4 * Copyright (C) 2013 Proxmox Server Solutions
9 * This work is licensed under the terms of the GNU GPL, version 2 or later.
10 * See the COPYING file in the top-level directory.
14 #include "qemu/osdep.h"
17 #include "block/block.h"
18 #include "block/block_int.h"
19 #include "block/blockjob_int.h"
20 #include "block/block_backup.h"
21 #include "qapi/error.h"
22 #include "qapi/qmp/qerror.h"
23 #include "qemu/ratelimit.h"
24 #include "qemu/cutils.h"
25 #include "sysemu/block-backend.h"
26 #include "qemu/bitmap.h"
27 #include "qemu/error-report.h"
29 #define BACKUP_CLUSTER_SIZE_DEFAULT (1 << 16)
31 typedef struct CowRequest {
34 QLIST_ENTRY(CowRequest) list;
35 CoQueue wait_queue; /* coroutines blocked on this request */
38 typedef struct BackupBlockJob {
41 /* bitmap for sync=incremental */
42 BdrvDirtyBitmap *sync_bitmap;
43 MirrorSyncMode sync_mode;
44 BlockdevOnError on_source_error;
45 BlockdevOnError on_target_error;
46 CoRwlock flush_rwlock;
51 NotifierWithReturn before_write;
52 QLIST_HEAD(, CowRequest) inflight_reqs;
56 int64_t copy_range_size;
58 bool serialize_target_writes;
61 static const BlockJobDriver backup_job_driver;
63 /* See if in-flight requests overlap and wait for them to complete */
64 static void coroutine_fn wait_for_overlapping_requests(BackupBlockJob *job,
73 QLIST_FOREACH(req, &job->inflight_reqs, list) {
74 if (end > req->start_byte && start < req->end_byte) {
75 qemu_co_queue_wait(&req->wait_queue, NULL);
83 /* Keep track of an in-flight request */
84 static void cow_request_begin(CowRequest *req, BackupBlockJob *job,
85 int64_t start, int64_t end)
87 req->start_byte = start;
89 qemu_co_queue_init(&req->wait_queue);
90 QLIST_INSERT_HEAD(&job->inflight_reqs, req, list);
93 /* Forget about a completed request */
94 static void cow_request_end(CowRequest *req)
96 QLIST_REMOVE(req, list);
97 qemu_co_queue_restart_all(&req->wait_queue);
100 /* Copy range to target with a bounce buffer and return the bytes copied. If
101 * error occurred, return a negative error number */
102 static int coroutine_fn backup_cow_with_bounce_buffer(BackupBlockJob *job,
105 bool is_write_notifier,
107 void **bounce_buffer)
110 BlockBackend *blk = job->common.blk;
112 int read_flags = is_write_notifier ? BDRV_REQ_NO_SERIALISING : 0;
113 int write_flags = job->serialize_target_writes ? BDRV_REQ_SERIALISING : 0;
115 hbitmap_reset(job->copy_bitmap, start / job->cluster_size, 1);
116 nbytes = MIN(job->cluster_size, job->len - start);
117 if (!*bounce_buffer) {
118 *bounce_buffer = blk_blockalign(blk, job->cluster_size);
121 ret = blk_co_pread(blk, start, nbytes, *bounce_buffer, read_flags);
123 trace_backup_do_cow_read_fail(job, start, ret);
125 *error_is_read = true;
130 if (buffer_is_zero(*bounce_buffer, nbytes)) {
131 ret = blk_co_pwrite_zeroes(job->target, start,
132 nbytes, write_flags | BDRV_REQ_MAY_UNMAP);
134 ret = blk_co_pwrite(job->target, start,
135 nbytes, *bounce_buffer, write_flags |
136 (job->compress ? BDRV_REQ_WRITE_COMPRESSED : 0));
139 trace_backup_do_cow_write_fail(job, start, ret);
141 *error_is_read = false;
148 hbitmap_set(job->copy_bitmap, start / job->cluster_size, 1);
153 /* Copy range to target and return the bytes copied. If error occurred, return a
154 * negative error number. */
155 static int coroutine_fn backup_cow_with_offload(BackupBlockJob *job,
158 bool is_write_notifier)
162 BlockBackend *blk = job->common.blk;
164 int read_flags = is_write_notifier ? BDRV_REQ_NO_SERIALISING : 0;
165 int write_flags = job->serialize_target_writes ? BDRV_REQ_SERIALISING : 0;
167 assert(QEMU_IS_ALIGNED(job->copy_range_size, job->cluster_size));
168 nbytes = MIN(job->copy_range_size, end - start);
169 nr_clusters = DIV_ROUND_UP(nbytes, job->cluster_size);
170 hbitmap_reset(job->copy_bitmap, start / job->cluster_size,
