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 page *page;
  96        struct r1bio *r1_bio;
  97        struct bio *bio;
  98        int i, j;
  99
 100        r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 101        if (!r1_bio)
 102                return NULL;
 103
 104        /*
 105         * Allocate bios : 1 for reading, n-1 for writing
 106         */
 107        for (j = pi->raid_disks ; j-- ; ) {
 108                bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 109                if (!bio)
 110                        goto out_free_bio;
 111                r1_bio->bios[j] = bio;
 112        }
 113        /*
 114         * Allocate RESYNC_PAGES data pages and attach them to
 115         * the first bio.
 116         * If this is a user-requested check/repair, allocate
 117         * RESYNC_PAGES for each bio.
 118         */
 119        if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 120                j = pi->raid_disks;
 121        else
 122                j = 1;
 123        while(j--) {
 124                bio = r1_bio->bios[j];
 125                for (i = 0; i < RESYNC_PAGES; i++) {
 126                        page = alloc_page(gfp_flags);
 127                        if (unlikely(!page))
 128                                goto out_free_pages;
 129
 130                        bio->bi_io_vec[i].bv_page = page;
 131                        bio->bi_vcnt = i+1;
 132                }
 133        }
 134        /* If not user-requests, copy the page pointers to all bios */
 135        if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 136                for (i=0; i<RESYNC_PAGES ; i++)
 137                        for (j=1; j<pi->raid_disks; j++)
 138                                r1_bio->bios[j]->bi_io_vec[i].bv_page =
 139                                        r1_bio->bios[0]->bi_io_vec[i].bv_page;
 140        }
 141
 142        r1_bio->master_bio = NULL;
 143
 144        return r1_bio;
 145
 146out_free_pages:
 147        for (j=0 ; j < pi->raid_disks; j++)
 148                for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 149                        put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 150        j = -1;
 151out_free_bio:
 152        while (++j < pi->raid_disks)
 153                bio_put(r1_bio->bios[j]);
 154        r1bio_pool_free(r1_bio, data);
 155        return NULL;
 156}
 157
 158static void r1buf_pool_free(void *__r1_bio, void *data)
 159{
 160        struct pool_info *pi = data;
 161        int i,j;
 162        struct r1bio *r1bio = __r1_bio;
 163
 164        for (i = 0; i < RESYNC_PAGES; i++)
 165                for (j = pi->raid_disks; j-- ;) {
 166                        if (j == 0 ||
 167                            r1bio->bios[j]->bi_io_vec[i].bv_page !=
 168                            r1bio->bios[0]->bi_io_vec[i].bv_page)
 169                                safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 170                }
 171        for (i=0 ; i < pi->raid_disks; i++)
 172                bio_put(r1bio->bios[i]);
 173
 174        r1bio_pool_free(r1bio, data);
 175}
 176
 177static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 178{
 179        int i;
 180
 181        for (i = 0; i < conf->raid_disks * 2; i++) {
 182                struct bio **bio = r1_bio->bios + i;
 183                if (!BIO_SPECIAL(*bio))
 184                        bio_put(*bio);
 185                *bio = NULL;
 186        }
 187}
 188
 189static void free_r1bio(struct r1bio *r1_bio)
 190{
 191        struct r1conf *conf = r1_bio->mddev->private;
 192
 193        put_all_bios(conf, r1_bio);
 194        mempool_free(r1_bio, conf->r1bio_pool);
 195}
 196
 197static void put_buf(struct r1bio *r1_bio)
 198{
 199        struct r1conf *conf = r1_bio->mddev->private;
 200        int i;
 201
 202        for (i = 0; i < conf->raid_disks * 2; i++) {
 203                struct bio *bio = r1_bio->bios[i];
 204                if (bio->bi_end_io)
 205                        rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 206        }
 207
 208        mempool_free(r1_bio, conf->r1buf_pool);
 209
 210        lower_barrier(conf);
 211}
 212
 213static void reschedule_retry(struct r1bio *r1_bio)
 214{
 215        unsigned long flags;
 216        struct mddev *mddev = r1_bio->mddev;
 217        struct r1conf *conf = mddev->private;
 218
 219        spin_lock_irqsave(&conf->device_lock, flags);
 220        list_add(&r1_bio->retry_list, &conf->retry_list);
 221        conf->nr_queued ++;
 222        spin_unlock_irqrestore(&conf->device_lock, flags);
 223
 224        wake_up(&conf->wait_barrier);
 225        md_wakeup_thread(mddev->thread);
 226}
 227
 228/*
 229 * raid_end_bio_io() is called when we have finished servicing a mirrored
 230 * operation and are ready to return a success/failure code to the buffer
 231 * cache layer.
 232 */
 233static void call_bio_endio(struct r1bio *r1_bio)
 234{
 235        struct bio *bio = r1_bio->master_bio;
 236        int done;
 237        struct r1conf *conf = r1_bio->mddev->private;
 238
 239        if (bio->bi_phys_segments) {
 240                unsigned long flags;
 241                spin_lock_irqsave(&conf->device_lock, flags);
 242                bio->bi_phys_segments--;
 243                done = (bio->bi_phys_segments == 0);
 244                spin_unlock_irqrestore(&conf->device_lock, flags);
 245        } else
 246                done = 1;
 247
 248        if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 249                clear_bit(BIO_UPTODATE, &bio->bi_flags);
 250        if (done) {
 251                bio_endio(bio, 0);
 252                /*
 253                 * Wake up any possible resync thread that waits for the device
 254                 * to go idle.
 255                 */
 256                allow_barrier(conf);
 257        }
 258}
 259
 260static void raid_end_bio_io(struct r1bio *r1_bio)
 261{
 262        struct bio *bio = r1_bio->master_bio;
 263
 264        /* if nobody has done the final endio yet, do it now */
 265        if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 266                pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 267                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 268                         (unsigned long long) bio->bi_sector,
 269                         (unsigned long long) bio->bi_sector +
 270                         (bio->bi_size >> 9) - 1);
 271
 272                call_bio_endio(r1_bio);
 273        }
 274        free_r1bio(r1_bio);
 275}
 276
 277/*
 278 * Update disk head position estimator based on IRQ completion info.
 279 */
 280static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 281{
 282        struct r1conf *conf = r1_bio->mddev->private;
 283
 284        conf->mirrors[disk].head_position =
 285                r1_bio->sector + (r1_bio->sectors);
 286}
 287
 288/*
 289 * Find the disk number which triggered given bio
 290 */
 291static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 292{
 293        int mirror;
 294        struct r1conf *conf = r1_bio->mddev->private;
 295        int raid_disks = conf->raid_disks;
 296
 297        for (mirror = 0; mirror < raid_disks * 2; mirror++)
 298                if (r1_bio->bios[mirror] == bio)
 299                        break;
 300
 301        BUG_ON(mirror == raid_disks * 2);
 302        update_head_pos(mirror, r1_bio);
 303
 304        return mirror;
 305}
 306
 307static void raid1_end_read_request(struct bio *bio, int error)
 308{
 309        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 310        struct r1bio *r1_bio = bio->bi_private;
 311        int mirror;
 312        struct r1conf *conf = r1_bio->mddev->private;
 313
 314        mirror = r1_bio->read_disk;
 315        /*
 316         * this branch is our 'one mirror IO has finished' event handler:
 317         */
 318        update_head_pos(mirror, r1_bio);
 319
 320        if (uptodate)
 321                set_bit(R1BIO_Uptodate, &r1_bio->state);
 322        else {
 323                /* If all other devices have failed, we want to return
 324                 * the error upwards rather than fail the last device.
 325                 * Here we redefine "uptodate" to mean "Don't want to retry"
 326                 */
 327                unsigned long flags;
 328                spin_lock_irqsave(&conf->device_lock, flags);
 329                if (r1_bio->mddev->degraded == conf->raid_disks ||
 330                    (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 331                     !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 332                        uptodate = 1;
 333                spin_unlock_irqrestore(&conf->device_lock, flags);
 334        }
 335
 336        if (uptodate) {
 337                raid_end_bio_io(r1_bio);
 338                rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 339        } else {
 340                /*
 341                 * oops, read error:
 342                 */
 343                char b[BDEVNAME_SIZE];
 344                printk_ratelimited(
 345                        KERN_ERR "md/raid1:%s: %s: "
 346                        "rescheduling sector %llu\n",
 347                        mdname(conf->mddev),
 348                        bdevname(conf->mirrors[mirror].rdev->bdev,
 349                                 b),
 350                        (unsigned long long)r1_bio->sector);
 351                set_bit(R1BIO_ReadError, &r1_bio->state);
 352                reschedule_retry(r1_bio);
 353                /* don't drop the reference on read_disk yet */
 354        }
 355}
 356
 357static void close_write(struct r1bio *r1_bio)
 358{
 359        /* it really is the end of this request */
 360        if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 361                /* free extra copy of the data pages */
 362                int i = r1_bio->behind_page_count;
 363                while (i--)
 364                        safe_put_page(r1_bio->behind_bvecs[i].bv_page);
 365                kfree(r1_bio->behind_bvecs);
 366                r1_bio->behind_bvecs = NULL;
 367        }
 368        /* clear the bitmap if all writes complete successfully */
 369        bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 370                        r1_bio->sectors,
 371                        !test_bit(R1BIO_Degraded, &r1_bio->state),
 372                        test_bit(R1BIO_BehindIO, &r1_bio->state));
 373        md_write_end(r1_bio->mddev);
 374}
 375
 376static void r1_bio_write_done(struct r1bio *r1_bio)
 377{
 378        if (!atomic_dec_and_test(&r1_bio->remaining))
 379                return;
 380
 381        if (test_bit(R1BIO_WriteError, &r1_bio->state))
 382                reschedule_retry(r1_bio);
 383        else {
 384                close_write(r1_bio);
 385                if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 386                        reschedule_retry(r1_bio);
 387                else
 388                        raid_end_bio_io(r1_bio);
 389        }
 390}
 391
 392static void raid1_end_write_request(struct bio *bio, int error)
 393{
 394        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 395        struct r1bio *r1_bio = bio->bi_private;
 396        int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 397        struct r1conf *conf = r1_bio->mddev->private;
 398        struct bio *to_put = NULL;
 399
 400        mirror = find_bio_disk(r1_bio, bio);
 401
 402        /*
 403         * 'one mirror IO has finished' event handler:
 404         */
 405        if (!uptodate) {
 406                set_bit(WriteErrorSeen,
 407                        &conf->mirrors[mirror].rdev->flags);
 408                if (!test_and_set_bit(WantReplacement,
 409                                      &conf->mirrors[mirror].rdev->flags))
 410                        set_bit(MD_RECOVERY_NEEDED, &
 411                                conf->mddev->recovery);
 412
 413                set_bit(R1BIO_WriteError, &r1_bio->state);
 414        } else {
 415                /*
 416                 * Set R1BIO_Uptodate in our master bio, so that we
 417                 * will return a good error code for to the higher
 418                 * levels even if IO on some other mirrored buffer
 419                 * fails.
