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)
112 BlockBackend *blk = job->common.blk;
114 int read_flags = is_write_notifier ? BDRV_REQ_NO_SERIALISING : 0;
115 int write_flags = job->serialize_target_writes ? BDRV_REQ_SERIALISING : 0;
117 hbitmap_reset(job->copy_bitmap, start / job->cluster_size, 1);
118 nbytes = MIN(job->cluster_size, job->len - start);
119 if (!*bounce_buffer) {
120 *bounce_buffer = blk_blockalign(blk, job->cluster_size);
122 iov.iov_base = *bounce_buffer;
123 iov.iov_len = nbytes;
124 qemu_iovec_init_external(&qiov, &iov, 1);
126 ret = blk_co_preadv(blk, start, qiov.size, &qiov, read_flags);
128 trace_backup_do_cow_read_fail(job, start, ret);
130 *error_is_read = true;
135 if (qemu_iovec_is_zero(&qiov)) {
136 ret = blk_co_pwrite_zeroes(job->target, start,
137 qiov.size, write_flags | BDRV_REQ_MAY_UNMAP);
139 ret = blk_co_pwritev(job->target, start,
140 qiov.size, &qiov, write_flags |
141 (job->compress ? BDRV_REQ_WRITE_COMPRESSED : 0));
144 trace_backup_do_cow_write_fail(job, start, ret);
146 *error_is_read = false;
153 hbitmap_set(job->copy_bitmap, start / job->cluster_size, 1);
158 /* Copy range to target and return the bytes copied. If error occurred, return a
159 * negative error number. */
160 static int coroutine_fn backup_cow_with_offload(BackupBlockJob *job,
163 bool is_write_notifier)
167 BlockBackend *blk = job->common.blk;
169 int read_flags = is_write_notifier ? BDRV_REQ_NO_SERIALISING : 0;
170 int write_flags = job->serialize_target_writes ? BDRV_REQ_SERIALISING : 0;
172 assert(QEMU_IS_ALIGNED(job->copy_range_size, job->cluster_size));
173 nbytes = MIN(job->copy_range_size, end - start);
174 nr_clusters = DIV_ROUND_UP(nbytes, job->cluster_size);
175 hbitmap_reset(job->copy_bitmap, start / job->cluster_size,
177 ret = blk_co_copy_range(blk, start, job->target, start, nbytes,
178 read_flags, write_flags);
180 trace_backup_do_cow_copy_range_fail(job, start, ret);
181 hbitmap_set(job->copy_bitmap, start / job->cluster_size,
189 static int coroutine_fn backup_do_cow(BackupBlockJob *job,
190 int64_t offset, uint64_t bytes,
192 bool is_write_notifier)
194 CowRequest cow_request;
196 int64_t start, end; /* bytes */
197 void *bounce_buffer = NULL;
199 qemu_co_rwlock_rdlock(&job->flush_rwlock);
201 start = QEMU_ALIGN_DOWN(offset, job->cluster_size);
202 end = QEMU_ALIGN_UP(bytes + offset, job->cluster_size);
204 trace_backup_do_cow_enter(job, start, offset, bytes);
206 wait_for_overlapping_requests(job, start, end);
207 cow_request_begin(&cow_request, job, start, end);
209 while (start < end) {
210 if (!hbitmap_get(job->copy_bitmap, start / job->cluster_size)) {
211 trace_backup_do_cow_skip(job, start);
212 start += job->cluster_size;
213 continue; /* already copied */
216 trace_backup_do_cow_process(job, start);
218 if (job->use_copy_range) {
219 ret = backup_cow_with_offload(job, start, end, is_write_notifier);
221 job->use_copy_range = false;
224 if (!job->use_copy_range) {
225 ret = backup_cow_with_bounce_buffer(job, start, end, is_write_notifier,
226 error_is_read, &bounce_buffer);
232 /* Publish progress, guest I/O counts as progress too. Note that the
233 * offset field is an opaque progress value, it is not a disk offset.
