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                        );
 869                conf->nr_waiting--;
 870        }
 871        conf->nr_pending++;
 872        spin_unlock_irq(&conf->resync_lock);
 873}
 874
 875static void allow_barrier(struct r1conf *conf)
 876{
 877        unsigned long flags;
 878        spin_lock_irqsave(&conf->resync_lock, flags);
 879        conf->nr_pending--;
 880        spin_unlock_irqrestore(&conf->resync_lock, flags);
 881        wake_up(&conf->wait_barrier);
 882}
 883
 884static void freeze_array(struct r1conf *conf)
 885{
 886        /* stop syncio and normal IO and wait for everything to
 887         * go quite.
 888         * We increment barrier and nr_waiting, and then
 889         * wait until nr_pending match nr_queued+1
 890         * This is called in the context of one normal IO request
 891         * that has failed. Thus any sync request that might be pending
 892         * will be blocked by nr_pending, and we need to wait for
 893         * pending IO requests to complete or be queued for re-try.
 894         * Thus the number queued (nr_queued) plus this request (1)
 895         * must match the number of pending IOs (nr_pending) before
 896         * we continue.
 897         */
 898        spin_lock_irq(&conf->resync_lock);
 899        conf->barrier++;
 900        conf->nr_waiting++;
 901        wait_event_lock_irq(conf->wait_barrier,
 902                            conf->nr_pending == conf->nr_queued+1,
 903                            conf->resync_lock,
 904                            flush_pending_writes(conf));
 905        spin_unlock_irq(&conf->resync_lock);
 906}
 907static void unfreeze_array(struct r1conf *conf)
 908{
 909        /* reverse the effect of the freeze */
 910        spin_lock_irq(&conf->resync_lock);
 911        conf->barrier--;
 912        conf->nr_waiting--;
 913        wake_up(&conf->wait_barrier);
 914        spin_unlock_irq(&conf->resync_lock);
 915}
 916
 917
 918/* duplicate the data pages for behind I/O 
 919 */
 920static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
 921{
 922        int i;
 923        struct bio_vec *bvec;
 924        struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
 925                                        GFP_NOIO);
 926        if (unlikely(!bvecs))
 927                return;
 928
 929        bio_for_each_segment(bvec, bio, i) {
 930                bvecs[i] = *bvec;
 931                bvecs[i].bv_page = alloc_page(GFP_NOIO);
 932                if (unlikely(!bvecs[i].bv_page))
 933                        goto do_sync_io;
 934                memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
 935                       kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 936                kunmap(bvecs[i].bv_page);
 937                kunmap(bvec->bv_page);
 938        }
 939        r1_bio->behind_bvecs = bvecs;
 940        r1_bio->behind_page_count = bio->bi_vcnt;
 941        set_bit(R1BIO_BehindIO, &r1_bio->state);
 942        return;
 943
 944do_sync_io:
 945        for (i = 0; i < bio->bi_vcnt; i++)
 946                if (bvecs[i].bv_page)
 947                        put_page(bvecs[i].bv_page);
 948        kfree(bvecs);
 949        pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 950}
 951
 952struct raid1_plug_cb {
 953        struct blk_plug_cb      cb;
 954        struct bio_list         pending;
 955        int                     pending_cnt;
 956};
 957
 958static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
 959{
 960        struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
 961                                                  cb);
 962        struct mddev *mddev = plug->cb.data;
 963        struct r1conf *conf = mddev->private;
 964        struct bio *bio;
 965
 966        if (from_schedule || current->bio_list) {
 967                spin_lock_irq(&conf->device_lock);
 968                bio_list_merge(&conf->pending_bio_list, &plug->pending);
 969                conf->pending_count += plug->pending_cnt;
 970                spin_unlock_irq(&conf->device_lock);
 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                generic_make_request(bio);
 985                bio = next;
 986        }
 987        kfree(plug);
 988}
 989
 990static void make_request(struct mddev *mddev, struct bio * bio)
 991{
 992        struct r1conf *conf = mddev->private;
 993        struct raid1_info *mirror;
 994        struct r1bio *r1_bio;
 995        struct bio *read_bio;
 996        int i, disks;
 997        struct bitmap *bitmap;
 998        unsigned long flags;
 999        const int rw = bio_data_dir(bio);
1000        const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1001        const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1002        const unsigned long do_discard = (bio->bi_rw
1003                                          & (REQ_DISCARD | REQ_SECURE));
1004        struct md_rdev *blocked_rdev;
1005        struct blk_plug_cb *cb;
1006        struct raid1_plug_cb *plug = NULL;
1007        int first_clone;
1008        int sectors_handled;
1009        int max_sectors;
1010
1011        /*
1012         * Register the new request and wait if the reconstruction
1013         * thread has put up a bar for new requests.
1014         * Continue immediately if no resync is active currently.
1015         */
1016
1017        md_write_start(mddev, bio); /* wait on superblock update early */
1018
1019        if (bio_data_dir(bio) == WRITE &&
1020            bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
1021            bio->bi_sector < mddev->suspend_hi) {
1022                /* As the suspend_* range is controlled by
1023                 * userspace, we want an interruptible
1024                 * wait.
1025                 */
1026                DEFINE_WAIT(w);
1027                for (;;) {
1028                        flush_signals(current);
1029                        prepare_to_wait(&conf->wait_barrier,
1030                                        &w, TASK_INTERRUPTIBLE);
1031                        if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
1032                            bio->bi_sector >= mddev->suspend_hi)
1033                                break;
1034                        schedule();
1035                }
1036                finish_wait(&conf->wait_barrier, &w);
1037        }
1038
1039        wait_barrier(conf);
1040
1041        bitmap = mddev->bitmap;
1042
1043        /*
1044         * make_request() can abort the operation when READA is being
1045         * used and no empty request is available.
1046         *
1047         */
1048        r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1049
1050        r1_bio->master_bio = bio;
1051        r1_bio->sectors = bio->bi_size >> 9;
1052        r1_bio->state = 0;
1053        r1_bio->mddev = mddev;
1054        r1_bio->sector = bio->bi_sector;
1055
1056        /* We might need to issue multiple reads to different
1057         * devices if there are bad blocks around, so we keep
1058         * track of the number of reads in bio->bi_phys_segments.
1059         * If this is 0, there is only one r1_bio and no locking
1060         * will be needed when requests complete.  If it is
1061         * non-zero, then it is the number of not-completed requests.
1062         */
1063        bio->bi_phys_segments = 0;
1064        clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1065
1066        if (rw == READ) {
1067                /*
1068                 * read balancing logic:
1069                 */
1070                int rdisk;
1071
1072read_again:
1073                rdisk = read_balance(conf, r1_bio, &max_sectors);
1074
1075                if (rdisk < 0) {
1076                        /* couldn't find anywhere to read from */
1077                        raid_end_bio_io(r1_bio);
1078                        return;
1079                }
1080                mirror = conf->mirrors + rdisk;
1081
1082                if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1083                    bitmap) {
1084                        /* Reading from a write-mostly device must
1085                         * take care not to over-take any writes
1086                         * that are 'behind'
1087                         */
1088                        wait_event(bitmap->behind_wait,
1089                                   atomic_read(&bitmap->behind_writes) == 0);
1090                }
1091                r1_bio->read_disk = rdisk;
1092
1093                read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1094                md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
1095                            max_sectors);
1096
1097                r1_bio->bios[rdisk] = read_bio;
1098
1099                read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
1100                read_bio->bi_bdev = mirror->rdev->bdev;
1101                read_bio->bi_end_io = raid1_end_read_request;
1102                read_bio->bi_rw = READ | do_sync;
1103                read_bio->bi_private = r1_bio;
1104
1105                if (max_sectors < r1_bio->sectors) {
1106                        /* could not read all from this device, so we will
1107                         * need another r1_bio.
1108                         */
1109
1110                        sectors_handled = (r1_bio->sector + max_sectors
1111                                           - bio->bi_sector);
1112                        r1_bio->sectors = max_sectors;
1113                        spin_lock_irq(&conf->device_lock);
1114                        if (bio->bi_phys_segments == 0)
1115                                bio->bi_phys_segments = 2;
1116                        else
1117                                bio->bi_phys_segments++;
1118                        spin_unlock_irq(&conf->device_lock);
1119                        /* Cannot call generic_make_request directly
1120                         * as that will be queued in __make_request
1121                         * and subsequent mempool_alloc might block waiting
1122                         * for it.  So hand bio over to raid1d.
1123                         */
1124                        reschedule_retry(r1_bio);
1125
1126                        r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1127
1128                        r1_bio->master_bio = bio;
1129                        r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1130                        r1_bio->state = 0;
1131                        r1_bio->mddev = mddev;
1132                        r1_bio->sector = bio->bi_sector + sectors_handled;
1133                        goto read_again;
1134                } else
1135                        generic_make_request(read_bio);
1136                return;
1137        }
1138
1139        /*
1140         * WRITE:
1141         */
1142        if (conf->pending_count >= max_queued_requests) {
1143                md_wakeup_thread(mddev->thread);
1144                wait_event(conf->wait_barrier,
1145                           conf->pending_count < max_queued_requests);
1146        }
1147        /* first select target devices under rcu_lock and
1148         * inc refcount on their rdev.  Record them by setting
1149         * bios[x] to bio
1150         * If there are known/acknowledged bad blocks on any device on
1151         * which we have seen a write error, we want to avoid writing those
1152         * blocks.
1153         * This potentially requires several writes to write around
1154         * the bad blocks.  Each set of writes gets it's own r1bio
1155         * with a set of bios attached.
