linux/drivers/md/raid1.c
<<
>>
Prefs
   1/*
   2 * raid1.c : Multiple Devices driver for Linux
   3 *
   4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5 *
   6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7 *
   8 * RAID-1 management functions.
   9 *
  10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11 *
  12 * Fixes to reconstruction by Jakob Ƙstergaard" <jakob@ostenfeld.dk>
  13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14 *
  15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  16 * bitmapped intelligence in resync:
  17 *
  18 *      - bitmap marked during normal i/o
  19 *      - bitmap used to skip nondirty blocks during sync
  20 *
  21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  22 * - persistent bitmap code
  23 *
  24 * This program is free software; you can redistribute it and/or modify
  25 * it under the terms of the GNU General Public License as published by
  26 * the Free Software Foundation; either version 2, or (at your option)
  27 * any later version.
  28 *
  29 * You should have received a copy of the GNU General Public License
  30 * (for example /usr/src/linux/COPYING); if not, write to the Free
  31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  32 */
  33
  34#include <linux/slab.h>
  35#include <linux/delay.h>
  36#include <linux/blkdev.h>
  37#include <linux/module.h>
  38#include <linux/seq_file.h>
  39#include <linux/ratelimit.h>
  40#include "md.h"
  41#include "raid1.h"
  42#include "bitmap.h"
  43
  44/*
  45 * Number of guaranteed r1bios in case of extreme VM load:
  46 */
  47#define NR_RAID1_BIOS 256
  48
  49/* when we get a read error on a read-only array, we redirect to another
  50 * device without failing the first device, or trying to over-write to
  51 * correct the read error.  To keep track of bad blocks on a per-bio
  52 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
  53 */
  54#define IO_BLOCKED ((struct bio *)1)
  55/* When we successfully write to a known bad-block, we need to remove the
  56 * bad-block marking which must be done from process context.  So we record
  57 * the success by setting devs[n].bio to IO_MADE_GOOD
  58 */
  59#define IO_MADE_GOOD ((struct bio *)2)
  60
  61#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
  62
  63/* When there are this many requests queue to be written by
  64 * the raid1 thread, we become 'congested' to provide back-pressure
  65 * for writeback.
  66 */
  67static int max_queued_requests = 1024;
  68
  69static void allow_barrier(struct r1conf *conf);
  70static void lower_barrier(struct r1conf *conf);
  71
  72static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  73{
  74        struct pool_info *pi = data;
  75        int size = offsetof(struct r1bio, bios[pi->raid_disks]);
  76
  77        /* allocate a r1bio with room for raid_disks entries in the bios array */
  78        return kzalloc(size, gfp_flags);
  79}
  80
  81static void r1bio_pool_free(void *r1_bio, void *data)
  82{
  83        kfree(r1_bio);
  84}
  85
  86#define RESYNC_BLOCK_SIZE (64*1024)
  87//#define RESYNC_BLOCK_SIZE PAGE_SIZE
  88#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  89#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  90#define RESYNC_WINDOW (2048*1024)
  91
  92static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  93{
  94        struct pool_info *pi = data;
  95        struct r1bio *r1_bio;
  96        struct bio *bio;
  97        int i, j;
  98
  99        r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 100        if (!r1_bio)
 101                return NULL;
 102
 103        /*
 104         * Allocate bios : 1 for reading, n-1 for writing
 105         */
 106        for (j = pi->raid_disks ; j-- ; ) {
 107                bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 108                if (!bio)
 109                        goto out_free_bio;
 110                r1_bio->bios[j] = bio;
 111        }
 112        /*
 113         * Allocate RESYNC_PAGES data pages and attach them to
 114         * the first bio.
 115         * If this is a user-requested check/repair, allocate
 116         * RESYNC_PAGES for each bio.
 117         */
 118        if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 119                j = pi->raid_disks;
 120        else
 121                j = 1;
 122        while(j--) {
 123                bio = r1_bio->bios[j];
 124                bio->bi_vcnt = RESYNC_PAGES;
 125
 126                if (bio_alloc_pages(bio, gfp_flags))
 127                        goto out_free_bio;
 128        }
 129        /* If not user-requests, copy the page pointers to all bios */
 130        if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 131                for (i=0; i<RESYNC_PAGES ; i++)
 132                        for (j=1; j<pi->raid_disks; j++)
 133                                r1_bio->bios[j]->bi_io_vec[i].bv_page =
 134                                        r1_bio->bios[0]->bi_io_vec[i].bv_page;
 135        }
 136
 137        r1_bio->master_bio = NULL;
 138
 139        return r1_bio;
 140
 141out_free_bio:
 142        while (++j < pi->raid_disks)
 143                bio_put(r1_bio->bios[j]);
 144        r1bio_pool_free(r1_bio, data);
 145        return NULL;
 146}
 147
 148static void r1buf_pool_free(void *__r1_bio, void *data)
 149{
 150        struct pool_info *pi = data;
 151        int i,j;
 152        struct r1bio *r1bio = __r1_bio;
 153
 154        for (i = 0; i < RESYNC_PAGES; i++)
 155                for (j = pi->raid_disks; j-- ;) {
 156                        if (j == 0 ||
 157                            r1bio->bios[j]->bi_io_vec[i].bv_page !=
 158                            r1bio->bios[0]->bi_io_vec[i].bv_page)
 159                                safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 160                }
 161        for (i=0 ; i < pi->raid_disks; i++)
 162                bio_put(r1bio->bios[i]);
 163
 164        r1bio_pool_free(r1bio, data);
 165}
 166
 167static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 168{
 169        int i;
 170
 171        for (i = 0; i < conf->raid_disks * 2; i++) {
 172                struct bio **bio = r1_bio->bios + i;
 173                if (!BIO_SPECIAL(*bio))
 174                        bio_put(*bio);
 175                *bio = NULL;
 176        }
 177}
 178
 179static void free_r1bio(struct r1bio *r1_bio)
 180{
 181        struct r1conf *conf = r1_bio->mddev->private;
 182
 183        put_all_bios(conf, r1_bio);
 184        mempool_free(r1_bio, conf->r1bio_pool);
 185}
 186
 187static void put_buf(struct r1bio *r1_bio)
 188{
 189        struct r1conf *conf = r1_bio->mddev->private;
 190        int i;
 191
 192        for (i = 0; i < conf->raid_disks * 2; i++) {
 193                struct bio *bio = r1_bio->bios[i];
 194                if (bio->bi_end_io)
 195                        rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 196        }
 197
 198        mempool_free(r1_bio, conf->r1buf_pool);
 199
 200        lower_barrier(conf);
 201}
 202
 203static void reschedule_retry(struct r1bio *r1_bio)
 204{
 205        unsigned long flags;
 206        struct mddev *mddev = r1_bio->mddev;
 207        struct r1conf *conf = mddev->private;
 208
 209        spin_lock_irqsave(&conf->device_lock, flags);
 210        list_add(&r1_bio->retry_list, &conf->retry_list);
 211        conf->nr_queued ++;
 212        spin_unlock_irqrestore(&conf->device_lock, flags);
 213
 214        wake_up(&conf->wait_barrier);
 215        md_wakeup_thread(mddev->thread);
 216}
 217
 218/*
 219 * raid_end_bio_io() is called when we have finished servicing a mirrored
 220 * operation and are ready to return a success/failure code to the buffer
 221 * cache layer.
 222 */
 223static void call_bio_endio(struct r1bio *r1_bio)
 224{
 225        struct bio *bio = r1_bio->master_bio;
 226        int done;
 227        struct r1conf *conf = r1_bio->mddev->private;
 228
 229        if (bio->bi_phys_segments) {
 230                unsigned long flags;
 231                spin_lock_irqsave(&conf->device_lock, flags);
 232                bio->bi_phys_segments--;
 233                done = (bio->bi_phys_segments == 0);
 234                spin_unlock_irqrestore(&conf->device_lock, flags);
 235        } else
 236                done = 1;
 237
 238        if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 239                clear_bit(BIO_UPTODATE, &bio->bi_flags);
 240        if (done) {
 241                bio_endio(bio, 0);
 242                /*
 243                 * Wake up any possible resync thread that waits for the device
 244                 * to go idle.
 245                 */
 246                allow_barrier(conf);
 247        }
 248}
 249
 250static void raid_end_bio_io(struct r1bio *r1_bio)
 251{
 252        struct bio *bio = r1_bio->master_bio;
 253
 254        /* if nobody has done the final endio yet, do it now */
 255        if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 256                pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 257                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 258                         (unsigned long long) bio->bi_sector,
 259                         (unsigned long long) bio->bi_sector +
 260                         bio_sectors(bio) - 1);
 261
 262                call_bio_endio(r1_bio);
 263        }
 264        free_r1bio(r1_bio);
 265}
 266
 267/*
 268 * Update disk head position estimator based on IRQ completion info.
 269 */
 270static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 271{
 272        struct r1conf *conf = r1_bio->mddev->private;
 273
 274        conf->mirrors[disk].head_position =
 275                r1_bio->sector + (r1_bio->sectors);
 276}
 277
 278/*
 279 * Find the disk number which triggered given bio
 280 */
 281static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 282{
 283        int mirror;
 284        struct r1conf *conf = r1_bio->mddev->private;
 285        int raid_disks = conf->raid_disks;
 286
 287        for (mirror = 0; mirror < raid_disks * 2; mirror++)
 288                if (r1_bio->bios[mirror] == bio)
 289                        break;
 290
 291        BUG_ON(mirror == raid_disks * 2);
 292        update_head_pos(mirror, r1_bio);
 293
 294        return mirror;
 295}
 296
 297static void raid1_end_read_request(struct bio *bio, int error)
 298{
 299        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 300        struct r1bio *r1_bio = bio->bi_private;
 301        int mirror;
 302        struct r1conf *conf = r1_bio->mddev->private;
 303
 304        mirror = r1_bio->read_disk;
 305        /*
 306         * this branch is our 'one mirror IO has finished' event handler:
 307         */
 308        update_head_pos(mirror, r1_bio);
 309
 310        if (uptodate)
 311                set_bit(R1BIO_Uptodate, &r1_bio->state);
 312        else {
 313                /* If all other devices have failed, we want to return
 314                 * the error upwards rather than fail the last device.
 315                 * Here we redefine "uptodate" to mean "Don't want to retry"
 316                 */
 317                unsigned long flags;
 318                spin_lock_irqsave(&conf->device_lock, flags);
 319                if (r1_bio->mddev->degraded == conf->raid_disks ||
 320                    (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 321                     !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 322                        uptodate = 1;
 323                spin_unlock_irqrestore(&conf->device_lock, flags);
 324        }
 325
 326        if (uptodate) {
 327                raid_end_bio_io(r1_bio);
 328                rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 329        } else {
 330                /*
 331                 * oops, read error:
 332                 */
 333                char b[BDEVNAME_SIZE];
 334                printk_ratelimited(
 335                        KERN_ERR "md/raid1:%s: %s: "
 336                        "rescheduling sector %llu\n",
 337                        mdname(conf->mddev),
 338                        bdevname(conf->mirrors[mirror].rdev->bdev,
 339                                 b),
 340                        (unsigned long long)r1_bio->sector);
 341                set_bit(R1BIO_ReadError, &r1_bio->state);
 342                reschedule_retry(r1_bio);
 343                /* don't drop the reference on read_disk yet */
 344        }
 345}
 346
 347static void close_write(struct r1bio *r1_bio)
 348{
 349        /* it really is the end of this request */
 350        if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 351                /* free extra copy of the data pages */
 352                int i = r1_bio->behind_page_count;
 353                while (i--)
 354                        safe_put_page(r1_bio->behind_bvecs[i].bv_page);
 355                kfree(r1_bio->behind_bvecs);
 356                r1_bio->behind_bvecs = NULL;
 357        }
 358        /* clear the bitmap if all writes complete successfully */
 359        bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 360                        r1_bio->sectors,
 361                        !test_bit(R1BIO_Degraded, &r1_bio->state),
 362                        test_bit(R1BIO_BehindIO, &r1_bio->state));
 363        md_write_end(r1_bio->mddev);
 364}
 365
 366static void r1_bio_write_done(struct r1bio *r1_bio)
 367{
 368        if (!atomic_dec_and_test(&r1_bio->remaining))
 369                return;
 370
 371        if (test_bit(R1BIO_WriteError, &r1_bio->state))
 372                reschedule_retry(r1_bio);
 373        else {
 374                close_write(r1_bio);
 375                if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 376                        reschedule_retry(r1_bio);
 377                else
 378                        raid_end_bio_io(r1_bio);
 379        }
 380}
 381
 382static void raid1_end_write_request(struct bio *bio, int error)
 383{
 384        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 385        struct r1bio *r1_bio = bio->bi_private;
 386        int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 387        struct r1conf *conf = r1_bio->mddev->private;
 388        struct bio *to_put = NULL;
 389
 390        mirror = find_bio_disk(r1_bio, bio);
 391
 392        /*
 393         * 'one mirror IO has finished' event handler:
 394         */
 395        if (!uptodate) {
 396                set_bit(WriteErrorSeen,
 397                        &conf->mirrors[mirror].rdev->flags);
 398                if (!test_and_set_bit(WantReplacement,
 399                                      &conf->mirrors[mirror].rdev->flags))
 400                        set_bit(MD_RECOVERY_NEEDED, &
 401                                conf->mddev->recovery);
 402
 403                set_bit(R1BIO_WriteError, &r1_bio->state);
 404        } else {
 405                /*
 406                 * Set R1BIO_Uptodate in our master bio, so that we
 407                 * will return a good error code for to the higher
 408                 * levels even if IO on some other mirrored buffer
 409                 * fails.
