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                           && !test_bit(Faulty, &tmp->rdev->flags)
1786                           && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1787                        count++;
1788                        sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1789                }
1790        }
1791        spin_lock_irqsave(&conf->device_lock, flags);
1792        mddev->degraded -= count;
1793        spin_unlock_irqrestore(&conf->device_lock, flags);
1794
1795        print_conf(conf);
1796        return count;
1797}
1798
1799
1800static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1801{
1802        struct r10conf *conf = mddev->private;
1803        int err = -EEXIST;
1804        int mirror;
1805        int first = 0;
1806        int last = conf->geo.raid_disks - 1;
1807        struct request_queue *q = bdev_get_queue(rdev->bdev);
1808
1809        if (mddev->recovery_cp < MaxSector)
1810                /* only hot-add to in-sync arrays, as recovery is
1811                 * very different from resync
1812                 */
1813                return -EBUSY;
1814        if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1815                return -EINVAL;
1816
1817        if (rdev->raid_disk >= 0)
1818                first = last = rdev->raid_disk;
1819
1820        if (q->merge_bvec_fn) {
1821                set_bit(Unmerged, &rdev->flags);
1822                mddev->merge_check_needed = 1;
1823        }
1824
1825        if (rdev->saved_raid_disk >= first &&
1826            conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1827                mirror = rdev->saved_raid_disk;
1828        else
1829                mirror = first;
1830        for ( ; mirror <= last ; mirror++) {
1831                struct raid10_info *p = &conf->mirrors[mirror];
1832                if (p->recovery_disabled == mddev->recovery_disabled)
1833                        continue;
1834                if (p->rdev) {
1835                        if (!test_bit(WantReplacement, &p->rdev->flags) ||
1836                            p->replacement != NULL)
1837                                continue;
1838                        clear_bit(In_sync, &rdev->flags);
1839                        set_bit(Replacement, &rdev->flags);
1840                        rdev->raid_disk = mirror;
1841                        err = 0;
1842                        if (mddev->gendisk)
1843                                disk_stack_limits(mddev->gendisk, rdev->bdev,
1844                                                  rdev->data_offset << 9);
1845                        conf->fullsync = 1;
1846                        rcu_assign_pointer(p->replacement, rdev);
1847                        break;
1848                }
1849
1850                if (mddev->gendisk)
1851                        disk_stack_limits(mddev->gendisk, rdev->bdev,
1852                                          rdev->data_offset << 9);
1853
1854                p->head_position = 0;
1855                p->recovery_disabled = mddev->recovery_disabled - 1;
1856                rdev->raid_disk = mirror;
1857                err = 0;
1858                if (rdev->saved_raid_disk != mirror)
1859                        conf->fullsync = 1;
1860                rcu_assign_pointer(p->rdev, rdev);
1861                break;
1862        }
1863        if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1864                /* Some requests might not have seen this new
1865                 * merge_bvec_fn.  We must wait for them to complete
1866                 * before merging the device fully.
1867                 * First we make sure any code which has tested
1868                 * our function has submitted the request, then
1869                 * we wait for all outstanding requests to complete.
1870                 */
1871                synchronize_sched();
1872                freeze_array(conf, 0);
1873                unfreeze_array(conf);
1874                clear_bit(Unmerged, &rdev->flags);
1875        }
1876        md_integrity_add_rdev(rdev, mddev);
1877        if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1878                queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1879
1880        print_conf(conf);
1881        return err;
1882}
1883
1884static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1885{
1886        struct r10conf *conf = mddev->private;
1887        int err = 0;
1888        int number = rdev->raid_disk;
1889        struct md_rdev **rdevp;
1890        struct raid10_info *p = conf->mirrors + number;
1891
1892        print_conf(conf);
1893        if (rdev == p->rdev)
1894                rdevp = &p->rdev;
1895        else if (rdev == p->replacement)
1896                rdevp = &p->replacement;
1897        else
1898                return 0;
1899
1900        if (test_bit(In_sync, &rdev->flags) ||
1901            atomic_read(&rdev->nr_pending)) {
1902                err = -EBUSY;
1903                goto abort;
1904        }
1905        /* Only remove faulty devices if recovery
1906         * is not possible.
1907         */
1908        if (!test_bit(Faulty, &rdev->flags) &&
1909            mddev->recovery_disabled != p->recovery_disabled &&
1910            (!p->replacement || p->replacement == rdev) &&
1911            number < conf->geo.raid_disks &&
1912            enough(conf, -1)) {
1913                err = -EBUSY;
1914                goto abort;
1915        }
1916        *rdevp = NULL;
1917        synchronize_rcu();
1918        if (atomic_read(&rdev->nr_pending)) {
1919                /* lost the race, try later */
1920                err = -EBUSY;
1921                *rdevp = rdev;
1922                goto abort;
1923        } else if (p->replacement) {
1924                /* We must have just cleared 'rdev' */
1925                p->rdev = p->replacement;
1926                clear_bit(Replacement, &p->replacement->flags);
1927                smp_mb(); /* Make sure other CPUs may see both as identical
1928                           * but will never see neither -- if they are careful.
1929                           */
1930                p->replacement = NULL;
1931                clear_bit(WantReplacement, &rdev->flags);
1932        } else
1933                /* We might have just remove the Replacement as faulty
1934                 * Clear the flag just in case
1935                 */
1936                clear_bit(WantReplacement, &rdev->flags);
1937
1938        err = md_integrity_register(mddev);
1939
1940abort:
1941
1942        print_conf(conf);
1943        return err;
1944}
1945
1946
1947static void end_sync_read(struct bio *bio, int error)
1948{
1949        struct r10bio *r10_bio = bio->bi_private;
1950        struct r10conf *conf = r10_bio->mddev->private;
1951        int d;
1952
1953        if (bio == r10_bio->master_bio) {
1954                /* this is a reshape read */
1955                d = r10_bio->read_slot; /* really the read dev */
1956        } else
1957                d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1958
1959        if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1960                set_bit(R10BIO_Uptodate, &r10_bio->state);
1961        else
1962                /* The write handler will notice the lack of
1963                 * R10BIO_Uptodate and record any errors etc
1964                 */
1965                atomic_add(r10_bio->sectors,
1966                           &conf->mirrors[d].rdev->corrected_errors);
1967
1968        /* for reconstruct, we always reschedule after a read.
1969         * for resync, only after all reads
1970         */
1971        rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1972        if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1973            atomic_dec_and_test(&r10_bio->remaining)) {
1974                /* we have read all the blocks,
1975                 * do the comparison in process context in raid10d
1976                 */
1977                reschedule_retry(r10_bio);
1978        }
1979}
1980
1981static void end_sync_request(struct r10bio *r10_bio)
1982{
1983        struct mddev *mddev = r10_bio->mddev;
1984
1985        while (atomic_dec_and_test(&r10_bio->remaining)) {
1986                if (r10_bio->master_bio == NULL) {
1987                        /* the primary of several recovery bios */
1988                        sector_t s = r10_bio->sectors;
1989                        if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1990                            test_bit(R10BIO_WriteError, &r10_bio->state))
1991                                reschedule_retry(r10_bio);
1992                        else
1993                                put_buf(r10_bio);
1994                        md_done_sync(mddev, s, 1);
1995                        break;
1996                } else {
1997                        struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1998                        if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1999                            test_bit(R10BIO_WriteError, &r10_bio->state))
2000                                reschedule_retry(r10_bio);
2001                        else
2002                                put_buf(r10_bio);
2003                        r10_bio = r10_bio2;
2004                }
2005        }
2006}
2007
2008static void end_sync_write(struct bio *bio, int error)
2009{
2010        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2011        struct r10bio *r10_bio = bio->bi_private;
2012        struct mddev *mddev = r10_bio->mddev;
2013        struct r10conf *conf = mddev->private;
2014        int d;
2015        sector_t first_bad;
2016        int bad_sectors;
2017        int slot;
2018        int repl;
2019        struct md_rdev *rdev = NULL;
2020
2021        d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2022        if (repl)
2023                rdev = conf->mirrors[d].replacement;
2024        else
2025                rdev = conf->mirrors[d].rdev;
2026
2027        if (!uptodate) {
2028                if (repl)
2029                        md_error(mddev, rdev);
2030                else {
2031                        set_bit(WriteErrorSeen, &rdev->flags);
2032                        if (!test_and_set_bit(WantReplacement, &rdev->flags))
2033                                set_bit(MD_RECOVERY_NEEDED,
2034                                        &rdev->mddev->recovery);
2035                        set_bit(R10BIO_WriteError, &r10_bio->state);
2036                }
2037        } else if (is_badblock(rdev,
2038                             r10_bio->devs[slot].addr,
2039                             r10_bio->sectors,
2040                             &first_bad, &bad_sectors))
2041                set_bit(R10BIO_MadeGood, &r10_bio->state);
2042
2043        rdev_dec_pending(rdev, mddev);
2044
2045        end_sync_request(r10_bio);
2046}
2047
2048/*
2049 * Note: sync and recover and handled very differently for raid10
2050 * This code is for resync.
2051 * For resync, we read through virtual addresses and read all blocks.
2052 * If there is any error, we schedule a write.  The lowest numbered
2053 * drive is authoritative.
2054 * However requests come for physical address, so we need to map.
2055 * For every physical address there are raid_disks/copies virtual addresses,
2056 * which is always are least one, but is not necessarly an integer.
2057 * This means that a physical address can span multiple chunks, so we may
2058 * have to submit multiple io requests for a single sync request.
2059 */
2060/*
2061 * We check if all blocks are in-sync and only write to blocks that
2062 * aren't in sync
2063 */
2064static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2065{
2066        struct r10conf *conf = mddev->private;
2067        int i, first;
2068        struct bio *tbio, *fbio;
2069        int vcnt;
2070
2071        atomic_set(&r10_bio->remaining, 1);
2072
2073        /* find the first device with a block */
2074        for (i=0; i<conf->copies; i++)
2075                if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
2076                        break;
2077
2078        if (i == conf->copies)
2079                goto done;
2080
2081        first = i;
2082        fbio = r10_bio->devs[i].bio;
2083
2084        vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2085        /* now find blocks with errors */
2086        for (i=0 ; i < conf->copies ; i++) {
2087                int  j, d;
2088
2089                tbio = r10_bio->devs[i].bio;
2090
2091                if (tbio->bi_end_io != end_sync_read)
2092                        continue;
2093                if (i == first)
2094                        continue;
2095                if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
2096                        /* We know that the bi_io_vec layout is the same for
2097                         * both 'first' and 'i', so we just compare them.
2098                         * All vec entries are PAGE_SIZE;
2099                         */
2100                        int sectors = r10_bio->sectors;
2101                        for (j = 0; j < vcnt; j++) {
2102                                int len = PAGE_SIZE;
2103                                if (sectors < (len / 512))
2104                                        len = sectors * 512;
2105                                if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2106                                           page_address(tbio->bi_io_vec[j].bv_page),
2107                                           len))
2108                                        break;
2109                                sectors -= len/512;
2110                        }
2111                        if (j == vcnt)
2112                                continue;
2113                        atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2114                        if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2115                                /* Don't fix anything. */
2116                                continue;
2117                }
2118                /* Ok, we need to write this bio, either to correct an
2119                 * inconsistency or to correct an unreadable block.
2120                 * First we need to fixup bv_offset, bv_len and
2121                 * bi_vecs, as the read request might have corrupted these
2122                 */
2123                bio_reset(tbio);
2124
2125                tbio->bi_vcnt = vcnt;
2126                tbio->bi_size = r10_bio->sectors << 9;
2127                tbio->bi_rw = WRITE;
2128                tbio->bi_private = r10_bio;
2129                tbio->bi_sector = r10_bio->devs[i].addr;
2130
2131                for (j=0; j < vcnt ; j++) {
2132                        tbio->bi_io_vec[j].bv_offset = 0;
2133                        tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
2134
2135                        memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2136                               page_address(fbio->bi_io_vec[j].bv_page),
2137                               PAGE_SIZE);
2138                }
2139                tbio->bi_end_io = end_sync_write;
2140
2141                d = r10_bio->devs[i].devnum;
2142                atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2143                atomic_inc(&r10_bio->remaining);
2144                md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2145
2146                tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
2147                tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2148                generic_make_request(tbio);
2149        }
2150
2151        /* Now write out to any replacement devices
2152         * that are active
2153         */
2154        for (i = 0; i < conf->copies; i++) {
2155                int j, d;
2156
2157                tbio = r10_bio->devs[i].repl_bio;
2158                if (!tbio || !tbio->bi_end_io)
2159                        continue;
2160                if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2161                    && r10_bio->devs[i].bio != fbio)
2162                        for (j = 0; j < vcnt; j++)
2163                                memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2164                                       page_address(fbio->bi_io_vec[j].bv_page),
2165                                       PAGE_SIZE);
2166                d = r10_bio->devs[i].devnum;
2167                atomic_inc(&r10_bio->remaining);
2168                md_sync_acct(conf->mirrors[d].replacement->bdev,
2169                             bio_sectors(tbio));
2170                generic_make_request(tbio);
2171        }
2172
2173done:
2174        if (atomic_dec_and_test(&r10_bio->remaining)) {
2175                md_done_sync(mddev, r10_bio->sectors, 1);
2176                put_buf(r10_bio);
2177        }
2178}
2179
2180/*
2181 * Now for the recovery code.
