master
c 3,827 lines 115 KB
Raw
1 /*
2 * Block layer I/O functions
3 *
4 * Copyright (c) 2003 Fabrice Bellard
5 *
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:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
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.
23 */
24
25 #include "qemu/osdep.h"
26 #include "trace.h"
27 #include "system/block-backend.h"
28 #include "qemu/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/dirty-bitmap.h"
34 #include "block/write-threshold.h"
35 #include "qemu/cutils.h"
36 #include "qemu/memalign.h"
37 #include "qapi/error.h"
38 #include "qemu/error-report.h"
39 #include "qemu/main-loop.h"
40 #include "system/replay.h"
41 #include "qemu/units.h"
42 #include "qemu/atomic.h"
43
44 /* Maximum bounce buffer for copy-on-read and write zeroes, in bytes */
45 #define MAX_BOUNCE_BUFFER (32768 << BDRV_SECTOR_BITS)
46
47 /* Maximum read size for checking if data reads as zero, in bytes */
48 #define MAX_ZERO_CHECK_BUFFER (128 * KiB)
49
50 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs,
51 int64_t offset, int64_t bytes, BdrvRequestFlags flags);
52
53 static void GRAPH_RDLOCK
54 bdrv_parent_drained_begin(BlockDriverState *bs, BdrvChild *ignore)
55 {
56 BdrvChild *c, *next;
57 IO_OR_GS_CODE();
58 assert_bdrv_graph_readable();
59
60 QLIST_FOREACH_SAFE(c, &bs->parents, next_parent, next) {
61 if (c == ignore) {
62 continue;
63 }
64 bdrv_parent_drained_begin_single(c);
65 }
66 }
67
68 void bdrv_parent_drained_end_single(BdrvChild *c)
69 {
70 GLOBAL_STATE_CODE();
71
72 assert(c->quiesced_parent);
73 c->quiesced_parent = false;
74
75 if (c->klass->drained_end) {
76 c->klass->drained_end(c);
77 }
78 }
79
80 static void GRAPH_RDLOCK
81 bdrv_parent_drained_end(BlockDriverState *bs, BdrvChild *ignore)
82 {
83 BdrvChild *c;
84 IO_OR_GS_CODE();
85 assert_bdrv_graph_readable();
86
87 QLIST_FOREACH(c, &bs->parents, next_parent) {
88 if (c == ignore) {
89 continue;
90 }
91 bdrv_parent_drained_end_single(c);
92 }
93 }
94
95 bool bdrv_parent_drained_poll_single(BdrvChild *c)
96 {
97 IO_OR_GS_CODE();
98
99 if (c->klass->drained_poll) {
100 return c->klass->drained_poll(c);
101 }
102 return false;
103 }
104
105 static bool GRAPH_RDLOCK
106 bdrv_parent_drained_poll(BlockDriverState *bs, BdrvChild *ignore,
107 bool ignore_bds_parents)
108 {
109 BdrvChild *c, *next;
110 bool busy = false;
111 IO_OR_GS_CODE();
112 assert_bdrv_graph_readable();
113
114 QLIST_FOREACH_SAFE(c, &bs->parents, next_parent, next) {
115 if (c == ignore || (ignore_bds_parents && c->klass->parent_is_bds)) {
116 continue;
117 }
118 busy |= bdrv_parent_drained_poll_single(c);
119 }
120
121 return busy;
122 }
123
124 void bdrv_parent_drained_begin_single(BdrvChild *c)
125 {
126 GLOBAL_STATE_CODE();
127
128 assert(!c->quiesced_parent);
129 c->quiesced_parent = true;
130
131 if (c->klass->drained_begin) {
132 /* called with rdlock taken, but it doesn't really need it. */
133 c->klass->drained_begin(c);
134 }
135 }
136
137 static void bdrv_merge_limits(BlockLimits *dst, const BlockLimits *src)
138 {
139 dst->pdiscard_alignment = MAX(dst->pdiscard_alignment,
140 src->pdiscard_alignment);
141 dst->opt_transfer = MAX(dst->opt_transfer, src->opt_transfer);
142 dst->max_transfer = MIN_NON_ZERO(dst->max_transfer, src->max_transfer);
143 dst->max_hw_transfer = MIN_NON_ZERO(dst->max_hw_transfer,
144 src->max_hw_transfer);
145 dst->opt_mem_alignment = MAX(dst->opt_mem_alignment,
146 src->opt_mem_alignment);
147 dst->min_mem_alignment = MAX(dst->min_mem_alignment,
148 src->min_mem_alignment);
149 dst->max_iov = MIN_NON_ZERO(dst->max_iov, src->max_iov);
150 dst->max_hw_iov = MIN_NON_ZERO(dst->max_hw_iov, src->max_hw_iov);
151 }
152
153 typedef struct BdrvRefreshLimitsState {
154 BlockDriverState *bs;
155 BlockLimits old_bl;
156 } BdrvRefreshLimitsState;
157
158 static void bdrv_refresh_limits_abort(void *opaque)
159 {
160 BdrvRefreshLimitsState *s = opaque;
161
162 s->bs->bl = s->old_bl;
163 }
164
165 static TransactionActionDrv bdrv_refresh_limits_drv = {
166 .abort = bdrv_refresh_limits_abort,
167 .clean = g_free,
168 };
169
170 /* @tran is allowed to be NULL, in this case no rollback is possible. */
171 void bdrv_refresh_limits(BlockDriverState *bs, Transaction *tran, Error **errp)
172 {
173 ERRP_GUARD();
174 BlockDriver *drv = bs->drv;
175 BdrvChild *c;
176 bool have_limits;
177
178 GLOBAL_STATE_CODE();
179
180 if (tran) {
181 BdrvRefreshLimitsState *s = g_new(BdrvRefreshLimitsState, 1);
182 *s = (BdrvRefreshLimitsState) {
183 .bs = bs,
184 .old_bl = bs->bl,
185 };
186 tran_add(tran, &bdrv_refresh_limits_drv, s);
187 }
188
189 memset(&bs->bl, 0, sizeof(bs->bl));
190
191 if (!drv) {
192 return;
193 }
194
195 /* Default alignment based on whether driver has byte interface */
196 bs->bl.request_alignment = (drv->bdrv_co_preadv ||
197 drv->bdrv_aio_preadv ||
198 drv->bdrv_co_preadv_part) ? 1 : 512;
199
200 /* Take some limits from the children as a default */
201 have_limits = false;
202 QLIST_FOREACH(c, &bs->children, next) {
203 if (c->role & (BDRV_CHILD_DATA | BDRV_CHILD_FILTERED | BDRV_CHILD_COW))
204 {
205 bdrv_merge_limits(&bs->bl, &c->bs->bl);
206 have_limits = true;
207 }
208
209 if (c->role & BDRV_CHILD_FILTERED) {
210 bs->bl.has_variable_length |= c->bs->bl.has_variable_length;
211 }
212 }
213
214 if (!have_limits) {
215 bs->bl.min_mem_alignment = 512;
216 bs->bl.opt_mem_alignment = qemu_real_host_page_size();
217
218 /* Safe default since most protocols use readv()/writev()/etc */
219 bs->bl.max_iov = IOV_MAX;
220 }
221
222 /* Then let the driver override it */
223 if (drv->bdrv_refresh_limits) {
224 drv->bdrv_refresh_limits(bs, errp);
225 if (*errp) {
226 return;
227 }
228 }
229
230 if (bs->bl.request_alignment > BDRV_MAX_ALIGNMENT) {
231 error_setg(errp, "Driver requires too large request alignment");
232 }
233 }
234
235 /**
236 * The copy-on-read flag is actually a reference count so multiple users may
237 * use the feature without worrying about clobbering its previous state.
238 * Copy-on-read stays enabled until all users have called to disable it.
239 */
240 void bdrv_enable_copy_on_read(BlockDriverState *bs)
241 {
242 IO_CODE();
243 qatomic_inc(&bs->copy_on_read);
244 }
245
246 void bdrv_disable_copy_on_read(BlockDriverState *bs)
247 {
248 int old = qatomic_fetch_dec(&bs->copy_on_read);
249 IO_CODE();
250 assert(old >= 1);
251 }
252
253 typedef struct {
254 Coroutine *co;
255 BlockDriverState *bs;
256 bool done;
257 bool begin;
258 bool poll;
259 BdrvChild *parent;
260 } BdrvCoDrainData;
261
262 /* Returns true if BDRV_POLL_WHILE() should go into a blocking aio_poll() */
263 bool bdrv_drain_poll(BlockDriverState *bs, BdrvChild *ignore_parent,
264 bool ignore_bds_parents)
265 {
266 GLOBAL_STATE_CODE();
267
268 if (bdrv_parent_drained_poll(bs, ignore_parent, ignore_bds_parents)) {
269 return true;
270 }
271
272 if (qatomic_read(&bs->in_flight)) {
273 return true;
274 }
275
276 return false;
277 }
278
279 static bool bdrv_drain_poll_top_level(BlockDriverState *bs,
280 BdrvChild *ignore_parent)
281 {
282 GLOBAL_STATE_CODE();
283 GRAPH_RDLOCK_GUARD_MAINLOOP();
284
285 return bdrv_drain_poll(bs, ignore_parent, false);
286 }
287
288 static void bdrv_do_drained_begin(BlockDriverState *bs, BdrvChild *parent,
289 bool poll);
290 static void bdrv_do_drained_end(BlockDriverState *bs, BdrvChild *parent);
291
292 static void bdrv_co_drain_bh_cb(void *opaque)
293 {
294 BdrvCoDrainData *data = opaque;
295 Coroutine *co = data->co;
296 BlockDriverState *bs = data->bs;
297
298 if (bs) {
299 bdrv_dec_in_flight(bs);
300 if (data->begin) {
301 bdrv_do_drained_begin(bs, data->parent, data->poll);
302 } else {
303 assert(!data->poll);
304 bdrv_do_drained_end(bs, data->parent);
305 }
306 } else {
307 assert(data->begin);
308 bdrv_drain_all_begin();
309 }
310
311 data->done = true;
312 aio_co_wake(co);
313 }
314
315 static void coroutine_fn bdrv_co_yield_to_drain(BlockDriverState *bs,
316 bool begin,
317 BdrvChild *parent,
318 bool poll)
319 {
320 BdrvCoDrainData data;
321 Coroutine *self = qemu_coroutine_self();
322
323 /* Calling bdrv_drain() from a BH ensures the current coroutine yields and
324 * other coroutines run if they were queued by aio_co_enter(). */
325
326 assert(qemu_in_coroutine());
327 data = (BdrvCoDrainData) {
328 .co = self,
329 .bs = bs,
330 .done = false,
331 .begin = begin,
332 .parent = parent,
333 .poll = poll,
334 };
335
336 if (bs) {
337 bdrv_inc_in_flight(bs);
338 }
339
340 replay_bh_schedule_oneshot_event(qemu_get_aio_context(),
341 bdrv_co_drain_bh_cb, &data);
342
343 qemu_coroutine_yield();
344 /* If we are resumed from some other event (such as an aio completion or a
345 * timer callback), it is a bug in the caller that should be fixed. */
346 assert(data.done);
347 }
348
349 static void bdrv_do_drained_begin(BlockDriverState *bs, BdrvChild *parent,
350 bool poll)
351 {
352 IO_OR_GS_CODE();
353
354 if (qemu_in_coroutine()) {
355 bdrv_co_yield_to_drain(bs, true, parent, poll);
356 return;
357 }
358
359 GLOBAL_STATE_CODE();
360
361 /* Stop things in parent-to-child order */
362 if (bs->quiesce_counter++ == 0) {
363 GRAPH_RDLOCK_GUARD_MAINLOOP();
364 bdrv_parent_drained_begin(bs, parent);
365 if (bs->drv && bs->drv->bdrv_drain_begin) {
366 bs->drv->bdrv_drain_begin(bs);
367 }
368 }
369
370 /*
371 * Wait for drained requests to finish.
372 *
373 * Calling BDRV_POLL_WHILE() only once for the top-level node is okay: The
374 * call is needed so things in this AioContext can make progress even
375 * though we don't return to the main AioContext loop - this automatically
376 * includes other nodes in the same AioContext and therefore all child
377 * nodes.
378 */
379 if (poll) {
380 BDRV_POLL_WHILE(bs, bdrv_drain_poll_top_level(bs, parent));
381 }
382 }
383
384 void bdrv_do_drained_begin_quiesce(BlockDriverState *bs, BdrvChild *parent)
385 {
386 bdrv_do_drained_begin(bs, parent, false);
387 }
388
389 void coroutine_mixed_fn
390 bdrv_drained_begin(BlockDriverState *bs)
391 {
392 IO_OR_GS_CODE();
393 bdrv_do_drained_begin(bs, NULL, true);
394 }
395
396 /**
397 * This function does not poll, nor must any of its recursively called
398 * functions.
399 */
400 static void bdrv_do_drained_end(BlockDriverState *bs, BdrvChild *parent)
401 {
402 IO_OR_GS_CODE();
403
404 if (qemu_in_coroutine()) {
405 bdrv_co_yield_to_drain(bs, false, parent, false);
406 return;
407 }
408
409 /* At this point, we should be always running in the main loop. */
410 GLOBAL_STATE_CODE();
411 assert(bs->quiesce_counter > 0);
412
413 /* Re-enable things in child-to-parent order */
414 if (--bs->quiesce_counter == 0) {
415 GRAPH_RDLOCK_GUARD_MAINLOOP();
416 if (bs->drv && bs->drv->bdrv_drain_end) {
417 bs->drv->bdrv_drain_end(bs);
418 }
419 bdrv_parent_drained_end(bs, parent);
420 }
421 }
422
423 void bdrv_drained_end(BlockDriverState *bs)
424 {
425 IO_OR_GS_CODE();
426 bdrv_do_drained_end(bs, NULL);
427 }
428
429 void bdrv_drain(BlockDriverState *bs)
430 {
431 IO_OR_GS_CODE();
432 bdrv_drained_begin(bs);
433 bdrv_drained_end(bs);
434 }
435
436 static void bdrv_drain_assert_idle(BlockDriverState *bs)
437 {
438 BdrvChild *child, *next;
439 GLOBAL_STATE_CODE();
440 GRAPH_RDLOCK_GUARD_MAINLOOP();
441
442 assert(qatomic_read(&bs->in_flight) == 0);
443 QLIST_FOREACH_SAFE(child, &bs->children, next, next) {
444 bdrv_drain_assert_idle(child->bs);
445 }
446 }
447
448 unsigned int bdrv_drain_all_count = 0;
449
450 static bool bdrv_drain_all_poll(void)
451 {
452 BlockDriverState *bs = NULL;
453 bool result = false;
454
455 GLOBAL_STATE_CODE();
456 GRAPH_RDLOCK_GUARD_MAINLOOP();
457
458 /*
459 * bdrv_drain_poll() can't make changes to the graph and we hold the BQL,
460 * so iterating bdrv_next_all_states() is safe.
461 */
462 while ((bs = bdrv_next_all_states(bs))) {
463 result |= bdrv_drain_poll(bs, NULL, true);
464 }
465
466 return result;
467 }
468
469 /*
470 * Wait for pending requests to complete across all BlockDriverStates
471 *
472 * This function does not flush data to disk, use bdrv_flush_all() for that
473 * after calling this function.
474 *
475 * This pauses all block jobs and disables external clients. It must
476 * be paired with bdrv_drain_all_end().
477 *
478 * NOTE: no new block jobs or BlockDriverStates can be created between
479 * the bdrv_drain_all_begin() and bdrv_drain_all_end() calls.
480 */
481 void bdrv_drain_all_begin_nopoll(void)
482 {
483 BlockDriverState *bs = NULL;
484 GLOBAL_STATE_CODE();
485
486 /*
487 * bdrv queue is managed by record/replay,
488 * waiting for finishing the I/O requests may
489 * be infinite
490 */
491 if (replay_events_enabled()) {
492 return;
493 }
494
495 /* AIO_WAIT_WHILE() with a NULL context can only be called from the main
496 * loop AioContext, so make sure we're in the main context. */
497 assert(qemu_get_current_aio_context() == qemu_get_aio_context());
498 assert(bdrv_drain_all_count < INT_MAX);
499 bdrv_drain_all_count++;
500
501 /* Quiesce all nodes, without polling in-flight requests yet. The graph
502 * cannot change during this loop. */
503 while ((bs = bdrv_next_all_states(bs))) {
504 bdrv_do_drained_begin(bs, NULL, false);
505 }
506 }
507
508 void coroutine_mixed_fn bdrv_drain_all_begin(void)
509 {
510 BlockDriverState *bs = NULL;
511
512 if (qemu_in_coroutine()) {
513 bdrv_co_yield_to_drain(NULL, true, NULL, true);
514 return;
515 }
516
517 /*
518 * bdrv queue is managed by record/replay,
519 * waiting for finishing the I/O requests may
520 * be infinite
521 */
522 if (replay_events_enabled()) {
523 return;
524 }
525
526 bdrv_drain_all_begin_nopoll();
527
528 /* Now poll the in-flight requests */
529 AIO_WAIT_WHILE_UNLOCKED(NULL, bdrv_drain_all_poll());
530
531 while ((bs = bdrv_next_all_states(bs))) {
532 bdrv_drain_assert_idle(bs);
533 }
534 }
535
536 void bdrv_drain_all_end_quiesce(BlockDriverState *bs)
537 {
538 GLOBAL_STATE_CODE();
539
540 g_assert(bs->quiesce_counter > 0);
541 g_assert(!bs->refcnt);
542
543 while (bs->quiesce_counter) {
544 bdrv_do_drained_end(bs, NULL);
545 }
546 }
547
548 void bdrv_drain_all_end(void)
549 {
550 BlockDriverState *bs = NULL;
551 GLOBAL_STATE_CODE();
552
553 /*
554 * bdrv queue is managed by record/replay,
555 * waiting for finishing the I/O requests may
556 * be endless
557 */
558 if (replay_events_enabled()) {
559 return;
560 }
561
562 while ((bs = bdrv_next_all_states(bs))) {
563 bdrv_do_drained_end(bs, NULL);
564 }
565
566 assert(qemu_get_current_aio_context() == qemu_get_aio_context());
567 assert(bdrv_drain_all_count > 0);
568 bdrv_drain_all_count--;
569 }
570
571 void bdrv_drain_all(void)
572 {
573 GLOBAL_STATE_CODE();
574 bdrv_drain_all_begin();
575 bdrv_drain_all_end();
576 }
577
578 /**
579 * Remove an active request from the tracked requests list
580 *
581 * This function should be called when a tracked request is completing.
