block/io: bdrv_aligned_preadv: use and support qiov_offset
[qemu.git] / block / io.c
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 "sysemu/block-backend.h"
28 #include "block/aio-wait.h"
29 #include "block/blockjob.h"
30 #include "block/blockjob_int.h"
31 #include "block/block_int.h"
32 #include "qemu/cutils.h"
33 #include "qapi/error.h"
34 #include "qemu/error-report.h"
35 #include "qemu/main-loop.h"
36
37 #define NOT_DONE 0x7fffffff /* used while emulated sync operation in progress */
38
39 /* Maximum bounce buffer for copy-on-read and write zeroes, in bytes */
40 #define MAX_BOUNCE_BUFFER (32768 << BDRV_SECTOR_BITS)
41
42 static void bdrv_parent_cb_resize(BlockDriverState *bs);
43 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs,
44 int64_t offset, int bytes, BdrvRequestFlags flags);
45
46 static void bdrv_parent_drained_begin(BlockDriverState *bs, BdrvChild *ignore,
47 bool ignore_bds_parents)
48 {
49 BdrvChild *c, *next;
50
51 QLIST_FOREACH_SAFE(c, &bs->parents, next_parent, next) {
52 if (c == ignore || (ignore_bds_parents && c->role->parent_is_bds)) {
53 continue;
54 }
55 bdrv_parent_drained_begin_single(c, false);
56 }
57 }
58
59 static void bdrv_parent_drained_end_single_no_poll(BdrvChild *c,
60 int *drained_end_counter)
61 {
62 assert(c->parent_quiesce_counter > 0);
63 c->parent_quiesce_counter--;
64 if (c->role->drained_end) {
65 c->role->drained_end(c, drained_end_counter);
66 }
67 }
68
69 void bdrv_parent_drained_end_single(BdrvChild *c)
70 {
71 int drained_end_counter = 0;
72 bdrv_parent_drained_end_single_no_poll(c, &drained_end_counter);
73 BDRV_POLL_WHILE(c->bs, atomic_read(&drained_end_counter) > 0);
74 }
75
76 static void bdrv_parent_drained_end(BlockDriverState *bs, BdrvChild *ignore,
77 bool ignore_bds_parents,
78 int *drained_end_counter)
79 {
80 BdrvChild *c;
81
82 QLIST_FOREACH(c, &bs->parents, next_parent) {
83 if (c == ignore || (ignore_bds_parents && c->role->parent_is_bds)) {
84 continue;
85 }
86 bdrv_parent_drained_end_single_no_poll(c, drained_end_counter);
87 }
88 }
89
90 static bool bdrv_parent_drained_poll_single(BdrvChild *c)
91 {
92 if (c->role->drained_poll) {
93 return c->role->drained_poll(c);
94 }
95 return false;
96 }
97
98 static bool bdrv_parent_drained_poll(BlockDriverState *bs, BdrvChild *ignore,
99 bool ignore_bds_parents)
100 {
101 BdrvChild *c, *next;
102 bool busy = false;
103
104 QLIST_FOREACH_SAFE(c, &bs->parents, next_parent, next) {
105 if (c == ignore || (ignore_bds_parents && c->role->parent_is_bds)) {
106 continue;
107 }
108 busy |= bdrv_parent_drained_poll_single(c);
109 }
110
111 return busy;
112 }
113
114 void bdrv_parent_drained_begin_single(BdrvChild *c, bool poll)
115 {
116 c->parent_quiesce_counter++;
117 if (c->role->drained_begin) {
118 c->role->drained_begin(c);
119 }
120 if (poll) {
121 BDRV_POLL_WHILE(c->bs, bdrv_parent_drained_poll_single(c));
122 }
123 }
124
125 static void bdrv_merge_limits(BlockLimits *dst, const BlockLimits *src)
126 {
127 dst->opt_transfer = MAX(dst->opt_transfer, src->opt_transfer);
128 dst->max_transfer = MIN_NON_ZERO(dst->max_transfer, src->max_transfer);
129 dst->opt_mem_alignment = MAX(dst->opt_mem_alignment,
130 src->opt_mem_alignment);
131 dst->min_mem_alignment = MAX(dst->min_mem_alignment,
132 src->min_mem_alignment);
133 dst->max_iov = MIN_NON_ZERO(dst->max_iov, src->max_iov);
134 }
135
136 void bdrv_refresh_limits(BlockDriverState *bs, Error **errp)
137 {
138 BlockDriver *drv = bs->drv;
139 Error *local_err = NULL;
140
141 memset(&bs->bl, 0, sizeof(bs->bl));
142
143 if (!drv) {
144 return;
145 }
146
147 /* Default alignment based on whether driver has byte interface */
148 bs->bl.request_alignment = (drv->bdrv_co_preadv ||
149 drv->bdrv_aio_preadv ||
150 drv->bdrv_co_preadv_part) ? 1 : 512;
151
152 /* Take some limits from the children as a default */
153 if (bs->file) {
154 bdrv_refresh_limits(bs->file->bs, &local_err);
155 if (local_err) {
156 error_propagate(errp, local_err);
157 return;
158 }
159 bdrv_merge_limits(&bs->bl, &bs->file->bs->bl);
160 } else {
161 bs->bl.min_mem_alignment = 512;
162 bs->bl.opt_mem_alignment = getpagesize();
163
164 /* Safe default since most protocols use readv()/writev()/etc */
165 bs->bl.max_iov = IOV_MAX;
166 }
167
168 if (bs->backing) {
169 bdrv_refresh_limits(bs->backing->bs, &local_err);
170 if (local_err) {
171 error_propagate(errp, local_err);
172 return;
173 }
174 bdrv_merge_limits(&bs->bl, &bs->backing->bs->bl);
175 }
176
177 /* Then let the driver override it */
178 if (drv->bdrv_refresh_limits) {
179 drv->bdrv_refresh_limits(bs, errp);
180 }
181 }
182
183 /**
184 * The copy-on-read flag is actually a reference count so multiple users may
185 * use the feature without worrying about clobbering its previous state.
186 * Copy-on-read stays enabled until all users have called to disable it.
187 */
188 void bdrv_enable_copy_on_read(BlockDriverState *bs)
189 {
190 atomic_inc(&bs->copy_on_read);
191 }
192
193 void bdrv_disable_copy_on_read(BlockDriverState *bs)
194 {
195 int old = atomic_fetch_dec(&bs->copy_on_read);
196 assert(old >= 1);
197 }
198
199 typedef struct {
200 Coroutine *co;
201 BlockDriverState *bs;
202 bool done;
203 bool begin;
204 bool recursive;
205 bool poll;
206 BdrvChild *parent;
207 bool ignore_bds_parents;
208 int *drained_end_counter;
209 } BdrvCoDrainData;
210
211 static void coroutine_fn bdrv_drain_invoke_entry(void *opaque)
212 {
213 BdrvCoDrainData *data = opaque;
214 BlockDriverState *bs = data->bs;
215
216 if (data->begin) {
217 bs->drv->bdrv_co_drain_begin(bs);
218 } else {
219 bs->drv->bdrv_co_drain_end(bs);
220 }
221
222 /* Set data->done and decrement drained_end_counter before bdrv_wakeup() */
223 atomic_mb_set(&data->done, true);
224 if (!data->begin) {
225 atomic_dec(data->drained_end_counter);
226 }
227 bdrv_dec_in_flight(bs);
228
229 g_free(data);
230 }
231
232 /* Recursively call BlockDriver.bdrv_co_drain_begin/end callbacks */
233 static void bdrv_drain_invoke(BlockDriverState *bs, bool begin,
234 int *drained_end_counter)
235 {
236 BdrvCoDrainData *data;
237
238 if (!bs->drv || (begin && !bs->drv->bdrv_co_drain_begin) ||
239 (!begin && !bs->drv->bdrv_co_drain_end)) {
240 return;
241 }
242
243 data = g_new(BdrvCoDrainData, 1);
244 *data = (BdrvCoDrainData) {
245 .bs = bs,
246 .done = false,
247 .begin = begin,
248 .drained_end_counter = drained_end_counter,
249 };
250
251 if (!begin) {
252 atomic_inc(drained_end_counter);
253 }
254
255 /* Make sure the driver callback completes during the polling phase for
256 * drain_begin. */
257 bdrv_inc_in_flight(bs);
258 data->co = qemu_coroutine_create(bdrv_drain_invoke_entry, data);
259 aio_co_schedule(bdrv_get_aio_context(bs), data->co);
260 }
261
262 /* Returns true if BDRV_POLL_WHILE() should go into a blocking aio_poll() */
263 bool bdrv_drain_poll(BlockDriverState *bs, bool recursive,
264 BdrvChild *ignore_parent, bool ignore_bds_parents)
265 {
266 BdrvChild *child, *next;
267
268 if (bdrv_parent_drained_poll(bs, ignore_parent, ignore_bds_parents)) {
269 return true;
270 }
271
272 if (atomic_read(&bs->in_flight)) {
273 return true;
274 }
275
276 if (recursive) {
277 assert(!ignore_bds_parents);
278 QLIST_FOREACH_SAFE(child, &bs->children, next, next) {
279 if (bdrv_drain_poll(child->bs, recursive, child, false)) {
280 return true;
281 }
282 }
283 }
284
285 return false;
286 }
287
288 static bool bdrv_drain_poll_top_level(BlockDriverState *bs, bool recursive,
289 BdrvChild *ignore_parent)
290 {
291 return bdrv_drain_poll(bs, recursive, ignore_parent, false);
292 }
293
294 static void bdrv_do_drained_begin(BlockDriverState *bs, bool recursive,
295 BdrvChild *parent, bool ignore_bds_parents,
296 bool poll);
297 static void bdrv_do_drained_end(BlockDriverState *bs, bool recursive,
298 BdrvChild *parent, bool ignore_bds_parents,
299 int *drained_end_counter);
300
301 static void bdrv_co_drain_bh_cb(void *opaque)
302 {
303 BdrvCoDrainData *data = opaque;
304 Coroutine *co = data->co;
305 BlockDriverState *bs = data->bs;
306
307 if (bs) {
308 AioContext *ctx = bdrv_get_aio_context(bs);
309 AioContext *co_ctx = qemu_coroutine_get_aio_context(co);
310
311 /*
312 * When the coroutine yielded, the lock for its home context was
313 * released, so we need to re-acquire it here. If it explicitly
314 * acquired a different context, the lock is still held and we don't
315 * want to lock it a second time (or AIO_WAIT_WHILE() would hang).
316 */
317 if (ctx == co_ctx) {
318 aio_context_acquire(ctx);
319 }
320 bdrv_dec_in_flight(bs);
321 if (data->begin) {
322 assert(!data->drained_end_counter);
323 bdrv_do_drained_begin(bs, data->recursive, data->parent,
324 data->ignore_bds_parents, data->poll);
325 } else {
326 assert(!data->poll);
327 bdrv_do_drained_end(bs, data->recursive, data->parent,
328 data->ignore_bds_parents,
329 data->drained_end_counter);
330 }
331 if (ctx == co_ctx) {
332 aio_context_release(ctx);
333 }
334 } else {
335 assert(data->begin);
336 bdrv_drain_all_begin();
337 }
338
339 data->done = true;
340 aio_co_wake(co);
341 }
342
343 static void coroutine_fn bdrv_co_yield_to_drain(BlockDriverState *bs,
344 bool begin, bool recursive,
345 BdrvChild *parent,
346 bool ignore_bds_parents,
347 bool poll,
348 int *drained_end_counter)
349 {
350 BdrvCoDrainData data;
351
352 /* Calling bdrv_drain() from a BH ensures the current coroutine yields and
353 * other coroutines run if they were queued by aio_co_enter(). */
354
355 assert(qemu_in_coroutine());
356 data = (BdrvCoDrainData) {
357 .co = qemu_coroutine_self(),
358 .bs = bs,
359 .done = false,
360 .begin = begin,
361 .recursive = recursive,
362 .parent = parent,
363 .ignore_bds_parents = ignore_bds_parents,
364 .poll = poll,
365 .drained_end_counter = drained_end_counter,
366 };
367
368 if (bs) {
369 bdrv_inc_in_flight(bs);
370 }
371 aio_bh_schedule_oneshot(bdrv_get_aio_context(bs),
372 bdrv_co_drain_bh_cb, &data);
373
374 qemu_coroutine_yield();
375 /* If we are resumed from some other event (such as an aio completion or a
376 * timer callback), it is a bug in the caller that should be fixed. */
377 assert(data.done);
378 }
379
380 void bdrv_do_drained_begin_quiesce(BlockDriverState *bs,
381 BdrvChild *parent, bool ignore_bds_parents)
382 {
383 assert(!qemu_in_coroutine());
384
385 /* Stop things in parent-to-child order */
386 if (atomic_fetch_inc(&bs->quiesce_counter) == 0) {
387 aio_disable_external(bdrv_get_aio_context(bs));
388 }
389
390 bdrv_parent_drained_begin(bs, parent, ignore_bds_parents);
391 bdrv_drain_invoke(bs, true, NULL);
392 }
393
394 static void bdrv_do_drained_begin(BlockDriverState *bs, bool recursive,
395 BdrvChild *parent, bool ignore_bds_parents,
396 bool poll)
397 {
398 BdrvChild *child, *next;
399
400 if (qemu_in_coroutine()) {
401 bdrv_co_yield_to_drain(bs, true, recursive, parent, ignore_bds_parents,
402 poll, NULL);
403 return;
404 }
405
406 bdrv_do_drained_begin_quiesce(bs, parent, ignore_bds_parents);
407
408 if (recursive) {
409 assert(!ignore_bds_parents);
410 bs->recursive_quiesce_counter++;
411 QLIST_FOREACH_SAFE(child, &bs->children, next, next) {
412 bdrv_do_drained_begin(child->bs, true, child, ignore_bds_parents,
413 false);
414 }
415 }
416
417 /*
418 * Wait for drained requests to finish.
419 *
420 * Calling BDRV_POLL_WHILE() only once for the top-level node is okay: The
421 * call is needed so things in this AioContext can make progress even
422 * though we don't return to the main AioContext loop - this automatically
423 * includes other nodes in the same AioContext and therefore all child
424 * nodes.
425 */
426 if (poll) {
427 assert(!ignore_bds_parents);
428 BDRV_POLL_WHILE(bs, bdrv_drain_poll_top_level(bs, recursive, parent));
429 }
430 }
431
432 void bdrv_drained_begin(BlockDriverState *bs)
433 {
434 bdrv_do_drained_begin(bs, false, NULL, false, true);
435 }
436
437 void bdrv_subtree_drained_begin(BlockDriverState *bs)
438 {
439 bdrv_do_drained_begin(bs, true, NULL, false, true);
440 }
441
442 /**
443 * This function does not poll, nor must any of its recursively called
444 * functions. The *drained_end_counter pointee will be incremented
445 * once for every background operation scheduled, and decremented once
446 * the operation settles. Therefore, the pointer must remain valid
447 * until the pointee reaches 0. That implies that whoever sets up the
448 * pointee has to poll until it is 0.
