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/seq_file.h>
  38#include "md.h"
  39#include "raid1.h"
  40#include "bitmap.h"
  41
  42#define DEBUG 0
  43#if DEBUG
  44#define PRINTK(x...) printk(x)
  45#else
  46#define PRINTK(x...)
  47#endif
  48
  49/*
  50 * Number of guaranteed r1bios in case of extreme VM load:
  51 */
  52#define NR_RAID1_BIOS 256
  53
  54
  55static void unplug_slaves(mddev_t *mddev);
  56
  57static void allow_barrier(conf_t *conf);
  58static void lower_barrier(conf_t *conf);
  59
  60static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  61{
  62        struct pool_info *pi = data;
  63        r1bio_t *r1_bio;
  64        int size = offsetof(r1bio_t, bios[pi->raid_disks]);
  65
  66        /* allocate a r1bio with room for raid_disks entries in the bios array */
  67        r1_bio = kzalloc(size, gfp_flags);
  68        if (!r1_bio && pi->mddev)
  69                unplug_slaves(pi->mddev);
  70
  71        return r1_bio;
  72}
  73
  74static void r1bio_pool_free(void *r1_bio, void *data)
  75{
  76        kfree(r1_bio);
  77}
  78
  79#define RESYNC_BLOCK_SIZE (64*1024)
  80//#define RESYNC_BLOCK_SIZE PAGE_SIZE
  81#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  82#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  83#define RESYNC_WINDOW (2048*1024)
  84
  85static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  86{
  87        struct pool_info *pi = data;
  88        struct page *page;
  89        r1bio_t *r1_bio;
  90        struct bio *bio;
  91        int i, j;
  92
  93        r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  94        if (!r1_bio) {
  95                unplug_slaves(pi->mddev);
  96                return NULL;
  97        }
  98
  99        /*
 100         * Allocate bios : 1 for reading, n-1 for writing
 101         */
 102        for (j = pi->raid_disks ; j-- ; ) {
 103                bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 104                if (!bio)
 105                        goto out_free_bio;
 106                r1_bio->bios[j] = bio;
 107        }
 108        /*
 109         * Allocate RESYNC_PAGES data pages and attach them to
 110         * the first bio.
 111         * If this is a user-requested check/repair, allocate
 112         * RESYNC_PAGES for each bio.
 113         */
 114        if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 115                j = pi->raid_disks;
 116        else
 117                j = 1;
 118        while(j--) {
 119                bio = r1_bio->bios[j];
 120                for (i = 0; i < RESYNC_PAGES; i++) {
 121                        page = alloc_page(gfp_flags);
 122                        if (unlikely(!page))
 123                                goto out_free_pages;
 124
 125                        bio->bi_io_vec[i].bv_page = page;
 126                        bio->bi_vcnt = i+1;
 127                }
 128        }
 129        /* If not user-requests, copy the page pointers to all bios */
 130        if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 131                for (i=0; i<RESYNC_PAGES ; i++)
 132                        for (j=1; j<pi->raid_disks; j++)
 133                                r1_bio->bios[j]->bi_io_vec[i].bv_page =
 134                                        r1_bio->bios[0]->bi_io_vec[i].bv_page;
 135        }
 136
 137        r1_bio->master_bio = NULL;
 138
 139        return r1_bio;
 140
 141out_free_pages:
 142        for (j=0 ; j < pi->raid_disks; j++)
 143                for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 144                        put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 145        j = -1;
 146out_free_bio:
 147        while ( ++j < pi->raid_disks )
 148                bio_put(r1_bio->bios[j]);
 149        r1bio_pool_free(r1_bio, data);
 150        return NULL;
 151}
 152
 153static void r1buf_pool_free(void *__r1_bio, void *data)
 154{
 155        struct pool_info *pi = data;
 156        int i,j;
 157        r1bio_t *r1bio = __r1_bio;
 158
 159        for (i = 0; i < RESYNC_PAGES; i++)
 160                for (j = pi->raid_disks; j-- ;) {
 161                        if (j == 0 ||
 162                            r1bio->bios[j]->bi_io_vec[i].bv_page !=
 163                            r1bio->bios[0]->bi_io_vec[i].bv_page)
 164                                safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 165                }
 166        for (i=0 ; i < pi->raid_disks; i++)
 167                bio_put(r1bio->bios[i]);
 168
 169        r1bio_pool_free(r1bio, data);
 170}
 171
 172static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 173{
 174        int i;
 175
 176        for (i = 0; i < conf->raid_disks; i++) {
 177                struct bio **bio = r1_bio->bios + i;
 178                if (*bio && *bio != IO_BLOCKED)
 179                        bio_put(*bio);
 180                *bio = NULL;
 181        }
 182}
 183
 184static void free_r1bio(r1bio_t *r1_bio)
 185{
 186        conf_t *conf = r1_bio->mddev->private;
 187
 188        /*
 189         * Wake up any possible resync thread that waits for the device
 190         * to go idle.
 191         */
 192        allow_barrier(conf);
 193
 194        put_all_bios(conf, r1_bio);
 195        mempool_free(r1_bio, conf->r1bio_pool);
 196}
 197
 198static void put_buf(r1bio_t *r1_bio)
 199{
 200        conf_t *conf = r1_bio->mddev->private;
 201        int i;
 202
 203        for (i=0; i<conf->raid_disks; i++) {
 204                struct bio *bio = r1_bio->bios[i];
 205                if (bio->bi_end_io)
 206                        rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 207        }
 208
 209        mempool_free(r1_bio, conf->r1buf_pool);
 210
 211        lower_barrier(conf);
 212}
 213
 214static void reschedule_retry(r1bio_t *r1_bio)
 215{
 216        unsigned long flags;
 217        mddev_t *mddev = r1_bio->mddev;
 218        conf_t *conf = mddev->private;
 219
 220        spin_lock_irqsave(&conf->device_lock, flags);
 221        list_add(&r1_bio->retry_list, &conf->retry_list);
 222        conf->nr_queued ++;
 223        spin_unlock_irqrestore(&conf->device_lock, flags);
 224
 225        wake_up(&conf->wait_barrier);
 226        md_wakeup_thread(mddev->thread);
 227}
 228
 229/*
 230 * raid_end_bio_io() is called when we have finished servicing a mirrored
 231 * operation and are ready to return a success/failure code to the buffer
 232 * cache layer.
 233 */
 234static void raid_end_bio_io(r1bio_t *r1_bio)
 235{
 236        struct bio *bio = r1_bio->master_bio;
 237
 238        /* if nobody has done the final endio yet, do it now */
 239        if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 240                PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
 241                        (bio_data_dir(bio) == WRITE) ? "write" : "read",
 242                        (unsigned long long) bio->bi_sector,
 243                        (unsigned long long) bio->bi_sector +
 244                                (bio->bi_size >> 9) - 1);
 245
 246                bio_endio(bio,
 247                        test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 248        }
 249        free_r1bio(r1_bio);
 250}
 251
 252/*
 253 * Update disk head position estimator based on IRQ completion info.
 254 */
 255static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 256{
 257        conf_t *conf = r1_bio->mddev->private;
 258
 259        conf->mirrors[disk].head_position =
 260                r1_bio->sector + (r1_bio->sectors);
 261}
 262
 263static void raid1_end_read_request(struct bio *bio, int error)
 264{
 265        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 266        r1bio_t *r1_bio = bio->bi_private;
 267        int mirror;
 268        conf_t *conf = r1_bio->mddev->private;
 269
 270        mirror = r1_bio->read_disk;
 271        /*
 272         * this branch is our 'one mirror IO has finished' event handler:
 273         */
 274        update_head_pos(mirror, r1_bio);
 275
 276        if (uptodate)
 277                set_bit(R1BIO_Uptodate, &r1_bio->state);
 278        else {
 279                /* If all other devices have failed, we want to return
 280                 * the error upwards rather than fail the last device.
 281                 * Here we redefine "uptodate" to mean "Don't want to retry"
 282                 */
 283                unsigned long flags;
 284                spin_lock_irqsave(&conf->device_lock, flags);
 285                if (r1_bio->mddev->degraded == conf->raid_disks ||
 286                    (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 287                     !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 288                        uptodate = 1;
 289                spin_unlock_irqrestore(&conf->device_lock, flags);
 290        }
 291
 292        if (uptodate)
 293                raid_end_bio_io(r1_bio);
 294        else {
 295                /*
 296                 * oops, read error:
 297                 */
 298                char b[BDEVNAME_SIZE];
 299                if (printk_ratelimit())
 300                        printk(KERN_ERR "md/raid1:%s: %s: rescheduling sector %llu\n",
 301                               mdname(conf->mddev),
 302                               bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 303                reschedule_retry(r1_bio);
 304        }
 305
 306        rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 307}
 308
 309static void raid1_end_write_request(struct bio *bio, int error)
 310{
 311        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 312        r1bio_t *r1_bio = bio->bi_private;
 313        int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 314        conf_t *conf = r1_bio->mddev->private;
 315        struct bio *to_put = NULL;
 316
 317
 318        for (mirror = 0; mirror < conf->raid_disks; mirror++)
 319                if (r1_bio->bios[mirror] == bio)
 320                        break;
 321
 322        if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
 323                set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
 324                set_bit(R1BIO_BarrierRetry, &r1_bio->state);
 325                r1_bio->mddev->barriers_work = 0;
 326                /* Don't rdev_dec_pending in this branch - keep it for the retry */
 327        } else {
 328                /*
 329                 * this branch is our 'one mirror IO has finished' event handler:
 330                 */
 331                r1_bio->bios[mirror] = NULL;
 332                to_put = bio;
 333                if (!uptodate) {
 334                        md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 335                        /* an I/O failed, we can't clear the bitmap */
 336                        set_bit(R1BIO_Degraded, &r1_bio->state);
 337                } else
 338                        /*
 339                         * Set R1BIO_Uptodate in our master bio, so that
 340                         * we will return a good error code for to the higher
 341                         * levels even if IO on some other mirrored buffer fails.
 342                         *
 343                         * The 'master' represents the composite IO operation to
 344                         * user-side. So if something waits for IO, then it will
 345                         * wait for the 'master' bio.
