block/mirror.c

   1 /*
   2  * Image mirroring
   3  *
   4  * Copyright Red Hat, Inc. 2012
   5  *
   6  * Authors:
   7  *  Paolo Bonzini  <pbonzini@redhat.com>
   8  *
   9  * This work is licensed under the terms of the GNU LGPL, version 2 or later.
  10  * See the COPYING.LIB file in the top-level directory.
  11  *
  12  */
  13
  14 #include "trace.h"
  15 #include "block/blockjob.h"
  16 #include "block/block_int.h"
  17 #include "qemu/ratelimit.h"
  18 #include "qemu/bitmap.h"
  19
  20 #define SLICE_TIME    100000000ULL /* ns */
  21 #define MAX_IN_FLIGHT 16
  22
  23 /* The mirroring buffer is a list of granularity-sized chunks.
  24  * Free chunks are organized in a list.
  25  */
  26 typedef struct MirrorBuffer {
  27     QSIMPLEQ_ENTRY(MirrorBuffer) next;
  28 } MirrorBuffer;
  29
  30 typedef struct MirrorBlockJob {
  31     BlockJob common;
  32     RateLimit limit;
  33     BlockDriverState *target;
  34     MirrorSyncMode mode;
  35     BlockdevOnError on_source_error, on_target_error;
  36     bool synced;
  37     bool should_complete;
  38     int64_t sector_num;
  39     int64_t granularity;
  40     size_t buf_size;
  41     unsigned long *cow_bitmap;
  42     BdrvDirtyBitmap *dirty_bitmap;
  43     HBitmapIter hbi;
  44     uint8_t *buf;
  45     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
  46     int buf_free_count;
  47
  48     unsigned long *in_flight_bitmap;
  49     int in_flight;
  50     int ret;
  51 } MirrorBlockJob;
  52
  53 typedef struct MirrorOp {
  54     MirrorBlockJob *s;
  55     QEMUIOVector qiov;
  56     int64_t sector_num;
  57     int nb_sectors;
  58 } MirrorOp;
  59
  60 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
  61                                             int error)
  62 {
  63     s->synced = false;
  64     if (read) {
  65         return block_job_error_action(&s->common, s->common.bs,
  66                                       s->on_source_error, true, error);
  67     } else {
  68         return block_job_error_action(&s->common, s->target,
  69                                       s->on_target_error, false, error);
  70     }
  71 }
  72
  73 static void mirror_iteration_done(MirrorOp *op, int ret)
  74 {
  75     MirrorBlockJob *s = op->s;
  76     struct iovec *iov;
  77     int64_t chunk_num;
  78     int i, nb_chunks, sectors_per_chunk;
  79
  80     trace_mirror_iteration_done(s, op->sector_num, op->nb_sectors, ret);
  81
  82     s->in_flight--;
  83     iov = op->qiov.iov;
  84     for (i = 0; i < op->qiov.niov; i++) {
  85         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
  86         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
  87         s->buf_free_count++;
  88     }
  89
  90     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
  91     chunk_num = op->sector_num / sectors_per_chunk;
  92     nb_chunks = op->nb_sectors / sectors_per_chunk;
  93     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
  94     if (s->cow_bitmap && ret >= 0) {
  95         bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
  96     }
  97
  98     g_slice_free(MirrorOp, op);
  99     qemu_coroutine_enter(s->common.co, NULL);
 100 }
 101
 102 static void mirror_write_complete(void *opaque, int ret)
 103 {
 104     MirrorOp *op = opaque;
 105     MirrorBlockJob *s = op->s;
 106     if (ret < 0) {
 107         BlockDriverState *source = s->common.bs;
 108         BlockErrorAction action;
 109
 110         bdrv_set_dirty(source, op->sector_num, op->nb_sectors);
 111         action = mirror_error_action(s, false, -ret);
 112         if (action == BDRV_ACTION_REPORT && s->ret >= 0) {
 113             s->ret = ret;
 114         }
 115     }
 116     mirror_iteration_done(op, ret);
 117 }
 118
 119 static void mirror_read_complete(void *opaque, int ret)
 120 {
 121     MirrorOp *op = opaque;
 122     MirrorBlockJob *s = op->s;
 123     if (ret < 0) {
 124         BlockDriverState *source = s->common.bs;
 125         BlockErrorAction action;
 126
 127         bdrv_set_dirty(source, op->sector_num, op->nb_sectors);
 128         action = mirror_error_action(s, true, -ret);
 129         if (action == BDRV_ACTION_REPORT && s->ret >= 0) {
 130             s->ret = ret;
 131         }
 132
 133         mirror_iteration_done(op, ret);
 134         return;
 135     }
 136     bdrv_aio_writev(s->target, op->sector_num, &op->qiov, op->nb_sectors,
 137                     mirror_write_complete, op);
