1 /* 2 * Block layer I/O functions 3 * 4 * Copyright (c) 2003 Fabrice Bellard 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a copy 7 * of this software and associated documentation files (the "Software"), to deal 8 * in the Software without restriction, including without limitation the rights 9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 10 * copies of the Software, and to permit persons to whom the Software is 11 * furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice shall be included in 14 * all copies or substantial portions of the Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 22 * THE SOFTWARE. 23 */ 24 25 #include "qemu/osdep.h" 26 #include "trace.h" 27 #include "sysemu/block-backend.h" 28 #include "block/aio-wait.h" 29 #include "block/blockjob.h" 30 #include "block/blockjob_int.h" 31 #include "block/block_int.h" 32 #include "block/coroutines.h" 33 #include "block/dirty-bitmap.h" 34 #include "block/write-threshold.h" 35 #include "qemu/cutils.h" 36 #include "qemu/memalign.h" 37 #include "qapi/error.h" 38 #include "qemu/error-report.h" 39 #include "qemu/main-loop.h" 40 #include "sysemu/replay.h" 41 42 /* Maximum bounce buffer for copy-on-read and write zeroes, in bytes */ 43 #define MAX_BOUNCE_BUFFER (32768 << BDRV_SECTOR_BITS) 44 45 static void bdrv_parent_cb_resize(BlockDriverState *bs); 46 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs, 47 int64_t offset, int64_t bytes, BdrvRequestFlags flags); 48 49 static void bdrv_parent_drained_begin(BlockDriverState *bs, BdrvChild *ignore) 50 { 51 BdrvChild *c, *next; 52 53 QLIST_FOREACH_SAFE(c, &bs->parents, next_parent, next) { 54 if (c == ignore) { 55 continue; 56 } 57 bdrv_parent_drained_begin_single(c); 58 } 59 } 60 61 void bdrv_parent_drained_end_single(BdrvChild *c) 62 { 63 IO_OR_GS_CODE(); 64 65 assert(c->quiesced_parent); 66 c->quiesced_parent = false; 67 68 if (c->klass->drained_end) { 69 c->klass->drained_end(c); 70 } 71 } 72 73 static void bdrv_parent_drained_end(BlockDriverState *bs, BdrvChild *ignore) 74 { 75 BdrvChild *c; 76 77 QLIST_FOREACH(c, &bs->parents, next_parent) { 78 if (c == ignore) { 79 continue; 80 } 81 bdrv_parent_drained_end_single(c); 82 } 83 } 84 85 bool bdrv_parent_drained_poll_single(BdrvChild *c) 86 { 87 if (c->klass->drained_poll) { 88 return c->klass->drained_poll(c); 89 } 90 return false; 91 } 92 93 static bool bdrv_parent_drained_poll(BlockDriverState *bs, BdrvChild *ignore, 94 bool ignore_bds_parents) 95 { 96 BdrvChild *c, *next; 97 bool busy = false; 98 99 QLIST_FOREACH_SAFE(c, &bs->parents, next_parent, next) { 100 if (c == ignore || (ignore_bds_parents && c->klass->parent_is_bds)) { 101 continue; 102 } 103 busy |= bdrv_parent_drained_poll_single(c); 104 } 105 106 return busy; 107 } 108 109 void bdrv_parent_drained_begin_single(BdrvChild *c) 110 { 111 IO_OR_GS_CODE(); 112 113 assert(!c->quiesced_parent); 114 c->quiesced_parent = true; 115 116 if (c->klass->drained_begin) { 117 c->klass->drained_begin(c); 118 } 119 } 120 121 static void bdrv_merge_limits(BlockLimits *dst, const BlockLimits *src) 122 { 123 dst->pdiscard_alignment = MAX(dst->pdiscard_alignment, 124 src->pdiscard_alignment); 125 dst->opt_transfer = MAX(dst->opt_transfer, src->opt_transfer); 126 dst->max_transfer = MIN_NON_ZERO(dst->max_transfer, src->max_transfer); 127 dst->max_hw_transfer = MIN_NON_ZERO(dst->max_hw_transfer, 128 src->max_hw_transfer); 129 dst->opt_mem_alignment = MAX(dst->opt_mem_alignment, 130 src->opt_mem_alignment); 131 dst->min_mem_alignment = MAX(dst->min_mem_alignment, 132 src->min_mem_alignment); 133 dst->max_iov = MIN_NON_ZERO(dst->max_iov, src->max_iov); 134 dst->max_hw_iov = MIN_NON_ZERO(dst->max_hw_iov, src->max_hw_iov); 135 } 136 137 typedef struct BdrvRefreshLimitsState { 138 