 410                 *
 411                 * The 'master' represents the composite IO operation
 412                 * to user-side. So if something waits for IO, then it
 413                 * will wait for the 'master' bio.
 414                 */
 415                sector_t first_bad;
 416                int bad_sectors;
 417
 418                r1_bio->bios[mirror] = NULL;
 419                to_put = bio;
 420                /*
 421                 * Do not set R1BIO_Uptodate if the current device is
 422                 * rebuilding or Faulty. This is because we cannot use
 423                 * such device for properly reading the data back (we could
 424                 * potentially use it, if the current write would have felt
 425                 * before rdev->recovery_offset, but for simplicity we don't
 426                 * check this here.
 427                 */
 428                if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) &&
 429                    !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))
 430                        set_bit(R1BIO_Uptodate, &r1_bio->state);
 431
 432                /* Maybe we can clear some bad blocks. */
 433                if (is_badblock(conf->mirrors[mirror].rdev,
 434                                r1_bio->sector, r1_bio->sectors,
 435                                &first_bad, &bad_sectors)) {
 436                        r1_bio->bios[mirror] = IO_MADE_GOOD;
 437                        set_bit(R1BIO_MadeGood, &r1_bio->state);
 438                }
 439        }
 440
 441        if (behind) {
 442                if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 443                        atomic_dec(&r1_bio->behind_remaining);
 444
 445                /*
 446                 * In behind mode, we ACK the master bio once the I/O
 447                 * has safely reached all non-writemostly
 448                 * disks. Setting the Returned bit ensures that this
 449                 * gets done only once -- we don't ever want to return
 450                 * -EIO here, instead we'll wait
 451                 */
 452                if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 453                    test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 454                        /* Maybe we can return now */
 455                        if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 456                                struct bio *mbio = r1_bio->master_bio;
 457                                pr_debug("raid1: behind end write sectors"
 458                                         " %llu-%llu\n",
 459                                         (unsigned long long) mbio->bi_sector,
 460                                         (unsigned long long) mbio->bi_sector +
 461                                         bio_sectors(mbio) - 1);
 462                                call_bio_endio(r1_bio);
 463                        }
 464                }
 465        }
 466        if (r1_bio->bios[mirror] == NULL)
 467                rdev_dec_pending(conf->mirrors[mirror].rdev,
 468                                 conf->mddev);
 469
 470        /*
 471         * Let's see if all mirrored write operations have finished
 472         * already.
 473         */
 474        r1_bio_write_done(r1_bio);
 475
 476        if (to_put)
 477                bio_put(to_put);
 478}
 479
 480
 481/*
 482 * This routine returns the disk from which the requested read should
 483 * be done. There is a per-array 'next expected sequential IO' sector
 484 * number - if this matches on the next IO then we use the last disk.
 485 * There is also a per-disk 'last know head position' sector that is
 486 * maintained from IRQ contexts, both the normal and the resync IO
 487 * completion handlers update this position correctly. If there is no
 488 * perfect sequential match then we pick the disk whose head is closest.
 489 *
 490 * If there are 2 mirrors in the same 2 devices, performance degrades
 491 * because position is mirror, not device based.
 492 *
 493 * The rdev for the device selected will have nr_pending incremented.
 494 */
 495static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 496{
 497        const sector_t this_sector = r1_bio->sector;
 498        int sectors;
 499        int best_good_sectors;
 500        int best_disk, best_dist_disk, best_pending_disk;
 501        int has_nonrot_disk;
 502        int disk;
 503        sector_t best_dist;
 504        unsigned int min_pending;
 505        struct md_rdev *rdev;
 506        int choose_first;
 507        int choose_next_idle;
 508
 509        rcu_read_lock();
 510        /*
 511         * Check if we can balance. We can balance on the whole
 512         * device if no resync is going on, or below the resync window.
 513         * We take the first readable disk when above the resync window.
 514         */
 515 retry:
 516        sectors = r1_bio->sectors;
 517        best_disk = -1;
 518        best_dist_disk = -1;
 519        best_dist = MaxSector;
 520        best_pending_disk = -1;
 521        min_pending = UINT_MAX;
 522        best_good_sectors = 0;
 523        has_nonrot_disk = 0;
 524        choose_next_idle = 0;
 525
 526        if (conf->mddev->recovery_cp < MaxSector &&
 527            (this_sector + sectors >= conf->next_resync))
 528                choose_first = 1;
 529        else
 530                choose_first = 0;
 531
 532        for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 533                sector_t dist;
 534                sector_t first_bad;
 535                int bad_sectors;
 536                unsigned int pending;
 537                bool nonrot;
 538
 539                rdev = rcu_dereference(conf->mirrors[disk].rdev);
 540                if (r1_bio->bios[disk] == IO_BLOCKED
 541                    || rdev == NULL
 542                    || test_bit(Unmerged, &rdev->flags)
 543                    || test_bit(Faulty, &rdev->flags))
 544                        continue;
 545                if (!test_bit(In_sync, &rdev->flags) &&
 546                    rdev->recovery_offset < this_sector + sectors)
 547                        continue;
 548                if (test_bit(WriteMostly, &rdev->flags)) {
 549                        /* Don't balance among write-mostly, just
 550                         * use the first as a last resort */
 551                        if (best_disk < 0) {
 552                                if (is_badblock(rdev, this_sector, sectors,
 553                                                &first_bad, &bad_sectors)) {
 554                                        if (first_bad < this_sector)
 555                                                /* Cannot use this */
 556                                                continue;
 557                                        best_good_sectors = first_bad - this_sector;
 558                                } else
 559                                        best_good_sectors = sectors;
 560                                best_disk = disk;
 561                        }
 562                        continue;
 563                }
 564                /* This is a reasonable device to use.  It might
 565                 * even be best.
 566                 */
 567                if (is_badblock(rdev, this_sector, sectors,
 568                                &first_bad, &bad_sectors)) {
 569                        if (best_dist < MaxSector)
 570                                /* already have a better device */
 571                                continue;
 572                        if (first_bad <= this_sector) {
 573                                /* cannot read here. If this is the 'primary'
 574                                 * device, then we must not read beyond
 575                                 * bad_sectors from another device..
 576                                 */
 577                                bad_sectors -= (this_sector - first_bad);
 578                                if (choose_first && sectors > bad_sectors)
 579                                        sectors = bad_sectors;
 580                                if (best_good_sectors > sectors)
 581                                        best_good_sectors = sectors;
 582
 583                        } else {
 584                                sector_t good_sectors = first_bad - this_sector;
 585                                if (good_sectors > best_good_sectors) {
 586                                        best_good_sectors = good_sectors;
 587                                        best_disk = disk;
 588                                }
 589                                if (choose_first)
 590                                        break;
 591                        }
 592                        continue;
 593                } else
 594                        best_good_sectors = sectors;
 595
 596                nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 597                has_nonrot_disk |= nonrot;
 598                pending = atomic_read(&rdev->nr_pending);
 599                dist = abs(this_sector - conf->mirrors[disk].head_position);
 600                if (choose_first) {
 601                        best_disk = disk;
 602                        break;
 603                }
 604                /* Don't change to another disk for sequential reads */
 605                if (conf->mirrors[disk].next_seq_sect == this_sector
 606                    || dist == 0) {
 607                        int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 608                        struct raid1_info *mirror = &conf->mirrors[disk];
 609
 610                        best_disk = disk;
 611                        /*
 612                         * If buffered sequential IO size exceeds optimal
 613                         * iosize, check if there is idle disk. If yes, choose
 614                         * the idle disk. read_balance could already choose an
 615                         * idle disk before noticing it's a sequential IO in
 616                         * this disk. This doesn't matter because this disk
 617                         * will idle, next time it will be utilized after the
 618                         * first disk has IO size exceeds optimal iosize. In
 619                         * this way, iosize of the first disk will be optimal
 620                         * iosize at least. iosize of the second disk might be
 621                         * small, but not a big deal since when the second disk
 622                         * starts IO, the first disk is likely still busy.
 623                         */
 624                        if (nonrot && opt_iosize > 0 &&
 625                            mirror->seq_start != MaxSector &&
 626                            mirror->next_seq_sect > opt_iosize &&
 627                            mirror->next_seq_sect - opt_iosize >=
 628                            mirror->seq_start) {
 629                                choose_next_idle = 1;
 630                                continue;
 631                        }
 632                        break;
 633                }
 634                /* If device is idle, use it */
 635                if (pending == 0) {
 636                        best_disk = disk;
 637                        break;
 638                }
 639
 640                if (choose_next_idle)
 641                        continue;
 642
 643                if (min_pending > pending) {
 644                        min_pending = pending;
 645                        best_pending_disk = disk;
 646                }
 647
 648                if (dist < best_dist) {
 649                        best_dist = dist;
 650                        best_dist_disk = disk;
 651                }
 652        }
 653
 654        /*
 655         * If all disks are rotational, choose the closest disk. If any disk is
 656         * non-rotational, choose the disk with less pending request even the
 657         * disk is rotational, which might/might not be optimal for raids with
 658         * mixed ratation/non-rotational disks depending on workload.
 659         */
 660        if (best_disk == -1) {
 661                if (has_nonrot_disk)
 662                        best_disk = best_pending_disk;
 663                else
 664                        best_disk = best_dist_disk;
 665        }
 666
 667        if (best_disk >= 0) {
 668                rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 669                if (!rdev)
 670                        goto retry;
 671                atomic_inc(&rdev->nr_pending);
 672                if (test_bit(Faulty, &rdev->flags)) {
 673                        /* cannot risk returning a device that failed
 674                         * before we inc'ed nr_pending
 675                         */
 676                        rdev_dec_pending(rdev, conf->mddev);
 677                        goto retry;
 678                }
 679                sectors = best_good_sectors;
 680
 681                if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 682                        conf->mirrors[best_disk].seq_start = this_sector;
 683
 684                conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 685        }
 686        rcu_read_unlock();
 687        *max_sectors = sectors;
 688
 689        return best_disk;
 690}
 691
 692static int raid1_mergeable_bvec(struct request_queue *q,
 693                                struct bvec_merge_data *bvm,
 694                                struct bio_vec *biovec)
 695{
 696        struct mddev *mddev = q->queuedata;
 697        struct r1conf *conf = mddev->private;
 698        sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 699        int max = biovec->bv_len;
 700
 701        if (mddev->merge_check_needed) {
 702                int disk;
 703                rcu_read_lock();
 704                for (disk = 0; disk < conf->raid_disks * 2; disk++) {
 705                        struct md_rdev *rdev = rcu_dereference(
 706                                conf->mirrors[disk].rdev);
 707                        if (rdev && !test_bit(Faulty, &rdev->flags)) {
 708                                struct request_queue *q =
 709                                        bdev_get_queue(rdev->bdev);
 710                                if (q->merge_bvec_fn) {
 711                                        bvm->bi_sector = sector +
 712                                                rdev->data_offset;
 713                                        bvm->bi_bdev = rdev->bdev;
 714                                        max = min(max, q->merge_bvec_fn(
 715                                                          q, bvm, biovec));
 716                                }
 717                        }
 718                }
 719                rcu_read_unlock();
 720        }
 721        return max;
 722
 723}
 724
 725int md_raid1_congested(struct mddev *mddev, int bits)
 726{
 727        struct r1conf *conf = mddev->private;
 728        int i, ret = 0;
 729
 730        if ((bits & (1 << BDI_async_congested)) &&
 731            conf->pending_count >= max_queued_requests)
 732                return 1;
 733
 734        rcu_read_lock();
 735        for (i = 0; i < conf->raid_disks * 2; i++) {
 736                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 737                if (rdev && !test_bit(Faulty, &rdev->flags)) {
 738                        struct request_queue *q = bdev_get_queue(rdev->bdev);
 739
 740                        BUG_ON(!q);
 741
 742                        /* Note the '|| 1' - when read_balance prefers
 743                         * non-congested targets, it can be removed
 744                         */
 745                        if ((bits & (1<<BDI_async_congested)) || 1)
 746                                ret |= bdi_congested(&q->backing_dev_info, bits);
 747                        else
 748                                ret &= bdi_congested(&q->backing_dev_info, bits);
 749                }
 750        }
 751        rcu_read_unlock();
 752        return ret;
 753}
 754EXPORT_SYMBOL_GPL(md_raid1_congested);
 755
 756static int raid1_congested(void *data, int bits)
 757{
 758        struct mddev *mddev = data;
 759
 760        return mddev_congested(mddev, bits) ||
 761                md_raid1_congested(mddev, bits);
 762}
 763
 764static void flush_pending_writes(struct r1conf *conf)
 765{
 766        /* Any writes that have been queued but are awaiting
 767         * bitmap updates get flushed here.
 768         */
 769        spin_lock_irq(&conf->device_lock);
 770
 771        if (conf->pending_bio_list.head) {
 772                struct bio *bio;
 773                bio = bio_list_get(&conf->pending_bio_list);
 774                conf->pending_count = 0;
 775                spin_unlock_irq(&conf->device_lock);
 776                /* flush any pending bitmap writes to
 777                 * disk before proceeding w/ I/O */
 778                bitmap_unplug(conf->mddev->bitmap);
 779                wake_up(&conf->wait_barrier);
 780
 781                while (bio) { /* submit pending writes */
 782                        struct bio *next = bio->bi_next;
 783                        bio->bi_next = NULL;
 784                        if (unlikely((bio->bi_rw & REQ_DISCARD) &&
 785                            !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 786                                /* Just ignore it */
 787                                bio_endio(bio, 0);
 788                        else
 789                                generic_make_request(bio);
 790                        bio = next;
 791                }
 792        } else
 793                spin_unlock_irq(&conf->device_lock);
 794}
 795
 796/* Barriers....
