block/io: bdrv_co_do_copy_on_readv: 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;
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 bounce_buffer = qemu_try_blockalign(bs,
1275 MIN(MIN(max_transfer, cluster_bytes),
1276 MAX_BOUNCE_BUFFER));
1277 if (bounce_buffer == NULL) {
1278 ret = -ENOMEM;
1279 goto err;
1280 }
1281
1282 while (cluster_bytes) {
1283 int64_t pnum;
1284
1285 ret = bdrv_is_allocated(bs, cluster_offset,
1286 MIN(cluster_bytes, max_transfer), &pnum);
1287 if (ret < 0) {
1288 /* Safe to treat errors in querying allocation as if
1289 * unallocated; we'll probably fail again soon on the
1290 * read, but at least that will set a decent errno.
1291 */
1292 pnum = MIN(cluster_bytes, max_transfer);
1293 }
1294
1295 /* Stop at EOF if the image ends in the middle of the cluster */
1296 if (ret == 0 && pnum == 0) {
1297 assert(progress >= bytes);
1298 break;
1299 }
1300
1301 assert(skip_bytes < pnum);
1302
1303 if (ret <= 0) {
1304 QEMUIOVector local_qiov;
1305
1306 /* Must copy-on-read; use the bounce buffer */
1307 pnum = MIN(pnum, MAX_BOUNCE_BUFFER);
1308 qemu_iovec_init_buf(&local_qiov, bounce_buffer, pnum);
1309
1310 ret = bdrv_driver_preadv(bs, cluster_offset, pnum,
1311 &local_qiov, 0, 0);
1312 if (ret < 0) {
1313 goto err;
1314 }
1315
1316 bdrv_debug_event(bs, BLKDBG_COR_WRITE);
1317 if (drv->bdrv_co_pwrite_zeroes &&
1318 buffer_is_zero(bounce_buffer, pnum)) {
1319 /* FIXME: Should we (perhaps conditionally) be setting
1320 * BDRV_REQ_MAY_UNMAP, if it will allow for a sparser copy
1321 * that still correctly reads as zero? */
1322 ret = bdrv_co_do_pwrite_zeroes(bs, cluster_offset, pnum,
1323 BDRV_REQ_WRITE_UNCHANGED);
1324 } else {
1325 /* This does not change the data on the disk, it is not
1326 * necessary to flush even in cache=writethrough mode.
1327 */
1328 ret = bdrv_driver_pwritev(bs, cluster_offset, pnum,
1329 &local_qiov, 0,
1330 BDRV_REQ_WRITE_UNCHANGED);
1331 }
1332
1333 if (ret < 0) {
1334 /* It might be okay to ignore write errors for guest
1335 * requests. If this is a deliberate copy-on-read
1336 * then we don't want to ignore the error. Simply
1337 * report it in all cases.
1338 */
1339 goto err;
1340 }
1341
1342 if (!(flags & BDRV_REQ_PREFETCH)) {
1343 qemu_iovec_from_buf(qiov, qiov_offset + progress,
1344 bounce_buffer + skip_bytes,
1345 pnum - skip_bytes);
1346 }
1347 } else if (!(flags & BDRV_REQ_PREFETCH)) {
1348 /* Read directly into the destination */
1349 ret = bdrv_driver_preadv(bs, offset + progress,
1350 MIN(pnum - skip_bytes, bytes - progress),
1351 qiov, qiov_offset + progress, 0);
1352 if (ret < 0) {
1353 goto err;
1354 }
1355 }
1356
1357 cluster_offset += pnum;
1358 cluster_bytes -= pnum;
1359 progress += pnum - skip_bytes;
1360 skip_bytes = 0;
1361 }
1362 ret = 0;
1363
1364 err:
1365 qemu_vfree(bounce_buffer);
1366 return ret;
1367 }
1368
1369 /*
1370 * Forwards an already correctly aligned request to the BlockDriver. This
1371 * handles copy on read, zeroing after EOF, and fragmentation of large
1372 * reads; any other features must be implemented by the caller.
1373 */
1374 static int coroutine_fn bdrv_aligned_preadv(BdrvChild *child,
1375 BdrvTrackedRequest *req, int64_t offset, unsigned int bytes,
1376 int64_t align, QEMUIOVector *qiov, int flags)
1377 {
1378 BlockDriverState *bs = child->bs;
1379 int64_t total_bytes, max_bytes;
1380 int ret = 0;
1381 uint64_t bytes_remaining = bytes;
1382 int max_transfer;
1383
1384 assert(is_power_of_2(align));
1385 assert((offset & (align - 1)) == 0);
1386 assert((bytes & (align - 1)) == 0);
1387 assert(!qiov || bytes == qiov->size);
1388 assert((bs->open_flags & BDRV_O_NO_IO) == 0);
1389 max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX),
1390 align);
1391
1392 /* TODO: We would need a per-BDS .supported_read_flags and
1393 * potential fallback support, if we ever implement any read flags
1394 * to pass through to drivers. For now, there aren't any
1395 * passthrough flags. */
1396 assert(!(flags & ~(BDRV_REQ_NO_SERIALISING | BDRV_REQ_COPY_ON_READ |
1397 BDRV_REQ_PREFETCH)));
1398
1399 /* Handle Copy on Read and associated serialisation */
1400 if (flags & BDRV_REQ_COPY_ON_READ) {
1401 /* If we touch the same cluster it counts as an overlap. This
1402 * guarantees that allocating writes will be serialized and not race
1403 * with each other for the same cluster. For example, in copy-on-read
1404 * it ensures that the CoR read and write operations are atomic and
1405 * guest writes cannot interleave between them. */
1406 mark_request_serialising(req, bdrv_get_cluster_size(bs));
1407 }
1408
1409 /* BDRV_REQ_SERIALISING is only for write operation */
1410 assert(!(flags & BDRV_REQ_SERIALISING));
1411
1412 if (!(flags & BDRV_REQ_NO_SERIALISING)) {
1413 wait_serialising_requests(req);
1414 }
1415
1416 if (flags & BDRV_REQ_COPY_ON_READ) {
1417 int64_t pnum;
1418
1419 ret = bdrv_is_allocated(bs, offset, bytes, &pnum);
1420 if (ret < 0) {
1421 goto out;
1422 }
1423
1424 if (!ret || pnum != bytes) {
1425 ret = bdrv_co_do_copy_on_readv(child, offset, bytes, qiov, 0, flags);
1426 goto out;
1427 } else if (flags & BDRV_REQ_PREFETCH) {
1428 goto out;
1429 }
1430 }
1431
1432 /* Forward the request to the BlockDriver, possibly fragmenting it */
1433 total_bytes = bdrv_getlength(bs);
1434 if (total_bytes < 0) {
1435 ret = total_bytes;
1436 goto out;
1437 }
1438
1439 max_bytes = ROUND_UP(MAX(0, total_bytes - offset), align);
1440 if (bytes <= max_bytes && bytes <= max_transfer) {
1441 ret = bdrv_driver_preadv(bs, offset, bytes, qiov, 0, 0);
1442 goto out;
1443 }
1444
1445 while (bytes_remaining) {
1446 int num;
1447
1448 if (max_bytes) {
1449 QEMUIOVector local_qiov;
1450
1451 num = MIN(bytes_remaining, MIN(max_bytes, max_transfer));
1452 assert(num);
1453 qemu_iovec_init(&local_qiov, qiov->niov);
1454 qemu_iovec_concat(&local_qiov, qiov, bytes - bytes_remaining, num);
1455
1456 ret = bdrv_driver_preadv(bs, offset + bytes - bytes_remaining,
1457 num, &local_qiov, 0, 0);
1458 max_bytes -= num;
1459 qemu_iovec_destroy(&local_qiov);
1460 } else {
1461 num = bytes_remaining;
1462 ret = qemu_iovec_memset(qiov, bytes - bytes_remaining, 0,
1463 bytes_remaining);
1464 }
1465 if (ret < 0) {
1466 goto out;
1467 }
1468 bytes_remaining -= num;
1469 }
1470
1471 out:
1472 return ret < 0 ? ret : 0;
1473 }
1474
1475 /*
1476 * Request padding
1477 *
1478 * |<---- align ----->| |<----- align ---->|
1479 * |<- head ->|<------------- bytes ------------->|<-- tail -->|
1480 * | | | | | |
1481 * -*----------$-------*-------- ... --------*-----$------------*---
1482 * | | | | | |
1483 * | offset | | end |
1484 * ALIGN_DOWN(offset) ALIGN_UP(offset) ALIGN_DOWN(end) ALIGN_UP(end)
1485 * [buf ... ) [tail_buf )
1486 *
1487 * @buf is an aligned allocation needed to store @head and @tail paddings. @head
1488 * is placed at the beginning of @buf and @tail at the @end.
1489 *
1490 * @tail_buf is a pointer to sub-buffer, corresponding to align-sized chunk
1491 * around tail, if tail exists.
1492 *
1493 * @merge_reads is true for small requests,
1494 * if @buf_len == @head + bytes + @tail. In this case it is possible that both
1495 * head and tail exist but @buf_len == align and @tail_buf == @buf.