 138 }
 139
 140 static void coroutine_fn mirror_iteration(MirrorBlockJob *s)
 141 {
 142     BlockDriverState *source = s->common.bs;
 143     int nb_sectors, sectors_per_chunk, nb_chunks;
 144     int64_t end, sector_num, next_chunk, next_sector, hbitmap_next_sector;
 145     MirrorOp *op;
 146
 147     s->sector_num = hbitmap_iter_next(&s->hbi);
 148     if (s->sector_num < 0) {
 149         bdrv_dirty_iter_init(source, s->dirty_bitmap, &s->hbi);
 150         s->sector_num = hbitmap_iter_next(&s->hbi);
 151         trace_mirror_restart_iter(s,
 152                                   bdrv_get_dirty_count(source, s->dirty_bitmap));
 153         assert(s->sector_num >= 0);
 154     }
 155
 156     hbitmap_next_sector = s->sector_num;
 157     sector_num = s->sector_num;
 158     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 159     end = s->common.len >> BDRV_SECTOR_BITS;
 160
 161     /* Extend the QEMUIOVector to include all adjacent blocks that will
 162      * be copied in this operation.
 163      *
 164      * We have to do this if we have no backing file yet in the destination,
 165      * and the cluster size is very large.  Then we need to do COW ourselves.
 166      * The first time a cluster is copied, copy it entirely.  Note that,
 167      * because both the granularity and the cluster size are powers of two,
 168      * the number of sectors to copy cannot exceed one cluster.
 169      *
 170      * We also want to extend the QEMUIOVector to include more adjacent
 171      * dirty blocks if possible, to limit the number of I/O operations and
 172      * run efficiently even with a small granularity.
 173      */
 174     nb_chunks = 0;
 175     nb_sectors = 0;
 176     next_sector = sector_num;
 177     next_chunk = sector_num / sectors_per_chunk;
 178
 179     /* Wait for I/O to this cluster (from a previous iteration) to be done.  */
 180     while (test_bit(next_chunk, s->in_flight_bitmap)) {
 181         trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
 182         qemu_coroutine_yield();
 183     }
 184
 185     do {
 186         int added_sectors, added_chunks;
 187
 188         if (!bdrv_get_dirty(source, s->dirty_bitmap, next_sector) ||
 189             test_bit(next_chunk, s->in_flight_bitmap)) {
 190             assert(nb_sectors > 0);
 191             break;
 192         }
 193
 194         added_sectors = sectors_per_chunk;
 195         if (s->cow_bitmap && !test_bit(next_chunk, s->cow_bitmap)) {
 196             bdrv_round_to_clusters(s->target,
 197                                    next_sector, added_sectors,
 198                                    &next_sector, &added_sectors);
 199
 200             /* On the first iteration, the rounding may make us copy
 201              * sectors before the first dirty one.
 202              */
 203             if (next_sector < sector_num) {
 204                 assert(nb_sectors == 0);
 205                 sector_num = next_sector;
 206                 next_chunk = next_sector / sectors_per_chunk;
 207             }
 208         }
 209
 210         added_sectors = MIN(added_sectors, end - (sector_num + nb_sectors));
 211         added_chunks = (added_sectors + sectors_per_chunk - 1) / sectors_per_chunk;
 212
 213         /* When doing COW, it may happen that there is not enough space for
 214          * a full cluster.  Wait if that is the case.
 215          */
 216         while (nb_chunks == 0 && s->buf_free_count < added_chunks) {
 217             trace_mirror_yield_buf_busy(s, nb_chunks, s->in_flight);
 218             qemu_coroutine_yield();
 219         }
 220         if (s->buf_free_count < nb_chunks + added_chunks) {
 221             trace_mirror_break_buf_busy(s, nb_chunks, s->in_flight);
 222             break;
 223         }
 224
 225         /* We have enough free space to copy these sectors.  */
 226         bitmap_set(s->in_flight_bitmap, next_chunk, added_chunks);
 227
 228         nb_sectors += added_sectors;
 229         nb_chunks += added_chunks;
 230         next_sector += added_sectors;
 231         next_chunk += added_chunks;
 232     } while (next_sector < end);
 233
 234     /* Allocate a MirrorOp that is used as an AIO callback.  */
 235     op = g_slice_new(MirrorOp);
 236     op->s = s;
 237     op->sector_num = sector_num;
 238     op->nb_sectors = nb_sectors;
 239
 240     /* Now make a QEMUIOVector taking enough granularity-sized chunks
 241      * from s->buf_free.