449 *
450 * We use atomic operations to access *drained_end_counter, because
451 * (1) when called from bdrv_set_aio_context_ignore(), the subgraph of
452 * @bs may contain nodes in different AioContexts,
453 * (2) bdrv_drain_all_end() uses the same counter for all nodes,
454 * regardless of which AioContext they are in.
455 */
456 static void bdrv_do_drained_end(BlockDriverState *bs, bool recursive,
457 BdrvChild *parent, bool ignore_bds_parents,
458 int *drained_end_counter)
459 {
460 BdrvChild *child;
461 int old_quiesce_counter;
462
463 assert(drained_end_counter != NULL);
464
465 if (qemu_in_coroutine()) {
466 bdrv_co_yield_to_drain(bs, false, recursive, parent, ignore_bds_parents,
467 false, drained_end_counter);
468 return;
469 }
470 assert(bs->quiesce_counter > 0);
471
472 /* Re-enable things in child-to-parent order */
473 bdrv_drain_invoke(bs, false, drained_end_counter);
474 bdrv_parent_drained_end(bs, parent, ignore_bds_parents,
475 drained_end_counter);
476
477 old_quiesce_counter = atomic_fetch_dec(&bs->quiesce_counter);
478 if (old_quiesce_counter == 1) {
479 aio_enable_external(bdrv_get_aio_context(bs));
480 }
481
482 if (recursive) {
483 assert(!ignore_bds_parents);
484 bs->recursive_quiesce_counter--;
485 QLIST_FOREACH(child, &bs->children, next) {
486 bdrv_do_drained_end(child->bs, true, child, ignore_bds_parents,
487 drained_end_counter);
488 }
489 }
490 }
491
492 void bdrv_drained_end(BlockDriverState *bs)
493 {
494 int drained_end_counter = 0;
495 bdrv_do_drained_end(bs, false, NULL, false, &drained_end_counter);
496 BDRV_POLL_WHILE(bs, atomic_read(&drained_end_counter) > 0);
497 }
498
499 void bdrv_drained_end_no_poll(BlockDriverState *bs, int *drained_end_counter)
500 {
501 bdrv_do_drained_end(bs, false, NULL, false, drained_end_counter);
502 }
503
504 void bdrv_subtree_drained_end(BlockDriverState *bs)
505 {
506 int drained_end_counter = 0;
507 bdrv_do_drained_end(bs, true, NULL, false, &drained_end_counter);
508 BDRV_POLL_WHILE(bs, atomic_read(&drained_end_counter) > 0);
509 }
510
511 void bdrv_apply_subtree_drain(BdrvChild *child, BlockDriverState *new_parent)
512 {
513 int i;
514
515 for (i = 0; i < new_parent->recursive_quiesce_counter; i++) {
516 bdrv_do_drained_begin(child->bs, true, child, false, true);
517 }
518 }
519
520 void bdrv_unapply_subtree_drain(BdrvChild *child, BlockDriverState *old_parent)
521 {
522 int drained_end_counter = 0;
523 int i;
524
525 for (i = 0; i < old_parent->recursive_quiesce_counter; i++) {
526 bdrv_do_drained_end(child->bs, true, child, false,
527 &drained_end_counter);
528 }
529
530 BDRV_POLL_WHILE(child->bs, atomic_read(&drained_end_counter) > 0);
531 }
532
533 /*
534 * Wait for pending requests to complete on a single BlockDriverState subtree,
535 * and suspend block driver's internal I/O until next request arrives.
536 *
537 * Note that unlike bdrv_drain_all(), the caller must hold the BlockDriverState
538 * AioContext.
539 */
540 void coroutine_fn bdrv_co_drain(BlockDriverState *bs)
541 {
542 assert(qemu_in_coroutine());
543 bdrv_drained_begin(bs);
544 bdrv_drained_end(bs);
545 }
546
547 void bdrv_drain(BlockDriverState *bs)
548 {
549 bdrv_drained_begin(bs);
550 bdrv_drained_end(bs);
551 }
552
553 static void bdrv_drain_assert_idle(BlockDriverState *bs)
554 {
555 BdrvChild *child, *next;
556
557 assert(atomic_read(&bs->in_flight) == 0);
558 QLIST_FOREACH_SAFE(child, &bs->children, next, next) {
559 bdrv_drain_assert_idle(child->bs);
560 }
561 }
562
563 unsigned int bdrv_drain_all_count = 0;
564
565 static bool bdrv_drain_all_poll(void)
566 {
567 BlockDriverState *bs = NULL;
568 bool result = false;
569
570 /* bdrv_drain_poll() can't make changes to the graph and we are holding the
571 * main AioContext lock, so iterating bdrv_next_all_states() is safe. */
572 while ((bs = bdrv_next_all_states(bs))) {
573 AioContext *aio_context = bdrv_get_aio_context(bs);
574 aio_context_acquire(aio_context);
575 result |= bdrv_drain_poll(bs, false, NULL, true);
576 aio_context_release(aio_context);
577 }
578
579 return result;
580 }
581
582 /*
583 * Wait for pending requests to complete across all BlockDriverStates
584 *
585 * This function does not flush data to disk, use bdrv_flush_all() for that
586 * after calling this function.
587 *
588 * This pauses all block jobs and disables external clients. It must
589 * be paired with bdrv_drain_all_end().
590 *
591 * NOTE: no new block jobs or BlockDriverStates can be created between
592 * the bdrv_drain_all_begin() and bdrv_drain_all_end() calls.
593 */
594 void bdrv_drain_all_begin(void)
595 {
596 BlockDriverState *bs = NULL;
597
598 if (qemu_in_coroutine()) {
599 bdrv_co_yield_to_drain(NULL, true, false, NULL, true, true, NULL);
600 return;
601 }
602
603 /* AIO_WAIT_WHILE() with a NULL context can only be called from the main
604 * loop AioContext, so make sure we're in the main context. */
605 assert(qemu_get_current_aio_context() == qemu_get_aio_context());
606 assert(bdrv_drain_all_count < INT_MAX);
607 bdrv_drain_all_count++;
608
609 /* Quiesce all nodes, without polling in-flight requests yet. The graph
610 * cannot change during this loop. */
611 while ((bs = bdrv_next_all_states(bs))) {
612 AioContext *aio_context = bdrv_get_aio_context(bs);
613
614 aio_context_acquire(aio_context);
615 bdrv_do_drained_begin(bs, false, NULL, true, false);
616 aio_context_release(aio_context);
617 }
618
619 /* Now poll the in-flight requests */
620 AIO_WAIT_WHILE(NULL, bdrv_drain_all_poll());
621
622 while ((bs = bdrv_next_all_states(bs))) {
623 bdrv_drain_assert_idle(bs);
624 }
625 }
626
627 void bdrv_drain_all_end(void)
628 {
629 BlockDriverState *bs = NULL;
630 int drained_end_counter = 0;
631
632 while ((bs = bdrv_next_all_states(bs))) {
633 AioContext *aio_context = bdrv_get_aio_context(bs);
634
635 aio_context_acquire(aio_context);
636 bdrv_do_drained_end(bs, false, NULL, true, &drained_end_counter);
637 aio_context_release(aio_context);
638 }
639
640 assert(qemu_get_current_aio_context() == qemu_get_aio_context());
641 AIO_WAIT_WHILE(NULL, atomic_read(&drained_end_counter) > 0);
642
643 assert(bdrv_drain_all_count > 0);
644 bdrv_drain_all_count--;
645 }
646
647 void bdrv_drain_all(void)
648 {
649 bdrv_drain_all_begin();
650 bdrv_drain_all_end();
651 }
652
653 /**
654 * Remove an active request from the tracked requests list
655 *
656 * This function should be called when a tracked request is completing.
657 */
658 static void tracked_request_end(BdrvTrackedRequest *req)
659 {
660 if (req->serialising) {
661 atomic_dec(&req->bs->serialising_in_flight);
662 }
663
664 qemu_co_mutex_lock(&req->bs->reqs_lock);
665 QLIST_REMOVE(req, list);
666 qemu_co_queue_restart_all(&req->wait_queue);
667 qemu_co_mutex_unlock(&req->bs->reqs_lock);
668 }
669
670 /**
671 * Add an active request to the tracked requests list
672 */
673 static void tracked_request_begin(BdrvTrackedRequest *req,
674 BlockDriverState *bs,
675 int64_t offset,
676 uint64_t bytes,
677 enum BdrvTrackedRequestType type)
678 {
679 assert(bytes <= INT64_MAX && offset <= INT64_MAX - bytes);
680
681 *req = (BdrvTrackedRequest){
682 .bs = bs,
683 .offset = offset,
684 .bytes = bytes,
685 .type = type,
686 .co = qemu_coroutine_self(),
687 .serialising = false,
688 .overlap_offset = offset,
689 .overlap_bytes = bytes,
690 };
691
692 qemu_co_queue_init(&req->wait_queue);
693
694 qemu_co_mutex_lock(&bs->reqs_lock);
695 QLIST_INSERT_HEAD(&bs->tracked_requests, req, list);
696 qemu_co_mutex_unlock(&bs->reqs_lock);
697 }
698
699 static void mark_request_serialising(BdrvTrackedRequest *req, uint64_t align)
700 {
701 int64_t overlap_offset = req->offset & ~(align - 1);
702 uint64_t overlap_bytes = ROUND_UP(req->offset + req->bytes, align)
703 - overlap_offset;
704
705 if (!req->serialising) {
706 atomic_inc(&req->bs->serialising_in_flight);
707 req->serialising = true;
708 }
709
710 req->overlap_offset = MIN(req->overlap_offset, overlap_offset);
711 req->overlap_bytes = MAX(req->overlap_bytes, overlap_bytes);
712 }
713
714 static bool is_request_serialising_and_aligned(BdrvTrackedRequest *req)
715 {
716 /*
717 * If the request is serialising, overlap_offset and overlap_bytes are set,
718 * so we can check if the request is aligned. Otherwise, don't care and
719 * return false.
720 */
721
722 return req->serialising && (req->offset == req->overlap_offset) &&
723 (req->bytes == req->overlap_bytes);
724 }
725
726 /**
727 * Round a region to cluster boundaries
728 */
729 void bdrv_round_to_clusters(BlockDriverState *bs,
730 int64_t offset, int64_t bytes,
731 int64_t *cluster_offset,
732 int64_t *cluster_bytes)
733 {
734 BlockDriverInfo bdi;
735
736 if (bdrv_get_info(bs, &bdi) < 0 || bdi.cluster_size == 0) {
737 *cluster_offset = offset;
738 *cluster_bytes = bytes;
739 } else {
740 int64_t c = bdi.cluster_size;
741 *cluster_offset = QEMU_ALIGN_DOWN(offset, c);
742 *cluster_bytes = QEMU_ALIGN_UP(offset - *cluster_offset + bytes, c);
743 }
744 }
745
746 static int bdrv_get_cluster_size(BlockDriverState *bs)
747 {
748 BlockDriverInfo bdi;
749 int ret;
750
751 ret = bdrv_get_info(bs, &bdi);
752 if (ret < 0 || bdi.cluster_size == 0) {
753 return bs->bl.request_alignment;
754 } else {
755 return bdi.cluster_size;
756 }
757 }
758
759 static bool tracked_request_overlaps(BdrvTrackedRequest *req,
760 int64_t offset, uint64_t bytes)
761 {
762 /* aaaa bbbb */
763 if (offset >= req->overlap_offset + req->overlap_bytes) {
764 return false;
765 }
766 /* bbbb aaaa */
767 if (req->overlap_offset >= offset + bytes) {
768 return false;
769 }
770 return true;
771 }
772
773 void bdrv_inc_in_flight(BlockDriverState *bs)
774 {
775 atomic_inc(&bs->in_flight);
776 }
777
778 void bdrv_wakeup(BlockDriverState *bs)
779 {
780 aio_wait_kick();
781 }
782
783 void bdrv_dec_in_flight(BlockDriverState *bs)
784 {
785 atomic_dec(&bs->in_flight);
786 bdrv_wakeup(bs);
787 }
788
789 static bool coroutine_fn wait_serialising_requests(BdrvTrackedRequest *self)
790 {
791 BlockDriverState *bs = self->bs;
792 BdrvTrackedRequest *req;
793 bool retry;
794 bool waited = false;
795
796 if (!atomic_read(&bs->serialising_in_flight)) {
797 return false;
798 }
799
800 do {
801 retry = false;
802 qemu_co_mutex_lock(&bs->reqs_lock);
803 QLIST_FOREACH(req, &bs->tracked_requests, list) {
804 if (req == self || (!req->serialising && !self->serialising)) {
805 continue;
806 }
807 if (tracked_request_overlaps(req, self->overlap_offset,
808 self->overlap_bytes))
809 {
810 /* Hitting this means there was a reentrant request, for
811 * example, a block driver issuing nested requests. This must
812 * never happen since it means deadlock.
813 */
814 assert(qemu_coroutine_self() != req->co);
815
816 /* If the request is already (indirectly) waiting for us, or
817 * will wait for us as soon as it wakes up, then just go on
818 * (instead of producing a deadlock in the former case). */
819 if (!req->waiting_for) {
820 self->waiting_for = req;
821 qemu_co_queue_wait(&req->wait_queue, &bs->reqs_lock);
822 self->waiting_for = NULL;
823 retry = true;
824 waited = true;
825 break;
826 }
827 }
828 }
829 qemu_co_mutex_unlock(&bs->reqs_lock);
830 } while (retry);
831
832 return waited;
833 }
834
835 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset,
836 size_t size)
837 {
838 if (size > BDRV_REQUEST_MAX_BYTES) {
839 return -EIO;
840 }
841
842 if (!bdrv_is_inserted(bs)) {
843 return -ENOMEDIUM;
844 }
845
846 if (offset < 0) {
847 return -EIO;
848 }
849
850 return 0;
851 }
852
853 typedef struct RwCo {
854 BdrvChild *child;
855 int64_t offset;
856 QEMUIOVector *qiov;
857 bool is_write;
858 int ret;
859 BdrvRequestFlags flags;
860 } RwCo;
861
862 static void coroutine_fn bdrv_rw_co_entry(void *opaque)
863 {
864 RwCo *rwco = opaque;
865
866 if (!rwco->is_write) {
867 rwco->ret = bdrv_co_preadv(rwco->child, rwco->offset,
868 rwco->qiov->size, rwco->qiov,
869 rwco->flags);
870 } else {
871 rwco->ret = bdrv_co_pwritev(rwco->child, rwco->offset,
872 rwco->qiov->size, rwco->qiov,
873 rwco->flags);
874 }
875 aio_wait_kick();
876 }
877
878 /*
879 * Process a vectored synchronous request using coroutines
880 */
881 static int bdrv_prwv_co(BdrvChild *child, int64_t offset,
882 QEMUIOVector *qiov, bool is_write,
883 BdrvRequestFlags flags)
884 {
885 Coroutine *co;
886 RwCo rwco = {
887 .child = child,
888 .offset = offset,
889 .qiov = qiov,
890 .is_write = is_write,
891 .ret = NOT_DONE,
892 .flags = flags,
893 };
894
895 if (qemu_in_coroutine()) {
896 /* Fast-path if already in coroutine context */
897 bdrv_rw_co_entry(&rwco);
898 } else {
899 co = qemu_coroutine_create(bdrv_rw_co_entry, &rwco);
900 bdrv_coroutine_enter(child->bs, co);
901 BDRV_POLL_WHILE(child->bs, rwco.ret == NOT_DONE);
902 }
903 return rwco.ret;
904 }
905
906 int bdrv_pwrite_zeroes(BdrvChild *child, int64_t offset,
907 int bytes, BdrvRequestFlags flags)
908 {
909 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, bytes);
910
911 return bdrv_prwv_co(child, offset, &qiov, true,
912 BDRV_REQ_ZERO_WRITE | flags);
913 }
914
915 /*
916 * Completely zero out a block device with the help of bdrv_pwrite_zeroes.