2182 * Recovery happens across physical sectors.
2183 * We recover all non-is_sync drives by finding the virtual address of
2184 * each, and then choose a working drive that also has that virt address.
2185 * There is a separate r10_bio for each non-in_sync drive.
2186 * Only the first two slots are in use. The first for reading,
2187 * The second for writing.
2188 *
2189 */
2190static void fix_recovery_read_error(struct r10bio *r10_bio)
2191{
2192        /* We got a read error during recovery.
2193         * We repeat the read in smaller page-sized sections.
2194         * If a read succeeds, write it to the new device or record
2195         * a bad block if we cannot.
2196         * If a read fails, record a bad block on both old and
2197         * new devices.
2198         */
2199        struct mddev *mddev = r10_bio->mddev;
2200        struct r10conf *conf = mddev->private;
2201        struct bio *bio = r10_bio->devs[0].bio;
2202        sector_t sect = 0;
2203        int sectors = r10_bio->sectors;
2204        int idx = 0;
2205        int dr = r10_bio->devs[0].devnum;
2206        int dw = r10_bio->devs[1].devnum;
2207
2208        while (sectors) {
2209                int s = sectors;
2210                struct md_rdev *rdev;
2211                sector_t addr;
2212                int ok;
2213
2214                if (s > (PAGE_SIZE>>9))
2215                        s = PAGE_SIZE >> 9;
2216
2217                rdev = conf->mirrors[dr].rdev;
2218                addr = r10_bio->devs[0].addr + sect,
2219                ok = sync_page_io(rdev,
2220                                  addr,
2221                                  s << 9,
2222                                  bio->bi_io_vec[idx].bv_page,
2223                                  READ, false);
2224                if (ok) {
2225                        rdev = conf->mirrors[dw].rdev;
2226                        addr = r10_bio->devs[1].addr + sect;
2227                        ok = sync_page_io(rdev,
2228                                          addr,
2229                                          s << 9,
2230                                          bio->bi_io_vec[idx].bv_page,
2231                                          WRITE, false);
2232                        if (!ok) {
2233                                set_bit(WriteErrorSeen, &rdev->flags);
2234                                if (!test_and_set_bit(WantReplacement,
2235                                                      &rdev->flags))
2236                                        set_bit(MD_RECOVERY_NEEDED,
2237                                                &rdev->mddev->recovery);
2238                        }
2239                }
2240                if (!ok) {
2241                        /* We don't worry if we cannot set a bad block -
2242                         * it really is bad so there is no loss in not
2243                         * recording it yet
2244                         */
2245                        rdev_set_badblocks(rdev, addr, s, 0);
2246
2247                        if (rdev != conf->mirrors[dw].rdev) {
2248                                /* need bad block on destination too */
2249                                struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2250                                addr = r10_bio->devs[1].addr + sect;
2251                                ok = rdev_set_badblocks(rdev2, addr, s, 0);
2252                                if (!ok) {
2253                                        /* just abort the recovery */
2254                                        printk(KERN_NOTICE
2255                                               "md/raid10:%s: recovery aborted"
2256                                               " due to read error\n",
2257                                               mdname(mddev));
2258
2259                                        conf->mirrors[dw].recovery_disabled
2260                                                = mddev->recovery_disabled;
2261                                        set_bit(MD_RECOVERY_INTR,
2262                                                &mddev->recovery);
2263                                        break;
2264                                }
2265                        }
2266                }
2267
2268                sectors -= s;
2269                sect += s;
2270                idx++;
2271        }
2272}
2273
2274static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2275{
2276        struct r10conf *conf = mddev->private;
2277        int d;
2278        struct bio *wbio, *wbio2;
2279
2280        if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2281                fix_recovery_read_error(r10_bio);
2282                end_sync_request(r10_bio);
2283                return;
2284        }
2285
2286        /*
2287         * share the pages with the first bio
2288         * and submit the write request
2289         */
2290        d = r10_bio->devs[1].devnum;
2291        wbio = r10_bio->devs[1].bio;
2292        wbio2 = r10_bio->devs[1].repl_bio;
2293        /* Need to test wbio2->bi_end_io before we call
2294         * generic_make_request as if the former is NULL,
2295         * the latter is free to free wbio2.
2296         */
2297        if (wbio2 && !wbio2->bi_end_io)
2298                wbio2 = NULL;
2299        if (wbio->bi_end_io) {
2300                atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2301                md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2302                generic_make_request(wbio);
2303        }
2304        if (wbio2) {
2305                atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2306                md_sync_acct(conf->mirrors[d].replacement->bdev,
2307                             bio_sectors(wbio2));
2308                generic_make_request(wbio2);
2309        }
2310}
2311
2312
2313/*
2314 * Used by fix_read_error() to decay the per rdev read_errors.
2315 * We halve the read error count for every hour that has elapsed
2316 * since the last recorded read error.
2317 *
2318 */
2319static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2320{
2321        struct timespec cur_time_mon;
2322        unsigned long hours_since_last;
2323        unsigned int read_errors = atomic_read(&rdev->read_errors);
2324
2325        ktime_get_ts(&cur_time_mon);
2326
2327        if (rdev->last_read_error.tv_sec == 0 &&
2328            rdev->last_read_error.tv_nsec == 0) {
2329                /* first time we've seen a read error */
2330                rdev->last_read_error = cur_time_mon;
2331                return;
2332        }
2333
2334        hours_since_last = (cur_time_mon.tv_sec -
2335                            rdev->last_read_error.tv_sec) / 3600;
2336
2337        rdev->last_read_error = cur_time_mon;
2338
2339        /*
2340         * if hours_since_last is > the number of bits in read_errors
2341         * just set read errors to 0. We do this to avoid
2342         * overflowing the shift of read_errors by hours_since_last.
2343         */
2344        if (hours_since_last >= 8 * sizeof(read_errors))
2345                atomic_set(&rdev->read_errors, 0);
2346        else
2347                atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2348}
2349
2350static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2351                            int sectors, struct page *page, int rw)
2352{
2353        sector_t first_bad;
2354        int bad_sectors;
2355
2356        if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2357            && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2358                return -1;
2359        if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2360                /* success */
2361                return 1;
2362        if (rw == WRITE) {
2363                set_bit(WriteErrorSeen, &rdev->flags);
2364                if (!test_and_set_bit(WantReplacement, &rdev->flags))
2365                        set_bit(MD_RECOVERY_NEEDED,
2366                                &rdev->mddev->recovery);
2367        }
2368        /* need to record an error - either for the block or the device */
2369        if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2370                md_error(rdev->mddev, rdev);
2371        return 0;
2372}
2373
2374/*
2375 * This is a kernel thread which:
2376 *
2377 *      1.      Retries failed read operations on working mirrors.
2378 *      2.      Updates the raid superblock when problems encounter.
2379 *      3.      Performs writes following reads for array synchronising.
2380 */
2381
2382static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2383{
2384        int sect = 0; /* Offset from r10_bio->sector */
2385        int sectors = r10_bio->sectors;
2386        struct md_rdev*rdev;
2387        int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2388        int d = r10_bio->devs[r10_bio->read_slot].devnum;
2389
2390        /* still own a reference to this rdev, so it cannot
2391         * have been cleared recently.
2392         */
2393        rdev = conf->mirrors[d].rdev;
2394
2395        if (test_bit(Faulty, &rdev->flags))
2396                /* drive has already been failed, just ignore any
2397                   more fix_read_error() attempts */
2398                return;
2399
2400        check_decay_read_errors(mddev, rdev);
2401        atomic_inc(&rdev->read_errors);
2402        if (atomic_read(&rdev->read_errors) > max_read_errors) {
2403                char b[BDEVNAME_SIZE];
2404                bdevname(rdev->bdev, b);
2405
2406                printk(KERN_NOTICE
2407                       "md/raid10:%s: %s: Raid device exceeded "
2408                       "read_error threshold [cur %d:max %d]\n",
2409                       mdname(mddev), b,
2410                       atomic_read(&rdev->read_errors), max_read_errors);
2411                printk(KERN_NOTICE
2412                       "md/raid10:%s: %s: Failing raid device\n",
2413                       mdname(mddev), b);
2414                md_error(mddev, conf->mirrors[d].rdev);
2415                r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2416                return;
2417        }
2418
2419        while(sectors) {
2420                int s = sectors;
2421                int sl = r10_bio->read_slot;
2422                int success = 0;
2423                int start;
2424
2425                if (s > (PAGE_SIZE>>9))
2426                        s = PAGE_SIZE >> 9;
2427
2428                rcu_read_lock();
2429                do {
2430                        sector_t first_bad;
2431                        int bad_sectors;
2432
2433                        d = r10_bio->devs[sl].devnum;
2434                        rdev = rcu_dereference(conf->mirrors[d].rdev);
2435                        if (rdev &&
2436                            !test_bit(Unmerged, &rdev->flags) &&
2437                            test_bit(In_sync, &rdev->flags) &&
2438                            is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2439                                        &first_bad, &bad_sectors) == 0) {
2440                                atomic_inc(&rdev->nr_pending);
2441                                rcu_read_unlock();
2442                                success = sync_page_io(rdev,
2443                                                       r10_bio->devs[sl].addr +
2444                                                       sect,
2445                                                       s<<9,
2446                                                       conf->tmppage, READ, false);
2447                                rdev_dec_pending(rdev, mddev);
2448                                rcu_read_lock();
2449                                if (success)
2450                                        break;
2451                        }
2452                        sl++;
2453                        if (sl == conf->copies)
2454                                sl = 0;
2455                } while (!success && sl != r10_bio->read_slot);
2456                rcu_read_unlock();
2457
2458                if (!success) {
2459                        /* Cannot read from anywhere, just mark the block
2460                         * as bad on the first device to discourage future
2461                         * reads.
2462                         */
2463                        int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2464                        rdev = conf->mirrors[dn].rdev;
2465
2466                        if (!rdev_set_badblocks(
2467                                    rdev,
2468                                    r10_bio->devs[r10_bio->read_slot].addr
2469                                    + sect,
2470                                    s, 0)) {
2471                                md_error(mddev, rdev);
2472                                r10_bio->devs[r10_bio->read_slot].bio
2473                                        = IO_BLOCKED;
2474                        }
2475                        break;
2476                }
2477
2478                start = sl;
2479                /* write it back and re-read */
2480                rcu_read_lock();
2481                while (sl != r10_bio->read_slot) {
2482                        char b[BDEVNAME_SIZE];
2483
2484                        if (sl==0)
2485                                sl = conf->copies;
2486                        sl--;
2487                        d = r10_bio->devs[sl].devnum;
2488                        rdev = rcu_dereference(conf->mirrors[d].rdev);
2489                        if (!rdev ||
2490                            test_bit(Unmerged, &rdev->flags) ||
2491                            !test_bit(In_sync, &rdev->flags))
2492                                continue;
2493
2494                        atomic_inc(&rdev->nr_pending);
2495                        rcu_read_unlock();
2496                        if (r10_sync_page_io(rdev,
2497                                             r10_bio->devs[sl].addr +
2498                                             sect,
2499                                             s, conf->tmppage, WRITE)
2500                            == 0) {
2501                                /* Well, this device is dead */
2502                                printk(KERN_NOTICE
2503                                       "md/raid10:%s: read correction "
2504                                       "write failed"
2505                                       " (%d sectors at %llu on %s)\n",
2506                                       mdname(mddev), s,
2507                                       (unsigned long long)(
2508                                               sect +
2509                                               choose_data_offset(r10_bio,
2510                                                                  rdev)),
2511                                       bdevname(rdev->bdev, b));
2512                                printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2513                                       "drive\n",
2514                                       mdname(mddev),
2515                                       bdevname(rdev->bdev, b));
2516                        }
2517                        rdev_dec_pending(rdev, mddev);
2518                        rcu_read_lock();
2519                }
2520                sl = start;
2521                while (sl != r10_bio->read_slot) {
2522                        char b[BDEVNAME_SIZE];
2523
2524                        if (sl==0)
2525                                sl = conf->copies;
2526                        sl--;
2527                        d = r10_bio->devs[sl].devnum;
2528                        rdev = rcu_dereference(conf->mirrors[d].rdev);
2529                        if (!rdev ||
2530                            !test_bit(In_sync, &rdev->flags))
2531                                continue;
2532
2533                        atomic_inc(&rdev->nr_pending);
2534                        rcu_read_unlock();
2535                        switch (r10_sync_page_io(rdev,
2536                                             r10_bio->devs[sl].addr +
2537                                             sect,
2538                                             s, conf->tmppage,
2539                                                 READ)) {
2540                        case 0:
2541                                /* Well, this device is dead */
2542                                printk(KERN_NOTICE
2543                                       "md/raid10:%s: unable to read back "
2544                                       "corrected sectors"
2545                                       " (%d sectors at %llu on %s)\n",
2546                                       mdname(mddev), s,
2547                                       (unsigned long long)(
2548                                               sect +
2549                                               choose_data_offset(r10_bio, rdev)),
2550                                       bdevname(rdev->bdev, b));
2551                                printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2552                                       "drive\n",
2553                                       mdname(mddev),
2554                                       bdevname(rdev->bdev, b));
2555                                break;
2556                        case 1:
2557                                printk(KERN_INFO
2558                                       "md/raid10:%s: read error corrected"
2559                                       " (%d sectors at %llu on %s)\n",
2560                                       mdname(mddev), s,
2561                                       (unsigned long long)(
2562                                               sect +
2563                                               choose_data_offset(r10_bio, rdev)),
2564                                       bdevname(rdev->bdev, b));
2565                                atomic_add(s, &rdev->corrected_errors);
2566                        }
2567
2568                        rdev_dec_pending(rdev, mddev);
2569                        rcu_read_lock();
2570                }
2571                rcu_read_unlock();
2572
2573                sectors -= s;
2574                sect += s;
2575        }
2576}
2577
2578static int narrow_write_error(struct r10bio *r10_bio, int i)
2579{
2580        struct bio *bio = r10_bio->master_bio;
2581        struct mddev *mddev = r10_bio->mddev;
2582        struct r10conf *conf = mddev->private;
2583        struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2584        /* bio has the data to be written to slot 'i' where
2585         * we just recently had a write error.