BlockDriverState *bs; 139 BlockLimits old_bl; 140 } BdrvRefreshLimitsState; 141 142 static void bdrv_refresh_limits_abort(void *opaque) 143 { 144 BdrvRefreshLimitsState *s = opaque; 145 146 s->bs->bl = s->old_bl; 147 } 148 149 static TransactionActionDrv bdrv_refresh_limits_drv = { 150 .abort = bdrv_refresh_limits_abort, 151 .clean = g_free, 152 }; 153 154 /* @tran is allowed to be NULL, in this case no rollback is possible. */ 155 void bdrv_refresh_limits(BlockDriverState *bs, Transaction *tran, Error **errp) 156 { 157 ERRP_GUARD(); 158 BlockDriver *drv = bs->drv; 159 BdrvChild *c; 160 bool have_limits; 161 162 GLOBAL_STATE_CODE(); 163 164 if (tran) { 165 BdrvRefreshLimitsState *s = g_new(BdrvRefreshLimitsState, 1); 166 *s = (BdrvRefreshLimitsState) { 167 .bs = bs, 168 .old_bl = bs->bl, 169 }; 170 tran_add(tran, &bdrv_refresh_limits_drv, s); 171 } 172 173 memset(&bs->bl, 0, sizeof(bs->bl)); 174 175 if (!drv) { 176 return; 177 } 178 179 /* Default alignment based on whether driver has byte interface */ 180 bs->bl.request_alignment = (drv->bdrv_co_preadv || 181 drv->bdrv_aio_preadv || 182 drv->bdrv_co_preadv_part) ? 1 : 512; 183 184 /* Take some limits from the children as a default */ 185 have_limits = false; 186 QLIST_FOREACH(c, &bs->children, next) { 187 if (c->role & (BDRV_CHILD_DATA | BDRV_CHILD_FILTERED | BDRV_CHILD_COW)) 188 { 189 bdrv_merge_limits(&bs->bl, &c->bs->bl); 190 have_limits = true; 191 } 192 } 193 194 if (!have_limits) { 195 bs->bl.min_mem_alignment = 512; 196 bs->bl.opt_mem_alignment = qemu_real_host_page_size(); 197 198 /* Safe default since most protocols use readv()/writev()/etc */ 199 bs->bl.max_iov = IOV_MAX; 200 } 201 202 /* Then let the driver override it */ 203 if (drv->bdrv_refresh_limits) { 204 drv->bdrv_refresh_limits(bs, errp); 205 if (*errp) { 206 return; 207 } 208 } 209 210 if (bs->bl.request_alignment > BDRV_MAX_ALIGNMENT) { 211 error_setg(errp, "Driver requires too large request alignment"); 212 } 213 } 214 215 /** 216 * The copy-on-read flag is actually a reference count so multiple users may 217 * use the feature without worrying about clobbering its previous state. 218 * Copy-on-read stays enabled until all users have called to disable it. 219 */ 220 void bdrv_enable_copy_on_read(BlockDriverState *bs) 221 { 222 IO_CODE(); 223 qatomic_inc(&bs->copy_on_read); 224 } 225 226 void bdrv_disable_copy_on_read(BlockDriverState *bs) 227 { 228 int old = qatomic_fetch_dec(&bs->copy_on_read); 229 IO_CODE(); 230 assert(old >= 1); 231 } 232 233 typedef struct { 234 Coroutine *co; 235 BlockDriverState *bs; 236 bool done; 237 bool begin; 238 bool poll; 239 BdrvChild *parent; 240 } BdrvCoDrainData; 241 242 /* Returns true if BDRV_POLL_WHILE() should go into a blocking aio_poll() */ 243 bool bdrv_drain_poll(BlockDriverState *bs, BdrvChild *ignore_parent, 244 bool ignore_bds_parents) 245 { 246 IO_OR_GS_CODE(); 247 248 if (bdrv_parent_drained_poll(bs, ignore_parent, ignore_bds_parents)) { 249 return true; 250 } 251 252 if (qatomic_read(&bs->in_flight)) { 253 return true; 254 } 255 256 return false; 257 } 258 259 static bool bdrv_drain_poll_top_level(BlockDriverState *bs, 260 BdrvChild *ignore_parent) 261 { 262 return bdrv_drain_poll(bs, ignore_parent, false); 263 } 264 265 static void bdrv_do_drained_begin(BlockDriverState *bs, BdrvChild *parent, 266 bool poll); 267 static void bdrv_do_drained_end(BlockDriverState *bs, BdrvChild *parent); 268 269 static void bdrv_co_drain_bh_cb(void *opaque) 270 { 271 BdrvCoDrainData *data = opaque; 272 Coroutine *co = data->co; 273 BlockDriverState *bs = data->bs; 274 275 if (bs) { 276 AioContext *ctx = bdrv_get_aio_context(bs); 277 aio_context_acquire(ctx); 278 bdrv_dec_in_flight(bs); 279 if (data->begin) { 280 bdrv_do_drained_begin(bs, data->parent, data->poll); 