917 * The operation is sped up by checking the block status and only writing
918 * zeroes to the device if they currently do not return zeroes. Optional
919 * flags are passed through to bdrv_pwrite_zeroes (e.g. BDRV_REQ_MAY_UNMAP,
920 * BDRV_REQ_FUA).
921 *
922 * Returns < 0 on error, 0 on success. For error codes see bdrv_write().
923 */
924 int bdrv_make_zero(BdrvChild *child, BdrvRequestFlags flags)
925 {
926 int ret;
927 int64_t target_size, bytes, offset = 0;
928 BlockDriverState *bs = child->bs;
929
930 target_size = bdrv_getlength(bs);
931 if (target_size < 0) {
932 return target_size;
933 }
934
935 for (;;) {
936 bytes = MIN(target_size - offset, BDRV_REQUEST_MAX_BYTES);
937 if (bytes <= 0) {
938 return 0;
939 }
940 ret = bdrv_block_status(bs, offset, bytes, &bytes, NULL, NULL);
941 if (ret < 0) {
942 return ret;
943 }
944 if (ret & BDRV_BLOCK_ZERO) {
945 offset += bytes;
946 continue;
947 }
948 ret = bdrv_pwrite_zeroes(child, offset, bytes, flags);
949 if (ret < 0) {
950 return ret;
951 }
952 offset += bytes;
953 }
954 }
955
956 int bdrv_preadv(BdrvChild *child, int64_t offset, QEMUIOVector *qiov)
957 {
958 int ret;
959
960 ret = bdrv_prwv_co(child, offset, qiov, false, 0);
961 if (ret < 0) {
962 return ret;
963 }
964
965 return qiov->size;
966 }
967
968 /* See bdrv_pwrite() for the return codes */
969 int bdrv_pread(BdrvChild *child, int64_t offset, void *buf, int bytes)
970 {
971 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, bytes);
972
973 if (bytes < 0) {
974 return -EINVAL;
975 }
976
977 return bdrv_preadv(child, offset, &qiov);
978 }
979
980 int bdrv_pwritev(BdrvChild *child, int64_t offset, QEMUIOVector *qiov)
981 {
982 int ret;
983
984 ret = bdrv_prwv_co(child, offset, qiov, true, 0);
985 if (ret < 0) {
986 return ret;
987 }
988
989 return qiov->size;
990 }
991
992 /* Return no. of bytes on success or < 0 on error. Important errors are:
993 -EIO generic I/O error (may happen for all errors)
994 -ENOMEDIUM No media inserted.
995 -EINVAL Invalid offset or number of bytes
996 -EACCES Trying to write a read-only device
997 */
998 int bdrv_pwrite(BdrvChild *child, int64_t offset, const void *buf, int bytes)
999 {
1000 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, bytes);
1001
1002 if (bytes < 0) {
1003 return -EINVAL;
1004 }
1005
1006 return bdrv_pwritev(child, offset, &qiov);
1007 }
1008
1009 /*
1010 * Writes to the file and ensures that no writes are reordered across this
1011 * request (acts as a barrier)
1012 *
1013 * Returns 0 on success, -errno in error cases.
1014 */
1015 int bdrv_pwrite_sync(BdrvChild *child, int64_t offset,
1016 const void *buf, int count)
1017 {
1018 int ret;
1019
1020 ret = bdrv_pwrite(child, offset, buf, count);
1021 if (ret < 0) {
1022 return ret;
1023 }
1024
1025 ret = bdrv_flush(child->bs);
1026 if (ret < 0) {
1027 return ret;
1028 }
1029
1030 return 0;
1031 }
1032
1033 typedef struct CoroutineIOCompletion {
1034 Coroutine *coroutine;
1035 int ret;
1036 } CoroutineIOCompletion;
1037
1038 static void bdrv_co_io_em_complete(void *opaque, int ret)
1039 {
1040 CoroutineIOCompletion *co = opaque;
1041
1042 co->ret = ret;
1043 aio_co_wake(co->coroutine);
1044 }
1045
1046 static int coroutine_fn bdrv_driver_preadv(BlockDriverState *bs,
1047 uint64_t offset, uint64_t bytes,
1048 QEMUIOVector *qiov,
1049 size_t qiov_offset, int flags)
1050 {
1051 BlockDriver *drv = bs->drv;
1052 int64_t sector_num;
1053 unsigned int nb_sectors;
1054 QEMUIOVector local_qiov;
1055 int ret;
1056
1057 assert(!(flags & ~BDRV_REQ_MASK));
1058 assert(!(flags & BDRV_REQ_NO_FALLBACK));
1059
1060 if (!drv) {
1061 return -ENOMEDIUM;
1062 }
1063
1064 if (drv->bdrv_co_preadv_part) {
1065 return drv->bdrv_co_preadv_part(bs, offset, bytes, qiov, qiov_offset,
1066 flags);
1067 }
1068
1069 if (qiov_offset > 0 || bytes != qiov->size) {
1070 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes);
1071 qiov = &local_qiov;
1072 }
1073
1074 if (drv->bdrv_co_preadv) {
1075 ret = drv->bdrv_co_preadv(bs, offset, bytes, qiov, flags);
1076 goto out;
1077 }
1078
1079 if (drv->bdrv_aio_preadv) {
1080 BlockAIOCB *acb;
1081 CoroutineIOCompletion co = {
1082 .coroutine = qemu_coroutine_self(),
1083 };
1084
1085 acb = drv->bdrv_aio_preadv(bs, offset, bytes, qiov, flags,
1086 bdrv_co_io_em_complete, &co);
1087 if (acb == NULL) {
1088 ret = -EIO;
1089 goto out;
1090 } else {
1091 qemu_coroutine_yield();
1092 ret = co.ret;
1093 goto out;
1094 }
1095 }
1096
1097 sector_num = offset >> BDRV_SECTOR_BITS;
1098 nb_sectors = bytes >> BDRV_SECTOR_BITS;
1099
1100 assert((offset & (BDRV_SECTOR_SIZE - 1)) == 0);
1101 assert((bytes & (BDRV_SECTOR_SIZE - 1)) == 0);
1102 assert(bytes <= BDRV_REQUEST_MAX_BYTES);
1103 assert(drv->bdrv_co_readv);
1104
1105 ret = drv->bdrv_co_readv(bs, sector_num, nb_sectors, qiov);
1106
1107 out:
1108 if (qiov == &local_qiov) {
1109 qemu_iovec_destroy(&local_qiov);
1110 }
1111
1112 return ret;
1113 }
1114
1115 static int coroutine_fn bdrv_driver_pwritev(BlockDriverState *bs,
1116 uint64_t offset, uint64_t bytes,
1117 QEMUIOVector *qiov,
1118 size_t qiov_offset, int flags)
1119 {
1120 BlockDriver *drv = bs->drv;
1121 int64_t sector_num;
1122 unsigned int nb_sectors;
1123 QEMUIOVector local_qiov;
1124 int ret;
1125
1126 assert(!(flags & ~BDRV_REQ_MASK));
1127 assert(!(flags & BDRV_REQ_NO_FALLBACK));
1128
1129 if (!drv) {
1130 return -ENOMEDIUM;
1131 }
1132
1133 if (drv->bdrv_co_pwritev_part) {
1134 ret = drv->bdrv_co_pwritev_part(bs, offset, bytes, qiov, qiov_offset,
1135 flags & bs->supported_write_flags);
1136 flags &= ~bs->supported_write_flags;
1137 goto emulate_flags;
1138 }
1139
1140 if (qiov_offset > 0 || bytes != qiov->size) {
1141 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes);
1142 qiov = &local_qiov;
1143 }
1144
1145 if (drv->bdrv_co_pwritev) {
1146 ret = drv->bdrv_co_pwritev(bs, offset, bytes, qiov,
1147 flags & bs->supported_write_flags);
1148 flags &= ~bs->supported_write_flags;
1149 goto emulate_flags;
1150 }
1151
1152 if (drv->bdrv_aio_pwritev) {
1153 BlockAIOCB *acb;
1154 CoroutineIOCompletion co = {
1155 .coroutine = qemu_coroutine_self(),
1156 };
1157
1158 acb = drv->bdrv_aio_pwritev(bs, offset, bytes, qiov,
1159 flags & bs->supported_write_flags,
1160 bdrv_co_io_em_complete, &co);
1161 flags &= ~bs->supported_write_flags;
1162 if (acb == NULL) {
1163 ret = -EIO;
1164 } else {
1165 qemu_coroutine_yield();
1166 ret = co.ret;
1167 }
1168 goto emulate_flags;
1169 }
1170
1171 sector_num = offset >> BDRV_SECTOR_BITS;
1172 nb_sectors = bytes >> BDRV_SECTOR_BITS;
1173
1174 assert((offset & (BDRV_SECTOR_SIZE - 1)) == 0);
1175 assert((bytes & (BDRV_SECTOR_SIZE - 1)) == 0);
1176 assert(bytes <= BDRV_REQUEST_MAX_BYTES);
1177
1178 assert(drv->bdrv_co_writev);
1179 ret = drv->bdrv_co_writev(bs, sector_num, nb_sectors, qiov,
1180 flags & bs->supported_write_flags);
1181 flags &= ~bs->supported_write_flags;
1182
1183 emulate_flags:
1184 if (ret == 0 && (flags & BDRV_REQ_FUA)) {
1185 ret = bdrv_co_flush(bs);
1186 }
1187
1188 if (qiov == &local_qiov) {
1189 qemu_iovec_destroy(&local_qiov);
1190 }
1191
1192 return ret;
1193 }
1194
1195 static int coroutine_fn
1196 bdrv_driver_pwritev_compressed(BlockDriverState *bs, uint64_t offset,
1197 uint64_t bytes, QEMUIOVector *qiov,
1198 size_t qiov_offset)
1199 {
1200 BlockDriver *drv = bs->drv;
1201 QEMUIOVector local_qiov;
1202 int ret;
1203
1204 if (!drv) {
1205 return -ENOMEDIUM;
1206 }
1207
1208 if (!block_driver_can_compress(drv)) {
1209 return -ENOTSUP;
1210 }
1211
1212 if (drv->bdrv_co_pwritev_compressed_part) {
1213 return drv->bdrv_co_pwritev_compressed_part(bs, offset, bytes,
1214 qiov, qiov_offset);
1215 }
1216
1217 if (qiov_offset == 0) {
1218 return drv->bdrv_co_pwritev_compressed(bs, offset, bytes, qiov);
1219 }
1220
1221 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes);
1222 ret = drv->bdrv_co_pwritev_compressed(bs, offset, bytes, &local_qiov);
1223 qemu_iovec_destroy(&local_qiov);
1224
1225 return ret;
1226 }
1227
1228 static int coroutine_fn bdrv_co_do_copy_on_readv(BdrvChild *child,
1229 int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
1230 size_t qiov_offset, int flags)
1231 {
1232 BlockDriverState *bs = child->bs;
1233
1234 /* Perform I/O through a temporary buffer so that users who scribble over
1235 * their read buffer while the operation is in progress do not end up
1236 * modifying the image file. This is critical for zero-copy guest I/O
1237 * where anything might happen inside guest memory.
1238 */
1239 void *bounce_buffer = NULL;
1240
1241 BlockDriver *drv = bs->drv;
1242 int64_t cluster_offset;
1243 int64_t cluster_bytes;
1244 size_t skip_bytes;
1245 int ret;
1246 int max_transfer = MIN_NON_ZERO(bs->bl.max_transfer,
1247 BDRV_REQUEST_MAX_BYTES);
1248 unsigned int progress = 0;
1249
1250 if (!drv) {
1251 return -ENOMEDIUM;
1252 }
1253
1254 /* FIXME We cannot require callers to have write permissions when all they
1255 * are doing is a read request. If we did things right, write permissions
1256 * would be obtained anyway, but internally by the copy-on-read code. As
1257 * long as it is implemented here rather than in a separate filter driver,
1258 * the copy-on-read code doesn't have its own BdrvChild, however, for which
1259 * it could request permissions. Therefore we have to bypass the permission
1260 * system for the moment. */
1261 // assert(child->perm & (BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE));
1262
1263 /* Cover entire cluster so no additional backing file I/O is required when
1264 * allocating cluster in the image file. Note that this value may exceed
1265 * BDRV_REQUEST_MAX_BYTES (even when the original read did not), which
1266 * is one reason we loop rather than doing it all at once.
1267 */
1268 bdrv_round_to_clusters(bs, offset, bytes, &cluster_offset, &cluster_bytes);
1269 skip_bytes = offset - cluster_offset;
1270
1271 trace_bdrv_co_do_copy_on_readv(bs, offset, bytes,
1272 cluster_offset, cluster_bytes);
1273
1274 while (cluster_bytes) {
1275 int64_t pnum;
1276
1277 ret = bdrv_is_allocated(bs, cluster_offset,
1278 MIN(cluster_bytes, max_transfer), &pnum);
1279 if (ret < 0) {
1280 /* Safe to treat errors in querying allocation as if
1281 * unallocated; we'll probably fail again soon on the
1282 * read, but at least that will set a decent errno.
1283 */
1284 pnum = MIN(cluster_bytes, max_transfer);
1285 }
1286
1287 /* Stop at EOF if the image ends in the middle of the cluster */
1288 if (ret == 0 && pnum == 0) {
1289 assert(progress >= bytes);
1290 break;
1291 }
1292
1293 assert(skip_bytes < pnum);
1294
1295 if (ret <= 0) {
1296 QEMUIOVector local_qiov;
1297
1298 /* Must copy-on-read; use the bounce buffer */
1299 pnum = MIN(pnum, MAX_BOUNCE_BUFFER);
1300 if (!bounce_buffer) {
1301 int64_t max_we_need = MAX(pnum, cluster_bytes - pnum);
1302 int64_t max_allowed = MIN(max_transfer, MAX_BOUNCE_BUFFER);
1303 int64_t bounce_buffer_len = MIN(max_we_need, max_allowed);
1304
1305 bounce_buffer = qemu_try_blockalign(bs, bounce_buffer_len);
1306 if (!bounce_buffer) {
1307 ret = -ENOMEM;
1308 goto err;
1309 }
1310 }
1311 qemu_iovec_init_buf(&local_qiov, bounce_buffer, pnum);
1312
1313 ret = bdrv_driver_preadv(bs, cluster_offset, pnum,
1314 &local_qiov, 0, 0);
1315 if (ret < 0) {
1316 goto err;
1317 }
1318
1319 bdrv_debug_event(bs, BLKDBG_COR_WRITE);
1320 if (drv->bdrv_co_pwrite_zeroes &&
1321 buffer_is_zero(bounce_buffer, pnum)) {
1322 /* FIXME: Should we (perhaps conditionally) be setting
1323 * BDRV_REQ_MAY_UNMAP, if it will allow for a sparser copy
1324 * that still correctly reads as zero? */
1325 ret = bdrv_co_do_pwrite_zeroes(bs, cluster_offset, pnum,
1326 BDRV_REQ_WRITE_UNCHANGED);
1327 } else {
1328 /* This does not change the data on the disk, it is not
1329 * necessary to flush even in cache=writethrough mode.