2586         * We repeatedly clone the bio and trim down to one block,
2587         * then try the write.  Where the write fails we record
2588         * a bad block.
2589         * It is conceivable that the bio doesn't exactly align with
2590         * blocks.  We must handle this.
2591         *
2592         * We currently own a reference to the rdev.
2593         */
2594
2595        int block_sectors;
2596        sector_t sector;
2597        int sectors;
2598        int sect_to_write = r10_bio->sectors;
2599        int ok = 1;
2600
2601        if (rdev->badblocks.shift < 0)
2602                return 0;
2603
2604        block_sectors = 1 << rdev->badblocks.shift;
2605        sector = r10_bio->sector;
2606        sectors = ((r10_bio->sector + block_sectors)
2607                   & ~(sector_t)(block_sectors - 1))
2608                - sector;
2609
2610        while (sect_to_write) {
2611                struct bio *wbio;
2612                if (sectors > sect_to_write)
2613                        sectors = sect_to_write;
2614                /* Write at 'sector' for 'sectors' */
2615                wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2616                md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2617                wbio->bi_sector = (r10_bio->devs[i].addr+
2618                                   choose_data_offset(r10_bio, rdev) +
2619                                   (sector - r10_bio->sector));
2620                wbio->bi_bdev = rdev->bdev;
2621                if (submit_bio_wait(WRITE, wbio) == 0)
2622                        /* Failure! */
2623                        ok = rdev_set_badblocks(rdev, sector,
2624                                                sectors, 0)
2625                                && ok;
2626
2627                bio_put(wbio);
2628                sect_to_write -= sectors;
2629                sector += sectors;
2630                sectors = block_sectors;
2631        }
2632        return ok;
2633}
2634
2635static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2636{
2637        int slot = r10_bio->read_slot;
2638        struct bio *bio;
2639        struct r10conf *conf = mddev->private;
2640        struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2641        char b[BDEVNAME_SIZE];
2642        unsigned long do_sync;
2643        int max_sectors;
2644
2645        /* we got a read error. Maybe the drive is bad.  Maybe just
2646         * the block and we can fix it.
2647         * We freeze all other IO, and try reading the block from
2648         * other devices.  When we find one, we re-write
2649         * and check it that fixes the read error.
2650         * This is all done synchronously while the array is
2651         * frozen.
2652         */
2653        bio = r10_bio->devs[slot].bio;
2654        bdevname(bio->bi_bdev, b);
2655        bio_put(bio);
2656        r10_bio->devs[slot].bio = NULL;
2657
2658        if (mddev->ro == 0) {
2659                freeze_array(conf, 1);
2660                fix_read_error(conf, mddev, r10_bio);
2661                unfreeze_array(conf);
2662        } else
2663                r10_bio->devs[slot].bio = IO_BLOCKED;
2664
2665        rdev_dec_pending(rdev, mddev);
2666
2667read_more:
2668        rdev = read_balance(conf, r10_bio, &max_sectors);
2669        if (rdev == NULL) {
2670                printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2671                       " read error for block %llu\n",
2672                       mdname(mddev), b,
2673                       (unsigned long long)r10_bio->sector);
2674                raid_end_bio_io(r10_bio);
2675                return;
2676        }
2677
2678        do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2679        slot = r10_bio->read_slot;
2680        printk_ratelimited(
2681                KERN_ERR
2682                "md/raid10:%s: %s: redirecting "
2683                "sector %llu to another mirror\n",
2684                mdname(mddev),
2685                bdevname(rdev->bdev, b),
2686                (unsigned long long)r10_bio->sector);
2687        bio = bio_clone_mddev(r10_bio->master_bio,
2688                              GFP_NOIO, mddev);
2689        md_trim_bio(bio,
2690                    r10_bio->sector - bio->bi_sector,
2691                    max_sectors);
2692        r10_bio->devs[slot].bio = bio;
2693        r10_bio->devs[slot].rdev = rdev;
2694        bio->bi_sector = r10_bio->devs[slot].addr
2695                + choose_data_offset(r10_bio, rdev);
2696        bio->bi_bdev = rdev->bdev;
2697        bio->bi_rw = READ | do_sync;
2698        bio->bi_private = r10_bio;
2699        bio->bi_end_io = raid10_end_read_request;
2700        if (max_sectors < r10_bio->sectors) {
2701                /* Drat - have to split this up more */
2702                struct bio *mbio = r10_bio->master_bio;
2703                int sectors_handled =
2704                        r10_bio->sector + max_sectors
2705                        - mbio->bi_sector;
2706                r10_bio->sectors = max_sectors;
2707                spin_lock_irq(&conf->device_lock);
2708                if (mbio->bi_phys_segments == 0)
2709                        mbio->bi_phys_segments = 2;
2710                else
2711                        mbio->bi_phys_segments++;
2712                spin_unlock_irq(&conf->device_lock);
2713                generic_make_request(bio);
2714
2715                r10_bio = mempool_alloc(conf->r10bio_pool,
2716                                        GFP_NOIO);
2717                r10_bio->master_bio = mbio;
2718                r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2719                r10_bio->state = 0;
2720                set_bit(R10BIO_ReadError,
2721                        &r10_bio->state);
2722                r10_bio->mddev = mddev;
2723                r10_bio->sector = mbio->bi_sector
2724                        + sectors_handled;
2725
2726                goto read_more;
2727        } else
2728                generic_make_request(bio);
2729}
2730
2731static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2732{
2733        /* Some sort of write request has finished and it
2734         * succeeded in writing where we thought there was a
2735         * bad block.  So forget the bad block.
2736         * Or possibly if failed and we need to record
2737         * a bad block.
2738         */
2739        int m;
2740        struct md_rdev *rdev;
2741
2742        if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2743            test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2744                for (m = 0; m < conf->copies; m++) {
2745                        int dev = r10_bio->devs[m].devnum;
2746                        rdev = conf->mirrors[dev].rdev;
2747                        if (r10_bio->devs[m].bio == NULL)
2748                                continue;
2749                        if (test_bit(BIO_UPTODATE,
2750                                     &r10_bio->devs[m].bio->bi_flags)) {
2751                                rdev_clear_badblocks(
2752                                        rdev,
2753                                        r10_bio->devs[m].addr,
2754                                        r10_bio->sectors, 0);
2755                        } else {
2756                                if (!rdev_set_badblocks(
2757                                            rdev,
2758                                            r10_bio->devs[m].addr,
2759                                            r10_bio->sectors, 0))
2760                                        md_error(conf->mddev, rdev);
2761                        }
2762                        rdev = conf->mirrors[dev].replacement;
2763                        if (r10_bio->devs[m].repl_bio == NULL)
2764                                continue;
2765                        if (test_bit(BIO_UPTODATE,
2766                                     &r10_bio->devs[m].repl_bio->bi_flags)) {
2767                                rdev_clear_badblocks(
2768                                        rdev,
2769                                        r10_bio->devs[m].addr,
2770                                        r10_bio->sectors, 0);
2771                        } else {
2772                                if (!rdev_set_badblocks(
2773                                            rdev,
2774                                            r10_bio->devs[m].addr,
2775                                            r10_bio->sectors, 0))
2776                                        md_error(conf->mddev, rdev);
2777                        }
2778                }
2779                put_buf(r10_bio);
2780        } else {
2781                for (m = 0; m < conf->copies; m++) {
2782                        int dev = r10_bio->devs[m].devnum;
2783                        struct bio *bio = r10_bio->devs[m].bio;
2784                        rdev = conf->mirrors[dev].rdev;
2785                        if (bio == IO_MADE_GOOD) {
2786                                rdev_clear_badblocks(
2787                                        rdev,
2788                                        r10_bio->devs[m].addr,
2789                                        r10_bio->sectors, 0);
2790                                rdev_dec_pending(rdev, conf->mddev);
2791                        } else if (bio != NULL &&
2792                                   !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2793                                if (!narrow_write_error(r10_bio, m)) {
2794                                        md_error(conf->mddev, rdev);
2795                                        set_bit(R10BIO_Degraded,
2796                                                &r10_bio->state);
2797                                }
2798                                rdev_dec_pending(rdev, conf->mddev);
2799                        }
2800                        bio = r10_bio->devs[m].repl_bio;
2801                        rdev = conf->mirrors[dev].replacement;
2802                        if (rdev && bio == IO_MADE_GOOD) {
2803                                rdev_clear_badblocks(
2804                                        rdev,
2805                                        r10_bio->devs[m].addr,
2806                                        r10_bio->sectors, 0);
2807                                rdev_dec_pending(rdev, conf->mddev);
2808                        }
2809                }
2810                if (test_bit(R10BIO_WriteError,
2811                             &r10_bio->state))
2812                        close_write(r10_bio);
2813                raid_end_bio_io(r10_bio);
2814        }
2815}
2816
2817static void raid10d(struct md_thread *thread)
2818{
2819        struct mddev *mddev = thread->mddev;
2820        struct r10bio *r10_bio;
2821        unsigned long flags;
2822        struct r10conf *conf = mddev->private;
2823        struct list_head *head = &conf->retry_list;
2824        struct blk_plug plug;
2825
2826        md_check_recovery(mddev);
2827
2828        blk_start_plug(&plug);
2829        for (;;) {
2830
2831                flush_pending_writes(conf);
2832
2833                spin_lock_irqsave(&conf->device_lock, flags);
2834                if (list_empty(head)) {
2835                        spin_unlock_irqrestore(&conf->device_lock, flags);
2836                        break;
2837                }
2838                r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2839                list_del(head->prev);
2840                conf->nr_queued--;
2841                spin_unlock_irqrestore(&conf->device_lock, flags);
2842
2843                mddev = r10_bio->mddev;
2844                conf = mddev->private;
2845                if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2846                    test_bit(R10BIO_WriteError, &r10_bio->state))
2847                        handle_write_completed(conf, r10_bio);
2848                else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2849                        reshape_request_write(mddev, r10_bio);
2850                else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2851                        sync_request_write(mddev, r10_bio);
2852                else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2853                        recovery_request_write(mddev, r10_bio);
2854                else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2855                        handle_read_error(mddev, r10_bio);
2856                else {
2857                        /* just a partial read to be scheduled from a
2858                         * separate context
2859                         */
2860                        int slot = r10_bio->read_slot;
2861                        generic_make_request(r10_bio->devs[slot].bio);
2862                }
2863
2864                cond_resched();
2865                if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2866                        md_check_recovery(mddev);
2867        }
2868        blk_finish_plug(&plug);
2869}
2870
2871
2872static int init_resync(struct r10conf *conf)
2873{
2874        int buffs;
2875        int i;
2876
2877        buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2878        BUG_ON(conf->r10buf_pool);
2879        conf->have_replacement = 0;
2880        for (i = 0; i < conf->geo.raid_disks; i++)
2881                if (conf->mirrors[i].replacement)
2882                        conf->have_replacement = 1;
2883        conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2884        if (!conf->r10buf_pool)
2885                return -ENOMEM;
2886        conf->next_resync = 0;
2887        return 0;
2888}
2889
2890/*
2891 * perform a "sync" on one "block"
2892 *
2893 * We need to make sure that no normal I/O request - particularly write
2894 * requests - conflict with active sync requests.
2895 *
2896 * This is achieved by tracking pending requests and a 'barrier' concept
2897 * that can be installed to exclude normal IO requests.
2898 *
2899 * Resync and recovery are handled very differently.
2900 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2901 *
2902 * For resync, we iterate over virtual addresses, read all copies,
2903 * and update if there are differences.  If only one copy is live,
2904 * skip it.
2905 * For recovery, we iterate over physical addresses, read a good
2906 * value for each non-in_sync drive, and over-write.
2907 *
2908 * So, for recovery we may have several outstanding complex requests for a
2909 * given address, one for each out-of-sync device.  We model this by allocating
2910 * a number of r10_bio structures, one for each out-of-sync device.
