta_disks;
3521                break;
3522        }
3523        if (layout >> 18)
3524                return -1;
3525        if (chunk < (PAGE_SIZE >> 9) ||
3526            !is_power_of_2(chunk))
3527                return -2;
3528        nc = layout & 255;
3529        fc = (layout >> 8) & 255;
3530        fo = layout & (1<<16);
3531        geo->raid_disks = disks;
3532        geo->near_copies = nc;
3533        geo->far_copies = fc;
3534        geo->far_offset = fo;
3535        geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks;
3536        geo->chunk_mask = chunk - 1;
3537        geo->chunk_shift = ffz(~chunk);
3538        return nc*fc;
3539}
3540
3541static struct r10conf *setup_conf(struct mddev *mddev)
3542{
3543        struct r10conf *conf = NULL;
3544        int err = -EINVAL;
3545        struct geom geo;
3546        int copies;
3547
3548        copies = setup_geo(&geo, mddev, geo_new);
3549
3550        if (copies == -2) {
3551                printk(KERN_ERR "md/raid10:%s: chunk size must be "
3552                       "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3553                       mdname(mddev), PAGE_SIZE);
3554                goto out;
3555        }
3556
3557        if (copies < 2 || copies > mddev->raid_disks) {
3558                printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3559                       mdname(mddev), mddev->new_layout);
3560                goto out;
3561        }
3562
3563        err = -ENOMEM;
3564        conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3565        if (!conf)
3566                goto out;
3567
3568        /* FIXME calc properly */
3569        conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3570                                                            max(0,-mddev->delta_disks)),
3571                                GFP_KERNEL);
3572        if (!conf->mirrors)
3573                goto out;
3574
3575        conf->tmppage = alloc_page(GFP_KERNEL);
3576        if (!conf->tmppage)
3577                goto out;
3578
3579        conf->geo = geo;
3580        conf->copies = copies;
3581        conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3582                                           r10bio_pool_free, conf);
3583        if (!conf->r10bio_pool)
3584                goto out;
3585
3586        calc_sectors(conf, mddev->dev_sectors);
3587        if (mddev->reshape_position == MaxSector) {
3588                conf->prev = conf->geo;
3589                conf->reshape_progress = MaxSector;
3590        } else {
3591                if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3592                        err = -EINVAL;
3593                        goto out;
3594                }
3595                conf->reshape_progress = mddev->reshape_position;
3596                if (conf->prev.far_offset)
3597                        conf->prev.stride = 1 << conf->prev.chunk_shift;
3598                else
3599                        /* far_copies must be 1 */
3600                        conf->prev.stride = conf->dev_sectors;
3601        }
3602        spin_lock_init(&conf->device_lock);
3603        INIT_LIST_HEAD(&conf->retry_list);
3604
3605        spin_lock_init(&conf->resync_lock);
3606        init_waitqueue_head(&conf->wait_barrier);
3607
3608        conf->thread = md_register_thread(raid10d, mddev, "raid10");
3609        if (!conf->thread)
3610                goto out;
3611
3612        conf->mddev = mddev;
3613        return conf;
3614
3615 out:
3616        if (err == -ENOMEM)
3617                printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3618                       mdname(mddev));
3619        if (conf) {
3620                if (conf->r10bio_pool)
3621                        mempool_destroy(conf->r10bio_pool);
3622                kfree(conf->mirrors);
3623                safe_put_page(conf->tmppage);
3624                kfree(conf);
3625        }
3626        return ERR_PTR(err);
3627}
3628
3629static int run(struct mddev *mddev)
3630{
3631        struct r10conf *conf;
3632        int i, disk_idx, chunk_size;
3633        struct raid10_info *disk;
3634        struct md_rdev *rdev;
3635        sector_t size;
3636        sector_t min_offset_diff = 0;
3637        int first = 1;
3638        bool discard_supported = false;
3639
3640        if (mddev->private == NULL) {
3641                conf = setup_conf(mddev);
3642                if (IS_ERR(conf))
3643                        return PTR_ERR(conf);
3644                mddev->private = conf;
3645        }
3646        conf = mddev->private;
3647        if (!conf)
3648                goto out;
3649
3650        mddev->thread = conf->thread;
3651        conf->thread = NULL;
3652
3653        chunk_size = mddev->chunk_sectors << 9;
3654        if (mddev->queue) {
3655                blk_queue_max_discard_sectors(mddev->queue,
3656                                              mddev->chunk_sectors);
3657                blk_queue_max_write_same_sectors(mddev->queue, 0);
3658                blk_queue_io_min(mddev->queue, chunk_size);
3659                if (conf->geo.raid_disks % conf->geo.near_copies)
3660                        blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3661                else
3662                        blk_queue_io_opt(mddev->queue, chunk_size *
3663                                         (conf->geo.raid_disks / conf->geo.near_copies));
3664        }
3665
3666        rdev_for_each(rdev, mddev) {
3667                long long diff;
3668                struct request_queue *q;
3669
3670                disk_idx = rdev->raid_disk;
3671                if (disk_idx < 0)
3672                        continue;
3673                if (disk_idx >= conf->geo.raid_disks &&
3674                    disk_idx >= conf->prev.raid_disks)
3675                        continue;
3676                disk = conf->mirrors + disk_idx;
3677
3678                if (test_bit(Replacement, &rdev->flags)) {
3679                        if (disk->replacement)
3680                                goto out_free_conf;
3681                        disk->replacement = rdev;
3682                } else {
3683                        if (disk->rdev)
3684                                goto out_free_conf;
3685                        disk->rdev = rdev;
3686                }
3687                q = bdev_get_queue(rdev->bdev);
3688                if (q->merge_bvec_fn)
3689                        mddev->merge_check_needed = 1;
3690                diff = (rdev->new_data_offset - rdev->data_offset);
3691                if (!mddev->reshape_backwards)
3692                        diff = -diff;
3693                if (diff < 0)
3694                        diff = 0;
3695                if (first || diff < min_offset_diff)
3696                        min_offset_diff = diff;
3697
3698                if (mddev->gendisk)
