2 * Block layer I/O functions
4 * Copyright (c) 2003 Fabrice Bellard
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
25 #include "qemu/osdep.h"
27 #include "sysemu/block-backend.h"
28 #include "block/aio-wait.h"
29 #include "block/blockjob.h"
30 #include "block/blockjob_int.h"
31 #include "block/block_int.h"
32 #include "block/coroutines.h"
33 #include "block/write-threshold.h"
34 #include "qemu/cutils.h"
35 #include "qemu/memalign.h"
36 #include "qapi/error.h"
37 #include "qemu/error-report.h"
38 #include "qemu/main-loop.h"
39 #include "sysemu/replay.h"
41 /* Maximum bounce buffer for copy-on-read and write zeroes, in bytes */
42 #define MAX_BOUNCE_BUFFER (32768 << BDRV_SECTOR_BITS)
44 static void bdrv_parent_cb_resize(BlockDriverState *bs);
45 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs,
46 int64_t offset, int64_t bytes, BdrvRequestFlags flags);
48 static void bdrv_parent_drained_begin(BlockDriverState *bs, BdrvChild *ignore,
49 bool ignore_bds_parents)
53 QLIST_FOREACH_SAFE(c, &bs->parents, next_parent, next) {
54 if (c == ignore || (ignore_bds_parents && c->klass->parent_is_bds)) {
57 bdrv_parent_drained_begin_single(c, false);
61 static void bdrv_parent_drained_end_single_no_poll(BdrvChild *c,
62 int *drained_end_counter)
64 assert(c->parent_quiesce_counter > 0);
65 c->parent_quiesce_counter--;
66 if (c->klass->drained_end) {
67 c->klass->drained_end(c, drained_end_counter);
71 void bdrv_parent_drained_end_single(BdrvChild *c)
73 int drained_end_counter = 0;
75 bdrv_parent_drained_end_single_no_poll(c, &drained_end_counter);
76 BDRV_POLL_WHILE(c->bs, qatomic_read(&drained_end_counter) > 0);
79 static void bdrv_parent_drained_end(BlockDriverState *bs, BdrvChild *ignore,
80 bool ignore_bds_parents,
81 int *drained_end_counter)
85 QLIST_FOREACH(c, &bs->parents, next_parent) {
86 if (c == ignore || (ignore_bds_parents && c->klass->parent_is_bds)) {
89 bdrv_parent_drained_end_single_no_poll(c, drained_end_counter);
93 static bool bdrv_parent_drained_poll_single(BdrvChild *c)
95 if (c->klass->drained_poll) {
96 return c->klass->drained_poll(c);
101 static bool bdrv_parent_drained_poll(BlockDriverState *bs, BdrvChild *ignore,
102 bool ignore_bds_parents)
107 QLIST_FOREACH_SAFE(c, &bs->parents, next_parent, next) {
108 if (c == ignore || (ignore_bds_parents && c->klass->parent_is_bds)) {
111 busy |= bdrv_parent_drained_poll_single(c);
117 void bdrv_parent_drained_begin_single(BdrvChild *c, bool poll)
120 c->parent_quiesce_counter++;
121 if (c->klass->drained_begin) {
122 c->klass->drained_begin(c);
125 BDRV_POLL_WHILE(c->bs, bdrv_parent_drained_poll_single(c));
129 static void bdrv_merge_limits(BlockLimits *dst, const BlockLimits *src)
131 dst->pdiscard_alignment = MAX(dst->pdiscard_alignment,
132 src->pdiscard_alignment);
133 dst->opt_transfer = MAX(dst->opt_transfer, src->opt_transfer);
134 dst->max_transfer = MIN_NON_ZERO(dst->max_transfer, src->max_transfer);
135 dst->max_hw_transfer = MIN_NON_ZERO(dst->max_hw_transfer,
136 src->max_hw_transfer);
137 dst->opt_mem_alignment = MAX(dst->opt_mem_alignment,
138 src->opt_mem_alignment);
139 dst->min_mem_alignment = MAX(dst->min_mem_alignment,
140 src->min_mem_alignment);
141 dst->max_iov = MIN_NON_ZERO(dst->max_iov, src->max_iov);
142 dst->max_hw_iov = MIN_NON_ZERO(dst->max_hw_iov, src->max_hw_iov);
145 typedef struct BdrvRefreshLimitsState {
146 BlockDriverState *bs;
148 } BdrvRefreshLimitsState;
150 static void bdrv_refresh_limits_abort(void *opaque)
152 BdrvRefreshLimitsState *s = opaque;
154 s->bs->bl = s->old_bl;
157 static TransactionActionDrv bdrv_refresh_limits_drv = {
158 .abort = bdrv_refresh_limits_abort,
162 /* @tran is allowed to be NULL, in this case no rollback is possible. */
163 void bdrv_refresh_limits(BlockDriverState *bs, Transaction *tran, Error **errp)
166 BlockDriver *drv = bs->drv;
173 BdrvRefreshLimitsState *s = g_new(BdrvRefreshLimitsState, 1);
174 *s = (BdrvRefreshLimitsState) {
178 tran_add(tran, &bdrv_refresh_limits_drv, s);
181 memset(&bs->bl, 0, sizeof(bs->bl));
187 /* Default alignment based on whether driver has byte interface */
188 bs->bl.request_alignment = (drv->bdrv_co_preadv ||
189 drv->bdrv_aio_preadv ||
190 drv->bdrv_co_preadv_part) ? 1 : 512;
192 /* Take some limits from the children as a default */
194 QLIST_FOREACH(c, &bs->children, next) {
195 if (c->role & (BDRV_CHILD_DATA | BDRV_CHILD_FILTERED | BDRV_CHILD_COW))
197 bdrv_merge_limits(&bs->bl, &c->bs->bl);
203 bs->bl.min_mem_alignment = 512;
204 bs->bl.opt_mem_alignment = qemu_real_host_page_size();
206 /* Safe default since most protocols use readv()/writev()/etc */
207 bs->bl.max_iov = IOV_MAX;
210 /* Then let the driver override it */
211 if (drv->bdrv_refresh_limits) {
212 drv->bdrv_refresh_limits(bs, errp);
218 if (bs->bl.request_alignment > BDRV_MAX_ALIGNMENT) {
219 error_setg(errp, "Driver requires too large request alignment");
224 * The copy-on-read flag is actually a reference count so multiple users may
225 * use the feature without worrying about clobbering its previous state.
226 * Copy-on-read stays enabled until all users have called to disable it.
228 void bdrv_enable_copy_on_read(BlockDriverState *bs)
231 qatomic_inc(&bs->copy_on_read);
234 void bdrv_disable_copy_on_read(BlockDriverState *bs)
236 int old = qatomic_fetch_dec(&bs->copy_on_read);
243 BlockDriverState *bs;
249 bool ignore_bds_parents;
250 int *drained_end_counter;
253 static void coroutine_fn bdrv_drain_invoke_entry(void *opaque)
255 BdrvCoDrainData *data = opaque;
256 BlockDriverState *bs = data->bs;
259 bs->drv->bdrv_co_drain_begin(bs);
261 bs->drv->bdrv_co_drain_end(bs);
264 /* Set data->done and decrement drained_end_counter before bdrv_wakeup() */
265 qatomic_mb_set(&data->done, true);
267 qatomic_dec(data->drained_end_counter);
269 bdrv_dec_in_flight(bs);
274 /* Recursively call BlockDriver.bdrv_co_drain_begin/end callbacks */
275 static void bdrv_drain_invoke(BlockDriverState *bs, bool begin,
276 int *drained_end_counter)
278 BdrvCoDrainData *data;
280 if (!bs->drv || (begin && !bs->drv->bdrv_co_drain_begin) ||
281 (!begin && !bs->drv->bdrv_co_drain_end)) {
285 data = g_new(BdrvCoDrainData, 1);
286 *data = (BdrvCoDrainData) {
290 .drained_end_counter = drained_end_counter,
294 qatomic_inc(drained_end_counter);
297 /* Make sure the driver callback completes during the polling phase for
299 bdrv_inc_in_flight(bs);
300 data->co = qemu_coroutine_create(bdrv_drain_invoke_entry, data);
301 aio_co_schedule(bdrv_get_aio_context(bs), data->co);
304 /* Returns true if BDRV_POLL_WHILE() should go into a blocking aio_poll() */
305 bool bdrv_drain_poll(BlockDriverState *bs, bool recursive,
306 BdrvChild *ignore_parent, bool ignore_bds_parents)
308 BdrvChild *child, *next;
311 if (bdrv_parent_drained_poll(bs, ignore_parent, ignore_bds_parents)) {
315 if (qatomic_read(&bs->in_flight)) {
320 assert(!ignore_bds_parents);
321 QLIST_FOREACH_SAFE(child, &bs->children, next, next) {
322 if (bdrv_drain_poll(child->bs, recursive, child, false)) {
331 static bool bdrv_drain_poll_top_level(BlockDriverState *bs, bool recursive,
332 BdrvChild *ignore_parent)
334 return bdrv_drain_poll(bs, recursive, ignore_parent, false);
337 static void bdrv_do_drained_begin(BlockDriverState *bs, bool recursive,
338 BdrvChild *parent, bool ignore_bds_parents,
340 static void bdrv_do_drained_end(BlockDriverState *bs, bool recursive,
341 BdrvChild *parent, bool ignore_bds_parents,
342 int *drained_end_counter);
344 static void bdrv_co_drain_bh_cb(void *opaque)
346 BdrvCoDrainData *data = opaque;
347 Coroutine *co = data->co;
348 BlockDriverState *bs = data->bs;
351 AioContext *ctx = bdrv_get_aio_context(bs);
352 aio_context_acquire(ctx);
353 bdrv_dec_in_flight(bs);
355 assert(!data->drained_end_counter);
356 bdrv_do_drained_begin(bs, data->recursive, data->parent,
357 data->ignore_bds_parents, data->poll);
360 bdrv_do_drained_end(bs, data->recursive, data->parent,
361 data->ignore_bds_parents,
362 data->drained_end_counter);
364 aio_context_release(ctx);
367 bdrv_drain_all_begin();
374 static void coroutine_fn bdrv_co_yield_to_drain(BlockDriverState *bs,
375 bool begin, bool recursive,
377 bool ignore_bds_parents,
379 int *drained_end_counter)
381 BdrvCoDrainData data;
382 Coroutine *self = qemu_coroutine_self();
383 AioContext *ctx = bdrv_get_aio_context(bs);
384 AioContext *co_ctx = qemu_coroutine_get_aio_context(self);
386 /* Calling bdrv_drain() from a BH ensures the current coroutine yields and
387 * other coroutines run if they were queued by aio_co_enter(). */
389 assert(qemu_in_coroutine());
390 data = (BdrvCoDrainData) {
395 .recursive = recursive,
397 .ignore_bds_parents = ignore_bds_parents,
399 .drained_end_counter = drained_end_counter,
403 bdrv_inc_in_flight(bs);
407 * Temporarily drop the lock across yield or we would get deadlocks.
408 * bdrv_co_drain_bh_cb() reaquires the lock as needed.
410 * When we yield below, the lock for the current context will be
411 * released, so if this is actually the lock that protects bs, don't drop
415 aio_context_release(ctx);
417 replay_bh_schedule_oneshot_event(ctx, bdrv_co_drain_bh_cb, &data);
419 qemu_coroutine_yield();
420 /* If we are resumed from some other event (such as an aio completion or a
421 * timer callback), it is a bug in the caller that should be fixed. */
424 /* Reaquire the AioContext of bs if we dropped it */
426 aio_context_acquire(ctx);
430 void bdrv_do_drained_begin_quiesce(BlockDriverState *bs,
431 BdrvChild *parent, bool ignore_bds_parents)
434 assert(!qemu_in_coroutine());
436 /* Stop things in parent-to-child order */
437 if (qatomic_fetch_inc(&bs->quiesce_counter) == 0) {
438 aio_disable_external(bdrv_get_aio_context(bs));
441 bdrv_parent_drained_begin(bs, parent, ignore_bds_parents);
442 bdrv_drain_invoke(bs, true, NULL);
445 static void bdrv_do_drained_begin(BlockDriverState *bs, bool recursive,
446 BdrvChild *parent, bool ignore_bds_parents,
449 BdrvChild *child, *next;
451 if (qemu_in_coroutine()) {
452 bdrv_co_yield_to_drain(bs, true, recursive, parent, ignore_bds_parents,
457 bdrv_do_drained_begin_quiesce(bs, parent, ignore_bds_parents);
460 assert(!ignore_bds_parents);
461 bs->recursive_quiesce_counter++;
462 QLIST_FOREACH_SAFE(child, &bs->children, next, next) {
463 bdrv_do_drained_begin(child->bs, true, child, ignore_bds_parents,
469 * Wait for drained requests to finish.