236 job->bytes_read += ret;
237 job_progress_update(&job->common.job, ret);
242 qemu_vfree(bounce_buffer);
245 cow_request_end(&cow_request);
247 trace_backup_do_cow_return(job, offset, bytes, ret);
249 qemu_co_rwlock_unlock(&job->flush_rwlock);
254 static int coroutine_fn backup_before_write_notify(
255 NotifierWithReturn *notifier,
258 BackupBlockJob *job = container_of(notifier, BackupBlockJob, before_write);
259 BdrvTrackedRequest *req = opaque;
261 assert(req->bs == blk_bs(job->common.blk));
262 assert(QEMU_IS_ALIGNED(req->offset, BDRV_SECTOR_SIZE));
263 assert(QEMU_IS_ALIGNED(req->bytes, BDRV_SECTOR_SIZE));
265 return backup_do_cow(job, req->offset, req->bytes, NULL, true);
268 static void backup_cleanup_sync_bitmap(BackupBlockJob *job, int ret)
271 BlockDriverState *bs = blk_bs(job->common.blk);
274 /* Merge the successor back into the parent, delete nothing. */
275 bm = bdrv_reclaim_dirty_bitmap(bs, job->sync_bitmap, NULL);
278 /* Everything is fine, delete this bitmap and install the backup. */
279 bm = bdrv_dirty_bitmap_abdicate(bs, job->sync_bitmap, NULL);
284 static void backup_commit(Job *job)
286 BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
287 if (s->sync_bitmap) {
288 backup_cleanup_sync_bitmap(s, 0);
292 static void backup_abort(Job *job)
294 BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
295 if (s->sync_bitmap) {
296 backup_cleanup_sync_bitmap(s, -1);
300 static void backup_clean(Job *job)
302 BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
304 blk_unref(s->target);
308 static void backup_attached_aio_context(BlockJob *job, AioContext *aio_context)
310 BackupBlockJob *s = container_of(job, BackupBlockJob, common);
312 blk_set_aio_context(s->target, aio_context);
315 void backup_do_checkpoint(BlockJob *job, Error **errp)
317 BackupBlockJob *backup_job = container_of(job, BackupBlockJob, common);
320 assert(block_job_driver(job) == &backup_job_driver);
322 if (backup_job->sync_mode != MIRROR_SYNC_MODE_NONE) {
323 error_setg(errp, "The backup job only supports block checkpoint in"
328 len = DIV_ROUND_UP(backup_job->len, backup_job->cluster_size);
329 hbitmap_set(backup_job->copy_bitmap, 0, len);
332 static void backup_drain(BlockJob *job)
334 BackupBlockJob *s = container_of(job, BackupBlockJob, common);
336 /* Need to keep a reference in case blk_drain triggers execution
337 * of backup_complete...
340 BlockBackend *target = s->target;
347 static BlockErrorAction backup_error_action(BackupBlockJob *job,
348 bool read, int error)
351 return block_job_error_action(&job->common, job->on_source_error,
354 return block_job_error_action(&job->common, job->on_target_error,
359 static bool coroutine_fn yield_and_check(BackupBlockJob *job)
363 if (job_is_cancelled(&job->common.job)) {
367 /* We need to yield even for delay_ns = 0 so that bdrv_drain_all() can
368 * return. Without a yield, the VM would not reboot. */
369 delay_ns = block_job_ratelimit_get_delay(&job->common, job->bytes_read);
371 job_sleep_ns(&job->common.job, delay_ns);
373 if (job_is_cancelled(&job->common.job)) {
380 static int coroutine_fn backup_run_incremental(BackupBlockJob *job)
387 hbitmap_iter_init(&hbi, job->copy_bitmap, 0);
388 while ((cluster = hbitmap_iter_next(&hbi, true)) != -1) {
390 if (yield_and_check(job)) {
393 ret = backup_do_cow(job, cluster * job->cluster_size,
394 job->cluster_size, &error_is_read, false);
395 if (ret < 0 && backup_error_action(job, error_is_read, -ret) ==
396 BLOCK_ERROR_ACTION_REPORT)
406 /* init copy_bitmap from sync_bitmap */
407 static void backup_incremental_init_copy_bitmap(BackupBlockJob *job)
409 BdrvDirtyBitmapIter *dbi;
411 int64_t end = DIV_ROUND_UP(bdrv_dirty_bitmap_size(job->sync_bitmap),
414 dbi = bdrv_dirty_iter_new(job->sync_bitmap);
415 while ((offset = bdrv_dirty_iter_next(dbi)) != -1) {
416 int64_t cluster = offset / job->cluster_size;
417 int64_t next_cluster;
419 offset += bdrv_dirty_bitmap_granularity(job->sync_bitmap);