582 */
583 static void coroutine_fn tracked_request_end(BdrvTrackedRequest *req)
584 {
585 if (req->serialising) {
586 qatomic_dec(&req->bs->serialising_in_flight);
587 }
588
589 qemu_mutex_lock(&req->bs->reqs_lock);
590 QLIST_REMOVE(req, list);
591 qemu_mutex_unlock(&req->bs->reqs_lock);
592
593 /*
594 * At this point qemu_co_queue_wait(&req->wait_queue, ...) won't be called
595 * anymore because the request has been removed from the list, so it's safe
596 * to restart the queue outside reqs_lock to minimize the critical section.
597 */
598 qemu_co_queue_restart_all(&req->wait_queue);
599 }
600
601 /**
602 * Add an active request to the tracked requests list
603 */
604 static void coroutine_fn tracked_request_begin(BdrvTrackedRequest *req,
605 BlockDriverState *bs,
606 int64_t offset,
607 int64_t bytes,
608 enum BdrvTrackedRequestType type)
609 {
610 bdrv_check_request(offset, bytes, &error_abort);
611
612 *req = (BdrvTrackedRequest){
613 .bs = bs,
614 .offset = offset,
615 .bytes = bytes,
616 .type = type,
617 .co = qemu_coroutine_self(),
618 .serialising = false,
619 .overlap_offset = offset,
620 .overlap_bytes = bytes,
621 };
622
623 qemu_co_queue_init(&req->wait_queue);
624
625 qemu_mutex_lock(&bs->reqs_lock);
626 QLIST_INSERT_HEAD(&bs->tracked_requests, req, list);
627 qemu_mutex_unlock(&bs->reqs_lock);
628 }
629
630 static bool tracked_request_overlaps(BdrvTrackedRequest *req,
631 int64_t offset, int64_t bytes)
632 {
633 bdrv_check_request(offset, bytes, &error_abort);
634
635 /* aaaa bbbb */
636 if (offset >= req->overlap_offset + req->overlap_bytes) {
637 return false;
638 }
639 /* bbbb aaaa */
640 if (req->overlap_offset >= offset + bytes) {
641 return false;
642 }
643 return true;
644 }
645
646 /* Called with self->bs->reqs_lock held */
647 static coroutine_fn BdrvTrackedRequest *
648 bdrv_find_conflicting_request(BdrvTrackedRequest *self)
649 {
650 BdrvTrackedRequest *req;
651
652 QLIST_FOREACH(req, &self->bs->tracked_requests, list) {
653 if (req == self || (!req->serialising && !self->serialising)) {
654 continue;
655 }
656 if (tracked_request_overlaps(req, self->overlap_offset,
657 self->overlap_bytes))
658 {
659 /*
660 * Hitting this means there was a reentrant request, for
661 * example, a block driver issuing nested requests. This must
662 * never happen since it means deadlock.
663 */
664 assert(qemu_coroutine_self() != req->co);
665
666 /*
667 * If the request is already (indirectly) waiting for us, or
668 * will wait for us as soon as it wakes up, then just go on
669 * (instead of producing a deadlock in the former case).
670 */
671 if (!req->waiting_for) {
672 return req;
673 }
674 }
675 }
676
677 return NULL;
678 }
679
680 /* Called with self->bs->reqs_lock held */
681 static void coroutine_fn
682 bdrv_wait_serialising_requests_locked(BdrvTrackedRequest *self)
683 {
684 BdrvTrackedRequest *req;
685
686 while ((req = bdrv_find_conflicting_request(self))) {
687 self->waiting_for = req;
688 qemu_co_queue_wait(&req->wait_queue, &self->bs->reqs_lock);
689 self->waiting_for = NULL;
690 }
691 }
692
693 /* Called with req->bs->reqs_lock held */
694 static void tracked_request_set_serialising(BdrvTrackedRequest *req,
695 uint64_t align)
696 {
697 int64_t overlap_offset = req->offset & ~(align - 1);
698 int64_t overlap_bytes =
699 ROUND_UP(req->offset + req->bytes, align) - overlap_offset;
700
701 bdrv_check_request(req->offset, req->bytes, &error_abort);
702
703 if (!req->serialising) {
704 qatomic_inc(&req->bs->serialising_in_flight);
705 req->serialising = true;
706 }
707
708 req->overlap_offset = MIN(req->overlap_offset, overlap_offset);
709 req->overlap_bytes = MAX(req->overlap_bytes, overlap_bytes);
710 }
711
712 /**
713 * Return the tracked request on @bs for the current coroutine, or
714 * NULL if there is none.
715 */
716 BdrvTrackedRequest *coroutine_fn bdrv_co_get_self_request(BlockDriverState *bs)
717 {
718 BdrvTrackedRequest *req;
719 Coroutine *self = qemu_coroutine_self();
720 IO_CODE();
721
722 QEMU_LOCK_GUARD(&bs->reqs_lock);
723 QLIST_FOREACH(req, &bs->tracked_requests, list) {
724 if (req->co == self) {
725 return req;
726 }
727 }
728
729 return NULL;
730 }
731
732 /**
733 * Round a region to subcluster (if supported) or cluster boundaries
734 */
735 void coroutine_fn GRAPH_RDLOCK
736 bdrv_round_to_subclusters(BlockDriverState *bs, int64_t offset, int64_t bytes,
737 int64_t *align_offset, int64_t *align_bytes)
738 {
739 BlockDriverInfo bdi;
740 IO_CODE();
741 if (bdrv_co_get_info(bs, &bdi) < 0 || bdi.subcluster_size == 0) {
742 *align_offset = offset;
743 *align_bytes = bytes;
744 } else {
745 int64_t c = bdi.subcluster_size;
746 *align_offset = QEMU_ALIGN_DOWN(offset, c);
747 *align_bytes = QEMU_ALIGN_UP(offset - *align_offset + bytes, c);
748 }
749 }
750
751 static int coroutine_fn GRAPH_RDLOCK bdrv_get_cluster_size(BlockDriverState *bs)
752 {
753 BlockDriverInfo bdi;
754 int ret;
755
756 ret = bdrv_co_get_info(bs, &bdi);
757 if (ret < 0 || bdi.cluster_size == 0) {
758 return bs->bl.request_alignment;
759 } else {
760 return bdi.cluster_size;
761 }
762 }
763
764 void bdrv_inc_in_flight(BlockDriverState *bs)
765 {
766 IO_CODE();
767 qatomic_inc(&bs->in_flight);
768 }
769
770 void bdrv_wakeup(BlockDriverState *bs)
771 {
772 IO_CODE();
773 aio_wait_kick();
774 }
775
776 void bdrv_dec_in_flight(BlockDriverState *bs)
777 {
778 IO_CODE();
779 qatomic_dec(&bs->in_flight);
780 bdrv_wakeup(bs);
781 }
782
783 static void coroutine_fn
784 bdrv_wait_serialising_requests(BdrvTrackedRequest *self)
785 {
786 BlockDriverState *bs = self->bs;
787
788 if (!qatomic_read(&bs->serialising_in_flight)) {
789 return;
790 }
791
792 qemu_mutex_lock(&bs->reqs_lock);
793 bdrv_wait_serialising_requests_locked(self);
794 qemu_mutex_unlock(&bs->reqs_lock);
795 }
796
797 void coroutine_fn bdrv_make_request_serialising(BdrvTrackedRequest *req,
798 uint64_t align)
799 {
800 IO_CODE();
801
802 qemu_mutex_lock(&req->bs->reqs_lock);
803
804 tracked_request_set_serialising(req, align);
805 bdrv_wait_serialising_requests_locked(req);
806
807 qemu_mutex_unlock(&req->bs->reqs_lock);
808 }
809
810 int bdrv_check_qiov_request(int64_t offset, int64_t bytes,
811 QEMUIOVector *qiov, size_t qiov_offset,
812 Error **errp)
813 {
814 /*
815 * Check generic offset/bytes correctness
816 */
817
818 if (offset < 0) {
819 error_setg(errp, "offset is negative: %" PRIi64, offset);
820 return -EIO;
821 }
822
823 if (bytes < 0) {
824 error_setg(errp, "bytes is negative: %" PRIi64, bytes);
825 return -EIO;
826 }
827
828 if (bytes > BDRV_MAX_LENGTH) {
829 error_setg(errp, "bytes(%" PRIi64 ") exceeds maximum(%" PRIi64 ")",
830 bytes, BDRV_MAX_LENGTH);
831 return -EIO;
832 }
833
834 if (offset > BDRV_MAX_LENGTH) {
835 error_setg(errp, "offset(%" PRIi64 ") exceeds maximum(%" PRIi64 ")",
836 offset, BDRV_MAX_LENGTH);
837 return -EIO;
838 }
839
840 if (offset > BDRV_MAX_LENGTH - bytes) {
841 error_setg(errp, "sum of offset(%" PRIi64 ") and bytes(%" PRIi64 ") "
842 "exceeds maximum(%" PRIi64 ")", offset, bytes,
843 BDRV_MAX_LENGTH);
844 return -EIO;
845 }
846
847 if (!qiov) {
848 return 0;
849 }
850
851 /*
852 * Check qiov and qiov_offset
853 */
854
855 if (qiov_offset > qiov->size) {
856 error_setg(errp, "qiov_offset(%zu) overflow io vector size(%zu)",
857 qiov_offset, qiov->size);
858 return -EIO;
859 }
860
861 if (bytes > qiov->size - qiov_offset) {
862 error_setg(errp, "bytes(%" PRIi64 ") + qiov_offset(%zu) overflow io "
863 "vector size(%zu)", bytes, qiov_offset, qiov->size);
864 return -EIO;
865 }
866
867 return 0;
868 }
869
870 int bdrv_check_request(int64_t offset, int64_t bytes, Error **errp)
871 {
872 return bdrv_check_qiov_request(offset, bytes, NULL, 0, errp);
873 }
874
875 static int bdrv_check_request32(int64_t offset, int64_t bytes,
876 QEMUIOVector *qiov, size_t qiov_offset)
877 {
878 int ret = bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, NULL);
879 if (ret < 0) {
880 return ret;
881 }
882
883 if (bytes > BDRV_REQUEST_MAX_BYTES) {
884 return -EIO;
885 }
886
887 return 0;
888 }
889
890 /*
891 * Completely zero out a block device with the help of bdrv_pwrite_zeroes.
892 * The operation is sped up by checking the block status and only writing
893 * zeroes to the device if they currently do not return zeroes. Optional
894 * flags are passed through to bdrv_pwrite_zeroes (e.g. BDRV_REQ_MAY_UNMAP,
895 * BDRV_REQ_FUA).
896 *
897 * Returns < 0 on error, 0 on success. For error codes see bdrv_pwrite().
898 */
899 int bdrv_make_zero(BdrvChild *child, BdrvRequestFlags flags)
900 {
901 int ret;
902 int64_t target_size, bytes, offset = 0;
903 BlockDriverState *bs = child->bs;
904 IO_CODE();
905
906 target_size = bdrv_getlength(bs);
907 if (target_size < 0) {
908 return target_size;
909 }
910
911 for (;;) {
912 bytes = MIN(target_size - offset, BDRV_REQUEST_MAX_BYTES);
913 if (bytes <= 0) {
914 return 0;
915 }
916 ret = bdrv_block_status(bs, offset, bytes, &bytes, NULL, NULL);
917 if (ret < 0) {
918 return ret;
919 }
920 if (ret & BDRV_BLOCK_ZERO) {
921 offset += bytes;
922 continue;
923 }
924 ret = bdrv_pwrite_zeroes(child, offset, bytes, flags);
925 if (ret < 0) {
926 return ret;
927 }
928 offset += bytes;
929 }
930 }
931
932 /*
933 * Writes to the file and ensures that no writes are reordered across this
934 * request (acts as a barrier)
935 *
936 * Returns 0 on success, -errno in error cases.
937 */
938 int coroutine_fn bdrv_co_pwrite_sync(BdrvChild *child, int64_t offset,
939 int64_t bytes, const void *buf,
940 BdrvRequestFlags flags)
941 {
942 int ret;
943 IO_CODE();
944 assert_bdrv_graph_readable();
945
946 ret = bdrv_co_pwrite(child, offset, bytes, buf, flags);
947 if (ret < 0) {
948 return ret;
949 }
950
951 ret = bdrv_co_flush(child->bs);
952 if (ret < 0) {
953 return ret;
954 }
955
956 return 0;
957 }
958
959 typedef struct CoroutineIOCompletion {
960 Coroutine *coroutine;
961 int ret;
962 } CoroutineIOCompletion;
963
964 static void bdrv_co_io_em_complete(void *opaque, int ret)
965 {
966 CoroutineIOCompletion *co = opaque;
967
968 co->ret = ret;
969 aio_co_wake(co->coroutine);
970 }
971
972 static int coroutine_fn GRAPH_RDLOCK
973 bdrv_driver_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
974 QEMUIOVector *qiov, size_t qiov_offset, int flags)
975 {
976 BlockDriver *drv = bs->drv;
977 int64_t sector_num;
978 unsigned int nb_sectors;
979 QEMUIOVector local_qiov;
980 int ret;
981 assert_bdrv_graph_readable();
982
983 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort);
984 assert(!(flags & ~bs->supported_read_flags));
985
986 if (!drv) {
987 return -ENOMEDIUM;
988 }
989
990 if (drv->bdrv_co_preadv_part) {
991 return drv->bdrv_co_preadv_part(bs, offset, bytes, qiov, qiov_offset,
992 flags);
993 }
994
995 if (qiov_offset > 0 || bytes != qiov->size) {
996 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes);
997 qiov = &local_qiov;
998 }
999
1000 if (drv->bdrv_co_preadv) {
1001 ret = drv->bdrv_co_preadv(bs, offset, bytes, qiov, flags);
1002 goto out;
1003 }
1004
1005 if (drv->bdrv_aio_preadv) {
1006 BlockAIOCB *acb;
1007 CoroutineIOCompletion co = {
1008 .coroutine = qemu_coroutine_self(),
1009 };
1010
1011 acb = drv->bdrv_aio_preadv(bs, offset, bytes, qiov, flags,
1012 bdrv_co_io_em_complete, &co);
1013 if (acb == NULL) {
1014 ret = -EIO;
1015 goto out;
1016 } else {
1017 qemu_coroutine_yield();
1018 ret = co.ret;
1019 goto out;
1020 }
1021 }
1022
1023 sector_num = offset >> BDRV_SECTOR_BITS;
1024 nb_sectors = bytes >> BDRV_SECTOR_BITS;
1025
1026 assert(QEMU_IS_ALIGNED(offset, BDRV_SECTOR_SIZE));
1027 assert(QEMU_IS_ALIGNED(bytes, BDRV_SECTOR_SIZE));
1028 assert(bytes <= BDRV_REQUEST_MAX_BYTES);
1029 assert(drv->bdrv_co_readv);
1030
1031 ret = drv->bdrv_co_readv(bs, sector_num, nb_sectors, qiov);
1032
1033 out:
1034 if (qiov == &local_qiov) {
1035 qemu_iovec_destroy(&local_qiov);
1036 }
1037
1038 return ret;
1039 }
1040
1041 static int coroutine_fn GRAPH_RDLOCK
1042 bdrv_driver_pwritev(BlockDriverState *bs, int64_t offset, int64_t bytes,
1043 QEMUIOVector *qiov, size_t qiov_offset,
1044 BdrvRequestFlags flags)
1045 {
1046 BlockDriver *drv = bs->drv;
1047 bool emulate_fua = false;
1048 int64_t sector_num;
1049 unsigned int nb_sectors;
1050 QEMUIOVector local_qiov;
1051 int ret;
1052 assert_bdrv_graph_readable();
1053
1054 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort);
1055
1056 if (!drv) {
1057 return -ENOMEDIUM;
1058 }
1059
1060 if (bs->open_flags & BDRV_O_NO_FLUSH) {
1061 flags &= ~BDRV_REQ_FUA;
1062 }
1063
1064 if ((flags & BDRV_REQ_FUA) &&
1065 (~bs->supported_write_flags & BDRV_REQ_FUA)) {
1066 flags &= ~BDRV_REQ_FUA;
1067 emulate_fua = true;
1068 }
1069
1070 flags &= bs->supported_write_flags;
1071
1072 if (drv->bdrv_co_pwritev_part) {
1073 ret = drv->bdrv_co_pwritev_part(bs, offset, bytes, qiov, qiov_offset,
1074 flags);
1075 goto emulate_flags;
1076 }
1077
1078 if (qiov_offset > 0 || bytes != qiov->size) {
1079 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes);
1080 qiov = &local_qiov;
1081 }
1082
1083 if (drv->bdrv_co_pwritev) {
1084 ret = drv->bdrv_co_pwritev(bs, offset, bytes, qiov, flags);
1085 goto emulate_flags;
1086 }
1087
1088 if (drv->bdrv_aio_pwritev) {
1089 BlockAIOCB *acb;
1090 CoroutineIOCompletion co = {
1091 .coroutine = qemu_coroutine_self(),
1092 };
1093
1094 acb = drv->bdrv_aio_pwritev(bs, offset, bytes, qiov, flags,
1095 bdrv_co_io_em_complete, &co);
1096 if (acb == NULL) {
1097 ret = -EIO;
1098 } else {
1099 qemu_coroutine_yield();
1100 ret = co.ret;
1101 }
1102 goto emulate_flags;
1103 }
1104
1105 sector_num = offset >> BDRV_SECTOR_BITS;
1106 nb_sectors = bytes >> BDRV_SECTOR_BITS;
1107
1108 assert(QEMU_IS_ALIGNED(offset, BDRV_SECTOR_SIZE));
1109 assert(QEMU_IS_ALIGNED(bytes, BDRV_SECTOR_SIZE));
1110 assert(bytes <= BDRV_REQUEST_MAX_BYTES);
1111
1112 assert(drv->bdrv_co_writev);
1113 ret = drv->bdrv_co_writev(bs, sector_num, nb_sectors, qiov, flags);
1114
1115 emulate_flags:
1116 if (ret == 0 && emulate_fua) {
1117 ret = bdrv_co_flush(bs);
1118 }
1119
1120 if (qiov == &local_qiov) {
1121 qemu_iovec_destroy(&local_qiov);
1122 }
1123
1124 return ret;
1125 }
1126
1127 static int coroutine_fn GRAPH_RDLOCK
1128 bdrv_driver_pwritev_compressed(BlockDriverState *bs, int64_t offset,
1129 int64_t bytes, QEMUIOVector *qiov,
1130 size_t qiov_offset)
1131 {
1132 BlockDriver *drv = bs->drv;
1133 QEMUIOVector local_qiov;
1134 int ret;
1135 assert_bdrv_graph_readable();
1136
1137 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort);
1138
1139 if (!drv) {
1140 return -ENOMEDIUM;
1141 }
1142
1143 if (!block_driver_can_compress(drv)) {
1144 return -ENOTSUP;
1145 }
1146
1147 if (drv->bdrv_co_pwritev_compressed_part) {
1148 return drv->bdrv_co_pwritev_compressed_part(bs, offset, bytes,
1149 qiov, qiov_offset);
1150 }
1151
1152 if (qiov_offset == 0) {
1153 return drv->bdrv_co_pwritev_compressed(bs, offset, bytes, qiov);
1154 }
1155
1156 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes);
1157 ret = drv->bdrv_co_pwritev_compressed(bs, offset, bytes, &local_qiov);
1158 qemu_iovec_destroy(&local_qiov);
1159
1160 return ret;
1161 }
1162
1163 static int coroutine_fn GRAPH_RDLOCK
1164 bdrv_co_do_copy_on_readv(BdrvChild *child, int64_t offset, int64_t bytes,
1165 QEMUIOVector *qiov, size_t qiov_offset, int flags)
1166 {
1167 BlockDriverState *bs = child->bs;
1168
1169 /* Perform I/O through a temporary buffer so that users who scribble over
1170 * their read buffer while the operation is in progress do not end up
1171 * modifying the image file. This is critical for zero-copy guest I/O
1172 * where anything might happen inside guest memory.
