s390x/tcg: Fault-safe memset
[qemu/ar7.git] / block / io.c
blobf8c359613160052840b535ff7bc3d514c278369e
1 /*
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
22 * THE SOFTWARE.
25 #include "qemu/osdep.h"
26 #include "trace.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 "qemu/cutils.h"
33 #include "qapi/error.h"
34 #include "qemu/error-report.h"
35 #include "qemu/main-loop.h"
37 #define NOT_DONE 0x7fffffff /* used while emulated sync operation in progress */
39 /* Maximum bounce buffer for copy-on-read and write zeroes, in bytes */
40 #define MAX_BOUNCE_BUFFER (32768 << BDRV_SECTOR_BITS)
42 static void bdrv_parent_cb_resize(BlockDriverState *bs);
43 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs,
44 int64_t offset, int bytes, BdrvRequestFlags flags);
46 static void bdrv_parent_drained_begin(BlockDriverState *bs, BdrvChild *ignore,
47 bool ignore_bds_parents)
49 BdrvChild *c, *next;
51 QLIST_FOREACH_SAFE(c, &bs->parents, next_parent, next) {
52 if (c == ignore || (ignore_bds_parents && c->role->parent_is_bds)) {
53 continue;
55 bdrv_parent_drained_begin_single(c, false);
59 static void bdrv_parent_drained_end_single_no_poll(BdrvChild *c,
60 int *drained_end_counter)
62 assert(c->parent_quiesce_counter > 0);
63 c->parent_quiesce_counter--;
64 if (c->role->drained_end) {
65 c->role->drained_end(c, drained_end_counter);
69 void bdrv_parent_drained_end_single(BdrvChild *c)
71 int drained_end_counter = 0;
72 bdrv_parent_drained_end_single_no_poll(c, &drained_end_counter);
73 BDRV_POLL_WHILE(c->bs, atomic_read(&drained_end_counter) > 0);
76 static void bdrv_parent_drained_end(BlockDriverState *bs, BdrvChild *ignore,
77 bool ignore_bds_parents,
78 int *drained_end_counter)
80 BdrvChild *c;
82 QLIST_FOREACH(c, &bs->parents, next_parent) {
83 if (c == ignore || (ignore_bds_parents && c->role->parent_is_bds)) {
84 continue;
86 bdrv_parent_drained_end_single_no_poll(c, drained_end_counter);
90 static bool bdrv_parent_drained_poll_single(BdrvChild *c)
92 if (c->role->drained_poll) {
93 return c->role->drained_poll(c);
95 return false;
98 static bool bdrv_parent_drained_poll(BlockDriverState *bs, BdrvChild *ignore,
99 bool ignore_bds_parents)
101 BdrvChild *c, *next;
102 bool busy = false;
104 QLIST_FOREACH_SAFE(c, &bs->parents, next_parent, next) {
105 if (c == ignore || (ignore_bds_parents && c->role->parent_is_bds)) {
106 continue;
108 busy |= bdrv_parent_drained_poll_single(c);
111 return busy;
114 void bdrv_parent_drained_begin_single(BdrvChild *c, bool poll)
116 c->parent_quiesce_counter++;
117 if (c->role->drained_begin) {
118 c->role->drained_begin(c);
120 if (poll) {
121 BDRV_POLL_WHILE(c->bs, bdrv_parent_drained_poll_single(c));
125 static void bdrv_merge_limits(BlockLimits *dst, const BlockLimits *src)
127 dst->opt_transfer = MAX(dst->opt_transfer, src->opt_transfer);
128 dst->max_transfer = MIN_NON_ZERO(dst->max_transfer, src->max_transfer);
129 dst->opt_mem_alignment = MAX(dst->opt_mem_alignment,
130 src->opt_mem_alignment);
131 dst->min_mem_alignment = MAX(dst->min_mem_alignment,
132 src->min_mem_alignment);
133 dst->max_iov = MIN_NON_ZERO(dst->max_iov, src->max_iov);
136 void bdrv_refresh_limits(BlockDriverState *bs, Error **errp)
138 BlockDriver *drv = bs->drv;
139 Error *local_err = NULL;
141 memset(&bs->bl, 0, sizeof(bs->bl));
143 if (!drv) {
144 return;
147 /* Default alignment based on whether driver has byte interface */
148 bs->bl.request_alignment = (drv->bdrv_co_preadv ||
149 drv->bdrv_aio_preadv ||
150 drv->bdrv_co_preadv_part) ? 1 : 512;
152 /* Take some limits from the children as a default */
153 if (bs->file) {
154 bdrv_refresh_limits(bs->file->bs, &local_err);
155 if (local_err) {
156 error_propagate(errp, local_err);
157 return;
159 bdrv_merge_limits(&bs->bl, &bs->file->bs->bl);
160 } else {
161 bs->bl.min_mem_alignment = 512;
162 bs->bl.opt_mem_alignment = getpagesize();
164 /* Safe default since most protocols use readv()/writev()/etc */
165 bs->bl.max_iov = IOV_MAX;
168 if (bs->backing) {
169 bdrv_refresh_limits(bs->backing->bs, &local_err);
170 if (local_err) {
171 error_propagate(errp, local_err);
172 return;
174 bdrv_merge_limits(&bs->bl, &bs->backing->bs->bl);
177 /* Then let the driver override it */
178 if (drv->bdrv_refresh_limits) {
179 drv->bdrv_refresh_limits(bs, errp);
184 * The copy-on-read flag is actually a reference count so multiple users may
185 * use the feature without worrying about clobbering its previous state.
186 * Copy-on-read stays enabled until all users have called to disable it.
188 void bdrv_enable_copy_on_read(BlockDriverState *bs)
190 atomic_inc(&bs->copy_on_read);
193 void bdrv_disable_copy_on_read(BlockDriverState *bs)
195 int old = atomic_fetch_dec(&bs->copy_on_read);
196 assert(old >= 1);
199 typedef struct {
200 Coroutine *co;
201 BlockDriverState *bs;
202 bool done;
203 bool begin;
204 bool recursive;
205 bool poll;
206 BdrvChild *parent;
207 bool ignore_bds_parents;
208 int *drained_end_counter;
209 } BdrvCoDrainData;
211 static void coroutine_fn bdrv_drain_invoke_entry(void *opaque)
213 BdrvCoDrainData *data = opaque;
214 BlockDriverState *bs = data->bs;
216 if (data->begin) {
217 bs->drv->bdrv_co_drain_begin(bs);
218 } else {
219 bs->drv->bdrv_co_drain_end(bs);
222 /* Set data->done and decrement drained_end_counter before bdrv_wakeup() */
223 atomic_mb_set(&data->done, true);
224 if (!data->begin) {
225 atomic_dec(data->drained_end_counter);
227 bdrv_dec_in_flight(bs);
229 g_free(data);
232 /* Recursively call BlockDriver.bdrv_co_drain_begin/end callbacks */
233 static void bdrv_drain_invoke(BlockDriverState *bs, bool begin,
234 int *drained_end_counter)
236 BdrvCoDrainData *data;
238 if (!bs->drv || (begin && !bs->drv->bdrv_co_drain_begin) ||
239 (!begin && !bs->drv->bdrv_co_drain_end)) {
240 return;
243 data = g_new(BdrvCoDrainData, 1);
244 *data = (BdrvCoDrainData) {
245 .bs = bs,
246 .done = false,
247 .begin = begin,
248 .drained_end_counter = drained_end_counter,
251 if (!begin) {
252 atomic_inc(drained_end_counter);
255 /* Make sure the driver callback completes during the polling phase for
256 * drain_begin. */
257 bdrv_inc_in_flight(bs);
258 data->co = qemu_coroutine_create(bdrv_drain_invoke_entry, data);
259 aio_co_schedule(bdrv_get_aio_context(bs), data->co);
262 /* Returns true if BDRV_POLL_WHILE() should go into a blocking aio_poll() */
263 bool bdrv_drain_poll(BlockDriverState *bs, bool recursive,
264 BdrvChild *ignore_parent, bool ignore_bds_parents)
266 BdrvChild *child, *next;
268 if (bdrv_parent_drained_poll(bs, ignore_parent, ignore_bds_parents)) {
269 return true;
272 if (atomic_read(&bs->in_flight)) {
273 return true;
276 if (recursive) {
277 assert(!ignore_bds_parents);
278 QLIST_FOREACH_SAFE(child, &bs->children, next, next) {
279 if (bdrv_drain_poll(child->bs, recursive, child, false)) {
280 return true;
285 return false;
288 static bool bdrv_drain_poll_top_level(BlockDriverState *bs, bool recursive,
289 BdrvChild *ignore_parent)
291 return bdrv_drain_poll(bs, recursive, ignore_parent, false);
294 static void bdrv_do_drained_begin(BlockDriverState *bs, bool recursive,
295 BdrvChild *parent, bool ignore_bds_parents,
296 bool poll);
297 static void bdrv_do_drained_end(BlockDriverState *bs, bool recursive,
298 BdrvChild *parent, bool ignore_bds_parents,
299 int *drained_end_counter);
301 static void bdrv_co_drain_bh_cb(void *opaque)
303 BdrvCoDrainData *data = opaque;
304 Coroutine *co = data->co;
305 BlockDriverState *bs = data->bs;
307 if (bs) {
308 AioContext *ctx = bdrv_get_aio_context(bs);
309 AioContext *co_ctx = qemu_coroutine_get_aio_context(co);
312 * When the coroutine yielded, the lock for its home context was
313 * released, so we need to re-acquire it here. If it explicitly
314 * acquired a different context, the lock is still held and we don't
315 * want to lock it a second time (or AIO_WAIT_WHILE() would hang).
317 if (ctx == co_ctx) {
318 aio_context_acquire(ctx);
320 bdrv_dec_in_flight(bs);
321 if (data->begin) {
322 assert(!data->drained_end_counter);
323 bdrv_do_drained_begin(bs, data->recursive, data->parent,
324 data->ignore_bds_parents, data->poll);
325 } else {
326 assert(!data->poll);
327 bdrv_do_drained_end(bs, data->recursive, data->parent,
328 data->ignore_bds_parents,
329 data->drained_end_counter);
331 if (ctx == co_ctx) {
332 aio_context_release(ctx);
334 } else {
335 assert(data->begin);
336 bdrv_drain_all_begin();
339 data->done = true;
340 aio_co_wake(co);
343 static void coroutine_fn bdrv_co_yield_to_drain(BlockDriverState *bs,
344 bool begin, bool recursive,
345 BdrvChild *parent,
346 bool ignore_bds_parents,
347 bool poll,
348 int *drained_end_counter)
350 BdrvCoDrainData data;
352 /* Calling bdrv_drain() from a BH ensures the current coroutine yields and
353 * other coroutines run if they were queued by aio_co_enter(). */
355 assert(qemu_in_coroutine());
356 data = (BdrvCoDrainData) {
357 .co = qemu_coroutine_self(),
358 .bs = bs,
359 .done = false,
360 .begin = begin,
361 .recursive = recursive,
362 .parent = parent,
363 .ignore_bds_parents = ignore_bds_parents,
364 .poll = poll,
365 .drained_end_counter = drained_end_counter,
368 if (bs) {
369 bdrv_inc_in_flight(bs);
371 aio_bh_schedule_oneshot(bdrv_get_aio_context(bs),
372 bdrv_co_drain_bh_cb, &data);
374 qemu_coroutine_yield();
375 /* If we are resumed from some other event (such as an aio completion or a
376 * timer callback), it is a bug in the caller that should be fixed. */
377 assert(data.done);
380 void bdrv_do_drained_begin_quiesce(BlockDriverState *bs,
381 BdrvChild *parent, bool ignore_bds_parents)
383 assert(!qemu_in_coroutine());
385 /* Stop things in parent-to-child order */
386 if (atomic_fetch_inc(&bs->quiesce_counter) == 0) {
387 aio_disable_external(bdrv_get_aio_context(bs));
390 bdrv_parent_drained_begin(bs, parent, ignore_bds_parents);
391 bdrv_drain_invoke(bs, true, NULL);
394 static void bdrv_do_drained_begin(BlockDriverState *bs, bool recursive,
395 BdrvChild *parent, bool ignore_bds_parents,
396 bool poll)
398 BdrvChild *child, *next;
400 if (qemu_in_coroutine()) {
401 bdrv_co_yield_to_drain(bs, true, recursive, parent, ignore_bds_parents,
402 poll, NULL);
403 return;
406 bdrv_do_drained_begin_quiesce(bs, parent, ignore_bds_parents);
408 if (recursive) {
409 assert(!ignore_bds_parents);
410 bs->recursive_quiesce_counter++;
411 QLIST_FOREACH_SAFE(child, &bs->children, next, next) {
412 bdrv_do_drained_begin(child->bs, true, child, ignore_bds_parents,
413 false);
418 * Wait for drained requests to finish.
420 * Calling BDRV_POLL_WHILE() only once for the top-level node is okay: The
421 * call is needed so things in this AioContext can make progress even
422 * though we don't return to the main AioContext loop - this automatically
423 * includes other nodes in the same AioContext and therefore all child
424 * nodes.
426 if (poll) {
427 assert(!ignore_bds_parents);
428 BDRV_POLL_WHILE(bs, bdrv_drain_poll_top_level(bs, recursive, parent));
432 void bdrv_drained_begin(BlockDriverState *bs)
434 bdrv_do_drained_begin(bs, false, NULL, false, true);
437 void bdrv_subtree_drained_begin(BlockDriverState *bs)
439 bdrv_do_drained_begin(bs, true, NULL, false, true);
443 * This function does not poll, nor must any of its recursively called
444 * functions. The *drained_end_counter pointee will be incremented
445 * once for every background operation scheduled, and decremented once
446 * the operation settles. Therefore, the pointer must remain valid
447 * until the pointee reaches 0. That implies that whoever sets up the
448 * pointee has to poll until it is 0.