2911 * As we setup these structures, we collect all bio's together into a list
2912 * which we then process collectively to add pages, and then process again
2913 * to pass to generic_make_request.
2914 *
2915 * The r10_bio structures are linked using a borrowed master_bio pointer.
2916 * This link is counted in ->remaining.  When the r10_bio that points to NULL
2917 * has its remaining count decremented to 0, the whole complex operation
2918 * is complete.
2919 *
2920 */
2921
2922static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2923                             int *skipped, int go_faster)
2924{
2925        struct r10conf *conf = mddev->private;
2926        struct r10bio *r10_bio;
2927        struct bio *biolist = NULL, *bio;
2928        sector_t max_sector, nr_sectors;
2929        int i;
2930        int max_sync;
2931        sector_t sync_blocks;
2932        sector_t sectors_skipped = 0;
2933        int chunks_skipped = 0;
2934        sector_t chunk_mask = conf->geo.chunk_mask;
2935
2936        if (!conf->r10buf_pool)
2937                if (init_resync(conf))
2938                        return 0;
2939
2940        /*
2941         * Allow skipping a full rebuild for incremental assembly
2942         * of a clean array, like RAID1 does.
2943         */
2944        if (mddev->bitmap == NULL &&
2945            mddev->recovery_cp == MaxSector &&
2946            mddev->reshape_position == MaxSector &&
2947            !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2948            !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2949            !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2950            conf->fullsync == 0) {
2951                *skipped = 1;
2952                return mddev->dev_sectors - sector_nr;
2953        }
2954
2955 skipped:
2956        max_sector = mddev->dev_sectors;
2957        if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2958            test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2959                max_sector = mddev->resync_max_sectors;
2960        if (sector_nr >= max_sector) {
2961                /* If we aborted, we need to abort the
2962                 * sync on the 'current' bitmap chucks (there can
2963                 * be several when recovering multiple devices).
2964                 * as we may have started syncing it but not finished.
2965                 * We can find the current address in
2966                 * mddev->curr_resync, but for recovery,
2967                 * we need to convert that to several
2968                 * virtual addresses.
2969                 */
2970                if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2971                        end_reshape(conf);
2972                        return 0;
2973                }
2974
2975                if (mddev->curr_resync < max_sector) { /* aborted */
2976                        if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2977                                bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2978                                                &sync_blocks, 1);
2979                        else for (i = 0; i < conf->geo.raid_disks; i++) {
2980                                sector_t sect =
2981                                        raid10_find_virt(conf, mddev->curr_resync, i);
2982                                bitmap_end_sync(mddev->bitmap, sect,
2983                                                &sync_blocks, 1);
2984                        }
2985                } else {
2986                        /* completed sync */
2987                        if ((!mddev->bitmap || conf->fullsync)
2988                            && conf->have_replacement
2989                            && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2990                                /* Completed a full sync so the replacements
2991                                 * are now fully recovered.
2992                                 */
2993                                for (i = 0; i < conf->geo.raid_disks; i++)
2994                                        if (conf->mirrors[i].replacement)
2995                                                conf->mirrors[i].replacement
2996                                                        ->recovery_offset
2997                                                        = MaxSector;
2998                        }
2999                        conf->fullsync = 0;
3000                }
3001                bitmap_close_sync(mddev->bitmap);
3002                close_sync(conf);
3003                *skipped = 1;
3004                return sectors_skipped;
3005        }
3006
3007        if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3008                return reshape_request(mddev, sector_nr, skipped);
3009
3010        if (chunks_skipped >= conf->geo.raid_disks) {
3011                /* if there has been nothing to do on any drive,
3012                 * then there is nothing to do at all..
3013                 */
3014                *skipped = 1;
3015                return (max_sector - sector_nr) + sectors_skipped;
3016        }
3017
3018        if (max_sector > mddev->resync_max)
3019                max_sector = mddev->resync_max; /* Don't do IO beyond here */
3020
3021        /* make sure whole request will fit in a chunk - if chunks
3022         * are meaningful
3023         */
3024        if (conf->geo.near_copies < conf->geo.raid_disks &&
3025            max_sector > (sector_nr | chunk_mask))
3026                max_sector = (sector_nr | chunk_mask) + 1;
3027        /*
3028         * If there is non-resync activity waiting for us then
3029         * put in a delay to throttle resync.
3030         */
3031        if (!go_faster && conf->nr_waiting)
3032                msleep_interruptible(1000);
3033
3034        /* Again, very different code for resync and recovery.
3035         * Both must result in an r10bio with a list of bios that
3036         * have bi_end_io, bi_sector, bi_bdev set,
3037         * and bi_private set to the r10bio.
3038         * For recovery, we may actually create several r10bios
3039         * with 2 bios in each, that correspond to the bios in the main one.
3040         * In this case, the subordinate r10bios link back through a
3041         * borrowed master_bio pointer, and the counter in the master
3042         * includes a ref from each subordinate.
3043         */
3044        /* First, we decide what to do and set ->bi_end_io
3045         * To end_sync_read if we want to read, and
3046         * end_sync_write if we will want to write.
3047         */
3048
3049        max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3050        if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3051                /* recovery... the complicated one */
3052                int j;
3053                r10_bio = NULL;
3054
3055                for (i = 0 ; i < conf->geo.raid_disks; i++) {
3056                        int still_degraded;
3057                        struct r10bio *rb2;
3058                        sector_t sect;
3059                        int must_sync;
3060                        int any_working;
3061                        struct raid10_info *mirror = &conf->mirrors[i];
3062
3063                        if ((mirror->rdev == NULL ||
3064                             test_bit(In_sync, &mirror->rdev->flags))
3065                            &&
3066                            (mirror->replacement == NULL ||
3067                             test_bit(Faulty,
3068                                      &mirror->replacement->flags)))
3069                                continue;
3070
3071                        still_degraded = 0;
3072                        /* want to reconstruct this device */
3073                        rb2 = r10_bio;
3074                        sect = raid10_find_virt(conf, sector_nr, i);
3075                        if (sect >= mddev->resync_max_sectors) {
3076                                /* last stripe is not complete - don't
3077                                 * try to recover this sector.
3078                                 */
3079                                continue;
3080                        }
3081                        /* Unless we are doing a full sync, or a replacement
3082                         * we only need to recover the block if it is set in
3083                         * the bitmap
3084                         */
3085                        must_sync = bitmap_start_sync(mddev->bitmap, sect,
3086                                                      &sync_blocks, 1);
3087                        if (sync_blocks < max_sync)
3088                                max_sync = sync_blocks;
3089                        if (!must_sync &&
3090                            mirror->replacement == NULL &&
3091                            !conf->fullsync) {
3092                                /* yep, skip the sync_blocks here, but don't assume
3093                                 * that there will never be anything to do here
3094                                 */
3095                                chunks_skipped = -1;
3096                                continue;
3097                        }
3098
3099                        r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3100                        raise_barrier(conf, rb2 != NULL);
3101                        atomic_set(&r10_bio->remaining, 0);
3102
3103                        r10_bio->master_bio = (struct bio*)rb2;
3104                        if (rb2)
3105                                atomic_inc(&rb2->remaining);
3106                        r10_bio->mddev = mddev;
3107                        set_bit(R10BIO_IsRecover, &r10_bio->state);
3108                        r10_bio->sector = sect;
3109
3110                        raid10_find_phys(conf, r10_bio);
3111
3112                        /* Need to check if the array will still be
3113                         * degraded
3114                         */
3115                        for (j = 0; j < conf->geo.raid_disks; j++)
3116                                if (conf->mirrors[j].rdev == NULL ||
3117                                    test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
3118                                        still_degraded = 1;
3119                                        break;
3120                                }
3121
3122                        must_sync = bitmap_start_sync(mddev->bitmap, sect,
3123                                                      &sync_blocks, still_degraded);
3124
3125                        any_working = 0;
3126                        for (j=0; j<conf->copies;j++) {
3127                                int k;
3128                                int d = r10_bio->devs[j].devnum;
3129                                sector_t from_addr, to_addr;
3130                                struct md_rdev *rdev;
3131                                sector_t sector, first_bad;
3132                                int bad_sectors;
3133                                if (!conf->mirrors[d].rdev ||
3134                                    !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3135                                        continue;
3136                                /* This is where we read from */
3137                                any_working = 1;
3138                                rdev = conf->mirrors[d].rdev;
3139                                sector = r10_bio->devs[j].addr;
3140
3141                                if (is_badblock(rdev, sector, max_sync,
3142                                                &first_bad, &bad_sectors)) {
3143                                        if (first_bad > sector)
3144                                                max_sync = first_bad - sector;
3145                                        else {
3146                                                bad_sectors -= (sector
3147                                                                - first_bad);
3148                                                if (max_sync > bad_sectors)
3149                                                        max_sync = bad_sectors;
3150                                                continue;
3151                                        }
3152                                }
3153                                bio = r10_bio->devs[0].bio;
3154                                bio_reset(bio);
3155                                bio->bi_next = biolist;
3156                                biolist = bio;
3157                                bio->bi_private = r10_bio;
3158                                bio->bi_end_io = end_sync_read;
3159                                bio->bi_rw = READ;
3160                                from_addr = r10_bio->devs[j].addr;
3161                                bio->bi_sector = from_addr + rdev->data_offset;
3162                                bio->bi_bdev = rdev->bdev;
3163                                atomic_inc(&rdev->nr_pending);
3164                                /* and we write to 'i' (if not in_sync) */
3165
3166                                for (k=0; k<conf->copies; k++)
3167                                        if (r10_bio->devs[k].devnum == i)
3168                                                break;
3169                                BUG_ON(k == conf->copies);
3170                                to_addr = r10_bio->devs[k].addr;
3171                                r10_bio->devs[0].devnum = d;
3172                                r10_bio->devs[0].addr = from_addr;
3173                                r10_bio->devs[1].devnum = i;
3174                                r10_bio->devs[1].addr = to_addr;
3175
3176                                rdev = mirror->rdev;
3177                                if (!test_bit(In_sync, &rdev->flags)) {
3178                                        bio = r10_bio->devs[1].bio;
3179                                        bio_reset(bio);
3180                                        bio->bi_next = biolist;
3181                                        biolist = bio;
3182                                        bio->bi_private = r10_bio;
3183                                        bio->bi_end_io = end_sync_write;
3184                                        bio->bi_rw = WRITE;
3185                                        bio->bi_sector = to_addr
3186                                                + rdev->data_offset;
3187                                        bio->bi_bdev = rdev->bdev;
3188                                        atomic_inc(&r10_bio->remaining);
3189                                } else
3190                                        r10_bio->devs[1].bio->bi_end_io = NULL;
3191
3192                                /* and maybe write to replacement */
3193                                bio = r10_bio->devs[1].repl_bio;
3194                                if (bio)
3195                                        bio->bi_end_io = NULL;
3196                                rdev = mirror->replacement;
3197                                /* Note: if rdev != NULL, then bio
3198                                 * cannot be NULL as r10buf_pool_alloc will
3199                                 * have allocated it.
3200                                 * So the second test here is pointless.
3201                                 * But it keeps semantic-checkers happy, and
3202                                 * this comment keeps human reviewers
3203                                 * happy.