 242      */
 243     qemu_iovec_init(&op->qiov, nb_chunks);
 244     next_sector = sector_num;
 245     while (nb_chunks-- > 0) {
 246         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
 247         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
 248         s->buf_free_count--;
 249         qemu_iovec_add(&op->qiov, buf, s->granularity);
 250
 251         /* Advance the HBitmapIter in parallel, so that we do not examine
 252          * the same sector twice.
 253          */
 254         if (next_sector > hbitmap_next_sector
 255             && bdrv_get_dirty(source, s->dirty_bitmap, next_sector)) {
 256             hbitmap_next_sector = hbitmap_iter_next(&s->hbi);
 257         }
 258
 259         next_sector += sectors_per_chunk;
 260     }
 261
 262     bdrv_reset_dirty(source, sector_num, nb_sectors);
 263
 264     /* Copy the dirty cluster.  */
 265     s->in_flight++;
 266     trace_mirror_one_iteration(s, sector_num, nb_sectors);
 267     bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors,
 268                    mirror_read_complete, op);
 269 }
 270
 271 static void mirror_free_init(MirrorBlockJob *s)
 272 {
 273     int granularity = s->granularity;
 274     size_t buf_size = s->buf_size;
 275     uint8_t *buf = s->buf;
 276
 277     assert(s->buf_free_count == 0);
 278     QSIMPLEQ_INIT(&s->buf_free);
 279     while (buf_size != 0) {
 280         MirrorBuffer *cur = (MirrorBuffer *)buf;
 281         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
 282         s->buf_free_count++;
 283         buf_size -= granularity;
 284         buf += granularity;
 285     }
 286 }
 287
 288 static void mirror_drain(MirrorBlockJob *s)
 289 {
 290     while (s->in_flight > 0) {
 291         qemu_coroutine_yield();
 292     }
 293 }
 294
 295 static void coroutine_fn mirror_run(void *opaque)
 296 {
 297     MirrorBlockJob *s = opaque;
 298     BlockDriverState *bs = s->common.bs;
 299     int64_t sector_num, end, sectors_per_chunk, length;
 300     uint64_t last_pause_ns;
 301     BlockDriverInfo bdi;
 302     char backing_filename[1024];
 303     int ret = 0;
 304     int n;
 305
 306     if (block_job_is_cancelled(&s->common)) {
 307         goto immediate_exit;
 308     }
 309
 310     s->common.len = bdrv_getlength(bs);
 311     if (s->common.len <= 0) {
 312         block_job_completed(&s->common, s->common.len);
 313         return;
 314     }
 315
 316     length = (bdrv_getlength(bs) + s->granularity - 1) / s->granularity;
 317     s->in_flight_bitmap = bitmap_new(length);
 318
 319     /* If we have no backing file yet in the destination, we cannot let
 320      * the destination do COW.  Instead, we copy sectors around the
 321      * dirty data if needed.  We need a bitmap to do that.