 346                         */
 347                        set_bit(R1BIO_Uptodate, &r1_bio->state);
 348
 349                update_head_pos(mirror, r1_bio);
 350
 351                if (behind) {
 352                        if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 353                                atomic_dec(&r1_bio->behind_remaining);
 354
 355                        /* In behind mode, we ACK the master bio once the I/O has safely
 356                         * reached all non-writemostly disks. Setting the Returned bit
 357                         * ensures that this gets done only once -- we don't ever want to
 358                         * return -EIO here, instead we'll wait */
 359
 360                        if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 361                            test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 362                                /* Maybe we can return now */
 363                                if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 364                                        struct bio *mbio = r1_bio->master_bio;
 365                                        PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 366                                               (unsigned long long) mbio->bi_sector,
 367                                               (unsigned long long) mbio->bi_sector +
 368                                               (mbio->bi_size >> 9) - 1);
 369                                        bio_endio(mbio, 0);
 370                                }
 371                        }
 372                }
 373                rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 374        }
 375        /*
 376         *
 377         * Let's see if all mirrored write operations have finished
 378         * already.
 379         */
 380        if (atomic_dec_and_test(&r1_bio->remaining)) {
 381                if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
 382                        reschedule_retry(r1_bio);
 383                else {
 384                        /* it really is the end of this request */
 385                        if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 386                                /* free extra copy of the data pages */
 387                                int i = bio->bi_vcnt;
 388                                while (i--)
 389                                        safe_put_page(bio->bi_io_vec[i].bv_page);
 390                        }
 391                        /* clear the bitmap if all writes complete successfully */
 392                        bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 393                                        r1_bio->sectors,
 394                                        !test_bit(R1BIO_Degraded, &r1_bio->state),
 395                                        behind);
 396                        md_write_end(r1_bio->mddev);
 397                        raid_end_bio_io(r1_bio);
 398                }
 399        }
 400
 401        if (to_put)
 402                bio_put(to_put);
 403}
 404
 405
 406/*
 407 * This routine returns the disk from which the requested read should
 408 * be done. There is a per-array 'next expected sequential IO' sector
 409 * number - if this matches on the next IO then we use the last disk.
 410 * There is also a per-disk 'last know head position' sector that is
 411 * maintained from IRQ contexts, both the normal and the resync IO
 412 * completion handlers update this position correctly. If there is no
 413 * perfect sequential match then we pick the disk whose head is closest.
 414 *
 415 * If there are 2 mirrors in the same 2 devices, performance degrades
 416 * because position is mirror, not device based.
 417 *
 418 * The rdev for the device selected will have nr_pending incremented.
 419 */
 420static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 421{
 422        const sector_t this_sector = r1_bio->sector;
 423        int new_disk = conf->last_used, disk = new_disk;
 424        int wonly_disk = -1;
 425        const int sectors = r1_bio->sectors;
 426        sector_t new_distance, current_distance;
 427        mdk_rdev_t *rdev;
 428
 429        rcu_read_lock();
 430        /*
 431         * Check if we can balance. We can balance on the whole
 432         * device if no resync is going on, or below the resync window.
 433         * We take the first readable disk when above the resync window.
 434         */
 435 retry:
 436        if (conf->mddev->recovery_cp < MaxSector &&
 437            (this_sector + sectors >= conf->next_resync)) {
 438                /* Choose the first operational device, for consistancy */
 439                new_disk = 0;
 440
 441                for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 442                     r1_bio->bios[new_disk] == IO_BLOCKED ||
 443                     !rdev || !test_bit(In_sync, &rdev->flags)
 444                             || test_bit(WriteMostly, &rdev->flags);
 445                     rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
 446
 447                        if (rdev && test_bit(In_sync, &rdev->flags) &&
 448                                r1_bio->bios[new_disk] != IO_BLOCKED)
 449                                wonly_disk = new_disk;
 450
 451                        if (new_disk == conf->raid_disks - 1) {
 452                                new_disk = wonly_disk;
 453                                break;
 454                        }
 455                }
 456                goto rb_out;
 457        }
 458
 459
 460        /* make sure the disk is operational */
 461        for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 462             r1_bio->bios[new_disk] == IO_BLOCKED ||
 463             !rdev || !test_bit(In_sync, &rdev->flags) ||
 464                     test_bit(WriteMostly, &rdev->flags);
 465             rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
 466
 467                if (rdev && test_bit(In_sync, &rdev->flags) &&
 468                    r1_bio->bios[new_disk] != IO_BLOCKED)
 469                        wonly_disk = new_disk;
 470
 471                if (new_disk <= 0)
 472                        new_disk = conf->raid_disks;
 473                new_disk--;
 474                if (new_disk == disk) {
 475                        new_disk = wonly_disk;
 476                        break;
 477                }
 478        }
 479
 480        if (new_disk < 0)
 481                goto rb_out;
 482
 483        disk = new_disk;
 484        /* now disk == new_disk == starting point for search */
 485
 486        /*
 487         * Don't change to another disk for sequential reads:
 488         */
 489        if (conf->next_seq_sect == this_sector)
 490                goto rb_out;
 491        if (this_sector == conf->mirrors[new_disk].head_position)
 492                goto rb_out;
 493
 494        current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 495
 496        /* Find the disk whose head is closest */
 497
 498        do {
 499                if (disk <= 0)
 500                        disk = conf->raid_disks;
 501                disk--;
 502
 503                rdev = rcu_dereference(conf->mirrors[disk].rdev);
 504
 505                if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
 506                    !test_bit(In_sync, &rdev->flags) ||
 507                    test_bit(WriteMostly, &rdev->flags))
 508                        continue;
 509
 510                if (!atomic_read(&rdev->nr_pending)) {
 511                        new_disk = disk;
 512                        break;
 513                }
 514                new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 515                if (new_distance < current_distance) {
 516                        current_distance = new_distance;
 517                        new_disk = disk;
 518                }
 519        } while (disk != conf->last_used);
 520
 521 rb_out:
 522
 523
 524        if (new_disk >= 0) {
 525                rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 526                if (!rdev)
 527                        goto retry;
 528                atomic_inc(&rdev->nr_pending);
 529                if (!test_bit(In_sync, &rdev->flags)) {
 530                        /* cannot risk returning a device that failed
 531                         * before we inc'ed nr_pending
 532                         */
 533                        rdev_dec_pending(rdev, conf->mddev);
 534                        goto retry;
 535                }
 536                conf->next_seq_sect = this_sector + sectors;
 537                conf->last_used = new_disk;
 538        }
 539        rcu_read_unlock();
 540
 541        return new_disk;
 542}
 543
 544static void unplug_slaves(mddev_t *mddev)
 545{
 546        conf_t *conf = mddev->private;
 547        int i;
 548
 549        rcu_read_lock();
 550        for (i=0; i<mddev->raid_disks; i++) {
 551                mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 552                if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 553                        struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 554
 555                        atomic_inc(&rdev->nr_pending);
 556                        rcu_read_unlock();
 557
 558                        blk_unplug(r_queue);
 559
 560                        rdev_dec_pending(rdev, mddev);
 561                        rcu_read_lock();
 562                }
 563        }
 564        rcu_read_unlock();
 565}
 566
 567static void raid1_unplug(struct request_queue *q)
 568{
 569        mddev_t *mddev = q->queuedata;
 570
 571        unplug_slaves(mddev);
 572        md_wakeup_thread(mddev->thread);
 573}
 574
 575static int raid1_congested(void *data, int bits)
 576{
 577        mddev_t *mddev = data;
 578        conf_t *conf = mddev->private;
 579        int i, ret = 0;
 580
 581        if (mddev_congested(mddev, bits))
 582                return 1;
 583
 584        rcu_read_lock();
 585        for (i = 0; i < mddev->raid_disks; i++) {
 586                mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 587                if (rdev && !test_bit(Faulty, &rdev->flags)) {
 588                        struct request_queue *q = bdev_get_queue(rdev->bdev);
 589
 590                        /* Note the '|| 1' - when read_balance prefers
 591                         * non-congested targets, it can be removed
 592                         */
 593                        if ((bits & (1<<BDI_async_congested)) || 1)
 594                                ret |= bdi_congested(&q->backing_dev_info, bits);
 595                        else
 596                                ret &= bdi_congested(&q->backing_dev_info, bits);
 597                }
 598        }
 599        rcu_read_unlock();
 600        return ret;
 601}
 602
 603
 604static int flush_pending_writes(conf_t *conf)
 605{
 606        /* Any writes that have been queued but are awaiting
 607         * bitmap updates get flushed here.
 608         * We return 1 if any requests were actually submitted.
 609         */
 610        int rv = 0;
 611
 612        spin_lock_irq(&conf->device_lock);
 613
 614        if (conf->pending_bio_list.head) {
 615                struct bio *bio;
 616                bio = bio_list_get(&conf->pending_bio_list);
 617                blk_remove_plug(conf->mddev->queue);
 618                spin_unlock_irq(&conf->device_lock);
 619                /* flush any pending bitmap writes to
 620                 * disk before proceeding w/ I/O */
 621                bitmap_unplug(conf->mddev->bitmap);
 622
 623                while (bio) { /* submit pending writes */
 624                        struct bio *next = bio->bi_next;
 625                        bio->bi_next = NULL;
 626                        generic_make_request(bio);
 627                        bio = next;
 628                }
 629                rv = 1;
 630        } else
 631                spin_unlock_irq(&conf->device_lock);
 632        return rv;
 633}
 634
 635/* Barriers....
 636 * Sometimes we need to suspend IO while we do something else,
 637 * either some resync/recovery, or reconfigure the array.
 638 * To do this we raise a 'barrier'.
 639 * The 'barrier' is a counter that can be raised multiple times
 640 * to count how many activities are happening which preclude
 641 * normal IO.
 642 * We can only raise the barrier if there is no pending IO.
 643 * i.e. if nr_pending == 0.
 644 * We choose only to raise the barrier if no-one is waiting for the
 645 * barrier to go down.  This means that as soon as an IO request
 646 * is ready, no other operations which require a barrier will start
 647 * until the IO request has had a chance.
 648 *
 649 * So: regular IO calls 'wait_barrier'.  When that returns there
 650 *    is no backgroup IO happening,  It must arrange to call
 651 *    allow_barrier when it has finished its IO.