1173 */
1174 void *bounce_buffer = NULL;
1175
1176 BlockDriver *drv = bs->drv;
1177 int64_t align_offset;
1178 int64_t align_bytes;
1179 int64_t skip_bytes;
1180 int ret;
1181 int max_transfer = MIN_NON_ZERO(bs->bl.max_transfer,
1182 BDRV_REQUEST_MAX_BYTES);
1183 int64_t progress = 0;
1184 bool skip_write;
1185
1186 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort);
1187
1188 if (!drv) {
1189 return -ENOMEDIUM;
1190 }
1191
1192 /*
1193 * Do not write anything when the BDS is inactive. That is not
1194 * allowed, and it would not help.
1195 */
1196 skip_write = (bs->open_flags & BDRV_O_INACTIVE);
1197
1198 /* FIXME We cannot require callers to have write permissions when all they
1199 * are doing is a read request. If we did things right, write permissions
1200 * would be obtained anyway, but internally by the copy-on-read code. As
1201 * long as it is implemented here rather than in a separate filter driver,
1202 * the copy-on-read code doesn't have its own BdrvChild, however, for which
1203 * it could request permissions. Therefore we have to bypass the permission
1204 * system for the moment. */
1205 // assert(child->perm & (BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE));
1206
1207 /* Cover entire cluster so no additional backing file I/O is required when
1208 * allocating cluster in the image file. Note that this value may exceed
1209 * BDRV_REQUEST_MAX_BYTES (even when the original read did not), which
1210 * is one reason we loop rather than doing it all at once.
1211 */
1212 bdrv_round_to_subclusters(bs, offset, bytes, &align_offset, &align_bytes);
1213 skip_bytes = offset - align_offset;
1214
1215 trace_bdrv_co_do_copy_on_readv(bs, offset, bytes,
1216 align_offset, align_bytes);
1217
1218 while (align_bytes) {
1219 int64_t pnum;
1220
1221 if (skip_write) {
1222 ret = 1; /* "already allocated", so nothing will be copied */
1223 pnum = MIN(align_bytes, max_transfer);
1224 } else {
1225 ret = bdrv_co_is_allocated(bs, align_offset,
1226 MIN(align_bytes, max_transfer), &pnum);
1227 if (ret < 0) {
1228 /*
1229 * Safe to treat errors in querying allocation as if
1230 * unallocated; we'll probably fail again soon on the
1231 * read, but at least that will set a decent errno.
1232 */
1233 pnum = MIN(align_bytes, max_transfer);
1234 }
1235
1236 /* Stop at EOF if the image ends in the middle of the cluster */
1237 if (ret == 0 && pnum == 0) {
1238 assert(progress >= bytes);
1239 break;
1240 }
1241
1242 assert(skip_bytes < pnum);
1243 }
1244
1245 if (ret <= 0) {
1246 QEMUIOVector local_qiov;
1247
1248 /* Must copy-on-read; use the bounce buffer */
1249 pnum = MIN(pnum, MAX_BOUNCE_BUFFER);
1250 if (!bounce_buffer) {
1251 int64_t max_we_need = MAX(pnum, align_bytes - pnum);
1252 int64_t max_allowed = MIN(max_transfer, MAX_BOUNCE_BUFFER);
1253 int64_t bounce_buffer_len = MIN(max_we_need, max_allowed);
1254
1255 bounce_buffer = qemu_try_blockalign(bs, bounce_buffer_len);
1256 if (!bounce_buffer) {
1257 ret = -ENOMEM;
1258 goto err;
1259 }
1260 }
1261 qemu_iovec_init_buf(&local_qiov, bounce_buffer, pnum);
1262
1263 ret = bdrv_driver_preadv(bs, align_offset, pnum,
1264 &local_qiov, 0, 0);
1265 if (ret < 0) {
1266 goto err;
1267 }
1268
1269 bdrv_co_debug_event(bs, BLKDBG_COR_WRITE);
1270 if (drv->bdrv_co_pwrite_zeroes &&
1271 buffer_is_zero(bounce_buffer, pnum)) {
1272 /* FIXME: Should we (perhaps conditionally) be setting
1273 * BDRV_REQ_MAY_UNMAP, if it will allow for a sparser copy
1274 * that still correctly reads as zero? */
1275 ret = bdrv_co_do_pwrite_zeroes(bs, align_offset, pnum,
1276 BDRV_REQ_WRITE_UNCHANGED);
1277 } else {
1278 /* This does not change the data on the disk, it is not
1279 * necessary to flush even in cache=writethrough mode.
1280 */
1281 ret = bdrv_driver_pwritev(bs, align_offset, pnum,
1282 &local_qiov, 0,
1283 BDRV_REQ_WRITE_UNCHANGED);
1284 }
1285
1286 if (ret < 0) {
1287 /* It might be okay to ignore write errors for guest
1288 * requests. If this is a deliberate copy-on-read
1289 * then we don't want to ignore the error. Simply
1290 * report it in all cases.
1291 */
1292 goto err;
1293 }
1294
1295 if (!(flags & BDRV_REQ_PREFETCH)) {
1296 qemu_iovec_from_buf(qiov, qiov_offset + progress,
1297 bounce_buffer + skip_bytes,
1298 MIN(pnum - skip_bytes, bytes - progress));
1299 }
1300 } else if (!(flags & BDRV_REQ_PREFETCH)) {
1301 /* Read directly into the destination */
1302 ret = bdrv_driver_preadv(bs, offset + progress,
1303 MIN(pnum - skip_bytes, bytes - progress),
1304 qiov, qiov_offset + progress, 0);
1305 if (ret < 0) {
1306 goto err;
1307 }
1308 }
1309
1310 align_offset += pnum;
1311 align_bytes -= pnum;
1312 progress += pnum - skip_bytes;
1313 skip_bytes = 0;
1314 }
1315 ret = 0;
1316
1317 err:
1318 qemu_vfree(bounce_buffer);
1319 return ret;
1320 }
1321
1322 /*
1323 * Forwards an already correctly aligned request to the BlockDriver. This
1324 * handles copy on read, zeroing after EOF, and fragmentation of large
1325 * reads; any other features must be implemented by the caller.
1326 */
1327 static int coroutine_fn GRAPH_RDLOCK
1328 bdrv_aligned_preadv(BdrvChild *child, BdrvTrackedRequest *req,
1329 int64_t offset, int64_t bytes, int64_t align,
1330 QEMUIOVector *qiov, size_t qiov_offset, int flags)
1331 {
1332 BlockDriverState *bs = child->bs;
1333 int64_t total_bytes, max_bytes;
1334 int ret = 0;
1335 int64_t bytes_remaining = bytes;
1336 int max_transfer;
1337
1338 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort);
1339 assert(is_power_of_2(align));
1340 assert((offset & (align - 1)) == 0);
1341 assert((bytes & (align - 1)) == 0);
1342 assert((bs->open_flags & BDRV_O_NO_IO) == 0);
1343 max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX),
1344 align);
1345
1346 /*
1347 * TODO: We would need a per-BDS .supported_read_flags and
1348 * potential fallback support, if we ever implement any read flags
1349 * to pass through to drivers. For now, there aren't any
1350 * passthrough flags except the BDRV_REQ_REGISTERED_BUF optimization hint.
1351 */
1352 assert(!(flags & ~(BDRV_REQ_COPY_ON_READ | BDRV_REQ_PREFETCH |
1353 BDRV_REQ_REGISTERED_BUF)));
1354
1355 /* Handle Copy on Read and associated serialisation */
1356 if (flags & BDRV_REQ_COPY_ON_READ) {
1357 /* If we touch the same cluster it counts as an overlap. This
1358 * guarantees that allocating writes will be serialized and not race
1359 * with each other for the same cluster. For example, in copy-on-read
1360 * it ensures that the CoR read and write operations are atomic and
1361 * guest writes cannot interleave between them. */
1362 bdrv_make_request_serialising(req, bdrv_get_cluster_size(bs));
1363 } else {
1364 bdrv_wait_serialising_requests(req);
1365 }
1366
1367 if (flags & BDRV_REQ_COPY_ON_READ) {
1368 int64_t pnum;
1369
1370 /* The flag BDRV_REQ_COPY_ON_READ has reached its addressee */
1371 flags &= ~BDRV_REQ_COPY_ON_READ;
1372
1373 ret = bdrv_co_is_allocated(bs, offset, bytes, &pnum);
1374 if (ret < 0) {
1375 goto out;
1376 }
1377
1378 if (!ret || pnum != bytes) {
1379 ret = bdrv_co_do_copy_on_readv(child, offset, bytes,
1380 qiov, qiov_offset, flags);
1381 goto out;
1382 } else if (flags & BDRV_REQ_PREFETCH) {
1383 goto out;
1384 }
1385 }
1386
1387 /* Forward the request to the BlockDriver, possibly fragmenting it */
1388 total_bytes = bdrv_co_getlength(bs);
1389 if (total_bytes < 0) {
1390 ret = total_bytes;
1391 goto out;
1392 }
1393
1394 assert(!(flags & ~(bs->supported_read_flags | BDRV_REQ_REGISTERED_BUF)));
1395
1396 max_bytes = ROUND_UP(MAX(0, total_bytes - offset), align);
1397 if (bytes <= max_bytes && bytes <= max_transfer) {
1398 ret = bdrv_driver_preadv(bs, offset, bytes, qiov, qiov_offset, flags);
1399 goto out;
1400 }
1401
1402 while (bytes_remaining) {
1403 int64_t num;
1404
1405 if (max_bytes) {
1406 num = MIN(bytes_remaining, MIN(max_bytes, max_transfer));
1407 assert(num);
1408
1409 ret = bdrv_driver_preadv(bs, offset + bytes - bytes_remaining,
1410 num, qiov,
1411 qiov_offset + bytes - bytes_remaining,
1412 flags);
1413 max_bytes -= num;
1414 } else {
1415 num = bytes_remaining;
1416 ret = qemu_iovec_memset(qiov, qiov_offset + bytes - bytes_remaining,
1417 0, bytes_remaining);
1418 }
1419 if (ret < 0) {
1420 goto out;
1421 }
1422 bytes_remaining -= num;
1423 }
1424
1425 out:
1426 return ret < 0 ? ret : 0;
1427 }
1428
1429 /*
1430 * Request padding
1431 *
1432 * |<---- align ----->| |<----- align ---->|
1433 * |<- head ->|<------------- bytes ------------->|<-- tail -->|
1434 * | | | | | |
1435 * -*----------$-------*-------- ... --------*-----$------------*---
1436 * | | | | | |
1437 * | offset | | end |
1438 * ALIGN_DOWN(offset) ALIGN_UP(offset) ALIGN_DOWN(end) ALIGN_UP(end)
1439 * [buf ... ) [tail_buf )
1440 *
1441 * @buf is an aligned allocation needed to store @head and @tail paddings. @head
1442 * is placed at the beginning of @buf and @tail at the @end.
1443 *
1444 * @tail_buf is a pointer to sub-buffer, corresponding to align-sized chunk
1445 * around tail, if tail exists.
1446 *
1447 * @merge_reads is true for small requests,
1448 * if @buf_len == @head + bytes + @tail. In this case it is possible that both
1449 * head and tail exist but @buf_len == align and @tail_buf == @buf.
1450 *
1451 * @write is true for write requests, false for read requests.
1452 *
1453 * If padding makes the vector too long (exceeding IOV_MAX), then we need to
1454 * merge existing vector elements into a single one. @collapse_bounce_buf acts
1455 * as the bounce buffer in such cases. @pre_collapse_qiov has the pre-collapse
1456 * I/O vector elements so for read requests, the data can be copied back after
1457 * the read is done.
1458 */
1459 typedef struct BdrvRequestPadding {
1460 uint8_t *buf;
1461 size_t buf_len;
1462 uint8_t *tail_buf;
1463 size_t head;
1464 size_t tail;
1465 bool merge_reads;
1466 bool write;
1467 QEMUIOVector local_qiov;
1468
1469 uint8_t *collapse_bounce_buf;
1470 size_t collapse_len;
1471 QEMUIOVector pre_collapse_qiov;
1472 } BdrvRequestPadding;
1473
1474 static bool bdrv_init_padding(BlockDriverState *bs,
1475 int64_t offset, int64_t bytes,
1476 bool write,
1477 BdrvRequestPadding *pad)
1478 {
1479 int64_t align = bs->bl.request_alignment;
1480 int64_t sum;
1481
1482 bdrv_check_request(offset, bytes, &error_abort);
1483 assert(align <= INT_MAX); /* documented in block/block_int.h */
1484 assert(align <= SIZE_MAX / 2); /* so we can allocate the buffer */
1485
1486 memset(pad, 0, sizeof(*pad));
1487
1488 pad->head = offset & (align - 1);
1489 pad->tail = ((offset + bytes) & (align - 1));
1490 if (pad->tail) {
1491 pad->tail = align - pad->tail;
1492 }
1493
1494 if (!pad->head && !pad->tail) {
1495 return false;
1496 }
1497
1498 assert(bytes); /* Nothing good in aligning zero-length requests */
1499
1500 sum = pad->head + bytes + pad->tail;
1501 pad->buf_len = (sum > align && pad->head && pad->tail) ? 2 * align : align;
1502 pad->buf = qemu_blockalign(bs, pad->buf_len);
1503 pad->merge_reads = sum == pad->buf_len;
1504 if (pad->tail) {
1505 pad->tail_buf = pad->buf + pad->buf_len - align;
1506 }
1507
1508 pad->write = write;
1509
1510 return true;
1511 }
1512
1513 static int coroutine_fn GRAPH_RDLOCK
1514 bdrv_padding_rmw_read(BdrvChild *child, BdrvTrackedRequest *req,
1515 BdrvRequestPadding *pad, bool zero_middle)
1516 {
1517 QEMUIOVector local_qiov;
1518 BlockDriverState *bs = child->bs;
1519 uint64_t align = bs->bl.request_alignment;
1520 int ret;
1521
1522 assert(req->serialising && pad->buf);
1523
1524 if (pad->head || pad->merge_reads) {
1525 int64_t bytes = pad->merge_reads ? pad->buf_len : align;
1526
1527 qemu_iovec_init_buf(&local_qiov, pad->buf, bytes);
1528
1529 if (pad->head) {
1530 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD);
1531 }
1532 if (pad->merge_reads && pad->tail) {
1533 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
1534 }
1535 ret = bdrv_aligned_preadv(child, req, req->overlap_offset, bytes,
1536 align, &local_qiov, 0, 0);
1537 if (ret < 0) {
1538 return ret;
1539 }
1540 if (pad->head) {
1541 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD);
1542 }
1543 if (pad->merge_reads && pad->tail) {
1544 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
1545 }
1546
1547 if (pad->merge_reads) {
1548 goto zero_mem;
1549 }
1550 }
1551
1552 if (pad->tail) {
1553 qemu_iovec_init_buf(&local_qiov, pad->tail_buf, align);
1554
1555 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
1556 ret = bdrv_aligned_preadv(
1557 child, req,
1558 req->overlap_offset + req->overlap_bytes - align,
1559 align, align, &local_qiov, 0, 0);
1560 if (ret < 0) {
1561 return ret;
1562 }
1563 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
1564 }
1565
1566 zero_mem:
1567 if (zero_middle) {
1568 memset(pad->buf + pad->head, 0, pad->buf_len - pad->head - pad->tail);
1569 }
1570
1571 return 0;
1572 }
1573
1574 /**
1575 * Free *pad's associated buffers, and perform any necessary finalization steps.