3204                                 */
3205                                if (rdev == NULL || bio == NULL ||
3206                                    test_bit(Faulty, &rdev->flags))
3207                                        break;
3208                                bio_reset(bio);
3209                                bio->bi_next = biolist;
3210                                biolist = bio;
3211                                bio->bi_private = r10_bio;
3212                                bio->bi_end_io = end_sync_write;
3213                                bio->bi_rw = WRITE;
3214                                bio->bi_sector = to_addr + rdev->data_offset;
3215                                bio->bi_bdev = rdev->bdev;
3216                                atomic_inc(&r10_bio->remaining);
3217                                break;
3218                        }
3219                        if (j == conf->copies) {
3220                                /* Cannot recover, so abort the recovery or
3221                                 * record a bad block */
3222                                put_buf(r10_bio);
3223                                if (rb2)
3224                                        atomic_dec(&rb2->remaining);
3225                                r10_bio = rb2;
3226                                if (any_working) {
3227                                        /* problem is that there are bad blocks
3228                                         * on other device(s)
3229                                         */
3230                                        int k;
3231                                        for (k = 0; k < conf->copies; k++)
3232                                                if (r10_bio->devs[k].devnum == i)
3233                                                        break;
3234                                        if (!test_bit(In_sync,
3235                                                      &mirror->rdev->flags)
3236                                            && !rdev_set_badblocks(
3237                                                    mirror->rdev,
3238                                                    r10_bio->devs[k].addr,
3239                                                    max_sync, 0))
3240                                                any_working = 0;
3241                                        if (mirror->replacement &&
3242                                            !rdev_set_badblocks(
3243                                                    mirror->replacement,
3244                                                    r10_bio->devs[k].addr,
3245                                                    max_sync, 0))
3246                                                any_working = 0;
3247                                }
3248                                if (!any_working)  {
3249                                        if (!test_and_set_bit(MD_RECOVERY_INTR,
3250                                                              &mddev->recovery))
3251                                                printk(KERN_INFO "md/raid10:%s: insufficient "
3252                                                       "working devices for recovery.\n",
3253                                                       mdname(mddev));
3254                                        mirror->recovery_disabled
3255                                                = mddev->recovery_disabled;
3256                                }
3257                                break;
3258                        }
3259                }
3260                if (biolist == NULL) {
3261                        while (r10_bio) {
3262                                struct r10bio *rb2 = r10_bio;
3263                                r10_bio = (struct r10bio*) rb2->master_bio;
3264                                rb2->master_bio = NULL;
3265                                put_buf(rb2);
3266                        }
3267                        goto giveup;
3268                }
3269        } else {
3270                /* resync. Schedule a read for every block at this virt offset */
3271                int count = 0;
3272
3273                bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3274
3275                if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3276                                       &sync_blocks, mddev->degraded) &&
3277                    !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3278                                                 &mddev->recovery)) {
3279                        /* We can skip this block */
3280                        *skipped = 1;
3281                        return sync_blocks + sectors_skipped;
3282                }
3283                if (sync_blocks < max_sync)
3284                        max_sync = sync_blocks;
3285                r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3286
3287                r10_bio->mddev = mddev;
3288                atomic_set(&r10_bio->remaining, 0);
3289                raise_barrier(conf, 0);
3290                conf->next_resync = sector_nr;
3291
3292                r10_bio->master_bio = NULL;
3293                r10_bio->sector = sector_nr;
3294                set_bit(R10BIO_IsSync, &r10_bio->state);
3295                raid10_find_phys(conf, r10_bio);
3296                r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3297
3298                for (i = 0; i < conf->copies; i++) {
3299                        int d = r10_bio->devs[i].devnum;
3300                        sector_t first_bad, sector;
3301                        int bad_sectors;
3302
3303                        if (r10_bio->devs[i].repl_bio)
3304                                r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3305
3306                        bio = r10_bio->devs[i].bio;
3307                        bio_reset(bio);
3308                        clear_bit(BIO_UPTODATE, &bio->bi_flags);
3309                        if (conf->mirrors[d].rdev == NULL ||
3310                            test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3311                                continue;
3312                        sector = r10_bio->devs[i].addr;
3313                        if (is_badblock(conf->mirrors[d].rdev,
3314                                        sector, max_sync,
3315                                        &first_bad, &bad_sectors)) {
3316                                if (first_bad > sector)
3317                                        max_sync = first_bad - sector;
3318                                else {
3319                                        bad_sectors -= (sector - first_bad);
3320                                        if (max_sync > bad_sectors)
3321                                                max_sync = bad_sectors;
3322                                        continue;
3323                                }
3324                        }
3325                        atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3326                        atomic_inc(&r10_bio->remaining);
3327                        bio->bi_next = biolist;
3328                        biolist = bio;
3329                        bio->bi_private = r10_bio;
3330                        bio->bi_end_io = end_sync_read;
3331                        bio->bi_rw = READ;
3332                        bio->bi_sector = sector +
3333                                conf->mirrors[d].rdev->data_offset;
3334                        bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3335                        count++;
3336
3337                        if (conf->mirrors[d].replacement == NULL ||
3338                            test_bit(Faulty,
3339                                     &conf->mirrors[d].replacement->flags))
3340                                continue;
3341
3342                        /* Need to set up for writing to the replacement */
3343                        bio = r10_bio->devs[i].repl_bio;
3344                        bio_reset(bio);
3345                        clear_bit(BIO_UPTODATE, &bio->bi_flags);
3346
3347                        sector = r10_bio->devs[i].addr;
3348                        atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3349                        bio->bi_next = biolist;
3350                        biolist = bio;
3351                        bio->bi_private = r10_bio;
3352                        bio->bi_end_io = end_sync_write;
3353                        bio->bi_rw = WRITE;
3354                        bio->bi_sector = sector +
3355                                conf->mirrors[d].replacement->data_offset;
3356                        bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3357                        count++;
3358                }
3359
3360                if (count < 2) {
3361                        for (i=0; i<conf->copies; i++) {
3362                                int d = r10_bio->devs[i].devnum;
3363                                if (r10_bio->devs[i].bio->bi_end_io)
3364                                        rdev_dec_pending(conf->mirrors[d].rdev,
3365                                                         mddev);
3366                                if (r10_bio->devs[i].repl_bio &&
3367                                    r10_bio->devs[i].repl_bio->bi_end_io)
3368                                        rdev_dec_pending(
3369                                                conf->mirrors[d].replacement,
3370                                                mddev);
3371                        }
3372                        put_buf(r10_bio);
3373                        biolist = NULL;
3374                        goto giveup;
3375                }
3376        }
3377
3378        nr_sectors = 0;
3379        if (sector_nr + max_sync < max_sector)
3380                max_sector = sector_nr + max_sync;
3381        do {
3382                struct page *page;
3383                int len = PAGE_SIZE;
3384                if (sector_nr + (len>>9) > max_sector)
3385                        len = (max_sector - sector_nr) << 9;
3386                if (len == 0)
3387                        break;
3388                for (bio= biolist ; bio ; bio=bio->bi_next) {
3389                        struct bio *bio2;
3390                        page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3391                        if (bio_add_page(bio, page, len, 0))
3392                                continue;
3393
3394                        /* stop here */
3395                        bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3396                        for (bio2 = biolist;
3397                             bio2 && bio2 != bio;
3398                             bio2 = bio2->bi_next) {
3399                                /* remove last page from this bio */
3400                                bio2->bi_vcnt--;
3401                                bio2->bi_size -= len;
3402                                bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3403                        }
3404                        goto bio_full;
3405                }
3406                nr_sectors += len>>9;
3407                sector_nr += len>>9;
3408        } while (biolist->bi_vcnt < RESYNC_PAGES);
3409 bio_full:
3410        r10_bio->sectors = nr_sectors;
3411
3412        while (biolist) {
3413                bio = biolist;
3414                biolist = biolist->bi_next;
3415
3416                bio->bi_next = NULL;
3417                r10_bio = bio->bi_private;
3418                r10_bio->sectors = nr_sectors;
3419
3420                if (bio->bi_end_io == end_sync_read) {
3421                        md_sync_acct(bio->bi_bdev, nr_sectors);
3422                        set_bit(BIO_UPTODATE, &bio->bi_flags);
3423                        generic_make_request(bio);
3424                }
3425        }
3426
3427        if (sectors_skipped)
3428                /* pretend they weren't skipped, it makes
3429                 * no important difference in this case
3430                 */
3431                md_done_sync(mddev, sectors_skipped, 1);
3432
3433        return sectors_skipped + nr_sectors;
3434 giveup:
3435        /* There is nowhere to write, so all non-sync
3436         * drives must be failed or in resync, all drives
3437         * have a bad block, so try the next chunk...
3438         */
3439        if (sector_nr + max_sync < max_sector)
3440                max_sector = sector_nr + max_sync;
3441
3442        sectors_skipped += (max_sector - sector_nr);
3443        chunks_skipped ++;
3444        sector_nr = max_sector;
3445        goto skipped;
3446}
3447
3448static sector_t
3449raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3450{
3451        sector_t size;
3452        struct r10conf *conf = mddev->private;
3453
3454        if (!raid_disks)
3455                raid_disks = min(conf->geo.raid_disks,
3456                                 conf->prev.raid_disks);
3457        if (!sectors)
3458                sectors = conf->dev_sectors;
3459
3460        size = sectors >> conf->geo.chunk_shift;
3461        sector_div(size, conf->geo.far_copies);
3462        size = size * raid_disks;
3463        sector_div(size, conf->geo.near_copies);
3464
3465        return size << conf->geo.chunk_shift;
3466}
3467
3468static void calc_sectors(struct r10conf *conf, sector_t size)
3469{
3470        /* Calculate the number of sectors-per-device that will
3471         * actually be used, and set conf->dev_sectors and
3472         * conf->stride
3473         */
3474
3475        size = size >> conf->geo.chunk_shift;
3476        sector_div(size, conf->geo.far_copies);
3477        size = size * conf->geo.raid_disks;
3478        sector_div(size, conf->geo.near_copies);
3479        /* 'size' is now the number of chunks in the array */
3480        /* calculate "used chunks per device" */
3481        size = size * conf->copies;
3482
3483        /* We need to round up when dividing by raid_disks to
3484         * get the stride size.
3485         */
3486        size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3487
3488        conf->dev_sectors = size << conf->geo.chunk_shift;
3489
3490        if (conf->geo.far_offset)
3491                conf->geo.stride = 1 << conf->geo.chunk_shift;
3492        else {
3493                sector_div(size, conf->geo.far_copies);
3494                conf->geo.stride = size << conf->geo.chunk_shift;
3495        }
3496}
3497
3498enum geo_type {geo_new, geo_old, geo_start};
3499static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3500{
3501        int nc, fc, fo;
3502        int layout, chunk, disks;
3503        switch (new) {
3504        case geo_old:
3505                layout = mddev->layout;
3506                chunk = mddev->chunk_sectors;
3507                disks = mddev->raid_disks - mddev->delta_disks;
3508                break;
3509        case geo_new:
3510                layout = mddev->new_layout;
3511                chunk = mddev->new_chunk_sectors;
3512                disks = mddev->raid_disks;
3513                break;
3514        default: /* avoid 'may be unused' warnings */
3515        case geo_start: /* new when starting reshape - raid_disks not
3516                         * updated yet. */
3517                layout = mddev->new_layout;
3518                chunk = mddev->new_chunk_sectors;
3519                disks = mddev->raid_disks + mddev->delta_disks;
3520                break;
3521        }
3522        if (layout >> 18)
3523                return -1;
3524        if (chunk < (PAGE_SIZE >> 9) ||
3525            !is_power_of_2(chunk))
3526                return -2;
3527        nc = layout & 255;
3528        fc = (layout >> 8) & 255;
3529        fo = layout & (1<<16);
3530        geo->raid_disks = disks;
3531        geo->near_copies = nc;
3532        geo->far_copies = fc;
3533        geo->far_offset = fo;
3534        geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks;
3535        geo->chunk_mask = chunk - 1;
3536        geo->chunk_shift = ffz(~chunk);
3537        return nc*fc;
3538}
3539
3540static struct r10conf *setup_conf(struct mddev *mddev)
3541{
3542        struct r10conf *conf = NULL;
3543        int err = -EINVAL;
3544        struct geom geo;
3545        int copies;
3546
3547        copies = setup_geo(&geo, mddev, geo_new);
3548
3549        if (copies == -2) {
3550                printk(KERN_ERR "md/raid10:%s: chunk size must be "
3551                       "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3552                       mdname(mddev), PAGE_SIZE);
3553                goto out;
3554        }
3555
3556        if (copies < 2 || copies > mddev->raid_disks) {
3557                printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3558                       mdname(mddev), mddev->new_layout);
3559                goto out;
3560        }
3561
3562        err = -ENOMEM;
3563        conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3564        if (!conf)
3565                goto out;
3566
3567        /* FIXME calc properly */
3568        conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3569                                                            max(0,-mddev->delta_disks)),
3570                                GFP_KERNEL);
3571        if (!conf->mirrors)
3572                goto out;
3573
3574        conf->tmppage = alloc_page(GFP_KERNEL);
3575        if (!conf->tmppage)
3576                goto out;
3577
3578        conf->geo = geo;
3579        conf->copies = copies;
3580        conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3581                                           r10bio_pool_free, conf);