1496 */
1497 typedef struct BdrvRequestPadding {
1498 uint8_t *buf;
1499 size_t buf_len;
1500 uint8_t *tail_buf;
1501 size_t head;
1502 size_t tail;
1503 bool merge_reads;
1504 QEMUIOVector local_qiov;
1505 } BdrvRequestPadding;
1506
1507 static bool bdrv_init_padding(BlockDriverState *bs,
1508 int64_t offset, int64_t bytes,
1509 BdrvRequestPadding *pad)
1510 {
1511 uint64_t align = bs->bl.request_alignment;
1512 size_t sum;
1513
1514 memset(pad, 0, sizeof(*pad));
1515
1516 pad->head = offset & (align - 1);
1517 pad->tail = ((offset + bytes) & (align - 1));
1518 if (pad->tail) {
1519 pad->tail = align - pad->tail;
1520 }
1521
1522 if ((!pad->head && !pad->tail) || !bytes) {
1523 return false;
1524 }
1525
1526 sum = pad->head + bytes + pad->tail;
1527 pad->buf_len = (sum > align && pad->head && pad->tail) ? 2 * align : align;
1528 pad->buf = qemu_blockalign(bs, pad->buf_len);
1529 pad->merge_reads = sum == pad->buf_len;
1530 if (pad->tail) {
1531 pad->tail_buf = pad->buf + pad->buf_len - align;
1532 }
1533
1534 return true;
1535 }
1536
1537 static int bdrv_padding_rmw_read(BdrvChild *child,
1538 BdrvTrackedRequest *req,
1539 BdrvRequestPadding *pad,
1540 bool zero_middle)
1541 {
1542 QEMUIOVector local_qiov;
1543 BlockDriverState *bs = child->bs;
1544 uint64_t align = bs->bl.request_alignment;
1545 int ret;
1546
1547 assert(req->serialising && pad->buf);
1548
1549 if (pad->head || pad->merge_reads) {
1550 uint64_t bytes = pad->merge_reads ? pad->buf_len : align;
1551
1552 qemu_iovec_init_buf(&local_qiov, pad->buf, bytes);
1553
1554 if (pad->head) {
1555 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD);
1556 }
1557 if (pad->merge_reads && pad->tail) {
1558 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
1559 }
1560 ret = bdrv_aligned_preadv(child, req, req->overlap_offset, bytes,
1561 align, &local_qiov, 0);
1562 if (ret < 0) {
1563 return ret;
1564 }
1565 if (pad->head) {
1566 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD);
1567 }
1568 if (pad->merge_reads && pad->tail) {
1569 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
1570 }
1571
1572 if (pad->merge_reads) {
1573 goto zero_mem;
1574 }
1575 }
1576
1577 if (pad->tail) {
1578 qemu_iovec_init_buf(&local_qiov, pad->tail_buf, align);
1579
1580 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
1581 ret = bdrv_aligned_preadv(
1582 child, req,
1583 req->overlap_offset + req->overlap_bytes - align,
1584 align, align, &local_qiov, 0);
1585 if (ret < 0) {
1586 return ret;
1587 }
1588 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
1589 }
1590
1591 zero_mem:
1592 if (zero_middle) {
1593 memset(pad->buf + pad->head, 0, pad->buf_len - pad->head - pad->tail);
1594 }
1595
1596 return 0;
1597 }
1598
1599 static void bdrv_padding_destroy(BdrvRequestPadding *pad)
1600 {
1601 if (pad->buf) {
1602 qemu_vfree(pad->buf);
1603 qemu_iovec_destroy(&pad->local_qiov);
1604 }
1605 }
1606
1607 /*
1608 * bdrv_pad_request
1609 *
1610 * Exchange request parameters with padded request if needed. Don't include RMW
1611 * read of padding, bdrv_padding_rmw_read() should be called separately if
1612 * needed.
1613 *
1614 * All parameters except @bs are in-out: they represent original request at
1615 * function call and padded (if padding needed) at function finish.
1616 *
1617 * Function always succeeds.
1618 */
1619 static bool bdrv_pad_request(BlockDriverState *bs, QEMUIOVector **qiov,
1620 int64_t *offset, unsigned int *bytes,
1621 BdrvRequestPadding *pad)
1622 {
1623 if (!bdrv_init_padding(bs, *offset, *bytes, pad)) {
1624 return false;
1625 }
1626
1627 qemu_iovec_init_extended(&pad->local_qiov, pad->buf, pad->head,
1628 *qiov, 0, *bytes,
1629 pad->buf + pad->buf_len - pad->tail, pad->tail);
1630 *bytes += pad->head + pad->tail;
1631 *offset -= pad->head;
1632 *qiov = &pad->local_qiov;
1633
1634 return true;
1635 }
1636
1637 int coroutine_fn bdrv_co_preadv(BdrvChild *child,
1638 int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
1639 BdrvRequestFlags flags)
1640 {
1641 BlockDriverState *bs = child->bs;
1642 BdrvTrackedRequest req;
1643 BdrvRequestPadding pad;
1644 int ret;
1645
1646 trace_bdrv_co_preadv(bs, offset, bytes, flags);
1647
1648 ret = bdrv_check_byte_request(bs, offset, bytes);
1649 if (ret < 0) {
1650 return ret;
1651 }
1652
1653 bdrv_inc_in_flight(bs);
1654
1655 /* Don't do copy-on-read if we read data before write operation */
1656 if (atomic_read(&bs->copy_on_read) && !(flags & BDRV_REQ_NO_SERIALISING)) {
1657 flags |= BDRV_REQ_COPY_ON_READ;
1658 }
1659
1660 bdrv_pad_request(bs, &qiov, &offset, &bytes, &pad);
1661
1662 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_READ);
1663 ret = bdrv_aligned_preadv(child, &req, offset, bytes,
1664 bs->bl.request_alignment,
1665 qiov, flags);
1666 tracked_request_end(&req);
1667 bdrv_dec_in_flight(bs);
1668
1669 bdrv_padding_destroy(&pad);
1670
1671 return ret;
1672 }
1673
1674 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs,
1675 int64_t offset, int bytes, BdrvRequestFlags flags)
1676 {
1677 BlockDriver *drv = bs->drv;
1678 QEMUIOVector qiov;
1679 void *buf = NULL;
1680 int ret = 0;
1681 bool need_flush = false;
1682 int head = 0;
1683 int tail = 0;
1684
1685 int max_write_zeroes = MIN_NON_ZERO(bs->bl.max_pwrite_zeroes, INT_MAX);
1686 int alignment = MAX(bs->bl.pwrite_zeroes_alignment,
1687 bs->bl.request_alignment);
1688 int max_transfer = MIN_NON_ZERO(bs->bl.max_transfer, MAX_BOUNCE_BUFFER);
1689
1690 if (!drv) {
1691 return -ENOMEDIUM;
1692 }
1693
1694 if ((flags & ~bs->supported_zero_flags) & BDRV_REQ_NO_FALLBACK) {
1695 return -ENOTSUP;
1696 }
1697
1698 assert(alignment % bs->bl.request_alignment == 0);
1699 head = offset % alignment;
1700 tail = (offset + bytes) % alignment;
1701 max_write_zeroes = QEMU_ALIGN_DOWN(max_write_zeroes, alignment);
1702 assert(max_write_zeroes >= bs->bl.request_alignment);
1703
1704 while (bytes > 0 && !ret) {
1705 int num = bytes;
1706
1707 /* Align request. Block drivers can expect the "bulk" of the request
1708 * to be aligned, and that unaligned requests do not cross cluster
1709 * boundaries.
1710 */
1711 if (head) {
1712 /* Make a small request up to the first aligned sector. For
1713 * convenience, limit this request to max_transfer even if
1714 * we don't need to fall back to writes. */
1715 num = MIN(MIN(bytes, max_transfer), alignment - head);
1716 head = (head + num) % alignment;
1717 assert(num < max_write_zeroes);
1718 } else if (tail && num > alignment) {
1719 /* Shorten the request to the last aligned sector. */
1720 num -= tail;
1721 }
1722
1723 /* limit request size */
1724 if (num > max_write_zeroes) {
1725 num = max_write_zeroes;
1726 }
1727
1728 ret = -ENOTSUP;
1729 /* First try the efficient write zeroes operation */
1730 if (drv->bdrv_co_pwrite_zeroes) {
1731 ret = drv->bdrv_co_pwrite_zeroes(bs, offset, num,
1732 flags & bs->supported_zero_flags);
1733 if (ret != -ENOTSUP && (flags & BDRV_REQ_FUA) &&
1734 !(bs->supported_zero_flags & BDRV_REQ_FUA)) {
1735 need_flush = true;
1736 }
1737 } else {
1738 assert(!bs->supported_zero_flags);
1739 }
1740
1741 if (ret < 0 && !(flags & BDRV_REQ_NO_FALLBACK)) {
1742 /* Fall back to bounce buffer if write zeroes is unsupported */
1743 BdrvRequestFlags write_flags = flags & ~BDRV_REQ_ZERO_WRITE;
1744
1745 if ((flags & BDRV_REQ_FUA) &&
1746 !(bs->supported_write_flags & BDRV_REQ_FUA)) {
1747 /* No need for bdrv_driver_pwrite() to do a fallback
1748 * flush on each chunk; use just one at the end */
1749 write_flags &= ~BDRV_REQ_FUA;
1750 need_flush = true;
1751 }
1752 num = MIN(num, max_transfer);
1753 if (buf == NULL) {
1754 buf = qemu_try_blockalign0(bs, num);
1755 if (buf == NULL) {
1756 ret = -ENOMEM;
1757 goto fail;
1758 }
1759 }
1760 qemu_iovec_init_buf(&qiov, buf, num);
1761
1762 ret = bdrv_driver_pwritev(bs, offset, num, &qiov, 0, write_flags);
1763
1764 /* Keep bounce buffer around if it is big enough for all
1765 * all future requests.