450 * We use atomic operations to access *drained_end_counter, because
451 * (1) when called from bdrv_set_aio_context_ignore(), the subgraph of
452 * @bs may contain nodes in different AioContexts,
453 * (2) bdrv_drain_all_end() uses the same counter for all nodes,
454 * regardless of which AioContext they are in.
456 static void bdrv_do_drained_end(BlockDriverState *bs, bool recursive,
457 BdrvChild *parent, bool ignore_bds_parents,
458 int *drained_end_counter)
460 BdrvChild *child;
461 int old_quiesce_counter;
463 assert(drained_end_counter != NULL);
465 if (qemu_in_coroutine()) {
466 bdrv_co_yield_to_drain(bs, false, recursive, parent, ignore_bds_parents,
467 false, drained_end_counter);
468 return;
470 assert(bs->quiesce_counter > 0);
472 /* Re-enable things in child-to-parent order */
473 bdrv_drain_invoke(bs, false, drained_end_counter);
474 bdrv_parent_drained_end(bs, parent, ignore_bds_parents,
475 drained_end_counter);
477 old_quiesce_counter = atomic_fetch_dec(&bs->quiesce_counter);
478 if (old_quiesce_counter == 1) {
479 aio_enable_external(bdrv_get_aio_context(bs));
482 if (recursive) {
483 assert(!ignore_bds_parents);
484 bs->recursive_quiesce_counter--;
485 QLIST_FOREACH(child, &bs->children, next) {
486 bdrv_do_drained_end(child->bs, true, child, ignore_bds_parents,
487 drained_end_counter);
492 void bdrv_drained_end(BlockDriverState *bs)
494 int drained_end_counter = 0;
495 bdrv_do_drained_end(bs, false, NULL, false, &drained_end_counter);
496 BDRV_POLL_WHILE(bs, atomic_read(&drained_end_counter) > 0);
499 void bdrv_drained_end_no_poll(BlockDriverState *bs, int *drained_end_counter)
501 bdrv_do_drained_end(bs, false, NULL, false, drained_end_counter);
504 void bdrv_subtree_drained_end(BlockDriverState *bs)
506 int drained_end_counter = 0;
507 bdrv_do_drained_end(bs, true, NULL, false, &drained_end_counter);
508 BDRV_POLL_WHILE(bs, atomic_read(&drained_end_counter) > 0);
511 void bdrv_apply_subtree_drain(BdrvChild *child, BlockDriverState *new_parent)
513 int i;
515 for (i = 0; i < new_parent->recursive_quiesce_counter; i++) {
516 bdrv_do_drained_begin(child->bs, true, child, false, true);
520 void bdrv_unapply_subtree_drain(BdrvChild *child, BlockDriverState *old_parent)
522 int drained_end_counter = 0;
523 int i;
525 for (i = 0; i < old_parent->recursive_quiesce_counter; i++) {
526 bdrv_do_drained_end(child->bs, true, child, false,
527 &drained_end_counter);
530 BDRV_POLL_WHILE(child->bs, atomic_read(&drained_end_counter) > 0);
534 * Wait for pending requests to complete on a single BlockDriverState subtree,
535 * and suspend block driver's internal I/O until next request arrives.
537 * Note that unlike bdrv_drain_all(), the caller must hold the BlockDriverState
538 * AioContext.
540 void coroutine_fn bdrv_co_drain(BlockDriverState *bs)
542 assert(qemu_in_coroutine());
543 bdrv_drained_begin(bs);
544 bdrv_drained_end(bs);
547 void bdrv_drain(BlockDriverState *bs)
549 bdrv_drained_begin(bs);
550 bdrv_drained_end(bs);
553 static void bdrv_drain_assert_idle(BlockDriverState *bs)
555 BdrvChild *child, *next;
557 assert(atomic_read(&bs->in_flight) == 0);
558 QLIST_FOREACH_SAFE(child, &bs->children, next, next) {
559 bdrv_drain_assert_idle(child->bs);
563 unsigned int bdrv_drain_all_count = 0;
565 static bool bdrv_drain_all_poll(void)
567 BlockDriverState *bs = NULL;
568 bool result = false;
570 /* bdrv_drain_poll() can't make changes to the graph and we are holding the
571 * main AioContext lock, so iterating bdrv_next_all_states() is safe. */
572 while ((bs = bdrv_next_all_states(bs))) {
573 AioContext *aio_context = bdrv_get_aio_context(bs);
574 aio_context_acquire(aio_context);
575 result |= bdrv_drain_poll(bs, false, NULL, true);
576 aio_context_release(aio_context);
579 return result;
583 * Wait for pending requests to complete across all BlockDriverStates
585 * This function does not flush data to disk, use bdrv_flush_all() for that
586 * after calling this function.
588 * This pauses all block jobs and disables external clients. It must
589 * be paired with bdrv_drain_all_end().
591 * NOTE: no new block jobs or BlockDriverStates can be created between
592 * the bdrv_drain_all_begin() and bdrv_drain_all_end() calls.
594 void bdrv_drain_all_begin(void)
596 BlockDriverState *bs = NULL;
598 if (qemu_in_coroutine()) {
599 bdrv_co_yield_to_drain(NULL, true, false, NULL, true, true, NULL);
600 return;
603 /* AIO_WAIT_WHILE() with a NULL context can only be called from the main
604 * loop AioContext, so make sure we're in the main context. */
605 assert(qemu_get_current_aio_context() == qemu_get_aio_context());
606 assert(bdrv_drain_all_count < INT_MAX);
607 bdrv_drain_all_count++;
609 /* Quiesce all nodes, without polling in-flight requests yet. The graph
610 * cannot change during this loop. */
611 while ((bs = bdrv_next_all_states(bs))) {
612 AioContext *aio_context = bdrv_get_aio_context(bs);
614 aio_context_acquire(aio_context);
615 bdrv_do_drained_begin(bs, false, NULL, true, false);
616 aio_context_release(aio_context);
619 /* Now poll the in-flight requests */
620 AIO_WAIT_WHILE(NULL, bdrv_drain_all_poll());
622 while ((bs = bdrv_next_all_states(bs))) {
623 bdrv_drain_assert_idle(bs);
627 void bdrv_drain_all_end(void)
629 BlockDriverState *bs = NULL;
630 int drained_end_counter = 0;
632 while ((bs = bdrv_next_all_states(bs))) {
633 AioContext *aio_context = bdrv_get_aio_context(bs);
635 aio_context_acquire(aio_context);
636 bdrv_do_drained_end(bs, false, NULL, true, &drained_end_counter);
637 aio_context_release(aio_context);
640 assert(qemu_get_current_aio_context() == qemu_get_aio_context());
641 AIO_WAIT_WHILE(NULL, atomic_read(&drained_end_counter) > 0);
643 assert(bdrv_drain_all_count > 0);
644 bdrv_drain_all_count--;
647 void bdrv_drain_all(void)
649 bdrv_drain_all_begin();
650 bdrv_drain_all_end();
654 * Remove an active request from the tracked requests list
656 * This function should be called when a tracked request is completing.
658 static void tracked_request_end(BdrvTrackedRequest *req)
660 if (req->serialising) {
661 atomic_dec(&req->bs->serialising_in_flight);
664 qemu_co_mutex_lock(&req->bs->reqs_lock);
665 QLIST_REMOVE(req, list);
666 qemu_co_queue_restart_all(&req->wait_queue);
667 qemu_co_mutex_unlock(&req->bs->reqs_lock);
671 * Add an active request to the tracked requests list
673 static void tracked_request_begin(BdrvTrackedRequest *req,
674 BlockDriverState *bs,
675 int64_t offset,
676 uint64_t bytes,
677 enum BdrvTrackedRequestType type)
679 assert(bytes <= INT64_MAX && offset <= INT64_MAX - bytes);
681 *req = (BdrvTrackedRequest){
682 .bs = bs,
683 .offset = offset,
684 .bytes = bytes,
685 .type = type,
686 .co = qemu_coroutine_self(),
687 .serialising = false,
688 .overlap_offset = offset,
689 .overlap_bytes = bytes,
692 qemu_co_queue_init(&req->wait_queue);
694 qemu_co_mutex_lock(&bs->reqs_lock);
695 QLIST_INSERT_HEAD(&bs->tracked_requests, req, list);
696 qemu_co_mutex_unlock(&bs->reqs_lock);
699 static void mark_request_serialising(BdrvTrackedRequest *req, uint64_t align)
701 int64_t overlap_offset = req->offset & ~(align - 1);
702 uint64_t overlap_bytes = ROUND_UP(req->offset + req->bytes, align)
703 - overlap_offset;
705 if (!req->serialising) {
706 atomic_inc(&req->bs->serialising_in_flight);
707 req->serialising = true;
710 req->overlap_offset = MIN(req->overlap_offset, overlap_offset);
711 req->overlap_bytes = MAX(req->overlap_bytes, overlap_bytes);
714 static bool is_request_serialising_and_aligned(BdrvTrackedRequest *req)
717 * If the request is serialising, overlap_offset and overlap_bytes are set,
718 * so we can check if the request is aligned. Otherwise, don't care and
719 * return false.
722 return req->serialising && (req->offset == req->overlap_offset) &&
723 (req->bytes == req->overlap_bytes);
727 * Round a region to cluster boundaries
729 void bdrv_round_to_clusters(BlockDriverState *bs,
730 int64_t offset, int64_t bytes,
731 int64_t *cluster_offset,
732 int64_t *cluster_bytes)
734 BlockDriverInfo bdi;
736 if (bdrv_get_info(bs, &bdi) < 0 || bdi.cluster_size == 0) {
737 *cluster_offset = offset;
738 *cluster_bytes = bytes;
739 } else {
740 int64_t c = bdi.cluster_size;
741 *cluster_offset = QEMU_ALIGN_DOWN(offset, c);
742 *cluster_bytes = QEMU_ALIGN_UP(offset - *cluster_offset + bytes, c);
746 static int bdrv_get_cluster_size(BlockDriverState *bs)
748 BlockDriverInfo bdi;
749 int ret;
751 ret = bdrv_get_info(bs, &bdi);
752 if (ret < 0 || bdi.cluster_size == 0) {
753 return bs->bl.request_alignment;
754 } else {
755 return bdi.cluster_size;
759 static bool tracked_request_overlaps(BdrvTrackedRequest *req,
760 int64_t offset, uint64_t bytes)
762 /* aaaa bbbb */
763 if (offset >= req->overlap_offset + req->overlap_bytes) {
764 return false;
766 /* bbbb aaaa */
767 if (req->overlap_offset >= offset + bytes) {
768 return false;
770 return true;
773 void bdrv_inc_in_flight(BlockDriverState *bs)
775 atomic_inc(&bs->in_flight);
778 void bdrv_wakeup(BlockDriverState *bs)
780 aio_wait_kick();
783 void bdrv_dec_in_flight(BlockDriverState *bs)
785 atomic_dec(&bs->in_flight);
786 bdrv_wakeup(bs);
789 static bool coroutine_fn wait_serialising_requests(BdrvTrackedRequest *self)
791 BlockDriverState *bs = self->bs;
792 BdrvTrackedRequest *req;
793 bool retry;
794 bool waited = false;
796 if (!atomic_read(&bs->serialising_in_flight)) {
797 return false;
800 do {
801 retry = false;
802 qemu_co_mutex_lock(&bs->reqs_lock);
803 QLIST_FOREACH(req, &bs->tracked_requests, list) {
804 if (req == self || (!req->serialising && !self->serialising)) {
805 continue;
807 if (tracked_request_overlaps(req, self->overlap_offset,
808 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);
816 /* If the request is already (indirectly) waiting for us, or
817 * will wait for us as soon as it wakes up, then just go on
818 * (instead of producing a deadlock in the former case). */
819 if (!req->waiting_for) {
820 self->waiting_for = req;
821 qemu_co_queue_wait(&req->wait_queue, &bs->reqs_lock);
822 self->waiting_for = NULL;
823 retry = true;
824 waited = true;
825 break;
829 qemu_co_mutex_unlock(&bs->reqs_lock);
830 } while (retry);
832 return waited;
835 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset,
836 size_t size)
838 if (size > BDRV_REQUEST_MAX_BYTES) {
839 return -EIO;
842 if (!bdrv_is_inserted(bs)) {
843 return -ENOMEDIUM;
846 if (offset < 0) {
847 return -EIO;
850 return 0;
853 typedef struct RwCo {
854 BdrvChild *child;
855 int64_t offset;
856 QEMUIOVector *qiov;
857 bool is_write;
858 int ret;
859 BdrvRequestFlags flags;
860 } RwCo;
862 static void coroutine_fn bdrv_rw_co_entry(void *opaque)
864 RwCo *rwco = opaque;
866 if (!rwco->is_write) {
867 rwco->ret = bdrv_co_preadv(rwco->child, rwco->offset,
868 rwco->qiov->size, rwco->qiov,
869 rwco->flags);
870 } else {
871 rwco->ret = bdrv_co_pwritev(rwco->child, rwco->offset,
872 rwco->qiov->size, rwco->qiov,
873 rwco->flags);
875 aio_wait_kick();
879 * Process a vectored synchronous request using coroutines
881 static int bdrv_prwv_co(BdrvChild *child, int64_t offset,
882 QEMUIOVector *qiov, bool is_write,
883 BdrvRequestFlags flags)
885 Coroutine *co;
886 RwCo rwco = {
887 .child = child,
888 .offset = offset,
889 .qiov = qiov,
890 .is_write = is_write,
891 .ret = NOT_DONE,
892 .flags = flags,
895 if (qemu_in_coroutine()) {
896 /* Fast-path if already in coroutine context */
897 bdrv_rw_co_entry(&rwco);
898 } else {
899 co = qemu_coroutine_create(bdrv_rw_co_entry, &rwco);
900 bdrv_coroutine_enter(child->bs, co);
901 BDRV_POLL_WHILE(child->bs, rwco.ret == NOT_DONE);
903 return rwco.ret;
906 int bdrv_pwrite_zeroes(BdrvChild *child, int64_t offset,
907 int bytes, BdrvRequestFlags flags)
909 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, bytes);
911 return bdrv_prwv_co(child, offset, &qiov, true,
912 BDRV_REQ_ZERO_WRITE | flags);
916 * Completely zero out a block device with the help of bdrv_pwrite_zeroes.