1330 */
1331 ret = bdrv_driver_pwritev(bs, cluster_offset, pnum,
1332 &local_qiov, 0,
1333 BDRV_REQ_WRITE_UNCHANGED);
1334 }
1335
1336 if (ret < 0) {
1337 /* It might be okay to ignore write errors for guest
1338 * requests. If this is a deliberate copy-on-read
1339 * then we don't want to ignore the error. Simply
1340 * report it in all cases.
1341 */
1342 goto err;
1343 }
1344
1345 if (!(flags & BDRV_REQ_PREFETCH)) {
1346 qemu_iovec_from_buf(qiov, qiov_offset + progress,
1347 bounce_buffer + skip_bytes,
1348 pnum - skip_bytes);
1349 }
1350 } else if (!(flags & BDRV_REQ_PREFETCH)) {
1351 /* Read directly into the destination */
1352 ret = bdrv_driver_preadv(bs, offset + progress,
1353 MIN(pnum - skip_bytes, bytes - progress),
1354 qiov, qiov_offset + progress, 0);
1355 if (ret < 0) {
1356 goto err;
1357 }
1358 }
1359
1360 cluster_offset += pnum;
1361 cluster_bytes -= pnum;
1362 progress += pnum - skip_bytes;
1363 skip_bytes = 0;
1364 }
1365 ret = 0;
1366
1367 err:
1368 qemu_vfree(bounce_buffer);
1369 return ret;
1370 }
1371
1372 /*
1373 * Forwards an already correctly aligned request to the BlockDriver. This
1374 * handles copy on read, zeroing after EOF, and fragmentation of large
1375 * reads; any other features must be implemented by the caller.
1376 */
1377 static int coroutine_fn bdrv_aligned_preadv(BdrvChild *child,
1378 BdrvTrackedRequest *req, int64_t offset, unsigned int bytes,
1379 int64_t align, QEMUIOVector *qiov, size_t qiov_offset, int flags)
1380 {
1381 BlockDriverState *bs = child->bs;
1382 int64_t total_bytes, max_bytes;
1383 int ret = 0;
1384 uint64_t bytes_remaining = bytes;
1385 int max_transfer;
1386
1387 assert(is_power_of_2(align));
1388 assert((offset & (align - 1)) == 0);
1389 assert((bytes & (align - 1)) == 0);
1390 assert((bs->open_flags & BDRV_O_NO_IO) == 0);
1391 max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX),
1392 align);
1393
1394 /* TODO: We would need a per-BDS .supported_read_flags and
1395 * potential fallback support, if we ever implement any read flags
1396 * to pass through to drivers. For now, there aren't any
1397 * passthrough flags. */
1398 assert(!(flags & ~(BDRV_REQ_NO_SERIALISING | BDRV_REQ_COPY_ON_READ |
1399 BDRV_REQ_PREFETCH)));
1400
1401 /* Handle Copy on Read and associated serialisation */
1402 if (flags & BDRV_REQ_COPY_ON_READ) {
1403 /* If we touch the same cluster it counts as an overlap. This
1404 * guarantees that allocating writes will be serialized and not race
1405 * with each other for the same cluster. For example, in copy-on-read
1406 * it ensures that the CoR read and write operations are atomic and
1407 * guest writes cannot interleave between them. */
1408 mark_request_serialising(req, bdrv_get_cluster_size(bs));
1409 }
1410
1411 /* BDRV_REQ_SERIALISING is only for write operation */
1412 assert(!(flags & BDRV_REQ_SERIALISING));
1413
1414 if (!(flags & BDRV_REQ_NO_SERIALISING)) {
1415 wait_serialising_requests(req);
1416 }
1417
1418 if (flags & BDRV_REQ_COPY_ON_READ) {
1419 int64_t pnum;
1420
1421 ret = bdrv_is_allocated(bs, offset, bytes, &pnum);
1422 if (ret < 0) {
1423 goto out;
1424 }
1425
1426 if (!ret || pnum != bytes) {
1427 ret = bdrv_co_do_copy_on_readv(child, offset, bytes,
1428 qiov, qiov_offset, flags);
1429 goto out;
1430 } else if (flags & BDRV_REQ_PREFETCH) {
1431 goto out;
1432 }
1433 }
1434
1435 /* Forward the request to the BlockDriver, possibly fragmenting it */
1436 total_bytes = bdrv_getlength(bs);
1437 if (total_bytes < 0) {
1438 ret = total_bytes;
1439 goto out;
1440 }
1441
1442 max_bytes = ROUND_UP(MAX(0, total_bytes - offset), align);
1443 if (bytes <= max_bytes && bytes <= max_transfer) {
1444 ret = bdrv_driver_preadv(bs, offset, bytes, qiov, qiov_offset, 0);
1445 goto out;
1446 }
1447
1448 while (bytes_remaining) {
1449 int num;
1450
1451 if (max_bytes) {
1452 num = MIN(bytes_remaining, MIN(max_bytes, max_transfer));
1453 assert(num);
1454
1455 ret = bdrv_driver_preadv(bs, offset + bytes - bytes_remaining,
1456 num, qiov, bytes - bytes_remaining, 0);
1457 max_bytes -= num;
1458 } else {
1459 num = bytes_remaining;
1460 ret = qemu_iovec_memset(qiov, bytes - bytes_remaining, 0,
1461 bytes_remaining);
1462 }
1463 if (ret < 0) {
1464 goto out;
1465 }
1466 bytes_remaining -= num;
1467 }
1468
1469 out:
1470 return ret < 0 ? ret : 0;
1471 }
1472
1473 /*
1474 * Request padding
1475 *
1476 * |<---- align ----->| |<----- align ---->|
1477 * |<- head ->|<------------- bytes ------------->|<-- tail -->|
1478 * | | | | | |
1479 * -*----------$-------*-------- ... --------*-----$------------*---
1480 * | | | | | |
1481 * | offset | | end |
1482 * ALIGN_DOWN(offset) ALIGN_UP(offset) ALIGN_DOWN(end) ALIGN_UP(end)
1483 * [buf ... ) [tail_buf )
1484 *
1485 * @buf is an aligned allocation needed to store @head and @tail paddings. @head
1486 * is placed at the beginning of @buf and @tail at the @end.
1487 *
1488 * @tail_buf is a pointer to sub-buffer, corresponding to align-sized chunk
1489 * around tail, if tail exists.
1490 *
1491 * @merge_reads is true for small requests,
1492 * if @buf_len == @head + bytes + @tail. In this case it is possible that both
1493 * head and tail exist but @buf_len == align and @tail_buf == @buf.
1494 */
1495 typedef struct BdrvRequestPadding {
1496 uint8_t *buf;
1497 size_t buf_len;
1498 uint8_t *tail_buf;
1499 size_t head;
1500 size_t tail;
1501 bool merge_reads;
1502 QEMUIOVector local_qiov;
1503 } BdrvRequestPadding;
1504
1505 static bool bdrv_init_padding(BlockDriverState *bs,
1506 int64_t offset, int64_t bytes,
1507 BdrvRequestPadding *pad)
1508 {
1509 uint64_t align = bs->bl.request_alignment;
1510 size_t sum;
1511
1512 memset(pad, 0, sizeof(*pad));
1513
1514 pad->head = offset & (align - 1);
1515 pad->tail = ((offset + bytes) & (align - 1));
1516 if (pad->tail) {
1517 pad->tail = align - pad->tail;
1518 }
1519
1520 if ((!pad->head && !pad->tail) || !bytes) {
1521 return false;
1522 }
1523
1524 sum = pad->head + bytes + pad->tail;
1525 pad->buf_len = (sum > align && pad->head && pad->tail) ? 2 * align : align;
1526 pad->buf = qemu_blockalign(bs, pad->buf_len);
1527 pad->merge_reads = sum == pad->buf_len;
1528 if (pad->tail) {
1529 pad->tail_buf = pad->buf + pad->buf_len - align;
1530 }
1531
1532 return true;
1533 }
1534
1535 static int bdrv_padding_rmw_read(BdrvChild *child,
1536 BdrvTrackedRequest *req,
1537 BdrvRequestPadding *pad,
1538 bool zero_middle)
1539 {
1540 QEMUIOVector local_qiov;
1541 BlockDriverState *bs = child->bs;
1542 uint64_t align = bs->bl.request_alignment;
1543 int ret;
1544
1545 assert(req->serialising && pad->buf);
1546
1547 if (pad->head || pad->merge_reads) {
1548 uint64_t bytes = pad->merge_reads ? pad->buf_len : align;
1549
1550 qemu_iovec_init_buf(&local_qiov, pad->buf, bytes);
1551
1552 if (pad->head) {
1553 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD);
1554 }
1555 if (pad->merge_reads && pad->tail) {
1556 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
1557 }
1558 ret = bdrv_aligned_preadv(child, req, req->overlap_offset, bytes,
1559 align, &local_qiov, 0, 0);
1560 if (ret < 0) {
1561 return ret;
1562 }
1563 if (pad->head) {
1564 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD);
1565 }
1566 if (pad->merge_reads && pad->tail) {
1567 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
1568 }
1569
1570 if (pad->merge_reads) {
1571 goto zero_mem;
1572 }
1573 }
1574
1575 if (pad->tail) {
1576 qemu_iovec_init_buf(&local_qiov, pad->tail_buf, align);
1577
1578 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
1579 ret = bdrv_aligned_preadv(
1580 child, req,
1581 req->overlap_offset + req->overlap_bytes - align,
1582 align, align, &local_qiov, 0, 0);
1583 if (ret < 0) {
1584 return ret;
1585 }
1586 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
1587 }
1588
1589 zero_mem:
1590 if (zero_middle) {
1591 memset(pad->buf + pad->head, 0, pad->buf_len - pad->head - pad->tail);
1592 }
1593
1594 return 0;
1595 }
1596
1597 static void bdrv_padding_destroy(BdrvRequestPadding *pad)
1598 {
1599 if (pad->buf) {
1600 qemu_vfree(pad->buf);
1601 qemu_iovec_destroy(&pad->local_qiov);
1602 }
1603 }
1604
1605 /*
1606 * bdrv_pad_request
1607 *
1608 * Exchange request parameters with padded request if needed. Don't include RMW
1609 * read of padding, bdrv_padding_rmw_read() should be called separately if
1610 * needed.
1611 *
1612 * All parameters except @bs are in-out: they represent original request at
1613 * function call and padded (if padding needed) at function finish.
1614 *
1615 * Function always succeeds.
1616 */
1617 static bool bdrv_pad_request(BlockDriverState *bs, QEMUIOVector **qiov,
1618 int64_t *offset, unsigned int *bytes,
1619 BdrvRequestPadding *pad)
1620 {
1621 if (!bdrv_init_padding(bs, *offset, *bytes, pad)) {
1622 return false;
1623 }
1624
1625 qemu_iovec_init_extended(&pad->local_qiov, pad->buf, pad->head,
1626 *qiov, 0, *bytes,
1627 pad->buf + pad->buf_len - pad->tail, pad->tail);
1628 *bytes += pad->head + pad->tail;
1629 *offset -= pad->head;
1630 *qiov = &pad->local_qiov;
1631
1632 return true;
1633 }
1634
1635 int coroutine_fn bdrv_co_preadv(BdrvChild *child,
1636 int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
1637 BdrvRequestFlags flags)
1638 {
1639 BlockDriverState *bs = child->bs;
1640 BdrvTrackedRequest req;
1641 BdrvRequestPadding pad;
1642 int ret;
1643
1644 trace_bdrv_co_preadv(bs, offset, bytes, flags);
1645
1646 ret = bdrv_check_byte_request(bs, offset, bytes);
1647 if (ret < 0) {
1648 return ret;
1649 }
1650
1651 bdrv_inc_in_flight(bs);
1652
1653 /* Don't do copy-on-read if we read data before write operation */
1654 if (atomic_read(&bs->copy_on_read) && !(flags & BDRV_REQ_NO_SERIALISING)) {
1655 flags |= BDRV_REQ_COPY_ON_READ;
1656 }
1657
1658 bdrv_pad_request(bs, &qiov, &offset, &bytes, &pad);
1659
1660 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_READ);
1661 ret = bdrv_aligned_preadv(child, &req, offset, bytes,
1662 bs->bl.request_alignment,
1663 qiov, 0, flags);
1664 tracked_request_end(&req);
1665 bdrv_dec_in_flight(bs);
1666
1667 bdrv_padding_destroy(&pad);
1668
1669 return ret;
1670 }
1671
1672 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs,
1673 int64_t offset, int bytes, BdrvRequestFlags flags)
1674 {
1675 BlockDriver *drv = bs->drv;
1676 QEMUIOVector qiov;
1677 void *buf = NULL;
1678 int ret = 0;
1679 bool need_flush = false;
1680 int head = 0;
1681 int tail = 0;
1682
1683 int max_write_zeroes = MIN_NON_ZERO(bs->bl.max_pwrite_zeroes, INT_MAX);
1684 int alignment = MAX(bs->bl.pwrite_zeroes_alignment,
1685 bs->bl.request_alignment);
1686 int max_transfer = MIN_NON_ZERO(bs->bl.max_transfer, MAX_BOUNCE_BUFFER);
1687
1688 if (!drv) {
1689 return -ENOMEDIUM;
1690 }
1691
1692 if ((flags & ~bs->supported_zero_flags) & BDRV_REQ_NO_FALLBACK) {
1693 return -ENOTSUP;
1694 }
1695
1696 assert(alignment % bs->bl.request_alignment == 0);
1697 head = offset % alignment;
1698 tail = (offset + bytes) % alignment;
1699 max_write_zeroes = QEMU_ALIGN_DOWN(max_write_zeroes, alignment);
1700 assert(max_write_zeroes >= bs->bl.request_alignment);
1701
1702 while (bytes > 0 && !ret) {
1703 int num = bytes;
1704
1705 /* Align request. Block drivers can expect the "bulk" of the request
1706 * to be aligned, and that unaligned requests do not cross cluster
1707 * boundaries.
