4 * Copyright Red Hat, Inc. 2012
7 * Paolo Bonzini <pbonzini@redhat.com>
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.
15 #include "block/blockjob.h"
16 #include "block/block_int.h"
17 #include "qemu/ratelimit.h"
18 #include "qemu/bitmap.h"
20 #define SLICE_TIME 100000000ULL /* ns */
21 #define MAX_IN_FLIGHT 16
23 /* The mirroring buffer is a list of granularity-sized chunks.
24 * Free chunks are organized in a list.
26 typedef struct MirrorBuffer {
27 QSIMPLEQ_ENTRY(MirrorBuffer) next;
30 typedef struct MirrorBlockJob {
33 BlockDriverState *target;
34 BlockDriverState *base;
35 /* The name of the graph node to replace */
37 /* The BDS to replace */
38 BlockDriverState *to_replace;
39 /* Used to block operations on the drive-mirror-replace target */
40 Error *replace_blocker;
42 BlockdevOnError on_source_error, on_target_error;
49 unsigned long *cow_bitmap;
50 BdrvDirtyBitmap *dirty_bitmap;
53 QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
56 unsigned long *in_flight_bitmap;
58 int sectors_in_flight;
62 typedef struct MirrorOp {
69 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
74 return block_job_error_action(&s->common, s->common.bs,
75 s->on_source_error, true, error);
77 return block_job_error_action(&s->common, s->target,
78 s->on_target_error, false, error);
82 static void mirror_iteration_done(MirrorOp *op, int ret)
84 MirrorBlockJob *s = op->s;
87 int i, nb_chunks, sectors_per_chunk;
89 trace_mirror_iteration_done(s, op->sector_num, op->nb_sectors, ret);
92 s->sectors_in_flight -= op->nb_sectors;
94 for (i = 0; i < op->qiov.niov; i++) {
95 MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
96 QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
100 sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
101 chunk_num = op->sector_num / sectors_per_chunk;
102 nb_chunks = op->nb_sectors / sectors_per_chunk;
103 bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
106 bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
108 s->common.offset += (uint64_t)op->nb_sectors * BDRV_SECTOR_SIZE;
111 qemu_iovec_destroy(&op->qiov);
112 g_slice_free(MirrorOp, op);
114 /* Enter coroutine when it is not sleeping. The coroutine sleeps to
115 * rate-limit itself. The coroutine will eventually resume since there is
116 * a sleep timeout so don't wake it early.
118 if (s->common.busy) {
119 qemu_coroutine_enter(s->common.co, NULL);
123 static void mirror_write_complete(void *opaque, int ret)
125 MirrorOp *op = opaque;
126 MirrorBlockJob *s = op->s;
128 BlockDriverState *source = s->common.bs;
129 BlockErrorAction action;
131 bdrv_set_dirty(source, op->sector_num, op->nb_sectors);
132 action = mirror_error_action(s, false, -ret);
133 if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
137 mirror_iteration_done(op, ret);
140 static void mirror_read_complete(void *opaque, int ret)
142 MirrorOp *op = opaque;
143 MirrorBlockJob *s = op->s;
145 BlockDriverState *source = s->common.bs;
146 BlockErrorAction action;
148 bdrv_set_dirty(source, op->sector_num, op->nb_sectors);
149 action = mirror_error_action(s, true, -ret);
150 if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
154 mirror_iteration_done(op, ret);
157 bdrv_aio_writev(s->target, op->sector_num, &op->qiov, op->nb_sectors,
158 mirror_write_complete, op);
161 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
163 BlockDriverState *source = s->common.bs;
164 int nb_sectors, sectors_per_chunk, nb_chunks;
165 int64_t end, sector_num, next_chunk, next_sector, hbitmap_next_sector;
166 uint64_t delay_ns = 0;
169 s->sector_num = hbitmap_iter_next(&s->hbi);
170 if (s->sector_num < 0) {
171 bdrv_dirty_iter_init(source, s->dirty_bitmap, &s->hbi);
172 s->sector_num = hbitmap_iter_next(&s->hbi);
173 trace_mirror_restart_iter(s,
174 bdrv_get_dirty_count(source, s->dirty_bitmap));
175 assert(s->sector_num >= 0);
178 hbitmap_next_sector = s->sector_num;
179 sector_num = s->sector_num;
180 sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
181 end = s->bdev_length / BDRV_SECTOR_SIZE;
183 /* Extend the QEMUIOVector to include all adjacent blocks that will
184 * be copied in this operation.