 652 * backgroup IO calls must call raise_barrier.  Once that returns
 653 *    there is no normal IO happeing.  It must arrange to call
 654 *    lower_barrier when the particular background IO completes.
 655 */
 656#define RESYNC_DEPTH 32
 657
 658static void raise_barrier(conf_t *conf)
 659{
 660        spin_lock_irq(&conf->resync_lock);
 661
 662        /* Wait until no block IO is waiting */
 663        wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 664                            conf->resync_lock,
 665                            raid1_unplug(conf->mddev->queue));
 666
 667        /* block any new IO from starting */
 668        conf->barrier++;
 669
 670        /* No wait for all pending IO to complete */
 671        wait_event_lock_irq(conf->wait_barrier,
 672                            !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 673                            conf->resync_lock,
 674                            raid1_unplug(conf->mddev->queue));
 675
 676        spin_unlock_irq(&conf->resync_lock);
 677}
 678
 679static void lower_barrier(conf_t *conf)
 680{
 681        unsigned long flags;
 682        BUG_ON(conf->barrier <= 0);
 683        spin_lock_irqsave(&conf->resync_lock, flags);
 684        conf->barrier--;
 685        spin_unlock_irqrestore(&conf->resync_lock, flags);
 686        wake_up(&conf->wait_barrier);
 687}
 688
 689static void wait_barrier(conf_t *conf)
 690{
 691        spin_lock_irq(&conf->resync_lock);
 692        if (conf->barrier) {
 693                conf->nr_waiting++;
 694                wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 695                                    conf->resync_lock,
 696                                    raid1_unplug(conf->mddev->queue));
 697                conf->nr_waiting--;
 698        }
 699        conf->nr_pending++;
 700        spin_unlock_irq(&conf->resync_lock);
 701}
 702
 703static void allow_barrier(conf_t *conf)
 704{
 705        unsigned long flags;
 706        spin_lock_irqsave(&conf->resync_lock, flags);
 707        conf->nr_pending--;
 708        spin_unlock_irqrestore(&conf->resync_lock, flags);
 709        wake_up(&conf->wait_barrier);
 710}
 711
 712static void freeze_array(conf_t *conf)
 713{
 714        /* stop syncio and normal IO and wait for everything to
 715         * go quite.
 716         * We increment barrier and nr_waiting, and then
 717         * wait until nr_pending match nr_queued+1
 718         * This is called in the context of one normal IO request
 719         * that has failed. Thus any sync request that might be pending
 720         * will be blocked by nr_pending, and we need to wait for
 721         * pending IO requests to complete or be queued for re-try.
 722         * Thus the number queued (nr_queued) plus this request (1)
 723         * must match the number of pending IOs (nr_pending) before
 724         * we continue.
 725         */
 726        spin_lock_irq(&conf->resync_lock);
 727        conf->barrier++;
 728        conf->nr_waiting++;
 729        wait_event_lock_irq(conf->wait_barrier,
 730                            conf->nr_pending == conf->nr_queued+1,
 731                            conf->resync_lock,
 732                            ({ flush_pending_writes(conf);
 733                               raid1_unplug(conf->mddev->queue); }));
 734        spin_unlock_irq(&conf->resync_lock);
 735}
 736static void unfreeze_array(conf_t *conf)
 737{
 738        /* reverse the effect of the freeze */
 739        spin_lock_irq(&conf->resync_lock);
 740        conf->barrier--;
 741        conf->nr_waiting--;
 742        wake_up(&conf->wait_barrier);
 743        spin_unlock_irq(&conf->resync_lock);
 744}
 745
 746
 747/* duplicate the data pages for behind I/O */
 748static struct page **alloc_behind_pages(struct bio *bio)
 749{
 750        int i;
 751        struct bio_vec *bvec;
 752        struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
 753                                        GFP_NOIO);
 754        if (unlikely(!pages))
 755                goto do_sync_io;
 756
 757        bio_for_each_segment(bvec, bio, i) {
 758                pages[i] = alloc_page(GFP_NOIO);
 759                if (unlikely(!pages[i]))
 760                        goto do_sync_io;
 761                memcpy(kmap(pages[i]) + bvec->bv_offset,
 762                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 763                kunmap(pages[i]);
 764                kunmap(bvec->bv_page);
 765        }
 766
 767        return pages;
 768
 769do_sync_io:
 770        if (pages)
 771                for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
 772                        put_page(pages[i]);
 773        kfree(pages);
 774        PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 775        return NULL;
 776}
 777
 778static int make_request(mddev_t *mddev, struct bio * bio)
 779{
 780        conf_t *conf = mddev->private;
 781        mirror_info_t *mirror;
 782        r1bio_t *r1_bio;
 783        struct bio *read_bio;
 784        int i, targets = 0, disks;
 785        struct bitmap *bitmap;
 786        unsigned long flags;
 787        struct bio_list bl;
 788        struct page **behind_pages = NULL;
 789        const int rw = bio_data_dir(bio);
 790        const bool do_sync = bio_rw_flagged(bio, BIO_RW_SYNCIO);
 791        bool do_barriers;
 792        mdk_rdev_t *blocked_rdev;
 793
 794        /*
 795         * Register the new request and wait if the reconstruction
 796         * thread has put up a bar for new requests.
 797         * Continue immediately if no resync is active currently.
 798         * We test barriers_work *after* md_write_start as md_write_start
 799         * may cause the first superblock write, and that will check out
 800         * if barriers work.
 801         */
 802
 803        md_write_start(mddev, bio); /* wait on superblock update early */
 804
 805        if (bio_data_dir(bio) == WRITE &&
 806            bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
 807            bio->bi_sector < mddev->suspend_hi) {
 808                /* As the suspend_* range is controlled by
 809                 * userspace, we want an interruptible
 810                 * wait.
 811                 */
 812                DEFINE_WAIT(w);
 813                for (;;) {
 814                        flush_signals(current);
 815                        prepare_to_wait(&conf->wait_barrier,
 816                                        &w, TASK_INTERRUPTIBLE);
 817                        if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
 818                            bio->bi_sector >= mddev->suspend_hi)
 819                                break;
 820                        schedule();
 821                }
 822                finish_wait(&conf->wait_barrier, &w);
 823        }
 824        if (unlikely(!mddev->barriers_work &&
 825                     bio_rw_flagged(bio, BIO_RW_BARRIER))) {
 826                if (rw == WRITE)
 827                        md_write_end(mddev);
 828                bio_endio(bio, -EOPNOTSUPP);
 829                return 0;
 830        }
 831
 832        wait_barrier(conf);
 833
 834        bitmap = mddev->bitmap;
 835
 836        /*
 837         * make_request() can abort the operation when READA is being
 838         * used and no empty request is available.
 839         *
 840         */
 841        r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 842
 843        r1_bio->master_bio = bio;
 844        r1_bio->sectors = bio->bi_size >> 9;
 845        r1_bio->state = 0;
 846        r1_bio->mddev = mddev;
 847        r1_bio->sector = bio->bi_sector;
 848
 849        if (rw == READ) {
 850                /*
 851                 * read balancing logic:
 852                 */
 853                int rdisk = read_balance(conf, r1_bio);
 854
 855                if (rdisk < 0) {
 856                        /* couldn't find anywhere to read from */
 857                        raid_end_bio_io(r1_bio);
 858                        return 0;
 859                }
 860                mirror = conf->mirrors + rdisk;
 861
 862                if (test_bit(WriteMostly, &mirror->rdev->flags) &&
 863                    bitmap) {
 864                        /* Reading from a write-mostly device must
 865                         * take care not to over-take any writes
 866                         * that are 'behind'
 867                         */
 868                        wait_event(bitmap->behind_wait,
 869                                   atomic_read(&bitmap->behind_writes) == 0);
 870                }
 871                r1_bio->read_disk = rdisk;
 872
 873                read_bio = bio_clone(bio, GFP_NOIO);
 874
 875                r1_bio->bios[rdisk] = read_bio;
 876
 877                read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 878                read_bio->bi_bdev = mirror->rdev->bdev;
 879                read_bio->bi_end_io = raid1_end_read_request;
 880                read_bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
 881                read_bio->bi_private = r1_bio;
 882
 883                generic_make_request(read_bio);
 884                return 0;
 885        }
 886
 887        /*
 888         * WRITE:
 889         */
 890        /* first select target devices under spinlock and
 891         * inc refcount on their rdev.  Record them by setting
 892         * bios[x] to bio
 893         */
 894        disks = conf->raid_disks;
 895#if 0
 896        { static int first=1;
 897        if (first) printk("First Write sector %llu disks %d\n",
 898                          (unsigned long long)r1_bio->sector, disks);
 899        first = 0;
 900        }
 901#endif
 902 retry_write:
 903        blocked_rdev = NULL;
 904        rcu_read_lock();
 905        for (i = 0;  i < disks; i++) {
 906                mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 907                if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 908                        atomic_inc(&rdev->nr_pending);
 909                        blocked_rdev = rdev;
 910                        break;
 911                }
 912                if (rdev && !test_bit(Faulty, &rdev->flags)) {
 913                        atomic_inc(&rdev->nr_pending);
 914                        if (test_bit(Faulty, &rdev->flags)) {
 915                                rdev_dec_pending(rdev, mddev);
 916                                r1_bio->bios[i] = NULL;
 917                        } else {
 918                                r1_bio->bios[i] = bio;
 919                                targets++;
 920                        }
 921                } else
 922                        r1_bio->bios[i] = NULL;
 923        }
 924        rcu_read_unlock();
 925
 926        if (unlikely(blocked_rdev)) {
 927                /* Wait for this device to become unblocked */
 928                int j;
 929
 930                for (j = 0; j < i; j++)
 931                        if (r1_bio->bios[j])
 932                                rdev_dec_pending(conf->mirrors[j].rdev, mddev);
 933
 934                allow_barrier(conf);
 935                md_wait_for_blocked_rdev(blocked_rdev, mddev);
 936                wait_barrier(conf);
 937                goto retry_write;
 938        }
 939
 940        BUG_ON(targets == 0); /* we never fail the last device */
 941
 942        if (targets < conf->raid_disks) {
 943                /* array is degraded, we will not clear the bitmap
 944                 * on I/O completion (see raid1_end_write_request) */
 945                set_bit(R1BIO_Degraded, &r1_bio->state);
 946        }
 947
 948        /* do behind I/O ?
