block/io.c

   1 /*
   2  * Block layer I/O functions
   3  *
   4  * Copyright (c) 2003 Fabrice Bellard
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a copy
   7  * of this software and associated documentation files (the "Software"), to deal
   8  * in the Software without restriction, including without limitation the rights
   9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  10  * copies of the Software, and to permit persons to whom the Software is
  11  * furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  22  * THE SOFTWARE.
  23  */
  24
  25 #include "trace.h"
  26 #include "sysemu/block-backend.h"
  27 #include "block/blockjob.h"
  28 #include "block/block_int.h"
  29 #include "block/throttle-groups.h"
  30 #include "qemu/error-report.h"
  31
  32 #define NOT_DONE 0x7fffffff /* used while emulated sync operation in progress */
  33
  34 static BlockAIOCB *bdrv_aio_readv_em(BlockDriverState *bs,
  35         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
  36         BlockCompletionFunc *cb, void *opaque);
  37 static BlockAIOCB *bdrv_aio_writev_em(BlockDriverState *bs,
  38         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
  39         BlockCompletionFunc *cb, void *opaque);
  40 static int coroutine_fn bdrv_co_readv_em(BlockDriverState *bs,
  41                                          int64_t sector_num, int nb_sectors,
  42                                          QEMUIOVector *iov);
  43 static int coroutine_fn bdrv_co_writev_em(BlockDriverState *bs,
  44                                          int64_t sector_num, int nb_sectors,
  45                                          QEMUIOVector *iov);
  46 static int coroutine_fn bdrv_co_do_preadv(BlockDriverState *bs,
  47     int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
  48     BdrvRequestFlags flags);
  49 static int coroutine_fn bdrv_co_do_pwritev(BlockDriverState *bs,
  50     int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
  51     BdrvRequestFlags flags);
  52 static BlockAIOCB *bdrv_co_aio_rw_vector(BlockDriverState *bs,
  53                                          int64_t sector_num,
  54                                          QEMUIOVector *qiov,
  55                                          int nb_sectors,
  56                                          BdrvRequestFlags flags,
  57                                          BlockCompletionFunc *cb,
  58                                          void *opaque,
  59                                          bool is_write);
  60 static void coroutine_fn bdrv_co_do_rw(void *opaque);
  61 static int coroutine_fn bdrv_co_do_write_zeroes(BlockDriverState *bs,
  62     int64_t sector_num, int nb_sectors, BdrvRequestFlags flags);
  63
  64 /* throttling disk I/O limits */
  65 void bdrv_set_io_limits(BlockDriverState *bs,
  66                         ThrottleConfig *cfg)
  67 {
  68     int i;
  69
  70     throttle_group_config(bs, cfg);
  71
  72     for (i = 0; i < 2; i++) {
  73         qemu_co_enter_next(&bs->throttled_reqs[i]);
  74     }
  75 }
  76
  77 /* this function drain all the throttled IOs */
  78 static bool bdrv_start_throttled_reqs(BlockDriverState *bs)
  79 {
  80     bool drained = false;
  81     bool enabled = bs->io_limits_enabled;
  82     int i;
  83
  84     bs->io_limits_enabled = false;
  85
  86     for (i = 0; i < 2; i++) {
  87         while (qemu_co_enter_next(&bs->throttled_reqs[i])) {
  88             drained = true;
  89         }
  90     }
  91
  92     bs->io_limits_enabled = enabled;
  93
  94     return drained;
  95 }
  96
  97 void bdrv_io_limits_disable(BlockDriverState *bs)
  98 {
  99     bs->io_limits_enabled = false;
 100     bdrv_start_throttled_reqs(bs);
 101     throttle_group_unregister_bs(bs);
 102 }
 103
 104 /* should be called before bdrv_set_io_limits if a limit is set */
 105 void bdrv_io_limits_enable(BlockDriverState *bs, const char *group)
 106 {
 107     assert(!bs->io_limits_enabled);
 108     throttle_group_register_bs(bs, group);
 109     bs->io_limits_enabled = true;
 110 }
 111
 112 void bdrv_io_limits_update_group(BlockDriverState *bs, const char *group)
 113 {
 114     /* this bs is not part of any group */
 115     if (!bs->throttle_state) {
 116         return;
 117     }
 118
 119     /* this bs is a part of the same group than the one we want */
 120     if (!g_strcmp0(throttle_group_get_name(bs), group)) {
 121         return;
 122     }
 123
 124     /* need to change the group this bs belong to */
 125     bdrv_io_limits_disable(bs);
 126     bdrv_io_limits_enable(bs, group);
 127 }
 128
 129 void bdrv_setup_io_funcs(BlockDriver *bdrv)
 130 {
 131     /* Block drivers without coroutine functions need emulation */
 132     if (!bdrv->bdrv_co_readv) {
 133         bdrv->bdrv_co_readv = bdrv_co_readv_em;
 134         bdrv->bdrv_co_writev = bdrv_co_writev_em;
 135
 136         /* bdrv_co_readv_em()/brdv_co_writev_em() work in terms of aio, so if
 137          * the block driver lacks aio we need to emulate that too.
 138          */
 139         if (!bdrv->bdrv_aio_readv) {
 140             /* add AIO emulation layer */
 141             bdrv->bdrv_aio_readv = bdrv_aio_readv_em;
 142             bdrv->bdrv_aio_writev = bdrv_aio_writev_em;
 143         }
 144     }
 145 }
 146
 147 void bdrv_refresh_limits(BlockDriverState *bs, Error **errp)
 148 {
 149     BlockDriver *drv = bs->drv;
 150     Error *local_err = NULL;
 151
 152     memset(&bs->bl, 0, sizeof(bs->bl));
 153
 154     if (!drv) {
 155         return;
 156     }
 157
 158     /* Take some limits from the children as a default */
 159     if (bs->file) {
 160         bdrv_refresh_limits(bs->file->bs, &local_err);
 161         if (local_err) {
 162             error_propagate(errp, local_err);
 163             return;
 164         }
 165         bs->bl.opt_transfer_length = bs->file->bs->bl.opt_transfer_length;
 166         bs->bl.max_transfer_length = bs->file->bs->bl.max_transfer_length;
 167         bs->bl.min_mem_alignment = bs->file->bs->bl.min_mem_alignment;
 168         bs->bl.opt_mem_alignment = bs->file->bs->bl.opt_mem_alignment;
 169     } else {
 170         bs->bl.min_mem_alignment = 512;
 171         bs->bl.opt_mem_alignment = getpagesize();
 172     }
 173
 174     if (bs->backing) {
 175         bdrv_refresh_limits(bs->backing->bs, &local_err);
 176         if (local_err) {
 177             error_propagate(errp, local_err);
 178             return;
 179         }
 180         bs->bl.opt_transfer_length =
 181             MAX(bs->bl.opt_transfer_length,
 182                 bs->backing->bs->bl.opt_transfer_length);
 183         bs->bl.max_transfer_length =
 184             MIN_NON_ZERO(bs->bl.max_transfer_length,
 185                          bs->backing->bs->bl.max_transfer_length);
 186         bs->bl.opt_mem_alignment =
 187             MAX(bs->bl.opt_mem_alignment,
 188                 bs->backing->bs->bl.opt_mem_alignment);
 189         bs->bl.min_mem_alignment =
 190             MAX(bs->bl.min_mem_alignment,
 191                 bs->backing->bs->bl.min_mem_alignment);
 192     }
 193
 194     /* Then let the driver override it */
 195     if (drv->bdrv_refresh_limits) {
 196         drv->bdrv_refresh_limits(bs, errp);
 197     }
 198 }
 199
 200 /**
 201  * The copy-on-read flag is actually a reference count so multiple users may
 202  * use the feature without worrying about clobbering its previous state.
 203  * Copy-on-read stays enabled until all users have called to disable it.
 204  */
 205 void bdrv_enable_copy_on_read(BlockDriverState *bs)
 206 {
 207     bs->copy_on_read++;
 208 }
 209
 210 void bdrv_disable_copy_on_read(BlockDriverState *bs)
 211 {
 212     assert(bs->copy_on_read > 0);
 213     bs->copy_on_read--;
 214 }
 215
 216 /* Check if any requests are in-flight (including throttled requests) */
 217 bool bdrv_requests_pending(BlockDriverState *bs)
 218 {
 219     BdrvChild *child;
 220
 221     if (!QLIST_EMPTY(&bs->tracked_requests)) {
 222         return true;
 223     }
 224     if (!qemu_co_queue_empty(&bs->throttled_reqs[0])) {
 225         return true;
 226     }
 227     if (!qemu_co_queue_empty(&bs->throttled_reqs[1])) {
 228         return true;
 229     }
 230
 231     QLIST_FOREACH(child, &bs->children, next) {
 232         if (bdrv_requests_pending(child->bs)) {
 233             return true;
 234         }
 235     }
 236
 237     return false;
 238 }
 239
 240 static void bdrv_drain_recurse(BlockDriverState *bs)
 241 {
 242     BdrvChild *child;
 243
 244     if (bs->drv && bs->drv->bdrv_drain) {
 245         bs->drv->bdrv_drain(bs);
 246     }
 247     QLIST_FOREACH(child, &bs->children, next) {
 248         bdrv_drain_recurse(child->bs);
 249     }
 250 }
 251
 252 /*
 253  * Wait for pending requests to complete on a single BlockDriverState subtree,
 254  * and suspend block driver's internal I/O until next request arrives.
 255  *
 256  * Note that unlike bdrv_drain_all(), the caller must hold the BlockDriverState
 257  * AioContext.
 258  *
 259  * Only this BlockDriverState's AioContext is run, so in-flight requests must
 260  * not depend on events in other AioContexts.  In that case, use
 261  * bdrv_drain_all() instead.
 262  */
 263 void bdrv_drain(BlockDriverState *bs)
 264 {
 265     bool busy = true;
 266
 267     bdrv_drain_recurse(bs);
 268     while (busy) {
 269         /* Keep iterating */
 270          bdrv_flush_io_queue(bs);
 271          busy = bdrv_requests_pending(bs);
 272          busy |= aio_poll(bdrv_get_aio_context(bs), busy);
 273     }
 274 }
 275
 276 /*
 277  * Wait for pending requests to complete across all BlockDriverStates
 278  *
 279  * This function does not flush data to disk, use bdrv_flush_all() for that
 280  * after calling this function.
 281  */
 282 void bdrv_drain_all(void)
 283 {
 284     /* Always run first iteration so any pending completion BHs run */
 285     bool busy = true;
 286     BlockDriverState *bs = NULL;
 287     GSList *aio_ctxs = NULL, *ctx;
 288
 289     while ((bs = bdrv_next(bs))) {
 290         AioContext *aio_context = bdrv_get_aio_context(bs);
 291
 292         aio_context_acquire(aio_context);
 293         if (bs->job) {
 294             block_job_pause(bs->job);
 295         }
 296         aio_context_release(aio_context);
 297
 298         if (!g_slist_find(aio_ctxs, aio_context)) {
 299             aio_ctxs = g_slist_prepend(aio_ctxs, aio_context);
 300         }
 301     }
 302
 303     /* Note that completion of an asynchronous I/O operation can trigger any
 304      * number of other I/O operations on other devices---for example a
 305      * coroutine can submit an I/O request to another device in response to
 306      * request completion.  Therefore we must keep looping until there was no
 307      * more activity rather than simply draining each device independently.
 308      */
 309     while (busy) {
 310         busy = false;
 311
 312         for (ctx = aio_ctxs; ctx != NULL; ctx = ctx->next) {
 313             AioContext *aio_context = ctx->data;
 314             bs = NULL;
 315
 316             aio_context_acquire(aio_context);
 317             while ((bs = bdrv_next(bs))) {
 318                 if (aio_context == bdrv_get_aio_context(bs)) {
 319                     bdrv_flush_io_queue(bs);
 320                     if (bdrv_requests_pending(bs)) {
 321                         busy = true;
 322                         aio_poll(aio_context, busy);
 323                     }
 324                 }
 325             }
 326             busy |= aio_poll(aio_context, false);
 327             aio_context_release(aio_context);
 328         }
 329     }
 330
 331     bs = NULL;
 332     while ((bs = bdrv_next(bs))) {
 333         AioContext *aio_context = bdrv_get_aio_context(bs);
 334
 335         aio_context_acquire(aio_context);
 336         if (bs->job) {
 337             block_job_resume(bs->job);
 338         }
 339         aio_context_release(aio_context);
 340     }
 341     g_slist_free(aio_ctxs);
 342 }
 343
 344 /**
 345  * Remove an active request from the tracked requests list
 346  *
 347  * This function should be called when a tracked request is completing.