471 * Calling BDRV_POLL_WHILE() only once for the top-level node is okay: The
472 * call is needed so things in this AioContext can make progress even
473 * though we don't return to the main AioContext loop - this automatically
474 * includes other nodes in the same AioContext and therefore all child
478 assert(!ignore_bds_parents);
479 BDRV_POLL_WHILE(bs, bdrv_drain_poll_top_level(bs, recursive, parent));
483 void bdrv_drained_begin(BlockDriverState *bs)
486 bdrv_do_drained_begin(bs, false, NULL, false, true);
489 void bdrv_subtree_drained_begin(BlockDriverState *bs)
492 bdrv_do_drained_begin(bs, true, NULL, false, true);
496 * This function does not poll, nor must any of its recursively called
497 * functions. The *drained_end_counter pointee will be incremented
498 * once for every background operation scheduled, and decremented once
499 * the operation settles. Therefore, the pointer must remain valid
500 * until the pointee reaches 0. That implies that whoever sets up the
501 * pointee has to poll until it is 0.
503 * We use atomic operations to access *drained_end_counter, because
504 * (1) when called from bdrv_set_aio_context_ignore(), the subgraph of
505 * @bs may contain nodes in different AioContexts,
506 * (2) bdrv_drain_all_end() uses the same counter for all nodes,
507 * regardless of which AioContext they are in.
509 static void bdrv_do_drained_end(BlockDriverState *bs, bool recursive,
510 BdrvChild *parent, bool ignore_bds_parents,
511 int *drained_end_counter)
514 int old_quiesce_counter;
516 assert(drained_end_counter != NULL);
518 if (qemu_in_coroutine()) {
519 bdrv_co_yield_to_drain(bs, false, recursive, parent, ignore_bds_parents,
520 false, drained_end_counter);
523 assert(bs->quiesce_counter > 0);
525 /* Re-enable things in child-to-parent order */
526 bdrv_drain_invoke(bs, false, drained_end_counter);
527 bdrv_parent_drained_end(bs, parent, ignore_bds_parents,
528 drained_end_counter);
530 old_quiesce_counter = qatomic_fetch_dec(&bs->quiesce_counter);
531 if (old_quiesce_counter == 1) {
532 aio_enable_external(bdrv_get_aio_context(bs));
536 assert(!ignore_bds_parents);
537 bs->recursive_quiesce_counter--;
538 QLIST_FOREACH(child, &bs->children, next) {
539 bdrv_do_drained_end(child->bs, true, child, ignore_bds_parents,
540 drained_end_counter);
545 void bdrv_drained_end(BlockDriverState *bs)
547 int drained_end_counter = 0;
549 bdrv_do_drained_end(bs, false, NULL, false, &drained_end_counter);
550 BDRV_POLL_WHILE(bs, qatomic_read(&drained_end_counter) > 0);
553 void bdrv_drained_end_no_poll(BlockDriverState *bs, int *drained_end_counter)
556 bdrv_do_drained_end(bs, false, NULL, false, drained_end_counter);
559 void bdrv_subtree_drained_end(BlockDriverState *bs)
561 int drained_end_counter = 0;
563 bdrv_do_drained_end(bs, true, NULL, false, &drained_end_counter);
564 BDRV_POLL_WHILE(bs, qatomic_read(&drained_end_counter) > 0);
567 void bdrv_apply_subtree_drain(BdrvChild *child, BlockDriverState *new_parent)
572 for (i = 0; i < new_parent->recursive_quiesce_counter; i++) {
573 bdrv_do_drained_begin(child->bs, true, child, false, true);
577 void bdrv_unapply_subtree_drain(BdrvChild *child, BlockDriverState *old_parent)
579 int drained_end_counter = 0;
583 for (i = 0; i < old_parent->recursive_quiesce_counter; i++) {
584 bdrv_do_drained_end(child->bs, true, child, false,
585 &drained_end_counter);
588 BDRV_POLL_WHILE(child->bs, qatomic_read(&drained_end_counter) > 0);
591 void bdrv_drain(BlockDriverState *bs)
594 bdrv_drained_begin(bs);
595 bdrv_drained_end(bs);
598 static void bdrv_drain_assert_idle(BlockDriverState *bs)
600 BdrvChild *child, *next;
602 assert(qatomic_read(&bs->in_flight) == 0);
603 QLIST_FOREACH_SAFE(child, &bs->children, next, next) {
604 bdrv_drain_assert_idle(child->bs);
608 unsigned int bdrv_drain_all_count = 0;
610 static bool bdrv_drain_all_poll(void)
612 BlockDriverState *bs = NULL;
616 /* bdrv_drain_poll() can't make changes to the graph and we are holding the
617 * main AioContext lock, so iterating bdrv_next_all_states() is safe. */
618 while ((bs = bdrv_next_all_states(bs))) {
619 AioContext *aio_context = bdrv_get_aio_context(bs);
620 aio_context_acquire(aio_context);
621 result |= bdrv_drain_poll(bs, false, NULL, true);
622 aio_context_release(aio_context);
629 * Wait for pending requests to complete across all BlockDriverStates
631 * This function does not flush data to disk, use bdrv_flush_all() for that
632 * after calling this function.
634 * This pauses all block jobs and disables external clients. It must
635 * be paired with bdrv_drain_all_end().
637 * NOTE: no new block jobs or BlockDriverStates can be created between
638 * the bdrv_drain_all_begin() and bdrv_drain_all_end() calls.
640 void bdrv_drain_all_begin(void)
642 BlockDriverState *bs = NULL;
645 if (qemu_in_coroutine()) {
646 bdrv_co_yield_to_drain(NULL, true, false, NULL, true, true, NULL);
651 * bdrv queue is managed by record/replay,
652 * waiting for finishing the I/O requests may
655 if (replay_events_enabled()) {
659 /* AIO_WAIT_WHILE() with a NULL context can only be called from the main
660 * loop AioContext, so make sure we're in the main context. */
661 assert(qemu_get_current_aio_context() == qemu_get_aio_context());
662 assert(bdrv_drain_all_count < INT_MAX);
663 bdrv_drain_all_count++;
665 /* Quiesce all nodes, without polling in-flight requests yet. The graph
666 * cannot change during this loop. */
667 while ((bs = bdrv_next_all_states(bs))) {
668 AioContext *aio_context = bdrv_get_aio_context(bs);
670 aio_context_acquire(aio_context);
671 bdrv_do_drained_begin(bs, false, NULL, true, false);
672 aio_context_release(aio_context);
675 /* Now poll the in-flight requests */
676 AIO_WAIT_WHILE(NULL, bdrv_drain_all_poll());
678 while ((bs = bdrv_next_all_states(bs))) {
679 bdrv_drain_assert_idle(bs);
683 void bdrv_drain_all_end_quiesce(BlockDriverState *bs)
685 int drained_end_counter = 0;
688 g_assert(bs->quiesce_counter > 0);
689 g_assert(!bs->refcnt);
691 while (bs->quiesce_counter) {
692 bdrv_do_drained_end(bs, false, NULL, true, &drained_end_counter);
694 BDRV_POLL_WHILE(bs, qatomic_read(&drained_end_counter) > 0);
697 void bdrv_drain_all_end(void)
699 BlockDriverState *bs = NULL;
700 int drained_end_counter = 0;
704 * bdrv queue is managed by record/replay,
705 * waiting for finishing the I/O requests may
708 if (replay_events_enabled()) {
712 while ((bs = bdrv_next_all_states(bs))) {
713 AioContext *aio_context = bdrv_get_aio_context(bs);
715 aio_context_acquire(aio_context);
716 bdrv_do_drained_end(bs, false, NULL, true, &drained_end_counter);
717 aio_context_release(aio_context);
720 assert(qemu_get_current_aio_context() == qemu_get_aio_context());
721 AIO_WAIT_WHILE(NULL, qatomic_read(&drained_end_counter) > 0);
723 assert(bdrv_drain_all_count > 0);
724 bdrv_drain_all_count--;
727 void bdrv_drain_all(void)
730 bdrv_drain_all_begin();
731 bdrv_drain_all_end();
735 * Remove an active request from the tracked requests list
737 * This function should be called when a tracked request is completing.
739 static void coroutine_fn tracked_request_end(BdrvTrackedRequest *req)
741 if (req->serialising) {
742 qatomic_dec(&req->bs->serialising_in_flight);
745 qemu_co_mutex_lock(&req->bs->reqs_lock);
746 QLIST_REMOVE(req, list);
747 qemu_co_queue_restart_all(&req->wait_queue);
748 qemu_co_mutex_unlock(&req->bs->reqs_lock);
752 * Add an active request to the tracked requests list
754 static void tracked_request_begin(BdrvTrackedRequest *req,
755 BlockDriverState *bs,
758 enum BdrvTrackedRequestType type)
760 bdrv_check_request(offset, bytes, &error_abort);
762 *req = (BdrvTrackedRequest){
767 .co = qemu_coroutine_self(),
768 .serialising = false,
769 .overlap_offset = offset,
770 .overlap_bytes = bytes,
773 qemu_co_queue_init(&req->wait_queue);
775 qemu_co_mutex_lock(&bs->reqs_lock);
776 QLIST_INSERT_HEAD(&bs->tracked_requests, req, list);
777 qemu_co_mutex_unlock(&bs->reqs_lock);
780 static bool tracked_request_overlaps(BdrvTrackedRequest *req,
781 int64_t offset, int64_t bytes)
783 bdrv_check_request(offset, bytes, &error_abort);
786 if (offset >= req->overlap_offset + req->overlap_bytes) {
790 if (req->overlap_offset >= offset + bytes) {
796 /* Called with self->bs->reqs_lock held */
797 static BdrvTrackedRequest *
798 bdrv_find_conflicting_request(BdrvTrackedRequest *self)
800 BdrvTrackedRequest *req;
802 QLIST_FOREACH(req, &self->bs->tracked_requests, list) {
803 if (req == self || (!req->serialising && !self->serialising)) {
806 if (tracked_request_overlaps(req, self->overlap_offset,
807 self->overlap_bytes))
810 * Hitting this means there was a reentrant request, for
811 * example, a block driver issuing nested requests. This must
812 * never happen since it means deadlock.
814 assert(qemu_coroutine_self() != req->co);
817 * If the request is already (indirectly) waiting for us, or
818 * will wait for us as soon as it wakes up, then just go on
819 * (instead of producing a deadlock in the former case).
821 if (!req->waiting_for) {
830 /* Called with self->bs->reqs_lock held */
831 static bool coroutine_fn
832 bdrv_wait_serialising_requests_locked(BdrvTrackedRequest *self)
834 BdrvTrackedRequest *req;
837 while ((req = bdrv_find_conflicting_request(self))) {
838 self->waiting_for = req;
839 qemu_co_queue_wait(&req->wait_queue, &self->bs->reqs_lock);
840 self->waiting_for = NULL;
847 /* Called with req->bs->reqs_lock held */
848 static void tracked_request_set_serialising(BdrvTrackedRequest *req,
851 int64_t overlap_offset = req->offset & ~(align - 1);
852 int64_t overlap_bytes =
853 ROUND_UP(req->offset + req->bytes, align) - overlap_offset;
855 bdrv_check_request(req->offset, req->bytes, &error_abort);
857 if (!req->serialising) {
858 qatomic_inc(&req->bs->serialising_in_flight);
859 req->serialising = true;
862 req->overlap_offset = MIN(req->overlap_offset, overlap_offset);
863 req->overlap_bytes = MAX(req->overlap_bytes, overlap_bytes);
867 * Return the tracked request on @bs for the current coroutine, or
868 * NULL if there is none.
