linux/drivers/md/raid1.c
<<
>>
Prefs
   1/*
   2 * raid1.c : Multiple Devices driver for Linux
   3 *
   4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5 *
   6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7 *
   8 * RAID-1 management functions.
   9 *
  10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11 *
  12 * Fixes to reconstruction by Jakob Ƙstergaard" <jakob@ostenfeld.dk>
  13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14 *
  15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  16 * bitmapped intelligence in resync:
  17 *
  18 *      - bitmap marked during normal i/o
  19 *      - bitmap used to skip nondirty blocks during sync
  20 *
  21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  22 * - persistent bitmap code
  23 *
  24 * This program is free software; you can redistribute it and/or modify
  25 * it under the terms of the GNU General Public License as published by
  26 * the Free Software Foundation; either version 2, or (at your option)
  27 * any later version.
  28 *
  29 * You should have received a copy of the GNU General Public License
  30 * (for example /usr/src/linux/COPYING); if not, write to the Free
  31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  32 */
  33
  34#include <linux/slab.h>
  35#include <linux/delay.h>
  36#include <linux/blkdev.h>
  37#include <linux/module.h>
  38#include <linux/seq_file.h>
  39#include <linux/ratelimit.h>
  40#include "md.h"
  41#include "raid1.h"
  42#include "bitmap.h"
  43
  44/*
  45 * Number of guaranteed r1bios in case of extreme VM load:
  46 */
  47#define NR_RAID1_BIOS 256
  48
  49/* when we get a read error on a read-only array, we redirect to another
  50 * device without failing the first device, or trying to over-write to
  51 * correct the read error.  To keep track of bad blocks on a per-bio
  52 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
  53 */
  54#define IO_BLOCKED ((struct bio *)1)
  55/* When we successfully write to a known bad-block, we need to remove the
  56 * bad-block marking which must be done from process context.  So we record
  57 * the success by setting devs[n].bio to IO_MADE_GOOD
  58 */
  59#define IO_MADE_GOOD ((struct bio *)2)
  60
  61#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
  62
  63/* When there are this many requests queue to be written by
  64 * the raid1 thread, we become 'congested' to provide back-pressure
  65 * for writeback.
  66 */
  67static int max_queued_requests = 1024;
  68
  69static void allow_barrier(struct r1conf *conf);
  70static void lower_barrier(struct r1conf *conf);
  71
  72static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  73{
  74        struct pool_info *pi = data;
  75        int size = offsetof(struct r1bio, bios[pi->raid_disks]);
  76
  77        /* allocate a r1bio with room for raid_disks entries in the bios array */
  78        return kzalloc(size, gfp_flags);
  79}
  80
  81static void r1bio_pool_free(void *r1_bio, void *data)
  82{
  83        kfree(r1_bio);
  84}
  85
  86#define RESYNC_BLOCK_SIZE (64*1024)
  87//#define RESYNC_BLOCK_SIZE PAGE_SIZE
  88#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  89#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  90#define RESYNC_WINDOW (2048*1024)
  91
  92static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  93{
  94        struct pool_info *pi = data;
  95        struct page *page;
  96        struct r1bio *r1_bio;
  97        struct bio *bio;
  98        int i, j;
  99
 100        r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 101        if (!r1_bio)
 102                return NULL;
 103
 104        /*
 105         * Allocate bios : 1 for reading, n-1 for writing
 106         */
 107        for (j = pi->raid_disks ; j-- ; ) {
 108                bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 109                if (!bio)
 110                        goto out_free_bio;
 111                r1_bio->bios[j] = bio;
 112        }
 113        /*
 114         * Allocate RESYNC_PAGES data pages and attach them to
 115         * the first bio.
 116         * If this is a user-requested check/repair, allocate
 117         * RESYNC_PAGES for each bio.
 118         */
 119        if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 120                j = pi->raid_disks;
 121        else
 122                j = 1;
 123        while(j--) {
 124                bio = r1_bio->bios[j];
 125                for (i = 0; i < RESYNC_PAGES; i++) {
 126                        page = alloc_page(gfp_flags);
 127                        if (unlikely(!page))
 128                                goto out_free_pages;
 129
 130                        bio->bi_io_vec[i].bv_page = page;
 131                        bio->bi_vcnt = i+1;
 132                }
 133        }
 134        /* If not user-requests, copy the page pointers to all bios */
 135        if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 136                for (i=0; i<RESYNC_PAGES ; i++)
 137                        for (j=1; j<pi->raid_disks; j++)
 138                                r1_bio->bios[j]->bi_io_vec[i].bv_page =
 139                                        r1_bio->bios[0]->bi_io_vec[i].bv_page;
 140        }
 141
 142        r1_bio->master_bio = NULL;
 143
 144        return r1_bio;
 145
 146out_free_pages:
 147        for (j=0 ; j < pi->raid_disks; j++)
 148                for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 149                        put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 150        j = -1;
 151out_free_bio:
 152        while (++j < pi->raid_disks)
 153                bio_put(r1_bio->bios[j]);
 154        r1bio_pool_free(r1_bio, data);
 155        return NULL;
 156}
 157
 158static void r1buf_pool_free(void *__r1_bio, void *data)
 159{
 160        struct pool_info *pi = data;
 161        int i,j;
 162        struct r1bio *r1bio = __r1_bio;
 163
 164        for (i = 0; i < RESYNC_PAGES; i++)
 165                for (j = pi->raid_disks; j-- ;) {
 166                        if (j == 0 ||
 167                            r1bio->bios[j]->bi_io_vec[i].bv_page !=
 168                            r1bio->bios[0]->bi_io_vec[i].bv_page)
 169                                safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 170                }
 171        for (i=0 ; i < pi->raid_disks; i++)
 172                bio_put(r1bio->bios[i]);
 173
 174        r1bio_pool_free(r1bio, data);
 175}
 176
 177static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 178{
 179        int i;
 180
 181        for (i = 0; i < conf->raid_disks * 2; i++) {
 182                struct bio **bio = r1_bio->bios + i;
 183                if (!BIO_SPECIAL(*bio))
 184                        bio_put(*bio);
 185                *bio = NULL;
 186        }
 187}
 188
 189static void free_r1bio(struct r1bio *r1_bio)
 190{
 191        struct r1conf *conf = r1_bio->mddev->private;
 192
 193        put_all_bios(conf, r1_bio);
 194        mempool_free(r1_bio, conf->r1bio_pool);
 195}
 196
 197static void put_buf(struct r1bio *r1_bio)
 198{
 199        struct r1conf *conf = r1_bio->mddev->private;
 200        int i;
 201
 202        for (i = 0; i < conf->raid_disks * 2; i++) {
 203                struct bio *bio = r1_bio->bios[i];
 204                if (bio->bi_end_io)
 205                        rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 206        }
 207
 208        mempool_free(r1_bio, conf->r1buf_pool);
 209
 210        lower_barrier(conf);
 211}
 212
 213static void reschedule_retry(struct r1bio *r1_bio)
 214{
 215        unsigned long flags;
 216        struct mddev *mddev = r1_bio->mddev;
 217        struct r1conf *conf = mddev->private;
 218
 219        spin_lock_irqsave(&conf->device_lock, flags);
 220        list_add(&r1_bio->retry_list, &conf->retry_list);
 221        conf->nr_queued ++;
 222        spin_unlock_irqrestore(&conf->device_lock, flags);
 223
 224        wake_up(&conf->wait_barrier);
 225        md_wakeup_thread(mddev->thread);
 226}
 227
 228/*
 229 * raid_end_bio_io() is called when we have finished servicing a mirrored
 230 * operation and are ready to return a success/failure code to the buffer
 231 * cache layer.
 232 */
 233static void call_bio_endio(struct r1bio *r1_bio)
 234{
 235        struct bio *bio = r1_bio->master_bio;
 236        int done;
 237        struct r1conf *conf = r1_bio->mddev->private;
 238
 239        if (bio->bi_phys_segments) {
 240                unsigned long flags;
 241                spin_lock_irqsave(&conf->device_lock, flags);
 242                bio->bi_phys_segments--;
 243                done = (bio->bi_phys_segments == 0);
 244                spin_unlock_irqrestore(&conf->device_lock, flags);
 245        } else
 246                done = 1;
 247
 248        if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 249                clear_bit(BIO_UPTODATE, &bio->bi_flags);
 250        if (done) {
 251                bio_endio(bio, 0);
 252                /*
 253                 * Wake up any possible resync thread that waits for the device
 254                 * to go idle.
 255                 */
 256                allow_barrier(conf);
 257        }
 258}
 259
 260static void raid_end_bio_io(struct r1bio *r1_bio)
 261{
 262        struct bio *bio = r1_bio->master_bio;
 263
 264        /* if nobody has done the final endio yet, do it now */
 265        if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 266                pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 267                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 268                         (unsigned long long) bio->bi_sector,
 269                         (unsigned long long) bio->bi_sector +
 270                         (bio->bi_size >> 9) - 1);
 271
 272                call_bio_endio(r1_bio);
 273        }
 274        free_r1bio(r1_bio);
 275}
 276
 277/*
 278 * Update disk head position estimator based on IRQ completion info.
 279 */
 280static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 281{
 282        struct r1conf *conf = r1_bio->mddev->private;
 283
 284        conf->mirrors[disk].head_position =
 285                r1_bio->sector + (r1_bio->sectors);
 286}
 287
 288/*
 289 * Find the disk number which triggered given bio
 290 */
 291static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 292{
 293        int mirror;
 294        struct r1conf *conf = r1_bio->mddev->private;
 295        int raid_disks = conf->raid_disks;
 296
 297        for (mirror = 0; mirror < raid_disks * 2; mirror++)
 298                if (r1_bio->bios[mirror] == bio)
 299                        break;
 300
 301        BUG_ON(mirror == raid_disks * 2);
 302        update_head_pos(mirror, r1_bio);
 303
 304        return mirror;
 305}
 306
 307static void raid1_end_read_request(struct bio *bio, int error)
 308{
 309        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 310        struct r1bio *r1_bio = bio->bi_private;
 311        int mirror;
 312        struct r1conf *conf = r1_bio->mddev->private;
 313
 314        mirror = r1_bio->read_disk;
 315        /*
 316         * this branch is our 'one mirror IO has finished' event handler:
 317         */
 318        update_head_pos(mirror, r1_bio);
 319
 320        if (uptodate)
 321                set_bit(R1BIO_Uptodate, &r1_bio->state);
 322        else {
 323                /* If all other devices have failed, we want to return
 324                 * the error upwards rather than fail the last device.
 325                 * Here we redefine "uptodate" to mean "Don't want to retry"
 326                 */
 327                unsigned long flags;
 328                spin_lock_irqsave(&conf->device_lock, flags);
 329                if (r1_bio->mddev->degraded == conf->raid_disks ||
 330                    (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 331                     !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 332                        uptodate = 1;
 333                spin_unlock_irqrestore(&conf->device_lock, flags);
 334        }
 335
 336        if (uptodate) {
 337                raid_end_bio_io(r1_bio);
 338                rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 339        } else {
 340                /*
 341                 * oops, read error:
 342                 */
 343                char b[BDEVNAME_SIZE];
 344                printk_ratelimited(
 345                        KERN_ERR "md/raid1:%s: %s: "
 346                        "rescheduling sector %llu\n",
 347                        mdname(conf->mddev),
 348                        bdevname(conf->mirrors[mirror].rdev->bdev,
 349                                 b),
 350                        (unsigned long long)r1_bio->sector);
 351                set_bit(R1BIO_ReadError, &r1_bio->state);
 352                reschedule_retry(r1_bio);
 353                /* don't drop the reference on read_disk yet */
 354        }
 355}
 356
 357static void close_write(struct r1bio *r1_bio)
 358{
 359        /* it really is the end of this request */
 360        if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 361                /* free extra copy of the data pages */
 362                int i = r1_bio->behind_page_count;
 363                while (i--)
 364                        safe_put_page(r1_bio->behind_bvecs[i].bv_page);
 365                kfree(r1_bio->behind_bvecs);
 366                r1_bio->behind_bvecs = NULL;
 367        }
 368        /* clear the bitmap if all writes complete successfully */
 369        bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 370                        r1_bio->sectors,
 371                        !test_bit(R1BIO_Degraded, &r1_bio->state),
 372                        test_bit(R1BIO_BehindIO, &r1_bio->state));
 373        md_write_end(r1_bio->mddev);
 374}
 375
 376static void r1_bio_write_done(struct r1bio *r1_bio)
 377{
 378        if (!atomic_dec_and_test(&r1_bio->remaining))
 379                return;
 380
 381        if (test_bit(R1BIO_WriteError, &r1_bio->state))
 382                reschedule_retry(r1_bio);
 383        else {
 384                close_write(r1_bio);
 385                if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 386                        reschedule_retry(r1_bio);
 387                else
 388                        raid_end_bio_io(r1_bio);
 389        }
 390}
 391
 392static void raid1_end_write_request(struct bio *bio, int error)
 393{
 394        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 395        struct r1bio *r1_bio = bio->bi_private;
 396        int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 397        struct r1conf *conf = r1_bio->mddev->private;
 398        struct bio *to_put = NULL;
 399
 400        mirror = find_bio_disk(r1_bio, bio);
 401
 402        /*
 403         * 'one mirror IO has finished' event handler:
 404         */
 405        if (!uptodate) {
 406                set_bit(WriteErrorSeen,
 407                        &conf->mirrors[mirror].rdev->flags);
 408                if (!test_and_set_bit(WantReplacement,
 409                                      &conf->mirrors[mirror].rdev->flags))
 410                        set_bit(MD_RECOVERY_NEEDED, &
 411                                conf->mddev->recovery);
 412
 413                set_bit(R1BIO_WriteError, &r1_bio->state);
 414        } else {
 415                /*
 416                 * Set R1BIO_Uptodate in our master bio, so that we
 417                 * will return a good error code for to the higher
 418                 * levels even if IO on some other mirrored buffer
 419                 * fails.