917 * The operation is sped up by checking the block status and only writing
918 * zeroes to the device if they currently do not return zeroes. Optional
919 * flags are passed through to bdrv_pwrite_zeroes (e.g. BDRV_REQ_MAY_UNMAP,
920 * BDRV_REQ_FUA).
922 * Returns < 0 on error, 0 on success. For error codes see bdrv_write().
924 int bdrv_make_zero(BdrvChild *child, BdrvRequestFlags flags)
926 int ret;
927 int64_t target_size, bytes, offset = 0;
928 BlockDriverState *bs = child->bs;
930 target_size = bdrv_getlength(bs);
931 if (target_size < 0) {
932 return target_size;
935 for (;;) {
936 bytes = MIN(target_size - offset, BDRV_REQUEST_MAX_BYTES);
937 if (bytes <= 0) {
938 return 0;
940 ret = bdrv_block_status(bs, offset, bytes, &bytes, NULL, NULL);
941 if (ret < 0) {
942 return ret;
944 if (ret & BDRV_BLOCK_ZERO) {
945 offset += bytes;
946 continue;
948 ret = bdrv_pwrite_zeroes(child, offset, bytes, flags);
949 if (ret < 0) {
950 return ret;
952 offset += bytes;
956 int bdrv_preadv(BdrvChild *child, int64_t offset, QEMUIOVector *qiov)
958 int ret;
960 ret = bdrv_prwv_co(child, offset, qiov, false, 0);
961 if (ret < 0) {
962 return ret;
965 return qiov->size;
968 /* See bdrv_pwrite() for the return codes */
969 int bdrv_pread(BdrvChild *child, int64_t offset, void *buf, int bytes)
971 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, bytes);
973 if (bytes < 0) {
974 return -EINVAL;
977 return bdrv_preadv(child, offset, &qiov);
980 int bdrv_pwritev(BdrvChild *child, int64_t offset, QEMUIOVector *qiov)
982 int ret;
984 ret = bdrv_prwv_co(child, offset, qiov, true, 0);
985 if (ret < 0) {
986 return ret;
989 return qiov->size;
992 /* Return no. of bytes on success or < 0 on error. Important errors are:
993 -EIO generic I/O error (may happen for all errors)
994 -ENOMEDIUM No media inserted.
995 -EINVAL Invalid offset or number of bytes
996 -EACCES Trying to write a read-only device
998 int bdrv_pwrite(BdrvChild *child, int64_t offset, const void *buf, int bytes)
1000 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, bytes);
1002 if (bytes < 0) {
1003 return -EINVAL;
1006 return bdrv_pwritev(child, offset, &qiov);
1010 * Writes to the file and ensures that no writes are reordered across this
1011 * request (acts as a barrier)
1013 * Returns 0 on success, -errno in error cases.
1015 int bdrv_pwrite_sync(BdrvChild *child, int64_t offset,
1016 const void *buf, int count)
1018 int ret;
1020 ret = bdrv_pwrite(child, offset, buf, count);
1021 if (ret < 0) {
1022 return ret;
1025 ret = bdrv_flush(child->bs);
1026 if (ret < 0) {
1027 return ret;
1030 return 0;
1033 typedef struct CoroutineIOCompletion {
1034 Coroutine *coroutine;
1035 int ret;
1036 } CoroutineIOCompletion;
1038 static void bdrv_co_io_em_complete(void *opaque, int ret)
1040 CoroutineIOCompletion *co = opaque;
1042 co->ret = ret;
1043 aio_co_wake(co->coroutine);
1046 static int coroutine_fn bdrv_driver_preadv(BlockDriverState *bs,
1047 uint64_t offset, uint64_t bytes,
1048 QEMUIOVector *qiov,
1049 size_t qiov_offset, int flags)
1051 BlockDriver *drv = bs->drv;
1052 int64_t sector_num;
1053 unsigned int nb_sectors;
1054 QEMUIOVector local_qiov;
1055 int ret;
1057 assert(!(flags & ~BDRV_REQ_MASK));
1058 assert(!(flags & BDRV_REQ_NO_FALLBACK));
1060 if (!drv) {
1061 return -ENOMEDIUM;
1064 if (drv->bdrv_co_preadv_part) {
1065 return drv->bdrv_co_preadv_part(bs, offset, bytes, qiov, qiov_offset,
1066 flags);
1069 if (qiov_offset > 0 || bytes != qiov->size) {
1070 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes);
1071 qiov = &local_qiov;
1074 if (drv->bdrv_co_preadv) {
1075 ret = drv->bdrv_co_preadv(bs, offset, bytes, qiov, flags);
1076 goto out;
1079 if (drv->bdrv_aio_preadv) {
1080 BlockAIOCB *acb;
1081 CoroutineIOCompletion co = {
1082 .coroutine = qemu_coroutine_self(),
1085 acb = drv->bdrv_aio_preadv(bs, offset, bytes, qiov, flags,
1086 bdrv_co_io_em_complete, &co);
1087 if (acb == NULL) {
1088 ret = -EIO;
1089 goto out;
1090 } else {
1091 qemu_coroutine_yield();
1092 ret = co.ret;
1093 goto out;
1097 sector_num = offset >> BDRV_SECTOR_BITS;
1098 nb_sectors = bytes >> BDRV_SECTOR_BITS;
1100 assert(QEMU_IS_ALIGNED(offset, BDRV_SECTOR_SIZE));
1101 assert(QEMU_IS_ALIGNED(bytes, BDRV_SECTOR_SIZE));
1102 assert(bytes <= BDRV_REQUEST_MAX_BYTES);
1103 assert(drv->bdrv_co_readv);
1105 ret = drv->bdrv_co_readv(bs, sector_num, nb_sectors, qiov);
1107 out:
1108 if (qiov == &local_qiov) {
1109 qemu_iovec_destroy(&local_qiov);
1112 return ret;
1115 static int coroutine_fn bdrv_driver_pwritev(BlockDriverState *bs,
1116 uint64_t offset, uint64_t bytes,
1117 QEMUIOVector *qiov,
1118 size_t qiov_offset, int flags)
1120 BlockDriver *drv = bs->drv;
1121 int64_t sector_num;
1122 unsigned int nb_sectors;
1123 QEMUIOVector local_qiov;
1124 int ret;
1126 assert(!(flags & ~BDRV_REQ_MASK));
1127 assert(!(flags & BDRV_REQ_NO_FALLBACK));
1129 if (!drv) {
1130 return -ENOMEDIUM;
1133 if (drv->bdrv_co_pwritev_part) {
1134 ret = drv->bdrv_co_pwritev_part(bs, offset, bytes, qiov, qiov_offset,
1135 flags & bs->supported_write_flags);
1136 flags &= ~bs->supported_write_flags;
1137 goto emulate_flags;
1140 if (qiov_offset > 0 || bytes != qiov->size) {
1141 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes);
1142 qiov = &local_qiov;
1145 if (drv->bdrv_co_pwritev) {
1146 ret = drv->bdrv_co_pwritev(bs, offset, bytes, qiov,
1147 flags & bs->supported_write_flags);
1148 flags &= ~bs->supported_write_flags;
1149 goto emulate_flags;
1152 if (drv->bdrv_aio_pwritev) {
1153 BlockAIOCB *acb;
1154 CoroutineIOCompletion co = {
1155 .coroutine = qemu_coroutine_self(),
1158 acb = drv->bdrv_aio_pwritev(bs, offset, bytes, qiov,
1159 flags & bs->supported_write_flags,
1160 bdrv_co_io_em_complete, &co);
1161 flags &= ~bs->supported_write_flags;
1162 if (acb == NULL) {
1163 ret = -EIO;
1164 } else {
1165 qemu_coroutine_yield();
1166 ret = co.ret;
1168 goto emulate_flags;
1171 sector_num = offset >> BDRV_SECTOR_BITS;
1172 nb_sectors = bytes >> BDRV_SECTOR_BITS;
1174 assert(QEMU_IS_ALIGNED(offset, BDRV_SECTOR_SIZE));
1175 assert(QEMU_IS_ALIGNED(bytes, BDRV_SECTOR_SIZE));
1176 assert(bytes <= BDRV_REQUEST_MAX_BYTES);
1178 assert(drv->bdrv_co_writev);
1179 ret = drv->bdrv_co_writev(bs, sector_num, nb_sectors, qiov,
1180 flags & bs->supported_write_flags);
1181 flags &= ~bs->supported_write_flags;
1183 emulate_flags:
1184 if (ret == 0 && (flags & BDRV_REQ_FUA)) {
1185 ret = bdrv_co_flush(bs);
1188 if (qiov == &local_qiov) {
1189 qemu_iovec_destroy(&local_qiov);
1192 return ret;
1195 static int coroutine_fn
1196 bdrv_driver_pwritev_compressed(BlockDriverState *bs, uint64_t offset,
1197 uint64_t bytes, QEMUIOVector *qiov,
1198 size_t qiov_offset)
1200 BlockDriver *drv = bs->drv;
1201 QEMUIOVector local_qiov;
1202 int ret;
1204 if (!drv) {
1205 return -ENOMEDIUM;
1208 if (!block_driver_can_compress(drv)) {
1209 return -ENOTSUP;
1212 if (drv->bdrv_co_pwritev_compressed_part) {
1213 return drv->bdrv_co_pwritev_compressed_part(bs, offset, bytes,
1214 qiov, qiov_offset);
1217 if (qiov_offset == 0) {
1218 return drv->bdrv_co_pwritev_compressed(bs, offset, bytes, qiov);
1221 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes);
1222 ret = drv->bdrv_co_pwritev_compressed(bs, offset, bytes, &local_qiov);
1223 qemu_iovec_destroy(&local_qiov);
1225 return ret;
1228 static int coroutine_fn bdrv_co_do_copy_on_readv(BdrvChild *child,
1229 int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
1230 size_t qiov_offset, int flags)
1232 BlockDriverState *bs = child->bs;
1234 /* Perform I/O through a temporary buffer so that users who scribble over
1235 * their read buffer while the operation is in progress do not end up
1236 * modifying the image file. This is critical for zero-copy guest I/O
1237 * where anything might happen inside guest memory.
1239 void *bounce_buffer = NULL;
1241 BlockDriver *drv = bs->drv;
1242 int64_t cluster_offset;
1243 int64_t cluster_bytes;
1244 size_t skip_bytes;
1245 int ret;
1246 int max_transfer = MIN_NON_ZERO(bs->bl.max_transfer,
1247 BDRV_REQUEST_MAX_BYTES);
1248 unsigned int progress = 0;
1250 if (!drv) {
1251 return -ENOMEDIUM;
1254 /* FIXME We cannot require callers to have write permissions when all they
1255 * are doing is a read request. If we did things right, write permissions
1256 * would be obtained anyway, but internally by the copy-on-read code. As
1257 * long as it is implemented here rather than in a separate filter driver,
1258 * the copy-on-read code doesn't have its own BdrvChild, however, for which
1259 * it could request permissions. Therefore we have to bypass the permission
1260 * system for the moment. */
1261 // assert(child->perm & (BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE));
1263 /* Cover entire cluster so no additional backing file I/O is required when
1264 * allocating cluster in the image file. Note that this value may exceed
1265 * BDRV_REQUEST_MAX_BYTES (even when the original read did not), which
1266 * is one reason we loop rather than doing it all at once.
1268 bdrv_round_to_clusters(bs, offset, bytes, &cluster_offset, &cluster_bytes);
1269 skip_bytes = offset - cluster_offset;
1271 trace_bdrv_co_do_copy_on_readv(bs, offset, bytes,
1272 cluster_offset, cluster_bytes);
1274 while (cluster_bytes) {
1275 int64_t pnum;
1277 ret = bdrv_is_allocated(bs, cluster_offset,
1278 MIN(cluster_bytes, max_transfer), &pnum);
1279 if (ret < 0) {
1280 /* Safe to treat errors in querying allocation as if
1281 * unallocated; we'll probably fail again soon on the
1282 * read, but at least that will set a decent errno.
1284 pnum = MIN(cluster_bytes, max_transfer);
1287 /* Stop at EOF if the image ends in the middle of the cluster */
1288 if (ret == 0 && pnum == 0) {
1289 assert(progress >= bytes);
1290 break;
1293 assert(skip_bytes < pnum);
1295 if (ret <= 0) {
1296 QEMUIOVector local_qiov;
1298 /* Must copy-on-read; use the bounce buffer */
1299 pnum = MIN(pnum, MAX_BOUNCE_BUFFER);
1300 if (!bounce_buffer) {
1301 int64_t max_we_need = MAX(pnum, cluster_bytes - pnum);
1302 int64_t max_allowed = MIN(max_transfer, MAX_BOUNCE_BUFFER);
1303 int64_t bounce_buffer_len = MIN(max_we_need, max_allowed);
1305 bounce_buffer = qemu_try_blockalign(bs, bounce_buffer_len);
1306 if (!bounce_buffer) {
1307 ret = -ENOMEM;
1308 goto err;
1311 qemu_iovec_init_buf(&local_qiov, bounce_buffer, pnum);
1313 ret = bdrv_driver_preadv(bs, cluster_offset, pnum,
1314 &local_qiov, 0, 0);
1315 if (ret < 0) {
1316 goto err;
1319 bdrv_debug_event(bs, BLKDBG_COR_WRITE);
1320 if (drv->bdrv_co_pwrite_zeroes &&
1321 buffer_is_zero(bounce_buffer, pnum)) {
1322 /* FIXME: Should we (perhaps conditionally) be setting
1323 * BDRV_REQ_MAY_UNMAP, if it will allow for a sparser copy
1324 * that still correctly reads as zero? */
1325 ret = bdrv_co_do_pwrite_zeroes(bs, cluster_offset, pnum,
1326 BDRV_REQ_WRITE_UNCHANGED);
1327 } else {
1328 /* This does not change the data on the disk, it is not
1329 * necessary to flush even in cache=writethrough mode.