 949         * Not if there are too many, or cannot allocate memory,
 950         * or a reader on WriteMostly is waiting for behind writes 
 951         * to flush */
 952        if (bitmap &&
 953            (atomic_read(&bitmap->behind_writes)
 954             < mddev->bitmap_info.max_write_behind) &&
 955            !waitqueue_active(&bitmap->behind_wait) &&
 956            (behind_pages = alloc_behind_pages(bio)) != NULL)
 957                set_bit(R1BIO_BehindIO, &r1_bio->state);
 958
 959        atomic_set(&r1_bio->remaining, 0);
 960        atomic_set(&r1_bio->behind_remaining, 0);
 961
 962        do_barriers = bio_rw_flagged(bio, BIO_RW_BARRIER);
 963        if (do_barriers)
 964                set_bit(R1BIO_Barrier, &r1_bio->state);
 965
 966        bio_list_init(&bl);
 967        for (i = 0; i < disks; i++) {
 968                struct bio *mbio;
 969                if (!r1_bio->bios[i])
 970                        continue;
 971
 972                mbio = bio_clone(bio, GFP_NOIO);
 973                r1_bio->bios[i] = mbio;
 974
 975                mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 976                mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 977                mbio->bi_end_io = raid1_end_write_request;
 978                mbio->bi_rw = WRITE | (do_barriers << BIO_RW_BARRIER) |
 979                        (do_sync << BIO_RW_SYNCIO);
 980                mbio->bi_private = r1_bio;
 981
 982                if (behind_pages) {
 983                        struct bio_vec *bvec;
 984                        int j;
 985
 986                        /* Yes, I really want the '__' version so that
 987                         * we clear any unused pointer in the io_vec, rather
 988                         * than leave them unchanged.  This is important
 989                         * because when we come to free the pages, we won't
 990                         * know the originial bi_idx, so we just free
 991                         * them all
 992                         */
 993                        __bio_for_each_segment(bvec, mbio, j, 0)
 994                                bvec->bv_page = behind_pages[j];
 995                        if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 996                                atomic_inc(&r1_bio->behind_remaining);
 997                }
 998
 999                atomic_inc(&r1_bio->remaining);
1000
1001                bio_list_add(&bl, mbio);
1002        }
1003        kfree(behind_pages); /* the behind pages are attached to the bios now */
1004
1005        bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
1006                                test_bit(R1BIO_BehindIO, &r1_bio->state));
1007        spin_lock_irqsave(&conf->device_lock, flags);
1008        bio_list_merge(&conf->pending_bio_list, &bl);
1009        bio_list_init(&bl);
1010
1011        blk_plug_device(mddev->queue);
1012        spin_unlock_irqrestore(&conf->device_lock, flags);
1013
1014        /* In case raid1d snuck into freeze_array */
1015        wake_up(&conf->wait_barrier);
1016
1017        if (do_sync)
1018                md_wakeup_thread(mddev->thread);
1019#if 0
1020        while ((bio = bio_list_pop(&bl)) != NULL)
1021                generic_make_request(bio);
1022#endif
1023
1024        return 0;
1025}
1026
1027static void status(struct seq_file *seq, mddev_t *mddev)
1028{
1029        conf_t *conf = mddev->private;
1030        int i;
1031
1032        seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1033                   conf->raid_disks - mddev->degraded);
1034        rcu_read_lock();
1035        for (i = 0; i < conf->raid_disks; i++) {
1036                mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1037                seq_printf(seq, "%s",
1038                           rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1039        }
1040        rcu_read_unlock();
1041        seq_printf(seq, "]");
1042}
1043
1044
1045static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1046{
1047        char b[BDEVNAME_SIZE];
1048        conf_t *conf = mddev->private;
1049
1050        /*
1051         * If it is not operational, then we have already marked it as dead
1052         * else if it is the last working disks, ignore the error, let the
1053         * next level up know.
1054         * else mark the drive as failed
1055         */
1056        if (test_bit(In_sync, &rdev->flags)
1057            && (conf->raid_disks - mddev->degraded) == 1) {
1058                /*
1059                 * Don't fail the drive, act as though we were just a
1060                 * normal single drive.
1061                 * However don't try a recovery from this drive as
1062                 * it is very likely to fail.
1063                 */
1064                mddev->recovery_disabled = 1;
1065                return;
1066        }
1067        if (test_and_clear_bit(In_sync, &rdev->flags)) {
1068                unsigned long flags;
1069                spin_lock_irqsave(&conf->device_lock, flags);
1070                mddev->degraded++;
1071                set_bit(Faulty, &rdev->flags);
1072                spin_unlock_irqrestore(&conf->device_lock, flags);
1073                /*
1074                 * if recovery is running, make sure it aborts.
1075                 */
1076                set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1077        } else
1078                set_bit(Faulty, &rdev->flags);
1079        set_bit(MD_CHANGE_DEVS, &mddev->flags);
1080        printk(KERN_ALERT "md/raid1:%s: Disk failure on %s, disabling device.\n"
1081               KERN_ALERT "md/raid1:%s: Operation continuing on %d devices.\n",
1082               mdname(mddev), bdevname(rdev->bdev, b),
1083               mdname(mddev), conf->raid_disks - mddev->degraded);
1084}
1085
1086static void print_conf(conf_t *conf)
1087{
1088        int i;
1089
1090        printk(KERN_DEBUG "RAID1 conf printout:\n");
1091        if (!conf) {
1092                printk(KERN_DEBUG "(!conf)\n");
1093                return;
1094        }
1095        printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1096                conf->raid_disks);
1097
1098        rcu_read_lock();
1099        for (i = 0; i < conf->raid_disks; i++) {
1100                char b[BDEVNAME_SIZE];
1101                mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1102                if (rdev)
1103                        printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1104                               i, !test_bit(In_sync, &rdev->flags),
1105                               !test_bit(Faulty, &rdev->flags),
1106                               bdevname(rdev->bdev,b));
1107        }
1108        rcu_read_unlock();
1109}
1110
1111static void close_sync(conf_t *conf)
1112{
1113        wait_barrier(conf);
1114        allow_barrier(conf);
1115
1116        mempool_destroy(conf->r1buf_pool);
1117        conf->r1buf_pool = NULL;
1118}
1119
1120static int raid1_spare_active(mddev_t *mddev)
1121{
1122        int i;
1123        conf_t *conf = mddev->private;
1124
1125        /*
1126         * Find all failed disks within the RAID1 configuration 
1127         * and mark them readable.
1128         * Called under mddev lock, so rcu protection not needed.
1129         */
1130        for (i = 0; i < conf->raid_disks; i++) {
1131                mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1132                if (rdev
1133                    && !test_bit(Faulty, &rdev->flags)
1134                    && !test_and_set_bit(In_sync, &rdev->flags)) {
1135                        unsigned long flags;
1136                        spin_lock_irqsave(&conf->device_lock, flags);
1137                        mddev->degraded--;
1138                        spin_unlock_irqrestore(&conf->device_lock, flags);
1139                }
1140        }
1141
1142        print_conf(conf);
1143        return 0;
1144}
1145
1146
1147static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1148{
1149        conf_t *conf = mddev->private;
1150        int err = -EEXIST;
1151        int mirror = 0;
1152        mirror_info_t *p;
1153        int first = 0;
1154        int last = mddev->raid_disks - 1;
1155
1156        if (rdev->raid_disk >= 0)
1157                first = last = rdev->raid_disk;
1158
1159        for (mirror = first; mirror <= last; mirror++)
1160                if ( !(p=conf->mirrors+mirror)->rdev) {
1161
1162                        disk_stack_limits(mddev->gendisk, rdev->bdev,
1163                                          rdev->data_offset << 9);
1164                        /* as we don't honour merge_bvec_fn, we must
1165                         * never risk violating it, so limit
1166                         * ->max_segments to one lying with a single
1167                         * page, as a one page request is never in
1168                         * violation.
1169                         */
1170                        if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1171                                blk_queue_max_segments(mddev->queue, 1);
1172                                blk_queue_segment_boundary(mddev->queue,
1173                                                           PAGE_CACHE_SIZE - 1);
1174                        }
1175
1176                        p->head_position = 0;
1177                        rdev->raid_disk = mirror;
1178                        err = 0;
1179                        /* As all devices are equivalent, we don't need a full recovery
1180                         * if this was recently any drive of the array
1181                         */
1182                        if (rdev->saved_raid_disk < 0)
1183                                conf->fullsync = 1;
1184                        rcu_assign_pointer(p->rdev, rdev);
1185                        break;
1186                }
1187        md_integrity_add_rdev(rdev, mddev);
1188        print_conf(conf);
1189        return err;
1190}
1191
1192static int raid1_remove_disk(mddev_t *mddev, int number)
1193{
1194        conf_t *conf = mddev->private;
1195        int err = 0;
1196        mdk_rdev_t *rdev;
1197        mirror_info_t *p = conf->mirrors+ number;
1198
1199        print_conf(conf);
1200        rdev = p->rdev;
1201        if (rdev) {
1202                if (test_bit(In_sync, &rdev->flags) ||
1203                    atomic_read(&rdev->nr_pending)) {
1204                        err = -EBUSY;
1205                        goto abort;
1206                }
1207                /* Only remove non-faulty devices is recovery
1208                 * is not possible.
1209                 */
1210                if (!test_bit(Faulty, &rdev->flags) &&
1211                    mddev->degraded < conf->raid_disks) {
1212                        err = -EBUSY;
1213                        goto abort;
1214                }
1215                p->rdev = NULL;
1216                synchronize_rcu();
1217                if (atomic_read(&rdev->nr_pending)) {
1218                        /* lost the race, try later */
1219                        err = -EBUSY;
1220                        p->rdev = rdev;
1221                        goto abort;
1222                }
1223                md_integrity_register(mddev);
1224        }
1225abort:
1226
1227        print_conf(conf);
1228        return err;
1229}
1230
1231
1232static void end_sync_read(struct bio *bio, int error)
1233{
1234        r1bio_t *r1_bio = bio->bi_private;
1235        int i;
1236
1237        for (i=r1_bio->mddev->raid_disks; i--; )
1238                if (r1_bio->bios[i] == bio)
1239                        break;
1240        BUG_ON(i < 0);
1241        update_head_pos(i, r1_bio);
1242        /*
1243         * we have read a block, now it needs to be re-written,
1244         * or re-read if the read failed.