1766 */
1767 if (num < max_transfer) {
1768 qemu_vfree(buf);
1769 buf = NULL;
1770 }
1771 }
1772
1773 offset += num;
1774 bytes -= num;
1775 }
1776
1777 fail:
1778 if (ret == 0 && need_flush) {
1779 ret = bdrv_co_flush(bs);
1780 }
1781 qemu_vfree(buf);
1782 return ret;
1783 }
1784
1785 static inline int coroutine_fn
1786 bdrv_co_write_req_prepare(BdrvChild *child, int64_t offset, uint64_t bytes,
1787 BdrvTrackedRequest *req, int flags)
1788 {
1789 BlockDriverState *bs = child->bs;
1790 bool waited;
1791 int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);
1792
1793 if (bs->read_only) {
1794 return -EPERM;
1795 }
1796
1797 /* BDRV_REQ_NO_SERIALISING is only for read operation */
1798 assert(!(flags & BDRV_REQ_NO_SERIALISING));
1799 assert(!(bs->open_flags & BDRV_O_INACTIVE));
1800 assert((bs->open_flags & BDRV_O_NO_IO) == 0);
1801 assert(!(flags & ~BDRV_REQ_MASK));
1802
1803 if (flags & BDRV_REQ_SERIALISING) {
1804 mark_request_serialising(req, bdrv_get_cluster_size(bs));
1805 }
1806
1807 waited = wait_serialising_requests(req);
1808
1809 assert(!waited || !req->serialising ||
1810 is_request_serialising_and_aligned(req));
1811 assert(req->overlap_offset <= offset);
1812 assert(offset + bytes <= req->overlap_offset + req->overlap_bytes);
1813 assert(end_sector <= bs->total_sectors || child->perm & BLK_PERM_RESIZE);
1814
1815 switch (req->type) {
1816 case BDRV_TRACKED_WRITE:
1817 case BDRV_TRACKED_DISCARD:
1818 if (flags & BDRV_REQ_WRITE_UNCHANGED) {
1819 assert(child->perm & (BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE));
1820 } else {
1821 assert(child->perm & BLK_PERM_WRITE);
1822 }
1823 return notifier_with_return_list_notify(&bs->before_write_notifiers,
1824 req);
1825 case BDRV_TRACKED_TRUNCATE:
1826 assert(child->perm & BLK_PERM_RESIZE);
1827 return 0;
1828 default:
1829 abort();
1830 }
1831 }
1832
1833 static inline void coroutine_fn
1834 bdrv_co_write_req_finish(BdrvChild *child, int64_t offset, uint64_t bytes,
1835 BdrvTrackedRequest *req, int ret)
1836 {
1837 int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);
1838 BlockDriverState *bs = child->bs;
1839
1840 atomic_inc(&bs->write_gen);
1841
1842 /*
1843 * Discard cannot extend the image, but in error handling cases, such as
1844 * when reverting a qcow2 cluster allocation, the discarded range can pass
1845 * the end of image file, so we cannot assert about BDRV_TRACKED_DISCARD
1846 * here. Instead, just skip it, since semantically a discard request
1847 * beyond EOF cannot expand the image anyway.
1848 */
1849 if (ret == 0 &&
1850 (req->type == BDRV_TRACKED_TRUNCATE ||
1851 end_sector > bs->total_sectors) &&
1852 req->type != BDRV_TRACKED_DISCARD) {
1853 bs->total_sectors = end_sector;
1854 bdrv_parent_cb_resize(bs);
1855 bdrv_dirty_bitmap_truncate(bs, end_sector << BDRV_SECTOR_BITS);
1856 }
1857 if (req->bytes) {
1858 switch (req->type) {
1859 case BDRV_TRACKED_WRITE:
1860 stat64_max(&bs->wr_highest_offset, offset + bytes);
1861 /* fall through, to set dirty bits */
1862 case BDRV_TRACKED_DISCARD:
1863 bdrv_set_dirty(bs, offset, bytes);
1864 break;
1865 default:
1866 break;
1867 }
1868 }
1869 }
1870
1871 /*
1872 * Forwards an already correctly aligned write request to the BlockDriver,
1873 * after possibly fragmenting it.
1874 */
1875 static int coroutine_fn bdrv_aligned_pwritev(BdrvChild *child,
1876 BdrvTrackedRequest *req, int64_t offset, unsigned int bytes,
1877 int64_t align, QEMUIOVector *qiov, int flags)
1878 {
1879 BlockDriverState *bs = child->bs;
1880 BlockDriver *drv = bs->drv;
1881 int ret;
1882
1883 uint64_t bytes_remaining = bytes;
1884 int max_transfer;
1885
1886 if (!drv) {
1887 return -ENOMEDIUM;
1888 }
1889
1890 if (bdrv_has_readonly_bitmaps(bs)) {
1891 return -EPERM;
1892 }
1893
1894 assert(is_power_of_2(align));
1895 assert((offset & (align - 1)) == 0);
1896 assert((bytes & (align - 1)) == 0);
1897 assert(!qiov || bytes == qiov->size);
1898 max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX),
1899 align);
1900
1901 ret = bdrv_co_write_req_prepare(child, offset, bytes, req, flags);
1902
1903 if (!ret && bs->detect_zeroes != BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF &&
1904 !(flags & BDRV_REQ_ZERO_WRITE) && drv->bdrv_co_pwrite_zeroes &&
1905 qemu_iovec_is_zero(qiov, 0, qiov->size)) {
1906 flags |= BDRV_REQ_ZERO_WRITE;
1907 if (bs->detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP) {
1908 flags |= BDRV_REQ_MAY_UNMAP;
1909 }
1910 }
1911
1912 if (ret < 0) {
1913 /* Do nothing, write notifier decided to fail this request */
1914 } else if (flags & BDRV_REQ_ZERO_WRITE) {
1915 bdrv_debug_event(bs, BLKDBG_PWRITEV_ZERO);
1916 ret = bdrv_co_do_pwrite_zeroes(bs, offset, bytes, flags);
1917 } else if (flags & BDRV_REQ_WRITE_COMPRESSED) {
1918 ret = bdrv_driver_pwritev_compressed(bs, offset, bytes, qiov, 0);
1919 } else if (bytes <= max_transfer) {
1920 bdrv_debug_event(bs, BLKDBG_PWRITEV);
1921 ret = bdrv_driver_pwritev(bs, offset, bytes, qiov, 0, flags);
1922 } else {
1923 bdrv_debug_event(bs, BLKDBG_PWRITEV);
1924 while (bytes_remaining) {
1925 int num = MIN(bytes_remaining, max_transfer);
1926 QEMUIOVector local_qiov;
1927 int local_flags = flags;
1928
1929 assert(num);
1930 if (num < bytes_remaining && (flags & BDRV_REQ_FUA) &&
1931 !(bs->supported_write_flags & BDRV_REQ_FUA)) {
1932 /* If FUA is going to be emulated by flush, we only
1933 * need to flush on the last iteration */
1934 local_flags &= ~BDRV_REQ_FUA;
1935 }
1936 qemu_iovec_init(&local_qiov, qiov->niov);
1937 qemu_iovec_concat(&local_qiov, qiov, bytes - bytes_remaining, num);
1938
1939 ret = bdrv_driver_pwritev(bs, offset + bytes - bytes_remaining,
1940 num, &local_qiov, 0, local_flags);
1941 qemu_iovec_destroy(&local_qiov);
1942 if (ret < 0) {
1943 break;
1944 }
1945 bytes_remaining -= num;
1946 }
1947 }
1948 bdrv_debug_event(bs, BLKDBG_PWRITEV_DONE);
1949
1950 if (ret >= 0) {
1951 ret = 0;
1952 }
1953 bdrv_co_write_req_finish(child, offset, bytes, req, ret);
1954
1955 return ret;
1956 }
1957
1958 static int coroutine_fn bdrv_co_do_zero_pwritev(BdrvChild *child,
1959 int64_t offset,
1960 unsigned int bytes,
1961 BdrvRequestFlags flags,
1962 BdrvTrackedRequest *req)
1963 {
1964 BlockDriverState *bs = child->bs;
1965 QEMUIOVector local_qiov;
1966 uint64_t align = bs->bl.request_alignment;
1967 int ret = 0;
1968 bool padding;
1969 BdrvRequestPadding pad;
1970
1971 padding = bdrv_init_padding(bs, offset, bytes, &pad);
1972 if (padding) {
1973 mark_request_serialising(req, align);
1974 wait_serialising_requests(req);
1975
1976 bdrv_padding_rmw_read(child, req, &pad, true);
1977
1978 if (pad.head || pad.merge_reads) {
1979 int64_t aligned_offset = offset & ~(align - 1);
1980 int64_t write_bytes = pad.merge_reads ? pad.buf_len : align;
1981
1982 qemu_iovec_init_buf(&local_qiov, pad.buf, write_bytes);
1983 ret = bdrv_aligned_pwritev(child, req, aligned_offset, write_bytes,
1984 align, &local_qiov,
1985 flags & ~BDRV_REQ_ZERO_WRITE);
1986 if (ret < 0 || pad.merge_reads) {
1987 /* Error or all work is done */
1988 goto out;
1989 }
1990 offset += write_bytes - pad.head;
1991 bytes -= write_bytes - pad.head;
1992 }
1993 }
1994
1995 assert(!bytes || (offset & (align - 1)) == 0);
1996 if (bytes >= align) {
1997 /* Write the aligned part in the middle. */
1998 uint64_t aligned_bytes = bytes & ~(align - 1);
1999 ret = bdrv_aligned_pwritev(child, req, offset, aligned_bytes, align,
2000 NULL, flags);
2001 if (ret < 0) {
2002 goto out;
2003 }
2004 bytes -= aligned_bytes;
2005 offset += aligned_bytes;
2006 }
2007
2008 assert(!bytes || (offset & (align - 1)) == 0);
2009 if (bytes) {
2010 assert(align == pad.tail + bytes);
2011
2012 qemu_iovec_init_buf(&local_qiov, pad.tail_buf, align);
2013 ret = bdrv_aligned_pwritev(child, req, offset, align, align,
2014 &local_qiov, flags & ~BDRV_REQ_ZERO_WRITE);
2015 }
2016
2017 out:
2018 bdrv_padding_destroy(&pad);
2019
2020 return ret;
2021 }
2022
2023 /*
2024 * Handle a write request in coroutine context
2025 */
2026 int coroutine_fn bdrv_co_pwritev(BdrvChild *child,
2027 int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
2028 BdrvRequestFlags flags)
2029 {
2030 BlockDriverState *bs = child->bs;
2031 BdrvTrackedRequest req;
2032 uint64_t align = bs->bl.request_alignment;
2033 BdrvRequestPadding pad;
2034 int ret;
2035
2036 trace_bdrv_co_pwritev(child->bs, offset, bytes, flags);
2037
2038 if (!bs->drv) {
2039 return -ENOMEDIUM;
2040 }
2041
2042 ret = bdrv_check_byte_request(bs, offset, bytes);
2043 if (ret < 0) {
2044 return ret;
2045 }
2046
2047 bdrv_inc_in_flight(bs);
2048 /*
2049 * Align write if necessary by performing a read-modify-write cycle.