 348  */
 349 static void tracked_request_end(BdrvTrackedRequest *req)
 350 {
 351     if (req->serialising) {
 352         req->bs->serialising_in_flight--;
 353     }
 354
 355     QLIST_REMOVE(req, list);
 356     qemu_co_queue_restart_all(&req->wait_queue);
 357 }
 358
 359 /**
 360  * Add an active request to the tracked requests list
 361  */
 362 static void tracked_request_begin(BdrvTrackedRequest *req,
 363                                   BlockDriverState *bs,
 364                                   int64_t offset,
 365                                   unsigned int bytes,
 366                                   enum BdrvTrackedRequestType type)
 367 {
 368     *req = (BdrvTrackedRequest){
 369         .bs = bs,
 370         .offset         = offset,
 371         .bytes          = bytes,
 372         .type           = type,
 373         .co             = qemu_coroutine_self(),
 374         .serialising    = false,
 375         .overlap_offset = offset,
 376         .overlap_bytes  = bytes,
 377     };
 378
 379     qemu_co_queue_init(&req->wait_queue);
 380
 381     QLIST_INSERT_HEAD(&bs->tracked_requests, req, list);
 382 }
 383
 384 static void mark_request_serialising(BdrvTrackedRequest *req, uint64_t align)
 385 {
 386     int64_t overlap_offset = req->offset & ~(align - 1);
 387     unsigned int overlap_bytes = ROUND_UP(req->offset + req->bytes, align)
 388                                - overlap_offset;
 389
 390     if (!req->serialising) {
 391         req->bs->serialising_in_flight++;
 392         req->serialising = true;
 393     }
 394
 395     req->overlap_offset = MIN(req->overlap_offset, overlap_offset);
 396     req->overlap_bytes = MAX(req->overlap_bytes, overlap_bytes);
 397 }
 398
 399 /**
 400  * Round a region to cluster boundaries
 401  */
 402 void bdrv_round_to_clusters(BlockDriverState *bs,
 403                             int64_t sector_num, int nb_sectors,
 404                             int64_t *cluster_sector_num,
 405                             int *cluster_nb_sectors)
 406 {
 407     BlockDriverInfo bdi;
 408
 409     if (bdrv_get_info(bs, &bdi) < 0 || bdi.cluster_size == 0) {
 410         *cluster_sector_num = sector_num;
 411         *cluster_nb_sectors = nb_sectors;
 412     } else {
 413         int64_t c = bdi.cluster_size / BDRV_SECTOR_SIZE;
 414         *cluster_sector_num = QEMU_ALIGN_DOWN(sector_num, c);
 415         *cluster_nb_sectors = QEMU_ALIGN_UP(sector_num - *cluster_sector_num +
 416                                             nb_sectors, c);
 417     }
 418 }
 419
 420 static int bdrv_get_cluster_size(BlockDriverState *bs)
 421 {
 422     BlockDriverInfo bdi;
 423     int ret;
 424
 425     ret = bdrv_get_info(bs, &bdi);
 426     if (ret < 0 || bdi.cluster_size == 0) {
 427         return bs->request_alignment;
 428     } else {
 429         return bdi.cluster_size;
 430     }
 431 }
 432
 433 static bool tracked_request_overlaps(BdrvTrackedRequest *req,
 434                                      int64_t offset, unsigned int bytes)
 435 {
 436     /*        aaaa   bbbb */
 437     if (offset >= req->overlap_offset + req->overlap_bytes) {
 438         return false;
 439     }
 440     /* bbbb   aaaa        */
 441     if (req->overlap_offset >= offset + bytes) {
 442         return false;
 443     }
 444     return true;
 445 }
 446
 447 static bool coroutine_fn wait_serialising_requests(BdrvTrackedRequest *self)
 448 {
 449     BlockDriverState *bs = self->bs;
 450     BdrvTrackedRequest *req;
 451     bool retry;
 452     bool waited = false;
 453
 454     if (!bs->serialising_in_flight) {
 455         return false;
 456     }
 457
 458     do {
 459         retry = false;
 460         QLIST_FOREACH(req, &bs->tracked_requests, list) {
 461             if (req == self || (!req->serialising && !self->serialising)) {
 462                 continue;
 463             }
 464             if (tracked_request_overlaps(req, self->overlap_offset,
 465                                          self->overlap_bytes))
 466             {
 467                 /* Hitting this means there was a reentrant request, for
 468                  * example, a block driver issuing nested requests.  This must
 469                  * never happen since it means deadlock.
 470                  */
 471                 assert(qemu_coroutine_self() != req->co);
 472
 473                 /* If the request is already (indirectly) waiting for us, or
 474                  * will wait for us as soon as it wakes up, then just go on
 475                  * (instead of producing a deadlock in the former case). */
 476                 if (!req->waiting_for) {
 477                     self->waiting_for = req;
 478                     qemu_co_queue_wait(&req->wait_queue);
 479                     self->waiting_for = NULL;
 480                     retry = true;
 481                     waited = true;
 482                     break;
 483                 }
 484             }
 485         }
 486     } while (retry);
 487
 488     return waited;
 489 }
 490
 491 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset,
 492                                    size_t size)
 493 {
 494     if (size > BDRV_REQUEST_MAX_SECTORS << BDRV_SECTOR_BITS) {
 495         return -EIO;
 496     }
 497
 498     if (!bdrv_is_inserted(bs)) {
 499         return -ENOMEDIUM;
 500     }
 501
 502     if (offset < 0) {
 503         return -EIO;
 504     }
 505
 506     return 0;
 507 }
 508
 509 static int bdrv_check_request(BlockDriverState *bs, int64_t sector_num,
 510                               int nb_sectors)
 511 {
 512     if (nb_sectors < 0 || nb_sectors > BDRV_REQUEST_MAX_SECTORS) {
 513         return -EIO;
 514     }
 515
 516     return bdrv_check_byte_request(bs, sector_num * BDRV_SECTOR_SIZE,
 517                                    nb_sectors * BDRV_SECTOR_SIZE);
 518 }
 519
 520 typedef struct RwCo {
 521     BlockDriverState *bs;
 522     int64_t offset;
 523     QEMUIOVector *qiov;
 524     bool is_write;
 525     int ret;
 526     BdrvRequestFlags flags;
 527 } RwCo;
 528
 529 static void coroutine_fn bdrv_rw_co_entry(void *opaque)
 530 {
 531     RwCo *rwco = opaque;
 532
 533     if (!rwco->is_write) {
 534         rwco->ret = bdrv_co_do_preadv(rwco->bs, rwco->offset,
 535                                       rwco->qiov->size, rwco->qiov,
 536                                       rwco->flags);
 537     } else {
 538         rwco->ret = bdrv_co_do_pwritev(rwco->bs, rwco->offset,
 539                                        rwco->qiov->size, rwco->qiov,
 540                                        rwco->flags);
 541     }
 542 }
 543
 544 /*
 545  * Process a vectored synchronous request using coroutines
 546  */
 547 static int bdrv_prwv_co(BlockDriverState *bs, int64_t offset,
 548                         QEMUIOVector *qiov, bool is_write,
 549                         BdrvRequestFlags flags)
 550 {
 551     Coroutine *co;
 552     RwCo rwco = {
 553         .bs = bs,
 554         .offset = offset,
 555         .qiov = qiov,
 556         .is_write = is_write,
 557         .ret = NOT_DONE,
 558         .flags = flags,
 559     };
 560
 561     /**
 562      * In sync call context, when the vcpu is blocked, this throttling timer
 563      * will not fire; so the I/O throttling function has to be disabled here
 564      * if it has been enabled.
 565      */
 566     if (bs->io_limits_enabled) {
 567         fprintf(stderr, "Disabling I/O throttling on '%s' due "
 568                         "to synchronous I/O.\n", bdrv_get_device_name(bs));
 569         bdrv_io_limits_disable(bs);
 570     }
 571
 572     if (qemu_in_coroutine()) {
 573         /* Fast-path if already in coroutine context */
 574         bdrv_rw_co_entry(&rwco);
 575     } else {
 576         AioContext *aio_context = bdrv_get_aio_context(bs);
 577
 578         co = qemu_coroutine_create(bdrv_rw_co_entry);
 579         qemu_coroutine_enter(co, &rwco);
 580         while (rwco.ret == NOT_DONE) {
 581             aio_poll(aio_context, true);
 582         }
 583     }
 584     return rwco.ret;
 585 }
 586
 587 /*
 588  * Process a synchronous request using coroutines
 589  */
 590 static int bdrv_rw_co(BlockDriverState *bs, int64_t sector_num, uint8_t *buf,
 591                       int nb_sectors, bool is_write, BdrvRequestFlags flags)
 592 {
 593     QEMUIOVector qiov;
 594     struct iovec iov = {
 595         .iov_base = (void *)buf,
 596         .iov_len = nb_sectors * BDRV_SECTOR_SIZE,
 597     };
 598
 599     if (nb_sectors < 0 || nb_sectors > BDRV_REQUEST_MAX_SECTORS) {
 600         return -EINVAL;
 601     }
 602
 603     qemu_iovec_init_external(&qiov, &iov, 1);
 604     return bdrv_prwv_co(bs, sector_num << BDRV_SECTOR_BITS,
 605                         &qiov, is_write, flags);
 606 }
 607
 608 /* return < 0 if error. See bdrv_write() for the return codes */
 609 int bdrv_read(BlockDriverState *bs, int64_t sector_num,
 610               uint8_t *buf, int nb_sectors)
 611 {
 612     return bdrv_rw_co(bs, sector_num, buf, nb_sectors, false, 0);
 613 }
 614
 615 /* Just like bdrv_read(), but with I/O throttling temporarily disabled */
 616 int bdrv_read_unthrottled(BlockDriverState *bs, int64_t sector_num,
 617                           uint8_t *buf, int nb_sectors)
 618 {
 619     bool enabled;
 620     int ret;
 621
 622     enabled = bs->io_limits_enabled;
 623     bs->io_limits_enabled = false;
 624     ret = bdrv_read(bs, sector_num, buf, nb_sectors);
 625     bs->io_limits_enabled = enabled;
 626     return ret;
 627 }
 628
 629 /* Return < 0 if error. Important errors are:
 630   -EIO         generic I/O error (may happen for all errors)
 631   -ENOMEDIUM   No media inserted.
 632   -EINVAL      Invalid sector number or nb_sectors
 633   -EACCES      Trying to write a read-only device
 634 */
 635 int bdrv_write(BlockDriverState *bs, int64_t sector_num,
 636                const uint8_t *buf, int nb_sectors)
 637 {
 638     return bdrv_rw_co(bs, sector_num, (uint8_t *)buf, nb_sectors, true, 0);
 639 }
 640
 641 int bdrv_write_zeroes(BlockDriverState *bs, int64_t sector_num,
 642                       int nb_sectors, BdrvRequestFlags flags)
 643 {
 644     return bdrv_rw_co(bs, sector_num, NULL, nb_sectors, true,
 645                       BDRV_REQ_ZERO_WRITE | flags);
 646 }
 647
 648 /*
 649  * Completely zero out a block device with the help of bdrv_write_zeroes.
 650  * The operation is sped up by checking the block status and only writing
 651  * zeroes to the device if they currently do not return zeroes. Optional
 652  * flags are passed through to bdrv_write_zeroes (e.g. BDRV_REQ_MAY_UNMAP).
 653  *
 654  * Returns < 0 on error, 0 on success. For error codes see bdrv_write().
 655  */
 656 int bdrv_make_zero(BlockDriverState *bs, BdrvRequestFlags flags)
 657 {
 658     int64_t target_sectors, ret, nb_sectors, sector_num = 0;
 659     int n;
 660
 661     target_sectors = bdrv_nb_sectors(bs);
 662     if (target_sectors < 0) {
 663         return target_sectors;
 664     }
 665
 666     for (;;) {
 667         nb_sectors = MIN(target_sectors - sector_num, BDRV_REQUEST_MAX_SECTORS);
 668         if (nb_sectors <= 0) {
 669             return 0;
 670         }
 671         ret = bdrv_get_block_status(bs, sector_num, nb_sectors, &n);
 672         if (ret < 0) {
 673             error_report("error getting block status at sector %" PRId64 ": %s",
 674                          sector_num, strerror(-ret));
 675             return ret;
 676         }
 677         if (ret & BDRV_BLOCK_ZERO) {
 678             sector_num += n;
 679             continue;
 680         }
 681         ret = bdrv_write_zeroes(bs, sector_num, n, flags);
 682         if (ret < 0) {
 683             error_report("error writing zeroes at sector %" PRId64 ": %s",
 684                          sector_num, strerror(-ret));
 685             return ret;
 686         }
 687         sector_num += n;
 688     }
 689 }
 690
 691 int bdrv_pread(BlockDriverState *bs, int64_t offset, void *buf, int bytes)
 692 {
 693     QEMUIOVector qiov;
 694     struct iovec iov = {
 695         .iov_base = (void *)buf,
 696         .iov_len = bytes,
 697     };
 698     int ret;
 699
 700     if (bytes < 0) {
 701         return -EINVAL;
 702     }
 703
 704     qemu_iovec_init_external(&qiov, &iov, 1);
 705     ret = bdrv_prwv_co(bs, offset, &qiov, false, 0);
 706     if (ret < 0) {
 707         return ret;
 708     }
 709
 710     return bytes;
 711 }
 712
 713 int bdrv_pwritev(BlockDriverState *bs, int64_t offset, QEMUIOVector *qiov)
 714 {
 715     int ret;
 716
 717     ret = bdrv_prwv_co(bs, offset, qiov, true, 0);
 718     if (ret < 0) {
 719         return ret;
 720     }
 721
 722     return qiov->size;
 723 }
 724
 725 int bdrv_pwrite(BlockDriverState *bs, int64_t offset,
 726                 const void *buf, int bytes)
 727 {
 728     QEMUIOVector qiov;
 729     struct iovec iov = {
 730         .iov_base   = (void *) buf,
 731         .iov_len    = bytes,
 732     };
 733
 734     if (bytes < 0) {
 735         return -EINVAL;
 736     }
 737
 738     qemu_iovec_init_external(&qiov, &iov, 1);
 739     return bdrv_pwritev(bs, offset, &qiov);
 740 }
 741
 742 /*
 743  * Writes to the file and ensures that no writes are reordered across this
 744  * request (acts as a barrier)
 745  *
 746  * Returns 0 on success, -errno in error cases.