186 * We have to do this if we have no backing file yet in the destination,
187 * and the cluster size is very large. Then we need to do COW ourselves.
188 * The first time a cluster is copied, copy it entirely. Note that,
189 * because both the granularity and the cluster size are powers of two,
190 * the number of sectors to copy cannot exceed one cluster.
192 * We also want to extend the QEMUIOVector to include more adjacent
193 * dirty blocks if possible, to limit the number of I/O operations and
194 * run efficiently even with a small granularity.
198 next_sector = sector_num;
199 next_chunk = sector_num / sectors_per_chunk;
201 /* Wait for I/O to this cluster (from a previous iteration) to be done. */
202 while (test_bit(next_chunk, s->in_flight_bitmap)) {
203 trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
204 qemu_coroutine_yield();
208 int added_sectors, added_chunks;
210 if (!bdrv_get_dirty(source, s->dirty_bitmap, next_sector) ||
211 test_bit(next_chunk, s->in_flight_bitmap)) {
212 assert(nb_sectors > 0);
216 added_sectors = sectors_per_chunk;
217 if (s->cow_bitmap && !test_bit(next_chunk, s->cow_bitmap)) {
218 bdrv_round_to_clusters(s->target,
219 next_sector, added_sectors,
220 &next_sector, &added_sectors);
222 /* On the first iteration, the rounding may make us copy
223 * sectors before the first dirty one.
225 if (next_sector < sector_num) {
226 assert(nb_sectors == 0);
227 sector_num = next_sector;
228 next_chunk = next_sector / sectors_per_chunk;
232 added_sectors = MIN(added_sectors, end - (sector_num + nb_sectors));
233 added_chunks = (added_sectors + sectors_per_chunk - 1) / sectors_per_chunk;
235 /* When doing COW, it may happen that there is not enough space for
236 * a full cluster. Wait if that is the case.
238 while (nb_chunks == 0 && s->buf_free_count < added_chunks) {
239 trace_mirror_yield_buf_busy(s, nb_chunks, s->in_flight);
240 qemu_coroutine_yield();
242 if (s->buf_free_count < nb_chunks + added_chunks) {
243 trace_mirror_break_buf_busy(s, nb_chunks, s->in_flight);
247 /* We have enough free space to copy these sectors. */
248 bitmap_set(s->in_flight_bitmap, next_chunk, added_chunks);
250 nb_sectors += added_sectors;
251 nb_chunks += added_chunks;
252 next_sector += added_sectors;
253 next_chunk += added_chunks;
254 if (!s->synced && s->common.speed) {
255 delay_ns = ratelimit_calculate_delay(&s->limit, added_sectors);
257 } while (delay_ns == 0 && next_sector < end);
259 /* Allocate a MirrorOp that is used as an AIO callback. */
260 op = g_slice_new(MirrorOp);
262 op->sector_num = sector_num;
263 op->nb_sectors = nb_sectors;
265 /* Now make a QEMUIOVector taking enough granularity-sized chunks
268 qemu_iovec_init(&op->qiov, nb_chunks);
269 next_sector = sector_num;
270 while (nb_chunks-- > 0) {
271 MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
272 size_t remaining = (nb_sectors * BDRV_SECTOR_SIZE) - op->qiov.size;
274 QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
276 qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
278 /* Advance the HBitmapIter in parallel, so that we do not examine
279 * the same sector twice.