 420                 *
 421                 * The 'master' represents the composite IO operation
 422                 * to user-side. So if something waits for IO, then it
 423                 * will wait for the 'master' bio.
 424                 */
 425                sector_t first_bad;
 426                int bad_sectors;
 427
 428                r1_bio->bios[mirror] = NULL;
 429                to_put = bio;
 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                                         (mbio->bi_size >> 9) - 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)
 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+1
 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 (1)
 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+1,
 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(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->bi_sector + bio->bi_size/512 > 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->bi_sector + bio->bi_size/512 <= 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->bi_size >> 9;
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->bi_size >> 9) - 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                        /* Yes, I really want the '__' version so that
1293                         * we clear any unused pointer in the io_vec, rather
1294                         * than leave them unchanged.  This is important
1295                         * because when we come to free the pages, we won't
1296                         * know the original bi_idx, so we just free
1297                         * them all
1298                         */
1299                        __bio_for_each_segment(bvec, mbio, j, 0)
1300                                bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1301                        if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1302                                atomic_inc(&r1_bio->behind_remaining);
1303                }
1304
1305                r1_bio->bios[i] = mbio;
1306
1307                mbio->bi_sector = (r1_bio->sector +
1308                                   conf->mirrors[i].rdev->data_offset);
1309                mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1310                mbio->bi_end_io = raid1_end_write_request;
1311                mbio->bi_rw =
1312                        WRITE | do_flush_fua | do_sync | do_discard | do_same;
1313                mbio->bi_private = r1_bio;
1314
1315                atomic_inc(&r1_bio->remaining);
1316
1317                cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1318                if (cb)
1319                        plug = container_of(cb, struct raid1_plug_cb, cb);
1320                else
1321                        plug = NULL;
1322                spin_lock_irqsave(&conf->device_lock, flags);
1323                if (plug) {
1324                        bio_list_add(&plug->pending, mbio);
1325                        plug->pending_cnt++;
1326                } else {
1327                        bio_list_add(&conf->pending_bio_list, mbio);
1328                        conf->pending_count++;
1329                }
1330                spin_unlock_irqrestore(&conf->device_lock, flags);
1331                if (!plug)
1332                        md_wakeup_thread(mddev->thread);
1333        }
1334        /* Mustn't call r1_bio_write_done before this next test,
1335         * as it could result in the bio being freed.
1336         */
1337        if (sectors_handled < (bio->bi_size >> 9)) {
1338                r1_bio_write_done(r1_bio);
1339                /* We need another r1_bio.  It has already been counted
1340                 * in bio->bi_phys_segments
1341                 */
1342                r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1343                r1_bio->master_bio = bio;
1344                r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1345                r1_bio->state = 0;
1346                r1_bio->mddev = mddev;
1347                r1_bio->sector = bio->bi_sector + sectors_handled;
1348                goto retry_write;
1349        }
1350
1351        r1_bio_write_done(r1_bio);
1352
1353        /* In case raid1d snuck in to freeze_array */
1354        wake_up(&conf->wait_barrier);
1355}
1356
1357static void status(struct seq_file *seq, struct mddev *mddev)
1358{
1359        struct r1conf *conf = mddev->private;
1360        int i;
1361
1362        seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1363                   conf->raid_disks - mddev->degraded);
1364        rcu_read_lock();
1365        for (i = 0; i < conf->raid_disks; i++) {
1366                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1367                seq_printf(seq, "%s",
1368                           rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1369        }
1370        rcu_read_unlock();
1371        seq_printf(seq, "]");
1372}
1373
1374
1375static void error(struct mddev *mddev, struct md_rdev *rdev)
1376{
1377        char b[BDEVNAME_SIZE];
1378        struct r1conf *conf = mddev->private;
1379
1380        /*
1381         * If it is not operational, then we have already marked it as dead
1382         * else if it is the last working disks, ignore the error, let the
1383         * next level up know.
1384         * else mark the drive as failed
1385         */
1386        if (test_bit(In_sync, &rdev->flags)
1387            && (conf->raid_disks - mddev->degraded) == 1) {
1388                /*
1389                 * Don't fail the drive, act as though we were just a
1390                 * normal single drive.
1391                 * However don't try a recovery from this drive as
1392                 * it is very likely to fail.
1393                 */
1394                conf->recovery_disabled = mddev->recovery_disabled;
1395                return;
1396        }
1397        set_bit(Blocked, &rdev->flags);
1398        if (test_and_clear_bit(In_sync, &rdev->flags)) {
1399                unsigned long flags;
1400                spin_lock_irqsave(&conf->device_lock, flags);
1401                mddev->degraded++;
1402                set_bit(Faulty, &rdev->flags);
1403                spin_unlock_irqrestore(&conf->device_lock, flags);
1404                /*
1405                 * if recovery is running, make sure it aborts.
1406                 */
1407                set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1408        } else
1409                set_bit(Faulty, &rdev->flags);
1410        set_bit(MD_CHANGE_DEVS, &mddev->flags);
1411        printk(KERN_ALERT
1412               "md/raid1:%s: Disk failure on %s, disabling device.\n"
1413               "md/raid1:%s: Operation continuing on %d devices.\n",
1414               mdname(mddev), bdevname(rdev->bdev, b),
1415               mdname(mddev), conf->raid_disks - mddev->degraded);
1416}
1417
1418static void print_conf(struct r1conf *conf)
1419{
1420        int i;
1421
1422        printk(KERN_DEBUG "RAID1 conf printout:\n");
1423        if (!conf) {
1424                printk(KERN_DEBUG "(!conf)\n");
1425                return;
1426        }
1427        printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1428                conf->raid_disks);
1429
1430        rcu_read_lock();
1431        for (i = 0; i < conf->raid_disks; i++) {
1432                char b[BDEVNAME_SIZE];
1433                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1434                if (rdev)
1435                        printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1436                               i, !test_bit(In_sync, &rdev->flags),
1437                               !test_bit(Faulty, &rdev->flags),
1438                               bdevname(rdev->bdev,b));
1439        }
1440        rcu_read_unlock();
1441}
1442
1443static void close_sync(struct r1conf *conf)
1444{
1445        wait_barrier(conf);
1446        allow_barrier(conf);
1447
1448        mempool_destroy(conf->r1buf_pool);
1449        conf->r1buf_pool = NULL;
1450}
1451
1452static int raid1_spare_active(struct mddev *mddev)
1453{
1454        int i;
1455        struct r1conf *conf = mddev->private;
1456        int count = 0;
1457        unsigned long flags;
1458
1459        /*
1460         * Find all failed disks within the RAID1 configuration 
1461         * and mark them readable.
1462         * Called under mddev lock, so rcu protection not needed.