1708 */
1709 if (head) {
1710 /* Make a small request up to the first aligned sector. For
1711 * convenience, limit this request to max_transfer even if
1712 * we don't need to fall back to writes. */
1713 num = MIN(MIN(bytes, max_transfer), alignment - head);
1714 head = (head + num) % alignment;
1715 assert(num < max_write_zeroes);
1716 } else if (tail && num > alignment) {
1717 /* Shorten the request to the last aligned sector. */
1718 num -= tail;
1719 }
1720
1721 /* limit request size */
1722 if (num > max_write_zeroes) {
1723 num = max_write_zeroes;
1724 }
1725
1726 ret = -ENOTSUP;
1727 /* First try the efficient write zeroes operation */
1728 if (drv->bdrv_co_pwrite_zeroes) {
1729 ret = drv->bdrv_co_pwrite_zeroes(bs, offset, num,
1730 flags & bs->supported_zero_flags);
1731 if (ret != -ENOTSUP && (flags & BDRV_REQ_FUA) &&
1732 !(bs->supported_zero_flags & BDRV_REQ_FUA)) {
1733 need_flush = true;
1734 }
1735 } else {
1736 assert(!bs->supported_zero_flags);
1737 }
1738
1739 if (ret < 0 && !(flags & BDRV_REQ_NO_FALLBACK)) {
1740 /* Fall back to bounce buffer if write zeroes is unsupported */
1741 BdrvRequestFlags write_flags = flags & ~BDRV_REQ_ZERO_WRITE;
1742
1743 if ((flags & BDRV_REQ_FUA) &&
1744 !(bs->supported_write_flags & BDRV_REQ_FUA)) {
1745 /* No need for bdrv_driver_pwrite() to do a fallback
1746 * flush on each chunk; use just one at the end */
1747 write_flags &= ~BDRV_REQ_FUA;
1748 need_flush = true;
1749 }
1750 num = MIN(num, max_transfer);
1751 if (buf == NULL) {
1752 buf = qemu_try_blockalign0(bs, num);
1753 if (buf == NULL) {
1754 ret = -ENOMEM;
1755 goto fail;
1756 }
1757 }
1758 qemu_iovec_init_buf(&qiov, buf, num);
1759
1760 ret = bdrv_driver_pwritev(bs, offset, num, &qiov, 0, write_flags);
1761
1762 /* Keep bounce buffer around if it is big enough for all
1763 * all future requests.
1764 */
1765 if (num < max_transfer) {
1766 qemu_vfree(buf);
1767 buf = NULL;
1768 }
1769 }
1770
1771 offset += num;
1772 bytes -= num;
1773 }
1774
1775 fail:
1776 if (ret == 0 && need_flush) {
1777 ret = bdrv_co_flush(bs);
1778 }
1779 qemu_vfree(buf);
1780 return ret;
1781 }
1782
1783 static inline int coroutine_fn
1784 bdrv_co_write_req_prepare(BdrvChild *child, int64_t offset, uint64_t bytes,
1785 BdrvTrackedRequest *req, int flags)
1786 {
1787 BlockDriverState *bs = child->bs;
1788 bool waited;
1789 int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);
1790
1791 if (bs->read_only) {
1792 return -EPERM;
1793 }
1794
1795 /* BDRV_REQ_NO_SERIALISING is only for read operation */
1796 assert(!(flags & BDRV_REQ_NO_SERIALISING));
1797 assert(!(bs->open_flags & BDRV_O_INACTIVE));
1798 assert((bs->open_flags & BDRV_O_NO_IO) == 0);
1799 assert(!(flags & ~BDRV_REQ_MASK));
1800
1801 if (flags & BDRV_REQ_SERIALISING) {
1802 mark_request_serialising(req, bdrv_get_cluster_size(bs));
1803 }
1804
1805 waited = wait_serialising_requests(req);
1806
1807 assert(!waited || !req->serialising ||
1808 is_request_serialising_and_aligned(req));
1809 assert(req->overlap_offset <= offset);
1810 assert(offset + bytes <= req->overlap_offset + req->overlap_bytes);
1811 assert(end_sector <= bs->total_sectors || child->perm & BLK_PERM_RESIZE);
1812
1813 switch (req->type) {
1814 case BDRV_TRACKED_WRITE:
1815 case BDRV_TRACKED_DISCARD:
1816 if (flags & BDRV_REQ_WRITE_UNCHANGED) {
1817 assert(child->perm & (BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE));
1818 } else {
1819 assert(child->perm & BLK_PERM_WRITE);
1820 }
1821 return notifier_with_return_list_notify(&bs->before_write_notifiers,
1822 req);
1823 case BDRV_TRACKED_TRUNCATE:
1824 assert(child->perm & BLK_PERM_RESIZE);
1825 return 0;
1826 default:
1827 abort();
1828 }
1829 }
1830
1831 static inline void coroutine_fn
1832 bdrv_co_write_req_finish(BdrvChild *child, int64_t offset, uint64_t bytes,
1833 BdrvTrackedRequest *req, int ret)
1834 {
1835 int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);
1836 BlockDriverState *bs = child->bs;
1837
1838 atomic_inc(&bs->write_gen);
1839
1840 /*
1841 * Discard cannot extend the image, but in error handling cases, such as
1842 * when reverting a qcow2 cluster allocation, the discarded range can pass
1843 * the end of image file, so we cannot assert about BDRV_TRACKED_DISCARD
1844 * here. Instead, just skip it, since semantically a discard request
1845 * beyond EOF cannot expand the image anyway.
1846 */
1847 if (ret == 0 &&
1848 (req->type == BDRV_TRACKED_TRUNCATE ||
1849 end_sector > bs->total_sectors) &&
1850 req->type != BDRV_TRACKED_DISCARD) {
1851 bs->total_sectors = end_sector;
1852 bdrv_parent_cb_resize(bs);
1853 bdrv_dirty_bitmap_truncate(bs, end_sector << BDRV_SECTOR_BITS);
1854 }
1855 if (req->bytes) {
1856 switch (req->type) {
1857 case BDRV_TRACKED_WRITE:
1858 stat64_max(&bs->wr_highest_offset, offset + bytes);
1859 /* fall through, to set dirty bits */
1860 case BDRV_TRACKED_DISCARD:
1861 bdrv_set_dirty(bs, offset, bytes);
1862 break;
1863 default:
1864 break;
1865 }
1866 }
1867 }
1868
1869 /*
1870 * Forwards an already correctly aligned write request to the BlockDriver,
1871 * after possibly fragmenting it.
1872 */
1873 static int coroutine_fn bdrv_aligned_pwritev(BdrvChild *child,
1874 BdrvTrackedRequest *req, int64_t offset, unsigned int bytes,
1875 int64_t align, QEMUIOVector *qiov, int flags)
1876 {
1877 BlockDriverState *bs = child->bs;
1878 BlockDriver *drv = bs->drv;
1879 int ret;
1880
1881 uint64_t bytes_remaining = bytes;
1882 int max_transfer;
1883
1884 if (!drv) {
1885 return -ENOMEDIUM;
1886 }
1887
1888 if (bdrv_has_readonly_bitmaps(bs)) {
1889 return -EPERM;
1890 }
1891
1892 assert(is_power_of_2(align));
1893 assert((offset & (align - 1)) == 0);
1894 assert((bytes & (align - 1)) == 0);
1895 assert(!qiov || bytes == qiov->size);
1896 max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX),
1897 align);
1898
1899 ret = bdrv_co_write_req_prepare(child, offset, bytes, req, flags);
1900
1901 if (!ret && bs->detect_zeroes != BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF &&
1902 !(flags & BDRV_REQ_ZERO_WRITE) && drv->bdrv_co_pwrite_zeroes &&
1903 qemu_iovec_is_zero(qiov, 0, qiov->size)) {
1904 flags |= BDRV_REQ_ZERO_WRITE;
1905 if (bs->detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP) {
1906 flags |= BDRV_REQ_MAY_UNMAP;
1907 }
1908 }
1909
1910 if (ret < 0) {
1911 /* Do nothing, write notifier decided to fail this request */
1912 } else if (flags & BDRV_REQ_ZERO_WRITE) {
1913 bdrv_debug_event(bs, BLKDBG_PWRITEV_ZERO);
1914 ret = bdrv_co_do_pwrite_zeroes(bs, offset, bytes, flags);
1915 } else if (flags & BDRV_REQ_WRITE_COMPRESSED) {
1916 ret = bdrv_driver_pwritev_compressed(bs, offset, bytes, qiov, 0);
1917 } else if (bytes <= max_transfer) {
1918 bdrv_debug_event(bs, BLKDBG_PWRITEV);
1919 ret = bdrv_driver_pwritev(bs, offset, bytes, qiov, 0, flags);
1920 } else {
1921 bdrv_debug_event(bs, BLKDBG_PWRITEV);
1922 while (bytes_remaining) {
1923 int num = MIN(bytes_remaining, max_transfer);
1924 QEMUIOVector local_qiov;
1925 int local_flags = flags;
1926
1927 assert(num);
1928 if (num < bytes_remaining && (flags & BDRV_REQ_FUA) &&
1929 !(bs->supported_write_flags & BDRV_REQ_FUA)) {
1930 /* If FUA is going to be emulated by flush, we only
1931 * need to flush on the last iteration */
1932 local_flags &= ~BDRV_REQ_FUA;
1933 }
1934 qemu_iovec_init(&local_qiov, qiov->niov);
1935 qemu_iovec_concat(&local_qiov, qiov, bytes - bytes_remaining, num);
1936
1937 ret = bdrv_driver_pwritev(bs, offset + bytes - bytes_remaining,
1938 num, &local_qiov, 0, local_flags);
1939 qemu_iovec_destroy(&local_qiov);
1940 if (ret < 0) {
1941 break;
1942 }
1943 bytes_remaining -= num;
1944 }
1945 }
1946 bdrv_debug_event(bs, BLKDBG_PWRITEV_DONE);
1947
1948 if (ret >= 0) {
1949 ret = 0;
1950 }
1951 bdrv_co_write_req_finish(child, offset, bytes, req, ret);
1952
1953 return ret;
1954 }
1955
1956 static int coroutine_fn bdrv_co_do_zero_pwritev(BdrvChild *child,
1957 int64_t offset,
1958 unsigned int bytes,
1959 BdrvRequestFlags flags,
1960 BdrvTrackedRequest *req)
1961 {
1962 BlockDriverState *bs = child->bs;
1963 QEMUIOVector local_qiov;
1964 uint64_t align = bs->bl.request_alignment;
1965 int ret = 0;
1966 bool padding;
1967 BdrvRequestPadding pad;
1968
1969 padding = bdrv_init_padding(bs, offset, bytes, &pad);
1970 if (padding) {
1971 mark_request_serialising(req, align);
1972 wait_serialising_requests(req);
1973
1974 bdrv_padding_rmw_read(child, req, &pad, true);
1975
1976 if (pad.head || pad.merge_reads) {
1977 int64_t aligned_offset = offset & ~(align - 1);
1978 int64_t write_bytes = pad.merge_reads ? pad.buf_len : align;
1979
1980 qemu_iovec_init_buf(&local_qiov, pad.buf, write_bytes);
1981 ret = bdrv_aligned_pwritev(child, req, aligned_offset, write_bytes,
1982 align, &local_qiov,
1983 flags & ~BDRV_REQ_ZERO_WRITE);
1984 if (ret < 0 || pad.merge_reads) {
1985 /* Error or all work is done */
1986 goto out;
1987 }
1988 offset += write_bytes - pad.head;
1989 bytes -= write_bytes - pad.head;
1990 }
1991 }
1992
1993 assert(!bytes || (offset & (align - 1)) == 0);
1994 if (bytes >= align) {
1995 /* Write the aligned part in the middle. */
1996 uint64_t aligned_bytes = bytes & ~(align - 1);
1997 ret = bdrv_aligned_pwritev(child, req, offset, aligned_bytes, align,
1998 NULL, flags);
1999 if (ret < 0) {
2000 goto out;
2001 }
2002 bytes -= aligned_bytes;
2003 offset += aligned_bytes;
2004 }
2005
2006 assert(!bytes || (offset & (align - 1)) == 0);
2007 if (bytes) {
2008 assert(align == pad.tail + bytes);
2009
2010 qemu_iovec_init_buf(&local_qiov, pad.tail_buf, align);
2011 ret = bdrv_aligned_pwritev(child, req, offset, align, align,
2012 &local_qiov, flags & ~BDRV_REQ_ZERO_WRITE);
2013 }
2014
2015 out:
2016 bdrv_padding_destroy(&pad);
2017
2018 return ret;
2019 }
2020
2021 /*
2022 * Handle a write request in coroutine context
2023 */
2024 int coroutine_fn bdrv_co_pwritev(BdrvChild *child,
2025 int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
2026 BdrvRequestFlags flags)
2027 {
2028 BlockDriverState *bs = child->bs;
2029 BdrvTrackedRequest req;
2030 uint64_t align = bs->bl.request_alignment;
2031 BdrvRequestPadding pad;
2032 int ret;
2033
2034 trace_bdrv_co_pwritev(child->bs, offset, bytes, flags);
2035
2036 if (!bs->drv) {
2037 return -ENOMEDIUM;
2038 }
2039
2040 ret = bdrv_check_byte_request(bs, offset, bytes);
2041 if (ret < 0) {
2042 return ret;
2043 }
2044
2045 bdrv_inc_in_flight(bs);
2046 /*
2047 * Align write if necessary by performing a read-modify-write cycle.
2048 * Pad qiov with the read parts and be sure to have a tracked request not
2049 * only for bdrv_aligned_pwritev, but also for the reads of the RMW cycle.
2050 */
2051 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_WRITE);
2052
2053 if (flags & BDRV_REQ_ZERO_WRITE) {
2054 ret = bdrv_co_do_zero_pwritev(child, offset, bytes, flags, &req);
2055 goto out;
2056 }
2057
2058 if (bdrv_pad_request(bs, &qiov, &offset, &bytes, &pad)) {
2059 mark_request_serialising(&req, align);
2060 wait_serialising_requests(&req);
2061 bdrv_padding_rmw_read(child, &req, &pad, false);
2062 }
2063
2064 ret = bdrv_aligned_pwritev(child, &req, offset, bytes, align,
2065 qiov, flags);
2066
2067 bdrv_padding_destroy(&pad);
2068
2069 out:
2070 tracked_request_end(&req);
2071 bdrv_dec_in_flight(bs);
2072
2073 return ret;
2074 }
2075
2076 int coroutine_fn bdrv_co_pwrite_zeroes(BdrvChild *child, int64_t offset,
2077 int bytes, BdrvRequestFlags flags)
2078 {
2079 trace_bdrv_co_pwrite_zeroes(child->bs, offset, bytes, flags);
2080
2081 if (!(child->bs->open_flags & BDRV_O_UNMAP)) {
2082 flags &= ~BDRV_REQ_MAY_UNMAP;
2083 }
2084
2085 return bdrv_co_pwritev(child, offset, bytes, NULL,
2086 BDRV_REQ_ZERO_WRITE | flags);
2087 }
2088
2089 /*
2090 * Flush ALL BDSes regardless of if they are reachable via a BlkBackend or not.