 420                 *
 421                 * The 'master' represents the composite IO operation
 422                 * to user-side. So if something waits for IO, then it
 423                 * will wait for the 'master' bio.
 424                 */
 425                sector_t first_bad;
 426                int bad_sectors;
 427
 428                r1_bio->bios[mirror] = NULL;
 429                to_put = bio;
 430                set_bit(R1BIO_Uptodate, &r1_bio->state);
 431
 432                /* Maybe we can clear some bad blocks. */
 433                if (is_badblock(conf->mirrors[mirror].rdev,
 434                                r1_bio->sector, r1_bio->sectors,
 435                                &first_bad, &bad_sectors)) {
 436                        r1_bio->bios[mirror] = IO_MADE_GOOD;
 437                        set_bit(R1BIO_MadeGood, &r1_bio->state);
 438                }
 439        }
 440
 441        if (behind) {
 442                if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 443                        atomic_dec(&r1_bio->behind_remaining);
 444
 445                /*
 446                 * In behind mode, we ACK the master bio once the I/O
 447                 * has safely reached all non-writemostly
 448                 * disks. Setting the Returned bit ensures that this
 449                 * gets done only once -- we don't ever want to return
 450                 * -EIO here, instead we'll wait
 451                 */
 452                if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 453                    test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 454                        /* Maybe we can return now */
 455                        if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 456                                struct bio *mbio = r1_bio->master_bio;
 457                                pr_debug("raid1: behind end write sectors"
 458                                         " %llu-%llu\n",
 459                                         (unsigned long long) mbio->bi_sector,
 460                                         (unsigned long long) mbio->bi_sector +
 461                                         (mbio->bi_size >> 9) - 1);
 462                                call_bio_endio(r1_bio);
 463                        }
 464                }
 465        }
 466        if (r1_bio->bios[mirror] == NULL)
 467                rdev_dec_pending(conf->mirrors[mirror].rdev,
 468                                 conf->mddev);
 469
 470        /*
 471         * Let's see if all mirrored write operations have finished
 472         * already.
 473         */
 474        r1_bio_write_done(r1_bio);
 475
 476        if (to_put)
 477                bio_put(to_put);
 478}
 479
 480
 481/*
 482 * This routine returns the disk from which the requested read should
 483 * be done. There is a per-array 'next expected sequential IO' sector
 484 * number - if this matches on the next IO then we use the last disk.
 485 * There is also a per-disk 'last know head position' sector that is
 486 * maintained from IRQ contexts, both the normal and the resync IO
 487 * completion handlers update this position correctly. If there is no
 488 * perfect sequential match then we pick the disk whose head is closest.
 489 *
 490 * If there are 2 mirrors in the same 2 devices, performance degrades
 491 * because position is mirror, not device based.
 492 *
 493 * The rdev for the device selected will have nr_pending incremented.
 494 */
 495static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 496{
 497        const sector_t this_sector = r1_bio->sector;
 498        int sectors;
 499        int best_good_sectors;
 500        int best_disk, best_dist_disk, best_pending_disk;
 501        int has_nonrot_disk;
 502        int disk;
 503        sector_t best_dist;
 504        unsigned int min_pending;
 505        struct md_rdev *rdev;
 506        int choose_first;
 507        int choose_next_idle;
 508
 509        rcu_read_lock();
 510        /*
 511         * Check if we can balance. We can balance on the whole
 512         * device if no resync is going on, or below the resync window.
 513         * We take the first readable disk when above the resync window.
 514         */
 515 retry:
 516        sectors = r1_bio->sectors;
 517        best_disk = -1;
 518        best_dist_disk = -1;
 519        best_dist = MaxSector;
 520        best_pending_disk = -1;
 521        min_pending = UINT_MAX;
 522        best_good_sectors = 0;
 523        has_nonrot_disk = 0;
 524        choose_next_idle = 0;
 525
 526        if (conf->mddev->recovery_cp < MaxSector &&
 527            (this_sector + sectors >= conf->next_resync))
 528                choose_first = 1;
 529        else
 530                choose_first = 0;
 531
 532        for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 533                sector_t dist;
 534                sector_t first_bad;
 535                int bad_sectors;
 536                unsigned int pending;
 537                bool nonrot;
 538
 539                rdev = rcu_dereference(conf->mirrors[disk].rdev);
 540                if (r1_bio->bios[disk] == IO_BLOCKED
 541                    || rdev == NULL
 542                    || test_bit(Unmerged, &rdev->flags)
 543                    || test_bit(Faulty, &rdev->flags))
 544                        continue;
 545                if (!test_bit(In_sync, &rdev->flags) &&
 546                    rdev->recovery_offset < this_sector + sectors)
 547                        continue;
 548                if (test_bit(WriteMostly, &rdev->flags)) {
 549                        /* Don't balance among write-mostly, just
 550                         * use the first as a last resort */
 551                        if (best_disk < 0) {
 552                                if (is_badblock(rdev, this_sector, sectors,
 553                                                &first_bad, &bad_sectors)) {
 554                                        if (first_bad < this_sector)
 555                                                /* Cannot use this */
 556                                                continue;
 557                                        best_good_sectors = first_bad - this_sector;
 558                                } else
 559                                        best_good_sectors = sectors;
 560                                best_disk = disk;
 561                        }
 562                        continue;
 563                }
 564                /* This is a reasonable device to use.  It might
 565                 * even be best.
 566                 */
 567                if (is_badblock(rdev, this_sector, sectors,
 568                                &first_bad, &bad_sectors)) {
 569                        if (best_dist < MaxSector)
 570                                /* already have a better device */
 571                                continue;
 572                        if (first_bad <= this_sector) {
 573                                /* cannot read here. If this is the 'primary'
 574                                 * device, then we must not read beyond
 575                                 * bad_sectors from another device..
 576                                 */
 577                                bad_sectors -= (this_sector - first_bad);
 578                                if (choose_first && sectors > bad_sectors)
 579                                        sectors = bad_sectors;
 580                                if (best_good_sectors > sectors)
 581                                        best_good_sectors = sectors;
 582
 583                        } else {
 584                                sector_t good_sectors = first_bad - this_sector;
 585                                if (good_sectors > best_good_sectors) {
 586                                        best_good_sectors = good_sectors;
 587                                        best_disk = disk;
 588                                }
 589                                if (choose_first)
 590                                        break;
 591                        }
 592                        continue;
 593                } else
 594                        best_good_sectors = sectors;
 595
 596                nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 597                has_nonrot_disk |= nonrot;
 598                pending = atomic_read(&rdev->nr_pending);
 599                dist = abs(this_sector - conf->mirrors[disk].head_position);
 600                if (choose_first) {
 601                        best_disk = disk;
 602                        break;
 603                }
 604                /* Don't change to another disk for sequential reads */
 605                if (conf->mirrors[disk].next_seq_sect == this_sector
 606                    || dist == 0) {
 607                        int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 608                        struct raid1_info *mirror = &conf->mirrors[disk];
 609
 610                        best_disk = disk;
 611                        /*
 612                         * If buffered sequential IO size exceeds optimal
 613                         * iosize, check if there is idle disk. If yes, choose
 614                         * the idle disk. read_balance could already choose an
 615                         * idle disk before noticing it's a sequential IO in
 616                         * this disk. This doesn't matter because this disk
 617                         * will idle, next time it will be utilized after the
 618                         * first disk has IO size exceeds optimal iosize. In
 619                         * this way, iosize of the first disk will be optimal
 620                         * iosize at least. iosize of the second disk might be
 621                         * small, but not a big deal since when the second disk
 622                         * starts IO, the first disk is likely still busy.
 623                         */
 624                        if (nonrot && opt_iosize > 0 &&
 625                            mirror->seq_start != MaxSector &&
 626                            mirror->next_seq_sect > opt_iosize &&
 627                            mirror->next_seq_sect - opt_iosize >=
 628                            mirror->seq_start) {
 629                                choose_next_idle = 1;
 630                                continue;
 631                        }
 632                        break;
 633                }
 634                /* If device is idle, use it */
 635                if (pending == 0) {
 636                        best_disk = disk;
 637                        break;
 638                }
 639
 640                if (choose_next_idle)
 641                        continue;
 642
 643                if (min_pending > pending) {
 644                        min_pending = pending;
 645                        best_pending_disk = disk;
 646                }
 647
 648                if (dist < best_dist) {
 649                        best_dist = dist;
 650                        best_dist_disk = disk;
 651                }
 652        }
 653
 654        /*
 655         * If all disks are rotational, choose the closest disk. If any disk is
 656         * non-rotational, choose the disk with less pending request even the
 657         * disk is rotational, which might/might not be optimal for raids with
 658         * mixed ratation/non-rotational disks depending on workload.
 659         */
 660        if (best_disk == -1) {
 661                if (has_nonrot_disk)
 662                        best_disk = best_pending_disk;
 663                else
 664                        best_disk = best_dist_disk;
 665        }
 666
 667        if (best_disk >= 0) {
 668                rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 669                if (!rdev)
 670                        goto retry;
 671                atomic_inc(&rdev->nr_pending);
 672                if (test_bit(Faulty, &rdev->flags)) {
 673                        /* cannot risk returning a device that failed
 674                         * before we inc'ed nr_pending
 675                         */
 676                        rdev_dec_pending(rdev, conf->mddev);
 677                        goto retry;
 678                }
 679                sectors = best_good_sectors;
 680
 681                if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 682                        conf->mirrors[best_disk].seq_start = this_sector;
 683
 684                conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 685        }
 686        rcu_read_unlock();
 687        *max_sectors = sectors;
 688
 689        return best_disk;
 690}
 691
 692static int raid1_mergeable_bvec(struct request_queue *q,
 693                                struct bvec_merge_data *bvm,
 694                                struct bio_vec *biovec)
 695{
 696        struct mddev *mddev = q->queuedata;
 697        struct r1conf *conf = mddev->private;
 698        sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 699        int max = biovec->bv_len;
 700
 701        if (mddev->merge_check_needed) {
 702                int disk;
 703                rcu_read_lock();
 704                for (disk = 0; disk < conf->raid_disks * 2; disk++) {
 705                        struct md_rdev *rdev = rcu_dereference(
 706                                conf->mirrors[disk].rdev);
 707                        if (rdev && !test_bit(Faulty, &rdev->flags)) {
 708                                struct request_queue *q =
 709                                        bdev_get_queue(rdev->bdev);
 710                                if (q->merge_bvec_fn) {
 711                                        bvm->bi_sector = sector +
 712                                                rdev->data_offset;
 713                                        bvm->bi_bdev = rdev->bdev;
 714                                        max = min(max, q->merge_bvec_fn(
 715                                                          q, bvm, biovec));
 716                                }
 717                        }
 718                }
 719                rcu_read_unlock();
 720        }
 721        return max;
 722
 723}
 724
 725int md_raid1_congested(struct mddev *mddev, int bits)
 726{
 727        struct r1conf *conf = mddev->private;
 728        int i, ret = 0;
 729
 730        if ((bits & (1 << BDI_async_congested)) &&
 731            conf->pending_count >= max_queued_requests)
 732                return 1;
 733
 734        rcu_read_lock();
 735        for (i = 0; i < conf->raid_disks * 2; i++) {
 736                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 737                if (rdev && !test_bit(Faulty, &rdev->flags)) {
 738                        struct request_queue *q = bdev_get_queue(rdev->bdev);
 739
 740                        BUG_ON(!q);
 741
 742                        /* Note the '|| 1' - when read_balance prefers
 743                         * non-congested targets, it can be removed
 744                         */
 745                        if ((bits & (1<<BDI_async_congested)) || 1)
 746                                ret |= bdi_congested(&q->backing_dev_info, bits);
 747                        else
 748                                ret &= bdi_congested(&q->backing_dev_info, bits);
 749                }
 750        }
 751        rcu_read_unlock();
 752        return ret;
 753}
 754EXPORT_SYMBOL_GPL(md_raid1_congested);
 755
 756static int raid1_congested(void *data, int bits)
 757{
 758        struct mddev *mddev = data;
 759
 760        return mddev_congested(mddev, bits) ||
 761                md_raid1_congested(mddev, bits);
 762}
 763
 764static void flush_pending_writes(struct r1conf *conf)
 765{
 766        /* Any writes that have been queued but are awaiting
 767         * bitmap updates get flushed here.
 768         */
 769        spin_lock_irq(&conf->device_lock);
 770
 771        if (conf->pending_bio_list.head) {
 772                struct bio *bio;
 773                bio = bio_list_get(&conf->pending_bio_list);
 774                conf->pending_count = 0;
 775                spin_unlock_irq(&conf->device_lock);
 776                /* flush any pending bitmap writes to
 777                 * disk before proceeding w/ I/O */
 778                bitmap_unplug(conf->mddev->bitmap);
 779                wake_up(&conf->wait_barrier);
 780
 781                while (bio) { /* submit pending writes */
 782                        struct bio *next = bio->bi_next;
 783                        bio->bi_next = NULL;
 784                        if (unlikely((bio->bi_rw & REQ_DISCARD) &&
 785                            !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 786                                /* Just ignore it */
 787                                bio_endio(bio, 0);
 788                        else
 789                                generic_make_request(bio);
 790                        bio = next;
 791                }
 792        } else
 793                spin_unlock_irq(&conf->device_lock);
 794}
 795
 796/* Barriers....
