linux/drivers/md/raid1.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * raid1.c : Multiple Devices driver for Linux
   4 *
   5 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   6 *
   7 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   8 *
   9 * RAID-1 management functions.
  10 *
  11 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  12 *
  13 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  14 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  15 *
  16 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  17 * bitmapped intelligence in resync:
  18 *
  19 *      - bitmap marked during normal i/o
  20 *      - bitmap used to skip nondirty blocks during sync
  21 *
  22 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  23 * - persistent bitmap code
  24 */
  25
  26#include <linux/slab.h>
  27#include <linux/delay.h>
  28#include <linux/blkdev.h>
  29#include <linux/module.h>
  30#include <linux/seq_file.h>
  31#include <linux/ratelimit.h>
  32#include <linux/interval_tree_generic.h>
  33
  34#include <trace/events/block.h>
  35
  36#include "md.h"
  37#include "raid1.h"
  38#include "md-bitmap.h"
  39
  40#define UNSUPPORTED_MDDEV_FLAGS         \
  41        ((1L << MD_HAS_JOURNAL) |       \
  42         (1L << MD_JOURNAL_CLEAN) |     \
  43         (1L << MD_HAS_PPL) |           \
  44         (1L << MD_HAS_MULTIPLE_PPLS))
  45
  46static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
  47static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
  48
  49#define raid1_log(md, fmt, args...)                             \
  50        do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
  51
  52#include "raid1-10.c"
  53
  54#define START(node) ((node)->start)
  55#define LAST(node) ((node)->last)
  56INTERVAL_TREE_DEFINE(struct serial_info, node, sector_t, _subtree_last,
  57                     START, LAST, static inline, raid1_rb);
  58
  59static int check_and_add_serial(struct md_rdev *rdev, struct r1bio *r1_bio,
  60                                struct serial_info *si, int idx)
  61{
  62        unsigned long flags;
  63        int ret = 0;
  64        sector_t lo = r1_bio->sector;
  65        sector_t hi = lo + r1_bio->sectors;
  66        struct serial_in_rdev *serial = &rdev->serial[idx];
  67
  68        spin_lock_irqsave(&serial->serial_lock, flags);
  69        /* collision happened */
  70        if (raid1_rb_iter_first(&serial->serial_rb, lo, hi))
  71                ret = -EBUSY;
  72        else {
  73                si->start = lo;
  74                si->last = hi;
  75                raid1_rb_insert(si, &serial->serial_rb);
  76        }
  77        spin_unlock_irqrestore(&serial->serial_lock, flags);
  78
  79        return ret;
  80}
  81
  82static void wait_for_serialization(struct md_rdev *rdev, struct r1bio *r1_bio)
  83{
  84        struct mddev *mddev = rdev->mddev;
  85        struct serial_info *si;
  86        int idx = sector_to_idx(r1_bio->sector);
  87        struct serial_in_rdev *serial = &rdev->serial[idx];
  88
  89        if (WARN_ON(!mddev->serial_info_pool))
  90                return;
  91        si = mempool_alloc(mddev->serial_info_pool, GFP_NOIO);
  92        wait_event(serial->serial_io_wait,
  93                   check_and_add_serial(rdev, r1_bio, si, idx) == 0);
  94}
  95
  96static void remove_serial(struct md_rdev *rdev, sector_t lo, sector_t hi)
  97{
  98        struct serial_info *si;
  99        unsigned long flags;
 100        int found = 0;
 101        struct mddev *mddev = rdev->mddev;
 102        int idx = sector_to_idx(lo);
 103        struct serial_in_rdev *serial = &rdev->serial[idx];
 104
 105        spin_lock_irqsave(&serial->serial_lock, flags);
 106        for (si = raid1_rb_iter_first(&serial->serial_rb, lo, hi);
 107             si; si = raid1_rb_iter_next(si, lo, hi)) {
 108                if (si->start == lo && si->last == hi) {
 109                        raid1_rb_remove(si, &serial->serial_rb);
 110                        mempool_free(si, mddev->serial_info_pool);
 111                        found = 1;
 112                        break;
 113                }
 114        }
 115        if (!found)
 116                WARN(1, "The write IO is not recorded for serialization\n");
 117        spin_unlock_irqrestore(&serial->serial_lock, flags);
 118        wake_up(&serial->serial_io_wait);
 119}
 120
 121/*
 122 * for resync bio, r1bio pointer can be retrieved from the per-bio
 123 * 'struct resync_pages'.
 124 */
 125static inline struct r1bio *get_resync_r1bio(struct bio *bio)
 126{
 127        return get_resync_pages(bio)->raid_bio;
 128}
 129
 130static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
 131{
 132        struct pool_info *pi = data;
 133        int size = offsetof(struct r1bio, bios[pi->raid_disks]);
 134
 135        /* allocate a r1bio with room for raid_disks entries in the bios array */
 136        return kzalloc(size, gfp_flags);
 137}
 138
 139#define RESYNC_DEPTH 32
 140#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
 141#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
 142#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
 143#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
 144#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
 145
 146static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
 147{
 148        struct pool_info *pi = data;
 149        struct r1bio *r1_bio;
 150        struct bio *bio;
 151        int need_pages;
 152        int j;
 153        struct resync_pages *rps;
 154
 155        r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 156        if (!r1_bio)
 157                return NULL;
 158
 159        rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
 160                            gfp_flags);
 161        if (!rps)
 162                goto out_free_r1bio;
 163
 164        /*
 165         * Allocate bios : 1 for reading, n-1 for writing
 166         */
 167        for (j = pi->raid_disks ; j-- ; ) {
 168                bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 169                if (!bio)
 170                        goto out_free_bio;
 171                r1_bio->bios[j] = bio;
 172        }
 173        /*
 174         * Allocate RESYNC_PAGES data pages and attach them to
 175         * the first bio.
 176         * If this is a user-requested check/repair, allocate
 177         * RESYNC_PAGES for each bio.
 178         */
 179        if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 180                need_pages = pi->raid_disks;
 181        else
 182                need_pages = 1;
 183        for (j = 0; j < pi->raid_disks; j++) {
 184                struct resync_pages *rp = &rps[j];
 185
 186                bio = r1_bio->bios[j];
 187
 188                if (j < need_pages) {
 189                        if (resync_alloc_pages(rp, gfp_flags))
 190                                goto out_free_pages;
 191                } else {
 192                        memcpy(rp, &rps[0], sizeof(*rp));
 193                        resync_get_all_pages(rp);
 194                }
 195
 196                rp->raid_bio = r1_bio;
 197                bio->bi_private = rp;
 198        }
 199
 200        r1_bio->master_bio = NULL;
 201
 202        return r1_bio;
 203
 204out_free_pages:
 205        while (--j >= 0)
 206                resync_free_pages(&rps[j]);
 207
 208out_free_bio:
 209        while (++j < pi->raid_disks)
 210                bio_put(r1_bio->bios[j]);
 211        kfree(rps);
 212
 213out_free_r1bio:
 214        rbio_pool_free(r1_bio, data);
 215        return NULL;
 216}
 217
 218static void r1buf_pool_free(void *__r1_bio, void *data)
 219{
 220        struct pool_info *pi = data;
 221        int i;
 222        struct r1bio *r1bio = __r1_bio;
 223        struct resync_pages *rp = NULL;
 224
 225        for (i = pi->raid_disks; i--; ) {
 226                rp = get_resync_pages(r1bio->bios[i]);
 227                resync_free_pages(rp);
 228                bio_put(r1bio->bios[i]);
 229        }
 230
 231        /* resync pages array stored in the 1st bio's .bi_private */
 232        kfree(rp);
 233
 234        rbio_pool_free(r1bio, data);
 235}
 236
 237static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 238{
 239        int i;
 240
 241        for (i = 0; i < conf->raid_disks * 2; i++) {
 242                struct bio **bio = r1_bio->bios + i;
 243                if (!BIO_SPECIAL(*bio))
 244                        bio_put(*bio);
 245                *bio = NULL;
 246        }
 247}
 248
 249static void free_r1bio(struct r1bio *r1_bio)
 250{
 251        struct r1conf *conf = r1_bio->mddev->private;
 252
 253        put_all_bios(conf, r1_bio);
 254        mempool_free(r1_bio, &conf->r1bio_pool);
 255}
 256
 257static void put_buf(struct r1bio *r1_bio)
 258{
 259        struct r1conf *conf = r1_bio->mddev->private;
 260        sector_t sect = r1_bio->sector;
 261        int i;
 262
 263        for (i = 0; i < conf->raid_disks * 2; i++) {
 264                struct bio *bio = r1_bio->bios[i];
 265                if (bio->bi_end_io)
 266                        rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 267        }
 268
 269        mempool_free(r1_bio, &conf->r1buf_pool);
 270
 271        lower_barrier(conf, sect);
 272}
 273
 274static void reschedule_retry(struct r1bio *r1_bio)
 275{
 276        unsigned long flags;
 277        struct mddev *mddev = r1_bio->mddev;
 278        struct r1conf *conf = mddev->private;
 279        int idx;
 280
 281        idx = sector_to_idx(r1_bio->sector);
 282        spin_lock_irqsave(&conf->device_lock, flags);
 283        list_add(&r1_bio->retry_list, &conf->retry_list);
 284        atomic_inc(&conf->nr_queued[idx]);
 285        spin_unlock_irqrestore(&conf->device_lock, flags);
 286
 287        wake_up(&conf->wait_barrier);
 288        md_wakeup_thread(mddev->thread);
 289}
 290
 291/*
 292 * raid_end_bio_io() is called when we have finished servicing a mirrored
 293 * operation and are ready to return a success/failure code to the buffer
 294 * cache layer.
 295 */
 296static void call_bio_endio(struct r1bio *r1_bio)
 297{
 298        struct bio *bio = r1_bio->master_bio;
 299
 300        if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 301                bio->bi_status = BLK_STS_IOERR;
 302
 303        bio_endio(bio);
 304}
 305
 306static void raid_end_bio_io(struct r1bio *r1_bio)
 307{
 308        struct bio *bio = r1_bio->master_bio;
 309        struct r1conf *conf = r1_bio->mddev->private;
 310
 311        /* if nobody has done the final endio yet, do it now */
 312        if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 313                pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 314                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 315                         (unsigned long long) bio->bi_iter.bi_sector,
 316                         (unsigned long long) bio_end_sector(bio) - 1);
 317
 318                call_bio_endio(r1_bio);
 319        }
 320        /*
 321         * Wake up any possible resync thread that waits for the device
 322         * to go idle.  All I/Os, even write-behind writes, are done.
 323         */
 324        allow_barrier(conf, r1_bio->sector);
 325
 326        free_r1bio(r1_bio);
 327}
 328
 329/*
 330 * Update disk head position estimator based on IRQ completion info.
 331 */
 332static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 333{
 334        struct r1conf *conf = r1_bio->mddev->private;
 335
 336        conf->mirrors[disk].head_position =
 337                r1_bio->sector + (r1_bio->sectors);
 338}
 339
 340/*
 341 * Find the disk number which triggered given bio
 342 */
 343static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 344{
 345        int mirror;
 346        struct r1conf *conf = r1_bio->mddev->private;
 347        int raid_disks = conf->raid_disks;
 348
 349        for (mirror = 0; mirror < raid_disks * 2; mirror++)
 350                if (r1_bio->bios[mirror] == bio)
 351                        break;
 352
 353        BUG_ON(mirror == raid_disks * 2);
 354        update_head_pos(mirror, r1_bio);
 355
 356        return mirror;
 357}
 358
 359static void raid1_end_read_request(struct bio *bio)
 360{
 361        int uptodate = !bio->bi_status;
 362        struct r1bio *r1_bio = bio->bi_private;
 363        struct r1conf *conf = r1_bio->mddev->private;
 364        struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 365
 366        /*
 367         * this branch is our 'one mirror IO has finished' event handler:
 368         */
 369        update_head_pos(r1_bio->read_disk, r1_bio);
 370
 371        if (uptodate)
 372                set_bit(R1BIO_Uptodate, &r1_bio->state);
 373        else if (test_bit(FailFast, &rdev->flags) &&
 374                 test_bit(R1BIO_FailFast, &r1_bio->state))
 375                /* This was a fail-fast read so we definitely
 376                 * want to retry */
 377                ;
 378        else {
 379                /* If all other devices have failed, we want to return
 380                 * the error upwards rather than fail the last device.
 381                 * Here we redefine "uptodate" to mean "Don't want to retry"
 382                 */
 383                unsigned long flags;
 384                spin_lock_irqsave(&conf->device_lock, flags);
 385                if (r1_bio->mddev->degraded == conf->raid_disks ||
 386                    (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 387                     test_bit(In_sync, &rdev->flags)))
 388                        uptodate = 1;
 389                spin_unlock_irqrestore(&conf->device_lock, flags);
 390        }
 391
 392        if (uptodate) {
 393                raid_end_bio_io(r1_bio);
 394                rdev_dec_pending(rdev, conf->mddev);
 395        } else {
 396                /*
 397                 * oops, read error:
 398                 */
 399                char b[BDEVNAME_SIZE];
 400                pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
 401                                   mdname(conf->mddev),
 402                                   bdevname(rdev->bdev, b),
 403                                   (unsigned long long)r1_bio->sector);
 404                set_bit(R1BIO_ReadError, &r1_bio->state);
 405                reschedule_retry(r1_bio);
 406                /* don't drop the reference on read_disk yet */
 407        }
 408}
 409
 410static void close_write(struct r1bio *r1_bio)
 411{
 412        /* it really is the end of this request */
 413        if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 414                bio_free_pages(r1_bio->behind_master_bio);
 415                bio_put(r1_bio->behind_master_bio);
 416                r1_bio->behind_master_bio = NULL;
 417        }
 418        /* clear the bitmap if all writes complete successfully */
 419        md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 420                           r1_bio->sectors,
 421                           !test_bit(R1BIO_Degraded, &r1_bio->state),
 422                           test_bit(R1BIO_BehindIO, &r1_bio->state));
 423        md_write_end(r1_bio->mddev);
 424}
 425
 426static void r1_bio_write_done(struct r1bio *r1_bio)
 427{
 428        if (!atomic_dec_and_test(&r1_bio->remaining))
 429                return;
 430
 431        if (test_bit(R1BIO_WriteError, &r1_bio->state))
 432                reschedule_retry(r1_bio);
 433        else {
 434                close_write(r1_bio);
 435                if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 436                        reschedule_retry(r1_bio);
 437                else
 438                        raid_end_bio_io(r1_bio);
 439        }
 440}
 441
 442static void raid1_end_write_request(struct bio *bio)
 443{
 444        struct r1bio *r1_bio = bio->bi_private;
 445        int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 446        struct r1conf *conf = r1_bio->mddev->private;
 447        struct bio *to_put = NULL;
 448        int mirror = find_bio_disk(r1_bio, bio);
 449        struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 450        bool discard_error;
 451        sector_t lo = r1_bio->sector;
 452        sector_t hi = r1_bio->sector + r1_bio->sectors;
 453
 454        discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 455
 456        /*
 457         * 'one mirror IO has finished' event handler:
 458         */
 459        if (bio->bi_status && !discard_error) {
 460                set_bit(WriteErrorSeen, &rdev->flags);
 461                if (!test_and_set_bit(WantReplacement, &rdev->flags))
 462                        set_bit(MD_RECOVERY_NEEDED, &
 463                                conf->mddev->recovery);
 464
 465                if (test_bit(FailFast, &rdev->flags) &&
 466                    (bio->bi_opf & MD_FAILFAST) &&
 467                    /* We never try FailFast to WriteMostly devices */
 468                    !test_bit(WriteMostly, &rdev->flags)) {
 469                        md_error(r1_bio->mddev, rdev);
 470                }
 471
 472                /*
 473                 * When the device is faulty, it is not necessary to
 474                 * handle write error.