2050 * Pad qiov with the read parts and be sure to have a tracked request not
2051 * only for bdrv_aligned_pwritev, but also for the reads of the RMW cycle.
2052 */
2053 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_WRITE);
2054
2055 if (flags & BDRV_REQ_ZERO_WRITE) {
2056 ret = bdrv_co_do_zero_pwritev(child, offset, bytes, flags, &req);
2057 goto out;
2058 }
2059
2060 if (bdrv_pad_request(bs, &qiov, &offset, &bytes, &pad)) {
2061 mark_request_serialising(&req, align);
2062 wait_serialising_requests(&req);
2063 bdrv_padding_rmw_read(child, &req, &pad, false);
2064 }
2065
2066 ret = bdrv_aligned_pwritev(child, &req, offset, bytes, align,
2067 qiov, flags);
2068
2069 bdrv_padding_destroy(&pad);
2070
2071 out:
2072 tracked_request_end(&req);
2073 bdrv_dec_in_flight(bs);
2074
2075 return ret;
2076 }
2077
2078 int coroutine_fn bdrv_co_pwrite_zeroes(BdrvChild *child, int64_t offset,
2079 int bytes, BdrvRequestFlags flags)
2080 {
2081 trace_bdrv_co_pwrite_zeroes(child->bs, offset, bytes, flags);
2082
2083 if (!(child->bs->open_flags & BDRV_O_UNMAP)) {
2084 flags &= ~BDRV_REQ_MAY_UNMAP;
2085 }
2086
2087 return bdrv_co_pwritev(child, offset, bytes, NULL,
2088 BDRV_REQ_ZERO_WRITE | flags);
2089 }
2090
2091 /*
2092 * Flush ALL BDSes regardless of if they are reachable via a BlkBackend or not.
2093 */
2094 int bdrv_flush_all(void)
2095 {
2096 BdrvNextIterator it;
2097 BlockDriverState *bs = NULL;
2098 int result = 0;
2099
2100 for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
2101 AioContext *aio_context = bdrv_get_aio_context(bs);
2102 int ret;
2103
2104 aio_context_acquire(aio_context);
2105 ret = bdrv_flush(bs);
2106 if (ret < 0 && !result) {
2107 result = ret;
2108 }
2109 aio_context_release(aio_context);
2110 }
2111
2112 return result;
2113 }
2114
2115
2116 typedef struct BdrvCoBlockStatusData {
2117 BlockDriverState *bs;
2118 BlockDriverState *base;
2119 bool want_zero;
2120 int64_t offset;
2121 int64_t bytes;
2122 int64_t *pnum;
2123 int64_t *map;
2124 BlockDriverState **file;
2125 int ret;
2126 bool done;
2127 } BdrvCoBlockStatusData;
2128
2129 int coroutine_fn bdrv_co_block_status_from_file(BlockDriverState *bs,
2130 bool want_zero,
2131 int64_t offset,
2132 int64_t bytes,
2133 int64_t *pnum,
2134 int64_t *map,
2135 BlockDriverState **file)
2136 {
2137 assert(bs->file && bs->file->bs);
2138 *pnum = bytes;
2139 *map = offset;
2140 *file = bs->file->bs;
2141 return BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID;
2142 }
2143
2144 int coroutine_fn bdrv_co_block_status_from_backing(BlockDriverState *bs,
2145 bool want_zero,
2146 int64_t offset,
2147 int64_t bytes,
2148 int64_t *pnum,
2149 int64_t *map,
2150 BlockDriverState **file)
2151 {
2152 assert(bs->backing && bs->backing->bs);
2153 *pnum = bytes;
2154 *map = offset;
2155 *file = bs->backing->bs;
2156 return BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID;
2157 }
2158
2159 /*
2160 * Returns the allocation status of the specified sectors.
2161 * Drivers not implementing the functionality are assumed to not support
2162 * backing files, hence all their sectors are reported as allocated.
2163 *
2164 * If 'want_zero' is true, the caller is querying for mapping
2165 * purposes, with a focus on valid BDRV_BLOCK_OFFSET_VALID, _DATA, and
2166 * _ZERO where possible; otherwise, the result favors larger 'pnum',
2167 * with a focus on accurate BDRV_BLOCK_ALLOCATED.
2168 *
2169 * If 'offset' is beyond the end of the disk image the return value is
2170 * BDRV_BLOCK_EOF and 'pnum' is set to 0.
2171 *
2172 * 'bytes' is the max value 'pnum' should be set to. If bytes goes
2173 * beyond the end of the disk image it will be clamped; if 'pnum' is set to
2174 * the end of the image, then the returned value will include BDRV_BLOCK_EOF.
2175 *
2176 * 'pnum' is set to the number of bytes (including and immediately
2177 * following the specified offset) that are easily known to be in the
2178 * same allocated/unallocated state. Note that a second call starting
2179 * at the original offset plus returned pnum may have the same status.
2180 * The returned value is non-zero on success except at end-of-file.
2181 *
2182 * Returns negative errno on failure. Otherwise, if the
2183 * BDRV_BLOCK_OFFSET_VALID bit is set, 'map' and 'file' (if non-NULL) are
2184 * set to the host mapping and BDS corresponding to the guest offset.
2185 */
2186 static int coroutine_fn bdrv_co_block_status(BlockDriverState *bs,
2187 bool want_zero,
2188 int64_t offset, int64_t bytes,
2189 int64_t *pnum, int64_t *map,
2190 BlockDriverState **file)
2191 {
2192 int64_t total_size;
2193 int64_t n; /* bytes */
2194 int ret;
2195 int64_t local_map = 0;
2196 BlockDriverState *local_file = NULL;
2197 int64_t aligned_offset, aligned_bytes;
2198 uint32_t align;
2199
2200 assert(pnum);
2201 *pnum = 0;
2202 total_size = bdrv_getlength(bs);
2203 if (total_size < 0) {
2204 ret = total_size;
2205 goto early_out;
2206 }
2207
2208 if (offset >= total_size) {
2209 ret = BDRV_BLOCK_EOF;
2210 goto early_out;
2211 }
2212 if (!bytes) {
2213 ret = 0;
2214 goto early_out;
2215 }
2216
2217 n = total_size - offset;
2218 if (n < bytes) {
2219 bytes = n;
2220 }
2221
2222 /* Must be non-NULL or bdrv_getlength() would have failed */
2223 assert(bs->drv);
2224 if (!bs->drv->bdrv_co_block_status) {
2225 *pnum = bytes;
2226 ret = BDRV_BLOCK_DATA | BDRV_BLOCK_ALLOCATED;
2227 if (offset + bytes == total_size) {
2228 ret |= BDRV_BLOCK_EOF;
2229 }
2230 if (bs->drv->protocol_name) {
2231 ret |= BDRV_BLOCK_OFFSET_VALID;
2232 local_map = offset;
2233 local_file = bs;
2234 }
2235 goto early_out;
2236 }
2237
2238 bdrv_inc_in_flight(bs);
2239
2240 /* Round out to request_alignment boundaries */
2241 align = bs->bl.request_alignment;
2242 aligned_offset = QEMU_ALIGN_DOWN(offset, align);
2243 aligned_bytes = ROUND_UP(offset + bytes, align) - aligned_offset;
2244
2245 ret = bs->drv->bdrv_co_block_status(bs, want_zero, aligned_offset,
2246 aligned_bytes, pnum, &local_map,
2247 &local_file);
2248 if (ret < 0) {
2249 *pnum = 0;
2250 goto out;
2251 }
2252
2253 /*
2254 * The driver's result must be a non-zero multiple of request_alignment.
2255 * Clamp pnum and adjust map to original request.
2256 */
2257 assert(*pnum && QEMU_IS_ALIGNED(*pnum, align) &&
2258 align > offset - aligned_offset);
2259 if (ret & BDRV_BLOCK_RECURSE) {
2260 assert(ret & BDRV_BLOCK_DATA);
2261 assert(ret & BDRV_BLOCK_OFFSET_VALID);
2262 assert(!(ret & BDRV_BLOCK_ZERO));
2263 }
2264
2265 *pnum -= offset - aligned_offset;
2266 if (*pnum > bytes) {
2267 *pnum = bytes;
2268 }
2269 if (ret & BDRV_BLOCK_OFFSET_VALID) {
2270 local_map += offset - aligned_offset;
2271 }
2272
2273 if (ret & BDRV_BLOCK_RAW) {
2274 assert(ret & BDRV_BLOCK_OFFSET_VALID && local_file);
2275 ret = bdrv_co_block_status(local_file, want_zero, local_map,
2276 *pnum, pnum, &local_map, &local_file);
2277 goto out;
2278 }
2279
2280 if (ret & (BDRV_BLOCK_DATA | BDRV_BLOCK_ZERO)) {
2281 ret |= BDRV_BLOCK_ALLOCATED;
2282 } else if (want_zero) {
2283 if (bdrv_unallocated_blocks_are_zero(bs)) {
2284 ret |= BDRV_BLOCK_ZERO;
2285 } else if (bs->backing) {
2286 BlockDriverState *bs2 = bs->backing->bs;
2287 int64_t size2 = bdrv_getlength(bs2);
2288
2289 if (size2 >= 0 && offset >= size2) {
2290 ret |= BDRV_BLOCK_ZERO;
2291 }
2292 }
2293 }
2294
2295 if (want_zero && ret & BDRV_BLOCK_RECURSE &&
2296 local_file && local_file != bs &&
2297 (ret & BDRV_BLOCK_DATA) && !(ret & BDRV_BLOCK_ZERO) &&
2298 (ret & BDRV_BLOCK_OFFSET_VALID)) {
2299 int64_t file_pnum;
2300 int ret2;
2301
2302 ret2 = bdrv_co_block_status(local_file, want_zero, local_map,
2303 *pnum, &file_pnum, NULL, NULL);
2304 if (ret2 >= 0) {
2305 /* Ignore errors. This is just providing extra information, it
2306 * is useful but not necessary.