2091 */
2092 int bdrv_flush_all(void)
2093 {
2094 BdrvNextIterator it;
2095 BlockDriverState *bs = NULL;
2096 int result = 0;
2097
2098 for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
2099 AioContext *aio_context = bdrv_get_aio_context(bs);
2100 int ret;
2101
2102 aio_context_acquire(aio_context);
2103 ret = bdrv_flush(bs);
2104 if (ret < 0 && !result) {
2105 result = ret;
2106 }
2107 aio_context_release(aio_context);
2108 }
2109
2110 return result;
2111 }
2112
2113
2114 typedef struct BdrvCoBlockStatusData {
2115 BlockDriverState *bs;
2116 BlockDriverState *base;
2117 bool want_zero;
2118 int64_t offset;
2119 int64_t bytes;
2120 int64_t *pnum;
2121 int64_t *map;
2122 BlockDriverState **file;
2123 int ret;
2124 bool done;
2125 } BdrvCoBlockStatusData;
2126
2127 int coroutine_fn bdrv_co_block_status_from_file(BlockDriverState *bs,
2128 bool want_zero,
2129 int64_t offset,
2130 int64_t bytes,
2131 int64_t *pnum,
2132 int64_t *map,
2133 BlockDriverState **file)
2134 {
2135 assert(bs->file && bs->file->bs);
2136 *pnum = bytes;
2137 *map = offset;
2138 *file = bs->file->bs;
2139 return BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID;
2140 }
2141
2142 int coroutine_fn bdrv_co_block_status_from_backing(BlockDriverState *bs,
2143 bool want_zero,
2144 int64_t offset,
2145 int64_t bytes,
2146 int64_t *pnum,
2147 int64_t *map,
2148 BlockDriverState **file)
2149 {
2150 assert(bs->backing && bs->backing->bs);
2151 *pnum = bytes;
2152 *map = offset;
2153 *file = bs->backing->bs;
2154 return BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID;
2155 }
2156
2157 /*
2158 * Returns the allocation status of the specified sectors.
2159 * Drivers not implementing the functionality are assumed to not support
2160 * backing files, hence all their sectors are reported as allocated.
2161 *
2162 * If 'want_zero' is true, the caller is querying for mapping
2163 * purposes, with a focus on valid BDRV_BLOCK_OFFSET_VALID, _DATA, and
2164 * _ZERO where possible; otherwise, the result favors larger 'pnum',
2165 * with a focus on accurate BDRV_BLOCK_ALLOCATED.
2166 *
2167 * If 'offset' is beyond the end of the disk image the return value is
2168 * BDRV_BLOCK_EOF and 'pnum' is set to 0.
2169 *
2170 * 'bytes' is the max value 'pnum' should be set to. If bytes goes
2171 * beyond the end of the disk image it will be clamped; if 'pnum' is set to
2172 * the end of the image, then the returned value will include BDRV_BLOCK_EOF.
2173 *
2174 * 'pnum' is set to the number of bytes (including and immediately
2175 * following the specified offset) that are easily known to be in the
2176 * same allocated/unallocated state. Note that a second call starting
2177 * at the original offset plus returned pnum may have the same status.
2178 * The returned value is non-zero on success except at end-of-file.
2179 *
2180 * Returns negative errno on failure. Otherwise, if the
2181 * BDRV_BLOCK_OFFSET_VALID bit is set, 'map' and 'file' (if non-NULL) are
2182 * set to the host mapping and BDS corresponding to the guest offset.
2183 */
2184 static int coroutine_fn bdrv_co_block_status(BlockDriverState *bs,
2185 bool want_zero,
2186 int64_t offset, int64_t bytes,
2187 int64_t *pnum, int64_t *map,
2188 BlockDriverState **file)
2189 {
2190 int64_t total_size;
2191 int64_t n; /* bytes */
2192 int ret;
2193 int64_t local_map = 0;
2194 BlockDriverState *local_file = NULL;
2195 int64_t aligned_offset, aligned_bytes;
2196 uint32_t align;
2197
2198 assert(pnum);
2199 *pnum = 0;
2200 total_size = bdrv_getlength(bs);
2201 if (total_size < 0) {
2202 ret = total_size;
2203 goto early_out;
2204 }
2205
2206 if (offset >= total_size) {
2207 ret = BDRV_BLOCK_EOF;
2208 goto early_out;
2209 }
2210 if (!bytes) {
2211 ret = 0;
2212 goto early_out;
2213 }
2214
2215 n = total_size - offset;
2216 if (n < bytes) {
2217 bytes = n;
2218 }
2219
2220 /* Must be non-NULL or bdrv_getlength() would have failed */
2221 assert(bs->drv);
2222 if (!bs->drv->bdrv_co_block_status) {
2223 *pnum = bytes;
2224 ret = BDRV_BLOCK_DATA | BDRV_BLOCK_ALLOCATED;
2225 if (offset + bytes == total_size) {
2226 ret |= BDRV_BLOCK_EOF;
2227 }
2228 if (bs->drv->protocol_name) {
2229 ret |= BDRV_BLOCK_OFFSET_VALID;
2230 local_map = offset;
2231 local_file = bs;
2232 }
2233 goto early_out;
2234 }
2235
2236 bdrv_inc_in_flight(bs);
2237
2238 /* Round out to request_alignment boundaries */
2239 align = bs->bl.request_alignment;
2240 aligned_offset = QEMU_ALIGN_DOWN(offset, align);
2241 aligned_bytes = ROUND_UP(offset + bytes, align) - aligned_offset;
2242
2243 ret = bs->drv->bdrv_co_block_status(bs, want_zero, aligned_offset,
2244 aligned_bytes, pnum, &local_map,
2245 &local_file);
2246 if (ret < 0) {
2247 *pnum = 0;
2248 goto out;
2249 }
2250
2251 /*
2252 * The driver's result must be a non-zero multiple of request_alignment.
2253 * Clamp pnum and adjust map to original request.
2254 */
2255 assert(*pnum && QEMU_IS_ALIGNED(*pnum, align) &&
2256 align > offset - aligned_offset);
2257 if (ret & BDRV_BLOCK_RECURSE) {
2258 assert(ret & BDRV_BLOCK_DATA);
2259 assert(ret & BDRV_BLOCK_OFFSET_VALID);
2260 assert(!(ret & BDRV_BLOCK_ZERO));
2261 }
2262
2263 *pnum -= offset - aligned_offset;
2264 if (*pnum > bytes) {
2265 *pnum = bytes;
2266 }
2267 if (ret & BDRV_BLOCK_OFFSET_VALID) {
2268 local_map += offset - aligned_offset;
2269 }
2270
2271 if (ret & BDRV_BLOCK_RAW) {
2272 assert(ret & BDRV_BLOCK_OFFSET_VALID && local_file);
2273 ret = bdrv_co_block_status(local_file, want_zero, local_map,
2274 *pnum, pnum, &local_map, &local_file);
2275 goto out;
2276 }
2277
2278 if (ret & (BDRV_BLOCK_DATA | BDRV_BLOCK_ZERO)) {
2279 ret |= BDRV_BLOCK_ALLOCATED;
2280 } else if (want_zero) {
2281 if (bdrv_unallocated_blocks_are_zero(bs)) {
2282 ret |= BDRV_BLOCK_ZERO;
2283 } else if (bs->backing) {
2284 BlockDriverState *bs2 = bs->backing->bs;
2285 int64_t size2 = bdrv_getlength(bs2);
2286
2287 if (size2 >= 0 && offset >= size2) {
2288 ret |= BDRV_BLOCK_ZERO;
2289 }
2290 }
2291 }
2292
2293 if (want_zero && ret & BDRV_BLOCK_RECURSE &&
2294 local_file && local_file != bs &&
2295 (ret & BDRV_BLOCK_DATA) && !(ret & BDRV_BLOCK_ZERO) &&
2296 (ret & BDRV_BLOCK_OFFSET_VALID)) {
2297 int64_t file_pnum;
2298 int ret2;
2299
2300 ret2 = bdrv_co_block_status(local_file, want_zero, local_map,
2301 *pnum, &file_pnum, NULL, NULL);
2302 if (ret2 >= 0) {
2303 /* Ignore errors. This is just providing extra information, it
2304 * is useful but not necessary.
2305 */
2306 if (ret2 & BDRV_BLOCK_EOF &&
2307 (!file_pnum || ret2 & BDRV_BLOCK_ZERO)) {
2308 /*
2309 * It is valid for the format block driver to read
2310 * beyond the end of the underlying file's current
2311 * size; such areas read as zero.
2312 */
2313 ret |= BDRV_BLOCK_ZERO;
2314 } else {
2315 /* Limit request to the range reported by the protocol driver */
2316 *pnum = file_pnum;
2317 ret |= (ret2 & BDRV_BLOCK_ZERO);
2318 }
2319 }
2320 }
2321
2322 out:
2323 bdrv_dec_in_flight(bs);
2324 if (ret >= 0 && offset + *pnum == total_size) {
2325 ret |= BDRV_BLOCK_EOF;
2326 }
2327 early_out:
2328 if (file) {
2329 *file = local_file;
2330 }
2331 if (map) {
2332 *map = local_map;
2333 }
2334 return ret;
2335 }
2336
2337 static int coroutine_fn bdrv_co_block_status_above(BlockDriverState *bs,
2338 BlockDriverState *base,
2339 bool want_zero,
2340 int64_t offset,
2341 int64_t bytes,
2342 int64_t *pnum,
2343 int64_t *map,
2344 BlockDriverState **file)
2345 {
2346 BlockDriverState *p;
2347 int ret = 0;
2348 bool first = true;
2349
2350 assert(bs != base);
2351 for (p = bs; p != base; p = backing_bs(p)) {
2352 ret = bdrv_co_block_status(p, want_zero, offset, bytes, pnum, map,
2353 file);
2354 if (ret < 0) {
2355 break;
2356 }
2357 if (ret & BDRV_BLOCK_ZERO && ret & BDRV_BLOCK_EOF && !first) {
2358 /*
2359 * Reading beyond the end of the file continues to read
2360 * zeroes, but we can only widen the result to the
2361 * unallocated length we learned from an earlier
2362 * iteration.
2363 */
2364 *pnum = bytes;
2365 }
2366 if (ret & (BDRV_BLOCK_ZERO | BDRV_BLOCK_DATA)) {
2367 break;
2368 }
2369 /* [offset, pnum] unallocated on this layer, which could be only
2370 * the first part of [offset, bytes]. */
2371 bytes = MIN(bytes, *pnum);
2372 first = false;
2373 }
2374 return ret;
2375 }
2376
2377 /* Coroutine wrapper for bdrv_block_status_above() */
2378 static void coroutine_fn bdrv_block_status_above_co_entry(void *opaque)
2379 {
2380 BdrvCoBlockStatusData *data = opaque;
2381
2382 data->ret = bdrv_co_block_status_above(data->bs, data->base,
2383 data->want_zero,
2384 data->offset, data->bytes,
2385 data->pnum, data->map, data->file);
2386 data->done = true;
2387 aio_wait_kick();
2388 }
2389
2390 /*
2391 * Synchronous wrapper around bdrv_co_block_status_above().
2392 *
2393 * See bdrv_co_block_status_above() for details.
2394 */
2395 static int bdrv_common_block_status_above(BlockDriverState *bs,
2396 BlockDriverState *base,
2397 bool want_zero, int64_t offset,
2398 int64_t bytes, int64_t *pnum,
2399 int64_t *map,
2400 BlockDriverState **file)
2401 {
2402 Coroutine *co;
2403 BdrvCoBlockStatusData data = {
2404 .bs = bs,
2405 .base = base,
2406 .want_zero = want_zero,
2407 .offset = offset,
2408 .bytes = bytes,
2409 .pnum = pnum,
2410 .map = map,
2411 .file = file,
2412 .done = false,
2413 };
2414
2415 if (qemu_in_coroutine()) {
2416 /* Fast-path if already in coroutine context */
2417 bdrv_block_status_above_co_entry(&data);
2418 } else {
2419 co = qemu_coroutine_create(bdrv_block_status_above_co_entry, &data);
2420 bdrv_coroutine_enter(bs, co);
2421 BDRV_POLL_WHILE(bs, !data.done);
2422 }
2423 return data.ret;
2424 }
2425
2426 int bdrv_block_status_above(BlockDriverState *bs, BlockDriverState *base,
2427 int64_t offset, int64_t bytes, int64_t *pnum,
2428 int64_t *map, BlockDriverState **file)
2429 {
2430 return bdrv_common_block_status_above(bs, base, true, offset, bytes,
2431 pnum, map, file);
2432 }
2433
2434 int bdrv_block_status(BlockDriverState *bs, int64_t offset, int64_t bytes,
2435 int64_t *pnum, int64_t *map, BlockDriverState **file)
2436 {
2437 return bdrv_block_status_above(bs, backing_bs(bs),
2438 offset, bytes, pnum, map, file);
2439 }
2440
2441 int coroutine_fn bdrv_is_allocated(BlockDriverState *bs, int64_t offset,
2442 int64_t bytes, int64_t *pnum)
2443 {
2444 int ret;
2445 int64_t dummy;
2446
2447 ret = bdrv_common_block_status_above(bs, backing_bs(bs), false, offset,
2448 bytes, pnum ? pnum : &dummy, NULL,
2449 NULL);
2450 if (ret < 0) {
2451 return ret;
2452 }
2453 return !!(ret & BDRV_BLOCK_ALLOCATED);
2454 }
2455
2456 /*
2457 * Given an image chain: ... -> [BASE] -> [INTER1] -> [INTER2] -> [TOP]
2458 *
2459 * Return 1 if (a prefix of) the given range is allocated in any image
2460 * between BASE and TOP (BASE is only included if include_base is set).
2461 * BASE can be NULL to check if the given offset is allocated in any
2462 * image of the chain. Return 0 otherwise, or negative errno on
2463 * failure.
2464 *
2465 * 'pnum' is set to the number of bytes (including and immediately
2466 * following the specified offset) that are known to be in the same
2467 * allocated/unallocated state. Note that a subsequent call starting
2468 * at 'offset + *pnum' may return the same allocation status (in other
2469 * words, the result is not necessarily the maximum possible range);
2470 * but 'pnum' will only be 0 when end of file is reached.