 797 * Sometimes we need to suspend IO while we do something else,
 798 * either some resync/recovery, or reconfigure the array.
 799 * To do this we raise a 'barrier'.
 800 * The 'barrier' is a counter that can be raised multiple times
 801 * to count how many activities are happening which preclude
 802 * normal IO.
 803 * We can only raise the barrier if there is no pending IO.
 804 * i.e. if nr_pending == 0.
 805 * We choose only to raise the barrier if no-one is waiting for the
 806 * barrier to go down.  This means that as soon as an IO request
 807 * is ready, no other operations which require a barrier will start
 808 * until the IO request has had a chance.
 809 *
 810 * So: regular IO calls 'wait_barrier'.  When that returns there
 811 *    is no backgroup IO happening,  It must arrange to call
 812 *    allow_barrier when it has finished its IO.
 813 * backgroup IO calls must call raise_barrier.  Once that returns
 814 *    there is no normal IO happeing.  It must arrange to call
 815 *    lower_barrier when the particular background IO completes.
 816 */
 817#define RESYNC_DEPTH 32
 818
 819static void raise_barrier(struct r1conf *conf)
 820{
 821        spin_lock_irq(&conf->resync_lock);
 822
 823        /* Wait until no block IO is waiting */
 824        wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 825                            conf->resync_lock);
 826
 827        /* block any new IO from starting */
 828        conf->barrier++;
 829
 830        /* Now wait for all pending IO to complete */
 831        wait_event_lock_irq(conf->wait_barrier,
 832                            !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 833                            conf->resync_lock);
 834
 835        spin_unlock_irq(&conf->resync_lock);
 836}
 837
 838static void lower_barrier(struct r1conf *conf)
 839{
 840        unsigned long flags;
 841        BUG_ON(conf->barrier <= 0);
 842        spin_lock_irqsave(&conf->resync_lock, flags);
 843        conf->barrier--;
 844        spin_unlock_irqrestore(&conf->resync_lock, flags);
 845        wake_up(&conf->wait_barrier);
 846}
 847
 848static void wait_barrier(struct r1conf *conf)
 849{
 850        spin_lock_irq(&conf->resync_lock);
 851        if (conf->barrier) {
 852                conf->nr_waiting++;
 853                /* Wait for the barrier to drop.
 854                 * However if there are already pending
 855                 * requests (preventing the barrier from
 856                 * rising completely), and the
 857                 * pre-process bio queue isn't empty,
 858                 * then don't wait, as we need to empty
 859                 * that queue to get the nr_pending
 860                 * count down.
 861                 */
 862                wait_event_lock_irq(conf->wait_barrier,
 863                                    !conf->barrier ||
 864                                    (conf->nr_pending &&
 865                                     current->bio_list &&
 866                                     !bio_list_empty(current->bio_list)),
 867                                    conf->resync_lock);
 868                conf->nr_waiting--;
 869        }
 870        conf->nr_pending++;
 871        spin_unlock_irq(&conf->resync_lock);
 872}
 873
 874static void allow_barrier(struct r1conf *conf)
 875{
 876        unsigned long flags;
 877        spin_lock_irqsave(&conf->resync_lock, flags);
 878        conf->nr_pending--;
 879        spin_unlock_irqrestore(&conf->resync_lock, flags);
 880        wake_up(&conf->wait_barrier);
 881}
 882
 883static void freeze_array(struct r1conf *conf)
 884{
 885        /* stop syncio and normal IO and wait for everything to
 886         * go quite.
 887         * We increment barrier and nr_waiting, and then
 888         * wait until nr_pending match nr_queued+1
 889         * This is called in the context of one normal IO request
 890         * that has failed. Thus any sync request that might be pending
 891         * will be blocked by nr_pending, and we need to wait for
 892         * pending IO requests to complete or be queued for re-try.
 893         * Thus the number queued (nr_queued) plus this request (1)
 894         * must match the number of pending IOs (nr_pending) before
 895         * we continue.
 896         */
 897        spin_lock_irq(&conf->resync_lock);
 898        conf->barrier++;
 899        conf->nr_waiting++;
 900        wait_event_lock_irq_cmd(conf->wait_barrier,
 901                                conf->nr_pending == conf->nr_queued+1,
 902                                conf->resync_lock,
 903                                flush_pending_writes(conf));
 904        spin_unlock_irq(&conf->resync_lock);
 905}
 906static void unfreeze_array(struct r1conf *conf)
 907{
 908        /* reverse the effect of the freeze */
 909        spin_lock_irq(&conf->resync_lock);
 910        conf->barrier--;
 911        conf->nr_waiting--;
 912        wake_up(&conf->wait_barrier);
 913        spin_unlock_irq(&conf->resync_lock);
 914}
 915
 916
 917/* duplicate the data pages for behind I/O 
 918 */
 919static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
 920{
 921        int i;
 922        struct bio_vec *bvec;
 923        struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
 924                                        GFP_NOIO);
 925        if (unlikely(!bvecs))
 926                return;
 927
 928        bio_for_each_segment(bvec, bio, i) {
 929                bvecs[i] = *bvec;
 930                bvecs[i].bv_page = alloc_page(GFP_NOIO);
 931                if (unlikely(!bvecs[i].bv_page))
 932                        goto do_sync_io;
 933                memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
 934                       kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 935                kunmap(bvecs[i].bv_page);
 936                kunmap(bvec->bv_page);
 937        }
 938        r1_bio->behind_bvecs = bvecs;
 939        r1_bio->behind_page_count = bio->bi_vcnt;
 940        set_bit(R1BIO_BehindIO, &r1_bio->state);
 941        return;
 942
 943do_sync_io:
 944        for (i = 0; i < bio->bi_vcnt; i++)
 945                if (bvecs[i].bv_page)
 946                        put_page(bvecs[i].bv_page);
 947        kfree(bvecs);
 948        pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 949}
 950
 951struct raid1_plug_cb {
 952        struct blk_plug_cb      cb;
 953        struct bio_list         pending;
 954        int                     pending_cnt;
 955};
 956
 957static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
 958{
 959        struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
 960                                                  cb);
 961        struct mddev *mddev = plug->cb.data;
 962        struct r1conf *conf = mddev->private;
 963        struct bio *bio;
 964
 965        if (from_schedule || current->bio_list) {
 966                spin_lock_irq(&conf->device_lock);
 967                bio_list_merge(&conf->pending_bio_list, &plug->pending);
 968                conf->pending_count += plug->pending_cnt;
 969                spin_unlock_irq(&conf->device_lock);
 970                md_wakeup_thread(mddev->thread);
 971                kfree(plug);
 972                return;
 973        }
 974
 975        /* we aren't scheduling, so we can do the write-out directly. */
 976        bio = bio_list_get(&plug->pending);
 977        bitmap_unplug(mddev->bitmap);
 978        wake_up(&conf->wait_barrier);
 979
 980        while (bio) { /* submit pending writes */
 981                struct bio *next = bio->bi_next;
 982                bio->bi_next = NULL;
 983                generic_make_request(bio);
 984                bio = next;
 985        }
 986        kfree(plug);
 987}
 988
 989static void make_request(struct mddev *mddev, struct bio * bio)
 990{
 991        struct r1conf *conf = mddev->private;
 992        struct raid1_info *mirror;
 993        struct r1bio *r1_bio;
 994        struct bio *read_bio;
 995        int i, disks;
 996        struct bitmap *bitmap;
 997        unsigned long flags;
 998        const int rw = bio_data_dir(bio);
 999        const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1000        const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1001        const unsigned long do_discard = (bio->bi_rw
1002                                          & (REQ_DISCARD | REQ_SECURE));
1003        struct md_rdev *blocked_rdev;
1004        struct blk_plug_cb *cb;
1005        struct raid1_plug_cb *plug = NULL;
1006        int first_clone;
1007        int sectors_handled;
1008        int max_sectors;
1009
1010        /*
1011         * Register the new request and wait if the reconstruction
1012         * thread has put up a bar for new requests.
1013         * Continue immediately if no resync is active currently.
1014         */
1015
1016        md_write_start(mddev, bio); /* wait on superblock update early */
1017
1018        if (bio_data_dir(bio) == WRITE &&
1019            bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
1020            bio->bi_sector < mddev->suspend_hi) {
1021                /* As the suspend_* range is controlled by
1022                 * userspace, we want an interruptible
1023                 * wait.
1024                 */
1025                DEFINE_WAIT(w);
1026                for (;;) {
1027                        flush_signals(current);
1028                        prepare_to_wait(&conf->wait_barrier,
1029                                        &w, TASK_INTERRUPTIBLE);
1030                        if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
1031                            bio->bi_sector >= mddev->suspend_hi)
1032                                break;
1033                        schedule();
1034                }
1035                finish_wait(&conf->wait_barrier, &w);
1036        }
1037
1038        wait_barrier(conf);
1039
1040        bitmap = mddev->bitmap;
1041
1042        /*
1043         * make_request() can abort the operation when READA is being
1044         * used and no empty request is available.
1045         *
1046         */
1047        r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1048
1049        r1_bio->master_bio = bio;
1050        r1_bio->sectors = bio->bi_size >> 9;
1051        r1_bio->state = 0;
1052        r1_bio->mddev = mddev;
1053        r1_bio->sector = bio->bi_sector;
1054
1055        /* We might need to issue multiple reads to different
1056         * devices if there are bad blocks around, so we keep
1057         * track of the number of reads in bio->bi_phys_segments.
1058         * If this is 0, there is only one r1_bio and no locking
1059         * will be needed when requests complete.  If it is
1060         * non-zero, then it is the number of not-completed requests.
1061         */
1062        bio->bi_phys_segments = 0;
1063        clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1064
1065        if (rw == READ) {
1066                /*
1067                 * read balancing logic:
1068                 */
1069                int rdisk;
1070
1071read_again:
1072                rdisk = read_balance(conf, r1_bio, &max_sectors);
1073
1074                if (rdisk < 0) {
1075                        /* couldn't find anywhere to read from */
1076                        raid_end_bio_io(r1_bio);
1077                        return;
1078                }
1079                mirror = conf->mirrors + rdisk;
1080
1081                if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1082                    bitmap) {
1083                        /* Reading from a write-mostly device must
1084                         * take care not to over-take any writes
1085                         * that are 'behind'
1086                         */
1087                        wait_event(bitmap->behind_wait,
1088                                   atomic_read(&bitmap->behind_writes) == 0);
1089                }
1090                r1_bio->read_disk = rdisk;
1091
1092                read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1093                md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
1094                            max_sectors);
1095
1096                r1_bio->bios[rdisk] = read_bio;
1097
1098                read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
1099                read_bio->bi_bdev = mirror->rdev->bdev;
1100                read_bio->bi_end_io = raid1_end_read_request;
1101                read_bio->bi_rw = READ | do_sync;
1102                read_bio->bi_private = r1_bio;
1103
1104                if (max_sectors < r1_bio->sectors) {
1105                        /* could not read all from this device, so we will
1106                         * need another r1_bio.
1107                         */
1108
1109                        sectors_handled = (r1_bio->sector + max_sectors
1110                                           - bio->bi_sector);
1111                        r1_bio->sectors = max_sectors;
1112                        spin_lock_irq(&conf->device_lock);
1113                        if (bio->bi_phys_segments == 0)
1114                                bio->bi_phys_segments = 2;
1115                        else
1116                                bio->bi_phys_segments++;
1117                        spin_unlock_irq(&conf->device_lock);
1118                        /* Cannot call generic_make_request directly
1119                         * as that will be queued in __make_request
1120                         * and subsequent mempool_alloc might block waiting
1121                         * for it.  So hand bio over to raid1d.
1122                         */
1123                        reschedule_retry(r1_bio);
1124
1125                        r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1126
1127                        r1_bio->master_bio = bio;
1128                        r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1129                        r1_bio->state = 0;
1130                        r1_bio->mddev = mddev;
1131                        r1_bio->sector = bio->bi_sector + sectors_handled;
1132                        goto read_again;
1133                } else
1134                        generic_make_request(read_bio);
1135                return;
1136        }
1137
1138        /*
1139         * WRITE:
1140         */
1141        if (conf->pending_count >= max_queued_requests) {
1142                md_wakeup_thread(mddev->thread);
1143                wait_event(conf->wait_barrier,
1144                           conf->pending_count < max_queued_requests);
1145        }
1146        /* first select target devices under rcu_lock and
1147         * inc refcount on their rdev.  Record them by setting
1148         * bios[x] to bio
1149         * If there are known/acknowledged bad blocks on any device on
1150         * which we have seen a write error, we want to avoid writing those
1151         * blocks.
1152         * This potentially requires several writes to write around
1153         * the bad blocks.  Each set of writes gets it's own r1bio
1154         * with a set of bios attached.