3582        if (!conf->r10bio_pool)
3583                goto out;
3584
3585        calc_sectors(conf, mddev->dev_sectors);
3586        if (mddev->reshape_position == MaxSector) {
3587                conf->prev = conf->geo;
3588                conf->reshape_progress = MaxSector;
3589        } else {
3590                if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3591                        err = -EINVAL;
3592                        goto out;
3593                }
3594                conf->reshape_progress = mddev->reshape_position;
3595                if (conf->prev.far_offset)
3596                        conf->prev.stride = 1 << conf->prev.chunk_shift;
3597                else
3598                        /* far_copies must be 1 */
3599                        conf->prev.stride = conf->dev_sectors;
3600        }
3601        spin_lock_init(&conf->device_lock);
3602        INIT_LIST_HEAD(&conf->retry_list);
3603
3604        spin_lock_init(&conf->resync_lock);
3605        init_waitqueue_head(&conf->wait_barrier);
3606
3607        conf->thread = md_register_thread(raid10d, mddev, "raid10");
3608        if (!conf->thread)
3609                goto out;
3610
3611        conf->mddev = mddev;
3612        return conf;
3613
3614 out:
3615        if (err == -ENOMEM)
3616                printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3617                       mdname(mddev));
3618        if (conf) {
3619                if (conf->r10bio_pool)
3620                        mempool_destroy(conf->r10bio_pool);
3621                kfree(conf->mirrors);
3622                safe_put_page(conf->tmppage);
3623                kfree(conf);
3624        }
3625        return ERR_PTR(err);
3626}
3627
3628static int run(struct mddev *mddev)
3629{
3630        struct r10conf *conf;
3631        int i, disk_idx, chunk_size;
3632        struct raid10_info *disk;
3633        struct md_rdev *rdev;
3634        sector_t size;
3635        sector_t min_offset_diff = 0;
3636        int first = 1;
3637        bool discard_supported = false;
3638
3639        if (mddev->private == NULL) {
3640                conf = setup_conf(mddev);
3641                if (IS_ERR(conf))
3642                        return PTR_ERR(conf);
3643                mddev->private = conf;
3644        }
3645        conf = mddev->private;
3646        if (!conf)
3647                goto out;
3648
3649        mddev->thread = conf->thread;
3650        conf->thread = NULL;
3651
3652        chunk_size = mddev->chunk_sectors << 9;
3653        if (mddev->queue) {
3654                blk_queue_max_discard_sectors(mddev->queue,
3655                                              mddev->chunk_sectors);
3656                blk_queue_max_write_same_sectors(mddev->queue, 0);
3657                blk_queue_io_min(mddev->queue, chunk_size);
3658                if (conf->geo.raid_disks % conf->geo.near_copies)
3659                        blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3660                else
3661                        blk_queue_io_opt(mddev->queue, chunk_size *
3662                                         (conf->geo.raid_disks / conf->geo.near_copies));
3663        }
3664
3665        rdev_for_each(rdev, mddev) {
3666                long long diff;
3667                struct request_queue *q;
3668
3669                disk_idx = rdev->raid_disk;
3670                if (disk_idx < 0)
3671                        continue;
3672                if (disk_idx >= conf->geo.raid_disks &&
3673                    disk_idx >= conf->prev.raid_disks)
3674                        continue;
3675                disk = conf->mirrors + disk_idx;
3676
3677                if (test_bit(Replacement, &rdev->flags)) {
3678                        if (disk->replacement)
3679                                goto out_free_conf;
3680                        disk->replacement = rdev;
3681                } else {
3682                        if (disk->rdev)
3683                                goto out_free_conf;
3684                        disk->rdev = rdev;
3685                }
3686                q = bdev_get_queue(rdev->bdev);
3687                if (q->merge_bvec_fn)
3688                        mddev->merge_check_needed = 1;
3689                diff = (rdev->new_data_offset - rdev->data_offset);
3690                if (!mddev->reshape_backwards)
3691                        diff = -diff;
3692                if (diff < 0)
3693                        diff = 0;
3694                if (first || diff < min_offset_diff)
3695                        min_offset_diff = diff;
3696
3697                if (mddev->gendisk)
3698                        disk_stack_limits(mddev->gendisk, rdev->bdev,
3699                                          rdev->data_offset << 9);
3700
3701                disk->head_position = 0;
3702
3703                if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3704                        discard_supported = true;
3705        }
3706
3707        if (mddev->queue) {
3708                if (discard_supported)
3709                        queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3710                                                mddev->queue);
3711                else
3712                        queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3713                                                  mddev->queue);
3714        }
3715        /* need to check that every block has at least one working mirror */
3716        if (!enough(conf, -1)) {
3717                printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3718                       mdname(mddev));
3719                goto out_free_conf;
3720        }
3721
3722        if (conf->reshape_progress != MaxSector) {
3723                /* must ensure that shape change is supported */
3724                if (conf->geo.far_copies != 1 &&
3725                    conf->geo.far_offset == 0)
3726                        goto out_free_conf;
3727                if (conf->prev.far_copies != 1 &&
3728                    conf->prev.far_offset == 0)
3729                        goto out_free_conf;
3730        }
3731
3732        mddev->degraded = 0;
3733        for (i = 0;
3734             i < conf->geo.raid_disks
3735                     || i < conf->prev.raid_disks;
3736             i++) {
3737
3738                disk = conf->mirrors + i;
3739
3740                if (!disk->rdev && disk->replacement) {
3741                        /* The replacement is all we have - use it */
3742                        disk->rdev = disk->replacement;
3743                        disk->replacement = NULL;
3744                        clear_bit(Replacement, &disk->rdev->flags);
3745                }
3746
3747                if (!disk->rdev ||
3748                    !test_bit(In_sync, &disk->rdev->flags)) {
3749                        disk->head_position = 0;
3750                        mddev->degraded++;
3751                        if (disk->rdev)
3752                                conf->fullsync = 1;
3753                }
3754                disk->recovery_disabled = mddev->recovery_disabled - 1;
3755        }
3756
3757        if (mddev->recovery_cp != MaxSector)
3758                printk(KERN_NOTICE "md/raid10:%s: not clean"
3759                       " -- starting background reconstruction\n",
3760                       mdname(mddev));
3761        printk(KERN_INFO
3762                "md/raid10:%s: active with %d out of %d devices\n",
3763                mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3764                conf->geo.raid_disks);
3765        /*
3766         * Ok, everything is just fine now
3767         */
3768        mddev->dev_sectors = conf->dev_sectors;
3769        size = raid10_size(mddev, 0, 0);
3770        md_set_array_sectors(mddev, size);
3771        mddev->resync_max_sectors = size;
3772
3773        if (mddev->queue) {
3774                int stripe = conf->geo.raid_disks *
3775                        ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3776                mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3777                mddev->queue->backing_dev_info.congested_data = mddev;
3778
3779                /* Calculate max read-ahead size.
3780                 * We need to readahead at least twice a whole stripe....
3781                 * maybe...
3782                 */
3783                stripe /= conf->geo.near_copies;
3784                if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3785                        mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3786                blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3787        }
3788
3789
3790        if (md_integrity_register(mddev))
3791                goto out_free_conf;
3792
3793        if (conf->reshape_progress != MaxSector) {
3794                unsigned long before_length, after_length;
3795
3796                before_length = ((1 << conf->prev.chunk_shift) *
3797                                 conf->prev.far_copies);
3798                after_length = ((1 << conf->geo.chunk_shift) *
3799                                conf->geo.far_copies);
3800
3801                if (max(before_length, after_length) > min_offset_diff) {
3802                        /* This cannot work */
3803                        printk("md/raid10: offset difference not enough to continue reshape\n");
3804                        goto out_free_conf;
3805                }
3806                conf->offset_diff = min_offset_diff;
3807
3808                conf->reshape_safe = conf->reshape_progress;
3809                clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3810                clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3811                set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3812                set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3813                mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3814                                                        "reshape");
3815        }
3816
3817        return 0;
3818
3819out_free_conf:
3820        md_unregister_thread(&mddev->thread);
3821        if (conf->r10bio_pool)
3822                mempool_destroy(conf->r10bio_pool);
3823        safe_put_page(conf->tmppage);
3824        kfree(conf->mirrors);
3825        kfree(conf);
3826        mddev->private = NULL;
3827out:
3828        return -EIO;
3829}
3830
3831static int stop(struct mddev *mddev)
3832{
3833        struct r10conf *conf = mddev->private;
3834
3835        raise_barrier(conf, 0);
3836        lower_barrier(conf);
3837
3838        md_unregister_thread(&mddev->thread);
3839        if (mddev->queue)
3840                /* the unplug fn references 'conf'*/
3841                blk_sync_queue(mddev->queue);
3842
3843        if (conf->r10bio_pool)
3844                mempool_destroy(conf->r10bio_pool);
3845        safe_put_page(conf->tmppage);
3846        kfree(conf->mirrors);
3847        kfree(conf);
3848        mddev->private = NULL;
3849        return 0;
3850}
3851
3852static void raid10_quiesce(struct mddev *mddev, int state)
3853{
3854        struct r10conf *conf = mddev->private;
3855
3856        switch(state) {
3857        case 1:
3858                raise_barrier(conf, 0);
3859                break;
3860        case 0:
3861                lower_barrier(conf);
3862                break;
3863        }
3864}
3865
3866static int raid10_resize(struct mddev *mddev, sector_t sectors)
3867{
3868        /* Resize of 'far' arrays is not supported.
3869         * For 'near' and 'offset' arrays we can set the
3870         * number of sectors used to be an appropriate multiple
3871         * of the chunk size.
3872         * For 'offset', this is far_copies*chunksize.
3873         * For 'near' the multiplier is the LCM of
3874         * near_copies and raid_disks.
3875         * So if far_copies > 1 && !far_offset, fail.
3876         * Else find LCM(raid_disks, near_copy)*far_copies and
3877         * multiply by chunk_size.  Then round to this number.
3878         * This is mostly done by raid10_size()
3879         */
3880        struct r10conf *conf = mddev->private;
3881        sector_t oldsize, size;
3882
3883        if (mddev->reshape_position != MaxSector)
3884                return -EBUSY;
3885
3886        if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3887                return -EINVAL;
3888
3889        oldsize = raid10_size(mddev, 0, 0);
3890        size = raid10_size(mddev, sectors, 0);
3891        if (mddev->external_size &&
3892            mddev->array_sectors > size)
3893                return -EINVAL;
3894        if (mddev->bitmap) {
3895                int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3896                if (ret)
3897                        return ret;
3898        }
3899        md_set_array_sectors(mddev, size);
3900        set_capacity(mddev->gendisk, mddev->array_sectors);
3901        revalidate_disk(mddev->gendisk);
3902        if (sectors > mddev->dev_sectors &&
3903            mddev->recovery_cp > oldsize) {
3904                mddev->recovery_cp = oldsize;
3905                set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3906        }
3907        calc_sectors(conf, sectors);
3908        mddev->dev_sectors = conf->dev_sectors;
3909        mddev->resync_max_sectors = size;
3910        return 0;
3911}
3912
3913static void *raid10_takeover_raid0(struct mddev *mddev)
3914{
3915        struct md_rdev *rdev;
3916        struct r10conf *conf;
3917
3918        if (mddev->degraded > 0) {
3919                printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3920                       mdname(mddev));
3921                return ERR_PTR(-EINVAL);
3922        }
3923
3924        /* Set new parameters */
3925        mddev->new_level = 10;
3926        /* new layout: far_copies = 1, near_copies = 2 */
3927        mddev->new_layout = (1<<8) + 2;
3928        mddev->new_chunk_sectors = mddev->chunk_sectors;
3929        mddev->delta_disks = mddev->raid_disks;
3930        mddev->raid_disks *= 2;
3931        /* make sure it will be not marked as dirty */
3932        mddev->recovery_cp = MaxSector;
3933
3934        conf = setup_conf(mddev);
3935        if (!IS_ERR(conf)) {
3936                rdev_for_each(rdev, mddev)
3937                        if (rdev->raid_disk >= 0)
3938                                rdev->new_raid_disk = rdev->raid_disk * 2;
3939                conf->barrier = 1;
3940        }
3941
3942        return conf;
3943}
3944
3945static void *raid10_takeover(struct mddev *mddev)
3946{
3947        struct r0conf *raid0_conf;
3948
3949        /* raid10 can take over:
3950         *  raid0 - providing it has only two drives
3951         */
3952        if (mddev->level == 0) {
3953                /* for raid0 takeover only one zone is supported */
3954                raid0_conf = mddev->private;
3955                if (raid0_conf->nr_strip_zones > 1) {
3956                        printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3957                               " with more than one zone.\n",
3958                               mdname(mddev));
3959                        return ERR_PTR(-EINVAL);
3960                }
3961                return raid10_takeover_raid0(mddev);
3962        }
3963        return ERR_PTR(-EINVAL);
3964}
3965
3966static int raid10_check_reshape(struct mddev *mddev)
3967{
3968        /* Called when there is a request to change
3969         * - layout (to ->new_layout)
3970         * - chunk size (to ->new_chunk_sectors)
3971         * - raid_disks (by delta_disks)
3972         * or when trying to restart a reshape that was ongoing.
3973         *
3974         * We need to validate the request and possibly allocate
3975         * space if that might be an issue later.
3976         *
3977         * Currently we reject any reshape of a 'far' mode array,
3978         * allow chunk size to change if new is generally acceptable,
3979         * allow raid_disks to increase, and allow
3980         * a switch between 'near' mode and 'offset' mode.
3981         */
3982        struct r10conf *conf = mddev->private;
3983        struct geom geo;
3984
3985        if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3986                return -EINVAL;
3987
3988        if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3989                /* mustn't change number of copies */
3990                return -EINVAL;
3991        if (geo.far_copies > 1 && !geo.far_offset)
3992                /* Cannot switch to 'far' mode */
3993                return -EINVAL;
3994
3995        if (mddev->array_sectors & geo.chunk_mask)
3996                        /* not factor of array size */
3997                        return -EINVAL;
3998
3999        if (!enough(conf, -1))
4000                return -EINVAL;
4001
4002        kfree(conf->mirrors_new);
4003        conf->mirrors_new = NULL;
4004        if (mddev->delta_disks > 0) {
4005                /* allocate new 'mirrors' list */
4006                conf->mirrors_new = kzalloc(
4007                        sizeof(struct raid10_info)
4008                        *(mddev->raid_disks +
4009                          mddev->delta_disks),
4010                        GFP_KERNEL);
4011                if (!conf->mirrors_new)
4012                        return -ENOMEM;
4013        }
4014        return 0;
4015}
4016
4017/*
4018 * Need to check if array has failed when deciding whether to:
4019 *  - start an array
4020 *  - remove non-faulty devices
4021 *  - add a spare
4022 *  - allow a reshape
4023 * This determination is simple when no reshape is happening.