 747  */
 748 int bdrv_pwrite_sync(BlockDriverState *bs, int64_t offset,
 749     const void *buf, int count)
 750 {
 751     int ret;
 752
 753     ret = bdrv_pwrite(bs, offset, buf, count);
 754     if (ret < 0) {
 755         return ret;
 756     }
 757
 758     /* No flush needed for cache modes that already do it */
 759     if (bs->enable_write_cache) {
 760         bdrv_flush(bs);
 761     }
 762
 763     return 0;
 764 }
 765
 766 static int coroutine_fn bdrv_co_do_copy_on_readv(BlockDriverState *bs,
 767         int64_t sector_num, int nb_sectors, QEMUIOVector *qiov)
 768 {
 769     /* Perform I/O through a temporary buffer so that users who scribble over
 770      * their read buffer while the operation is in progress do not end up
 771      * modifying the image file.  This is critical for zero-copy guest I/O
 772      * where anything might happen inside guest memory.
 773      */
 774     void *bounce_buffer;
 775
 776     BlockDriver *drv = bs->drv;
 777     struct iovec iov;
 778     QEMUIOVector bounce_qiov;
 779     int64_t cluster_sector_num;
 780     int cluster_nb_sectors;
 781     size_t skip_bytes;
 782     int ret;
 783
 784     /* Cover entire cluster so no additional backing file I/O is required when
 785      * allocating cluster in the image file.
 786      */
 787     bdrv_round_to_clusters(bs, sector_num, nb_sectors,
 788                            &cluster_sector_num, &cluster_nb_sectors);
 789
 790     trace_bdrv_co_do_copy_on_readv(bs, sector_num, nb_sectors,
 791                                    cluster_sector_num, cluster_nb_sectors);
 792
 793     iov.iov_len = cluster_nb_sectors * BDRV_SECTOR_SIZE;
 794     iov.iov_base = bounce_buffer = qemu_try_blockalign(bs, iov.iov_len);
 795     if (bounce_buffer == NULL) {
 796         ret = -ENOMEM;
 797         goto err;
 798     }
 799
 800     qemu_iovec_init_external(&bounce_qiov, &iov, 1);
 801
 802     ret = drv->bdrv_co_readv(bs, cluster_sector_num, cluster_nb_sectors,
 803                              &bounce_qiov);
 804     if (ret < 0) {
 805         goto err;
 806     }
 807
 808     if (drv->bdrv_co_write_zeroes &&
 809         buffer_is_zero(bounce_buffer, iov.iov_len)) {
 810         ret = bdrv_co_do_write_zeroes(bs, cluster_sector_num,
 811                                       cluster_nb_sectors, 0);
 812     } else {
 813         /* This does not change the data on the disk, it is not necessary
 814          * to flush even in cache=writethrough mode.
 815          */
 816         ret = drv->bdrv_co_writev(bs, cluster_sector_num, cluster_nb_sectors,
 817                                   &bounce_qiov);
 818     }
 819
 820     if (ret < 0) {
 821         /* It might be okay to ignore write errors for guest requests.  If this
 822          * is a deliberate copy-on-read then we don't want to ignore the error.
 823          * Simply report it in all cases.
 824          */
 825         goto err;
 826     }
 827
 828     skip_bytes = (sector_num - cluster_sector_num) * BDRV_SECTOR_SIZE;
 829     qemu_iovec_from_buf(qiov, 0, bounce_buffer + skip_bytes,
 830                         nb_sectors * BDRV_SECTOR_SIZE);
 831
 832 err:
 833     qemu_vfree(bounce_buffer);
 834     return ret;
 835 }
 836
 837 /*
 838  * Forwards an already correctly aligned request to the BlockDriver. This
 839  * handles copy on read and zeroing after EOF; any other features must be
 840  * implemented by the caller.
 841  */
 842 static int coroutine_fn bdrv_aligned_preadv(BlockDriverState *bs,
 843     BdrvTrackedRequest *req, int64_t offset, unsigned int bytes,
 844     int64_t align, QEMUIOVector *qiov, int flags)
 845 {
 846     BlockDriver *drv = bs->drv;
 847     int ret;
 848
 849     int64_t sector_num = offset >> BDRV_SECTOR_BITS;
 850     unsigned int nb_sectors = bytes >> BDRV_SECTOR_BITS;
 851
 852     assert((offset & (BDRV_SECTOR_SIZE - 1)) == 0);
 853     assert((bytes & (BDRV_SECTOR_SIZE - 1)) == 0);
 854     assert(!qiov || bytes == qiov->size);
 855
 856     /* Handle Copy on Read and associated serialisation */
 857     if (flags & BDRV_REQ_COPY_ON_READ) {
 858         /* If we touch the same cluster it counts as an overlap.  This
 859          * guarantees that allocating writes will be serialized and not race
 860          * with each other for the same cluster.  For example, in copy-on-read
 861          * it ensures that the CoR read and write operations are atomic and
 862          * guest writes cannot interleave between them. */
 863         mark_request_serialising(req, bdrv_get_cluster_size(bs));
 864     }
 865
 866     if (!(flags & BDRV_REQ_NO_SERIALISING)) {
 867         wait_serialising_requests(req);
 868     }
 869
 870     if (flags & BDRV_REQ_COPY_ON_READ) {
 871         int pnum;
 872
 873         ret = bdrv_is_allocated(bs, sector_num, nb_sectors, &pnum);
 874         if (ret < 0) {
 875             goto out;
 876         }
 877
 878         if (!ret || pnum != nb_sectors) {
 879             ret = bdrv_co_do_copy_on_readv(bs, sector_num, nb_sectors, qiov);
 880             goto out;
 881         }
 882     }
 883
 884     /* Forward the request to the BlockDriver */
 885     if (!bs->zero_beyond_eof) {
 886         ret = drv->bdrv_co_readv(bs, sector_num, nb_sectors, qiov);
 887     } else {
 888         /* Read zeros after EOF */
 889         int64_t total_sectors, max_nb_sectors;
 890
 891         total_sectors = bdrv_nb_sectors(bs);
 892         if (total_sectors < 0) {
 893             ret = total_sectors;
 894             goto out;
 895         }
 896
 897         max_nb_sectors = ROUND_UP(MAX(0, total_sectors - sector_num),
 898                                   align >> BDRV_SECTOR_BITS);
 899         if (nb_sectors < max_nb_sectors) {
 900             ret = drv->bdrv_co_readv(bs, sector_num, nb_sectors, qiov);
 901         } else if (max_nb_sectors > 0) {
 902             QEMUIOVector local_qiov;
 903
 904             qemu_iovec_init(&local_qiov, qiov->niov);
 905             qemu_iovec_concat(&local_qiov, qiov, 0,
 906                               max_nb_sectors * BDRV_SECTOR_SIZE);
 907
 908             ret = drv->bdrv_co_readv(bs, sector_num, max_nb_sectors,
 909                                      &local_qiov);
 910
 911             qemu_iovec_destroy(&local_qiov);
 912         } else {
 913             ret = 0;
 914         }
 915
 916         /* Reading beyond end of file is supposed to produce zeroes */
 917         if (ret == 0 && total_sectors < sector_num + nb_sectors) {
 918             uint64_t offset = MAX(0, total_sectors - sector_num);
 919             uint64_t bytes = (sector_num + nb_sectors - offset) *
 920                               BDRV_SECTOR_SIZE;
 921             qemu_iovec_memset(qiov, offset * BDRV_SECTOR_SIZE, 0, bytes);
 922         }
 923     }
 924
 925 out:
 926     return ret;
 927 }
 928
 929 /*
 930  * Handle a read request in coroutine context
 931  */
 932 static int coroutine_fn bdrv_co_do_preadv(BlockDriverState *bs,
 933     int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
 934     BdrvRequestFlags flags)
 935 {
 936     BlockDriver *drv = bs->drv;
 937     BdrvTrackedRequest req;
 938
 939     /* TODO Lift BDRV_SECTOR_SIZE restriction in BlockDriver interface */
 940     uint64_t align = MAX(BDRV_SECTOR_SIZE, bs->request_alignment);
 941     uint8_t *head_buf = NULL;
 942     uint8_t *tail_buf = NULL;
 943     QEMUIOVector local_qiov;
 944     bool use_local_qiov = false;
 945     int ret;
 946
 947     if (!drv) {
 948         return -ENOMEDIUM;
 949     }
 950
 951     ret = bdrv_check_byte_request(bs, offset, bytes);
 952     if (ret < 0) {
 953         return ret;
 954     }
 955
 956     /* Don't do copy-on-read if we read data before write operation */
 957     if (bs->copy_on_read && !(flags & BDRV_REQ_NO_SERIALISING)) {
 958         flags |= BDRV_REQ_COPY_ON_READ;
 959     }
 960
 961     /* throttling disk I/O */
 962     if (bs->io_limits_enabled) {
 963         throttle_group_co_io_limits_intercept(bs, bytes, false);
 964     }
 965
 966     /* Align read if necessary by padding qiov */
 967     if (offset & (align - 1)) {
 968         head_buf = qemu_blockalign(bs, align);
 969         qemu_iovec_init(&local_qiov, qiov->niov + 2);
 970         qemu_iovec_add(&local_qiov, head_buf, offset & (align - 1));
 971         qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size);
 972         use_local_qiov = true;
 973
 974         bytes += offset & (align - 1);
 975         offset = offset & ~(align - 1);
 976     }
 977
 978     if ((offset + bytes) & (align - 1)) {
 979         if (!use_local_qiov) {
 980             qemu_iovec_init(&local_qiov, qiov->niov + 1);
 981             qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size);
 982             use_local_qiov = true;
 983         }
 984         tail_buf = qemu_blockalign(bs, align);
 985         qemu_iovec_add(&local_qiov, tail_buf,
 986                        align - ((offset + bytes) & (align - 1)));
 987
 988         bytes = ROUND_UP(bytes, align);
 989     }
 990
 991     tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_READ);
 992     ret = bdrv_aligned_preadv(bs, &req, offset, bytes, align,
 993                               use_local_qiov ? &local_qiov : qiov,
 994                               flags);
 995     tracked_request_end(&req);
 996
 997     if (use_local_qiov) {
 998         qemu_iovec_destroy(&local_qiov);
 999         qemu_vfree(head_buf);
1000         qemu_vfree(tail_buf);
1001     }
1002
1003     return ret;
1004 }
1005
1006 static int coroutine_fn bdrv_co_do_readv(BlockDriverState *bs,
1007     int64_t sector_num, int nb_sectors, QEMUIOVector *qiov,
1008     BdrvRequestFlags flags)
1009 {
1010     if (nb_sectors < 0 || nb_sectors > BDRV_REQUEST_MAX_SECTORS) {
1011         return -EINVAL;
1012     }
1013
1014     return bdrv_co_do_preadv(bs, sector_num << BDRV_SECTOR_BITS,
1015                              nb_sectors << BDRV_SECTOR_BITS, qiov, flags);
1016 }
1017
1018 int coroutine_fn bdrv_co_readv(BlockDriverState *bs, int64_t sector_num,
1019     int nb_sectors, QEMUIOVector *qiov)
1020 {
1021     trace_bdrv_co_readv(bs, sector_num, nb_sectors);
1022
1023     return bdrv_co_do_readv(bs, sector_num, nb_sectors, qiov, 0);
1024 }
1025
1026 int coroutine_fn bdrv_co_readv_no_serialising(BlockDriverState *bs,
1027     int64_t sector_num, int nb_sectors, QEMUIOVector *qiov)
1028 {
1029     trace_bdrv_co_readv_no_serialising(bs, sector_num, nb_sectors);
1030
1031     return bdrv_co_do_readv(bs, sector_num, nb_sectors, qiov,
1032                             BDRV_REQ_NO_SERIALISING);
1033 }
1034
1035 int coroutine_fn bdrv_co_copy_on_readv(BlockDriverState *bs,
1036     int64_t sector_num, int nb_sectors, QEMUIOVector *qiov)
1037 {
1038     trace_bdrv_co_copy_on_readv(bs, sector_num, nb_sectors);
1039
1040     return bdrv_co_do_readv(bs, sector_num, nb_sectors, qiov,
1041                             BDRV_REQ_COPY_ON_READ);
1042 }
1043
1044 #define MAX_WRITE_ZEROES_BOUNCE_BUFFER 32768
1045
1046 static int coroutine_fn bdrv_co_do_write_zeroes(BlockDriverState *bs,
1047     int64_t sector_num, int nb_sectors, BdrvRequestFlags flags)
1048 {
1049     BlockDriver *drv = bs->drv;
1050     QEMUIOVector qiov;
1051     struct iovec iov = {0};
1052     int ret = 0;
1053
1054     int max_write_zeroes = MIN_NON_ZERO(bs->bl.max_write_zeroes,
1055                                         BDRV_REQUEST_MAX_SECTORS);
1056
1057     while (nb_sectors > 0 && !ret) {
1058         int num = nb_sectors;
1059
1060         /* Align request.  Block drivers can expect the "bulk" of the request
1061          * to be aligned.