1155         */
1156
1157        disks = conf->raid_disks * 2;
1158 retry_write:
1159        blocked_rdev = NULL;
1160        rcu_read_lock();
1161        max_sectors = r1_bio->sectors;
1162        for (i = 0;  i < disks; i++) {
1163                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1164                if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1165                        atomic_inc(&rdev->nr_pending);
1166                        blocked_rdev = rdev;
1167                        break;
1168                }
1169                r1_bio->bios[i] = NULL;
1170                if (!rdev || test_bit(Faulty, &rdev->flags)
1171                    || test_bit(Unmerged, &rdev->flags)) {
1172                        if (i < conf->raid_disks)
1173                                set_bit(R1BIO_Degraded, &r1_bio->state);
1174                        continue;
1175                }
1176
1177                atomic_inc(&rdev->nr_pending);
1178                if (test_bit(WriteErrorSeen, &rdev->flags)) {
1179                        sector_t first_bad;
1180                        int bad_sectors;
1181                        int is_bad;
1182
1183                        is_bad = is_badblock(rdev, r1_bio->sector,
1184                                             max_sectors,
1185                                             &first_bad, &bad_sectors);
1186                        if (is_bad < 0) {
1187                                /* mustn't write here until the bad block is
1188                                 * acknowledged*/
1189                                set_bit(BlockedBadBlocks, &rdev->flags);
1190                                blocked_rdev = rdev;
1191                                break;
1192                        }
1193                        if (is_bad && first_bad <= r1_bio->sector) {
1194                                /* Cannot write here at all */
1195                                bad_sectors -= (r1_bio->sector - first_bad);
1196                                if (bad_sectors < max_sectors)
1197                                        /* mustn't write more than bad_sectors
1198                                         * to other devices yet
1199                                         */
1200                                        max_sectors = bad_sectors;
1201                                rdev_dec_pending(rdev, mddev);
1202                                /* We don't set R1BIO_Degraded as that
1203                                 * only applies if the disk is
1204                                 * missing, so it might be re-added,
1205                                 * and we want to know to recover this
1206                                 * chunk.
1207                                 * In this case the device is here,
1208                                 * and the fact that this chunk is not
1209                                 * in-sync is recorded in the bad
1210                                 * block log
1211                                 */
1212                                continue;
1213                        }
1214                        if (is_bad) {
1215                                int good_sectors = first_bad - r1_bio->sector;
1216                                if (good_sectors < max_sectors)
1217                                        max_sectors = good_sectors;
1218                        }
1219                }
1220                r1_bio->bios[i] = bio;
1221        }
1222        rcu_read_unlock();
1223
1224        if (unlikely(blocked_rdev)) {
1225                /* Wait for this device to become unblocked */
1226                int j;
1227
1228                for (j = 0; j < i; j++)
1229                        if (r1_bio->bios[j])
1230                                rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1231                r1_bio->state = 0;
1232                allow_barrier(conf);
1233                md_wait_for_blocked_rdev(blocked_rdev, mddev);
1234                wait_barrier(conf);
1235                goto retry_write;
1236        }
1237
1238        if (max_sectors < r1_bio->sectors) {
1239                /* We are splitting this write into multiple parts, so
1240                 * we need to prepare for allocating another r1_bio.
1241                 */
1242                r1_bio->sectors = max_sectors;
1243                spin_lock_irq(&conf->device_lock);
1244                if (bio->bi_phys_segments == 0)
1245                        bio->bi_phys_segments = 2;
1246                else
1247                        bio->bi_phys_segments++;
1248                spin_unlock_irq(&conf->device_lock);
1249        }
1250        sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1251
1252        atomic_set(&r1_bio->remaining, 1);
1253        atomic_set(&r1_bio->behind_remaining, 0);
1254
1255        first_clone = 1;
1256        for (i = 0; i < disks; i++) {
1257                struct bio *mbio;
1258                if (!r1_bio->bios[i])
1259                        continue;
1260
1261                mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1262                md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1263
1264                if (first_clone) {
1265                        /* do behind I/O ?
1266                         * Not if there are too many, or cannot
1267                         * allocate memory, or a reader on WriteMostly
1268                         * is waiting for behind writes to flush */
1269                        if (bitmap &&
1270                            (atomic_read(&bitmap->behind_writes)
1271                             < mddev->bitmap_info.max_write_behind) &&
1272                            !waitqueue_active(&bitmap->behind_wait))
1273                                alloc_behind_pages(mbio, r1_bio);
1274
1275                        bitmap_startwrite(bitmap, r1_bio->sector,
1276                                          r1_bio->sectors,
1277                                          test_bit(R1BIO_BehindIO,
1278                                                   &r1_bio->state));
1279                        first_clone = 0;
1280                }
1281                if (r1_bio->behind_bvecs) {
1282                        struct bio_vec *bvec;
1283                        int j;
1284
1285                        /* Yes, I really want the '__' version so that
1286                         * we clear any unused pointer in the io_vec, rather
1287                         * than leave them unchanged.  This is important
1288                         * because when we come to free the pages, we won't
1289                         * know the original bi_idx, so we just free
1290                         * them all
1291                         */
1292                        __bio_for_each_segment(bvec, mbio, j, 0)
1293                                bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1294                        if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1295                                atomic_inc(&r1_bio->behind_remaining);
1296                }
1297
1298                r1_bio->bios[i] = mbio;
1299
1300                mbio->bi_sector = (r1_bio->sector +
1301                                   conf->mirrors[i].rdev->data_offset);
1302                mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1303                mbio->bi_end_io = raid1_end_write_request;
1304                mbio->bi_rw = WRITE | do_flush_fua | do_sync | do_discard;
1305                mbio->bi_private = r1_bio;
1306
1307                atomic_inc(&r1_bio->remaining);
1308
1309                cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1310                if (cb)
1311                        plug = container_of(cb, struct raid1_plug_cb, cb);
1312                else
1313                        plug = NULL;
1314                spin_lock_irqsave(&conf->device_lock, flags);
1315                if (plug) {
1316                        bio_list_add(&plug->pending, mbio);
1317                        plug->pending_cnt++;
1318                } else {
1319                        bio_list_add(&conf->pending_bio_list, mbio);
1320                        conf->pending_count++;
1321                }
1322                spin_unlock_irqrestore(&conf->device_lock, flags);
1323                if (!plug)
1324                        md_wakeup_thread(mddev->thread);
1325        }
1326        /* Mustn't call r1_bio_write_done before this next test,
1327         * as it could result in the bio being freed.
1328         */
1329        if (sectors_handled < (bio->bi_size >> 9)) {
1330                r1_bio_write_done(r1_bio);
1331                /* We need another r1_bio.  It has already been counted
1332                 * in bio->bi_phys_segments
1333                 */
1334                r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1335                r1_bio->master_bio = bio;
1336                r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1337                r1_bio->state = 0;
1338                r1_bio->mddev = mddev;
1339                r1_bio->sector = bio->bi_sector + sectors_handled;
1340                goto retry_write;
1341        }
1342
1343        r1_bio_write_done(r1_bio);
1344
1345        /* In case raid1d snuck in to freeze_array */
1346        wake_up(&conf->wait_barrier);
1347}
1348
1349static void status(struct seq_file *seq, struct mddev *mddev)
1350{
1351        struct r1conf *conf = mddev->private;
1352        int i;
1353
1354        seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1355                   conf->raid_disks - mddev->degraded);
1356        rcu_read_lock();
1357        for (i = 0; i < conf->raid_disks; i++) {
1358                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1359                seq_printf(seq, "%s",
1360                           rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1361        }
1362        rcu_read_unlock();
1363        seq_printf(seq, "]");
1364}
1365
1366
1367static void error(struct mddev *mddev, struct md_rdev *rdev)
1368{
1369        char b[BDEVNAME_SIZE];
1370        struct r1conf *conf = mddev->private;
1371
1372        /*
1373         * If it is not operational, then we have already marked it as dead
1374         * else if it is the last working disks, ignore the error, let the
1375         * next level up know.
1376         * else mark the drive as failed
1377         */
1378        if (test_bit(In_sync, &rdev->flags)
1379            && (conf->raid_disks - mddev->degraded) == 1) {
1380                /*
1381                 * Don't fail the drive, act as though we were just a
1382                 * normal single drive.
1383                 * However don't try a recovery from this drive as
1384                 * it is very likely to fail.
1385                 */
1386                conf->recovery_disabled = mddev->recovery_disabled;
1387                return;
1388        }
1389        set_bit(Blocked, &rdev->flags);
1390        if (test_and_clear_bit(In_sync, &rdev->flags)) {
1391                unsigned long flags;
1392                spin_lock_irqsave(&conf->device_lock, flags);
1393                mddev->degraded++;
1394                set_bit(Faulty, &rdev->flags);
1395                spin_unlock_irqrestore(&conf->device_lock, flags);
1396                /*
1397                 * if recovery is running, make sure it aborts.
1398                 */
1399                set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1400        } else
1401                set_bit(Faulty, &rdev->flags);
1402        set_bit(MD_CHANGE_DEVS, &mddev->flags);
1403        printk(KERN_ALERT
1404               "md/raid1:%s: Disk failure on %s, disabling device.\n"
1405               "md/raid1:%s: Operation continuing on %d devices.\n",
1406               mdname(mddev), bdevname(rdev->bdev, b),
1407               mdname(mddev), conf->raid_disks - mddev->degraded);
1408}
1409
1410static void print_conf(struct r1conf *conf)
1411{
1412        int i;
1413
1414        printk(KERN_DEBUG "RAID1 conf printout:\n");
1415        if (!conf) {
1416                printk(KERN_DEBUG "(!conf)\n");
1417                return;
1418        }
1419        printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1420                conf->raid_disks);
1421
1422        rcu_read_lock();
1423        for (i = 0; i < conf->raid_disks; i++) {
1424                char b[BDEVNAME_SIZE];
1425                struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1426                if (rdev)
1427                        printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1428                               i, !test_bit(In_sync, &rdev->flags),
1429                               !test_bit(Faulty, &rdev->flags),
1430                               bdevname(rdev->bdev,b));
1431        }
1432        rcu_read_unlock();
1433}
1434
1435static void close_sync(struct r1conf *conf)
1436{
1437        wait_barrier(conf);
1438        allow_barrier(conf);
1439
1440        mempool_destroy(conf->r1buf_pool);
1441        conf->r1buf_pool = NULL;
1442}
1443
1444static int raid1_spare_active(struct mddev *mddev)
1445{
1446        int i;
1447        struct r1conf *conf = mddev->private;
1448        int count = 0;
1449        unsigned long flags;
1450
1451        /*
1452         * Find all failed disks within the RAID1 configuration 
1453         * and mark them readable.
1454         * Called under mddev lock, so rcu protection not needed.
1455         */
1456        for (i = 0; i < conf->raid_disks; i++) {
1457                struct md_rdev *rdev = conf->mirrors[i].rdev;
1458                struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1459                if (repl
1460                    && repl->recovery_offset == MaxSector
1461                    && !test_bit(Faulty, &repl->flags)
1462                    && !test_and_set_bit(In_sync, &repl->flags)) {
1463                        /* replacement has just become active */
1464                        if (!rdev ||
1465                            !test_and_clear_bit(In_sync, &rdev->flags))
1466                                count++;
1467                        if (rdev) {
1468                                /* Replaced device not technically
1469                                 * faulty, but we need to be sure
1470                                 * it gets removed and never re-added
1471                                 */
1472                                set_bit(Faulty, &rdev->flags);
1473                                sysfs_notify_dirent_safe(
1474                                        rdev->sysfs_state);
1475                        }
1476                }
1477                if (rdev
1478                    && !test_bit(Faulty, &rdev->flags)
1479                    && !test_and_set_bit(In_sync, &rdev->flags)) {
1480                        count++;
1481                        sysfs_notify_dirent_safe(rdev->sysfs_state);
1482                }
1483        }
1484        spin_lock_irqsave(&conf->device_lock, flags);
1485        mddev->degraded -= count;
1486        spin_unlock_irqrestore(&conf->device_lock, flags);
1487
1488        print_conf(conf);
1489        return count;
1490}
1491
1492
1493static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1494{
1495        struct r1conf *conf = mddev->private;
1496        int err = -EEXIST;
1497        int mirror = 0;
1498        struct raid1_info *p;
1499        int first = 0;
1500        int last = conf->raid_disks - 1;
1501        struct request_queue *q = bdev_get_queue(rdev->bdev);
1502
1503        if (mddev->recovery_disabled == conf->recovery_disabled)
1504                return -EBUSY;
1505
1506        if (rdev->raid_disk >= 0)
1507                first = last = rdev->raid_disk;
1508
1509        if (q->merge_bvec_fn) {
1510                set_bit(Unmerged, &rdev->flags);
1511                mddev->merge_check_needed = 1;
1512        }
1513
1514        for (mirror = first; mirror <= last; mirror++) {
1515                p = conf->mirrors+mirror;
1516                if (!p->rdev) {
1517
1518                        disk_stack_limits(mddev->gendisk, rdev->bdev,
1519                                          rdev->data_offset << 9);
1520
1521                        p->head_position = 0;
1522                        rdev->raid_disk = mirror;
1523                        err = 0;
1524                        /* As all devices are equivalent, we don't need a full recovery
1525                         * if this was recently any drive of the array
1526                         */
1527                        if (rdev->saved_raid_disk < 0)
1528                                conf->fullsync = 1;
1529                        rcu_assign_pointer(p->rdev, rdev);
1530                        break;
1531                }
1532                if (test_bit(WantReplacement, &p->rdev->flags) &&
1533                    p[conf->raid_disks].rdev == NULL) {
1534                        /* Add this device as a replacement */
1535                        clear_bit(In_sync, &rdev->flags);
1536                        set_bit(Replacement, &rdev->flags);
1537                        rdev->raid_disk = mirror;
1538                        err = 0;
1539                        conf->fullsync = 1;
1540                        rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1541                        break;
1542                }
1543        }
1544        if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1545                /* Some requests might not have seen this new
1546                 * merge_bvec_fn.  We must wait for them to complete
1547                 * before merging the device fully.