3699                        disk_stack_limits(mddev->gendisk, rdev->bdev,
3700                                          rdev->data_offset << 9);
3701
3702                disk->head_position = 0;
3703
3704                if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3705                        discard_supported = true;
3706        }
3707
3708        if (mddev->queue) {
3709                if (discard_supported)
3710                        queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3711                                                mddev->queue);
3712                else
3713                        queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3714                                                  mddev->queue);
3715        }
3716        /* need to check that every block has at least one working mirror */
3717        if (!enough(conf, -1)) {
3718                printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3719                       mdname(mddev));
3720                goto out_free_conf;
3721        }
3722
3723        if (conf->reshape_progress != MaxSector) {
3724                /* must ensure that shape change is supported */
3725                if (conf->geo.far_copies != 1 &&
3726                    conf->geo.far_offset == 0)
3727                        goto out_free_conf;
3728                if (conf->prev.far_copies != 1 &&
3729                    conf->prev.far_offset == 0)
3730                        goto out_free_conf;
3731        }
3732
3733        mddev->degraded = 0;
3734        for (i = 0;
3735             i < conf->geo.raid_disks
3736                     || i < conf->prev.raid_disks;
3737             i++) {
3738
3739                disk = conf->mirrors + i;
3740
3741                if (!disk->rdev && disk->replacement) {
3742                        /* The replacement is all we have - use it */
3743                        disk->rdev = disk->replacement;
3744                        disk->replacement = NULL;
3745                        clear_bit(Replacement, &disk->rdev->flags);
3746                }
3747
3748                if (!disk->rdev ||
3749                    !test_bit(In_sync, &disk->rdev->flags)) {
3750                        disk->head_position = 0;
3751                        mddev->degraded++;
3752                        if (disk->rdev)
3753                                conf->fullsync = 1;
3754                }
3755                disk->recovery_disabled = mddev->recovery_disabled - 1;
3756        }
3757
3758        if (mddev->recovery_cp != MaxSector)
3759                printk(KERN_NOTICE "md/raid10:%s: not clean"
3760                       " -- starting background reconstruction\n",
3761                       mdname(mddev));
3762        printk(KERN_INFO
3763                "md/raid10:%s: active with %d out of %d devices\n",
3764                mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3765                conf->geo.raid_disks);
3766        /*
3767         * Ok, everything is just fine now
3768         */
3769        mddev->dev_sectors = conf->dev_sectors;
3770        size = raid10_size(mddev, 0, 0);
3771        md_set_array_sectors(mddev, size);
3772        mddev->resync_max_sectors = size;
3773
3774        if (mddev->queue) {
3775                int stripe = conf->geo.raid_disks *
3776                        ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3777                mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3778                mddev->queue->backing_dev_info.congested_data = mddev;
3779
3780                /* Calculate max read-ahead size.
3781                 * We need to readahead at least twice a whole stripe....
3782                 * maybe...
3783                 */
3784                stripe /= conf->geo.near_copies;
3785                if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3786                        mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3787                blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3788        }
3789
3790
3791        if (md_integrity_register(mddev))
3792                goto out_free_conf;
3793
3794        if (conf->reshape_progress != MaxSector) {
3795                unsigned long before_length, after_length;
3796
3797                before_length = ((1 << conf->prev.chunk_shift) *
3798                                 conf->prev.far_copies);
3799                after_length = ((1 << conf->geo.chunk_shift) *
3800                                conf->geo.far_copies);
3801
3802                if (max(before_length, after_length) > min_offset_diff) {
3803                        /* This cannot work */
3804                        printk("md/raid10: offset difference not enough to continue reshape\n");
3805                        goto out_free_conf;
3806                }
3807                conf->offset_diff = min_offset_diff;
3808
3809                conf->reshape_safe = conf->reshape_progress;
3810                clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3811                clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3812                set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3813                set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3814                mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3815                                                        "reshape");
3816        }
3817
3818        return 0;
3819
3820out_free_conf:
3821        md_unregister_thread(&mddev->thread);
3822        if (conf->r10bio_pool)
3823                mempool_destroy(conf->r10bio_pool);
3824        safe_put_page(conf->tmppage);
3825        kfree(conf->mirrors);
3826        kfree(conf);
3827        mddev->private = NULL;
3828out:
3829        return -EIO;
3830}
3831
3832static int stop(struct mddev *mddev)
3833{
3834        struct r10conf *conf = mddev->private;
3835
3836        raise_barrier(conf, 0);
3837        lower_barrier(conf);
3838
3839        md_unregister_thread(&mddev->thread);
3840        if (mddev->queue)
3841                /* the unplug fn references 'conf'*/
3842                blk_sync_queue(mddev->queue);
3843
3844        if (conf->r10bio_pool)
3845                mempool_destroy(conf->r10bio_pool);
3846        safe_put_page(conf->tmppage);
3847        kfree(conf->mirrors);
3848        kfree(conf);
3849        mddev->private = NULL;
3850        return 0;
3851}
3852
3853static void raid10_quiesce(struct mddev *mddev, int state)
3854{
3855        struct r10conf *conf = mddev->private;
3856
3857        switch(state) {
3858        case 1:
3859                raise_barrier(conf, 0);
3860                break;
3861        case 0:
3862                lower_barrier(conf);
3863                break;
3864        }
3865}
3866
3867static int raid10_resize(struct mddev *mddev, sector_t sectors)
3868{
3869        /* Resize of 'far' arrays is not supported.