2307 */
2308 if (ret2 & BDRV_BLOCK_EOF &&
2309 (!file_pnum || ret2 & BDRV_BLOCK_ZERO)) {
2310 /*
2311 * It is valid for the format block driver to read
2312 * beyond the end of the underlying file's current
2313 * size; such areas read as zero.
2314 */
2315 ret |= BDRV_BLOCK_ZERO;
2316 } else {
2317 /* Limit request to the range reported by the protocol driver */
2318 *pnum = file_pnum;
2319 ret |= (ret2 & BDRV_BLOCK_ZERO);
2320 }
2321 }
2322 }
2323
2324 out:
2325 bdrv_dec_in_flight(bs);
2326 if (ret >= 0 && offset + *pnum == total_size) {
2327 ret |= BDRV_BLOCK_EOF;
2328 }
2329 early_out:
2330 if (file) {
2331 *file = local_file;
2332 }
2333 if (map) {
2334 *map = local_map;
2335 }
2336 return ret;
2337 }
2338
2339 static int coroutine_fn bdrv_co_block_status_above(BlockDriverState *bs,
2340 BlockDriverState *base,
2341 bool want_zero,
2342 int64_t offset,
2343 int64_t bytes,
2344 int64_t *pnum,
2345 int64_t *map,
2346 BlockDriverState **file)
2347 {
2348 BlockDriverState *p;
2349 int ret = 0;
2350 bool first = true;
2351
2352 assert(bs != base);
2353 for (p = bs; p != base; p = backing_bs(p)) {
2354 ret = bdrv_co_block_status(p, want_zero, offset, bytes, pnum, map,
2355 file);
2356 if (ret < 0) {
2357 break;
2358 }
2359 if (ret & BDRV_BLOCK_ZERO && ret & BDRV_BLOCK_EOF && !first) {
2360 /*
2361 * Reading beyond the end of the file continues to read
2362 * zeroes, but we can only widen the result to the
2363 * unallocated length we learned from an earlier
2364 * iteration.
2365 */
2366 *pnum = bytes;
2367 }
2368 if (ret & (BDRV_BLOCK_ZERO | BDRV_BLOCK_DATA)) {
2369 break;
2370 }
2371 /* [offset, pnum] unallocated on this layer, which could be only
2372 * the first part of [offset, bytes]. */
2373 bytes = MIN(bytes, *pnum);
2374 first = false;
2375 }
2376 return ret;
2377 }
2378
2379 /* Coroutine wrapper for bdrv_block_status_above() */
2380 static void coroutine_fn bdrv_block_status_above_co_entry(void *opaque)
2381 {
2382 BdrvCoBlockStatusData *data = opaque;
2383
2384 data->ret = bdrv_co_block_status_above(data->bs, data->base,
2385 data->want_zero,
2386 data->offset, data->bytes,
2387 data->pnum, data->map, data->file);
2388 data->done = true;
2389 aio_wait_kick();
2390 }
2391
2392 /*
2393 * Synchronous wrapper around bdrv_co_block_status_above().
2394 *
2395 * See bdrv_co_block_status_above() for details.
2396 */
2397 static int bdrv_common_block_status_above(BlockDriverState *bs,
2398 BlockDriverState *base,
2399 bool want_zero, int64_t offset,
2400 int64_t bytes, int64_t *pnum,
2401 int64_t *map,
2402 BlockDriverState **file)
2403 {
2404 Coroutine *co;
2405 BdrvCoBlockStatusData data = {
2406 .bs = bs,
2407 .base = base,
2408 .want_zero = want_zero,
2409 .offset = offset,
2410 .bytes = bytes,
2411 .pnum = pnum,
2412 .map = map,
2413 .file = file,
2414 .done = false,
2415 };
2416
2417 if (qemu_in_coroutine()) {
2418 /* Fast-path if already in coroutine context */
2419 bdrv_block_status_above_co_entry(&data);
2420 } else {
2421 co = qemu_coroutine_create(bdrv_block_status_above_co_entry, &data);
2422 bdrv_coroutine_enter(bs, co);
2423 BDRV_POLL_WHILE(bs, !data.done);
2424 }
2425 return data.ret;
2426 }
2427
2428 int bdrv_block_status_above(BlockDriverState *bs, BlockDriverState *base,
2429 int64_t offset, int64_t bytes, int64_t *pnum,
2430 int64_t *map, BlockDriverState **file)
2431 {
2432 return bdrv_common_block_status_above(bs, base, true, offset, bytes,
2433 pnum, map, file);
2434 }
2435
2436 int bdrv_block_status(BlockDriverState *bs, int64_t offset, int64_t bytes,
2437 int64_t *pnum, int64_t *map, BlockDriverState **file)
2438 {
2439 return bdrv_block_status_above(bs, backing_bs(bs),
2440 offset, bytes, pnum, map, file);
2441 }
2442
2443 int coroutine_fn bdrv_is_allocated(BlockDriverState *bs, int64_t offset,
2444 int64_t bytes, int64_t *pnum)
2445 {
2446 int ret;
2447 int64_t dummy;
2448
2449 ret = bdrv_common_block_status_above(bs, backing_bs(bs), false, offset,
2450 bytes, pnum ? pnum : &dummy, NULL,
2451 NULL);
2452 if (ret < 0) {
2453 return ret;
2454 }
2455 return !!(ret & BDRV_BLOCK_ALLOCATED);
2456 }
2457
2458 /*
2459 * Given an image chain: ... -> [BASE] -> [INTER1] -> [INTER2] -> [TOP]
2460 *
2461 * Return 1 if (a prefix of) the given range is allocated in any image
2462 * between BASE and TOP (BASE is only included if include_base is set).
2463 * BASE can be NULL to check if the given offset is allocated in any
2464 * image of the chain. Return 0 otherwise, or negative errno on
2465 * failure.
2466 *
2467 * 'pnum' is set to the number of bytes (including and immediately
2468 * following the specified offset) that are known to be in the same
2469 * allocated/unallocated state. Note that a subsequent call starting
2470 * at 'offset + *pnum' may return the same allocation status (in other
2471 * words, the result is not necessarily the maximum possible range);
2472 * but 'pnum' will only be 0 when end of file is reached.