 797 * Sometimes we need to suspend IO while we do something else,
 798 * either some resync/recovery, or reconfigure the array.
 799 * To do this we raise a 'barrier'.
 800 * The 'barrier' is a counter that can be raised multiple times
 801 * to count how many activities are happening which preclude
 802 * normal IO.
 803 * We can only raise the barrier if there is no pending IO.
 804 * i.e. if nr_pending == 0.
 805 * We choose only to raise the barrier if no-one is waiting for the
 806 * barrier to go down.  This means that as soon as an IO request
 807 * is ready, no other operations which require a barrier will start
 808 * until the IO request has had a chance.
 809 *
 810 * So: regular IO calls 'wait_barrier'.  When that returns there
 811 *    is no backgroup IO happening,  It must arrange to call
 812 *    allow_barrier when it has finished its IO.
 813 * backgroup IO calls must call raise_barrier.  Once that returns
 814 *    there is no normal IO happeing.  It must arrange to call
 815 *    lower_barrier when the particular background IO completes.
 816 */
 817#define RESYNC_DEPTH 32
 818
 819static void raise_barrier(struct r1conf *conf)
 820{
 821        spin_lock_irq(&conf->resync_lock);
 822
 823        /* Wait until no block IO is waiting */
 824        wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 825                            conf->resync_lock);
 826
 827        /* block any new IO from starting */
 828        conf->barrier++;
 829
 830        /* Now wait for all pending IO to complete */
 831        wait_event_lock_irq(conf->wait_barrier,
 832                            !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 833                            conf->resync_lock);
 834
 835        spin_unlock_irq(&conf->resync_lock);
 836}
 837
 838static void lower_barrier(struct r1conf *conf)
 839{
 840        unsigned long flags;
 841        BUG_ON(conf->barrier <= 0);
 842        spin_lock_irqsave(&conf->resync_lock, flags);
 843        conf->barrier--;
 844        spin_unlock_irqrestore(&conf->resync_lock, flags);
 845        wake_up(&conf->wait_barrier);
 846}
 847
 848static void wait_barrier(struct r1conf *conf)
 849{
 850        spin_lock_irq(&conf->resync_lock);
 851        if (conf->barrier) {
 852                conf->nr_waiting++;
 853                /* Wait for the barrier to drop.
 854                 * However if there are already pending
 855                 * requests (preventing the barrier from
 856                 * rising completely), and the
 857                 * pre-process bio queue isn't empty,
 858                 * then don't wait, as we need to empty
 859                 * that queue to get the nr_pending
 860                 * count down.
 861                 */
 862                wait_event_lock_irq(conf->wait_barrier,
 863                                    !conf->barrier ||
 864                                    (conf->nr_pending &&
 865                                     current->bio_list &&
 866                                     !bio_list_empty(current->bio_list)),
 867                                    conf->resync_lock);
 868                conf->nr_waiting--;
 869        }
 870        conf->nr_pending++;
 871        spin_unlock_irq(&conf->resync_lock);
 872}
 873
 874static void allow_barrier(struct r1conf *conf)
 875{
 876        unsigned long flags;
 877        spin_lock_irqsave(&conf->resync_lock, flags);
 878        conf->nr_pending--;
 879        spin_unlock_irqrestore(&conf->resync_lock, flags);
 880        wake_up(&conf->wait_barrier);
 881}
 882
 883static void freeze_array(struct r1conf *conf, int extra)
 884{
 885        /* stop syncio and normal IO and wait for everything to
 886         * go quite.
 887         * We increment barrier and nr_waiting, and then
 888         * wait until nr_pending match nr_queued+extra
 889         * This is called in the context of one normal IO request
 890         * that has failed. Thus any sync request that might be pending
 891         * will be blocked by nr_pending, and we need to wait for
 892         * pending IO requests to complete or be queued for re-try.
 893         * Thus the number queued (nr_queued) plus this request (extra)
 894         * must match the number of pending IOs (nr_pending) before
 895         * we continue.
 896         */
 897        spin_lock_irq(&conf->resync_lock);
 898        conf->barrier++;
 899        conf->nr_waiting++;
 900        wait_event_lock_irq_cmd(conf->wait_barrier,
 901                                conf->nr_pending == conf->nr_queued+extra,
 902                                conf->resync_lock,
 903                                flush_pending_writes(conf));
 904        spin_unlock_irq(&conf->resync_lock);
 905}
 906static void unfreeze_array(struct r1conf *conf)
 907{
 908        /* reverse the effect of the freeze */
 909        spin_lock_irq(&conf->resync_lock);
 910        conf->barrier--;
 911        conf->nr_waiting--;
 912        wake_up(&conf->wait_barrier);
 913        spin_unlock_irq(&conf->resync_lock);
 914}
 915
 916
 917/* duplicate the data pages for behind I/O 
 918 */
 919static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
 920{
 921        int i;
 922        struct bio_vec *bvec;
 923        struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
 924                                        GFP_NOIO);
 925        if (unlikely(!bvecs))
 926                return;
 927
 928        bio_for_each_segment_all(bvec, bio, i) {
 929                bvecs[i] = *bvec;
 930                bvecs[i].bv_page = alloc_page(GFP_NOIO);
 931                if (unlikely(!bvecs[i].bv_page))
 932                        goto do_sync_io;
 933                memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
 934                       kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 935                kunmap(bvecs[i].bv_page);
 936                kunmap(bvec->bv_page);
 937        }
 938        r1_bio->behind_bvecs = bvecs;
 939        r1_bio->behind_page_count = bio->bi_vcnt;
 940        set_bit(R1BIO_BehindIO, &r1_bio->state);
 941        return;
 942
 943do_sync_io:
 944        for (i = 0; i < bio->bi_vcnt; i++)
 945                if (bvecs[i].bv_page)
 946                        put_page(bvecs[i].bv_page);
 947        kfree(bvecs);
 948        pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 949}
 950
 951struct raid1_plug_cb {
 952        struct blk_plug_cb      cb;
 953        struct bio_list         pending;
 954        int                     pending_cnt;
 955};
 956
 957static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
 958{
 959        struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
 960                                                  cb);
 961        struct mddev *mddev = plug->cb.data;
 962        struct r1conf *conf = mddev->private;
 963        struct bio *bio;
 964
 965        if (from_schedule || current->bio_list) {
 966                spin_lock_irq(&conf->device_lock);
 967                bio_list_merge(&conf->pending_bio_list, &plug->pending);
 968                conf->pending_count += plug->pending_cnt;
 969                spin_unlock_irq(&conf->device_lock);
 970                wake_up(&conf->wait_barrier);
 971                md_wakeup_thread(mddev->thread);
 972                kfree(plug);
 973                return;
 974        }
 975
 976        /* we aren't scheduling, so we can do the write-out directly. */
 977        bio = bio_list_get(&plug->pending);
 978        bitmap_unplug(mddev->bitmap);
 979        wake_up(&conf->wait_barrier);
 980
 981        while (bio) { /* submit pending writes */
 982                struct bio *next = bio->bi_next;
 983                bio->bi_next = NULL;
 984                if (unlikely((bio->bi_rw & REQ_DISCARD) &&
 985                    !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 986                        /* Just ignore it */
 987                        bio_endio(bio, 0);
 988                else
 989                        generic_make_request(bio);
 990                bio = next;
 991        }
 992        kfree(plug);
 993}
 994
 995static void make_request(struct mddev *mddev, struct bio * bio)
 996{
 997        struct r1conf *conf = mddev->private;
 998        struct raid1_info *mirror;
 999        struct r1bio *r1_bio;
1000        struct bio *read_bio;
1001        int i, disks;
1002        struct bitmap *bitmap;
1003        unsigned long flags;
1004        const int rw = bio_data_dir(bio);
1005        const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1006        const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1007        const unsigned long do_discard = (bio->bi_rw
1008                                          & (REQ_DISCARD | REQ_SECURE));
1009        const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1010        struct md_rdev *blocked_rdev;
1011        struct blk_plug_cb *cb;
1012        struct raid1_plug_cb *plug = NULL;
1013        int first_clone;
1014        int sectors_handled;
1015        int max_sectors;
1016
1017        /*
1018         * Register the new request and wait if the reconstruction
1019         * thread has put up a bar for new requests.
1020         * Continue immediately if no resync is active currently.
1021         */
1022
1023        md_write_start(mddev, bio); /* wait on superblock update early */
1024
1025        if (bio_data_dir(bio) == WRITE &&
1026            bio_end_sector(bio) > mddev->suspend_lo &&
1027            bio->bi_sector < mddev->suspend_hi) {
1028                /* As the suspend_* range is controlled by
1029                 * userspace, we want an interruptible
1030                 * wait.
1031                 */
1032                DEFINE_WAIT(w);
1033                for (;;) {
1034                        flush_signals(current);
1035                        prepare_to_wait(&conf->wait_barrier,
1036                                        &w, TASK_INTERRUPTIBLE);
1037                        if (bio_end_sector(bio) <= mddev->suspend_lo ||
1038                            bio->bi_sector >= mddev->suspend_hi)
1039                                break;
1040                        schedule();
1041                }
1042                finish_wait(&conf->wait_barrier, &w);
1043        }
1044
1045        wait_barrier(conf);
1046
1047        bitmap = mddev->bitmap;
1048
1049        /*
1050         * make_request() can abort the operation when READA is being
1051         * used and no empty request is available.
1052         *
1053         */
1054        r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1055
1056        r1_bio->master_bio = bio;
1057        r1_bio->sectors = bio_sectors(bio);
1058        r1_bio->state = 0;
1059        r1_bio->mddev = mddev;
1060        r1_bio->sector = bio->bi_sector;
1061
1062        /* We might need to issue multiple reads to different
1063         * devices if there are bad blocks around, so we keep
1064         * track of the number of reads in bio->bi_phys_segments.
1065         * If this is 0, there is only one r1_bio and no locking
1066         * will be needed when requests complete.  If it is
1067         * non-zero, then it is the number of not-completed requests.
1068         */
1069        bio->bi_phys_segments = 0;
1070        clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1071
1072        if (rw == READ) {
1073                /*
1074                 * read balancing logic:
1075                 */
1076                int rdisk;
1077
1078read_again:
1079                rdisk = read_balance(conf, r1_bio, &max_sectors);
1080
1081                if (rdisk < 0) {
1082                        /* couldn't find anywhere to read from */
1083                        raid_end_bio_io(r1_bio);
1084                        return;
1085                }
1086                mirror = conf->mirrors + rdisk;
1087
1088                if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1089                    bitmap) {
1090                        /* Reading from a write-mostly device must
1091                         * take care not to over-take any writes
1092                         * that are 'behind'
1093                         */
1094                        wait_event(bitmap->behind_wait,
1095                                   atomic_read(&bitmap->behind_writes) == 0);
1096                }
1097                r1_bio->read_disk = rdisk;
1098
1099                read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1100                md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
1101                            max_sectors);
1102
1103                r1_bio->bios[rdisk] = read_bio;
1104
1105                read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
1106                read_bio->bi_bdev = mirror->rdev->bdev;
1107                read_bio->bi_end_io = raid1_end_read_request;
1108                read_bio->bi_rw = READ | do_sync;
1109                read_bio->bi_private = r1_bio;
1110
1111                if (max_sectors < r1_bio->sectors) {
1112                        /* could not read all from this device, so we will
1113                         * need another r1_bio.
1114                         */
1115
1116                        sectors_handled = (r1_bio->sector + max_sectors
1117                                           - bio->bi_sector);
1118                        r1_bio->sectors = max_sectors;
1119                        spin_lock_irq(&conf->device_lock);
1120                        if (bio->bi_phys_segments == 0)
1121                                bio->bi_phys_segments = 2;
1122                        else
1123                                bio->bi_phys_segments++;
1124                        spin_unlock_irq(&conf->device_lock);
1125                        /* Cannot call generic_make_request directly
1126                         * as that will be queued in __make_request
1127                         * and subsequent mempool_alloc might block waiting
1128                         * for it.  So hand bio over to raid1d.
1129                         */
1130                        reschedule_retry(r1_bio);
1131
1132                        r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1133
1134                        r1_bio->master_bio = bio;
1135                        r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1136                        r1_bio->state = 0;
1137                        r1_bio->mddev = mddev;
1138                        r1_bio->sector = bio->bi_sector + sectors_handled;
1139                        goto read_again;
1140                } else
1141                        generic_make_request(read_bio);
1142                return;
1143        }
1144
1145        /*
1146         * WRITE:
1147         */
1148        if (conf->pending_count >= max_queued_requests) {
1149                md_wakeup_thread(mddev->thread);
1150                wait_event(conf->wait_barrier,
1151                           conf->pending_count < max_queued_requests);
1152        }
1153        /* first select target devices under rcu_lock and
1154         * inc refcount on their rdev.  Record them by setting
1155         * bios[x] to bio
1156         * If there are known/acknowledged bad blocks on any device on
1157         * which we have seen a write error, we want to avoid writing those
1158         * blocks.