1156         */
1157
1158        disks = conf->raid_disks * 2;
1159 retry_write:
1160        blocked_rdev = NULL;
1161        rcu_read_lock();
1162        max_sectors = r1_bio->sectors;
1163        for (i = 0;  i < disks; i++) {
1164                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1165                if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1166                        atomic_inc(&rdev->nr_pending);
1167                        blocked_rdev = rdev;
1168                        break;
1169                }
1170                r1_bio->bios[i] = NULL;
1171                if (!rdev || test_bit(Faulty, &rdev->flags)
1172                    || test_bit(Unmerged, &rdev->flags)) {
1173                        if (i < conf->raid_disks)
1174                                set_bit(R1BIO_Degraded, &r1_bio->state);
1175                        continue;
1176                }
1177
1178                atomic_inc(&rdev->nr_pending);
1179                if (test_bit(WriteErrorSeen, &rdev->flags)) {
1180                        sector_t first_bad;
1181                        int bad_sectors;
1182                        int is_bad;
1183
1184                        is_bad = is_badblock(rdev, r1_bio->sector,
1185                                             max_sectors,
1186                                             &first_bad, &bad_sectors);
1187                        if (is_bad < 0) {
1188                                /* mustn't write here until the bad block is
1189                                 * acknowledged*/
1190                                set_bit(BlockedBadBlocks, &rdev->flags);
1191                                blocked_rdev = rdev;
1192                                break;
1193                        }
1194                        if (is_bad && first_bad <= r1_bio->sector) {
1195                                /* Cannot write here at all */
1196                                bad_sectors -= (r1_bio->sector - first_bad);
1197                                if (bad_sectors < max_sectors)
1198                                        /* mustn't write more than bad_sectors
1199                                         * to other devices yet
1200                                         */
1201                                        max_sectors = bad_sectors;
1202                                rdev_dec_pending(rdev, mddev);
1203                                /* We don't set R1BIO_Degraded as that
1204                                 * only applies if the disk is
1205                                 * missing, so it might be re-added,
1206                                 * and we want to know to recover this
1207                                 * chunk.
1208                                 * In this case the device is here,
1209                                 * and the fact that this chunk is not
1210                                 * in-sync is recorded in the bad
1211                                 * block log
1212                                 */
1213                                continue;
1214                        }
1215                        if (is_bad) {
1216                                int good_sectors = first_bad - r1_bio->sector;
1217                                if (good_sectors < max_sectors)
1218                                        max_sectors = good_sectors;
1219                        }
1220                }
1221                r1_bio->bios[i] = bio;
1222        }
1223        rcu_read_unlock();
1224
1225        if (unlikely(blocked_rdev)) {
1226                /* Wait for this device to become unblocked */
1227                int j;
1228
1229                for (j = 0; j < i; j++)
1230                        if (r1_bio->bios[j])
1231                                rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1232                r1_bio->state = 0;
1233                allow_barrier(conf);
1234                md_wait_for_blocked_rdev(blocked_rdev, mddev);
1235                wait_barrier(conf);
1236                goto retry_write;
1237        }
1238
1239        if (max_sectors < r1_bio->sectors) {
1240                /* We are splitting this write into multiple parts, so
1241                 * we need to prepare for allocating another r1_bio.
1242                 */
1243                r1_bio->sectors = max_sectors;
1244                spin_lock_irq(&conf->device_lock);
1245                if (bio->bi_phys_segments == 0)
1246                        bio->bi_phys_segments = 2;
1247                else
1248                        bio->bi_phys_segments++;
1249                spin_unlock_irq(&conf->device_lock);
1250        }
1251        sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1252
1253        atomic_set(&r1_bio->remaining, 1);
1254        atomic_set(&r1_bio->behind_remaining, 0);
1255
1256        first_clone = 1;
1257        for (i = 0; i < disks; i++) {
1258                struct bio *mbio;
1259                if (!r1_bio->bios[i])
1260                        continue;
1261
1262                mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1263                md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1264
1265                if (first_clone) {
1266                        /* do behind I/O ?
1267                         * Not if there are too many, or cannot
1268                         * allocate memory, or a reader on WriteMostly
1269                         * is waiting for behind writes to flush */
1270                        if (bitmap &&
1271                            (atomic_read(&bitmap->behind_writes)
1272                             < mddev->bitmap_info.max_write_behind) &&
1273                            !waitqueue_active(&bitmap->behind_wait))
1274                                alloc_behind_pages(mbio, r1_bio);
1275
1276                        bitmap_startwrite(bitmap, r1_bio->sector,
1277                                          r1_bio->sectors,
1278                                          test_bit(R1BIO_BehindIO,
1279                                                   &r1_bio->state));
1280                        first_clone = 0;
1281                }
1282                if (r1_bio->behind_bvecs) {
1283                        struct bio_vec *bvec;
1284                        int j;
1285
1286                        /* Yes, I really want the '__' version so that
1287                         * we clear any unused pointer in the io_vec, rather
1288                         * than leave them unchanged.  This is important
1289                         * because when we come to free the pages, we won't
1290                         * know the original bi_idx, so we just free
1291                         * them all
1292                         */
1293                        __bio_for_each_segment(bvec, mbio, j, 0)
1294                                bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1295                        if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1296                                atomic_inc(&r1_bio->behind_remaining);
1297                }
1298
1299                r1_bio->bios[i] = mbio;
1300
1301                mbio->bi_sector = (r1_bio->sector +
1302                                   conf->mirrors[i].rdev->data_offset);
1303                mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1304                mbio->bi_end_io = raid1_end_write_request;
1305                mbio->bi_rw = WRITE | do_flush_fua | do_sync | do_discard;
1306                mbio->bi_private = r1_bio;
1307
1308                atomic_inc(&r1_bio->remaining);
1309
1310                cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1311                if (cb)
1312                        plug = container_of(cb, struct raid1_plug_cb, cb);
1313                else
1314                        plug = NULL;
1315                spin_lock_irqsave(&conf->device_lock, flags);
1316                if (plug) {
1317                        bio_list_add(&plug->pending, mbio);
1318                        plug->pending_cnt++;
1319                } else {
1320                        bio_list_add(&conf->pending_bio_list, mbio);
1321                        conf->pending_count++;
1322                }
1323                spin_unlock_irqrestore(&conf->device_lock, flags);
1324                if (!plug)
1325                        md_wakeup_thread(mddev->thread);
1326        }
1327        /* Mustn't call r1_bio_write_done before this next test,
1328         * as it could result in the bio being freed.
1329         */
1330        if (sectors_handled < (bio->bi_size >> 9)) {
1331                r1_bio_write_done(r1_bio);
1332                /* We need another r1_bio.  It has already been counted
1333                 * in bio->bi_phys_segments
1334                 */
1335                r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1336                r1_bio->master_bio = bio;
1337                r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1338                r1_bio->state = 0;
1339                r1_bio->mddev = mddev;
1340                r1_bio->sector = bio->bi_sector + sectors_handled;
1341                goto retry_write;
1342        }
1343
1344        r1_bio_write_done(r1_bio);
1345
1346        /* In case raid1d snuck in to freeze_array */
1347        wake_up(&conf->wait_barrier);
1348}
1349
1350static void status(struct seq_file *seq, struct mddev *mddev)
1351{
1352        struct r1conf *conf = mddev->private;
1353        int i;
1354
1355        seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1356                   conf->raid_disks - mddev->degraded);
1357        rcu_read_lock();
1358        for (i = 0; i < conf->raid_disks; i++) {
1359                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1360                seq_printf(seq, "%s",
1361                           rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1362        }
1363        rcu_read_unlock();
1364        seq_printf(seq, "]");
1365}
1366
1367
1368static void error(struct mddev *mddev, struct md_rdev *rdev)
1369{
1370        char b[BDEVNAME_SIZE];
1371        struct r1conf *conf = mddev->private;
1372
1373        /*
1374         * If it is not operational, then we have already marked it as dead
1375         * else if it is the last working disks, ignore the error, let the
1376         * next level up know.
1377         * else mark the drive as failed
1378         */
1379        if (test_bit(In_sync, &rdev->flags)
1380            && (conf->raid_disks - mddev->degraded) == 1) {
1381                /*
1382                 * Don't fail the drive, act as though we were just a
1383                 * normal single drive.
1384                 * However don't try a recovery from this drive as
1385                 * it is very likely to fail.
1386                 */
1387                conf->recovery_disabled = mddev->recovery_disabled;
1388                return;
1389        }
1390        set_bit(Blocked, &rdev->flags);
1391        if (test_and_clear_bit(In_sync, &rdev->flags)) {
1392                unsigned long flags;
1393                spin_lock_irqsave(&conf->device_lock, flags);
1394                mddev->degraded++;
1395                set_bit(Faulty, &rdev->flags);
1396                spin_unlock_irqrestore(&conf->device_lock, flags);
1397                /*
1398                 * if recovery is running, make sure it aborts.
1399                 */
1400                set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1401        } else
1402                set_bit(Faulty, &rdev->flags);
1403        set_bit(MD_CHANGE_DEVS, &mddev->flags);
1404        printk(KERN_ALERT
1405               "md/raid1:%s: Disk failure on %s, disabling device.\n"
1406               "md/raid1:%s: Operation continuing on %d devices.\n",
1407               mdname(mddev), bdevname(rdev->bdev, b),
1408               mdname(mddev), conf->raid_disks - mddev->degraded);
1409}
1410
1411static void print_conf(struct r1conf *conf)
1412{
1413        int i;
1414
1415        printk(KERN_DEBUG "RAID1 conf printout:\n");
1416        if (!conf) {
1417                printk(KERN_DEBUG "(!conf)\n");
1418                return;
1419        }
1420        printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1421                conf->raid_disks);
1422
1423        rcu_read_lock();
1424        for (i = 0; i < conf->raid_disks; i++) {
1425                char b[BDEVNAME_SIZE];
1426                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1427                if (rdev)
1428                        printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1429                               i, !test_bit(In_sync, &rdev->flags),
1430                               !test_bit(Faulty, &rdev->flags),
1431                               bdevname(rdev->bdev,b));
1432        }
1433        rcu_read_unlock();
1434}
1435
1436static void close_sync(struct r1conf *conf)
1437{
1438        wait_barrier(conf);
1439        allow_barrier(conf);
1440
1441        mempool_destroy(conf->r1buf_pool);
1442        conf->r1buf_pool = NULL;
1443}
1444
1445static int raid1_spare_active(struct mddev *mddev)
1446{
1447        int i;
1448        struct r1conf *conf = mddev->private;
1449        int count = 0;
1450        unsigned long flags;
1451
1452        /*
1453         * Find all failed disks within the RAID1 configuration 
1454         * and mark them readable.