1062          */
1063         if (bs->bl.write_zeroes_alignment
1064             && num > bs->bl.write_zeroes_alignment) {
1065             if (sector_num % bs->bl.write_zeroes_alignment != 0) {
1066                 /* Make a small request up to the first aligned sector.  */
1067                 num = bs->bl.write_zeroes_alignment;
1068                 num -= sector_num % bs->bl.write_zeroes_alignment;
1069             } else if ((sector_num + num) % bs->bl.write_zeroes_alignment != 0) {
1070                 /* Shorten the request to the last aligned sector.  num cannot
1071                  * underflow because num > bs->bl.write_zeroes_alignment.
1072                  */
1073                 num -= (sector_num + num) % bs->bl.write_zeroes_alignment;
1074             }
1075         }
1076
1077         /* limit request size */
1078         if (num > max_write_zeroes) {
1079             num = max_write_zeroes;
1080         }
1081
1082         ret = -ENOTSUP;
1083         /* First try the efficient write zeroes operation */
1084         if (drv->bdrv_co_write_zeroes) {
1085             ret = drv->bdrv_co_write_zeroes(bs, sector_num, num, flags);
1086         }
1087
1088         if (ret == -ENOTSUP) {
1089             /* Fall back to bounce buffer if write zeroes is unsupported */
1090             int max_xfer_len = MIN_NON_ZERO(bs->bl.max_transfer_length,
1091                                             MAX_WRITE_ZEROES_BOUNCE_BUFFER);
1092             num = MIN(num, max_xfer_len);
1093             iov.iov_len = num * BDRV_SECTOR_SIZE;
1094             if (iov.iov_base == NULL) {
1095                 iov.iov_base = qemu_try_blockalign(bs, num * BDRV_SECTOR_SIZE);
1096                 if (iov.iov_base == NULL) {
1097                     ret = -ENOMEM;
1098                     goto fail;
1099                 }
1100                 memset(iov.iov_base, 0, num * BDRV_SECTOR_SIZE);
1101             }
1102             qemu_iovec_init_external(&qiov, &iov, 1);
1103
1104             ret = drv->bdrv_co_writev(bs, sector_num, num, &qiov);
1105
1106             /* Keep bounce buffer around if it is big enough for all
1107              * all future requests.
1108              */
1109             if (num < max_xfer_len) {
1110                 qemu_vfree(iov.iov_base);
1111                 iov.iov_base = NULL;
1112             }
1113         }
1114
1115         sector_num += num;
1116         nb_sectors -= num;
1117     }
1118
1119 fail:
1120     qemu_vfree(iov.iov_base);
1121     return ret;
1122 }
1123
1124 /*
1125  * Forwards an already correctly aligned write request to the BlockDriver.
1126  */
1127 static int coroutine_fn bdrv_aligned_pwritev(BlockDriverState *bs,
1128     BdrvTrackedRequest *req, int64_t offset, unsigned int bytes,
1129     QEMUIOVector *qiov, int flags)
1130 {
1131     BlockDriver *drv = bs->drv;
1132     bool waited;
1133     int ret;
1134
1135     int64_t sector_num = offset >> BDRV_SECTOR_BITS;
1136     unsigned int nb_sectors = bytes >> BDRV_SECTOR_BITS;
1137
1138     assert((offset & (BDRV_SECTOR_SIZE - 1)) == 0);
1139     assert((bytes & (BDRV_SECTOR_SIZE - 1)) == 0);
1140     assert(!qiov || bytes == qiov->size);
1141
1142     waited = wait_serialising_requests(req);
1143     assert(!waited || !req->serialising);
1144     assert(req->overlap_offset <= offset);
1145     assert(offset + bytes <= req->overlap_offset + req->overlap_bytes);
1146
1147     ret = notifier_with_return_list_notify(&bs->before_write_notifiers, req);
1148
1149     if (!ret && bs->detect_zeroes != BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF &&
1150         !(flags & BDRV_REQ_ZERO_WRITE) && drv->bdrv_co_write_zeroes &&
1151         qemu_iovec_is_zero(qiov)) {
1152         flags |= BDRV_REQ_ZERO_WRITE;
1153         if (bs->detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP) {
1154             flags |= BDRV_REQ_MAY_UNMAP;
1155         }
1156     }
1157
1158     if (ret < 0) {
1159         /* Do nothing, write notifier decided to fail this request */
1160     } else if (flags & BDRV_REQ_ZERO_WRITE) {
1161         bdrv_debug_event(bs, BLKDBG_PWRITEV_ZERO);
1162         ret = bdrv_co_do_write_zeroes(bs, sector_num, nb_sectors, flags);
1163     } else {
1164         bdrv_debug_event(bs, BLKDBG_PWRITEV);
1165         ret = drv->bdrv_co_writev(bs, sector_num, nb_sectors, qiov);
1166     }
1167     bdrv_debug_event(bs, BLKDBG_PWRITEV_DONE);
1168
1169     if (ret == 0 && !bs->enable_write_cache) {
1170         ret = bdrv_co_flush(bs);
1171     }
1172
1173     bdrv_set_dirty(bs, sector_num, nb_sectors);
1174
1175     if (bs->wr_highest_offset < offset + bytes) {
1176         bs->wr_highest_offset = offset + bytes;
1177     }
1178
1179     if (ret >= 0) {
1180         bs->total_sectors = MAX(bs->total_sectors, sector_num + nb_sectors);
1181     }
1182
1183     return ret;
1184 }
1185
1186 static int coroutine_fn bdrv_co_do_zero_pwritev(BlockDriverState *bs,
1187                                                 int64_t offset,
1188                                                 unsigned int bytes,
1189                                                 BdrvRequestFlags flags,
1190                                                 BdrvTrackedRequest *req)
1191 {
1192     uint8_t *buf = NULL;
1193     QEMUIOVector local_qiov;
1194     struct iovec iov;
1195     uint64_t align = MAX(BDRV_SECTOR_SIZE, bs->request_alignment);
1196     unsigned int head_padding_bytes, tail_padding_bytes;
1197     int ret = 0;
1198
1199     head_padding_bytes = offset & (align - 1);
1200     tail_padding_bytes = align - ((offset + bytes) & (align - 1));
1201
1202
1203     assert(flags & BDRV_REQ_ZERO_WRITE);
1204     if (head_padding_bytes || tail_padding_bytes) {
1205         buf = qemu_blockalign(bs, align);
1206         iov = (struct iovec) {
1207             .iov_base   = buf,
1208             .iov_len    = align,
1209         };
1210         qemu_iovec_init_external(&local_qiov, &iov, 1);
1211     }
1212     if (head_padding_bytes) {
1213         uint64_t zero_bytes = MIN(bytes, align - head_padding_bytes);
1214
1215         /* RMW the unaligned part before head. */
1216         mark_request_serialising(req, align);
1217         wait_serialising_requests(req);
1218         bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD);
1219         ret = bdrv_aligned_preadv(bs, req, offset & ~(align - 1), align,
1220                                   align, &local_qiov, 0);
1221         if (ret < 0) {
1222             goto fail;
1223         }
1224         bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD);
1225
1226         memset(buf + head_padding_bytes, 0, zero_bytes);
1227         ret = bdrv_aligned_pwritev(bs, req, offset & ~(align - 1), align,
1228                                    &local_qiov,
1229                                    flags & ~BDRV_REQ_ZERO_WRITE);
1230         if (ret < 0) {
1231             goto fail;
1232         }
1233         offset += zero_bytes;
1234         bytes -= zero_bytes;
1235     }
1236
1237     assert(!bytes || (offset & (align - 1)) == 0);
1238     if (bytes >= align) {
1239         /* Write the aligned part in the middle. */
1240         uint64_t aligned_bytes = bytes & ~(align - 1);
1241         ret = bdrv_aligned_pwritev(bs, req, offset, aligned_bytes,
1242                                    NULL, flags);
1243         if (ret < 0) {
1244             goto fail;
1245         }
1246         bytes -= aligned_bytes;
1247         offset += aligned_bytes;
1248     }
1249
1250     assert(!bytes || (offset & (align - 1)) == 0);
1251     if (bytes) {
1252         assert(align == tail_padding_bytes + bytes);
1253         /* RMW the unaligned part after tail. */
1254         mark_request_serialising(req, align);
1255         wait_serialising_requests(req);
1256         bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
1257         ret = bdrv_aligned_preadv(bs, req, offset, align,
1258                                   align, &local_qiov, 0);
1259         if (ret < 0) {
1260             goto fail;
1261         }
1262         bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
1263
1264         memset(buf, 0, bytes);
1265         ret = bdrv_aligned_pwritev(bs, req, offset, align,
1266                                    &local_qiov, flags & ~BDRV_REQ_ZERO_WRITE);
1267     }
1268 fail:
1269     qemu_vfree(buf);
1270     return ret;
1271
1272 }
1273
1274 /*
1275  * Handle a write request in coroutine context
1276  */
1277 static int coroutine_fn bdrv_co_do_pwritev(BlockDriverState *bs,
1278     int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
1279     BdrvRequestFlags flags)
1280 {
1281     BdrvTrackedRequest req;
1282     /* TODO Lift BDRV_SECTOR_SIZE restriction in BlockDriver interface */
1283     uint64_t align = MAX(BDRV_SECTOR_SIZE, bs->request_alignment);
1284     uint8_t *head_buf = NULL;
1285     uint8_t *tail_buf = NULL;
1286     QEMUIOVector local_qiov;
1287     bool use_local_qiov = false;
1288     int ret;
1289
1290     if (!bs->drv) {
1291         return -ENOMEDIUM;
1292     }
1293     if (bs->read_only) {
1294         return -EPERM;
1295     }
1296
1297     ret = bdrv_check_byte_request(bs, offset, bytes);
1298     if (ret < 0) {
1299         return ret;
1300     }
1301
1302     /* throttling disk I/O */
1303     if (bs->io_limits_enabled) {
1304         throttle_group_co_io_limits_intercept(bs, bytes, true);
1305     }
1306
1307     /*
1308      * Align write if necessary by performing a read-modify-write cycle.
1309      * Pad qiov with the read parts and be sure to have a tracked request not
1310      * only for bdrv_aligned_pwritev, but also for the reads of the RMW cycle.