1463         */
1464        for (i = 0; i < conf->raid_disks; i++) {
1465                struct md_rdev *rdev = conf->mirrors[i].rdev;
1466                struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1467                if (repl
1468                    && repl->recovery_offset == MaxSector
1469                    && !test_bit(Faulty, &repl->flags)
1470                    && !test_and_set_bit(In_sync, &repl->flags)) {
1471                        /* replacement has just become active */
1472                        if (!rdev ||
1473                            !test_and_clear_bit(In_sync, &rdev->flags))
1474                                count++;
1475                        if (rdev) {
1476                                /* Replaced device not technically
1477                                 * faulty, but we need to be sure
1478                                 * it gets removed and never re-added
1479                                 */
1480                                set_bit(Faulty, &rdev->flags);
1481                                sysfs_notify_dirent_safe(
1482                                        rdev->sysfs_state);
1483                        }
1484                }
1485                if (rdev
1486                    && !test_bit(Faulty, &rdev->flags)
1487                    && !test_and_set_bit(In_sync, &rdev->flags)) {
1488                        count++;
1489                        sysfs_notify_dirent_safe(rdev->sysfs_state);
1490                }
1491        }
1492        spin_lock_irqsave(&conf->device_lock, flags);
1493        mddev->degraded -= count;
1494        spin_unlock_irqrestore(&conf->device_lock, flags);
1495
1496        print_conf(conf);
1497        return count;
1498}
1499
1500
1501static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1502{
1503        struct r1conf *conf = mddev->private;
1504        int err = -EEXIST;
1505        int mirror = 0;
1506        struct raid1_info *p;
1507        int first = 0;
1508        int last = conf->raid_disks - 1;
1509        struct request_queue *q = bdev_get_queue(rdev->bdev);
1510
1511        if (mddev->recovery_disabled == conf->recovery_disabled)
1512                return -EBUSY;
1513
1514        if (rdev->raid_disk >= 0)
1515                first = last = rdev->raid_disk;
1516
1517        if (q->merge_bvec_fn) {
1518                set_bit(Unmerged, &rdev->flags);
1519                mddev->merge_check_needed = 1;
1520        }
1521
1522        for (mirror = first; mirror <= last; mirror++) {
1523                p = conf->mirrors+mirror;
1524                if (!p->rdev) {
1525
1526                        disk_stack_limits(mddev->gendisk, rdev->bdev,
1527                                          rdev->data_offset << 9);
1528
1529                        p->head_position = 0;
1530                        rdev->raid_disk = mirror;
1531                        err = 0;
1532                        /* As all devices are equivalent, we don't need a full recovery
1533                         * if this was recently any drive of the array
1534                         */
1535                        if (rdev->saved_raid_disk < 0)
1536                                conf->fullsync = 1;
1537                        rcu_assign_pointer(p->rdev, rdev);
1538                        break;
1539                }
1540                if (test_bit(WantReplacement, &p->rdev->flags) &&
1541                    p[conf->raid_disks].rdev == NULL) {
1542                        /* Add this device as a replacement */
1543                        clear_bit(In_sync, &rdev->flags);
1544                        set_bit(Replacement, &rdev->flags);
1545                        rdev->raid_disk = mirror;
1546                        err = 0;
1547                        conf->fullsync = 1;
1548                        rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1549                        break;
1550                }
1551        }
1552        if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1553                /* Some requests might not have seen this new
1554                 * merge_bvec_fn.  We must wait for them to complete
1555                 * before merging the device fully.
1556                 * First we make sure any code which has tested
1557                 * our function has submitted the request, then
1558                 * we wait for all outstanding requests to complete.
1559                 */
1560                synchronize_sched();
1561                raise_barrier(conf);
1562                lower_barrier(conf);
1563                clear_bit(Unmerged, &rdev->flags);
1564        }
1565        md_integrity_add_rdev(rdev, mddev);
1566        if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
1567                queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1568        print_conf(conf);
1569        return err;
1570}
1571
1572static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1573{
1574        struct r1conf *conf = mddev->private;
1575        int err = 0;
1576        int number = rdev->raid_disk;
1577        struct raid1_info *p = conf->mirrors + number;
1578
1579        if (rdev != p->rdev)
1580                p = conf->mirrors + conf->raid_disks + number;
1581
1582        print_conf(conf);
1583        if (rdev == p->rdev) {
1584                if (test_bit(In_sync, &rdev->flags) ||
1585                    atomic_read(&rdev->nr_pending)) {
1586                        err = -EBUSY;
1587                        goto abort;
1588                }
1589                /* Only remove non-faulty devices if recovery
1590                 * is not possible.
1591                 */
1592                if (!test_bit(Faulty, &rdev->flags) &&
1593                    mddev->recovery_disabled != conf->recovery_disabled &&
1594                    mddev->degraded < conf->raid_disks) {
1595                        err = -EBUSY;
1596                        goto abort;
1597                }
1598                p->rdev = NULL;
1599                synchronize_rcu();
1600                if (atomic_read(&rdev->nr_pending)) {
1601                        /* lost the race, try later */
1602                        err = -EBUSY;
1603                        p->rdev = rdev;
1604                        goto abort;
1605                } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1606                        /* We just removed a device that is being replaced.
1607                         * Move down the replacement.  We drain all IO before
1608                         * doing this to avoid confusion.
1609                         */
1610                        struct md_rdev *repl =
1611                                conf->mirrors[conf->raid_disks + number].rdev;
1612                        raise_barrier(conf);
1613                        clear_bit(Replacement, &repl->flags);
1614                        p->rdev = repl;
1615                        conf->mirrors[conf->raid_disks + number].rdev = NULL;
1616                        lower_barrier(conf);
1617                        clear_bit(WantReplacement, &rdev->flags);
1618                } else
1619                        clear_bit(WantReplacement, &rdev->flags);
1620                err = md_integrity_register(mddev);
1621        }
1622abort:
1623
1624        print_conf(conf);
1625        return err;
1626}
1627
1628
1629static void end_sync_read(struct bio *bio, int error)
1630{
1631        struct r1bio *r1_bio = bio->bi_private;
1632
1633        update_head_pos(r1_bio->read_disk, r1_bio);
1634
1635        /*
1636         * we have read a block, now it needs to be re-written,
1637         * or re-read if the read failed.
1638         * We don't do much here, just schedule handling by raid1d
1639         */
1640        if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1641                set_bit(R1BIO_Uptodate, &r1_bio->state);
1642
1643        if (atomic_dec_and_test(&r1_bio->remaining))
1644                reschedule_retry(r1_bio);
1645}
1646
1647static void end_sync_write(struct bio *bio, int error)
1648{
1649        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1650        struct r1bio *r1_bio = bio->bi_private;
1651        struct mddev *mddev = r1_bio->mddev;
1652        struct r1conf *conf = mddev->private;
1653        int mirror=0;
1654        sector_t first_bad;
1655        int bad_sectors;
1656
1657        mirror = find_bio_disk(r1_bio, bio);
1658
1659        if (!uptodate) {
1660                sector_t sync_blocks = 0;
1661                sector_t s = r1_bio->sector;
1662                long sectors_to_go = r1_bio->sectors;
1663                /* make sure these bits doesn't get cleared. */
1664                do {
1665                        bitmap_end_sync(mddev->bitmap, s,
1666                                        &sync_blocks, 1);
1667                        s += sync_blocks;
1668                        sectors_to_go -= sync_blocks;
1669                } while (sectors_to_go > 0);
1670                set_bit(WriteErrorSeen,
1671                        &conf->mirrors[mirror].rdev->flags);
1672                if (!test_and_set_bit(WantReplacement,
1673                                      &conf->mirrors[mirror].rdev->flags))
1674                        set_bit(MD_RECOVERY_NEEDED, &
1675                                mddev->recovery);
1676                set_bit(R1BIO_WriteError, &r1_bio->state);
1677        } else if (is_badblock(conf->mirrors[mirror].rdev,
1678                               r1_bio->sector,
1679                               r1_bio->sectors,
1680                               &first_bad, &bad_sectors) &&
1681                   !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1682                                r1_bio->sector,
1683                                r1_bio->sectors,
1684                                &first_bad, &bad_sectors)
1685                )
1686                set_bit(R1BIO_MadeGood, &r1_bio->state);
1687
1688        if (atomic_dec_and_test(&r1_bio->remaining)) {
1689                int s = r1_bio->sectors;
1690                if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1691                    test_bit(R1BIO_WriteError, &r1_bio->state))
1692                        reschedule_retry(r1_bio);
1693                else {
1694                        put_buf(r1_bio);
1695                        md_done_sync(mddev, s, uptodate);
1696                }
1697        }
1698}
1699
1700static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1701                            int sectors, struct page *page, int rw)
1702{
1703        if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1704                /* success */
1705                return 1;
1706        if (rw == WRITE) {
1707                set_bit(WriteErrorSeen, &rdev->flags);
1708                if (!test_and_set_bit(WantReplacement,
1709                                      &rdev->flags))
1710                        set_bit(MD_RECOVERY_NEEDED, &
1711                                rdev->mddev->recovery);
1712        }
1713        /* need to record an error - either for the block or the device */
1714        if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1715                md_error(rdev->mddev, rdev);
1716        return 0;
1717}
1718
1719static int fix_sync_read_error(struct r1bio *r1_bio)
1720{
1721        /* Try some synchronous reads of other devices to get
1722         * good data, much like with normal read errors.  Only
1723         * read into the pages we already have so we don't
1724         * need to re-issue the read request.
1725         * We don't need to freeze the array, because being in an
1726         * active sync request, there is no normal IO, and
1727         * no overlapping syncs.
1728         * We don't need to check is_badblock() again as we
1729         * made sure that anything with a bad block in range
1730         * will have bi_end_io clear.
1731         */
1732        struct mddev *mddev = r1_bio->mddev;
1733        struct r1conf *conf = mddev->private;
1734        struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1735        sector_t sect = r1_bio->sector;
1736        int sectors = r1_bio->sectors;
1737        int idx = 0;
1738
1739        while(sectors) {
1740                int s = sectors;
1741                int d = r1_bio->read_disk;
1742                int success = 0;
1743                struct md_rdev *rdev;
1744                int start;
1745
1746                if (s > (PAGE_SIZE>>9))
1747                        s = PAGE_SIZE >> 9;
1748                do {
1749                        if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1750                                /* No rcu protection needed here devices
1751                                 * can only be removed when no resync is
1752                                 * active, and resync is currently active
1753                                 */
1754                                rdev = conf->mirrors[d].rdev;
1755                                if (sync_page_io(rdev, sect, s<<9,
1756                                                 bio->bi_io_vec[idx].bv_page,
1757                                                 READ, false)) {
1758                                        success = 1;
1759                                        break;
1760                                }
1761                        }
1762                        d++;
1763                        if (d == conf->raid_disks * 2)
1764                                d = 0;
1765                } while (!success && d != r1_bio->read_disk);
1766
1767                if (!success) {
1768                        char b[BDEVNAME_SIZE];
1769                        int abort = 0;
1770                        /* Cannot read from anywhere, this block is lost.
