xref: /openbmc/qemu/block/io.c (revision 8c95e1f20c18007f30bdba6a60e99aea37cc473b)
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/blockjob.h"
29 #include "block/block_int.h"
30 #include "qemu/cutils.h"
31 #include "qapi/error.h"
32 #include "qemu/error-report.h"
33 
34 #define NOT_DONE 0x7fffffff /* used while emulated sync operation in progress */
35 
36 static BlockAIOCB *bdrv_co_aio_prw_vector(BdrvChild *child,
37                                           int64_t offset,
38                                           QEMUIOVector *qiov,
39                                           BdrvRequestFlags flags,
40                                           BlockCompletionFunc *cb,
41                                           void *opaque,
42                                           bool is_write);
43 static void coroutine_fn bdrv_co_do_rw(void *opaque);
44 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs,
45     int64_t offset, int count, BdrvRequestFlags flags);
46 
47 void bdrv_parent_drained_begin(BlockDriverState *bs)
48 {
49     BdrvChild *c;
50 
51     QLIST_FOREACH(c, &bs->parents, next_parent) {
52         if (c->role->drained_begin) {
53             c->role->drained_begin(c);
54         }
55     }
56 }
57 
58 void bdrv_parent_drained_end(BlockDriverState *bs)
59 {
60     BdrvChild *c;
61 
62     QLIST_FOREACH(c, &bs->parents, next_parent) {
63         if (c->role->drained_end) {
64             c->role->drained_end(c);
65         }
66     }
67 }
68 
69 static void bdrv_merge_limits(BlockLimits *dst, const BlockLimits *src)
70 {
71     dst->opt_transfer = MAX(dst->opt_transfer, src->opt_transfer);
72     dst->max_transfer = MIN_NON_ZERO(dst->max_transfer, src->max_transfer);
73     dst->opt_mem_alignment = MAX(dst->opt_mem_alignment,
74                                  src->opt_mem_alignment);
75     dst->min_mem_alignment = MAX(dst->min_mem_alignment,
76                                  src->min_mem_alignment);
77     dst->max_iov = MIN_NON_ZERO(dst->max_iov, src->max_iov);
78 }
79 
80 void bdrv_refresh_limits(BlockDriverState *bs, Error **errp)
81 {
82     BlockDriver *drv = bs->drv;
83     Error *local_err = NULL;
84 
85     memset(&bs->bl, 0, sizeof(bs->bl));
86 
87     if (!drv) {
88         return;
89     }
90 
91     /* Default alignment based on whether driver has byte interface */
92     bs->bl.request_alignment = drv->bdrv_co_preadv ? 1 : 512;
93 
94     /* Take some limits from the children as a default */
95     if (bs->file) {
96         bdrv_refresh_limits(bs->file->bs, &local_err);
97         if (local_err) {
98             error_propagate(errp, local_err);
99             return;
100         }
101         bdrv_merge_limits(&bs->bl, &bs->file->bs->bl);
102     } else {
103         bs->bl.min_mem_alignment = 512;
104         bs->bl.opt_mem_alignment = getpagesize();
105 
106         /* Safe default since most protocols use readv()/writev()/etc */
107         bs->bl.max_iov = IOV_MAX;
108     }
109 
110     if (bs->backing) {
111         bdrv_refresh_limits(bs->backing->bs, &local_err);
112         if (local_err) {
113             error_propagate(errp, local_err);
114             return;
115         }
116         bdrv_merge_limits(&bs->bl, &bs->backing->bs->bl);
117     }
118 
119     /* Then let the driver override it */
120     if (drv->bdrv_refresh_limits) {
121         drv->bdrv_refresh_limits(bs, errp);
122     }
123 }
124 
125 /**
126  * The copy-on-read flag is actually a reference count so multiple users may
127  * use the feature without worrying about clobbering its previous state.
128  * Copy-on-read stays enabled until all users have called to disable it.
129  */
130 void bdrv_enable_copy_on_read(BlockDriverState *bs)
131 {
132     bs->copy_on_read++;
133 }
134 
135 void bdrv_disable_copy_on_read(BlockDriverState *bs)
136 {
137     assert(bs->copy_on_read > 0);
138     bs->copy_on_read--;
139 }
140 
141 /* Check if any requests are in-flight (including throttled requests) */
142 bool bdrv_requests_pending(BlockDriverState *bs)
143 {
144     BdrvChild *child;
145 
146     if (atomic_read(&bs->in_flight)) {
147         return true;
148     }
149 
150     QLIST_FOREACH(child, &bs->children, next) {
151         if (bdrv_requests_pending(child->bs)) {
152             return true;
153         }
154     }
155 
156     return false;
157 }
158 
159 static bool bdrv_drain_recurse(BlockDriverState *bs)
160 {
161     BdrvChild *child, *tmp;
162     bool waited;
163 
164     waited = BDRV_POLL_WHILE(bs, atomic_read(&bs->in_flight) > 0);
165 
166     if (bs->drv && bs->drv->bdrv_drain) {
167         bs->drv->bdrv_drain(bs);
168     }
169 
170     QLIST_FOREACH_SAFE(child, &bs->children, next, tmp) {
171         BlockDriverState *bs = child->bs;
172         bool in_main_loop =
173             qemu_get_current_aio_context() == qemu_get_aio_context();
174         assert(bs->refcnt > 0);
175         if (in_main_loop) {
176             /* In case the recursive bdrv_drain_recurse processes a
177              * block_job_defer_to_main_loop BH and modifies the graph,
178              * let's hold a reference to bs until we are done.
179              *
180              * IOThread doesn't have such a BH, and it is not safe to call
181              * bdrv_unref without BQL, so skip doing it there.
182              */
183             bdrv_ref(bs);
184         }
185         waited |= bdrv_drain_recurse(bs);
186         if (in_main_loop) {
187             bdrv_unref(bs);
188         }
189     }
190 
191     return waited;
192 }
193 
194 typedef struct {
195     Coroutine *co;
196     BlockDriverState *bs;
197     bool done;
198 } BdrvCoDrainData;
199 
200 static void bdrv_co_drain_bh_cb(void *opaque)
201 {
202     BdrvCoDrainData *data = opaque;
203     Coroutine *co = data->co;
204     BlockDriverState *bs = data->bs;
205 
206     bdrv_dec_in_flight(bs);
207     bdrv_drained_begin(bs);
208     data->done = true;
209     aio_co_wake(co);
210 }
211 
212 static void coroutine_fn bdrv_co_yield_to_drain(BlockDriverState *bs)
213 {
214     BdrvCoDrainData data;
215 
216     /* Calling bdrv_drain() from a BH ensures the current coroutine yields and
217      * other coroutines run if they were queued from
218      * qemu_co_queue_run_restart(). */
219 
220     assert(qemu_in_coroutine());
221     data = (BdrvCoDrainData) {
222         .co = qemu_coroutine_self(),
223         .bs = bs,
224         .done = false,
225     };
226     bdrv_inc_in_flight(bs);
227     aio_bh_schedule_oneshot(bdrv_get_aio_context(bs),
228                             bdrv_co_drain_bh_cb, &data);
229 
230     qemu_coroutine_yield();
231     /* If we are resumed from some other event (such as an aio completion or a
232      * timer callback), it is a bug in the caller that should be fixed. */
233     assert(data.done);
234 }
235 
236 void bdrv_drained_begin(BlockDriverState *bs)
237 {
238     if (qemu_in_coroutine()) {
239         bdrv_co_yield_to_drain(bs);
240         return;
241     }
242 
243     if (!bs->quiesce_counter++) {
244         aio_disable_external(bdrv_get_aio_context(bs));
245         bdrv_parent_drained_begin(bs);
246     }
247 
248     bdrv_drain_recurse(bs);
249 }
250 
251 void bdrv_drained_end(BlockDriverState *bs)
252 {
253     assert(bs->quiesce_counter > 0);
254     if (--bs->quiesce_counter > 0) {
255         return;
256     }
257 
258     bdrv_parent_drained_end(bs);
259     aio_enable_external(bdrv_get_aio_context(bs));
260 }
261 
262 /*
263  * Wait for pending requests to complete on a single BlockDriverState subtree,
264  * and suspend block driver's internal I/O until next request arrives.
265  *
266  * Note that unlike bdrv_drain_all(), the caller must hold the BlockDriverState
267  * AioContext.
268  *
269  * Only this BlockDriverState's AioContext is run, so in-flight requests must
270  * not depend on events in other AioContexts.  In that case, use
271  * bdrv_drain_all() instead.
272  */
273 void coroutine_fn bdrv_co_drain(BlockDriverState *bs)
274 {
275     assert(qemu_in_coroutine());
276     bdrv_drained_begin(bs);
277     bdrv_drained_end(bs);
278 }
279 
280 void bdrv_drain(BlockDriverState *bs)
281 {
282     bdrv_drained_begin(bs);
283     bdrv_drained_end(bs);
284 }
285 
286 /*
287  * Wait for pending requests to complete across all BlockDriverStates
288  *
289  * This function does not flush data to disk, use bdrv_flush_all() for that
290  * after calling this function.
291  *
292  * This pauses all block jobs and disables external clients. It must
293  * be paired with bdrv_drain_all_end().
294  *
295  * NOTE: no new block jobs or BlockDriverStates can be created between
296  * the bdrv_drain_all_begin() and bdrv_drain_all_end() calls.
297  */
298 void bdrv_drain_all_begin(void)
299 {
300     /* Always run first iteration so any pending completion BHs run */
301     bool waited = true;
302     BlockDriverState *bs;
303     BdrvNextIterator it;
304     BlockJob *job = NULL;
305     GSList *aio_ctxs = NULL, *ctx;
306 
307     while ((job = block_job_next(job))) {
308         AioContext *aio_context = blk_get_aio_context(job->blk);
309 
310         aio_context_acquire(aio_context);
311         block_job_pause(job);
312         aio_context_release(aio_context);
313     }
314 
315     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
316         AioContext *aio_context = bdrv_get_aio_context(bs);
317 
318         aio_context_acquire(aio_context);
319         bdrv_parent_drained_begin(bs);
320         aio_disable_external(aio_context);
321         aio_context_release(aio_context);
322 
323         if (!g_slist_find(aio_ctxs, aio_context)) {
324             aio_ctxs = g_slist_prepend(aio_ctxs, aio_context);
325         }
326     }
327 
328     /* Note that completion of an asynchronous I/O operation can trigger any
329      * number of other I/O operations on other devices---for example a
330      * coroutine can submit an I/O request to another device in response to
331      * request completion.  Therefore we must keep looping until there was no
332      * more activity rather than simply draining each device independently.
333      */
334     while (waited) {
335         waited = false;
336 
337         for (ctx = aio_ctxs; ctx != NULL; ctx = ctx->next) {
338             AioContext *aio_context = ctx->data;
339 
340             aio_context_acquire(aio_context);
341             for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
342                 if (aio_context == bdrv_get_aio_context(bs)) {
343                     waited |= bdrv_drain_recurse(bs);
344                 }
345             }
346             aio_context_release(aio_context);
347         }
348     }
349 
350     g_slist_free(aio_ctxs);
351 }
352 
353 void bdrv_drain_all_end(void)
354 {
355     BlockDriverState *bs;
356     BdrvNextIterator it;
357     BlockJob *job = NULL;
358 
359     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
360         AioContext *aio_context = bdrv_get_aio_context(bs);
361 
362         aio_context_acquire(aio_context);
363         aio_enable_external(aio_context);
364         bdrv_parent_drained_end(bs);
365         aio_context_release(aio_context);
366     }
367 
368     while ((job = block_job_next(job))) {
369         AioContext *aio_context = blk_get_aio_context(job->blk);
370 
371         aio_context_acquire(aio_context);
372         block_job_resume(job);
373         aio_context_release(aio_context);
374     }
375 }
376 
377 void bdrv_drain_all(void)
378 {
379     bdrv_drain_all_begin();
380     bdrv_drain_all_end();
381 }
382 
383 /**
384  * Remove an active request from the tracked requests list
385  *
386  * This function should be called when a tracked request is completing.