1455         * Called under mddev lock, so rcu protection not needed.
1456         */
1457        for (i = 0; i < conf->raid_disks; i++) {
1458                struct md_rdev *rdev = conf->mirrors[i].rdev;
1459                struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1460                if (repl
1461                    && repl->recovery_offset == MaxSector
1462                    && !test_bit(Faulty, &repl->flags)
1463                    && !test_and_set_bit(In_sync, &repl->flags)) {
1464                        /* replacement has just become active */
1465                        if (!rdev ||
1466                            !test_and_clear_bit(In_sync, &rdev->flags))
1467                                count++;
1468                        if (rdev) {
1469                                /* Replaced device not technically
1470                                 * faulty, but we need to be sure
1471                                 * it gets removed and never re-added
1472                                 */
1473                                set_bit(Faulty, &rdev->flags);
1474                                sysfs_notify_dirent_safe(
1475                                        rdev->sysfs_state);
1476                        }
1477                }
1478                if (rdev
1479                    && !test_bit(Faulty, &rdev->flags)
1480                    && !test_and_set_bit(In_sync, &rdev->flags)) {
1481                        count++;
1482                        sysfs_notify_dirent_safe(rdev->sysfs_state);
1483                }
1484        }
1485        spin_lock_irqsave(&conf->device_lock, flags);
1486        mddev->degraded -= count;
1487        spin_unlock_irqrestore(&conf->device_lock, flags);
1488
1489        print_conf(conf);
1490        return count;
1491}
1492
1493
1494static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1495{
1496        struct r1conf *conf = mddev->private;
1497        int err = -EEXIST;
1498        int mirror = 0;
1499        struct raid1_info *p;
1500        int first = 0;
1501        int last = conf->raid_disks - 1;
1502        struct request_queue *q = bdev_get_queue(rdev->bdev);
1503
1504        if (mddev->recovery_disabled == conf->recovery_disabled)
1505                return -EBUSY;
1506
1507        if (rdev->raid_disk >= 0)
1508                first = last = rdev->raid_disk;
1509
1510        if (q->merge_bvec_fn) {
1511                set_bit(Unmerged, &rdev->flags);
1512                mddev->merge_check_needed = 1;
1513        }
1514
1515        for (mirror = first; mirror <= last; mirror++) {
1516                p = conf->mirrors+mirror;
1517                if (!p->rdev) {
1518
1519                        disk_stack_limits(mddev->gendisk, rdev->bdev,
1520                                          rdev->data_offset << 9);
1521
1522                        p->head_position = 0;
1523                        rdev->raid_disk = mirror;
1524                        err = 0;
1525                        /* As all devices are equivalent, we don't need a full recovery
1526                         * if this was recently any drive of the array
1527                         */
1528                        if (rdev->saved_raid_disk < 0)
1529                                conf->fullsync = 1;
1530                        rcu_assign_pointer(p->rdev, rdev);
1531                        break;
1532                }
1533                if (test_bit(WantReplacement, &p->rdev->flags) &&
1534                    p[conf->raid_disks].rdev == NULL) {
1535                        /* Add this device as a replacement */
1536                        clear_bit(In_sync, &rdev->flags);
1537                        set_bit(Replacement, &rdev->flags);
1538                        rdev->raid_disk = mirror;
1539                        err = 0;
1540                        conf->fullsync = 1;
1541                        rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1542                        break;
1543                }
1544        }
1545        if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1546                /* Some requests might not have seen this new
1547                 * merge_bvec_fn.  We must wait for them to complete
1548                 * before merging the device fully.
1549                 * First we make sure any code which has tested
1550                 * our function has submitted the request, then
1551                 * we wait for all outstanding requests to complete.
1552                 */
1553                synchronize_sched();
1554                raise_barrier(conf);
1555                lower_barrier(conf);
1556                clear_bit(Unmerged, &rdev->flags);
1557        }
1558        md_integrity_add_rdev(rdev, mddev);
1559        if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
1560                queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1561        print_conf(conf);
1562        return err;
1563}
1564
1565static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1566{
1567        struct r1conf *conf = mddev->private;
1568        int err = 0;
1569        int number = rdev->raid_disk;
1570        struct raid1_info *p = conf->mirrors + number;
1571
1572        if (rdev != p->rdev)
1573                p = conf->mirrors + conf->raid_disks + number;
1574
1575        print_conf(conf);
1576        if (rdev == p->rdev) {
1577                if (test_bit(In_sync, &rdev->flags) ||
1578                    atomic_read(&rdev->nr_pending)) {
1579                        err = -EBUSY;
1580                        goto abort;
1581                }
1582                /* Only remove non-faulty devices if recovery
1583                 * is not possible.
1584                 */
1585                if (!test_bit(Faulty, &rdev->flags) &&
1586                    mddev->recovery_disabled != conf->recovery_disabled &&
1587                    mddev->degraded < conf->raid_disks) {
1588                        err = -EBUSY;
1589                        goto abort;
1590                }
1591                p->rdev = NULL;
1592                synchronize_rcu();
1593                if (atomic_read(&rdev->nr_pending)) {
1594                        /* lost the race, try later */
1595                        err = -EBUSY;
1596                        p->rdev = rdev;
1597                        goto abort;
1598                } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1599                        /* We just removed a device that is being replaced.
1600                         * Move down the replacement.  We drain all IO before
1601                         * doing this to avoid confusion.
1602                         */
1603                        struct md_rdev *repl =
1604                                conf->mirrors[conf->raid_disks + number].rdev;
1605                        raise_barrier(conf);
1606                        clear_bit(Replacement, &repl->flags);
1607                        p->rdev = repl;
1608                        conf->mirrors[conf->raid_disks + number].rdev = NULL;
1609                        lower_barrier(conf);
1610                        clear_bit(WantReplacement, &rdev->flags);
1611                } else
1612                        clear_bit(WantReplacement, &rdev->flags);
1613                err = md_integrity_register(mddev);
1614        }
1615abort:
1616
1617        print_conf(conf);
1618        return err;
1619}
1620
1621
1622static void end_sync_read(struct bio *bio, int error)
1623{
1624        struct r1bio *r1_bio = bio->bi_private;
1625
1626        update_head_pos(r1_bio->read_disk, r1_bio);
1627
1628        /*
1629         * we have read a block, now it needs to be re-written,
1630         * or re-read if the read failed.
1631         * We don't do much here, just schedule handling by raid1d
1632         */
1633        if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1634                set_bit(R1BIO_Uptodate, &r1_bio->state);
1635
1636        if (atomic_dec_and_test(&r1_bio->remaining))
1637                reschedule_retry(r1_bio);
1638}
1639
1640static void end_sync_write(struct bio *bio, int error)
1641{
1642        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1643        struct r1bio *r1_bio = bio->bi_private;
1644        struct mddev *mddev = r1_bio->mddev;
1645        struct r1conf *conf = mddev->private;
1646        int mirror=0;
1647        sector_t first_bad;
1648        int bad_sectors;
1649
1650        mirror = find_bio_disk(r1_bio, bio);
1651
1652        if (!uptodate) {
1653                sector_t sync_blocks = 0;
1654                sector_t s = r1_bio->sector;
1655                long sectors_to_go = r1_bio->sectors;
1656                /* make sure these bits doesn't get cleared. */
1657                do {
1658                        bitmap_end_sync(mddev->bitmap, s,
1659                                        &sync_blocks, 1);
1660                        s += sync_blocks;
1661                        sectors_to_go -= sync_blocks;
1662                } while (sectors_to_go > 0);
1663                set_bit(WriteErrorSeen,
1664                        &conf->mirrors[mirror].rdev->flags);
1665                if (!test_and_set_bit(WantReplacement,
1666                                      &conf->mirrors[mirror].rdev->flags))
1667                        set_bit(MD_RECOVERY_NEEDED, &
1668                                mddev->recovery);
1669                set_bit(R1BIO_WriteError, &r1_bio->state);
1670        } else if (is_badblock(conf->mirrors[mirror].rdev,
1671                               r1_bio->sector,
1672                               r1_bio->sectors,
1673                               &first_bad, &bad_sectors) &&
1674                   !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1675                                r1_bio->sector,
1676                                r1_bio->sectors,
1677                                &first_bad, &bad_sectors)
1678                )
1679                set_bit(R1BIO_MadeGood, &r1_bio->state);
1680
1681        if (atomic_dec_and_test(&r1_bio->remaining)) {
1682                int s = r1_bio->sectors;
1683                if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1684                    test_bit(R1BIO_WriteError, &r1_bio->state))
1685                        reschedule_retry(r1_bio);
1686                else {
1687                        put_buf(r1_bio);
1688                        md_done_sync(mddev, s, uptodate);
1689                }
1690        }
1691}
1692
1693static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1694                            int sectors, struct page *page, int rw)
1695{
1696        if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1697                /* success */
1698                return 1;
1699        if (rw == WRITE) {
1700                set_bit(WriteErrorSeen, &rdev->flags);
1701                if (!test_and_set_bit(WantReplacement,
1702                                      &rdev->flags))
1703                        set_bit(MD_RECOVERY_NEEDED, &
1704                                rdev->mddev->recovery);
1705        }
1706        /* need to record an error - either for the block or the device */
1707        if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1708                md_error(rdev->mddev, rdev);
1709        return 0;
1710}
1711
1712static int fix_sync_read_error(struct r1bio *r1_bio)
1713{
1714        /* Try some synchronous reads of other devices to get
1715         * good data, much like with normal read errors.  Only
1716         * read into the pages we already have so we don't
1717         * need to re-issue the read request.