1576 */
1577 static void bdrv_padding_finalize(BdrvRequestPadding *pad)
1578 {
1579 if (pad->collapse_bounce_buf) {
1580 if (!pad->write) {
1581 /*
1582 * If padding required elements in the vector to be collapsed into a
1583 * bounce buffer, copy the bounce buffer content back
1584 */
1585 qemu_iovec_from_buf(&pad->pre_collapse_qiov, 0,
1586 pad->collapse_bounce_buf, pad->collapse_len);
1587 }
1588 qemu_vfree(pad->collapse_bounce_buf);
1589 qemu_iovec_destroy(&pad->pre_collapse_qiov);
1590 }
1591 if (pad->buf) {
1592 qemu_vfree(pad->buf);
1593 qemu_iovec_destroy(&pad->local_qiov);
1594 }
1595 memset(pad, 0, sizeof(*pad));
1596 }
1597
1598 /*
1599 * Create pad->local_qiov by wrapping @iov in the padding head and tail, while
1600 * ensuring that the resulting vector will not exceed IOV_MAX elements.
1601 *
1602 * To ensure this, when necessary, the first two or three elements of @iov are
1603 * merged into pad->collapse_bounce_buf and replaced by a reference to that
1604 * bounce buffer in pad->local_qiov.
1605 *
1606 * After performing a read request, the data from the bounce buffer must be
1607 * copied back into pad->pre_collapse_qiov (e.g. by bdrv_padding_finalize()).
1608 */
1609 static int bdrv_create_padded_qiov(BlockDriverState *bs,
1610 BdrvRequestPadding *pad,
1611 struct iovec *iov, int niov,
1612 size_t iov_offset, size_t bytes)
1613 {
1614 int padded_niov, surplus_count, collapse_count;
1615
1616 /* Assert this invariant */
1617 assert(niov <= IOV_MAX);
1618
1619 /*
1620 * Cannot pad if resulting length would exceed SIZE_MAX. Returning an error
1621 * to the guest is not ideal, but there is little else we can do. At least
1622 * this will practically never happen on 64-bit systems.
1623 */
1624 if (SIZE_MAX - pad->head < bytes ||
1625 SIZE_MAX - pad->head - bytes < pad->tail)
1626 {
1627 return -EINVAL;
1628 }
1629
1630 /* Length of the resulting IOV if we just concatenated everything */
1631 padded_niov = !!pad->head + niov + !!pad->tail;
1632
1633 qemu_iovec_init(&pad->local_qiov, MIN(padded_niov, IOV_MAX));
1634
1635 if (pad->head) {
1636 qemu_iovec_add(&pad->local_qiov, pad->buf, pad->head);
1637 }
1638
1639 /*
1640 * If padded_niov > IOV_MAX, we cannot just concatenate everything.
1641 * Instead, merge the first two or three elements of @iov to reduce the
1642 * number of vector elements as necessary.
1643 */
1644 if (padded_niov > IOV_MAX) {
1645 /*
1646 * Only head and tail can have lead to the number of entries exceeding
1647 * IOV_MAX, so we can exceed it by the head and tail at most. We need
1648 * to reduce the number of elements by `surplus_count`, so we merge that
1649 * many elements plus one into one element.
1650 */
1651 surplus_count = padded_niov - IOV_MAX;
1652 assert(surplus_count <= !!pad->head + !!pad->tail);
1653 collapse_count = surplus_count + 1;
1654
1655 /*
1656 * Move the elements to collapse into `pad->pre_collapse_qiov`, then
1657 * advance `iov` (and associated variables) by those elements.
1658 */
1659 qemu_iovec_init(&pad->pre_collapse_qiov, collapse_count);
1660 qemu_iovec_concat_iov(&pad->pre_collapse_qiov, iov,
1661 collapse_count, iov_offset, SIZE_MAX);
1662 iov += collapse_count;
1663 iov_offset = 0;
1664 niov -= collapse_count;
1665 bytes -= pad->pre_collapse_qiov.size;
1666
1667 /*
1668 * Construct the bounce buffer to match the length of the to-collapse
1669 * vector elements, and for write requests, initialize it with the data
1670 * from those elements. Then add it to `pad->local_qiov`.
1671 */
1672 pad->collapse_len = pad->pre_collapse_qiov.size;
1673 pad->collapse_bounce_buf = qemu_blockalign(bs, pad->collapse_len);
1674 if (pad->write) {
1675 qemu_iovec_to_buf(&pad->pre_collapse_qiov, 0,
1676 pad->collapse_bounce_buf, pad->collapse_len);
1677 }
1678 qemu_iovec_add(&pad->local_qiov,
1679 pad->collapse_bounce_buf, pad->collapse_len);
1680 }
1681
1682 qemu_iovec_concat_iov(&pad->local_qiov, iov, niov, iov_offset, bytes);
1683
1684 if (pad->tail) {
1685 qemu_iovec_add(&pad->local_qiov,
1686 pad->buf + pad->buf_len - pad->tail, pad->tail);
1687 }
1688
1689 assert(pad->local_qiov.niov == MIN(padded_niov, IOV_MAX));
1690 return 0;
1691 }
1692
1693 /*
1694 * bdrv_pad_request
1695 *
1696 * Exchange request parameters with padded request if needed. Don't include RMW
1697 * read of padding, bdrv_padding_rmw_read() should be called separately if
1698 * needed.
1699 *
1700 * @write is true for write requests, false for read requests.
1701 *
1702 * Request parameters (@qiov, &qiov_offset, &offset, &bytes) are in-out:
1703 * - on function start they represent original request
1704 * - on failure or when padding is not needed they are unchanged
1705 * - on success when padding is needed they represent padded request
1706 */
1707 static int bdrv_pad_request(BlockDriverState *bs,
1708 QEMUIOVector **qiov, size_t *qiov_offset,
1709 int64_t *offset, int64_t *bytes,
1710 bool write,
1711 BdrvRequestPadding *pad, bool *padded,
1712 BdrvRequestFlags *flags)
1713 {
1714 int ret;
1715 struct iovec *sliced_iov;
1716 int sliced_niov;
1717 size_t sliced_head, sliced_tail;
1718
1719 /* Should have been checked by the caller already */
1720 ret = bdrv_check_request32(*offset, *bytes, *qiov, *qiov_offset);
1721 if (ret < 0) {
1722 return ret;
1723 }
1724
1725 if (!bdrv_init_padding(bs, *offset, *bytes, write, pad)) {
1726 if (padded) {
1727 *padded = false;
1728 }
1729 return 0;
1730 }
1731
1732 /*
1733 * For prefetching in stream_populate(), no qiov is passed along, because
1734 * only copy-on-read matters.
1735 */
1736 if (*qiov) {
1737 sliced_iov = qemu_iovec_slice(*qiov, *qiov_offset, *bytes,
1738 &sliced_head, &sliced_tail,
1739 &sliced_niov);
1740
1741 /* Guaranteed by bdrv_check_request32() */
1742 assert(*bytes <= SIZE_MAX);
1743 ret = bdrv_create_padded_qiov(bs, pad, sliced_iov, sliced_niov,
1744 sliced_head, *bytes);
1745 if (ret < 0) {
1746 bdrv_padding_finalize(pad);
1747 return ret;
1748 }
1749 *qiov = &pad->local_qiov;
1750 *qiov_offset = 0;
1751 }
1752
1753 *bytes += pad->head + pad->tail;
1754 *offset -= pad->head;
1755 if (padded) {
1756 *padded = true;
1757 }
1758 if (flags) {
1759 /* Can't use optimization hint with bounce buffer */
1760 *flags &= ~BDRV_REQ_REGISTERED_BUF;
1761 }
1762
1763 return 0;
1764 }
1765
1766 int coroutine_fn bdrv_co_preadv(BdrvChild *child,
1767 int64_t offset, int64_t bytes, QEMUIOVector *qiov,
1768 BdrvRequestFlags flags)
1769 {
1770 IO_CODE();
1771 return bdrv_co_preadv_part(child, offset, bytes, qiov, 0, flags);
1772 }
1773
1774 int coroutine_fn bdrv_co_preadv_part(BdrvChild *child,
1775 int64_t offset, int64_t bytes,
1776 QEMUIOVector *qiov, size_t qiov_offset,
1777 BdrvRequestFlags flags)
1778 {
1779 BlockDriverState *bs = child->bs;
1780 BdrvTrackedRequest req;
1781 BdrvRequestPadding pad;
1782 int ret;
1783 IO_CODE();
1784
1785 trace_bdrv_co_preadv_part(bs, offset, bytes, flags);
1786
1787 if (!bdrv_co_is_inserted(bs)) {
1788 return -ENOMEDIUM;
1789 }
1790
1791 ret = bdrv_check_request32(offset, bytes, qiov, qiov_offset);
1792 if (ret < 0) {
1793 return ret;
1794 }
1795
1796 if (bytes == 0 && !QEMU_IS_ALIGNED(offset, bs->bl.request_alignment)) {
1797 /*
1798 * Aligning zero request is nonsense. Even if driver has special meaning
1799 * of zero-length (like qcow2_co_pwritev_compressed_part), we can't pass
1800 * it to driver due to request_alignment.
1801 *
1802 * Still, no reason to return an error if someone do unaligned
1803 * zero-length read occasionally.
1804 */
1805 return 0;
1806 }
1807
1808 bdrv_inc_in_flight(bs);
1809
1810 /* Don't do copy-on-read if we read data before write operation */
1811 if (qatomic_read(&bs->copy_on_read)) {
1812 flags |= BDRV_REQ_COPY_ON_READ;
1813 }
1814
1815 ret = bdrv_pad_request(bs, &qiov, &qiov_offset, &offset, &bytes, false,
1816 &pad, NULL, &flags);
1817 if (ret < 0) {
1818 goto fail;
1819 }
1820
1821 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_READ);
1822 ret = bdrv_aligned_preadv(child, &req, offset, bytes,
1823 bs->bl.request_alignment,
1824 qiov, qiov_offset, flags);
1825 tracked_request_end(&req);
1826 bdrv_padding_finalize(&pad);
1827
1828 fail:
1829 bdrv_dec_in_flight(bs);
1830
1831 return ret;
1832 }
1833
1834 static int coroutine_fn GRAPH_RDLOCK
1835 bdrv_co_do_pwrite_zeroes(BlockDriverState *bs, int64_t offset, int64_t bytes,
1836 BdrvRequestFlags flags)
1837 {
1838 BlockDriver *drv = bs->drv;
1839 QEMUIOVector qiov;
1840 void *buf = NULL;
1841 int ret = 0;
1842 bool need_flush = false;
1843 int head = 0;
1844 int tail = 0;
1845
1846 int64_t max_write_zeroes = MIN_NON_ZERO(bs->bl.max_pwrite_zeroes,
1847 INT64_MAX);
1848 int alignment = MAX(bs->bl.pwrite_zeroes_alignment,
1849 bs->bl.request_alignment);
1850 int max_transfer = MIN_NON_ZERO(bs->bl.max_transfer, MAX_BOUNCE_BUFFER);
1851
1852 assert_bdrv_graph_readable();
1853 bdrv_check_request(offset, bytes, &error_abort);
1854
1855 if (!drv) {
1856 return -ENOMEDIUM;
1857 }
1858
1859 if ((flags & ~bs->supported_zero_flags) & BDRV_REQ_NO_FALLBACK) {
1860 return -ENOTSUP;
1861 }
1862
1863 /* By definition there is no user buffer so this flag doesn't make sense */
1864 if (flags & BDRV_REQ_REGISTERED_BUF) {
1865 return -EINVAL;
1866 }
1867
1868 /* If opened with discard=off we should never unmap. */
1869 if (!(bs->open_flags & BDRV_O_UNMAP)) {
1870 flags &= ~BDRV_REQ_MAY_UNMAP;
1871 }
1872
1873 /* Invalidate the cached block-status data range if this write overlaps */
1874 bdrv_bsc_invalidate_range(bs, offset, bytes);
1875
1876 assert(alignment % bs->bl.request_alignment == 0);
1877 head = offset % alignment;
1878 tail = (offset + bytes) % alignment;
1879 max_write_zeroes = QEMU_ALIGN_DOWN(max_write_zeroes, alignment);
1880 assert(max_write_zeroes >= bs->bl.request_alignment);
1881
1882 while (bytes > 0 && !ret) {
1883 int64_t num = bytes;
1884
1885 /* Align request. Block drivers can expect the "bulk" of the request
1886 * to be aligned, and that unaligned requests do not cross cluster
1887 * boundaries.
1888 */
1889 if (head) {
1890 /* Make a small request up to the first aligned sector. For
1891 * convenience, limit this request to max_transfer even if
1892 * we don't need to fall back to writes. */
1893 num = MIN(MIN(bytes, max_transfer), alignment - head);
1894 head = (head + num) % alignment;
1895 assert(num < max_write_zeroes);
1896 } else if (tail && num > alignment) {
1897 /* Shorten the request to the last aligned sector. */
1898 num -= tail;
1899 }
1900
1901 /* limit request size */
1902 if (num > max_write_zeroes) {
1903 num = max_write_zeroes;
1904 }
1905
1906 ret = -ENOTSUP;
1907 /* First try the efficient write zeroes operation */
1908 if (drv->bdrv_co_pwrite_zeroes) {
1909 ret = drv->bdrv_co_pwrite_zeroes(bs, offset, num,
1910 flags & bs->supported_zero_flags);
1911 if (ret != -ENOTSUP && (flags & BDRV_REQ_FUA) &&
1912 !(bs->supported_zero_flags & BDRV_REQ_FUA)) {
1913 need_flush = true;
1914 }
1915 } else {
1916 assert(!bs->supported_zero_flags);
1917 }
1918
1919 /*
1920 * TODO The ret == -EINVAL && num < alignment case is a workaround for
1921 * when request_alignment is 1 on files with cache=writeback. The Linux
1922 * ioctl(BLKZEROOUT) requires block alignment and will fail with
1923 * EINVAL. The block layer should align the request to
1924 * write_zeroes_alignment instead of trying the syscall, failing, and
1925 * falling back to a bounce buffer. Doing that is not easy so for now
1926 * we use a bounce buffer:
1927 * https://lore.kernel.org/qemu-devel/20260109120837.2772961-1-f.ebner@proxmox.com/
1928 */
1929 if ((ret == -ENOTSUP || (ret == -EINVAL && num < alignment)) &&
1930 !(flags & BDRV_REQ_NO_FALLBACK)) {
1931 /* Fall back to bounce buffer if write zeroes is unsupported */
1932 BdrvRequestFlags write_flags = flags & ~BDRV_REQ_ZERO_WRITE;
1933
1934 if ((flags & BDRV_REQ_FUA) &&
1935 !(bs->supported_write_flags & BDRV_REQ_FUA)) {
1936 /* No need for bdrv_driver_pwrite() to do a fallback
1937 * flush on each chunk; use just one at the end */
1938 write_flags &= ~BDRV_REQ_FUA;
1939 need_flush = true;
1940 }
1941 num = MIN(num, max_transfer);
1942 if (buf == NULL) {
1943 buf = qemu_try_blockalign0(bs, num);
1944 if (buf == NULL) {
1945 ret = -ENOMEM;
1946 goto fail;
1947 }
1948 }
1949 qemu_iovec_init_buf(&qiov, buf, num);
1950
1951 ret = bdrv_driver_pwritev(bs, offset, num, &qiov, 0, write_flags);
1952
1953 /* Keep bounce buffer around if it is big enough for all
1954 * all future requests.