1548                 * First we make sure any code which has tested
1549                 * our function has submitted the request, then
1550                 * we wait for all outstanding requests to complete.
1551                 */
1552                synchronize_sched();
1553                raise_barrier(conf);
1554                lower_barrier(conf);
1555                clear_bit(Unmerged, &rdev->flags);
1556        }
1557        md_integrity_add_rdev(rdev, mddev);
1558        if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
1559                queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1560        print_conf(conf);
1561        return err;
1562}
1563
1564static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1565{
1566        struct r1conf *conf = mddev->private;
1567        int err = 0;
1568        int number = rdev->raid_disk;
1569        struct raid1_info *p = conf->mirrors + number;
1570
1571        if (rdev != p->rdev)
1572                p = conf->mirrors + conf->raid_disks + number;
1573
1574        print_conf(conf);
1575        if (rdev == p->rdev) {
1576                if (test_bit(In_sync, &rdev->flags) ||
1577                    atomic_read(&rdev->nr_pending)) {
1578                        err = -EBUSY;
1579                        goto abort;
1580                }
1581                /* Only remove non-faulty devices if recovery
1582                 * is not possible.
1583                 */
1584                if (!test_bit(Faulty, &rdev->flags) &&
1585                    mddev->recovery_disabled != conf->recovery_disabled &&
1586                    mddev->degraded < conf->raid_disks) {
1587                        err = -EBUSY;
1588                        goto abort;
1589                }
1590                p->rdev = NULL;
1591                synchronize_rcu();
1592                if (atomic_read(&rdev->nr_pending)) {
1593                        /* lost the race, try later */
1594                        err = -EBUSY;
1595                        p->rdev = rdev;
1596                        goto abort;
1597                } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1598                        /* We just removed a device that is being replaced.
1599                         * Move down the replacement.  We drain all IO before
1600                         * doing this to avoid confusion.
1601                         */
1602                        struct md_rdev *repl =
1603                                conf->mirrors[conf->raid_disks + number].rdev;
1604                        raise_barrier(conf);
1605                        clear_bit(Replacement, &repl->flags);
1606                        p->rdev = repl;
1607                        conf->mirrors[conf->raid_disks + number].rdev = NULL;
1608                        lower_barrier(conf);
1609                        clear_bit(WantReplacement, &rdev->flags);
1610                } else
1611                        clear_bit(WantReplacement, &rdev->flags);
1612                err = md_integrity_register(mddev);
1613        }
1614abort:
1615
1616        print_conf(conf);
1617        return err;
1618}
1619
1620
1621static void end_sync_read(struct bio *bio, int error)
1622{
1623        struct r1bio *r1_bio = bio->bi_private;
1624
1625        update_head_pos(r1_bio->read_disk, r1_bio);
1626
1627        /*
1628         * we have read a block, now it needs to be re-written,
1629         * or re-read if the read failed.
1630         * We don't do much here, just schedule handling by raid1d
1631         */
1632        if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1633                set_bit(R1BIO_Uptodate, &r1_bio->state);
1634
1635        if (atomic_dec_and_test(&r1_bio->remaining))
1636                reschedule_retry(r1_bio);
1637}
1638
1639static void end_sync_write(struct bio *bio, int error)
1640{
1641        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1642        struct r1bio *r1_bio = bio->bi_private;
1643        struct mddev *mddev = r1_bio->mddev;
1644        struct r1conf *conf = mddev->private;
1645        int mirror=0;
1646        sector_t first_bad;
1647        int bad_sectors;
1648
1649        mirror = find_bio_disk(r1_bio, bio);
1650
1651        if (!uptodate) {
1652                sector_t sync_blocks = 0;
1653                sector_t s = r1_bio->sector;
1654                long sectors_to_go = r1_bio->sectors;
1655                /* make sure these bits doesn't get cleared. */
1656                do {
1657                        bitmap_end_sync(mddev->bitmap, s,
1658                                        &sync_blocks, 1);
1659                        s += sync_blocks;
1660                        sectors_to_go -= sync_blocks;
1661                } while (sectors_to_go > 0);
1662                set_bit(WriteErrorSeen,
1663                        &conf->mirrors[mirror].rdev->flags);
1664                if (!test_and_set_bit(WantReplacement,
1665                                      &conf->mirrors[mirror].rdev->flags))
1666                        set_bit(MD_RECOVERY_NEEDED, &
1667                                mddev->recovery);
1668                set_bit(R1BIO_WriteError, &r1_bio->state);
1669        } else if (is_badblock(conf->mirrors[mirror].rdev,
1670                               r1_bio->sector,
1671                               r1_bio->sectors,
1672                               &first_bad, &bad_sectors) &&
1673                   !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1674                                r1_bio->sector,
1675                                r1_bio->sectors,
1676                                &first_bad, &bad_sectors)
1677                )
1678                set_bit(R1BIO_MadeGood, &r1_bio->state);
1679
1680        if (atomic_dec_and_test(&r1_bio->remaining)) {
1681                int s = r1_bio->sectors;
1682                if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1683                    test_bit(R1BIO_WriteError, &r1_bio->state))
1684                        reschedule_retry(r1_bio);
1685                else {
1686                        put_buf(r1_bio);
1687                        md_done_sync(mddev, s, uptodate);
1688                }
1689        }
1690}
1691
1692static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1693                            int sectors, struct page *page, int rw)
1694{
1695        if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1696                /* success */
1697                return 1;
1698        if (rw == WRITE) {
1699                set_bit(WriteErrorSeen, &rdev->flags);
1700                if (!test_and_set_bit(WantReplacement,
1701                                      &rdev->flags))
1702                        set_bit(MD_RECOVERY_NEEDED, &
1703                                rdev->mddev->recovery);
1704        }
1705        /* need to record an error - either for the block or the device */
1706        if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1707                md_error(rdev->mddev, rdev);
1708        return 0;
1709}
1710
1711static int fix_sync_read_error(struct r1bio *r1_bio)
1712{
1713        /* Try some synchronous reads of other devices to get
1714         * good data, much like with normal read errors.  Only
1715         * read into the pages we already have so we don't
1716         * need to re-issue the read request.
1717         * We don't need to freeze the array, because being in an
1718         * active sync request, there is no normal IO, and
1719         * no overlapping syncs.
1720         * We don't need to check is_badblock() again as we
1721         * made sure that anything with a bad block in range
1722         * will have bi_end_io clear.
1723         */
1724        struct mddev *mddev = r1_bio->mddev;
1725        struct r1conf *conf = mddev->private;
1726        struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1727        sector_t sect = r1_bio->sector;
1728        int sectors = r1_bio->sectors;
1729        int idx = 0;
1730
1731        while(sectors) {
1732                int s = sectors;
1733                int d = r1_bio->read_disk;
1734                int success = 0;
1735                struct md_rdev *rdev;
1736                int start;
1737
1738                if (s > (PAGE_SIZE>>9))
1739                        s = PAGE_SIZE >> 9;
1740                do {
1741                        if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1742                                /* No rcu protection needed here devices
1743                                 * can only be removed when no resync is
1744                                 * active, and resync is currently active
1745                                 */
1746                                rdev = conf->mirrors[d].rdev;
1747                                if (sync_page_io(rdev, sect, s<<9,
1748                                                 bio->bi_io_vec[idx].bv_page,
1749                                                 READ, false)) {
1750                                        success = 1;
1751                                        break;
1752                                }
1753                        }
1754                        d++;
1755                        if (d == conf->raid_disks * 2)
1756                                d = 0;
1757                } while (!success && d != r1_bio->read_disk);
1758
1759                if (!success) {
1760                        char b[BDEVNAME_SIZE];
1761                        int abort = 0;
1762                        /* Cannot read from anywhere, this block is lost.
1763                         * Record a bad block on each device.  If that doesn't
1764                         * work just disable and interrupt the recovery.
1765                         * Don't fail devices as that won't really help.
1766                         */
1767                        printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1768                               " for block %llu\n",
1769                               mdname(mddev),
1770                               bdevname(bio->bi_bdev, b),
1771                               (unsigned long long)r1_bio->sector);
1772                        for (d = 0; d < conf->raid_disks * 2; d++) {
1773                                rdev = conf->mirrors[d].rdev;
1774                                if (!rdev || test_bit(Faulty, &rdev->flags))
1775                                        continue;
1776                                if (!rdev_set_badblocks(rdev, sect, s, 0))
1777                                        abort = 1;
1778                        }
1779                        if (abort) {
1780                                conf->recovery_disabled =
1781                                        mddev->recovery_disabled;
1782                                set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1783                                md_done_sync(mddev, r1_bio->sectors, 0);
1784                                put_buf(r1_bio);
1785                                return 0;
1786                        }
1787                        /* Try next page */
1788                        sectors -= s;
1789                        sect += s;
1790                        idx++;
1791                        continue;
1792                }
1793
1794                start = d;
1795                /* write it back and re-read */
1796                while (d != r1_bio->read_disk) {
1797                        if (d == 0)
1798                                d = conf->raid_disks * 2;
1799                        d--;
1800                        if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1801                                continue;
1802                        rdev = conf->mirrors[d].rdev;
1803                        if (r1_sync_page_io(rdev, sect, s,
1804                                            bio->bi_io_vec[idx].bv_page,
1805                                            WRITE) == 0) {
1806                                r1_bio->bios[d]->bi_end_io = NULL;
1807                                rdev_dec_pending(rdev, mddev);
1808                        }
1809                }
1810                d = start;
1811                while (d != r1_bio->read_disk) {
1812                        if (d == 0)
1813                                d = conf->raid_disks * 2;
1814                        d--;
1815                        if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1816                                continue;
1817                        rdev = conf->mirrors[d].rdev;
1818                        if (r1_sync_page_io(rdev, sect, s,
1819                                            bio->bi_io_vec[idx].bv_page,
1820                                            READ) != 0)
1821                                atomic_add(s, &rdev->corrected_errors);
1822                }
1823                sectors -= s;
1824                sect += s;
1825                idx ++;
1826        }
1827        set_bit(R1BIO_Uptodate, &r1_bio->state);
1828        set_bit(BIO_UPTODATE, &bio->bi_flags);
1829        return 1;
1830}
1831
1832static int process_checks(struct r1bio *r1_bio)
1833{
1834        /* We have read all readable devices.  If we haven't
1835         * got the block, then there is no hope left.
1836         * If we have, then we want to do a comparison
1837         * and skip the write if everything is the same.
1838         * If any blocks failed to read, then we need to
1839         * attempt an over-write
1840         */
1841        struct mddev *mddev = r1_bio->mddev;
1842        struct r1conf *conf = mddev->private;
1843        int primary;
1844        int i;
1845        int vcnt;
1846
1847        for (primary = 0; primary < conf->raid_disks * 2; primary++)
1848                if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1849                    test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1850                        r1_bio->bios[primary]->bi_end_io = NULL;
1851                        rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1852                        break;
1853                }
1854        r1_bio->read_disk = primary;
1855        vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1856        for (i = 0; i < conf->raid_disks * 2; i++) {
1857                int j;
1858                struct bio *pbio = r1_bio->bios[primary];
1859                struct bio *sbio = r1_bio->bios[i];
1860                int size;
1861
1862                if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1863                        continue;
1864
1865                if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1866                        for (j = vcnt; j-- ; ) {
1867                                struct page *p, *s;
1868                                p = pbio->bi_io_vec[j].bv_page;
1869                                s = sbio->bi_io_vec[j].bv_page;
1870                                if (memcmp(page_address(p),
1871                                           page_address(s),
1872                                           sbio->bi_io_vec[j].bv_len))
1873                                        break;
1874                        }
1875                } else
1876                        j = 0;
1877                if (j >= 0)
1878                        atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
1879                if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1880                              && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1881                        /* No need to write to this device. */
1882                        sbio->bi_end_io = NULL;
1883                        rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1884                        continue;
1885                }
1886                /* fixup the bio for reuse */
1887                sbio->bi_vcnt = vcnt;
1888                sbio->bi_size = r1_bio->sectors << 9;
1889                sbio->bi_idx = 0;
1890                sbio->bi_phys_segments = 0;
1891                sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1892                sbio->bi_flags |= 1 << BIO_UPTODATE;
1893                sbio->bi_next = NULL;
1894                sbio->bi_sector = r1_bio->sector +
1895                        conf->mirrors[i].rdev->data_offset;
1896                sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1897                size = sbio->bi_size;
1898                for (j = 0; j < vcnt ; j++) {
1899                        struct bio_vec *bi;
1900                        bi = &sbio->bi_io_vec[j];
1901                        bi->bv_offset = 0;
1902                        if (size > PAGE_SIZE)
1903                                bi->bv_len = PAGE_SIZE;
1904                        else
1905                                bi->bv_len = size;
1906                        size -= PAGE_SIZE;
1907                        memcpy(page_address(bi->bv_page),
1908                               page_address(pbio->bi_io_vec[j].bv_page),
1909                               PAGE_SIZE);
1910                }
1911        }
1912        return 0;
1913}
1914
1915static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1916{
1917        struct r1conf *conf = mddev->private;
1918        int i;
1919        int disks = conf->raid_disks * 2;
1920        struct bio *bio, *wbio;
1921
1922        bio = r1_bio->bios[r1_bio->read_disk];
1923
1924        if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1925                /* ouch - failed to read all of that. */
1926                if (!fix_sync_read_error(r1_bio))
1927                        return;
1928
1929        if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1930                if (process_checks(r1_bio) < 0)
1931                        return;
1932        /*
1933         * schedule writes
1934         */
1935        atomic_set(&r1_bio->remaining, 1);
1936        for (i = 0; i < disks ; i++) {
1937                wbio = r1_bio->bios[i];
1938                if (wbio->bi_end_io == NULL ||
1939                    (wbio->bi_end_io == end_sync_read &&
1940                     (i == r1_bio->read_disk ||
1941                      !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1942                        continue;
1943
1944                wbio->bi_rw = WRITE;
1945                wbio->bi_end_io = end_sync_write;
1946                atomic_inc(&r1_bio->remaining);
1947                md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1948
1949                generic_make_request(wbio);
1950        }
1951
1952        if (atomic_dec_and_test(&r1_bio->remaining)) {
1953                /* if we're here, all write(s) have completed, so clean up */
1954                int s = r1_bio->sectors;
1955                if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1956                    test_bit(R1BIO_WriteError, &r1_bio->state))
1957                        reschedule_retry(r1_bio);
1958                else {
1959                        put_buf(r1_bio);
1960                        md_done_sync(mddev, s, 1);
1961                }
1962        }
1963}
1964
1965/*
1966 * This is a kernel thread which:
1967 *
1968 *      1.      Retries failed read operations on working mirrors.