1245         * We don't do much here, just schedule handling by raid1d
1246         */
1247        if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1248                set_bit(R1BIO_Uptodate, &r1_bio->state);
1249
1250        if (atomic_dec_and_test(&r1_bio->remaining))
1251                reschedule_retry(r1_bio);
1252}
1253
1254static void end_sync_write(struct bio *bio, int error)
1255{
1256        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1257        r1bio_t *r1_bio = bio->bi_private;
1258        mddev_t *mddev = r1_bio->mddev;
1259        conf_t *conf = mddev->private;
1260        int i;
1261        int mirror=0;
1262
1263        for (i = 0; i < conf->raid_disks; i++)
1264                if (r1_bio->bios[i] == bio) {
1265                        mirror = i;
1266                        break;
1267                }
1268        if (!uptodate) {
1269                int sync_blocks = 0;
1270                sector_t s = r1_bio->sector;
1271                long sectors_to_go = r1_bio->sectors;
1272                /* make sure these bits doesn't get cleared. */
1273                do {
1274                        bitmap_end_sync(mddev->bitmap, s,
1275                                        &sync_blocks, 1);
1276                        s += sync_blocks;
1277                        sectors_to_go -= sync_blocks;
1278                } while (sectors_to_go > 0);
1279                md_error(mddev, conf->mirrors[mirror].rdev);
1280        }
1281
1282        update_head_pos(mirror, r1_bio);
1283
1284        if (atomic_dec_and_test(&r1_bio->remaining)) {
1285                sector_t s = r1_bio->sectors;
1286                put_buf(r1_bio);
1287                md_done_sync(mddev, s, uptodate);
1288        }
1289}
1290
1291static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1292{
1293        conf_t *conf = mddev->private;
1294        int i;
1295        int disks = conf->raid_disks;
1296        struct bio *bio, *wbio;
1297
1298        bio = r1_bio->bios[r1_bio->read_disk];
1299
1300
1301        if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1302                /* We have read all readable devices.  If we haven't
1303                 * got the block, then there is no hope left.
1304                 * If we have, then we want to do a comparison
1305                 * and skip the write if everything is the same.
1306                 * If any blocks failed to read, then we need to
1307                 * attempt an over-write
1308                 */
1309                int primary;
1310                if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1311                        for (i=0; i<mddev->raid_disks; i++)
1312                                if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1313                                        md_error(mddev, conf->mirrors[i].rdev);
1314
1315                        md_done_sync(mddev, r1_bio->sectors, 1);
1316                        put_buf(r1_bio);
1317                        return;
1318                }
1319                for (primary=0; primary<mddev->raid_disks; primary++)
1320                        if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1321                            test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1322                                r1_bio->bios[primary]->bi_end_io = NULL;
1323                                rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1324                                break;
1325                        }
1326                r1_bio->read_disk = primary;
1327                for (i=0; i<mddev->raid_disks; i++)
1328                        if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1329                                int j;
1330                                int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1331                                struct bio *pbio = r1_bio->bios[primary];
1332                                struct bio *sbio = r1_bio->bios[i];
1333
1334                                if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1335                                        for (j = vcnt; j-- ; ) {
1336                                                struct page *p, *s;
1337                                                p = pbio->bi_io_vec[j].bv_page;
1338                                                s = sbio->bi_io_vec[j].bv_page;
1339                                                if (memcmp(page_address(p),
1340                                                           page_address(s),
1341                                                           PAGE_SIZE))
1342                                                        break;
1343                                        }
1344                                } else
1345                                        j = 0;
1346                                if (j >= 0)
1347                                        mddev->resync_mismatches += r1_bio->sectors;
1348                                if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1349                                              && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1350                                        sbio->bi_end_io = NULL;
1351                                        rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1352                                } else {
1353                                        /* fixup the bio for reuse */
1354                                        int size;
1355                                        sbio->bi_vcnt = vcnt;
1356                                        sbio->bi_size = r1_bio->sectors << 9;
1357                                        sbio->bi_idx = 0;
1358                                        sbio->bi_phys_segments = 0;
1359                                        sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1360                                        sbio->bi_flags |= 1 << BIO_UPTODATE;
1361                                        sbio->bi_next = NULL;
1362                                        sbio->bi_sector = r1_bio->sector +
1363                                                conf->mirrors[i].rdev->data_offset;
1364                                        sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1365                                        size = sbio->bi_size;
1366                                        for (j = 0; j < vcnt ; j++) {
1367                                                struct bio_vec *bi;
1368                                                bi = &sbio->bi_io_vec[j];
1369                                                bi->bv_offset = 0;
1370                                                if (size > PAGE_SIZE)
1371                                                        bi->bv_len = PAGE_SIZE;
1372                                                else
1373                                                        bi->bv_len = size;
1374                                                size -= PAGE_SIZE;
1375                                                memcpy(page_address(bi->bv_page),
1376                                                       page_address(pbio->bi_io_vec[j].bv_page),
1377                                                       PAGE_SIZE);
1378                                        }
1379
1380                                }
1381                        }
1382        }
1383        if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1384                /* ouch - failed to read all of that.
1385                 * Try some synchronous reads of other devices to get
1386                 * good data, much like with normal read errors.  Only
1387                 * read into the pages we already have so we don't
1388                 * need to re-issue the read request.
1389                 * We don't need to freeze the array, because being in an
1390                 * active sync request, there is no normal IO, and
1391                 * no overlapping syncs.
1392                 */
1393                sector_t sect = r1_bio->sector;
1394                int sectors = r1_bio->sectors;
1395                int idx = 0;
1396
1397                while(sectors) {
1398                        int s = sectors;
1399                        int d = r1_bio->read_disk;
1400                        int success = 0;
1401                        mdk_rdev_t *rdev;
1402
1403                        if (s > (PAGE_SIZE>>9))
1404                                s = PAGE_SIZE >> 9;
1405                        do {
1406                                if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1407                                        /* No rcu protection needed here devices
1408                                         * can only be removed when no resync is
1409                                         * active, and resync is currently active
1410                                         */
1411                                        rdev = conf->mirrors[d].rdev;
1412                                        if (sync_page_io(rdev->bdev,
1413                                                         sect + rdev->data_offset,
1414                                                         s<<9,
1415                                                         bio->bi_io_vec[idx].bv_page,
1416                                                         READ)) {
1417                                                success = 1;
1418                                                break;
1419                                        }
1420                                }
1421                                d++;
1422                                if (d == conf->raid_disks)
1423                                        d = 0;
1424                        } while (!success && d != r1_bio->read_disk);
1425
1426                        if (success) {
1427                                int start = d;
1428                                /* write it back and re-read */
1429                                set_bit(R1BIO_Uptodate, &r1_bio->state);
1430                                while (d != r1_bio->read_disk) {
1431                                        if (d == 0)
1432                                                d = conf->raid_disks;
1433                                        d--;
1434                                        if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1435                                                continue;
1436                                        rdev = conf->mirrors[d].rdev;
1437                                        atomic_add(s, &rdev->corrected_errors);
1438                                        if (sync_page_io(rdev->bdev,
1439                                                         sect + rdev->data_offset,
1440                                                         s<<9,
1441                                                         bio->bi_io_vec[idx].bv_page,
1442                                                         WRITE) == 0)
1443                                                md_error(mddev, rdev);
1444                                }
1445                                d = start;
1446                                while (d != r1_bio->read_disk) {
1447                                        if (d == 0)
1448                                                d = conf->raid_disks;
1449                                        d--;
1450                                        if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1451                                                continue;
1452                                        rdev = conf->mirrors[d].rdev;
1453                                        if (sync_page_io(rdev->bdev,
1454                                                         sect + rdev->data_offset,
1455                                                         s<<9,
1456                                                         bio->bi_io_vec[idx].bv_page,
1457                                                         READ) == 0)
1458                                                md_error(mddev, rdev);
1459                                }
1460                        } else {
1461                                char b[BDEVNAME_SIZE];
1462                                /* Cannot read from anywhere, array is toast */
1463                                md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1464                                printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1465                                       " for block %llu\n",
1466                                       mdname(mddev),
1467                                       bdevname(bio->bi_bdev, b),
1468                                       (unsigned long long)r1_bio->sector);
1469                                md_done_sync(mddev, r1_bio->sectors, 0);
1470                                put_buf(r1_bio);
1471                                return;
1472                        }
1473                        sectors -= s;
1474                        sect += s;
1475                        idx ++;
1476                }
1477        }
1478
1479        /*
1480         * schedule writes
1481         */
1482        atomic_set(&r1_bio->remaining, 1);
1483        for (i = 0; i < disks ; i++) {
1484                wbio = r1_bio->bios[i];
1485                if (wbio->bi_end_io == NULL ||
1486                    (wbio->bi_end_io == end_sync_read &&
1487                     (i == r1_bio->read_disk ||
1488                      !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1489                        continue;
1490
1491                wbio->bi_rw = WRITE;
1492                wbio->bi_end_io = end_sync_write;
1493                atomic_inc(&r1_bio->remaining);
1494                md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1495
1496                generic_make_request(wbio);
1497        }
1498
1499        if (atomic_dec_and_test(&r1_bio->remaining)) {
1500                /* if we're here, all write(s) have completed, so clean up */
1501                md_done_sync(mddev, r1_bio->sectors, 1);
1502                put_buf(r1_bio);
1503        }
1504}
1505
1506/*
1507 * This is a kernel thread which:
1508 *
1509 *      1.      Retries failed read operations on working mirrors.
1510 *      2.      Updates the raid superblock when problems encounter.
1511 *      3.      Performs writes following reads for array syncronising.