2471 *
2472 */
2473 int bdrv_is_allocated_above(BlockDriverState *top,
2474 BlockDriverState *base,
2475 bool include_base, int64_t offset,
2476 int64_t bytes, int64_t *pnum)
2477 {
2478 BlockDriverState *intermediate;
2479 int ret;
2480 int64_t n = bytes;
2481
2482 assert(base || !include_base);
2483
2484 intermediate = top;
2485 while (include_base || intermediate != base) {
2486 int64_t pnum_inter;
2487 int64_t size_inter;
2488
2489 assert(intermediate);
2490 ret = bdrv_is_allocated(intermediate, offset, bytes, &pnum_inter);
2491 if (ret < 0) {
2492 return ret;
2493 }
2494 if (ret) {
2495 *pnum = pnum_inter;
2496 return 1;
2497 }
2498
2499 size_inter = bdrv_getlength(intermediate);
2500 if (size_inter < 0) {
2501 return size_inter;
2502 }
2503 if (n > pnum_inter &&
2504 (intermediate == top || offset + pnum_inter < size_inter)) {
2505 n = pnum_inter;
2506 }
2507
2508 if (intermediate == base) {
2509 break;
2510 }
2511
2512 intermediate = backing_bs(intermediate);
2513 }
2514
2515 *pnum = n;
2516 return 0;
2517 }
2518
2519 typedef struct BdrvVmstateCo {
2520 BlockDriverState *bs;
2521 QEMUIOVector *qiov;
2522 int64_t pos;
2523 bool is_read;
2524 int ret;
2525 } BdrvVmstateCo;
2526
2527 static int coroutine_fn
2528 bdrv_co_rw_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos,
2529 bool is_read)
2530 {
2531 BlockDriver *drv = bs->drv;
2532 int ret = -ENOTSUP;
2533
2534 bdrv_inc_in_flight(bs);
2535
2536 if (!drv) {
2537 ret = -ENOMEDIUM;
2538 } else if (drv->bdrv_load_vmstate) {
2539 if (is_read) {
2540 ret = drv->bdrv_load_vmstate(bs, qiov, pos);
2541 } else {
2542 ret = drv->bdrv_save_vmstate(bs, qiov, pos);
2543 }
2544 } else if (bs->file) {
2545 ret = bdrv_co_rw_vmstate(bs->file->bs, qiov, pos, is_read);
2546 }
2547
2548 bdrv_dec_in_flight(bs);
2549 return ret;
2550 }
2551
2552 static void coroutine_fn bdrv_co_rw_vmstate_entry(void *opaque)
2553 {
2554 BdrvVmstateCo *co = opaque;
2555 co->ret = bdrv_co_rw_vmstate(co->bs, co->qiov, co->pos, co->is_read);
2556 aio_wait_kick();
2557 }
2558
2559 static inline int
2560 bdrv_rw_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos,
2561 bool is_read)
2562 {
2563 if (qemu_in_coroutine()) {
2564 return bdrv_co_rw_vmstate(bs, qiov, pos, is_read);
2565 } else {
2566 BdrvVmstateCo data = {
2567 .bs = bs,
2568 .qiov = qiov,
2569 .pos = pos,
2570 .is_read = is_read,
2571 .ret = -EINPROGRESS,
2572 };
2573 Coroutine *co = qemu_coroutine_create(bdrv_co_rw_vmstate_entry, &data);
2574
2575 bdrv_coroutine_enter(bs, co);
2576 BDRV_POLL_WHILE(bs, data.ret == -EINPROGRESS);
2577 return data.ret;
2578 }
2579 }
2580
2581 int bdrv_save_vmstate(BlockDriverState *bs, const uint8_t *buf,
2582 int64_t pos, int size)
2583 {
2584 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, size);
2585 int ret;
2586
2587 ret = bdrv_writev_vmstate(bs, &qiov, pos);
2588 if (ret < 0) {
2589 return ret;
2590 }
2591
2592 return size;
2593 }
2594
2595 int bdrv_writev_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos)
2596 {
2597 return bdrv_rw_vmstate(bs, qiov, pos, false);
2598 }
2599
2600 int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf,
2601 int64_t pos, int size)
2602 {
2603 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, size);
2604 int ret;
2605
2606 ret = bdrv_readv_vmstate(bs, &qiov, pos);
2607 if (ret < 0) {
2608 return ret;
2609 }
2610
2611 return size;
2612 }
2613
2614 int bdrv_readv_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos)
2615 {
2616 return bdrv_rw_vmstate(bs, qiov, pos, true);
2617 }
2618
2619 /**************************************************************/
2620 /* async I/Os */
2621
2622 void bdrv_aio_cancel(BlockAIOCB *acb)
2623 {
2624 qemu_aio_ref(acb);
2625 bdrv_aio_cancel_async(acb);
2626 while (acb->refcnt > 1) {
2627 if (acb->aiocb_info->get_aio_context) {
2628 aio_poll(acb->aiocb_info->get_aio_context(acb), true);
2629 } else if (acb->bs) {
2630 /* qemu_aio_ref and qemu_aio_unref are not thread-safe, so
2631 * assert that we're not using an I/O thread. Thread-safe
2632 * code should use bdrv_aio_cancel_async exclusively.
2633 */
2634 assert(bdrv_get_aio_context(acb->bs) == qemu_get_aio_context());
2635 aio_poll(bdrv_get_aio_context(acb->bs), true);
2636 } else {
2637 abort();
2638 }
2639 }
2640 qemu_aio_unref(acb);
2641 }
2642
2643 /* Async version of aio cancel. The caller is not blocked if the acb implements
2644 * cancel_async, otherwise we do nothing and let the request normally complete.
2645 * In either case the completion callback must be called. */
2646 void bdrv_aio_cancel_async(BlockAIOCB *acb)
2647 {
2648 if (acb->aiocb_info->cancel_async) {
2649 acb->aiocb_info->cancel_async(acb);
2650 }
2651 }
2652
2653 /**************************************************************/
2654 /* Coroutine block device emulation */
2655
2656 typedef struct FlushCo {
2657 BlockDriverState *bs;
2658 int ret;
2659 } FlushCo;
2660
2661
2662 static void coroutine_fn bdrv_flush_co_entry(void *opaque)
2663 {
2664 FlushCo *rwco = opaque;
2665
2666 rwco->ret = bdrv_co_flush(rwco->bs);
2667 aio_wait_kick();
2668 }
2669
2670 int coroutine_fn bdrv_co_flush(BlockDriverState *bs)
2671 {
2672 int current_gen;
2673 int ret = 0;
2674
2675 bdrv_inc_in_flight(bs);
2676
2677 if (!bdrv_is_inserted(bs) || bdrv_is_read_only(bs) ||
2678 bdrv_is_sg(bs)) {
2679 goto early_exit;
2680 }
2681
2682 qemu_co_mutex_lock(&bs->reqs_lock);
2683 current_gen = atomic_read(&bs->write_gen);
2684
2685 /* Wait until any previous flushes are completed */
2686 while (bs->active_flush_req) {
2687 qemu_co_queue_wait(&bs->flush_queue, &bs->reqs_lock);
2688 }
2689
2690 /* Flushes reach this point in nondecreasing current_gen order. */
2691 bs->active_flush_req = true;
2692 qemu_co_mutex_unlock(&bs->reqs_lock);
2693
2694 /* Write back all layers by calling one driver function */
2695 if (bs->drv->bdrv_co_flush) {
2696 ret = bs->drv->bdrv_co_flush(bs);
2697 goto out;
2698 }
2699
2700 /* Write back cached data to the OS even with cache=unsafe */
2701 BLKDBG_EVENT(bs->file, BLKDBG_FLUSH_TO_OS);
2702 if (bs->drv->bdrv_co_flush_to_os) {
2703 ret = bs->drv->bdrv_co_flush_to_os(bs);
2704 if (ret < 0) {
2705 goto out;
2706 }
2707 }
2708
2709 /* But don't actually force it to the disk with cache=unsafe */
2710 if (bs->open_flags & BDRV_O_NO_FLUSH) {
2711 goto flush_parent;
2712 }
2713
2714 /* Check if we really need to flush anything */
2715 if (bs->flushed_gen == current_gen) {
2716 goto flush_parent;
2717 }
2718
2719 BLKDBG_EVENT(bs->file, BLKDBG_FLUSH_TO_DISK);
2720 if (!bs->drv) {
2721 /* bs->drv->bdrv_co_flush() might have ejected the BDS
2722 * (even in case of apparent success) */
2723 ret = -ENOMEDIUM;
2724 goto out;
2725 }
2726 if (bs->drv->bdrv_co_flush_to_disk) {
2727 ret = bs->drv->bdrv_co_flush_to_disk(bs);
2728 } else if (bs->drv->bdrv_aio_flush) {
2729 BlockAIOCB *acb;
2730 CoroutineIOCompletion co = {
2731 .coroutine = qemu_coroutine_self(),
2732 };
2733
2734 acb = bs->drv->bdrv_aio_flush(bs, bdrv_co_io_em_complete, &co);
2735 if (acb == NULL) {
2736 ret = -EIO;
2737 } else {
2738 qemu_coroutine_yield();
2739 ret = co.ret;
2740 }
2741 } else {
2742 /*
2743 * Some block drivers always operate in either writethrough or unsafe
2744 * mode and don't support bdrv_flush therefore. Usually qemu doesn't
2745 * know how the server works (because the behaviour is hardcoded or
2746 * depends on server-side configuration), so we can't ensure that
2747 * everything is safe on disk. Returning an error doesn't work because
2748 * that would break guests even if the server operates in writethrough
2749 * mode.
2750 *
2751 * Let's hope the user knows what he's doing.
2752 */
2753 ret = 0;
2754 }
2755
2756 if (ret < 0) {
2757 goto out;
2758 }
2759
2760 /* Now flush the underlying protocol. It will also have BDRV_O_NO_FLUSH
2761 * in the case of cache=unsafe, so there are no useless flushes.
2762 */
2763 flush_parent:
2764 ret = bs->file ? bdrv_co_flush(bs->file->bs) : 0;
2765 out:
2766 /* Notify any pending flushes that we have completed */
2767 if (ret == 0) {
2768 bs->flushed_gen = current_gen;
2769 }
2770
2771 qemu_co_mutex_lock(&bs->reqs_lock);
2772 bs->active_flush_req = false;
2773 /* Return value is ignored - it's ok if wait queue is empty */
2774 qemu_co_queue_next(&bs->flush_queue);
2775 qemu_co_mutex_unlock(&bs->reqs_lock);
2776
2777 early_exit:
2778 bdrv_dec_in_flight(bs);
2779 return ret;
2780 }
2781
2782 int bdrv_flush(BlockDriverState *bs)
2783 {
2784 Coroutine *co;
2785 FlushCo flush_co = {
2786 .bs = bs,
2787 .ret = NOT_DONE,
2788 };
2789
2790 if (qemu_in_coroutine()) {
2791 /* Fast-path if already in coroutine context */
2792 bdrv_flush_co_entry(&flush_co);
2793 } else {
2794 co = qemu_coroutine_create(bdrv_flush_co_entry, &flush_co);
2795 bdrv_coroutine_enter(bs, co);
2796 BDRV_POLL_WHILE(bs, flush_co.ret == NOT_DONE);
2797 }
2798
2799 return flush_co.ret;
2800 }
2801
2802 typedef struct DiscardCo {
2803 BdrvChild *child;
2804 int64_t offset;
2805 int64_t bytes;
2806 int ret;
2807 } DiscardCo;
2808 static void coroutine_fn bdrv_pdiscard_co_entry(void *opaque)
2809 {
2810 DiscardCo *rwco = opaque;
2811
2812 rwco->ret = bdrv_co_pdiscard(rwco->child, rwco->offset, rwco->bytes);
2813 aio_wait_kick();
2814 }
2815
2816 int coroutine_fn bdrv_co_pdiscard(BdrvChild *child, int64_t offset,
2817 int64_t bytes)
2818 {
2819 BdrvTrackedRequest req;
2820 int max_pdiscard, ret;
2821 int head, tail, align;
2822 BlockDriverState *bs = child->bs;
2823
2824 if (!bs || !bs->drv || !bdrv_is_inserted(bs)) {
2825 return -ENOMEDIUM;
2826 }
2827
2828 if (bdrv_has_readonly_bitmaps(bs)) {
2829 return -EPERM;
2830 }
2831
2832 if (offset < 0 || bytes < 0 || bytes > INT64_MAX - offset) {
2833 return -EIO;
2834 }
2835
2836 /* Do nothing if disabled. */
2837 if (!(bs->open_flags & BDRV_O_UNMAP)) {
2838 return 0;
2839 }
2840
2841 if (!bs->drv->bdrv_co_pdiscard && !bs->drv->bdrv_aio_pdiscard) {
2842 return 0;
2843 }
2844
2845 /* Discard is advisory, but some devices track and coalesce
2846 * unaligned requests, so we must pass everything down rather than
2847 * round here. Still, most devices will just silently ignore
2848 * unaligned requests (by returning -ENOTSUP), so we must fragment
2849 * the request accordingly. */
2850 align = MAX(bs->bl.pdiscard_alignment, bs->bl.request_alignment);
2851 assert(align % bs->bl.request_alignment == 0);
2852 head = offset % align;
2853 tail = (offset + bytes) % align;
2854
2855 bdrv_inc_in_flight(bs);
2856 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_DISCARD);
2857
2858 ret = bdrv_co_write_req_prepare(child, offset, bytes, &req, 0);
2859 if (ret < 0) {
2860 goto out;
2861 }
2862
2863 max_pdiscard = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_pdiscard, INT_MAX),
2864 align);
2865 assert(max_pdiscard >= bs->bl.request_alignment);
2866
2867 while (bytes > 0) {
2868 int64_t num = bytes;
2869
2870 if (head) {
2871 /* Make small requests to get to alignment boundaries. */
2872 num = MIN(bytes, align - head);
2873 if (!QEMU_IS_ALIGNED(num, bs->bl.request_alignment)) {
2874 num %= bs->bl.request_alignment;
2875 }
2876 head = (head + num) % align;
2877 assert(num < max_pdiscard);
2878 } else if (tail) {
2879 if (num > align) {
2880 /* Shorten the request to the last aligned cluster. */
2881 num -= tail;
2882 } else if (!QEMU_IS_ALIGNED(tail, bs->bl.request_alignment) &&
2883 tail > bs->bl.request_alignment) {
2884 tail %= bs->bl.request_alignment;
2885 num -= tail;
2886 }
2887 }
2888 /* limit request size */
2889 if (num > max_pdiscard) {
2890 num = max_pdiscard;
2891 }
2892
2893 if (!bs->drv) {
2894 ret = -ENOMEDIUM;
2895 goto out;
2896 }
2897 if (bs->drv->bdrv_co_pdiscard) {
2898 ret = bs->drv->bdrv_co_pdiscard(bs, offset, num);
2899 } else {
2900 BlockAIOCB *acb;
2901 CoroutineIOCompletion co = {
2902 .coroutine = qemu_coroutine_self(),
2903 };
2904
2905 acb = bs->drv->bdrv_aio_pdiscard(bs, offset, num,
2906 bdrv_co_io_em_complete, &co);
2907 if (acb == NULL) {
2908 ret = -EIO;
2909 goto out;
2910 } else {
2911 qemu_coroutine_yield();
2912 ret = co.ret;
2913 }
2914 }
2915 if (ret && ret != -ENOTSUP) {
2916 goto out;
2917 }
2918
2919 offset += num;
2920 bytes -= num;
2921 }
2922 ret = 0;
2923 out:
2924 bdrv_co_write_req_finish(child, req.offset, req.bytes, &req, ret);
2925 tracked_request_end(&req);
2926 bdrv_dec_in_flight(bs);
2927 return ret;
2928 }
2929
2930 int bdrv_pdiscard(BdrvChild *child, int64_t offset, int64_t bytes)
2931 {
2932 Coroutine *co;
2933 DiscardCo rwco = {
2934 .child = child,
2935 .offset = offset,
2936 .bytes = bytes,
2937 .ret = NOT_DONE,
2938 };
2939
2940 if (qemu_in_coroutine()) {
2941 /* Fast-path if already in coroutine context */
2942 bdrv_pdiscard_co_entry(&rwco);
2943 } else {
2944 co = qemu_coroutine_create(bdrv_pdiscard_co_entry, &rwco);
2945 bdrv_coroutine_enter(child->bs, co);
2946 BDRV_POLL_WHILE(child->bs, rwco.ret == NOT_DONE);
2947 }
2948
2949 return rwco.ret;
2950 }
2951
2952 int bdrv_co_ioctl(BlockDriverState *bs, int req, void *buf)
2953 {
2954 BlockDriver *drv = bs->drv;
2955 CoroutineIOCompletion co = {
2956 .coroutine = qemu_coroutine_self(),
2957 };
2958 BlockAIOCB *acb;
2959
2960 bdrv_inc_in_flight(bs);
2961 if (!drv || (!drv->bdrv_aio_ioctl && !drv->bdrv_co_ioctl)) {
2962 co.ret = -ENOTSUP;
2963 goto out;
2964 }
2965
2966 if (drv->bdrv_co_ioctl) {
2967 co.ret = drv->bdrv_co_ioctl(bs, req, buf);
2968 } else {
2969 acb = drv->bdrv_aio_ioctl(bs, req, buf, bdrv_co_io_em_complete, &co);
2970 if (!acb) {
2971 co.ret = -ENOTSUP;
2972 goto out;
2973 }
2974 qemu_coroutine_yield();
2975 }
2976 out:
2977 bdrv_dec_in_flight(bs);
2978 return co.ret;
2979 }
2980
2981 void *qemu_blockalign(BlockDriverState *bs, size_t size)
2982 {
2983 return qemu_memalign(bdrv_opt_mem_align(bs), size);
2984 }
2985
2986 void *qemu_blockalign0(BlockDriverState *bs, size_t size)
2987 {
2988 return memset(qemu_blockalign(bs, size), 0, size);
2989 }
2990
2991 void *qemu_try_blockalign(BlockDriverState *bs, size_t size)
2992 {
2993 size_t align = bdrv_opt_mem_align(bs);
2994
2995 /* Ensure that NULL is never returned on success */
2996 assert(align > 0);
2997 if (size == 0) {
2998 size = align;
2999 }
3000
3001 return qemu_try_memalign(align, size);
3002 }
3003
3004 void *qemu_try_blockalign0(BlockDriverState *bs, size_t size)
3005 {
3006 void *mem = qemu_try_blockalign(bs, size);
3007
3008 if (mem) {
3009 memset(mem, 0, size);
3010 }
3011
3012 return mem;
3013 }
3014
3015 /*
3016 * Check if all memory in this vector is sector aligned.