 322      */
 323     bdrv_get_backing_filename(s->target, backing_filename,
 324                               sizeof(backing_filename));
 325     if (backing_filename[0] && !s->target->backing_hd) {
 326         bdrv_get_info(s->target, &bdi);
 327         if (s->granularity < bdi.cluster_size) {
 328             s->buf_size = MAX(s->buf_size, bdi.cluster_size);
 329             s->cow_bitmap = bitmap_new(length);
 330         }
 331     }
 332
 333     end = s->common.len >> BDRV_SECTOR_BITS;
 334     s->buf = qemu_blockalign(bs, s->buf_size);
 335     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 336     mirror_free_init(s);
 337
 338     if (s->mode != MIRROR_SYNC_MODE_NONE) {
 339         /* First part, loop on the sectors and initialize the dirty bitmap.  */
 340         BlockDriverState *base;
 341         base = s->mode == MIRROR_SYNC_MODE_FULL ? NULL : bs->backing_hd;
 342         for (sector_num = 0; sector_num < end; ) {
 343             int64_t next = (sector_num | (sectors_per_chunk - 1)) + 1;
 344             ret = bdrv_is_allocated_above(bs, base,
 345                                           sector_num, next - sector_num, &n);
 346
 347             if (ret < 0) {
 348                 goto immediate_exit;
 349             }
 350
 351             assert(n > 0);
 352             if (ret == 1) {
 353                 bdrv_set_dirty(bs, sector_num, n);
 354                 sector_num = next;
 355             } else {
 356                 sector_num += n;
 357             }
 358         }
 359     }
 360
 361     bdrv_dirty_iter_init(bs, s->dirty_bitmap, &s->hbi);
 362     last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 363     for (;;) {
 364         uint64_t delay_ns;
 365         int64_t cnt;
 366         bool should_complete;
 367
 368         if (s->ret < 0) {
 369             ret = s->ret;
 370             goto immediate_exit;
 371         }
 372
 373         cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);
 374
 375         /* Note that even when no rate limit is applied we need to yield
 376          * periodically with no pending I/O so that qemu_aio_flush() returns.
 377          * We do so every SLICE_TIME nanoseconds, or when there is an error,
 378          * or when the source is clean, whichever comes first.
 379          */
 380         if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - last_pause_ns < SLICE_TIME &&
 381             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
 382             if (s->in_flight == MAX_IN_FLIGHT || s->buf_free_count == 0 ||
 383                 (cnt == 0 && s->in_flight > 0)) {
 384                 trace_mirror_yield(s, s->in_flight, s->buf_free_count, cnt);
 385                 qemu_coroutine_yield();
 386                 continue;
 387             } else if (cnt != 0) {
 388                 mirror_iteration(s);
 389                 continue;
 390             }
 391         }
 392
 393         should_complete = false;
 394         if (s->in_flight == 0 && cnt == 0) {
 395             trace_mirror_before_flush(s);
 396             ret = bdrv_flush(s->target);
 397             if (ret < 0) {
 398                 if (mirror_error_action(s, false, -ret) == BDRV_ACTION_REPORT) {
 399                     goto immediate_exit;
 400                 }
 401             } else {
 402                 /* We're out of the streaming phase.  From now on, if the job
 403                  * is cancelled we will actually complete all pending I/O and
 404                  * report completion.  This way, block-job-cancel will leave
 405                  * the target in a consistent state.
 406                  */
 407                 s->common.offset = end * BDRV_SECTOR_SIZE;
 408                 if (!s->synced) {
 409                     block_job_ready(&s->common);
 410                     s->synced = true;
 411                 }
 412
 413                 should_complete = s->should_complete ||
 414                     block_job_is_cancelled(&s->common);
 415                 cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);
 416             }
 417         }
 418
 419         if (cnt == 0 && should_complete) {
 420             /* The dirty bitmap is not updated while operations are pending.
 421              * If we're about to exit, wait for pending operations before
 422              * calling bdrv_get_dirty_count(bs), or we may exit while the
 423              * source has dirty data to copy!
 424              *
 425              * Note that I/O can be submitted by the guest while
 426              * mirror_populate runs.
 427              */
 428             trace_mirror_before_drain(s, cnt);
 429             bdrv_drain_all();
 430             cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);
 431         }
 432
 433         ret = 0;
 434         trace_mirror_before_sleep(s, cnt, s->synced);
 435         if (!s->synced) {
 436             /* Publish progress */
 437             s->common.offset = (end - cnt) * BDRV_SECTOR_SIZE;
 438
 439             if (s->common.speed) {
 440                 delay_ns = ratelimit_calculate_delay(&s->limit, sectors_per_chunk);
 441             } else {
 442                 delay_ns = 0;
 443             }
 444
 445             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 446             if (block_job_is_cancelled(&s->common)) {
 447                 break;
 448             }
 449         } else if (!should_complete) {
 450             delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);
 451             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 452         } else if (cnt == 0) {
 453             /* The two disks are in sync.  Exit and report successful
 454              * completion.
 455              */
 456             assert(QLIST_EMPTY(&bs->tracked_requests));
 457             s->common.cancelled = false;
 458             break;
 459         }
 460         last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 461     }
 462
 463 immediate_exit:
 464     if (s->in_flight > 0) {
 465         /* We get here only if something went wrong.  Either the job failed,
 466          * or it was cancelled prematurely so that we do not guarantee that
 467          * the target is a copy of the source.