870 BdrvTrackedRequest *coroutine_fn bdrv_co_get_self_request(BlockDriverState *bs)
872 BdrvTrackedRequest *req;
873 Coroutine *self = qemu_coroutine_self();
876 QLIST_FOREACH(req, &bs->tracked_requests, list) {
877 if (req->co == self) {
886 * Round a region to cluster boundaries
888 void bdrv_round_to_clusters(BlockDriverState *bs,
889 int64_t offset, int64_t bytes,
890 int64_t *cluster_offset,
891 int64_t *cluster_bytes)
895 if (bdrv_get_info(bs, &bdi) < 0 || bdi.cluster_size == 0) {
896 *cluster_offset = offset;
897 *cluster_bytes = bytes;
899 int64_t c = bdi.cluster_size;
900 *cluster_offset = QEMU_ALIGN_DOWN(offset, c);
901 *cluster_bytes = QEMU_ALIGN_UP(offset - *cluster_offset + bytes, c);
905 static int bdrv_get_cluster_size(BlockDriverState *bs)
910 ret = bdrv_get_info(bs, &bdi);
911 if (ret < 0 || bdi.cluster_size == 0) {
912 return bs->bl.request_alignment;
914 return bdi.cluster_size;
918 void bdrv_inc_in_flight(BlockDriverState *bs)
921 qatomic_inc(&bs->in_flight);
924 void bdrv_wakeup(BlockDriverState *bs)
930 void bdrv_dec_in_flight(BlockDriverState *bs)
933 qatomic_dec(&bs->in_flight);
937 static bool coroutine_fn bdrv_wait_serialising_requests(BdrvTrackedRequest *self)
939 BlockDriverState *bs = self->bs;
942 if (!qatomic_read(&bs->serialising_in_flight)) {
946 qemu_co_mutex_lock(&bs->reqs_lock);
947 waited = bdrv_wait_serialising_requests_locked(self);
948 qemu_co_mutex_unlock(&bs->reqs_lock);
953 bool coroutine_fn bdrv_make_request_serialising(BdrvTrackedRequest *req,
959 qemu_co_mutex_lock(&req->bs->reqs_lock);
961 tracked_request_set_serialising(req, align);
962 waited = bdrv_wait_serialising_requests_locked(req);
964 qemu_co_mutex_unlock(&req->bs->reqs_lock);
969 int bdrv_check_qiov_request(int64_t offset, int64_t bytes,
970 QEMUIOVector *qiov, size_t qiov_offset,
974 * Check generic offset/bytes correctness
978 error_setg(errp, "offset is negative: %" PRIi64, offset);
983 error_setg(errp, "bytes is negative: %" PRIi64, bytes);
987 if (bytes > BDRV_MAX_LENGTH) {
988 error_setg(errp, "bytes(%" PRIi64 ") exceeds maximum(%" PRIi64 ")",
989 bytes, BDRV_MAX_LENGTH);
993 if (offset > BDRV_MAX_LENGTH) {
994 error_setg(errp, "offset(%" PRIi64 ") exceeds maximum(%" PRIi64 ")",
995 offset, BDRV_MAX_LENGTH);
999 if (offset > BDRV_MAX_LENGTH - bytes) {
1000 error_setg(errp, "sum of offset(%" PRIi64 ") and bytes(%" PRIi64 ") "
1001 "exceeds maximum(%" PRIi64 ")", offset, bytes,
1011 * Check qiov and qiov_offset
1014 if (qiov_offset > qiov->size) {
1015 error_setg(errp, "qiov_offset(%zu) overflow io vector size(%zu)",
1016 qiov_offset, qiov->size);
1020 if (bytes > qiov->size - qiov_offset) {
1021 error_setg(errp, "bytes(%" PRIi64 ") + qiov_offset(%zu) overflow io "
1022 "vector size(%zu)", bytes, qiov_offset, qiov->size);
1029 int bdrv_check_request(int64_t offset, int64_t bytes, Error **errp)
1031 return bdrv_check_qiov_request(offset, bytes, NULL, 0, errp);
1034 static int bdrv_check_request32(int64_t offset, int64_t bytes,
1035 QEMUIOVector *qiov, size_t qiov_offset)
1037 int ret = bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, NULL);
1042 if (bytes > BDRV_REQUEST_MAX_BYTES) {
1050 * Completely zero out a block device with the help of bdrv_pwrite_zeroes.
1051 * The operation is sped up by checking the block status and only writing
1052 * zeroes to the device if they currently do not return zeroes. Optional
1053 * flags are passed through to bdrv_pwrite_zeroes (e.g. BDRV_REQ_MAY_UNMAP,
1056 * Returns < 0 on error, 0 on success. For error codes see bdrv_pwrite().
1058 int bdrv_make_zero(BdrvChild *child, BdrvRequestFlags flags)
1061 int64_t target_size, bytes, offset = 0;
1062 BlockDriverState *bs = child->bs;
1065 target_size = bdrv_getlength(bs);
1066 if (target_size < 0) {
1071 bytes = MIN(target_size - offset, BDRV_REQUEST_MAX_BYTES);
1075 ret = bdrv_block_status(bs, offset, bytes, &bytes, NULL, NULL);
1079 if (ret & BDRV_BLOCK_ZERO) {
1083 ret = bdrv_pwrite_zeroes(child, offset, bytes, flags);
1092 * Writes to the file and ensures that no writes are reordered across this
1093 * request (acts as a barrier)
1095 * Returns 0 on success, -errno in error cases.
1097 int coroutine_fn bdrv_co_pwrite_sync(BdrvChild *child, int64_t offset,
1098 int64_t bytes, const void *buf,
1099 BdrvRequestFlags flags)
1104 ret = bdrv_co_pwrite(child, offset, bytes, buf, flags);
1109 ret = bdrv_co_flush(child->bs);
1117 typedef struct CoroutineIOCompletion {
1118 Coroutine *coroutine;
1120 } CoroutineIOCompletion;
1122 static void bdrv_co_io_em_complete(void *opaque, int ret)
1124 CoroutineIOCompletion *co = opaque;
1127 aio_co_wake(co->coroutine);
1130 static int coroutine_fn bdrv_driver_preadv(BlockDriverState *bs,
1131 int64_t offset, int64_t bytes,
1133 size_t qiov_offset, int flags)
1135 BlockDriver *drv = bs->drv;
1137 unsigned int nb_sectors;
1138 QEMUIOVector local_qiov;
1141 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort);
1142 assert(!(flags & ~BDRV_REQ_MASK));
1143 assert(!(flags & BDRV_REQ_NO_FALLBACK));
1149 if (drv->bdrv_co_preadv_part) {
1150 return drv->bdrv_co_preadv_part(bs, offset, bytes, qiov, qiov_offset,
1154 if (qiov_offset > 0 || bytes != qiov->size) {
1155 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes);
1159 if (drv->bdrv_co_preadv) {
1160 ret = drv->bdrv_co_preadv(bs, offset, bytes, qiov, flags);
1164 if (drv->bdrv_aio_preadv) {
1166 CoroutineIOCompletion co = {
1167 .coroutine = qemu_coroutine_self(),
1170 acb = drv->bdrv_aio_preadv(bs, offset, bytes, qiov, flags,
1171 bdrv_co_io_em_complete, &co);
1176 qemu_coroutine_yield();
1182 sector_num = offset >> BDRV_SECTOR_BITS;
1183 nb_sectors = bytes >> BDRV_SECTOR_BITS;
1185 assert(QEMU_IS_ALIGNED(offset, BDRV_SECTOR_SIZE));
1186 assert(QEMU_IS_ALIGNED(bytes, BDRV_SECTOR_SIZE));
1187 assert(bytes <= BDRV_REQUEST_MAX_BYTES);
1188 assert(drv->bdrv_co_readv);
1190 ret = drv->bdrv_co_readv(bs, sector_num, nb_sectors, qiov);
1193 if (qiov == &local_qiov) {
1194 qemu_iovec_destroy(&local_qiov);
1200 static int coroutine_fn bdrv_driver_pwritev(BlockDriverState *bs,
1201 int64_t offset, int64_t bytes,
1204 BdrvRequestFlags flags)
1206 BlockDriver *drv = bs->drv;
1208 unsigned int nb_sectors;
1209 QEMUIOVector local_qiov;
1212 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort);
1213 assert(!(flags & ~BDRV_REQ_MASK));
1214 assert(!(flags & BDRV_REQ_NO_FALLBACK));
1220 if (drv->bdrv_co_pwritev_part) {
1221 ret = drv->bdrv_co_pwritev_part(bs, offset, bytes, qiov, qiov_offset,
1222 flags & bs->supported_write_flags);
1223 flags &= ~bs->supported_write_flags;
1227 if (qiov_offset > 0 || bytes != qiov->size) {
1228 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes);
1232 if (drv->bdrv_co_pwritev) {
1233 ret = drv->bdrv_co_pwritev(bs, offset, bytes, qiov,
1234 flags & bs->supported_write_flags);
1235 flags &= ~bs->supported_write_flags;
1239 if (drv->bdrv_aio_pwritev) {
1241 CoroutineIOCompletion co = {
1242 .coroutine = qemu_coroutine_self(),
1245 acb = drv->bdrv_aio_pwritev(bs, offset, bytes, qiov,
1246 flags & bs->supported_write_flags,
1247 bdrv_co_io_em_complete, &co);
1248 flags &= ~bs->supported_write_flags;
1252 qemu_coroutine_yield();
1258 sector_num = offset >> BDRV_SECTOR_BITS;
1259 nb_sectors = bytes >> BDRV_SECTOR_BITS;
1261 assert(QEMU_IS_ALIGNED(offset, BDRV_SECTOR_SIZE));
1262 assert(QEMU_IS_ALIGNED(bytes, BDRV_SECTOR_SIZE));
1263 assert(bytes <= BDRV_REQUEST_MAX_BYTES);
1265 assert(drv->bdrv_co_writev);
1266 ret = drv->bdrv_co_writev(bs, sector_num, nb_sectors, qiov,
1267 flags & bs->supported_write_flags);
1268 flags &= ~bs->supported_write_flags;
1271 if (ret == 0 && (flags & BDRV_REQ_FUA)) {
1272 ret = bdrv_co_flush(bs);
1275 if (qiov == &local_qiov) {
1276 qemu_iovec_destroy(&local_qiov);
1282 static int coroutine_fn
1283 bdrv_driver_pwritev_compressed(BlockDriverState *bs, int64_t offset,
1284 int64_t bytes, QEMUIOVector *qiov,
1287 BlockDriver *drv = bs->drv;
1288 QEMUIOVector local_qiov;
1291 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort);
1297 if (!block_driver_can_compress(drv)) {
1301 if (drv->bdrv_co_pwritev_compressed_part) {
1302 return drv->bdrv_co_pwritev_compressed_part(bs, offset, bytes,
1306 if (qiov_offset == 0) {
1307 return drv->bdrv_co_pwritev_compressed(bs, offset, bytes, qiov);
1310 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes);
1311 ret = drv->bdrv_co_pwritev_compressed(bs, offset, bytes, &local_qiov);
1312 qemu_iovec_destroy(&local_qiov);
1317 static int coroutine_fn bdrv_co_do_copy_on_readv(BdrvChild *child,
1318 int64_t offset, int64_t bytes, QEMUIOVector *qiov,
1319 size_t qiov_offset, int flags)
1321 BlockDriverState *bs = child->bs;
1323 /* Perform I/O through a temporary buffer so that users who scribble over
1324 * their read buffer while the operation is in progress do not end up
1325 * modifying the image file. This is critical for zero-copy guest I/O
1326 * where anything might happen inside guest memory.
1328 void *bounce_buffer = NULL;
1330 BlockDriver *drv = bs->drv;
1331 int64_t cluster_offset;
1332 int64_t cluster_bytes;
1335 int max_transfer = MIN_NON_ZERO(bs->bl.max_transfer,
1336 BDRV_REQUEST_MAX_BYTES);
1337 int64_t progress = 0;
1340 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort);
1347 * Do not write anything when the BDS is inactive. That is not
1348 * allowed, and it would not help.
1350 skip_write = (bs->open_flags & BDRV_O_INACTIVE);
1352 /* FIXME We cannot require callers to have write permissions when all they
1353 * are doing is a read request. If we did things right, write permissions
1354 * would be obtained anyway, but internally by the copy-on-read code. As
1355 * long as it is implemented here rather than in a separate filter driver,
1356 * the copy-on-read code doesn't have its own BdrvChild, however, for which
1357 * it could request permissions. Therefore we have to bypass the permission
1358 * system for the moment. */
1359 // assert(child->perm & (BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE));
1361 /* Cover entire cluster so no additional backing file I/O is required when
1362 * allocating cluster in the image file. Note that this value may exceed
1363 * BDRV_REQUEST_MAX_BYTES (even when the original read did not), which
1364 * is one reason we loop rather than doing it all at once.
1366 bdrv_round_to_clusters(bs, offset, bytes, &cluster_offset, &cluster_bytes);
1367 skip_bytes = offset - cluster_offset;
1369 trace_bdrv_co_do_copy_on_readv(bs, offset, bytes,
1370 cluster_offset, cluster_bytes);
1372 while (cluster_bytes) {
1376 ret = 1; /* "already allocated", so nothing will be copied */
1377 pnum = MIN(cluster_bytes, max_transfer);
1379 ret = bdrv_is_allocated(bs, cluster_offset,
1380 MIN(cluster_bytes, max_transfer), &pnum);
1383 * Safe to treat errors in querying allocation as if
1384 * unallocated; we'll probably fail again soon on the
1385 * read, but at least that will set a decent errno.