3017 */
3018 bool bdrv_qiov_is_aligned(BlockDriverState *bs, QEMUIOVector *qiov)
3019 {
3020 int i;
3021 size_t alignment = bdrv_min_mem_align(bs);
3022
3023 for (i = 0; i < qiov->niov; i++) {
3024 if ((uintptr_t) qiov->iov[i].iov_base % alignment) {
3025 return false;
3026 }
3027 if (qiov->iov[i].iov_len % alignment) {
3028 return false;
3029 }
3030 }
3031
3032 return true;
3033 }
3034
3035 void bdrv_add_before_write_notifier(BlockDriverState *bs,
3036 NotifierWithReturn *notifier)
3037 {
3038 notifier_with_return_list_add(&bs->before_write_notifiers, notifier);
3039 }
3040
3041 void bdrv_io_plug(BlockDriverState *bs)
3042 {
3043 BdrvChild *child;
3044
3045 QLIST_FOREACH(child, &bs->children, next) {
3046 bdrv_io_plug(child->bs);
3047 }
3048
3049 if (atomic_fetch_inc(&bs->io_plugged) == 0) {
3050 BlockDriver *drv = bs->drv;
3051 if (drv && drv->bdrv_io_plug) {
3052 drv->bdrv_io_plug(bs);
3053 }
3054 }
3055 }
3056
3057 void bdrv_io_unplug(BlockDriverState *bs)
3058 {
3059 BdrvChild *child;
3060
3061 assert(bs->io_plugged);
3062 if (atomic_fetch_dec(&bs->io_plugged) == 1) {
3063 BlockDriver *drv = bs->drv;
3064 if (drv && drv->bdrv_io_unplug) {
3065 drv->bdrv_io_unplug(bs);
3066 }
3067 }
3068
3069 QLIST_FOREACH(child, &bs->children, next) {
3070 bdrv_io_unplug(child->bs);
3071 }
3072 }
3073
3074 void bdrv_register_buf(BlockDriverState *bs, void *host, size_t size)
3075 {
3076 BdrvChild *child;
3077
3078 if (bs->drv && bs->drv->bdrv_register_buf) {
3079 bs->drv->bdrv_register_buf(bs, host, size);
3080 }
3081 QLIST_FOREACH(child, &bs->children, next) {
3082 bdrv_register_buf(child->bs, host, size);
3083 }
3084 }
3085
3086 void bdrv_unregister_buf(BlockDriverState *bs, void *host)
3087 {
3088 BdrvChild *child;
3089
3090 if (bs->drv && bs->drv->bdrv_unregister_buf) {
3091 bs->drv->bdrv_unregister_buf(bs, host);
3092 }
3093 QLIST_FOREACH(child, &bs->children, next) {
3094 bdrv_unregister_buf(child->bs, host);
3095 }
3096 }
3097
3098 static int coroutine_fn bdrv_co_copy_range_internal(
3099 BdrvChild *src, uint64_t src_offset, BdrvChild *dst,
3100 uint64_t dst_offset, uint64_t bytes,
3101 BdrvRequestFlags read_flags, BdrvRequestFlags write_flags,
3102 bool recurse_src)
3103 {
3104 BdrvTrackedRequest req;
3105 int ret;
3106
3107 /* TODO We can support BDRV_REQ_NO_FALLBACK here */
3108 assert(!(read_flags & BDRV_REQ_NO_FALLBACK));
3109 assert(!(write_flags & BDRV_REQ_NO_FALLBACK));
3110
3111 if (!dst || !dst->bs) {
3112 return -ENOMEDIUM;
3113 }
3114 ret = bdrv_check_byte_request(dst->bs, dst_offset, bytes);
3115 if (ret) {
3116 return ret;
3117 }
3118 if (write_flags & BDRV_REQ_ZERO_WRITE) {
3119 return bdrv_co_pwrite_zeroes(dst, dst_offset, bytes, write_flags);
3120 }
3121
3122 if (!src || !src->bs) {
3123 return -ENOMEDIUM;
3124 }
3125 ret = bdrv_check_byte_request(src->bs, src_offset, bytes);
3126 if (ret) {
3127 return ret;
3128 }
3129
3130 if (!src->bs->drv->bdrv_co_copy_range_from
3131 || !dst->bs->drv->bdrv_co_copy_range_to
3132 || src->bs->encrypted || dst->bs->encrypted) {
3133 return -ENOTSUP;
3134 }
3135
3136 if (recurse_src) {
3137 bdrv_inc_in_flight(src->bs);
3138 tracked_request_begin(&req, src->bs, src_offset, bytes,
3139 BDRV_TRACKED_READ);
3140
3141 /* BDRV_REQ_SERIALISING is only for write operation */
3142 assert(!(read_flags & BDRV_REQ_SERIALISING));
3143 if (!(read_flags & BDRV_REQ_NO_SERIALISING)) {
3144 wait_serialising_requests(&req);
3145 }
3146
3147 ret = src->bs->drv->bdrv_co_copy_range_from(src->bs,
3148 src, src_offset,
3149 dst, dst_offset,
3150 bytes,
3151 read_flags, write_flags);
3152
3153 tracked_request_end(&req);
3154 bdrv_dec_in_flight(src->bs);
3155 } else {
3156 bdrv_inc_in_flight(dst->bs);
3157 tracked_request_begin(&req, dst->bs, dst_offset, bytes,
3158 BDRV_TRACKED_WRITE);
3159 ret = bdrv_co_write_req_prepare(dst, dst_offset, bytes, &req,
3160 write_flags);
3161 if (!ret) {
3162 ret = dst->bs->drv->bdrv_co_copy_range_to(dst->bs,
3163 src, src_offset,
3164 dst, dst_offset,
3165 bytes,
3166 read_flags, write_flags);
3167 }
3168 bdrv_co_write_req_finish(dst, dst_offset, bytes, &req, ret);
3169 tracked_request_end(&req);
3170 bdrv_dec_in_flight(dst->bs);
3171 }
3172
3173 return ret;
3174 }
3175
3176 /* Copy range from @src to @dst.
3177 *
3178 * See the comment of bdrv_co_copy_range for the parameter and return value
3179 * semantics. */
3180 int coroutine_fn bdrv_co_copy_range_from(BdrvChild *src, uint64_t src_offset,
3181 BdrvChild *dst, uint64_t dst_offset,
3182 uint64_t bytes,
3183 BdrvRequestFlags read_flags,
3184 BdrvRequestFlags write_flags)
3185 {
3186 trace_bdrv_co_copy_range_from(src, src_offset, dst, dst_offset, bytes,
3187 read_flags, write_flags);
3188 return bdrv_co_copy_range_internal(src, src_offset, dst, dst_offset,
3189 bytes, read_flags, write_flags, true);
3190 }
3191
3192 /* Copy range from @src to @dst.
3193 *
3194 * See the comment of bdrv_co_copy_range for the parameter and return value
3195 * semantics. */
3196 int coroutine_fn bdrv_co_copy_range_to(BdrvChild *src, uint64_t src_offset,
3197 BdrvChild *dst, uint64_t dst_offset,
3198 uint64_t bytes,
3199 BdrvRequestFlags read_flags,
3200 BdrvRequestFlags write_flags)
3201 {
3202 trace_bdrv_co_copy_range_to(src, src_offset, dst, dst_offset, bytes,
3203 read_flags, write_flags);
3204 return bdrv_co_copy_range_internal(src, src_offset, dst, dst_offset,
3205 bytes, read_flags, write_flags, false);
3206 }
3207
3208 int coroutine_fn bdrv_co_copy_range(BdrvChild *src, uint64_t src_offset,
3209 BdrvChild *dst, uint64_t dst_offset,
3210 uint64_t bytes, BdrvRequestFlags read_flags,
3211 BdrvRequestFlags write_flags)
3212 {
3213 return bdrv_co_copy_range_from(src, src_offset,
3214 dst, dst_offset,
3215 bytes, read_flags, write_flags);
3216 }
3217
3218 static void bdrv_parent_cb_resize(BlockDriverState *bs)
3219 {
3220 BdrvChild *c;
3221 QLIST_FOREACH(c, &bs->parents, next_parent) {
3222 if (c->role->resize) {
3223 c->role->resize(c);
3224 }
3225 }
3226 }
3227
3228 /**
3229 * Truncate file to 'offset' bytes (needed only for file protocols)
3230 */
3231 int coroutine_fn bdrv_co_truncate(BdrvChild *child, int64_t offset,
3232 PreallocMode prealloc, Error **errp)
3233 {
3234 BlockDriverState *bs = child->bs;
3235 BlockDriver *drv = bs->drv;
3236 BdrvTrackedRequest req;
3237 int64_t old_size, new_bytes;
3238 int ret;
3239
3240
3241 /* if bs->drv == NULL, bs is closed, so there's nothing to do here */
3242 if (!drv) {
3243 error_setg(errp, "No medium inserted");
3244 return -ENOMEDIUM;
3245 }
3246 if (offset < 0) {
3247 error_setg(errp, "Image size cannot be negative");
3248 return -EINVAL;
3249 }
3250
3251 old_size = bdrv_getlength(bs);
3252 if (old_size < 0) {
3253 error_setg_errno(errp, -old_size, "Failed to get old image size");
3254 return old_size;
3255 }
3256
3257 if (offset > old_size) {
3258 new_bytes = offset - old_size;
3259 } else {
3260 new_bytes = 0;
3261 }
3262
3263 bdrv_inc_in_flight(bs);
3264 tracked_request_begin(&req, bs, offset - new_bytes, new_bytes,
3265 BDRV_TRACKED_TRUNCATE);
3266
3267 /* If we are growing the image and potentially using preallocation for the
3268 * new area, we need to make sure that no write requests are made to it
3269 * concurrently or they might be overwritten by preallocation. */
3270 if (new_bytes) {
3271 mark_request_serialising(&req, 1);
3272 }
3273 if (bs->read_only) {
3274 error_setg(errp, "Image is read-only");
3275 ret = -EACCES;
3276 goto out;
3277 }
3278 ret = bdrv_co_write_req_prepare(child, offset - new_bytes, new_bytes, &req,
3279 0);
3280 if (ret < 0) {
3281 error_setg_errno(errp, -ret,
3282 "Failed to prepare request for truncation");
3283 goto out;
3284 }
3285
3286 if (!drv->bdrv_co_truncate) {
3287 if (bs->file && drv->is_filter) {
3288 ret = bdrv_co_truncate(bs->file, offset, prealloc, errp);
3289 goto out;
3290 }
3291 error_setg(errp, "Image format driver does not support resize");
3292 ret = -ENOTSUP;
3293 goto out;
3294 }
3295
3296 ret = drv->bdrv_co_truncate(bs, offset, prealloc, errp);
3297 if (ret < 0) {
3298 goto out;
3299 }
3300 ret = refresh_total_sectors(bs, offset >> BDRV_SECTOR_BITS);
3301 if (ret < 0) {
3302 error_setg_errno(errp, -ret, "Could not refresh total sector count");
3303 } else {
3304 offset = bs->total_sectors * BDRV_SECTOR_SIZE;
3305 }
3306 /* It's possible that truncation succeeded but refresh_total_sectors
3307 * failed, but the latter doesn't affect how we should finish the request.
3308 * Pass 0 as the last parameter so that dirty bitmaps etc. are handled. */
3309 bdrv_co_write_req_finish(child, offset - new_bytes, new_bytes, &req, 0);
3310
3311 out:
3312 tracked_request_end(&req);
3313 bdrv_dec_in_flight(bs);
3314
3315 return ret;
3316 }
3317
3318 typedef struct TruncateCo {
3319 BdrvChild *child;
3320 int64_t offset;
3321 PreallocMode prealloc;
3322 Error **errp;
3323 int ret;
3324 } TruncateCo;
3325
3326 static void coroutine_fn bdrv_truncate_co_entry(void *opaque)
3327 {
3328 TruncateCo *tco = opaque;
3329 tco->ret = bdrv_co_truncate(tco->child, tco->offset, tco->prealloc,
3330 tco->errp);
3331 aio_wait_kick();
3332 }
3333
3334 int bdrv_truncate(BdrvChild *child, int64_t offset, PreallocMode prealloc,
3335 Error **errp)
3336 {
3337 Coroutine *co;
3338 TruncateCo tco = {
3339 .child = child,
3340 .offset = offset,
3341 .prealloc = prealloc,
3342 .errp = errp,
3343 .ret = NOT_DONE,
3344 };
3345
3346 if (qemu_in_coroutine()) {
3347 /* Fast-path if already in coroutine context */
3348 bdrv_truncate_co_entry(&tco);
3349 } else {
3350 co = qemu_coroutine_create(bdrv_truncate_co_entry, &tco);
3351 bdrv_coroutine_enter(child->bs, co);
3352 BDRV_POLL_WHILE(child->bs, tco.ret == NOT_DONE);
3353 }
3354
3355 return tco.ret;
3356 }