1331 ret = bdrv_driver_pwritev(bs, cluster_offset, pnum,
1332 &local_qiov, 0,
1333 BDRV_REQ_WRITE_UNCHANGED);
1336 if (ret < 0) {
1337 /* It might be okay to ignore write errors for guest
1338 * requests. If this is a deliberate copy-on-read
1339 * then we don't want to ignore the error. Simply
1340 * report it in all cases.
1342 goto err;
1345 if (!(flags & BDRV_REQ_PREFETCH)) {
1346 qemu_iovec_from_buf(qiov, qiov_offset + progress,
1347 bounce_buffer + skip_bytes,
1348 pnum - skip_bytes);
1350 } else if (!(flags & BDRV_REQ_PREFETCH)) {
1351 /* Read directly into the destination */
1352 ret = bdrv_driver_preadv(bs, offset + progress,
1353 MIN(pnum - skip_bytes, bytes - progress),
1354 qiov, qiov_offset + progress, 0);
1355 if (ret < 0) {
1356 goto err;
1360 cluster_offset += pnum;
1361 cluster_bytes -= pnum;
1362 progress += pnum - skip_bytes;
1363 skip_bytes = 0;
1365 ret = 0;
1367 err:
1368 qemu_vfree(bounce_buffer);
1369 return ret;
1373 * Forwards an already correctly aligned request to the BlockDriver. This
1374 * handles copy on read, zeroing after EOF, and fragmentation of large
1375 * reads; any other features must be implemented by the caller.
1377 static int coroutine_fn bdrv_aligned_preadv(BdrvChild *child,
1378 BdrvTrackedRequest *req, int64_t offset, unsigned int bytes,
1379 int64_t align, QEMUIOVector *qiov, size_t qiov_offset, int flags)
1381 BlockDriverState *bs = child->bs;
1382 int64_t total_bytes, max_bytes;
1383 int ret = 0;
1384 uint64_t bytes_remaining = bytes;
1385 int max_transfer;
1387 assert(is_power_of_2(align));
1388 assert((offset & (align - 1)) == 0);
1389 assert((bytes & (align - 1)) == 0);
1390 assert((bs->open_flags & BDRV_O_NO_IO) == 0);
1391 max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX),
1392 align);
1394 /* TODO: We would need a per-BDS .supported_read_flags and
1395 * potential fallback support, if we ever implement any read flags
1396 * to pass through to drivers. For now, there aren't any
1397 * passthrough flags. */
1398 assert(!(flags & ~(BDRV_REQ_NO_SERIALISING | BDRV_REQ_COPY_ON_READ |
1399 BDRV_REQ_PREFETCH)));
1401 /* Handle Copy on Read and associated serialisation */
1402 if (flags & BDRV_REQ_COPY_ON_READ) {
1403 /* If we touch the same cluster it counts as an overlap. This
1404 * guarantees that allocating writes will be serialized and not race
1405 * with each other for the same cluster. For example, in copy-on-read
1406 * it ensures that the CoR read and write operations are atomic and
1407 * guest writes cannot interleave between them. */
1408 mark_request_serialising(req, bdrv_get_cluster_size(bs));
1411 /* BDRV_REQ_SERIALISING is only for write operation */
1412 assert(!(flags & BDRV_REQ_SERIALISING));
1414 if (!(flags & BDRV_REQ_NO_SERIALISING)) {
1415 wait_serialising_requests(req);
1418 if (flags & BDRV_REQ_COPY_ON_READ) {
1419 int64_t pnum;
1421 ret = bdrv_is_allocated(bs, offset, bytes, &pnum);
1422 if (ret < 0) {
1423 goto out;
1426 if (!ret || pnum != bytes) {
1427 ret = bdrv_co_do_copy_on_readv(child, offset, bytes,
1428 qiov, qiov_offset, flags);
1429 goto out;
1430 } else if (flags & BDRV_REQ_PREFETCH) {
1431 goto out;
1435 /* Forward the request to the BlockDriver, possibly fragmenting it */
1436 total_bytes = bdrv_getlength(bs);
1437 if (total_bytes < 0) {
1438 ret = total_bytes;
1439 goto out;
1442 max_bytes = ROUND_UP(MAX(0, total_bytes - offset), align);
1443 if (bytes <= max_bytes && bytes <= max_transfer) {
1444 ret = bdrv_driver_preadv(bs, offset, bytes, qiov, qiov_offset, 0);
1445 goto out;
1448 while (bytes_remaining) {
1449 int num;
1451 if (max_bytes) {
1452 num = MIN(bytes_remaining, MIN(max_bytes, max_transfer));
1453 assert(num);
1455 ret = bdrv_driver_preadv(bs, offset + bytes - bytes_remaining,
1456 num, qiov, bytes - bytes_remaining, 0);
1457 max_bytes -= num;
1458 } else {
1459 num = bytes_remaining;
1460 ret = qemu_iovec_memset(qiov, bytes - bytes_remaining, 0,
1461 bytes_remaining);
1463 if (ret < 0) {
1464 goto out;
1466 bytes_remaining -= num;
1469 out:
1470 return ret < 0 ? ret : 0;
1474 * Request padding
1476 * |<---- align ----->| |<----- align ---->|
1477 * |<- head ->|<------------- bytes ------------->|<-- tail -->|
1478 * | | | | | |
1479 * -*----------$-------*-------- ... --------*-----$------------*---
1480 * | | | | | |
1481 * | offset | | end |
1482 * ALIGN_DOWN(offset) ALIGN_UP(offset) ALIGN_DOWN(end) ALIGN_UP(end)
1483 * [buf ... ) [tail_buf )
1485 * @buf is an aligned allocation needed to store @head and @tail paddings. @head
1486 * is placed at the beginning of @buf and @tail at the @end.
1488 * @tail_buf is a pointer to sub-buffer, corresponding to align-sized chunk
1489 * around tail, if tail exists.
1491 * @merge_reads is true for small requests,
1492 * if @buf_len == @head + bytes + @tail. In this case it is possible that both
1493 * head and tail exist but @buf_len == align and @tail_buf == @buf.
1495 typedef struct BdrvRequestPadding {
1496 uint8_t *buf;
1497 size_t buf_len;
1498 uint8_t *tail_buf;
1499 size_t head;
1500 size_t tail;
1501 bool merge_reads;
1502 QEMUIOVector local_qiov;
1503 } BdrvRequestPadding;
1505 static bool bdrv_init_padding(BlockDriverState *bs,
1506 int64_t offset, int64_t bytes,
1507 BdrvRequestPadding *pad)
1509 uint64_t align = bs->bl.request_alignment;
1510 size_t sum;
1512 memset(pad, 0, sizeof(*pad));
1514 pad->head = offset & (align - 1);
1515 pad->tail = ((offset + bytes) & (align - 1));
1516 if (pad->tail) {
1517 pad->tail = align - pad->tail;
1520 if ((!pad->head && !pad->tail) || !bytes) {
1521 return false;
1524 sum = pad->head + bytes + pad->tail;
1525 pad->buf_len = (sum > align && pad->head && pad->tail) ? 2 * align : align;
1526 pad->buf = qemu_blockalign(bs, pad->buf_len);
1527 pad->merge_reads = sum == pad->buf_len;
1528 if (pad->tail) {
1529 pad->tail_buf = pad->buf + pad->buf_len - align;
1532 return true;
1535 static int bdrv_padding_rmw_read(BdrvChild *child,
1536 BdrvTrackedRequest *req,
1537 BdrvRequestPadding *pad,
1538 bool zero_middle)
1540 QEMUIOVector local_qiov;
1541 BlockDriverState *bs = child->bs;
1542 uint64_t align = bs->bl.request_alignment;
1543 int ret;
1545 assert(req->serialising && pad->buf);
1547 if (pad->head || pad->merge_reads) {
1548 uint64_t bytes = pad->merge_reads ? pad->buf_len : align;
1550 qemu_iovec_init_buf(&local_qiov, pad->buf, bytes);
1552 if (pad->head) {
1553 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD);
1555 if (pad->merge_reads && pad->tail) {
1556 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
1558 ret = bdrv_aligned_preadv(child, req, req->overlap_offset, bytes,
1559 align, &local_qiov, 0, 0);
1560 if (ret < 0) {
1561 return ret;
1563 if (pad->head) {
1564 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD);
1566 if (pad->merge_reads && pad->tail) {
1567 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
1570 if (pad->merge_reads) {
1571 goto zero_mem;
1575 if (pad->tail) {
1576 qemu_iovec_init_buf(&local_qiov, pad->tail_buf, align);
1578 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
1579 ret = bdrv_aligned_preadv(
1580 child, req,
1581 req->overlap_offset + req->overlap_bytes - align,
1582 align, align, &local_qiov, 0, 0);
1583 if (ret < 0) {
1584 return ret;
1586 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
1589 zero_mem:
1590 if (zero_middle) {
1591 memset(pad->buf + pad->head, 0, pad->buf_len - pad->head - pad->tail);
1594 return 0;
1597 static void bdrv_padding_destroy(BdrvRequestPadding *pad)
1599 if (pad->buf) {
1600 qemu_vfree(pad->buf);
1601 qemu_iovec_destroy(&pad->local_qiov);
1606 * bdrv_pad_request
1608 * Exchange request parameters with padded request if needed. Don't include RMW
1609 * read of padding, bdrv_padding_rmw_read() should be called separately if
1610 * needed.
1612 * All parameters except @bs are in-out: they represent original request at
1613 * function call and padded (if padding needed) at function finish.
1615 * Function always succeeds.
1617 static bool bdrv_pad_request(BlockDriverState *bs,
1618 QEMUIOVector **qiov, size_t *qiov_offset,
1619 int64_t *offset, unsigned int *bytes,
1620 BdrvRequestPadding *pad)
1622 if (!bdrv_init_padding(bs, *offset, *bytes, pad)) {
1623 return false;
1626 qemu_iovec_init_extended(&pad->local_qiov, pad->buf, pad->head,
1627 *qiov, *qiov_offset, *bytes,
1628 pad->buf + pad->buf_len - pad->tail, pad->tail);
1629 *bytes += pad->head + pad->tail;
1630 *offset -= pad->head;
1631 *qiov = &pad->local_qiov;
1632 *qiov_offset = 0;
1634 return true;
1637 int coroutine_fn bdrv_co_preadv(BdrvChild *child,
1638 int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
1639 BdrvRequestFlags flags)
1641 return bdrv_co_preadv_part(child, offset, bytes, qiov, 0, flags);
1644 int coroutine_fn bdrv_co_preadv_part(BdrvChild *child,
1645 int64_t offset, unsigned int bytes,
1646 QEMUIOVector *qiov, size_t qiov_offset,
1647 BdrvRequestFlags flags)
1649 BlockDriverState *bs = child->bs;
1650 BdrvTrackedRequest req;
1651 BdrvRequestPadding pad;
1652 int ret;
1654 trace_bdrv_co_preadv(bs, offset, bytes, flags);
1656 ret = bdrv_check_byte_request(bs, offset, bytes);
1657 if (ret < 0) {
1658 return ret;
1661 bdrv_inc_in_flight(bs);
1663 /* Don't do copy-on-read if we read data before write operation */
1664 if (atomic_read(&bs->copy_on_read) && !(flags & BDRV_REQ_NO_SERIALISING)) {
1665 flags |= BDRV_REQ_COPY_ON_READ;
1668 bdrv_pad_request(bs, &qiov, &qiov_offset, &offset, &bytes, &pad);
1670 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_READ);
1671 ret = bdrv_aligned_preadv(child, &req, offset, bytes,
1672 bs->bl.request_alignment,
1673 qiov, qiov_offset, flags);
1674 tracked_request_end(&req);
1675 bdrv_dec_in_flight(bs);
1677 bdrv_padding_destroy(&pad);
1679 return ret;
1682 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs,
1683 int64_t offset, int bytes, BdrvRequestFlags flags)
1685 BlockDriver *drv = bs->drv;
1686 QEMUIOVector qiov;
1687 void *buf = NULL;
1688 int ret = 0;
1689 bool need_flush = false;
1690 int head = 0;
1691 int tail = 0;
1693 int max_write_zeroes = MIN_NON_ZERO(bs->bl.max_pwrite_zeroes, INT_MAX);
1694 int alignment = MAX(bs->bl.pwrite_zeroes_alignment,
1695 bs->bl.request_alignment);
1696 int max_transfer = MIN_NON_ZERO(bs->bl.max_transfer, MAX_BOUNCE_BUFFER);
1698 if (!drv) {
1699 return -ENOMEDIUM;
1702 if ((flags & ~bs->supported_zero_flags) & BDRV_REQ_NO_FALLBACK) {
1703 return -ENOTSUP;
1706 assert(alignment % bs->bl.request_alignment == 0);
1707 head = offset % alignment;
1708 tail = (offset + bytes) % alignment;
1709 max_write_zeroes = QEMU_ALIGN_DOWN(max_write_zeroes, alignment);
1710 assert(max_write_zeroes >= bs->bl.request_alignment);
1712 while (bytes > 0 && !ret) {
1713 int num = bytes;
1715 /* Align request. Block drivers can expect the "bulk" of the request
1716 * to be aligned, and that unaligned requests do not cross cluster
1717 * boundaries.