 475                 * For failfast, this is the only remaining device,
 476                 * We need to retry the write without FailFast.
 477                 */
 478                if (!test_bit(Faulty, &rdev->flags))
 479                        set_bit(R1BIO_WriteError, &r1_bio->state);
 480                else {
 481                        /* Finished with this branch */
 482                        r1_bio->bios[mirror] = NULL;
 483                        to_put = bio;
 484                }
 485        } else {
 486                /*
 487                 * Set R1BIO_Uptodate in our master bio, so that we
 488                 * will return a good error code for to the higher
 489                 * levels even if IO on some other mirrored buffer
 490                 * fails.
 491                 *
 492                 * The 'master' represents the composite IO operation
 493                 * to user-side. So if something waits for IO, then it
 494                 * will wait for the 'master' bio.
 495                 */
 496                sector_t first_bad;
 497                int bad_sectors;
 498
 499                r1_bio->bios[mirror] = NULL;
 500                to_put = bio;
 501                /*
 502                 * Do not set R1BIO_Uptodate if the current device is
 503                 * rebuilding or Faulty. This is because we cannot use
 504                 * such device for properly reading the data back (we could
 505                 * potentially use it, if the current write would have felt
 506                 * before rdev->recovery_offset, but for simplicity we don't
 507                 * check this here.
 508                 */
 509                if (test_bit(In_sync, &rdev->flags) &&
 510                    !test_bit(Faulty, &rdev->flags))
 511                        set_bit(R1BIO_Uptodate, &r1_bio->state);
 512
 513                /* Maybe we can clear some bad blocks. */
 514                if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 515                                &first_bad, &bad_sectors) && !discard_error) {
 516                        r1_bio->bios[mirror] = IO_MADE_GOOD;
 517                        set_bit(R1BIO_MadeGood, &r1_bio->state);
 518                }
 519        }
 520
 521        if (behind) {
 522                if (test_bit(CollisionCheck, &rdev->flags))
 523                        remove_serial(rdev, lo, hi);
 524                if (test_bit(WriteMostly, &rdev->flags))
 525                        atomic_dec(&r1_bio->behind_remaining);
 526
 527                /*
 528                 * In behind mode, we ACK the master bio once the I/O
 529                 * has safely reached all non-writemostly
 530                 * disks. Setting the Returned bit ensures that this
 531                 * gets done only once -- we don't ever want to return
 532                 * -EIO here, instead we'll wait
 533                 */
 534                if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 535                    test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 536                        /* Maybe we can return now */
 537                        if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 538                                struct bio *mbio = r1_bio->master_bio;
 539                                pr_debug("raid1: behind end write sectors"
 540                                         " %llu-%llu\n",
 541                                         (unsigned long long) mbio->bi_iter.bi_sector,
 542                                         (unsigned long long) bio_end_sector(mbio) - 1);
 543                                call_bio_endio(r1_bio);
 544                        }
 545                }
 546        } else if (rdev->mddev->serialize_policy)
 547                remove_serial(rdev, lo, hi);
 548        if (r1_bio->bios[mirror] == NULL)
 549                rdev_dec_pending(rdev, conf->mddev);
 550
 551        /*
 552         * Let's see if all mirrored write operations have finished
 553         * already.
 554         */
 555        r1_bio_write_done(r1_bio);
 556
 557        if (to_put)
 558                bio_put(to_put);
 559}
 560
 561static sector_t align_to_barrier_unit_end(sector_t start_sector,
 562                                          sector_t sectors)
 563{
 564        sector_t len;
 565
 566        WARN_ON(sectors == 0);
 567        /*
 568         * len is the number of sectors from start_sector to end of the
 569         * barrier unit which start_sector belongs to.
 570         */
 571        len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 572              start_sector;
 573
 574        if (len > sectors)
 575                len = sectors;
 576
 577        return len;
 578}
 579
 580/*
 581 * This routine returns the disk from which the requested read should
 582 * be done. There is a per-array 'next expected sequential IO' sector
 583 * number - if this matches on the next IO then we use the last disk.
 584 * There is also a per-disk 'last know head position' sector that is
 585 * maintained from IRQ contexts, both the normal and the resync IO
 586 * completion handlers update this position correctly. If there is no
 587 * perfect sequential match then we pick the disk whose head is closest.
 588 *
 589 * If there are 2 mirrors in the same 2 devices, performance degrades
 590 * because position is mirror, not device based.
 591 *
 592 * The rdev for the device selected will have nr_pending incremented.
 593 */
 594static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 595{
 596        const sector_t this_sector = r1_bio->sector;
 597        int sectors;
 598        int best_good_sectors;
 599        int best_disk, best_dist_disk, best_pending_disk;
 600        int has_nonrot_disk;
 601        int disk;
 602        sector_t best_dist;
 603        unsigned int min_pending;
 604        struct md_rdev *rdev;
 605        int choose_first;
 606        int choose_next_idle;
 607
 608        rcu_read_lock();
 609        /*
 610         * Check if we can balance. We can balance on the whole
 611         * device if no resync is going on, or below the resync window.
 612         * We take the first readable disk when above the resync window.
 613         */
 614 retry:
 615        sectors = r1_bio->sectors;
 616        best_disk = -1;
 617        best_dist_disk = -1;
 618        best_dist = MaxSector;
 619        best_pending_disk = -1;
 620        min_pending = UINT_MAX;
 621        best_good_sectors = 0;
 622        has_nonrot_disk = 0;
 623        choose_next_idle = 0;
 624        clear_bit(R1BIO_FailFast, &r1_bio->state);
 625
 626        if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 627            (mddev_is_clustered(conf->mddev) &&
 628            md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 629                    this_sector + sectors)))
 630                choose_first = 1;
 631        else
 632                choose_first = 0;
 633
 634        for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 635                sector_t dist;
 636                sector_t first_bad;
 637                int bad_sectors;
 638                unsigned int pending;
 639                bool nonrot;
 640
 641                rdev = rcu_dereference(conf->mirrors[disk].rdev);
 642                if (r1_bio->bios[disk] == IO_BLOCKED
 643                    || rdev == NULL
 644                    || test_bit(Faulty, &rdev->flags))
 645                        continue;
 646                if (!test_bit(In_sync, &rdev->flags) &&
 647                    rdev->recovery_offset < this_sector + sectors)
 648                        continue;
 649                if (test_bit(WriteMostly, &rdev->flags)) {
 650                        /* Don't balance among write-mostly, just
 651                         * use the first as a last resort */
 652                        if (best_dist_disk < 0) {
 653                                if (is_badblock(rdev, this_sector, sectors,
 654                                                &first_bad, &bad_sectors)) {
 655                                        if (first_bad <= this_sector)
 656                                                /* Cannot use this */
 657                                                continue;
 658                                        best_good_sectors = first_bad - this_sector;
 659                                } else
 660                                        best_good_sectors = sectors;
 661                                best_dist_disk = disk;
 662                                best_pending_disk = disk;
 663                        }
 664                        continue;
 665                }
 666                /* This is a reasonable device to use.  It might
 667                 * even be best.
 668                 */
 669                if (is_badblock(rdev, this_sector, sectors,
 670                                &first_bad, &bad_sectors)) {
 671                        if (best_dist < MaxSector)
 672                                /* already have a better device */
 673                                continue;
 674                        if (first_bad <= this_sector) {
 675                                /* cannot read here. If this is the 'primary'
 676                                 * device, then we must not read beyond
 677                                 * bad_sectors from another device..
 678                                 */
 679                                bad_sectors -= (this_sector - first_bad);
 680                                if (choose_first && sectors > bad_sectors)
 681                                        sectors = bad_sectors;
 682                                if (best_good_sectors > sectors)
 683                                        best_good_sectors = sectors;
 684
 685                        } else {
 686                                sector_t good_sectors = first_bad - this_sector;
 687                                if (good_sectors > best_good_sectors) {
 688                                        best_good_sectors = good_sectors;
 689                                        best_disk = disk;
 690                                }
 691                                if (choose_first)
 692                                        break;
 693                        }
 694                        continue;
 695                } else {
 696                        if ((sectors > best_good_sectors) && (best_disk >= 0))
 697                                best_disk = -1;
 698                        best_good_sectors = sectors;
 699                }
 700
 701                if (best_disk >= 0)
 702                        /* At least two disks to choose from so failfast is OK */
 703                        set_bit(R1BIO_FailFast, &r1_bio->state);
 704
 705                nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 706                has_nonrot_disk |= nonrot;
 707                pending = atomic_read(&rdev->nr_pending);
 708                dist = abs(this_sector - conf->mirrors[disk].head_position);
 709                if (choose_first) {
 710                        best_disk = disk;
 711                        break;
 712                }
 713                /* Don't change to another disk for sequential reads */
 714                if (conf->mirrors[disk].next_seq_sect == this_sector
 715                    || dist == 0) {
 716                        int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 717                        struct raid1_info *mirror = &conf->mirrors[disk];
 718
 719                        best_disk = disk;
 720                        /*
 721                         * If buffered sequential IO size exceeds optimal
 722                         * iosize, check if there is idle disk. If yes, choose
 723                         * the idle disk. read_balance could already choose an
 724                         * idle disk before noticing it's a sequential IO in
 725                         * this disk. This doesn't matter because this disk
 726                         * will idle, next time it will be utilized after the
 727                         * first disk has IO size exceeds optimal iosize. In
 728                         * this way, iosize of the first disk will be optimal
 729                         * iosize at least. iosize of the second disk might be
 730                         * small, but not a big deal since when the second disk
 731                         * starts IO, the first disk is likely still busy.
 732                         */
 733                        if (nonrot && opt_iosize > 0 &&
 734                            mirror->seq_start != MaxSector &&
 735                            mirror->next_seq_sect > opt_iosize &&
 736                            mirror->next_seq_sect - opt_iosize >=
 737                            mirror->seq_start) {
 738                                choose_next_idle = 1;
 739                                continue;
 740                        }
 741                        break;
 742                }
 743
 744                if (choose_next_idle)
 745                        continue;
 746
 747                if (min_pending > pending) {
 748                        min_pending = pending;
 749                        best_pending_disk = disk;
 750                }
 751
 752                if (dist < best_dist) {
 753                        best_dist = dist;
 754                        best_dist_disk = disk;
 755                }
 756        }
 757
 758        /*
 759         * If all disks are rotational, choose the closest disk. If any disk is
 760         * non-rotational, choose the disk with less pending request even the
 761         * disk is rotational, which might/might not be optimal for raids with
 762         * mixed ratation/non-rotational disks depending on workload.
 763         */
 764        if (best_disk == -1) {
 765                if (has_nonrot_disk || min_pending == 0)
 766                        best_disk = best_pending_disk;
 767                else
 768                        best_disk = best_dist_disk;
 769        }
 770
 771        if (best_disk >= 0) {
 772                rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 773                if (!rdev)
 774                        goto retry;
 775                atomic_inc(&rdev->nr_pending);
 776                sectors = best_good_sectors;
 777
 778                if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 779                        conf->mirrors[best_disk].seq_start = this_sector;
 780
 781                conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 782        }
 783        rcu_read_unlock();
 784        *max_sectors = sectors;
 785
 786        return best_disk;
 787}
 788
 789static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 790{
 791        /* flush any pending bitmap writes to disk before proceeding w/ I/O */
 792        md_bitmap_unplug(conf->mddev->bitmap);
 793        wake_up(&conf->wait_barrier);
 794
 795        while (bio) { /* submit pending writes */
 796                struct bio *next = bio->bi_next;
 797                struct md_rdev *rdev = (void *)bio->bi_bdev;
 798                bio->bi_next = NULL;
 799                bio_set_dev(bio, rdev->bdev);
 800                if (test_bit(Faulty, &rdev->flags)) {
 801                        bio_io_error(bio);
 802                } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
 803                                    !blk_queue_discard(bio->bi_bdev->bd_disk->queue)))
 804                        /* Just ignore it */
 805                        bio_endio(bio);
 806                else
 807                        submit_bio_noacct(bio);
 808                bio = next;
 809                cond_resched();
 810        }
 811}
 812
 813static void flush_pending_writes(struct r1conf *conf)
 814{
 815        /* Any writes that have been queued but are awaiting
 816         * bitmap updates get flushed here.
 817         */
 818        spin_lock_irq(&conf->device_lock);
 819
 820        if (conf->pending_bio_list.head) {
 821                struct blk_plug plug;
 822                struct bio *bio;
 823
 824                bio = bio_list_get(&conf->pending_bio_list);
 825                conf->pending_count = 0;
 826                spin_unlock_irq(&conf->device_lock);
 827
 828                /*
 829                 * As this is called in a wait_event() loop (see freeze_array),
 830                 * current->state might be TASK_UNINTERRUPTIBLE which will
 831                 * cause a warning when we prepare to wait again.  As it is
 832                 * rare that this path is taken, it is perfectly safe to force
 833                 * us to go around the wait_event() loop again, so the warning
 834                 * is a false-positive.  Silence the warning by resetting
 835                 * thread state
 836                 */
 837                __set_current_state(TASK_RUNNING);
 838                blk_start_plug(&plug);
 839                flush_bio_list(conf, bio);
 840                blk_finish_plug(&plug);
 841        } else
 842                spin_unlock_irq(&conf->device_lock);
 843}
 844
 845/* Barriers....
 846 * Sometimes we need to suspend IO while we do something else,
 847 * either some resync/recovery, or reconfigure the array.
 848 * To do this we raise a 'barrier'.
 849 * The 'barrier' is a counter that can be raised multiple times
 850 * to count how many activities are happening which preclude
 851 * normal IO.