172 ret = blk_co_copy_range(blk, start, job->target, start, nbytes,
173 read_flags, write_flags);
175 trace_backup_do_cow_copy_range_fail(job, start, ret);
176 hbitmap_set(job->copy_bitmap, start / job->cluster_size,
184 static int coroutine_fn backup_do_cow(BackupBlockJob *job,
185 int64_t offset, uint64_t bytes,
187 bool is_write_notifier)
189 CowRequest cow_request;
191 int64_t start, end; /* bytes */
192 void *bounce_buffer = NULL;
194 qemu_co_rwlock_rdlock(&job->flush_rwlock);
196 start = QEMU_ALIGN_DOWN(offset, job->cluster_size);
197 end = QEMU_ALIGN_UP(bytes + offset, job->cluster_size);
199 trace_backup_do_cow_enter(job, start, offset, bytes);
201 wait_for_overlapping_requests(job, start, end);
202 cow_request_begin(&cow_request, job, start, end);
204 while (start < end) {
205 if (!hbitmap_get(job->copy_bitmap, start / job->cluster_size)) {
206 trace_backup_do_cow_skip(job, start);
207 start += job->cluster_size;
208 continue; /* already copied */
211 trace_backup_do_cow_process(job, start);
213 if (job->use_copy_range) {
214 ret = backup_cow_with_offload(job, start, end, is_write_notifier);
216 job->use_copy_range = false;
219 if (!job->use_copy_range) {
220 ret = backup_cow_with_bounce_buffer(job, start, end, is_write_notifier,
221 error_is_read, &bounce_buffer);
227 /* Publish progress, guest I/O counts as progress too. Note that the
228 * offset field is an opaque progress value, it is not a disk offset.
231 job->bytes_read += ret;
232 job_progress_update(&job->common.job, ret);
237 qemu_vfree(bounce_buffer);
240 cow_request_end(&cow_request);
242 trace_backup_do_cow_return(job, offset, bytes, ret);
244 qemu_co_rwlock_unlock(&job->flush_rwlock);
249 static int coroutine_fn backup_before_write_notify(
250 NotifierWithReturn *notifier,
253 BackupBlockJob *job = container_of(notifier, BackupBlockJob, before_write);
254 BdrvTrackedRequest *req = opaque;
256 assert(req->bs == blk_bs(job->common.blk));
257 assert(QEMU_IS_ALIGNED(req->offset, BDRV_SECTOR_SIZE));
258 assert(QEMU_IS_ALIGNED(req->bytes, BDRV_SECTOR_SIZE));
260 return backup_do_cow(job, req->offset, req->bytes, NULL, true);
263 static void backup_cleanup_sync_bitmap(BackupBlockJob *job, int ret)
266 BlockDriverState *bs = blk_bs(job->common.blk);
269 /* Merge the successor back into the parent, delete nothing. */
270 bm = bdrv_reclaim_dirty_bitmap(bs, job->sync_bitmap, NULL);
273 /* Everything is fine, delete this bitmap and install the backup. */
274 bm = bdrv_dirty_bitmap_abdicate(bs, job->sync_bitmap, NULL);
279 static void backup_commit(Job *job)
281 BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
282 if (s->sync_bitmap) {
283 backup_cleanup_sync_bitmap(s, 0);
287 static void backup_abort(Job *job)
289 BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
290 if (s->sync_bitmap) {
291 backup_cleanup_sync_bitmap(s, -1);
295 static void backup_clean(Job *job)
297 BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
299 blk_unref(s->target);
303 static void backup_attached_aio_context(BlockJob *job, AioContext *aio_context)
305 BackupBlockJob *s = container_of(job, BackupBlockJob, common);
307 blk_set_aio_context(s->target, aio_context);
310 void backup_do_checkpoint(BlockJob *job, Error **errp)
312 BackupBlockJob *backup_job = container_of(job, BackupBlockJob, common);
315 assert(block_job_driver(job) == &backup_job_driver);
317 if (backup_job->sync_mode != MIRROR_SYNC_MODE_NONE) {
318 error_setg(errp, "The backup job only supports block checkpoint in"
323 len = DIV_ROUND_UP(backup_job->len, backup_job->cluster_size);
324 hbitmap_set(backup_job->copy_bitmap, 0, len);
327 static void backup_drain(BlockJob *job)
329 BackupBlockJob *s = container_of(job, BackupBlockJob, common);
331 /* Need to keep a reference in case blk_drain triggers execution
332 * of backup_complete...