281 if (next_sector > hbitmap_next_sector
282 && bdrv_get_dirty(source, s->dirty_bitmap, next_sector)) {
283 hbitmap_next_sector = hbitmap_iter_next(&s->hbi);
286 next_sector += sectors_per_chunk;
289 bdrv_reset_dirty(source, sector_num, nb_sectors);
291 /* Copy the dirty cluster. */
293 s->sectors_in_flight += nb_sectors;
294 trace_mirror_one_iteration(s, sector_num, nb_sectors);
295 bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors,
296 mirror_read_complete, op);
300 static void mirror_free_init(MirrorBlockJob *s)
302 int granularity = s->granularity;
303 size_t buf_size = s->buf_size;
304 uint8_t *buf = s->buf;
306 assert(s->buf_free_count == 0);
307 QSIMPLEQ_INIT(&s->buf_free);
308 while (buf_size != 0) {
309 MirrorBuffer *cur = (MirrorBuffer *)buf;
310 QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
312 buf_size -= granularity;
317 static void mirror_drain(MirrorBlockJob *s)
319 while (s->in_flight > 0) {
320 qemu_coroutine_yield();
328 static void mirror_exit(BlockJob *job, void *opaque)
330 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
331 MirrorExitData *data = opaque;
332 AioContext *replace_aio_context = NULL;
335 replace_aio_context = bdrv_get_aio_context(s->to_replace);
336 aio_context_acquire(replace_aio_context);
339 if (s->should_complete && data->ret == 0) {
340 BlockDriverState *to_replace = s->common.bs;
342 to_replace = s->to_replace;
344 if (bdrv_get_flags(s->target) != bdrv_get_flags(to_replace)) {
345 bdrv_reopen(s->target, bdrv_get_flags(to_replace), NULL);
347 bdrv_swap(s->target, to_replace);
348 if (s->common.driver->job_type == BLOCK_JOB_TYPE_COMMIT) {
349 /* drop the bs loop chain formed by the swap: break the loop then
350 * trigger the unref from the top one */
351 BlockDriverState *p = s->base->backing_hd;
352 bdrv_set_backing_hd(s->base, NULL);
357 bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
358 error_free(s->replace_blocker);
359 bdrv_unref(s->to_replace);
361 if (replace_aio_context) {
362 aio_context_release(replace_aio_context);
365 bdrv_unref(s->target);
366 block_job_completed(&s->common, data->ret);
370 static void coroutine_fn mirror_run(void *opaque)
372 MirrorBlockJob *s = opaque;
373 MirrorExitData *data;
374 BlockDriverState *bs = s->common.bs;
375 int64_t sector_num, end, sectors_per_chunk, length;
376 uint64_t last_pause_ns;
378 char backing_filename[1024];
382 if (block_job_is_cancelled(&s->common)) {
386 s->bdev_length = bdrv_getlength(bs);
387 if (s->bdev_length < 0) {
388 ret = s->bdev_length;
390 } else if (s->bdev_length == 0) {
391 /* Report BLOCK_JOB_READY and wait for complete. */
392 block_job_event_ready(&s->common);
394 while (!block_job_is_cancelled(&s->common) && !s->should_complete) {
395 block_job_yield(&s->common);
397 s->common.cancelled = false;
401 length = DIV_ROUND_UP(s->bdev_length, s->granularity);
402 s->in_flight_bitmap = bitmap_new(length);
404 /* If we have no backing file yet in the destination, we cannot let
405 * the destination do COW. Instead, we copy sectors around the
406 * dirty data if needed. We need a bitmap to do that.
408 bdrv_get_backing_filename(s->target, backing_filename,
409 sizeof(backing_filename));
410 if (backing_filename[0] && !s->target->backing_hd) {
411 ret = bdrv_get_info(s->target, &bdi);
415 if (s->granularity < bdi.cluster_size) {
416 s->buf_size = MAX(s->buf_size, bdi.cluster_size);
417 s->cow_bitmap = bitmap_new(length);
421 end = s->bdev_length / BDRV_SECTOR_SIZE;
422 s->buf = qemu_try_blockalign(bs, s->buf_size);
423 if (s->buf == NULL) {
428 sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
431 if (!s->is_none_mode) {
432 /* First part, loop on the sectors and initialize the dirty bitmap. */
433 BlockDriverState *base = s->base;
434 for (sector_num = 0; sector_num < end; ) {
435 int64_t next = (sector_num | (sectors_per_chunk - 1)) + 1;
436 ret = bdrv_is_allocated_above(bs, base,
437 sector_num, next - sector_num, &n);
445 bdrv_set_dirty(bs, sector_num, n);
453 bdrv_dirty_iter_init(bs, s->dirty_bitmap, &s->hbi);
454 last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
456 uint64_t delay_ns = 0;
458 bool should_complete;
465 cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);
466 /* s->common.offset contains the number of bytes already processed so
467 * far, cnt is the number of dirty sectors remaining and
468 * s->sectors_in_flight is the number of sectors currently being
469 * processed; together those are the current total operation length */
470 s->common.len = s->common.offset +
471 (cnt + s->sectors_in_flight) * BDRV_SECTOR_SIZE;
473 /* Note that even when no rate limit is applied we need to yield
474 * periodically with no pending I/O so that qemu_aio_flush() returns.