1387 pnum = MIN(cluster_bytes, max_transfer);
1390 /* Stop at EOF if the image ends in the middle of the cluster */
1391 if (ret == 0 && pnum == 0) {
1392 assert(progress >= bytes);
1396 assert(skip_bytes < pnum);
1400 QEMUIOVector local_qiov;
1402 /* Must copy-on-read; use the bounce buffer */
1403 pnum = MIN(pnum, MAX_BOUNCE_BUFFER);
1404 if (!bounce_buffer) {
1405 int64_t max_we_need = MAX(pnum, cluster_bytes - pnum);
1406 int64_t max_allowed = MIN(max_transfer, MAX_BOUNCE_BUFFER);
1407 int64_t bounce_buffer_len = MIN(max_we_need, max_allowed);
1409 bounce_buffer = qemu_try_blockalign(bs, bounce_buffer_len);
1410 if (!bounce_buffer) {
1415 qemu_iovec_init_buf(&local_qiov, bounce_buffer, pnum);
1417 ret = bdrv_driver_preadv(bs, cluster_offset, pnum,
1423 bdrv_debug_event(bs, BLKDBG_COR_WRITE);
1424 if (drv->bdrv_co_pwrite_zeroes &&
1425 buffer_is_zero(bounce_buffer, pnum)) {
1426 /* FIXME: Should we (perhaps conditionally) be setting
1427 * BDRV_REQ_MAY_UNMAP, if it will allow for a sparser copy
1428 * that still correctly reads as zero? */
1429 ret = bdrv_co_do_pwrite_zeroes(bs, cluster_offset, pnum,
1430 BDRV_REQ_WRITE_UNCHANGED);
1432 /* This does not change the data on the disk, it is not
1433 * necessary to flush even in cache=writethrough mode.
1435 ret = bdrv_driver_pwritev(bs, cluster_offset, pnum,
1437 BDRV_REQ_WRITE_UNCHANGED);
1441 /* It might be okay to ignore write errors for guest
1442 * requests. If this is a deliberate copy-on-read
1443 * then we don't want to ignore the error. Simply
1444 * report it in all cases.
1449 if (!(flags & BDRV_REQ_PREFETCH)) {
1450 qemu_iovec_from_buf(qiov, qiov_offset + progress,
1451 bounce_buffer + skip_bytes,
1452 MIN(pnum - skip_bytes, bytes - progress));
1454 } else if (!(flags & BDRV_REQ_PREFETCH)) {
1455 /* Read directly into the destination */
1456 ret = bdrv_driver_preadv(bs, offset + progress,
1457 MIN(pnum - skip_bytes, bytes - progress),
1458 qiov, qiov_offset + progress, 0);
1464 cluster_offset += pnum;
1465 cluster_bytes -= pnum;
1466 progress += pnum - skip_bytes;
1472 qemu_vfree(bounce_buffer);
1477 * Forwards an already correctly aligned request to the BlockDriver. This
1478 * handles copy on read, zeroing after EOF, and fragmentation of large
1479 * reads; any other features must be implemented by the caller.
1481 static int coroutine_fn bdrv_aligned_preadv(BdrvChild *child,
1482 BdrvTrackedRequest *req, int64_t offset, int64_t bytes,
1483 int64_t align, QEMUIOVector *qiov, size_t qiov_offset, int flags)
1485 BlockDriverState *bs = child->bs;
1486 int64_t total_bytes, max_bytes;
1488 int64_t bytes_remaining = bytes;
1491 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort);
1492 assert(is_power_of_2(align));
1493 assert((offset & (align - 1)) == 0);
1494 assert((bytes & (align - 1)) == 0);
1495 assert((bs->open_flags & BDRV_O_NO_IO) == 0);
1496 max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX),
1499 /* TODO: We would need a per-BDS .supported_read_flags and
1500 * potential fallback support, if we ever implement any read flags
1501 * to pass through to drivers. For now, there aren't any
1502 * passthrough flags. */
1503 assert(!(flags & ~(BDRV_REQ_COPY_ON_READ | BDRV_REQ_PREFETCH)));
1505 /* Handle Copy on Read and associated serialisation */
1506 if (flags & BDRV_REQ_COPY_ON_READ) {
1507 /* If we touch the same cluster it counts as an overlap. This
1508 * guarantees that allocating writes will be serialized and not race
1509 * with each other for the same cluster. For example, in copy-on-read
1510 * it ensures that the CoR read and write operations are atomic and
1511 * guest writes cannot interleave between them. */
1512 bdrv_make_request_serialising(req, bdrv_get_cluster_size(bs));
1514 bdrv_wait_serialising_requests(req);
1517 if (flags & BDRV_REQ_COPY_ON_READ) {
1520 /* The flag BDRV_REQ_COPY_ON_READ has reached its addressee */
1521 flags &= ~BDRV_REQ_COPY_ON_READ;
1523 ret = bdrv_is_allocated(bs, offset, bytes, &pnum);
1528 if (!ret || pnum != bytes) {
1529 ret = bdrv_co_do_copy_on_readv(child, offset, bytes,
1530 qiov, qiov_offset, flags);
1532 } else if (flags & BDRV_REQ_PREFETCH) {
1537 /* Forward the request to the BlockDriver, possibly fragmenting it */
1538 total_bytes = bdrv_getlength(bs);
1539 if (total_bytes < 0) {
1544 assert(!(flags & ~bs->supported_read_flags));
1546 max_bytes = ROUND_UP(MAX(0, total_bytes - offset), align);
1547 if (bytes <= max_bytes && bytes <= max_transfer) {
1548 ret = bdrv_driver_preadv(bs, offset, bytes, qiov, qiov_offset, flags);
1552 while (bytes_remaining) {
1556 num = MIN(bytes_remaining, MIN(max_bytes, max_transfer));
1559 ret = bdrv_driver_preadv(bs, offset + bytes - bytes_remaining,
1561 qiov_offset + bytes - bytes_remaining,
1565 num = bytes_remaining;
1566 ret = qemu_iovec_memset(qiov, qiov_offset + bytes - bytes_remaining,
1567 0, bytes_remaining);
1572 bytes_remaining -= num;
1576 return ret < 0 ? ret : 0;
1582 * |<---- align ----->| |<----- align ---->|
1583 * |<- head ->|<------------- bytes ------------->|<-- tail -->|
1585 * -*----------$-------*-------- ... --------*-----$------------*---
1587 * | offset | | end |
1588 * ALIGN_DOWN(offset) ALIGN_UP(offset) ALIGN_DOWN(end) ALIGN_UP(end)
1589 * [buf ... ) [tail_buf )
1591 * @buf is an aligned allocation needed to store @head and @tail paddings. @head
1592 * is placed at the beginning of @buf and @tail at the @end.
1594 * @tail_buf is a pointer to sub-buffer, corresponding to align-sized chunk
1595 * around tail, if tail exists.
1597 * @merge_reads is true for small requests,
1598 * if @buf_len == @head + bytes + @tail. In this case it is possible that both
1599 * head and tail exist but @buf_len == align and @tail_buf == @buf.
1601 typedef struct BdrvRequestPadding {
1608 QEMUIOVector local_qiov;
1609 } BdrvRequestPadding;
1611 static bool bdrv_init_padding(BlockDriverState *bs,
1612 int64_t offset, int64_t bytes,
1613 BdrvRequestPadding *pad)
1615 int64_t align = bs->bl.request_alignment;
1618 bdrv_check_request(offset, bytes, &error_abort);
1619 assert(align <= INT_MAX); /* documented in block/block_int.h */
1620 assert(align <= SIZE_MAX / 2); /* so we can allocate the buffer */
1622 memset(pad, 0, sizeof(*pad));
1624 pad->head = offset & (align - 1);
1625 pad->tail = ((offset + bytes) & (align - 1));
1627 pad->tail = align - pad->tail;
1630 if (!pad->head && !pad->tail) {
1634 assert(bytes); /* Nothing good in aligning zero-length requests */
1636 sum = pad->head + bytes + pad->tail;
1637 pad->buf_len = (sum > align && pad->head && pad->tail) ? 2 * align : align;
1638 pad->buf = qemu_blockalign(bs, pad->buf_len);
1639 pad->merge_reads = sum == pad->buf_len;
1641 pad->tail_buf = pad->buf + pad->buf_len - align;
1647 static int bdrv_padding_rmw_read(BdrvChild *child,
1648 BdrvTrackedRequest *req,
1649 BdrvRequestPadding *pad,
1652 QEMUIOVector local_qiov;
1653 BlockDriverState *bs = child->bs;
1654 uint64_t align = bs->bl.request_alignment;
1657 assert(req->serialising && pad->buf);
1659 if (pad->head || pad->merge_reads) {
1660 int64_t bytes = pad->merge_reads ? pad->buf_len : align;
1662 qemu_iovec_init_buf(&local_qiov, pad->buf, bytes);
1665 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD);
1667 if (pad->merge_reads && pad->tail) {
1668 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
1670 ret = bdrv_aligned_preadv(child, req, req->overlap_offset, bytes,
1671 align, &local_qiov, 0, 0);
1676 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD);
1678 if (pad->merge_reads && pad->tail) {
1679 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
1682 if (pad->merge_reads) {
1688 qemu_iovec_init_buf(&local_qiov, pad->tail_buf, align);
1690 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
1691 ret = bdrv_aligned_preadv(
1693 req->overlap_offset + req->overlap_bytes - align,
1694 align, align, &local_qiov, 0, 0);
1698 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
1703 memset(pad->buf + pad->head, 0, pad->buf_len - pad->head - pad->tail);
1709 static void bdrv_padding_destroy(BdrvRequestPadding *pad)
1712 qemu_vfree(pad->buf);
1713 qemu_iovec_destroy(&pad->local_qiov);
1715 memset(pad, 0, sizeof(*pad));
1721 * Exchange request parameters with padded request if needed. Don't include RMW
1722 * read of padding, bdrv_padding_rmw_read() should be called separately if
1725 * Request parameters (@qiov, &qiov_offset, &offset, &bytes) are in-out:
1726 * - on function start they represent original request
1727 * - on failure or when padding is not needed they are unchanged
1728 * - on success when padding is needed they represent padded request
1730 static int bdrv_pad_request(BlockDriverState *bs,
1731 QEMUIOVector **qiov, size_t *qiov_offset,
1732 int64_t *offset, int64_t *bytes,
1733 BdrvRequestPadding *pad, bool *padded)
1737 bdrv_check_qiov_request(*offset, *bytes, *qiov, *qiov_offset, &error_abort);
1739 if (!bdrv_init_padding(bs, *offset, *bytes, pad)) {
1746 ret = qemu_iovec_init_extended(&pad->local_qiov, pad->buf, pad->head,
1747 *qiov, *qiov_offset, *bytes,
1748 pad->buf + pad->buf_len - pad->tail,
1751 bdrv_padding_destroy(pad);
1754 *bytes += pad->head + pad->tail;
1755 *offset -= pad->head;
1756 *qiov = &pad->local_qiov;
1765 int coroutine_fn bdrv_co_preadv(BdrvChild *child,
1766 int64_t offset, int64_t bytes, QEMUIOVector *qiov,
1767 BdrvRequestFlags flags)
1770 return bdrv_co_preadv_part(child, offset, bytes, qiov, 0, flags);
1773 int coroutine_fn bdrv_co_preadv_part(BdrvChild *child,
1774 int64_t offset, int64_t bytes,
1775 QEMUIOVector *qiov, size_t qiov_offset,
1776 BdrvRequestFlags flags)
1778 BlockDriverState *bs = child->bs;
1779 BdrvTrackedRequest req;
1780 BdrvRequestPadding pad;
1784 trace_bdrv_co_preadv_part(bs, offset, bytes, flags);
1786 if (!bdrv_is_inserted(bs)) {
1790 ret = bdrv_check_request32(offset, bytes, qiov, qiov_offset);
1795 if (bytes == 0 && !QEMU_IS_ALIGNED(offset, bs->bl.request_alignment)) {
1797 * Aligning zero request is nonsense. Even if driver has special meaning
1798 * of zero-length (like qcow2_co_pwritev_compressed_part), we can't pass
1799 * it to driver due to request_alignment.
1801 * Still, no reason to return an error if someone do unaligned
1802 * zero-length read occasionally.