1955 */
1956 if (num < max_transfer) {
1957 qemu_vfree(buf);
1958 buf = NULL;
1959 }
1960 }
1961
1962 offset += num;
1963 bytes -= num;
1964 }
1965
1966 fail:
1967 if (ret == 0 && need_flush) {
1968 ret = bdrv_co_flush(bs);
1969 }
1970 qemu_vfree(buf);
1971 return ret;
1972 }
1973
1974 static inline int coroutine_fn GRAPH_RDLOCK
1975 bdrv_co_write_req_prepare(BdrvChild *child, int64_t offset, int64_t bytes,
1976 BdrvTrackedRequest *req, int flags)
1977 {
1978 BlockDriverState *bs = child->bs;
1979
1980 bdrv_check_request(offset, bytes, &error_abort);
1981
1982 if (bdrv_is_read_only(bs)) {
1983 return -EPERM;
1984 }
1985
1986 assert(!(bs->open_flags & BDRV_O_INACTIVE));
1987 assert((bs->open_flags & BDRV_O_NO_IO) == 0);
1988 assert(!(flags & ~BDRV_REQ_MASK));
1989 assert(!((flags & BDRV_REQ_NO_WAIT) && !(flags & BDRV_REQ_SERIALISING)));
1990
1991 if (flags & BDRV_REQ_SERIALISING) {
1992 QEMU_LOCK_GUARD(&bs->reqs_lock);
1993
1994 tracked_request_set_serialising(req, bdrv_get_cluster_size(bs));
1995
1996 if ((flags & BDRV_REQ_NO_WAIT) && bdrv_find_conflicting_request(req)) {
1997 return -EBUSY;
1998 }
1999
2000 bdrv_wait_serialising_requests_locked(req);
2001 } else {
2002 bdrv_wait_serialising_requests(req);
2003 }
2004
2005 assert(req->overlap_offset <= offset);
2006 assert(offset + bytes <= req->overlap_offset + req->overlap_bytes);
2007 assert(offset + bytes <= bs->total_sectors * BDRV_SECTOR_SIZE ||
2008 child->perm & BLK_PERM_RESIZE);
2009
2010 switch (req->type) {
2011 case BDRV_TRACKED_WRITE:
2012 case BDRV_TRACKED_DISCARD:
2013 if (flags & BDRV_REQ_WRITE_UNCHANGED) {
2014 assert(child->perm & (BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE));
2015 } else {
2016 assert(child->perm & BLK_PERM_WRITE);
2017 }
2018 bdrv_write_threshold_check_write(bs, offset, bytes);
2019 return 0;
2020 case BDRV_TRACKED_TRUNCATE:
2021 assert(child->perm & BLK_PERM_RESIZE);
2022 return 0;
2023 default:
2024 abort();
2025 }
2026 }
2027
2028 static inline void coroutine_fn GRAPH_RDLOCK
2029 bdrv_co_write_req_finish(BdrvChild *child, int64_t offset, int64_t bytes,
2030 BdrvTrackedRequest *req, int ret)
2031 {
2032 int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);
2033 BlockDriverState *bs = child->bs;
2034
2035 bdrv_check_request(offset, bytes, &error_abort);
2036
2037 qatomic_inc(&bs->write_gen);
2038
2039 /*
2040 * Discard cannot extend the image, but in error handling cases, such as
2041 * when reverting a qcow2 cluster allocation, the discarded range can pass
2042 * the end of image file, so we cannot assert about BDRV_TRACKED_DISCARD
2043 * here. Instead, just skip it, since semantically a discard request
2044 * beyond EOF cannot expand the image anyway.
2045 */
2046 if (ret == 0 &&
2047 (req->type == BDRV_TRACKED_TRUNCATE ||
2048 end_sector > bs->total_sectors) &&
2049 req->type != BDRV_TRACKED_DISCARD) {
2050 bs->total_sectors = end_sector;
2051 bdrv_co_parent_cb_resize(bs);
2052 bdrv_dirty_bitmap_truncate(bs, end_sector << BDRV_SECTOR_BITS);
2053 }
2054 if (req->bytes) {
2055 switch (req->type) {
2056 case BDRV_TRACKED_WRITE:
2057 {
2058 uint64_t new = offset + bytes;
2059 uint64_t old = qatomic_read(&bs->wr_highest_offset);
2060
2061 while (old < new) {
2062 old = qatomic_cmpxchg(&bs->wr_highest_offset, old, new);
2063 }
2064 }
2065 /* fall through, to set dirty bits */
2066 case BDRV_TRACKED_DISCARD:
2067 bdrv_set_dirty(bs, offset, bytes);
2068 break;
2069 default:
2070 break;
2071 }
2072 }
2073 }
2074
2075 /*
2076 * Forwards an already correctly aligned write request to the BlockDriver,
2077 * after possibly fragmenting it.
2078 */
2079 static int coroutine_fn GRAPH_RDLOCK
2080 bdrv_aligned_pwritev(BdrvChild *child, BdrvTrackedRequest *req,
2081 int64_t offset, int64_t bytes, int64_t align,
2082 QEMUIOVector *qiov, size_t qiov_offset,
2083 BdrvRequestFlags flags)
2084 {
2085 BlockDriverState *bs = child->bs;
2086 BlockDriver *drv = bs->drv;
2087 int ret;
2088
2089 int64_t bytes_remaining = bytes;
2090 int max_transfer;
2091
2092 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort);
2093
2094 if (!drv) {
2095 return -ENOMEDIUM;
2096 }
2097
2098 if (bdrv_has_readonly_bitmaps(bs)) {
2099 return -EPERM;
2100 }
2101
2102 assert(is_power_of_2(align));
2103 assert((offset & (align - 1)) == 0);
2104 assert((bytes & (align - 1)) == 0);
2105 max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX),
2106 align);
2107
2108 ret = bdrv_co_write_req_prepare(child, offset, bytes, req, flags);
2109
2110 if (!ret && bs->detect_zeroes != BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF &&
2111 !(flags & BDRV_REQ_ZERO_WRITE) && drv->bdrv_co_pwrite_zeroes &&
2112 qemu_iovec_is_zero(qiov, qiov_offset, bytes)) {
2113 flags |= BDRV_REQ_ZERO_WRITE;
2114 if (bs->detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP) {
2115 flags |= BDRV_REQ_MAY_UNMAP;
2116 }
2117
2118 /* Can't use optimization hint with bufferless zero write */
2119 flags &= ~BDRV_REQ_REGISTERED_BUF;
2120 }
2121
2122 if (ret < 0) {
2123 /* Do nothing, write notifier decided to fail this request */
2124 } else if (flags & BDRV_REQ_ZERO_WRITE) {
2125 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_ZERO);
2126 ret = bdrv_co_do_pwrite_zeroes(bs, offset, bytes, flags);
2127 } else if (flags & BDRV_REQ_WRITE_COMPRESSED) {
2128 ret = bdrv_driver_pwritev_compressed(bs, offset, bytes,
2129 qiov, qiov_offset);
2130 } else if (bytes <= max_transfer) {
2131 bdrv_co_debug_event(bs, BLKDBG_PWRITEV);
2132 ret = bdrv_driver_pwritev(bs, offset, bytes, qiov, qiov_offset, flags);
2133 } else {
2134 bdrv_co_debug_event(bs, BLKDBG_PWRITEV);
2135 while (bytes_remaining) {
2136 int num = MIN(bytes_remaining, max_transfer);
2137 int local_flags = flags;
2138
2139 assert(num);
2140 if (num < bytes_remaining && (flags & BDRV_REQ_FUA) &&
2141 !(bs->supported_write_flags & BDRV_REQ_FUA)) {
2142 /* If FUA is going to be emulated by flush, we only
2143 * need to flush on the last iteration */
2144 local_flags &= ~BDRV_REQ_FUA;
2145 }
2146
2147 ret = bdrv_driver_pwritev(bs, offset + bytes - bytes_remaining,
2148 num, qiov,
2149 qiov_offset + bytes - bytes_remaining,
2150 local_flags);
2151 if (ret < 0) {
2152 break;
2153 }
2154 bytes_remaining -= num;
2155 }
2156 }
2157 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_DONE);
2158
2159 if (ret >= 0) {
2160 ret = 0;
2161 }
2162 bdrv_co_write_req_finish(child, offset, bytes, req, ret);
2163
2164 return ret;
2165 }
2166
2167 static int coroutine_fn GRAPH_RDLOCK
2168 bdrv_co_do_zero_pwritev(BdrvChild *child, int64_t offset, int64_t bytes,
2169 BdrvRequestFlags flags, BdrvTrackedRequest *req)
2170 {
2171 BlockDriverState *bs = child->bs;
2172 QEMUIOVector local_qiov;
2173 uint64_t align = bs->bl.request_alignment;
2174 int ret = 0;
2175 bool padding;
2176 BdrvRequestPadding pad;
2177
2178 /* This flag doesn't make sense for padding or zero writes */
2179 flags &= ~BDRV_REQ_REGISTERED_BUF;
2180
2181 padding = bdrv_init_padding(bs, offset, bytes, true, &pad);
2182 if (padding) {
2183 assert(!(flags & BDRV_REQ_NO_WAIT));
2184 bdrv_make_request_serialising(req, align);
2185
2186 bdrv_padding_rmw_read(child, req, &pad, true);
2187
2188 if (pad.head || pad.merge_reads) {
2189 int64_t aligned_offset = offset & ~(align - 1);
2190 int64_t write_bytes = pad.merge_reads ? pad.buf_len : align;
2191
2192 qemu_iovec_init_buf(&local_qiov, pad.buf, write_bytes);
2193 ret = bdrv_aligned_pwritev(child, req, aligned_offset, write_bytes,
2194 align, &local_qiov, 0,
2195 flags & ~BDRV_REQ_ZERO_WRITE);
2196 if (ret < 0 || pad.merge_reads) {
2197 /* Error or all work is done */
2198 goto out;
2199 }
2200 offset += write_bytes - pad.head;
2201 bytes -= write_bytes - pad.head;
2202 }
2203 }
2204
2205 assert(!bytes || (offset & (align - 1)) == 0);
2206 if (bytes >= align) {
2207 /* Write the aligned part in the middle. */
2208 int64_t aligned_bytes = bytes & ~(align - 1);
2209 ret = bdrv_aligned_pwritev(child, req, offset, aligned_bytes, align,
2210 NULL, 0, flags);
2211 if (ret < 0) {
2212 goto out;
2213 }
2214 bytes -= aligned_bytes;
2215 offset += aligned_bytes;
2216 }
2217
2218 assert(!bytes || (offset & (align - 1)) == 0);
2219 if (bytes) {
2220 assert(align == pad.tail + bytes);
2221
2222 qemu_iovec_init_buf(&local_qiov, pad.tail_buf, align);
2223 ret = bdrv_aligned_pwritev(child, req, offset, align, align,
2224 &local_qiov, 0,
2225 flags & ~BDRV_REQ_ZERO_WRITE);
2226 }
2227
2228 out:
2229 bdrv_padding_finalize(&pad);
2230
2231 return ret;
2232 }
2233
2234 /*
2235 * Handle a write request in coroutine context
2236 */
2237 int coroutine_fn bdrv_co_pwritev(BdrvChild *child,
2238 int64_t offset, int64_t bytes, QEMUIOVector *qiov,
2239 BdrvRequestFlags flags)
2240 {
2241 IO_CODE();
2242 return bdrv_co_pwritev_part(child, offset, bytes, qiov, 0, flags);
2243 }
2244
2245 int coroutine_fn bdrv_co_pwritev_part(BdrvChild *child,
2246 int64_t offset, int64_t bytes, QEMUIOVector *qiov, size_t qiov_offset,
2247 BdrvRequestFlags flags)
2248 {
2249 BlockDriverState *bs = child->bs;
2250 BdrvTrackedRequest req;
2251 uint64_t align = bs->bl.request_alignment;
2252 BdrvRequestPadding pad;
2253 int ret;
2254 bool padded = false;
2255 IO_CODE();
2256
2257 trace_bdrv_co_pwritev_part(child->bs, offset, bytes, flags);
2258
2259 if (!bdrv_co_is_inserted(bs)) {
2260 return -ENOMEDIUM;
2261 }
2262
2263 if (flags & BDRV_REQ_ZERO_WRITE) {
2264 ret = bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, NULL);
2265 } else {
2266 ret = bdrv_check_request32(offset, bytes, qiov, qiov_offset);
2267 }
2268 if (ret < 0) {
2269 return ret;
2270 }
2271
2272 /* If the request is misaligned then we can't make it efficient */
2273 if ((flags & BDRV_REQ_NO_FALLBACK) &&
2274 !QEMU_IS_ALIGNED(offset | bytes, align))
2275 {
2276 return -ENOTSUP;
2277 }
2278
2279 if (bytes == 0 && !QEMU_IS_ALIGNED(offset, bs->bl.request_alignment)) {
2280 /*
2281 * Aligning zero request is nonsense. Even if driver has special meaning
2282 * of zero-length (like qcow2_co_pwritev_compressed_part), we can't pass
2283 * it to driver due to request_alignment.
2284 *
2285 * Still, no reason to return an error if someone do unaligned
2286 * zero-length write occasionally.
2287 */
2288 return 0;
2289 }
2290
2291 if (!(flags & BDRV_REQ_ZERO_WRITE)) {
2292 /*
2293 * Pad request for following read-modify-write cycle.
2294 * bdrv_co_do_zero_pwritev() does aligning by itself, so, we do
2295 * alignment only if there is no ZERO flag.
2296 */
2297 ret = bdrv_pad_request(bs, &qiov, &qiov_offset, &offset, &bytes, true,
2298 &pad, &padded, &flags);
2299 if (ret < 0) {
2300 return ret;
2301 }
2302 }
2303
2304 bdrv_inc_in_flight(bs);
2305 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_WRITE);
2306
2307 if (flags & BDRV_REQ_ZERO_WRITE) {
2308 assert(!padded);
2309 ret = bdrv_co_do_zero_pwritev(child, offset, bytes, flags, &req);
2310 goto out;
2311 }
2312
2313 if (padded) {
2314 /*
2315 * Request was unaligned to request_alignment and therefore
2316 * padded. We are going to do read-modify-write, and must
2317 * serialize the request to prevent interactions of the
2318 * widened region with other transactions.
2319 */
2320 assert(!(flags & BDRV_REQ_NO_WAIT));
2321 bdrv_make_request_serialising(&req, align);
2322 bdrv_padding_rmw_read(child, &req, &pad, false);
2323 }
2324
2325 ret = bdrv_aligned_pwritev(child, &req, offset, bytes, align,
2326 qiov, qiov_offset, flags);
2327
2328 bdrv_padding_finalize(&pad);
2329
2330 out:
2331 tracked_request_end(&req);
2332 bdrv_dec_in_flight(bs);
2333
2334 return ret;
2335 }
2336
2337 int coroutine_fn bdrv_co_pwrite_zeroes(BdrvChild *child, int64_t offset,
2338 int64_t bytes, BdrvRequestFlags flags)
2339 {
2340 IO_CODE();
2341 trace_bdrv_co_pwrite_zeroes(child->bs, offset, bytes, flags);
2342 assert_bdrv_graph_readable();
2343
2344 return bdrv_co_pwritev(child, offset, bytes, NULL,
2345 BDRV_REQ_ZERO_WRITE | flags);
2346 }
2347
2348 /*
2349 * Flush ALL BDSes regardless of if they are reachable via a BlkBackend or not.
2350 */
2351 int bdrv_flush_all(void)
2352 {
2353 BdrvNextIterator it;
2354 BlockDriverState *bs = NULL;
2355 int result = 0;
2356
2357 GLOBAL_STATE_CODE();
2358 GRAPH_RDLOCK_GUARD_MAINLOOP();
2359
2360 /*
2361 * bdrv queue is managed by record/replay,
2362 * creating new flush request for stopping
2363 * the VM may break the determinism
2364 */
2365 if (replay_events_enabled()) {
2366 return result;
2367 }
2368
2369 for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
2370 int ret = bdrv_flush(bs);
2371 if (ret < 0 && !result) {
2372 result = ret;
2373 }
2374 }
2375
2376 return result;
2377 }
2378
2379 /*
2380 * Returns the allocation status of the specified sectors.
2381 * Drivers not implementing the functionality are assumed to not support
2382 * backing files, hence all their sectors are reported as allocated.
2383 *
2384 * 'mode' serves as a hint as to which results are favored; see the
2385 * BDRV_WANT_* macros for details.
2386 *
2387 * If 'offset' is beyond the end of the disk image the return value is
2388 * BDRV_BLOCK_EOF and 'pnum' is set to 0.
2389 *
2390 * 'bytes' is the max value 'pnum' should be set to. If bytes goes
2391 * beyond the end of the disk image it will be clamped; if 'pnum' is set to
2392 * the end of the image, then the returned value will include BDRV_BLOCK_EOF.
2393 *
2394 * 'pnum' is set to the number of bytes (including and immediately
2395 * following the specified offset) that are easily known to be in the
2396 * same allocated/unallocated state. Note that a second call starting
2397 * at the original offset plus returned pnum may have the same status.
2398 * The returned value is non-zero on success except at end-of-file.
2399 *
2400 * Returns negative errno on failure. Otherwise, if the
2401 * BDRV_BLOCK_OFFSET_VALID bit is set, 'map' and 'file' (if non-NULL) are
2402 * set to the host mapping and BDS corresponding to the guest offset.