3870         * For 'near' and 'offset' arrays we can set the
3871         * number of sectors used to be an appropriate multiple
3872         * of the chunk size.
3873         * For 'offset', this is far_copies*chunksize.
3874         * For 'near' the multiplier is the LCM of
3875         * near_copies and raid_disks.
3876         * So if far_copies > 1 && !far_offset, fail.
3877         * Else find LCM(raid_disks, near_copy)*far_copies and
3878         * multiply by chunk_size.  Then round to this number.
3879         * This is mostly done by raid10_size()
3880         */
3881        struct r10conf *conf = mddev->private;
3882        sector_t oldsize, size;
3883
3884        if (mddev->reshape_position != MaxSector)
3885                return -EBUSY;
3886
3887        if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3888                return -EINVAL;
3889
3890        oldsize = raid10_size(mddev, 0, 0);
3891        size = raid10_size(mddev, sectors, 0);
3892        if (mddev->external_size &&
3893            mddev->array_sectors > size)
3894                return -EINVAL;
3895        if (mddev->bitmap) {
3896                int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3897                if (ret)
3898                        return ret;
3899        }
3900        md_set_array_sectors(mddev, size);
3901        set_capacity(mddev->gendisk, mddev->array_sectors);
3902        revalidate_disk(mddev->gendisk);
3903        if (sectors > mddev->dev_sectors &&
3904            mddev->recovery_cp > oldsize) {
3905                mddev->recovery_cp = oldsize;
3906                set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3907        }
3908        calc_sectors(conf, sectors);
3909        mddev->dev_sectors = conf->dev_sectors;
3910        mddev->resync_max_sectors = size;
3911        return 0;
3912}
3913
3914static void *raid10_takeover_raid0(struct mddev *mddev)
3915{
3916        struct md_rdev *rdev;
3917        struct r10conf *conf;
3918
3919        if (mddev->degraded > 0) {
3920                printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3921                       mdname(mddev));
3922                return ERR_PTR(-EINVAL);
3923        }
3924
3925        /* Set new parameters */
3926        mddev->new_level = 10;
3927        /* new layout: far_copies = 1, near_copies = 2 */
3928        mddev->new_layout = (1<<8) + 2;
3929        mddev->new_chunk_sectors = mddev->chunk_sectors;
3930        mddev->delta_disks = mddev->raid_disks;
3931        mddev->raid_disks *= 2;
3932        /* make sure it will be not marked as dirty */
3933        mddev->recovery_cp = MaxSector;
3934
3935        conf = setup_conf(mddev);
3936        if (!IS_ERR(conf)) {
3937                rdev_for_each(rdev, mddev)
3938                        if (rdev->raid_disk >= 0)
3939                                rdev->new_raid_disk = rdev->raid_disk * 2;
3940                conf->barrier = 1;
3941        }
3942
3943        return conf;
3944}
3945
3946static void *raid10_takeover(struct mddev *mddev)
3947{
3948        struct r0conf *raid0_conf;
3949
3950        /* raid10 can take over:
3951         *  raid0 - providing it has only two drives
3952         */
3953        if (mddev->level == 0) {
3954                /* for raid0 takeover only one zone is supported */
3955                raid0_conf = mddev->private;
3956                if (raid0_conf->nr_strip_zones > 1) {
3957                        printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3958                               " with more than one zone.\n",
3959                               mdname(mddev));
3960                        return ERR_PTR(-EINVAL);
3961                }
3962                return raid10_takeover_raid0(mddev);
3963        }
3964        return ERR_PTR(-EINVAL);
3965}
3966
3967static int raid10_check_reshape(struct mddev *mddev)
3968{
3969        /* Called when there is a request to change
3970         * - layout (to ->new_layout)
3971         * - chunk size (to ->new_chunk_sectors)
3972         * - raid_disks (by delta_disks)
3973         * or when trying to restart a reshape that was ongoing.
3974         *
3975         * We need to validate the request and possibly allocate
3976         * space if that might be an issue later.
3977         *
3978         * Currently we reject any reshape of a 'far' mode array,
3979         * allow chunk size to change if new is generally acceptable,
3980         * allow raid_disks to increase, and allow
3981         * a switch between 'near' mode and 'offset' mode.