 852 * We can only raise the barrier if there is no pending IO.
 853 * i.e. if nr_pending == 0.
 854 * We choose only to raise the barrier if no-one is waiting for the
 855 * barrier to go down.  This means that as soon as an IO request
 856 * is ready, no other operations which require a barrier will start
 857 * until the IO request has had a chance.
 858 *
 859 * So: regular IO calls 'wait_barrier'.  When that returns there
 860 *    is no backgroup IO happening,  It must arrange to call
 861 *    allow_barrier when it has finished its IO.
 862 * backgroup IO calls must call raise_barrier.  Once that returns
 863 *    there is no normal IO happeing.  It must arrange to call
 864 *    lower_barrier when the particular background IO completes.
 865 *
 866 * If resync/recovery is interrupted, returns -EINTR;
 867 * Otherwise, returns 0.
 868 */
 869static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
 870{
 871        int idx = sector_to_idx(sector_nr);
 872
 873        spin_lock_irq(&conf->resync_lock);
 874
 875        /* Wait until no block IO is waiting */
 876        wait_event_lock_irq(conf->wait_barrier,
 877                            !atomic_read(&conf->nr_waiting[idx]),
 878                            conf->resync_lock);
 879
 880        /* block any new IO from starting */
 881        atomic_inc(&conf->barrier[idx]);
 882        /*
 883         * In raise_barrier() we firstly increase conf->barrier[idx] then
 884         * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 885         * increase conf->nr_pending[idx] then check conf->barrier[idx].
 886         * A memory barrier here to make sure conf->nr_pending[idx] won't
 887         * be fetched before conf->barrier[idx] is increased. Otherwise
 888         * there will be a race between raise_barrier() and _wait_barrier().
 889         */
 890        smp_mb__after_atomic();
 891
 892        /* For these conditions we must wait:
 893         * A: while the array is in frozen state
 894         * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 895         *    existing in corresponding I/O barrier bucket.
 896         * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 897         *    max resync count which allowed on current I/O barrier bucket.
 898         */
 899        wait_event_lock_irq(conf->wait_barrier,
 900                            (!conf->array_frozen &&
 901                             !atomic_read(&conf->nr_pending[idx]) &&
 902                             atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
 903                                test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
 904                            conf->resync_lock);
 905
 906        if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
 907                atomic_dec(&conf->barrier[idx]);
 908                spin_unlock_irq(&conf->resync_lock);
 909                wake_up(&conf->wait_barrier);
 910                return -EINTR;
 911        }
 912
 913        atomic_inc(&conf->nr_sync_pending);
 914        spin_unlock_irq(&conf->resync_lock);
 915
 916        return 0;
 917}
 918
 919static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 920{
 921        int idx = sector_to_idx(sector_nr);
 922
 923        BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 924
 925        atomic_dec(&conf->barrier[idx]);
 926        atomic_dec(&conf->nr_sync_pending);
 927        wake_up(&conf->wait_barrier);
 928}
 929
 930static void _wait_barrier(struct r1conf *conf, int idx)
 931{
 932        /*
 933         * We need to increase conf->nr_pending[idx] very early here,
 934         * then raise_barrier() can be blocked when it waits for
 935         * conf->nr_pending[idx] to be 0. Then we can avoid holding
 936         * conf->resync_lock when there is no barrier raised in same
 937         * barrier unit bucket. Also if the array is frozen, I/O
 938         * should be blocked until array is unfrozen.
 939         */
 940        atomic_inc(&conf->nr_pending[idx]);
 941        /*
 942         * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
 943         * check conf->barrier[idx]. In raise_barrier() we firstly increase
 944         * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
 945         * barrier is necessary here to make sure conf->barrier[idx] won't be
 946         * fetched before conf->nr_pending[idx] is increased. Otherwise there
 947         * will be a race between _wait_barrier() and raise_barrier().
 948         */
 949        smp_mb__after_atomic();
 950
 951        /*
 952         * Don't worry about checking two atomic_t variables at same time
 953         * here. If during we check conf->barrier[idx], the array is
 954         * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
 955         * 0, it is safe to return and make the I/O continue. Because the
 956         * array is frozen, all I/O returned here will eventually complete
 957         * or be queued, no race will happen. See code comment in
 958         * frozen_array().
 959         */
 960        if (!READ_ONCE(conf->array_frozen) &&
 961            !atomic_read(&conf->barrier[idx]))
 962                return;
 963
 964        /*
 965         * After holding conf->resync_lock, conf->nr_pending[idx]
 966         * should be decreased before waiting for barrier to drop.
 967         * Otherwise, we may encounter a race condition because
 968         * raise_barrer() might be waiting for conf->nr_pending[idx]
 969         * to be 0 at same time.
 970         */
 971        spin_lock_irq(&conf->resync_lock);
 972        atomic_inc(&conf->nr_waiting[idx]);
 973        atomic_dec(&conf->nr_pending[idx]);
 974        /*
 975         * In case freeze_array() is waiting for
 976         * get_unqueued_pending() == extra
 977         */
 978        wake_up(&conf->wait_barrier);
 979        /* Wait for the barrier in same barrier unit bucket to drop. */
 980        wait_event_lock_irq(conf->wait_barrier,
 981                            !conf->array_frozen &&
 982                             !atomic_read(&conf->barrier[idx]),
 983                            conf->resync_lock);
 984        atomic_inc(&conf->nr_pending[idx]);
 985        atomic_dec(&conf->nr_waiting[idx]);
 986        spin_unlock_irq(&conf->resync_lock);
 987}
 988
 989static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
 990{
 991        int idx = sector_to_idx(sector_nr);
 992
 993        /*
 994         * Very similar to _wait_barrier(). The difference is, for read
 995         * I/O we don't need wait for sync I/O, but if the whole array
 996         * is frozen, the read I/O still has to wait until the array is
 997         * unfrozen. Since there is no ordering requirement with
 998         * conf->barrier[idx] here, memory barrier is unnecessary as well.
 999         */
1000        atomic_inc(&conf->nr_pending[idx]);
1001
1002        if (!READ_ONCE(conf->array_frozen))
1003                return;
1004
1005        spin_lock_irq(&conf->resync_lock);
1006        atomic_inc(&conf->nr_waiting[idx]);
1007        atomic_dec(&conf->nr_pending[idx]);
1008        /*
1009         * In case freeze_array() is waiting for
1010         * get_unqueued_pending() == extra
1011         */
1012        wake_up(&conf->wait_barrier);
1013        /* Wait for array to be unfrozen */
1014        wait_event_lock_irq(conf->wait_barrier,
1015                            !conf->array_frozen,
1016                            conf->resync_lock);
1017        atomic_inc(&conf->nr_pending[idx]);
1018        atomic_dec(&conf->nr_waiting[idx]);
1019        spin_unlock_irq(&conf->resync_lock);
1020}
1021
1022static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1023{
1024        int idx = sector_to_idx(sector_nr);
1025
1026        _wait_barrier(conf, idx);
1027}
1028
1029static void _allow_barrier(struct r1conf *conf, int idx)
1030{
1031        atomic_dec(&conf->nr_pending[idx]);
1032        wake_up(&conf->wait_barrier);
1033}
1034
1035static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1036{
1037        int idx = sector_to_idx(sector_nr);
1038
1039        _allow_barrier(conf, idx);
1040}
1041
1042/* conf->resync_lock should be held */
1043static int get_unqueued_pending(struct r1conf *conf)
1044{
1045        int idx, ret;
1046
1047        ret = atomic_read(&conf->nr_sync_pending);
1048        for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1049                ret += atomic_read(&conf->nr_pending[idx]) -
1050                        atomic_read(&conf->nr_queued[idx]);
1051
1052        return ret;
1053}
1054
1055static void freeze_array(struct r1conf *conf, int extra)
1056{
1057        /* Stop sync I/O and normal I/O and wait for everything to
1058         * go quiet.
1059         * This is called in two situations:
1060         * 1) management command handlers (reshape, remove disk, quiesce).
1061         * 2) one normal I/O request failed.
1062
1063         * After array_frozen is set to 1, new sync IO will be blocked at
1064         * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1065         * or wait_read_barrier(). The flying I/Os will either complete or be
1066         * queued. When everything goes quite, there are only queued I/Os left.
1067
1068         * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1069         * barrier bucket index which this I/O request hits. When all sync and
1070         * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1071         * of all conf->nr_queued[]. But normal I/O failure is an exception,
1072         * in handle_read_error(), we may call freeze_array() before trying to
1073         * fix the read error. In this case, the error read I/O is not queued,
1074         * so get_unqueued_pending() == 1.
1075         *
1076         * Therefore before this function returns, we need to wait until
1077         * get_unqueued_pendings(conf) gets equal to extra. For
1078         * normal I/O context, extra is 1, in rested situations extra is 0.
1079         */
1080        spin_lock_irq(&conf->resync_lock);
1081        conf->array_frozen = 1;
1082        raid1_log(conf->mddev, "wait freeze");
1083        wait_event_lock_irq_cmd(
1084                conf->wait_barrier,
1085                get_unqueued_pending(conf) == extra,
1086                conf->resync_lock,
1087                flush_pending_writes(conf));
1088        spin_unlock_irq(&conf->resync_lock);
1089}
1090static void unfreeze_array(struct r1conf *conf)
1091{
1092        /* reverse the effect of the freeze */
1093        spin_lock_irq(&conf->resync_lock);
1094        conf->array_frozen = 0;
1095        spin_unlock_irq(&conf->resync_lock);
1096        wake_up(&conf->wait_barrier);
1097}
1098
1099static void alloc_behind_master_bio(struct r1bio *r1_bio,
1100                                           struct bio *bio)
1101{
1102        int size = bio->bi_iter.bi_size;
1103        unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1104        int i = 0;
1105        struct bio *behind_bio = NULL;
1106
1107        behind_bio = bio_alloc_bioset(GFP_NOIO, vcnt, &r1_bio->mddev->bio_set);
1108        if (!behind_bio)
1109                return;
1110
1111        /* discard op, we don't support writezero/writesame yet */
1112        if (!bio_has_data(bio)) {
1113                behind_bio->bi_iter.bi_size = size;
1114                goto skip_copy;
1115        }
1116
1117        behind_bio->bi_write_hint = bio->bi_write_hint;
1118
1119        while (i < vcnt && size) {
1120                struct page *page;
1121                int len = min_t(int, PAGE_SIZE, size);
1122
1123                page = alloc_page(GFP_NOIO);
1124                if (unlikely(!page))
1125                        goto free_pages;
1126
1127                bio_add_page(behind_bio, page, len, 0);
1128
1129                size -= len;
1130                i++;
1131        }
1132
1133        bio_copy_data(behind_bio, bio);
1134skip_copy:
1135        r1_bio->behind_master_bio = behind_bio;
1136        set_bit(R1BIO_BehindIO, &r1_bio->state);
1137
1138        return;
1139
1140free_pages:
1141        pr_debug("%dB behind alloc failed, doing sync I/O\n",
1142                 bio->bi_iter.bi_size);
1143        bio_free_pages(behind_bio);
1144        bio_put(behind_bio);
1145}
1146
1147struct raid1_plug_cb {
1148        struct blk_plug_cb      cb;
1149        struct bio_list         pending;
1150        int                     pending_cnt;
1151};
1152
1153static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1154{
1155        struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1156                                                  cb);
1157        struct mddev *mddev = plug->cb.data;
1158        struct r1conf *conf = mddev->private;
1159        struct bio *bio;
1160
1161        if (from_schedule || current->bio_list) {
1162                spin_lock_irq(&conf->device_lock);
1163                bio_list_merge(&conf->pending_bio_list, &plug->pending);
1164                conf->pending_count += plug->pending_cnt;
1165                spin_unlock_irq(&conf->device_lock);
1166                wake_up(&conf->wait_barrier);
1167                md_wakeup_thread(mddev->thread);
1168                kfree(plug);
1169                return;
1170        }
1171
1172        /* we aren't scheduling, so we can do the write-out directly. */
1173        bio = bio_list_get(&plug->pending);
1174        flush_bio_list(conf, bio);
1175        kfree(plug);
1176}
1177
1178static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1179{
1180        r1_bio->master_bio = bio;
1181        r1_bio->sectors = bio_sectors(bio);
1182        r1_bio->state = 0;
1183        r1_bio->mddev = mddev;
1184        r1_bio->sector = bio->bi_iter.bi_sector;
1185}
1186
1187static inline struct r1bio *
1188alloc_r1bio(struct mddev *mddev, struct bio *bio)
1189{
1190        struct r1conf *conf = mddev->private;
1191        struct r1bio *r1_bio;
1192
1193        r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1194        /* Ensure no bio records IO_BLOCKED */
1195        memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1196        init_r1bio(r1_bio, mddev, bio);
1197        return r1_bio;
1198}
1199
1200static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1201                               int max_read_sectors, struct r1bio *r1_bio)
1202{
1203        struct r1conf *conf = mddev->private;
1204        struct raid1_info *mirror;
1205        struct bio *read_bio;
1206        struct bitmap *bitmap = mddev->bitmap;
1207        const int op = bio_op(bio);
1208        const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1209        int max_sectors;
1210        int rdisk;
1211        bool print_msg = !!r1_bio;
1212        char b[BDEVNAME_SIZE];
1213
1214        /*
1215         * If r1_bio is set, we are blocking the raid1d thread
1216         * so there is a tiny risk of deadlock.  So ask for
1217         * emergency memory if needed.
1218         */
1219        gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1220
1221        if (print_msg) {
1222                /* Need to get the block device name carefully */
1223                struct md_rdev *rdev;
1224                rcu_read_lock();
1225                rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1226                if (rdev)
1227                        bdevname(rdev->bdev, b);
1228                else
1229                        strcpy(b, "???");
1230                rcu_read_unlock();
1231        }
1232
1233        /*
1234         * Still need barrier for READ in case that whole
1235         * array is frozen.
1236         */
1237        wait_read_barrier(conf, bio->bi_iter.bi_sector);
1238
1239        if (!r1_bio)
1240                r1_bio = alloc_r1bio(mddev, bio);
1241        else
1242                init_r1bio(r1_bio, mddev, bio);
1243        r1_bio->sectors = max_read_sectors;
1244
1245        /*
1246         * make_request() can abort the operation when read-ahead is being
1247         * used and no empty request is available.