1718         * We don't need to freeze the array, because being in an
1719         * active sync request, there is no normal IO, and
1720         * no overlapping syncs.
1721         * We don't need to check is_badblock() again as we
1722         * made sure that anything with a bad block in range
1723         * will have bi_end_io clear.
1724         */
1725        struct mddev *mddev = r1_bio->mddev;
1726        struct r1conf *conf = mddev->private;
1727        struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1728        sector_t sect = r1_bio->sector;
1729        int sectors = r1_bio->sectors;
1730        int idx = 0;
1731
1732        while(sectors) {
1733                int s = sectors;
1734                int d = r1_bio->read_disk;
1735                int success = 0;
1736                struct md_rdev *rdev;
1737                int start;
1738
1739                if (s > (PAGE_SIZE>>9))
1740                        s = PAGE_SIZE >> 9;
1741                do {
1742                        if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1743                                /* No rcu protection needed here devices
1744                                 * can only be removed when no resync is
1745                                 * active, and resync is currently active
1746                                 */
1747                                rdev = conf->mirrors[d].rdev;
1748                                if (sync_page_io(rdev, sect, s<<9,
1749                                                 bio->bi_io_vec[idx].bv_page,
1750                                                 READ, false)) {
1751                                        success = 1;
1752                                        break;
1753                                }
1754                        }
1755                        d++;
1756                        if (d == conf->raid_disks * 2)
1757                                d = 0;
1758                } while (!success && d != r1_bio->read_disk);
1759
1760                if (!success) {
1761                        char b[BDEVNAME_SIZE];
1762                        int abort = 0;
1763                        /* Cannot read from anywhere, this block is lost.
1764                         * Record a bad block on each device.  If that doesn't
1765                         * work just disable and interrupt the recovery.
1766                         * Don't fail devices as that won't really help.
1767                         */
1768                        printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1769                               " for block %llu\n",
1770                               mdname(mddev),
1771                               bdevname(bio->bi_bdev, b),
1772                               (unsigned long long)r1_bio->sector);
1773                        for (d = 0; d < conf->raid_disks * 2; d++) {
1774                                rdev = conf->mirrors[d].rdev;
1775                                if (!rdev || test_bit(Faulty, &rdev->flags))
1776                                        continue;
1777                                if (!rdev_set_badblocks(rdev, sect, s, 0))
1778                                        abort = 1;
1779                        }
1780                        if (abort) {
1781                                conf->recovery_disabled =
1782                                        mddev->recovery_disabled;
1783                                set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1784                                md_done_sync(mddev, r1_bio->sectors, 0);
1785                                put_buf(r1_bio);
1786                                return 0;
1787                        }
1788                        /* Try next page */
1789                        sectors -= s;
1790                        sect += s;
1791                        idx++;
1792                        continue;
1793                }
1794
1795                start = d;
1796                /* write it back and re-read */
1797                while (d != r1_bio->read_disk) {
1798                        if (d == 0)
1799                                d = conf->raid_disks * 2;
1800                        d--;
1801                        if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1802                                continue;
1803                        rdev = conf->mirrors[d].rdev;
1804                        if (r1_sync_page_io(rdev, sect, s,
1805                                            bio->bi_io_vec[idx].bv_page,
1806                                            WRITE) == 0) {
1807                                r1_bio->bios[d]->bi_end_io = NULL;
1808                                rdev_dec_pending(rdev, mddev);
1809                        }
1810                }
1811                d = start;
1812                while (d != r1_bio->read_disk) {
1813                        if (d == 0)
1814                                d = conf->raid_disks * 2;
1815                        d--;
1816                        if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1817                                continue;
1818                        rdev = conf->mirrors[d].rdev;
1819                        if (r1_sync_page_io(rdev, sect, s,
1820                                            bio->bi_io_vec[idx].bv_page,
1821                                            READ) != 0)
1822                                atomic_add(s, &rdev->corrected_errors);
1823                }
1824                sectors -= s;
1825                sect += s;
1826                idx ++;
1827        }
1828        set_bit(R1BIO_Uptodate, &r1_bio->state);
1829        set_bit(BIO_UPTODATE, &bio->bi_flags);
1830        return 1;
1831}
1832
1833static int process_checks(struct r1bio *r1_bio)
1834{
1835        /* We have read all readable devices.  If we haven't
1836         * got the block, then there is no hope left.
1837         * If we have, then we want to do a comparison
1838         * and skip the write if everything is the same.
1839         * If any blocks failed to read, then we need to
1840         * attempt an over-write
1841         */
1842        struct mddev *mddev = r1_bio->mddev;
1843        struct r1conf *conf = mddev->private;
1844        int primary;
1845        int i;
1846        int vcnt;
1847
1848        for (primary = 0; primary < conf->raid_disks * 2; primary++)
1849                if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1850                    test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1851                        r1_bio->bios[primary]->bi_end_io = NULL;
1852                        rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1853                        break;
1854                }
1855        r1_bio->read_disk = primary;
1856        vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1857        for (i = 0; i < conf->raid_disks * 2; i++) {
1858                int j;
1859                struct bio *pbio = r1_bio->bios[primary];
1860                struct bio *sbio = r1_bio->bios[i];
1861                int size;
1862
1863                if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1864                        continue;
1865
1866                if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1867                        for (j = vcnt; j-- ; ) {
1868                                struct page *p, *s;
1869                                p = pbio->bi_io_vec[j].bv_page;
1870                                s = sbio->bi_io_vec[j].bv_page;
1871                                if (memcmp(page_address(p),
1872                                           page_address(s),
1873                                           sbio->bi_io_vec[j].bv_len))
1874                                        break;
1875                        }
1876                } else
1877                        j = 0;
1878                if (j >= 0)
1879                        atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
1880                if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1881                              && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1882                        /* No need to write to this device. */
1883                        sbio->bi_end_io = NULL;
1884                        rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1885                        continue;
1886                }
1887                /* fixup the bio for reuse */
1888                sbio->bi_vcnt = vcnt;
1889                sbio->bi_size = r1_bio->sectors << 9;
1890                sbio->bi_idx = 0;
1891                sbio->bi_phys_segments = 0;
1892                sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1893                sbio->bi_flags |= 1 << BIO_UPTODATE;
1894                sbio->bi_next = NULL;
1895                sbio->bi_sector = r1_bio->sector +
1896                        conf->mirrors[i].rdev->data_offset;
1897                sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1898                size = sbio->bi_size;
1899                for (j = 0; j < vcnt ; j++) {
1900                        struct bio_vec *bi;
1901                        bi = &sbio->bi_io_vec[j];
1902                        bi->bv_offset = 0;
1903                        if (size > PAGE_SIZE)
1904                                bi->bv_len = PAGE_SIZE;
1905                        else
1906                                bi->bv_len = size;
1907                        size -= PAGE_SIZE;
1908                        memcpy(page_address(bi->bv_page),
1909                               page_address(pbio->bi_io_vec[j].bv_page),
1910                               PAGE_SIZE);
1911                }
1912        }
1913        return 0;
1914}
1915
1916static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1917{
1918        struct r1conf *conf = mddev->private;
1919        int i;
1920        int disks = conf->raid_disks * 2;
1921        struct bio *bio, *wbio;
1922
1923        bio = r1_bio->bios[r1_bio->read_disk];
1924
1925        if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1926                /* ouch - failed to read all of that. */
1927                if (!fix_sync_read_error(r1_bio))
1928                        return;
1929
1930        if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1931                if (process_checks(r1_bio) < 0)
1932                        return;
1933        /*
1934         * schedule writes
1935         */
1936        atomic_set(&r1_bio->remaining, 1);
1937        for (i = 0; i < disks ; i++) {
1938                wbio = r1_bio->bios[i];
1939                if (wbio->bi_end_io == NULL ||
1940                    (wbio->bi_end_io == end_sync_read &&
1941                     (i == r1_bio->read_disk ||
1942                      !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1943                        continue;
1944
1945                wbio->bi_rw = WRITE;
1946                wbio->bi_end_io = end_sync_write;
1947                atomic_inc(&r1_bio->remaining);
1948                md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1949
1950                generic_make_request(wbio);
1951        }
1952
1953        if (atomic_dec_and_test(&r1_bio->remaining)) {
1954                /* if we're here, all write(s) have completed, so clean up */
1955                int s = r1_bio->sectors;
1956                if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1957                    test_bit(R1BIO_WriteError, &r1_bio->state))
1958                        reschedule_retry(r1_bio);
1959                else {
1960                        put_buf(r1_bio);
1961                        md_done_sync(mddev, s, 1);
1962                }
1963        }
1964}
1965
1966/*
1967 * This is a kernel thread which:
1968 *
1969 *      1.      Retries failed read operations on working mirrors.
1970 *      2.      Updates the raid superblock when problems encounter.
1971 *      3.      Performs writes following reads for array synchronising.
1972 */
1973
1974static void fix_read_error(struct r1conf *conf, int read_disk,
1975                           sector_t sect, int sectors)
1976{
1977        struct mddev *mddev = conf->mddev;
1978        while(sectors) {
1979                int s = sectors;
1980                int d = read_disk;
1981                int success = 0;
1982                int start;
1983                struct md_rdev *rdev;
1984
1985                if (s > (PAGE_SIZE>>9))
1986                        s = PAGE_SIZE >> 9;
1987
1988                do {
1989                        /* Note: no rcu protection needed here
1990                         * as this is synchronous in the raid1d thread
1991                         * which is the thread that might remove
1992                         * a device.  If raid1d ever becomes multi-threaded....