2473 *
2474 */
2475 int bdrv_is_allocated_above(BlockDriverState *top,
2476 BlockDriverState *base,
2477 bool include_base, int64_t offset,
2478 int64_t bytes, int64_t *pnum)
2479 {
2480 BlockDriverState *intermediate;
2481 int ret;
2482 int64_t n = bytes;
2483
2484 assert(base || !include_base);
2485
2486 intermediate = top;
2487 while (include_base || intermediate != base) {
2488 int64_t pnum_inter;
2489 int64_t size_inter;
2490
2491 assert(intermediate);
2492 ret = bdrv_is_allocated(intermediate, offset, bytes, &pnum_inter);
2493 if (ret < 0) {
2494 return ret;
2495 }
2496 if (ret) {
2497 *pnum = pnum_inter;
2498 return 1;
2499 }
2500
2501 size_inter = bdrv_getlength(intermediate);
2502 if (size_inter < 0) {
2503 return size_inter;
2504 }
2505 if (n > pnum_inter &&
2506 (intermediate == top || offset + pnum_inter < size_inter)) {
2507 n = pnum_inter;
2508 }
2509
2510 if (intermediate == base) {
2511 break;
2512 }
2513
2514 intermediate = backing_bs(intermediate);
2515 }
2516
2517 *pnum = n;
2518 return 0;
2519 }
2520
2521 typedef struct BdrvVmstateCo {
2522 BlockDriverState *bs;
2523 QEMUIOVector *qiov;
2524 int64_t pos;
2525 bool is_read;
2526 int ret;
2527 } BdrvVmstateCo;
2528
2529 static int coroutine_fn
2530 bdrv_co_rw_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos,
2531 bool is_read)
2532 {
2533 BlockDriver *drv = bs->drv;
2534 int ret = -ENOTSUP;
2535
2536 bdrv_inc_in_flight(bs);
2537
2538 if (!drv) {
2539 ret = -ENOMEDIUM;
2540 } else if (drv->bdrv_load_vmstate) {
2541 if (is_read) {
2542 ret = drv->bdrv_load_vmstate(bs, qiov, pos);
2543 } else {
2544 ret = drv->bdrv_save_vmstate(bs, qiov, pos);
2545 }
2546 } else if (bs->file) {
2547 ret = bdrv_co_rw_vmstate(bs->file->bs, qiov, pos, is_read);
2548 }
2549
2550 bdrv_dec_in_flight(bs);
2551 return ret;
2552 }
2553
2554 static void coroutine_fn bdrv_co_rw_vmstate_entry(void *opaque)
2555 {
2556 BdrvVmstateCo *co = opaque;
2557 co->ret = bdrv_co_rw_vmstate(co->bs, co->qiov, co->pos, co->is_read);
2558 aio_wait_kick();
2559 }
2560
2561 static inline int
2562 bdrv_rw_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos,
2563 bool is_read)
2564 {
2565 if (qemu_in_coroutine()) {
2566 return bdrv_co_rw_vmstate(bs, qiov, pos, is_read);
2567 } else {
2568 BdrvVmstateCo data = {
2569 .bs = bs,
2570 .qiov = qiov,
2571 .pos = pos,
2572 .is_read = is_read,
2573 .ret = -EINPROGRESS,
2574 };
2575 Coroutine *co = qemu_coroutine_create(bdrv_co_rw_vmstate_entry, &data);
2576
2577 bdrv_coroutine_enter(bs, co);
2578 BDRV_POLL_WHILE(bs, data.ret == -EINPROGRESS);
2579 return data.ret;
2580 }
2581 }
2582
2583 int bdrv_save_vmstate(BlockDriverState *bs, const uint8_t *buf,
2584 int64_t pos, int size)
2585 {
2586 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, size);
2587 int ret;
2588
2589 ret = bdrv_writev_vmstate(bs, &qiov, pos);
2590 if (ret < 0) {
2591 return ret;
2592 }
2593
2594 return size;
2595 }
2596
2597 int bdrv_writev_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos)
2598 {
2599 return bdrv_rw_vmstate(bs, qiov, pos, false);
2600 }
2601
2602 int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf,
2603 int64_t pos, int size)
2604 {
2605 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, size);
2606 int ret;
2607
2608 ret = bdrv_readv_vmstate(bs, &qiov, pos);
2609 if (ret < 0) {
2610 return ret;
2611 }
2612
2613 return size;
2614 }
2615
2616 int bdrv_readv_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos)
2617 {
2618 return bdrv_rw_vmstate(bs, qiov, pos, true);
2619 }
2620
2621 /**************************************************************/
2622 /* async I/Os */
2623
2624 void bdrv_aio_cancel(BlockAIOCB *acb)
2625 {
2626 qemu_aio_ref(acb);
2627 bdrv_aio_cancel_async(acb);
2628 while (acb->refcnt > 1) {
2629 if (acb->aiocb_info->get_aio_context) {
2630 aio_poll(acb->aiocb_info->get_aio_context(acb), true);
2631 } else if (acb->bs) {
2632 /* qemu_aio_ref and qemu_aio_unref are not thread-safe, so
2633 * assert that we're not using an I/O thread. Thread-safe
2634 * code should use bdrv_aio_cancel_async exclusively.
2635 */
2636 assert(bdrv_get_aio_context(acb->bs) == qemu_get_aio_context());
2637 aio_poll(bdrv_get_aio_context(acb->bs), true);
2638 } else {
2639 abort();
2640 }
2641 }
2642 qemu_aio_unref(acb);
2643 }
2644
2645 /* Async version of aio cancel. The caller is not blocked if the acb implements
2646 * cancel_async, otherwise we do nothing and let the request normally complete.
2647 * In either case the completion callback must be called. */
2648 void bdrv_aio_cancel_async(BlockAIOCB *acb)
2649 {
2650 if (acb->aiocb_info->cancel_async) {
2651 acb->aiocb_info->cancel_async(acb);
2652 }
2653 }
2654
2655 /**************************************************************/
2656 /* Coroutine block device emulation */
2657
2658 typedef struct FlushCo {
2659 BlockDriverState *bs;
2660 int ret;
2661 } FlushCo;
2662
2663
2664 static void coroutine_fn bdrv_flush_co_entry(void *opaque)
2665 {
2666 FlushCo *rwco = opaque;
2667
2668 rwco->ret = bdrv_co_flush(rwco->bs);
2669 aio_wait_kick();
2670 }
2671
2672 int coroutine_fn bdrv_co_flush(BlockDriverState *bs)
2673 {
2674 int current_gen;
2675 int ret = 0;
2676
2677 bdrv_inc_in_flight(bs);
2678
2679 if (!bdrv_is_inserted(bs) || bdrv_is_read_only(bs) ||
2680 bdrv_is_sg(bs)) {
2681 goto early_exit;
2682 }
2683
2684 qemu_co_mutex_lock(&bs->reqs_lock);
2685 current_gen = atomic_read(&bs->write_gen);
2686
2687 /* Wait until any previous flushes are completed */
2688 while (bs->active_flush_req) {
2689 qemu_co_queue_wait(&bs->flush_queue, &bs->reqs_lock);
2690 }
2691
2692 /* Flushes reach this point in nondecreasing current_gen order. */
2693 bs->active_flush_req = true;
2694 qemu_co_mutex_unlock(&bs->reqs_lock);
2695
2696 /* Write back all layers by calling one driver function */
2697 if (bs->drv->bdrv_co_flush) {
2698 ret = bs->drv->bdrv_co_flush(bs);
2699 goto out;
2700 }
2701
2702 /* Write back cached data to the OS even with cache=unsafe */
2703 BLKDBG_EVENT(bs->file, BLKDBG_FLUSH_TO_OS);
2704 if (bs->drv->bdrv_co_flush_to_os) {
2705 ret = bs->drv->bdrv_co_flush_to_os(bs);
2706 if (ret < 0) {
2707 goto out;
2708 }
2709 }
2710
2711 /* But don't actually force it to the disk with cache=unsafe */
2712 if (bs->open_flags & BDRV_O_NO_FLUSH) {
2713 goto flush_parent;
2714 }
2715
2716 /* Check if we really need to flush anything */
2717 if (bs->flushed_gen == current_gen) {
2718 goto flush_parent;
2719 }
2720
2721 BLKDBG_EVENT(bs->file, BLKDBG_FLUSH_TO_DISK);
2722 if (!bs->drv) {
2723 /* bs->drv->bdrv_co_flush() might have ejected the BDS
2724 * (even in case of apparent success) */
2725 ret = -ENOMEDIUM;
2726 goto out;
2727 }
2728 if (bs->drv->bdrv_co_flush_to_disk) {
2729 ret = bs->drv->bdrv_co_flush_to_disk(bs);
2730 } else if (bs->drv->bdrv_aio_flush) {
2731 BlockAIOCB *acb;
2732 CoroutineIOCompletion co = {
2733 .coroutine = qemu_coroutine_self(),
2734 };
2735
2736 acb = bs->drv->bdrv_aio_flush(bs, bdrv_co_io_em_complete, &co);
2737 if (acb == NULL) {
2738 ret = -EIO;
2739 } else {
2740 qemu_coroutine_yield();
2741 ret = co.ret;
2742 }
2743 } else {
2744 /*
2745 * Some block drivers always operate in either writethrough or unsafe
2746 * mode and don't support bdrv_flush therefore. Usually qemu doesn't
2747 * know how the server works (because the behaviour is hardcoded or
2748 * depends on server-side configuration), so we can't ensure that
2749 * everything is safe on disk. Returning an error doesn't work because
2750 * that would break guests even if the server operates in writethrough
2751 * mode.
2752 *
2753 * Let's hope the user knows what he's doing.
2754 */
2755 ret = 0;
2756 }
2757
2758 if (ret < 0) {
2759 goto out;
2760 }
2761
2762 /* Now flush the underlying protocol. It will also have BDRV_O_NO_FLUSH
2763 * in the case of cache=unsafe, so there are no useless flushes.