387  */
388 static void tracked_request_end(BdrvTrackedRequest *req)
389 {
390     if (req->serialising) {
391         req->bs->serialising_in_flight--;
392     }
393 
394     QLIST_REMOVE(req, list);
395     qemu_co_queue_restart_all(&req->wait_queue);
396 }
397 
398 /**
399  * Add an active request to the tracked requests list
400  */
401 static void tracked_request_begin(BdrvTrackedRequest *req,
402                                   BlockDriverState *bs,
403                                   int64_t offset,
404                                   unsigned int bytes,
405                                   enum BdrvTrackedRequestType type)
406 {
407     *req = (BdrvTrackedRequest){
408         .bs = bs,
409         .offset         = offset,
410         .bytes          = bytes,
411         .type           = type,
412         .co             = qemu_coroutine_self(),
413         .serialising    = false,
414         .overlap_offset = offset,
415         .overlap_bytes  = bytes,
416     };
417 
418     qemu_co_queue_init(&req->wait_queue);
419 
420     QLIST_INSERT_HEAD(&bs->tracked_requests, req, list);
421 }
422 
423 static void mark_request_serialising(BdrvTrackedRequest *req, uint64_t align)
424 {
425     int64_t overlap_offset = req->offset & ~(align - 1);
426     unsigned int overlap_bytes = ROUND_UP(req->offset + req->bytes, align)
427                                - overlap_offset;
428 
429     if (!req->serialising) {
430         req->bs->serialising_in_flight++;
431         req->serialising = true;
432     }
433 
434     req->overlap_offset = MIN(req->overlap_offset, overlap_offset);
435     req->overlap_bytes = MAX(req->overlap_bytes, overlap_bytes);
436 }
437 
438 /**
439  * Round a region to cluster boundaries (sector-based)
440  */
441 void bdrv_round_sectors_to_clusters(BlockDriverState *bs,
442                                     int64_t sector_num, int nb_sectors,
443                                     int64_t *cluster_sector_num,
444                                     int *cluster_nb_sectors)
445 {
446     BlockDriverInfo bdi;
447 
448     if (bdrv_get_info(bs, &bdi) < 0 || bdi.cluster_size == 0) {
449         *cluster_sector_num = sector_num;
450         *cluster_nb_sectors = nb_sectors;
451     } else {
452         int64_t c = bdi.cluster_size / BDRV_SECTOR_SIZE;
453         *cluster_sector_num = QEMU_ALIGN_DOWN(sector_num, c);
454         *cluster_nb_sectors = QEMU_ALIGN_UP(sector_num - *cluster_sector_num +
455                                             nb_sectors, c);
456     }
457 }
458 
459 /**
460  * Round a region to cluster boundaries
461  */
462 void bdrv_round_to_clusters(BlockDriverState *bs,
463                             int64_t offset, unsigned int bytes,
464                             int64_t *cluster_offset,
465                             unsigned int *cluster_bytes)
466 {
467     BlockDriverInfo bdi;
468 
469     if (bdrv_get_info(bs, &bdi) < 0 || bdi.cluster_size == 0) {
470         *cluster_offset = offset;
471         *cluster_bytes = bytes;
472     } else {
473         int64_t c = bdi.cluster_size;
474         *cluster_offset = QEMU_ALIGN_DOWN(offset, c);
475         *cluster_bytes = QEMU_ALIGN_UP(offset - *cluster_offset + bytes, c);
476     }
477 }
478 
479 static int bdrv_get_cluster_size(BlockDriverState *bs)
480 {
481     BlockDriverInfo bdi;
482     int ret;
483 
484     ret = bdrv_get_info(bs, &bdi);
485     if (ret < 0 || bdi.cluster_size == 0) {
486         return bs->bl.request_alignment;
487     } else {
488         return bdi.cluster_size;
489     }
490 }
491 
492 static bool tracked_request_overlaps(BdrvTrackedRequest *req,
493                                      int64_t offset, unsigned int bytes)
494 {
495     /*        aaaa   bbbb */
496     if (offset >= req->overlap_offset + req->overlap_bytes) {
497         return false;
498     }
499     /* bbbb   aaaa        */
500     if (req->overlap_offset >= offset + bytes) {
501         return false;
502     }
503     return true;
504 }
505 
506 void bdrv_inc_in_flight(BlockDriverState *bs)
507 {
508     atomic_inc(&bs->in_flight);
509 }
510 
511 static void dummy_bh_cb(void *opaque)
512 {
513 }
514 
515 void bdrv_wakeup(BlockDriverState *bs)
516 {
517     if (bs->wakeup) {
518         aio_bh_schedule_oneshot(qemu_get_aio_context(), dummy_bh_cb, NULL);
519     }
520 }
521 
522 void bdrv_dec_in_flight(BlockDriverState *bs)
523 {
524     atomic_dec(&bs->in_flight);
525     bdrv_wakeup(bs);
526 }
527 
528 static bool coroutine_fn wait_serialising_requests(BdrvTrackedRequest *self)
529 {
530     BlockDriverState *bs = self->bs;
531     BdrvTrackedRequest *req;
532     bool retry;
533     bool waited = false;
534 
535     if (!bs->serialising_in_flight) {
536         return false;
537     }
538 
539     do {
540         retry = false;
541         QLIST_FOREACH(req, &bs->tracked_requests, list) {
542             if (req == self || (!req->serialising && !self->serialising)) {
543                 continue;
544             }
545             if (tracked_request_overlaps(req, self->overlap_offset,
546                                          self->overlap_bytes))
547             {
548                 /* Hitting this means there was a reentrant request, for
549                  * example, a block driver issuing nested requests.  This must
550                  * never happen since it means deadlock.
551                  */
552                 assert(qemu_coroutine_self() != req->co);
553 
554                 /* If the request is already (indirectly) waiting for us, or
555                  * will wait for us as soon as it wakes up, then just go on
556                  * (instead of producing a deadlock in the former case). */
557                 if (!req->waiting_for) {
558                     self->waiting_for = req;
559                     qemu_co_queue_wait(&req->wait_queue, NULL);
560                     self->waiting_for = NULL;
561                     retry = true;
562                     waited = true;
563                     break;
564                 }
565             }
566         }
567     } while (retry);
568 
569     return waited;
570 }
571 
572 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset,
573                                    size_t size)
574 {
575     if (size > BDRV_REQUEST_MAX_SECTORS << BDRV_SECTOR_BITS) {
576         return -EIO;
577     }
578 
579     if (!bdrv_is_inserted(bs)) {
580         return -ENOMEDIUM;
581     }
582 
583     if (offset < 0) {
584         return -EIO;
585     }
586 
587     return 0;
588 }
589 
590 typedef struct RwCo {
591     BdrvChild *child;
592     int64_t offset;
593     QEMUIOVector *qiov;
594     bool is_write;
595     int ret;
596     BdrvRequestFlags flags;
597 } RwCo;
598 
599 static void coroutine_fn bdrv_rw_co_entry(void *opaque)
600 {
601     RwCo *rwco = opaque;
602 
603     if (!rwco->is_write) {
604         rwco->ret = bdrv_co_preadv(rwco->child, rwco->offset,
605                                    rwco->qiov->size, rwco->qiov,
606                                    rwco->flags);
607     } else {
608         rwco->ret = bdrv_co_pwritev(rwco->child, rwco->offset,
609                                     rwco->qiov->size, rwco->qiov,
610                                     rwco->flags);
611     }
612 }
613 
614 /*
615  * Process a vectored synchronous request using coroutines
616  */
617 static int bdrv_prwv_co(BdrvChild *child, int64_t offset,
618                         QEMUIOVector *qiov, bool is_write,
619                         BdrvRequestFlags flags)
620 {
621     Coroutine *co;
622     RwCo rwco = {
623         .child = child,
624         .offset = offset,
625         .qiov = qiov,
626         .is_write = is_write,
627         .ret = NOT_DONE,
628         .flags = flags,
629     };
630 
631     if (qemu_in_coroutine()) {
632         /* Fast-path if already in coroutine context */
633         bdrv_rw_co_entry(&rwco);
634     } else {
635         co = qemu_coroutine_create(bdrv_rw_co_entry, &rwco);
636         bdrv_coroutine_enter(child->bs, co);
637         BDRV_POLL_WHILE(child->bs, rwco.ret == NOT_DONE);
638     }
639     return rwco.ret;
640 }
641 
642 /*
643  * Process a synchronous request using coroutines
644  */
645 static int bdrv_rw_co(BdrvChild *child, int64_t sector_num, uint8_t *buf,
646                       int nb_sectors, bool is_write, BdrvRequestFlags flags)
647 {
648     QEMUIOVector qiov;
649     struct iovec iov = {
650         .iov_base = (void *)buf,
651         .iov_len = nb_sectors * BDRV_SECTOR_SIZE,
652     };
653 
654     if (nb_sectors < 0 || nb_sectors > BDRV_REQUEST_MAX_SECTORS) {
655         return -EINVAL;
656     }
657 
658     qemu_iovec_init_external(&qiov, &iov, 1);
659     return bdrv_prwv_co(child, sector_num << BDRV_SECTOR_BITS,
660                         &qiov, is_write, flags);
661 }
662 
663 /* return < 0 if error. See bdrv_write() for the return codes */
664 int bdrv_read(BdrvChild *child, int64_t sector_num,
665               uint8_t *buf, int nb_sectors)
666 {
667     return bdrv_rw_co(child, sector_num, buf, nb_sectors, false, 0);
668 }
669 
670 /* Return < 0 if error. Important errors are:
671   -EIO         generic I/O error (may happen for all errors)
672   -ENOMEDIUM   No media inserted.
673   -EINVAL      Invalid sector number or nb_sectors
674   -EACCES      Trying to write a read-only device
675 */
676 int bdrv_write(BdrvChild *child, int64_t sector_num,
677                const uint8_t *buf, int nb_sectors)
678 {
679     return bdrv_rw_co(child, sector_num, (uint8_t *)buf, nb_sectors, true, 0);
680 }
681 
682 int bdrv_pwrite_zeroes(BdrvChild *child, int64_t offset,
683                        int count, BdrvRequestFlags flags)
684 {
685     QEMUIOVector qiov;
686     struct iovec iov = {
687         .iov_base = NULL,
688         .iov_len = count,
689     };
690 
691     qemu_iovec_init_external(&qiov, &iov, 1);
692     return bdrv_prwv_co(child, offset, &qiov, true,
693                         BDRV_REQ_ZERO_WRITE | flags);
694 }
695 
696 /*
697  * Completely zero out a block device with the help of bdrv_pwrite_zeroes.
698  * The operation is sped up by checking the block status and only writing
699  * zeroes to the device if they currently do not return zeroes. Optional
700  * flags are passed through to bdrv_pwrite_zeroes (e.g. BDRV_REQ_MAY_UNMAP,
701  * BDRV_REQ_FUA).
702  *
703  * Returns < 0 on error, 0 on success. For error codes see bdrv_write().