1807 bdrv_inc_in_flight(bs);
1809 /* Don't do copy-on-read if we read data before write operation */
1810 if (qatomic_read(&bs->copy_on_read)) {
1811 flags |= BDRV_REQ_COPY_ON_READ;
1814 ret = bdrv_pad_request(bs, &qiov, &qiov_offset, &offset, &bytes, &pad,
1820 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_READ);
1821 ret = bdrv_aligned_preadv(child, &req, offset, bytes,
1822 bs->bl.request_alignment,
1823 qiov, qiov_offset, flags);
1824 tracked_request_end(&req);
1825 bdrv_padding_destroy(&pad);
1828 bdrv_dec_in_flight(bs);
1833 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs,
1834 int64_t offset, int64_t bytes, BdrvRequestFlags flags)
1836 BlockDriver *drv = bs->drv;
1840 bool need_flush = false;
1844 int64_t max_write_zeroes = MIN_NON_ZERO(bs->bl.max_pwrite_zeroes,
1846 int alignment = MAX(bs->bl.pwrite_zeroes_alignment,
1847 bs->bl.request_alignment);
1848 int max_transfer = MIN_NON_ZERO(bs->bl.max_transfer, MAX_BOUNCE_BUFFER);
1850 bdrv_check_request(offset, bytes, &error_abort);
1856 if ((flags & ~bs->supported_zero_flags) & BDRV_REQ_NO_FALLBACK) {
1860 /* Invalidate the cached block-status data range if this write overlaps */
1861 bdrv_bsc_invalidate_range(bs, offset, bytes);
1863 assert(alignment % bs->bl.request_alignment == 0);
1864 head = offset % alignment;
1865 tail = (offset + bytes) % alignment;
1866 max_write_zeroes = QEMU_ALIGN_DOWN(max_write_zeroes, alignment);
1867 assert(max_write_zeroes >= bs->bl.request_alignment);
1869 while (bytes > 0 && !ret) {
1870 int64_t num = bytes;
1872 /* Align request. Block drivers can expect the "bulk" of the request
1873 * to be aligned, and that unaligned requests do not cross cluster
1877 /* Make a small request up to the first aligned sector. For
1878 * convenience, limit this request to max_transfer even if
1879 * we don't need to fall back to writes. */
1880 num = MIN(MIN(bytes, max_transfer), alignment - head);
1881 head = (head + num) % alignment;
1882 assert(num < max_write_zeroes);
1883 } else if (tail && num > alignment) {
1884 /* Shorten the request to the last aligned sector. */
1888 /* limit request size */
1889 if (num > max_write_zeroes) {
1890 num = max_write_zeroes;
1894 /* First try the efficient write zeroes operation */
1895 if (drv->bdrv_co_pwrite_zeroes) {
1896 ret = drv->bdrv_co_pwrite_zeroes(bs, offset, num,
1897 flags & bs->supported_zero_flags);
1898 if (ret != -ENOTSUP && (flags & BDRV_REQ_FUA) &&
1899 !(bs->supported_zero_flags & BDRV_REQ_FUA)) {
1903 assert(!bs->supported_zero_flags);
1906 if (ret == -ENOTSUP && !(flags & BDRV_REQ_NO_FALLBACK)) {
1907 /* Fall back to bounce buffer if write zeroes is unsupported */
1908 BdrvRequestFlags write_flags = flags & ~BDRV_REQ_ZERO_WRITE;
1910 if ((flags & BDRV_REQ_FUA) &&
1911 !(bs->supported_write_flags & BDRV_REQ_FUA)) {
1912 /* No need for bdrv_driver_pwrite() to do a fallback
1913 * flush on each chunk; use just one at the end */
1914 write_flags &= ~BDRV_REQ_FUA;
1917 num = MIN(num, max_transfer);
1919 buf = qemu_try_blockalign0(bs, num);
1925 qemu_iovec_init_buf(&qiov, buf, num);
1927 ret = bdrv_driver_pwritev(bs, offset, num, &qiov, 0, write_flags);
1929 /* Keep bounce buffer around if it is big enough for all
1930 * all future requests.
1932 if (num < max_transfer) {
1943 if (ret == 0 && need_flush) {
1944 ret = bdrv_co_flush(bs);
1950 static inline int coroutine_fn
1951 bdrv_co_write_req_prepare(BdrvChild *child, int64_t offset, int64_t bytes,
1952 BdrvTrackedRequest *req, int flags)
1954 BlockDriverState *bs = child->bs;
1956 bdrv_check_request(offset, bytes, &error_abort);
1958 if (bdrv_is_read_only(bs)) {
1962 assert(!(bs->open_flags & BDRV_O_INACTIVE));
1963 assert((bs->open_flags & BDRV_O_NO_IO) == 0);
1964 assert(!(flags & ~BDRV_REQ_MASK));
1965 assert(!((flags & BDRV_REQ_NO_WAIT) && !(flags & BDRV_REQ_SERIALISING)));
1967 if (flags & BDRV_REQ_SERIALISING) {
1968 QEMU_LOCK_GUARD(&bs->reqs_lock);
1970 tracked_request_set_serialising(req, bdrv_get_cluster_size(bs));
1972 if ((flags & BDRV_REQ_NO_WAIT) && bdrv_find_conflicting_request(req)) {
1976 bdrv_wait_serialising_requests_locked(req);
1978 bdrv_wait_serialising_requests(req);
1981 assert(req->overlap_offset <= offset);
1982 assert(offset + bytes <= req->overlap_offset + req->overlap_bytes);
1983 assert(offset + bytes <= bs->total_sectors * BDRV_SECTOR_SIZE ||
1984 child->perm & BLK_PERM_RESIZE);
1986 switch (req->type) {
1987 case BDRV_TRACKED_WRITE:
1988 case BDRV_TRACKED_DISCARD:
1989 if (flags & BDRV_REQ_WRITE_UNCHANGED) {
1990 assert(child->perm & (BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE));
1992 assert(child->perm & BLK_PERM_WRITE);
1994 bdrv_write_threshold_check_write(bs, offset, bytes);
1996 case BDRV_TRACKED_TRUNCATE:
1997 assert(child->perm & BLK_PERM_RESIZE);
2004 static inline void coroutine_fn
2005 bdrv_co_write_req_finish(BdrvChild *child, int64_t offset, int64_t bytes,
2006 BdrvTrackedRequest *req, int ret)
2008 int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);
2009 BlockDriverState *bs = child->bs;
2011 bdrv_check_request(offset, bytes, &error_abort);
2013 qatomic_inc(&bs->write_gen);
2016 * Discard cannot extend the image, but in error handling cases, such as
2017 * when reverting a qcow2 cluster allocation, the discarded range can pass
2018 * the end of image file, so we cannot assert about BDRV_TRACKED_DISCARD
2019 * here. Instead, just skip it, since semantically a discard request
2020 * beyond EOF cannot expand the image anyway.
2023 (req->type == BDRV_TRACKED_TRUNCATE ||
2024 end_sector > bs->total_sectors) &&
2025 req->type != BDRV_TRACKED_DISCARD) {
2026 bs->total_sectors = end_sector;
2027 bdrv_parent_cb_resize(bs);
2028 bdrv_dirty_bitmap_truncate(bs, end_sector << BDRV_SECTOR_BITS);
2031 switch (req->type) {
2032 case BDRV_TRACKED_WRITE:
2033 stat64_max(&bs->wr_highest_offset, offset + bytes);
2034 /* fall through, to set dirty bits */
2035 case BDRV_TRACKED_DISCARD:
2036 bdrv_set_dirty(bs, offset, bytes);
2045 * Forwards an already correctly aligned write request to the BlockDriver,
2046 * after possibly fragmenting it.
2048 static int coroutine_fn bdrv_aligned_pwritev(BdrvChild *child,
2049 BdrvTrackedRequest *req, int64_t offset, int64_t bytes,
2050 int64_t align, QEMUIOVector *qiov, size_t qiov_offset,
2051 BdrvRequestFlags flags)
2053 BlockDriverState *bs = child->bs;
2054 BlockDriver *drv = bs->drv;
2057 int64_t bytes_remaining = bytes;
2060 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort);
2066 if (bdrv_has_readonly_bitmaps(bs)) {
2070 assert(is_power_of_2(align));
2071 assert((offset & (align - 1)) == 0);
2072 assert((bytes & (align - 1)) == 0);
2073 max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX),
2076 ret = bdrv_co_write_req_prepare(child, offset, bytes, req, flags);
2078 if (!ret && bs->detect_zeroes != BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF &&
2079 !(flags & BDRV_REQ_ZERO_WRITE) && drv->bdrv_co_pwrite_zeroes &&
2080 qemu_iovec_is_zero(qiov, qiov_offset, bytes)) {
2081 flags |= BDRV_REQ_ZERO_WRITE;
2082 if (bs->detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP) {
2083 flags |= BDRV_REQ_MAY_UNMAP;
2088 /* Do nothing, write notifier decided to fail this request */
2089 } else if (flags & BDRV_REQ_ZERO_WRITE) {
2090 bdrv_debug_event(bs, BLKDBG_PWRITEV_ZERO);
2091 ret = bdrv_co_do_pwrite_zeroes(bs, offset, bytes, flags);
2092 } else if (flags & BDRV_REQ_WRITE_COMPRESSED) {
2093 ret = bdrv_driver_pwritev_compressed(bs, offset, bytes,
2095 } else if (bytes <= max_transfer) {
2096 bdrv_debug_event(bs, BLKDBG_PWRITEV);
2097 ret = bdrv_driver_pwritev(bs, offset, bytes, qiov, qiov_offset, flags);
2099 bdrv_debug_event(bs, BLKDBG_PWRITEV);
2100 while (bytes_remaining) {
2101 int num = MIN(bytes_remaining, max_transfer);
2102 int local_flags = flags;
2105 if (num < bytes_remaining && (flags & BDRV_REQ_FUA) &&
2106 !(bs->supported_write_flags & BDRV_REQ_FUA)) {
2107 /* If FUA is going to be emulated by flush, we only
2108 * need to flush on the last iteration */
2109 local_flags &= ~BDRV_REQ_FUA;
2112 ret = bdrv_driver_pwritev(bs, offset + bytes - bytes_remaining,
2114 qiov_offset + bytes - bytes_remaining,
2119 bytes_remaining -= num;
2122 bdrv_debug_event(bs, BLKDBG_PWRITEV_DONE);
2127 bdrv_co_write_req_finish(child, offset, bytes, req, ret);
2132 static int coroutine_fn bdrv_co_do_zero_pwritev(BdrvChild *child,
2135 BdrvRequestFlags flags,
2136 BdrvTrackedRequest *req)
2138 BlockDriverState *bs = child->bs;
2139 QEMUIOVector local_qiov;
2140 uint64_t align = bs->bl.request_alignment;
2143 BdrvRequestPadding pad;
2145 padding = bdrv_init_padding(bs, offset, bytes, &pad);
2147 assert(!(flags & BDRV_REQ_NO_WAIT));
2148 bdrv_make_request_serialising(req, align);
2150 bdrv_padding_rmw_read(child, req, &pad, true);
2152 if (pad.head || pad.merge_reads) {
2153 int64_t aligned_offset = offset & ~(align - 1);
2154 int64_t write_bytes = pad.merge_reads ? pad.buf_len : align;
2156 qemu_iovec_init_buf(&local_qiov, pad.buf, write_bytes);
2157 ret = bdrv_aligned_pwritev(child, req, aligned_offset, write_bytes,
2158 align, &local_qiov, 0,
2159 flags & ~BDRV_REQ_ZERO_WRITE);
2160 if (ret < 0 || pad.merge_reads) {
2161 /* Error or all work is done */
2164 offset += write_bytes - pad.head;
2165 bytes -= write_bytes - pad.head;
2169 assert(!bytes || (offset & (align - 1)) == 0);
2170 if (bytes >= align) {
2171 /* Write the aligned part in the middle. */
2172 int64_t aligned_bytes = bytes & ~(align - 1);
2173 ret = bdrv_aligned_pwritev(child, req, offset, aligned_bytes, align,
2178 bytes -= aligned_bytes;
2179 offset += aligned_bytes;
2182 assert(!bytes || (offset & (align - 1)) == 0);
2184 assert(align == pad.tail + bytes);
2186 qemu_iovec_init_buf(&local_qiov, pad.tail_buf, align);
2187 ret = bdrv_aligned_pwritev(child, req, offset, align, align,
2189 flags & ~BDRV_REQ_ZERO_WRITE);
2193 bdrv_padding_destroy(&pad);
2199 * Handle a write request in coroutine context
2201 int coroutine_fn bdrv_co_pwritev(BdrvChild *child,
2202 int64_t offset, int64_t bytes, QEMUIOVector *qiov,
2203 BdrvRequestFlags flags)
2206 return bdrv_co_pwritev_part(child, offset, bytes, qiov, 0, flags);
2209 int coroutine_fn bdrv_co_pwritev_part(BdrvChild *child,
2210 int64_t offset, int64_t bytes, QEMUIOVector *qiov, size_t qiov_offset,
2211 BdrvRequestFlags flags)
2213 BlockDriverState *bs = child->bs;
2214 BdrvTrackedRequest req;
2215 uint64_t align = bs->bl.request_alignment;
2216 BdrvRequestPadding pad;
2218 bool padded = false;
2221 trace_bdrv_co_pwritev_part(child->bs, offset, bytes, flags);
2223 if (!bdrv_is_inserted(bs)) {
2227 if (flags & BDRV_REQ_ZERO_WRITE) {
2228 ret = bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, NULL);
2230 ret = bdrv_check_request32(offset, bytes, qiov, qiov_offset);
2236 /* If the request is misaligned then we can't make it efficient */
2237 if ((flags & BDRV_REQ_NO_FALLBACK) &&
2238 !QEMU_IS_ALIGNED(offset | bytes, align))
2243 if (bytes == 0 && !QEMU_IS_ALIGNED(offset, bs->bl.request_alignment)) {
2245 * Aligning zero request is nonsense. Even if driver has special meaning
2246 * of zero-length (like qcow2_co_pwritev_compressed_part), we can't pass
2247 * it to driver due to request_alignment.