1311      */
1312     tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_WRITE);
1313
1314     if (!qiov) {
1315         ret = bdrv_co_do_zero_pwritev(bs, offset, bytes, flags, &req);
1316         goto out;
1317     }
1318
1319     if (offset & (align - 1)) {
1320         QEMUIOVector head_qiov;
1321         struct iovec head_iov;
1322
1323         mark_request_serialising(&req, align);
1324         wait_serialising_requests(&req);
1325
1326         head_buf = qemu_blockalign(bs, align);
1327         head_iov = (struct iovec) {
1328             .iov_base   = head_buf,
1329             .iov_len    = align,
1330         };
1331         qemu_iovec_init_external(&head_qiov, &head_iov, 1);
1332
1333         bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD);
1334         ret = bdrv_aligned_preadv(bs, &req, offset & ~(align - 1), align,
1335                                   align, &head_qiov, 0);
1336         if (ret < 0) {
1337             goto fail;
1338         }
1339         bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD);
1340
1341         qemu_iovec_init(&local_qiov, qiov->niov + 2);
1342         qemu_iovec_add(&local_qiov, head_buf, offset & (align - 1));
1343         qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size);
1344         use_local_qiov = true;
1345
1346         bytes += offset & (align - 1);
1347         offset = offset & ~(align - 1);
1348     }
1349
1350     if ((offset + bytes) & (align - 1)) {
1351         QEMUIOVector tail_qiov;
1352         struct iovec tail_iov;
1353         size_t tail_bytes;
1354         bool waited;
1355
1356         mark_request_serialising(&req, align);
1357         waited = wait_serialising_requests(&req);
1358         assert(!waited || !use_local_qiov);
1359
1360         tail_buf = qemu_blockalign(bs, align);
1361         tail_iov = (struct iovec) {
1362             .iov_base   = tail_buf,
1363             .iov_len    = align,
1364         };
1365         qemu_iovec_init_external(&tail_qiov, &tail_iov, 1);
1366
1367         bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
1368         ret = bdrv_aligned_preadv(bs, &req, (offset + bytes) & ~(align - 1), align,
1369                                   align, &tail_qiov, 0);
1370         if (ret < 0) {
1371             goto fail;
1372         }
1373         bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
1374
1375         if (!use_local_qiov) {
1376             qemu_iovec_init(&local_qiov, qiov->niov + 1);
1377             qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size);
1378             use_local_qiov = true;
1379         }
1380
1381         tail_bytes = (offset + bytes) & (align - 1);
1382         qemu_iovec_add(&local_qiov, tail_buf + tail_bytes, align - tail_bytes);
1383
1384         bytes = ROUND_UP(bytes, align);
1385     }
1386
1387     ret = bdrv_aligned_pwritev(bs, &req, offset, bytes,
1388                                use_local_qiov ? &local_qiov : qiov,
1389                                flags);
1390
1391 fail:
1392
1393     if (use_local_qiov) {
1394         qemu_iovec_destroy(&local_qiov);
1395     }
1396     qemu_vfree(head_buf);
1397     qemu_vfree(tail_buf);
1398 out:
1399     tracked_request_end(&req);
1400     return ret;
1401 }
1402
1403 static int coroutine_fn bdrv_co_do_writev(BlockDriverState *bs,
1404     int64_t sector_num, int nb_sectors, QEMUIOVector *qiov,
1405     BdrvRequestFlags flags)
1406 {
1407     if (nb_sectors < 0 || nb_sectors > BDRV_REQUEST_MAX_SECTORS) {
1408         return -EINVAL;
1409     }
1410
1411     return bdrv_co_do_pwritev(bs, sector_num << BDRV_SECTOR_BITS,
1412                               nb_sectors << BDRV_SECTOR_BITS, qiov, flags);
1413 }
1414
1415 int coroutine_fn bdrv_co_writev(BlockDriverState *bs, int64_t sector_num,
1416     int nb_sectors, QEMUIOVector *qiov)
1417 {
1418     trace_bdrv_co_writev(bs, sector_num, nb_sectors);
1419
1420     return bdrv_co_do_writev(bs, sector_num, nb_sectors, qiov, 0);
1421 }
1422
1423 int coroutine_fn bdrv_co_write_zeroes(BlockDriverState *bs,
1424                                       int64_t sector_num, int nb_sectors,
1425                                       BdrvRequestFlags flags)
1426 {
1427     trace_bdrv_co_write_zeroes(bs, sector_num, nb_sectors, flags);
1428
1429     if (!(bs->open_flags & BDRV_O_UNMAP)) {
1430         flags &= ~BDRV_REQ_MAY_UNMAP;
1431     }
1432
1433     return bdrv_co_do_writev(bs, sector_num, nb_sectors, NULL,
1434                              BDRV_REQ_ZERO_WRITE | flags);
1435 }
1436
1437 int bdrv_flush_all(void)
1438 {
1439     BlockDriverState *bs = NULL;
1440     int result = 0;
1441
1442     while ((bs = bdrv_next(bs))) {
1443         AioContext *aio_context = bdrv_get_aio_context(bs);
1444         int ret;
1445
1446         aio_context_acquire(aio_context);
1447         ret = bdrv_flush(bs);
1448         if (ret < 0 && !result) {
1449             result = ret;
1450         }
1451         aio_context_release(aio_context);
1452     }
1453
1454     return result;
1455 }
1456
1457 typedef struct BdrvCoGetBlockStatusData {
1458     BlockDriverState *bs;
1459     BlockDriverState *base;
1460     int64_t sector_num;
1461     int nb_sectors;
1462     int *pnum;
1463     int64_t ret;
1464     bool done;
1465 } BdrvCoGetBlockStatusData;
1466
1467 /*
1468  * Returns the allocation status of the specified sectors.
1469  * Drivers not implementing the functionality are assumed to not support
1470  * backing files, hence all their sectors are reported as allocated.
1471  *
1472  * If 'sector_num' is beyond the end of the disk image the return value is 0
1473  * and 'pnum' is set to 0.
1474  *
1475  * 'pnum' is set to the number of sectors (including and immediately following
1476  * the specified sector) that are known to be in the same
1477  * allocated/unallocated state.
1478  *
1479  * 'nb_sectors' is the max value 'pnum' should be set to.  If nb_sectors goes
1480  * beyond the end of the disk image it will be clamped.
1481  */
1482 static int64_t coroutine_fn bdrv_co_get_block_status(BlockDriverState *bs,
1483                                                      int64_t sector_num,
1484                                                      int nb_sectors, int *pnum)
1485 {
1486     int64_t total_sectors;
1487     int64_t n;
1488     int64_t ret, ret2;
1489
1490     total_sectors = bdrv_nb_sectors(bs);
1491     if (total_sectors < 0) {
1492         return total_sectors;
1493     }
1494
1495     if (sector_num >= total_sectors) {
1496         *pnum = 0;
1497         return 0;
1498     }
1499
1500     n = total_sectors - sector_num;
1501     if (n < nb_sectors) {
1502         nb_sectors = n;
1503     }
1504
1505     if (!bs->drv->bdrv_co_get_block_status) {
1506         *pnum = nb_sectors;
1507         ret = BDRV_BLOCK_DATA | BDRV_BLOCK_ALLOCATED;
1508         if (bs->drv->protocol_name) {
1509             ret |= BDRV_BLOCK_OFFSET_VALID | (sector_num * BDRV_SECTOR_SIZE);
1510         }
1511         return ret;
1512     }
1513
1514     ret = bs->drv->bdrv_co_get_block_status(bs, sector_num, nb_sectors, pnum);
1515     if (ret < 0) {
1516         *pnum = 0;
1517         return ret;
1518     }
1519
1520     if (ret & BDRV_BLOCK_RAW) {
1521         assert(ret & BDRV_BLOCK_OFFSET_VALID);
1522         return bdrv_get_block_status(bs->file->bs, ret >> BDRV_SECTOR_BITS,
1523                                      *pnum, pnum);
1524     }
1525
1526     if (ret & (BDRV_BLOCK_DATA | BDRV_BLOCK_ZERO)) {
1527         ret |= BDRV_BLOCK_ALLOCATED;
1528     } else {
1529         if (bdrv_unallocated_blocks_are_zero(bs)) {
1530             ret |= BDRV_BLOCK_ZERO;
1531         } else if (bs->backing) {
1532             BlockDriverState *bs2 = bs->backing->bs;
1533             int64_t nb_sectors2 = bdrv_nb_sectors(bs2);
1534             if (nb_sectors2 >= 0 && sector_num >= nb_sectors2) {
1535                 ret |= BDRV_BLOCK_ZERO;
1536             }
1537         }
1538     }
1539
1540     if (bs->file &&
1541         (ret & BDRV_BLOCK_DATA) && !(ret & BDRV_BLOCK_ZERO) &&
1542         (ret & BDRV_BLOCK_OFFSET_VALID)) {
1543         int file_pnum;
1544
1545         ret2 = bdrv_co_get_block_status(bs->file->bs, ret >> BDRV_SECTOR_BITS,
1546                                         *pnum, &file_pnum);
1547         if (ret2 >= 0) {
1548             /* Ignore errors.  This is just providing extra information, it
1549              * is useful but not necessary.
1550              */
1551             if (!file_pnum) {
1552                 /* !file_pnum indicates an offset at or beyond the EOF; it is
1553                  * perfectly valid for the format block driver to point to such
1554                  * offsets, so catch it and mark everything as zero */
1555                 ret |= BDRV_BLOCK_ZERO;
1556             } else {
1557                 /* Limit request to the range reported by the protocol driver */
1558                 *pnum = file_pnum;
1559                 ret |= (ret2 & BDRV_BLOCK_ZERO);
1560             }
1561         }
1562     }
1563
1564     return ret;
1565 }
1566
1567 static int64_t coroutine_fn bdrv_co_get_block_status_above(BlockDriverState *bs,
1568         BlockDriverState *base,
1569         int64_t sector_num,
1570         int nb_sectors,
1571         int *pnum)
1572 {
1573     BlockDriverState *p;
1574     int64_t ret = 0;
1575
1576     assert(bs != base);
1577     for (p = bs; p != base; p = backing_bs(p)) {
1578         ret = bdrv_co_get_block_status(p, sector_num, nb_sectors, pnum);
1579         if (ret < 0 || ret & BDRV_BLOCK_ALLOCATED) {
1580             break;
1581         }
1582         /* [sector_num, pnum] unallocated on this layer, which could be only
1583          * the first part of [sector_num, nb_sectors].  */
1584         nb_sectors = MIN(nb_sectors, *pnum);
1585     }
1586     return ret;
1587 }
1588
1589 /* Coroutine wrapper for bdrv_get_block_status_above() */
1590 static void coroutine_fn bdrv_get_block_status_above_co_entry(void *opaque)
1591 {
1592     BdrvCoGetBlockStatusData *data = opaque;
1593
1594     data->ret = bdrv_co_get_block_status_above(data->bs, data->base,
1595                                                data->sector_num,
1596                                                data->nb_sectors,
1597                                                data->pnum);
1598     data->done = true;
1599 }
1600
1601 /*
1602  * Synchronous wrapper around bdrv_co_get_block_status_above().
1603  *
1604  * See bdrv_co_get_block_status_above() for details.
1605  */
1606 int64_t bdrv_get_block_status_above(BlockDriverState *bs,
1607                                     BlockDriverState *base,
1608                                     int64_t sector_num,
1609                                     int nb_sectors, int *pnum)
1610 {
1611     Coroutine *co;
1612     BdrvCoGetBlockStatusData data = {
1613         .bs = bs,
1614         .base = base,
1615         .sector_num = sector_num,
1616         .nb_sectors = nb_sectors,
1617         .pnum = pnum,
1618         .done = false,
1619     };
1620
1621     if (qemu_in_coroutine()) {
1622         /* Fast-path if already in coroutine context */
1623         bdrv_get_block_status_above_co_entry(&data);
1624     } else {
1625         AioContext *aio_context = bdrv_get_aio_context(bs);
1626
1627         co = qemu_coroutine_create(bdrv_get_block_status_above_co_entry);
1628         qemu_coroutine_enter(co, &data);
1629         while (!data.done) {
1630             aio_poll(aio_context, true);
1631         }
1632     }
1633     return data.ret;
1634 }
1635
1636 int64_t bdrv_get_block_status(BlockDriverState *bs,
1637                               int64_t sector_num,
1638                               int nb_sectors, int *pnum)
1639 {
1640     return bdrv_get_block_status_above(bs, backing_bs(bs),
1641                                        sector_num, nb_sectors, pnum);
1642 }
1643
1644 int coroutine_fn bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num,
1645                                    int nb_sectors, int *pnum)
1646 {
1647     int64_t ret = bdrv_get_block_status(bs, sector_num, nb_sectors, pnum);
1648     if (ret < 0) {
1649         return ret;
1650     }
1651     return !!(ret & BDRV_BLOCK_ALLOCATED);
1652 }
1653
1654 /*
1655  * Given an image chain: ... -> [BASE] -> [INTER1] -> [INTER2] -> [TOP]
1656  *
1657  * Return true if the given sector is allocated in any image between
1658  * BASE and TOP (inclusive).  BASE can be NULL to check if the given
1659  * sector is allocated in any image of the chain.  Return false otherwise.
1660  *
1661  * 'pnum' is set to the number of sectors (including and immediately following
1662  *  the specified sector) that are known to be in the same
1663  *  allocated/unallocated state.
1664  *
1665  */
1666 int bdrv_is_allocated_above(BlockDriverState *top,
1667                             BlockDriverState *base,
1668                             int64_t sector_num,
1669                             int nb_sectors, int *pnum)
1670 {
1671     BlockDriverState *intermediate;
1672     int ret, n = nb_sectors;
1673
1674     intermediate = top;
1675     while (intermediate && intermediate != base) {
1676         int pnum_inter;
1677         ret = bdrv_is_allocated(intermediate, sector_num, nb_sectors,
1678                                 &pnum_inter);
1679         if (ret < 0) {
1680             return ret;
1681         } else if (ret) {
1682             *pnum = pnum_inter;
1683             return 1;
1684         }
1685
1686         /*
1687          * [sector_num, nb_sectors] is unallocated on top but intermediate
1688          * might have
1689          *
1690          * [sector_num+x, nr_sectors] allocated.