704  */
705 int bdrv_make_zero(BdrvChild *child, BdrvRequestFlags flags)
706 {
707     int64_t target_sectors, ret, nb_sectors, sector_num = 0;
708     BlockDriverState *bs = child->bs;
709     BlockDriverState *file;
710     int n;
711 
712     target_sectors = bdrv_nb_sectors(bs);
713     if (target_sectors < 0) {
714         return target_sectors;
715     }
716 
717     for (;;) {
718         nb_sectors = MIN(target_sectors - sector_num, BDRV_REQUEST_MAX_SECTORS);
719         if (nb_sectors <= 0) {
720             return 0;
721         }
722         ret = bdrv_get_block_status(bs, sector_num, nb_sectors, &n, &file);
723         if (ret < 0) {
724             error_report("error getting block status at sector %" PRId64 ": %s",
725                          sector_num, strerror(-ret));
726             return ret;
727         }
728         if (ret & BDRV_BLOCK_ZERO) {
729             sector_num += n;
730             continue;
731         }
732         ret = bdrv_pwrite_zeroes(child, sector_num << BDRV_SECTOR_BITS,
733                                  n << BDRV_SECTOR_BITS, flags);
734         if (ret < 0) {
735             error_report("error writing zeroes at sector %" PRId64 ": %s",
736                          sector_num, strerror(-ret));
737             return ret;
738         }
739         sector_num += n;
740     }
741 }
742 
743 int bdrv_preadv(BdrvChild *child, int64_t offset, QEMUIOVector *qiov)
744 {
745     int ret;
746 
747     ret = bdrv_prwv_co(child, offset, qiov, false, 0);
748     if (ret < 0) {
749         return ret;
750     }
751 
752     return qiov->size;
753 }
754 
755 int bdrv_pread(BdrvChild *child, int64_t offset, void *buf, int bytes)
756 {
757     QEMUIOVector qiov;
758     struct iovec iov = {
759         .iov_base = (void *)buf,
760         .iov_len = bytes,
761     };
762 
763     if (bytes < 0) {
764         return -EINVAL;
765     }
766 
767     qemu_iovec_init_external(&qiov, &iov, 1);
768     return bdrv_preadv(child, offset, &qiov);
769 }
770 
771 int bdrv_pwritev(BdrvChild *child, int64_t offset, QEMUIOVector *qiov)
772 {
773     int ret;
774 
775     ret = bdrv_prwv_co(child, offset, qiov, true, 0);
776     if (ret < 0) {
777         return ret;
778     }
779 
780     return qiov->size;
781 }
782 
783 int bdrv_pwrite(BdrvChild *child, int64_t offset, const void *buf, int bytes)
784 {
785     QEMUIOVector qiov;
786     struct iovec iov = {
787         .iov_base   = (void *) buf,
788         .iov_len    = bytes,
789     };
790 
791     if (bytes < 0) {
792         return -EINVAL;
793     }
794 
795     qemu_iovec_init_external(&qiov, &iov, 1);
796     return bdrv_pwritev(child, offset, &qiov);
797 }
798 
799 /*
800  * Writes to the file and ensures that no writes are reordered across this
801  * request (acts as a barrier)
802  *
803  * Returns 0 on success, -errno in error cases.
804  */
805 int bdrv_pwrite_sync(BdrvChild *child, int64_t offset,
806                      const void *buf, int count)
807 {
808     int ret;
809 
810     ret = bdrv_pwrite(child, offset, buf, count);
811     if (ret < 0) {
812         return ret;
813     }
814 
815     ret = bdrv_flush(child->bs);
816     if (ret < 0) {
817         return ret;
818     }
819 
820     return 0;
821 }
822 
823 typedef struct CoroutineIOCompletion {
824     Coroutine *coroutine;
825     int ret;
826 } CoroutineIOCompletion;
827 
828 static void bdrv_co_io_em_complete(void *opaque, int ret)
829 {
830     CoroutineIOCompletion *co = opaque;
831 
832     co->ret = ret;
833     aio_co_wake(co->coroutine);
834 }
835 
836 static int coroutine_fn bdrv_driver_preadv(BlockDriverState *bs,
837                                            uint64_t offset, uint64_t bytes,
838                                            QEMUIOVector *qiov, int flags)
839 {
840     BlockDriver *drv = bs->drv;
841     int64_t sector_num;
842     unsigned int nb_sectors;
843 
844     assert(!(flags & ~BDRV_REQ_MASK));
845 
846     if (drv->bdrv_co_preadv) {
847         return drv->bdrv_co_preadv(bs, offset, bytes, qiov, flags);
848     }
849 
850     sector_num = offset >> BDRV_SECTOR_BITS;
851     nb_sectors = bytes >> BDRV_SECTOR_BITS;
852 
853     assert((offset & (BDRV_SECTOR_SIZE - 1)) == 0);
854     assert((bytes & (BDRV_SECTOR_SIZE - 1)) == 0);
855     assert((bytes >> BDRV_SECTOR_BITS) <= BDRV_REQUEST_MAX_SECTORS);
856 
857     if (drv->bdrv_co_readv) {
858         return drv->bdrv_co_readv(bs, sector_num, nb_sectors, qiov);
859     } else {
860         BlockAIOCB *acb;
861         CoroutineIOCompletion co = {
862             .coroutine = qemu_coroutine_self(),
863         };
864 
865         acb = bs->drv->bdrv_aio_readv(bs, sector_num, qiov, nb_sectors,
866                                       bdrv_co_io_em_complete, &co);
867         if (acb == NULL) {
868             return -EIO;
869         } else {
870             qemu_coroutine_yield();
871             return co.ret;
872         }
873     }
874 }
875 
876 static int coroutine_fn bdrv_driver_pwritev(BlockDriverState *bs,
877                                             uint64_t offset, uint64_t bytes,
878                                             QEMUIOVector *qiov, int flags)
879 {
880     BlockDriver *drv = bs->drv;
881     int64_t sector_num;
882     unsigned int nb_sectors;
883     int ret;
884 
885     assert(!(flags & ~BDRV_REQ_MASK));
886 
887     if (drv->bdrv_co_pwritev) {
888         ret = drv->bdrv_co_pwritev(bs, offset, bytes, qiov,
889                                    flags & bs->supported_write_flags);
890         flags &= ~bs->supported_write_flags;
891         goto emulate_flags;
892     }
893 
894     sector_num = offset >> BDRV_SECTOR_BITS;
895     nb_sectors = bytes >> BDRV_SECTOR_BITS;
896 
897     assert((offset & (BDRV_SECTOR_SIZE - 1)) == 0);
898     assert((bytes & (BDRV_SECTOR_SIZE - 1)) == 0);
899     assert((bytes >> BDRV_SECTOR_BITS) <= BDRV_REQUEST_MAX_SECTORS);
900 
901     if (drv->bdrv_co_writev_flags) {
902         ret = drv->bdrv_co_writev_flags(bs, sector_num, nb_sectors, qiov,
903                                         flags & bs->supported_write_flags);
904         flags &= ~bs->supported_write_flags;
905     } else if (drv->bdrv_co_writev) {
906         assert(!bs->supported_write_flags);
907         ret = drv->bdrv_co_writev(bs, sector_num, nb_sectors, qiov);
908     } else {
909         BlockAIOCB *acb;
910         CoroutineIOCompletion co = {
911             .coroutine = qemu_coroutine_self(),
912         };
913 
914         acb = bs->drv->bdrv_aio_writev(bs, sector_num, qiov, nb_sectors,
915                                        bdrv_co_io_em_complete, &co);
916         if (acb == NULL) {
917             ret = -EIO;
918         } else {
919             qemu_coroutine_yield();
920             ret = co.ret;
921         }
922     }
923 
924 emulate_flags:
925     if (ret == 0 && (flags & BDRV_REQ_FUA)) {
926         ret = bdrv_co_flush(bs);
927     }
928 
929     return ret;
930 }
931 
932 static int coroutine_fn
933 bdrv_driver_pwritev_compressed(BlockDriverState *bs, uint64_t offset,
934                                uint64_t bytes, QEMUIOVector *qiov)
935 {
936     BlockDriver *drv = bs->drv;
937 
938     if (!drv->bdrv_co_pwritev_compressed) {
939         return -ENOTSUP;
940     }
941 
942     return drv->bdrv_co_pwritev_compressed(bs, offset, bytes, qiov);
943 }
944 
945 static int coroutine_fn bdrv_co_do_copy_on_readv(BdrvChild *child,
946         int64_t offset, unsigned int bytes, QEMUIOVector *qiov)
947 {
948     BlockDriverState *bs = child->bs;
949 
950     /* Perform I/O through a temporary buffer so that users who scribble over
951      * their read buffer while the operation is in progress do not end up
952      * modifying the image file.  This is critical for zero-copy guest I/O
953      * where anything might happen inside guest memory.
954      */
955     void *bounce_buffer;
956 
957     BlockDriver *drv = bs->drv;
958     struct iovec iov;
959     QEMUIOVector bounce_qiov;
960     int64_t cluster_offset;
961     unsigned int cluster_bytes;
962     size_t skip_bytes;
963     int ret;
964 
965     /* FIXME We cannot require callers to have write permissions when all they
966      * are doing is a read request. If we did things right, write permissions
967      * would be obtained anyway, but internally by the copy-on-read code. As
968      * long as it is implemented here rather than in a separat filter driver,
969      * the copy-on-read code doesn't have its own BdrvChild, however, for which
970      * it could request permissions. Therefore we have to bypass the permission
971      * system for the moment. */
972     // assert(child->perm & (BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE));
973 
974     /* Cover entire cluster so no additional backing file I/O is required when
975      * allocating cluster in the image file.
976      */
977     bdrv_round_to_clusters(bs, offset, bytes, &cluster_offset, &cluster_bytes);
978 
979     trace_bdrv_co_do_copy_on_readv(bs, offset, bytes,
980                                    cluster_offset, cluster_bytes);
981 
982     iov.iov_len = cluster_bytes;
983     iov.iov_base = bounce_buffer = qemu_try_blockalign(bs, iov.iov_len);
984     if (bounce_buffer == NULL) {
985         ret = -ENOMEM;
986         goto err;
987     }
988 
989     qemu_iovec_init_external(&bounce_qiov, &iov, 1);
990 
991     ret = bdrv_driver_preadv(bs, cluster_offset, cluster_bytes,
992                              &bounce_qiov, 0);
993     if (ret < 0) {
994         goto err;
995     }
996 
997     if (drv->bdrv_co_pwrite_zeroes &&
998         buffer_is_zero(bounce_buffer, iov.iov_len)) {
999         /* FIXME: Should we (perhaps conditionally) be setting
1000          * BDRV_REQ_MAY_UNMAP, if it will allow for a sparser copy
1001          * that still correctly reads as zero? */
1002         ret = bdrv_co_do_pwrite_zeroes(bs, cluster_offset, cluster_bytes, 0);
1003     } else {
1004         /* This does not change the data on the disk, it is not necessary
1005          * to flush even in cache=writethrough mode.
1006          */
1007         ret = bdrv_driver_pwritev(bs, cluster_offset, cluster_bytes,
1008                                   &bounce_qiov, 0);
1009     }
1010 
1011     if (ret < 0) {
1012         /* It might be okay to ignore write errors for guest requests.  If this
1013          * is a deliberate copy-on-read then we don't want to ignore the error.
1014          * Simply report it in all cases.
1015          */
1016         goto err;
1017     }
1018 
1019     skip_bytes = offset - cluster_offset;
1020     qemu_iovec_from_buf(qiov, 0, bounce_buffer + skip_bytes, bytes);
1021 
1022 err:
1023     qemu_vfree(bounce_buffer);
1024     return ret;
1025 }
1026 
1027 /*
1028  * Forwards an already correctly aligned request to the BlockDriver. This
1029  * handles copy on read, zeroing after EOF, and fragmentation of large
1030  * reads; any other features must be implemented by the caller.