1771                         * Record a bad block on each device.  If that doesn't
1772                         * work just disable and interrupt the recovery.
1773                         * Don't fail devices as that won't really help.
1774                         */
1775                        printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1776                               " for block %llu\n",
1777                               mdname(mddev),
1778                               bdevname(bio->bi_bdev, b),
1779                               (unsigned long long)r1_bio->sector);
1780                        for (d = 0; d < conf->raid_disks * 2; d++) {
1781                                rdev = conf->mirrors[d].rdev;
1782                                if (!rdev || test_bit(Faulty, &rdev->flags))
1783                                        continue;
1784                                if (!rdev_set_badblocks(rdev, sect, s, 0))
1785                                        abort = 1;
1786                        }
1787                        if (abort) {
1788                                conf->recovery_disabled =
1789                                        mddev->recovery_disabled;
1790                                set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1791                                md_done_sync(mddev, r1_bio->sectors, 0);
1792                                put_buf(r1_bio);
1793                                return 0;
1794                        }
1795                        /* Try next page */
1796                        sectors -= s;
1797                        sect += s;
1798                        idx++;
1799                        continue;
1800                }
1801
1802                start = d;
1803                /* write it back and re-read */
1804                while (d != r1_bio->read_disk) {
1805                        if (d == 0)
1806                                d = conf->raid_disks * 2;
1807                        d--;
1808                        if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1809                                continue;
1810                        rdev = conf->mirrors[d].rdev;
1811                        if (r1_sync_page_io(rdev, sect, s,
1812                                            bio->bi_io_vec[idx].bv_page,
1813                                            WRITE) == 0) {
1814                                r1_bio->bios[d]->bi_end_io = NULL;
1815                                rdev_dec_pending(rdev, mddev);
1816                        }
1817                }
1818                d = start;
1819                while (d != r1_bio->read_disk) {
1820                        if (d == 0)
1821                                d = conf->raid_disks * 2;
1822                        d--;
1823                        if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1824                                continue;
1825                        rdev = conf->mirrors[d].rdev;
1826                        if (r1_sync_page_io(rdev, sect, s,
1827                                            bio->bi_io_vec[idx].bv_page,
1828                                            READ) != 0)
1829                                atomic_add(s, &rdev->corrected_errors);
1830                }
1831                sectors -= s;
1832                sect += s;
1833                idx ++;
1834        }
1835        set_bit(R1BIO_Uptodate, &r1_bio->state);
1836        set_bit(BIO_UPTODATE, &bio->bi_flags);
1837        return 1;
1838}
1839
1840static int process_checks(struct r1bio *r1_bio)
1841{
1842        /* We have read all readable devices.  If we haven't
1843         * got the block, then there is no hope left.
1844         * If we have, then we want to do a comparison
1845         * and skip the write if everything is the same.
1846         * If any blocks failed to read, then we need to
1847         * attempt an over-write
1848         */
1849        struct mddev *mddev = r1_bio->mddev;
1850        struct r1conf *conf = mddev->private;
1851        int primary;
1852        int i;
1853        int vcnt;
1854
1855        for (primary = 0; primary < conf->raid_disks * 2; primary++)
1856                if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1857                    test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1858                        r1_bio->bios[primary]->bi_end_io = NULL;
1859                        rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1860                        break;
1861                }
1862        r1_bio->read_disk = primary;
1863        vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1864        for (i = 0; i < conf->raid_disks * 2; i++) {
1865                int j;
1866                struct bio *pbio = r1_bio->bios[primary];
1867                struct bio *sbio = r1_bio->bios[i];
1868                int size;
1869
1870                if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1871                        continue;
1872
1873                if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1874                        for (j = vcnt; j-- ; ) {
1875                                struct page *p, *s;
1876                                p = pbio->bi_io_vec[j].bv_page;
1877                                s = sbio->bi_io_vec[j].bv_page;
1878                                if (memcmp(page_address(p),
1879                                           page_address(s),
1880                                           sbio->bi_io_vec[j].bv_len))
1881                                        break;
1882                        }
1883                } else
1884                        j = 0;
1885                if (j >= 0)
1886                        atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
1887                if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1888                              && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1889                        /* No need to write to this device. */
1890                        sbio->bi_end_io = NULL;
1891                        rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1892                        continue;
1893                }
1894                /* fixup the bio for reuse */
1895                sbio->bi_vcnt = vcnt;
1896                sbio->bi_size = r1_bio->sectors << 9;
1897                sbio->bi_idx = 0;
1898                sbio->bi_phys_segments = 0;
1899                sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1900                sbio->bi_flags |= 1 << BIO_UPTODATE;
1901                sbio->bi_next = NULL;
1902                sbio->bi_sector = r1_bio->sector +
1903                        conf->mirrors[i].rdev->data_offset;
1904                sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1905                size = sbio->bi_size;
1906                for (j = 0; j < vcnt ; j++) {
1907                        struct bio_vec *bi;
1908                        bi = &sbio->bi_io_vec[j];
1909                        bi->bv_offset = 0;
1910                        if (size > PAGE_SIZE)
1911                                bi->bv_len = PAGE_SIZE;
1912                        else
1913                                bi->bv_len = size;
1914                        size -= PAGE_SIZE;
1915                        memcpy(page_address(bi->bv_page),
1916                               page_address(pbio->bi_io_vec[j].bv_page),
1917                               PAGE_SIZE);
1918                }
1919        }
1920        return 0;
1921}
1922
1923static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1924{
1925        struct r1conf *conf = mddev->private;
1926        int i;
1927        int disks = conf->raid_disks * 2;
1928        struct bio *bio, *wbio;
1929
1930        bio = r1_bio->bios[r1_bio->read_disk];
1931
1932        if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1933                /* ouch - failed to read all of that. */
1934                if (!fix_sync_read_error(r1_bio))
1935                        return;
1936
1937        if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1938                if (process_checks(r1_bio) < 0)
1939                        return;
1940        /*
1941         * schedule writes
1942         */
1943        atomic_set(&r1_bio->remaining, 1);
1944        for (i = 0; i < disks ; i++) {
1945                wbio = r1_bio->bios[i];
1946                if (wbio->bi_end_io == NULL ||
1947                    (wbio->bi_end_io == end_sync_read &&
1948                     (i == r1_bio->read_disk ||
1949                      !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1950                        continue;
1951
1952                wbio->bi_rw = WRITE;
1953                wbio->bi_end_io = end_sync_write;
1954                atomic_inc(&r1_bio->remaining);
1955                md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1956
1957                generic_make_request(wbio);
1958        }
1959
1960        if (atomic_dec_and_test(&r1_bio->remaining)) {
1961                /* if we're here, all write(s) have completed, so clean up */
1962                int s = r1_bio->sectors;
1963                if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1964                    test_bit(R1BIO_WriteError, &r1_bio->state))
1965                        reschedule_retry(r1_bio);
1966                else {
1967                        put_buf(r1_bio);
1968                        md_done_sync(mddev, s, 1);
1969                }
1970        }
1971}
1972
1973/*
1974 * This is a kernel thread which:
1975 *
1976 *      1.      Retries failed read operations on working mirrors.
1977 *      2.      Updates the raid superblock when problems encounter.
1978 *      3.      Performs writes following reads for array synchronising.
1979 */
1980
1981static void fix_read_error(struct r1conf *conf, int read_disk,
1982                           sector_t sect, int sectors)
1983{
1984        struct mddev *mddev = conf->mddev;
1985        while(sectors) {
1986                int s = sectors;
1987                int d = read_disk;
1988                int success = 0;
1989                int start;
1990                struct md_rdev *rdev;
1991
1992                if (s > (PAGE_SIZE>>9))
1993                        s = PAGE_SIZE >> 9;
1994
1995                do {
1996                        /* Note: no rcu protection needed here
1997                         * as this is synchronous in the raid1d thread
1998                         * which is the thread that might remove
1999                         * a device.  If raid1d ever becomes multi-threaded....