3982         */
3983        struct r10conf *conf = mddev->private;
3984        struct geom geo;
3985
3986        if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3987                return -EINVAL;
3988
3989        if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3990                /* mustn't change number of copies */
3991                return -EINVAL;
3992        if (geo.far_copies > 1 && !geo.far_offset)
3993                /* Cannot switch to 'far' mode */
3994                return -EINVAL;
3995
3996        if (mddev->array_sectors & geo.chunk_mask)
3997                        /* not factor of array size */
3998                        return -EINVAL;
3999
4000        if (!enough(conf, -1))
4001                return -EINVAL;
4002
4003        kfree(conf->mirrors_new);
4004        conf->mirrors_new = NULL;
4005        if (mddev->delta_disks > 0) {
4006                /* allocate new 'mirrors' list */
4007                conf->mirrors_new = kzalloc(
4008                        sizeof(struct raid10_info)
4009                        *(mddev->raid_disks +
4010                          mddev->delta_disks),
4011                        GFP_KERNEL);
4012                if (!conf->mirrors_new)
4013                        return -ENOMEM;
4014        }
4015        return 0;
4016}
4017
4018/*
4019 * Need to check if array has failed when deciding whether to:
4020 *  - start an array
4021 *  - remove non-faulty devices
4022 *  - add a spare
4023 *  - allow a reshape
4024 * This determination is simple when no reshape is happening.
4025 * However if there is a reshape, we need to carefully check
4026 * both the before and after sections.
4027 * This is because some failed devices may only affect one
4028 * of the two sections, and some non-in_sync devices may
4029 * be insync in the section most affected by failed devices.
4030 */
4031static int calc_degraded(struct r10conf *conf)
4032{
4033        int degraded, degraded2;
4034        int i;
4035
4036        rcu_read_lock();
4037        degraded = 0;
4038        /* 'prev' section first */
4039        for (i = 0; i < conf->prev.raid_disks; i++) {
4040                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4041                if (!rdev || test_bit(Faulty, &rdev->flags))
4042                        degraded++;
4043                else if (!test_bit(In_sync, &rdev->flags))
4044                        /* When we can reduce the number of devices in
4045                         * an array, this might not contribute to
4046                         * 'degraded'.  It does now.
4047                         */
4048                        degraded++;
4049        }
4050        rcu_read_unlock();
4051        if (conf->geo.raid_disks == conf->prev.raid_disks)
4052                return degraded;
4053        rcu_read_lock();
4054        degraded2 = 0;
4055        for (i = 0; i < conf->geo.raid_disks; i++) {
4056                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4057                if (!rdev || test_bit(Faulty, &rdev->flags))
4058                        degraded2++;
4059                else if (!test_bit(In_sync, &rdev->flags)) {
4060                        /* If reshape is increasing the number of devices,
4061                         * this section has already been recovered, so
4062                         * it doesn't contribute to degraded.
4063                         * else it does.
4064                         */
4065                        if (conf->geo.raid_disks <= conf->prev.raid_disks)
4066                                degraded2++;
4067                }
4068        }
4069        rcu_read_unlock();
4070        if (degraded2 > degraded)
4071                return degraded2;
4072        return degraded;
4073}
4074
4075static int raid10_start_reshape(struct mddev *mddev)
4076{
4077        /* A 'reshape' has been requested. This commits
4078         * the various 'new' fields and sets MD_RECOVER_RESHAPE
4079         * This also checks if there are enough spares and adds them
4080         * to the array.
4081         * We currently require enough spares to make the final
4082         * array non-degraded.  We also require that the difference
4083         * between old and new data_offset - on each device - is
4084         * enough that we never risk over-writing.
4085         */
4086
4087        unsigned long before_length, after_length;
4088        sector_t min_offset_diff = 0;
4089        int first = 1;
4090        struct geom new;
4091        struct r10conf *conf = mddev->private;
4092        struct md_rdev *rdev;
4093        int spares = 0;
4094        int ret;
4095
4096        if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4097                return -EBUSY;
4098
4099        if (setup_geo(&new, mddev, geo_start) != conf->copies)
4100                return -EINVAL;
4101
4102        before_length = ((1 << conf->prev.chunk_shift) *
4103                         conf->prev.far_copies);
4104        after_length = ((1 << conf->geo.chunk_shift) *
4105                        conf->geo.far_copies);
4106
4107        rdev_for_each(rdev, mddev) {
4108                if (!test_bit(In_sync, &rdev->flags)
4109                    && !test_bit(Faulty, &rdev->flags))
4110                        spares++;
4111                if (rdev->raid_disk >= 0) {
4112                        long long diff = (rdev->new_data_offset
4113                                          - rdev->data_offset);
4114                        if (!mddev->reshape_backwards)
4115                                diff = -diff;
4116                        if (diff < 0)
4117                                diff = 0;
4118                        if (first || diff < min_offset_diff)
4119                                min_offset_diff = diff;
4120                }
4121        }
4122
4123        if (max(before_length, after_length) > min_offset_diff)
4124                return -EINVAL;
4125
4126        if (spares < mddev->delta_disks)
4127                return -EINVAL;
4128
4129        conf->offset_diff = min_offset_diff;
4130        spin_lock_irq(&conf->device_lock);
4131        if (conf->mirrors_new) {
4132                memcpy(conf->mirrors_new, conf->mirrors,
4133                       sizeof(struct raid10_info)*conf->prev.raid_disks);
4134                smp_mb();
4135                kfree(conf->mirrors_old); /* FIXME and elsewhere */
4136                conf->mirrors_old = conf->mirrors;
4137                conf->mirrors = conf->mirrors_new;
4138                conf->mirrors_new = NULL;
4139        }
4140        setup_geo(&conf->geo, mddev, geo_start);
4141        smp_mb();
4142        if (mddev->reshape_backwards) {
4143                sector_t size = raid10_size(mddev, 0, 0);
4144                if (size < mddev->array_sectors) {
4145                        spin_unlock_irq(&conf->device_lock);
4146                        printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4147                               mdname(mddev));
4148                        return -EINVAL;
4149                }
4150                mddev->resync_max_sectors = size;
4151                conf->reshape_progress = size;
4152        } else
4153                conf->reshape_progress = 0;
4154        spin_unlock_irq(&conf->device_lock);
4155
4156        if (mddev->delta_disks && mddev->bitmap) {
4157                ret = bitmap_resize(mddev->bitmap,
4158                                    raid10_size(mddev, 0,
4159                                                conf->geo.raid_disks),
4160                                    0, 0);
4161                if (ret)
4162                        goto abort;
4163        }
4164        if (mddev->delta_disks > 0) {
4165                rdev_for_each(rdev, mddev)
4166                        if (rdev->raid_disk < 0 &&
4167                            !test_bit(Faulty, &rdev->flags)) {
4168                                if (raid10_add_disk(mddev, rdev) == 0) {
4169                                        if (rdev->raid_disk >=
4170                                            conf->prev.raid_disks)
4171                                                set_bit(In_sync, &rdev->flags);
4172                                        else
4173                                                rdev->recovery_offset = 0;
4174
4175                                        if (sysfs_link_rdev(mddev, rdev))
4176                                                /* Failure here  is OK */;
4177                                }
4178                        } else if (rdev->raid_disk >= conf->prev.raid_disks
4179                                   && !test_bit(Faulty, &rdev->flags)) {
4180                                /* This is a spare that was manually added */
4181                                set_bit(In_sync, &rdev->flags);
4182                        }
4183        }
4184        /* When a reshape changes the number of devices,
4185         * ->degraded is measured against the larger of the
4186         * pre and  post numbers.