2249 * Still, no reason to return an error if someone do unaligned
2250 * zero-length write occasionally.
2255 if (!(flags & BDRV_REQ_ZERO_WRITE)) {
2257 * Pad request for following read-modify-write cycle.
2258 * bdrv_co_do_zero_pwritev() does aligning by itself, so, we do
2259 * alignment only if there is no ZERO flag.
2261 ret = bdrv_pad_request(bs, &qiov, &qiov_offset, &offset, &bytes, &pad,
2268 bdrv_inc_in_flight(bs);
2269 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_WRITE);
2271 if (flags & BDRV_REQ_ZERO_WRITE) {
2273 ret = bdrv_co_do_zero_pwritev(child, offset, bytes, flags, &req);
2279 * Request was unaligned to request_alignment and therefore
2280 * padded. We are going to do read-modify-write, and must
2281 * serialize the request to prevent interactions of the
2282 * widened region with other transactions.
2284 assert(!(flags & BDRV_REQ_NO_WAIT));
2285 bdrv_make_request_serialising(&req, align);
2286 bdrv_padding_rmw_read(child, &req, &pad, false);
2289 ret = bdrv_aligned_pwritev(child, &req, offset, bytes, align,
2290 qiov, qiov_offset, flags);
2292 bdrv_padding_destroy(&pad);
2295 tracked_request_end(&req);
2296 bdrv_dec_in_flight(bs);
2301 int coroutine_fn bdrv_co_pwrite_zeroes(BdrvChild *child, int64_t offset,
2302 int64_t bytes, BdrvRequestFlags flags)
2305 trace_bdrv_co_pwrite_zeroes(child->bs, offset, bytes, flags);
2307 if (!(child->bs->open_flags & BDRV_O_UNMAP)) {
2308 flags &= ~BDRV_REQ_MAY_UNMAP;
2311 return bdrv_co_pwritev(child, offset, bytes, NULL,
2312 BDRV_REQ_ZERO_WRITE | flags);
2316 * Flush ALL BDSes regardless of if they are reachable via a BlkBackend or not.
2318 int bdrv_flush_all(void)
2320 BdrvNextIterator it;
2321 BlockDriverState *bs = NULL;
2324 GLOBAL_STATE_CODE();
2327 * bdrv queue is managed by record/replay,
2328 * creating new flush request for stopping
2329 * the VM may break the determinism
2331 if (replay_events_enabled()) {
2335 for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
2336 AioContext *aio_context = bdrv_get_aio_context(bs);
2339 aio_context_acquire(aio_context);
2340 ret = bdrv_flush(bs);
2341 if (ret < 0 && !result) {
2344 aio_context_release(aio_context);
2351 * Returns the allocation status of the specified sectors.
2352 * Drivers not implementing the functionality are assumed to not support
2353 * backing files, hence all their sectors are reported as allocated.
2355 * If 'want_zero' is true, the caller is querying for mapping
2356 * purposes, with a focus on valid BDRV_BLOCK_OFFSET_VALID, _DATA, and
2357 * _ZERO where possible; otherwise, the result favors larger 'pnum',
2358 * with a focus on accurate BDRV_BLOCK_ALLOCATED.
2360 * If 'offset' is beyond the end of the disk image the return value is
2361 * BDRV_BLOCK_EOF and 'pnum' is set to 0.
2363 * 'bytes' is the max value 'pnum' should be set to. If bytes goes
2364 * beyond the end of the disk image it will be clamped; if 'pnum' is set to
2365 * the end of the image, then the returned value will include BDRV_BLOCK_EOF.
2367 * 'pnum' is set to the number of bytes (including and immediately
2368 * following the specified offset) that are easily known to be in the
2369 * same allocated/unallocated state. Note that a second call starting
2370 * at the original offset plus returned pnum may have the same status.
2371 * The returned value is non-zero on success except at end-of-file.
2373 * Returns negative errno on failure. Otherwise, if the
2374 * BDRV_BLOCK_OFFSET_VALID bit is set, 'map' and 'file' (if non-NULL) are
2375 * set to the host mapping and BDS corresponding to the guest offset.
2377 static int coroutine_fn bdrv_co_block_status(BlockDriverState *bs,
2379 int64_t offset, int64_t bytes,
2380 int64_t *pnum, int64_t *map,
2381 BlockDriverState **file)
2384 int64_t n; /* bytes */
2386 int64_t local_map = 0;
2387 BlockDriverState *local_file = NULL;
2388 int64_t aligned_offset, aligned_bytes;
2390 bool has_filtered_child;
2394 total_size = bdrv_getlength(bs);
2395 if (total_size < 0) {
2400 if (offset >= total_size) {
2401 ret = BDRV_BLOCK_EOF;
2409 n = total_size - offset;
2414 /* Must be non-NULL or bdrv_getlength() would have failed */
2416 has_filtered_child = bdrv_filter_child(bs);
2417 if (!bs->drv->bdrv_co_block_status && !has_filtered_child) {
2419 ret = BDRV_BLOCK_DATA | BDRV_BLOCK_ALLOCATED;
2420 if (offset + bytes == total_size) {
2421 ret |= BDRV_BLOCK_EOF;
2423 if (bs->drv->protocol_name) {
2424 ret |= BDRV_BLOCK_OFFSET_VALID;
2431 bdrv_inc_in_flight(bs);
2433 /* Round out to request_alignment boundaries */
2434 align = bs->bl.request_alignment;
2435 aligned_offset = QEMU_ALIGN_DOWN(offset, align);
2436 aligned_bytes = ROUND_UP(offset + bytes, align) - aligned_offset;
2438 if (bs->drv->bdrv_co_block_status) {
2440 * Use the block-status cache only for protocol nodes: Format
2441 * drivers are generally quick to inquire the status, but protocol
2442 * drivers often need to get information from outside of qemu, so
2443 * we do not have control over the actual implementation. There
2444 * have been cases where inquiring the status took an unreasonably
2445 * long time, and we can do nothing in qemu to fix it.
2446 * This is especially problematic for images with large data areas,
2447 * because finding the few holes in them and giving them special
2448 * treatment does not gain much performance. Therefore, we try to
2449 * cache the last-identified data region.
2451 * Second, limiting ourselves to protocol nodes allows us to assume
2452 * the block status for data regions to be DATA | OFFSET_VALID, and
2453 * that the host offset is the same as the guest offset.
2455 * Note that it is possible that external writers zero parts of
2456 * the cached regions without the cache being invalidated, and so
2457 * we may report zeroes as data. This is not catastrophic,
2458 * however, because reporting zeroes as data is fine.
2460 if (QLIST_EMPTY(&bs->children) &&
2461 bdrv_bsc_is_data(bs, aligned_offset, pnum))
2463 ret = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
2465 local_map = aligned_offset;
2467 ret = bs->drv->bdrv_co_block_status(bs, want_zero, aligned_offset,
2468 aligned_bytes, pnum, &local_map,
2472 * Note that checking QLIST_EMPTY(&bs->children) is also done when
2473 * the cache is queried above. Technically, we do not need to check
2474 * it here; the worst that can happen is that we fill the cache for
2475 * non-protocol nodes, and then it is never used. However, filling
2476 * the cache requires an RCU update, so double check here to avoid
2477 * such an update if possible.
2479 * Check want_zero, because we only want to update the cache when we
2480 * have accurate information about what is zero and what is data.
2483 ret == (BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID) &&
2484 QLIST_EMPTY(&bs->children))
2487 * When a protocol driver reports BLOCK_OFFSET_VALID, the
2488 * returned local_map value must be the same as the offset we
2489 * have passed (aligned_offset), and local_bs must be the node
2491 * Assert this, because we follow this rule when reading from
2492 * the cache (see the `local_file = bs` and
2493 * `local_map = aligned_offset` assignments above), and the
2494 * result the cache delivers must be the same as the driver
2497 assert(local_file == bs);
2498 assert(local_map == aligned_offset);
2499 bdrv_bsc_fill(bs, aligned_offset, *pnum);
2503 /* Default code for filters */
2505 local_file = bdrv_filter_bs(bs);
2508 *pnum = aligned_bytes;
2509 local_map = aligned_offset;
2510 ret = BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID;
2518 * The driver's result must be a non-zero multiple of request_alignment.
2519 * Clamp pnum and adjust map to original request.
2521 assert(*pnum && QEMU_IS_ALIGNED(*pnum, align) &&
2522 align > offset - aligned_offset);
2523 if (ret & BDRV_BLOCK_RECURSE) {
2524 assert(ret & BDRV_BLOCK_DATA);
2525 assert(ret & BDRV_BLOCK_OFFSET_VALID);
2526 assert(!(ret & BDRV_BLOCK_ZERO));
2529 *pnum -= offset - aligned_offset;
2530 if (*pnum > bytes) {
2533 if (ret & BDRV_BLOCK_OFFSET_VALID) {
2534 local_map += offset - aligned_offset;
2537 if (ret & BDRV_BLOCK_RAW) {
2538 assert(ret & BDRV_BLOCK_OFFSET_VALID && local_file);
2539 ret = bdrv_co_block_status(local_file, want_zero, local_map,
2540 *pnum, pnum, &local_map, &local_file);
2544 if (ret & (BDRV_BLOCK_DATA | BDRV_BLOCK_ZERO)) {
2545 ret |= BDRV_BLOCK_ALLOCATED;
2546 } else if (bs->drv->supports_backing) {
2547 BlockDriverState *cow_bs = bdrv_cow_bs(bs);
2550 ret |= BDRV_BLOCK_ZERO;
2551 } else if (want_zero) {
2552 int64_t size2 = bdrv_getlength(cow_bs);
2554 if (size2 >= 0 && offset >= size2) {
2555 ret |= BDRV_BLOCK_ZERO;
2560 if (want_zero && ret & BDRV_BLOCK_RECURSE &&
2561 local_file && local_file != bs &&
2562 (ret & BDRV_BLOCK_DATA) && !(ret & BDRV_BLOCK_ZERO) &&
2563 (ret & BDRV_BLOCK_OFFSET_VALID)) {
2567 ret2 = bdrv_co_block_status(local_file, want_zero, local_map,
2568 *pnum, &file_pnum, NULL, NULL);
2570 /* Ignore errors. This is just providing extra information, it
2571 * is useful but not necessary.
2573 if (ret2 & BDRV_BLOCK_EOF &&
2574 (!file_pnum || ret2 & BDRV_BLOCK_ZERO)) {
2576 * It is valid for the format block driver to read
2577 * beyond the end of the underlying file's current
2578 * size; such areas read as zero.
2580 ret |= BDRV_BLOCK_ZERO;
2582 /* Limit request to the range reported by the protocol driver */
2584 ret |= (ret2 & BDRV_BLOCK_ZERO);
2590 bdrv_dec_in_flight(bs);
2591 if (ret >= 0 && offset + *pnum == total_size) {
2592 ret |= BDRV_BLOCK_EOF;
2605 bdrv_co_common_block_status_above(BlockDriverState *bs,
2606 BlockDriverState *base,
2613 BlockDriverState **file,
2617 BlockDriverState *p;
2622 assert(!include_base || base); /* Can't include NULL base */
2629 if (!include_base && bs == base) {
2634 ret = bdrv_co_block_status(bs, want_zero, offset, bytes, pnum, map, file);
2636 if (ret < 0 || *pnum == 0 || ret & BDRV_BLOCK_ALLOCATED || bs == base) {
2640 if (ret & BDRV_BLOCK_EOF) {
2641 eof = offset + *pnum;
2644 assert(*pnum <= bytes);
2647 for (p = bdrv_filter_or_cow_bs(bs); include_base || p != base;
2648 p = bdrv_filter_or_cow_bs(p))
2650 ret = bdrv_co_block_status(p, want_zero, offset, bytes, pnum, map,
2658 * The top layer deferred to this layer, and because this layer is
2659 * short, any zeroes that we synthesize beyond EOF behave as if they
2660 * were allocated at this layer.
2662 * We don't include BDRV_BLOCK_EOF into ret, as upper layer may be
2663 * larger. We'll add BDRV_BLOCK_EOF if needed at function end, see
2666 assert(ret & BDRV_BLOCK_EOF);
2671 ret = BDRV_BLOCK_ZERO | BDRV_BLOCK_ALLOCATED;
2674 if (ret & BDRV_BLOCK_ALLOCATED) {
2676 * We've found the node and the status, we must break.
2678 * Drop BDRV_BLOCK_EOF, as it's not for upper layer, which may be
2679 * larger. We'll add BDRV_BLOCK_EOF if needed at function end, see
2682 ret &= ~BDRV_BLOCK_EOF;
2687 assert(include_base);
2692 * OK, [offset, offset + *pnum) region is unallocated on this layer,
2693 * let's continue the diving.