1993                         */
1994                        sector_t first_bad;
1995                        int bad_sectors;
1996
1997                        rdev = conf->mirrors[d].rdev;
1998                        if (rdev &&
1999                            (test_bit(In_sync, &rdev->flags) ||
2000                             (!test_bit(Faulty, &rdev->flags) &&
2001                              rdev->recovery_offset >= sect + s)) &&
2002                            is_badblock(rdev, sect, s,
2003                                        &first_bad, &bad_sectors) == 0 &&
2004                            sync_page_io(rdev, sect, s<<9,
2005                                         conf->tmppage, READ, false))
2006                                success = 1;
2007                        else {
2008                                d++;
2009                                if (d == conf->raid_disks * 2)
2010                                        d = 0;
2011                        }
2012                } while (!success && d != read_disk);
2013
2014                if (!success) {
2015                        /* Cannot read from anywhere - mark it bad */
2016                        struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2017                        if (!rdev_set_badblocks(rdev, sect, s, 0))
2018                                md_error(mddev, rdev);
2019                        break;
2020                }
2021                /* write it back and re-read */
2022                start = d;
2023                while (d != read_disk) {
2024                        if (d==0)
2025                                d = conf->raid_disks * 2;
2026                        d--;
2027                        rdev = conf->mirrors[d].rdev;
2028                        if (rdev &&
2029                            test_bit(In_sync, &rdev->flags))
2030                                r1_sync_page_io(rdev, sect, s,
2031                                                conf->tmppage, WRITE);
2032                }
2033                d = start;
2034                while (d != read_disk) {
2035                        char b[BDEVNAME_SIZE];
2036                        if (d==0)
2037                                d = conf->raid_disks * 2;
2038                        d--;
2039                        rdev = conf->mirrors[d].rdev;
2040                        if (rdev &&
2041                            test_bit(In_sync, &rdev->flags)) {
2042                                if (r1_sync_page_io(rdev, sect, s,
2043                                                    conf->tmppage, READ)) {
2044                                        atomic_add(s, &rdev->corrected_errors);
2045                                        printk(KERN_INFO
2046                                               "md/raid1:%s: read error corrected "
2047                                               "(%d sectors at %llu on %s)\n",
2048                                               mdname(mddev), s,
2049                                               (unsigned long long)(sect +
2050                                                   rdev->data_offset),
2051                                               bdevname(rdev->bdev, b));
2052                                }
2053                        }
2054                }
2055                sectors -= s;
2056                sect += s;
2057        }
2058}
2059
2060static void bi_complete(struct bio *bio, int error)
2061{
2062        complete((struct completion *)bio->bi_private);
2063}
2064
2065static int submit_bio_wait(int rw, struct bio *bio)
2066{
2067        struct completion event;
2068        rw |= REQ_SYNC;
2069
2070        init_completion(&event);
2071        bio->bi_private = &event;
2072        bio->bi_end_io = bi_complete;
2073        submit_bio(rw, bio);
2074        wait_for_completion(&event);
2075
2076        return test_bit(BIO_UPTODATE, &bio->bi_flags);
2077}
2078
2079static int narrow_write_error(struct r1bio *r1_bio, int i)
2080{
2081        struct mddev *mddev = r1_bio->mddev;
2082        struct r1conf *conf = mddev->private;
2083        struct md_rdev *rdev = conf->mirrors[i].rdev;
2084        int vcnt, idx;
2085        struct bio_vec *vec;
2086
2087        /* bio has the data to be written to device 'i' where
2088         * we just recently had a write error.
2089         * We repeatedly clone the bio and trim down to one block,
2090         * then try the write.  Where the write fails we record
2091         * a bad block.
2092         * It is conceivable that the bio doesn't exactly align with
2093         * blocks.  We must handle this somehow.
2094         *
2095         * We currently own a reference on the rdev.
2096         */
2097
2098        int block_sectors;
2099        sector_t sector;
2100        int sectors;
2101        int sect_to_write = r1_bio->sectors;
2102        int ok = 1;
2103
2104        if (rdev->badblocks.shift < 0)
2105                return 0;
2106
2107        block_sectors = 1 << rdev->badblocks.shift;
2108        sector = r1_bio->sector;
2109        sectors = ((sector + block_sectors)
2110                   & ~(sector_t)(block_sectors - 1))
2111                - sector;
2112
2113        if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2114                vcnt = r1_bio->behind_page_count;
2115                vec = r1_bio->behind_bvecs;
2116                idx = 0;
2117                while (vec[idx].bv_page == NULL)
2118                        idx++;
2119        } else {
2120                vcnt = r1_bio->master_bio->bi_vcnt;
2121                vec = r1_bio->master_bio->bi_io_vec;
2122                idx = r1_bio->master_bio->bi_idx;
2123        }
2124        while (sect_to_write) {
2125                struct bio *wbio;
2126                if (sectors > sect_to_write)
2127                        sectors = sect_to_write;
2128                /* Write at 'sector' for 'sectors'*/
2129
2130                wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2131                memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2132                wbio->bi_sector = r1_bio->sector;
2133                wbio->bi_rw = WRITE;
2134                wbio->bi_vcnt = vcnt;
2135                wbio->bi_size = r1_bio->sectors << 9;
2136                wbio->bi_idx = idx;
2137
2138                md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2139                wbio->bi_sector += rdev->data_offset;
2140                wbio->bi_bdev = rdev->bdev;
2141                if (submit_bio_wait(WRITE, wbio) == 0)
2142                        /* failure! */
2143                        ok = rdev_set_badblocks(rdev, sector,
2144                                                sectors, 0)
2145                                && ok;
2146
2147                bio_put(wbio);
2148                sect_to_write -= sectors;
2149                sector += sectors;
2150                sectors = block_sectors;
2151        }
2152        return ok;
2153}
2154
2155static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2156{
2157        int m;
2158        int s = r1_bio->sectors;
2159        for (m = 0; m < conf->raid_disks * 2 ; m++) {
2160                struct md_rdev *rdev = conf->mirrors[m].rdev;
2161                struct bio *bio = r1_bio->bios[m];
2162                if (bio->bi_end_io == NULL)
2163                        continue;
2164                if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2165                    test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2166                        rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2167                }
2168                if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2169                    test_bit(R1BIO_WriteError, &r1_bio->state)) {
2170                        if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2171                                md_error(conf->mddev, rdev);
2172                }
2173        }
2174        put_buf(r1_bio);
2175        md_done_sync(conf->mddev, s, 1);
2176}
2177
2178static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2179{
2180        int m;
2181        for (m = 0; m < conf->raid_disks * 2 ; m++)
2182                if (r1_bio->bios[m] == IO_MADE_GOOD) {
2183                        struct md_rdev *rdev = conf->mirrors[m].rdev;
2184                        rdev_clear_badblocks(rdev,
2185                                             r1_bio->sector,
2186                                             r1_bio->sectors, 0);
2187                        rdev_dec_pending(rdev, conf->mddev);
2188                } else if (r1_bio->bios[m] != NULL) {
2189                        /* This drive got a write error.  We need to
2190                         * narrow down and record precise write
2191                         * errors.
2192                         */
2193                        if (!narrow_write_error(r1_bio, m)) {
2194                                md_error(conf->mddev,
2195                                         conf->mirrors[m].rdev);
2196                                /* an I/O failed, we can't clear the bitmap */
2197                                set_bit(R1BIO_Degraded, &r1_bio->state);
2198                        }
2199                        rdev_dec_pending(conf->mirrors[m].rdev,
2200                                         conf->mddev);
2201                }
2202        if (test_bit(R1BIO_WriteError, &r1_bio->state))
2203                close_write(r1_bio);
2204        raid_end_bio_io(r1_bio);
2205}
2206
2207static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2208{
2209        int disk;
2210        int max_sectors;
2211        struct mddev *mddev = conf->mddev;
2212        struct bio *bio;
2213        char b[BDEVNAME_SIZE];
2214        struct md_rdev *rdev;
2215
2216        clear_bit(R1BIO_ReadError, &r1_bio->state);
2217        /* we got a read error. Maybe the drive is bad.  Maybe just
2218         * the block and we can fix it.
2219         * We freeze all other IO, and try reading the block from
2220         * other devices.  When we find one, we re-write
2221         * and check it that fixes the read error.