4187         */
4188        spin_lock_irq(&conf->device_lock);
4189        mddev->degraded = calc_degraded(conf);
4190        spin_unlock_irq(&conf->device_lock);
4191        mddev->raid_disks = conf->geo.raid_disks;
4192        mddev->reshape_position = conf->reshape_progress;
4193        set_bit(MD_CHANGE_DEVS, &mddev->flags);
4194
4195        clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4196        clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4197        set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4198        set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4199
4200        mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4201                                                "reshape");
4202        if (!mddev->sync_thread) {
4203                ret = -EAGAIN;
4204                goto abort;
4205        }
4206        conf->reshape_checkpoint = jiffies;
4207        md_wakeup_thread(mddev->sync_thread);
4208        md_new_event(mddev);
4209        return 0;
4210
4211abort:
4212        mddev->recovery = 0;
4213        spin_lock_irq(&conf->device_lock);
4214        conf->geo = conf->prev;
4215        mddev->raid_disks = conf->geo.raid_disks;
4216        rdev_for_each(rdev, mddev)
4217                rdev->new_data_offset = rdev->data_offset;
4218        smp_wmb();
4219        conf->reshape_progress = MaxSector;
4220        mddev->reshape_position = MaxSector;
4221        spin_unlock_irq(&conf->device_lock);
4222        return ret;
4223}
4224
4225/* Calculate the last device-address that could contain
4226 * any block from the chunk that includes the array-address 's'
4227 * and report the next address.
4228 * i.e. the address returned will be chunk-aligned and after
4229 * any data that is in the chunk containing 's'.
4230 */
4231static sector_t last_dev_address(sector_t s, struct geom *geo)
4232{
4233        s = (s | geo->chunk_mask) + 1;
4234        s >>= geo->chunk_shift;
4235        s *= geo->near_copies;
4236        s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4237        s *= geo->far_copies;
4238        s <<= geo->chunk_shift;
4239        return s;
4240}
4241
4242/* Calculate the first device-address that could contain
4243 * any block from the chunk that includes the array-address 's'.
4244 * This too will be the start of a chunk
4245 */
4246static sector_t first_dev_address(sector_t s, struct geom *geo)
4247{
4248        s >>= geo->chunk_shift;
4249        s *= geo->near_copies;
4250        sector_div(s, geo->raid_disks);
4251        s *= geo->far_copies;
4252        s <<= geo->chunk_shift;
4253        return s;
4254}
4255
4256static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4257                                int *skipped)
4258{
4259        /* We simply copy at most one chunk (smallest of old and new)
4260         * at a time, possibly less if that exceeds RESYNC_PAGES,
4261         * or we hit a bad block or something.
4262         * This might mean we pause for normal IO in the middle of
4263         * a chunk, but that is not a problem was mddev->reshape_position
4264         * can record any location.
4265         *
4266         * If we will want to write to a location that isn't
4267         * yet recorded as 'safe' (i.e. in metadata on disk) then
4268         * we need to flush all reshape requests and update the metadata.
4269         *
4270         * When reshaping forwards (e.g. to more devices), we interpret
4271         * 'safe' as the earliest block which might not have been copied
4272         * down yet.  We divide this by previous stripe size and multiply
4273         * by previous stripe length to get lowest device offset that we
4274         * cannot write to yet.
4275         * We interpret 'sector_nr' as an address that we want to write to.
4276         * From this we use last_device_address() to find where we might
4277         * write to, and first_device_address on the  'safe' position.
4278         * If this 'next' write position is after the 'safe' position,
4279         * we must update the metadata to increase the 'safe' position.
4280         *
4281         * When reshaping backwards, we round in the opposite direction
4282         * and perform the reverse test:  next write position must not be
4283         * less than current safe position.
4284         *
4285         * In all this the minimum difference in data offsets
4286         * (conf->offset_diff - always positive) allows a bit of slack,
4287         * so next can be after 'safe', but not by more than offset_disk
4288         *
4289         * We need to prepare all the bios here before we start any IO
4290         * to ensure the size we choose is acceptable to all devices.