1719 if (head) {
1720 /* Make a small request up to the first aligned sector. For
1721 * convenience, limit this request to max_transfer even if
1722 * we don't need to fall back to writes. */
1723 num = MIN(MIN(bytes, max_transfer), alignment - head);
1724 head = (head + num) % alignment;
1725 assert(num < max_write_zeroes);
1726 } else if (tail && num > alignment) {
1727 /* Shorten the request to the last aligned sector. */
1728 num -= tail;
1731 /* limit request size */
1732 if (num > max_write_zeroes) {
1733 num = max_write_zeroes;
1736 ret = -ENOTSUP;
1737 /* First try the efficient write zeroes operation */
1738 if (drv->bdrv_co_pwrite_zeroes) {
1739 ret = drv->bdrv_co_pwrite_zeroes(bs, offset, num,
1740 flags & bs->supported_zero_flags);
1741 if (ret != -ENOTSUP && (flags & BDRV_REQ_FUA) &&
1742 !(bs->supported_zero_flags & BDRV_REQ_FUA)) {
1743 need_flush = true;
1745 } else {
1746 assert(!bs->supported_zero_flags);
1749 if (ret == -ENOTSUP && !(flags & BDRV_REQ_NO_FALLBACK)) {
1750 /* Fall back to bounce buffer if write zeroes is unsupported */
1751 BdrvRequestFlags write_flags = flags & ~BDRV_REQ_ZERO_WRITE;
1753 if ((flags & BDRV_REQ_FUA) &&
1754 !(bs->supported_write_flags & BDRV_REQ_FUA)) {
1755 /* No need for bdrv_driver_pwrite() to do a fallback
1756 * flush on each chunk; use just one at the end */
1757 write_flags &= ~BDRV_REQ_FUA;
1758 need_flush = true;
1760 num = MIN(num, max_transfer);
1761 if (buf == NULL) {
1762 buf = qemu_try_blockalign0(bs, num);
1763 if (buf == NULL) {
1764 ret = -ENOMEM;
1765 goto fail;
1768 qemu_iovec_init_buf(&qiov, buf, num);
1770 ret = bdrv_driver_pwritev(bs, offset, num, &qiov, 0, write_flags);
1772 /* Keep bounce buffer around if it is big enough for all
1773 * all future requests.
1775 if (num < max_transfer) {
1776 qemu_vfree(buf);
1777 buf = NULL;
1781 offset += num;
1782 bytes -= num;
1785 fail:
1786 if (ret == 0 && need_flush) {
1787 ret = bdrv_co_flush(bs);
1789 qemu_vfree(buf);
1790 return ret;
1793 static inline int coroutine_fn
1794 bdrv_co_write_req_prepare(BdrvChild *child, int64_t offset, uint64_t bytes,
1795 BdrvTrackedRequest *req, int flags)
1797 BlockDriverState *bs = child->bs;
1798 bool waited;
1799 int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);
1801 if (bs->read_only) {
1802 return -EPERM;
1805 /* BDRV_REQ_NO_SERIALISING is only for read operation */
1806 assert(!(flags & BDRV_REQ_NO_SERIALISING));
1807 assert(!(bs->open_flags & BDRV_O_INACTIVE));
1808 assert((bs->open_flags & BDRV_O_NO_IO) == 0);
1809 assert(!(flags & ~BDRV_REQ_MASK));
1811 if (flags & BDRV_REQ_SERIALISING) {
1812 mark_request_serialising(req, bdrv_get_cluster_size(bs));
1815 waited = wait_serialising_requests(req);
1817 assert(!waited || !req->serialising ||
1818 is_request_serialising_and_aligned(req));
1819 assert(req->overlap_offset <= offset);
1820 assert(offset + bytes <= req->overlap_offset + req->overlap_bytes);
1821 assert(end_sector <= bs->total_sectors || child->perm & BLK_PERM_RESIZE);
1823 switch (req->type) {
1824 case BDRV_TRACKED_WRITE:
1825 case BDRV_TRACKED_DISCARD:
1826 if (flags & BDRV_REQ_WRITE_UNCHANGED) {
1827 assert(child->perm & (BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE));
1828 } else {
1829 assert(child->perm & BLK_PERM_WRITE);
1831 return notifier_with_return_list_notify(&bs->before_write_notifiers,
1832 req);
1833 case BDRV_TRACKED_TRUNCATE:
1834 assert(child->perm & BLK_PERM_RESIZE);
1835 return 0;
1836 default:
1837 abort();
1841 static inline void coroutine_fn
1842 bdrv_co_write_req_finish(BdrvChild *child, int64_t offset, uint64_t bytes,
1843 BdrvTrackedRequest *req, int ret)
1845 int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);
1846 BlockDriverState *bs = child->bs;
1848 atomic_inc(&bs->write_gen);
1851 * Discard cannot extend the image, but in error handling cases, such as
1852 * when reverting a qcow2 cluster allocation, the discarded range can pass
1853 * the end of image file, so we cannot assert about BDRV_TRACKED_DISCARD
1854 * here. Instead, just skip it, since semantically a discard request
1855 * beyond EOF cannot expand the image anyway.
1857 if (ret == 0 &&
1858 (req->type == BDRV_TRACKED_TRUNCATE ||
1859 end_sector > bs->total_sectors) &&
1860 req->type != BDRV_TRACKED_DISCARD) {
1861 bs->total_sectors = end_sector;
1862 bdrv_parent_cb_resize(bs);
1863 bdrv_dirty_bitmap_truncate(bs, end_sector << BDRV_SECTOR_BITS);
1865 if (req->bytes) {
1866 switch (req->type) {
1867 case BDRV_TRACKED_WRITE:
1868 stat64_max(&bs->wr_highest_offset, offset + bytes);
1869 /* fall through, to set dirty bits */
1870 case BDRV_TRACKED_DISCARD:
1871 bdrv_set_dirty(bs, offset, bytes);
1872 break;
1873 default:
1874 break;
1880 * Forwards an already correctly aligned write request to the BlockDriver,
1881 * after possibly fragmenting it.
1883 static int coroutine_fn bdrv_aligned_pwritev(BdrvChild *child,
1884 BdrvTrackedRequest *req, int64_t offset, unsigned int bytes,
1885 int64_t align, QEMUIOVector *qiov, size_t qiov_offset, int flags)
1887 BlockDriverState *bs = child->bs;
1888 BlockDriver *drv = bs->drv;
1889 int ret;
1891 uint64_t bytes_remaining = bytes;
1892 int max_transfer;
1894 if (!drv) {
1895 return -ENOMEDIUM;
1898 if (bdrv_has_readonly_bitmaps(bs)) {
1899 return -EPERM;
1902 assert(is_power_of_2(align));
1903 assert((offset & (align - 1)) == 0);
1904 assert((bytes & (align - 1)) == 0);
1905 assert(!qiov || qiov_offset + bytes <= qiov->size);
1906 max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX),
1907 align);
1909 ret = bdrv_co_write_req_prepare(child, offset, bytes, req, flags);
1911 if (!ret && bs->detect_zeroes != BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF &&
1912 !(flags & BDRV_REQ_ZERO_WRITE) && drv->bdrv_co_pwrite_zeroes &&
1913 qemu_iovec_is_zero(qiov, qiov_offset, bytes)) {
1914 flags |= BDRV_REQ_ZERO_WRITE;
1915 if (bs->detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP) {
1916 flags |= BDRV_REQ_MAY_UNMAP;
1920 if (ret < 0) {
1921 /* Do nothing, write notifier decided to fail this request */
1922 } else if (flags & BDRV_REQ_ZERO_WRITE) {
1923 bdrv_debug_event(bs, BLKDBG_PWRITEV_ZERO);
1924 ret = bdrv_co_do_pwrite_zeroes(bs, offset, bytes, flags);
1925 } else if (flags & BDRV_REQ_WRITE_COMPRESSED) {
1926 ret = bdrv_driver_pwritev_compressed(bs, offset, bytes,
1927 qiov, qiov_offset);
1928 } else if (bytes <= max_transfer) {
1929 bdrv_debug_event(bs, BLKDBG_PWRITEV);
1930 ret = bdrv_driver_pwritev(bs, offset, bytes, qiov, qiov_offset, flags);
1931 } else {
1932 bdrv_debug_event(bs, BLKDBG_PWRITEV);
1933 while (bytes_remaining) {
1934 int num = MIN(bytes_remaining, max_transfer);
1935 int local_flags = flags;
1937 assert(num);
1938 if (num < bytes_remaining && (flags & BDRV_REQ_FUA) &&
1939 !(bs->supported_write_flags & BDRV_REQ_FUA)) {
1940 /* If FUA is going to be emulated by flush, we only
1941 * need to flush on the last iteration */
1942 local_flags &= ~BDRV_REQ_FUA;
1945 ret = bdrv_driver_pwritev(bs, offset + bytes - bytes_remaining,
1946 num, qiov, bytes - bytes_remaining,
1947 local_flags);
1948 if (ret < 0) {
1949 break;
1951 bytes_remaining -= num;
1954 bdrv_debug_event(bs, BLKDBG_PWRITEV_DONE);
1956 if (ret >= 0) {
1957 ret = 0;
1959 bdrv_co_write_req_finish(child, offset, bytes, req, ret);
1961 return ret;
1964 static int coroutine_fn bdrv_co_do_zero_pwritev(BdrvChild *child,
1965 int64_t offset,
1966 unsigned int bytes,
1967 BdrvRequestFlags flags,
1968 BdrvTrackedRequest *req)
1970 BlockDriverState *bs = child->bs;
1971 QEMUIOVector local_qiov;
1972 uint64_t align = bs->bl.request_alignment;
1973 int ret = 0;
1974 bool padding;
1975 BdrvRequestPadding pad;
1977 padding = bdrv_init_padding(bs, offset, bytes, &pad);
1978 if (padding) {
1979 mark_request_serialising(req, align);
1980 wait_serialising_requests(req);
1982 bdrv_padding_rmw_read(child, req, &pad, true);
1984 if (pad.head || pad.merge_reads) {
1985 int64_t aligned_offset = offset & ~(align - 1);
1986 int64_t write_bytes = pad.merge_reads ? pad.buf_len : align;
1988 qemu_iovec_init_buf(&local_qiov, pad.buf, write_bytes);
1989 ret = bdrv_aligned_pwritev(child, req, aligned_offset, write_bytes,
1990 align, &local_qiov, 0,
1991 flags & ~BDRV_REQ_ZERO_WRITE);
1992 if (ret < 0 || pad.merge_reads) {
1993 /* Error or all work is done */
1994 goto out;
1996 offset += write_bytes - pad.head;
1997 bytes -= write_bytes - pad.head;
2001 assert(!bytes || (offset & (align - 1)) == 0);
2002 if (bytes >= align) {
2003 /* Write the aligned part in the middle. */
2004 uint64_t aligned_bytes = bytes & ~(align - 1);
2005 ret = bdrv_aligned_pwritev(child, req, offset, aligned_bytes, align,
2006 NULL, 0, flags);
2007 if (ret < 0) {
2008 goto out;
2010 bytes -= aligned_bytes;
2011 offset += aligned_bytes;
2014 assert(!bytes || (offset & (align - 1)) == 0);
2015 if (bytes) {
2016 assert(align == pad.tail + bytes);
2018 qemu_iovec_init_buf(&local_qiov, pad.tail_buf, align);
2019 ret = bdrv_aligned_pwritev(child, req, offset, align, align,
2020 &local_qiov, 0,
2021 flags & ~BDRV_REQ_ZERO_WRITE);
2024 out:
2025 bdrv_padding_destroy(&pad);
2027 return ret;
2031 * Handle a write request in coroutine context
2033 int coroutine_fn bdrv_co_pwritev(BdrvChild *child,
2034 int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
2035 BdrvRequestFlags flags)
2037 return bdrv_co_pwritev_part(child, offset, bytes, qiov, 0, flags);
2040 int coroutine_fn bdrv_co_pwritev_part(BdrvChild *child,
2041 int64_t offset, unsigned int bytes, QEMUIOVector *qiov, size_t qiov_offset,
2042 BdrvRequestFlags flags)
2044 BlockDriverState *bs = child->bs;
2045 BdrvTrackedRequest req;
2046 uint64_t align = bs->bl.request_alignment;
2047 BdrvRequestPadding pad;
2048 int ret;
2050 trace_bdrv_co_pwritev(child->bs, offset, bytes, flags);
2052 if (!bs->drv) {
2053 return -ENOMEDIUM;
2056 ret = bdrv_check_byte_request(bs, offset, bytes);
2057 if (ret < 0) {
2058 return ret;
2061 bdrv_inc_in_flight(bs);
2063 * Align write if necessary by performing a read-modify-write cycle.
2064 * Pad qiov with the read parts and be sure to have a tracked request not
2065 * only for bdrv_aligned_pwritev, but also for the reads of the RMW cycle.
2067 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_WRITE);
2069 if (flags & BDRV_REQ_ZERO_WRITE) {
2070 ret = bdrv_co_do_zero_pwritev(child, offset, bytes, flags, &req);
2071 goto out;
2074 if (bdrv_pad_request(bs, &qiov, &qiov_offset, &offset, &bytes, &pad)) {
2075 mark_request_serialising(&req, align);
2076 wait_serialising_requests(&req);
2077 bdrv_padding_rmw_read(child, &req, &pad, false);
2080 ret = bdrv_aligned_pwritev(child, &req, offset, bytes, align,
2081 qiov, qiov_offset, flags);
2083 bdrv_padding_destroy(&pad);
2085 out:
2086 tracked_request_end(&req);
2087 bdrv_dec_in_flight(bs);
2089 return ret;
2092 int coroutine_fn bdrv_co_pwrite_zeroes(BdrvChild *child, int64_t offset,
2093 int bytes, BdrvRequestFlags flags)
2095 trace_bdrv_co_pwrite_zeroes(child->bs, offset, bytes, flags);
2097 if (!(child->bs->open_flags & BDRV_O_UNMAP)) {
2098 flags &= ~BDRV_REQ_MAY_UNMAP;
2101 return bdrv_co_pwritev(child, offset, bytes, NULL,
2102 BDRV_REQ_ZERO_WRITE | flags);
2106 * Flush ALL BDSes regardless of if they are reachable via a BlkBackend or not.