2695 assert(*pnum <= bytes);
2699 if (offset + *pnum == eof) {
2700 ret |= BDRV_BLOCK_EOF;
2706 int bdrv_block_status_above(BlockDriverState *bs, BlockDriverState *base,
2707 int64_t offset, int64_t bytes, int64_t *pnum,
2708 int64_t *map, BlockDriverState **file)
2711 return bdrv_common_block_status_above(bs, base, false, true, offset, bytes,
2712 pnum, map, file, NULL);
2715 int bdrv_block_status(BlockDriverState *bs, int64_t offset, int64_t bytes,
2716 int64_t *pnum, int64_t *map, BlockDriverState **file)
2719 return bdrv_block_status_above(bs, bdrv_filter_or_cow_bs(bs),
2720 offset, bytes, pnum, map, file);
2724 * Check @bs (and its backing chain) to see if the range defined
2725 * by @offset and @bytes is known to read as zeroes.
2726 * Return 1 if that is the case, 0 otherwise and -errno on error.
2727 * This test is meant to be fast rather than accurate so returning 0
2728 * does not guarantee non-zero data.
2730 int coroutine_fn bdrv_co_is_zero_fast(BlockDriverState *bs, int64_t offset,
2734 int64_t pnum = bytes;
2741 ret = bdrv_common_block_status_above(bs, NULL, false, false, offset,
2742 bytes, &pnum, NULL, NULL, NULL);
2748 return (pnum == bytes) && (ret & BDRV_BLOCK_ZERO);
2751 int coroutine_fn bdrv_is_allocated(BlockDriverState *bs, int64_t offset,
2752 int64_t bytes, int64_t *pnum)
2758 ret = bdrv_common_block_status_above(bs, bs, true, false, offset,
2759 bytes, pnum ? pnum : &dummy, NULL,
2764 return !!(ret & BDRV_BLOCK_ALLOCATED);
2768 * Given an image chain: ... -> [BASE] -> [INTER1] -> [INTER2] -> [TOP]
2770 * Return a positive depth if (a prefix of) the given range is allocated
2771 * in any image between BASE and TOP (BASE is only included if include_base
2772 * is set). Depth 1 is TOP, 2 is the first backing layer, and so forth.
2773 * BASE can be NULL to check if the given offset is allocated in any
2774 * image of the chain. Return 0 otherwise, or negative errno on
2777 * 'pnum' is set to the number of bytes (including and immediately
2778 * following the specified offset) that are known to be in the same
2779 * allocated/unallocated state. Note that a subsequent call starting
2780 * at 'offset + *pnum' may return the same allocation status (in other
2781 * words, the result is not necessarily the maximum possible range);
2782 * but 'pnum' will only be 0 when end of file is reached.
2784 int bdrv_is_allocated_above(BlockDriverState *top,
2785 BlockDriverState *base,
2786 bool include_base, int64_t offset,
2787 int64_t bytes, int64_t *pnum)
2790 int ret = bdrv_common_block_status_above(top, base, include_base, false,
2791 offset, bytes, pnum, NULL, NULL,
2798 if (ret & BDRV_BLOCK_ALLOCATED) {
2805 bdrv_co_readv_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos)
2807 BlockDriver *drv = bs->drv;
2808 BlockDriverState *child_bs = bdrv_primary_bs(bs);
2812 ret = bdrv_check_qiov_request(pos, qiov->size, qiov, 0, NULL);
2821 bdrv_inc_in_flight(bs);
2823 if (drv->bdrv_load_vmstate) {
2824 ret = drv->bdrv_load_vmstate(bs, qiov, pos);
2825 } else if (child_bs) {
2826 ret = bdrv_co_readv_vmstate(child_bs, qiov, pos);
2831 bdrv_dec_in_flight(bs);
2837 bdrv_co_writev_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos)
2839 BlockDriver *drv = bs->drv;
2840 BlockDriverState *child_bs = bdrv_primary_bs(bs);
2844 ret = bdrv_check_qiov_request(pos, qiov->size, qiov, 0, NULL);
2853 bdrv_inc_in_flight(bs);
2855 if (drv->bdrv_save_vmstate) {
2856 ret = drv->bdrv_save_vmstate(bs, qiov, pos);
2857 } else if (child_bs) {
2858 ret = bdrv_co_writev_vmstate(child_bs, qiov, pos);
2863 bdrv_dec_in_flight(bs);
2868 int bdrv_save_vmstate(BlockDriverState *bs, const uint8_t *buf,
2869 int64_t pos, int size)
2871 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, size);
2872 int ret = bdrv_writev_vmstate(bs, &qiov, pos);
2875 return ret < 0 ? ret : size;
2878 int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf,
2879 int64_t pos, int size)
2881 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, size);
2882 int ret = bdrv_readv_vmstate(bs, &qiov, pos);
2885 return ret < 0 ? ret : size;
2888 /**************************************************************/
2891 void bdrv_aio_cancel(BlockAIOCB *acb)
2895 bdrv_aio_cancel_async(acb);
2896 while (acb->refcnt > 1) {
2897 if (acb->aiocb_info->get_aio_context) {
2898 aio_poll(acb->aiocb_info->get_aio_context(acb), true);
2899 } else if (acb->bs) {
2900 /* qemu_aio_ref and qemu_aio_unref are not thread-safe, so
2901 * assert that we're not using an I/O thread. Thread-safe
2902 * code should use bdrv_aio_cancel_async exclusively.
2904 assert(bdrv_get_aio_context(acb->bs) == qemu_get_aio_context());
2905 aio_poll(bdrv_get_aio_context(acb->bs), true);
2910 qemu_aio_unref(acb);
2913 /* Async version of aio cancel. The caller is not blocked if the acb implements
2914 * cancel_async, otherwise we do nothing and let the request normally complete.
2915 * In either case the completion callback must be called. */
2916 void bdrv_aio_cancel_async(BlockAIOCB *acb)
2919 if (acb->aiocb_info->cancel_async) {
2920 acb->aiocb_info->cancel_async(acb);
2924 /**************************************************************/
2925 /* Coroutine block device emulation */
2927 int coroutine_fn bdrv_co_flush(BlockDriverState *bs)
2929 BdrvChild *primary_child = bdrv_primary_child(bs);
2935 bdrv_inc_in_flight(bs);
2937 if (!bdrv_is_inserted(bs) || bdrv_is_read_only(bs) ||
2942 qemu_co_mutex_lock(&bs->reqs_lock);
2943 current_gen = qatomic_read(&bs->write_gen);
2945 /* Wait until any previous flushes are completed */
2946 while (bs->active_flush_req) {
2947 qemu_co_queue_wait(&bs->flush_queue, &bs->reqs_lock);
2950 /* Flushes reach this point in nondecreasing current_gen order. */
2951 bs->active_flush_req = true;
2952 qemu_co_mutex_unlock(&bs->reqs_lock);
2954 /* Write back all layers by calling one driver function */
2955 if (bs->drv->bdrv_co_flush) {
2956 ret = bs->drv->bdrv_co_flush(bs);
2960 /* Write back cached data to the OS even with cache=unsafe */
2961 BLKDBG_EVENT(primary_child, BLKDBG_FLUSH_TO_OS);
2962 if (bs->drv->bdrv_co_flush_to_os) {
2963 ret = bs->drv->bdrv_co_flush_to_os(bs);
2969 /* But don't actually force it to the disk with cache=unsafe */
2970 if (bs->open_flags & BDRV_O_NO_FLUSH) {
2971 goto flush_children;
2974 /* Check if we really need to flush anything */
2975 if (bs->flushed_gen == current_gen) {
2976 goto flush_children;
2979 BLKDBG_EVENT(primary_child, BLKDBG_FLUSH_TO_DISK);
2981 /* bs->drv->bdrv_co_flush() might have ejected the BDS
2982 * (even in case of apparent success) */
2986 if (bs->drv->bdrv_co_flush_to_disk) {
2987 ret = bs->drv->bdrv_co_flush_to_disk(bs);
2988 } else if (bs->drv->bdrv_aio_flush) {
2990 CoroutineIOCompletion co = {
2991 .coroutine = qemu_coroutine_self(),
2994 acb = bs->drv->bdrv_aio_flush(bs, bdrv_co_io_em_complete, &co);
2998 qemu_coroutine_yield();
3003 * Some block drivers always operate in either writethrough or unsafe
3004 * mode and don't support bdrv_flush therefore. Usually qemu doesn't
3005 * know how the server works (because the behaviour is hardcoded or
3006 * depends on server-side configuration), so we can't ensure that
3007 * everything is safe on disk. Returning an error doesn't work because
3008 * that would break guests even if the server operates in writethrough
3011 * Let's hope the user knows what he's doing.
3020 /* Now flush the underlying protocol. It will also have BDRV_O_NO_FLUSH
3021 * in the case of cache=unsafe, so there are no useless flushes.
3025 QLIST_FOREACH(child, &bs->children, next) {
3026 if (child->perm & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED)) {
3027 int this_child_ret = bdrv_co_flush(child->bs);
3029 ret = this_child_ret;
3035 /* Notify any pending flushes that we have completed */
3037 bs->flushed_gen = current_gen;
3040 qemu_co_mutex_lock(&bs->reqs_lock);
3041 bs->active_flush_req = false;
3042 /* Return value is ignored - it's ok if wait queue is empty */
3043 qemu_co_queue_next(&bs->flush_queue);
3044 qemu_co_mutex_unlock(&bs->reqs_lock);
3047 bdrv_dec_in_flight(bs);
3051 int coroutine_fn bdrv_co_pdiscard(BdrvChild *child, int64_t offset,
3054 BdrvTrackedRequest req;
3056 int64_t max_pdiscard;
3057 int head, tail, align;
3058 BlockDriverState *bs = child->bs;
3061 if (!bs || !bs->drv || !bdrv_is_inserted(bs)) {
3065 if (bdrv_has_readonly_bitmaps(bs)) {
3069 ret = bdrv_check_request(offset, bytes, NULL);
3074 /* Do nothing if disabled. */
3075 if (!(bs->open_flags & BDRV_O_UNMAP)) {
3079 if (!bs->drv->bdrv_co_pdiscard && !bs->drv->bdrv_aio_pdiscard) {
3083 /* Invalidate the cached block-status data range if this discard overlaps */
3084 bdrv_bsc_invalidate_range(bs, offset, bytes);
3086 /* Discard is advisory, but some devices track and coalesce
3087 * unaligned requests, so we must pass everything down rather than
3088 * round here. Still, most devices will just silently ignore
3089 * unaligned requests (by returning -ENOTSUP), so we must fragment
3090 * the request accordingly. */
3091 align = MAX(bs->bl.pdiscard_alignment, bs->bl.request_alignment);
3092 assert(align % bs->bl.request_alignment == 0);
3093 head = offset % align;
3094 tail = (offset + bytes) % align;
3096 bdrv_inc_in_flight(bs);
3097 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_DISCARD);
3099 ret = bdrv_co_write_req_prepare(child, offset, bytes, &req, 0);
3104 max_pdiscard = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_pdiscard, INT64_MAX),
3106 assert(max_pdiscard >= bs->bl.request_alignment);
3109 int64_t num = bytes;
3112 /* Make small requests to get to alignment boundaries. */
3113 num = MIN(bytes, align - head);
3114 if (!QEMU_IS_ALIGNED(num, bs->bl.request_alignment)) {
3115 num %= bs->bl.request_alignment;
3117 head = (head + num) % align;
3118 assert(num < max_pdiscard);
3121 /* Shorten the request to the last aligned cluster. */
3123 } else if (!QEMU_IS_ALIGNED(tail, bs->bl.request_alignment) &&
3124 tail > bs->bl.request_alignment) {
3125 tail %= bs->bl.request_alignment;
3129 /* limit request size */
3130 if (num > max_pdiscard) {
3138 if (bs->drv->bdrv_co_pdiscard) {
3139 ret = bs->drv->bdrv_co_pdiscard(bs, offset, num);
3142 CoroutineIOCompletion co = {
3143 .coroutine = qemu_coroutine_self(),
3146 acb = bs->drv->bdrv_aio_pdiscard(bs, offset, num,
3147 bdrv_co_io_em_complete, &co);
3152 qemu_coroutine_yield();
3156 if (ret && ret != -ENOTSUP) {
3165 bdrv_co_write_req_finish(child, req.offset, req.bytes, &req, ret);
3166 tracked_request_end(&req);
3167 bdrv_dec_in_flight(bs);
3171 int bdrv_co_ioctl(BlockDriverState *bs, int req, void *buf)
3173 BlockDriver *drv = bs->drv;
3174 CoroutineIOCompletion co = {
3175 .coroutine = qemu_coroutine_self(),
3180 bdrv_inc_in_flight(bs);
3181 if (!drv || (!drv->bdrv_aio_ioctl && !drv->bdrv_co_ioctl)) {
3186 if (drv->bdrv_co_ioctl) {
3187 co.ret = drv->bdrv_co_ioctl(bs, req, buf);
3189 acb = drv->bdrv_aio_ioctl(bs, req, buf, bdrv_co_io_em_complete, &co);
3194 qemu_coroutine_yield();
3197 bdrv_dec_in_flight(bs);
3201 void *qemu_blockalign(BlockDriverState *bs, size_t size)
3204 return qemu_memalign(bdrv_opt_mem_align(bs), size);
3207 void *qemu_blockalign0(BlockDriverState *bs, size_t size)
3210 return memset(qemu_blockalign(bs, size), 0, size);
3213 void *qemu_try_blockalign(BlockDriverState *bs, size_t size)
3215 size_t align = bdrv_opt_mem_align(bs);
3218 /* Ensure that NULL is never returned on success */
3224 return qemu_try_memalign(align, size);
3227 void *qemu_try_blockalign0(BlockDriverState *bs, size_t size)
3229 void *mem = qemu_try_blockalign(bs, size);
3233 memset(mem, 0, size);
3240 * Check if all memory in this vector is sector aligned.