 468          */
 469         assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common)));
 470         mirror_drain(s);
 471     }
 472
 473     assert(s->in_flight == 0);
 474     qemu_vfree(s->buf);
 475     g_free(s->cow_bitmap);
 476     g_free(s->in_flight_bitmap);
 477     bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
 478     bdrv_iostatus_disable(s->target);
 479     if (s->should_complete && ret == 0) {
 480         if (bdrv_get_flags(s->target) != bdrv_get_flags(s->common.bs)) {
 481             bdrv_reopen(s->target, bdrv_get_flags(s->common.bs), NULL);
 482         }
 483         bdrv_swap(s->target, s->common.bs);
 484     }
 485     bdrv_close(s->target);
 486     bdrv_unref(s->target);
 487     block_job_completed(&s->common, ret);
 488 }
 489
 490 static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp)
 491 {
 492     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 493
 494     if (speed < 0) {
 495         error_set(errp, QERR_INVALID_PARAMETER, "speed");
 496         return;
 497     }
 498     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 499 }
 500
 501 static void mirror_iostatus_reset(BlockJob *job)
 502 {
 503     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 504
 505     bdrv_iostatus_reset(s->target);
 506 }
 507
 508 static void mirror_complete(BlockJob *job, Error **errp)
 509 {
 510     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 511     Error *local_err = NULL;
 512     int ret;
 513
 514     ret = bdrv_open_backing_file(s->target, NULL, &local_err);
 515     if (ret < 0) {
 516         char backing_filename[PATH_MAX];
 517         bdrv_get_full_backing_filename(s->target, backing_filename,
 518                                        sizeof(backing_filename));
 519         error_propagate(errp, local_err);
 520         return;
 521     }
 522     if (!s->synced) {
 523         error_set(errp, QERR_BLOCK_JOB_NOT_READY, job->bs->device_name);
 524         return;
 525     }
 526
 527     s->should_complete = true;
 528     block_job_resume(job);
 529 }
 530
 531 static const BlockJobDriver mirror_job_driver = {
 532     .instance_size = sizeof(MirrorBlockJob),
 533     .job_type      = BLOCK_JOB_TYPE_MIRROR,
 534     .set_speed     = mirror_set_speed,
 535     .iostatus_reset= mirror_iostatus_reset,
 536     .complete      = mirror_complete,
 537 };
 538
 539 void mirror_start(BlockDriverState *bs, BlockDriverState *target,
 540                   int64_t speed, int64_t granularity, int64_t buf_size,
 541                   MirrorSyncMode mode, BlockdevOnError on_source_error,
 542                   BlockdevOnError on_target_error,
 543                   BlockDriverCompletionFunc *cb,
 544                   void *opaque, Error **errp)
 545 {
 546     MirrorBlockJob *s;
 547
 548     if (granularity == 0) {
 549         /* Choose the default granularity based on the target file's cluster
 550          * size, clamped between 4k and 64k.  */
 551         BlockDriverInfo bdi;
 552         if (bdrv_get_info(target, &bdi) >= 0 && bdi.cluster_size != 0) {
 553             granularity = MAX(4096, bdi.cluster_size);
 554             granularity = MIN(65536, granularity);
 555         } else {
 556             granularity = 65536;
 557         }
 558     }
 559
 560     assert ((granularity & (granularity - 1)) == 0);
 561
 562     if ((on_source_error == BLOCKDEV_ON_ERROR_STOP ||
 563          on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
 564         !bdrv_iostatus_is_enabled(bs)) {
 565         error_set(errp, QERR_INVALID_PARAMETER, "on-source-error");
 566         return;
 567     }
 568
 569     s = block_job_create(&mirror_job_driver, bs, speed, cb, opaque, errp);
 570     if (!s) {
 571         return;
 572     }
 573
 574     s->on_source_error = on_source_error;
 575     s->on_target_error = on_target_error;
 576     s->target = target;
 577     s->mode = mode;
 578     s->granularity = granularity;
 579     s->buf_size = MAX(buf_size, granularity);
 580
 581     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity);
 582     bdrv_set_enable_write_cache(s->target, true);
 583     bdrv_set_on_error(s->target, on_target_error, on_target_error);
 584     bdrv_iostatus_enable(s->target);
 585     s->common.co = qemu_coroutine_create(mirror_run);
 586     trace_mirror_start(bs, s, s->common.co, opaque);
 587     qemu_coroutine_enter(s->common.co, s);
 588 }