1512 */
1513
1514static void fix_read_error(conf_t *conf, int read_disk,
1515                           sector_t sect, int sectors)
1516{
1517        mddev_t *mddev = conf->mddev;
1518        while(sectors) {
1519                int s = sectors;
1520                int d = read_disk;
1521                int success = 0;
1522                int start;
1523                mdk_rdev_t *rdev;
1524
1525                if (s > (PAGE_SIZE>>9))
1526                        s = PAGE_SIZE >> 9;
1527
1528                do {
1529                        /* Note: no rcu protection needed here
1530                         * as this is synchronous in the raid1d thread
1531                         * which is the thread that might remove
1532                         * a device.  If raid1d ever becomes multi-threaded....
1533                         */
1534                        rdev = conf->mirrors[d].rdev;
1535                        if (rdev &&
1536                            test_bit(In_sync, &rdev->flags) &&
1537                            sync_page_io(rdev->bdev,
1538                                         sect + rdev->data_offset,
1539                                         s<<9,
1540                                         conf->tmppage, READ))
1541                                success = 1;
1542                        else {
1543                                d++;
1544                                if (d == conf->raid_disks)
1545                                        d = 0;
1546                        }
1547                } while (!success && d != read_disk);
1548
1549                if (!success) {
1550                        /* Cannot read from anywhere -- bye bye array */
1551                        md_error(mddev, conf->mirrors[read_disk].rdev);
1552                        break;
1553                }
1554                /* write it back and re-read */
1555                start = d;
1556                while (d != read_disk) {
1557                        if (d==0)
1558                                d = conf->raid_disks;
1559                        d--;
1560                        rdev = conf->mirrors[d].rdev;
1561                        if (rdev &&
1562                            test_bit(In_sync, &rdev->flags)) {
1563                                if (sync_page_io(rdev->bdev,
1564                                                 sect + rdev->data_offset,
1565                                                 s<<9, conf->tmppage, WRITE)
1566                                    == 0)
1567                                        /* Well, this device is dead */
1568                                        md_error(mddev, rdev);
1569                        }
1570                }
1571                d = start;
1572                while (d != read_disk) {
1573                        char b[BDEVNAME_SIZE];
1574                        if (d==0)
1575                                d = conf->raid_disks;
1576                        d--;
1577                        rdev = conf->mirrors[d].rdev;
1578                        if (rdev &&
1579                            test_bit(In_sync, &rdev->flags)) {
1580                                if (sync_page_io(rdev->bdev,
1581                                                 sect + rdev->data_offset,
1582                                                 s<<9, conf->tmppage, READ)
1583                                    == 0)
1584                                        /* Well, this device is dead */
1585                                        md_error(mddev, rdev);
1586                                else {
1587                                        atomic_add(s, &rdev->corrected_errors);
1588                                        printk(KERN_INFO
1589                                               "md/raid1:%s: read error corrected "
1590                                               "(%d sectors at %llu on %s)\n",
1591                                               mdname(mddev), s,
1592                                               (unsigned long long)(sect +
1593                                                   rdev->data_offset),
1594                                               bdevname(rdev->bdev, b));
1595                                }
1596                        }
1597                }
1598                sectors -= s;
1599                sect += s;
1600        }
1601}
1602
1603static void raid1d(mddev_t *mddev)
1604{
1605        r1bio_t *r1_bio;
1606        struct bio *bio;
1607        unsigned long flags;
1608        conf_t *conf = mddev->private;
1609        struct list_head *head = &conf->retry_list;
1610        int unplug=0;
1611        mdk_rdev_t *rdev;
1612
1613        md_check_recovery(mddev);
1614        
1615        for (;;) {
1616                char b[BDEVNAME_SIZE];
1617
1618                unplug += flush_pending_writes(conf);
1619
1620                spin_lock_irqsave(&conf->device_lock, flags);
1621                if (list_empty(head)) {
1622                        spin_unlock_irqrestore(&conf->device_lock, flags);
1623                        break;
1624                }
1625                r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1626                list_del(head->prev);
1627                conf->nr_queued--;
1628                spin_unlock_irqrestore(&conf->device_lock, flags);
1629
1630                mddev = r1_bio->mddev;
1631                conf = mddev->private;
1632                if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1633                        sync_request_write(mddev, r1_bio);
1634                        unplug = 1;
1635                } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1636                        /* some requests in the r1bio were BIO_RW_BARRIER
1637                         * requests which failed with -EOPNOTSUPP.  Hohumm..
1638                         * Better resubmit without the barrier.
1639                         * We know which devices to resubmit for, because
1640                         * all others have had their bios[] entry cleared.
1641                         * We already have a nr_pending reference on these rdevs.
1642                         */
1643                        int i;
1644                        const bool do_sync = bio_rw_flagged(r1_bio->master_bio, BIO_RW_SYNCIO);
1645                        clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1646                        clear_bit(R1BIO_Barrier, &r1_bio->state);
1647                        for (i=0; i < conf->raid_disks; i++)
1648                                if (r1_bio->bios[i])
1649                                        atomic_inc(&r1_bio->remaining);
1650                        for (i=0; i < conf->raid_disks; i++)
1651                                if (r1_bio->bios[i]) {
1652                                        struct bio_vec *bvec;
1653                                        int j;
1654
1655                                        bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1656                                        /* copy pages from the failed bio, as
1657                                         * this might be a write-behind device */
1658                                        __bio_for_each_segment(bvec, bio, j, 0)
1659                                                bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1660                                        bio_put(r1_bio->bios[i]);
1661                                        bio->bi_sector = r1_bio->sector +
1662                                                conf->mirrors[i].rdev->data_offset;
1663                                        bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1664                                        bio->bi_end_io = raid1_end_write_request;
1665                                        bio->bi_rw = WRITE |
1666                                                (do_sync << BIO_RW_SYNCIO);
1667                                        bio->bi_private = r1_bio;
1668                                        r1_bio->bios[i] = bio;
1669                                        generic_make_request(bio);
1670                                }
1671                } else {
1672                        int disk;
1673
1674                        /* we got a read error. Maybe the drive is bad.  Maybe just
1675                         * the block and we can fix it.
1676                         * We freeze all other IO, and try reading the block from
1677                         * other devices.  When we find one, we re-write
1678                         * and check it that fixes the read error.
1679                         * This is all done synchronously while the array is
1680                         * frozen
1681                         */
1682                        if (mddev->ro == 0) {
1683                                freeze_array(conf);
1684                                fix_read_error(conf, r1_bio->read_disk,
1685                                               r1_bio->sector,
1686                                               r1_bio->sectors);
1687                                unfreeze_array(conf);
1688                        } else
1689                                md_error(mddev,
1690                                         conf->mirrors[r1_bio->read_disk].rdev);
1691
1692                        bio = r1_bio->bios[r1_bio->read_disk];
1693                        if ((disk=read_balance(conf, r1_bio)) == -1) {
1694                                printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1695                                       " read error for block %llu\n",
1696                                       mdname(mddev),
1697                                       bdevname(bio->bi_bdev,b),
1698                                       (unsigned long long)r1_bio->sector);
1699                                raid_end_bio_io(r1_bio);
1700                        } else {
1701                                const bool do_sync = bio_rw_flagged(r1_bio->master_bio, BIO_RW_SYNCIO);
1702                                r1_bio->bios[r1_bio->read_disk] =
1703                                        mddev->ro ? IO_BLOCKED : NULL;
1704                                r1_bio->read_disk = disk;
1705                                bio_put(bio);
1706                                bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1707                                r1_bio->bios[r1_bio->read_disk] = bio;
1708                                rdev = conf->mirrors[disk].rdev;
1709                                if (printk_ratelimit())
1710                                        printk(KERN_ERR "md/raid1:%s: redirecting sector %llu to"
1711                                               " other mirror: %s\n",
1712                                               mdname(mddev),
1713                                               (unsigned long long)r1_bio->sector,
1714                                               bdevname(rdev->bdev,b));
1715                                bio->bi_sector = r1_bio->sector + rdev->data_offset;
1716                                bio->bi_bdev = rdev->bdev;
1717                                bio->bi_end_io = raid1_end_read_request;
1718                                bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
1719                                bio->bi_private = r1_bio;
1720                                unplug = 1;
1721                                generic_make_request(bio);
1722                        }
1723                }
1724                cond_resched();
1725        }
1726        if (unplug)
1727                unplug_slaves(mddev);
1728}
1729
1730
1731static int init_resync(conf_t *conf)
1732{
1733        int buffs;
1734
1735        buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1736        BUG_ON(conf->r1buf_pool);
1737        conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1738                                          conf->poolinfo);
1739        if (!conf->r1buf_pool)
1740                return -ENOMEM;
1741        conf->next_resync = 0;
1742        return 0;
1743}
1744
1745/*
1746 * perform a "sync" on one "block"
1747 *
1748 * We need to make sure that no normal I/O request - particularly write
1749 * requests - conflict with active sync requests.
1750 *
1751 * This is achieved by tracking pending requests and a 'barrier' concept
1752 * that can be installed to exclude normal IO requests.
1753 */
1754
1755static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1756{
1757        conf_t *conf = mddev->private;
1758        r1bio_t *r1_bio;
1759        struct bio *bio;
1760        sector_t max_sector, nr_sectors;
1761        int disk = -1;
1762        int i;
1763        int wonly = -1;
1764        int write_targets = 0, read_targets = 0;
1765        int sync_blocks;
1766        int still_degraded = 0;
1767
1768        if (!conf->r1buf_pool)
1769                if (init_resync(conf))
1770                        return 0;
1771
1772        max_sector = mddev->dev_sectors;
1773        if (sector_nr >= max_sector) {
1774                /* If we aborted, we need to abort the
1775                 * sync on the 'current' bitmap chunk (there will
1776                 * only be one in raid1 resync.
1777                 * We can find the current addess in mddev->curr_resync
1778                 */
1779                if (mddev->curr_resync < max_sector) /* aborted */
1780                        bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1781                                                &sync_blocks, 1);
1782                else /* completed sync */
1783                        conf->fullsync = 0;
1784
1785                bitmap_close_sync(mddev->bitmap);
1786                close_sync(conf);
1787                return 0;
1788        }
1789
1790        if (mddev->bitmap == NULL &&
1791            mddev->recovery_cp == MaxSector &&
1792            !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1793            conf->fullsync == 0) {
1794                *skipped = 1;
1795                return max_sector - sector_nr;
1796        }
1797        /* before building a request, check if we can skip these blocks..
