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