2000                         */
2001                        sector_t first_bad;
2002                        int bad_sectors;
2003
2004                        rdev = conf->mirrors[d].rdev;
2005                        if (rdev &&
2006                            (test_bit(In_sync, &rdev->flags) ||
2007                             (!test_bit(Faulty, &rdev->flags) &&
2008                              rdev->recovery_offset >= sect + s)) &&
2009                            is_badblock(rdev, sect, s,
2010                                        &first_bad, &bad_sectors) == 0 &&
2011                            sync_page_io(rdev, sect, s<<9,
2012                                         conf->tmppage, READ, false))
2013                                success = 1;
2014                        else {
2015                                d++;
2016                                if (d == conf->raid_disks * 2)
2017                                        d = 0;
2018                        }
2019                } while (!success && d != read_disk);
2020
2021                if (!success) {
2022                        /* Cannot read from anywhere - mark it bad */
2023                        struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2024                        if (!rdev_set_badblocks(rdev, sect, s, 0))
2025                                md_error(mddev, rdev);
2026                        break;
2027                }
2028                /* write it back and re-read */
2029                start = d;
2030                while (d != read_disk) {
2031                        if (d==0)
2032                                d = conf->raid_disks * 2;
2033                        d--;
2034                        rdev = conf->mirrors[d].rdev;
2035                        if (rdev &&
2036                            test_bit(In_sync, &rdev->flags))
2037                                r1_sync_page_io(rdev, sect, s,
2038                                                conf->tmppage, WRITE);
2039                }
2040                d = start;
2041                while (d != read_disk) {
2042                        char b[BDEVNAME_SIZE];
2043                        if (d==0)
2044                                d = conf->raid_disks * 2;
2045                        d--;
2046                        rdev = conf->mirrors[d].rdev;
2047                        if (rdev &&
2048                            test_bit(In_sync, &rdev->flags)) {
2049                                if (r1_sync_page_io(rdev, sect, s,
2050                                                    conf->tmppage, READ)) {
2051                                        atomic_add(s, &rdev->corrected_errors);
2052                                        printk(KERN_INFO
2053                                               "md/raid1:%s: read error corrected "
2054                                               "(%d sectors at %llu on %s)\n",
2055                                               mdname(mddev), s,
2056                                               (unsigned long long)(sect +
2057                                                   rdev->data_offset),
2058                                               bdevname(rdev->bdev, b));
2059                                }
2060                        }
2061                }
2062                sectors -= s;
2063                sect += s;
2064        }
2065}
2066
2067static void bi_complete(struct bio *bio, int error)
2068{
2069        complete((struct completion *)bio->bi_private);
2070}
2071
2072static int submit_bio_wait(int rw, struct bio *bio)
2073{
2074        struct completion event;
2075        rw |= REQ_SYNC;
2076
2077        init_completion(&event);
2078        bio->bi_private = &event;
2079        bio->bi_end_io = bi_complete;
2080        submit_bio(rw, bio);
2081        wait_for_completion(&event);
2082
2083        return test_bit(BIO_UPTODATE, &bio->bi_flags);
2084}
2085
2086static int narrow_write_error(struct r1bio *r1_bio, int i)
2087{
2088        struct mddev *mddev = r1_bio->mddev;
2089        struct r1conf *conf = mddev->private;
2090        struct md_rdev *rdev = conf->mirrors[i].rdev;
2091        int vcnt, idx;
2092        struct bio_vec *vec;
2093
2094        /* bio has the data to be written to device 'i' where
2095         * we just recently had a write error.
2096         * We repeatedly clone the bio and trim down to one block,
2097         * then try the write.  Where the write fails we record
2098         * a bad block.
2099         * It is conceivable that the bio doesn't exactly align with
2100         * blocks.  We must handle this somehow.
2101         *
2102         * We currently own a reference on the rdev.
2103         */
2104
2105        int block_sectors;
2106        sector_t sector;
2107        int sectors;
2108        int sect_to_write = r1_bio->sectors;
2109        int ok = 1;
2110
2111        if (rdev->badblocks.shift < 0)
2112                return 0;
2113
2114        block_sectors = 1 << rdev->badblocks.shift;
2115        sector = r1_bio->sector;
2116        sectors = ((sector + block_sectors)
2117                   & ~(sector_t)(block_sectors - 1))
2118                - sector;
2119
2120        if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2121                vcnt = r1_bio->behind_page_count;
2122                vec = r1_bio->behind_bvecs;
2123                idx = 0;
2124                while (vec[idx].bv_page == NULL)
2125                        idx++;
2126        } else {
2127                vcnt = r1_bio->master_bio->bi_vcnt;
2128                vec = r1_bio->master_bio->bi_io_vec;
2129                idx = r1_bio->master_bio->bi_idx;
2130        }
2131        while (sect_to_write) {
2132                struct bio *wbio;
2133                if (sectors > sect_to_write)
2134                        sectors = sect_to_write;
2135                /* Write at 'sector' for 'sectors'*/
2136
2137                wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2138                memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2139                wbio->bi_sector = r1_bio->sector;
2140                wbio->bi_rw = WRITE;
2141                wbio->bi_vcnt = vcnt;
2142                wbio->bi_size = r1_bio->sectors << 9;
2143                wbio->bi_idx = idx;
2144
2145                md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2146                wbio->bi_sector += rdev->data_offset;
2147                wbio->bi_bdev = rdev->bdev;
2148                if (submit_bio_wait(WRITE, wbio) == 0)
2149                        /* failure! */
2150                        ok = rdev_set_badblocks(rdev, sector,
2151                                                sectors, 0)
2152                                && ok;
2153
2154                bio_put(wbio);
2155                sect_to_write -= sectors;
2156                sector += sectors;
2157                sectors = block_sectors;
2158        }
2159        return ok;
2160}
2161
2162static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2163{
2164        int m;
2165        int s = r1_bio->sectors;
2166        for (m = 0; m < conf->raid_disks * 2 ; m++) {
2167                struct md_rdev *rdev = conf->mirrors[m].rdev;
2168                struct bio *bio = r1_bio->bios[m];
2169                if (bio->bi_end_io == NULL)
2170                        continue;
2171                if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2172                    test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2173                        rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2174                }
2175                if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2176                    test_bit(R1BIO_WriteError, &r1_bio->state)) {
2177                        if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2178                                md_error(conf->mddev, rdev);
2179                }
2180        }
2181        put_buf(r1_bio);
2182        md_done_sync(conf->mddev, s, 1);
2183}
2184
2185static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2186{
2187        int m;
2188        for (m = 0; m < conf->raid_disks * 2 ; m++)
2189                if (r1_bio->bios[m] == IO_MADE_GOOD) {
2190                        struct md_rdev *rdev = conf->mirrors[m].rdev;
2191                        rdev_clear_badblocks(rdev,
2192                                             r1_bio->sector,
2193                                             r1_bio->sectors, 0);
2194                        rdev_dec_pending(rdev, conf->mddev);
2195                } else if (r1_bio->bios[m] != NULL) {
2196                        /* This drive got a write error.  We need to
2197                         * narrow down and record precise write
2198                         * errors.
2199                         */
2200                        if (!narrow_write_error(r1_bio, m)) {
2201                                md_error(conf->mddev,
2202                                         conf->mirrors[m].rdev);
2203                                /* an I/O failed, we can't clear the bitmap */
2204                                set_bit(R1BIO_Degraded, &r1_bio->state);
2205                        }
2206                        rdev_dec_pending(conf->mirrors[m].rdev,
2207                                         conf->mddev);
2208                }
2209        if (test_bit(R1BIO_WriteError, &r1_bio->state))
2210                close_write(r1_bio);
2211        raid_end_bio_io(r1_bio);
2212}
2213
2214static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2215{
2216        int disk;
2217        int max_sectors;
2218        struct mddev *mddev = conf->mddev;
2219        struct bio *bio;
2220        char b[BDEVNAME_SIZE];
2221        struct md_rdev *rdev;
2222
2223        clear_bit(R1BIO_ReadError, &r1_bio->state);
2224        /* we got a read error. Maybe the drive is bad.  Maybe just
2225         * the block and we can fix it.
2226         * We freeze all other IO, and try reading the block from
2227         * other devices.  When we find one, we re-write
2228         * and check it that fixes the read error.