1031  */
1032 static int coroutine_fn bdrv_aligned_preadv(BdrvChild *child,
1033     BdrvTrackedRequest *req, int64_t offset, unsigned int bytes,
1034     int64_t align, QEMUIOVector *qiov, int flags)
1035 {
1036     BlockDriverState *bs = child->bs;
1037     int64_t total_bytes, max_bytes;
1038     int ret = 0;
1039     uint64_t bytes_remaining = bytes;
1040     int max_transfer;
1041 
1042     assert(is_power_of_2(align));
1043     assert((offset & (align - 1)) == 0);
1044     assert((bytes & (align - 1)) == 0);
1045     assert(!qiov || bytes == qiov->size);
1046     assert((bs->open_flags & BDRV_O_NO_IO) == 0);
1047     max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX),
1048                                    align);
1049 
1050     /* TODO: We would need a per-BDS .supported_read_flags and
1051      * potential fallback support, if we ever implement any read flags
1052      * to pass through to drivers.  For now, there aren't any
1053      * passthrough flags.  */
1054     assert(!(flags & ~(BDRV_REQ_NO_SERIALISING | BDRV_REQ_COPY_ON_READ)));
1055 
1056     /* Handle Copy on Read and associated serialisation */
1057     if (flags & BDRV_REQ_COPY_ON_READ) {
1058         /* If we touch the same cluster it counts as an overlap.  This
1059          * guarantees that allocating writes will be serialized and not race
1060          * with each other for the same cluster.  For example, in copy-on-read
1061          * it ensures that the CoR read and write operations are atomic and
1062          * guest writes cannot interleave between them. */
1063         mark_request_serialising(req, bdrv_get_cluster_size(bs));
1064     }
1065 
1066     if (!(flags & BDRV_REQ_NO_SERIALISING)) {
1067         wait_serialising_requests(req);
1068     }
1069 
1070     if (flags & BDRV_REQ_COPY_ON_READ) {
1071         int64_t start_sector = offset >> BDRV_SECTOR_BITS;
1072         int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);
1073         unsigned int nb_sectors = end_sector - start_sector;
1074         int pnum;
1075 
1076         ret = bdrv_is_allocated(bs, start_sector, nb_sectors, &pnum);
1077         if (ret < 0) {
1078             goto out;
1079         }
1080 
1081         if (!ret || pnum != nb_sectors) {
1082             ret = bdrv_co_do_copy_on_readv(child, offset, bytes, qiov);
1083             goto out;
1084         }
1085     }
1086 
1087     /* Forward the request to the BlockDriver, possibly fragmenting it */
1088     total_bytes = bdrv_getlength(bs);
1089     if (total_bytes < 0) {
1090         ret = total_bytes;
1091         goto out;
1092     }
1093 
1094     max_bytes = ROUND_UP(MAX(0, total_bytes - offset), align);
1095     if (bytes <= max_bytes && bytes <= max_transfer) {
1096         ret = bdrv_driver_preadv(bs, offset, bytes, qiov, 0);
1097         goto out;
1098     }
1099 
1100     while (bytes_remaining) {
1101         int num;
1102 
1103         if (max_bytes) {
1104             QEMUIOVector local_qiov;
1105 
1106             num = MIN(bytes_remaining, MIN(max_bytes, max_transfer));
1107             assert(num);
1108             qemu_iovec_init(&local_qiov, qiov->niov);
1109             qemu_iovec_concat(&local_qiov, qiov, bytes - bytes_remaining, num);
1110 
1111             ret = bdrv_driver_preadv(bs, offset + bytes - bytes_remaining,
1112                                      num, &local_qiov, 0);
1113             max_bytes -= num;
1114             qemu_iovec_destroy(&local_qiov);
1115         } else {
1116             num = bytes_remaining;
1117             ret = qemu_iovec_memset(qiov, bytes - bytes_remaining, 0,
1118                                     bytes_remaining);
1119         }
1120         if (ret < 0) {
1121             goto out;
1122         }
1123         bytes_remaining -= num;
1124     }
1125 
1126 out:
1127     return ret < 0 ? ret : 0;
1128 }
1129 
1130 /*
1131  * Handle a read request in coroutine context
1132  */
1133 int coroutine_fn bdrv_co_preadv(BdrvChild *child,
1134     int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
1135     BdrvRequestFlags flags)
1136 {
1137     BlockDriverState *bs = child->bs;
1138     BlockDriver *drv = bs->drv;
1139     BdrvTrackedRequest req;
1140 
1141     uint64_t align = bs->bl.request_alignment;
1142     uint8_t *head_buf = NULL;
1143     uint8_t *tail_buf = NULL;
1144     QEMUIOVector local_qiov;
1145     bool use_local_qiov = false;
1146     int ret;
1147 
1148     if (!drv) {
1149         return -ENOMEDIUM;
1150     }
1151 
1152     ret = bdrv_check_byte_request(bs, offset, bytes);
1153     if (ret < 0) {
1154         return ret;
1155     }
1156 
1157     bdrv_inc_in_flight(bs);
1158 
1159     /* Don't do copy-on-read if we read data before write operation */
1160     if (bs->copy_on_read && !(flags & BDRV_REQ_NO_SERIALISING)) {
1161         flags |= BDRV_REQ_COPY_ON_READ;
1162     }
1163 
1164     /* Align read if necessary by padding qiov */
1165     if (offset & (align - 1)) {
1166         head_buf = qemu_blockalign(bs, align);
1167         qemu_iovec_init(&local_qiov, qiov->niov + 2);
1168         qemu_iovec_add(&local_qiov, head_buf, offset & (align - 1));
1169         qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size);
1170         use_local_qiov = true;
1171 
1172         bytes += offset & (align - 1);
1173         offset = offset & ~(align - 1);
1174     }
1175 
1176     if ((offset + bytes) & (align - 1)) {
1177         if (!use_local_qiov) {
1178             qemu_iovec_init(&local_qiov, qiov->niov + 1);
1179             qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size);
1180             use_local_qiov = true;
1181         }
1182         tail_buf = qemu_blockalign(bs, align);
1183         qemu_iovec_add(&local_qiov, tail_buf,
1184                        align - ((offset + bytes) & (align - 1)));
1185 
1186         bytes = ROUND_UP(bytes, align);
1187     }
1188 
1189     tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_READ);
1190     ret = bdrv_aligned_preadv(child, &req, offset, bytes, align,
1191                               use_local_qiov ? &local_qiov : qiov,
1192                               flags);
1193     tracked_request_end(&req);
1194     bdrv_dec_in_flight(bs);
1195 
1196     if (use_local_qiov) {
1197         qemu_iovec_destroy(&local_qiov);
1198         qemu_vfree(head_buf);
1199         qemu_vfree(tail_buf);
1200     }
1201 
1202     return ret;
1203 }
1204 
1205 static int coroutine_fn bdrv_co_do_readv(BdrvChild *child,
1206     int64_t sector_num, int nb_sectors, QEMUIOVector *qiov,
1207     BdrvRequestFlags flags)
1208 {
1209     if (nb_sectors < 0 || nb_sectors > BDRV_REQUEST_MAX_SECTORS) {
1210         return -EINVAL;
1211     }
1212 
1213     return bdrv_co_preadv(child, sector_num << BDRV_SECTOR_BITS,
1214                           nb_sectors << BDRV_SECTOR_BITS, qiov, flags);
1215 }
1216 
1217 int coroutine_fn bdrv_co_readv(BdrvChild *child, int64_t sector_num,
1218                                int nb_sectors, QEMUIOVector *qiov)
1219 {
1220     trace_bdrv_co_readv(child->bs, sector_num, nb_sectors);
1221 
1222     return bdrv_co_do_readv(child, sector_num, nb_sectors, qiov, 0);
1223 }
1224 
1225 /* Maximum buffer for write zeroes fallback, in bytes */
1226 #define MAX_WRITE_ZEROES_BOUNCE_BUFFER (32768 << BDRV_SECTOR_BITS)
1227 
1228 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs,
1229     int64_t offset, int count, BdrvRequestFlags flags)
1230 {
1231     BlockDriver *drv = bs->drv;
1232     QEMUIOVector qiov;
1233     struct iovec iov = {0};
1234     int ret = 0;
1235     bool need_flush = false;
1236     int head = 0;
1237     int tail = 0;
1238 
1239     int max_write_zeroes = MIN_NON_ZERO(bs->bl.max_pwrite_zeroes, INT_MAX);
1240     int alignment = MAX(bs->bl.pwrite_zeroes_alignment,
1241                         bs->bl.request_alignment);
1242     int max_transfer = MIN_NON_ZERO(bs->bl.max_transfer,
1243                                     MAX_WRITE_ZEROES_BOUNCE_BUFFER);
1244 
1245     assert(alignment % bs->bl.request_alignment == 0);
1246     head = offset % alignment;
1247     tail = (offset + count) % alignment;
1248     max_write_zeroes = QEMU_ALIGN_DOWN(max_write_zeroes, alignment);
1249     assert(max_write_zeroes >= bs->bl.request_alignment);
1250 
1251     while (count > 0 && !ret) {
1252         int num = count;
1253 
1254         /* Align request.  Block drivers can expect the "bulk" of the request
1255          * to be aligned, and that unaligned requests do not cross cluster
1256          * boundaries.
1257          */
1258         if (head) {
1259             /* Make a small request up to the first aligned sector. For
1260              * convenience, limit this request to max_transfer even if
1261              * we don't need to fall back to writes.  */
1262             num = MIN(MIN(count, max_transfer), alignment - head);
1263             head = (head + num) % alignment;
1264             assert(num < max_write_zeroes);
1265         } else if (tail && num > alignment) {
1266             /* Shorten the request to the last aligned sector.  */
1267             num -= tail;
1268         }
1269 
1270         /* limit request size */
1271         if (num > max_write_zeroes) {
1272             num = max_write_zeroes;
1273         }
1274 
1275         ret = -ENOTSUP;
1276         /* First try the efficient write zeroes operation */
1277         if (drv->bdrv_co_pwrite_zeroes) {
1278             ret = drv->bdrv_co_pwrite_zeroes(bs, offset, num,
1279                                              flags & bs->supported_zero_flags);
1280             if (ret != -ENOTSUP && (flags & BDRV_REQ_FUA) &&
1281                 !(bs->supported_zero_flags & BDRV_REQ_FUA)) {
1282                 need_flush = true;
1283             }
1284         } else {
1285             assert(!bs->supported_zero_flags);
1286         }
1287 
1288         if (ret == -ENOTSUP) {
1289             /* Fall back to bounce buffer if write zeroes is unsupported */
1290             BdrvRequestFlags write_flags = flags & ~BDRV_REQ_ZERO_WRITE;
1291 
1292             if ((flags & BDRV_REQ_FUA) &&
1293                 !(bs->supported_write_flags & BDRV_REQ_FUA)) {
1294                 /* No need for bdrv_driver_pwrite() to do a fallback
1295                  * flush on each chunk; use just one at the end */
1296                 write_flags &= ~BDRV_REQ_FUA;
1297                 need_flush = true;
1298             }
1299             num = MIN(num, max_transfer);
1300             iov.iov_len = num;
1301             if (iov.iov_base == NULL) {
1302                 iov.iov_base = qemu_try_blockalign(bs, num);
1303                 if (iov.iov_base == NULL) {
1304                     ret = -ENOMEM;
1305                     goto fail;
1306                 }
1307                 memset(iov.iov_base, 0, num);
1308             }
1309             qemu_iovec_init_external(&qiov, &iov, 1);
1310 
1311             ret = bdrv_driver_pwritev(bs, offset, num, &qiov, write_flags);
1312 
1313             /* Keep bounce buffer around if it is big enough for all
1314              * all future requests.
1315              */
1316             if (num < max_transfer) {
1317                 qemu_vfree(iov.iov_base);
1318                 iov.iov_base = NULL;
1319             }
1320         }
1321 
1322         offset += num;
1323         count -= num;
1324     }
1325 
1326 fail:
1327     if (ret == 0 && need_flush) {
1328         ret = bdrv_co_flush(bs);
1329     }
1330     qemu_vfree(iov.iov_base);
1331     return ret;
1332 }
1333 
1334 /*
1335  * Forwards an already correctly aligned write request to the BlockDriver,
1336  * after possibly fragmenting it.