3242 bool bdrv_qiov_is_aligned(BlockDriverState *bs, QEMUIOVector *qiov)
3245 size_t alignment = bdrv_min_mem_align(bs);
3248 for (i = 0; i < qiov->niov; i++) {
3249 if ((uintptr_t) qiov->iov[i].iov_base % alignment) {
3252 if (qiov->iov[i].iov_len % alignment) {
3260 void bdrv_io_plug(BlockDriverState *bs)
3265 QLIST_FOREACH(child, &bs->children, next) {
3266 bdrv_io_plug(child->bs);
3269 if (qatomic_fetch_inc(&bs->io_plugged) == 0) {
3270 BlockDriver *drv = bs->drv;
3271 if (drv && drv->bdrv_io_plug) {
3272 drv->bdrv_io_plug(bs);
3277 void bdrv_io_unplug(BlockDriverState *bs)
3282 assert(bs->io_plugged);
3283 if (qatomic_fetch_dec(&bs->io_plugged) == 1) {
3284 BlockDriver *drv = bs->drv;
3285 if (drv && drv->bdrv_io_unplug) {
3286 drv->bdrv_io_unplug(bs);
3290 QLIST_FOREACH(child, &bs->children, next) {
3291 bdrv_io_unplug(child->bs);
3295 void bdrv_register_buf(BlockDriverState *bs, void *host, size_t size)
3299 GLOBAL_STATE_CODE();
3300 if (bs->drv && bs->drv->bdrv_register_buf) {
3301 bs->drv->bdrv_register_buf(bs, host, size);
3303 QLIST_FOREACH(child, &bs->children, next) {
3304 bdrv_register_buf(child->bs, host, size);
3308 void bdrv_unregister_buf(BlockDriverState *bs, void *host)
3312 GLOBAL_STATE_CODE();
3313 if (bs->drv && bs->drv->bdrv_unregister_buf) {
3314 bs->drv->bdrv_unregister_buf(bs, host);
3316 QLIST_FOREACH(child, &bs->children, next) {
3317 bdrv_unregister_buf(child->bs, host);
3321 static int coroutine_fn bdrv_co_copy_range_internal(
3322 BdrvChild *src, int64_t src_offset, BdrvChild *dst,
3323 int64_t dst_offset, int64_t bytes,
3324 BdrvRequestFlags read_flags, BdrvRequestFlags write_flags,
3327 BdrvTrackedRequest req;
3330 /* TODO We can support BDRV_REQ_NO_FALLBACK here */
3331 assert(!(read_flags & BDRV_REQ_NO_FALLBACK));
3332 assert(!(write_flags & BDRV_REQ_NO_FALLBACK));
3333 assert(!(read_flags & BDRV_REQ_NO_WAIT));
3334 assert(!(write_flags & BDRV_REQ_NO_WAIT));
3336 if (!dst || !dst->bs || !bdrv_is_inserted(dst->bs)) {
3339 ret = bdrv_check_request32(dst_offset, bytes, NULL, 0);
3343 if (write_flags & BDRV_REQ_ZERO_WRITE) {
3344 return bdrv_co_pwrite_zeroes(dst, dst_offset, bytes, write_flags);
3347 if (!src || !src->bs || !bdrv_is_inserted(src->bs)) {
3350 ret = bdrv_check_request32(src_offset, bytes, NULL, 0);
3355 if (!src->bs->drv->bdrv_co_copy_range_from
3356 || !dst->bs->drv->bdrv_co_copy_range_to
3357 || src->bs->encrypted || dst->bs->encrypted) {
3362 bdrv_inc_in_flight(src->bs);
3363 tracked_request_begin(&req, src->bs, src_offset, bytes,
3366 /* BDRV_REQ_SERIALISING is only for write operation */
3367 assert(!(read_flags & BDRV_REQ_SERIALISING));
3368 bdrv_wait_serialising_requests(&req);
3370 ret = src->bs->drv->bdrv_co_copy_range_from(src->bs,
3374 read_flags, write_flags);
3376 tracked_request_end(&req);
3377 bdrv_dec_in_flight(src->bs);
3379 bdrv_inc_in_flight(dst->bs);
3380 tracked_request_begin(&req, dst->bs, dst_offset, bytes,
3381 BDRV_TRACKED_WRITE);
3382 ret = bdrv_co_write_req_prepare(dst, dst_offset, bytes, &req,
3385 ret = dst->bs->drv->bdrv_co_copy_range_to(dst->bs,
3389 read_flags, write_flags);
3391 bdrv_co_write_req_finish(dst, dst_offset, bytes, &req, ret);
3392 tracked_request_end(&req);
3393 bdrv_dec_in_flight(dst->bs);
3399 /* Copy range from @src to @dst.
3401 * See the comment of bdrv_co_copy_range for the parameter and return value
3403 int coroutine_fn bdrv_co_copy_range_from(BdrvChild *src, int64_t src_offset,
3404 BdrvChild *dst, int64_t dst_offset,
3406 BdrvRequestFlags read_flags,
3407 BdrvRequestFlags write_flags)
3410 trace_bdrv_co_copy_range_from(src, src_offset, dst, dst_offset, bytes,
3411 read_flags, write_flags);
3412 return bdrv_co_copy_range_internal(src, src_offset, dst, dst_offset,
3413 bytes, read_flags, write_flags, true);
3416 /* Copy range from @src to @dst.
3418 * See the comment of bdrv_co_copy_range for the parameter and return value
3420 int coroutine_fn bdrv_co_copy_range_to(BdrvChild *src, int64_t src_offset,
3421 BdrvChild *dst, int64_t dst_offset,
3423 BdrvRequestFlags read_flags,
3424 BdrvRequestFlags write_flags)
3427 trace_bdrv_co_copy_range_to(src, src_offset, dst, dst_offset, bytes,
3428 read_flags, write_flags);
3429 return bdrv_co_copy_range_internal(src, src_offset, dst, dst_offset,
3430 bytes, read_flags, write_flags, false);
3433 int coroutine_fn bdrv_co_copy_range(BdrvChild *src, int64_t src_offset,
3434 BdrvChild *dst, int64_t dst_offset,
3435 int64_t bytes, BdrvRequestFlags read_flags,
3436 BdrvRequestFlags write_flags)
3439 return bdrv_co_copy_range_from(src, src_offset,
3441 bytes, read_flags, write_flags);
3444 static void bdrv_parent_cb_resize(BlockDriverState *bs)
3447 QLIST_FOREACH(c, &bs->parents, next_parent) {
3448 if (c->klass->resize) {
3449 c->klass->resize(c);
3455 * Truncate file to 'offset' bytes (needed only for file protocols)
3457 * If 'exact' is true, the file must be resized to exactly the given
3458 * 'offset'. Otherwise, it is sufficient for the node to be at least
3459 * 'offset' bytes in length.
3461 int coroutine_fn bdrv_co_truncate(BdrvChild *child, int64_t offset, bool exact,
3462 PreallocMode prealloc, BdrvRequestFlags flags,
3465 BlockDriverState *bs = child->bs;
3466 BdrvChild *filtered, *backing;
3467 BlockDriver *drv = bs->drv;
3468 BdrvTrackedRequest req;
3469 int64_t old_size, new_bytes;
3473 /* if bs->drv == NULL, bs is closed, so there's nothing to do here */
3475 error_setg(errp, "No medium inserted");
3479 error_setg(errp, "Image size cannot be negative");
3483 ret = bdrv_check_request(offset, 0, errp);
3488 old_size = bdrv_getlength(bs);
3490 error_setg_errno(errp, -old_size, "Failed to get old image size");
3494 if (bdrv_is_read_only(bs)) {
3495 error_setg(errp, "Image is read-only");
3499 if (offset > old_size) {
3500 new_bytes = offset - old_size;
3505 bdrv_inc_in_flight(bs);
3506 tracked_request_begin(&req, bs, offset - new_bytes, new_bytes,
3507 BDRV_TRACKED_TRUNCATE);
3509 /* If we are growing the image and potentially using preallocation for the
3510 * new area, we need to make sure that no write requests are made to it
3511 * concurrently or they might be overwritten by preallocation. */
3513 bdrv_make_request_serialising(&req, 1);
3515 ret = bdrv_co_write_req_prepare(child, offset - new_bytes, new_bytes, &req,
3518 error_setg_errno(errp, -ret,
3519 "Failed to prepare request for truncation");
3523 filtered = bdrv_filter_child(bs);
3524 backing = bdrv_cow_child(bs);
3527 * If the image has a backing file that is large enough that it would
3528 * provide data for the new area, we cannot leave it unallocated because
3529 * then the backing file content would become visible. Instead, zero-fill
3532 * Note that if the image has a backing file, but was opened without the
3533 * backing file, taking care of keeping things consistent with that backing
3534 * file is the user's responsibility.
3536 if (new_bytes && backing) {
3537 int64_t backing_len;
3539 backing_len = bdrv_getlength(backing->bs);
3540 if (backing_len < 0) {
3542 error_setg_errno(errp, -ret, "Could not get backing file size");
3546 if (backing_len > old_size) {
3547 flags |= BDRV_REQ_ZERO_WRITE;
3551 if (drv->bdrv_co_truncate) {
3552 if (flags & ~bs->supported_truncate_flags) {
3553 error_setg(errp, "Block driver does not support requested flags");
3557 ret = drv->bdrv_co_truncate(bs, offset, exact, prealloc, flags, errp);
3558 } else if (filtered) {
3559 ret = bdrv_co_truncate(filtered, offset, exact, prealloc, flags, errp);
3561 error_setg(errp, "Image format driver does not support resize");
3569 ret = refresh_total_sectors(bs, offset >> BDRV_SECTOR_BITS);
3571 error_setg_errno(errp, -ret, "Could not refresh total sector count");
3573 offset = bs->total_sectors * BDRV_SECTOR_SIZE;
3575 /* It's possible that truncation succeeded but refresh_total_sectors
3576 * failed, but the latter doesn't affect how we should finish the request.
3577 * Pass 0 as the last parameter so that dirty bitmaps etc. are handled. */
3578 bdrv_co_write_req_finish(child, offset - new_bytes, new_bytes, &req, 0);
3581 tracked_request_end(&req);
3582 bdrv_dec_in_flight(bs);
3587 void bdrv_cancel_in_flight(BlockDriverState *bs)
3589 GLOBAL_STATE_CODE();
3590 if (!bs || !bs->drv) {
3594 if (bs->drv->bdrv_cancel_in_flight) {
3595 bs->drv->bdrv_cancel_in_flight(bs);
3600 bdrv_co_preadv_snapshot(BdrvChild *child, int64_t offset, int64_t bytes,
3601 QEMUIOVector *qiov, size_t qiov_offset)
3603 BlockDriverState *bs = child->bs;
3604 BlockDriver *drv = bs->drv;
3612 if (!drv->bdrv_co_preadv_snapshot) {
3616 bdrv_inc_in_flight(bs);
3617 ret = drv->bdrv_co_preadv_snapshot(bs, offset, bytes, qiov, qiov_offset);
3618 bdrv_dec_in_flight(bs);
3624 bdrv_co_snapshot_block_status(BlockDriverState *bs,
3625 bool want_zero, int64_t offset, int64_t bytes,
3626 int64_t *pnum, int64_t *map,
3627 BlockDriverState **file)
3629 BlockDriver *drv = bs->drv;
3637 if (!drv->bdrv_co_snapshot_block_status) {
3641 bdrv_inc_in_flight(bs);
3642 ret = drv->bdrv_co_snapshot_block_status(bs, want_zero, offset, bytes,
3644 bdrv_dec_in_flight(bs);
3650 bdrv_co_pdiscard_snapshot(BlockDriverState *bs, int64_t offset, int64_t bytes)
3652 BlockDriver *drv = bs->drv;
3660 if (!drv->bdrv_co_pdiscard_snapshot) {
3664 bdrv_inc_in_flight(bs);
3665 ret = drv->bdrv_co_pdiscard_snapshot(bs, offset, bytes);
3666 bdrv_dec_in_flight(bs);