1248         */
1249        rdisk = read_balance(conf, r1_bio, &max_sectors);
1250
1251        if (rdisk < 0) {
1252                /* couldn't find anywhere to read from */
1253                if (print_msg) {
1254                        pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1255                                            mdname(mddev),
1256                                            b,
1257                                            (unsigned long long)r1_bio->sector);
1258                }
1259                raid_end_bio_io(r1_bio);
1260                return;
1261        }
1262        mirror = conf->mirrors + rdisk;
1263
1264        if (print_msg)
1265                pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1266                                    mdname(mddev),
1267                                    (unsigned long long)r1_bio->sector,
1268                                    bdevname(mirror->rdev->bdev, b));
1269
1270        if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1271            bitmap) {
1272                /*
1273                 * Reading from a write-mostly device must take care not to
1274                 * over-take any writes that are 'behind'
1275                 */
1276                raid1_log(mddev, "wait behind writes");
1277                wait_event(bitmap->behind_wait,
1278                           atomic_read(&bitmap->behind_writes) == 0);
1279        }
1280
1281        if (max_sectors < bio_sectors(bio)) {
1282                struct bio *split = bio_split(bio, max_sectors,
1283                                              gfp, &conf->bio_split);
1284                bio_chain(split, bio);
1285                submit_bio_noacct(bio);
1286                bio = split;
1287                r1_bio->master_bio = bio;
1288                r1_bio->sectors = max_sectors;
1289        }
1290
1291        r1_bio->read_disk = rdisk;
1292
1293        read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1294
1295        r1_bio->bios[rdisk] = read_bio;
1296
1297        read_bio->bi_iter.bi_sector = r1_bio->sector +
1298                mirror->rdev->data_offset;
1299        bio_set_dev(read_bio, mirror->rdev->bdev);
1300        read_bio->bi_end_io = raid1_end_read_request;
1301        bio_set_op_attrs(read_bio, op, do_sync);
1302        if (test_bit(FailFast, &mirror->rdev->flags) &&
1303            test_bit(R1BIO_FailFast, &r1_bio->state))
1304                read_bio->bi_opf |= MD_FAILFAST;
1305        read_bio->bi_private = r1_bio;
1306
1307        if (mddev->gendisk)
1308                trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1309                                      r1_bio->sector);
1310
1311        submit_bio_noacct(read_bio);
1312}
1313
1314static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1315                                int max_write_sectors)
1316{
1317        struct r1conf *conf = mddev->private;
1318        struct r1bio *r1_bio;
1319        int i, disks;
1320        struct bitmap *bitmap = mddev->bitmap;
1321        unsigned long flags;
1322        struct md_rdev *blocked_rdev;
1323        struct blk_plug_cb *cb;
1324        struct raid1_plug_cb *plug = NULL;
1325        int first_clone;
1326        int max_sectors;
1327
1328        if (mddev_is_clustered(mddev) &&
1329             md_cluster_ops->area_resyncing(mddev, WRITE,
1330                     bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1331
1332                DEFINE_WAIT(w);
1333                for (;;) {
1334                        prepare_to_wait(&conf->wait_barrier,
1335                                        &w, TASK_IDLE);
1336                        if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1337                                                        bio->bi_iter.bi_sector,
1338                                                        bio_end_sector(bio)))
1339                                break;
1340                        schedule();
1341                }
1342                finish_wait(&conf->wait_barrier, &w);
1343        }
1344
1345        /*
1346         * Register the new request and wait if the reconstruction
1347         * thread has put up a bar for new requests.
1348         * Continue immediately if no resync is active currently.
1349         */
1350        wait_barrier(conf, bio->bi_iter.bi_sector);
1351
1352        r1_bio = alloc_r1bio(mddev, bio);
1353        r1_bio->sectors = max_write_sectors;
1354
1355        if (conf->pending_count >= max_queued_requests) {
1356                md_wakeup_thread(mddev->thread);
1357                raid1_log(mddev, "wait queued");
1358                wait_event(conf->wait_barrier,
1359                           conf->pending_count < max_queued_requests);
1360        }
1361        /* first select target devices under rcu_lock and
1362         * inc refcount on their rdev.  Record them by setting
1363         * bios[x] to bio
1364         * If there are known/acknowledged bad blocks on any device on
1365         * which we have seen a write error, we want to avoid writing those
1366         * blocks.
1367         * This potentially requires several writes to write around
1368         * the bad blocks.  Each set of writes gets it's own r1bio
1369         * with a set of bios attached.
1370         */
1371
1372        disks = conf->raid_disks * 2;
1373 retry_write:
1374        blocked_rdev = NULL;
1375        rcu_read_lock();
1376        max_sectors = r1_bio->sectors;
1377        for (i = 0;  i < disks; i++) {
1378                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1379                if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1380                        atomic_inc(&rdev->nr_pending);
1381                        blocked_rdev = rdev;
1382                        break;
1383                }
1384                r1_bio->bios[i] = NULL;
1385                if (!rdev || test_bit(Faulty, &rdev->flags)) {
1386                        if (i < conf->raid_disks)
1387                                set_bit(R1BIO_Degraded, &r1_bio->state);
1388                        continue;
1389                }
1390
1391                atomic_inc(&rdev->nr_pending);
1392                if (test_bit(WriteErrorSeen, &rdev->flags)) {
1393                        sector_t first_bad;
1394                        int bad_sectors;
1395                        int is_bad;
1396
1397                        is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1398                                             &first_bad, &bad_sectors);
1399                        if (is_bad < 0) {
1400                                /* mustn't write here until the bad block is
1401                                 * acknowledged*/
1402                                set_bit(BlockedBadBlocks, &rdev->flags);
1403                                blocked_rdev = rdev;
1404                                break;
1405                        }
1406                        if (is_bad && first_bad <= r1_bio->sector) {
1407                                /* Cannot write here at all */
1408                                bad_sectors -= (r1_bio->sector - first_bad);
1409                                if (bad_sectors < max_sectors)
1410                                        /* mustn't write more than bad_sectors
1411                                         * to other devices yet
1412                                         */
1413                                        max_sectors = bad_sectors;
1414                                rdev_dec_pending(rdev, mddev);
1415                                /* We don't set R1BIO_Degraded as that
1416                                 * only applies if the disk is
1417                                 * missing, so it might be re-added,
1418                                 * and we want to know to recover this
1419                                 * chunk.
1420                                 * In this case the device is here,
1421                                 * and the fact that this chunk is not
1422                                 * in-sync is recorded in the bad
1423                                 * block log
1424                                 */
1425                                continue;
1426                        }
1427                        if (is_bad) {
1428                                int good_sectors = first_bad - r1_bio->sector;
1429                                if (good_sectors < max_sectors)
1430                                        max_sectors = good_sectors;
1431                        }
1432                }
1433                r1_bio->bios[i] = bio;
1434        }
1435        rcu_read_unlock();
1436
1437        if (unlikely(blocked_rdev)) {
1438                /* Wait for this device to become unblocked */
1439                int j;
1440
1441                for (j = 0; j < i; j++)
1442                        if (r1_bio->bios[j])
1443                                rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1444                r1_bio->state = 0;
1445                allow_barrier(conf, bio->bi_iter.bi_sector);
1446                raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1447                md_wait_for_blocked_rdev(blocked_rdev, mddev);
1448                wait_barrier(conf, bio->bi_iter.bi_sector);
1449                goto retry_write;
1450        }
1451
1452        if (max_sectors < bio_sectors(bio)) {
1453                struct bio *split = bio_split(bio, max_sectors,
1454                                              GFP_NOIO, &conf->bio_split);
1455                bio_chain(split, bio);
1456                submit_bio_noacct(bio);
1457                bio = split;
1458                r1_bio->master_bio = bio;
1459                r1_bio->sectors = max_sectors;
1460        }
1461
1462        atomic_set(&r1_bio->remaining, 1);
1463        atomic_set(&r1_bio->behind_remaining, 0);
1464
1465        first_clone = 1;
1466
1467        for (i = 0; i < disks; i++) {
1468                struct bio *mbio = NULL;
1469                struct md_rdev *rdev = conf->mirrors[i].rdev;
1470                if (!r1_bio->bios[i])
1471                        continue;
1472
1473                if (first_clone) {
1474                        /* do behind I/O ?
1475                         * Not if there are too many, or cannot
1476                         * allocate memory, or a reader on WriteMostly
1477                         * is waiting for behind writes to flush */
1478                        if (bitmap &&
1479                            (atomic_read(&bitmap->behind_writes)
1480                             < mddev->bitmap_info.max_write_behind) &&
1481                            !waitqueue_active(&bitmap->behind_wait)) {
1482                                alloc_behind_master_bio(r1_bio, bio);
1483                        }
1484
1485                        md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1486                                             test_bit(R1BIO_BehindIO, &r1_bio->state));
1487                        first_clone = 0;
1488                }
1489
1490                if (r1_bio->behind_master_bio)
1491                        mbio = bio_clone_fast(r1_bio->behind_master_bio,
1492                                              GFP_NOIO, &mddev->bio_set);
1493                else
1494                        mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1495
1496                if (r1_bio->behind_master_bio) {
1497                        if (test_bit(CollisionCheck, &rdev->flags))
1498                                wait_for_serialization(rdev, r1_bio);
1499                        if (test_bit(WriteMostly, &rdev->flags))
1500                                atomic_inc(&r1_bio->behind_remaining);
1501                } else if (mddev->serialize_policy)
1502                        wait_for_serialization(rdev, r1_bio);
1503
1504                r1_bio->bios[i] = mbio;
1505
1506                mbio->bi_iter.bi_sector = (r1_bio->sector +
1507                                   conf->mirrors[i].rdev->data_offset);
1508                bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1509                mbio->bi_end_io = raid1_end_write_request;
1510                mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1511                if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1512                    !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1513                    conf->raid_disks - mddev->degraded > 1)
1514                        mbio->bi_opf |= MD_FAILFAST;
1515                mbio->bi_private = r1_bio;
1516
1517                atomic_inc(&r1_bio->remaining);
1518
1519                if (mddev->gendisk)
1520                        trace_block_bio_remap(mbio, disk_devt(mddev->gendisk),
1521                                              r1_bio->sector);
1522                /* flush_pending_writes() needs access to the rdev so...*/
1523                mbio->bi_bdev = (void *)conf->mirrors[i].rdev;
1524
1525                cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1526                if (cb)
1527                        plug = container_of(cb, struct raid1_plug_cb, cb);
1528                else
1529                        plug = NULL;
1530                if (plug) {
1531                        bio_list_add(&plug->pending, mbio);
1532                        plug->pending_cnt++;
1533                } else {
1534                        spin_lock_irqsave(&conf->device_lock, flags);
1535                        bio_list_add(&conf->pending_bio_list, mbio);
1536                        conf->pending_count++;
1537                        spin_unlock_irqrestore(&conf->device_lock, flags);
1538                        md_wakeup_thread(mddev->thread);
1539                }
1540        }
1541
1542        r1_bio_write_done(r1_bio);
1543
1544        /* In case raid1d snuck in to freeze_array */
1545        wake_up(&conf->wait_barrier);
1546}
1547
1548static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1549{
1550        sector_t sectors;
1551
1552        if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1553            && md_flush_request(mddev, bio))
1554                return true;
1555
1556        /*
1557         * There is a limit to the maximum size, but
1558         * the read/write handler might find a lower limit
1559         * due to bad blocks.  To avoid multiple splits,
1560         * we pass the maximum number of sectors down
1561         * and let the lower level perform the split.
1562         */
1563        sectors = align_to_barrier_unit_end(
1564                bio->bi_iter.bi_sector, bio_sectors(bio));
1565
1566        if (bio_data_dir(bio) == READ)
1567                raid1_read_request(mddev, bio, sectors, NULL);
1568        else {
1569                if (!md_write_start(mddev,bio))
1570                        return false;
1571                raid1_write_request(mddev, bio, sectors);
1572        }
1573        return true;
1574}
1575
1576static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1577{
1578        struct r1conf *conf = mddev->private;
1579        int i;
1580
1581        seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1582                   conf->raid_disks - mddev->degraded);
1583        rcu_read_lock();
1584        for (i = 0; i < conf->raid_disks; i++) {
1585                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1586                seq_printf(seq, "%s",
1587                           rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1588        }
1589        rcu_read_unlock();
1590        seq_printf(seq, "]");
1591}
1592
1593static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1594{
1595        char b[BDEVNAME_SIZE];
1596        struct r1conf *conf = mddev->private;
1597        unsigned long flags;
1598
1599        /*
1600         * If it is not operational, then we have already marked it as dead
1601         * else if it is the last working disks with "fail_last_dev == false",
1602         * ignore the error, let the next level up know.
1603         * else mark the drive as failed
1604         */
1605        spin_lock_irqsave(&conf->device_lock, flags);
1606        if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1607            && (conf->raid_disks - mddev->degraded) == 1) {
1608                /*
1609                 * Don't fail the drive, act as though we were just a
1610                 * normal single drive.
1611                 * However don't try a recovery from this drive as
1612                 * it is very likely to fail.
1613                 */
1614                conf->recovery_disabled = mddev->recovery_disabled;
1615                spin_unlock_irqrestore(&conf->device_lock, flags);
1616                return;
1617        }
1618        set_bit(Blocked, &rdev->flags);
1619        if (test_and_clear_bit(In_sync, &rdev->flags))
1620                mddev->degraded++;
1621        set_bit(Faulty, &rdev->flags);
1622        spin_unlock_irqrestore(&conf->device_lock, flags);
1623        /*
1624         * if recovery is running, make sure it aborts.
1625         */
1626        set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1627        set_mask_bits(&mddev->sb_flags, 0,
1628                      BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1629        pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1630                "md/raid1:%s: Operation continuing on %d devices.\n",
1631                mdname(mddev), bdevname(rdev->bdev, b),
1632                mdname(mddev), conf->raid_disks - mddev->degraded);
1633}
1634
1635static void print_conf(struct r1conf *conf)
1636{
1637        int i;
1638
1639        pr_debug("RAID1 conf printout:\n");
1640        if (!conf) {
1641                pr_debug("(!conf)\n");
1642                return;
1643        }
1644        pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1645                 conf->raid_disks);
1646
1647        rcu_read_lock();
1648        for (i = 0; i < conf->raid_disks; i++) {
1649                char b[BDEVNAME_SIZE];
1650                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1651                if (rdev)
1652                        pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1653                                 i, !test_bit(In_sync, &rdev->flags),
1654                                 !test_bit(Faulty, &rdev->flags),
1655                                 bdevname(rdev->bdev,b));
1656        }
1657        rcu_read_unlock();
1658}
1659
1660static void close_sync(struct r1conf *conf)
1661{
1662        int idx;
1663
1664        for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1665                _wait_barrier(conf, idx);
1666                _allow_barrier(conf, idx);
1667        }
1668
1669        mempool_exit(&conf->r1buf_pool);
1670}
1671
1672static int raid1_spare_active(struct mddev *mddev)
1673{
1674        int i;
1675        struct r1conf *conf = mddev->private;
1676        int count = 0;
1677        unsigned long flags;
1678
1679        /*
1680         * Find all failed disks within the RAID1 configuration
1681         * and mark them readable.