1337  */
1338 static int coroutine_fn bdrv_aligned_pwritev(BdrvChild *child,
1339     BdrvTrackedRequest *req, int64_t offset, unsigned int bytes,
1340     int64_t align, QEMUIOVector *qiov, int flags)
1341 {
1342     BlockDriverState *bs = child->bs;
1343     BlockDriver *drv = bs->drv;
1344     bool waited;
1345     int ret;
1346 
1347     int64_t start_sector = offset >> BDRV_SECTOR_BITS;
1348     int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);
1349     uint64_t bytes_remaining = bytes;
1350     int max_transfer;
1351 
1352     assert(is_power_of_2(align));
1353     assert((offset & (align - 1)) == 0);
1354     assert((bytes & (align - 1)) == 0);
1355     assert(!qiov || bytes == qiov->size);
1356     assert((bs->open_flags & BDRV_O_NO_IO) == 0);
1357     assert(!(flags & ~BDRV_REQ_MASK));
1358     max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX),
1359                                    align);
1360 
1361     waited = wait_serialising_requests(req);
1362     assert(!waited || !req->serialising);
1363     assert(req->overlap_offset <= offset);
1364     assert(offset + bytes <= req->overlap_offset + req->overlap_bytes);
1365     /* FIXME: Block migration uses the BlockBackend of the guest device at a
1366      *        point when it has not yet taken write permissions. This will be
1367      *        fixed by a future patch, but for now we have to bypass this
1368      *        assertion for block migration to work. */
1369     // assert(child->perm & BLK_PERM_WRITE);
1370     /* FIXME: Because of the above, we also cannot guarantee that all format
1371      *        BDS take the BLK_PERM_RESIZE permission on their file BDS, since
1372      *        they are not obligated to do so if they do not have any parent
1373      *        that has taken the permission to write to them. */
1374     // assert(end_sector <= bs->total_sectors || child->perm & BLK_PERM_RESIZE);
1375 
1376     ret = notifier_with_return_list_notify(&bs->before_write_notifiers, req);
1377 
1378     if (!ret && bs->detect_zeroes != BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF &&
1379         !(flags & BDRV_REQ_ZERO_WRITE) && drv->bdrv_co_pwrite_zeroes &&
1380         qemu_iovec_is_zero(qiov)) {
1381         flags |= BDRV_REQ_ZERO_WRITE;
1382         if (bs->detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP) {
1383             flags |= BDRV_REQ_MAY_UNMAP;
1384         }
1385     }
1386 
1387     if (ret < 0) {
1388         /* Do nothing, write notifier decided to fail this request */
1389     } else if (flags & BDRV_REQ_ZERO_WRITE) {
1390         bdrv_debug_event(bs, BLKDBG_PWRITEV_ZERO);
1391         ret = bdrv_co_do_pwrite_zeroes(bs, offset, bytes, flags);
1392     } else if (flags & BDRV_REQ_WRITE_COMPRESSED) {
1393         ret = bdrv_driver_pwritev_compressed(bs, offset, bytes, qiov);
1394     } else if (bytes <= max_transfer) {
1395         bdrv_debug_event(bs, BLKDBG_PWRITEV);
1396         ret = bdrv_driver_pwritev(bs, offset, bytes, qiov, flags);
1397     } else {
1398         bdrv_debug_event(bs, BLKDBG_PWRITEV);
1399         while (bytes_remaining) {
1400             int num = MIN(bytes_remaining, max_transfer);
1401             QEMUIOVector local_qiov;
1402             int local_flags = flags;
1403 
1404             assert(num);
1405             if (num < bytes_remaining && (flags & BDRV_REQ_FUA) &&
1406                 !(bs->supported_write_flags & BDRV_REQ_FUA)) {
1407                 /* If FUA is going to be emulated by flush, we only
1408                  * need to flush on the last iteration */
1409                 local_flags &= ~BDRV_REQ_FUA;
1410             }
1411             qemu_iovec_init(&local_qiov, qiov->niov);
1412             qemu_iovec_concat(&local_qiov, qiov, bytes - bytes_remaining, num);
1413 
1414             ret = bdrv_driver_pwritev(bs, offset + bytes - bytes_remaining,
1415                                       num, &local_qiov, local_flags);
1416             qemu_iovec_destroy(&local_qiov);
1417             if (ret < 0) {
1418                 break;
1419             }
1420             bytes_remaining -= num;
1421         }
1422     }
1423     bdrv_debug_event(bs, BLKDBG_PWRITEV_DONE);
1424 
1425     ++bs->write_gen;
1426     bdrv_set_dirty(bs, start_sector, end_sector - start_sector);
1427 
1428     if (bs->wr_highest_offset < offset + bytes) {
1429         bs->wr_highest_offset = offset + bytes;
1430     }
1431 
1432     if (ret >= 0) {
1433         bs->total_sectors = MAX(bs->total_sectors, end_sector);
1434         ret = 0;
1435     }
1436 
1437     return ret;
1438 }
1439 
1440 static int coroutine_fn bdrv_co_do_zero_pwritev(BdrvChild *child,
1441                                                 int64_t offset,
1442                                                 unsigned int bytes,
1443                                                 BdrvRequestFlags flags,
1444                                                 BdrvTrackedRequest *req)
1445 {
1446     BlockDriverState *bs = child->bs;
1447     uint8_t *buf = NULL;
1448     QEMUIOVector local_qiov;
1449     struct iovec iov;
1450     uint64_t align = bs->bl.request_alignment;
1451     unsigned int head_padding_bytes, tail_padding_bytes;
1452     int ret = 0;
1453 
1454     head_padding_bytes = offset & (align - 1);
1455     tail_padding_bytes = align - ((offset + bytes) & (align - 1));
1456 
1457 
1458     assert(flags & BDRV_REQ_ZERO_WRITE);
1459     if (head_padding_bytes || tail_padding_bytes) {
1460         buf = qemu_blockalign(bs, align);
1461         iov = (struct iovec) {
1462             .iov_base   = buf,
1463             .iov_len    = align,
1464         };
1465         qemu_iovec_init_external(&local_qiov, &iov, 1);
1466     }
1467     if (head_padding_bytes) {
1468         uint64_t zero_bytes = MIN(bytes, align - head_padding_bytes);
1469 
1470         /* RMW the unaligned part before head. */
1471         mark_request_serialising(req, align);
1472         wait_serialising_requests(req);
1473         bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD);
1474         ret = bdrv_aligned_preadv(child, req, offset & ~(align - 1), align,
1475                                   align, &local_qiov, 0);
1476         if (ret < 0) {
1477             goto fail;
1478         }
1479         bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD);
1480 
1481         memset(buf + head_padding_bytes, 0, zero_bytes);
1482         ret = bdrv_aligned_pwritev(child, req, offset & ~(align - 1), align,
1483                                    align, &local_qiov,
1484                                    flags & ~BDRV_REQ_ZERO_WRITE);
1485         if (ret < 0) {
1486             goto fail;
1487         }
1488         offset += zero_bytes;
1489         bytes -= zero_bytes;
1490     }
1491 
1492     assert(!bytes || (offset & (align - 1)) == 0);
1493     if (bytes >= align) {
1494         /* Write the aligned part in the middle. */
1495         uint64_t aligned_bytes = bytes & ~(align - 1);
1496         ret = bdrv_aligned_pwritev(child, req, offset, aligned_bytes, align,
1497                                    NULL, flags);
1498         if (ret < 0) {
1499             goto fail;
1500         }
1501         bytes -= aligned_bytes;
1502         offset += aligned_bytes;
1503     }
1504 
1505     assert(!bytes || (offset & (align - 1)) == 0);
1506     if (bytes) {
1507         assert(align == tail_padding_bytes + bytes);
1508         /* RMW the unaligned part after tail. */
1509         mark_request_serialising(req, align);
1510         wait_serialising_requests(req);
1511         bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
1512         ret = bdrv_aligned_preadv(child, req, offset, align,
1513                                   align, &local_qiov, 0);
1514         if (ret < 0) {
1515             goto fail;
1516         }
1517         bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
1518 
1519         memset(buf, 0, bytes);
1520         ret = bdrv_aligned_pwritev(child, req, offset, align, align,
1521                                    &local_qiov, flags & ~BDRV_REQ_ZERO_WRITE);
1522     }
1523 fail:
1524     qemu_vfree(buf);
1525     return ret;
1526 
1527 }
1528 
1529 /*
1530  * Handle a write request in coroutine context
1531  */
1532 int coroutine_fn bdrv_co_pwritev(BdrvChild *child,
1533     int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
1534     BdrvRequestFlags flags)
1535 {
1536     BlockDriverState *bs = child->bs;
1537     BdrvTrackedRequest req;
1538     uint64_t align = bs->bl.request_alignment;
1539     uint8_t *head_buf = NULL;
1540     uint8_t *tail_buf = NULL;
1541     QEMUIOVector local_qiov;
1542     bool use_local_qiov = false;
1543     int ret;
1544 
1545     if (!bs->drv) {
1546         return -ENOMEDIUM;
1547     }
1548     if (bs->read_only) {
1549         return -EPERM;
1550     }
1551     assert(!(bs->open_flags & BDRV_O_INACTIVE));
1552 
1553     ret = bdrv_check_byte_request(bs, offset, bytes);
1554     if (ret < 0) {
1555         return ret;
1556     }
1557 
1558     bdrv_inc_in_flight(bs);
1559     /*
1560      * Align write if necessary by performing a read-modify-write cycle.
1561      * Pad qiov with the read parts and be sure to have a tracked request not
1562      * only for bdrv_aligned_pwritev, but also for the reads of the RMW cycle.
1563      */
1564     tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_WRITE);
1565 
1566     if (!qiov) {
1567         ret = bdrv_co_do_zero_pwritev(child, offset, bytes, flags, &req);
1568         goto out;
1569     }
1570 
1571     if (offset & (align - 1)) {
1572         QEMUIOVector head_qiov;
1573         struct iovec head_iov;
1574 
1575         mark_request_serialising(&req, align);
1576         wait_serialising_requests(&req);
1577 
1578         head_buf = qemu_blockalign(bs, align);
1579         head_iov = (struct iovec) {
1580             .iov_base   = head_buf,
1581             .iov_len    = align,
1582         };
1583         qemu_iovec_init_external(&head_qiov, &head_iov, 1);
1584 
1585         bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD);
1586         ret = bdrv_aligned_preadv(child, &req, offset & ~(align - 1), align,
1587                                   align, &head_qiov, 0);
1588         if (ret < 0) {
1589             goto fail;
1590         }
1591         bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD);
1592 
1593         qemu_iovec_init(&local_qiov, qiov->niov + 2);
1594         qemu_iovec_add(&local_qiov, head_buf, offset & (align - 1));
1595         qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size);
1596         use_local_qiov = true;
1597 
1598         bytes += offset & (align - 1);
1599         offset = offset & ~(align - 1);
1600 
1601         /* We have read the tail already if the request is smaller
1602          * than one aligned block.
1603          */
1604         if (bytes < align) {
1605             qemu_iovec_add(&local_qiov, head_buf + bytes, align - bytes);
1606             bytes = align;
1607         }
1608     }
1609 
1610     if ((offset + bytes) & (align - 1)) {
1611         QEMUIOVector tail_qiov;
1612         struct iovec tail_iov;
1613         size_t tail_bytes;
1614         bool waited;
1615 
1616         mark_request_serialising(&req, align);
1617         waited = wait_serialising_requests(&req);
1618         assert(!waited || !use_local_qiov);
1619 
1620         tail_buf = qemu_blockalign(bs, align);
1621         tail_iov = (struct iovec) {
1622             .iov_base   = tail_buf,
1623             .iov_len    = align,
1624         };
1625         qemu_iovec_init_external(&tail_qiov, &tail_iov, 1);
1626 
1627         bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
1628         ret = bdrv_aligned_preadv(child, &req, (offset + bytes) & ~(align - 1),
1629                                   align, align, &tail_qiov, 0);
1630         if (ret < 0) {
1631             goto fail;
1632         }
1633         bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
1634 
1635         if (!use_local_qiov) {
1636             qemu_iovec_init(&local_qiov, qiov->niov + 1);
1637             qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size);
1638             use_local_qiov = true;
1639         }
1640 
1641         tail_bytes = (offset + bytes) & (align - 1);
1642         qemu_iovec_add(&local_qiov, tail_buf + tail_bytes, align - tail_bytes);
1643 
1644         bytes = ROUND_UP(bytes, align);
1645     }
1646 
1647     ret = bdrv_aligned_pwritev(child, &req, offset, bytes, align,
1648                                use_local_qiov ? &local_qiov : qiov,
1649                                flags);
1650 
1651 fail:
1652 
1653     if (use_local_qiov) {
1654         qemu_iovec_destroy(&local_qiov);
1655     }
1656     qemu_vfree(head_buf);
1657     qemu_vfree(tail_buf);
1658 out:
1659     tracked_request_end(&req);
1660     bdrv_dec_in_flight(bs);
1661     return ret;
1662 }
1663 
1664 static int coroutine_fn bdrv_co_do_writev(BdrvChild *child,
1665     int64_t sector_num, int nb_sectors, QEMUIOVector *qiov,
1666     BdrvRequestFlags flags)
1667 {
1668     if (nb_sectors < 0 || nb_sectors > BDRV_REQUEST_MAX_SECTORS) {
1669         return -EINVAL;
1670     }
1671 
1672     return bdrv_co_pwritev(child, sector_num << BDRV_SECTOR_BITS,
1673                            nb_sectors << BDRV_SECTOR_BITS, qiov, flags);
1674 }
1675 
1676 int coroutine_fn bdrv_co_writev(BdrvChild *child, int64_t sector_num,
1677     int nb_sectors, QEMUIOVector *qiov)
1678 {
1679     trace_bdrv_co_writev(child->bs, sector_num, nb_sectors);
1680 
1681     return bdrv_co_do_writev(child, sector_num, nb_sectors, qiov, 0);
1682 }
1683 
1684 int coroutine_fn bdrv_co_pwrite_zeroes(BdrvChild *child, int64_t offset,
1685                                        int count, BdrvRequestFlags flags)
1686 {
1687     trace_bdrv_co_pwrite_zeroes(child->bs, offset, count, flags);
1688 
1689     if (!(child->bs->open_flags & BDRV_O_UNMAP)) {
1690         flags &= ~BDRV_REQ_MAY_UNMAP;
1691     }
1692 
1693     return bdrv_co_pwritev(child, offset, count, NULL,
1694                            BDRV_REQ_ZERO_WRITE | flags);
1695 }
1696 
1697 /*
1698  * Flush ALL BDSes regardless of if they are reachable via a BlkBackend or not.