1798         * This call the bitmap_start_sync doesn't actually record anything
1799         */
1800        if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1801            !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1802                /* We can skip this block, and probably several more */
1803                *skipped = 1;
1804                return sync_blocks;
1805        }
1806        /*
1807         * If there is non-resync activity waiting for a turn,
1808         * and resync is going fast enough,
1809         * then let it though before starting on this new sync request.
1810         */
1811        if (!go_faster && conf->nr_waiting)
1812                msleep_interruptible(1000);
1813
1814        bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1815        raise_barrier(conf);
1816
1817        conf->next_resync = sector_nr;
1818
1819        r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1820        rcu_read_lock();
1821        /*
1822         * If we get a correctably read error during resync or recovery,
1823         * we might want to read from a different device.  So we
1824         * flag all drives that could conceivably be read from for READ,
1825         * and any others (which will be non-In_sync devices) for WRITE.
1826         * If a read fails, we try reading from something else for which READ
1827         * is OK.
1828         */
1829
1830        r1_bio->mddev = mddev;
1831        r1_bio->sector = sector_nr;
1832        r1_bio->state = 0;
1833        set_bit(R1BIO_IsSync, &r1_bio->state);
1834
1835        for (i=0; i < conf->raid_disks; i++) {
1836                mdk_rdev_t *rdev;
1837                bio = r1_bio->bios[i];
1838
1839                /* take from bio_init */
1840                bio->bi_next = NULL;
1841                bio->bi_flags |= 1 << BIO_UPTODATE;
1842                bio->bi_rw = READ;
1843                bio->bi_vcnt = 0;
1844                bio->bi_idx = 0;
1845                bio->bi_phys_segments = 0;
1846                bio->bi_size = 0;
1847                bio->bi_end_io = NULL;
1848                bio->bi_private = NULL;
1849
1850                rdev = rcu_dereference(conf->mirrors[i].rdev);
1851                if (rdev == NULL ||
1852                           test_bit(Faulty, &rdev->flags)) {
1853                        still_degraded = 1;
1854                        continue;
1855                } else if (!test_bit(In_sync, &rdev->flags)) {
1856                        bio->bi_rw = WRITE;
1857                        bio->bi_end_io = end_sync_write;
1858                        write_targets ++;
1859                } else {
1860                        /* may need to read from here */
1861                        bio->bi_rw = READ;
1862                        bio->bi_end_io = end_sync_read;
1863                        if (test_bit(WriteMostly, &rdev->flags)) {
1864                                if (wonly < 0)
1865                                        wonly = i;
1866                        } else {
1867                                if (disk < 0)
1868                                        disk = i;
1869                        }
1870                        read_targets++;
1871                }
1872                atomic_inc(&rdev->nr_pending);
1873                bio->bi_sector = sector_nr + rdev->data_offset;
1874                bio->bi_bdev = rdev->bdev;
1875                bio->bi_private = r1_bio;
1876        }
1877        rcu_read_unlock();
1878        if (disk < 0)
1879                disk = wonly;
1880        r1_bio->read_disk = disk;
1881
1882        if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1883                /* extra read targets are also write targets */
1884                write_targets += read_targets-1;
1885
1886        if (write_targets == 0 || read_targets == 0) {
1887                /* There is nowhere to write, so all non-sync
1888                 * drives must be failed - so we are finished
1889                 */
1890                sector_t rv = max_sector - sector_nr;
1891                *skipped = 1;
1892                put_buf(r1_bio);
1893                return rv;
1894        }
1895
1896        if (max_sector > mddev->resync_max)
1897                max_sector = mddev->resync_max; /* Don't do IO beyond here */
1898        nr_sectors = 0;
1899        sync_blocks = 0;
1900        do {
1901                struct page *page;
1902                int len = PAGE_SIZE;
1903                if (sector_nr + (len>>9) > max_sector)
1904                        len = (max_sector - sector_nr) << 9;
1905                if (len == 0)
1906                        break;
1907                if (sync_blocks == 0) {
1908                        if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1909                                               &sync_blocks, still_degraded) &&
1910                            !conf->fullsync &&
1911                            !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1912                                break;
1913                        BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1914                        if (len > (sync_blocks<<9))
1915                                len = sync_blocks<<9;
1916                }
1917
1918                for (i=0 ; i < conf->raid_disks; i++) {
1919                        bio = r1_bio->bios[i];
1920                        if (bio->bi_end_io) {
1921                                page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1922                                if (bio_add_page(bio, page, len, 0) == 0) {
1923                                        /* stop here */
1924                                        bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1925                                        while (i > 0) {
1926                                                i--;
1927                                                bio = r1_bio->bios[i];
1928                                                if (bio->bi_end_io==NULL)
1929                                                        continue;
1930                                                /* remove last page from this bio */
1931                                                bio->bi_vcnt--;
1932                                                bio->bi_size -= len;
1933                                                bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1934                                        }
1935                                        goto bio_full;
1936                                }
1937                        }
1938                }
1939                nr_sectors += len>>9;
1940                sector_nr += len>>9;
1941                sync_blocks -= (len>>9);
1942        } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1943 bio_full:
1944        r1_bio->sectors = nr_sectors;
1945
1946        /* For a user-requested sync, we read all readable devices and do a
1947         * compare
1948         */
1949        if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1950                atomic_set(&r1_bio->remaining, read_targets);
1951                for (i=0; i<conf->raid_disks; i++) {
1952                        bio = r1_bio->bios[i];
1953                        if (bio->bi_end_io == end_sync_read) {
1954                                md_sync_acct(bio->bi_bdev, nr_sectors);
1955                                generic_make_request(bio);
1956                        }
1957                }
1958        } else {
1959                atomic_set(&r1_bio->remaining, 1);
1960                bio = r1_bio->bios[r1_bio->read_disk];
1961                md_sync_acct(bio->bi_bdev, nr_sectors);
1962                generic_make_request(bio);
1963
1964        }
1965        return nr_sectors;
1966}
1967
1968static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1969{
1970        if (sectors)
1971                return sectors;
1972
1973        return mddev->dev_sectors;
1974}
1975
1976static conf_t *setup_conf(mddev_t *mddev)
1977{
1978        conf_t *conf;
1979        int i;
1980        mirror_info_t *disk;
1981        mdk_rdev_t *rdev;
1982        int err = -ENOMEM;
1983
1984        conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1985        if (!conf)
1986                goto abort;
1987
1988        conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1989                                 GFP_KERNEL);
1990        if (!conf->mirrors)
1991                goto abort;
1992
1993        conf->tmppage = alloc_page(GFP_KERNEL);
1994        if (!conf->tmppage)
1995                goto abort;
1996
1997        conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1998        if (!conf->poolinfo)
1999                goto abort;
2000        conf->poolinfo->raid_disks = mddev->raid_disks;
2001        conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2002                                          r1bio_pool_free,
2003                                          conf->poolinfo);
2004        if (!conf->r1bio_pool)
2005                goto abort;
2006
2007        conf->poolinfo->mddev = mddev;
2008
2009        spin_lock_init(&conf->device_lock);
2010        list_for_each_entry(rdev, &mddev->disks, same_set) {
2011                int disk_idx = rdev->raid_disk;
2012                if (disk_idx >= mddev->raid_disks
2013                    || disk_idx < 0)
2014                        continue;
2015                disk = conf->mirrors + disk_idx;
2016
2017                disk->rdev = rdev;
2018
2019                disk->head_position = 0;
2020        }
2021        conf->raid_disks = mddev->raid_disks;
2022        conf->mddev = mddev;
2023        INIT_LIST_HEAD(&conf->retry_list);
2024
2025        spin_lock_init(&conf->resync_lock);
2026        init_waitqueue_head(&conf->wait_barrier);
2027
2028        bio_list_init(&conf->pending_bio_list);
2029        bio_list_init(&conf->flushing_bio_list);
2030
2031        conf->last_used = -1;
2032        for (i = 0; i < conf->raid_disks; i++) {
2033
2034                disk = conf->mirrors + i;
2035
2036                if (!disk->rdev ||
2037                    !test_bit(In_sync, &disk->rdev->flags)) {
2038                        disk->head_position = 0;
2039                        if (disk->rdev)
2040                                conf->fullsync = 1;
2041                } else if (conf->last_used < 0)
2042                        /*
2043                         * The first working device is used as a
2044                         * starting point to read balancing.
2045                         */
2046                        conf->last_used = i;
2047        }
2048
2049        err = -EIO;
2050        if (conf->last_used < 0) {
2051                printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
2052                       mdname(mddev));
2053                goto abort;
2054        }
2055        err = -ENOMEM;
2056        conf->thread = md_register_thread(raid1d, mddev, NULL);
2057        if (!conf->thread) {
2058                printk(KERN_ERR
2059                       "md/raid1:%s: couldn't allocate thread\n",
2060                       mdname(mddev));
2061                goto abort;
2062        }
2063
2064        return conf;
2065
2066 abort:
2067        if (conf) {
2068                if (conf->r1bio_pool)
2069                        mempool_destroy(conf->r1bio_pool);
2070                kfree(conf->mirrors);
2071                safe_put_page(conf->tmppage);
2072                kfree(conf->poolinfo);
2073                kfree(conf);
2074        }
2075        return ERR_PTR(err);
2076}
2077
2078static int run(mddev_t *mddev)
2079{
2080        conf_t *conf;
2081        int i;
2082        mdk_rdev_t *rdev;
2083
2084        if (mddev->level != 1) {
2085                printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2086                       mdname(mddev), mddev->level);
2087                return -EIO;
2088        }
2089        if (mddev->reshape_position != MaxSector) {
2090                printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2091                       mdname(mddev));
2092                return -EIO;
2093        }
2094        /*
2095         * copy the already verified devices into our private RAID1
2096         * bookkeeping area. [whatever we allocate in run(),
2097         * should be freed in stop()]
2098         */
2099        if (mddev->private == NULL)
2100                conf = setup_conf(mddev);
2101        else
2102                conf = mddev->private;
2103
2104        if (IS_ERR(conf))
2105                return PTR_ERR(conf);
2106
2107        mddev->queue->queue_lock = &conf->device_lock;
2108        list_for_each_entry(rdev, &mddev->disks, same_set) {
2109                disk_stack_limits(mddev->gendisk, rdev->bdev,
2110                                  rdev->data_offset << 9);
2111                /* as we don't honour merge_bvec_fn, we must never risk
2112                 * violating it, so limit ->max_segments to 1 lying within
2113                 * a single page, as a one page request is never in violation.
