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