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