4291         * The means one for each copy for write-out and an extra one for
4292         * read-in.
4293         * We store the read-in bio in ->master_bio and the others in
4294         * ->devs[x].bio and ->devs[x].repl_bio.
4295         */
4296        struct r10conf *conf = mddev->private;
4297        struct r10bio *r10_bio;
4298        sector_t next, safe, last;
4299        int max_sectors;
4300        int nr_sectors;
4301        int s;
4302        struct md_rdev *rdev;
4303        int need_flush = 0;
4304        struct bio *blist;
4305        struct bio *bio, *read_bio;
4306        int sectors_done = 0;
4307
4308        if (sector_nr == 0) {
4309                /* If restarting in the middle, skip the initial sectors */
4310                if (mddev->reshape_backwards &&
4311                    conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4312                        sector_nr = (raid10_size(mddev, 0, 0)
4313                                     - conf->reshape_progress);
4314                } else if (!mddev->reshape_backwards &&
4315                           conf->reshape_progress > 0)
4316                        sector_nr = conf->reshape_progress;
4317                if (sector_nr) {
4318                        mddev->curr_resync_completed = sector_nr;
4319                        sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4320                        *skipped = 1;
4321                        return sector_nr;
4322                }
4323        }
4324
4325        /* We don't use sector_nr to track where we are up to
4326         * as that doesn't work well for ->reshape_backwards.
4327         * So just use ->reshape_progress.
4328         */
4329        if (mddev->reshape_backwards) {
4330                /* 'next' is the earliest device address that we might
4331                 * write to for this chunk in the new layout
4332                 */
4333                next = first_dev_address(conf->reshape_progress - 1,
4334                                         &conf->geo);
4335
4336                /* 'safe' is the last device address that we might read from
4337                 * in the old layout after a restart
4338                 */
4339                safe = last_dev_address(conf->reshape_safe - 1,
4340                                        &conf->prev);
4341
4342                if (next + conf->offset_diff < safe)
4343                        need_flush = 1;
4344
4345                last = conf->reshape_progress - 1;
4346                sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4347                                               & conf->prev.chunk_mask);
4348                if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4349                        sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4350        } else {
4351                /* 'next' is after the last device address that we
4352                 * might write to for this chunk in the new layout
4353                 */
4354                next = last_dev_address(conf->reshape_progress, &conf->geo);
4355
4356                /* 'safe' is the earliest device address that we might
4357                 * read from in the old layout after a restart
4358                 */
4359                safe = first_dev_address(conf->reshape_safe, &conf->prev);
4360
4361                /* Need to update metadata if 'next' might be beyond 'safe'
4362                 * as that would possibly corrupt data
4363                 */
4364                if (next > safe + conf->offset_diff)
4365                        need_flush = 1;
4366
4367                sector_nr = conf->reshape_progress;
4368                last  = sector_nr | (conf->geo.chunk_mask
4369                                     & conf->prev.chunk_mask);
4370
4371                if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4372                        last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4373        }
4374
4375        if (need_flush ||
4376            time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4377                /* Need to update reshape_position in metadata */
4378                wait_barrier(conf);
4379                mddev->reshape_position = conf->reshape_progress;
4380                if (mddev->reshape_backwards)
4381                        mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4382                                - conf->reshape_progress;
4383                else
4384                        mddev->curr_resync_completed = conf->reshape_progress;
4385                conf->reshape_checkpoint = jiffies;
4386                set_bit(MD_CHANGE_DEVS, &mddev->flags);
4387                md_wakeup_thread(mddev->thread);
4388                wait_event(mddev->sb_wait, mddev->flags == 0 ||
4389                           kthread_should_stop());
4390                conf->reshape_safe = mddev->reshape_position;
4391                allow_barrier(conf);
4392        }
4393
4394read_more:
4395        /* Now schedule reads for blocks from sector_nr to last */
4396        r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4397        raise_barrier(conf, sectors_done != 0);
4398        atomic_set(&r10_bio->remaining, 0);
4399        r10_bio->mddev = mddev;
4400        r10_bio->sector = sector_nr;
4401        set_bit(R10BIO_IsReshape, &r10_bio->state);
4402        r10_bio->sectors = last - sector_nr + 1;
4403        rdev = read_balance(conf, r10_bio, &max_sectors);
4404        BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4405
4406        if (!rdev) {
4407                /* Cannot read from here, so need to record bad blocks
4408                 * on all the target devices.