2229         * This is all done synchronously while the array is
2230         * frozen
2231         */
2232        if (mddev->ro == 0) {
2233                freeze_array(conf);
2234                fix_read_error(conf, r1_bio->read_disk,
2235                               r1_bio->sector, r1_bio->sectors);
2236                unfreeze_array(conf);
2237        } else
2238                md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2239        rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2240
2241        bio = r1_bio->bios[r1_bio->read_disk];
2242        bdevname(bio->bi_bdev, b);
2243read_more:
2244        disk = read_balance(conf, r1_bio, &max_sectors);
2245        if (disk == -1) {
2246                printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2247                       " read error for block %llu\n",
2248                       mdname(mddev), b, (unsigned long long)r1_bio->sector);
2249                raid_end_bio_io(r1_bio);
2250        } else {
2251                const unsigned long do_sync
2252                        = r1_bio->master_bio->bi_rw & REQ_SYNC;
2253                if (bio) {
2254                        r1_bio->bios[r1_bio->read_disk] =
2255                                mddev->ro ? IO_BLOCKED : NULL;
2256                        bio_put(bio);
2257                }
2258                r1_bio->read_disk = disk;
2259                bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2260                md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2261                r1_bio->bios[r1_bio->read_disk] = bio;
2262                rdev = conf->mirrors[disk].rdev;
2263                printk_ratelimited(KERN_ERR
2264                                   "md/raid1:%s: redirecting sector %llu"
2265                                   " to other mirror: %s\n",
2266                                   mdname(mddev),
2267                                   (unsigned long long)r1_bio->sector,
2268                                   bdevname(rdev->bdev, b));
2269                bio->bi_sector = r1_bio->sector + rdev->data_offset;
2270                bio->bi_bdev = rdev->bdev;
2271                bio->bi_end_io = raid1_end_read_request;
2272                bio->bi_rw = READ | do_sync;
2273                bio->bi_private = r1_bio;
2274                if (max_sectors < r1_bio->sectors) {
2275                        /* Drat - have to split this up more */
2276                        struct bio *mbio = r1_bio->master_bio;
2277                        int sectors_handled = (r1_bio->sector + max_sectors
2278                                               - mbio->bi_sector);
2279                        r1_bio->sectors = max_sectors;
2280                        spin_lock_irq(&conf->device_lock);
2281                        if (mbio->bi_phys_segments == 0)
2282                                mbio->bi_phys_segments = 2;
2283                        else
2284                                mbio->bi_phys_segments++;
2285                        spin_unlock_irq(&conf->device_lock);
2286                        generic_make_request(bio);
2287                        bio = NULL;
2288
2289                        r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2290
2291                        r1_bio->master_bio = mbio;
2292                        r1_bio->sectors = (mbio->bi_size >> 9)
2293                                          - sectors_handled;
2294                        r1_bio->state = 0;
2295                        set_bit(R1BIO_ReadError, &r1_bio->state);
2296                        r1_bio->mddev = mddev;
2297                        r1_bio->sector = mbio->bi_sector + sectors_handled;
2298
2299                        goto read_more;
2300                } else
2301                        generic_make_request(bio);
2302        }
2303}
2304
2305static void raid1d(struct md_thread *thread)
2306{
2307        struct mddev *mddev = thread->mddev;
2308        struct r1bio *r1_bio;
2309        unsigned long flags;
2310        struct r1conf *conf = mddev->private;
2311        struct list_head *head = &conf->retry_list;
2312        struct blk_plug plug;
2313
2314        md_check_recovery(mddev);
2315
2316        blk_start_plug(&plug);
2317        for (;;) {
2318
2319                flush_pending_writes(conf);
2320
2321                spin_lock_irqsave(&conf->device_lock, flags);
2322                if (list_empty(head)) {
2323                        spin_unlock_irqrestore(&conf->device_lock, flags);
2324                        break;
2325                }
2326                r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2327                list_del(head->prev);
2328                conf->nr_queued--;
2329                spin_unlock_irqrestore(&conf->device_lock, flags);
2330
2331                mddev = r1_bio->mddev;
2332                conf = mddev->private;
2333                if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2334                        if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2335                            test_bit(R1BIO_WriteError, &r1_bio->state))
2336                                handle_sync_write_finished(conf, r1_bio);
2337                        else
2338                                sync_request_write(mddev, r1_bio);
2339                } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2340                           test_bit(R1BIO_WriteError, &r1_bio->state))
2341                        handle_write_finished(conf, r1_bio);
2342                else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2343                        handle_read_error(conf, r1_bio);
2344                else
2345                        /* just a partial read to be scheduled from separate
2346                         * context
2347                         */
2348                        generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2349
2350                cond_resched();
2351                if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2352                        md_check_recovery(mddev);
2353        }
2354        blk_finish_plug(&plug);
2355}
2356
2357
2358static int init_resync(struct r1conf *conf)
2359{
2360        int buffs;
2361
2362        buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2363        BUG_ON(conf->r1buf_pool);
2364        conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2365                                          conf->poolinfo);
2366        if (!conf->r1buf_pool)
2367                return -ENOMEM;
2368        conf->next_resync = 0;
2369        return 0;
2370}
2371
2372/*
2373 * perform a "sync" on one "block"
2374 *
2375 * We need to make sure that no normal I/O request - particularly write
2376 * requests - conflict with active sync requests.
2377 *
2378 * This is achieved by tracking pending requests and a 'barrier' concept
2379 * that can be installed to exclude normal IO requests.
2380 */
2381
2382static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2383{
2384        struct r1conf *conf = mddev->private;
2385        struct r1bio *r1_bio;
2386        struct bio *bio;
2387        sector_t max_sector, nr_sectors;
2388        int disk = -1;
2389        int i;
2390        int wonly = -1;
2391        int write_targets = 0, read_targets = 0;
2392        sector_t sync_blocks;
2393        int still_degraded = 0;
2394        int good_sectors = RESYNC_SECTORS;
2395        int min_bad = 0; /* number of sectors that are bad in all devices */
2396
2397        if (!conf->r1buf_pool)
2398                if (init_resync(conf))
2399                        return 0;
2400
2401        max_sector = mddev->dev_sectors;
2402        if (sector_nr >= max_sector) {
2403                /* If we aborted, we need to abort the
2404                 * sync on the 'current' bitmap chunk (there will
2405                 * only be one in raid1 resync.
2406                 * We can find the current addess in mddev->curr_resync
2407                 */
2408                if (mddev->curr_resync < max_sector) /* aborted */
2409                        bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2410                                                &sync_blocks, 1);
2411                else /* completed sync */
2412                        conf->fullsync = 0;
2413
2414                bitmap_close_sync(mddev->bitmap);
2415                close_sync(conf);
2416                return 0;
2417        }
2418
2419        if (mddev->bitmap == NULL &&
2420            mddev->recovery_cp == MaxSector &&
2421            !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2422            conf->fullsync == 0) {
2423                *skipped = 1;
2424                return max_sector - sector_nr;
2425        }
2426        /* before building a request, check if we can skip these blocks..
2427         * This call the bitmap_start_sync doesn't actually record anything
2428         */
2429        if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2430            !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2431                /* We can skip this block, and probably several more */
2432                *skipped = 1;
2433                return sync_blocks;
2434        }
2435        /*
2436         * If there is non-resync activity waiting for a turn,
2437         * and resync is going fast enough,
2438         * then let it though before starting on this new sync request.
2439         */
2440        if (!go_faster && conf->nr_waiting)
2441                msleep_interruptible(1000);
2442
2443        bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2444        r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2445        raise_barrier(conf);
2446
2447        conf->next_resync = sector_nr;
2448
2449        rcu_read_lock();
2450        /*
2451         * If we get a correctably read error during resync or recovery,
2452         * we might want to read from a different device.  So we
2453         * flag all drives that could conceivably be read from for READ,
2454         * and any others (which will be non-In_sync devices) for WRITE.
2455         * If a read fails, we try reading from something else for which READ
2456         * is OK.
2457         */
2458
2459        r1_bio->mddev = mddev;
2460        r1_bio->sector = sector_nr;
2461        r1_bio->state = 0;
2462        set_bit(R1BIO_IsSync, &r1_bio->state);
2463
2464        for (i = 0; i < conf->raid_disks * 2; i++) {
2465                struct md_rdev *rdev;
2466                bio = r1_bio->bios[i];
2467
2468                /* take from bio_init */
2469                bio->bi_next = NULL;
2470                bio->bi_flags &= ~(BIO_POOL_MASK-1);
2471                bio->bi_flags |= 1 << BIO_UPTODATE;
2472                bio->bi_rw = READ;
2473                bio->bi_vcnt = 0;
2474                bio->bi_idx = 0;
2475                bio->bi_phys_segments = 0;
2476                bio->bi_size = 0;
2477                bio->bi_end_io = NULL;
2478                bio->bi_private = NULL;
2479
2480                rdev = rcu_dereference(conf->mirrors[i].rdev);
2481                if (rdev == NULL ||
2482                    test_bit(Faulty, &rdev->flags)) {
2483                        if (i < conf->raid_disks)
2484                                still_degraded = 1;
2485                } else if (!test_bit(In_sync, &rdev->flags)) {
2486                        bio->bi_rw = WRITE;
2487                        bio->bi_end_io = end_sync_write;
2488                        write_targets ++;
2489                } else {
2490                        /* may need to read from here */
2491                        sector_t first_bad = MaxSector;
2492                        int bad_sectors;
2493
2494                        if (is_badblock(rdev, sector_nr, good_sectors,
2495                                        &first_bad, &bad_sectors)) {
2496                                if (first_bad > sector_nr)
2497                                        good_sectors = first_bad - sector_nr;
2498                                else {
2499                                        bad_sectors -= (sector_nr - first_bad);
2500                                        if (min_bad == 0 ||
2501                                            min_bad > bad_sectors)
2502                                                min_bad = bad_sectors;
2503                                }
2504                        }
2505                        if (sector_nr < first_bad) {
2506                                if (test_bit(WriteMostly, &rdev->flags)) {
2507                                        if (wonly < 0)
2508                                                wonly = i;
2509                                } else {
2510                                        if (disk < 0)
2511                                                disk = i;
2512                                }
2513                                bio->bi_rw = READ;
2514                                bio->bi_end_io = end_sync_read;
2515                                read_targets++;
2516                        } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2517                                test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2518                                !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2519                                /*
2520                                 * The device is suitable for reading (InSync),
2521                                 * but has bad block(s) here. Let's try to correct them,
2522                                 * if we are doing resync or repair. Otherwise, leave
2523                                 * this device alone for this sync request.
2524                                 */
2525                                bio->bi_rw = WRITE;
2526                                bio->bi_end_io = end_sync_write;
2527                                write_targets++;
2528                        }
2529                }
2530                if (bio->bi_end_io) {
2531                        atomic_inc(&rdev->nr_pending);
2532                        bio->bi_sector = sector_nr + rdev->data_offset;
2533                        bio->bi_bdev = rdev->bdev;
2534                        bio->bi_private = r1_bio;
2535                }
2536        }
2537        rcu_read_unlock();
2538        if (disk < 0)
2539                disk = wonly;
2540        r1_bio->read_disk = disk;
2541
2542        if (read_targets == 0 && min_bad > 0) {
2543                /* These sectors are bad on all InSync devices, so we
2544                 * need to mark them bad on all write targets
2545                 */
2546                int ok = 1;
2547                for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2548                        if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2549                                struct md_rdev *rdev = conf->mirrors[i].rdev;
2550                                ok = rdev_set_badblocks(rdev, sector_nr,
2551                                                        min_bad, 0
2552                                        ) && ok;
2553                        }
2554                set_bit(MD_CHANGE_DEVS, &mddev->flags);
2555                *skipped = 1;
2556                put_buf(r1_bio);
2557
2558                if (!ok) {
2559                        /* Cannot record the badblocks, so need to
2560                         * abort the resync.