2114                 */
2115                if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2116                        blk_queue_max_segments(mddev->queue, 1);
2117                        blk_queue_segment_boundary(mddev->queue,
2118                                                   PAGE_CACHE_SIZE - 1);
2119                }
2120        }
2121
2122        mddev->degraded = 0;
2123        for (i=0; i < conf->raid_disks; i++)
2124                if (conf->mirrors[i].rdev == NULL ||
2125                    !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2126                    test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2127                        mddev->degraded++;
2128
2129        if (conf->raid_disks - mddev->degraded == 1)
2130                mddev->recovery_cp = MaxSector;
2131
2132        if (mddev->recovery_cp != MaxSector)
2133                printk(KERN_NOTICE "md/raid1:%s: not clean"
2134                       " -- starting background reconstruction\n",
2135                       mdname(mddev));
2136        printk(KERN_INFO 
2137                "md/raid1:%s: active with %d out of %d mirrors\n",
2138                mdname(mddev), mddev->raid_disks - mddev->degraded, 
2139                mddev->raid_disks);
2140
2141        /*
2142         * Ok, everything is just fine now
2143         */
2144        mddev->thread = conf->thread;
2145        conf->thread = NULL;
2146        mddev->private = conf;
2147
2148        md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2149
2150        mddev->queue->unplug_fn = raid1_unplug;
2151        mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2152        mddev->queue->backing_dev_info.congested_data = mddev;
2153        md_integrity_register(mddev);
2154        return 0;
2155}
2156
2157static int stop(mddev_t *mddev)
2158{
2159        conf_t *conf = mddev->private;
2160        struct bitmap *bitmap = mddev->bitmap;
2161
2162        /* wait for behind writes to complete */
2163        if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2164                printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2165                       mdname(mddev));
2166                /* need to kick something here to make sure I/O goes? */
2167                wait_event(bitmap->behind_wait,
2168                           atomic_read(&bitmap->behind_writes) == 0);
2169        }
2170
2171        raise_barrier(conf);
2172        lower_barrier(conf);
2173
2174        md_unregister_thread(mddev->thread);
2175        mddev->thread = NULL;
2176        blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2177        if (conf->r1bio_pool)
2178                mempool_destroy(conf->r1bio_pool);
2179        kfree(conf->mirrors);
2180        kfree(conf->poolinfo);
2181        kfree(conf);
2182        mddev->private = NULL;
2183        return 0;
2184}
2185
2186static int raid1_resize(mddev_t *mddev, sector_t sectors)
2187{
2188        /* no resync is happening, and there is enough space
2189         * on all devices, so we can resize.
2190         * We need to make sure resync covers any new space.
2191         * If the array is shrinking we should possibly wait until
2192         * any io in the removed space completes, but it hardly seems
2193         * worth it.
2194         */
2195        md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2196        if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2197                return -EINVAL;
2198        set_capacity(mddev->gendisk, mddev->array_sectors);
2199        revalidate_disk(mddev->gendisk);
2200        if (sectors > mddev->dev_sectors &&
2201            mddev->recovery_cp == MaxSector) {
2202                mddev->recovery_cp = mddev->dev_sectors;
2203                set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2204        }
2205        mddev->dev_sectors = sectors;
2206        mddev->resync_max_sectors = sectors;
2207        return 0;
2208}
2209
2210static int raid1_reshape(mddev_t *mddev)
2211{
2212        /* We need to:
2213         * 1/ resize the r1bio_pool
2214         * 2/ resize conf->mirrors
2215         *
2216         * We allocate a new r1bio_pool if we can.
2217         * Then raise a device barrier and wait until all IO stops.
2218         * Then resize conf->mirrors and swap in the new r1bio pool.
2219         *
2220         * At the same time, we "pack" the devices so that all the missing
2221         * devices have the higher raid_disk numbers.
2222         */
2223        mempool_t *newpool, *oldpool;
2224        struct pool_info *newpoolinfo;
2225        mirror_info_t *newmirrors;
2226        conf_t *conf = mddev->private;
2227        int cnt, raid_disks;
2228        unsigned long flags;
2229        int d, d2, err;
2230
2231        /* Cannot change chunk_size, layout, or level */
2232        if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2233            mddev->layout != mddev->new_layout ||
2234            mddev->level != mddev->new_level) {
2235                mddev->new_chunk_sectors = mddev->chunk_sectors;
2236                mddev->new_layout = mddev->layout;
2237                mddev->new_level = mddev->level;
2238                return -EINVAL;
2239        }
2240
2241        err = md_allow_write(mddev);
2242        if (err)
2243                return err;
2244
2245        raid_disks = mddev->raid_disks + mddev->delta_disks;
2246
2247        if (raid_disks < conf->raid_disks) {
2248                cnt=0;
2249                for (d= 0; d < conf->raid_disks; d++)
2250                        if (conf->mirrors[d].rdev)
2251                                cnt++;
2252                if (cnt > raid_disks)
2253                        return -EBUSY;
2254        }
2255
2256        newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2257        if (!newpoolinfo)
2258                return -ENOMEM;
2259        newpoolinfo->mddev = mddev;
2260        newpoolinfo->raid_disks = raid_disks;
2261
2262        newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2263                                 r1bio_pool_free, newpoolinfo);
2264        if (!newpool) {
2265                kfree(newpoolinfo);
2266                return -ENOMEM;
2267        }
2268        newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2269        if (!newmirrors) {
2270                kfree(newpoolinfo);
2271                mempool_destroy(newpool);
2272                return -ENOMEM;
2273        }
2274
2275        raise_barrier(conf);
2276
2277        /* ok, everything is stopped */
2278        oldpool = conf->r1bio_pool;
2279        conf->r1bio_pool = newpool;
2280
2281        for (d = d2 = 0; d < conf->raid_disks; d++) {
2282                mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2283                if (rdev && rdev->raid_disk != d2) {
2284                        char nm[20];
2285                        sprintf(nm, "rd%d", rdev->raid_disk);
2286                        sysfs_remove_link(&mddev->kobj, nm);
2287                        rdev->raid_disk = d2;
2288                        sprintf(nm, "rd%d", rdev->raid_disk);
2289                        sysfs_remove_link(&mddev->kobj, nm);
2290                        if (sysfs_create_link(&mddev->kobj,
2291                                              &rdev->kobj, nm))
2292                                printk(KERN_WARNING
2293                                       "md/raid1:%s: cannot register "
2294                                       "%s\n",
2295                                       mdname(mddev), nm);
2296                }
2297                if (rdev)
2298                        newmirrors[d2++].rdev = rdev;
2299        }
2300        kfree(conf->mirrors);
2301        conf->mirrors = newmirrors;
2302        kfree(conf->poolinfo);
2303        conf->poolinfo = newpoolinfo;
2304
2305        spin_lock_irqsave(&conf->device_lock, flags);
2306        mddev->degraded += (raid_disks - conf->raid_disks);
2307        spin_unlock_irqrestore(&conf->device_lock, flags);
2308        conf->raid_disks = mddev->raid_disks = raid_disks;
2309        mddev->delta_disks = 0;
2310
2311        conf->last_used = 0; /* just make sure it is in-range */
2312        lower_barrier(conf);
2313
2314        set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2315        md_wakeup_thread(mddev->thread);
2316
2317        mempool_destroy(oldpool);
2318        return 0;
2319}
2320
2321static void raid1_quiesce(mddev_t *mddev, int state)
2322{
2323        conf_t *conf = mddev->private;
2324
2325        switch(state) {
2326        case 2: /* wake for suspend */
2327                wake_up(&conf->wait_barrier);
2328                break;
2329        case 1:
2330                raise_barrier(conf);
2331                break;
2332        case 0:
2333                lower_barrier(conf);
2334                break;
2335        }
2336}
2337
2338static void *raid1_takeover(mddev_t *mddev)
2339{
2340        /* raid1 can take over:
2341         *  raid5 with 2 devices, any layout or chunk size
2342         */
2343        if (mddev->level == 5 && mddev->raid_disks == 2) {
2344                conf_t *conf;
2345                mddev->new_level = 1;
2346                mddev->new_layout = 0;
2347                mddev->new_chunk_sectors = 0;
2348                conf = setup_conf(mddev);
2349                if (!IS_ERR(conf))
2350                        conf->barrier = 1;
2351                return conf;
2352        }
2353        return ERR_PTR(-EINVAL);
2354}
2355
2356static struct mdk_personality raid1_personality =
2357{
2358        .name           = "raid1",
2359        .level          = 1,
2360        .owner          = THIS_MODULE,
2361        .make_request   = make_request,
2362        .run            = run,
2363        .stop           = stop,
2364        .status         = status,
2365        .error_handler  = error,
2366        .hot_add_disk   = raid1_add_disk,
2367        .hot_remove_disk= raid1_remove_disk,
2368        .spare_active   = raid1_spare_active,
2369        .sync_request   = sync_request,
2370        .resize         = raid1_resize,
2371        .size           = raid1_size,
2372        .check_reshape  = raid1_reshape,
2373        .quiesce        = raid1_quiesce,
2374        .takeover       = raid1_takeover,
2375};
2376
2377static int __init raid_init(void)
2378{
2379        return register_md_personality(&raid1_personality);
2380}
2381
2382static void raid_exit(void)
2383{
2384        unregister_md_personality(&raid1_personality);
2385}
2386
2387module_init(raid_init);
2388module_exit(raid_exit);
2389MODULE_LICENSE("GPL");
2390MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2391MODULE_ALIAS("md-personality-3"); /* RAID1 */
2392MODULE_ALIAS("md-raid1");
2393MODULE_ALIAS("md-level-1");
2394
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.