1159         * This potentially requires several writes to write around
1160         * the bad blocks.  Each set of writes gets it's own r1bio
1161         * with a set of bios attached.
1162         */
1163
1164        disks = conf->raid_disks * 2;
1165 retry_write:
1166        blocked_rdev = NULL;
1167        rcu_read_lock();
1168        max_sectors = r1_bio->sectors;
1169        for (i = 0;  i < disks; i++) {
1170                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1171                if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1172                        atomic_inc(&rdev->nr_pending);
1173                        blocked_rdev = rdev;
1174                        break;
1175                }
1176                r1_bio->bios[i] = NULL;
1177                if (!rdev || test_bit(Faulty, &rdev->flags)
1178                    || test_bit(Unmerged, &rdev->flags)) {
1179                        if (i < conf->raid_disks)
1180                                set_bit(R1BIO_Degraded, &r1_bio->state);
1181                        continue;
1182                }
1183
1184                atomic_inc(&rdev->nr_pending);
1185                if (test_bit(WriteErrorSeen, &rdev->flags)) {
1186                        sector_t first_bad;
1187                        int bad_sectors;
1188                        int is_bad;
1189
1190                        is_bad = is_badblock(rdev, r1_bio->sector,
1191                                             max_sectors,
1192                                             &first_bad, &bad_sectors);
1193                        if (is_bad < 0) {
1194                                /* mustn't write here until the bad block is
1195                                 * acknowledged*/
1196                                set_bit(BlockedBadBlocks, &rdev->flags);
1197                                blocked_rdev = rdev;
1198                                break;
1199                        }
1200                        if (is_bad && first_bad <= r1_bio->sector) {
1201                                /* Cannot write here at all */
1202                                bad_sectors -= (r1_bio->sector - first_bad);
1203                                if (bad_sectors < max_sectors)
1204                                        /* mustn't write more than bad_sectors
1205                                         * to other devices yet
1206                                         */
1207                                        max_sectors = bad_sectors;
1208                                rdev_dec_pending(rdev, mddev);
1209                                /* We don't set R1BIO_Degraded as that
1210                                 * only applies if the disk is
1211                                 * missing, so it might be re-added,
1212                                 * and we want to know to recover this
1213                                 * chunk.
1214                                 * In this case the device is here,
1215                                 * and the fact that this chunk is not
1216                                 * in-sync is recorded in the bad
1217                                 * block log
1218                                 */
1219                                continue;
1220                        }
1221                        if (is_bad) {
1222                                int good_sectors = first_bad - r1_bio->sector;
1223                                if (good_sectors < max_sectors)
1224                                        max_sectors = good_sectors;
1225                        }
1226                }
1227                r1_bio->bios[i] = bio;
1228        }
1229        rcu_read_unlock();
1230
1231        if (unlikely(blocked_rdev)) {
1232                /* Wait for this device to become unblocked */
1233                int j;
1234
1235                for (j = 0; j < i; j++)
1236                        if (r1_bio->bios[j])
1237                                rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1238                r1_bio->state = 0;
1239                allow_barrier(conf);
1240                md_wait_for_blocked_rdev(blocked_rdev, mddev);
1241                wait_barrier(conf);
1242                goto retry_write;
1243        }
1244
1245        if (max_sectors < r1_bio->sectors) {
1246                /* We are splitting this write into multiple parts, so
1247                 * we need to prepare for allocating another r1_bio.
1248                 */
1249                r1_bio->sectors = max_sectors;
1250                spin_lock_irq(&conf->device_lock);
1251                if (bio->bi_phys_segments == 0)
1252                        bio->bi_phys_segments = 2;
1253                else
1254                        bio->bi_phys_segments++;
1255                spin_unlock_irq(&conf->device_lock);
1256        }
1257        sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1258
1259        atomic_set(&r1_bio->remaining, 1);
1260        atomic_set(&r1_bio->behind_remaining, 0);
1261
1262        first_clone = 1;
1263        for (i = 0; i < disks; i++) {
1264                struct bio *mbio;
1265                if (!r1_bio->bios[i])
1266                        continue;
1267
1268                mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1269                md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1270
1271                if (first_clone) {
1272                        /* do behind I/O ?
1273                         * Not if there are too many, or cannot
1274                         * allocate memory, or a reader on WriteMostly
1275                         * is waiting for behind writes to flush */
1276                        if (bitmap &&
1277                            (atomic_read(&bitmap->behind_writes)
1278                             < mddev->bitmap_info.max_write_behind) &&
1279                            !waitqueue_active(&bitmap->behind_wait))
1280                                alloc_behind_pages(mbio, r1_bio);
1281
1282                        bitmap_startwrite(bitmap, r1_bio->sector,
1283                                          r1_bio->sectors,
1284                                          test_bit(R1BIO_BehindIO,
1285                                                   &r1_bio->state));
1286                        first_clone = 0;
1287                }
1288                if (r1_bio->behind_bvecs) {
1289                        struct bio_vec *bvec;
1290                        int j;
1291
1292                        /*
1293                         * We trimmed the bio, so _all is legit
1294                         */
1295                        bio_for_each_segment_all(bvec, mbio, j)
1296                                bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1297                        if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1298                                atomic_inc(&r1_bio->behind_remaining);
1299                }
1300
1301                r1_bio->bios[i] = mbio;
1302
1303                mbio->bi_sector = (r1_bio->sector +
1304                                   conf->mirrors[i].rdev->data_offset);
1305                mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1306                mbio->bi_end_io = raid1_end_write_request;
1307                mbio->bi_rw =
1308                        WRITE | do_flush_fua | do_sync | do_discard | do_same;
1309                mbio->bi_private = r1_bio;
1310
1311                atomic_inc(&r1_bio->remaining);
1312
1313                cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1314                if (cb)
1315                        plug = container_of(cb, struct raid1_plug_cb, cb);
1316                else
1317                        plug = NULL;
1318                spin_lock_irqsave(&conf->device_lock, flags);
1319                if (plug) {
1320                        bio_list_add(&plug->pending, mbio);
1321                        plug->pending_cnt++;
1322                } else {
1323                        bio_list_add(&conf->pending_bio_list, mbio);
1324                        conf->pending_count++;
1325                }
1326                spin_unlock_irqrestore(&conf->device_lock, flags);
1327                if (!plug)
1328                        md_wakeup_thread(mddev->thread);
1329        }
1330        /* Mustn't call r1_bio_write_done before this next test,
1331         * as it could result in the bio being freed.
1332         */
1333        if (sectors_handled < bio_sectors(bio)) {
1334                r1_bio_write_done(r1_bio);
1335                /* We need another r1_bio.  It has already been counted
1336                 * in bio->bi_phys_segments
1337                 */
1338                r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1339                r1_bio->master_bio = bio;
1340                r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1341                r1_bio->state = 0;
1342                r1_bio->mddev = mddev;
1343                r1_bio->sector = bio->bi_sector + sectors_handled;
1344                goto retry_write;
1345        }
1346
1347        r1_bio_write_done(r1_bio);
1348
1349        /* In case raid1d snuck in to freeze_array */
1350        wake_up(&conf->wait_barrier);
1351}
1352
1353static void status(struct seq_file *seq, struct mddev *mddev)
1354{
1355        struct r1conf *conf = mddev->private;
1356        int i;
1357
1358        seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1359                   conf->raid_disks - mddev->degraded);
1360        rcu_read_lock();
1361        for (i = 0; i < conf->raid_disks; i++) {
1362                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1363                seq_printf(seq, "%s",
1364                           rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1365        }
1366        rcu_read_unlock();
1367        seq_printf(seq, "]");
1368}
1369
1370
1371static void error(struct mddev *mddev, struct md_rdev *rdev)
1372{
1373        char b[BDEVNAME_SIZE];
1374        struct r1conf *conf = mddev->private;
1375
1376        /*
1377         * If it is not operational, then we have already marked it as dead
1378         * else if it is the last working disks, ignore the error, let the
1379         * next level up know.
1380         * else mark the drive as failed
1381         */
1382        if (test_bit(In_sync, &rdev->flags)
1383            && (conf->raid_disks - mddev->degraded) == 1) {
1384                /*
1385                 * Don't fail the drive, act as though we were just a
1386                 * normal single drive.
1387                 * However don't try a recovery from this drive as
1388                 * it is very likely to fail.
1389                 */
1390                conf->recovery_disabled = mddev->recovery_disabled;
1391                return;
1392        }
1393        set_bit(Blocked, &rdev->flags);
1394        if (test_and_clear_bit(In_sync, &rdev->flags)) {
1395                unsigned long flags;
1396                spin_lock_irqsave(&conf->device_lock, flags);
1397                mddev->degraded++;
1398                set_bit(Faulty, &rdev->flags);
1399                spin_unlock_irqrestore(&conf->device_lock, flags);
1400                /*
1401                 * if recovery is running, make sure it aborts.
1402                 */
1403                set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1404        } else
1405                set_bit(Faulty, &rdev->flags);
1406        set_bit(MD_CHANGE_DEVS, &mddev->flags);
1407        printk(KERN_ALERT
1408               "md/raid1:%s: Disk failure on %s, disabling device.\n"
1409               "md/raid1:%s: Operation continuing on %d devices.\n",
1410               mdname(mddev), bdevname(rdev->bdev, b),
1411               mdname(mddev), conf->raid_disks - mddev->degraded);
1412}
1413
1414static void print_conf(struct r1conf *conf)
1415{
1416        int i;
1417
1418        printk(KERN_DEBUG "RAID1 conf printout:\n");
1419        if (!conf) {
1420                printk(KERN_DEBUG "(!conf)\n");
1421                return;
1422        }
1423        printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1424                conf->raid_disks);
1425
1426        rcu_read_lock();
1427        for (i = 0; i < conf->raid_disks; i++) {
1428                char b[BDEVNAME_SIZE];
1429                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1430                if (rdev)
1431                        printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1432                               i, !test_bit(In_sync, &rdev->flags),
1433                               !test_bit(Faulty, &rdev->flags),
1434                               bdevname(rdev->bdev,b));
1435        }
1436        rcu_read_unlock();
1437}
1438
1439static void close_sync(struct r1conf *conf)
1440{
1441        wait_barrier(conf);
1442        allow_barrier(conf);
1443
1444        mempool_destroy(conf->r1buf_pool);
1445        conf->r1buf_pool = NULL;
1446}
1447
1448static int raid1_spare_active(struct mddev *mddev)
1449{
1450        int i;
1451        struct r1conf *conf = mddev->private;
1452        int count = 0;
1453        unsigned long flags;
1454
1455        /*
1456         * Find all failed disks within the RAID1 configuration 
1457         * and mark them readable.
1458         * Called under mddev lock, so rcu protection not needed.
1459         */
1460        for (i = 0; i < conf->raid_disks; i++) {
1461                struct md_rdev *rdev = conf->mirrors[i].rdev;
1462                struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1463                if (repl
1464                    && repl->recovery_offset == MaxSector
1465                    && !test_bit(Faulty, &repl->flags)
1466                    && !test_and_set_bit(In_sync, &repl->flags)) {
1467                        /* replacement has just become active */
1468                        if (!rdev ||
1469                            !test_and_clear_bit(In_sync, &rdev->flags))
1470                                count++;
1471                        if (rdev) {
1472                                /* Replaced device not technically
1473                                 * faulty, but we need to be sure
1474                                 * it gets removed and never re-added
1475                                 */
1476                                set_bit(Faulty, &rdev->flags);
1477                                sysfs_notify_dirent_safe(
1478                                        rdev->sysfs_state);
1479                        }
1480                }
1481                if (rdev
1482                    && rdev->recovery_offset == MaxSector
1483                    && !test_bit(Faulty, &rdev->flags)
1484                    && !test_and_set_bit(In_sync, &rdev->flags)) {
1485                        count++;
1486                        sysfs_notify_dirent_safe(rdev->sysfs_state);
1487                }
1488        }
1489        spin_lock_irqsave(&conf->device_lock, flags);
1490        mddev->degraded -= count;
1491        spin_unlock_irqrestore(&conf->device_lock, flags);
1492
1493        print_conf(conf);
1494        return count;
1495}
1496
1497
1498static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1499{
1500        struct r1conf *conf = mddev->private;
1501        int err = -EEXIST;
1502        int mirror = 0;
1503        struct raid1_info *p;
1504        int first = 0;
1505        int last = conf->raid_disks - 1;
1506        struct request_queue *q = bdev_get_queue(rdev->bdev);
1507
1508        if (mddev->recovery_disabled == conf->recovery_disabled)
1509                return -EBUSY;
1510
1511        if (rdev->raid_disk >= 0)
1512                first = last = rdev->raid_disk;
1513
1514        if (q->merge_bvec_fn) {
1515                set_bit(Unmerged, &rdev->flags);
1516                mddev->merge_check_needed = 1;
1517        }
1518
1519        for (mirror = first; mirror <= last; mirror++) {
1520                p = conf->mirrors+mirror;
1521                if (!p->rdev) {
1522
1523                        if (mddev->gendisk)
1524                                disk_stack_limits(mddev->gendisk, rdev->bdev,
1525                                                  rdev->data_offset << 9);
1526
1527                        p->head_position = 0;
1528                        rdev->raid_disk = mirror;
1529                        err = 0;
1530                        /* As all devices are equivalent, we don't need a full recovery
1531                         * if this was recently any drive of the array
1532                         */
1533                        if (rdev->saved_raid_disk < 0)
1534                                conf->fullsync = 1;
1535                        rcu_assign_pointer(p->rdev, rdev);
1536                        break;
1537                }
1538                if (test_bit(WantReplacement, &p->rdev->flags) &&
1539                    p[conf->raid_disks].rdev == NULL) {
1540                        /* Add this device as a replacement */
1541                        clear_bit(In_sync, &rdev->flags);
1542                        set_bit(Replacement, &rdev->flags);
1543                        rdev->raid_disk = mirror;
1544                        err = 0;
1545                        conf->fullsync = 1;
1546                        rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1547                        break;
1548                }
1549        }
1550        if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1551                /* Some requests might not have seen this new
1552                 * merge_bvec_fn.  We must wait for them to complete
1553                 * before merging the device fully.