1691          */
1692         if (n > pnum_inter &&
1693             (intermediate == top ||
1694              sector_num + pnum_inter < intermediate->total_sectors)) {
1695             n = pnum_inter;
1696         }
1697
1698         intermediate = backing_bs(intermediate);
1699     }
1700
1701     *pnum = n;
1702     return 0;
1703 }
1704
1705 int bdrv_write_compressed(BlockDriverState *bs, int64_t sector_num,
1706                           const uint8_t *buf, int nb_sectors)
1707 {
1708     BlockDriver *drv = bs->drv;
1709     int ret;
1710
1711     if (!drv) {
1712         return -ENOMEDIUM;
1713     }
1714     if (!drv->bdrv_write_compressed) {
1715         return -ENOTSUP;
1716     }
1717     ret = bdrv_check_request(bs, sector_num, nb_sectors);
1718     if (ret < 0) {
1719         return ret;
1720     }
1721
1722     assert(QLIST_EMPTY(&bs->dirty_bitmaps));
1723
1724     return drv->bdrv_write_compressed(bs, sector_num, buf, nb_sectors);
1725 }
1726
1727 int bdrv_save_vmstate(BlockDriverState *bs, const uint8_t *buf,
1728                       int64_t pos, int size)
1729 {
1730     QEMUIOVector qiov;
1731     struct iovec iov = {
1732         .iov_base   = (void *) buf,
1733         .iov_len    = size,
1734     };
1735
1736     qemu_iovec_init_external(&qiov, &iov, 1);
1737     return bdrv_writev_vmstate(bs, &qiov, pos);
1738 }
1739
1740 int bdrv_writev_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos)
1741 {
1742     BlockDriver *drv = bs->drv;
1743
1744     if (!drv) {
1745         return -ENOMEDIUM;
1746     } else if (drv->bdrv_save_vmstate) {
1747         return drv->bdrv_save_vmstate(bs, qiov, pos);
1748     } else if (bs->file) {
1749         return bdrv_writev_vmstate(bs->file->bs, qiov, pos);
1750     }
1751
1752     return -ENOTSUP;
1753 }
1754
1755 int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf,
1756                       int64_t pos, int size)
1757 {
1758     BlockDriver *drv = bs->drv;
1759     if (!drv)
1760         return -ENOMEDIUM;
1761     if (drv->bdrv_load_vmstate)
1762         return drv->bdrv_load_vmstate(bs, buf, pos, size);
1763     if (bs->file)
1764         return bdrv_load_vmstate(bs->file->bs, buf, pos, size);
1765     return -ENOTSUP;
1766 }
1767
1768 /**************************************************************/
1769 /* async I/Os */
1770
1771 BlockAIOCB *bdrv_aio_readv(BlockDriverState *bs, int64_t sector_num,
1772                            QEMUIOVector *qiov, int nb_sectors,
1773                            BlockCompletionFunc *cb, void *opaque)
1774 {
1775     trace_bdrv_aio_readv(bs, sector_num, nb_sectors, opaque);
1776
1777     return bdrv_co_aio_rw_vector(bs, sector_num, qiov, nb_sectors, 0,
1778                                  cb, opaque, false);
1779 }
1780
1781 BlockAIOCB *bdrv_aio_writev(BlockDriverState *bs, int64_t sector_num,
1782                             QEMUIOVector *qiov, int nb_sectors,
1783                             BlockCompletionFunc *cb, void *opaque)
1784 {
1785     trace_bdrv_aio_writev(bs, sector_num, nb_sectors, opaque);
1786
1787     return bdrv_co_aio_rw_vector(bs, sector_num, qiov, nb_sectors, 0,
1788                                  cb, opaque, true);
1789 }
1790
1791 BlockAIOCB *bdrv_aio_write_zeroes(BlockDriverState *bs,
1792         int64_t sector_num, int nb_sectors, BdrvRequestFlags flags,
1793         BlockCompletionFunc *cb, void *opaque)
1794 {
1795     trace_bdrv_aio_write_zeroes(bs, sector_num, nb_sectors, flags, opaque);
1796
1797     return bdrv_co_aio_rw_vector(bs, sector_num, NULL, nb_sectors,
1798                                  BDRV_REQ_ZERO_WRITE | flags,
1799                                  cb, opaque, true);
1800 }
1801
1802
1803 typedef struct MultiwriteCB {
1804     int error;
1805     int num_requests;
1806     int num_callbacks;
1807     struct {
1808         BlockCompletionFunc *cb;
1809         void *opaque;
1810         QEMUIOVector *free_qiov;
1811     } callbacks[];
1812 } MultiwriteCB;
1813
1814 static void multiwrite_user_cb(MultiwriteCB *mcb)
1815 {
1816     int i;
1817
1818     for (i = 0; i < mcb->num_callbacks; i++) {
1819         mcb->callbacks[i].cb(mcb->callbacks[i].opaque, mcb->error);
1820         if (mcb->callbacks[i].free_qiov) {
1821             qemu_iovec_destroy(mcb->callbacks[i].free_qiov);
1822         }
1823         g_free(mcb->callbacks[i].free_qiov);
1824     }
1825 }
1826
1827 static void multiwrite_cb(void *opaque, int ret)
1828 {
1829     MultiwriteCB *mcb = opaque;
1830
1831     trace_multiwrite_cb(mcb, ret);
1832
1833     if (ret < 0 && !mcb->error) {
1834         mcb->error = ret;
1835     }
1836
1837     mcb->num_requests--;
1838     if (mcb->num_requests == 0) {
1839         multiwrite_user_cb(mcb);
1840         g_free(mcb);
1841     }
1842 }
1843
1844 static int multiwrite_req_compare(const void *a, const void *b)
1845 {
1846     const BlockRequest *req1 = a, *req2 = b;
1847
1848     /*
1849      * Note that we can't simply subtract req2->sector from req1->sector
1850      * here as that could overflow the return value.
1851      */
1852     if (req1->sector > req2->sector) {
1853         return 1;
1854     } else if (req1->sector < req2->sector) {
1855         return -1;
1856     } else {
1857         return 0;
1858     }
1859 }
1860
1861 /*
1862  * Takes a bunch of requests and tries to merge them. Returns the number of
1863  * requests that remain after merging.
1864  */
1865 static int multiwrite_merge(BlockDriverState *bs, BlockRequest *reqs,
1866     int num_reqs, MultiwriteCB *mcb)
1867 {
1868     int i, outidx;
1869
1870     // Sort requests by start sector
1871     qsort(reqs, num_reqs, sizeof(*reqs), &multiwrite_req_compare);
1872
1873     // Check if adjacent requests touch the same clusters. If so, combine them,
1874     // filling up gaps with zero sectors.
1875     outidx = 0;
1876     for (i = 1; i < num_reqs; i++) {
1877         int merge = 0;
1878         int64_t oldreq_last = reqs[outidx].sector + reqs[outidx].nb_sectors;
1879
1880         // Handle exactly sequential writes and overlapping writes.
1881         if (reqs[i].sector <= oldreq_last) {
1882             merge = 1;
1883         }
1884
1885         if (reqs[outidx].qiov->niov + reqs[i].qiov->niov + 1 > IOV_MAX) {
1886             merge = 0;
1887         }
1888
1889         if (bs->bl.max_transfer_length && reqs[outidx].nb_sectors +
1890             reqs[i].nb_sectors > bs->bl.max_transfer_length) {
1891             merge = 0;
1892         }
1893
1894         if (merge) {
1895             size_t size;
1896             QEMUIOVector *qiov = g_malloc0(sizeof(*qiov));
1897             qemu_iovec_init(qiov,
1898                 reqs[outidx].qiov->niov + reqs[i].qiov->niov + 1);
1899
1900             // Add the first request to the merged one. If the requests are
1901             // overlapping, drop the last sectors of the first request.
1902             size = (reqs[i].sector - reqs[outidx].sector) << 9;
1903             qemu_iovec_concat(qiov, reqs[outidx].qiov, 0, size);
1904
1905             // We should need to add any zeros between the two requests
1906             assert (reqs[i].sector <= oldreq_last);
1907
1908             // Add the second request
1909             qemu_iovec_concat(qiov, reqs[i].qiov, 0, reqs[i].qiov->size);
1910
1911             // Add tail of first request, if necessary
1912             if (qiov->size < reqs[outidx].qiov->size) {
1913                 qemu_iovec_concat(qiov, reqs[outidx].qiov, qiov->size,
1914                                   reqs[outidx].qiov->size - qiov->size);
1915             }
1916
1917             reqs[outidx].nb_sectors = qiov->size >> 9;
1918             reqs[outidx].qiov = qiov;
1919
1920             mcb->callbacks[i].free_qiov = reqs[outidx].qiov;
1921         } else {
1922             outidx++;
1923             reqs[outidx].sector     = reqs[i].sector;
1924             reqs[outidx].nb_sectors = reqs[i].nb_sectors;
1925             reqs[outidx].qiov       = reqs[i].qiov;
1926         }
1927     }
1928
1929     if (bs->blk) {
1930         block_acct_merge_done(blk_get_stats(bs->blk), BLOCK_ACCT_WRITE,
1931                               num_reqs - outidx - 1);
1932     }
1933
1934     return outidx + 1;
1935 }
1936
1937 /*
1938  * Submit multiple AIO write requests at once.
1939  *
1940  * On success, the function returns 0 and all requests in the reqs array have
1941  * been submitted. In error case this function returns -1, and any of the
1942  * requests may or may not be submitted yet. In particular, this means that the
1943  * callback will be called for some of the requests, for others it won't. The
1944  * caller must check the error field of the BlockRequest to wait for the right
1945  * callbacks (if error != 0, no callback will be called).
1946  *
1947  * The implementation may modify the contents of the reqs array, e.g. to merge
1948  * requests. However, the fields opaque and error are left unmodified as they
1949  * are used to signal failure for a single request to the caller.
1950  */
1951 int bdrv_aio_multiwrite(BlockDriverState *bs, BlockRequest *reqs, int num_reqs)
1952 {
1953     MultiwriteCB *mcb;
1954     int i;
1955
1956     /* don't submit writes if we don't have a medium */
1957     if (bs->drv == NULL) {
1958         for (i = 0; i < num_reqs; i++) {
1959             reqs[i].error = -ENOMEDIUM;
1960         }
1961         return -1;
1962     }
1963
1964     if (num_reqs == 0) {
1965         return 0;
1966     }
1967
1968     // Create MultiwriteCB structure
1969     mcb = g_malloc0(sizeof(*mcb) + num_reqs * sizeof(*mcb->callbacks));
1970     mcb->num_requests = 0;
1971     mcb->num_callbacks = num_reqs;
1972
1973     for (i = 0; i < num_reqs; i++) {
1974         mcb->callbacks[i].cb = reqs[i].cb;
1975         mcb->callbacks[i].opaque = reqs[i].opaque;
1976     }
1977
1978     // Check for mergable requests
1979     num_reqs = multiwrite_merge(bs, reqs, num_reqs, mcb);
1980
1981     trace_bdrv_aio_multiwrite(mcb, mcb->num_callbacks, num_reqs);
1982
1983     /* Run the aio requests. */
1984     mcb->num_requests = num_reqs;
1985     for (i = 0; i < num_reqs; i++) {
1986         bdrv_co_aio_rw_vector(bs, reqs[i].sector, reqs[i].qiov,
1987                               reqs[i].nb_sectors, reqs[i].flags,
1988                               multiwrite_cb, mcb,
1989                               true);
1990     }
1991
1992     return 0;
1993 }
1994
1995 void bdrv_aio_cancel(BlockAIOCB *acb)
1996 {
1997     qemu_aio_ref(acb);
1998     bdrv_aio_cancel_async(acb);
1999     while (acb->refcnt > 1) {
2000         if (acb->aiocb_info->get_aio_context) {
2001             aio_poll(acb->aiocb_info->get_aio_context(acb), true);
2002         } else if (acb->bs) {
2003             aio_poll(bdrv_get_aio_context(acb->bs), true);
2004         } else {
2005             abort();
2006         }
2007     }
2008     qemu_aio_unref(acb);
2009 }
2010
2011 /* Async version of aio cancel. The caller is not blocked if the acb implements
2012  * cancel_async, otherwise we do nothing and let the request normally complete.