2403 */
2404 static int coroutine_fn GRAPH_RDLOCK
2405 bdrv_co_do_block_status(BlockDriverState *bs, unsigned int mode,
2406 int64_t offset, int64_t bytes,
2407 int64_t *pnum, int64_t *map, BlockDriverState **file)
2408 {
2409 int64_t total_size;
2410 int64_t n; /* bytes */
2411 int ret;
2412 int64_t local_map = 0;
2413 BlockDriverState *local_file = NULL;
2414 int64_t aligned_offset, aligned_bytes;
2415 uint32_t align;
2416 bool has_filtered_child;
2417
2418 assert(pnum);
2419 assert_bdrv_graph_readable();
2420 *pnum = 0;
2421 total_size = bdrv_co_getlength(bs);
2422 if (total_size < 0) {
2423 ret = total_size;
2424 goto early_out;
2425 }
2426
2427 if (offset >= total_size) {
2428 ret = BDRV_BLOCK_EOF;
2429 goto early_out;
2430 }
2431 if (!bytes) {
2432 ret = 0;
2433 goto early_out;
2434 }
2435
2436 n = total_size - offset;
2437 if (n < bytes) {
2438 bytes = n;
2439 }
2440
2441 /* Must be non-NULL or bdrv_co_getlength() would have failed */
2442 assert(bs->drv);
2443 has_filtered_child = bdrv_filter_child(bs);
2444 if (!bs->drv->bdrv_co_block_status && !has_filtered_child) {
2445 *pnum = bytes;
2446 ret = BDRV_BLOCK_DATA | BDRV_BLOCK_ALLOCATED;
2447 if (offset + bytes == total_size) {
2448 ret |= BDRV_BLOCK_EOF;
2449 }
2450 if (bs->drv->protocol_name) {
2451 ret |= BDRV_BLOCK_OFFSET_VALID;
2452 local_map = offset;
2453 local_file = bs;
2454 }
2455 goto early_out;
2456 }
2457
2458 bdrv_inc_in_flight(bs);
2459
2460 /* Round out to request_alignment boundaries */
2461 align = bs->bl.request_alignment;
2462 aligned_offset = QEMU_ALIGN_DOWN(offset, align);
2463 aligned_bytes = ROUND_UP(offset + bytes, align) - aligned_offset;
2464
2465 if (bs->drv->bdrv_co_block_status) {
2466 /*
2467 * Use the block-status cache only for protocol nodes: Format
2468 * drivers are generally quick to inquire the status, but protocol
2469 * drivers often need to get information from outside of qemu, so
2470 * we do not have control over the actual implementation. There
2471 * have been cases where inquiring the status took an unreasonably
2472 * long time, and we can do nothing in qemu to fix it.
2473 * This is especially problematic for images with large data areas,
2474 * because finding the few holes in them and giving them special
2475 * treatment does not gain much performance. Therefore, we try to
2476 * cache the last-identified data region.
2477 *
2478 * Second, limiting ourselves to protocol nodes allows us to assume
2479 * the block status for data regions to be DATA | OFFSET_VALID, and
2480 * that the host offset is the same as the guest offset.
2481 *
2482 * Note that it is possible that external writers zero parts of
2483 * the cached regions without the cache being invalidated, and so
2484 * we may report zeroes as data. This is not catastrophic,
2485 * however, because reporting zeroes as data is fine.
2486 */
2487 if (QLIST_EMPTY(&bs->children) &&
2488 bdrv_bsc_is_data(bs, aligned_offset, pnum))
2489 {
2490 ret = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
2491 local_file = bs;
2492 local_map = aligned_offset;
2493 } else {
2494 ret = bs->drv->bdrv_co_block_status(bs, mode, aligned_offset,
2495 aligned_bytes, pnum, &local_map,
2496 &local_file);
2497
2498 /*
2499 * Note that checking QLIST_EMPTY(&bs->children) is also done when
2500 * the cache is queried above. Technically, we do not need to check
2501 * it here; the worst that can happen is that we fill the cache for
2502 * non-protocol nodes, and then it is never used. However, filling
2503 * the cache requires an RCU update, so double check here to avoid
2504 * such an update if possible.
2505 *
2506 * Check mode, because we only want to update the cache when we
2507 * have accurate information about what is zero and what is data.
2508 */
2509 if (mode == BDRV_WANT_PRECISE &&
2510 ret == (BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID) &&
2511 QLIST_EMPTY(&bs->children))
2512 {
2513 /*
2514 * When a protocol driver reports BLOCK_OFFSET_VALID, the
2515 * returned local_map value must be the same as the offset we
2516 * have passed (aligned_offset), and local_bs must be the node
2517 * itself.
2518 * Assert this, because we follow this rule when reading from
2519 * the cache (see the `local_file = bs` and
2520 * `local_map = aligned_offset` assignments above), and the
2521 * result the cache delivers must be the same as the driver
2522 * would deliver.
2523 */
2524 assert(local_file == bs);
2525 assert(local_map == aligned_offset);
2526 bdrv_bsc_fill(bs, aligned_offset, *pnum);
2527 }
2528 }
2529 } else {
2530 /* Default code for filters */
2531
2532 local_file = bdrv_filter_bs(bs);
2533 assert(local_file);
2534
2535 *pnum = aligned_bytes;
2536 local_map = aligned_offset;
2537 ret = BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID;
2538 }
2539 if (ret < 0) {
2540 *pnum = 0;
2541 goto out;
2542 }
2543
2544 /*
2545 * The driver's result must be a non-zero multiple of request_alignment.
2546 * Clamp pnum and adjust map to original request.
2547 */
2548 assert(*pnum && QEMU_IS_ALIGNED(*pnum, align) &&
2549 align > offset - aligned_offset);
2550 if (ret & BDRV_BLOCK_RECURSE) {
2551 assert(ret & BDRV_BLOCK_DATA);
2552 assert(ret & BDRV_BLOCK_OFFSET_VALID);
2553 assert(!(ret & BDRV_BLOCK_ZERO));
2554 }
2555
2556 *pnum -= offset - aligned_offset;
2557 if (*pnum > bytes) {
2558 *pnum = bytes;
2559 }
2560 if (ret & BDRV_BLOCK_OFFSET_VALID) {
2561 local_map += offset - aligned_offset;
2562 }
2563
2564 if (ret & BDRV_BLOCK_RAW) {
2565 assert(ret & BDRV_BLOCK_OFFSET_VALID && local_file);
2566 ret = bdrv_co_do_block_status(local_file, mode, local_map,
2567 *pnum, pnum, &local_map, &local_file);
2568 goto out;
2569 }
2570
2571 if (ret & (BDRV_BLOCK_DATA | BDRV_BLOCK_ZERO)) {
2572 ret |= BDRV_BLOCK_ALLOCATED;
2573 } else if (bs->drv->supports_backing) {
2574 BlockDriverState *cow_bs = bdrv_cow_bs(bs);
2575
2576 if (!cow_bs) {
2577 ret |= BDRV_BLOCK_ZERO;
2578 } else if (mode == BDRV_WANT_PRECISE) {
2579 int64_t size2 = bdrv_co_getlength(cow_bs);
2580
2581 if (size2 >= 0 && offset >= size2) {
2582 ret |= BDRV_BLOCK_ZERO;
2583 }
2584 }
2585 }
2586
2587 if (mode == BDRV_WANT_PRECISE && ret & BDRV_BLOCK_RECURSE &&
2588 local_file && local_file != bs &&
2589 (ret & BDRV_BLOCK_DATA) && !(ret & BDRV_BLOCK_ZERO) &&
2590 (ret & BDRV_BLOCK_OFFSET_VALID)) {
2591 int64_t file_pnum;
2592 int ret2;
2593
2594 ret2 = bdrv_co_do_block_status(local_file, mode, local_map,
2595 *pnum, &file_pnum, NULL, NULL);
2596 if (ret2 >= 0) {
2597 /* Ignore errors. This is just providing extra information, it
2598 * is useful but not necessary.
2599 */
2600 if (ret2 & BDRV_BLOCK_EOF &&
2601 (!file_pnum || ret2 & BDRV_BLOCK_ZERO)) {
2602 /*
2603 * It is valid for the format block driver to read
2604 * beyond the end of the underlying file's current
2605 * size; such areas read as zero.
2606 */
2607 ret |= BDRV_BLOCK_ZERO;
2608 } else {
2609 /* Limit request to the range reported by the protocol driver */
2610 *pnum = file_pnum;
2611 ret |= (ret2 & BDRV_BLOCK_ZERO);
2612 }
2613 }
2614
2615 /*
2616 * Now that the recursive search was done, clear the flag. Otherwise,
2617 * with more complicated block graphs like snapshot-access ->
2618 * copy-before-write -> qcow2, where the return value will be propagated
2619 * further up to a parent bdrv_co_do_block_status() call, both the
2620 * BDRV_BLOCK_RECURSE and BDRV_BLOCK_ZERO flags would be set, which is
2621 * not allowed.
2622 */
2623 ret &= ~BDRV_BLOCK_RECURSE;
2624 }
2625
2626 out:
2627 bdrv_dec_in_flight(bs);
2628 if (ret >= 0 && offset + *pnum == total_size) {
2629 ret |= BDRV_BLOCK_EOF;
2630 }
2631 early_out:
2632 if (file) {
2633 *file = local_file;
2634 }
2635 if (map) {
2636 *map = local_map;
2637 }
2638 return ret;
2639 }
2640
2641 int coroutine_fn
2642 bdrv_co_common_block_status_above(BlockDriverState *bs,
2643 BlockDriverState *base,
2644 bool include_base,
2645 unsigned int mode,
2646 int64_t offset,
2647 int64_t bytes,
2648 int64_t *pnum,
2649 int64_t *map,
2650 BlockDriverState **file,
2651 int *depth)
2652 {
2653 int ret;
2654 BlockDriverState *p;
2655 int64_t eof = 0;
2656 int dummy;
2657 IO_CODE();
2658
2659 assert(!include_base || base); /* Can't include NULL base */
2660 assert_bdrv_graph_readable();
2661
2662 if (!depth) {
2663 depth = &dummy;
2664 }
2665 *depth = 0;
2666
2667 if (!include_base && bs == base) {
2668 *pnum = bytes;
2669 return 0;
2670 }
2671
2672 ret = bdrv_co_do_block_status(bs, mode, offset, bytes, pnum,
2673 map, file);
2674 ++*depth;
2675 if (ret < 0 || *pnum == 0 || ret & BDRV_BLOCK_ALLOCATED || bs == base) {
2676 return ret;
2677 }
2678
2679 if (ret & BDRV_BLOCK_EOF) {
2680 eof = offset + *pnum;
2681 }
2682
2683 assert(*pnum <= bytes);
2684 bytes = *pnum;
2685
2686 for (p = bdrv_filter_or_cow_bs(bs); include_base || p != base;
2687 p = bdrv_filter_or_cow_bs(p))
2688 {
2689 ret = bdrv_co_do_block_status(p, mode, offset, bytes, pnum,
2690 map, file);
2691 ++*depth;
2692 if (ret < 0) {
2693 return ret;
2694 }
2695 if (*pnum == 0) {
2696 /*
2697 * The top layer deferred to this layer, and because this layer is
2698 * short, any zeroes that we synthesize beyond EOF behave as if they
2699 * were allocated at this layer.
2700 *
2701 * We don't include BDRV_BLOCK_EOF into ret, as upper layer may be
2702 * larger. We'll add BDRV_BLOCK_EOF if needed at function end, see
2703 * below.
2704 */
2705 assert(ret & BDRV_BLOCK_EOF);
2706 *pnum = bytes;
2707 if (file) {
2708 *file = p;
2709 }
2710 ret = BDRV_BLOCK_ZERO | BDRV_BLOCK_ALLOCATED;
2711 break;
2712 }
2713 if (ret & BDRV_BLOCK_ALLOCATED) {
2714 /*
2715 * We've found the node and the status, we must break.
2716 *
2717 * Drop BDRV_BLOCK_EOF, as it's not for upper layer, which may be
2718 * larger. We'll add BDRV_BLOCK_EOF if needed at function end, see
2719 * below.
2720 */
2721 ret &= ~BDRV_BLOCK_EOF;
2722 break;
2723 }
2724
2725 if (p == base) {
2726 assert(include_base);
2727 break;
2728 }
2729
2730 /*
2731 * OK, [offset, offset + *pnum) region is unallocated on this layer,
2732 * let's continue the diving.
2733 */
2734 assert(*pnum <= bytes);
2735 bytes = *pnum;
2736 }
2737
2738 if (offset + *pnum == eof) {
2739 ret |= BDRV_BLOCK_EOF;
2740 }
2741
2742 return ret;
2743 }
2744
2745 int coroutine_fn bdrv_co_block_status_above(BlockDriverState *bs,
2746 BlockDriverState *base,
2747 int64_t offset, int64_t bytes,
2748 int64_t *pnum, int64_t *map,
2749 BlockDriverState **file)
2750 {
2751 IO_CODE();
2752 return bdrv_co_common_block_status_above(bs, base, false,
2753 BDRV_WANT_PRECISE, offset,
2754 bytes, pnum, map, file, NULL);
2755 }
2756
2757 int coroutine_fn bdrv_co_block_status(BlockDriverState *bs, int64_t offset,
2758 int64_t bytes, int64_t *pnum,
2759 int64_t *map, BlockDriverState **file)
2760 {
2761 IO_CODE();
2762 return bdrv_co_block_status_above(bs, bdrv_filter_or_cow_bs(bs),
2763 offset, bytes, pnum, map, file);
2764 }
2765
2766 /*
2767 * Check @bs (and its backing chain) to see if the range defined
2768 * by @offset and @bytes is known to read as zeroes.
2769 * Return 1 if that is the case, 0 otherwise and -errno on error.
2770 * This test is meant to be fast rather than accurate so returning 0
2771 * does not guarantee non-zero data; but a return of 1 is reliable.
2772 */
2773 int coroutine_fn bdrv_co_is_zero_fast(BlockDriverState *bs, int64_t offset,
2774 int64_t bytes)
2775 {
2776 int ret;
2777 int64_t pnum;
2778 IO_CODE();
2779
2780 while (bytes) {
2781 ret = bdrv_co_common_block_status_above(bs, NULL, false,
2782 BDRV_WANT_ZERO, offset, bytes,
2783 &pnum, NULL, NULL, NULL);
2784
2785 if (ret < 0) {
2786 return ret;
2787 }
2788 if (!(ret & BDRV_BLOCK_ZERO)) {
2789 return 0;
2790 }
2791 offset += pnum;
2792 bytes -= pnum;
2793 }
2794
2795 return 1;
2796 }
2797
2798 /*
2799 * Check @bs (and its backing chain) to see if the entire image is known
2800 * to read as zeroes.
2801 * Return 1 if that is the case, 0 otherwise and -errno on error.
2802 * This test is meant to be fast rather than accurate so returning 0
2803 * does not guarantee non-zero data; however, a return of 1 is reliable,
2804 * and this function can report 1 in more cases than bdrv_co_is_zero_fast.
2805 */
2806 int coroutine_fn bdrv_co_is_all_zeroes(BlockDriverState *bs)
2807 {
2808 int ret;
2809 int64_t pnum, bytes;
2810 char *buf;
2811 QEMUIOVector local_qiov;
2812 IO_CODE();
2813
2814 bytes = bdrv_co_getlength(bs);
2815 if (bytes < 0) {
2816 return bytes;
2817 }
2818
2819 /* First probe - see if the entire image reads as zero */
2820 ret = bdrv_co_common_block_status_above(bs, NULL, false, BDRV_WANT_ZERO,
2821 0, bytes, &pnum, NULL, NULL,
2822 NULL);
2823 if (ret < 0) {
2824 return ret;
2825 }
2826 if (ret & BDRV_BLOCK_ZERO) {
2827 return bdrv_co_is_zero_fast(bs, pnum, bytes - pnum);
2828 }
2829
2830 /*
2831 * Because of the way 'blockdev-create' works, raw files tend to
2832 * be created with a non-sparse region at the front to make
2833 * alignment probing easier. If the block starts with only a
2834 * small allocated region, it is still worth the effort to see if
2835 * the rest of the image is still sparse, coupled with manually
2836 * reading the first region to see if it reads zero after all.
2837 */
2838 if (pnum > MAX_ZERO_CHECK_BUFFER) {
2839 return 0;
2840 }
2841 ret = bdrv_co_is_zero_fast(bs, pnum, bytes - pnum);
2842 if (ret <= 0) {
2843 return ret;
2844 }
2845 /* Only the head of the image is unknown, and it's small. Read it. */
2846 buf = qemu_blockalign(bs, pnum);
2847 qemu_iovec_init_buf(&local_qiov, buf, pnum);
2848 ret = bdrv_driver_preadv(bs, 0, pnum, &local_qiov, 0, 0);
2849 if (ret >= 0) {
2850 ret = buffer_is_zero(buf, pnum);
2851 }
2852 qemu_vfree(buf);
2853 return ret;
2854 }
2855
2856 int coroutine_fn bdrv_co_is_allocated(BlockDriverState *bs, int64_t offset,
2857 int64_t bytes, int64_t *pnum)
2858 {
2859 int ret;
2860 int64_t dummy;
2861 IO_CODE();
2862
2863 ret = bdrv_co_common_block_status_above(bs, bs, true, BDRV_WANT_ALLOCATED,
2864 offset, bytes, pnum ? pnum : &dummy,
2865 NULL, NULL, NULL);
2866 if (ret < 0) {
2867 return ret;
2868 }
2869 return !!(ret & BDRV_BLOCK_ALLOCATED);
2870 }
2871
2872 /*
2873 * Given an image chain: ... -> [BASE] -> [INTER1] -> [INTER2] -> [TOP]
2874 *
2875 * Return a positive depth if (a prefix of) the given range is allocated
2876 * in any image between BASE and TOP (BASE is only included if include_base
2877 * is set). Depth 1 is TOP, 2 is the first backing layer, and so forth.