1969 *      2.      Updates the raid superblock when problems encounter.
1970 *      3.      Performs writes following reads for array synchronising.
1971 */
1972
1973static void fix_read_error(struct r1conf *conf, int read_disk,
1974                           sector_t sect, int sectors)
1975{
1976        struct mddev *mddev = conf->mddev;
1977        while(sectors) {
1978                int s = sectors;
1979                int d = read_disk;
1980                int success = 0;
1981                int start;
1982                struct md_rdev *rdev;
1983
1984                if (s > (PAGE_SIZE>>9))
1985                        s = PAGE_SIZE >> 9;
1986
1987                do {
1988                        /* Note: no rcu protection needed here
1989                         * as this is synchronous in the raid1d thread
1990                         * which is the thread that might remove
1991                         * a device.  If raid1d ever becomes multi-threaded....
1992                         */
1993                        sector_t first_bad;
1994                        int bad_sectors;
1995
1996                        rdev = conf->mirrors[d].rdev;
1997                        if (rdev &&
1998                            (test_bit(In_sync, &rdev->flags) ||
1999                             (!test_bit(Faulty, &rdev->flags) &&
2000                              rdev->recovery_offset >= sect + s)) &&
2001                            is_badblock(rdev, sect, s,
2002                                        &first_bad, &bad_sectors) == 0 &&
2003                            sync_page_io(rdev, sect, s<<9,
2004                                         conf->tmppage, READ, false))
2005                                success = 1;
2006                        else {
2007                                d++;
2008                                if (d == conf->raid_disks * 2)
2009                                        d = 0;
2010                        }
2011                } while (!success && d != read_disk);
2012
2013                if (!success) {
2014                        /* Cannot read from anywhere - mark it bad */
2015                        struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2016                        if (!rdev_set_badblocks(rdev, sect, s, 0))
2017                                md_error(mddev, rdev);
2018                        break;
2019                }
2020                /* write it back and re-read */
2021                start = d;
2022                while (d != read_disk) {
2023                        if (d==0)
2024                                d = conf->raid_disks * 2;
2025                        d--;
2026                        rdev = conf->mirrors[d].rdev;
2027                        if (rdev &&
2028                            test_bit(In_sync, &rdev->flags))
2029                                r1_sync_page_io(rdev, sect, s,
2030                                                conf->tmppage, WRITE);
2031                }
2032                d = start;
2033                while (d != read_disk) {
2034                        char b[BDEVNAME_SIZE];
2035                        if (d==0)
2036                                d = conf->raid_disks * 2;
2037                        d--;
2038                        rdev = conf->mirrors[d].rdev;
2039                        if (rdev &&
2040                            test_bit(In_sync, &rdev->flags)) {
2041                                if (r1_sync_page_io(rdev, sect, s,
2042                                                    conf->tmppage, READ)) {
2043                                        atomic_add(s, &rdev->corrected_errors);
2044                                        printk(KERN_INFO
2045                                               "md/raid1:%s: read error corrected "
2046                                               "(%d sectors at %llu on %s)\n",
2047                                               mdname(mddev), s,
2048                                               (unsigned long long)(sect +
2049                                                   rdev->data_offset),
2050                                               bdevname(rdev->bdev, b));
2051                                }
2052                        }
2053                }
2054                sectors -= s;
2055                sect += s;
2056        }
2057}
2058
2059static void bi_complete(struct bio *bio, int error)
2060{
2061        complete((struct completion *)bio->bi_private);
2062}
2063
2064static int submit_bio_wait(int rw, struct bio *bio)
2065{
2066        struct completion event;
2067        rw |= REQ_SYNC;
2068
2069        init_completion(&event);
2070        bio->bi_private = &event;
2071        bio->bi_end_io = bi_complete;
2072        submit_bio(rw, bio);
2073        wait_for_completion(&event);
2074
2075        return test_bit(BIO_UPTODATE, &bio->bi_flags);
2076}
2077
2078static int narrow_write_error(struct r1bio *r1_bio, int i)
2079{
2080        struct mddev *mddev = r1_bio->mddev;
2081        struct r1conf *conf = mddev->private;
2082        struct md_rdev *rdev = conf->mirrors[i].rdev;
2083        int vcnt, idx;
2084        struct bio_vec *vec;
2085
2086        /* bio has the data to be written to device 'i' where
2087         * we just recently had a write error.
2088         * We repeatedly clone the bio and trim down to one block,
2089         * then try the write.  Where the write fails we record
2090         * a bad block.
2091         * It is conceivable that the bio doesn't exactly align with
2092         * blocks.  We must handle this somehow.
2093         *
2094         * We currently own a reference on the rdev.
2095         */
2096
2097        int block_sectors;
2098        sector_t sector;
2099        int sectors;
2100        int sect_to_write = r1_bio->sectors;
2101        int ok = 1;
2102
2103        if (rdev->badblocks.shift < 0)
2104                return 0;
2105
2106        block_sectors = 1 << rdev->badblocks.shift;
2107        sector = r1_bio->sector;
2108        sectors = ((sector + block_sectors)
2109                   & ~(sector_t)(block_sectors - 1))
2110                - sector;
2111
2112        if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2113                vcnt = r1_bio->behind_page_count;
2114                vec = r1_bio->behind_bvecs;
2115                idx = 0;
2116                while (vec[idx].bv_page == NULL)
2117                        idx++;
2118        } else {
2119                vcnt = r1_bio->master_bio->bi_vcnt;
2120                vec = r1_bio->master_bio->bi_io_vec;
2121                idx = r1_bio->master_bio->bi_idx;
2122        }
2123        while (sect_to_write) {
2124                struct bio *wbio;
2125                if (sectors > sect_to_write)
2126                        sectors = sect_to_write;
2127                /* Write at 'sector' for 'sectors'*/
2128
2129                wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2130                memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2131                wbio->bi_sector = r1_bio->sector;
2132                wbio->bi_rw = WRITE;
2133                wbio->bi_vcnt = vcnt;
2134                wbio->bi_size = r1_bio->sectors << 9;
2135                wbio->bi_idx = idx;
2136
2137                md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2138                wbio->bi_sector += rdev->data_offset;
2139                wbio->bi_bdev = rdev->bdev;
2140                if (submit_bio_wait(WRITE, wbio) == 0)
2141                        /* failure! */
2142                        ok = rdev_set_badblocks(rdev, sector,
2143                                                sectors, 0)
2144                                && ok;
2145
2146                bio_put(wbio);
2147                sect_to_write -= sectors;
2148                sector += sectors;
2149                sectors = block_sectors;
2150        }
2151        return ok;
2152}
2153
2154static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2155{
2156        int m;
2157        int s = r1_bio->sectors;
2158        for (m = 0; m < conf->raid_disks * 2 ; m++) {
2159                struct md_rdev *rdev = conf->mirrors[m].rdev;
2160                struct bio *bio = r1_bio->bios[m];
2161                if (bio->bi_end_io == NULL)
2162                        continue;
2163                if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2164                    test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2165                        rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2166                }
2167                if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2168                    test_bit(R1BIO_WriteError, &r1_bio->state)) {
2169                        if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2170                                md_error(conf->mddev, rdev);
2171                }
2172        }
2173        put_buf(r1_bio);
2174        md_done_sync(conf->mddev, s, 1);
2175}
2176
2177static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2178{
2179        int m;
2180        for (m = 0; m < conf->raid_disks * 2 ; m++)
2181                if (r1_bio->bios[m] == IO_MADE_GOOD) {
2182                        struct md_rdev *rdev = conf->mirrors[m].rdev;
2183                        rdev_clear_badblocks(rdev,
2184                                             r1_bio->sector,
2185                                             r1_bio->sectors, 0);
2186                        rdev_dec_pending(rdev, conf->mddev);
2187                } else if (r1_bio->bios[m] != NULL) {
2188                        /* This drive got a write error.  We need to
2189                         * narrow down and record precise write
2190                         * errors.
2191                         */
2192                        if (!narrow_write_error(r1_bio, m)) {
2193                                md_error(conf->mddev,
2194                                         conf->mirrors[m].rdev);
2195                                /* an I/O failed, we can't clear the bitmap */
2196                                set_bit(R1BIO_Degraded, &r1_bio->state);
2197                        }
2198                        rdev_dec_pending(conf->mirrors[m].rdev,
2199                                         conf->mddev);
2200                }
2201        if (test_bit(R1BIO_WriteError, &r1_bio->state))
2202                close_write(r1_bio);
2203        raid_end_bio_io(r1_bio);
2204}
2205
2206static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2207{
2208        int disk;
2209        int max_sectors;
2210        struct mddev *mddev = conf->mddev;
2211        struct bio *bio;
2212        char b[BDEVNAME_SIZE];
2213        struct md_rdev *rdev;
2214
2215        clear_bit(R1BIO_ReadError, &r1_bio->state);
2216        /* we got a read error. Maybe the drive is bad.  Maybe just
2217         * the block and we can fix it.
2218         * We freeze all other IO, and try reading the block from
2219         * other devices.  When we find one, we re-write
2220         * and check it that fixes the read error.
2221         * This is all done synchronously while the array is
2222         * frozen
2223         */
2224        if (mddev->ro == 0) {
2225                freeze_array(conf);
2226                fix_read_error(conf, r1_bio->read_disk,
2227                               r1_bio->sector, r1_bio->sectors);
2228                unfreeze_array(conf);
2229        } else
2230                md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2231        rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2232
2233        bio = r1_bio->bios[r1_bio->read_disk];
2234        bdevname(bio->bi_bdev, b);
2235read_more:
2236        disk = read_balance(conf, r1_bio, &max_sectors);
2237        if (disk == -1) {
2238                printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2239                       " read error for block %llu\n",
2240                       mdname(mddev), b, (unsigned long long)r1_bio->sector);
2241                raid_end_bio_io(r1_bio);
2242        } else {
2243                const unsigned long do_sync
2244                        = r1_bio->master_bio->bi_rw & REQ_SYNC;
2245                if (bio) {
2246                        r1_bio->bios[r1_bio->read_disk] =
2247                                mddev->ro ? IO_BLOCKED : NULL;
2248                        bio_put(bio);
2249                }
2250                r1_bio->read_disk = disk;
2251                bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2252                md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2253                r1_bio->bios[r1_bio->read_disk] = bio;
2254                rdev = conf->mirrors[disk].rdev;
2255                printk_ratelimited(KERN_ERR
2256                                   "md/raid1:%s: redirecting sector %llu"
2257                                   " to other mirror: %s\n",
2258                                   mdname(mddev),
2259                                   (unsigned long long)r1_bio->sector,
2260                                   bdevname(rdev->bdev, b));
2261                bio->bi_sector = r1_bio->sector + rdev->data_offset;
2262                bio->bi_bdev = rdev->bdev;
2263                bio->bi_end_io = raid1_end_read_request;
2264                bio->bi_rw = READ | do_sync;
2265                bio->bi_private = r1_bio;
2266                if (max_sectors < r1_bio->sectors) {
2267                        /* Drat - have to split this up more */
2268                        struct bio *mbio = r1_bio->master_bio;
2269                        int sectors_handled = (r1_bio->sector + max_sectors
2270                                               - mbio->bi_sector);
2271                        r1_bio->sectors = max_sectors;
2272                        spin_lock_irq(&conf->device_lock);
2273                        if (mbio->bi_phys_segments == 0)
2274                                mbio->bi_phys_segments = 2;
2275                        else
2276                                mbio->bi_phys_segments++;
2277                        spin_unlock_irq(&conf->device_lock);
2278                        generic_make_request(bio);
2279                        bio = NULL;
2280
2281                        r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2282
2283                        r1_bio->master_bio = mbio;
2284                        r1_bio->sectors = (mbio->bi_size >> 9)
2285                                          - sectors_handled;
2286                        r1_bio->state = 0;
2287                        set_bit(R1BIO_ReadError, &r1_bio->state);
2288                        r1_bio->mddev = mddev;
2289                        r1_bio->sector = mbio->bi_sector + sectors_handled;
2290
2291                        goto read_more;
2292                } else
2293                        generic_make_request(bio);
2294        }
2295}
2296
2297static void raid1d(struct md_thread *thread)
2298{
2299        struct mddev *mddev = thread->mddev;
2300        struct r1bio *r1_bio;
2301        unsigned long flags;
2302        struct r1conf *conf = mddev->private;
2303        struct list_head *head = &conf->retry_list;
2304        struct blk_plug plug;
2305
2306        md_check_recovery(mddev);
2307
2308        blk_start_plug(&plug);
2309        for (;;) {
2310
2311                flush_pending_writes(conf);
2312
2313                spin_lock_irqsave(&conf->device_lock, flags);
2314                if (list_empty(head)) {
2315                        spin_unlock_irqrestore(&conf->device_lock, flags);
2316                        break;
2317                }
2318                r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2319                list_del(head->prev);
2320                conf->nr_queued--;
2321                spin_unlock_irqrestore(&conf->device_lock, flags);
2322
2323                mddev = r1_bio->mddev;
2324                conf = mddev->private;
2325                if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2326                        if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2327                            test_bit(R1BIO_WriteError, &r1_bio->state))
2328                                handle_sync_write_finished(conf, r1_bio);
2329                        else
2330                                sync_request_write(mddev, r1_bio);
2331                } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2332                           test_bit(R1BIO_WriteError, &r1_bio->state))
2333                        handle_write_finished(conf, r1_bio);
2334                else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2335                        handle_read_error(conf, r1_bio);
2336                else
2337                        /* just a partial read to be scheduled from separate
2338                         * context
2339                         */
2340                        generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2341
2342                cond_resched();
2343                if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2344                        md_check_recovery(mddev);
2345        }
2346        blk_finish_plug(&plug);
2347}
2348
2349
2350static int init_resync(struct r1conf *conf)
2351{
2352        int buffs;
2353
2354        buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2355        BUG_ON(conf->r1buf_pool);
2356        conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2357                                          conf->poolinfo);
2358        if (!conf->r1buf_pool)
2359                return -ENOMEM;
2360        conf->next_resync = 0;
2361        return 0;
2362}
2363
2364/*
2365 * perform a "sync" on one "block"
2366 *
2367 * We need to make sure that no normal I/O request - particularly write
2368 * requests - conflict with active sync requests.