4409                 */
4410                // FIXME
4411                set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4412                return sectors_done;
4413        }
4414
4415        read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4416
4417        read_bio->bi_bdev = rdev->bdev;
4418        read_bio->bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4419                               + rdev->data_offset);
4420        read_bio->bi_private = r10_bio;
4421        read_bio->bi_end_io = end_sync_read;
4422        read_bio->bi_rw = READ;
4423        read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
4424        read_bio->bi_flags |= 1 << BIO_UPTODATE;
4425        read_bio->bi_vcnt = 0;
4426        read_bio->bi_size = 0;
4427        r10_bio->master_bio = read_bio;
4428        r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4429
4430        /* Now find the locations in the new layout */
4431        __raid10_find_phys(&conf->geo, r10_bio);
4432
4433        blist = read_bio;
4434        read_bio->bi_next = NULL;
4435
4436        for (s = 0; s < conf->copies*2; s++) {
4437                struct bio *b;
4438                int d = r10_bio->devs[s/2].devnum;
4439                struct md_rdev *rdev2;
4440                if (s&1) {
4441                        rdev2 = conf->mirrors[d].replacement;
4442                        b = r10_bio->devs[s/2].repl_bio;
4443                } else {
4444                        rdev2 = conf->mirrors[d].rdev;
4445                        b = r10_bio->devs[s/2].bio;
4446                }
4447                if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4448                        continue;
4449
4450                bio_reset(b);
4451                b->bi_bdev = rdev2->bdev;
4452                b->bi_sector = r10_bio->devs[s/2].addr + rdev2->new_data_offset;
4453                b->bi_private = r10_bio;
4454                b->bi_end_io = end_reshape_write;
4455                b->bi_rw = WRITE;
4456                b->bi_next = blist;
4457                blist = b;
4458        }
4459
4460        /* Now add as many pages as possible to all of these bios. */
4461
4462        nr_sectors = 0;
4463        for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4464                struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4465                int len = (max_sectors - s) << 9;
4466                if (len > PAGE_SIZE)
4467                        len = PAGE_SIZE;
4468                for (bio = blist; bio ; bio = bio->bi_next) {
4469                        struct bio *bio2;
4470                        if (bio_add_page(bio, page, len, 0))
4471                                continue;
4472
4473                        /* Didn't fit, must stop */
4474                        for (bio2 = blist;
4475                             bio2 && bio2 != bio;
4476                             bio2 = bio2->bi_next) {
4477                                /* Remove last page from this bio */
4478                                bio2->bi_vcnt--;
4479                                bio2->bi_size -= len;
4480                                bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4481                        }
4482                        goto bio_full;
4483                }
4484                sector_nr += len >> 9;
4485                nr_sectors += len >> 9;
4486        }
4487bio_full:
4488        r10_bio->sectors = nr_sectors;
4489
4490        /* Now submit the read */
4491        md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4492        atomic_inc(&r10_bio->remaining);
4493        read_bio->bi_next = NULL;
4494        generic_make_request(read_bio);
4495        sector_nr += nr_sectors;
4496        sectors_done += nr_sectors;
4497        if (sector_nr <= last)
4498                goto read_more;
4499
4500        /* Now that we have done the whole section we can
4501         * update reshape_progress
4502         */
4503        if (mddev->reshape_backwards)
4504                conf->reshape_progress -= sectors_done;
4505        else
4506                conf->reshape_progress += sectors_done;
4507
4508        return sectors_done;
4509}
4510
4511static void end_reshape_request(struct r10bio *r10_bio);
4512static int handle_reshape_read_error(struct mddev *mddev,
4513                                     struct r10bio *r10_bio);
4514static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4515{
4516        /* Reshape read completed.  Hopefully we have a block
4517         * to write out.
4518         * If we got a read error then we do sync 1-page reads from
4519         * elsewhere until we find the data - or give up.
4520         */
4521        struct r10conf *conf = mddev->private;
4522        int s;
4523
4524        if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4525                if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4526                        /* Reshape has been aborted */
4527                        md_done_sync(mddev, r10_bio->sectors, 0);
4528                        return;
4529                }
4530
4531        /* We definitely have the data in the pages, schedule the
4532         * writes.
4533         */
4534        atomic_set(&r10_bio->remaining, 1);
4535        for (s = 0; s < conf->copies*2; s++) {
4536                struct bio *b;
4537                int d = r10_bio->devs[s/2].devnum;
4538                struct md_rdev *rdev;
4539                if (s&1) {
4540                        rdev = conf->mirrors[d].replacement;
4541                        b = r10_bio->devs[s/2].repl_bio;
4542                } else {
4543                        rdev = conf->mirrors[d].rdev;
4544                        b = r10_bio->devs[s/2].bio;
4545                }
4546                if (!rdev || test_bit(Faulty, &rdev->flags))
4547                        continue;
4548                atomic_inc(&rdev->nr_pending);
4549                md_sync_acct(b->bi_bdev, r10_bio->sectors);
4550                atomic_inc(&r10_bio->remaining);
4551                b->bi_next = NULL;
4552                generic_make_request(b);
4553        }
4554        end_reshape_request(r10_bio);
4555}
4556
4557static void end_reshape(struct r10conf *conf)
4558{
4559        if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4560                return;
4561
4562        spin_lock_irq(&conf->device_lock);
4563        conf->prev = conf->geo;
4564        md_finish_reshape(conf->mddev);
4565        smp_wmb();
4566        conf->reshape_progress = MaxSector;
4567        spin_unlock_irq(&conf->device_lock);
4568
4569        /* read-ahead size must cover two whole stripes, which is
4570         * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4571         */
4572        if (conf->mddev->queue) {
4573                int stripe = conf->geo.raid_disks *
4574                        ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4575                stripe /= conf->geo.near_copies;
4576                if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4577                        conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4578        }