2878 * BASE can be NULL to check if the given offset is allocated in any
2879 * image of the chain. Return 0 otherwise, or negative errno on
2880 * failure.
2881 *
2882 * 'pnum' is set to the number of bytes (including and immediately
2883 * following the specified offset) that are known to be in the same
2884 * allocated/unallocated state. Note that a subsequent call starting
2885 * at 'offset + *pnum' may return the same allocation status (in other
2886 * words, the result is not necessarily the maximum possible range);
2887 * but 'pnum' will only be 0 when end of file is reached.
2888 */
2889 int coroutine_fn bdrv_co_is_allocated_above(BlockDriverState *bs,
2890 BlockDriverState *base,
2891 bool include_base, int64_t offset,
2892 int64_t bytes, int64_t *pnum)
2893 {
2894 int depth;
2895 int ret;
2896 IO_CODE();
2897
2898 ret = bdrv_co_common_block_status_above(bs, base, include_base,
2899 BDRV_WANT_ALLOCATED,
2900 offset, bytes, pnum, NULL, NULL,
2901 &depth);
2902 if (ret < 0) {
2903 return ret;
2904 }
2905
2906 if (ret & BDRV_BLOCK_ALLOCATED) {
2907 return depth;
2908 }
2909 return 0;
2910 }
2911
2912 int coroutine_fn
2913 bdrv_co_readv_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos)
2914 {
2915 BlockDriver *drv = bs->drv;
2916 BlockDriverState *child_bs = bdrv_primary_bs(bs);
2917 int ret;
2918 IO_CODE();
2919 assert_bdrv_graph_readable();
2920
2921 ret = bdrv_check_qiov_request(pos, qiov->size, qiov, 0, NULL);
2922 if (ret < 0) {
2923 return ret;
2924 }
2925
2926 if (!drv) {
2927 return -ENOMEDIUM;
2928 }
2929
2930 bdrv_inc_in_flight(bs);
2931
2932 if (drv->bdrv_co_load_vmstate) {
2933 ret = drv->bdrv_co_load_vmstate(bs, qiov, pos);
2934 } else if (child_bs) {
2935 ret = bdrv_co_readv_vmstate(child_bs, qiov, pos);
2936 } else {
2937 ret = -ENOTSUP;
2938 }
2939
2940 bdrv_dec_in_flight(bs);
2941
2942 return ret;
2943 }
2944
2945 int coroutine_fn
2946 bdrv_co_writev_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos)
2947 {
2948 BlockDriver *drv = bs->drv;
2949 BlockDriverState *child_bs = bdrv_primary_bs(bs);
2950 int ret;
2951 IO_CODE();
2952 assert_bdrv_graph_readable();
2953
2954 ret = bdrv_check_qiov_request(pos, qiov->size, qiov, 0, NULL);
2955 if (ret < 0) {
2956 return ret;
2957 }
2958
2959 if (!drv) {
2960 return -ENOMEDIUM;
2961 }
2962
2963 bdrv_inc_in_flight(bs);
2964
2965 if (drv->bdrv_co_save_vmstate) {
2966 ret = drv->bdrv_co_save_vmstate(bs, qiov, pos);
2967 } else if (child_bs) {
2968 ret = bdrv_co_writev_vmstate(child_bs, qiov, pos);
2969 } else {
2970 ret = -ENOTSUP;
2971 }
2972
2973 bdrv_dec_in_flight(bs);
2974
2975 return ret;
2976 }
2977
2978 int bdrv_save_vmstate(BlockDriverState *bs, const uint8_t *buf,
2979 int64_t pos, int size)
2980 {
2981 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, size);
2982 int ret = bdrv_writev_vmstate(bs, &qiov, pos);
2983 IO_CODE();
2984
2985 return ret < 0 ? ret : size;
2986 }
2987
2988 int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf,
2989 int64_t pos, int size)
2990 {
2991 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, size);
2992 int ret = bdrv_readv_vmstate(bs, &qiov, pos);
2993 IO_CODE();
2994
2995 return ret < 0 ? ret : size;
2996 }
2997
2998 /**************************************************************/
2999 /* async I/Os */
3000
3001 /**
3002 * Synchronously cancels an acb. Must be called with the BQL held and the acb
3003 * must be processed with the BQL held too (IOThreads are not allowed).
3004 *
3005 * Use bdrv_aio_cancel_async() instead when possible.
3006 */
3007 void bdrv_aio_cancel(BlockAIOCB *acb)
3008 {
3009 GLOBAL_STATE_CODE();
3010 qemu_aio_ref(acb);
3011 bdrv_aio_cancel_async(acb);
3012 AIO_WAIT_WHILE_UNLOCKED(NULL, acb->refcnt > 1);
3013 qemu_aio_unref(acb);
3014 }
3015
3016 /* Async version of aio cancel. The caller is not blocked if the acb implements
3017 * cancel_async, otherwise we do nothing and let the request normally complete.
3018 * In either case the completion callback must be called. */
3019 void bdrv_aio_cancel_async(BlockAIOCB *acb)
3020 {
3021 IO_CODE();
3022 if (acb->aiocb_info->cancel_async) {
3023 acb->aiocb_info->cancel_async(acb);
3024 }
3025 }
3026
3027 /**************************************************************/
3028 /* Coroutine block device emulation */
3029
3030 int coroutine_fn bdrv_co_flush(BlockDriverState *bs)
3031 {
3032 BdrvChild *primary_child = bdrv_primary_child(bs);
3033 BdrvChild *child;
3034 int current_gen;
3035 int ret = 0;
3036 IO_CODE();
3037
3038 assert_bdrv_graph_readable();
3039 bdrv_inc_in_flight(bs);
3040
3041 if (!bdrv_co_is_inserted(bs) || bdrv_is_read_only(bs) ||
3042 bdrv_is_sg(bs)) {
3043 goto early_exit;
3044 }
3045
3046 qemu_mutex_lock(&bs->reqs_lock);
3047 current_gen = qatomic_read(&bs->write_gen);
3048
3049 /* Wait until any previous flushes are completed */
3050 while (bs->active_flush_req) {
3051 qemu_co_queue_wait(&bs->flush_queue, &bs->reqs_lock);
3052 }
3053
3054 /* Flushes reach this point in nondecreasing current_gen order. */
3055 bs->active_flush_req = true;
3056 qemu_mutex_unlock(&bs->reqs_lock);
3057
3058 /* Write back all layers by calling one driver function */
3059 if (bs->drv->bdrv_co_flush) {
3060 ret = bs->drv->bdrv_co_flush(bs);
3061 goto out;
3062 }
3063
3064 /* Write back cached data to the OS even with cache=unsafe */
3065 BLKDBG_CO_EVENT(primary_child, BLKDBG_FLUSH_TO_OS);
3066 if (bs->drv->bdrv_co_flush_to_os) {
3067 ret = bs->drv->bdrv_co_flush_to_os(bs);
3068 if (ret < 0) {
3069 goto out;
3070 }
3071 }
3072
3073 /* But don't actually force it to the disk with cache=unsafe */
3074 if (bs->open_flags & BDRV_O_NO_FLUSH) {
3075 goto flush_children;
3076 }
3077
3078 /* Check if we really need to flush anything */
3079 if (bs->flushed_gen == current_gen) {
3080 goto flush_children;
3081 }
3082
3083 BLKDBG_CO_EVENT(primary_child, BLKDBG_FLUSH_TO_DISK);
3084 if (!bs->drv) {
3085 /* bs->drv->bdrv_co_flush() might have ejected the BDS
3086 * (even in case of apparent success) */
3087 ret = -ENOMEDIUM;
3088 goto out;
3089 }
3090 if (bs->drv->bdrv_co_flush_to_disk) {
3091 ret = bs->drv->bdrv_co_flush_to_disk(bs);
3092 } else if (bs->drv->bdrv_aio_flush) {
3093 BlockAIOCB *acb;
3094 CoroutineIOCompletion co = {
3095 .coroutine = qemu_coroutine_self(),
3096 };
3097
3098 acb = bs->drv->bdrv_aio_flush(bs, bdrv_co_io_em_complete, &co);
3099 if (acb == NULL) {
3100 ret = -EIO;
3101 } else {
3102 qemu_coroutine_yield();
3103 ret = co.ret;
3104 }
3105 } else {
3106 /*
3107 * Some block drivers always operate in either writethrough or unsafe
3108 * mode and don't support bdrv_flush therefore. Usually qemu doesn't
3109 * know how the server works (because the behaviour is hardcoded or
3110 * depends on server-side configuration), so we can't ensure that
3111 * everything is safe on disk. Returning an error doesn't work because
3112 * that would break guests even if the server operates in writethrough
3113 * mode.
3114 *
3115 * Let's hope the user knows what he's doing.
3116 */
3117 ret = 0;
3118 }
3119
3120 if (ret < 0) {
3121 goto out;
3122 }
3123
3124 /* Now flush the underlying protocol. It will also have BDRV_O_NO_FLUSH
3125 * in the case of cache=unsafe, so there are no useless flushes.
3126 */
3127 flush_children:
3128 ret = 0;
3129 QLIST_FOREACH(child, &bs->children, next) {
3130 if (child->perm & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED)) {
3131 int this_child_ret = bdrv_co_flush(child->bs);
3132 if (!ret) {
3133 ret = this_child_ret;
3134 }
3135 }
3136 }
3137
3138 out:
3139 /* Notify any pending flushes that we have completed */
3140 if (ret == 0) {
3141 bs->flushed_gen = current_gen;
3142 }
3143
3144 qemu_mutex_lock(&bs->reqs_lock);
3145 bs->active_flush_req = false;
3146 /* Return value is ignored - it's ok if wait queue is empty */
3147 qemu_co_queue_next(&bs->flush_queue);
3148 qemu_mutex_unlock(&bs->reqs_lock);
3149
3150 early_exit:
3151 bdrv_dec_in_flight(bs);
3152 return ret;
3153 }
3154
3155 int coroutine_fn bdrv_co_pdiscard(BdrvChild *child, int64_t offset,
3156 int64_t bytes)
3157 {
3158 BdrvTrackedRequest req;
3159 int ret;
3160 int64_t max_pdiscard;
3161 int head, tail, align;
3162 BlockDriverState *bs = child->bs;
3163 IO_CODE();
3164 assert_bdrv_graph_readable();
3165
3166 if (!bs || !bs->drv || !bdrv_co_is_inserted(bs)) {
3167 return -ENOMEDIUM;
3168 }
3169
3170 if (bdrv_has_readonly_bitmaps(bs)) {
3171 return -EPERM;
3172 }
3173
3174 ret = bdrv_check_request(offset, bytes, NULL);
3175 if (ret < 0) {
3176 return ret;
3177 }
3178
3179 /* Do nothing if disabled. */
3180 if (!(bs->open_flags & BDRV_O_UNMAP)) {
3181 return 0;
3182 }
3183
3184 if (!bs->drv->bdrv_co_pdiscard) {
3185 return 0;
3186 }
3187
3188 /* Invalidate the cached block-status data range if this discard overlaps */
3189 bdrv_bsc_invalidate_range(bs, offset, bytes);
3190
3191 /*
3192 * Discard is advisory, but some devices track and coalesce
3193 * unaligned requests, so we must pass everything down rather than
3194 * round here. Still, most devices reject unaligned requests with
3195 * -EINVAL or -ENOTSUP, so we must fragment the request accordingly.
3196 */
3197 align = MAX(bs->bl.pdiscard_alignment, bs->bl.request_alignment);
3198 assert(align % bs->bl.request_alignment == 0);
3199 head = offset % align;
3200 tail = (offset + bytes) % align;
3201
3202 bdrv_inc_in_flight(bs);
3203 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_DISCARD);
3204
3205 ret = bdrv_co_write_req_prepare(child, offset, bytes, &req, 0);
3206 if (ret < 0) {
3207 goto out;
3208 }
3209
3210 max_pdiscard = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_pdiscard, INT64_MAX),
3211 align);
3212 assert(max_pdiscard >= bs->bl.request_alignment);
3213
3214 while (bytes > 0) {
3215 int64_t num = bytes;
3216
3217 if (head) {
3218 /* Make small requests to get to alignment boundaries. */
3219 num = MIN(bytes, align - head);
3220 if (!QEMU_IS_ALIGNED(num, bs->bl.request_alignment)) {
3221 num %= bs->bl.request_alignment;
3222 }
3223 head = (head + num) % align;
3224 assert(num < max_pdiscard);
3225 } else if (tail) {
3226 if (num > align) {
3227 /* Shorten the request to the last aligned cluster. */
3228 num -= tail;
3229 } else if (!QEMU_IS_ALIGNED(tail, bs->bl.request_alignment) &&
3230 tail > bs->bl.request_alignment) {
3231 tail %= bs->bl.request_alignment;
3232 num -= tail;
3233 }
3234 }
3235 /* limit request size */
3236 if (num > max_pdiscard) {
3237 num = max_pdiscard;
3238 }
3239
3240 if (!bs->drv) {
3241 ret = -ENOMEDIUM;
3242 goto out;
3243 }
3244
3245 ret = bs->drv->bdrv_co_pdiscard(bs, offset, num);
3246 if (ret && ret != -ENOTSUP) {
3247 if (ret == -EINVAL && (offset % align != 0 || num % align != 0)) {
3248 /* Silently skip rejected unaligned head/tail requests */
3249 } else {
3250 goto out; /* bail out */
3251 }
3252 }
3253
3254 offset += num;
3255 bytes -= num;
3256 }
3257 ret = 0;
3258 out:
3259 bdrv_co_write_req_finish(child, req.offset, req.bytes, &req, ret);
3260 tracked_request_end(&req);
3261 bdrv_dec_in_flight(bs);
3262 return ret;
3263 }
3264
3265 int coroutine_fn bdrv_co_ioctl(BlockDriverState *bs, int req, void *buf)
3266 {
3267 BlockDriver *drv = bs->drv;
3268 CoroutineIOCompletion co = {
3269 .coroutine = qemu_coroutine_self(),
3270 };
3271 BlockAIOCB *acb;
3272 IO_CODE();
3273 assert_bdrv_graph_readable();
3274
3275 bdrv_inc_in_flight(bs);
3276 if (!drv || (!drv->bdrv_aio_ioctl && !drv->bdrv_co_ioctl)) {
3277 co.ret = -ENOTSUP;
3278 goto out;
3279 }
3280
3281 if (drv->bdrv_co_ioctl) {
3282 co.ret = drv->bdrv_co_ioctl(bs, req, buf);
3283 } else {
3284 acb = drv->bdrv_aio_ioctl(bs, req, buf, bdrv_co_io_em_complete, &co);
3285 if (!acb) {
3286 co.ret = -ENOTSUP;
3287 goto out;
3288 }
3289 qemu_coroutine_yield();
3290 }
3291 out:
3292 bdrv_dec_in_flight(bs);
3293 return co.ret;
3294 }
3295
3296 int coroutine_fn bdrv_co_zone_report(BlockDriverState *bs, int64_t offset,
3297 unsigned int *nr_zones,
3298 BlockZoneDescriptor *zones)
3299 {
3300 BlockDriver *drv = bs->drv;
3301 CoroutineIOCompletion co = {
3302 .coroutine = qemu_coroutine_self(),
3303 };
3304 IO_CODE();
3305
3306 bdrv_inc_in_flight(bs);
3307 if (!drv || !drv->bdrv_co_zone_report || bs->bl.zoned == BLK_Z_NONE) {
3308 co.ret = -ENOTSUP;
3309 goto out;
3310 }
3311 co.ret = drv->bdrv_co_zone_report(bs, offset, nr_zones, zones);
3312 out:
3313 bdrv_dec_in_flight(bs);
3314 return co.ret;
3315 }
3316
3317 int coroutine_fn bdrv_co_zone_mgmt(BlockDriverState *bs, BlockZoneOp op,
3318 int64_t offset, int64_t len)
3319 {
3320 BlockDriver *drv = bs->drv;
3321 CoroutineIOCompletion co = {
3322 .coroutine = qemu_coroutine_self(),
3323 };
3324 IO_CODE();
3325
3326 bdrv_inc_in_flight(bs);
3327 if (!drv || !drv->bdrv_co_zone_mgmt || bs->bl.zoned == BLK_Z_NONE) {
3328 co.ret = -ENOTSUP;
3329 goto out;
3330 }
3331 co.ret = drv->bdrv_co_zone_mgmt(bs, op, offset, len);
3332 out:
3333 bdrv_dec_in_flight(bs);
3334 return co.ret;
3335 }
3336
3337 int coroutine_fn bdrv_co_zone_append(BlockDriverState *bs, int64_t *offset,
3338 QEMUIOVector *qiov,
3339 BdrvRequestFlags flags)
3340 {
3341 int ret;
3342 BlockDriver *drv = bs->drv;
3343 CoroutineIOCompletion co = {
3344 .coroutine = qemu_coroutine_self(),
3345 };
3346 IO_CODE();
3347
3348 ret = bdrv_check_qiov_request(*offset, qiov->size, qiov, 0, NULL);
3349 if (ret < 0) {
3350 return ret;
3351 }
3352