2561                         * If there are multiple read targets, could just
2562                         * fail the really bad ones ???
2563                         */
2564                        conf->recovery_disabled = mddev->recovery_disabled;
2565                        set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2566                        return 0;
2567                } else
2568                        return min_bad;
2569
2570        }
2571        if (min_bad > 0 && min_bad < good_sectors) {
2572                /* only resync enough to reach the next bad->good
2573                 * transition */
2574                good_sectors = min_bad;
2575        }
2576
2577        if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2578                /* extra read targets are also write targets */
2579                write_targets += read_targets-1;
2580
2581        if (write_targets == 0 || read_targets == 0) {
2582                /* There is nowhere to write, so all non-sync
2583                 * drives must be failed - so we are finished
2584                 */
2585                sector_t rv;
2586                if (min_bad > 0)
2587                        max_sector = sector_nr + min_bad;
2588                rv = max_sector - sector_nr;
2589                *skipped = 1;
2590                put_buf(r1_bio);
2591                return rv;
2592        }
2593
2594        if (max_sector > mddev->resync_max)
2595                max_sector = mddev->resync_max; /* Don't do IO beyond here */
2596        if (max_sector > sector_nr + good_sectors)
2597                max_sector = sector_nr + good_sectors;
2598        nr_sectors = 0;
2599        sync_blocks = 0;
2600        do {
2601                struct page *page;
2602                int len = PAGE_SIZE;
2603                if (sector_nr + (len>>9) > max_sector)
2604                        len = (max_sector - sector_nr) << 9;
2605                if (len == 0)
2606                        break;
2607                if (sync_blocks == 0) {
2608                        if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2609                                               &sync_blocks, still_degraded) &&
2610                            !conf->fullsync &&
2611                            !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2612                                break;
2613                        BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2614                        if ((len >> 9) > sync_blocks)
2615                                len = sync_blocks<<9;
2616                }
2617
2618                for (i = 0 ; i < conf->raid_disks * 2; i++) {
2619                        bio = r1_bio->bios[i];
2620                        if (bio->bi_end_io) {
2621                                page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2622                                if (bio_add_page(bio, page, len, 0) == 0) {
2623                                        /* stop here */
2624                                        bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2625                                        while (i > 0) {
2626                                                i--;
2627                                                bio = r1_bio->bios[i];
2628                                                if (bio->bi_end_io==NULL)
2629                                                        continue;
2630                                                /* remove last page from this bio */
2631                                                bio->bi_vcnt--;
2632                                                bio->bi_size -= len;
2633                                                bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2634                                        }
2635                                        goto bio_full;
2636                                }
2637                        }
2638                }
2639                nr_sectors += len>>9;
2640                sector_nr += len>>9;
2641                sync_blocks -= (len>>9);
2642        } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2643 bio_full:
2644        r1_bio->sectors = nr_sectors;
2645
2646        /* For a user-requested sync, we read all readable devices and do a
2647         * compare
2648         */
2649        if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2650                atomic_set(&r1_bio->remaining, read_targets);
2651                for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2652                        bio = r1_bio->bios[i];
2653                        if (bio->bi_end_io == end_sync_read) {
2654                                read_targets--;
2655                                md_sync_acct(bio->bi_bdev, nr_sectors);
2656                                generic_make_request(bio);
2657                        }
2658                }
2659        } else {
2660                atomic_set(&r1_bio->remaining, 1);
2661                bio = r1_bio->bios[r1_bio->read_disk];
2662                md_sync_acct(bio->bi_bdev, nr_sectors);
2663                generic_make_request(bio);
2664
2665        }
2666        return nr_sectors;
2667}
2668
2669static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2670{
2671        if (sectors)
2672                return sectors;
2673
2674        return mddev->dev_sectors;
2675}
2676
2677static struct r1conf *setup_conf(struct mddev *mddev)
2678{
2679        struct r1conf *conf;
2680        int i;
2681        struct raid1_info *disk;
2682        struct md_rdev *rdev;
2683        int err = -ENOMEM;
2684
2685        conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2686        if (!conf)
2687                goto abort;
2688
2689        conf->mirrors = kzalloc(sizeof(struct raid1_info)
2690                                * mddev->raid_disks * 2,
2691                                 GFP_KERNEL);
2692        if (!conf->mirrors)
2693                goto abort;
2694
2695        conf->tmppage = alloc_page(GFP_KERNEL);
2696        if (!conf->tmppage)
2697                goto abort;
2698
2699        conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2700        if (!conf->poolinfo)
2701                goto abort;
2702        conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2703        conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2704                                          r1bio_pool_free,
2705                                          conf->poolinfo);
2706        if (!conf->r1bio_pool)
2707                goto abort;
2708
2709        conf->poolinfo->mddev = mddev;
2710
2711        err = -EINVAL;
2712        spin_lock_init(&conf->device_lock);
2713        rdev_for_each(rdev, mddev) {
2714                struct request_queue *q;
2715                int disk_idx = rdev->raid_disk;
2716                if (disk_idx >= mddev->raid_disks
2717                    || disk_idx < 0)
2718                        continue;
2719                if (test_bit(Replacement, &rdev->flags))
2720                        disk = conf->mirrors + mddev->raid_disks + disk_idx;
2721                else
2722                        disk = conf->mirrors + disk_idx;
2723
2724                if (disk->rdev)
2725                        goto abort;
2726                disk->rdev = rdev;
2727                q = bdev_get_queue(rdev->bdev);
2728                if (q->merge_bvec_fn)
2729                        mddev->merge_check_needed = 1;
2730
2731                disk->head_position = 0;
2732                disk->seq_start = MaxSector;
2733        }
2734        conf->raid_disks = mddev->raid_disks;
2735        conf->mddev = mddev;
2736        INIT_LIST_HEAD(&conf->retry_list);
2737
2738        spin_lock_init(&conf->resync_lock);
2739        init_waitqueue_head(&conf->wait_barrier);
2740
2741        bio_list_init(&conf->pending_bio_list);
2742        conf->pending_count = 0;
2743        conf->recovery_disabled = mddev->recovery_disabled - 1;
2744
2745        err = -EIO;
2746        for (i = 0; i < conf->raid_disks * 2; i++) {
2747
2748                disk = conf->mirrors + i;
2749
2750                if (i < conf->raid_disks &&
2751                    disk[conf->raid_disks].rdev) {
2752                        /* This slot has a replacement. */
2753                        if (!disk->rdev) {
2754                                /* No original, just make the replacement
2755                                 * a recovering spare
2756                                 */
2757                                disk->rdev =
2758                                        disk[conf->raid_disks].rdev;
2759                                disk[conf->raid_disks].rdev = NULL;
2760                        } else if (!test_bit(In_sync, &disk->rdev->flags))
2761                                /* Original is not in_sync - bad */
2762                                goto abort;
2763                }
2764
2765                if (!disk->rdev ||
2766                    !test_bit(In_sync, &disk->rdev->flags)) {
2767                        disk->head_position = 0;
2768                        if (disk->rdev &&
2769                            (disk->rdev->saved_raid_disk < 0))
2770                                conf->fullsync = 1;
2771                }
2772        }
2773
2774        err = -ENOMEM;
2775        conf->thread = md_register_thread(raid1d, mddev, "raid1");
2776        if (!conf->thread) {
2777                printk(KERN_ERR
2778                       "md/raid1:%s: couldn't allocate thread\n",
2779                       mdname(mddev));
2780                goto abort;
2781        }
2782
2783        return conf;
2784
2785 abort:
2786        if (conf) {
2787                if (conf->r1bio_pool)
2788                        mempool_destroy(conf->r1bio_pool);
2789                kfree(conf->mirrors);
2790                safe_put_page(conf->tmppage);
2791                kfree(conf->poolinfo);
2792                kfree(conf);
2793        }
2794        return ERR_PTR(err);
2795}
2796
2797static int stop(struct mddev *mddev);
2798static int run(struct mddev *mddev)
2799{
2800        struct r1conf *conf;
2801        int i;
2802        struct md_rdev *rdev;
2803        int ret;
2804        bool discard_supported = false;
2805
2806        if (mddev->level != 1) {
2807                printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2808                       mdname(mddev), mddev->level);
2809                return -EIO;
2810        }
2811        if (mddev->reshape_position != MaxSector) {
2812                printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2813                       mdname(mddev));
2814                return -EIO;
2815        }
2816        /*
2817         * copy the already verified devices into our private RAID1
2818         * bookkeeping area. [whatever we allocate in run(),
2819         * should be freed in stop()]
2820         */
2821        if (mddev->private == NULL)
2822                conf = setup_conf(mddev);
2823        else
2824                conf = mddev->private;
2825
2826        if (IS_ERR(conf))
2827                return PTR_ERR(conf);
2828
2829        if (mddev->queue)
2830                blk_queue_max_write_same_sectors(mddev->queue,
2831                                                 mddev->chunk_sectors);
2832        rdev_for_each(rdev, mddev) {
2833                if (!mddev->gendisk)
2834                        continue;
2835                disk_stack_limits(mddev->gendisk, rdev->bdev,
2836                                  rdev->data_offset << 9);
2837                if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2838                        discard_supported = true;
2839        }
2840
2841        mddev->degraded = 0;
2842        for (i=0; i < conf->raid_disks; i++)
2843                if (conf->mirrors[i].rdev == NULL ||
2844                    !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2845                    test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2846                        mddev->degraded++;
2847
2848        if (conf->raid_disks - mddev->degraded == 1)
2849                mddev->recovery_cp = MaxSector;
2850
2851        if (mddev->recovery_cp != MaxSector)
2852                printk(KERN_NOTICE "md/raid1:%s: not clean"
2853                       " -- starting background reconstruction\n",
2854                       mdname(mddev));
2855        printk(KERN_INFO 
2856                "md/raid1:%s: active with %d out of %d mirrors\n",
2857                mdname(mddev), mddev->raid_disks - mddev->degraded, 
2858                mddev->raid_disks);
2859
2860        /*
2861         * Ok, everything is just fine now
2862         */
2863        mddev->thread = conf->thread;
2864        conf->thread = NULL;
2865        mddev->private = conf;
2866
2867        md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2868
2869        if (mddev->queue) {
2870                mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2871                mddev->queue->backing_dev_info.congested_data = mddev;
2872                blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2873
2874                if (discard_supported)
2875                        queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2876                                                mddev->queue);
2877                else
2878                        queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2879                                                  mddev->queue);
2880        }
2881
2882        ret =  md_integrity_register(mddev);
2883        if (ret)
2884                stop(mddev);
2885        return ret;
2886}
2887
2888static int stop(struct mddev *mddev)
2889{
2890        struct r1conf *conf = mddev->private;
2891        struct bitmap *bitmap = mddev->bitmap;
2892
2893        /* wait for behind writes to complete */
2894        if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2895                printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2896                       mdname(mddev));
2897                /* need to kick something here to make sure I/O goes? */
2898                wait_event(bitmap->behind_wait,
2899                           atomic_read(&bitmap->behind_writes) == 0);
2900        }
2901
2902        raise_barrier(conf);
2903        lower_barrier(conf);
2904
2905        md_unregister_thread(&mddev->thread);
2906        if (conf->r1bio_pool)
2907                mempool_destroy(conf->r1bio_pool);
2908        kfree(conf->mirrors);
2909        kfree(conf->poolinfo);
2910        kfree(conf);
2911        mddev->private = NULL;
2912        return 0;
2913}
2914
2915static int raid1_resize(struct mddev *mddev, sector_t sectors)
2916{
2917        /* no resync is happening, and there is enough space
2918         * on all devices, so we can resize.