1554                 * First we make sure any code which has tested
1555                 * our function has submitted the request, then
1556                 * we wait for all outstanding requests to complete.
1557                 */
1558                synchronize_sched();
1559                freeze_array(conf, 0);
1560                unfreeze_array(conf);
1561                clear_bit(Unmerged, &rdev->flags);
1562        }
1563        md_integrity_add_rdev(rdev, mddev);
1564        if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1565                queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1566        print_conf(conf);
1567        return err;
1568}
1569
1570static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1571{
1572        struct r1conf *conf = mddev->private;
1573        int err = 0;
1574        int number = rdev->raid_disk;
1575        struct raid1_info *p = conf->mirrors + number;
1576
1577        if (rdev != p->rdev)
1578                p = conf->mirrors + conf->raid_disks + number;
1579
1580        print_conf(conf);
1581        if (rdev == p->rdev) {
1582                if (test_bit(In_sync, &rdev->flags) ||
1583                    atomic_read(&rdev->nr_pending)) {
1584                        err = -EBUSY;
1585                        goto abort;
1586                }
1587                /* Only remove non-faulty devices if recovery
1588                 * is not possible.
1589                 */
1590                if (!test_bit(Faulty, &rdev->flags) &&
1591                    mddev->recovery_disabled != conf->recovery_disabled &&
1592                    mddev->degraded < conf->raid_disks) {
1593                        err = -EBUSY;
1594                        goto abort;
1595                }
1596                p->rdev = NULL;
1597                synchronize_rcu();
1598                if (atomic_read(&rdev->nr_pending)) {
1599                        /* lost the race, try later */
1600                        err = -EBUSY;
1601                        p->rdev = rdev;
1602                        goto abort;
1603                } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1604                        /* We just removed a device that is being replaced.
1605                         * Move down the replacement.  We drain all IO before
1606                         * doing this to avoid confusion.
1607                         */
1608                        struct md_rdev *repl =
1609                                conf->mirrors[conf->raid_disks + number].rdev;
1610                        freeze_array(conf, 0);
1611                        clear_bit(Replacement, &repl->flags);
1612                        p->rdev = repl;
1613                        conf->mirrors[conf->raid_disks + number].rdev = NULL;
1614                        unfreeze_array(conf);
1615                        clear_bit(WantReplacement, &rdev->flags);
1616                } else
1617                        clear_bit(WantReplacement, &rdev->flags);
1618                err = md_integrity_register(mddev);
1619        }
1620abort:
1621
1622        print_conf(conf);
1623        return err;
1624}
1625
1626
1627static void end_sync_read(struct bio *bio, int error)
1628{
1629        struct r1bio *r1_bio = bio->bi_private;
1630
1631        update_head_pos(r1_bio->read_disk, r1_bio);
1632
1633        /*
1634         * we have read a block, now it needs to be re-written,
1635         * or re-read if the read failed.
1636         * We don't do much here, just schedule handling by raid1d
1637         */
1638        if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1639                set_bit(R1BIO_Uptodate, &r1_bio->state);
1640
1641        if (atomic_dec_and_test(&r1_bio->remaining))
1642                reschedule_retry(r1_bio);
1643}
1644
1645static void end_sync_write(struct bio *bio, int error)
1646{
1647        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1648        struct r1bio *r1_bio = bio->bi_private;
1649        struct mddev *mddev = r1_bio->mddev;
1650        struct r1conf *conf = mddev->private;
1651        int mirror=0;
1652        sector_t first_bad;
1653        int bad_sectors;
1654
1655        mirror = find_bio_disk(r1_bio, bio);
1656
1657        if (!uptodate) {
1658                sector_t sync_blocks = 0;
1659                sector_t s = r1_bio->sector;
1660                long sectors_to_go = r1_bio->sectors;
1661                /* make sure these bits doesn't get cleared. */
1662                do {
1663                        bitmap_end_sync(mddev->bitmap, s,
1664                                        &sync_blocks, 1);
1665                        s += sync_blocks;
1666                        sectors_to_go -= sync_blocks;
1667                } while (sectors_to_go > 0);
1668                set_bit(WriteErrorSeen,
1669                        &conf->mirrors[mirror].rdev->flags);
1670                if (!test_and_set_bit(WantReplacement,
1671                                      &conf->mirrors[mirror].rdev->flags))
1672                        set_bit(MD_RECOVERY_NEEDED, &
1673                                mddev->recovery);
1674                set_bit(R1BIO_WriteError, &r1_bio->state);
1675        } else if (is_badblock(conf->mirrors[mirror].rdev,
1676                               r1_bio->sector,
1677                               r1_bio->sectors,
1678                               &first_bad, &bad_sectors) &&
1679                   !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1680                                r1_bio->sector,
1681                                r1_bio->sectors,
1682                                &first_bad, &bad_sectors)
1683                )
1684                set_bit(R1BIO_MadeGood, &r1_bio->state);
1685
1686        if (atomic_dec_and_test(&r1_bio->remaining)) {
1687                int s = r1_bio->sectors;
1688                if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1689                    test_bit(R1BIO_WriteError, &r1_bio->state))
1690                        reschedule_retry(r1_bio);
1691                else {
1692                        put_buf(r1_bio);
1693                        md_done_sync(mddev, s, uptodate);
1694                }
1695        }
1696}
1697
1698static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1699                            int sectors, struct page *page, int rw)
1700{
1701        if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1702                /* success */
1703                return 1;
1704        if (rw == WRITE) {
1705                set_bit(WriteErrorSeen, &rdev->flags);
1706                if (!test_and_set_bit(WantReplacement,
1707                                      &rdev->flags))
1708                        set_bit(MD_RECOVERY_NEEDED, &
1709                                rdev->mddev->recovery);
1710        }
1711        /* need to record an error - either for the block or the device */
1712        if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1713                md_error(rdev->mddev, rdev);
1714        return 0;
1715}
1716
1717static int fix_sync_read_error(struct r1bio *r1_bio)
1718{
1719        /* Try some synchronous reads of other devices to get
1720         * good data, much like with normal read errors.  Only
1721         * read into the pages we already have so we don't
1722         * need to re-issue the read request.
1723         * We don't need to freeze the array, because being in an
1724         * active sync request, there is no normal IO, and
1725         * no overlapping syncs.
1726         * We don't need to check is_badblock() again as we
1727         * made sure that anything with a bad block in range
1728         * will have bi_end_io clear.
1729         */
1730        struct mddev *mddev = r1_bio->mddev;
1731        struct r1conf *conf = mddev->private;
1732        struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1733        sector_t sect = r1_bio->sector;
1734        int sectors = r1_bio->sectors;
1735        int idx = 0;
1736
1737        while(sectors) {
1738                int s = sectors;
1739                int d = r1_bio->read_disk;
1740                int success = 0;
1741                struct md_rdev *rdev;
1742                int start;
1743
1744                if (s > (PAGE_SIZE>>9))
1745                        s = PAGE_SIZE >> 9;
1746                do {
1747                        if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1748                                /* No rcu protection needed here devices
1749                                 * can only be removed when no resync is
1750                                 * active, and resync is currently active
1751                                 */
1752                                rdev = conf->mirrors[d].rdev;
1753                                if (sync_page_io(rdev, sect, s<<9,
1754                                                 bio->bi_io_vec[idx].bv_page,
1755                                                 READ, false)) {
1756                                        success = 1;
1757                                        break;
1758                                }
1759                        }
1760                        d++;
1761                        if (d == conf->raid_disks * 2)
1762                                d = 0;
1763                } while (!success && d != r1_bio->read_disk);
1764
1765                if (!success) {
1766                        char b[BDEVNAME_SIZE];
1767                        int abort = 0;
1768                        /* Cannot read from anywhere, this block is lost.
1769                         * Record a bad block on each device.  If that doesn't
1770                         * work just disable and interrupt the recovery.
1771                         * Don't fail devices as that won't really help.
1772                         */
1773                        printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1774                               " for block %llu\n",
1775                               mdname(mddev),
1776                               bdevname(bio->bi_bdev, b),
1777                               (unsigned long long)r1_bio->sector);
1778                        for (d = 0; d < conf->raid_disks * 2; d++) {
1779                                rdev = conf->mirrors[d].rdev;
1780                                if (!rdev || test_bit(Faulty, &rdev->flags))
1781                                        continue;
1782                                if (!rdev_set_badblocks(rdev, sect, s, 0))
1783                                        abort = 1;
1784                        }
1785                        if (abort) {
1786                                conf->recovery_disabled =
1787                                        mddev->recovery_disabled;
1788                                set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1789                                md_done_sync(mddev, r1_bio->sectors, 0);
1790                                put_buf(r1_bio);
1791                                return 0;
1792                        }
1793                        /* Try next page */
1794                        sectors -= s;
1795                        sect += s;
1796                        idx++;
1797                        continue;
1798                }
1799
1800                start = d;
1801                /* write it back and re-read */
1802                while (d != r1_bio->read_disk) {
1803                        if (d == 0)
1804                                d = conf->raid_disks * 2;
1805                        d--;
1806                        if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1807                                continue;
1808                        rdev = conf->mirrors[d].rdev;
1809                        if (r1_sync_page_io(rdev, sect, s,
1810                                            bio->bi_io_vec[idx].bv_page,
1811                                            WRITE) == 0) {
1812                                r1_bio->bios[d]->bi_end_io = NULL;
1813                                rdev_dec_pending(rdev, mddev);
1814                        }
1815                }
1816                d = start;
1817                while (d != r1_bio->read_disk) {
1818                        if (d == 0)
1819                                d = conf->raid_disks * 2;
1820                        d--;
1821                        if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1822                                continue;
1823                        rdev = conf->mirrors[d].rdev;
1824                        if (r1_sync_page_io(rdev, sect, s,
1825                                            bio->bi_io_vec[idx].bv_page,
1826                                            READ) != 0)
1827                                atomic_add(s, &rdev->corrected_errors);
1828                }
1829                sectors -= s;
1830                sect += s;
1831                idx ++;
1832        }
1833        set_bit(R1BIO_Uptodate, &r1_bio->state);
1834        set_bit(BIO_UPTODATE, &bio->bi_flags);
1835        return 1;
1836}
1837
1838static int process_checks(struct r1bio *r1_bio)
1839{
1840        /* We have read all readable devices.  If we haven't
1841         * got the block, then there is no hope left.
1842         * If we have, then we want to do a comparison
1843         * and skip the write if everything is the same.
1844         * If any blocks failed to read, then we need to
1845         * attempt an over-write
1846         */
1847        struct mddev *mddev = r1_bio->mddev;
1848        struct r1conf *conf = mddev->private;
1849        int primary;
1850        int i;
1851        int vcnt;
1852
1853        /* Fix variable parts of all bios */
1854        vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1855        for (i = 0; i < conf->raid_disks * 2; i++) {
1856                int j;
1857                int size;
1858                struct bio *b = r1_bio->bios[i];
1859                if (b->bi_end_io != end_sync_read)
1860                        continue;
1861                /* fixup the bio for reuse */
1862                bio_reset(b);
1863                b->bi_vcnt = vcnt;
1864                b->bi_size = r1_bio->sectors << 9;
1865                b->bi_sector = r1_bio->sector +
1866                        conf->mirrors[i].rdev->data_offset;
1867                b->bi_bdev = conf->mirrors[i].rdev->bdev;
1868                b->bi_end_io = end_sync_read;
1869                b->bi_private = r1_bio;
1870
1871                size = b->bi_size;
1872                for (j = 0; j < vcnt ; j++) {
1873                        struct bio_vec *bi;
1874                        bi = &b->bi_io_vec[j];
1875                        bi->bv_offset = 0;
1876                        if (size > PAGE_SIZE)
1877                                bi->bv_len = PAGE_SIZE;
1878                        else
1879                                bi->bv_len = size;
1880                        size -= PAGE_SIZE;
1881                }
1882        }
1883        for (primary = 0; primary < conf->raid_disks * 2; primary++)
1884                if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1885                    test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1886                        r1_bio->bios[primary]->bi_end_io = NULL;
1887                        rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1888                        break;
1889                }
1890        r1_bio->read_disk = primary;
1891        for (i = 0; i < conf->raid_disks * 2; i++) {
1892                int j;
1893                struct bio *pbio = r1_bio->bios[primary];
1894                struct bio *sbio = r1_bio->bios[i];
1895
1896                if (sbio->bi_end_io != end_sync_read)
1897                        continue;
1898
1899                if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1900                        for (j = vcnt; j-- ; ) {
1901                                struct page *p, *s;
1902                                p = pbio->bi_io_vec[j].bv_page;
1903                                s = sbio->bi_io_vec[j].bv_page;
1904                                if (memcmp(page_address(p),
1905                                           page_address(s),
1906                                           sbio->bi_io_vec[j].bv_len))
1907                                        break;
1908                        }
1909                } else
1910                        j = 0;
1911                if (j >= 0)
1912                        atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
1913                if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1914                              && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1915                        /* No need to write to this device. */
1916                        sbio->bi_end_io = NULL;
1917                        rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1918                        continue;
1919                }
1920
1921                bio_copy_data(sbio, pbio);
1922        }
1923        return 0;
1924}
1925
1926static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1927{
1928        struct r1conf *conf = mddev->private;
1929        int i;
1930        int disks = conf->raid_disks * 2;
1931        struct bio *bio, *wbio;
1932
1933        bio = r1_bio->bios[r1_bio->read_disk];
1934
1935        if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1936                /* ouch - failed to read all of that. */
1937                if (!fix_sync_read_error(r1_bio))
1938                        return;
1939
1940        if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1941                if (process_checks(r1_bio) < 0)
1942                        return;
1943        /*
1944         * schedule writes
1945         */
1946        atomic_set(&r1_bio->remaining, 1);
1947        for (i = 0; i < disks ; i++) {
1948                wbio = r1_bio->bios[i];
1949                if (wbio->bi_end_io == NULL ||
1950                    (wbio->bi_end_io == end_sync_read &&
1951                     (i == r1_bio->read_disk ||
1952                      !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1953                        continue;
1954
1955                wbio->bi_rw = WRITE;
1956                wbio->bi_end_io = end_sync_write;
1957                atomic_inc(&r1_bio->remaining);
1958                md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
1959
1960                generic_make_request(wbio);
1961        }
1962
1963        if (atomic_dec_and_test(&r1_bio->remaining)) {
1964                /* if we're here, all write(s) have completed, so clean up */
1965                int s = r1_bio->sectors;
1966                if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1967                    test_bit(R1BIO_WriteError, &r1_bio->state))
1968                        reschedule_retry(r1_bio);
1969                else {
1970                        put_buf(r1_bio);
1971                        md_done_sync(mddev, s, 1);
1972                }
1973        }
1974}
1975
1976/*
1977 * This is a kernel thread which:
1978 *
1979 *      1.      Retries failed read operations on working mirrors.