2108 int bdrv_flush_all(void)
2110 BdrvNextIterator it;
2111 BlockDriverState *bs = NULL;
2112 int result = 0;
2114 for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
2115 AioContext *aio_context = bdrv_get_aio_context(bs);
2116 int ret;
2118 aio_context_acquire(aio_context);
2119 ret = bdrv_flush(bs);
2120 if (ret < 0 && !result) {
2121 result = ret;
2123 aio_context_release(aio_context);
2126 return result;
2130 typedef struct BdrvCoBlockStatusData {
2131 BlockDriverState *bs;
2132 BlockDriverState *base;
2133 bool want_zero;
2134 int64_t offset;
2135 int64_t bytes;
2136 int64_t *pnum;
2137 int64_t *map;
2138 BlockDriverState **file;
2139 int ret;
2140 bool done;
2141 } BdrvCoBlockStatusData;
2143 int coroutine_fn bdrv_co_block_status_from_file(BlockDriverState *bs,
2144 bool want_zero,
2145 int64_t offset,
2146 int64_t bytes,
2147 int64_t *pnum,
2148 int64_t *map,
2149 BlockDriverState **file)
2151 assert(bs->file && bs->file->bs);
2152 *pnum = bytes;
2153 *map = offset;
2154 *file = bs->file->bs;
2155 return BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID;
2158 int coroutine_fn bdrv_co_block_status_from_backing(BlockDriverState *bs,
2159 bool want_zero,
2160 int64_t offset,
2161 int64_t bytes,
2162 int64_t *pnum,
2163 int64_t *map,
2164 BlockDriverState **file)
2166 assert(bs->backing && bs->backing->bs);
2167 *pnum = bytes;
2168 *map = offset;
2169 *file = bs->backing->bs;
2170 return BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID;
2174 * Returns the allocation status of the specified sectors.
2175 * Drivers not implementing the functionality are assumed to not support
2176 * backing files, hence all their sectors are reported as allocated.
2178 * If 'want_zero' is true, the caller is querying for mapping
2179 * purposes, with a focus on valid BDRV_BLOCK_OFFSET_VALID, _DATA, and
2180 * _ZERO where possible; otherwise, the result favors larger 'pnum',
2181 * with a focus on accurate BDRV_BLOCK_ALLOCATED.
2183 * If 'offset' is beyond the end of the disk image the return value is
2184 * BDRV_BLOCK_EOF and 'pnum' is set to 0.
2186 * 'bytes' is the max value 'pnum' should be set to. If bytes goes
2187 * beyond the end of the disk image it will be clamped; if 'pnum' is set to
2188 * the end of the image, then the returned value will include BDRV_BLOCK_EOF.
2190 * 'pnum' is set to the number of bytes (including and immediately
2191 * following the specified offset) that are easily known to be in the
2192 * same allocated/unallocated state. Note that a second call starting
2193 * at the original offset plus returned pnum may have the same status.
2194 * The returned value is non-zero on success except at end-of-file.
2196 * Returns negative errno on failure. Otherwise, if the
2197 * BDRV_BLOCK_OFFSET_VALID bit is set, 'map' and 'file' (if non-NULL) are
2198 * set to the host mapping and BDS corresponding to the guest offset.
2200 static int coroutine_fn bdrv_co_block_status(BlockDriverState *bs,
2201 bool want_zero,
2202 int64_t offset, int64_t bytes,
2203 int64_t *pnum, int64_t *map,
2204 BlockDriverState **file)
2206 int64_t total_size;
2207 int64_t n; /* bytes */
2208 int ret;
2209 int64_t local_map = 0;
2210 BlockDriverState *local_file = NULL;
2211 int64_t aligned_offset, aligned_bytes;
2212 uint32_t align;
2214 assert(pnum);
2215 *pnum = 0;
2216 total_size = bdrv_getlength(bs);
2217 if (total_size < 0) {
2218 ret = total_size;
2219 goto early_out;
2222 if (offset >= total_size) {
2223 ret = BDRV_BLOCK_EOF;
2224 goto early_out;
2226 if (!bytes) {
2227 ret = 0;
2228 goto early_out;
2231 n = total_size - offset;
2232 if (n < bytes) {
2233 bytes = n;
2236 /* Must be non-NULL or bdrv_getlength() would have failed */
2237 assert(bs->drv);
2238 if (!bs->drv->bdrv_co_block_status) {
2239 *pnum = bytes;
2240 ret = BDRV_BLOCK_DATA | BDRV_BLOCK_ALLOCATED;
2241 if (offset + bytes == total_size) {
2242 ret |= BDRV_BLOCK_EOF;
2244 if (bs->drv->protocol_name) {
2245 ret |= BDRV_BLOCK_OFFSET_VALID;
2246 local_map = offset;
2247 local_file = bs;
2249 goto early_out;
2252 bdrv_inc_in_flight(bs);
2254 /* Round out to request_alignment boundaries */
2255 align = bs->bl.request_alignment;
2256 aligned_offset = QEMU_ALIGN_DOWN(offset, align);
2257 aligned_bytes = ROUND_UP(offset + bytes, align) - aligned_offset;
2259 ret = bs->drv->bdrv_co_block_status(bs, want_zero, aligned_offset,
2260 aligned_bytes, pnum, &local_map,
2261 &local_file);
2262 if (ret < 0) {
2263 *pnum = 0;
2264 goto out;
2268 * The driver's result must be a non-zero multiple of request_alignment.
2269 * Clamp pnum and adjust map to original request.
2271 assert(*pnum && QEMU_IS_ALIGNED(*pnum, align) &&
2272 align > offset - aligned_offset);
2273 if (ret & BDRV_BLOCK_RECURSE) {
2274 assert(ret & BDRV_BLOCK_DATA);
2275 assert(ret & BDRV_BLOCK_OFFSET_VALID);
2276 assert(!(ret & BDRV_BLOCK_ZERO));
2279 *pnum -= offset - aligned_offset;
2280 if (*pnum > bytes) {
2281 *pnum = bytes;
2283 if (ret & BDRV_BLOCK_OFFSET_VALID) {
2284 local_map += offset - aligned_offset;
2287 if (ret & BDRV_BLOCK_RAW) {
2288 assert(ret & BDRV_BLOCK_OFFSET_VALID && local_file);
2289 ret = bdrv_co_block_status(local_file, want_zero, local_map,
2290 *pnum, pnum, &local_map, &local_file);
2291 goto out;
2294 if (ret & (BDRV_BLOCK_DATA | BDRV_BLOCK_ZERO)) {
2295 ret |= BDRV_BLOCK_ALLOCATED;
2296 } else if (want_zero) {
2297 if (bdrv_unallocated_blocks_are_zero(bs)) {
2298 ret |= BDRV_BLOCK_ZERO;
2299 } else if (bs->backing) {
2300 BlockDriverState *bs2 = bs->backing->bs;
2301 int64_t size2 = bdrv_getlength(bs2);
2303 if (size2 >= 0 && offset >= size2) {
2304 ret |= BDRV_BLOCK_ZERO;
2309 if (want_zero && ret & BDRV_BLOCK_RECURSE &&
2310 local_file && local_file != bs &&
2311 (ret & BDRV_BLOCK_DATA) && !(ret & BDRV_BLOCK_ZERO) &&
2312 (ret & BDRV_BLOCK_OFFSET_VALID)) {
2313 int64_t file_pnum;
2314 int ret2;
2316 ret2 = bdrv_co_block_status(local_file, want_zero, local_map,
2317 *pnum, &file_pnum, NULL, NULL);
2318 if (ret2 >= 0) {
2319 /* Ignore errors. This is just providing extra information, it
2320 * is useful but not necessary.
2322 if (ret2 & BDRV_BLOCK_EOF &&
2323 (!file_pnum || ret2 & BDRV_BLOCK_ZERO)) {
2325 * It is valid for the format block driver to read
2326 * beyond the end of the underlying file's current
2327 * size; such areas read as zero.
2329 ret |= BDRV_BLOCK_ZERO;
2330 } else {
2331 /* Limit request to the range reported by the protocol driver */
2332 *pnum = file_pnum;
2333 ret |= (ret2 & BDRV_BLOCK_ZERO);
2338 out:
2339 bdrv_dec_in_flight(bs);
2340 if (ret >= 0 && offset + *pnum == total_size) {
2341 ret |= BDRV_BLOCK_EOF;
2343 early_out:
2344 if (file) {
2345 *file = local_file;
2347 if (map) {
2348 *map = local_map;
2350 return ret;
2353 static int coroutine_fn bdrv_co_block_status_above(BlockDriverState *bs,
2354 BlockDriverState *base,
2355 bool want_zero,
2356 int64_t offset,
2357 int64_t bytes,
2358 int64_t *pnum,
2359 int64_t *map,
2360 BlockDriverState **file)
2362 BlockDriverState *p;
2363 int ret = 0;
2364 bool first = true;
2366 assert(bs != base);
2367 for (p = bs; p != base; p = backing_bs(p)) {
2368 ret = bdrv_co_block_status(p, want_zero, offset, bytes, pnum, map,
2369 file);
2370 if (ret < 0) {
2371 break;
2373 if (ret & BDRV_BLOCK_ZERO && ret & BDRV_BLOCK_EOF && !first) {
2375 * Reading beyond the end of the file continues to read
2376 * zeroes, but we can only widen the result to the
2377 * unallocated length we learned from an earlier
2378 * iteration.
2380 *pnum = bytes;
2382 if (ret & (BDRV_BLOCK_ZERO | BDRV_BLOCK_DATA)) {
2383 break;
2385 /* [offset, pnum] unallocated on this layer, which could be only
2386 * the first part of [offset, bytes]. */
2387 bytes = MIN(bytes, *pnum);
2388 first = false;
2390 return ret;
2393 /* Coroutine wrapper for bdrv_block_status_above() */
2394 static void coroutine_fn bdrv_block_status_above_co_entry(void *opaque)
2396 BdrvCoBlockStatusData *data = opaque;
2398 data->ret = bdrv_co_block_status_above(data->bs, data->base,
2399 data->want_zero,
2400 data->offset, data->bytes,
2401 data->pnum, data->map, data->file);
2402 data->done = true;
2403 aio_wait_kick();
2407 * Synchronous wrapper around bdrv_co_block_status_above().
2409 * See bdrv_co_block_status_above() for details.
2411 static int bdrv_common_block_status_above(BlockDriverState *bs,
2412 BlockDriverState *base,
2413 bool want_zero, int64_t offset,
2414 int64_t bytes, int64_t *pnum,
2415 int64_t *map,
2416 BlockDriverState **file)
2418 Coroutine *co;
2419 BdrvCoBlockStatusData data = {
2420 .bs = bs,
2421 .base = base,
2422 .want_zero = want_zero,
2423 .offset = offset,
2424 .bytes = bytes,
2425 .pnum = pnum,
2426 .map = map,
2427 .file = file,
2428 .done = false,
2431 if (qemu_in_coroutine()) {
2432 /* Fast-path if already in coroutine context */
2433 bdrv_block_status_above_co_entry(&data);
2434 } else {
2435 co = qemu_coroutine_create(bdrv_block_status_above_co_entry, &data);
2436 bdrv_coroutine_enter(bs, co);
2437 BDRV_POLL_WHILE(bs, !data.done);
2439 return data.ret;
2442 int bdrv_block_status_above(BlockDriverState *bs, BlockDriverState *base,
2443 int64_t offset, int64_t bytes, int64_t *pnum,
2444 int64_t *map, BlockDriverState **file)
2446 return bdrv_common_block_status_above(bs, base, true, offset, bytes,
2447 pnum, map, file);
2450 int bdrv_block_status(BlockDriverState *bs, int64_t offset, int64_t bytes,
2451 int64_t *pnum, int64_t *map, BlockDriverState **file)
2453 return bdrv_block_status_above(bs, backing_bs(bs),
2454 offset, bytes, pnum, map, file);
2457 int coroutine_fn bdrv_is_allocated(BlockDriverState *bs, int64_t offset,
2458 int64_t bytes, int64_t *pnum)
2460 int ret;
2461 int64_t dummy;
2463 ret = bdrv_common_block_status_above(bs, backing_bs(bs), false, offset,
2464 bytes, pnum ? pnum : &dummy, NULL,
2465 NULL);
2466 if (ret < 0) {
2467 return ret;
2469 return !!(ret & BDRV_BLOCK_ALLOCATED);
2473 * Given an image chain: ... -> [BASE] -> [INTER1] -> [INTER2] -> [TOP]
2475 * Return 1 if (a prefix of) the given range is allocated in any image
2476 * between BASE and TOP (BASE is only included if include_base is set).
2477 * BASE can be NULL to check if the given offset is allocated in any
2478 * image of the chain. Return 0 otherwise, or negative errno on
2479 * failure.
2481 * 'pnum' is set to the number of bytes (including and immediately
2482 * following the specified offset) that are known to be in the same
2483 * allocated/unallocated state. Note that a subsequent call starting
2484 * at 'offset + *pnum' may return the same allocation status (in other
2485 * words, the result is not necessarily the maximum possible range);
2486 * but 'pnum' will only be 0 when end of file is reached.