1682         * Called under mddev lock, so rcu protection not needed.
1683         * device_lock used to avoid races with raid1_end_read_request
1684         * which expects 'In_sync' flags and ->degraded to be consistent.
1685         */
1686        spin_lock_irqsave(&conf->device_lock, flags);
1687        for (i = 0; i < conf->raid_disks; i++) {
1688                struct md_rdev *rdev = conf->mirrors[i].rdev;
1689                struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1690                if (repl
1691                    && !test_bit(Candidate, &repl->flags)
1692                    && repl->recovery_offset == MaxSector
1693                    && !test_bit(Faulty, &repl->flags)
1694                    && !test_and_set_bit(In_sync, &repl->flags)) {
1695                        /* replacement has just become active */
1696                        if (!rdev ||
1697                            !test_and_clear_bit(In_sync, &rdev->flags))
1698                                count++;
1699                        if (rdev) {
1700                                /* Replaced device not technically
1701                                 * faulty, but we need to be sure
1702                                 * it gets removed and never re-added
1703                                 */
1704                                set_bit(Faulty, &rdev->flags);
1705                                sysfs_notify_dirent_safe(
1706                                        rdev->sysfs_state);
1707                        }
1708                }
1709                if (rdev
1710                    && rdev->recovery_offset == MaxSector
1711                    && !test_bit(Faulty, &rdev->flags)
1712                    && !test_and_set_bit(In_sync, &rdev->flags)) {
1713                        count++;
1714                        sysfs_notify_dirent_safe(rdev->sysfs_state);
1715                }
1716        }
1717        mddev->degraded -= count;
1718        spin_unlock_irqrestore(&conf->device_lock, flags);
1719
1720        print_conf(conf);
1721        return count;
1722}
1723
1724static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1725{
1726        struct r1conf *conf = mddev->private;
1727        int err = -EEXIST;
1728        int mirror = 0;
1729        struct raid1_info *p;
1730        int first = 0;
1731        int last = conf->raid_disks - 1;
1732
1733        if (mddev->recovery_disabled == conf->recovery_disabled)
1734                return -EBUSY;
1735
1736        if (md_integrity_add_rdev(rdev, mddev))
1737                return -ENXIO;
1738
1739        if (rdev->raid_disk >= 0)
1740                first = last = rdev->raid_disk;
1741
1742        /*
1743         * find the disk ... but prefer rdev->saved_raid_disk
1744         * if possible.
1745         */
1746        if (rdev->saved_raid_disk >= 0 &&
1747            rdev->saved_raid_disk >= first &&
1748            rdev->saved_raid_disk < conf->raid_disks &&
1749            conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1750                first = last = rdev->saved_raid_disk;
1751
1752        for (mirror = first; mirror <= last; mirror++) {
1753                p = conf->mirrors + mirror;
1754                if (!p->rdev) {
1755                        if (mddev->gendisk)
1756                                disk_stack_limits(mddev->gendisk, rdev->bdev,
1757                                                  rdev->data_offset << 9);
1758
1759                        p->head_position = 0;
1760                        rdev->raid_disk = mirror;
1761                        err = 0;
1762                        /* As all devices are equivalent, we don't need a full recovery
1763                         * if this was recently any drive of the array
1764                         */
1765                        if (rdev->saved_raid_disk < 0)
1766                                conf->fullsync = 1;
1767                        rcu_assign_pointer(p->rdev, rdev);
1768                        break;
1769                }
1770                if (test_bit(WantReplacement, &p->rdev->flags) &&
1771                    p[conf->raid_disks].rdev == NULL) {
1772                        /* Add this device as a replacement */
1773                        clear_bit(In_sync, &rdev->flags);
1774                        set_bit(Replacement, &rdev->flags);
1775                        rdev->raid_disk = mirror;
1776                        err = 0;
1777                        conf->fullsync = 1;
1778                        rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1779                        break;
1780                }
1781        }
1782        if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1783                blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1784        print_conf(conf);
1785        return err;
1786}
1787
1788static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1789{
1790        struct r1conf *conf = mddev->private;
1791        int err = 0;
1792        int number = rdev->raid_disk;
1793        struct raid1_info *p = conf->mirrors + number;
1794
1795        if (rdev != p->rdev)
1796                p = conf->mirrors + conf->raid_disks + number;
1797
1798        print_conf(conf);
1799        if (rdev == p->rdev) {
1800                if (test_bit(In_sync, &rdev->flags) ||
1801                    atomic_read(&rdev->nr_pending)) {
1802                        err = -EBUSY;
1803                        goto abort;
1804                }
1805                /* Only remove non-faulty devices if recovery
1806                 * is not possible.
1807                 */
1808                if (!test_bit(Faulty, &rdev->flags) &&
1809                    mddev->recovery_disabled != conf->recovery_disabled &&
1810                    mddev->degraded < conf->raid_disks) {
1811                        err = -EBUSY;
1812                        goto abort;
1813                }
1814                p->rdev = NULL;
1815                if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1816                        synchronize_rcu();
1817                        if (atomic_read(&rdev->nr_pending)) {
1818                                /* lost the race, try later */
1819                                err = -EBUSY;
1820                                p->rdev = rdev;
1821                                goto abort;
1822                        }
1823                }
1824                if (conf->mirrors[conf->raid_disks + number].rdev) {
1825                        /* We just removed a device that is being replaced.
1826                         * Move down the replacement.  We drain all IO before
1827                         * doing this to avoid confusion.
1828                         */
1829                        struct md_rdev *repl =
1830                                conf->mirrors[conf->raid_disks + number].rdev;
1831                        freeze_array(conf, 0);
1832                        if (atomic_read(&repl->nr_pending)) {
1833                                /* It means that some queued IO of retry_list
1834                                 * hold repl. Thus, we cannot set replacement
1835                                 * as NULL, avoiding rdev NULL pointer
1836                                 * dereference in sync_request_write and
1837                                 * handle_write_finished.
1838                                 */
1839                                err = -EBUSY;
1840                                unfreeze_array(conf);
1841                                goto abort;
1842                        }
1843                        clear_bit(Replacement, &repl->flags);
1844                        p->rdev = repl;
1845                        conf->mirrors[conf->raid_disks + number].rdev = NULL;
1846                        unfreeze_array(conf);
1847                }
1848
1849                clear_bit(WantReplacement, &rdev->flags);
1850                err = md_integrity_register(mddev);
1851        }
1852abort:
1853
1854        print_conf(conf);
1855        return err;
1856}
1857
1858static void end_sync_read(struct bio *bio)
1859{
1860        struct r1bio *r1_bio = get_resync_r1bio(bio);
1861
1862        update_head_pos(r1_bio->read_disk, r1_bio);
1863
1864        /*
1865         * we have read a block, now it needs to be re-written,
1866         * or re-read if the read failed.
1867         * We don't do much here, just schedule handling by raid1d
1868         */
1869        if (!bio->bi_status)
1870                set_bit(R1BIO_Uptodate, &r1_bio->state);
1871
1872        if (atomic_dec_and_test(&r1_bio->remaining))
1873                reschedule_retry(r1_bio);
1874}
1875
1876static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1877{
1878        sector_t sync_blocks = 0;
1879        sector_t s = r1_bio->sector;
1880        long sectors_to_go = r1_bio->sectors;
1881
1882        /* make sure these bits don't get cleared. */
1883        do {
1884                md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1885                s += sync_blocks;
1886                sectors_to_go -= sync_blocks;
1887        } while (sectors_to_go > 0);
1888}
1889
1890static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1891{
1892        if (atomic_dec_and_test(&r1_bio->remaining)) {
1893                struct mddev *mddev = r1_bio->mddev;
1894                int s = r1_bio->sectors;
1895
1896                if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1897                    test_bit(R1BIO_WriteError, &r1_bio->state))
1898                        reschedule_retry(r1_bio);
1899                else {
1900                        put_buf(r1_bio);
1901                        md_done_sync(mddev, s, uptodate);
1902                }
1903        }
1904}
1905
1906static void end_sync_write(struct bio *bio)
1907{
1908        int uptodate = !bio->bi_status;
1909        struct r1bio *r1_bio = get_resync_r1bio(bio);
1910        struct mddev *mddev = r1_bio->mddev;
1911        struct r1conf *conf = mddev->private;
1912        sector_t first_bad;
1913        int bad_sectors;
1914        struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1915
1916        if (!uptodate) {
1917                abort_sync_write(mddev, r1_bio);
1918                set_bit(WriteErrorSeen, &rdev->flags);
1919                if (!test_and_set_bit(WantReplacement, &rdev->flags))
1920                        set_bit(MD_RECOVERY_NEEDED, &
1921                                mddev->recovery);
1922                set_bit(R1BIO_WriteError, &r1_bio->state);
1923        } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1924                               &first_bad, &bad_sectors) &&
1925                   !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1926                                r1_bio->sector,
1927                                r1_bio->sectors,
1928                                &first_bad, &bad_sectors)
1929                )
1930                set_bit(R1BIO_MadeGood, &r1_bio->state);
1931
1932        put_sync_write_buf(r1_bio, uptodate);
1933}
1934
1935static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1936                            int sectors, struct page *page, int rw)
1937{
1938        if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1939                /* success */
1940                return 1;
1941        if (rw == WRITE) {
1942                set_bit(WriteErrorSeen, &rdev->flags);
1943                if (!test_and_set_bit(WantReplacement,
1944                                      &rdev->flags))
1945                        set_bit(MD_RECOVERY_NEEDED, &
1946                                rdev->mddev->recovery);
1947        }
1948        /* need to record an error - either for the block or the device */
1949        if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1950                md_error(rdev->mddev, rdev);
1951        return 0;
1952}
1953
1954static int fix_sync_read_error(struct r1bio *r1_bio)
1955{
1956        /* Try some synchronous reads of other devices to get
1957         * good data, much like with normal read errors.  Only
1958         * read into the pages we already have so we don't
1959         * need to re-issue the read request.
1960         * We don't need to freeze the array, because being in an
1961         * active sync request, there is no normal IO, and
1962         * no overlapping syncs.
1963         * We don't need to check is_badblock() again as we
1964         * made sure that anything with a bad block in range
1965         * will have bi_end_io clear.
1966         */
1967        struct mddev *mddev = r1_bio->mddev;
1968        struct r1conf *conf = mddev->private;
1969        struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1970        struct page **pages = get_resync_pages(bio)->pages;
1971        sector_t sect = r1_bio->sector;
1972        int sectors = r1_bio->sectors;
1973        int idx = 0;
1974        struct md_rdev *rdev;
1975
1976        rdev = conf->mirrors[r1_bio->read_disk].rdev;
1977        if (test_bit(FailFast, &rdev->flags)) {
1978                /* Don't try recovering from here - just fail it
1979                 * ... unless it is the last working device of course */
1980                md_error(mddev, rdev);
1981                if (test_bit(Faulty, &rdev->flags))
1982                        /* Don't try to read from here, but make sure
1983                         * put_buf does it's thing
1984                         */
1985                        bio->bi_end_io = end_sync_write;
1986        }
1987
1988        while(sectors) {
1989                int s = sectors;
1990                int d = r1_bio->read_disk;
1991                int success = 0;
1992                int start;
1993
1994                if (s > (PAGE_SIZE>>9))
1995                        s = PAGE_SIZE >> 9;
1996                do {
1997                        if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1998                                /* No rcu protection needed here devices
1999                                 * can only be removed when no resync is
2000                                 * active, and resync is currently active
2001                                 */
2002                                rdev = conf->mirrors[d].rdev;
2003                                if (sync_page_io(rdev, sect, s<<9,
2004                                                 pages[idx],
2005                                                 REQ_OP_READ, 0, false)) {
2006                                        success = 1;
2007                                        break;
2008                                }
2009                        }
2010                        d++;
2011                        if (d == conf->raid_disks * 2)
2012                                d = 0;
2013                } while (!success && d != r1_bio->read_disk);
2014
2015                if (!success) {
2016                        char b[BDEVNAME_SIZE];
2017                        int abort = 0;
2018                        /* Cannot read from anywhere, this block is lost.
2019                         * Record a bad block on each device.  If that doesn't
2020                         * work just disable and interrupt the recovery.
2021                         * Don't fail devices as that won't really help.
2022                         */
2023                        pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2024                                            mdname(mddev), bio_devname(bio, b),
2025                                            (unsigned long long)r1_bio->sector);
2026                        for (d = 0; d < conf->raid_disks * 2; d++) {
2027                                rdev = conf->mirrors[d].rdev;
2028                                if (!rdev || test_bit(Faulty, &rdev->flags))
2029                                        continue;
2030                                if (!rdev_set_badblocks(rdev, sect, s, 0))
2031                                        abort = 1;
2032                        }
2033                        if (abort) {
2034                                conf->recovery_disabled =
2035                                        mddev->recovery_disabled;
2036                                set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2037                                md_done_sync(mddev, r1_bio->sectors, 0);
2038                                put_buf(r1_bio);
2039                                return 0;
2040                        }
2041                        /* Try next page */
2042                        sectors -= s;
2043                        sect += s;
2044                        idx++;
2045                        continue;
2046                }
2047
2048                start = d;
2049                /* write it back and re-read */
2050                while (d != r1_bio->read_disk) {
2051                        if (d == 0)
2052                                d = conf->raid_disks * 2;
2053                        d--;
2054                        if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2055                                continue;
2056                        rdev = conf->mirrors[d].rdev;
2057                        if (r1_sync_page_io(rdev, sect, s,
2058                                            pages[idx],
2059                                            WRITE) == 0) {
2060                                r1_bio->bios[d]->bi_end_io = NULL;
2061                                rdev_dec_pending(rdev, mddev);
2062                        }
2063                }
2064                d = start;
2065                while (d != r1_bio->read_disk) {
2066                        if (d == 0)
2067                                d = conf->raid_disks * 2;
2068                        d--;
2069                        if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2070                                continue;
2071                        rdev = conf->mirrors[d].rdev;
2072                        if (r1_sync_page_io(rdev, sect, s,
2073                                            pages[idx],
2074                                            READ) != 0)
2075                                atomic_add(s, &rdev->corrected_errors);
2076                }
2077                sectors -= s;
2078                sect += s;
2079                idx ++;
2080        }
2081        set_bit(R1BIO_Uptodate, &r1_bio->state);
2082        bio->bi_status = 0;
2083        return 1;
2084}
2085
2086static void process_checks(struct r1bio *r1_bio)
2087{
2088        /* We have read all readable devices.  If we haven't
2089         * got the block, then there is no hope left.
2090         * If we have, then we want to do a comparison
2091         * and skip the write if everything is the same.