4024 * However if there is a reshape, we need to carefully check
4025 * both the before and after sections.
4026 * This is because some failed devices may only affect one
4027 * of the two sections, and some non-in_sync devices may
4028 * be insync in the section most affected by failed devices.
4029 */
4030static int calc_degraded(struct r10conf *conf)
4031{
4032        int degraded, degraded2;
4033        int i;
4034
4035        rcu_read_lock();
4036        degraded = 0;
4037        /* 'prev' section first */
4038        for (i = 0; i < conf->prev.raid_disks; i++) {
4039                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4040                if (!rdev || test_bit(Faulty, &rdev->flags))
4041                        degraded++;
4042                else if (!test_bit(In_sync, &rdev->flags))
4043                        /* When we can reduce the number of devices in
4044                         * an array, this might not contribute to
4045                         * 'degraded'.  It does now.
4046                         */
4047                        degraded++;
4048        }
4049        rcu_read_unlock();
4050        if (conf->geo.raid_disks == conf->prev.raid_disks)
4051                return degraded;
4052        rcu_read_lock();
4053        degraded2 = 0;
4054        for (i = 0; i < conf->geo.raid_disks; i++) {
4055                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4056                if (!rdev || test_bit(Faulty, &rdev->flags))
4057                        degraded2++;
4058                else if (!test_bit(In_sync, &rdev->flags)) {
4059                        /* If reshape is increasing the number of devices,
4060                         * this section has already been recovered, so
4061                         * it doesn't contribute to degraded.
4062                         * else it does.
4063                         */
4064                        if (conf->geo.raid_disks <= conf->prev.raid_disks)
4065                                degraded2++;
4066                }
4067        }
4068        rcu_read_unlock();
4069        if (degraded2 > degraded)
4070                return degraded2;
4071        return degraded;
4072}
4073
4074static int raid10_start_reshape(struct mddev *mddev)
4075{
4076        /* A 'reshape' has been requested. This commits
4077         * the various 'new' fields and sets MD_RECOVER_RESHAPE
4078         * This also checks if there are enough spares and adds them
4079         * to the array.
4080         * We currently require enough spares to make the final
4081         * array non-degraded.  We also require that the difference
4082         * between old and new data_offset - on each device - is
4083         * enough that we never risk over-writing.
4084         */
4085
4086        unsigned long before_length, after_length;
4087        sector_t min_offset_diff = 0;
4088        int first = 1;
4089        struct geom new;
4090        struct r10conf *conf = mddev->private;
4091        struct md_rdev *rdev;
4092        int spares = 0;
4093        int ret;
4094
4095        if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4096                return -EBUSY;
4097
4098        if (setup_geo(&new, mddev, geo_start) != conf->copies)
4099                return -EINVAL;
4100
4101        before_length = ((1 << conf->prev.chunk_shift) *
4102                         conf->prev.far_copies);
4103        after_length = ((1 << conf->geo.chunk_shift) *
4104                        conf->geo.far_copies);
4105
4106        rdev_for_each(rdev, mddev) {
4107                if (!test_bit(In_sync, &rdev->flags)
4108                    && !test_bit(Faulty, &rdev->flags))
4109                        spares++;
4110                if (rdev->raid_disk >= 0) {
4111                        long long diff = (rdev->new_data_offset
4112                                          - rdev->data_offset);
4113                        if (!mddev->reshape_backwards)
4114                                diff = -diff;
4115                        if (diff < 0)
4116                                diff = 0;
4117                        if (first || diff < min_offset_diff)
4118                                min_offset_diff = diff;
4119                }
4120        }
4121
4122        if (max(before_length, after_length) > min_offset_diff)
4123                return -EINVAL;
4124
4125        if (spares < mddev->delta_disks)
4126                return -EINVAL;
4127
4128        conf->offset_diff = min_offset_diff;
4129        spin_lock_irq(&conf->device_lock);
4130        if (conf->mirrors_new) {
4131                memcpy(conf->mirrors_new, conf->mirrors,
4132                       sizeof(struct raid10_info)*conf->prev.raid_disks);
4133                smp_mb();
4134                kfree(conf->mirrors_old); /* FIXME and elsewhere */
4135                conf->mirrors_old = conf->mirrors;
4136                conf->mirrors = conf->mirrors_new;
4137                conf->mirrors_new = NULL;
4138        }
4139        setup_geo(&conf->geo, mddev, geo_start);
4140        smp_mb();
4141        if (mddev->reshape_backwards) {
4142                sector_t size = raid10_size(mddev, 0, 0);
4143                if (size < mddev->array_sectors) {
4144                        spin_unlock_irq(&conf->device_lock);
4145                        printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4146                               mdname(mddev));
4147                        return -EINVAL;
4148                }
4149                mddev->resync_max_sectors = size;
4150                conf->reshape_progress = size;
4151        } else
4152                conf->reshape_progress = 0;
4153        spin_unlock_irq(&conf->device_lock);
4154
4155        if (mddev->delta_disks && mddev->bitmap) {
4156                ret = bitmap_resize(mddev->bitmap,
4157                                    raid10_size(mddev, 0,
4158                                                conf->geo.raid_disks),
4159                                    0, 0);
4160                if (ret)
4161                        goto abort;
4162        }
4163        if (mddev->delta_disks > 0) {
4164                rdev_for_each(rdev, mddev)
4165                        if (rdev->raid_disk < 0 &&
4166                            !test_bit(Faulty, &rdev->flags)) {
4167                                if (raid10_add_disk(mddev, rdev) == 0) {
4168                                        if (rdev->raid_disk >=
4169                                            conf->prev.raid_disks)
4170                                                set_bit(In_sync, &rdev->flags);
4171                                        else
4172                                                rdev->recovery_offset = 0;
4173
4174                                        if (sysfs_link_rdev(mddev, rdev))
4175                                                /* Failure here  is OK */;
4176                                }
4177                        } else if (rdev->raid_disk >= conf->prev.raid_disks
4178                                   && !test_bit(Faulty, &rdev->flags)) {
4179                                /* This is a spare that was manually added */
4180                                set_bit(In_sync, &rdev->flags);
4181                        }
4182        }
4183        /* When a reshape changes the number of devices,
4184         * ->degraded is measured against the larger of the
4185         * pre and  post numbers.
4186         */
4187        spin_lock_irq(&conf->device_lock);
4188        mddev->degraded = calc_degraded(conf);
4189        spin_unlock_irq(&conf->device_lock);
4190        mddev->raid_disks = conf->geo.raid_disks;
4191        mddev->reshape_position = conf->reshape_progress;
4192        set_bit(MD_CHANGE_DEVS, &mddev->flags);
4193
4194        clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4195        clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4196        set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4197        set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4198
4199        mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4200                                                "reshape");
4201        if (!mddev->sync_thread) {
4202                ret = -EAGAIN;
4203                goto abort;
4204        }
4205        conf->reshape_checkpoint = jiffies;
4206        md_wakeup_thread(mddev->sync_thread);
4207        md_new_event(mddev);
4208        return 0;
4209
4210abort:
4211        mddev->recovery = 0;
4212        spin_lock_irq(&conf->device_lock);
4213        conf->geo = conf->prev;
4214        mddev->raid_disks = conf->geo.raid_disks;
4215        rdev_for_each(rdev, mddev)
4216                rdev->new_data_offset = rdev->data_offset;
4217        smp_wmb();
4218        conf->reshape_progress = MaxSector;
4219        mddev->reshape_position = MaxSector;
4220        spin_unlock_irq(&conf->device_lock);
4221        return ret;
4222}
4223
4224/* Calculate the last device-address that could contain
4225 * any block from the chunk that includes the array-address 's'
4226 * and report the next address.
4227 * i.e. the address returned will be chunk-aligned and after
4228 * any data that is in the chunk containing 's'.
4229 */
4230static sector_t last_dev_address(sector_t s, struct geom *geo)
4231{
4232        s = (s | geo->chunk_mask) + 1;
4233        s >>= geo->chunk_shift;
4234        s *= geo->near_copies;
4235        s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4236        s *= geo->far_copies;
4237        s <<= geo->chunk_shift;
4238        return s;
4239}
4240
4241/* Calculate the first device-address that could contain
4242 * any block from the chunk that includes the array-address 's'.
4243 * This too will be the start of a chunk
4244 */
4245static sector_t first_dev_address(sector_t s, struct geom *geo)
4246{
4247        s >>= geo->chunk_shift;
4248        s *= geo->near_copies;
4249        sector_div(s, geo->raid_disks);
4250        s *= geo->far_copies;
4251        s <<= geo->chunk_shift;
4252        return s;
4253}
4254
4255static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4256                                int *skipped)
4257{
4258        /* We simply copy at most one chunk (smallest of old and new)
4259         * at a time, possibly less if that exceeds RESYNC_PAGES,
4260         * or we hit a bad block or something.
4261         * This might mean we pause for normal IO in the middle of
4262         * a chunk, but that is not a problem was mddev->reshape_position
4263         * can record any location.
4264         *
4265         * If we will want to write to a location that isn't
4266         * yet recorded as 'safe' (i.e. in metadata on disk) then
4267         * we need to flush all reshape requests and update the metadata.
4268         *
4269         * When reshaping forwards (e.g. to more devices), we interpret
4270         * 'safe' as the earliest block which might not have been copied
4271         * down yet.  We divide this by previous stripe size and multiply
4272         * by previous stripe length to get lowest device offset that we
4273         * cannot write to yet.
4274         * We interpret 'sector_nr' as an address that we want to write to.
4275         * From this we use last_device_address() to find where we might
4276         * write to, and first_device_address on the  'safe' position.
4277         * If this 'next' write position is after the 'safe' position,
4278         * we must update the metadata to increase the 'safe' position.
4279         *
4280         * When reshaping backwards, we round in the opposite direction
4281         * and perform the reverse test:  next write position must not be
4282         * less than current safe position.
4283         *
4284         * In all this the minimum difference in data offsets
4285         * (conf->offset_diff - always positive) allows a bit of slack,
4286         * so next can be after 'safe', but not by more than offset_disk
4287         *
4288         * We need to prepare all the bios here before we start any IO
4289         * to ensure the size we choose is acceptable to all devices.
4290         * The means one for each copy for write-out and an extra one for
4291         * read-in.
4292         * We store the read-in bio in ->master_bio and the others in
4293         * ->devs[x].bio and ->devs[x].repl_bio.
4294         */
4295        struct r10conf *conf = mddev->private;
4296        struct r10bio *r10_bio;
4297        sector_t next, safe, last;
4298        int max_sectors;
4299        int nr_sectors;
4300        int s;
4301        struct md_rdev *rdev;
4302        int need_flush = 0;
4303        struct bio *blist;
4304        struct bio *bio, *read_bio;
4305        int sectors_done = 0;
4306
4307        if (sector_nr == 0) {
4308                /* If restarting in the middle, skip the initial sectors */
4309                if (mddev->reshape_backwards &&
4310                    conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4311                        sector_nr = (raid10_size(mddev, 0, 0)
4312                                     - conf->reshape_progress);
4313                } else if (!mddev->reshape_backwards &&
4314                           conf->reshape_progress > 0)
4315                        sector_nr = conf->reshape_progress;
4316                if (sector_nr) {
4317                        mddev->curr_resync_completed = sector_nr;
4318                        sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4319                        *skipped = 1;
4320                        return sector_nr;
4321                }
4322        }
4323
4324        /* We don't use sector_nr to track where we are up to
4325         * as that doesn't work well for ->reshape_backwards.
4326         * So just use ->reshape_progress.