2764 */
2765 flush_parent:
2766 ret = bs->file ? bdrv_co_flush(bs->file->bs) : 0;
2767 out:
2768 /* Notify any pending flushes that we have completed */
2769 if (ret == 0) {
2770 bs->flushed_gen = current_gen;
2771 }
2772
2773 qemu_co_mutex_lock(&bs->reqs_lock);
2774 bs->active_flush_req = false;
2775 /* Return value is ignored - it's ok if wait queue is empty */
2776 qemu_co_queue_next(&bs->flush_queue);
2777 qemu_co_mutex_unlock(&bs->reqs_lock);
2778
2779 early_exit:
2780 bdrv_dec_in_flight(bs);
2781 return ret;
2782 }
2783
2784 int bdrv_flush(BlockDriverState *bs)
2785 {
2786 Coroutine *co;
2787 FlushCo flush_co = {
2788 .bs = bs,
2789 .ret = NOT_DONE,
2790 };
2791
2792 if (qemu_in_coroutine()) {
2793 /* Fast-path if already in coroutine context */
2794 bdrv_flush_co_entry(&flush_co);
2795 } else {
2796 co = qemu_coroutine_create(bdrv_flush_co_entry, &flush_co);
2797 bdrv_coroutine_enter(bs, co);
2798 BDRV_POLL_WHILE(bs, flush_co.ret == NOT_DONE);
2799 }
2800
2801 return flush_co.ret;
2802 }
2803
2804 typedef struct DiscardCo {
2805 BdrvChild *child;
2806 int64_t offset;
2807 int64_t bytes;
2808 int ret;
2809 } DiscardCo;
2810 static void coroutine_fn bdrv_pdiscard_co_entry(void *opaque)
2811 {
2812 DiscardCo *rwco = opaque;
2813
2814 rwco->ret = bdrv_co_pdiscard(rwco->child, rwco->offset, rwco->bytes);
2815 aio_wait_kick();
2816 }
2817
2818 int coroutine_fn bdrv_co_pdiscard(BdrvChild *child, int64_t offset,
2819 int64_t bytes)
2820 {
2821 BdrvTrackedRequest req;
2822 int max_pdiscard, ret;
2823 int head, tail, align;
2824 BlockDriverState *bs = child->bs;
2825
2826 if (!bs || !bs->drv || !bdrv_is_inserted(bs)) {
2827 return -ENOMEDIUM;
2828 }
2829
2830 if (bdrv_has_readonly_bitmaps(bs)) {
2831 return -EPERM;
2832 }
2833
2834 if (offset < 0 || bytes < 0 || bytes > INT64_MAX - offset) {
2835 return -EIO;
2836 }
2837
2838 /* Do nothing if disabled. */
2839 if (!(bs->open_flags & BDRV_O_UNMAP)) {
2840 return 0;
2841 }
2842
2843 if (!bs->drv->bdrv_co_pdiscard && !bs->drv->bdrv_aio_pdiscard) {
2844 return 0;
2845 }
2846
2847 /* Discard is advisory, but some devices track and coalesce
2848 * unaligned requests, so we must pass everything down rather than
2849 * round here. Still, most devices will just silently ignore
2850 * unaligned requests (by returning -ENOTSUP), so we must fragment
2851 * the request accordingly. */
2852 align = MAX(bs->bl.pdiscard_alignment, bs->bl.request_alignment);
2853 assert(align % bs->bl.request_alignment == 0);
2854 head = offset % align;
2855 tail = (offset + bytes) % align;
2856
2857 bdrv_inc_in_flight(bs);
2858 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_DISCARD);
2859
2860 ret = bdrv_co_write_req_prepare(child, offset, bytes, &req, 0);
2861 if (ret < 0) {
2862 goto out;
2863 }
2864
2865 max_pdiscard = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_pdiscard, INT_MAX),
2866 align);
2867 assert(max_pdiscard >= bs->bl.request_alignment);
2868
2869 while (bytes > 0) {
2870 int64_t num = bytes;
2871
2872 if (head) {
2873 /* Make small requests to get to alignment boundaries. */
2874 num = MIN(bytes, align - head);
2875 if (!QEMU_IS_ALIGNED(num, bs->bl.request_alignment)) {
2876 num %= bs->bl.request_alignment;
2877 }
2878 head = (head + num) % align;
2879 assert(num < max_pdiscard);
2880 } else if (tail) {
2881 if (num > align) {
2882 /* Shorten the request to the last aligned cluster. */
2883 num -= tail;
2884 } else if (!QEMU_IS_ALIGNED(tail, bs->bl.request_alignment) &&
2885 tail > bs->bl.request_alignment) {
2886 tail %= bs->bl.request_alignment;
2887 num -= tail;
2888 }
2889 }
2890 /* limit request size */
2891 if (num > max_pdiscard) {
2892 num = max_pdiscard;
2893 }
2894
2895 if (!bs->drv) {
2896 ret = -ENOMEDIUM;
2897 goto out;
2898 }
2899 if (bs->drv->bdrv_co_pdiscard) {
2900 ret = bs->drv->bdrv_co_pdiscard(bs, offset, num);
2901 } else {
2902 BlockAIOCB *acb;
2903 CoroutineIOCompletion co = {
2904 .coroutine = qemu_coroutine_self(),
2905 };
2906
2907 acb = bs->drv->bdrv_aio_pdiscard(bs, offset, num,
2908 bdrv_co_io_em_complete, &co);
2909 if (acb == NULL) {
2910 ret = -EIO;
2911 goto out;
2912 } else {
2913 qemu_coroutine_yield();
2914 ret = co.ret;
2915 }
2916 }
2917 if (ret && ret != -ENOTSUP) {
2918 goto out;
2919 }
2920
2921 offset += num;
2922 bytes -= num;
2923 }
2924 ret = 0;
2925 out:
2926 bdrv_co_write_req_finish(child, req.offset, req.bytes, &req, ret);
2927 tracked_request_end(&req);
2928 bdrv_dec_in_flight(bs);
2929 return ret;
2930 }
2931
2932 int bdrv_pdiscard(BdrvChild *child, int64_t offset, int64_t bytes)
2933 {
2934 Coroutine *co;
2935 DiscardCo rwco = {
2936 .child = child,
2937 .offset = offset,
2938 .bytes = bytes,
2939 .ret = NOT_DONE,
2940 };
2941
2942 if (qemu_in_coroutine()) {
2943 /* Fast-path if already in coroutine context */
2944 bdrv_pdiscard_co_entry(&rwco);
2945 } else {
2946 co = qemu_coroutine_create(bdrv_pdiscard_co_entry, &rwco);
2947 bdrv_coroutine_enter(child->bs, co);
2948 BDRV_POLL_WHILE(child->bs, rwco.ret == NOT_DONE);
2949 }
2950
2951 return rwco.ret;
2952 }
2953
2954 int bdrv_co_ioctl(BlockDriverState *bs, int req, void *buf)
2955 {
2956 BlockDriver *drv = bs->drv;
2957 CoroutineIOCompletion co = {
2958 .coroutine = qemu_coroutine_self(),
2959 };
2960 BlockAIOCB *acb;
2961
2962 bdrv_inc_in_flight(bs);
2963 if (!drv || (!drv->bdrv_aio_ioctl && !drv->bdrv_co_ioctl)) {
2964 co.ret = -ENOTSUP;
2965 goto out;
2966 }
2967
2968 if (drv->bdrv_co_ioctl) {
2969 co.ret = drv->bdrv_co_ioctl(bs, req, buf);
2970 } else {
2971 acb = drv->bdrv_aio_ioctl(bs, req, buf, bdrv_co_io_em_complete, &co);
2972 if (!acb) {
2973 co.ret = -ENOTSUP;
2974 goto out;
2975 }
2976 qemu_coroutine_yield();
2977 }
2978 out:
2979 bdrv_dec_in_flight(bs);
2980 return co.ret;
2981 }
2982
2983 void *qemu_blockalign(BlockDriverState *bs, size_t size)
2984 {
2985 return qemu_memalign(bdrv_opt_mem_align(bs), size);
2986 }
2987
2988 void *qemu_blockalign0(BlockDriverState *bs, size_t size)
2989 {
2990 return memset(qemu_blockalign(bs, size), 0, size);
2991 }
2992
2993 void *qemu_try_blockalign(BlockDriverState *bs, size_t size)
2994 {
2995 size_t align = bdrv_opt_mem_align(bs);
2996
2997 /* Ensure that NULL is never returned on success */
2998 assert(align > 0);
2999 if (size == 0) {
3000 size = align;
3001 }
3002
3003 return qemu_try_memalign(align, size);
3004 }
3005
3006 void *qemu_try_blockalign0(BlockDriverState *bs, size_t size)
3007 {
3008 void *mem = qemu_try_blockalign(bs, size);
3009
3010 if (mem) {
3011 memset(mem, 0, size);
3012 }
3013
3014 return mem;
3015 }
3016
3017 /*
3018 * Check if all memory in this vector is sector aligned.
3019 */
3020 bool bdrv_qiov_is_aligned(BlockDriverState *bs, QEMUIOVector *qiov)
3021 {
3022 int i;
3023 size_t alignment = bdrv_min_mem_align(bs);
3024
3025 for (i = 0; i < qiov->niov; i++) {
3026 if ((uintptr_t) qiov->iov[i].iov_base % alignment) {
3027 return false;
3028 }
3029 if (qiov->iov[i].iov_len % alignment) {
3030 return false;
3031 }
3032 }
3033
3034 return true;
3035 }
3036
3037 void bdrv_add_before_write_notifier(BlockDriverState *bs,
3038 NotifierWithReturn *notifier)
3039 {
3040 notifier_with_return_list_add(&bs->before_write_notifiers, notifier);
3041 }
3042
3043 void bdrv_io_plug(BlockDriverState *bs)
3044 {
3045 BdrvChild *child;
3046
3047 QLIST_FOREACH(child, &bs->children, next) {
3048 bdrv_io_plug(child->bs);
3049 }
3050
3051 if (atomic_fetch_inc(&bs->io_plugged) == 0) {
3052 BlockDriver *drv = bs->drv;
3053 if (drv && drv->bdrv_io_plug) {
3054 drv->bdrv_io_plug(bs);
3055 }
3056 }
3057 }
3058
3059 void bdrv_io_unplug(BlockDriverState *bs)
3060 {
3061 BdrvChild *child;
3062
3063 assert(bs->io_plugged);
3064 if (atomic_fetch_dec(&bs->io_plugged) == 1) {
3065 BlockDriver *drv = bs->drv;
3066 if (drv && drv->bdrv_io_unplug) {
3067 drv->bdrv_io_unplug(bs);
3068 }
3069 }
3070
3071 QLIST_FOREACH(child, &bs->children, next) {
3072 bdrv_io_unplug(child->bs);
3073 }
3074 }
3075
3076 void bdrv_register_buf(BlockDriverState *bs, void *host, size_t size)
3077 {
3078 BdrvChild *child;
3079
3080 if (bs->drv && bs->drv->bdrv_register_buf) {
3081 bs->drv->bdrv_register_buf(bs, host, size);
3082 }
3083 QLIST_FOREACH(child, &bs->children, next) {
3084 bdrv_register_buf(child->bs, host, size);
3085 }
3086 }
3087
3088 void bdrv_unregister_buf(BlockDriverState *bs, void *host)
3089 {
3090 BdrvChild *child;
3091
3092 if (bs->drv && bs->drv->bdrv_unregister_buf) {
3093 bs->drv->bdrv_unregister_buf(bs, host);
3094 }
3095 QLIST_FOREACH(child, &bs->children, next) {
3096 bdrv_unregister_buf(child->bs, host);
3097 }
3098 }
3099
3100 static int coroutine_fn bdrv_co_copy_range_internal(
3101 BdrvChild *src, uint64_t src_offset, BdrvChild *dst,
3102 uint64_t dst_offset, uint64_t bytes,
3103 BdrvRequestFlags read_flags, BdrvRequestFlags write_flags,
3104 bool recurse_src)
3105 {
3106 BdrvTrackedRequest req;
3107 int ret;
3108
3109 /* TODO We can support BDRV_REQ_NO_FALLBACK here */
3110 assert(!(read_flags & BDRV_REQ_NO_FALLBACK));
3111 assert(!(write_flags & BDRV_REQ_NO_FALLBACK));
3112
3113 if (!dst || !dst->bs) {
3114 return -ENOMEDIUM;
3115 }
3116 ret = bdrv_check_byte_request(dst->bs, dst_offset, bytes);
3117 if (ret) {
3118 return ret;
3119 }
3120 if (write_flags & BDRV_REQ_ZERO_WRITE) {
3121 return bdrv_co_pwrite_zeroes(dst, dst_offset, bytes, write_flags);
3122 }
3123
3124 if (!src || !src->bs) {
3125 return -ENOMEDIUM;
3126 }
3127 ret = bdrv_check_byte_request(src->bs, src_offset, bytes);
3128 if (ret) {
3129 return ret;
3130 }
3131
3132 if (!src->bs->drv->bdrv_co_copy_range_from
3133 || !dst->bs->drv->bdrv_co_copy_range_to
3134 || src->bs->encrypted || dst->bs->encrypted) {
3135 return -ENOTSUP;
3136 }
3137
3138 if (recurse_src) {
3139 bdrv_inc_in_flight(src->bs);
3140 tracked_request_begin(&req, src->bs, src_offset, bytes,
3141 BDRV_TRACKED_READ);
3142
3143 /* BDRV_REQ_SERIALISING is only for write operation */
3144 assert(!(read_flags & BDRV_REQ_SERIALISING));
3145 if (!(read_flags & BDRV_REQ_NO_SERIALISING)) {
3146 wait_serialising_requests(&req);
3147 }
3148
3149 ret = src->bs->drv->bdrv_co_copy_range_from(src->bs,
3150 src, src_offset,
3151 dst, dst_offset,
3152 bytes,
3153 read_flags, write_flags);
3154
3155 tracked_request_end(&req);
3156 bdrv_dec_in_flight(src->bs);
3157 } else {
3158 bdrv_inc_in_flight(dst->bs);
3159 tracked_request_begin(&req, dst->bs, dst_offset, bytes,
3160 BDRV_TRACKED_WRITE);
3161 ret = bdrv_co_write_req_prepare(dst, dst_offset, bytes, &req,
3162 write_flags);
3163 if (!ret) {
3164 ret = dst->bs->drv->bdrv_co_copy_range_to(dst->bs,
3165 src, src_offset,
3166 dst, dst_offset,
3167 bytes,
3168 read_flags, write_flags);
3169 }
3170 bdrv_co_write_req_finish(dst, dst_offset, bytes, &req, ret);
3171 tracked_request_end(&req);
3172 bdrv_dec_in_flight(dst->bs);
3173 }
3174
3175 return ret;
3176 }
3177
3178 /* Copy range from @src to @dst.