1980 *      2.      Updates the raid superblock when problems encounter.
1981 *      3.      Performs writes following reads for array synchronising.
1982 */
1983
1984static void fix_read_error(struct r1conf *conf, int read_disk,
1985                           sector_t sect, int sectors)
1986{
1987        struct mddev *mddev = conf->mddev;
1988        while(sectors) {
1989                int s = sectors;
1990                int d = read_disk;
1991                int success = 0;
1992                int start;
1993                struct md_rdev *rdev;
1994
1995                if (s > (PAGE_SIZE>>9))
1996                        s = PAGE_SIZE >> 9;
1997
1998                do {
1999                        /* Note: no rcu protection needed here
2000                         * as this is synchronous in the raid1d thread
2001                         * which is the thread that might remove
2002                         * a device.  If raid1d ever becomes multi-threaded....
2003                         */
2004                        sector_t first_bad;
2005                        int bad_sectors;
2006
2007                        rdev = conf->mirrors[d].rdev;
2008                        if (rdev &&
2009                            (test_bit(In_sync, &rdev->flags) ||
2010                             (!test_bit(Faulty, &rdev->flags) &&
2011                              rdev->recovery_offset >= sect + s)) &&
2012                            is_badblock(rdev, sect, s,
2013                                        &first_bad, &bad_sectors) == 0 &&
2014                            sync_page_io(rdev, sect, s<<9,
2015                                         conf->tmppage, READ, false))
2016                                success = 1;
2017                        else {
2018                                d++;
2019                                if (d == conf->raid_disks * 2)
2020                                        d = 0;
2021                        }
2022                } while (!success && d != read_disk);
2023
2024                if (!success) {
2025                        /* Cannot read from anywhere - mark it bad */
2026                        struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2027                        if (!rdev_set_badblocks(rdev, sect, s, 0))
2028                                md_error(mddev, rdev);
2029                        break;
2030                }
2031                /* write it back and re-read */
2032                start = d;
2033                while (d != read_disk) {
2034                        if (d==0)
2035                                d = conf->raid_disks * 2;
2036                        d--;
2037                        rdev = conf->mirrors[d].rdev;
2038                        if (rdev &&
2039                            test_bit(In_sync, &rdev->flags))
2040                                r1_sync_page_io(rdev, sect, s,
2041                                                conf->tmppage, WRITE);
2042                }
2043                d = start;
2044                while (d != read_disk) {
2045                        char b[BDEVNAME_SIZE];
2046                        if (d==0)
2047                                d = conf->raid_disks * 2;
2048                        d--;
2049                        rdev = conf->mirrors[d].rdev;
2050                        if (rdev &&
2051                            test_bit(In_sync, &rdev->flags)) {
2052                                if (r1_sync_page_io(rdev, sect, s,
2053                                                    conf->tmppage, READ)) {
2054                                        atomic_add(s, &rdev->corrected_errors);
2055                                        printk(KERN_INFO
2056                                               "md/raid1:%s: read error corrected "
2057                                               "(%d sectors at %llu on %s)\n",
2058                                               mdname(mddev), s,
2059                                               (unsigned long long)(sect +
2060                                                   rdev->data_offset),
2061                                               bdevname(rdev->bdev, b));
2062                                }
2063                        }
2064                }
2065                sectors -= s;
2066                sect += s;
2067        }
2068}
2069
2070static int narrow_write_error(struct r1bio *r1_bio, int i)
2071{
2072        struct mddev *mddev = r1_bio->mddev;
2073        struct r1conf *conf = mddev->private;
2074        struct md_rdev *rdev = conf->mirrors[i].rdev;
2075
2076        /* bio has the data to be written to device 'i' where
2077         * we just recently had a write error.
2078         * We repeatedly clone the bio and trim down to one block,
2079         * then try the write.  Where the write fails we record
2080         * a bad block.
2081         * It is conceivable that the bio doesn't exactly align with
2082         * blocks.  We must handle this somehow.
2083         *
2084         * We currently own a reference on the rdev.
2085         */
2086
2087        int block_sectors;
2088        sector_t sector;
2089        int sectors;
2090        int sect_to_write = r1_bio->sectors;
2091        int ok = 1;
2092
2093        if (rdev->badblocks.shift < 0)
2094                return 0;
2095
2096        block_sectors = 1 << rdev->badblocks.shift;
2097        sector = r1_bio->sector;
2098        sectors = ((sector + block_sectors)
2099                   & ~(sector_t)(block_sectors - 1))
2100                - sector;
2101
2102        while (sect_to_write) {
2103                struct bio *wbio;
2104                if (sectors > sect_to_write)
2105                        sectors = sect_to_write;
2106                /* Write at 'sector' for 'sectors'*/
2107
2108                if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2109                        unsigned vcnt = r1_bio->behind_page_count;
2110                        struct bio_vec *vec = r1_bio->behind_bvecs;
2111
2112                        while (!vec->bv_page) {
2113                                vec++;
2114                                vcnt--;
2115                        }
2116
2117                        wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2118                        memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2119
2120                        wbio->bi_vcnt = vcnt;
2121                } else {
2122                        wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2123                }
2124
2125                wbio->bi_rw = WRITE;
2126                wbio->bi_sector = r1_bio->sector;
2127                wbio->bi_size = r1_bio->sectors << 9;
2128
2129                md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2130                wbio->bi_sector += rdev->data_offset;
2131                wbio->bi_bdev = rdev->bdev;
2132                if (submit_bio_wait(WRITE, wbio) == 0)
2133                        /* failure! */
2134                        ok = rdev_set_badblocks(rdev, sector,
2135                                                sectors, 0)
2136                                && ok;
2137
2138                bio_put(wbio);
2139                sect_to_write -= sectors;
2140                sector += sectors;
2141                sectors = block_sectors;
2142        }
2143        return ok;
2144}
2145
2146static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2147{
2148        int m;
2149        int s = r1_bio->sectors;
2150        for (m = 0; m < conf->raid_disks * 2 ; m++) {
2151                struct md_rdev *rdev = conf->mirrors[m].rdev;
2152                struct bio *bio = r1_bio->bios[m];
2153                if (bio->bi_end_io == NULL)
2154                        continue;
2155                if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2156                    test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2157                        rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2158                }
2159                if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2160                    test_bit(R1BIO_WriteError, &r1_bio->state)) {
2161                        if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2162                                md_error(conf->mddev, rdev);
2163                }
2164        }
2165        put_buf(r1_bio);
2166        md_done_sync(conf->mddev, s, 1);
2167}
2168
2169static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2170{
2171        int m;
2172        for (m = 0; m < conf->raid_disks * 2 ; m++)
2173                if (r1_bio->bios[m] == IO_MADE_GOOD) {
2174                        struct md_rdev *rdev = conf->mirrors[m].rdev;
2175                        rdev_clear_badblocks(rdev,
2176                                             r1_bio->sector,
2177                                             r1_bio->sectors, 0);
2178                        rdev_dec_pending(rdev, conf->mddev);
2179                } else if (r1_bio->bios[m] != NULL) {
2180                        /* This drive got a write error.  We need to
2181                         * narrow down and record precise write
2182                         * errors.
2183                         */
2184                        if (!narrow_write_error(r1_bio, m)) {
2185                                md_error(conf->mddev,
2186                                         conf->mirrors[m].rdev);
2187                                /* an I/O failed, we can't clear the bitmap */
2188                                set_bit(R1BIO_Degraded, &r1_bio->state);
2189                        }
2190                        rdev_dec_pending(conf->mirrors[m].rdev,
2191                                         conf->mddev);
2192                }
2193        if (test_bit(R1BIO_WriteError, &r1_bio->state))
2194                close_write(r1_bio);
2195        raid_end_bio_io(r1_bio);
2196}
2197
2198static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2199{
2200        int disk;
2201        int max_sectors;
2202        struct mddev *mddev = conf->mddev;
2203        struct bio *bio;
2204        char b[BDEVNAME_SIZE];
2205        struct md_rdev *rdev;
2206
2207        clear_bit(R1BIO_ReadError, &r1_bio->state);
2208        /* we got a read error. Maybe the drive is bad.  Maybe just
2209         * the block and we can fix it.
2210         * We freeze all other IO, and try reading the block from
2211         * other devices.  When we find one, we re-write
2212         * and check it that fixes the read error.
2213         * This is all done synchronously while the array is
2214         * frozen
2215         */
2216        if (mddev->ro == 0) {
2217                freeze_array(conf, 1);
2218                fix_read_error(conf, r1_bio->read_disk,
2219                               r1_bio->sector, r1_bio->sectors);
2220                unfreeze_array(conf);
2221        } else
2222                md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2223        rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2224
2225        bio = r1_bio->bios[r1_bio->read_disk];
2226        bdevname(bio->bi_bdev, b);
2227read_more:
2228        disk = read_balance(conf, r1_bio, &max_sectors);
2229        if (disk == -1) {
2230                printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2231                       " read error for block %llu\n",
2232                       mdname(mddev), b, (unsigned long long)r1_bio->sector);
2233                raid_end_bio_io(r1_bio);
2234        } else {
2235                const unsigned long do_sync
2236                        = r1_bio->master_bio->bi_rw & REQ_SYNC;
2237                if (bio) {
2238                        r1_bio->bios[r1_bio->read_disk] =
2239                                mddev->ro ? IO_BLOCKED : NULL;
2240                        bio_put(bio);
2241                }
2242                r1_bio->read_disk = disk;
2243                bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2244                md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2245                r1_bio->bios[r1_bio->read_disk] = bio;
2246                rdev = conf->mirrors[disk].rdev;
2247                printk_ratelimited(KERN_ERR
2248                                   "md/raid1:%s: redirecting sector %llu"
2249                                   " to other mirror: %s\n",
2250                                   mdname(mddev),
2251                                   (unsigned long long)r1_bio->sector,
2252                                   bdevname(rdev->bdev, b));
2253                bio->bi_sector = r1_bio->sector + rdev->data_offset;
2254                bio->bi_bdev = rdev->bdev;
2255                bio->bi_end_io = raid1_end_read_request;
2256                bio->bi_rw = READ | do_sync;
2257                bio->bi_private = r1_bio;
2258                if (max_sectors < r1_bio->sectors) {
2259                        /* Drat - have to split this up more */
2260                        struct bio *mbio = r1_bio->master_bio;
2261                        int sectors_handled = (r1_bio->sector + max_sectors
2262                                               - mbio->bi_sector);
2263                        r1_bio->sectors = max_sectors;
2264                        spin_lock_irq(&conf->device_lock);
2265                        if (mbio->bi_phys_segments == 0)
2266                                mbio->bi_phys_segments = 2;
2267                        else
2268                                mbio->bi_phys_segments++;
2269                        spin_unlock_irq(&conf->device_lock);
2270                        generic_make_request(bio);
2271                        bio = NULL;
2272
2273                        r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2274
2275                        r1_bio->master_bio = mbio;
2276                        r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2277                        r1_bio->state = 0;
2278                        set_bit(R1BIO_ReadError, &r1_bio->state);
2279                        r1_bio->mddev = mddev;
2280                        r1_bio->sector = mbio->bi_sector + sectors_handled;
2281
2282                        goto read_more;
2283                } else
2284                        generic_make_request(bio);
2285        }
2286}
2287
2288static void raid1d(struct md_thread *thread)
2289{
2290        struct mddev *mddev = thread->mddev;
2291        struct r1bio *r1_bio;
2292        unsigned long flags;
2293        struct r1conf *conf = mddev->private;
2294        struct list_head *head = &conf->retry_list;
2295        struct blk_plug plug;
2296
2297        md_check_recovery(mddev);
2298
2299        blk_start_plug(&plug);
2300        for (;;) {
2301
2302                flush_pending_writes(conf);
2303
2304                spin_lock_irqsave(&conf->device_lock, flags);
2305                if (list_empty(head)) {
2306                        spin_unlock_irqrestore(&conf->device_lock, flags);
2307                        break;
2308                }
2309                r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2310                list_del(head->prev);
2311                conf->nr_queued--;
2312                spin_unlock_irqrestore(&conf->device_lock, flags);
2313
2314                mddev = r1_bio->mddev;
2315                conf = mddev->private;
2316                if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2317                        if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2318                            test_bit(R1BIO_WriteError, &r1_bio->state))
2319                                handle_sync_write_finished(conf, r1_bio);
2320                        else
2321                                sync_request_write(mddev, r1_bio);
2322                } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2323                           test_bit(R1BIO_WriteError, &r1_bio->state))
2324                        handle_write_finished(conf, r1_bio);
2325                else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2326                        handle_read_error(conf, r1_bio);
2327                else
2328                        /* just a partial read to be scheduled from separate
2329                         * context
2330                         */
2331                        generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2332
2333                cond_resched();
2334                if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2335                        md_check_recovery(mddev);
2336        }
2337        blk_finish_plug(&plug);
2338}
2339
2340
2341static int init_resync(struct r1conf *conf)
2342{
2343        int buffs;
2344
2345        buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2346        BUG_ON(conf->r1buf_pool);
2347        conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2348                                          conf->poolinfo);
2349        if (!conf->r1buf_pool)
2350                return -ENOMEM;
2351        conf->next_resync = 0;
2352        return 0;
2353}
2354
2355/*
2356 * perform a "sync" on one "block"
2357 *
2358 * We need to make sure that no normal I/O request - particularly write
2359 * requests - conflict with active sync requests.