335 BlockBackend *target = s->target;
342 static BlockErrorAction backup_error_action(BackupBlockJob *job,
343 bool read, int error)
346 return block_job_error_action(&job->common, job->on_source_error,
349 return block_job_error_action(&job->common, job->on_target_error,
354 static bool coroutine_fn yield_and_check(BackupBlockJob *job)
358 if (job_is_cancelled(&job->common.job)) {
362 /* We need to yield even for delay_ns = 0 so that bdrv_drain_all() can
363 * return. Without a yield, the VM would not reboot. */
364 delay_ns = block_job_ratelimit_get_delay(&job->common, job->bytes_read);
366 job_sleep_ns(&job->common.job, delay_ns);
368 if (job_is_cancelled(&job->common.job)) {
375 static int coroutine_fn backup_run_incremental(BackupBlockJob *job)
382 hbitmap_iter_init(&hbi, job->copy_bitmap, 0);
383 while ((cluster = hbitmap_iter_next(&hbi)) != -1) {
385 if (yield_and_check(job)) {
388 ret = backup_do_cow(job, cluster * job->cluster_size,
389 job->cluster_size, &error_is_read, false);
390 if (ret < 0 && backup_error_action(job, error_is_read, -ret) ==
391 BLOCK_ERROR_ACTION_REPORT)
401 /* init copy_bitmap from sync_bitmap */
402 static void backup_incremental_init_copy_bitmap(BackupBlockJob *job)
404 BdrvDirtyBitmapIter *dbi;
406 int64_t end = DIV_ROUND_UP(bdrv_dirty_bitmap_size(job->sync_bitmap),
409 dbi = bdrv_dirty_iter_new(job->sync_bitmap);
410 while ((offset = bdrv_dirty_iter_next(dbi)) != -1) {
411 int64_t cluster = offset / job->cluster_size;
412 int64_t next_cluster;
414 offset += bdrv_dirty_bitmap_granularity(job->sync_bitmap);
415 if (offset >= bdrv_dirty_bitmap_size(job->sync_bitmap)) {
416 hbitmap_set(job->copy_bitmap, cluster, end - cluster);
420 offset = bdrv_dirty_bitmap_next_zero(job->sync_bitmap, offset,
423 hbitmap_set(job->copy_bitmap, cluster, end - cluster);
427 next_cluster = DIV_ROUND_UP(offset, job->cluster_size);
428 hbitmap_set(job->copy_bitmap, cluster, next_cluster - cluster);
429 if (next_cluster >= end) {
433 bdrv_set_dirty_iter(dbi, next_cluster * job->cluster_size);
436 /* TODO job_progress_set_remaining() would make more sense */
437 job_progress_update(&job->common.job,
438 job->len - hbitmap_count(job->copy_bitmap) * job->cluster_size);
440 bdrv_dirty_iter_free(dbi);
443 static int coroutine_fn backup_run(Job *job, Error **errp)
445 BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
446 BlockDriverState *bs = blk_bs(s->common.blk);
447 int64_t offset, nb_clusters;
450 QLIST_INIT(&s->inflight_reqs);
451 qemu_co_rwlock_init(&s->flush_rwlock);
453 nb_clusters = DIV_ROUND_UP(s->len, s->cluster_size);
454 job_progress_set_remaining(job, s->len);
456 s->copy_bitmap = hbitmap_alloc(nb_clusters, 0);
457 if (s->sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
458 backup_incremental_init_copy_bitmap(s);
460 hbitmap_set(s->copy_bitmap, 0, nb_clusters);
464 s->before_write.notify = backup_before_write_notify;
465 bdrv_add_before_write_notifier(bs, &s->before_write);
467 if (s->sync_mode == MIRROR_SYNC_MODE_NONE) {
468 /* All bits are set in copy_bitmap to allow any cluster to be copied.