2092         * If any blocks failed to read, then we need to
2093         * attempt an over-write
2094         */
2095        struct mddev *mddev = r1_bio->mddev;
2096        struct r1conf *conf = mddev->private;
2097        int primary;
2098        int i;
2099        int vcnt;
2100
2101        /* Fix variable parts of all bios */
2102        vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2103        for (i = 0; i < conf->raid_disks * 2; i++) {
2104                blk_status_t status;
2105                struct bio *b = r1_bio->bios[i];
2106                struct resync_pages *rp = get_resync_pages(b);
2107                if (b->bi_end_io != end_sync_read)
2108                        continue;
2109                /* fixup the bio for reuse, but preserve errno */
2110                status = b->bi_status;
2111                bio_reset(b);
2112                b->bi_status = status;
2113                b->bi_iter.bi_sector = r1_bio->sector +
2114                        conf->mirrors[i].rdev->data_offset;
2115                bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2116                b->bi_end_io = end_sync_read;
2117                rp->raid_bio = r1_bio;
2118                b->bi_private = rp;
2119
2120                /* initialize bvec table again */
2121                md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2122        }
2123        for (primary = 0; primary < conf->raid_disks * 2; primary++)
2124                if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2125                    !r1_bio->bios[primary]->bi_status) {
2126                        r1_bio->bios[primary]->bi_end_io = NULL;
2127                        rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2128                        break;
2129                }
2130        r1_bio->read_disk = primary;
2131        for (i = 0; i < conf->raid_disks * 2; i++) {
2132                int j = 0;
2133                struct bio *pbio = r1_bio->bios[primary];
2134                struct bio *sbio = r1_bio->bios[i];
2135                blk_status_t status = sbio->bi_status;
2136                struct page **ppages = get_resync_pages(pbio)->pages;
2137                struct page **spages = get_resync_pages(sbio)->pages;
2138                struct bio_vec *bi;
2139                int page_len[RESYNC_PAGES] = { 0 };
2140                struct bvec_iter_all iter_all;
2141
2142                if (sbio->bi_end_io != end_sync_read)
2143                        continue;
2144                /* Now we can 'fixup' the error value */
2145                sbio->bi_status = 0;
2146
2147                bio_for_each_segment_all(bi, sbio, iter_all)
2148                        page_len[j++] = bi->bv_len;
2149
2150                if (!status) {
2151                        for (j = vcnt; j-- ; ) {
2152                                if (memcmp(page_address(ppages[j]),
2153                                           page_address(spages[j]),
2154                                           page_len[j]))
2155                                        break;
2156                        }
2157                } else
2158                        j = 0;
2159                if (j >= 0)
2160                        atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2161                if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2162                              && !status)) {
2163                        /* No need to write to this device. */
2164                        sbio->bi_end_io = NULL;
2165                        rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2166                        continue;
2167                }
2168
2169                bio_copy_data(sbio, pbio);
2170        }
2171}
2172
2173static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2174{
2175        struct r1conf *conf = mddev->private;
2176        int i;
2177        int disks = conf->raid_disks * 2;
2178        struct bio *wbio;
2179
2180        if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2181                /* ouch - failed to read all of that. */
2182                if (!fix_sync_read_error(r1_bio))
2183                        return;
2184
2185        if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2186                process_checks(r1_bio);
2187
2188        /*
2189         * schedule writes
2190         */
2191        atomic_set(&r1_bio->remaining, 1);
2192        for (i = 0; i < disks ; i++) {
2193                wbio = r1_bio->bios[i];
2194                if (wbio->bi_end_io == NULL ||
2195                    (wbio->bi_end_io == end_sync_read &&
2196                     (i == r1_bio->read_disk ||
2197                      !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2198                        continue;
2199                if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2200                        abort_sync_write(mddev, r1_bio);
2201                        continue;
2202                }
2203
2204                bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2205                if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2206                        wbio->bi_opf |= MD_FAILFAST;
2207
2208                wbio->bi_end_io = end_sync_write;
2209                atomic_inc(&r1_bio->remaining);
2210                md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2211
2212                submit_bio_noacct(wbio);
2213        }
2214
2215        put_sync_write_buf(r1_bio, 1);
2216}
2217
2218/*
2219 * This is a kernel thread which:
2220 *
2221 *      1.      Retries failed read operations on working mirrors.
2222 *      2.      Updates the raid superblock when problems encounter.
2223 *      3.      Performs writes following reads for array synchronising.
2224 */
2225
2226static void fix_read_error(struct r1conf *conf, int read_disk,
2227                           sector_t sect, int sectors)
2228{
2229        struct mddev *mddev = conf->mddev;
2230        while(sectors) {
2231                int s = sectors;
2232                int d = read_disk;
2233                int success = 0;
2234                int start;
2235                struct md_rdev *rdev;
2236
2237                if (s > (PAGE_SIZE>>9))
2238                        s = PAGE_SIZE >> 9;
2239
2240                do {
2241                        sector_t first_bad;
2242                        int bad_sectors;
2243
2244                        rcu_read_lock();
2245                        rdev = rcu_dereference(conf->mirrors[d].rdev);
2246                        if (rdev &&
2247                            (test_bit(In_sync, &rdev->flags) ||
2248                             (!test_bit(Faulty, &rdev->flags) &&
2249                              rdev->recovery_offset >= sect + s)) &&
2250                            is_badblock(rdev, sect, s,
2251                                        &first_bad, &bad_sectors) == 0) {
2252                                atomic_inc(&rdev->nr_pending);
2253                                rcu_read_unlock();
2254                                if (sync_page_io(rdev, sect, s<<9,
2255                                         conf->tmppage, REQ_OP_READ, 0, false))
2256                                        success = 1;
2257                                rdev_dec_pending(rdev, mddev);
2258                                if (success)
2259                                        break;
2260                        } else
2261                                rcu_read_unlock();
2262                        d++;
2263                        if (d == conf->raid_disks * 2)
2264                                d = 0;
2265                } while (!success && d != read_disk);
2266
2267                if (!success) {
2268                        /* Cannot read from anywhere - mark it bad */
2269                        struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2270                        if (!rdev_set_badblocks(rdev, sect, s, 0))
2271                                md_error(mddev, rdev);
2272                        break;
2273                }
2274                /* write it back and re-read */
2275                start = d;
2276                while (d != read_disk) {
2277                        if (d==0)
2278                                d = conf->raid_disks * 2;
2279                        d--;
2280                        rcu_read_lock();
2281                        rdev = rcu_dereference(conf->mirrors[d].rdev);
2282                        if (rdev &&
2283                            !test_bit(Faulty, &rdev->flags)) {
2284                                atomic_inc(&rdev->nr_pending);
2285                                rcu_read_unlock();
2286                                r1_sync_page_io(rdev, sect, s,
2287                                                conf->tmppage, WRITE);
2288                                rdev_dec_pending(rdev, mddev);
2289                        } else
2290                                rcu_read_unlock();
2291                }
2292                d = start;
2293                while (d != read_disk) {
2294                        char b[BDEVNAME_SIZE];
2295                        if (d==0)
2296                                d = conf->raid_disks * 2;
2297                        d--;
2298                        rcu_read_lock();
2299                        rdev = rcu_dereference(conf->mirrors[d].rdev);
2300                        if (rdev &&
2301                            !test_bit(Faulty, &rdev->flags)) {
2302                                atomic_inc(&rdev->nr_pending);
2303                                rcu_read_unlock();
2304                                if (r1_sync_page_io(rdev, sect, s,
2305                                                    conf->tmppage, READ)) {
2306                                        atomic_add(s, &rdev->corrected_errors);
2307                                        pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2308                                                mdname(mddev), s,
2309                                                (unsigned long long)(sect +
2310                                                                     rdev->data_offset),
2311                                                bdevname(rdev->bdev, b));
2312                                }
2313                                rdev_dec_pending(rdev, mddev);
2314                        } else
2315                                rcu_read_unlock();
2316                }
2317                sectors -= s;
2318                sect += s;
2319        }
2320}
2321
2322static int narrow_write_error(struct r1bio *r1_bio, int i)
2323{
2324        struct mddev *mddev = r1_bio->mddev;
2325        struct r1conf *conf = mddev->private;
2326        struct md_rdev *rdev = conf->mirrors[i].rdev;
2327
2328        /* bio has the data to be written to device 'i' where
2329         * we just recently had a write error.
2330         * We repeatedly clone the bio and trim down to one block,
2331         * then try the write.  Where the write fails we record
2332         * a bad block.
2333         * It is conceivable that the bio doesn't exactly align with
2334         * blocks.  We must handle this somehow.
2335         *
2336         * We currently own a reference on the rdev.
2337         */
2338
2339        int block_sectors;
2340        sector_t sector;
2341        int sectors;
2342        int sect_to_write = r1_bio->sectors;
2343        int ok = 1;
2344
2345        if (rdev->badblocks.shift < 0)
2346                return 0;
2347
2348        block_sectors = roundup(1 << rdev->badblocks.shift,
2349                                bdev_logical_block_size(rdev->bdev) >> 9);
2350        sector = r1_bio->sector;
2351        sectors = ((sector + block_sectors)
2352                   & ~(sector_t)(block_sectors - 1))
2353                - sector;
2354
2355        while (sect_to_write) {
2356                struct bio *wbio;
2357                if (sectors > sect_to_write)
2358                        sectors = sect_to_write;
2359                /* Write at 'sector' for 'sectors'*/
2360
2361                if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2362                        wbio = bio_clone_fast(r1_bio->behind_master_bio,
2363                                              GFP_NOIO,
2364                                              &mddev->bio_set);
2365                } else {
2366                        wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2367                                              &mddev->bio_set);
2368                }
2369
2370                bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2371                wbio->bi_iter.bi_sector = r1_bio->sector;
2372                wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2373
2374                bio_trim(wbio, sector - r1_bio->sector, sectors);
2375                wbio->bi_iter.bi_sector += rdev->data_offset;
2376                bio_set_dev(wbio, rdev->bdev);
2377
2378                if (submit_bio_wait(wbio) < 0)
2379                        /* failure! */
2380                        ok = rdev_set_badblocks(rdev, sector,
2381                                                sectors, 0)
2382                                && ok;
2383
2384                bio_put(wbio);
2385                sect_to_write -= sectors;
2386                sector += sectors;
2387                sectors = block_sectors;
2388        }
2389        return ok;
2390}
2391
2392static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2393{
2394        int m;
2395        int s = r1_bio->sectors;
2396        for (m = 0; m < conf->raid_disks * 2 ; m++) {
2397                struct md_rdev *rdev = conf->mirrors[m].rdev;
2398                struct bio *bio = r1_bio->bios[m];
2399                if (bio->bi_end_io == NULL)
2400                        continue;
2401                if (!bio->bi_status &&
2402                    test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2403                        rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2404                }
2405                if (bio->bi_status &&
2406                    test_bit(R1BIO_WriteError, &r1_bio->state)) {
2407                        if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2408                                md_error(conf->mddev, rdev);
2409                }
2410        }
2411        put_buf(r1_bio);
2412        md_done_sync(conf->mddev, s, 1);
2413}
2414
2415static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2416{
2417        int m, idx;
2418        bool fail = false;
2419
2420        for (m = 0; m < conf->raid_disks * 2 ; m++)
2421                if (r1_bio->bios[m] == IO_MADE_GOOD) {
2422                        struct md_rdev *rdev = conf->mirrors[m].rdev;
2423                        rdev_clear_badblocks(rdev,
2424                                             r1_bio->sector,
2425                                             r1_bio->sectors, 0);
2426                        rdev_dec_pending(rdev, conf->mddev);
2427                } else if (r1_bio->bios[m] != NULL) {
2428                        /* This drive got a write error.  We need to
2429                         * narrow down and record precise write
2430                         * errors.
2431                         */
2432                        fail = true;
2433                        if (!narrow_write_error(r1_bio, m)) {
2434                                md_error(conf->mddev,
2435                                         conf->mirrors[m].rdev);
2436                                /* an I/O failed, we can't clear the bitmap */
2437                                set_bit(R1BIO_Degraded, &r1_bio->state);
2438                        }
2439                        rdev_dec_pending(conf->mirrors[m].rdev,
2440                                         conf->mddev);
2441                }
2442        if (fail) {
2443                spin_lock_irq(&conf->device_lock);
2444                list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2445                idx = sector_to_idx(r1_bio->sector);
2446                atomic_inc(&conf->nr_queued[idx]);
2447                spin_unlock_irq(&conf->device_lock);
2448                /*
2449                 * In case freeze_array() is waiting for condition
2450                 * get_unqueued_pending() == extra to be true.
2451                 */
2452                wake_up(&conf->wait_barrier);
2453                md_wakeup_thread(conf->mddev->thread);
2454        } else {
2455                if (test_bit(R1BIO_WriteError, &r1_bio->state))
2456                        close_write(r1_bio);
2457                raid_end_bio_io(r1_bio);
2458        }
2459}
2460
2461static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2462{
2463        struct mddev *mddev = conf->mddev;
2464        struct bio *bio;
2465        struct md_rdev *rdev;
2466
2467        clear_bit(R1BIO_ReadError, &r1_bio->state);
2468        /* we got a read error. Maybe the drive is bad.  Maybe just
2469         * the block and we can fix it.
2470         * We freeze all other IO, and try reading the block from
2471         * other devices.  When we find one, we re-write
2472         * and check it that fixes the read error.