475 * We do so every SLICE_TIME nanoseconds, or when there is an error,
476 * or when the source is clean, whichever comes first.
478 if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - last_pause_ns < SLICE_TIME &&
479 s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
480 if (s->in_flight == MAX_IN_FLIGHT || s->buf_free_count == 0 ||
481 (cnt == 0 && s->in_flight > 0)) {
482 trace_mirror_yield(s, s->in_flight, s->buf_free_count, cnt);
483 qemu_coroutine_yield();
485 } else if (cnt != 0) {
486 delay_ns = mirror_iteration(s);
493 should_complete = false;
494 if (s->in_flight == 0 && cnt == 0) {
495 trace_mirror_before_flush(s);
496 ret = bdrv_flush(s->target);
498 if (mirror_error_action(s, false, -ret) ==
499 BLOCK_ERROR_ACTION_REPORT) {
503 /* We're out of the streaming phase. From now on, if the job
504 * is cancelled we will actually complete all pending I/O and
505 * report completion. This way, block-job-cancel will leave
506 * the target in a consistent state.
509 block_job_event_ready(&s->common);
513 should_complete = s->should_complete ||
514 block_job_is_cancelled(&s->common);
515 cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);
519 if (cnt == 0 && should_complete) {
520 /* The dirty bitmap is not updated while operations are pending.
521 * If we're about to exit, wait for pending operations before
522 * calling bdrv_get_dirty_count(bs), or we may exit while the
523 * source has dirty data to copy!
525 * Note that I/O can be submitted by the guest while
526 * mirror_populate runs.
528 trace_mirror_before_drain(s, cnt);
530 cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);
534 trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
536 block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
537 if (block_job_is_cancelled(&s->common)) {
540 } else if (!should_complete) {
541 delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);
542 block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
543 } else if (cnt == 0) {
544 /* The two disks are in sync. Exit and report successful
547 assert(QLIST_EMPTY(&bs->tracked_requests));
548 s->common.cancelled = false;
551 last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
555 if (s->in_flight > 0) {
556 /* We get here only if something went wrong. Either the job failed,
557 * or it was cancelled prematurely so that we do not guarantee that
558 * the target is a copy of the source.
560 assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common)));
564 assert(s->in_flight == 0);
566 g_free(s->cow_bitmap);
567 g_free(s->in_flight_bitmap);
568 bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
569 bdrv_iostatus_disable(s->target);
571 data = g_malloc(sizeof(*data));
573 block_job_defer_to_main_loop(&s->common, mirror_exit, data);
576 static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp)
578 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
581 error_set(errp, QERR_INVALID_PARAMETER, "speed");
584 ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
587 static void mirror_iostatus_reset(BlockJob *job)
589 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
591 bdrv_iostatus_reset(s->target);
594 static void mirror_complete(BlockJob *job, Error **errp)
596 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
597 Error *local_err = NULL;
600 ret = bdrv_open_backing_file(s->target, NULL, &local_err);
602 error_propagate(errp, local_err);
606 error_set(errp, QERR_BLOCK_JOB_NOT_READY,
607 bdrv_get_device_name(job->bs));
611 /* check the target bs is not blocked and block all operations on it */
613 AioContext *replace_aio_context;
615 s->to_replace = check_to_replace_node(s->replaces, &local_err);
616 if (!s->to_replace) {
617 error_propagate(errp, local_err);
621 replace_aio_context = bdrv_get_aio_context(s->to_replace);
622 aio_context_acquire(replace_aio_context);
624 error_setg(&s->replace_blocker,
625 "block device is in use by block-job-complete");
626 bdrv_op_block_all(s->to_replace, s->replace_blocker);
627 bdrv_ref(s->to_replace);
629 aio_context_release(replace_aio_context);
632 s->should_complete = true;