2222         * This is all done synchronously while the array is
2223         * frozen
2224         */
2225        if (mddev->ro == 0) {
2226                freeze_array(conf);
2227                fix_read_error(conf, r1_bio->read_disk,
2228                               r1_bio->sector, r1_bio->sectors);
2229                unfreeze_array(conf);
2230        } else
2231                md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2232        rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2233
2234        bio = r1_bio->bios[r1_bio->read_disk];
2235        bdevname(bio->bi_bdev, b);
2236read_more:
2237        disk = read_balance(conf, r1_bio, &max_sectors);
2238        if (disk == -1) {
2239                printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2240                       " read error for block %llu\n",
2241                       mdname(mddev), b, (unsigned long long)r1_bio->sector);
2242                raid_end_bio_io(r1_bio);
2243        } else {
2244                const unsigned long do_sync
2245                        = r1_bio->master_bio->bi_rw & REQ_SYNC;
2246                if (bio) {
2247                        r1_bio->bios[r1_bio->read_disk] =
2248                                mddev->ro ? IO_BLOCKED : NULL;
2249                        bio_put(bio);
2250                }
2251                r1_bio->read_disk = disk;
2252                bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2253                md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2254                r1_bio->bios[r1_bio->read_disk] = bio;
2255                rdev = conf->mirrors[disk].rdev;
2256                printk_ratelimited(KERN_ERR
2257                                   "md/raid1:%s: redirecting sector %llu"
2258                                   " to other mirror: %s\n",
2259                                   mdname(mddev),
2260                                   (unsigned long long)r1_bio->sector,
2261                                   bdevname(rdev->bdev, b));
2262                bio->bi_sector = r1_bio->sector + rdev->data_offset;
2263                bio->bi_bdev = rdev->bdev;
2264                bio->bi_end_io = raid1_end_read_request;
2265                bio->bi_rw = READ | do_sync;
2266                bio->bi_private = r1_bio;
2267                if (max_sectors < r1_bio->sectors) {
2268                        /* Drat - have to split this up more */
2269                        struct bio *mbio = r1_bio->master_bio;
2270                        int sectors_handled = (r1_bio->sector + max_sectors
2271                                               - mbio->bi_sector);
2272                        r1_bio->sectors = max_sectors;
2273                        spin_lock_irq(&conf->device_lock);
2274                        if (mbio->bi_phys_segments == 0)
2275                                mbio->bi_phys_segments = 2;
2276                        else
2277                                mbio->bi_phys_segments++;
2278                        spin_unlock_irq(&conf->device_lock);
2279                        generic_make_request(bio);
2280                        bio = NULL;
2281
2282                        r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2283
2284                        r1_bio->master_bio = mbio;
2285                        r1_bio->sectors = (mbio->bi_size >> 9)
2286                                          - sectors_handled;
2287                        r1_bio->state = 0;
2288                        set_bit(R1BIO_ReadError, &r1_bio->state);
2289                        r1_bio->mddev = mddev;
2290                        r1_bio->sector = mbio->bi_sector + sectors_handled;
2291
2292                        goto read_more;
2293                } else
2294                        generic_make_request(bio);
2295        }
2296}
2297
2298static void raid1d(struct md_thread *thread)
2299{
2300        struct mddev *mddev = thread->mddev;
2301        struct r1bio *r1_bio;
2302        unsigned long flags;
2303        struct r1conf *conf = mddev->private;
2304        struct list_head *head = &conf->retry_list;
2305        struct blk_plug plug;
2306
2307        md_check_recovery(mddev);
2308
2309        blk_start_plug(&plug);
2310        for (;;) {
2311
2312                flush_pending_writes(conf);
2313
2314                spin_lock_irqsave(&conf->device_lock, flags);
2315                if (list_empty(head)) {
2316                        spin_unlock_irqrestore(&conf->device_lock, flags);
2317                        break;
2318                }
2319                r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2320                list_del(head->prev);
2321                conf->nr_queued--;
2322                spin_unlock_irqrestore(&conf->device_lock, flags);
2323
2324                mddev = r1_bio->mddev;
2325                conf = mddev->private;
2326                if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2327                        if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2328                            test_bit(R1BIO_WriteError, &r1_bio->state))
2329                                handle_sync_write_finished(conf, r1_bio);
2330                        else
2331                                sync_request_write(mddev, r1_bio);
2332                } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2333                           test_bit(R1BIO_WriteError, &r1_bio->state))
2334                        handle_write_finished(conf, r1_bio);
2335                else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2336                        handle_read_error(conf, r1_bio);
2337                else
2338                        /* just a partial read to be scheduled from separate
2339                         * context
2340                         */
2341                        generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2342
2343                cond_resched();
2344                if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2345                        md_check_recovery(mddev);
2346        }
2347        blk_finish_plug(&plug);
2348}
2349
2350
2351static int init_resync(struct r1conf *conf)
2352{
2353        int buffs;
2354
2355        buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2356        BUG_ON(conf->r1buf_pool);
2357        conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2358                                          conf->poolinfo);
2359        if (!conf->r1buf_pool)
2360                return -ENOMEM;
2361        conf->next_resync = 0;
2362        return 0;
2363}
2364
2365/*
2366 * perform a "sync" on one "block"
2367 *
2368 * We need to make sure that no normal I/O request - particularly write
2369 * requests - conflict with active sync requests.
2370 *
2371 * This is achieved by tracking pending requests and a 'barrier' concept
2372 * that can be installed to exclude normal IO requests.
2373 */
2374
2375static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2376{
2377        struct r1conf *conf = mddev->private;
2378        struct r1bio *r1_bio;
2379        struct bio *bio;
2380        sector_t max_sector, nr_sectors;
2381        int disk = -1;
2382        int i;
2383        int wonly = -1;
2384        int write_targets = 0, read_targets = 0;
2385        sector_t sync_blocks;
2386        int still_degraded = 0;
2387        int good_sectors = RESYNC_SECTORS;
2388        int min_bad = 0; /* number of sectors that are bad in all devices */
2389
2390        if (!conf->r1buf_pool)
2391                if (init_resync(conf))
2392                        return 0;
2393
2394        max_sector = mddev->dev_sectors;
2395        if (sector_nr >= max_sector) {
2396                /* If we aborted, we need to abort the
2397                 * sync on the 'current' bitmap chunk (there will
2398                 * only be one in raid1 resync.
2399                 * We can find the current addess in mddev->curr_resync
2400                 */
2401                if (mddev->curr_resync < max_sector) /* aborted */
2402                        bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2403                                                &sync_blocks, 1);
2404                else /* completed sync */
2405                        conf->fullsync = 0;
2406
2407                bitmap_close_sync(mddev->bitmap);
2408                close_sync(conf);
2409                return 0;
2410        }
2411
2412        if (mddev->bitmap == NULL &&
2413            mddev->recovery_cp == MaxSector &&
2414            !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2415            conf->fullsync == 0) {
2416                *skipped = 1;
2417                return max_sector - sector_nr;
2418        }
2419        /* before building a request, check if we can skip these blocks..
2420         * This call the bitmap_start_sync doesn't actually record anything
2421         */
2422        if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2423            !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2424                /* We can skip this block, and probably several more */
2425                *skipped = 1;
2426                return sync_blocks;
2427        }
2428        /*
2429         * If there is non-resync activity waiting for a turn,
2430         * and resync is going fast enough,
2431         * then let it though before starting on this new sync request.
2432         */
2433        if (!go_faster && conf->nr_waiting)
2434                msleep_interruptible(1000);
2435
2436        bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2437        r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2438        raise_barrier(conf);
2439
2440        conf->next_resync = sector_nr;
2441
2442        rcu_read_lock();
2443        /*
2444         * If we get a correctably read error during resync or recovery,
2445         * we might want to read from a different device.  So we
2446         * flag all drives that could conceivably be read from for READ,
2447         * and any others (which will be non-In_sync devices) for WRITE.
2448         * If a read fails, we try reading from something else for which READ
2449         * is OK.
2450         */
2451
2452        r1_bio->mddev = mddev;
2453        r1_bio->sector = sector_nr;
2454        r1_bio->state = 0;
2455        set_bit(R1BIO_IsSync, &r1_bio->state);
2456
2457        for (i = 0; i < conf->raid_disks * 2; i++) {
2458                struct md_rdev *rdev;
2459                bio = r1_bio->bios[i];
2460
2461                /* take from bio_init */
2462                bio->bi_next = NULL;
2463                bio->bi_flags &= ~(BIO_POOL_MASK-1);
2464                bio->bi_flags |= 1 << BIO_UPTODATE;
2465                bio->bi_rw = READ;
2466                bio->bi_vcnt = 0;
2467                bio->bi_idx = 0;
2468                bio->bi_phys_segments = 0;
2469                bio->bi_size = 0;
2470                bio->bi_end_io = NULL;
2471                bio->bi_private = NULL;
2472
2473                rdev = rcu_dereference(conf->mirrors[i].rdev);
2474                if (rdev == NULL ||
2475                    test_bit(Faulty, &rdev->flags)) {
2476                        if (i < conf->raid_disks)
2477                                still_degraded = 1;
2478                } else if (!test_bit(In_sync, &rdev->flags)) {
2479                        bio->bi_rw = WRITE;
2480                        bio->bi_end_io = end_sync_write;
2481                        write_targets ++;
2482                } else {
2483                        /* may need to read from here */
2484                        sector_t first_bad = MaxSector;
2485                        int bad_sectors;
2486
2487                        if (is_badblock(rdev, sector_nr, good_sectors,
2488                                        &first_bad, &bad_sectors)) {
2489                                if (first_bad > sector_nr)
2490                                        good_sectors = first_bad - sector_nr;
2491                                else {
2492                                        bad_sectors -= (sector_nr - first_bad);
2493                                        if (min_bad == 0 ||
2494                                            min_bad > bad_sectors)
2495                                                min_bad = bad_sectors;
2496                                }
2497                        }
2498                        if (sector_nr < first_bad) {
2499                                if (test_bit(WriteMostly, &rdev->flags)) {
2500                                        if (wonly < 0)
2501                                                wonly = i;
2502                                } else {
2503                                        if (disk < 0)
2504                                                disk = i;
2505                                }
2506                                bio->bi_rw = READ;
2507                                bio->bi_end_io = end_sync_read;
2508                                read_targets++;
2509                        } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2510                                test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2511                                !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2512                                /*
2513                                 * The device is suitable for reading (InSync),
2514                                 * but has bad block(s) here. Let's try to correct them,
2515                                 * if we are doing resync or repair. Otherwise, leave
2516                                 * this device alone for this sync request.