2369 *
2370 * This is achieved by tracking pending requests and a 'barrier' concept
2371 * that can be installed to exclude normal IO requests.
2372 */
2373
2374static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2375{
2376        struct r1conf *conf = mddev->private;
2377        struct r1bio *r1_bio;
2378        struct bio *bio;
2379        sector_t max_sector, nr_sectors;
2380        int disk = -1;
2381        int i;
2382        int wonly = -1;
2383        int write_targets = 0, read_targets = 0;
2384        sector_t sync_blocks;
2385        int still_degraded = 0;
2386        int good_sectors = RESYNC_SECTORS;
2387        int min_bad = 0; /* number of sectors that are bad in all devices */
2388
2389        if (!conf->r1buf_pool)
2390                if (init_resync(conf))
2391                        return 0;
2392
2393        max_sector = mddev->dev_sectors;
2394        if (sector_nr >= max_sector) {
2395                /* If we aborted, we need to abort the
2396                 * sync on the 'current' bitmap chunk (there will
2397                 * only be one in raid1 resync.
2398                 * We can find the current addess in mddev->curr_resync
2399                 */
2400                if (mddev->curr_resync < max_sector) /* aborted */
2401                        bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2402                                                &sync_blocks, 1);
2403                else /* completed sync */
2404                        conf->fullsync = 0;
2405
2406                bitmap_close_sync(mddev->bitmap);
2407                close_sync(conf);
2408                return 0;
2409        }
2410
2411        if (mddev->bitmap == NULL &&
2412            mddev->recovery_cp == MaxSector &&
2413            !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2414            conf->fullsync == 0) {
2415                *skipped = 1;
2416                return max_sector - sector_nr;
2417        }
2418        /* before building a request, check if we can skip these blocks..
2419         * This call the bitmap_start_sync doesn't actually record anything
2420         */
2421        if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2422            !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2423                /* We can skip this block, and probably several more */
2424                *skipped = 1;
2425                return sync_blocks;
2426        }
2427        /*
2428         * If there is non-resync activity waiting for a turn,
2429         * and resync is going fast enough,
2430         * then let it though before starting on this new sync request.
2431         */
2432        if (!go_faster && conf->nr_waiting)
2433                msleep_interruptible(1000);
2434
2435        bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2436        r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2437        raise_barrier(conf);
2438
2439        conf->next_resync = sector_nr;
2440
2441        rcu_read_lock();
2442        /*
2443         * If we get a correctably read error during resync or recovery,
2444         * we might want to read from a different device.  So we
2445         * flag all drives that could conceivably be read from for READ,
2446         * and any others (which will be non-In_sync devices) for WRITE.
2447         * If a read fails, we try reading from something else for which READ
2448         * is OK.
2449         */
2450
2451        r1_bio->mddev = mddev;
2452        r1_bio->sector = sector_nr;
2453        r1_bio->state = 0;
2454        set_bit(R1BIO_IsSync, &r1_bio->state);
2455
2456        for (i = 0; i < conf->raid_disks * 2; i++) {
2457                struct md_rdev *rdev;
2458                bio = r1_bio->bios[i];
2459
2460                /* take from bio_init */
2461                bio->bi_next = NULL;
2462                bio->bi_flags &= ~(BIO_POOL_MASK-1);
2463                bio->bi_flags |= 1 << BIO_UPTODATE;
2464                bio->bi_rw = READ;
2465                bio->bi_vcnt = 0;
2466                bio->bi_idx = 0;
2467                bio->bi_phys_segments = 0;
2468                bio->bi_size = 0;
2469                bio->bi_end_io = NULL;
2470                bio->bi_private = NULL;
2471
2472                rdev = rcu_dereference(conf->mirrors[i].rdev);
2473                if (rdev == NULL ||
2474                    test_bit(Faulty, &rdev->flags)) {
2475                        if (i < conf->raid_disks)
2476                                still_degraded = 1;
2477                } else if (!test_bit(In_sync, &rdev->flags)) {
2478                        bio->bi_rw = WRITE;
2479                        bio->bi_end_io = end_sync_write;
2480                        write_targets ++;
2481                } else {
2482                        /* may need to read from here */
2483                        sector_t first_bad = MaxSector;
2484                        int bad_sectors;
2485
2486                        if (is_badblock(rdev, sector_nr, good_sectors,
2487                                        &first_bad, &bad_sectors)) {
2488                                if (first_bad > sector_nr)
2489                                        good_sectors = first_bad - sector_nr;
2490                                else {
2491                                        bad_sectors -= (sector_nr - first_bad);
2492                                        if (min_bad == 0 ||
2493                                            min_bad > bad_sectors)
2494                                                min_bad = bad_sectors;
2495                                }
2496                        }
2497                        if (sector_nr < first_bad) {
2498                                if (test_bit(WriteMostly, &rdev->flags)) {
2499                                        if (wonly < 0)
2500                                                wonly = i;
2501                                } else {
2502                                        if (disk < 0)
2503                                                disk = i;
2504                                }
2505                                bio->bi_rw = READ;
2506                                bio->bi_end_io = end_sync_read;
2507                                read_targets++;
2508                        } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2509                                test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2510                                !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2511                                /*
2512                                 * The device is suitable for reading (InSync),
2513                                 * but has bad block(s) here. Let's try to correct them,
2514                                 * if we are doing resync or repair. Otherwise, leave
2515                                 * this device alone for this sync request.
2516                                 */
2517                                bio->bi_rw = WRITE;
2518                                bio->bi_end_io = end_sync_write;
2519                                write_targets++;
2520                        }
2521                }
2522                if (bio->bi_end_io) {
2523                        atomic_inc(&rdev->nr_pending);
2524                        bio->bi_sector = sector_nr + rdev->data_offset;
2525                        bio->bi_bdev = rdev->bdev;
2526                        bio->bi_private = r1_bio;
2527                }
2528        }
2529        rcu_read_unlock();
2530        if (disk < 0)
2531                disk = wonly;
2532        r1_bio->read_disk = disk;
2533
2534        if (read_targets == 0 && min_bad > 0) {
2535                /* These sectors are bad on all InSync devices, so we
2536                 * need to mark them bad on all write targets
2537                 */
2538                int ok = 1;
2539                for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2540                        if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2541                                struct md_rdev *rdev = conf->mirrors[i].rdev;
2542                                ok = rdev_set_badblocks(rdev, sector_nr,
2543                                                        min_bad, 0
2544                                        ) && ok;
2545                        }
2546                set_bit(MD_CHANGE_DEVS, &mddev->flags);
2547                *skipped = 1;
2548                put_buf(r1_bio);
2549
2550                if (!ok) {
2551                        /* Cannot record the badblocks, so need to
2552                         * abort the resync.
2553                         * If there are multiple read targets, could just
2554                         * fail the really bad ones ???
2555                         */
2556                        conf->recovery_disabled = mddev->recovery_disabled;
2557                        set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2558                        return 0;
2559                } else
2560                        return min_bad;
2561
2562        }
2563        if (min_bad > 0 && min_bad < good_sectors) {
2564                /* only resync enough to reach the next bad->good
2565                 * transition */
2566                good_sectors = min_bad;
2567        }
2568
2569        if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2570                /* extra read targets are also write targets */
2571                write_targets += read_targets-1;
2572
2573        if (write_targets == 0 || read_targets == 0) {
2574                /* There is nowhere to write, so all non-sync
2575                 * drives must be failed - so we are finished
2576                 */
2577                sector_t rv;
2578                if (min_bad > 0)
2579                        max_sector = sector_nr + min_bad;
2580                rv = max_sector - sector_nr;
2581                *skipped = 1;
2582                put_buf(r1_bio);
2583                return rv;
2584        }
2585
2586        if (max_sector > mddev->resync_max)
2587                max_sector = mddev->resync_max; /* Don't do IO beyond here */
2588        if (max_sector > sector_nr + good_sectors)
2589                max_sector = sector_nr + good_sectors;
2590        nr_sectors = 0;
2591        sync_blocks = 0;
2592        do {
2593                struct page *page;
2594                int len = PAGE_SIZE;
2595                if (sector_nr + (len>>9) > max_sector)
2596                        len = (max_sector - sector_nr) << 9;
2597                if (len == 0)
2598                        break;
2599                if (sync_blocks == 0) {
2600                        if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2601                                               &sync_blocks, still_degraded) &&
2602                            !conf->fullsync &&
2603                            !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2604                                break;
2605                        BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2606                        if ((len >> 9) > sync_blocks)
2607                                len = sync_blocks<<9;
2608                }
2609
2610                for (i = 0 ; i < conf->raid_disks * 2; i++) {
2611                        bio = r1_bio->bios[i];
2612                        if (bio->bi_end_io) {
2613                                page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2614                                if (bio_add_page(bio, page, len, 0) == 0) {
2615                                        /* stop here */
2616                                        bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2617                                        while (i > 0) {
2618                                                i--;
2619                                                bio = r1_bio->bios[i];
2620                                                if (bio->bi_end_io==NULL)
2621                                                        continue;
2622                                                /* remove last page from this bio */
2623                                                bio->bi_vcnt--;
2624                                                bio->bi_size -= len;
2625                                                bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2626                                        }
2627                                        goto bio_full;
2628                                }
2629                        }
2630                }
2631                nr_sectors += len>>9;
2632                sector_nr += len>>9;
2633                sync_blocks -= (len>>9);
2634        } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2635 bio_full:
2636        r1_bio->sectors = nr_sectors;
2637
2638        /* For a user-requested sync, we read all readable devices and do a
2639         * compare
2640         */
2641        if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2642                atomic_set(&r1_bio->remaining, read_targets);
2643                for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2644                        bio = r1_bio->bios[i];
2645                        if (bio->bi_end_io == end_sync_read) {
2646                                read_targets--;
2647                                md_sync_acct(bio->bi_bdev, nr_sectors);
2648                                generic_make_request(bio);
2649                        }
2650                }
2651        } else {
2652                atomic_set(&r1_bio->remaining, 1);
2653                bio = r1_bio->bios[r1_bio->read_disk];
2654                md_sync_acct(bio->bi_bdev, nr_sectors);
2655                generic_make_request(bio);
2656
2657        }
2658        return nr_sectors;
2659}
2660
2661static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2662{
2663        if (sectors)
2664                return sectors;
2665
2666        return mddev->dev_sectors;
2667}
2668
2669static struct r1conf *setup_conf(struct mddev *mddev)
2670{
2671        struct r1conf *conf;
2672        int i;
2673        struct raid1_info *disk;
2674        struct md_rdev *rdev;
2675        int err = -ENOMEM;
2676
2677        conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2678        if (!conf)
2679                goto abort;
2680
2681        conf->mirrors = kzalloc(sizeof(struct raid1_info)
2682                                * mddev->raid_disks * 2,
2683                                 GFP_KERNEL);
2684        if (!conf->mirrors)
2685                goto abort;
2686
2687        conf->tmppage = alloc_page(GFP_KERNEL);
2688        if (!conf->tmppage)
2689                goto abort;
2690
2691        conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2692        if (!conf->poolinfo)
2693                goto abort;
2694        conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2695        conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2696                                          r1bio_pool_free,
2697                                          conf->poolinfo);
2698        if (!conf->r1bio_pool)
2699                goto abort;
2700
2701        conf->poolinfo->mddev = mddev;
2702
2703        err = -EINVAL;
2704        spin_lock_init(&conf->device_lock);
2705        rdev_for_each(rdev, mddev) {
2706                struct request_queue *q;
2707                int disk_idx = rdev->raid_disk;
2708                if (disk_idx >= mddev->raid_disks
2709                    || disk_idx < 0)