2919         * We need to make sure resync covers any new space.
2920         * If the array is shrinking we should possibly wait until
2921         * any io in the removed space completes, but it hardly seems
2922         * worth it.
2923         */
2924        sector_t newsize = raid1_size(mddev, sectors, 0);
2925        if (mddev->external_size &&
2926            mddev->array_sectors > newsize)
2927                return -EINVAL;
2928        if (mddev->bitmap) {
2929                int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2930                if (ret)
2931                        return ret;
2932        }
2933        md_set_array_sectors(mddev, newsize);
2934        set_capacity(mddev->gendisk, mddev->array_sectors);
2935        revalidate_disk(mddev->gendisk);
2936        if (sectors > mddev->dev_sectors &&
2937            mddev->recovery_cp > mddev->dev_sectors) {
2938                mddev->recovery_cp = mddev->dev_sectors;
2939                set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2940        }
2941        mddev->dev_sectors = sectors;
2942        mddev->resync_max_sectors = sectors;
2943        return 0;
2944}
2945
2946static int raid1_reshape(struct mddev *mddev)
2947{
2948        /* We need to:
2949         * 1/ resize the r1bio_pool
2950         * 2/ resize conf->mirrors
2951         *
2952         * We allocate a new r1bio_pool if we can.
2953         * Then raise a device barrier and wait until all IO stops.
2954         * Then resize conf->mirrors and swap in the new r1bio pool.
2955         *
2956         * At the same time, we "pack" the devices so that all the missing
2957         * devices have the higher raid_disk numbers.
2958         */
2959        mempool_t *newpool, *oldpool;
2960        struct pool_info *newpoolinfo;
2961        struct raid1_info *newmirrors;
2962        struct r1conf *conf = mddev->private;
2963        int cnt, raid_disks;
2964        unsigned long flags;
2965        int d, d2, err;
2966
2967        /* Cannot change chunk_size, layout, or level */
2968        if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2969            mddev->layout != mddev->new_layout ||
2970            mddev->level != mddev->new_level) {
2971                mddev->new_chunk_sectors = mddev->chunk_sectors;
2972                mddev->new_layout = mddev->layout;
2973                mddev->new_level = mddev->level;
2974                return -EINVAL;
2975        }
2976
2977        err = md_allow_write(mddev);
2978        if (err)
2979                return err;
2980
2981        raid_disks = mddev->raid_disks + mddev->delta_disks;
2982
2983        if (raid_disks < conf->raid_disks) {
2984                cnt=0;
2985                for (d= 0; d < conf->raid_disks; d++)
2986                        if (conf->mirrors[d].rdev)
2987                                cnt++;
2988                if (cnt > raid_disks)
2989                        return -EBUSY;
2990        }
2991
2992        newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2993        if (!newpoolinfo)
2994                return -ENOMEM;
2995        newpoolinfo->mddev = mddev;
2996        newpoolinfo->raid_disks = raid_disks * 2;
2997
2998        newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2999                                 r1bio_pool_free, newpoolinfo);
3000        if (!newpool) {
3001                kfree(newpoolinfo);
3002                return -ENOMEM;
3003        }
3004        newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3005                             GFP_KERNEL);
3006        if (!newmirrors) {
3007                kfree(newpoolinfo);
3008                mempool_destroy(newpool);
3009                return -ENOMEM;
3010        }
3011
3012        raise_barrier(conf);
3013
3014        /* ok, everything is stopped */
3015        oldpool = conf->r1bio_pool;
3016        conf->r1bio_pool = newpool;
3017
3018        for (d = d2 = 0; d < conf->raid_disks; d++) {
3019                struct md_rdev *rdev = conf->mirrors[d].rdev;
3020                if (rdev && rdev->raid_disk != d2) {
3021                        sysfs_unlink_rdev(mddev, rdev);
3022                        rdev->raid_disk = d2;
3023                        sysfs_unlink_rdev(mddev, rdev);
3024                        if (sysfs_link_rdev(mddev, rdev))
3025                                printk(KERN_WARNING
3026                                       "md/raid1:%s: cannot register rd%d\n",
3027                                       mdname(mddev), rdev->raid_disk);
3028                }
3029                if (rdev)
3030                        newmirrors[d2++].rdev = rdev;
3031        }
3032        kfree(conf->mirrors);
3033        conf->mirrors = newmirrors;
3034        kfree(conf->poolinfo);
3035        conf->poolinfo = newpoolinfo;
3036
3037        spin_lock_irqsave(&conf->device_lock, flags);
3038        mddev->degraded += (raid_disks - conf->raid_disks);
3039        spin_unlock_irqrestore(&conf->device_lock, flags);
3040        conf->raid_disks = mddev->raid_disks = raid_disks;
3041        mddev->delta_disks = 0;
3042
3043        lower_barrier(conf);
3044
3045        set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3046        md_wakeup_thread(mddev->thread);
3047
3048        mempool_destroy(oldpool);
3049        return 0;
3050}
3051
3052static void raid1_quiesce(struct mddev *mddev, int state)
3053{
3054        struct r1conf *conf = mddev->private;
3055
3056        switch(state) {
3057        case 2: /* wake for suspend */
3058                wake_up(&conf->wait_barrier);
3059                break;
3060        case 1:
3061                raise_barrier(conf);
3062                break;
3063        case 0:
3064                lower_barrier(conf);
3065                break;
3066        }
3067}
3068
3069static void *raid1_takeover(struct mddev *mddev)
3070{
3071        /* raid1 can take over:
3072         *  raid5 with 2 devices, any layout or chunk size
3073         */
3074        if (mddev->level == 5 && mddev->raid_disks == 2) {
3075                struct r1conf *conf;
3076                mddev->new_level = 1;
3077                mddev->new_layout = 0;
3078                mddev->new_chunk_sectors = 0;
3079                conf = setup_conf(mddev);
3080                if (!IS_ERR(conf))
3081                        conf->barrier = 1;
3082                return conf;
3083        }
3084        return ERR_PTR(-EINVAL);
3085}
3086
3087static struct md_personality raid1_personality =
3088{
3089        .name           = "raid1",
3090        .level          = 1,
3091        .owner          = THIS_MODULE,
3092        .make_request   = make_request,
3093        .run            = run,
3094        .stop           = stop,
3095        .status         = status,
3096        .error_handler  = error,
3097        .hot_add_disk   = raid1_add_disk,
3098        .hot_remove_disk= raid1_remove_disk,
3099        .spare_active   = raid1_spare_active,
3100        .sync_request   = sync_request,
3101        .resize         = raid1_resize,
3102        .size           = raid1_size,
3103        .check_reshape  = raid1_reshape,
3104        .quiesce        = raid1_quiesce,
3105        .takeover       = raid1_takeover,
3106};
3107
3108static int __init raid_init(void)
3109{
3110        return register_md_personality(&raid1_personality);
3111}
3112
3113static void raid_exit(void)
3114{
3115        unregister_md_personality(&raid1_personality);
3116}
3117
3118module_init(raid_init);
3119module_exit(raid_exit);
3120MODULE_LICENSE("GPL");
3121MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3122MODULE_ALIAS("md-personality-3"); /* RAID1 */
3123MODULE_ALIAS("md-raid1");
3124MODULE_ALIAS("md-level-1");
3125
3126module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
3127
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.