2517                                 */
2518                                bio->bi_rw = WRITE;
2519                                bio->bi_end_io = end_sync_write;
2520                                write_targets++;
2521                        }
2522                }
2523                if (bio->bi_end_io) {
2524                        atomic_inc(&rdev->nr_pending);
2525                        bio->bi_sector = sector_nr + rdev->data_offset;
2526                        bio->bi_bdev = rdev->bdev;
2527                        bio->bi_private = r1_bio;
2528                }
2529        }
2530        rcu_read_unlock();
2531        if (disk < 0)
2532                disk = wonly;
2533        r1_bio->read_disk = disk;
2534
2535        if (read_targets == 0 && min_bad > 0) {
2536                /* These sectors are bad on all InSync devices, so we
2537                 * need to mark them bad on all write targets
2538                 */
2539                int ok = 1;
2540                for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2541                        if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2542                                struct md_rdev *rdev = conf->mirrors[i].rdev;
2543                                ok = rdev_set_badblocks(rdev, sector_nr,
2544                                                        min_bad, 0
2545                                        ) && ok;
2546                        }
2547                set_bit(MD_CHANGE_DEVS, &mddev->flags);
2548                *skipped = 1;
2549                put_buf(r1_bio);
2550
2551                if (!ok) {
2552                        /* Cannot record the badblocks, so need to
2553                         * abort the resync.
2554                         * If there are multiple read targets, could just
2555                         * fail the really bad ones ???
2556                         */
2557                        conf->recovery_disabled = mddev->recovery_disabled;
2558                        set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2559                        return 0;
2560                } else
2561                        return min_bad;
2562
2563        }
2564        if (min_bad > 0 && min_bad < good_sectors) {
2565                /* only resync enough to reach the next bad->good
2566                 * transition */
2567                good_sectors = min_bad;
2568        }
2569
2570        if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2571                /* extra read targets are also write targets */
2572                write_targets += read_targets-1;
2573
2574        if (write_targets == 0 || read_targets == 0) {
2575                /* There is nowhere to write, so all non-sync
2576                 * drives must be failed - so we are finished
2577                 */
2578                sector_t rv;
2579                if (min_bad > 0)
2580                        max_sector = sector_nr + min_bad;
2581                rv = max_sector - sector_nr;
2582                *skipped = 1;
2583                put_buf(r1_bio);
2584                return rv;
2585        }
2586
2587        if (max_sector > mddev->resync_max)
2588                max_sector = mddev->resync_max; /* Don't do IO beyond here */
2589        if (max_sector > sector_nr + good_sectors)
2590                max_sector = sector_nr + good_sectors;
2591        nr_sectors = 0;
2592        sync_blocks = 0;
2593        do {
2594                struct page *page;
2595                int len = PAGE_SIZE;
2596                if (sector_nr + (len>>9) > max_sector)
2597                        len = (max_sector - sector_nr) << 9;
2598                if (len == 0)
2599                        break;
2600                if (sync_blocks == 0) {
2601                        if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2602                                               &sync_blocks, still_degraded) &&
2603                            !conf->fullsync &&
2604                            !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2605                                break;
2606                        BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2607                        if ((len >> 9) > sync_blocks)
2608                                len = sync_blocks<<9;
2609                }
2610
2611                for (i = 0 ; i < conf->raid_disks * 2; i++) {
2612                        bio = r1_bio->bios[i];
2613                        if (bio->bi_end_io) {
2614                                page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2615                                if (bio_add_page(bio, page, len, 0) == 0) {
2616                                        /* stop here */
2617                                        bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2618                                        while (i > 0) {
2619                                                i--;
2620                                                bio = r1_bio->bios[i];
2621                                                if (bio->bi_end_io==NULL)
2622                                                        continue;
2623                                                /* remove last page from this bio */
2624                                                bio->bi_vcnt--;
2625                                                bio->bi_size -= len;
2626                                                bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2627                                        }
2628                                        goto bio_full;
2629                                }
2630                        }
2631                }
2632                nr_sectors += len>>9;
2633                sector_nr += len>>9;
2634                sync_blocks -= (len>>9);
2635        } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2636 bio_full:
2637        r1_bio->sectors = nr_sectors;
2638
2639        /* For a user-requested sync, we read all readable devices and do a
2640         * compare
2641         */
2642        if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2643                atomic_set(&r1_bio->remaining, read_targets);
2644                for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2645                        bio = r1_bio->bios[i];
2646                        if (bio->bi_end_io == end_sync_read) {
2647                                read_targets--;
2648                                md_sync_acct(bio->bi_bdev, nr_sectors);
2649                                generic_make_request(bio);
2650                        }
2651                }
2652        } else {
2653                atomic_set(&r1_bio->remaining, 1);
2654                bio = r1_bio->bios[r1_bio->read_disk];
2655                md_sync_acct(bio->bi_bdev, nr_sectors);
2656                generic_make_request(bio);
2657
2658        }
2659        return nr_sectors;
2660}
2661
2662static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2663{
2664        if (sectors)
2665                return sectors;
2666
2667        return mddev->dev_sectors;
2668}
2669
2670static struct r1conf *setup_conf(struct mddev *mddev)
2671{
2672        struct r1conf *conf;
2673        int i;
2674        struct raid1_info *disk;
2675        struct md_rdev *rdev;
2676        int err = -ENOMEM;
2677
2678        conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2679        if (!conf)
2680                goto abort;
2681
2682        conf->mirrors = kzalloc(sizeof(struct raid1_info)
2683                                * mddev->raid_disks * 2,
2684                                 GFP_KERNEL);
2685        if (!conf->mirrors)
2686                goto abort;
2687
2688        conf->tmppage = alloc_page(GFP_KERNEL);
2689        if (!conf->tmppage)
2690                goto abort;
2691
2692        conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2693        if (!conf->poolinfo)
2694                goto abort;
2695        conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2696        conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2697                                          r1bio_pool_free,
2698                                          conf->poolinfo);
2699        if (!conf->r1bio_pool)
2700                goto abort;
2701
2702        conf->poolinfo->mddev = mddev;
2703
2704        err = -EINVAL;
2705        spin_lock_init(&conf->device_lock);
2706        rdev_for_each(rdev, mddev) {
2707                struct request_queue *q;
2708                int disk_idx = rdev->raid_disk;
2709                if (disk_idx >= mddev->raid_disks
2710                    || disk_idx < 0)
2711                        continue;
2712                if (test_bit(Replacement, &rdev->flags))
2713                        disk = conf->mirrors + mddev->raid_disks + disk_idx;
2714                else
2715                        disk = conf->mirrors + disk_idx;
2716
2717                if (disk->rdev)
2718                        goto abort;
2719                disk->rdev = rdev;
2720                q = bdev_get_queue(rdev->bdev);
2721                if (q->merge_bvec_fn)
2722                        mddev->merge_check_needed = 1;
2723
2724                disk->head_position = 0;
2725                disk->seq_start = MaxSector;
2726        }
2727        conf->raid_disks = mddev->raid_disks;
2728        conf->mddev = mddev;
2729        INIT_LIST_HEAD(&conf->retry_list);
2730
2731        spin_lock_init(&conf->resync_lock);
2732        init_waitqueue_head(&conf->wait_barrier);
2733
2734        bio_list_init(&conf->pending_bio_list);
2735        conf->pending_count = 0;
2736        conf->recovery_disabled = mddev->recovery_disabled - 1;
2737
2738        err = -EIO;
2739        for (i = 0; i < conf->raid_disks * 2; i++) {
2740
2741                disk = conf->mirrors + i;
2742
2743                if (i < conf->raid_disks &&
2744                    disk[conf->raid_disks].rdev) {
2745                        /* This slot has a replacement. */
2746                        if (!disk->rdev) {
2747                                /* No original, just make the replacement
2748                                 * a recovering spare
2749                                 */
2750                                disk->rdev =
2751                                        disk[conf->raid_disks].rdev;
2752                                disk[conf->raid_disks].rdev = NULL;
2753                        } else if (!test_bit(In_sync, &disk->rdev->flags))
2754                                /* Original is not in_sync - bad */
2755                                goto abort;
2756                }
2757
2758                if (!disk->rdev ||
2759                    !test_bit(In_sync, &disk->rdev->flags)) {
2760                        disk->head_position = 0;
2761                        if (disk->rdev &&
2762                            (disk->rdev->saved_raid_disk < 0))
2763                                conf->fullsync = 1;
2764                }
2765        }
2766
2767        err = -ENOMEM;
2768        conf->thread = md_register_thread(raid1d, mddev, "raid1");
2769        if (!conf->thread) {
2770                printk(KERN_ERR
2771                       "md/raid1:%s: couldn't allocate thread\n",
2772                       mdname(mddev));
2773                goto abort;
2774        }
2775
2776        return conf;
2777
2778 abort:
2779        if (conf) {
2780                if (conf->r1bio_pool)
2781                        mempool_destroy(conf->r1bio_pool);
2782                kfree(conf->mirrors);
2783                safe_put_page(conf->tmppage);
2784                kfree(conf->poolinfo);
2785                kfree(conf);
2786        }
2787        return ERR_PTR(err);
2788}
2789
2790static int stop(struct mddev *mddev);
2791static int run(struct mddev *mddev)
2792{
2793        struct r1conf *conf;
2794        int i;
2795        struct md_rdev *rdev;
2796        int ret;
2797        bool discard_supported = false;
2798
2799        if (mddev->level != 1) {
2800                printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2801                       mdname(mddev), mddev->level);
2802                return -EIO;
2803        }
2804        if (mddev->reshape_position != MaxSector) {
2805                printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2806                       mdname(mddev));
2807                return -EIO;
2808        }
2809        /*
2810         * copy the already verified devices into our private RAID1
2811         * bookkeeping area. [whatever we allocate in run(),
2812         * should be freed in stop()]
2813         */
2814        if (mddev->private == NULL)
2815                conf = setup_conf(mddev);
2816        else
2817                conf = mddev->private;
2818
2819        if (IS_ERR(conf))
2820                return PTR_ERR(conf);
2821
2822        rdev_for_each(rdev, mddev) {
2823                if (!mddev->gendisk)
2824                        continue;
2825                disk_stack_limits(mddev->gendisk, rdev->bdev,
2826                                  rdev->data_offset << 9);
2827                if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2828                        discard_supported = true;
2829        }
2830
2831        mddev->degraded = 0;
2832        for (i=0; i < conf->raid_disks; i++)
2833                if (conf->mirrors[i].rdev == NULL ||
2834                    !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2835                    test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2836                        mddev->degraded++;
2837
2838        if (conf->raid_disks - mddev->degraded == 1)
2839                mddev->recovery_cp = MaxSector;
2840
2841        if (mddev->recovery_cp != MaxSector)
2842                printk(KERN_NOTICE "md/raid1:%s: not clean"
2843                       " -- starting background reconstruction\n",
2844                       mdname(mddev));
2845        printk(KERN_INFO 
2846                "md/raid1:%s: active with %d out of %d mirrors\n",
2847                mdname(mddev), mddev->raid_disks - mddev->degraded, 
2848                mddev->raid_disks);
2849
2850        /*
2851         * Ok, everything is just fine now
2852         */
2853        mddev->thread = conf->thread;
2854        conf->thread = NULL;
2855        mddev->private = conf;
2856
2857        md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2858
2859        if (mddev->queue) {
2860                mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2861                mddev->queue->backing_dev_info.congested_data = mddev;
2862                blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2863
2864                if (discard_supported)
2865                        queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2866                                                mddev->queue);
2867                else
2868                        queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2869                                                  mddev->queue);
2870        }
2871
2872        ret =  md_integrity_register(mddev);
2873        if (ret)
2874                stop(mddev);
2875        return ret;
2876}
2877
2878static int stop(struct mddev *mddev)
2879{
2880        struct r1conf *conf = mddev->private;
2881        struct bitmap *bitmap = mddev->bitmap;
2882
2883        /* wait for behind writes to complete */
2884        if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2885                printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2886                       mdname(mddev));
2887                /* need to kick something here to make sure I/O goes? */
2888                wait_event(bitmap->behind_wait,
2889                           atomic_read(&bitmap->behind_writes) == 0);
2890        }
2891
2892        raise_barrier(conf);
2893        lower_barrier(conf);
2894
2895        md_unregister_thread(&mddev->thread);
2896        if (conf->r1bio_pool)
2897                mempool_destroy(conf->r1bio_pool);
2898        kfree(conf->mirrors);
2899        kfree(conf->poolinfo);
2900        kfree(conf);
2901        mddev->private = NULL;
2902        return 0;
2903}
2904
2905static int raid1_resize(struct mddev *mddev, sector_t sectors)
2906{
2907        /* no resync is happening, and there is enough space
2908         * on all devices, so we can resize.