2360 *
2361 * This is achieved by tracking pending requests and a 'barrier' concept
2362 * that can be installed to exclude normal IO requests.
2363 */
2364
2365static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2366{
2367        struct r1conf *conf = mddev->private;
2368        struct r1bio *r1_bio;
2369        struct bio *bio;
2370        sector_t max_sector, nr_sectors;
2371        int disk = -1;
2372        int i;
2373        int wonly = -1;
2374        int write_targets = 0, read_targets = 0;
2375        sector_t sync_blocks;
2376        int still_degraded = 0;
2377        int good_sectors = RESYNC_SECTORS;
2378        int min_bad = 0; /* number of sectors that are bad in all devices */
2379
2380        if (!conf->r1buf_pool)
2381                if (init_resync(conf))
2382                        return 0;
2383
2384        max_sector = mddev->dev_sectors;
2385        if (sector_nr >= max_sector) {
2386                /* If we aborted, we need to abort the
2387                 * sync on the 'current' bitmap chunk (there will
2388                 * only be one in raid1 resync.
2389                 * We can find the current addess in mddev->curr_resync
2390                 */
2391                if (mddev->curr_resync < max_sector) /* aborted */
2392                        bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2393                                                &sync_blocks, 1);
2394                else /* completed sync */
2395                        conf->fullsync = 0;
2396
2397                bitmap_close_sync(mddev->bitmap);
2398                close_sync(conf);
2399                return 0;
2400        }
2401
2402        if (mddev->bitmap == NULL &&
2403            mddev->recovery_cp == MaxSector &&
2404            !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2405            conf->fullsync == 0) {
2406                *skipped = 1;
2407                return max_sector - sector_nr;
2408        }
2409        /* before building a request, check if we can skip these blocks..
2410         * This call the bitmap_start_sync doesn't actually record anything
2411         */
2412        if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2413            !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2414                /* We can skip this block, and probably several more */
2415                *skipped = 1;
2416                return sync_blocks;
2417        }
2418        /*
2419         * If there is non-resync activity waiting for a turn,
2420         * and resync is going fast enough,
2421         * then let it though before starting on this new sync request.
2422         */
2423        if (!go_faster && conf->nr_waiting)
2424                msleep_interruptible(1000);
2425
2426        bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2427        r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2428        raise_barrier(conf);
2429
2430        conf->next_resync = sector_nr;
2431
2432        rcu_read_lock();
2433        /*
2434         * If we get a correctably read error during resync or recovery,
2435         * we might want to read from a different device.  So we
2436         * flag all drives that could conceivably be read from for READ,
2437         * and any others (which will be non-In_sync devices) for WRITE.
2438         * If a read fails, we try reading from something else for which READ
2439         * is OK.
2440         */
2441
2442        r1_bio->mddev = mddev;
2443        r1_bio->sector = sector_nr;
2444        r1_bio->state = 0;
2445        set_bit(R1BIO_IsSync, &r1_bio->state);
2446
2447        for (i = 0; i < conf->raid_disks * 2; i++) {
2448                struct md_rdev *rdev;
2449                bio = r1_bio->bios[i];
2450                bio_reset(bio);
2451
2452                rdev = rcu_dereference(conf->mirrors[i].rdev);
2453                if (rdev == NULL ||
2454                    test_bit(Faulty, &rdev->flags)) {
2455                        if (i < conf->raid_disks)
2456                                still_degraded = 1;
2457                } else if (!test_bit(In_sync, &rdev->flags)) {
2458                        bio->bi_rw = WRITE;
2459                        bio->bi_end_io = end_sync_write;
2460                        write_targets ++;
2461                } else {
2462                        /* may need to read from here */
2463                        sector_t first_bad = MaxSector;
2464                        int bad_sectors;
2465
2466                        if (is_badblock(rdev, sector_nr, good_sectors,
2467                                        &first_bad, &bad_sectors)) {
2468                                if (first_bad > sector_nr)
2469                                        good_sectors = first_bad - sector_nr;
2470                                else {
2471                                        bad_sectors -= (sector_nr - first_bad);
2472                                        if (min_bad == 0 ||
2473                                            min_bad > bad_sectors)
2474                                                min_bad = bad_sectors;
2475                                }
2476                        }
2477                        if (sector_nr < first_bad) {
2478                                if (test_bit(WriteMostly, &rdev->flags)) {
2479                                        if (wonly < 0)
2480                                                wonly = i;
2481                                } else {
2482                                        if (disk < 0)
2483                                                disk = i;
2484                                }
2485                                bio->bi_rw = READ;
2486                                bio->bi_end_io = end_sync_read;
2487                                read_targets++;
2488                        } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2489                                test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2490                                !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2491                                /*
2492                                 * The device is suitable for reading (InSync),
2493                                 * but has bad block(s) here. Let's try to correct them,
2494                                 * if we are doing resync or repair. Otherwise, leave
2495                                 * this device alone for this sync request.
2496                                 */
2497                                bio->bi_rw = WRITE;
2498                                bio->bi_end_io = end_sync_write;
2499                                write_targets++;
2500                        }
2501                }
2502                if (bio->bi_end_io) {
2503                        atomic_inc(&rdev->nr_pending);
2504                        bio->bi_sector = sector_nr + rdev->data_offset;
2505                        bio->bi_bdev = rdev->bdev;
2506                        bio->bi_private = r1_bio;
2507                }
2508        }
2509        rcu_read_unlock();
2510        if (disk < 0)
2511                disk = wonly;
2512        r1_bio->read_disk = disk;
2513
2514        if (read_targets == 0 && min_bad > 0) {
2515                /* These sectors are bad on all InSync devices, so we
2516                 * need to mark them bad on all write targets
2517                 */
2518                int ok = 1;
2519                for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2520                        if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2521                                struct md_rdev *rdev = conf->mirrors[i].rdev;
2522                                ok = rdev_set_badblocks(rdev, sector_nr,
2523                                                        min_bad, 0
2524                                        ) && ok;
2525                        }
2526                set_bit(MD_CHANGE_DEVS, &mddev->flags);
2527                *skipped = 1;
2528                put_buf(r1_bio);
2529
2530                if (!ok) {
2531                        /* Cannot record the badblocks, so need to
2532                         * abort the resync.
2533                         * If there are multiple read targets, could just
2534                         * fail the really bad ones ???
2535                         */
2536                        conf->recovery_disabled = mddev->recovery_disabled;
2537                        set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2538                        return 0;
2539                } else
2540                        return min_bad;
2541
2542        }
2543        if (min_bad > 0 && min_bad < good_sectors) {
2544                /* only resync enough to reach the next bad->good
2545                 * transition */
2546                good_sectors = min_bad;
2547        }
2548
2549        if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2550                /* extra read targets are also write targets */
2551                write_targets += read_targets-1;
2552
2553        if (write_targets == 0 || read_targets == 0) {
2554                /* There is nowhere to write, so all non-sync
2555                 * drives must be failed - so we are finished
2556                 */
2557                sector_t rv;
2558                if (min_bad > 0)
2559                        max_sector = sector_nr + min_bad;
2560                rv = max_sector - sector_nr;
2561                *skipped = 1;
2562                put_buf(r1_bio);
2563                return rv;
2564        }
2565
2566        if (max_sector > mddev->resync_max)
2567                max_sector = mddev->resync_max; /* Don't do IO beyond here */
2568        if (max_sector > sector_nr + good_sectors)
2569                max_sector = sector_nr + good_sectors;
2570        nr_sectors = 0;
2571        sync_blocks = 0;
2572        do {
2573                struct page *page;
2574                int len = PAGE_SIZE;
2575                if (sector_nr + (len>>9) > max_sector)
2576                        len = (max_sector - sector_nr) << 9;
2577                if (len == 0)
2578                        break;
2579                if (sync_blocks == 0) {
2580                        if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2581                                               &sync_blocks, still_degraded) &&
2582                            !conf->fullsync &&
2583                            !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2584                                break;
2585                        BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2586                        if ((len >> 9) > sync_blocks)
2587                                len = sync_blocks<<9;
2588                }
2589
2590                for (i = 0 ; i < conf->raid_disks * 2; i++) {
2591                        bio = r1_bio->bios[i];
2592                        if (bio->bi_end_io) {
2593                                page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2594                                if (bio_add_page(bio, page, len, 0) == 0) {
2595                                        /* stop here */
2596                                        bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2597                                        while (i > 0) {
2598                                                i--;
2599                                                bio = r1_bio->bios[i];
2600                                                if (bio->bi_end_io==NULL)
2601                                                        continue;
2602                                                /* remove last page from this bio */
2603                                                bio->bi_vcnt--;
2604                                                bio->bi_size -= len;
2605                                                bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2606                                        }
2607                                        goto bio_full;
2608                                }
2609                        }
2610                }
2611                nr_sectors += len>>9;
2612                sector_nr += len>>9;
2613                sync_blocks -= (len>>9);
2614        } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2615 bio_full:
2616        r1_bio->sectors = nr_sectors;
2617
2618        /* For a user-requested sync, we read all readable devices and do a
2619         * compare
2620         */
2621        if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2622                atomic_set(&r1_bio->remaining, read_targets);
2623                for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2624                        bio = r1_bio->bios[i];
2625                        if (bio->bi_end_io == end_sync_read) {
2626                                read_targets--;
2627                                md_sync_acct(bio->bi_bdev, nr_sectors);
2628                                generic_make_request(bio);
2629                        }
2630                }
2631        } else {
2632                atomic_set(&r1_bio->remaining, 1);
2633                bio = r1_bio->bios[r1_bio->read_disk];
2634                md_sync_acct(bio->bi_bdev, nr_sectors);
2635                generic_make_request(bio);
2636
2637        }
2638        return nr_sectors;
2639}
2640
2641static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2642{
2643        if (sectors)
2644                return sectors;
2645
2646        return mddev->dev_sectors;
2647}
2648
2649static struct r1conf *setup_conf(struct mddev *mddev)
2650{
2651        struct r1conf *conf;
2652        int i;
2653        struct raid1_info *disk;
2654        struct md_rdev *rdev;
2655        int err = -ENOMEM;
2656
2657        conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2658        if (!conf)
2659                goto abort;
2660
2661        conf->mirrors = kzalloc(sizeof(struct raid1_info)
2662                                * mddev->raid_disks * 2,
2663                                 GFP_KERNEL);
2664        if (!conf->mirrors)
2665                goto abort;
2666
2667        conf->tmppage = alloc_page(GFP_KERNEL);
2668        if (!conf->tmppage)
2669                goto abort;
2670
2671        conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2672        if (!conf->poolinfo)
2673                goto abort;
2674        conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2675        conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2676                                          r1bio_pool_free,
2677                                          conf->poolinfo);
2678        if (!conf->r1bio_pool)
2679                goto abort;
2680
2681        conf->poolinfo->mddev = mddev;
2682
2683        err = -EINVAL;
2684        spin_lock_init(&conf->device_lock);
2685        rdev_for_each(rdev, mddev) {
2686                struct request_queue *q;
2687                int disk_idx = rdev->raid_disk;
2688                if (disk_idx >= mddev->raid_disks
2689                    || disk_idx < 0)
2690                        continue;
2691                if (test_bit(Replacement, &rdev->flags))
2692                        disk = conf->mirrors + mddev->raid_disks + disk_idx;
2693                else
2694                        disk = conf->mirrors + disk_idx;
2695
2696                if (disk->rdev)