2013  * In either case the completion callback must be called. */
2014 void bdrv_aio_cancel_async(BlockAIOCB *acb)
2015 {
2016     if (acb->aiocb_info->cancel_async) {
2017         acb->aiocb_info->cancel_async(acb);
2018     }
2019 }
2020
2021 /**************************************************************/
2022 /* async block device emulation */
2023
2024 typedef struct BlockAIOCBSync {
2025     BlockAIOCB common;
2026     QEMUBH *bh;
2027     int ret;
2028     /* vector translation state */
2029     QEMUIOVector *qiov;
2030     uint8_t *bounce;
2031     int is_write;
2032 } BlockAIOCBSync;
2033
2034 static const AIOCBInfo bdrv_em_aiocb_info = {
2035     .aiocb_size         = sizeof(BlockAIOCBSync),
2036 };
2037
2038 static void bdrv_aio_bh_cb(void *opaque)
2039 {
2040     BlockAIOCBSync *acb = opaque;
2041
2042     if (!acb->is_write && acb->ret >= 0) {
2043         qemu_iovec_from_buf(acb->qiov, 0, acb->bounce, acb->qiov->size);
2044     }
2045     qemu_vfree(acb->bounce);
2046     acb->common.cb(acb->common.opaque, acb->ret);
2047     qemu_bh_delete(acb->bh);
2048     acb->bh = NULL;
2049     qemu_aio_unref(acb);
2050 }
2051
2052 static BlockAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
2053                                       int64_t sector_num,
2054                                       QEMUIOVector *qiov,
2055                                       int nb_sectors,
2056                                       BlockCompletionFunc *cb,
2057                                       void *opaque,
2058                                       int is_write)
2059
2060 {
2061     BlockAIOCBSync *acb;
2062
2063     acb = qemu_aio_get(&bdrv_em_aiocb_info, bs, cb, opaque);
2064     acb->is_write = is_write;
2065     acb->qiov = qiov;
2066     acb->bounce = qemu_try_blockalign(bs, qiov->size);
2067     acb->bh = aio_bh_new(bdrv_get_aio_context(bs), bdrv_aio_bh_cb, acb);
2068
2069     if (acb->bounce == NULL) {
2070         acb->ret = -ENOMEM;
2071     } else if (is_write) {
2072         qemu_iovec_to_buf(acb->qiov, 0, acb->bounce, qiov->size);
2073         acb->ret = bs->drv->bdrv_write(bs, sector_num, acb->bounce, nb_sectors);
2074     } else {
2075         acb->ret = bs->drv->bdrv_read(bs, sector_num, acb->bounce, nb_sectors);
2076     }
2077
2078     qemu_bh_schedule(acb->bh);
2079
2080     return &acb->common;
2081 }
2082
2083 static BlockAIOCB *bdrv_aio_readv_em(BlockDriverState *bs,
2084         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
2085         BlockCompletionFunc *cb, void *opaque)
2086 {
2087     return bdrv_aio_rw_vector(bs, sector_num, qiov, nb_sectors, cb, opaque, 0);
2088 }
2089
2090 static BlockAIOCB *bdrv_aio_writev_em(BlockDriverState *bs,
2091         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
2092         BlockCompletionFunc *cb, void *opaque)
2093 {
2094     return bdrv_aio_rw_vector(bs, sector_num, qiov, nb_sectors, cb, opaque, 1);
2095 }
2096
2097
2098 typedef struct BlockAIOCBCoroutine {
2099     BlockAIOCB common;
2100     BlockRequest req;
2101     bool is_write;
2102     bool need_bh;
2103     bool *done;
2104     QEMUBH* bh;
2105 } BlockAIOCBCoroutine;
2106
2107 static const AIOCBInfo bdrv_em_co_aiocb_info = {
2108     .aiocb_size         = sizeof(BlockAIOCBCoroutine),
2109 };
2110
2111 static void bdrv_co_complete(BlockAIOCBCoroutine *acb)
2112 {
2113     if (!acb->need_bh) {
2114         acb->common.cb(acb->common.opaque, acb->req.error);
2115         qemu_aio_unref(acb);
2116     }
2117 }
2118
2119 static void bdrv_co_em_bh(void *opaque)
2120 {
2121     BlockAIOCBCoroutine *acb = opaque;
2122
2123     assert(!acb->need_bh);
2124     qemu_bh_delete(acb->bh);
2125     bdrv_co_complete(acb);
2126 }
2127
2128 static void bdrv_co_maybe_schedule_bh(BlockAIOCBCoroutine *acb)
2129 {
2130     acb->need_bh = false;
2131     if (acb->req.error != -EINPROGRESS) {
2132         BlockDriverState *bs = acb->common.bs;
2133
2134         acb->bh = aio_bh_new(bdrv_get_aio_context(bs), bdrv_co_em_bh, acb);
2135         qemu_bh_schedule(acb->bh);
2136     }
2137 }
2138
2139 /* Invoke bdrv_co_do_readv/bdrv_co_do_writev */
2140 static void coroutine_fn bdrv_co_do_rw(void *opaque)
2141 {
2142     BlockAIOCBCoroutine *acb = opaque;
2143     BlockDriverState *bs = acb->common.bs;
2144
2145     if (!acb->is_write) {
2146         acb->req.error = bdrv_co_do_readv(bs, acb->req.sector,
2147             acb->req.nb_sectors, acb->req.qiov, acb->req.flags);
2148     } else {
2149         acb->req.error = bdrv_co_do_writev(bs, acb->req.sector,
2150             acb->req.nb_sectors, acb->req.qiov, acb->req.flags);
2151     }
2152
2153     bdrv_co_complete(acb);
2154 }
2155
2156 static BlockAIOCB *bdrv_co_aio_rw_vector(BlockDriverState *bs,
2157                                          int64_t sector_num,
2158                                          QEMUIOVector *qiov,
2159                                          int nb_sectors,
2160                                          BdrvRequestFlags flags,
2161                                          BlockCompletionFunc *cb,
2162                                          void *opaque,
2163                                          bool is_write)
2164 {
2165     Coroutine *co;
2166     BlockAIOCBCoroutine *acb;
2167
2168     acb = qemu_aio_get(&bdrv_em_co_aiocb_info, bs, cb, opaque);
2169     acb->need_bh = true;
2170     acb->req.error = -EINPROGRESS;
2171     acb->req.sector = sector_num;
2172     acb->req.nb_sectors = nb_sectors;
2173     acb->req.qiov = qiov;
2174     acb->req.flags = flags;
2175     acb->is_write = is_write;
2176
2177     co = qemu_coroutine_create(bdrv_co_do_rw);
2178     qemu_coroutine_enter(co, acb);
2179
2180     bdrv_co_maybe_schedule_bh(acb);
2181     return &acb->common;
2182 }
2183
2184 static void coroutine_fn bdrv_aio_flush_co_entry(void *opaque)
2185 {
2186     BlockAIOCBCoroutine *acb = opaque;
2187     BlockDriverState *bs = acb->common.bs;
2188
2189     acb->req.error = bdrv_co_flush(bs);
2190     bdrv_co_complete(acb);
2191 }
2192
2193 BlockAIOCB *bdrv_aio_flush(BlockDriverState *bs,
2194         BlockCompletionFunc *cb, void *opaque)
2195 {
2196     trace_bdrv_aio_flush(bs, opaque);
2197
2198     Coroutine *co;
2199     BlockAIOCBCoroutine *acb;
2200
2201     acb = qemu_aio_get(&bdrv_em_co_aiocb_info, bs, cb, opaque);
2202     acb->need_bh = true;
2203     acb->req.error = -EINPROGRESS;
2204
2205     co = qemu_coroutine_create(bdrv_aio_flush_co_entry);
2206     qemu_coroutine_enter(co, acb);
2207
2208     bdrv_co_maybe_schedule_bh(acb);
2209     return &acb->common;
2210 }
2211
2212 static void coroutine_fn bdrv_aio_discard_co_entry(void *opaque)
2213 {
2214     BlockAIOCBCoroutine *acb = opaque;
2215     BlockDriverState *bs = acb->common.bs;
2216
2217     acb->req.error = bdrv_co_discard(bs, acb->req.sector, acb->req.nb_sectors);
2218     bdrv_co_complete(acb);
2219 }
2220
2221 BlockAIOCB *bdrv_aio_discard(BlockDriverState *bs,
2222         int64_t sector_num, int nb_sectors,
2223         BlockCompletionFunc *cb, void *opaque)
2224 {
2225     Coroutine *co;
2226     BlockAIOCBCoroutine *acb;
2227
2228     trace_bdrv_aio_discard(bs, sector_num, nb_sectors, opaque);
2229
2230     acb = qemu_aio_get(&bdrv_em_co_aiocb_info, bs, cb, opaque);
2231     acb->need_bh = true;
2232     acb->req.error = -EINPROGRESS;
2233     acb->req.sector = sector_num;
2234     acb->req.nb_sectors = nb_sectors;
2235     co = qemu_coroutine_create(bdrv_aio_discard_co_entry);
2236     qemu_coroutine_enter(co, acb);
2237
2238     bdrv_co_maybe_schedule_bh(acb);
2239     return &acb->common;
2240 }
2241
2242 void *qemu_aio_get(const AIOCBInfo *aiocb_info, BlockDriverState *bs,
2243                    BlockCompletionFunc *cb, void *opaque)
2244 {
2245     BlockAIOCB *acb;
2246
2247     acb = g_malloc(aiocb_info->aiocb_size);
2248     acb->aiocb_info = aiocb_info;
2249     acb->bs = bs;
2250     acb->cb = cb;
2251     acb->opaque = opaque;
2252     acb->refcnt = 1;
2253     return acb;
2254 }
2255
2256 void qemu_aio_ref(void *p)
2257 {
2258     BlockAIOCB *acb = p;
2259     acb->refcnt++;
2260 }
2261
2262 void qemu_aio_unref(void *p)
2263 {
2264     BlockAIOCB *acb = p;
2265     assert(acb->refcnt > 0);
2266     if (--acb->refcnt == 0) {
2267         g_free(acb);
2268     }
2269 }
2270
2271 /**************************************************************/
2272 /* Coroutine block device emulation */
2273
2274 typedef struct CoroutineIOCompletion {
2275     Coroutine *coroutine;
2276     int ret;
2277 } CoroutineIOCompletion;
2278
2279 static void bdrv_co_io_em_complete(void *opaque, int ret)
2280 {
2281     CoroutineIOCompletion *co = opaque;
2282
2283     co->ret = ret;
2284     qemu_coroutine_enter(co->coroutine, NULL);
2285 }
2286
2287 static int coroutine_fn bdrv_co_io_em(BlockDriverState *bs, int64_t sector_num,
2288                                       int nb_sectors, QEMUIOVector *iov,
2289                                       bool is_write)
2290 {
2291     CoroutineIOCompletion co = {
2292         .coroutine = qemu_coroutine_self(),
2293     };
2294     BlockAIOCB *acb;
2295
2296     if (is_write) {
2297         acb = bs->drv->bdrv_aio_writev(bs, sector_num, iov, nb_sectors,
2298                                        bdrv_co_io_em_complete, &co);
2299     } else {
2300         acb = bs->drv->bdrv_aio_readv(bs, sector_num, iov, nb_sectors,
2301                                       bdrv_co_io_em_complete, &co);
2302     }
2303
2304     trace_bdrv_co_io_em(bs, sector_num, nb_sectors, is_write, acb);
2305     if (!acb) {
2306         return -EIO;
2307     }
2308     qemu_coroutine_yield();
2309
2310     return co.ret;
2311 }
2312
2313 static int coroutine_fn bdrv_co_readv_em(BlockDriverState *bs,
2314                                          int64_t sector_num, int nb_sectors,
2315                                          QEMUIOVector *iov)
2316 {
2317     return bdrv_co_io_em(bs, sector_num, nb_sectors, iov, false);
2318 }
2319
2320 static int coroutine_fn bdrv_co_writev_em(BlockDriverState *bs,
2321                                          int64_t sector_num, int nb_sectors,
2322                                          QEMUIOVector *iov)
2323 {
2324     return bdrv_co_io_em(bs, sector_num, nb_sectors, iov, true);
2325 }
2326
2327 static void coroutine_fn bdrv_flush_co_entry(void *opaque)
2328 {
2329     RwCo *rwco = opaque;
2330
2331     rwco->ret = bdrv_co_flush(rwco->bs);
2332 }
2333
2334 int coroutine_fn bdrv_co_flush(BlockDriverState *bs)
2335 {
2336     int ret;
2337     BdrvTrackedRequest req;
2338
2339     if (!bs || !bdrv_is_inserted(bs) || bdrv_is_read_only(bs) ||
2340         bdrv_is_sg(bs)) {
2341         return 0;
2342     }
2343
2344     tracked_request_begin(&req, bs, 0, 0, BDRV_TRACKED_FLUSH);
2345     /* Write back cached data to the OS even with cache=unsafe */
2346     BLKDBG_EVENT(bs->file, BLKDBG_FLUSH_TO_OS);
2347     if (bs->drv->bdrv_co_flush_to_os) {
2348         ret = bs->drv->bdrv_co_flush_to_os(bs);
2349         if (ret < 0) {
2350             goto out;
2351         }
2352     }
2353
2354     /* But don't actually force it to the disk with cache=unsafe */
2355     if (bs->open_flags & BDRV_O_NO_FLUSH) {
2356         goto flush_parent;
2357     }
2358
2359     BLKDBG_EVENT(bs->file, BLKDBG_FLUSH_TO_DISK);
2360     if (bs->drv->bdrv_co_flush_to_disk) {
2361         ret = bs->drv->bdrv_co_flush_to_disk(bs);
2362     } else if (bs->drv->bdrv_aio_flush) {
2363         BlockAIOCB *acb;
2364         CoroutineIOCompletion co = {
2365             .coroutine = qemu_coroutine_self(),
2366         };
2367
2368         acb = bs->drv->bdrv_aio_flush(bs, bdrv_co_io_em_complete, &co);
2369         if (acb == NULL) {
2370             ret = -EIO;
2371         } else {
2372             qemu_coroutine_yield();
2373             ret = co.ret;
2374         }
2375     } else {
2376         /*
2377          * Some block drivers always operate in either writethrough or unsafe
2378          * mode and don't support bdrv_flush therefore. Usually qemu doesn't
2379          * know how the server works (because the behaviour is hardcoded or
2380          * depends on server-side configuration), so we can't ensure that
2381          * everything is safe on disk. Returning an error doesn't work because
2382          * that would break guests even if the server operates in writethrough
2383          * mode.