420 if (offset >= bdrv_dirty_bitmap_size(job->sync_bitmap)) {
421 hbitmap_set(job->copy_bitmap, cluster, end - cluster);
425 offset = bdrv_dirty_bitmap_next_zero(job->sync_bitmap, offset);
427 hbitmap_set(job->copy_bitmap, cluster, end - cluster);
431 next_cluster = DIV_ROUND_UP(offset, job->cluster_size);
432 hbitmap_set(job->copy_bitmap, cluster, next_cluster - cluster);
433 if (next_cluster >= end) {
437 bdrv_set_dirty_iter(dbi, next_cluster * job->cluster_size);
440 /* TODO job_progress_set_remaining() would make more sense */
441 job_progress_update(&job->common.job,
442 job->len - hbitmap_count(job->copy_bitmap) * job->cluster_size);
444 bdrv_dirty_iter_free(dbi);
447 static int coroutine_fn backup_run(Job *job, Error **errp)
449 BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
450 BlockDriverState *bs = blk_bs(s->common.blk);
451 int64_t offset, nb_clusters;
454 QLIST_INIT(&s->inflight_reqs);
455 qemu_co_rwlock_init(&s->flush_rwlock);
457 nb_clusters = DIV_ROUND_UP(s->len, s->cluster_size);
458 job_progress_set_remaining(job, s->len);
460 s->copy_bitmap = hbitmap_alloc(nb_clusters, 0);
461 if (s->sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
462 backup_incremental_init_copy_bitmap(s);
464 hbitmap_set(s->copy_bitmap, 0, nb_clusters);
468 s->before_write.notify = backup_before_write_notify;
469 bdrv_add_before_write_notifier(bs, &s->before_write);
471 if (s->sync_mode == MIRROR_SYNC_MODE_NONE) {
472 /* All bits are set in copy_bitmap to allow any cluster to be copied.
473 * This does not actually require them to be copied. */
474 while (!job_is_cancelled(job)) {
475 /* Yield until the job is cancelled. We just let our before_write
476 * notify callback service CoW requests. */
479 } else if (s->sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
480 ret = backup_run_incremental(s);
482 /* Both FULL and TOP SYNC_MODE's require copying.. */
483 for (offset = 0; offset < s->len;
484 offset += s->cluster_size) {
488 if (yield_and_check(s)) {
492 if (s->sync_mode == MIRROR_SYNC_MODE_TOP) {
496 /* Check to see if these blocks are already in the
499 for (i = 0; i < s->cluster_size;) {
500 /* bdrv_is_allocated() only returns true/false based
501 * on the first set of sectors it comes across that
502 * are are all in the same state.
503 * For that reason we must verify each sector in the
504 * backup cluster length. We end up copying more than
505 * needed but at some point that is always the case. */
507 bdrv_is_allocated(bs, offset + i,
508 s->cluster_size - i, &n);
511 if (alloced || n == 0) {
516 /* If the above loop never found any sectors that are in
517 * the topmost image, skip this backup. */
522 /* FULL sync mode we copy the whole drive. */
526 ret = backup_do_cow(s, offset, s->cluster_size,
527 &error_is_read, false);
530 /* Depending on error action, fail now or retry cluster */
531 BlockErrorAction action =
532 backup_error_action(s, error_is_read, -ret);
533 if (action == BLOCK_ERROR_ACTION_REPORT) {
536 offset -= s->cluster_size;
543 notifier_with_return_remove(&s->before_write);
545 /* wait until pending backup_do_cow() calls have completed */
546 qemu_co_rwlock_wrlock(&s->flush_rwlock);
547 qemu_co_rwlock_unlock(&s->flush_rwlock);
548 hbitmap_free(s->copy_bitmap);
553 static const BlockJobDriver backup_job_driver = {
555 .instance_size = sizeof(BackupBlockJob),
556 .job_type = JOB_TYPE_BACKUP,
557 .free = block_job_free,
558 .user_resume = block_job_user_resume,
559 .drain = block_job_drain,
561 .commit = backup_commit,
562 .abort = backup_abort,
563 .clean = backup_clean,
565 .attached_aio_context = backup_attached_aio_context,
566 .drain = backup_drain,
569 BlockJob *backup_job_create(const char *job_id, BlockDriverState *bs,
570 BlockDriverState *target, int64_t speed,
571 MirrorSyncMode sync_mode, BdrvDirtyBitmap *sync_bitmap,
573 BlockdevOnError on_source_error,
574 BlockdevOnError on_target_error,
576 BlockCompletionFunc *cb, void *opaque,