2473         * This is all done synchronously while the array is
2474         * frozen
2475         */
2476
2477        bio = r1_bio->bios[r1_bio->read_disk];
2478        bio_put(bio);
2479        r1_bio->bios[r1_bio->read_disk] = NULL;
2480
2481        rdev = conf->mirrors[r1_bio->read_disk].rdev;
2482        if (mddev->ro == 0
2483            && !test_bit(FailFast, &rdev->flags)) {
2484                freeze_array(conf, 1);
2485                fix_read_error(conf, r1_bio->read_disk,
2486                               r1_bio->sector, r1_bio->sectors);
2487                unfreeze_array(conf);
2488        } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2489                md_error(mddev, rdev);
2490        } else {
2491                r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2492        }
2493
2494        rdev_dec_pending(rdev, conf->mddev);
2495        allow_barrier(conf, r1_bio->sector);
2496        bio = r1_bio->master_bio;
2497
2498        /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2499        r1_bio->state = 0;
2500        raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2501}
2502
2503static void raid1d(struct md_thread *thread)
2504{
2505        struct mddev *mddev = thread->mddev;
2506        struct r1bio *r1_bio;
2507        unsigned long flags;
2508        struct r1conf *conf = mddev->private;
2509        struct list_head *head = &conf->retry_list;
2510        struct blk_plug plug;
2511        int idx;
2512
2513        md_check_recovery(mddev);
2514
2515        if (!list_empty_careful(&conf->bio_end_io_list) &&
2516            !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2517                LIST_HEAD(tmp);
2518                spin_lock_irqsave(&conf->device_lock, flags);
2519                if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2520                        list_splice_init(&conf->bio_end_io_list, &tmp);
2521                spin_unlock_irqrestore(&conf->device_lock, flags);
2522                while (!list_empty(&tmp)) {
2523                        r1_bio = list_first_entry(&tmp, struct r1bio,
2524                                                  retry_list);
2525                        list_del(&r1_bio->retry_list);
2526                        idx = sector_to_idx(r1_bio->sector);
2527                        atomic_dec(&conf->nr_queued[idx]);
2528                        if (mddev->degraded)
2529                                set_bit(R1BIO_Degraded, &r1_bio->state);
2530                        if (test_bit(R1BIO_WriteError, &r1_bio->state))
2531                                close_write(r1_bio);
2532                        raid_end_bio_io(r1_bio);
2533                }
2534        }
2535
2536        blk_start_plug(&plug);
2537        for (;;) {
2538
2539                flush_pending_writes(conf);
2540
2541                spin_lock_irqsave(&conf->device_lock, flags);
2542                if (list_empty(head)) {
2543                        spin_unlock_irqrestore(&conf->device_lock, flags);
2544                        break;
2545                }
2546                r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2547                list_del(head->prev);
2548                idx = sector_to_idx(r1_bio->sector);
2549                atomic_dec(&conf->nr_queued[idx]);
2550                spin_unlock_irqrestore(&conf->device_lock, flags);
2551
2552                mddev = r1_bio->mddev;
2553                conf = mddev->private;
2554                if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2555                        if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2556                            test_bit(R1BIO_WriteError, &r1_bio->state))
2557                                handle_sync_write_finished(conf, r1_bio);
2558                        else
2559                                sync_request_write(mddev, r1_bio);
2560                } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2561                           test_bit(R1BIO_WriteError, &r1_bio->state))
2562                        handle_write_finished(conf, r1_bio);
2563                else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2564                        handle_read_error(conf, r1_bio);
2565                else
2566                        WARN_ON_ONCE(1);
2567
2568                cond_resched();
2569                if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2570                        md_check_recovery(mddev);
2571        }
2572        blk_finish_plug(&plug);
2573}
2574
2575static int init_resync(struct r1conf *conf)
2576{
2577        int buffs;
2578
2579        buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2580        BUG_ON(mempool_initialized(&conf->r1buf_pool));
2581
2582        return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2583                            r1buf_pool_free, conf->poolinfo);
2584}
2585
2586static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2587{
2588        struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2589        struct resync_pages *rps;
2590        struct bio *bio;
2591        int i;
2592
2593        for (i = conf->poolinfo->raid_disks; i--; ) {
2594                bio = r1bio->bios[i];
2595                rps = bio->bi_private;
2596                bio_reset(bio);
2597                bio->bi_private = rps;
2598        }
2599        r1bio->master_bio = NULL;
2600        return r1bio;
2601}
2602
2603/*
2604 * perform a "sync" on one "block"
2605 *
2606 * We need to make sure that no normal I/O request - particularly write
2607 * requests - conflict with active sync requests.
2608 *
2609 * This is achieved by tracking pending requests and a 'barrier' concept
2610 * that can be installed to exclude normal IO requests.
2611 */
2612
2613static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2614                                   int *skipped)
2615{
2616        struct r1conf *conf = mddev->private;
2617        struct r1bio *r1_bio;
2618        struct bio *bio;
2619        sector_t max_sector, nr_sectors;
2620        int disk = -1;
2621        int i;
2622        int wonly = -1;
2623        int write_targets = 0, read_targets = 0;
2624        sector_t sync_blocks;
2625        int still_degraded = 0;
2626        int good_sectors = RESYNC_SECTORS;
2627        int min_bad = 0; /* number of sectors that are bad in all devices */
2628        int idx = sector_to_idx(sector_nr);
2629        int page_idx = 0;
2630
2631        if (!mempool_initialized(&conf->r1buf_pool))
2632                if (init_resync(conf))
2633                        return 0;
2634
2635        max_sector = mddev->dev_sectors;
2636        if (sector_nr >= max_sector) {
2637                /* If we aborted, we need to abort the
2638                 * sync on the 'current' bitmap chunk (there will
2639                 * only be one in raid1 resync.
2640                 * We can find the current addess in mddev->curr_resync
2641                 */
2642                if (mddev->curr_resync < max_sector) /* aborted */
2643                        md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2644                                           &sync_blocks, 1);
2645                else /* completed sync */
2646                        conf->fullsync = 0;
2647
2648                md_bitmap_close_sync(mddev->bitmap);
2649                close_sync(conf);
2650
2651                if (mddev_is_clustered(mddev)) {
2652                        conf->cluster_sync_low = 0;
2653                        conf->cluster_sync_high = 0;
2654                }
2655                return 0;
2656        }
2657
2658        if (mddev->bitmap == NULL &&
2659            mddev->recovery_cp == MaxSector &&
2660            !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2661            conf->fullsync == 0) {
2662                *skipped = 1;
2663                return max_sector - sector_nr;
2664        }
2665        /* before building a request, check if we can skip these blocks..
2666         * This call the bitmap_start_sync doesn't actually record anything
2667         */
2668        if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2669            !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2670                /* We can skip this block, and probably several more */
2671                *skipped = 1;
2672                return sync_blocks;
2673        }
2674
2675        /*
2676         * If there is non-resync activity waiting for a turn, then let it
2677         * though before starting on this new sync request.
2678         */
2679        if (atomic_read(&conf->nr_waiting[idx]))
2680                schedule_timeout_uninterruptible(1);
2681
2682        /* we are incrementing sector_nr below. To be safe, we check against
2683         * sector_nr + two times RESYNC_SECTORS
2684         */
2685
2686        md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2687                mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2688
2689
2690        if (raise_barrier(conf, sector_nr))
2691                return 0;
2692
2693        r1_bio = raid1_alloc_init_r1buf(conf);
2694
2695        rcu_read_lock();
2696        /*
2697         * If we get a correctably read error during resync or recovery,
2698         * we might want to read from a different device.  So we
2699         * flag all drives that could conceivably be read from for READ,
2700         * and any others (which will be non-In_sync devices) for WRITE.
2701         * If a read fails, we try reading from something else for which READ
2702         * is OK.
2703         */
2704
2705        r1_bio->mddev = mddev;
2706        r1_bio->sector = sector_nr;
2707        r1_bio->state = 0;
2708        set_bit(R1BIO_IsSync, &r1_bio->state);
2709        /* make sure good_sectors won't go across barrier unit boundary */
2710        good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2711
2712        for (i = 0; i < conf->raid_disks * 2; i++) {
2713                struct md_rdev *rdev;
2714                bio = r1_bio->bios[i];
2715
2716                rdev = rcu_dereference(conf->mirrors[i].rdev);
2717                if (rdev == NULL ||
2718                    test_bit(Faulty, &rdev->flags)) {
2719                        if (i < conf->raid_disks)
2720                                still_degraded = 1;
2721                } else if (!test_bit(In_sync, &rdev->flags)) {
2722                        bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2723                        bio->bi_end_io = end_sync_write;
2724                        write_targets ++;
2725                } else {
2726                        /* may need to read from here */
2727                        sector_t first_bad = MaxSector;
2728                        int bad_sectors;
2729
2730                        if (is_badblock(rdev, sector_nr, good_sectors,
2731                                        &first_bad, &bad_sectors)) {
2732                                if (first_bad > sector_nr)
2733                                        good_sectors = first_bad - sector_nr;
2734                                else {
2735                                        bad_sectors -= (sector_nr - first_bad);
2736                                        if (min_bad == 0 ||
2737                                            min_bad > bad_sectors)
2738                                                min_bad = bad_sectors;
2739                                }
2740                        }
2741                        if (sector_nr < first_bad) {
2742                                if (test_bit(WriteMostly, &rdev->flags)) {
2743                                        if (wonly < 0)
2744                                                wonly = i;
2745                                } else {
2746                                        if (disk < 0)
2747                                                disk = i;
2748                                }
2749                                bio_set_op_attrs(bio, REQ_OP_READ, 0);
2750                                bio->bi_end_io = end_sync_read;
2751                                read_targets++;
2752                        } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2753                                test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2754                                !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2755                                /*
2756                                 * The device is suitable for reading (InSync),
2757                                 * but has bad block(s) here. Let's try to correct them,
2758                                 * if we are doing resync or repair. Otherwise, leave
2759                                 * this device alone for this sync request.
2760                                 */
2761                                bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2762                                bio->bi_end_io = end_sync_write;
2763                                write_targets++;
2764                        }
2765                }
2766                if (rdev && bio->bi_end_io) {
2767                        atomic_inc(&rdev->nr_pending);
2768                        bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2769                        bio_set_dev(bio, rdev->bdev);
2770                        if (test_bit(FailFast, &rdev->flags))
2771                                bio->bi_opf |= MD_FAILFAST;
2772                }
2773        }
2774        rcu_read_unlock();
2775        if (disk < 0)
2776                disk = wonly;
2777        r1_bio->read_disk = disk;
2778
2779        if (read_targets == 0 && min_bad > 0) {
2780                /* These sectors are bad on all InSync devices, so we
2781                 * need to mark them bad on all write targets
2782                 */
2783                int ok = 1;
2784                for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2785                        if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2786                                struct md_rdev *rdev = conf->mirrors[i].rdev;
2787                                ok = rdev_set_badblocks(rdev, sector_nr,
2788                                                        min_bad, 0
2789                                        ) && ok;
2790                        }
2791                set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2792                *skipped = 1;
2793                put_buf(r1_bio);
2794
2795                if (!ok) {
2796                        /* Cannot record the badblocks, so need to
2797                         * abort the resync.
2798                         * If there are multiple read targets, could just
2799                         * fail the really bad ones ???
2800                         */
2801                        conf->recovery_disabled = mddev->recovery_disabled;
2802                        set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2803                        return 0;
2804                } else
2805                        return min_bad;
2806
2807        }
2808        if (min_bad > 0 && min_bad < good_sectors) {
2809                /* only resync enough to reach the next bad->good
2810                 * transition */
2811                good_sectors = min_bad;
2812        }
2813
2814        if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2815                /* extra read targets are also write targets */
2816                write_targets += read_targets-1;
2817
2818        if (write_targets == 0 || read_targets == 0) {
2819                /* There is nowhere to write, so all non-sync
2820                 * drives must be failed - so we are finished
2821                 */
2822                sector_t rv;
2823                if (min_bad > 0)
2824                        max_sector = sector_nr + min_bad;
2825                rv = max_sector - sector_nr;
2826                *skipped = 1;
2827                put_buf(r1_bio);
2828                return rv;
2829        }
2830
2831        if (max_sector > mddev->resync_max)
2832                max_sector = mddev->resync_max; /* Don't do IO beyond here */
2833        if (max_sector > sector_nr + good_sectors)
2834                max_sector = sector_nr + good_sectors;
2835        nr_sectors = 0;
2836        sync_blocks = 0;
2837        do {
2838                struct page *page;
2839                int len = PAGE_SIZE;
2840                if (sector_nr + (len>>9) > max_sector)
2841                        len = (max_sector - sector_nr) << 9;
2842                if (len == 0)
2843                        break;
2844                if (sync_blocks == 0) {
2845                        if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2846                                                  &sync_blocks, still_degraded) &&
2847                            !conf->fullsync &&
2848                            !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2849                                break;
2850                        if ((len >> 9) > sync_blocks)
2851                                len = sync_blocks<<9;
2852                }
2853
2854                for (i = 0 ; i < conf->raid_disks * 2; i++) {
2855                        struct resync_pages *rp;
2856
2857                        bio = r1_bio->bios[i];
2858                        rp = get_resync_pages(bio);
2859                        if (bio->bi_end_io) {
2860                                page = resync_fetch_page(rp, page_idx);
2861
2862                                /*
2863                                 * won't fail because the vec table is big
2864                                 * enough to hold all these pages
2865                                 */
2866                                bio_add_page(bio, page, len, 0);
2867                        }
2868                }
2869                nr_sectors += len>>9;
2870                sector_nr += len>>9;
2871                sync_blocks -= (len>>9);
2872        } while (++page_idx < RESYNC_PAGES);
2873
2874        r1_bio->sectors = nr_sectors;
2875
2876        if (mddev_is_clustered(mddev) &&
2877                        conf->cluster_sync_high < sector_nr + nr_sectors) {
2878                conf->cluster_sync_low = mddev->curr_resync_completed;
2879                conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2880                /* Send resync message */
2881                md_cluster_ops->resync_info_update(mddev,
2882                                conf->cluster_sync_low,
2883                                conf->cluster_sync_high);
2884        }
2885
2886        /* For a user-requested sync, we read all readable devices and do a
2887         * compare
2888         */
2889        if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2890                atomic_set(&r1_bio->remaining, read_targets);