1699  */
1700 int bdrv_flush_all(void)
1701 {
1702     BdrvNextIterator it;
1703     BlockDriverState *bs = NULL;
1704     int result = 0;
1705 
1706     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
1707         AioContext *aio_context = bdrv_get_aio_context(bs);
1708         int ret;
1709 
1710         aio_context_acquire(aio_context);
1711         ret = bdrv_flush(bs);
1712         if (ret < 0 && !result) {
1713             result = ret;
1714         }
1715         aio_context_release(aio_context);
1716     }
1717 
1718     return result;
1719 }
1720 
1721 
1722 typedef struct BdrvCoGetBlockStatusData {
1723     BlockDriverState *bs;
1724     BlockDriverState *base;
1725     BlockDriverState **file;
1726     int64_t sector_num;
1727     int nb_sectors;
1728     int *pnum;
1729     int64_t ret;
1730     bool done;
1731 } BdrvCoGetBlockStatusData;
1732 
1733 /*
1734  * Returns the allocation status of the specified sectors.
1735  * Drivers not implementing the functionality are assumed to not support
1736  * backing files, hence all their sectors are reported as allocated.
1737  *
1738  * If 'sector_num' is beyond the end of the disk image the return value is 0
1739  * and 'pnum' is set to 0.
1740  *
1741  * 'pnum' is set to the number of sectors (including and immediately following
1742  * the specified sector) that are known to be in the same
1743  * allocated/unallocated state.
1744  *
1745  * 'nb_sectors' is the max value 'pnum' should be set to.  If nb_sectors goes
1746  * beyond the end of the disk image it will be clamped.
1747  *
1748  * If returned value is positive and BDRV_BLOCK_OFFSET_VALID bit is set, 'file'
1749  * points to the BDS which the sector range is allocated in.
1750  */
1751 static int64_t coroutine_fn bdrv_co_get_block_status(BlockDriverState *bs,
1752                                                      int64_t sector_num,
1753                                                      int nb_sectors, int *pnum,
1754                                                      BlockDriverState **file)
1755 {
1756     int64_t total_sectors;
1757     int64_t n;
1758     int64_t ret, ret2;
1759 
1760     total_sectors = bdrv_nb_sectors(bs);
1761     if (total_sectors < 0) {
1762         return total_sectors;
1763     }
1764 
1765     if (sector_num >= total_sectors) {
1766         *pnum = 0;
1767         return 0;
1768     }
1769 
1770     n = total_sectors - sector_num;
1771     if (n < nb_sectors) {
1772         nb_sectors = n;
1773     }
1774 
1775     if (!bs->drv->bdrv_co_get_block_status) {
1776         *pnum = nb_sectors;
1777         ret = BDRV_BLOCK_DATA | BDRV_BLOCK_ALLOCATED;
1778         if (bs->drv->protocol_name) {
1779             ret |= BDRV_BLOCK_OFFSET_VALID | (sector_num * BDRV_SECTOR_SIZE);
1780         }
1781         return ret;
1782     }
1783 
1784     *file = NULL;
1785     bdrv_inc_in_flight(bs);
1786     ret = bs->drv->bdrv_co_get_block_status(bs, sector_num, nb_sectors, pnum,
1787                                             file);
1788     if (ret < 0) {
1789         *pnum = 0;
1790         goto out;
1791     }
1792 
1793     if (ret & BDRV_BLOCK_RAW) {
1794         assert(ret & BDRV_BLOCK_OFFSET_VALID);
1795         ret = bdrv_get_block_status(*file, ret >> BDRV_SECTOR_BITS,
1796                                     *pnum, pnum, file);
1797         goto out;
1798     }
1799 
1800     if (ret & (BDRV_BLOCK_DATA | BDRV_BLOCK_ZERO)) {
1801         ret |= BDRV_BLOCK_ALLOCATED;
1802     } else {
1803         if (bdrv_unallocated_blocks_are_zero(bs)) {
1804             ret |= BDRV_BLOCK_ZERO;
1805         } else if (bs->backing) {
1806             BlockDriverState *bs2 = bs->backing->bs;
1807             int64_t nb_sectors2 = bdrv_nb_sectors(bs2);
1808             if (nb_sectors2 >= 0 && sector_num >= nb_sectors2) {
1809                 ret |= BDRV_BLOCK_ZERO;
1810             }
1811         }
1812     }
1813 
1814     if (*file && *file != bs &&
1815         (ret & BDRV_BLOCK_DATA) && !(ret & BDRV_BLOCK_ZERO) &&
1816         (ret & BDRV_BLOCK_OFFSET_VALID)) {
1817         BlockDriverState *file2;
1818         int file_pnum;
1819 
1820         ret2 = bdrv_co_get_block_status(*file, ret >> BDRV_SECTOR_BITS,
1821                                         *pnum, &file_pnum, &file2);
1822         if (ret2 >= 0) {
1823             /* Ignore errors.  This is just providing extra information, it
1824              * is useful but not necessary.
1825              */
1826             if (!file_pnum) {
1827                 /* !file_pnum indicates an offset at or beyond the EOF; it is
1828                  * perfectly valid for the format block driver to point to such
1829                  * offsets, so catch it and mark everything as zero */
1830                 ret |= BDRV_BLOCK_ZERO;
1831             } else {
1832                 /* Limit request to the range reported by the protocol driver */
1833                 *pnum = file_pnum;
1834                 ret |= (ret2 & BDRV_BLOCK_ZERO);
1835             }
1836         }
1837     }
1838 
1839 out:
1840     bdrv_dec_in_flight(bs);
1841     return ret;
1842 }
1843 
1844 static int64_t coroutine_fn bdrv_co_get_block_status_above(BlockDriverState *bs,
1845         BlockDriverState *base,
1846         int64_t sector_num,
1847         int nb_sectors,
1848         int *pnum,
1849         BlockDriverState **file)
1850 {
1851     BlockDriverState *p;
1852     int64_t ret = 0;
1853 
1854     assert(bs != base);
1855     for (p = bs; p != base; p = backing_bs(p)) {
1856         ret = bdrv_co_get_block_status(p, sector_num, nb_sectors, pnum, file);
1857         if (ret < 0 || ret & BDRV_BLOCK_ALLOCATED) {
1858             break;
1859         }
1860         /* [sector_num, pnum] unallocated on this layer, which could be only
1861          * the first part of [sector_num, nb_sectors].  */
1862         nb_sectors = MIN(nb_sectors, *pnum);
1863     }
1864     return ret;
1865 }
1866 
1867 /* Coroutine wrapper for bdrv_get_block_status_above() */
1868 static void coroutine_fn bdrv_get_block_status_above_co_entry(void *opaque)
1869 {
1870     BdrvCoGetBlockStatusData *data = opaque;
1871 
1872     data->ret = bdrv_co_get_block_status_above(data->bs, data->base,
1873                                                data->sector_num,
1874                                                data->nb_sectors,
1875                                                data->pnum,
1876                                                data->file);
1877     data->done = true;
1878 }
1879 
1880 /*
1881  * Synchronous wrapper around bdrv_co_get_block_status_above().
1882  *
1883  * See bdrv_co_get_block_status_above() for details.
1884  */
1885 int64_t bdrv_get_block_status_above(BlockDriverState *bs,
1886                                     BlockDriverState *base,
1887                                     int64_t sector_num,
1888                                     int nb_sectors, int *pnum,
1889                                     BlockDriverState **file)
1890 {
1891     Coroutine *co;
1892     BdrvCoGetBlockStatusData data = {
1893         .bs = bs,
1894         .base = base,
1895         .file = file,
1896         .sector_num = sector_num,
1897         .nb_sectors = nb_sectors,
1898         .pnum = pnum,
1899         .done = false,
1900     };
1901 
1902     if (qemu_in_coroutine()) {
1903         /* Fast-path if already in coroutine context */
1904         bdrv_get_block_status_above_co_entry(&data);
1905     } else {
1906         co = qemu_coroutine_create(bdrv_get_block_status_above_co_entry,
1907                                    &data);
1908         bdrv_coroutine_enter(bs, co);
1909         BDRV_POLL_WHILE(bs, !data.done);
1910     }
1911     return data.ret;
1912 }
1913 
1914 int64_t bdrv_get_block_status(BlockDriverState *bs,
1915                               int64_t sector_num,
1916                               int nb_sectors, int *pnum,
1917                               BlockDriverState **file)
1918 {
1919     return bdrv_get_block_status_above(bs, backing_bs(bs),
1920                                        sector_num, nb_sectors, pnum, file);
1921 }
1922 
1923 int coroutine_fn bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num,
1924                                    int nb_sectors, int *pnum)
1925 {
1926     BlockDriverState *file;
1927     int64_t ret = bdrv_get_block_status(bs, sector_num, nb_sectors, pnum,
1928                                         &file);
1929     if (ret < 0) {
1930         return ret;
1931     }
1932     return !!(ret & BDRV_BLOCK_ALLOCATED);
1933 }
1934 
1935 /*
1936  * Given an image chain: ... -> [BASE] -> [INTER1] -> [INTER2] -> [TOP]
1937  *
1938  * Return true if the given sector is allocated in any image between
1939  * BASE and TOP (inclusive).  BASE can be NULL to check if the given
1940  * sector is allocated in any image of the chain.  Return false otherwise.
1941  *
1942  * 'pnum' is set to the number of sectors (including and immediately following
1943  *  the specified sector) that are known to be in the same
1944  *  allocated/unallocated state.
1945  *
1946  */
1947 int bdrv_is_allocated_above(BlockDriverState *top,
1948                             BlockDriverState *base,
1949                             int64_t sector_num,
1950                             int nb_sectors, int *pnum)
1951 {
1952     BlockDriverState *intermediate;
1953     int ret, n = nb_sectors;
1954 
1955     intermediate = top;
1956     while (intermediate && intermediate != base) {
1957         int pnum_inter;
1958         ret = bdrv_is_allocated(intermediate, sector_num, nb_sectors,
1959                                 &pnum_inter);
1960         if (ret < 0) {
1961             return ret;
1962         } else if (ret) {
1963             *pnum = pnum_inter;
1964             return 1;
1965         }
1966 
1967         /*
1968          * [sector_num, nb_sectors] is unallocated on top but intermediate
1969          * might have
1970          *
1971          * [sector_num+x, nr_sectors] allocated.