2489 int bdrv_is_allocated_above(BlockDriverState *top,
2490 BlockDriverState *base,
2491 bool include_base, int64_t offset,
2492 int64_t bytes, int64_t *pnum)
2494 BlockDriverState *intermediate;
2495 int ret;
2496 int64_t n = bytes;
2498 assert(base || !include_base);
2500 intermediate = top;
2501 while (include_base || intermediate != base) {
2502 int64_t pnum_inter;
2503 int64_t size_inter;
2505 assert(intermediate);
2506 ret = bdrv_is_allocated(intermediate, offset, bytes, &pnum_inter);
2507 if (ret < 0) {
2508 return ret;
2510 if (ret) {
2511 *pnum = pnum_inter;
2512 return 1;
2515 size_inter = bdrv_getlength(intermediate);
2516 if (size_inter < 0) {
2517 return size_inter;
2519 if (n > pnum_inter &&
2520 (intermediate == top || offset + pnum_inter < size_inter)) {
2521 n = pnum_inter;
2524 if (intermediate == base) {
2525 break;
2528 intermediate = backing_bs(intermediate);
2531 *pnum = n;
2532 return 0;
2535 typedef struct BdrvVmstateCo {
2536 BlockDriverState *bs;
2537 QEMUIOVector *qiov;
2538 int64_t pos;
2539 bool is_read;
2540 int ret;
2541 } BdrvVmstateCo;
2543 static int coroutine_fn
2544 bdrv_co_rw_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos,
2545 bool is_read)
2547 BlockDriver *drv = bs->drv;
2548 int ret = -ENOTSUP;
2550 bdrv_inc_in_flight(bs);
2552 if (!drv) {
2553 ret = -ENOMEDIUM;
2554 } else if (drv->bdrv_load_vmstate) {
2555 if (is_read) {
2556 ret = drv->bdrv_load_vmstate(bs, qiov, pos);
2557 } else {
2558 ret = drv->bdrv_save_vmstate(bs, qiov, pos);
2560 } else if (bs->file) {
2561 ret = bdrv_co_rw_vmstate(bs->file->bs, qiov, pos, is_read);
2564 bdrv_dec_in_flight(bs);
2565 return ret;
2568 static void coroutine_fn bdrv_co_rw_vmstate_entry(void *opaque)
2570 BdrvVmstateCo *co = opaque;
2571 co->ret = bdrv_co_rw_vmstate(co->bs, co->qiov, co->pos, co->is_read);
2572 aio_wait_kick();
2575 static inline int
2576 bdrv_rw_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos,
2577 bool is_read)
2579 if (qemu_in_coroutine()) {
2580 return bdrv_co_rw_vmstate(bs, qiov, pos, is_read);
2581 } else {
2582 BdrvVmstateCo data = {
2583 .bs = bs,
2584 .qiov = qiov,
2585 .pos = pos,
2586 .is_read = is_read,
2587 .ret = -EINPROGRESS,
2589 Coroutine *co = qemu_coroutine_create(bdrv_co_rw_vmstate_entry, &data);
2591 bdrv_coroutine_enter(bs, co);
2592 BDRV_POLL_WHILE(bs, data.ret == -EINPROGRESS);
2593 return data.ret;
2597 int bdrv_save_vmstate(BlockDriverState *bs, const uint8_t *buf,
2598 int64_t pos, int size)
2600 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, size);
2601 int ret;
2603 ret = bdrv_writev_vmstate(bs, &qiov, pos);
2604 if (ret < 0) {
2605 return ret;
2608 return size;
2611 int bdrv_writev_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos)
2613 return bdrv_rw_vmstate(bs, qiov, pos, false);
2616 int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf,
2617 int64_t pos, int size)
2619 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, size);
2620 int ret;
2622 ret = bdrv_readv_vmstate(bs, &qiov, pos);
2623 if (ret < 0) {
2624 return ret;
2627 return size;
2630 int bdrv_readv_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos)
2632 return bdrv_rw_vmstate(bs, qiov, pos, true);
2635 /**************************************************************/
2636 /* async I/Os */
2638 void bdrv_aio_cancel(BlockAIOCB *acb)
2640 qemu_aio_ref(acb);
2641 bdrv_aio_cancel_async(acb);
2642 while (acb->refcnt > 1) {
2643 if (acb->aiocb_info->get_aio_context) {
2644 aio_poll(acb->aiocb_info->get_aio_context(acb), true);
2645 } else if (acb->bs) {
2646 /* qemu_aio_ref and qemu_aio_unref are not thread-safe, so
2647 * assert that we're not using an I/O thread. Thread-safe
2648 * code should use bdrv_aio_cancel_async exclusively.
2650 assert(bdrv_get_aio_context(acb->bs) == qemu_get_aio_context());
2651 aio_poll(bdrv_get_aio_context(acb->bs), true);
2652 } else {
2653 abort();
2656 qemu_aio_unref(acb);
2659 /* Async version of aio cancel. The caller is not blocked if the acb implements
2660 * cancel_async, otherwise we do nothing and let the request normally complete.
2661 * In either case the completion callback must be called. */
2662 void bdrv_aio_cancel_async(BlockAIOCB *acb)
2664 if (acb->aiocb_info->cancel_async) {
2665 acb->aiocb_info->cancel_async(acb);
2669 /**************************************************************/
2670 /* Coroutine block device emulation */
2672 typedef struct FlushCo {
2673 BlockDriverState *bs;
2674 int ret;
2675 } FlushCo;
2678 static void coroutine_fn bdrv_flush_co_entry(void *opaque)
2680 FlushCo *rwco = opaque;
2682 rwco->ret = bdrv_co_flush(rwco->bs);
2683 aio_wait_kick();
2686 int coroutine_fn bdrv_co_flush(BlockDriverState *bs)
2688 int current_gen;
2689 int ret = 0;
2691 bdrv_inc_in_flight(bs);
2693 if (!bdrv_is_inserted(bs) || bdrv_is_read_only(bs) ||
2694 bdrv_is_sg(bs)) {
2695 goto early_exit;
2698 qemu_co_mutex_lock(&bs->reqs_lock);
2699 current_gen = atomic_read(&bs->write_gen);
2701 /* Wait until any previous flushes are completed */
2702 while (bs->active_flush_req) {
2703 qemu_co_queue_wait(&bs->flush_queue, &bs->reqs_lock);
2706 /* Flushes reach this point in nondecreasing current_gen order. */
2707 bs->active_flush_req = true;
2708 qemu_co_mutex_unlock(&bs->reqs_lock);
2710 /* Write back all layers by calling one driver function */
2711 if (bs->drv->bdrv_co_flush) {
2712 ret = bs->drv->bdrv_co_flush(bs);
2713 goto out;
2716 /* Write back cached data to the OS even with cache=unsafe */
2717 BLKDBG_EVENT(bs->file, BLKDBG_FLUSH_TO_OS);
2718 if (bs->drv->bdrv_co_flush_to_os) {
2719 ret = bs->drv->bdrv_co_flush_to_os(bs);
2720 if (ret < 0) {
2721 goto out;
2725 /* But don't actually force it to the disk with cache=unsafe */
2726 if (bs->open_flags & BDRV_O_NO_FLUSH) {
2727 goto flush_parent;
2730 /* Check if we really need to flush anything */
2731 if (bs->flushed_gen == current_gen) {
2732 goto flush_parent;
2735 BLKDBG_EVENT(bs->file, BLKDBG_FLUSH_TO_DISK);
2736 if (!bs->drv) {
2737 /* bs->drv->bdrv_co_flush() might have ejected the BDS
2738 * (even in case of apparent success) */
2739 ret = -ENOMEDIUM;
2740 goto out;
2742 if (bs->drv->bdrv_co_flush_to_disk) {
2743 ret = bs->drv->bdrv_co_flush_to_disk(bs);
2744 } else if (bs->drv->bdrv_aio_flush) {
2745 BlockAIOCB *acb;
2746 CoroutineIOCompletion co = {
2747 .coroutine = qemu_coroutine_self(),
2750 acb = bs->drv->bdrv_aio_flush(bs, bdrv_co_io_em_complete, &co);
2751 if (acb == NULL) {
2752 ret = -EIO;
2753 } else {
2754 qemu_coroutine_yield();
2755 ret = co.ret;
2757 } else {
2759 * Some block drivers always operate in either writethrough or unsafe
2760 * mode and don't support bdrv_flush therefore. Usually qemu doesn't
2761 * know how the server works (because the behaviour is hardcoded or
2762 * depends on server-side configuration), so we can't ensure that
2763 * everything is safe on disk. Returning an error doesn't work because
2764 * that would break guests even if the server operates in writethrough
2765 * mode.
2767 * Let's hope the user knows what he's doing.
2769 ret = 0;
2772 if (ret < 0) {
2773 goto out;
2776 /* Now flush the underlying protocol. It will also have BDRV_O_NO_FLUSH
2777 * in the case of cache=unsafe, so there are no useless flushes.
2779 flush_parent:
2780 ret = bs->file ? bdrv_co_flush(bs->file->bs) : 0;
2781 out:
2782 /* Notify any pending flushes that we have completed */
2783 if (ret == 0) {
2784 bs->flushed_gen = current_gen;
2787 qemu_co_mutex_lock(&bs->reqs_lock);
2788 bs->active_flush_req = false;
2789 /* Return value is ignored - it's ok if wait queue is empty */
2790 qemu_co_queue_next(&bs->flush_queue);
2791 qemu_co_mutex_unlock(&bs->reqs_lock);
2793 early_exit:
2794 bdrv_dec_in_flight(bs);
2795 return ret;
2798 int bdrv_flush(BlockDriverState *bs)
2800 Coroutine *co;
2801 FlushCo flush_co = {
2802 .bs = bs,
2803 .ret = NOT_DONE,
2806 if (qemu_in_coroutine()) {
2807 /* Fast-path if already in coroutine context */
2808 bdrv_flush_co_entry(&flush_co);
2809 } else {
2810 co = qemu_coroutine_create(bdrv_flush_co_entry, &flush_co);
2811 bdrv_coroutine_enter(bs, co);
2812 BDRV_POLL_WHILE(bs, flush_co.ret == NOT_DONE);
2815 return flush_co.ret;
2818 typedef struct DiscardCo {
2819 BdrvChild *child;
2820 int64_t offset;
2821 int64_t bytes;
2822 int ret;
2823 } DiscardCo;
2824 static void coroutine_fn bdrv_pdiscard_co_entry(void *opaque)
2826 DiscardCo *rwco = opaque;
2828 rwco->ret = bdrv_co_pdiscard(rwco->child, rwco->offset, rwco->bytes);
2829 aio_wait_kick();
2832 int coroutine_fn bdrv_co_pdiscard(BdrvChild *child, int64_t offset,
2833 int64_t bytes)
2835 BdrvTrackedRequest req;
2836 int max_pdiscard, ret;
2837 int head, tail, align;
2838 BlockDriverState *bs = child->bs;
2840 if (!bs || !bs->drv || !bdrv_is_inserted(bs)) {
2841 return -ENOMEDIUM;
2844 if (bdrv_has_readonly_bitmaps(bs)) {
2845 return -EPERM;
2848 if (offset < 0 || bytes < 0 || bytes > INT64_MAX - offset) {
2849 return -EIO;
2852 /* Do nothing if disabled. */
2853 if (!(bs->open_flags & BDRV_O_UNMAP)) {
2854 return 0;
2857 if (!bs->drv->bdrv_co_pdiscard && !bs->drv->bdrv_aio_pdiscard) {
2858 return 0;
2861 /* Discard is advisory, but some devices track and coalesce
2862 * unaligned requests, so we must pass everything down rather than
2863 * round here. Still, most devices will just silently ignore
2864 * unaligned requests (by returning -ENOTSUP), so we must fragment
2865 * the request accordingly. */
2866 align = MAX(bs->bl.pdiscard_alignment, bs->bl.request_alignment);
2867 assert(align % bs->bl.request_alignment == 0);
2868 head = offset % align;
2869 tail = (offset + bytes) % align;
2871 bdrv_inc_in_flight(bs);
2872 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_DISCARD);
2874 ret = bdrv_co_write_req_prepare(child, offset, bytes, &req, 0);
2875 if (ret < 0) {
2876 goto out;
2879 max_pdiscard = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_pdiscard, INT_MAX),
2880 align);
2881 assert(max_pdiscard >= bs->bl.request_alignment);
2883 while (bytes > 0) {
2884 int64_t num = bytes;
2886 if (head) {
2887 /* Make small requests to get to alignment boundaries. */
2888 num = MIN(bytes, align - head);
2889 if (!QEMU_IS_ALIGNED(num, bs->bl.request_alignment)) {
2890 num %= bs->bl.request_alignment;
2892 head = (head + num) % align;
2893 assert(num < max_pdiscard);
2894 } else if (tail) {
2895 if (num > align) {
2896 /* Shorten the request to the last aligned cluster. */
2897 num -= tail;
2898 } else if (!QEMU_IS_ALIGNED(tail, bs->bl.request_alignment) &&
2899 tail > bs->bl.request_alignment) {
2900 tail %= bs->bl.request_alignment;
2901 num -= tail;
2904 /* limit request size */
2905 if (num > max_pdiscard) {
2906 num = max_pdiscard;
2909 if (!bs->drv) {
2910 ret = -ENOMEDIUM;
2911 goto out;
2913 if (bs->drv->bdrv_co_pdiscard) {
2914 ret = bs->drv->bdrv_co_pdiscard(bs, offset, num);
2915 } else {
2916 BlockAIOCB *acb;
2917 CoroutineIOCompletion co = {
2918 .coroutine = qemu_coroutine_self(),
2921 acb = bs->drv->bdrv_aio_pdiscard(bs, offset, num,
2922 bdrv_co_io_em_complete, &co);
2923 if (acb == NULL) {
2924 ret = -EIO;
2925 goto out;
2926 } else {
2927 qemu_coroutine_yield();
2928 ret = co.ret;
2931 if (ret && ret != -ENOTSUP) {
2932 goto out;
2935 offset += num;
2936 bytes -= num;
2938 ret = 0;
2939 out:
2940 bdrv_co_write_req_finish(child, req.offset, req.bytes, &req, ret);
2941 tracked_request_end(&req);
2942 bdrv_dec_in_flight(bs);
2943 return ret;
2946 int bdrv_pdiscard(BdrvChild *child, int64_t offset, int64_t bytes)
2948 Coroutine *co;
2949 DiscardCo rwco = {
2950 .child = child,
2951 .offset = offset,
2952 .bytes = bytes,
2953 .ret = NOT_DONE,
2956 if (qemu_in_coroutine()) {
2957 /* Fast-path if already in coroutine context */
2958 bdrv_pdiscard_co_entry(&rwco);
2959 } else {
2960 co = qemu_coroutine_create(bdrv_pdiscard_co_entry, &rwco);
2961 bdrv_coroutine_enter(child->bs, co);
2962 BDRV_POLL_WHILE(child->bs, rwco.ret == NOT_DONE);
2965 return rwco.ret;
2968 int bdrv_co_ioctl(BlockDriverState *bs, int req, void *buf)
2970 BlockDriver *drv = bs->drv;
2971 CoroutineIOCompletion co = {
2972 .coroutine = qemu_coroutine_self(),
2974 BlockAIOCB *acb;
2976 bdrv_inc_in_flight(bs);
2977 if (!drv || (!drv->bdrv_aio_ioctl && !drv->bdrv_co_ioctl)) {
2978 co.ret = -ENOTSUP;
2979 goto out;
2982 if (drv->bdrv_co_ioctl) {
2983 co.ret = drv->bdrv_co_ioctl(bs, req, buf);
2984 } else {
2985 acb = drv->bdrv_aio_ioctl(bs, req, buf, bdrv_co_io_em_complete, &co);
2986 if (!acb) {
2987 co.ret = -ENOTSUP;
2988 goto out;
2990 qemu_coroutine_yield();
2992 out:
2993 bdrv_dec_in_flight(bs);
2994 return co.ret;
2997 void *qemu_blockalign(BlockDriverState *bs, size_t size)
2999 return qemu_memalign(bdrv_opt_mem_align(bs), size);
3002 void *qemu_blockalign0(BlockDriverState *bs, size_t size)
3004 return memset(qemu_blockalign(bs, size), 0, size);
3007 void *qemu_try_blockalign(BlockDriverState *bs, size_t size)
3009 size_t align = bdrv_opt_mem_align(bs);
3011 /* Ensure that NULL is never returned on success */
3012 assert(align > 0);
3013 if (size == 0) {
3014 size = align;
3017 return qemu_try_memalign(align, size);
3020 void *qemu_try_blockalign0(BlockDriverState *bs, size_t size)
3022 void *mem = qemu_try_blockalign(bs, size);
3024 if (mem) {
3025 memset(mem, 0, size);
3028 return mem;
3032 * Check if all memory in this vector is sector aligned.
