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