1972          */
1973         if (n > pnum_inter &&
1974             (intermediate == top ||
1975              sector_num + pnum_inter < intermediate->total_sectors)) {
1976             n = pnum_inter;
1977         }
1978 
1979         intermediate = backing_bs(intermediate);
1980     }
1981 
1982     *pnum = n;
1983     return 0;
1984 }
1985 
1986 typedef struct BdrvVmstateCo {
1987     BlockDriverState    *bs;
1988     QEMUIOVector        *qiov;
1989     int64_t             pos;
1990     bool                is_read;
1991     int                 ret;
1992 } BdrvVmstateCo;
1993 
1994 static int coroutine_fn
1995 bdrv_co_rw_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos,
1996                    bool is_read)
1997 {
1998     BlockDriver *drv = bs->drv;
1999 
2000     if (!drv) {
2001         return -ENOMEDIUM;
2002     } else if (drv->bdrv_load_vmstate) {
2003         return is_read ? drv->bdrv_load_vmstate(bs, qiov, pos)
2004                        : drv->bdrv_save_vmstate(bs, qiov, pos);
2005     } else if (bs->file) {
2006         return bdrv_co_rw_vmstate(bs->file->bs, qiov, pos, is_read);
2007     }
2008 
2009     return -ENOTSUP;
2010 }
2011 
2012 static void coroutine_fn bdrv_co_rw_vmstate_entry(void *opaque)
2013 {
2014     BdrvVmstateCo *co = opaque;
2015     co->ret = bdrv_co_rw_vmstate(co->bs, co->qiov, co->pos, co->is_read);
2016 }
2017 
2018 static inline int
2019 bdrv_rw_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos,
2020                 bool is_read)
2021 {
2022     if (qemu_in_coroutine()) {
2023         return bdrv_co_rw_vmstate(bs, qiov, pos, is_read);
2024     } else {
2025         BdrvVmstateCo data = {
2026             .bs         = bs,
2027             .qiov       = qiov,
2028             .pos        = pos,
2029             .is_read    = is_read,
2030             .ret        = -EINPROGRESS,
2031         };
2032         Coroutine *co = qemu_coroutine_create(bdrv_co_rw_vmstate_entry, &data);
2033 
2034         bdrv_coroutine_enter(bs, co);
2035         while (data.ret == -EINPROGRESS) {
2036             aio_poll(bdrv_get_aio_context(bs), true);
2037         }
2038         return data.ret;
2039     }
2040 }
2041 
2042 int bdrv_save_vmstate(BlockDriverState *bs, const uint8_t *buf,
2043                       int64_t pos, int size)
2044 {
2045     QEMUIOVector qiov;
2046     struct iovec iov = {
2047         .iov_base   = (void *) buf,
2048         .iov_len    = size,
2049     };
2050     int ret;
2051 
2052     qemu_iovec_init_external(&qiov, &iov, 1);
2053 
2054     ret = bdrv_writev_vmstate(bs, &qiov, pos);
2055     if (ret < 0) {
2056         return ret;
2057     }
2058 
2059     return size;
2060 }
2061 
2062 int bdrv_writev_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos)
2063 {
2064     return bdrv_rw_vmstate(bs, qiov, pos, false);
2065 }
2066 
2067 int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf,
2068                       int64_t pos, int size)
2069 {
2070     QEMUIOVector qiov;
2071     struct iovec iov = {
2072         .iov_base   = buf,
2073         .iov_len    = size,
2074     };
2075     int ret;
2076 
2077     qemu_iovec_init_external(&qiov, &iov, 1);
2078     ret = bdrv_readv_vmstate(bs, &qiov, pos);
2079     if (ret < 0) {
2080         return ret;
2081     }
2082 
2083     return size;
2084 }
2085 
2086 int bdrv_readv_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos)
2087 {
2088     return bdrv_rw_vmstate(bs, qiov, pos, true);
2089 }
2090 
2091 /**************************************************************/
2092 /* async I/Os */
2093 
2094 BlockAIOCB *bdrv_aio_readv(BdrvChild *child, int64_t sector_num,
2095                            QEMUIOVector *qiov, int nb_sectors,
2096                            BlockCompletionFunc *cb, void *opaque)
2097 {
2098     trace_bdrv_aio_readv(child->bs, sector_num, nb_sectors, opaque);
2099 
2100     assert(nb_sectors << BDRV_SECTOR_BITS == qiov->size);
2101     return bdrv_co_aio_prw_vector(child, sector_num << BDRV_SECTOR_BITS, qiov,
2102                                   0, cb, opaque, false);
2103 }
2104 
2105 BlockAIOCB *bdrv_aio_writev(BdrvChild *child, int64_t sector_num,
2106                             QEMUIOVector *qiov, int nb_sectors,
2107                             BlockCompletionFunc *cb, void *opaque)
2108 {
2109     trace_bdrv_aio_writev(child->bs, sector_num, nb_sectors, opaque);
2110 
2111     assert(nb_sectors << BDRV_SECTOR_BITS == qiov->size);
2112     return bdrv_co_aio_prw_vector(child, sector_num << BDRV_SECTOR_BITS, qiov,
2113                                   0, cb, opaque, true);
2114 }
2115 
2116 void bdrv_aio_cancel(BlockAIOCB *acb)
2117 {
2118     qemu_aio_ref(acb);
2119     bdrv_aio_cancel_async(acb);
2120     while (acb->refcnt > 1) {
2121         if (acb->aiocb_info->get_aio_context) {
2122             aio_poll(acb->aiocb_info->get_aio_context(acb), true);
2123         } else if (acb->bs) {
2124             /* qemu_aio_ref and qemu_aio_unref are not thread-safe, so
2125              * assert that we're not using an I/O thread.  Thread-safe
2126              * code should use bdrv_aio_cancel_async exclusively.
2127              */
2128             assert(bdrv_get_aio_context(acb->bs) == qemu_get_aio_context());
2129             aio_poll(bdrv_get_aio_context(acb->bs), true);
2130         } else {
2131             abort();
2132         }
2133     }
2134     qemu_aio_unref(acb);
2135 }
2136 
2137 /* Async version of aio cancel. The caller is not blocked if the acb implements
2138  * cancel_async, otherwise we do nothing and let the request normally complete.
2139  * In either case the completion callback must be called. */
2140 void bdrv_aio_cancel_async(BlockAIOCB *acb)
2141 {
2142     if (acb->aiocb_info->cancel_async) {
2143         acb->aiocb_info->cancel_async(acb);
2144     }
2145 }
2146 
2147 /**************************************************************/
2148 /* async block device emulation */
2149 
2150 typedef struct BlockRequest {
2151     union {
2152         /* Used during read, write, trim */
2153         struct {
2154             int64_t offset;
2155             int bytes;
2156             int flags;
2157             QEMUIOVector *qiov;
2158         };
2159         /* Used during ioctl */
2160         struct {
2161             int req;
2162             void *buf;
2163         };
2164     };
2165     BlockCompletionFunc *cb;
2166     void *opaque;
2167 
2168     int error;
2169 } BlockRequest;
2170 
2171 typedef struct BlockAIOCBCoroutine {
2172     BlockAIOCB common;
2173     BdrvChild *child;
2174     BlockRequest req;
2175     bool is_write;
2176     bool need_bh;
2177     bool *done;
2178 } BlockAIOCBCoroutine;
2179 
2180 static const AIOCBInfo bdrv_em_co_aiocb_info = {
2181     .aiocb_size         = sizeof(BlockAIOCBCoroutine),
2182 };
2183 
2184 static void bdrv_co_complete(BlockAIOCBCoroutine *acb)
2185 {
2186     if (!acb->need_bh) {
2187         bdrv_dec_in_flight(acb->common.bs);
2188         acb->common.cb(acb->common.opaque, acb->req.error);
2189         qemu_aio_unref(acb);
2190     }
2191 }
2192 
2193 static void bdrv_co_em_bh(void *opaque)
2194 {
2195     BlockAIOCBCoroutine *acb = opaque;
2196 
2197     assert(!acb->need_bh);
2198     bdrv_co_complete(acb);
2199 }
2200 
2201 static void bdrv_co_maybe_schedule_bh(BlockAIOCBCoroutine *acb)
2202 {
2203     acb->need_bh = false;
2204     if (acb->req.error != -EINPROGRESS) {
2205         BlockDriverState *bs = acb->common.bs;
2206 
2207         aio_bh_schedule_oneshot(bdrv_get_aio_context(bs), bdrv_co_em_bh, acb);
2208     }
2209 }
2210 
2211 /* Invoke bdrv_co_do_readv/bdrv_co_do_writev */
2212 static void coroutine_fn bdrv_co_do_rw(void *opaque)
2213 {
2214     BlockAIOCBCoroutine *acb = opaque;
2215 
2216     if (!acb->is_write) {
2217         acb->req.error = bdrv_co_preadv(acb->child, acb->req.offset,
2218             acb->req.qiov->size, acb->req.qiov, acb->req.flags);
2219     } else {
2220         acb->req.error = bdrv_co_pwritev(acb->child, acb->req.offset,
2221             acb->req.qiov->size, acb->req.qiov, acb->req.flags);
2222     }
2223 
2224     bdrv_co_complete(acb);
2225 }
2226 
2227 static BlockAIOCB *bdrv_co_aio_prw_vector(BdrvChild *child,
2228                                           int64_t offset,
2229                                           QEMUIOVector *qiov,
2230                                           BdrvRequestFlags flags,
2231                                           BlockCompletionFunc *cb,
2232                                           void *opaque,
2233                                           bool is_write)
2234 {
2235     Coroutine *co;
2236     BlockAIOCBCoroutine *acb;
2237 
2238     /* Matched by bdrv_co_complete's bdrv_dec_in_flight.  */
2239     bdrv_inc_in_flight(child->bs);
2240 
2241     acb = qemu_aio_get(&bdrv_em_co_aiocb_info, child->bs, cb, opaque);
2242     acb->child = child;
2243     acb->need_bh = true;
2244     acb->req.error = -EINPROGRESS;
2245     acb->req.offset = offset;
2246     acb->req.qiov = qiov;
2247     acb->req.flags = flags;
2248     acb->is_write = is_write;
2249 
2250     co = qemu_coroutine_create(bdrv_co_do_rw, acb);
2251     bdrv_coroutine_enter(child->bs, co);
2252 
2253     bdrv_co_maybe_schedule_bh(acb);
2254     return &acb->common;
2255 }
2256 
2257 static void coroutine_fn bdrv_aio_flush_co_entry(void *opaque)
2258 {
2259     BlockAIOCBCoroutine *acb = opaque;
2260     BlockDriverState *bs = acb->common.bs;
2261 
2262     acb->req.error = bdrv_co_flush(bs);
2263     bdrv_co_complete(acb);
2264 }
2265 
2266 BlockAIOCB *bdrv_aio_flush(BlockDriverState *bs,
2267         BlockCompletionFunc *cb, void *opaque)
2268 {
2269     trace_bdrv_aio_flush(bs, opaque);
2270 
2271     Coroutine *co;
2272     BlockAIOCBCoroutine *acb;
2273 
2274     /* Matched by bdrv_co_complete's bdrv_dec_in_flight.  */
2275     bdrv_inc_in_flight(bs);
2276 
2277     acb = qemu_aio_get(&bdrv_em_co_aiocb_info, bs, cb, opaque);
2278     acb->need_bh = true;
2279     acb->req.error = -EINPROGRESS;
2280 
2281     co = qemu_coroutine_create(bdrv_aio_flush_co_entry, acb);
2282     bdrv_coroutine_enter(bs, co);
2283 
2284     bdrv_co_maybe_schedule_bh(acb);
2285     return &acb->common;
2286 }
2287 
2288 /**************************************************************/
2289 /* Coroutine block device emulation */
2290 
2291 typedef struct FlushCo {
2292     BlockDriverState *bs;
2293     int ret;
2294 } FlushCo;
2295 
2296 
2297 static void coroutine_fn bdrv_flush_co_entry(void *opaque)
2298 {
2299     FlushCo *rwco = opaque;
2300 
2301     rwco->ret = bdrv_co_flush(rwco->bs);
2302 }
2303 
2304 int coroutine_fn bdrv_co_flush(BlockDriverState *bs)
2305 {
2306     int current_gen;
2307     int ret = 0;
2308 
2309     bdrv_inc_in_flight(bs);
2310 
2311     if (!bs || !bdrv_is_inserted(bs) || bdrv_is_read_only(bs) ||
2312         bdrv_is_sg(bs)) {
2313         goto early_exit;
2314     }
2315 
2316     current_gen = bs->write_gen;
2317 
2318     /* Wait until any previous flushes are completed */
2319     while (bs->active_flush_req) {
2320         qemu_co_queue_wait(&bs->flush_queue, NULL);
2321     }
2322 
2323     bs->active_flush_req = true;
2324 
2325     /* Write back all layers by calling one driver function */
2326     if (bs->drv->bdrv_co_flush) {
2327         ret = bs->drv->bdrv_co_flush(bs);
2328         goto out;
2329     }
2330 
2331     /* Write back cached data to the OS even with cache=unsafe */
2332     BLKDBG_EVENT(bs->file, BLKDBG_FLUSH_TO_OS);
2333     if (bs->drv->bdrv_co_flush_to_os) {
2334         ret = bs->drv->bdrv_co_flush_to_os(bs);
2335         if (ret < 0) {
2336             goto out;
2337         }
2338     }
2339 
2340     /* But don't actually force it to the disk with cache=unsafe */
2341     if (bs->open_flags & BDRV_O_NO_FLUSH) {
2342         goto flush_parent;
2343     }
2344 
2345     /* Check if we really need to flush anything */
2346     if (bs->flushed_gen == current_gen) {
2347         goto flush_parent;
2348     }
2349 
2350     BLKDBG_EVENT(bs->file, BLKDBG_FLUSH_TO_DISK);
2351     if (bs->drv->bdrv_co_flush_to_disk) {
2352         ret = bs->drv->bdrv_co_flush_to_disk(bs);
2353     } else if (bs->drv->bdrv_aio_flush) {
2354         BlockAIOCB *acb;
2355         CoroutineIOCompletion co = {
2356             .coroutine = qemu_coroutine_self(),
2357         };
2358 
2359         acb = bs->drv->bdrv_aio_flush(bs, bdrv_co_io_em_complete, &co);
2360         if (acb == NULL) {
2361             ret = -EIO;
2362         } else {
2363             qemu_coroutine_yield();
2364             ret = co.ret;
2365         }
2366     } else {
2367         /*
2368          * Some block drivers always operate in either writethrough or unsafe
2369          * mode and don't support bdrv_flush therefore. Usually qemu doesn't
2370          * know how the server works (because the behaviour is hardcoded or
2371          * depends on server-side configuration), so we can't ensure that
2372          * everything is safe on disk. Returning an error doesn't work because
2373          * that would break guests even if the server operates in writethrough
2374          * mode.