469 * This does not actually require them to be copied. */
470 while (!job_is_cancelled(job)) {
471 /* Yield until the job is cancelled. We just let our before_write
472 * notify callback service CoW requests. */
475 } else if (s->sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
476 ret = backup_run_incremental(s);
478 /* Both FULL and TOP SYNC_MODE's require copying.. */
479 for (offset = 0; offset < s->len;
480 offset += s->cluster_size) {
484 if (yield_and_check(s)) {
488 if (s->sync_mode == MIRROR_SYNC_MODE_TOP) {
492 /* Check to see if these blocks are already in the
495 for (i = 0; i < s->cluster_size;) {
496 /* bdrv_is_allocated() only returns true/false based
497 * on the first set of sectors it comes across that
498 * are are all in the same state.
499 * For that reason we must verify each sector in the
500 * backup cluster length. We end up copying more than
501 * needed but at some point that is always the case. */
503 bdrv_is_allocated(bs, offset + i,
504 s->cluster_size - i, &n);
507 if (alloced || n == 0) {
512 /* If the above loop never found any sectors that are in
513 * the topmost image, skip this backup. */
518 /* FULL sync mode we copy the whole drive. */
522 ret = backup_do_cow(s, offset, s->cluster_size,
523 &error_is_read, false);
526 /* Depending on error action, fail now or retry cluster */
527 BlockErrorAction action =
528 backup_error_action(s, error_is_read, -ret);
529 if (action == BLOCK_ERROR_ACTION_REPORT) {
532 offset -= s->cluster_size;
539 notifier_with_return_remove(&s->before_write);
541 /* wait until pending backup_do_cow() calls have completed */
542 qemu_co_rwlock_wrlock(&s->flush_rwlock);
543 qemu_co_rwlock_unlock(&s->flush_rwlock);
544 hbitmap_free(s->copy_bitmap);
549 static const BlockJobDriver backup_job_driver = {
551 .instance_size = sizeof(BackupBlockJob),
552 .job_type = JOB_TYPE_BACKUP,
553 .free = block_job_free,
554 .user_resume = block_job_user_resume,
555 .drain = block_job_drain,
557 .commit = backup_commit,
558 .abort = backup_abort,
559 .clean = backup_clean,
561 .attached_aio_context = backup_attached_aio_context,
562 .drain = backup_drain,
565 BlockJob *backup_job_create(const char *job_id, BlockDriverState *bs,
566 BlockDriverState *target, int64_t speed,
567 MirrorSyncMode sync_mode, BdrvDirtyBitmap *sync_bitmap,
569 BlockdevOnError on_source_error,
570 BlockdevOnError on_target_error,
572 BlockCompletionFunc *cb, void *opaque,
573 JobTxn *txn, Error **errp)
577 BackupBlockJob *job = NULL;
584 error_setg(errp, "Source and target cannot be the same");
588 if (!bdrv_is_inserted(bs)) {
589 error_setg(errp, "Device is not inserted: %s",
590 bdrv_get_device_name(bs));
594 if (!bdrv_is_inserted(target)) {
595 error_setg(errp, "Device is not inserted: %s",
596 bdrv_get_device_name(target));
600 if (compress && target->drv->bdrv_co_pwritev_compressed == NULL) {
601 error_setg(errp, "Compression is not supported for this drive %s",
602 bdrv_get_device_name(target));
606 if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_BACKUP_SOURCE, errp)) {
610 if (bdrv_op_is_blocked(target, BLOCK_OP_TYPE_BACKUP_TARGET, errp)) {
614 if (sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