3353 bdrv_inc_in_flight(bs);
3354 if (!drv || !drv->bdrv_co_zone_append || bs->bl.zoned == BLK_Z_NONE) {
3355 co.ret = -ENOTSUP;
3356 goto out;
3357 }
3358 co.ret = drv->bdrv_co_zone_append(bs, offset, qiov, flags);
3359 out:
3360 bdrv_dec_in_flight(bs);
3361 return co.ret;
3362 }
3363
3364 void *qemu_blockalign(BlockDriverState *bs, size_t size)
3365 {
3366 IO_CODE();
3367 return qemu_memalign(bdrv_opt_mem_align(bs), size);
3368 }
3369
3370 void *qemu_blockalign0(BlockDriverState *bs, size_t size)
3371 {
3372 IO_CODE();
3373 return memset(qemu_blockalign(bs, size), 0, size);
3374 }
3375
3376 void *qemu_try_blockalign(BlockDriverState *bs, size_t size)
3377 {
3378 size_t align = bdrv_opt_mem_align(bs);
3379 IO_CODE();
3380
3381 /* Ensure that NULL is never returned on success */
3382 assert(align > 0);
3383 if (size == 0) {
3384 size = align;
3385 }
3386
3387 return qemu_try_memalign(align, size);
3388 }
3389
3390 void *qemu_try_blockalign0(BlockDriverState *bs, size_t size)
3391 {
3392 void *mem = qemu_try_blockalign(bs, size);
3393 IO_CODE();
3394
3395 if (mem) {
3396 memset(mem, 0, size);
3397 }
3398
3399 return mem;
3400 }
3401
3402 /* Helper that undoes bdrv_register_buf() when it fails partway through */
3403 static void GRAPH_RDLOCK
3404 bdrv_register_buf_rollback(BlockDriverState *bs, void *host, size_t size,
3405 BdrvChild *final_child)
3406 {
3407 BdrvChild *child;
3408
3409 GLOBAL_STATE_CODE();
3410 assert_bdrv_graph_readable();
3411
3412 QLIST_FOREACH(child, &bs->children, next) {
3413 if (child == final_child) {
3414 break;
3415 }
3416
3417 bdrv_unregister_buf(child->bs, host, size);
3418 }
3419
3420 if (bs->drv && bs->drv->bdrv_unregister_buf) {
3421 bs->drv->bdrv_unregister_buf(bs, host, size);
3422 }
3423 }
3424
3425 bool bdrv_register_buf(BlockDriverState *bs, void *host, size_t size,
3426 Error **errp)
3427 {
3428 BdrvChild *child;
3429
3430 GLOBAL_STATE_CODE();
3431 GRAPH_RDLOCK_GUARD_MAINLOOP();
3432
3433 if (bs->drv && bs->drv->bdrv_register_buf) {
3434 if (!bs->drv->bdrv_register_buf(bs, host, size, errp)) {
3435 return false;
3436 }
3437 }
3438 QLIST_FOREACH(child, &bs->children, next) {
3439 if (!bdrv_register_buf(child->bs, host, size, errp)) {
3440 bdrv_register_buf_rollback(bs, host, size, child);
3441 return false;
3442 }
3443 }
3444 return true;
3445 }
3446
3447 void bdrv_unregister_buf(BlockDriverState *bs, void *host, size_t size)
3448 {
3449 BdrvChild *child;
3450
3451 GLOBAL_STATE_CODE();
3452 GRAPH_RDLOCK_GUARD_MAINLOOP();
3453
3454 if (bs->drv && bs->drv->bdrv_unregister_buf) {
3455 bs->drv->bdrv_unregister_buf(bs, host, size);
3456 }
3457 QLIST_FOREACH(child, &bs->children, next) {
3458 bdrv_unregister_buf(child->bs, host, size);
3459 }
3460 }
3461
3462 static int coroutine_fn GRAPH_RDLOCK bdrv_co_copy_range_internal(
3463 BdrvChild *src, int64_t src_offset, BdrvChild *dst,
3464 int64_t dst_offset, int64_t bytes,
3465 BdrvRequestFlags read_flags, BdrvRequestFlags write_flags,
3466 bool recurse_src)
3467 {
3468 BdrvTrackedRequest req;
3469 int ret;
3470 assert_bdrv_graph_readable();
3471
3472 /* TODO We can support BDRV_REQ_NO_FALLBACK here */
3473 assert(!(read_flags & BDRV_REQ_NO_FALLBACK));
3474 assert(!(write_flags & BDRV_REQ_NO_FALLBACK));
3475 assert(!(read_flags & BDRV_REQ_NO_WAIT));
3476 assert(!(write_flags & BDRV_REQ_NO_WAIT));
3477
3478 if (!dst || !dst->bs || !bdrv_co_is_inserted(dst->bs)) {
3479 return -ENOMEDIUM;
3480 }
3481 ret = bdrv_check_request32(dst_offset, bytes, NULL, 0);
3482 if (ret) {
3483 return ret;
3484 }
3485 if (write_flags & BDRV_REQ_ZERO_WRITE) {
3486 return bdrv_co_pwrite_zeroes(dst, dst_offset, bytes, write_flags);
3487 }
3488
3489 if (!src || !src->bs || !bdrv_co_is_inserted(src->bs)) {
3490 return -ENOMEDIUM;
3491 }
3492 ret = bdrv_check_request32(src_offset, bytes, NULL, 0);
3493 if (ret) {
3494 return ret;
3495 }
3496
3497 if (!src->bs->drv->bdrv_co_copy_range_from
3498 || !dst->bs->drv->bdrv_co_copy_range_to
3499 || src->bs->encrypted || dst->bs->encrypted) {
3500 return -ENOTSUP;
3501 }
3502
3503 if (recurse_src) {
3504 bdrv_inc_in_flight(src->bs);
3505 tracked_request_begin(&req, src->bs, src_offset, bytes,
3506 BDRV_TRACKED_READ);
3507
3508 /* BDRV_REQ_SERIALISING is only for write operation */
3509 assert(!(read_flags & BDRV_REQ_SERIALISING));
3510 bdrv_wait_serialising_requests(&req);
3511
3512 ret = src->bs->drv->bdrv_co_copy_range_from(src->bs,
3513 src, src_offset,
3514 dst, dst_offset,
3515 bytes,
3516 read_flags, write_flags);
3517
3518 tracked_request_end(&req);
3519 bdrv_dec_in_flight(src->bs);
3520 } else {
3521 bdrv_inc_in_flight(dst->bs);
3522 tracked_request_begin(&req, dst->bs, dst_offset, bytes,
3523 BDRV_TRACKED_WRITE);
3524 ret = bdrv_co_write_req_prepare(dst, dst_offset, bytes, &req,
3525 write_flags);
3526 if (!ret) {
3527 ret = dst->bs->drv->bdrv_co_copy_range_to(dst->bs,
3528 src, src_offset,
3529 dst, dst_offset,
3530 bytes,
3531 read_flags, write_flags);
3532 }
3533 bdrv_co_write_req_finish(dst, dst_offset, bytes, &req, ret);
3534 tracked_request_end(&req);
3535 bdrv_dec_in_flight(dst->bs);
3536 }
3537
3538 return ret;
3539 }
3540
3541 /* Copy range from @src to @dst.
3542 *
3543 * See the comment of bdrv_co_copy_range for the parameter and return value
3544 * semantics. */
3545 int coroutine_fn bdrv_co_copy_range_from(BdrvChild *src, int64_t src_offset,
3546 BdrvChild *dst, int64_t dst_offset,
3547 int64_t bytes,
3548 BdrvRequestFlags read_flags,
3549 BdrvRequestFlags write_flags)
3550 {
3551 IO_CODE();
3552 assert_bdrv_graph_readable();
3553 trace_bdrv_co_copy_range_from(src, src_offset, dst, dst_offset, bytes,
3554 read_flags, write_flags);
3555 return bdrv_co_copy_range_internal(src, src_offset, dst, dst_offset,
3556 bytes, read_flags, write_flags, true);
3557 }
3558
3559 /* Copy range from @src to @dst.
3560 *
3561 * See the comment of bdrv_co_copy_range for the parameter and return value
3562 * semantics. */
3563 int coroutine_fn bdrv_co_copy_range_to(BdrvChild *src, int64_t src_offset,
3564 BdrvChild *dst, int64_t dst_offset,
3565 int64_t bytes,
3566 BdrvRequestFlags read_flags,
3567 BdrvRequestFlags write_flags)
3568 {
3569 IO_CODE();
3570 assert_bdrv_graph_readable();
3571 trace_bdrv_co_copy_range_to(src, src_offset, dst, dst_offset, bytes,
3572 read_flags, write_flags);
3573 return bdrv_co_copy_range_internal(src, src_offset, dst, dst_offset,
3574 bytes, read_flags, write_flags, false);
3575 }
3576
3577 int coroutine_fn bdrv_co_copy_range(BdrvChild *src, int64_t src_offset,
3578 BdrvChild *dst, int64_t dst_offset,
3579 int64_t bytes, BdrvRequestFlags read_flags,
3580 BdrvRequestFlags write_flags)
3581 {
3582 IO_CODE();
3583 assert_bdrv_graph_readable();
3584
3585 return bdrv_co_copy_range_from(src, src_offset,
3586 dst, dst_offset,
3587 bytes, read_flags, write_flags);
3588 }
3589
3590 void coroutine_fn bdrv_co_parent_cb_resize(BlockDriverState *bs)
3591 {
3592 BdrvChild *c;
3593
3594 IO_CODE();
3595 assert_bdrv_graph_readable();
3596
3597 QLIST_FOREACH(c, &bs->parents, next_parent) {
3598 if (c->klass->resize) {
3599 c->klass->resize(c);
3600 }
3601 }
3602 }
3603
3604 /**
3605 * Truncate file to 'offset' bytes (needed only for file protocols)
3606 *
3607 * If 'exact' is true, the file must be resized to exactly the given
3608 * 'offset'. Otherwise, it is sufficient for the node to be at least
3609 * 'offset' bytes in length.
3610 */
3611 int coroutine_fn bdrv_co_truncate(BdrvChild *child, int64_t offset, bool exact,
3612 PreallocMode prealloc, BdrvRequestFlags flags,
3613 Error **errp)
3614 {
3615 BlockDriverState *bs = child->bs;
3616 BdrvChild *filtered, *backing;
3617 BlockDriver *drv = bs->drv;
3618 BdrvTrackedRequest req;
3619 int64_t old_size, new_bytes;
3620 int ret;
3621 IO_CODE();
3622 assert_bdrv_graph_readable();
3623
3624 /* if bs->drv == NULL, bs is closed, so there's nothing to do here */
3625 if (!drv) {
3626 error_setg(errp, "No medium inserted");
3627 return -ENOMEDIUM;
3628 }
3629 if (offset < 0) {
3630 error_setg(errp, "Image size cannot be negative");
3631 return -EINVAL;
3632 }
3633
3634 ret = bdrv_check_request(offset, 0, errp);
3635 if (ret < 0) {
3636 return ret;
3637 }
3638
3639 old_size = bdrv_co_getlength(bs);
3640 if (old_size < 0) {
3641 error_setg_errno(errp, -old_size, "Failed to get old image size");
3642 return old_size;
3643 }
3644
3645 if (bdrv_is_read_only(bs)) {
3646 error_setg(errp, "Image is read-only");
3647 return -EACCES;
3648 }
3649
3650 if (offset > old_size) {
3651 new_bytes = offset - old_size;
3652 } else {
3653 new_bytes = 0;
3654 }
3655
3656 bdrv_inc_in_flight(bs);
3657 tracked_request_begin(&req, bs, offset - new_bytes, new_bytes,
3658 BDRV_TRACKED_TRUNCATE);
3659
3660 /* If we are growing the image and potentially using preallocation for the
3661 * new area, we need to make sure that no write requests are made to it
3662 * concurrently or they might be overwritten by preallocation. */
3663 if (new_bytes) {
3664 bdrv_make_request_serialising(&req, 1);
3665 }
3666 ret = bdrv_co_write_req_prepare(child, offset - new_bytes, new_bytes, &req,
3667 0);
3668 if (ret < 0) {
3669 error_setg_errno(errp, -ret,
3670 "Failed to prepare request for truncation");
3671 goto out;
3672 }
3673
3674 filtered = bdrv_filter_child(bs);
3675 backing = bdrv_cow_child(bs);
3676
3677 /*
3678 * If the image has a backing file that is large enough that it would
3679 * provide data for the new area, we cannot leave it unallocated because
3680 * then the backing file content would become visible. Instead, zero-fill
3681 * the new area.
3682 *
3683 * Note that if the image has a backing file, but was opened without the
3684 * backing file, taking care of keeping things consistent with that backing
3685 * file is the user's responsibility.
3686 */
3687 if (new_bytes && backing) {
3688 int64_t backing_len;
3689
3690 backing_len = bdrv_co_getlength(backing->bs);
3691 if (backing_len < 0) {
3692 ret = backing_len;
3693 error_setg_errno(errp, -ret, "Could not get backing file size");
3694 goto out;
3695 }
3696
3697 if (backing_len > old_size) {
3698 flags |= BDRV_REQ_ZERO_WRITE;
3699 }
3700 }
3701
3702 if (drv->bdrv_co_truncate) {
3703 if (flags & ~bs->supported_truncate_flags) {
3704 error_setg(errp, "Block driver does not support requested flags");
3705 ret = -ENOTSUP;
3706 goto out;
3707 }
3708 ret = drv->bdrv_co_truncate(bs, offset, exact, prealloc, flags, errp);
3709 } else if (filtered) {
3710 ret = bdrv_co_truncate(filtered, offset, exact, prealloc, flags, errp);
3711 } else {
3712 error_setg(errp, "Image format driver does not support resize");
3713 ret = -ENOTSUP;
3714 goto out;
3715 }
3716 if (ret < 0) {
3717 goto out;
3718 }
3719
3720 ret = bdrv_co_refresh_total_sectors(bs, offset >> BDRV_SECTOR_BITS);
3721 if (ret < 0) {
3722 error_setg_errno(errp, -ret, "Could not refresh total sector count");
3723 } else {
3724 offset = bs->total_sectors * BDRV_SECTOR_SIZE;
3725 }
3726 /*
3727 * It's possible that truncation succeeded but bdrv_refresh_total_sectors
3728 * failed, but the latter doesn't affect how we should finish the request.
3729 * Pass 0 as the last parameter so that dirty bitmaps etc. are handled.
3730 */
3731 bdrv_co_write_req_finish(child, offset - new_bytes, new_bytes, &req, 0);
3732
3733 out:
3734 tracked_request_end(&req);
3735 bdrv_dec_in_flight(bs);
3736
3737 return ret;
3738 }
3739
3740 void bdrv_cancel_in_flight(BlockDriverState *bs)
3741 {
3742 GLOBAL_STATE_CODE();
3743 GRAPH_RDLOCK_GUARD_MAINLOOP();
3744
3745 if (!bs || !bs->drv) {
3746 return;
3747 }
3748
3749 if (bs->drv->bdrv_cancel_in_flight) {
3750 bs->drv->bdrv_cancel_in_flight(bs);
3751 }
3752 }
3753
3754 int coroutine_fn
3755 bdrv_co_preadv_snapshot(BdrvChild *child, int64_t offset, int64_t bytes,
3756 QEMUIOVector *qiov, size_t qiov_offset)
3757 {
3758 BlockDriverState *bs = child->bs;
3759 BlockDriver *drv = bs->drv;
3760 int ret;
3761 IO_CODE();
3762 assert_bdrv_graph_readable();
3763
3764 if (!drv) {
3765 return -ENOMEDIUM;
3766 }
3767
3768 if (!drv->bdrv_co_preadv_snapshot) {
3769 return -ENOTSUP;
3770 }
3771
3772 bdrv_inc_in_flight(bs);
3773 ret = drv->bdrv_co_preadv_snapshot(bs, offset, bytes, qiov, qiov_offset);
3774 bdrv_dec_in_flight(bs);
3775
3776 return ret;
3777 }
3778
3779 int coroutine_fn
3780 bdrv_co_snapshot_block_status(BlockDriverState *bs, unsigned int mode,
3781 int64_t offset, int64_t bytes,
3782 int64_t *pnum, int64_t *map,
3783 BlockDriverState **file)
3784 {
3785 BlockDriver *drv = bs->drv;
3786 int ret;
3787 IO_CODE();
3788 assert_bdrv_graph_readable();
3789
3790 if (!drv) {
3791 return -ENOMEDIUM;
3792 }
3793
3794 if (!drv->bdrv_co_snapshot_block_status) {
3795 return -ENOTSUP;
3796 }
3797
3798 bdrv_inc_in_flight(bs);
3799 ret = drv->bdrv_co_snapshot_block_status(bs, mode, offset, bytes,
3800 pnum, map, file);
3801 bdrv_dec_in_flight(bs);
3802
3803 return ret;
3804 }
3805
3806 int coroutine_fn
3807 bdrv_co_pdiscard_snapshot(BlockDriverState *bs, int64_t offset, int64_t bytes)
3808 {
3809 BlockDriver *drv = bs->drv;
3810 int ret;
3811 IO_CODE();
3812 assert_bdrv_graph_readable();
3813
3814 if (!drv) {
3815 return -ENOMEDIUM;
3816 }
3817
3818 if (!drv->bdrv_co_pdiscard_snapshot) {
3819 return -ENOTSUP;
3820 }
3821
3822 bdrv_inc_in_flight(bs);
3823 ret = drv->bdrv_co_pdiscard_snapshot(bs, offset, bytes);
3824 bdrv_dec_in_flight(bs);
3825
3826 return ret;
3827 }