2384          *
2385          * Let's hope the user knows what he's doing.
2386          */
2387         ret = 0;
2388     }
2389     if (ret < 0) {
2390         goto out;
2391     }
2392
2393     /* Now flush the underlying protocol.  It will also have BDRV_O_NO_FLUSH
2394      * in the case of cache=unsafe, so there are no useless flushes.
2395      */
2396 flush_parent:
2397     ret = bs->file ? bdrv_co_flush(bs->file->bs) : 0;
2398 out:
2399     tracked_request_end(&req);
2400     return ret;
2401 }
2402
2403 int bdrv_flush(BlockDriverState *bs)
2404 {
2405     Coroutine *co;
2406     RwCo rwco = {
2407         .bs = bs,
2408         .ret = NOT_DONE,
2409     };
2410
2411     if (qemu_in_coroutine()) {
2412         /* Fast-path if already in coroutine context */
2413         bdrv_flush_co_entry(&rwco);
2414     } else {
2415         AioContext *aio_context = bdrv_get_aio_context(bs);
2416
2417         co = qemu_coroutine_create(bdrv_flush_co_entry);
2418         qemu_coroutine_enter(co, &rwco);
2419         while (rwco.ret == NOT_DONE) {
2420             aio_poll(aio_context, true);
2421         }
2422     }
2423
2424     return rwco.ret;
2425 }
2426
2427 typedef struct DiscardCo {
2428     BlockDriverState *bs;
2429     int64_t sector_num;
2430     int nb_sectors;
2431     int ret;
2432 } DiscardCo;
2433 static void coroutine_fn bdrv_discard_co_entry(void *opaque)
2434 {
2435     DiscardCo *rwco = opaque;
2436
2437     rwco->ret = bdrv_co_discard(rwco->bs, rwco->sector_num, rwco->nb_sectors);
2438 }
2439
2440 int coroutine_fn bdrv_co_discard(BlockDriverState *bs, int64_t sector_num,
2441                                  int nb_sectors)
2442 {
2443     BdrvTrackedRequest req;
2444     int max_discard, ret;
2445
2446     if (!bs->drv) {
2447         return -ENOMEDIUM;
2448     }
2449
2450     ret = bdrv_check_request(bs, sector_num, nb_sectors);
2451     if (ret < 0) {
2452         return ret;
2453     } else if (bs->read_only) {
2454         return -EPERM;
2455     }
2456
2457     /* Do nothing if disabled.  */
2458     if (!(bs->open_flags & BDRV_O_UNMAP)) {
2459         return 0;
2460     }
2461
2462     if (!bs->drv->bdrv_co_discard && !bs->drv->bdrv_aio_discard) {
2463         return 0;
2464     }
2465
2466     tracked_request_begin(&req, bs, sector_num, nb_sectors,
2467                           BDRV_TRACKED_DISCARD);
2468     bdrv_set_dirty(bs, sector_num, nb_sectors);
2469
2470     max_discard = MIN_NON_ZERO(bs->bl.max_discard, BDRV_REQUEST_MAX_SECTORS);
2471     while (nb_sectors > 0) {
2472         int ret;
2473         int num = nb_sectors;
2474
2475         /* align request */
2476         if (bs->bl.discard_alignment &&
2477             num >= bs->bl.discard_alignment &&
2478             sector_num % bs->bl.discard_alignment) {
2479             if (num > bs->bl.discard_alignment) {
2480                 num = bs->bl.discard_alignment;
2481             }
2482             num -= sector_num % bs->bl.discard_alignment;
2483         }
2484
2485         /* limit request size */
2486         if (num > max_discard) {
2487             num = max_discard;
2488         }
2489
2490         if (bs->drv->bdrv_co_discard) {
2491             ret = bs->drv->bdrv_co_discard(bs, sector_num, num);
2492         } else {
2493             BlockAIOCB *acb;
2494             CoroutineIOCompletion co = {
2495                 .coroutine = qemu_coroutine_self(),
2496             };
2497
2498             acb = bs->drv->bdrv_aio_discard(bs, sector_num, nb_sectors,
2499                                             bdrv_co_io_em_complete, &co);
2500             if (acb == NULL) {
2501                 ret = -EIO;
2502                 goto out;
2503             } else {
2504                 qemu_coroutine_yield();
2505                 ret = co.ret;
2506             }
2507         }
2508         if (ret && ret != -ENOTSUP) {
2509             goto out;
2510         }
2511
2512         sector_num += num;
2513         nb_sectors -= num;
2514     }
2515     ret = 0;
2516 out:
2517     tracked_request_end(&req);
2518     return ret;
2519 }
2520
2521 int bdrv_discard(BlockDriverState *bs, int64_t sector_num, int nb_sectors)
2522 {
2523     Coroutine *co;
2524     DiscardCo rwco = {
2525         .bs = bs,
2526         .sector_num = sector_num,
2527         .nb_sectors = nb_sectors,
2528         .ret = NOT_DONE,
2529     };
2530
2531     if (qemu_in_coroutine()) {
2532         /* Fast-path if already in coroutine context */
2533         bdrv_discard_co_entry(&rwco);
2534     } else {
2535         AioContext *aio_context = bdrv_get_aio_context(bs);
2536
2537         co = qemu_coroutine_create(bdrv_discard_co_entry);
2538         qemu_coroutine_enter(co, &rwco);
2539         while (rwco.ret == NOT_DONE) {
2540             aio_poll(aio_context, true);
2541         }
2542     }
2543
2544     return rwco.ret;
2545 }
2546
2547 typedef struct {
2548     CoroutineIOCompletion *co;
2549     QEMUBH *bh;
2550 } BdrvIoctlCompletionData;
2551
2552 static void bdrv_ioctl_bh_cb(void *opaque)
2553 {
2554     BdrvIoctlCompletionData *data = opaque;
2555
2556     bdrv_co_io_em_complete(data->co, -ENOTSUP);
2557     qemu_bh_delete(data->bh);
2558 }
2559
2560 static int bdrv_co_do_ioctl(BlockDriverState *bs, int req, void *buf)
2561 {
2562     BlockDriver *drv = bs->drv;
2563     BdrvTrackedRequest tracked_req;
2564     CoroutineIOCompletion co = {
2565         .coroutine = qemu_coroutine_self(),
2566     };
2567     BlockAIOCB *acb;
2568
2569     tracked_request_begin(&tracked_req, bs, 0, 0, BDRV_TRACKED_IOCTL);
2570     if (!drv || !drv->bdrv_aio_ioctl) {
2571         co.ret = -ENOTSUP;
2572         goto out;
2573     }
2574
2575     acb = drv->bdrv_aio_ioctl(bs, req, buf, bdrv_co_io_em_complete, &co);
2576     if (!acb) {
2577         BdrvIoctlCompletionData *data = g_new(BdrvIoctlCompletionData, 1);
2578         data->bh = aio_bh_new(bdrv_get_aio_context(bs),
2579                                 bdrv_ioctl_bh_cb, data);
2580         data->co = &co;
2581         qemu_bh_schedule(data->bh);
2582     }
2583     qemu_coroutine_yield();
2584 out:
2585     tracked_request_end(&tracked_req);
2586     return co.ret;
2587 }
2588
2589 typedef struct {
2590     BlockDriverState *bs;
2591     int req;
2592     void *buf;
2593     int ret;
2594 } BdrvIoctlCoData;
2595
2596 static void coroutine_fn bdrv_co_ioctl_entry(void *opaque)
2597 {
2598     BdrvIoctlCoData *data = opaque;
2599     data->ret = bdrv_co_do_ioctl(data->bs, data->req, data->buf);
2600 }
2601
2602 /* needed for generic scsi interface */
2603 int bdrv_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
2604 {
2605     BdrvIoctlCoData data = {
2606         .bs = bs,
2607         .req = req,
2608         .buf = buf,
2609         .ret = -EINPROGRESS,
2610     };
2611
2612     if (qemu_in_coroutine()) {
2613         /* Fast-path if already in coroutine context */
2614         bdrv_co_ioctl_entry(&data);
2615     } else {
2616         Coroutine *co = qemu_coroutine_create(bdrv_co_ioctl_entry);
2617
2618         qemu_coroutine_enter(co, &data);
2619         while (data.ret == -EINPROGRESS) {
2620             aio_poll(bdrv_get_aio_context(bs), true);
2621         }
2622     }
2623     return data.ret;
2624 }
2625
2626 static void coroutine_fn bdrv_co_aio_ioctl_entry(void *opaque)
2627 {
2628     BlockAIOCBCoroutine *acb = opaque;
2629     acb->req.error = bdrv_co_do_ioctl(acb->common.bs,
2630                                       acb->req.req, acb->req.buf);
2631     bdrv_co_complete(acb);
2632 }
2633
2634 BlockAIOCB *bdrv_aio_ioctl(BlockDriverState *bs,
2635         unsigned long int req, void *buf,
2636         BlockCompletionFunc *cb, void *opaque)
2637 {
2638     BlockAIOCBCoroutine *acb = qemu_aio_get(&bdrv_em_co_aiocb_info,
2639                                             bs, cb, opaque);
2640     Coroutine *co;
2641
2642     acb->need_bh = true;
2643     acb->req.error = -EINPROGRESS;
2644     acb->req.req = req;
2645     acb->req.buf = buf;
2646     co = qemu_coroutine_create(bdrv_co_aio_ioctl_entry);
2647     qemu_coroutine_enter(co, acb);
2648
2649     bdrv_co_maybe_schedule_bh(acb);
2650     return &acb->common;
2651 }
2652
2653 void *qemu_blockalign(BlockDriverState *bs, size_t size)
2654 {
2655     return qemu_memalign(bdrv_opt_mem_align(bs), size);
2656 }
2657
2658 void *qemu_blockalign0(BlockDriverState *bs, size_t size)
2659 {
2660     return memset(qemu_blockalign(bs, size), 0, size);
2661 }
2662
2663 void *qemu_try_blockalign(BlockDriverState *bs, size_t size)
2664 {
2665     size_t align = bdrv_opt_mem_align(bs);
2666
2667     /* Ensure that NULL is never returned on success */
2668     assert(align > 0);
2669     if (size == 0) {
2670         size = align;
2671     }
2672
2673     return qemu_try_memalign(align, size);
2674 }
2675
2676 void *qemu_try_blockalign0(BlockDriverState *bs, size_t size)
2677 {
2678     void *mem = qemu_try_blockalign(bs, size);
2679
2680     if (mem) {
2681         memset(mem, 0, size);
2682     }
2683
2684     return mem;
2685 }
2686
2687 /*
2688  * Check if all memory in this vector is sector aligned.
2689  */
2690 bool bdrv_qiov_is_aligned(BlockDriverState *bs, QEMUIOVector *qiov)
2691 {
2692     int i;
2693     size_t alignment = bdrv_min_mem_align(bs);
2694
2695     for (i = 0; i < qiov->niov; i++) {
2696         if ((uintptr_t) qiov->iov[i].iov_base % alignment) {
2697             return false;
2698         }
2699         if (qiov->iov[i].iov_len % alignment) {
2700             return false;
2701         }
2702     }
2703
2704     return true;
2705 }
2706
2707 void bdrv_add_before_write_notifier(BlockDriverState *bs,
2708                                     NotifierWithReturn *notifier)
2709 {
2710     notifier_with_return_list_add(&bs->before_write_notifiers, notifier);
2711 }
2712
2713 void bdrv_io_plug(BlockDriverState *bs)
2714 {
2715     BlockDriver *drv = bs->drv;
2716     if (drv && drv->bdrv_io_plug) {
2717         drv->bdrv_io_plug(bs);
2718     } else if (bs->file) {
2719         bdrv_io_plug(bs->file->bs);
2720     }
2721 }
2722
2723 void bdrv_io_unplug(BlockDriverState *bs)
2724 {
2725     BlockDriver *drv = bs->drv;
2726     if (drv && drv->bdrv_io_unplug) {
2727         drv->bdrv_io_unplug(bs);
2728     } else if (bs->file) {
2729         bdrv_io_unplug(bs->file->bs);
2730     }
2731 }
2732
2733 void bdrv_flush_io_queue(BlockDriverState *bs)
2734 {
2735     BlockDriver *drv = bs->drv;
2736     if (drv && drv->bdrv_flush_io_queue) {
2737         drv->bdrv_flush_io_queue(bs);
2738     } else if (bs->file) {
2739         bdrv_flush_io_queue(bs->file->bs);
2740     }
2741     bdrv_start_throttled_reqs(bs);
2742 }
2743
2744 void bdrv_drained_begin(BlockDriverState *bs)
2745 {
2746     if (!bs->quiesce_counter++) {
2747         aio_disable_external(bdrv_get_aio_context(bs));
2748     }
2749     bdrv_drain(bs);
2750 }
2751
2752 void bdrv_drained_end(BlockDriverState *bs)
2753 {
2754     assert(bs->quiesce_counter > 0);
2755     if (--bs->quiesce_counter > 0) {
2756         return;
2757     }
2758     aio_enable_external(bdrv_get_aio_context(bs));
2759 }