4327         */
4328        if (mddev->reshape_backwards) {
4329                /* 'next' is the earliest device address that we might
4330                 * write to for this chunk in the new layout
4331                 */
4332                next = first_dev_address(conf->reshape_progress - 1,
4333                                         &conf->geo);
4334
4335                /* 'safe' is the last device address that we might read from
4336                 * in the old layout after a restart
4337                 */
4338                safe = last_dev_address(conf->reshape_safe - 1,
4339                                        &conf->prev);
4340
4341                if (next + conf->offset_diff < safe)
4342                        need_flush = 1;
4343
4344                last = conf->reshape_progress - 1;
4345                sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4346                                               & conf->prev.chunk_mask);
4347                if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4348                        sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4349        } else {
4350                /* 'next' is after the last device address that we
4351                 * might write to for this chunk in the new layout
4352                 */
4353                next = last_dev_address(conf->reshape_progress, &conf->geo);
4354
4355                /* 'safe' is the earliest device address that we might
4356                 * read from in the old layout after a restart
4357                 */
4358                safe = first_dev_address(conf->reshape_safe, &conf->prev);
4359
4360                /* Need to update metadata if 'next' might be beyond 'safe'
4361                 * as that would possibly corrupt data
4362                 */
4363                if (next > safe + conf->offset_diff)
4364                        need_flush = 1;
4365
4366                sector_nr = conf->reshape_progress;
4367                last  = sector_nr | (conf->geo.chunk_mask
4368                                     & conf->prev.chunk_mask);
4369
4370                if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4371                        last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4372        }
4373
4374        if (need_flush ||
4375            time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4376                /* Need to update reshape_position in metadata */
4377                wait_barrier(conf);
4378                mddev->reshape_position = conf->reshape_progress;
4379                if (mddev->reshape_backwards)
4380                        mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4381                                - conf->reshape_progress;
4382                else
4383                        mddev->curr_resync_completed = conf->reshape_progress;
4384                conf->reshape_checkpoint = jiffies;
4385                set_bit(MD_CHANGE_DEVS, &mddev->flags);
4386                md_wakeup_thread(mddev->thread);
4387                wait_event(mddev->sb_wait, mddev->flags == 0 ||
4388                           kthread_should_stop());
4389                conf->reshape_safe = mddev->reshape_position;
4390                allow_barrier(conf);
4391        }
4392
4393read_more:
4394        /* Now schedule reads for blocks from sector_nr to last */
4395        r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4396        raise_barrier(conf, sectors_done != 0);
4397        atomic_set(&r10_bio->remaining, 0);
4398        r10_bio->mddev = mddev;
4399        r10_bio->sector = sector_nr;
4400        set_bit(R10BIO_IsReshape, &r10_bio->state);
4401        r10_bio->sectors = last - sector_nr + 1;
4402        rdev = read_balance(conf, r10_bio, &max_sectors);
4403        BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4404
4405        if (!rdev) {
4406                /* Cannot read from here, so need to record bad blocks
4407                 * on all the target devices.
4408                 */
4409                // FIXME
4410                set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4411                return sectors_done;
4412        }
4413
4414        read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4415
4416        read_bio->bi_bdev = rdev->bdev;
4417        read_bio->bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4418                               + rdev->data_offset);
4419        read_bio->bi_private = r10_bio;
4420        read_bio->bi_end_io = end_sync_read;
4421        read_bio->bi_rw = READ;
4422        read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
4423        read_bio->bi_flags |= 1 << BIO_UPTODATE;
4424        read_bio->bi_vcnt = 0;
4425        read_bio->bi_size = 0;
4426        r10_bio->master_bio = read_bio;
4427        r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4428
4429        /* Now find the locations in the new layout */
4430        __raid10_find_phys(&conf->geo, r10_bio);
4431
4432        blist = read_bio;
4433        read_bio->bi_next = NULL;
4434
4435        for (s = 0; s < conf->copies*2; s++) {
4436                struct bio *b;
4437                int d = r10_bio->devs[s/2].devnum;
4438                struct md_rdev *rdev2;
4439                if (s&1) {
4440                        rdev2 = conf->mirrors[d].replacement;
4441                        b = r10_bio->devs[s/2].repl_bio;
4442                } else {
4443                        rdev2 = conf->mirrors[d].rdev;
4444                        b = r10_bio->devs[s/2].bio;
4445                }
4446                if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4447                        continue;
4448
4449                bio_reset(b);
4450                b->bi_bdev = rdev2->bdev;
4451                b->bi_sector = r10_bio->devs[s/2].addr + rdev2->new_data_offset;
4452                b->bi_private = r10_bio;
4453                b->bi_end_io = end_reshape_write;
4454                b->bi_rw = WRITE;
4455                b->bi_next = blist;
4456                blist = b;
4457        }
4458
4459        /* Now add as many pages as possible to all of these bios. */
4460
4461        nr_sectors = 0;
4462        for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4463                struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4464                int len = (max_sectors - s) << 9;
4465                if (len > PAGE_SIZE)
4466                        len = PAGE_SIZE;
4467                for (bio = blist; bio ; bio = bio->bi_next) {
4468                        struct bio *bio2;
4469                        if (bio_add_page(bio, page, len, 0))
4470                                continue;
4471
4472                        /* Didn't fit, must stop */
4473                        for (bio2 = blist;
4474                             bio2 && bio2 != bio;
4475                             bio2 = bio2->bi_next) {
4476                                /* Remove last page from this bio */
4477                                bio2->bi_vcnt--;
4478                                bio2->bi_size -= len;
4479                                bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4480                        }
4481                        goto bio_full;
4482                }
4483                sector_nr += len >> 9;
4484                nr_sectors += len >> 9;
4485        }
4486bio_full:
4487        r10_bio->sectors = nr_sectors;
4488
4489        /* Now submit the read */
4490        md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4491        atomic_inc(&r10_bio->remaining);
4492        read_bio->bi_next = NULL;
4493        generic_make_request(read_bio);
4494        sector_nr += nr_sectors;
4495        sectors_done += nr_sectors;
4496        if (sector_nr <= last)
4497                goto read_more;
4498
4499        /* Now that we have done the whole section we can
4500         * update reshape_progress
4501         */
4502        if (mddev->reshape_backwards)
4503                conf->reshape_progress -= sectors_done;
4504        else
4505                conf->reshape_progress += sectors_done;
4506
4507        return sectors_done;
4508}
4509
4510static void end_reshape_request(struct r10bio *r10_bio);
4511static int handle_reshape_read_error(struct mddev *mddev,
4512                                     struct r10bio *r10_bio);
4513static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4514{
4515        /* Reshape read completed.  Hopefully we have a block
4516         * to write out.
4517         * If we got a read error then we do sync 1-page reads from
4518         * elsewhere until we find the data - or give up.
4519         */
4520        struct r10conf *conf = mddev->private;
4521        int s;
4522
4523        if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4524                if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4525                        /* Reshape has been aborted */
4526                        md_done_sync(mddev, r10_bio->sectors, 0);
4527                        return;
4528                }
4529
4530        /* We definitely have the data in the pages, schedule the
4531         * writes.
4532         */
4533        atomic_set(&r10_bio->remaining, 1);
4534        for (s = 0; s < conf->copies*2; s++) {
4535                struct bio *b;
4536                int d = r10_bio->devs[s/2].devnum;
4537                struct md_rdev *rdev;
4538                if (s&1) {
4539                        rdev = conf->mirrors[d].replacement;
4540                        b = r10_bio->devs[s/2].repl_bio;
4541                } else {
4542                        rdev = conf->mirrors[d].rdev;
4543                        b = r10_bio->devs[s/2].bio;
4544                }
4545                if (!rdev || test_bit(Faulty, &rdev->flags))
4546                        continue;
4547                atomic_inc(&rdev->nr_pending);
4548                md_sync_acct(b->bi_bdev, r10_bio->sectors);
4549                atomic_inc(&r10_bio->remaining);
4550                b->bi_next = NULL;
4551                generic_make_request(b);
4552        }
4553        end_reshape_request(r10_bio);
4554}
4555
4556static void end_reshape(struct r10conf *conf)
4557{
4558        if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4559                return;
4560
4561        spin_lock_irq(&conf->device_lock);
4562        conf->prev = conf->geo;
4563        md_finish_reshape(conf->mddev);
4564        smp_wmb();
4565        conf->reshape_progress = MaxSector;
4566        spin_unlock_irq(&conf->device_lock);
4567
4568        /* read-ahead size must cover two whole stripes, which is
4569         * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4570         */
4571        if (conf->mddev->queue) {
4572                int stripe = conf->geo.raid_disks *
4573                        ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4574                stripe /= conf->geo.near_copies;
4575                if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4576                        conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4577        }
4578        conf->fullsync = 0;
4579}
4580
4581
4582static int handle_reshape_read_error(struct mddev *mddev,
4583                                     struct r10bio *r10_bio)
4584{
4585        /* Use sync reads to get the blocks from somewhere else */
4586        int sectors = r10_bio->sectors;
4587        struct r10conf *conf = mddev->private;
4588        struct {
4589                struct r10bio r10_bio;
4590                struct r10dev devs[conf->copies];
4591        } on_stack;
4592        struct r10bio *r10b = &on_stack.r10_bio;
4593        int slot = 0;
4594        int idx = 0;
4595        struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4596
4597        r10b->sector = r10_bio->sector;
4598        __raid10_find_phys(&conf->prev, r10b);
4599
4600        while (sectors) {
4601                int s = sectors;
4602                int success = 0;
4603                int first_slot = slot;
4604
4605                if (s > (PAGE_SIZE >> 9))
4606                        s = PAGE_SIZE >> 9;
4607
4608                while (!success) {
4609                        int d = r10b->devs[slot].devnum;
4610                        struct md_rdev *rdev = conf->mirrors[d].rdev;
4611                        sector_t addr;
4612                        if (rdev == NULL ||
4613                            test_bit(Faulty, &rdev->flags) ||
4614                            !test_bit(In_sync, &rdev->flags))
4615                                goto failed;
4616
4617                        addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4618                        success = sync_page_io(rdev,
4619                                               addr,
4620                                               s << 9,
4621                                               bvec[idx].bv_page,
4622                                               READ, false);
4623                        if (success)
4624                                break;
4625                failed:
4626                        slot++;
4627                        if (slot >= conf->copies)
4628                                slot = 0;
4629                        if (slot == first_slot)
4630                                break;
4631                }
4632                if (!success) {
4633                        /* couldn't read this block, must give up */
4634                        set_bit(MD_RECOVERY_INTR,
4635                                &mddev->recovery);
4636                        return -EIO;
4637                }
4638                sectors -= s;
4639                idx++;
4640        }
4641        return 0;
4642}
4643
4644static void end_reshape_write(struct bio *bio, int error)
4645{
4646        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4647        struct r10bio *r10_bio = bio->bi_private;
4648        struct mddev *mddev = r10_bio->mddev;
4649        struct r10conf *conf = mddev->private;
4650        int d;
4651        int slot;
4652        int repl;
4653        struct md_rdev *rdev = NULL;
4654
4655        d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4656        if (repl)
4657                rdev = conf->mirrors[d].replacement;
4658        if (!rdev) {
4659                smp_mb();
4660                rdev = conf->mirrors[d].rdev;
4661        }
4662
4663        if (!uptodate) {
4664                /* FIXME should record badblock */
4665                md_error(mddev, rdev);
4666        }
4667
4668        rdev_dec_pending(rdev, mddev);
4669        end_reshape_request(r10_bio);
4670}
4671
4672static void end_reshape_request(struct r10bio *r10_bio)
4673{
4674        if (!atomic_dec_and_test(&r10_bio->remaining))
4675                return;
4676        md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4677        bio_put(r10_bio->master_bio);
4678        put_buf(r10_bio);
4679}
4680
4681static void raid10_finish_reshape(struct mddev *mddev)
4682{
4683        struct r10conf *conf = mddev->private;
4684
4685        if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4686                return;
4687
4688        if (mddev->delta_disks > 0) {
4689                sector_t size = raid10_size(mddev, 0, 0);
4690                md_set_array_sectors(mddev, size);
4691                if (mddev->recovery_cp > mddev->resync_max_sectors) {
4692                        mddev->recovery_cp = mddev->resync_max_sectors;
4693                        set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4694                }
4695                mddev->resync_max_sectors = size;
4696                set_capacity(mddev->gendisk, mddev->array_sectors);
4697                revalidate_disk(mddev->gendisk);
4698        } else {
4699                int d;
4700                for (d = conf->geo.raid_disks ;
4701                     d < conf->geo.raid_disks - mddev->delta_disks;
4702                     d++) {
4703                        struct md_rdev *rdev = conf->mirrors[d].rdev;
4704                        if (rdev)
4705                                clear_bit(In_sync, &rdev->flags);
4706                        rdev = conf->mirrors[d].replacement;
4707                        if (rdev)
4708                                clear_bit(In_sync, &rdev->flags);
4709                }
4710        }
4711        mddev->layout = mddev->new_layout;
4712        mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4713        mddev->reshape_position = MaxSector;
4714        mddev->delta_disks = 0;
4715        mddev->reshape_backwards = 0;
4716}
4717
4718static struct md_personality raid10_personality =
4719{
4720        .name           = "raid10",
4721        .level          = 10,
4722        .owner          = THIS_MODULE,
4723        .make_request   = make_request,
4724        .run            = run,
4725        .stop           = stop,
4726        .status         = status,
4727        .error_handler  = error,
4728        .hot_add_disk   = raid10_add_disk,
4729        .hot_remove_disk= raid10_remove_disk,
4730        .spare_active   = raid10_spare_active,
4731        .sync_request   = sync_request,
4732        .quiesce        = raid10_quiesce,
4733        .size           = raid10_size,
4734        .resize         = raid10_resize,
4735        .takeover       = raid10_takeover,
4736        .check_reshape  = raid10_check_reshape,
4737        .start_reshape  = raid10_start_reshape,
4738        .finish_reshape = raid10_finish_reshape,
4739};
4740
4741static int __init raid_init(void)
4742{
4743        return register_md_personality(&raid10_personality);
4744}
4745
4746static void raid_exit(void)
4747{
4748        unregister_md_personality(&raid10_personality);
4749}
4750
4751module_init(raid_init);
4752module_exit(raid_exit);
4753MODULE_LICENSE("GPL");
4754MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4755MODULE_ALIAS("md-personality-9"); /* RAID10 */
4756MODULE_ALIAS("md-raid10");
4757MODULE_ALIAS("md-level-10");
4758
4759module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
4760
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.