3034 bool bdrv_qiov_is_aligned(BlockDriverState *bs, QEMUIOVector *qiov)
3036 int i;
3037 size_t alignment = bdrv_min_mem_align(bs);
3039 for (i = 0; i < qiov->niov; i++) {
3040 if ((uintptr_t) qiov->iov[i].iov_base % alignment) {
3041 return false;
3043 if (qiov->iov[i].iov_len % alignment) {
3044 return false;
3048 return true;
3051 void bdrv_add_before_write_notifier(BlockDriverState *bs,
3052 NotifierWithReturn *notifier)
3054 notifier_with_return_list_add(&bs->before_write_notifiers, notifier);
3057 void bdrv_io_plug(BlockDriverState *bs)
3059 BdrvChild *child;
3061 QLIST_FOREACH(child, &bs->children, next) {
3062 bdrv_io_plug(child->bs);
3065 if (atomic_fetch_inc(&bs->io_plugged) == 0) {
3066 BlockDriver *drv = bs->drv;
3067 if (drv && drv->bdrv_io_plug) {
3068 drv->bdrv_io_plug(bs);
3073 void bdrv_io_unplug(BlockDriverState *bs)
3075 BdrvChild *child;
3077 assert(bs->io_plugged);
3078 if (atomic_fetch_dec(&bs->io_plugged) == 1) {
3079 BlockDriver *drv = bs->drv;
3080 if (drv && drv->bdrv_io_unplug) {
3081 drv->bdrv_io_unplug(bs);
3085 QLIST_FOREACH(child, &bs->children, next) {
3086 bdrv_io_unplug(child->bs);
3090 void bdrv_register_buf(BlockDriverState *bs, void *host, size_t size)
3092 BdrvChild *child;
3094 if (bs->drv && bs->drv->bdrv_register_buf) {
3095 bs->drv->bdrv_register_buf(bs, host, size);
3097 QLIST_FOREACH(child, &bs->children, next) {
3098 bdrv_register_buf(child->bs, host, size);
3102 void bdrv_unregister_buf(BlockDriverState *bs, void *host)
3104 BdrvChild *child;
3106 if (bs->drv && bs->drv->bdrv_unregister_buf) {
3107 bs->drv->bdrv_unregister_buf(bs, host);
3109 QLIST_FOREACH(child, &bs->children, next) {
3110 bdrv_unregister_buf(child->bs, host);
3114 static int coroutine_fn bdrv_co_copy_range_internal(
3115 BdrvChild *src, uint64_t src_offset, BdrvChild *dst,
3116 uint64_t dst_offset, uint64_t bytes,
3117 BdrvRequestFlags read_flags, BdrvRequestFlags write_flags,
3118 bool recurse_src)
3120 BdrvTrackedRequest req;
3121 int ret;
3123 /* TODO We can support BDRV_REQ_NO_FALLBACK here */
3124 assert(!(read_flags & BDRV_REQ_NO_FALLBACK));
3125 assert(!(write_flags & BDRV_REQ_NO_FALLBACK));
3127 if (!dst || !dst->bs) {
3128 return -ENOMEDIUM;
3130 ret = bdrv_check_byte_request(dst->bs, dst_offset, bytes);
3131 if (ret) {
3132 return ret;
3134 if (write_flags & BDRV_REQ_ZERO_WRITE) {
3135 return bdrv_co_pwrite_zeroes(dst, dst_offset, bytes, write_flags);
3138 if (!src || !src->bs) {
3139 return -ENOMEDIUM;
3141 ret = bdrv_check_byte_request(src->bs, src_offset, bytes);
3142 if (ret) {
3143 return ret;
3146 if (!src->bs->drv->bdrv_co_copy_range_from
3147 || !dst->bs->drv->bdrv_co_copy_range_to
3148 || src->bs->encrypted || dst->bs->encrypted) {
3149 return -ENOTSUP;
3152 if (recurse_src) {
3153 bdrv_inc_in_flight(src->bs);
3154 tracked_request_begin(&req, src->bs, src_offset, bytes,
3155 BDRV_TRACKED_READ);
3157 /* BDRV_REQ_SERIALISING is only for write operation */
3158 assert(!(read_flags & BDRV_REQ_SERIALISING));
3159 if (!(read_flags & BDRV_REQ_NO_SERIALISING)) {
3160 wait_serialising_requests(&req);
3163 ret = src->bs->drv->bdrv_co_copy_range_from(src->bs,
3164 src, src_offset,
3165 dst, dst_offset,
3166 bytes,
3167 read_flags, write_flags);
3169 tracked_request_end(&req);
3170 bdrv_dec_in_flight(src->bs);
3171 } else {
3172 bdrv_inc_in_flight(dst->bs);
3173 tracked_request_begin(&req, dst->bs, dst_offset, bytes,
3174 BDRV_TRACKED_WRITE);
3175 ret = bdrv_co_write_req_prepare(dst, dst_offset, bytes, &req,
3176 write_flags);
3177 if (!ret) {
3178 ret = dst->bs->drv->bdrv_co_copy_range_to(dst->bs,
3179 src, src_offset,
3180 dst, dst_offset,
3181 bytes,
3182 read_flags, write_flags);
3184 bdrv_co_write_req_finish(dst, dst_offset, bytes, &req, ret);
3185 tracked_request_end(&req);
3186 bdrv_dec_in_flight(dst->bs);
3189 return ret;
3192 /* Copy range from @src to @dst.
3194 * See the comment of bdrv_co_copy_range for the parameter and return value
3195 * semantics. */
3196 int coroutine_fn bdrv_co_copy_range_from(BdrvChild *src, uint64_t src_offset,
3197 BdrvChild *dst, uint64_t dst_offset,
3198 uint64_t bytes,
3199 BdrvRequestFlags read_flags,
3200 BdrvRequestFlags write_flags)
3202 trace_bdrv_co_copy_range_from(src, src_offset, dst, dst_offset, bytes,
3203 read_flags, write_flags);
3204 return bdrv_co_copy_range_internal(src, src_offset, dst, dst_offset,
3205 bytes, read_flags, write_flags, true);
3208 /* Copy range from @src to @dst.
3210 * See the comment of bdrv_co_copy_range for the parameter and return value
3211 * semantics. */
3212 int coroutine_fn bdrv_co_copy_range_to(BdrvChild *src, uint64_t src_offset,
3213 BdrvChild *dst, uint64_t dst_offset,
3214 uint64_t bytes,
3215 BdrvRequestFlags read_flags,
3216 BdrvRequestFlags write_flags)
3218 trace_bdrv_co_copy_range_to(src, src_offset, dst, dst_offset, bytes,
3219 read_flags, write_flags);
3220 return bdrv_co_copy_range_internal(src, src_offset, dst, dst_offset,
3221 bytes, read_flags, write_flags, false);
3224 int coroutine_fn bdrv_co_copy_range(BdrvChild *src, uint64_t src_offset,
3225 BdrvChild *dst, uint64_t dst_offset,
3226 uint64_t bytes, BdrvRequestFlags read_flags,
3227 BdrvRequestFlags write_flags)
3229 return bdrv_co_copy_range_from(src, src_offset,
3230 dst, dst_offset,
3231 bytes, read_flags, write_flags);
3234 static void bdrv_parent_cb_resize(BlockDriverState *bs)
3236 BdrvChild *c;
3237 QLIST_FOREACH(c, &bs->parents, next_parent) {
3238 if (c->role->resize) {
3239 c->role->resize(c);
3245 * Truncate file to 'offset' bytes (needed only for file protocols)
3247 int coroutine_fn bdrv_co_truncate(BdrvChild *child, int64_t offset,
3248 PreallocMode prealloc, Error **errp)
3250 BlockDriverState *bs = child->bs;
3251 BlockDriver *drv = bs->drv;
3252 BdrvTrackedRequest req;
3253 int64_t old_size, new_bytes;
3254 int ret;
3257 /* if bs->drv == NULL, bs is closed, so there's nothing to do here */
3258 if (!drv) {
3259 error_setg(errp, "No medium inserted");
3260 return -ENOMEDIUM;
3262 if (offset < 0) {
3263 error_setg(errp, "Image size cannot be negative");
3264 return -EINVAL;
3267 old_size = bdrv_getlength(bs);
3268 if (old_size < 0) {
3269 error_setg_errno(errp, -old_size, "Failed to get old image size");
3270 return old_size;
3273 if (offset > old_size) {
3274 new_bytes = offset - old_size;
3275 } else {
3276 new_bytes = 0;
3279 bdrv_inc_in_flight(bs);
3280 tracked_request_begin(&req, bs, offset - new_bytes, new_bytes,
3281 BDRV_TRACKED_TRUNCATE);
3283 /* If we are growing the image and potentially using preallocation for the
3284 * new area, we need to make sure that no write requests are made to it
3285 * concurrently or they might be overwritten by preallocation. */
3286 if (new_bytes) {
3287 mark_request_serialising(&req, 1);
3289 if (bs->read_only) {
3290 error_setg(errp, "Image is read-only");
3291 ret = -EACCES;
3292 goto out;
3294 ret = bdrv_co_write_req_prepare(child, offset - new_bytes, new_bytes, &req,
3296 if (ret < 0) {
3297 error_setg_errno(errp, -ret,
3298 "Failed to prepare request for truncation");
3299 goto out;
3302 if (!drv->bdrv_co_truncate) {
3303 if (bs->file && drv->is_filter) {
3304 ret = bdrv_co_truncate(bs->file, offset, prealloc, errp);
3305 goto out;
3307 error_setg(errp, "Image format driver does not support resize");
3308 ret = -ENOTSUP;
3309 goto out;
3312 ret = drv->bdrv_co_truncate(bs, offset, prealloc, errp);
3313 if (ret < 0) {
3314 goto out;
3316 ret = refresh_total_sectors(bs, offset >> BDRV_SECTOR_BITS);
3317 if (ret < 0) {
3318 error_setg_errno(errp, -ret, "Could not refresh total sector count");
3319 } else {
3320 offset = bs->total_sectors * BDRV_SECTOR_SIZE;
3322 /* It's possible that truncation succeeded but refresh_total_sectors
3323 * failed, but the latter doesn't affect how we should finish the request.
3324 * Pass 0 as the last parameter so that dirty bitmaps etc. are handled. */
3325 bdrv_co_write_req_finish(child, offset - new_bytes, new_bytes, &req, 0);
3327 out:
3328 tracked_request_end(&req);
3329 bdrv_dec_in_flight(bs);
3331 return ret;
3334 typedef struct TruncateCo {
3335 BdrvChild *child;
3336 int64_t offset;
3337 PreallocMode prealloc;
3338 Error **errp;
3339 int ret;
3340 } TruncateCo;
3342 static void coroutine_fn bdrv_truncate_co_entry(void *opaque)
3344 TruncateCo *tco = opaque;
3345 tco->ret = bdrv_co_truncate(tco->child, tco->offset, tco->prealloc,
3346 tco->errp);
3347 aio_wait_kick();
3350 int bdrv_truncate(BdrvChild *child, int64_t offset, PreallocMode prealloc,
3351 Error **errp)
3353 Coroutine *co;
3354 TruncateCo tco = {
3355 .child = child,
3356 .offset = offset,
3357 .prealloc = prealloc,
3358 .errp = errp,
3359 .ret = NOT_DONE,
3362 if (qemu_in_coroutine()) {
3363 /* Fast-path if already in coroutine context */
3364 bdrv_truncate_co_entry(&tco);
3365 } else {
3366 co = qemu_coroutine_create(bdrv_truncate_co_entry, &tco);
3367 bdrv_coroutine_enter(child->bs, co);
3368 BDRV_POLL_WHILE(child->bs, tco.ret == NOT_DONE);
3371 return tco.ret;