2710                        continue;
2711                if (test_bit(Replacement, &rdev->flags))
2712                        disk = conf->mirrors + mddev->raid_disks + disk_idx;
2713                else
2714                        disk = conf->mirrors + disk_idx;
2715
2716                if (disk->rdev)
2717                        goto abort;
2718                disk->rdev = rdev;
2719                q = bdev_get_queue(rdev->bdev);
2720                if (q->merge_bvec_fn)
2721                        mddev->merge_check_needed = 1;
2722
2723                disk->head_position = 0;
2724                disk->seq_start = MaxSector;
2725        }
2726        conf->raid_disks = mddev->raid_disks;
2727        conf->mddev = mddev;
2728        INIT_LIST_HEAD(&conf->retry_list);
2729
2730        spin_lock_init(&conf->resync_lock);
2731        init_waitqueue_head(&conf->wait_barrier);
2732
2733        bio_list_init(&conf->pending_bio_list);
2734        conf->pending_count = 0;
2735        conf->recovery_disabled = mddev->recovery_disabled - 1;
2736
2737        err = -EIO;
2738        for (i = 0; i < conf->raid_disks * 2; i++) {
2739
2740                disk = conf->mirrors + i;
2741
2742                if (i < conf->raid_disks &&
2743                    disk[conf->raid_disks].rdev) {
2744                        /* This slot has a replacement. */
2745                        if (!disk->rdev) {
2746                                /* No original, just make the replacement
2747                                 * a recovering spare
2748                                 */
2749                                disk->rdev =
2750                                        disk[conf->raid_disks].rdev;
2751                                disk[conf->raid_disks].rdev = NULL;
2752                        } else if (!test_bit(In_sync, &disk->rdev->flags))
2753                                /* Original is not in_sync - bad */
2754                                goto abort;
2755                }
2756
2757                if (!disk->rdev ||
2758                    !test_bit(In_sync, &disk->rdev->flags)) {
2759                        disk->head_position = 0;
2760                        if (disk->rdev &&
2761                            (disk->rdev->saved_raid_disk < 0))
2762                                conf->fullsync = 1;
2763                }
2764        }
2765
2766        err = -ENOMEM;
2767        conf->thread = md_register_thread(raid1d, mddev, "raid1");
2768        if (!conf->thread) {
2769                printk(KERN_ERR
2770                       "md/raid1:%s: couldn't allocate thread\n",
2771                       mdname(mddev));
2772                goto abort;
2773        }
2774
2775        return conf;
2776
2777 abort:
2778        if (conf) {
2779                if (conf->r1bio_pool)
2780                        mempool_destroy(conf->r1bio_pool);
2781                kfree(conf->mirrors);
2782                safe_put_page(conf->tmppage);
2783                kfree(conf->poolinfo);
2784                kfree(conf);
2785        }
2786        return ERR_PTR(err);
2787}
2788
2789static int stop(struct mddev *mddev);
2790static int run(struct mddev *mddev)
2791{
2792        struct r1conf *conf;
2793        int i;
2794        struct md_rdev *rdev;
2795        int ret;
2796        bool discard_supported = false;
2797
2798        if (mddev->level != 1) {
2799                printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2800                       mdname(mddev), mddev->level);
2801                return -EIO;
2802        }
2803        if (mddev->reshape_position != MaxSector) {
2804                printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2805                       mdname(mddev));
2806                return -EIO;
2807        }
2808        /*
2809         * copy the already verified devices into our private RAID1
2810         * bookkeeping area. [whatever we allocate in run(),
2811         * should be freed in stop()]
2812         */
2813        if (mddev->private == NULL)
2814                conf = setup_conf(mddev);
2815        else
2816                conf = mddev->private;
2817
2818        if (IS_ERR(conf))
2819                return PTR_ERR(conf);
2820
2821        rdev_for_each(rdev, mddev) {
2822                if (!mddev->gendisk)
2823                        continue;
2824                disk_stack_limits(mddev->gendisk, rdev->bdev,
2825                                  rdev->data_offset << 9);
2826                if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2827                        discard_supported = true;
2828        }
2829
2830        mddev->degraded = 0;
2831        for (i=0; i < conf->raid_disks; i++)
2832                if (conf->mirrors[i].rdev == NULL ||
2833                    !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2834                    test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2835                        mddev->degraded++;
2836
2837        if (conf->raid_disks - mddev->degraded == 1)
2838                mddev->recovery_cp = MaxSector;
2839
2840        if (mddev->recovery_cp != MaxSector)
2841                printk(KERN_NOTICE "md/raid1:%s: not clean"
2842                       " -- starting background reconstruction\n",
2843                       mdname(mddev));
2844        printk(KERN_INFO 
2845                "md/raid1:%s: active with %d out of %d mirrors\n",
2846                mdname(mddev), mddev->raid_disks - mddev->degraded, 
2847                mddev->raid_disks);
2848
2849        /*
2850         * Ok, everything is just fine now
2851         */
2852        mddev->thread = conf->thread;
2853        conf->thread = NULL;
2854        mddev->private = conf;
2855
2856        md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2857
2858        if (mddev->queue) {
2859                mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2860                mddev->queue->backing_dev_info.congested_data = mddev;
2861                blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2862
2863                if (discard_supported)
2864                        queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2865                                                mddev->queue);
2866                else
2867                        queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2868                                                  mddev->queue);
2869        }
2870
2871        ret =  md_integrity_register(mddev);
2872        if (ret)
2873                stop(mddev);
2874        return ret;
2875}
2876
2877static int stop(struct mddev *mddev)
2878{
2879        struct r1conf *conf = mddev->private;
2880        struct bitmap *bitmap = mddev->bitmap;
2881
2882        /* wait for behind writes to complete */
2883        if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2884                printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2885                       mdname(mddev));
2886                /* need to kick something here to make sure I/O goes? */
2887                wait_event(bitmap->behind_wait,
2888                           atomic_read(&bitmap->behind_writes) == 0);
2889        }
2890
2891        raise_barrier(conf);
2892        lower_barrier(conf);
2893
2894        md_unregister_thread(&mddev->thread);
2895        if (conf->r1bio_pool)
2896                mempool_destroy(conf->r1bio_pool);
2897        kfree(conf->mirrors);
2898        kfree(conf->poolinfo);
2899        kfree(conf);
2900        mddev->private = NULL;
2901        return 0;
2902}
2903
2904static int raid1_resize(struct mddev *mddev, sector_t sectors)
2905{
2906        /* no resync is happening, and there is enough space
2907         * on all devices, so we can resize.
2908         * We need to make sure resync covers any new space.
2909         * If the array is shrinking we should possibly wait until
2910         * any io in the removed space completes, but it hardly seems
2911         * worth it.
2912         */
2913        sector_t newsize = raid1_size(mddev, sectors, 0);
2914        if (mddev->external_size &&
2915            mddev->array_sectors > newsize)
2916                return -EINVAL;
2917        if (mddev->bitmap) {
2918                int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2919                if (ret)
2920                        return ret;
2921        }
2922        md_set_array_sectors(mddev, newsize);
2923        set_capacity(mddev->gendisk, mddev->array_sectors);
2924        revalidate_disk(mddev->gendisk);
2925        if (sectors > mddev->dev_sectors &&
2926            mddev->recovery_cp > mddev->dev_sectors) {
2927                mddev->recovery_cp = mddev->dev_sectors;
2928                set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2929        }
2930        mddev->dev_sectors = sectors;
2931        mddev->resync_max_sectors = sectors;
2932        return 0;
2933}
2934
2935static int raid1_reshape(struct mddev *mddev)
2936{
2937        /* We need to:
2938         * 1/ resize the r1bio_pool
2939         * 2/ resize conf->mirrors
2940         *
2941         * We allocate a new r1bio_pool if we can.
2942         * Then raise a device barrier and wait until all IO stops.
2943         * Then resize conf->mirrors and swap in the new r1bio pool.
2944         *
2945         * At the same time, we "pack" the devices so that all the missing
2946         * devices have the higher raid_disk numbers.
2947         */
2948        mempool_t *newpool, *oldpool;
2949        struct pool_info *newpoolinfo;
2950        struct raid1_info *newmirrors;
2951        struct r1conf *conf = mddev->private;
2952        int cnt, raid_disks;
2953        unsigned long flags;
2954        int d, d2, err;
2955
2956        /* Cannot change chunk_size, layout, or level */
2957        if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2958            mddev->layout != mddev->new_layout ||
2959            mddev->level != mddev->new_level) {
2960                mddev->new_chunk_sectors = mddev->chunk_sectors;
2961                mddev->new_layout = mddev->layout;
2962                mddev->new_level = mddev->level;
2963                return -EINVAL;
2964        }
2965
2966        err = md_allow_write(mddev);
2967        if (err)
2968                return err;
2969
2970        raid_disks = mddev->raid_disks + mddev->delta_disks;
2971
2972        if (raid_disks < conf->raid_disks) {
2973                cnt=0;
2974                for (d= 0; d < conf->raid_disks; d++)
2975                        if (conf->mirrors[d].rdev)
2976                                cnt++;
2977                if (cnt > raid_disks)
2978                        return -EBUSY;
2979        }
2980
2981        newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2982        if (!newpoolinfo)
2983                return -ENOMEM;
2984        newpoolinfo->mddev = mddev;
2985        newpoolinfo->raid_disks = raid_disks * 2;
2986
2987        newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2988                                 r1bio_pool_free, newpoolinfo);
2989        if (!newpool) {
2990                kfree(newpoolinfo);
2991                return -ENOMEM;
2992        }
2993        newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
2994                             GFP_KERNEL);
2995        if (!newmirrors) {
2996                kfree(newpoolinfo);
2997                mempool_destroy(newpool);
2998                return -ENOMEM;
2999        }
3000
3001        raise_barrier(conf);
3002
3003        /* ok, everything is stopped */
3004        oldpool = conf->r1bio_pool;
3005        conf->r1bio_pool = newpool;
3006
3007        for (d = d2 = 0; d < conf->raid_disks; d++) {
3008                struct md_rdev *rdev = conf->mirrors[d].rdev;
3009                if (rdev && rdev->raid_disk != d2) {
3010                        sysfs_unlink_rdev(mddev, rdev);
3011                        rdev->raid_disk = d2;
3012                        sysfs_unlink_rdev(mddev, rdev);
3013                        if (sysfs_link_rdev(mddev, rdev))
3014                                printk(KERN_WARNING
3015                                       "md/raid1:%s: cannot register rd%d\n",
3016                                       mdname(mddev), rdev->raid_disk);
3017                }
3018                if (rdev)
3019                        newmirrors[d2++].rdev = rdev;
3020        }
3021        kfree(conf->mirrors);
3022        conf->mirrors = newmirrors;
3023        kfree(conf->poolinfo);
3024        conf->poolinfo = newpoolinfo;
3025
3026        spin_lock_irqsave(&conf->device_lock, flags);
3027        mddev->degraded += (raid_disks - conf->raid_disks);
3028        spin_unlock_irqrestore(&conf->device_lock, flags);
3029        conf->raid_disks = mddev->raid_disks = raid_disks;
3030        mddev->delta_disks = 0;
3031
3032        lower_barrier(conf);
3033
3034        set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3035        md_wakeup_thread(mddev->thread);
3036
3037        mempool_destroy(oldpool);
3038        return 0;
3039}
3040
3041static void raid1_quiesce(struct mddev *mddev, int state)
3042{
3043        struct r1conf *conf = mddev->private;
3044
3045        switch(state) {
3046        case 2: /* wake for suspend */
3047                wake_up(&conf->wait_barrier);
3048                break;
3049        case 1:
3050                raise_barrier(conf);
3051                break;
3052        case 0:
3053                lower_barrier(conf);
3054                break;
3055        }
3056}
3057
3058static void *raid1_takeover(struct mddev *mddev)
3059{
3060        /* raid1 can take over:
3061         *  raid5 with 2 devices, any layout or chunk size
3062         */
3063        if (mddev->level == 5 && mddev->raid_disks == 2) {
3064                struct r1conf *conf;
3065                mddev->new_level = 1;
3066                mddev->new_layout = 0;
3067                mddev->new_chunk_sectors = 0;
3068                conf = setup_conf(mddev);
3069                if (!IS_ERR(conf))
3070                        conf->barrier = 1;
3071                return conf;
3072        }
3073        return ERR_PTR(-EINVAL);
3074}
3075
3076static struct md_personality raid1_personality =
3077{
3078        .name           = "raid1",
3079        .level          = 1,
3080        .owner          = THIS_MODULE,
3081        .make_request   = make_request,
3082        .run            = run,
3083        .stop           = stop,
3084        .status         = status,
3085        .error_handler  = error,
3086        .hot_add_disk   = raid1_add_disk,
3087        .hot_remove_disk= raid1_remove_disk,
3088        .spare_active   = raid1_spare_active,
3089        .sync_request   = sync_request,
3090        .resize         = raid1_resize,
3091        .size           = raid1_size,
3092        .check_reshape  = raid1_reshape,
3093        .quiesce        = raid1_quiesce,
3094        .takeover       = raid1_takeover,
3095};
3096
3097static int __init raid_init(void)
3098{
3099        return register_md_personality(&raid1_personality);
3100}
3101
3102static void raid_exit(void)
3103{
3104        unregister_md_personality(&raid1_personality);
3105}
3106
3107module_init(raid_init);
3108module_exit(raid_exit);
3109MODULE_LICENSE("GPL");
3110MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3111MODULE_ALIAS("md-personality-3"); /* RAID1 */
3112MODULE_ALIAS("md-raid1");
3113MODULE_ALIAS("md-level-1");
3114
3115module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
3116
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.