3179 *
3180 * See the comment of bdrv_co_copy_range for the parameter and return value
3181 * semantics. */
3182 int coroutine_fn bdrv_co_copy_range_from(BdrvChild *src, uint64_t src_offset,
3183 BdrvChild *dst, uint64_t dst_offset,
3184 uint64_t bytes,
3185 BdrvRequestFlags read_flags,
3186 BdrvRequestFlags write_flags)
3187 {
3188 trace_bdrv_co_copy_range_from(src, src_offset, dst, dst_offset, bytes,
3189 read_flags, write_flags);
3190 return bdrv_co_copy_range_internal(src, src_offset, dst, dst_offset,
3191 bytes, read_flags, write_flags, true);
3192 }
3193
3194 /* Copy range from @src to @dst.
3195 *
3196 * See the comment of bdrv_co_copy_range for the parameter and return value
3197 * semantics. */
3198 int coroutine_fn bdrv_co_copy_range_to(BdrvChild *src, uint64_t src_offset,
3199 BdrvChild *dst, uint64_t dst_offset,
3200 uint64_t bytes,
3201 BdrvRequestFlags read_flags,
3202 BdrvRequestFlags write_flags)
3203 {
3204 trace_bdrv_co_copy_range_to(src, src_offset, dst, dst_offset, bytes,
3205 read_flags, write_flags);
3206 return bdrv_co_copy_range_internal(src, src_offset, dst, dst_offset,
3207 bytes, read_flags, write_flags, false);
3208 }
3209
3210 int coroutine_fn bdrv_co_copy_range(BdrvChild *src, uint64_t src_offset,
3211 BdrvChild *dst, uint64_t dst_offset,
3212 uint64_t bytes, BdrvRequestFlags read_flags,
3213 BdrvRequestFlags write_flags)
3214 {
3215 return bdrv_co_copy_range_from(src, src_offset,
3216 dst, dst_offset,
3217 bytes, read_flags, write_flags);
3218 }
3219
3220 static void bdrv_parent_cb_resize(BlockDriverState *bs)
3221 {
3222 BdrvChild *c;
3223 QLIST_FOREACH(c, &bs->parents, next_parent) {
3224 if (c->role->resize) {
3225 c->role->resize(c);
3226 }
3227 }
3228 }
3229
3230 /**
3231 * Truncate file to 'offset' bytes (needed only for file protocols)
3232 */
3233 int coroutine_fn bdrv_co_truncate(BdrvChild *child, int64_t offset,
3234 PreallocMode prealloc, Error **errp)
3235 {
3236 BlockDriverState *bs = child->bs;
3237 BlockDriver *drv = bs->drv;
3238 BdrvTrackedRequest req;
3239 int64_t old_size, new_bytes;
3240 int ret;
3241
3242
3243 /* if bs->drv == NULL, bs is closed, so there's nothing to do here */
3244 if (!drv) {
3245 error_setg(errp, "No medium inserted");
3246 return -ENOMEDIUM;
3247 }
3248 if (offset < 0) {
3249 error_setg(errp, "Image size cannot be negative");
3250 return -EINVAL;
3251 }
3252
3253 old_size = bdrv_getlength(bs);
3254 if (old_size < 0) {
3255 error_setg_errno(errp, -old_size, "Failed to get old image size");
3256 return old_size;
3257 }
3258
3259 if (offset > old_size) {
3260 new_bytes = offset - old_size;
3261 } else {
3262 new_bytes = 0;
3263 }
3264
3265 bdrv_inc_in_flight(bs);
3266 tracked_request_begin(&req, bs, offset - new_bytes, new_bytes,
3267 BDRV_TRACKED_TRUNCATE);
3268
3269 /* If we are growing the image and potentially using preallocation for the
3270 * new area, we need to make sure that no write requests are made to it
3271 * concurrently or they might be overwritten by preallocation. */
3272 if (new_bytes) {
3273 mark_request_serialising(&req, 1);
3274 }
3275 if (bs->read_only) {
3276 error_setg(errp, "Image is read-only");
3277 ret = -EACCES;
3278 goto out;
3279 }
3280 ret = bdrv_co_write_req_prepare(child, offset - new_bytes, new_bytes, &req,
3281 0);
3282 if (ret < 0) {
3283 error_setg_errno(errp, -ret,
3284 "Failed to prepare request for truncation");
3285 goto out;
3286 }
3287
3288 if (!drv->bdrv_co_truncate) {
3289 if (bs->file && drv->is_filter) {
3290 ret = bdrv_co_truncate(bs->file, offset, prealloc, errp);
3291 goto out;
3292 }
3293 error_setg(errp, "Image format driver does not support resize");
3294 ret = -ENOTSUP;
3295 goto out;
3296 }
3297
3298 ret = drv->bdrv_co_truncate(bs, offset, prealloc, errp);
3299 if (ret < 0) {
3300 goto out;
3301 }
3302 ret = refresh_total_sectors(bs, offset >> BDRV_SECTOR_BITS);
3303 if (ret < 0) {
3304 error_setg_errno(errp, -ret, "Could not refresh total sector count");
3305 } else {
3306 offset = bs->total_sectors * BDRV_SECTOR_SIZE;
3307 }
3308 /* It's possible that truncation succeeded but refresh_total_sectors
3309 * failed, but the latter doesn't affect how we should finish the request.
3310 * Pass 0 as the last parameter so that dirty bitmaps etc. are handled. */
3311 bdrv_co_write_req_finish(child, offset - new_bytes, new_bytes, &req, 0);
3312
3313 out:
3314 tracked_request_end(&req);
3315 bdrv_dec_in_flight(bs);
3316
3317 return ret;
3318 }
3319
3320 typedef struct TruncateCo {
3321 BdrvChild *child;
3322 int64_t offset;
3323 PreallocMode prealloc;
3324 Error **errp;
3325 int ret;
3326 } TruncateCo;
3327
3328 static void coroutine_fn bdrv_truncate_co_entry(void *opaque)
3329 {
3330 TruncateCo *tco = opaque;
3331 tco->ret = bdrv_co_truncate(tco->child, tco->offset, tco->prealloc,
3332 tco->errp);
3333 aio_wait_kick();
3334 }
3335
3336 int bdrv_truncate(BdrvChild *child, int64_t offset, PreallocMode prealloc,
3337 Error **errp)
3338 {
3339 Coroutine *co;
3340 TruncateCo tco = {
3341 .child = child,
3342 .offset = offset,
3343 .prealloc = prealloc,
3344 .errp = errp,
3345 .ret = NOT_DONE,
3346 };
3347
3348 if (qemu_in_coroutine()) {
3349 /* Fast-path if already in coroutine context */
3350 bdrv_truncate_co_entry(&tco);
3351 } else {
3352 co = qemu_coroutine_create(bdrv_truncate_co_entry, &tco);
3353 bdrv_coroutine_enter(child->bs, co);
3354 BDRV_POLL_WHILE(child->bs, tco.ret == NOT_DONE);
3355 }
3356
3357 return tco.ret;
3358 }