2375          *
2376          * Let's hope the user knows what he's doing.
2377          */
2378         ret = 0;
2379     }
2380 
2381     if (ret < 0) {
2382         goto out;
2383     }
2384 
2385     /* Now flush the underlying protocol.  It will also have BDRV_O_NO_FLUSH
2386      * in the case of cache=unsafe, so there are no useless flushes.
2387      */
2388 flush_parent:
2389     ret = bs->file ? bdrv_co_flush(bs->file->bs) : 0;
2390 out:
2391     /* Notify any pending flushes that we have completed */
2392     if (ret == 0) {
2393         bs->flushed_gen = current_gen;
2394     }
2395     bs->active_flush_req = false;
2396     /* Return value is ignored - it's ok if wait queue is empty */
2397     qemu_co_queue_next(&bs->flush_queue);
2398 
2399 early_exit:
2400     bdrv_dec_in_flight(bs);
2401     return ret;
2402 }
2403 
2404 int bdrv_flush(BlockDriverState *bs)
2405 {
2406     Coroutine *co;
2407     FlushCo flush_co = {
2408         .bs = bs,
2409         .ret = NOT_DONE,
2410     };
2411 
2412     if (qemu_in_coroutine()) {
2413         /* Fast-path if already in coroutine context */
2414         bdrv_flush_co_entry(&flush_co);
2415     } else {
2416         co = qemu_coroutine_create(bdrv_flush_co_entry, &flush_co);
2417         bdrv_coroutine_enter(bs, co);
2418         BDRV_POLL_WHILE(bs, flush_co.ret == NOT_DONE);
2419     }
2420 
2421     return flush_co.ret;
2422 }
2423 
2424 typedef struct DiscardCo {
2425     BlockDriverState *bs;
2426     int64_t offset;
2427     int count;
2428     int ret;
2429 } DiscardCo;
2430 static void coroutine_fn bdrv_pdiscard_co_entry(void *opaque)
2431 {
2432     DiscardCo *rwco = opaque;
2433 
2434     rwco->ret = bdrv_co_pdiscard(rwco->bs, rwco->offset, rwco->count);
2435 }
2436 
2437 int coroutine_fn bdrv_co_pdiscard(BlockDriverState *bs, int64_t offset,
2438                                   int count)
2439 {
2440     BdrvTrackedRequest req;
2441     int max_pdiscard, ret;
2442     int head, tail, align;
2443 
2444     if (!bs->drv) {
2445         return -ENOMEDIUM;
2446     }
2447 
2448     ret = bdrv_check_byte_request(bs, offset, count);
2449     if (ret < 0) {
2450         return ret;
2451     } else if (bs->read_only) {
2452         return -EPERM;
2453     }
2454     assert(!(bs->open_flags & BDRV_O_INACTIVE));
2455 
2456     /* Do nothing if disabled.  */
2457     if (!(bs->open_flags & BDRV_O_UNMAP)) {
2458         return 0;
2459     }
2460 
2461     if (!bs->drv->bdrv_co_pdiscard && !bs->drv->bdrv_aio_pdiscard) {
2462         return 0;
2463     }
2464 
2465     /* Discard is advisory, but some devices track and coalesce
2466      * unaligned requests, so we must pass everything down rather than
2467      * round here.  Still, most devices will just silently ignore
2468      * unaligned requests (by returning -ENOTSUP), so we must fragment
2469      * the request accordingly.  */
2470     align = MAX(bs->bl.pdiscard_alignment, bs->bl.request_alignment);
2471     assert(align % bs->bl.request_alignment == 0);
2472     head = offset % align;
2473     tail = (offset + count) % align;
2474 
2475     bdrv_inc_in_flight(bs);
2476     tracked_request_begin(&req, bs, offset, count, BDRV_TRACKED_DISCARD);
2477 
2478     ret = notifier_with_return_list_notify(&bs->before_write_notifiers, &req);
2479     if (ret < 0) {
2480         goto out;
2481     }
2482 
2483     max_pdiscard = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_pdiscard, INT_MAX),
2484                                    align);
2485     assert(max_pdiscard >= bs->bl.request_alignment);
2486 
2487     while (count > 0) {
2488         int ret;
2489         int num = count;
2490 
2491         if (head) {
2492             /* Make small requests to get to alignment boundaries. */
2493             num = MIN(count, align - head);
2494             if (!QEMU_IS_ALIGNED(num, bs->bl.request_alignment)) {
2495                 num %= bs->bl.request_alignment;
2496             }
2497             head = (head + num) % align;
2498             assert(num < max_pdiscard);
2499         } else if (tail) {
2500             if (num > align) {
2501                 /* Shorten the request to the last aligned cluster.  */
2502                 num -= tail;
2503             } else if (!QEMU_IS_ALIGNED(tail, bs->bl.request_alignment) &&
2504                        tail > bs->bl.request_alignment) {
2505                 tail %= bs->bl.request_alignment;
2506                 num -= tail;
2507             }
2508         }
2509         /* limit request size */
2510         if (num > max_pdiscard) {
2511             num = max_pdiscard;
2512         }
2513 
2514         if (bs->drv->bdrv_co_pdiscard) {
2515             ret = bs->drv->bdrv_co_pdiscard(bs, offset, num);
2516         } else {
2517             BlockAIOCB *acb;
2518             CoroutineIOCompletion co = {
2519                 .coroutine = qemu_coroutine_self(),
2520             };
2521 
2522             acb = bs->drv->bdrv_aio_pdiscard(bs, offset, num,
2523                                              bdrv_co_io_em_complete, &co);
2524             if (acb == NULL) {
2525                 ret = -EIO;
2526                 goto out;
2527             } else {
2528                 qemu_coroutine_yield();
2529                 ret = co.ret;
2530             }
2531         }
2532         if (ret && ret != -ENOTSUP) {
2533             goto out;
2534         }
2535 
2536         offset += num;
2537         count -= num;
2538     }
2539     ret = 0;
2540 out:
2541     ++bs->write_gen;
2542     bdrv_set_dirty(bs, req.offset >> BDRV_SECTOR_BITS,
2543                    req.bytes >> BDRV_SECTOR_BITS);
2544     tracked_request_end(&req);
2545     bdrv_dec_in_flight(bs);
2546     return ret;
2547 }
2548 
2549 int bdrv_pdiscard(BlockDriverState *bs, int64_t offset, int count)
2550 {
2551     Coroutine *co;
2552     DiscardCo rwco = {
2553         .bs = bs,
2554         .offset = offset,
2555         .count = count,
2556         .ret = NOT_DONE,
2557     };
2558 
2559     if (qemu_in_coroutine()) {
2560         /* Fast-path if already in coroutine context */
2561         bdrv_pdiscard_co_entry(&rwco);
2562     } else {
2563         co = qemu_coroutine_create(bdrv_pdiscard_co_entry, &rwco);
2564         bdrv_coroutine_enter(bs, co);
2565         BDRV_POLL_WHILE(bs, rwco.ret == NOT_DONE);
2566     }
2567 
2568     return rwco.ret;
2569 }
2570 
2571 int bdrv_co_ioctl(BlockDriverState *bs, int req, void *buf)
2572 {
2573     BlockDriver *drv = bs->drv;
2574     CoroutineIOCompletion co = {
2575         .coroutine = qemu_coroutine_self(),
2576     };
2577     BlockAIOCB *acb;
2578 
2579     bdrv_inc_in_flight(bs);
2580     if (!drv || (!drv->bdrv_aio_ioctl && !drv->bdrv_co_ioctl)) {
2581         co.ret = -ENOTSUP;
2582         goto out;
2583     }
2584 
2585     if (drv->bdrv_co_ioctl) {
2586         co.ret = drv->bdrv_co_ioctl(bs, req, buf);
2587     } else {
2588         acb = drv->bdrv_aio_ioctl(bs, req, buf, bdrv_co_io_em_complete, &co);
2589         if (!acb) {
2590             co.ret = -ENOTSUP;
2591             goto out;
2592         }
2593         qemu_coroutine_yield();
2594     }
2595 out:
2596     bdrv_dec_in_flight(bs);
2597     return co.ret;
2598 }
2599 
2600 void *qemu_blockalign(BlockDriverState *bs, size_t size)
2601 {
2602     return qemu_memalign(bdrv_opt_mem_align(bs), size);
2603 }
2604 
2605 void *qemu_blockalign0(BlockDriverState *bs, size_t size)
2606 {
2607     return memset(qemu_blockalign(bs, size), 0, size);
2608 }
2609 
2610 void *qemu_try_blockalign(BlockDriverState *bs, size_t size)
2611 {
2612     size_t align = bdrv_opt_mem_align(bs);
2613 
2614     /* Ensure that NULL is never returned on success */
2615     assert(align > 0);
2616     if (size == 0) {
2617         size = align;
2618     }
2619 
2620     return qemu_try_memalign(align, size);
2621 }
2622 
2623 void *qemu_try_blockalign0(BlockDriverState *bs, size_t size)
2624 {
2625     void *mem = qemu_try_blockalign(bs, size);
2626 
2627     if (mem) {
2628         memset(mem, 0, size);
2629     }
2630 
2631     return mem;
2632 }
2633 
2634 /*
2635  * Check if all memory in this vector is sector aligned.
2636  */
2637 bool bdrv_qiov_is_aligned(BlockDriverState *bs, QEMUIOVector *qiov)
2638 {
2639     int i;
2640     size_t alignment = bdrv_min_mem_align(bs);
2641 
2642     for (i = 0; i < qiov->niov; i++) {
2643         if ((uintptr_t) qiov->iov[i].iov_base % alignment) {
2644             return false;
2645         }
2646         if (qiov->iov[i].iov_len % alignment) {
2647             return false;
2648         }
2649     }
2650 
2651     return true;
2652 }
2653 
2654 void bdrv_add_before_write_notifier(BlockDriverState *bs,
2655                                     NotifierWithReturn *notifier)
2656 {
2657     notifier_with_return_list_add(&bs->before_write_notifiers, notifier);
2658 }
2659 
2660 void bdrv_io_plug(BlockDriverState *bs)
2661 {
2662     BdrvChild *child;
2663 
2664     QLIST_FOREACH(child, &bs->children, next) {
2665         bdrv_io_plug(child->bs);
2666     }
2667 
2668     if (bs->io_plugged++ == 0) {
2669         BlockDriver *drv = bs->drv;
2670         if (drv && drv->bdrv_io_plug) {
2671             drv->bdrv_io_plug(bs);
2672         }
2673     }
2674 }
2675 
2676 void bdrv_io_unplug(BlockDriverState *bs)
2677 {
2678     BdrvChild *child;
2679 
2680     assert(bs->io_plugged);
2681     if (--bs->io_plugged == 0) {
2682         BlockDriver *drv = bs->drv;
2683         if (drv && drv->bdrv_io_unplug) {
2684             drv->bdrv_io_unplug(bs);
2685         }
2686     }
2687 
2688     QLIST_FOREACH(child, &bs->children, next) {
2689         bdrv_io_unplug(child->bs);
2690     }
2691 }
2692