2891                for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2892                        bio = r1_bio->bios[i];
2893                        if (bio->bi_end_io == end_sync_read) {
2894                                read_targets--;
2895                                md_sync_acct_bio(bio, nr_sectors);
2896                                if (read_targets == 1)
2897                                        bio->bi_opf &= ~MD_FAILFAST;
2898                                submit_bio_noacct(bio);
2899                        }
2900                }
2901        } else {
2902                atomic_set(&r1_bio->remaining, 1);
2903                bio = r1_bio->bios[r1_bio->read_disk];
2904                md_sync_acct_bio(bio, nr_sectors);
2905                if (read_targets == 1)
2906                        bio->bi_opf &= ~MD_FAILFAST;
2907                submit_bio_noacct(bio);
2908        }
2909        return nr_sectors;
2910}
2911
2912static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2913{
2914        if (sectors)
2915                return sectors;
2916
2917        return mddev->dev_sectors;
2918}
2919
2920static struct r1conf *setup_conf(struct mddev *mddev)
2921{
2922        struct r1conf *conf;
2923        int i;
2924        struct raid1_info *disk;
2925        struct md_rdev *rdev;
2926        int err = -ENOMEM;
2927
2928        conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2929        if (!conf)
2930                goto abort;
2931
2932        conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2933                                   sizeof(atomic_t), GFP_KERNEL);
2934        if (!conf->nr_pending)
2935                goto abort;
2936
2937        conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2938                                   sizeof(atomic_t), GFP_KERNEL);
2939        if (!conf->nr_waiting)
2940                goto abort;
2941
2942        conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2943                                  sizeof(atomic_t), GFP_KERNEL);
2944        if (!conf->nr_queued)
2945                goto abort;
2946
2947        conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2948                                sizeof(atomic_t), GFP_KERNEL);
2949        if (!conf->barrier)
2950                goto abort;
2951
2952        conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2953                                            mddev->raid_disks, 2),
2954                                GFP_KERNEL);
2955        if (!conf->mirrors)
2956                goto abort;
2957
2958        conf->tmppage = alloc_page(GFP_KERNEL);
2959        if (!conf->tmppage)
2960                goto abort;
2961
2962        conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2963        if (!conf->poolinfo)
2964                goto abort;
2965        conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2966        err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2967                           rbio_pool_free, conf->poolinfo);
2968        if (err)
2969                goto abort;
2970
2971        err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2972        if (err)
2973                goto abort;
2974
2975        conf->poolinfo->mddev = mddev;
2976
2977        err = -EINVAL;
2978        spin_lock_init(&conf->device_lock);
2979        rdev_for_each(rdev, mddev) {
2980                int disk_idx = rdev->raid_disk;
2981                if (disk_idx >= mddev->raid_disks
2982                    || disk_idx < 0)
2983                        continue;
2984                if (test_bit(Replacement, &rdev->flags))
2985                        disk = conf->mirrors + mddev->raid_disks + disk_idx;
2986                else
2987                        disk = conf->mirrors + disk_idx;
2988
2989                if (disk->rdev)
2990                        goto abort;
2991                disk->rdev = rdev;
2992                disk->head_position = 0;
2993                disk->seq_start = MaxSector;
2994        }
2995        conf->raid_disks = mddev->raid_disks;
2996        conf->mddev = mddev;
2997        INIT_LIST_HEAD(&conf->retry_list);
2998        INIT_LIST_HEAD(&conf->bio_end_io_list);
2999
3000        spin_lock_init(&conf->resync_lock);
3001        init_waitqueue_head(&conf->wait_barrier);
3002
3003        bio_list_init(&conf->pending_bio_list);
3004        conf->pending_count = 0;
3005        conf->recovery_disabled = mddev->recovery_disabled - 1;
3006
3007        err = -EIO;
3008        for (i = 0; i < conf->raid_disks * 2; i++) {
3009
3010                disk = conf->mirrors + i;
3011
3012                if (i < conf->raid_disks &&
3013                    disk[conf->raid_disks].rdev) {
3014                        /* This slot has a replacement. */
3015                        if (!disk->rdev) {
3016                                /* No original, just make the replacement
3017                                 * a recovering spare
3018                                 */
3019                                disk->rdev =
3020                                        disk[conf->raid_disks].rdev;
3021                                disk[conf->raid_disks].rdev = NULL;
3022                        } else if (!test_bit(In_sync, &disk->rdev->flags))
3023                                /* Original is not in_sync - bad */
3024                                goto abort;
3025                }
3026
3027                if (!disk->rdev ||
3028                    !test_bit(In_sync, &disk->rdev->flags)) {
3029                        disk->head_position = 0;
3030                        if (disk->rdev &&
3031                            (disk->rdev->saved_raid_disk < 0))
3032                                conf->fullsync = 1;
3033                }
3034        }
3035
3036        err = -ENOMEM;
3037        conf->thread = md_register_thread(raid1d, mddev, "raid1");
3038        if (!conf->thread)
3039                goto abort;
3040
3041        return conf;
3042
3043 abort:
3044        if (conf) {
3045                mempool_exit(&conf->r1bio_pool);
3046                kfree(conf->mirrors);
3047                safe_put_page(conf->tmppage);
3048                kfree(conf->poolinfo);
3049                kfree(conf->nr_pending);
3050                kfree(conf->nr_waiting);
3051                kfree(conf->nr_queued);
3052                kfree(conf->barrier);
3053                bioset_exit(&conf->bio_split);
3054                kfree(conf);
3055        }
3056        return ERR_PTR(err);
3057}
3058
3059static void raid1_free(struct mddev *mddev, void *priv);
3060static int raid1_run(struct mddev *mddev)
3061{
3062        struct r1conf *conf;
3063        int i;
3064        struct md_rdev *rdev;
3065        int ret;
3066        bool discard_supported = false;
3067
3068        if (mddev->level != 1) {
3069                pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3070                        mdname(mddev), mddev->level);
3071                return -EIO;
3072        }
3073        if (mddev->reshape_position != MaxSector) {
3074                pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3075                        mdname(mddev));
3076                return -EIO;
3077        }
3078        if (mddev_init_writes_pending(mddev) < 0)
3079                return -ENOMEM;
3080        /*
3081         * copy the already verified devices into our private RAID1
3082         * bookkeeping area. [whatever we allocate in run(),
3083         * should be freed in raid1_free()]
3084         */
3085        if (mddev->private == NULL)
3086                conf = setup_conf(mddev);
3087        else
3088                conf = mddev->private;
3089
3090        if (IS_ERR(conf))
3091                return PTR_ERR(conf);
3092
3093        if (mddev->queue) {
3094                blk_queue_max_write_same_sectors(mddev->queue, 0);
3095                blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3096        }
3097
3098        rdev_for_each(rdev, mddev) {
3099                if (!mddev->gendisk)
3100                        continue;
3101                disk_stack_limits(mddev->gendisk, rdev->bdev,
3102                                  rdev->data_offset << 9);
3103                if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3104                        discard_supported = true;
3105        }
3106
3107        mddev->degraded = 0;
3108        for (i = 0; i < conf->raid_disks; i++)
3109                if (conf->mirrors[i].rdev == NULL ||
3110                    !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3111                    test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3112                        mddev->degraded++;
3113        /*
3114         * RAID1 needs at least one disk in active
3115         */
3116        if (conf->raid_disks - mddev->degraded < 1) {
3117                ret = -EINVAL;
3118                goto abort;
3119        }
3120
3121        if (conf->raid_disks - mddev->degraded == 1)
3122                mddev->recovery_cp = MaxSector;
3123
3124        if (mddev->recovery_cp != MaxSector)
3125                pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3126                        mdname(mddev));
3127        pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3128                mdname(mddev), mddev->raid_disks - mddev->degraded,
3129                mddev->raid_disks);
3130
3131        /*
3132         * Ok, everything is just fine now
3133         */
3134        mddev->thread = conf->thread;
3135        conf->thread = NULL;
3136        mddev->private = conf;
3137        set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3138
3139        md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3140
3141        if (mddev->queue) {
3142                if (discard_supported)
3143                        blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3144                                                mddev->queue);
3145                else
3146                        blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3147                                                  mddev->queue);
3148        }
3149
3150        ret = md_integrity_register(mddev);
3151        if (ret) {
3152                md_unregister_thread(&mddev->thread);
3153                goto abort;
3154        }
3155        return 0;
3156
3157abort:
3158        raid1_free(mddev, conf);
3159        return ret;
3160}
3161
3162static void raid1_free(struct mddev *mddev, void *priv)
3163{
3164        struct r1conf *conf = priv;
3165
3166        mempool_exit(&conf->r1bio_pool);
3167        kfree(conf->mirrors);
3168        safe_put_page(conf->tmppage);
3169        kfree(conf->poolinfo);
3170        kfree(conf->nr_pending);
3171        kfree(conf->nr_waiting);
3172        kfree(conf->nr_queued);
3173        kfree(conf->barrier);
3174        bioset_exit(&conf->bio_split);
3175        kfree(conf);
3176}
3177
3178static int raid1_resize(struct mddev *mddev, sector_t sectors)
3179{
3180        /* no resync is happening, and there is enough space
3181         * on all devices, so we can resize.
3182         * We need to make sure resync covers any new space.
3183         * If the array is shrinking we should possibly wait until
3184         * any io in the removed space completes, but it hardly seems
3185         * worth it.
3186         */
3187        sector_t newsize = raid1_size(mddev, sectors, 0);
3188        if (mddev->external_size &&
3189            mddev->array_sectors > newsize)
3190                return -EINVAL;
3191        if (mddev->bitmap) {
3192                int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3193                if (ret)
3194                        return ret;
3195        }
3196        md_set_array_sectors(mddev, newsize);
3197        if (sectors > mddev->dev_sectors &&
3198            mddev->recovery_cp > mddev->dev_sectors) {
3199                mddev->recovery_cp = mddev->dev_sectors;
3200                set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3201        }
3202        mddev->dev_sectors = sectors;
3203        mddev->resync_max_sectors = sectors;
3204        return 0;
3205}
3206
3207static int raid1_reshape(struct mddev *mddev)
3208{
3209        /* We need to:
3210         * 1/ resize the r1bio_pool
3211         * 2/ resize conf->mirrors
3212         *
3213         * We allocate a new r1bio_pool if we can.
3214         * Then raise a device barrier and wait until all IO stops.
3215         * Then resize conf->mirrors and swap in the new r1bio pool.
3216         *
3217         * At the same time, we "pack" the devices so that all the missing
3218         * devices have the higher raid_disk numbers.
3219         */
3220        mempool_t newpool, oldpool;
3221        struct pool_info *newpoolinfo;
3222        struct raid1_info *newmirrors;
3223        struct r1conf *conf = mddev->private;
3224        int cnt, raid_disks;
3225        unsigned long flags;
3226        int d, d2;
3227        int ret;
3228
3229        memset(&newpool, 0, sizeof(newpool));
3230        memset(&oldpool, 0, sizeof(oldpool));
3231
3232        /* Cannot change chunk_size, layout, or level */
3233        if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3234            mddev->layout != mddev->new_layout ||
3235            mddev->level != mddev->new_level) {
3236                mddev->new_chunk_sectors = mddev->chunk_sectors;
3237                mddev->new_layout = mddev->layout;
3238                mddev->new_level = mddev->level;
3239                return -EINVAL;
3240        }
3241
3242        if (!mddev_is_clustered(mddev))
3243                md_allow_write(mddev);
3244
3245        raid_disks = mddev->raid_disks + mddev->delta_disks;
3246
3247        if (raid_disks < conf->raid_disks) {
3248                cnt=0;
3249                for (d= 0; d < conf->raid_disks; d++)
3250                        if (conf->mirrors[d].rdev)
3251                                cnt++;
3252                if (cnt > raid_disks)
3253                        return -EBUSY;
3254        }
3255
3256        newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3257        if (!newpoolinfo)
3258                return -ENOMEM;
3259        newpoolinfo->mddev = mddev;
3260        newpoolinfo->raid_disks = raid_disks * 2;
3261
3262        ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3263                           rbio_pool_free, newpoolinfo);
3264        if (ret) {
3265                kfree(newpoolinfo);
3266                return ret;
3267        }
3268        newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3269                                         raid_disks, 2),
3270                             GFP_KERNEL);
3271        if (!newmirrors) {
3272                kfree(newpoolinfo);
3273                mempool_exit(&newpool);
3274                return -ENOMEM;
3275        }
3276
3277        freeze_array(conf, 0);
3278
3279        /* ok, everything is stopped */
3280        oldpool = conf->r1bio_pool;
3281        conf->r1bio_pool = newpool;
3282
3283        for (d = d2 = 0; d < conf->raid_disks; d++) {
3284                struct md_rdev *rdev = conf->mirrors[d].rdev;
3285                if (rdev && rdev->raid_disk != d2) {
3286                        sysfs_unlink_rdev(mddev, rdev);
3287                        rdev->raid_disk = d2;
3288                        sysfs_unlink_rdev(mddev, rdev);
3289                        if (sysfs_link_rdev(mddev, rdev))
3290                                pr_warn("md/raid1:%s: cannot register rd%d\n",
3291                                        mdname(mddev), rdev->raid_disk);
3292                }
3293                if (rdev)
3294                        newmirrors[d2++].rdev = rdev;
3295        }
3296        kfree(conf->mirrors);
3297        conf->mirrors = newmirrors;
3298        kfree(conf->poolinfo);
3299        conf->poolinfo = newpoolinfo;
3300
3301        spin_lock_irqsave(&conf->device_lock, flags);
3302        mddev->degraded += (raid_disks - conf->raid_disks);
3303        spin_unlock_irqrestore(&conf->device_lock, flags);
3304        conf->raid_disks = mddev->raid_disks = raid_disks;
3305        mddev->delta_disks = 0;
3306
3307        unfreeze_array(conf);
3308
3309        set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3310        set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3311        md_wakeup_thread(mddev->thread);
3312
3313        mempool_exit(&oldpool);
3314        return 0;
3315}
3316
3317static void raid1_quiesce(struct mddev *mddev, int quiesce)
3318{
3319        struct r1conf *conf = mddev->private;
3320
3321        if (quiesce)
3322                freeze_array(conf, 0);
3323        else
3324                unfreeze_array(conf);
3325}
3326
3327static void *raid1_takeover(struct mddev *mddev)
3328{
3329        /* raid1 can take over:
3330         *  raid5 with 2 devices, any layout or chunk size
3331         */
3332        if (mddev->level == 5 && mddev->raid_disks == 2) {
3333                struct r1conf *conf;
3334                mddev->new_level = 1;
3335                mddev->new_layout = 0;
3336                mddev->new_chunk_sectors = 0;
3337                conf = setup_conf(mddev);
3338                if (!IS_ERR(conf)) {
3339                        /* Array must appear to be quiesced */
3340                        conf->array_frozen = 1;
3341                        mddev_clear_unsupported_flags(mddev,
3342                                UNSUPPORTED_MDDEV_FLAGS);
3343                }
3344                return conf;
3345        }
3346        return ERR_PTR(-EINVAL);
3347}
3348
3349static struct md_personality raid1_personality =
3350{
3351        .name           = "raid1",
3352        .level          = 1,
3353        .owner          = THIS_MODULE,
3354        .make_request   = raid1_make_request,
3355        .run            = raid1_run,
3356        .free           = raid1_free,
3357        .status         = raid1_status,
3358        .error_handler  = raid1_error,
3359        .hot_add_disk   = raid1_add_disk,
3360        .hot_remove_disk= raid1_remove_disk,
3361        .spare_active   = raid1_spare_active,
3362        .sync_request   = raid1_sync_request,
3363        .resize         = raid1_resize,
3364        .size           = raid1_size,
3365        .check_reshape  = raid1_reshape,
3366        .quiesce        = raid1_quiesce,
3367        .takeover       = raid1_takeover,
3368};
3369
3370static int __init raid_init(void)
3371{
3372        return register_md_personality(&raid1_personality);
3373}
3374
3375static void raid_exit(void)
3376{
3377        unregister_md_personality(&raid1_personality);
3378}
3379
3380module_init(raid_init);
3381module_exit(raid_exit);
3382MODULE_LICENSE("GPL");
3383MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3384MODULE_ALIAS("md-personality-3"); /* RAID1 */
3385MODULE_ALIAS("md-raid1");
3386MODULE_ALIAS("md-level-1");
3387
3388module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
3389