2697                        goto abort;
2698                disk->rdev = rdev;
2699                q = bdev_get_queue(rdev->bdev);
2700                if (q->merge_bvec_fn)
2701                        mddev->merge_check_needed = 1;
2702
2703                disk->head_position = 0;
2704                disk->seq_start = MaxSector;
2705        }
2706        conf->raid_disks = mddev->raid_disks;
2707        conf->mddev = mddev;
2708        INIT_LIST_HEAD(&conf->retry_list);
2709
2710        spin_lock_init(&conf->resync_lock);
2711        init_waitqueue_head(&conf->wait_barrier);
2712
2713        bio_list_init(&conf->pending_bio_list);
2714        conf->pending_count = 0;
2715        conf->recovery_disabled = mddev->recovery_disabled - 1;
2716
2717        err = -EIO;
2718        for (i = 0; i < conf->raid_disks * 2; i++) {
2719
2720                disk = conf->mirrors + i;
2721
2722                if (i < conf->raid_disks &&
2723                    disk[conf->raid_disks].rdev) {
2724                        /* This slot has a replacement. */
2725                        if (!disk->rdev) {
2726                                /* No original, just make the replacement
2727                                 * a recovering spare
2728                                 */
2729                                disk->rdev =
2730                                        disk[conf->raid_disks].rdev;
2731                                disk[conf->raid_disks].rdev = NULL;
2732                        } else if (!test_bit(In_sync, &disk->rdev->flags))
2733                                /* Original is not in_sync - bad */
2734                                goto abort;
2735                }
2736
2737                if (!disk->rdev ||
2738                    !test_bit(In_sync, &disk->rdev->flags)) {
2739                        disk->head_position = 0;
2740                        if (disk->rdev &&
2741                            (disk->rdev->saved_raid_disk < 0))
2742                                conf->fullsync = 1;
2743                }
2744        }
2745
2746        err = -ENOMEM;
2747        conf->thread = md_register_thread(raid1d, mddev, "raid1");
2748        if (!conf->thread) {
2749                printk(KERN_ERR
2750                       "md/raid1:%s: couldn't allocate thread\n",
2751                       mdname(mddev));
2752                goto abort;
2753        }
2754
2755        return conf;
2756
2757 abort:
2758        if (conf) {
2759                if (conf->r1bio_pool)
2760                        mempool_destroy(conf->r1bio_pool);
2761                kfree(conf->mirrors);
2762                safe_put_page(conf->tmppage);
2763                kfree(conf->poolinfo);
2764                kfree(conf);
2765        }
2766        return ERR_PTR(err);
2767}
2768
2769static int stop(struct mddev *mddev);
2770static int run(struct mddev *mddev)
2771{
2772        struct r1conf *conf;
2773        int i;
2774        struct md_rdev *rdev;
2775        int ret;
2776        bool discard_supported = false;
2777
2778        if (mddev->level != 1) {
2779                printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2780                       mdname(mddev), mddev->level);
2781                return -EIO;
2782        }
2783        if (mddev->reshape_position != MaxSector) {
2784                printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2785                       mdname(mddev));
2786                return -EIO;
2787        }
2788        /*
2789         * copy the already verified devices into our private RAID1
2790         * bookkeeping area. [whatever we allocate in run(),
2791         * should be freed in stop()]
2792         */
2793        if (mddev->private == NULL)
2794                conf = setup_conf(mddev);
2795        else
2796                conf = mddev->private;
2797
2798        if (IS_ERR(conf))
2799                return PTR_ERR(conf);
2800
2801        if (mddev->queue)
2802                blk_queue_max_write_same_sectors(mddev->queue, 0);
2803
2804        rdev_for_each(rdev, mddev) {
2805                if (!mddev->gendisk)
2806                        continue;
2807                disk_stack_limits(mddev->gendisk, rdev->bdev,
2808                                  rdev->data_offset << 9);
2809                if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2810                        discard_supported = true;
2811        }
2812
2813        mddev->degraded = 0;
2814        for (i=0; i < conf->raid_disks; i++)
2815                if (conf->mirrors[i].rdev == NULL ||
2816                    !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2817                    test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2818                        mddev->degraded++;
2819
2820        if (conf->raid_disks - mddev->degraded == 1)
2821                mddev->recovery_cp = MaxSector;
2822
2823        if (mddev->recovery_cp != MaxSector)
2824                printk(KERN_NOTICE "md/raid1:%s: not clean"
2825                       " -- starting background reconstruction\n",
2826                       mdname(mddev));
2827        printk(KERN_INFO 
2828                "md/raid1:%s: active with %d out of %d mirrors\n",
2829                mdname(mddev), mddev->raid_disks - mddev->degraded, 
2830                mddev->raid_disks);
2831
2832        /*
2833         * Ok, everything is just fine now
2834         */
2835        mddev->thread = conf->thread;
2836        conf->thread = NULL;
2837        mddev->private = conf;
2838
2839        md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2840
2841        if (mddev->queue) {
2842                mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2843                mddev->queue->backing_dev_info.congested_data = mddev;
2844                blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2845
2846                if (discard_supported)
2847                        queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2848                                                mddev->queue);
2849                else
2850                        queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2851                                                  mddev->queue);
2852        }
2853
2854        ret =  md_integrity_register(mddev);
2855        if (ret)
2856                stop(mddev);
2857        return ret;
2858}
2859
2860static int stop(struct mddev *mddev)
2861{
2862        struct r1conf *conf = mddev->private;
2863        struct bitmap *bitmap = mddev->bitmap;
2864
2865        /* wait for behind writes to complete */
2866        if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2867                printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2868                       mdname(mddev));
2869                /* need to kick something here to make sure I/O goes? */
2870                wait_event(bitmap->behind_wait,
2871                           atomic_read(&bitmap->behind_writes) == 0);
2872        }
2873
2874        raise_barrier(conf);
2875        lower_barrier(conf);
2876
2877        md_unregister_thread(&mddev->thread);
2878        if (conf->r1bio_pool)
2879                mempool_destroy(conf->r1bio_pool);
2880        kfree(conf->mirrors);
2881        safe_put_page(conf->tmppage);
2882        kfree(conf->poolinfo);
2883        kfree(conf);
2884        mddev->private = NULL;
2885        return 0;
2886}
2887
2888static int raid1_resize(struct mddev *mddev, sector_t sectors)
2889{
2890        /* no resync is happening, and there is enough space
2891         * on all devices, so we can resize.
2892         * We need to make sure resync covers any new space.
2893         * If the array is shrinking we should possibly wait until
2894         * any io in the removed space completes, but it hardly seems
2895         * worth it.
2896         */
2897        sector_t newsize = raid1_size(mddev, sectors, 0);
2898        if (mddev->external_size &&
2899            mddev->array_sectors > newsize)
2900                return -EINVAL;
2901        if (mddev->bitmap) {
2902                int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2903                if (ret)
2904                        return ret;
2905        }
2906        md_set_array_sectors(mddev, newsize);
2907        set_capacity(mddev->gendisk, mddev->array_sectors);
2908        revalidate_disk(mddev->gendisk);
2909        if (sectors > mddev->dev_sectors &&
2910            mddev->recovery_cp > mddev->dev_sectors) {
2911                mddev->recovery_cp = mddev->dev_sectors;
2912                set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2913        }
2914        mddev->dev_sectors = sectors;
2915        mddev->resync_max_sectors = sectors;
2916        return 0;
2917}
2918
2919static int raid1_reshape(struct mddev *mddev)
2920{
2921        /* We need to:
2922         * 1/ resize the r1bio_pool
2923         * 2/ resize conf->mirrors
2924         *
2925         * We allocate a new r1bio_pool if we can.
2926         * Then raise a device barrier and wait until all IO stops.
2927         * Then resize conf->mirrors and swap in the new r1bio pool.
2928         *
2929         * At the same time, we "pack" the devices so that all the missing
2930         * devices have the higher raid_disk numbers.
2931         */
2932        mempool_t *newpool, *oldpool;
2933        struct pool_info *newpoolinfo;
2934        struct raid1_info *newmirrors;
2935        struct r1conf *conf = mddev->private;
2936        int cnt, raid_disks;
2937        unsigned long flags;
2938        int d, d2, err;
2939
2940        /* Cannot change chunk_size, layout, or level */
2941        if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2942            mddev->layout != mddev->new_layout ||
2943            mddev->level != mddev->new_level) {
2944                mddev->new_chunk_sectors = mddev->chunk_sectors;
2945                mddev->new_layout = mddev->layout;
2946                mddev->new_level = mddev->level;
2947                return -EINVAL;
2948        }
2949
2950        err = md_allow_write(mddev);
2951        if (err)
2952                return err;
2953
2954        raid_disks = mddev->raid_disks + mddev->delta_disks;
2955
2956        if (raid_disks < conf->raid_disks) {
2957                cnt=0;
2958                for (d= 0; d < conf->raid_disks; d++)
2959                        if (conf->mirrors[d].rdev)
2960                                cnt++;
2961                if (cnt > raid_disks)
2962                        return -EBUSY;
2963        }
2964
2965        newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2966        if (!newpoolinfo)
2967                return -ENOMEM;
2968        newpoolinfo->mddev = mddev;
2969        newpoolinfo->raid_disks = raid_disks * 2;
2970
2971        newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2972                                 r1bio_pool_free, newpoolinfo);
2973        if (!newpool) {
2974                kfree(newpoolinfo);
2975                return -ENOMEM;
2976        }
2977        newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
2978                             GFP_KERNEL);
2979        if (!newmirrors) {
2980                kfree(newpoolinfo);
2981                mempool_destroy(newpool);
2982                return -ENOMEM;
2983        }
2984
2985        freeze_array(conf, 0);
2986
2987        /* ok, everything is stopped */
2988        oldpool = conf->r1bio_pool;
2989        conf->r1bio_pool = newpool;
2990
2991        for (d = d2 = 0; d < conf->raid_disks; d++) {
2992                struct md_rdev *rdev = conf->mirrors[d].rdev;
2993                if (rdev && rdev->raid_disk != d2) {
2994                        sysfs_unlink_rdev(mddev, rdev);
2995                        rdev->raid_disk = d2;
2996                        sysfs_unlink_rdev(mddev, rdev);
2997                        if (sysfs_link_rdev(mddev, rdev))
2998                                printk(KERN_WARNING
2999                                       "md/raid1:%s: cannot register rd%d\n",
3000                                       mdname(mddev), rdev->raid_disk);
3001                }
3002                if (rdev)
3003                        newmirrors[d2++].rdev = rdev;
3004        }
3005        kfree(conf->mirrors);
3006        conf->mirrors = newmirrors;
3007        kfree(conf->poolinfo);
3008        conf->poolinfo = newpoolinfo;
3009
3010        spin_lock_irqsave(&conf->device_lock, flags);
3011        mddev->degraded += (raid_disks - conf->raid_disks);
3012        spin_unlock_irqrestore(&conf->device_lock, flags);
3013        conf->raid_disks = mddev->raid_disks = raid_disks;
3014        mddev->delta_disks = 0;
3015
3016        unfreeze_array(conf);
3017
3018        set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3019        md_wakeup_thread(mddev->thread);
3020
3021        mempool_destroy(oldpool);
3022        return 0;
3023}
3024
3025static void raid1_quiesce(struct mddev *mddev, int state)
3026{
3027        struct r1conf *conf = mddev->private;
3028
3029        switch(state) {
3030        case 2: /* wake for suspend */
3031                wake_up(&conf->wait_barrier);
3032                break;
3033        case 1:
3034                raise_barrier(conf);
3035                break;
3036        case 0:
3037                lower_barrier(conf);
3038                break;
3039        }
3040}
3041
3042static void *raid1_takeover(struct mddev *mddev)
3043{
3044        /* raid1 can take over:
3045         *  raid5 with 2 devices, any layout or chunk size
3046         */
3047        if (mddev->level == 5 && mddev->raid_disks == 2) {
3048                struct r1conf *conf;
3049                mddev->new_level = 1;
3050                mddev->new_layout = 0;
3051                mddev->new_chunk_sectors = 0;
3052                conf = setup_conf(mddev);
3053                if (!IS_ERR(conf))
3054                        conf->barrier = 1;
3055                return conf;
3056        }
3057        return ERR_PTR(-EINVAL);
3058}
3059
3060static struct md_personality raid1_personality =
3061{
3062        .name           = "raid1",
3063        .level          = 1,
3064        .owner          = THIS_MODULE,
3065        .make_request   = make_request,
3066        .run            = run,
3067        .stop           = stop,
3068        .status         = status,
3069        .error_handler  = error,
3070        .hot_add_disk   = raid1_add_disk,
3071        .hot_remove_disk= raid1_remove_disk,
3072        .spare_active   = raid1_spare_active,
3073        .sync_request   = sync_request,
3074        .resize         = raid1_resize,
3075        .size           = raid1_size,
3076        .check_reshape  = raid1_reshape,
3077        .quiesce        = raid1_quiesce,
3078        .takeover       = raid1_takeover,
3079};
3080
3081static int __init raid_init(void)
3082{
3083        return register_md_personality(&raid1_personality);
3084}
3085
3086static void raid_exit(void)
3087{
3088        unregister_md_personality(&raid1_personality);
3089}
3090
3091module_init(raid_init);
3092module_exit(raid_exit);
3093MODULE_LICENSE("GPL");
3094MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3095MODULE_ALIAS("md-personality-3"); /* RAID1 */
3096MODULE_ALIAS("md-raid1");
3097MODULE_ALIAS("md-level-1");
3098
3099module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
3100
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.