4579        conf->fullsync = 0;
4580}
4581
4582
4583static int handle_reshape_read_error(struct mddev *mddev,
4584                                     struct r10bio *r10_bio)
4585{
4586        /* Use sync reads to get the blocks from somewhere else */
4587        int sectors = r10_bio->sectors;
4588        struct r10conf *conf = mddev->private;
4589        struct {
4590                struct r10bio r10_bio;
4591                struct r10dev devs[conf->copies];
4592        } on_stack;
4593        struct r10bio *r10b = &on_stack.r10_bio;
4594        int slot = 0;
4595        int idx = 0;
4596        struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4597
4598        r10b->sector = r10_bio->sector;
4599        __raid10_find_phys(&conf->prev, r10b);
4600
4601        while (sectors) {
4602                int s = sectors;
4603                int success = 0;
4604                int first_slot = slot;
4605
4606                if (s > (PAGE_SIZE >> 9))
4607                        s = PAGE_SIZE >> 9;
4608
4609                while (!success) {
4610                        int d = r10b->devs[slot].devnum;
4611                        struct md_rdev *rdev = conf->mirrors[d].rdev;
4612                        sector_t addr;
4613                        if (rdev == NULL ||
4614                            test_bit(Faulty, &rdev->flags) ||
4615                            !test_bit(In_sync, &rdev->flags))
4616                                goto failed;
4617
4618                        addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4619                        success = sync_page_io(rdev,
4620                                               addr,
4621                                               s << 9,
4622                                               bvec[idx].bv_page,
4623                                               READ, false);
4624                        if (success)
4625                                break;
4626                failed:
4627                        slot++;
4628                        if (slot >= conf->copies)
4629                                slot = 0;
4630                        if (slot == first_slot)
4631                                break;
4632                }
4633                if (!success) {
4634                        /* couldn't read this block, must give up */
4635                        set_bit(MD_RECOVERY_INTR,
4636                                &mddev->recovery);
4637                        return -EIO;
4638                }
4639                sectors -= s;
4640                idx++;
4641        }
4642        return 0;
4643}
4644
4645static void end_reshape_write(struct bio *bio, int error)
4646{
4647        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4648        struct r10bio *r10_bio = bio->bi_private;
4649        struct mddev *mddev = r10_bio->mddev;
4650        struct r10conf *conf = mddev->private;
4651        int d;
4652        int slot;
4653        int repl;
4654        struct md_rdev *rdev = NULL;
4655
4656        d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4657        if (repl)
4658                rdev = conf->mirrors[d].replacement;
4659        if (!rdev) {
4660                smp_mb();
4661                rdev = conf->mirrors[d].rdev;
4662        }
4663
4664        if (!uptodate) {
4665                /* FIXME should record badblock */
4666                md_error(mddev, rdev);
4667        }
4668
4669        rdev_dec_pending(rdev, mddev);
4670        end_reshape_request(r10_bio);
4671}
4672
4673static void end_reshape_request(struct r10bio *r10_bio)
4674{
4675        if (!atomic_dec_and_test(&r10_bio->remaining))
4676                return;
4677        md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4678        bio_put(r10_bio->master_bio);
4679        put_buf(r10_bio);
4680}
4681
4682static void raid10_finish_reshape(struct mddev *mddev)
4683{
4684        struct r10conf *conf = mddev->private;
4685
4686        if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4687                return;
4688
4689        if (mddev->delta_disks > 0) {
4690                sector_t size = raid10_size(mddev, 0, 0);
4691                md_set_array_sectors(mddev, size);
4692                if (mddev->recovery_cp > mddev->resync_max_sectors) {
4693                        mddev->recovery_cp = mddev->resync_max_sectors;
4694                        set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4695                }
4696                mddev->resync_max_sectors = size;
4697                set_capacity(mddev->gendisk, mddev->array_sectors);
4698                revalidate_disk(mddev->gendisk);
4699        } else {
4700                int d;
4701                for (d = conf->geo.raid_disks ;
4702                     d < conf->geo.raid_disks - mddev->delta_disks;
4703                     d++) {
4704                        struct md_rdev *rdev = conf->mirrors[d].rdev;
4705                        if (rdev)
4706                                clear_bit(In_sync, &rdev->flags);
4707                        rdev = conf->mirrors[d].replacement;
4708                        if (rdev)
4709                                clear_bit(In_sync, &rdev->flags);
4710                }
4711        }
4712        mddev->layout = mddev->new_layout;
4713        mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4714        mddev->reshape_position = MaxSector;
4715        mddev->delta_disks = 0;
4716        mddev->reshape_backwards = 0;
4717}
4718
4719static struct md_personality raid10_personality =
4720{
4721        .name           = "raid10",
4722        .level          = 10,
4723        .owner          = THIS_MODULE,
4724        .make_request   = make_request,
4725        .run            = run,
4726        .stop           = stop,
4727        .status         = status,
4728        .error_handler  = error,
4729        .hot_add_disk   = raid10_add_disk,
4730        .hot_remove_disk= raid10_remove_disk,
4731        .spare_active   = raid10_spare_active,
4732        .sync_request   = sync_request,
4733        .quiesce        = raid10_quiesce,
4734        .size           = raid10_size,
4735        .resize         = raid10_resize,
4736        .takeover       = raid10_takeover,
4737        .check_reshape  = raid10_check_reshape,
4738        .start_reshape  = raid10_start_reshape,
4739        .finish_reshape = raid10_finish_reshape,
4740};
4741
4742static int __init raid_init(void)
4743{
4744        return register_md_personality(&raid10_personality);
4745}
4746
4747static void raid_exit(void)
4748{
4749        unregister_md_personality(&raid10_personality);
4750}
4751
4752module_init(raid_init);
4753module_exit(raid_exit);
4754MODULE_LICENSE("GPL");
4755MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4756MODULE_ALIAS("md-personality-9"); /* RAID10 */
4757MODULE_ALIAS("md-raid10");
4758MODULE_ALIAS("md-level-10");
4759
4760module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
4761
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.