2909         * We need to make sure resync covers any new space.
2910         * If the array is shrinking we should possibly wait until
2911         * any io in the removed space completes, but it hardly seems
2912         * worth it.
2913         */
2914        sector_t newsize = raid1_size(mddev, sectors, 0);
2915        if (mddev->external_size &&
2916            mddev->array_sectors > newsize)
2917                return -EINVAL;
2918        if (mddev->bitmap) {
2919                int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2920                if (ret)
2921                        return ret;
2922        }
2923        md_set_array_sectors(mddev, newsize);
2924        set_capacity(mddev->gendisk, mddev->array_sectors);
2925        revalidate_disk(mddev->gendisk);
2926        if (sectors > mddev->dev_sectors &&
2927            mddev->recovery_cp > mddev->dev_sectors) {
2928                mddev->recovery_cp = mddev->dev_sectors;
2929                set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2930        }
2931        mddev->dev_sectors = sectors;
2932        mddev->resync_max_sectors = sectors;
2933        return 0;
2934}
2935
2936static int raid1_reshape(struct mddev *mddev)
2937{
2938        /* We need to:
2939         * 1/ resize the r1bio_pool
2940         * 2/ resize conf->mirrors
2941         *
2942         * We allocate a new r1bio_pool if we can.
2943         * Then raise a device barrier and wait until all IO stops.
2944         * Then resize conf->mirrors and swap in the new r1bio pool.
2945         *
2946         * At the same time, we "pack" the devices so that all the missing
2947         * devices have the higher raid_disk numbers.
2948         */
2949        mempool_t *newpool, *oldpool;
2950        struct pool_info *newpoolinfo;
2951        struct raid1_info *newmirrors;
2952        struct r1conf *conf = mddev->private;
2953        int cnt, raid_disks;
2954        unsigned long flags;
2955        int d, d2, err;
2956
2957        /* Cannot change chunk_size, layout, or level */
2958        if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2959            mddev->layout != mddev->new_layout ||
2960            mddev->level != mddev->new_level) {
2961                mddev->new_chunk_sectors = mddev->chunk_sectors;
2962                mddev->new_layout = mddev->layout;
2963                mddev->new_level = mddev->level;
2964                return -EINVAL;
2965        }
2966
2967        err = md_allow_write(mddev);
2968        if (err)
2969                return err;
2970
2971        raid_disks = mddev->raid_disks + mddev->delta_disks;
2972
2973        if (raid_disks < conf->raid_disks) {
2974                cnt=0;
2975                for (d= 0; d < conf->raid_disks; d++)
2976                        if (conf->mirrors[d].rdev)
2977                                cnt++;
2978                if (cnt > raid_disks)
2979                        return -EBUSY;
2980        }
2981
2982        newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2983        if (!newpoolinfo)
2984                return -ENOMEM;
2985        newpoolinfo->mddev = mddev;
2986        newpoolinfo->raid_disks = raid_disks * 2;
2987
2988        newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2989                                 r1bio_pool_free, newpoolinfo);
2990        if (!newpool) {
2991                kfree(newpoolinfo);
2992                return -ENOMEM;
2993        }
2994        newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
2995                             GFP_KERNEL);
2996        if (!newmirrors) {
2997                kfree(newpoolinfo);
2998                mempool_destroy(newpool);
2999                return -ENOMEM;
3000        }
3001
3002        raise_barrier(conf);
3003
3004        /* ok, everything is stopped */
3005        oldpool = conf->r1bio_pool;
3006        conf->r1bio_pool = newpool;
3007
3008        for (d = d2 = 0; d < conf->raid_disks; d++) {
3009                struct md_rdev *rdev = conf->mirrors[d].rdev;
3010                if (rdev && rdev->raid_disk != d2) {
3011                        sysfs_unlink_rdev(mddev, rdev);
3012                        rdev->raid_disk = d2;
3013                        sysfs_unlink_rdev(mddev, rdev);
3014                        if (sysfs_link_rdev(mddev, rdev))
3015                                printk(KERN_WARNING
3016                                       "md/raid1:%s: cannot register rd%d\n",
3017                                       mdname(mddev), rdev->raid_disk);
3018                }
3019                if (rdev)
3020                        newmirrors[d2++].rdev = rdev;
3021        }
3022        kfree(conf->mirrors);
3023        conf->mirrors = newmirrors;
3024        kfree(conf->poolinfo);
3025        conf->poolinfo = newpoolinfo;
3026
3027        spin_lock_irqsave(&conf->device_lock, flags);
3028        mddev->degraded += (raid_disks - conf->raid_disks);
3029        spin_unlock_irqrestore(&conf->device_lock, flags);
3030        conf->raid_disks = mddev->raid_disks = raid_disks;
3031        mddev->delta_disks = 0;
3032
3033        lower_barrier(conf);
3034
3035        set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3036        md_wakeup_thread(mddev->thread);
3037
3038        mempool_destroy(oldpool);
3039        return 0;
3040}
3041
3042static void raid1_quiesce(struct mddev *mddev, int state)
3043{
3044        struct r1conf *conf = mddev->private;
3045
3046        switch(state) {
3047        case 2: /* wake for suspend */
3048                wake_up(&conf->wait_barrier);
3049                break;
3050        case 1:
3051                raise_barrier(conf);
3052                break;
3053        case 0:
3054                lower_barrier(conf);
3055                break;
3056        }
3057}
3058
3059static void *raid1_takeover(struct mddev *mddev)
3060{
3061        /* raid1 can take over:
3062         *  raid5 with 2 devices, any layout or chunk size
3063         */
3064        if (mddev->level == 5 && mddev->raid_disks == 2) {
3065                struct r1conf *conf;
3066                mddev->new_level = 1;
3067                mddev->new_layout = 0;
3068                mddev->new_chunk_sectors = 0;
3069                conf = setup_conf(mddev);
3070                if (!IS_ERR(conf))
3071                        conf->barrier = 1;
3072                return conf;
3073        }
3074        return ERR_PTR(-EINVAL);
3075}
3076
3077static struct md_personality raid1_personality =
3078{
3079        .name           = "raid1",
3080        .level          = 1,
3081        .owner          = THIS_MODULE,
3082        .make_request   = make_request,
3083        .run            = run,
3084        .stop           = stop,
3085        .status         = status,
3086        .error_handler  = error,
3087        .hot_add_disk   = raid1_add_disk,
3088        .hot_remove_disk= raid1_remove_disk,
3089        .spare_active   = raid1_spare_active,
3090        .sync_request   = sync_request,
3091        .resize         = raid1_resize,
3092        .size           = raid1_size,
3093        .check_reshape  = raid1_reshape,
3094        .quiesce        = raid1_quiesce,
3095        .takeover       = raid1_takeover,
3096};
3097
3098static int __init raid_init(void)
3099{
3100        return register_md_personality(&raid1_personality);
3101}
3102
3103static void raid_exit(void)
3104{
3105        unregister_md_personality(&raid1_personality);
3106}
3107
3108module_init(raid_init);
3109module_exit(raid_exit);
3110MODULE_LICENSE("GPL");
3111MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3112MODULE_ALIAS("md-personality-3"); /* RAID1 */
3113MODULE_ALIAS("md-raid1");
3114MODULE_ALIAS("md-level-1");
3115
3116module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
3117
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.