616 error_setg(errp, "must provide a valid bitmap name for "
617 "\"incremental\" sync mode");
621 /* Create a new bitmap, and freeze/disable this one. */
622 if (bdrv_dirty_bitmap_create_successor(bs, sync_bitmap, errp) < 0) {
625 } else if (sync_bitmap) {
627 "a sync_bitmap was provided to backup_run, "
628 "but received an incompatible sync_mode (%s)",
629 MirrorSyncMode_str(sync_mode));
633 len = bdrv_getlength(bs);
635 error_setg_errno(errp, -len, "unable to get length for '%s'",
636 bdrv_get_device_name(bs));
640 /* job->len is fixed, so we can't allow resize */
641 job = block_job_create(job_id, &backup_job_driver, txn, bs,
642 BLK_PERM_CONSISTENT_READ,
643 BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE |
644 BLK_PERM_WRITE_UNCHANGED | BLK_PERM_GRAPH_MOD,
645 speed, creation_flags, cb, opaque, errp);
650 /* The target must match the source in size, so no resize here either */
651 job->target = blk_new(BLK_PERM_WRITE,
652 BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE |
653 BLK_PERM_WRITE_UNCHANGED | BLK_PERM_GRAPH_MOD);
654 ret = blk_insert_bs(job->target, target, errp);
659 job->on_source_error = on_source_error;
660 job->on_target_error = on_target_error;
661 job->sync_mode = sync_mode;
662 job->sync_bitmap = sync_mode == MIRROR_SYNC_MODE_INCREMENTAL ?
664 job->compress = compress;
666 /* Detect image-fleecing (and similar) schemes */
667 job->serialize_target_writes = bdrv_chain_contains(target, bs);
669 /* If there is no backing file on the target, we cannot rely on COW if our
670 * backup cluster size is smaller than the target cluster size. Even for
671 * targets with a backing file, try to avoid COW if possible. */
672 ret = bdrv_get_info(target, &bdi);
673 if (ret == -ENOTSUP && !target->backing) {
674 /* Cluster size is not defined */
675 warn_report("The target block device doesn't provide "
676 "information about the block size and it doesn't have a "
677 "backing file. The default block size of %u bytes is "
678 "used. If the actual block size of the target exceeds "
679 "this default, the backup may be unusable",
680 BACKUP_CLUSTER_SIZE_DEFAULT);
681 job->cluster_size = BACKUP_CLUSTER_SIZE_DEFAULT;
682 } else if (ret < 0 && !target->backing) {
683 error_setg_errno(errp, -ret,
684 "Couldn't determine the cluster size of the target image, "
685 "which has no backing file");
686 error_append_hint(errp,
687 "Aborting, since this may create an unusable destination image\n");
689 } else if (ret < 0 && target->backing) {
690 /* Not fatal; just trudge on ahead. */
691 job->cluster_size = BACKUP_CLUSTER_SIZE_DEFAULT;
693 job->cluster_size = MAX(BACKUP_CLUSTER_SIZE_DEFAULT, bdi.cluster_size);
695 job->use_copy_range = true;
696 job->copy_range_size = MIN_NON_ZERO(blk_get_max_transfer(job->common.blk),
697 blk_get_max_transfer(job->target));
698 job->copy_range_size = MAX(job->cluster_size,
699 QEMU_ALIGN_UP(job->copy_range_size,
702 /* Required permissions are already taken with target's blk_new() */
703 block_job_add_bdrv(&job->common, "target", target, 0, BLK_PERM_ALL,
711 bdrv_reclaim_dirty_bitmap(bs, sync_bitmap, NULL);
714 backup_clean(&job->common.job);
715 job_early_fail(&job->common.job);