633 block_job_resume(job);
636 static const BlockJobDriver mirror_job_driver = {
637 .instance_size = sizeof(MirrorBlockJob),
638 .job_type = BLOCK_JOB_TYPE_MIRROR,
639 .set_speed = mirror_set_speed,
640 .iostatus_reset= mirror_iostatus_reset,
641 .complete = mirror_complete,
644 static const BlockJobDriver commit_active_job_driver = {
645 .instance_size = sizeof(MirrorBlockJob),
646 .job_type = BLOCK_JOB_TYPE_COMMIT,
647 .set_speed = mirror_set_speed,
649 = mirror_iostatus_reset,
650 .complete = mirror_complete,
653 static void mirror_start_job(BlockDriverState *bs, BlockDriverState *target,
654 const char *replaces,
655 int64_t speed, int64_t granularity,
657 BlockdevOnError on_source_error,
658 BlockdevOnError on_target_error,
659 BlockCompletionFunc *cb,
660 void *opaque, Error **errp,
661 const BlockJobDriver *driver,
662 bool is_none_mode, BlockDriverState *base)
666 if (granularity == 0) {
667 /* Choose the default granularity based on the target file's cluster
668 * size, clamped between 4k and 64k. */
670 if (bdrv_get_info(target, &bdi) >= 0 && bdi.cluster_size != 0) {
671 granularity = MAX(4096, bdi.cluster_size);
672 granularity = MIN(65536, granularity);
678 assert ((granularity & (granularity - 1)) == 0);
680 if ((on_source_error == BLOCKDEV_ON_ERROR_STOP ||
681 on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
682 !bdrv_iostatus_is_enabled(bs)) {
683 error_set(errp, QERR_INVALID_PARAMETER, "on-source-error");
688 s = block_job_create(driver, bs, speed, cb, opaque, errp);
693 s->replaces = g_strdup(replaces);
694 s->on_source_error = on_source_error;
695 s->on_target_error = on_target_error;
697 s->is_none_mode = is_none_mode;
699 s->granularity = granularity;
700 s->buf_size = MAX(buf_size, granularity);
702 s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, errp);
703 if (!s->dirty_bitmap) {
706 bdrv_set_enable_write_cache(s->target, true);
707 bdrv_set_on_error(s->target, on_target_error, on_target_error);
708 bdrv_iostatus_enable(s->target);
709 s->common.co = qemu_coroutine_create(mirror_run);
710 trace_mirror_start(bs, s, s->common.co, opaque);
711 qemu_coroutine_enter(s->common.co, s);
714 void mirror_start(BlockDriverState *bs, BlockDriverState *target,
715 const char *replaces,
716 int64_t speed, int64_t granularity, int64_t buf_size,
717 MirrorSyncMode mode, BlockdevOnError on_source_error,
718 BlockdevOnError on_target_error,
719 BlockCompletionFunc *cb,
720 void *opaque, Error **errp)
723 BlockDriverState *base;
725 is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
726 base = mode == MIRROR_SYNC_MODE_TOP ? bs->backing_hd : NULL;
727 mirror_start_job(bs, target, replaces,
728 speed, granularity, buf_size,
729 on_source_error, on_target_error, cb, opaque, errp,
730 &mirror_job_driver, is_none_mode, base);
733 void commit_active_start(BlockDriverState *bs, BlockDriverState *base,
735 BlockdevOnError on_error,
736 BlockCompletionFunc *cb,
737 void *opaque, Error **errp)
739 int64_t length, base_length;
742 Error *local_err = NULL;
744 orig_base_flags = bdrv_get_flags(base);
746 if (bdrv_reopen(base, bs->open_flags, errp)) {
750 length = bdrv_getlength(bs);
752 error_setg_errno(errp, -length,
753 "Unable to determine length of %s", bs->filename);
754 goto error_restore_flags;
757 base_length = bdrv_getlength(base);
758 if (base_length < 0) {
759 error_setg_errno(errp, -base_length,
760 "Unable to determine length of %s", base->filename);
761 goto error_restore_flags;
764 if (length > base_length) {
765 ret = bdrv_truncate(base, length);
767 error_setg_errno(errp, -ret,
768 "Top image %s is larger than base image %s, and "
769 "resize of base image failed",
770 bs->filename, base->filename);
771 goto error_restore_flags;
776 mirror_start_job(bs, base, NULL, speed, 0, 0,
777 on_error, on_error, cb, opaque, &local_err,
778 &commit_active_job_driver, false, base);
780 error_propagate(errp, local_err);
781 goto error_restore_flags;
787 /* ignore error and errp for bdrv_reopen, because we want to propagate
788 * the original error */
789 bdrv_reopen(base, orig_base_flags, NULL);