577 JobTxn *txn, Error **errp)
581 BackupBlockJob *job = NULL;
588 error_setg(errp, "Source and target cannot be the same");
592 if (!bdrv_is_inserted(bs)) {
593 error_setg(errp, "Device is not inserted: %s",
594 bdrv_get_device_name(bs));
598 if (!bdrv_is_inserted(target)) {
599 error_setg(errp, "Device is not inserted: %s",
600 bdrv_get_device_name(target));
604 if (compress && target->drv->bdrv_co_pwritev_compressed == NULL) {
605 error_setg(errp, "Compression is not supported for this drive %s",
606 bdrv_get_device_name(target));
610 if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_BACKUP_SOURCE, errp)) {
614 if (bdrv_op_is_blocked(target, BLOCK_OP_TYPE_BACKUP_TARGET, errp)) {
618 if (sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
620 error_setg(errp, "must provide a valid bitmap name for "
621 "\"incremental\" sync mode");
625 /* Create a new bitmap, and freeze/disable this one. */
626 if (bdrv_dirty_bitmap_create_successor(bs, sync_bitmap, errp) < 0) {
629 } else if (sync_bitmap) {
631 "a sync_bitmap was provided to backup_run, "
632 "but received an incompatible sync_mode (%s)",
633 MirrorSyncMode_str(sync_mode));
637 len = bdrv_getlength(bs);
639 error_setg_errno(errp, -len, "unable to get length for '%s'",
640 bdrv_get_device_name(bs));
644 /* job->len is fixed, so we can't allow resize */
645 job = block_job_create(job_id, &backup_job_driver, txn, bs,
646 BLK_PERM_CONSISTENT_READ,
647 BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE |
648 BLK_PERM_WRITE_UNCHANGED | BLK_PERM_GRAPH_MOD,
649 speed, creation_flags, cb, opaque, errp);
654 /* The target must match the source in size, so no resize here either */
655 job->target = blk_new(BLK_PERM_WRITE,
656 BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE |
657 BLK_PERM_WRITE_UNCHANGED | BLK_PERM_GRAPH_MOD);
658 ret = blk_insert_bs(job->target, target, errp);
663 job->on_source_error = on_source_error;
664 job->on_target_error = on_target_error;
665 job->sync_mode = sync_mode;
666 job->sync_bitmap = sync_mode == MIRROR_SYNC_MODE_INCREMENTAL ?
668 job->compress = compress;
670 /* Detect image-fleecing (and similar) schemes */
671 job->serialize_target_writes = bdrv_chain_contains(target, bs);
673 /* If there is no backing file on the target, we cannot rely on COW if our
674 * backup cluster size is smaller than the target cluster size. Even for
675 * targets with a backing file, try to avoid COW if possible. */
676 ret = bdrv_get_info(target, &bdi);
677 if (ret == -ENOTSUP && !target->backing) {
678 /* Cluster size is not defined */
679 warn_report("The target block device doesn't provide "
680 "information about the block size and it doesn't have a "
681 "backing file. The default block size of %u bytes is "
682 "used. If the actual block size of the target exceeds "
683 "this default, the backup may be unusable",
684 BACKUP_CLUSTER_SIZE_DEFAULT);
685 job->cluster_size = BACKUP_CLUSTER_SIZE_DEFAULT;
686 } else if (ret < 0 && !target->backing) {
687 error_setg_errno(errp, -ret,
688 "Couldn't determine the cluster size of the target image, "
689 "which has no backing file");
690 error_append_hint(errp,
691 "Aborting, since this may create an unusable destination image\n");
693 } else if (ret < 0 && target->backing) {
694 /* Not fatal; just trudge on ahead. */
695 job->cluster_size = BACKUP_CLUSTER_SIZE_DEFAULT;
697 job->cluster_size = MAX(BACKUP_CLUSTER_SIZE_DEFAULT, bdi.cluster_size);
699 job->use_copy_range = true;
700 job->copy_range_size = MIN_NON_ZERO(blk_get_max_transfer(job->common.blk),
701 blk_get_max_transfer(job->target));
702 job->copy_range_size = MAX(job->cluster_size,
703 QEMU_ALIGN_UP(job->copy_range_size,
706 /* Required permissions are already taken with target's blk_new() */
707 block_job_add_bdrv(&job->common, "target", target, 0, BLK_PERM_ALL,
715 bdrv_reclaim_dirty_bitmap(bs, sync_bitmap, NULL);
718 backup_clean(&job->common.job);
719 job_early_fail(&job->common.job);