281 } else { 282 assert(!data->poll); 283 bdrv_do_drained_end(bs, data->parent); 284 } 285 aio_context_release(ctx); 286 } else { 287 assert(data->begin); 288 bdrv_drain_all_begin(); 289 } 290 291 data->done = true; 292 aio_co_wake(co); 293 } 294 295 static void coroutine_fn bdrv_co_yield_to_drain(BlockDriverState *bs, 296 bool begin, 297 BdrvChild *parent, 298 bool poll) 299 { 300 BdrvCoDrainData data; 301 Coroutine *self = qemu_coroutine_self(); 302 AioContext *ctx = bdrv_get_aio_context(bs); 303 AioContext *co_ctx = qemu_coroutine_get_aio_context(self); 304 305 /* Calling bdrv_drain() from a BH ensures the current coroutine yields and 306 * other coroutines run if they were queued by aio_co_enter(). */ 307 308 assert(qemu_in_coroutine()); 309 data = (BdrvCoDrainData) { 310 .co = self, 311 .bs = bs, 312 .done = false, 313 .begin = begin, 314 .parent = parent, 315 .poll = poll, 316 }; 317 318 if (bs) { 319 bdrv_inc_in_flight(bs); 320 } 321 322 /* 323 * Temporarily drop the lock across yield or we would get deadlocks. 324 * bdrv_co_drain_bh_cb() reaquires the lock as needed. 325 * 326 * When we yield below, the lock for the current context will be 327 * released, so if this is actually the lock that protects bs, don't drop 328 * it a second time. 329 */ 330 if (ctx != co_ctx) { 331 aio_context_release(ctx); 332 } 333 replay_bh_schedule_oneshot_event(ctx, bdrv_co_drain_bh_cb, &data); 334 335 qemu_coroutine_yield(); 336 /* If we are resumed from some other event (such as an aio completion or a 337 * timer callback), it is a bug in the caller that should be fixed. */ 338 assert(data.done); 339 340 /* Reaquire the AioContext of bs if we dropped it */ 341 if (ctx != co_ctx) { 342 aio_context_acquire(ctx); 343 } 344 } 345 346 static void bdrv_do_drained_begin(BlockDriverState *bs, BdrvChild *parent, 347 bool poll) 348 { 349 IO_OR_GS_CODE(); 350 351 if (qemu_in_coroutine()) { 352 bdrv_co_yield_to_drain(bs, true, parent, poll); 353 return; 354 } 355 356 /* Stop things in parent-to-child order */ 357 if (qatomic_fetch_inc(&bs->quiesce_counter) == 0) { 358 aio_disable_external(bdrv_get_aio_context(bs)); 359 bdrv_parent_drained_begin(bs, parent); 360 if (bs->drv && bs->drv->bdrv_drain_begin) { 361 bs->drv->bdrv_drain_begin(bs); 362 } 363 } 364 365 /* 366 * Wait for drained requests to finish. 367 * 368 * Calling BDRV_POLL_WHILE() only once for the top-level node is okay: The 369 * call is needed so things in this AioContext can make progress even 370 * though we don't return to the main AioContext loop - this automatically 371 * includes other nodes in the same AioContext and therefore all child 372 * nodes. 373 */ 374 if (poll) { 375 BDRV_POLL_WHILE(bs, bdrv_drain_poll_top_level(bs, parent)); 376 } 377 } 378 379 void bdrv_do_drained_begin_quiesce(BlockDriverState *bs, BdrvChild *parent) 380 { 381 bdrv_do_drained_begin(bs, parent, false); 382 } 383 384 void bdrv_drained_begin(BlockDriverState *bs) 385 { 386 IO_OR_GS_CODE(); 387 bdrv_do_drained_begin(bs, NULL, true); 388 } 389 390 /** 391 * This function does not poll, nor must any of its recursively called 392 * functions. 393 */ 394 static void bdrv_do_drained_end(BlockDriverState *bs, BdrvChild *parent) 395 { 396 int old_quiesce_counter; 397 398 if (qemu_in_coroutine()) { 399 bdrv_co_yield_to_drain(bs, false, parent, false); 400 return; 401 } 402 assert(bs->quiesce_counter > 0); 403 404 /* Re-enable things in child-to-parent order */ 405 old_quiesce_counter = qatomic_fetch_dec(&bs->quiesce_counter); 406 if (old_quiesce_counter == 1) { 407 if (bs->drv && bs->drv->bdrv_drain_end) { 408 bs->drv->bdrv_drain_end(bs); 409 } 410 bdrv_parent_drained_end(bs, parent); 411 aio_enable_external(bdrv_get_aio_context(bs)); 412 } 413 } 414 415 void bdrv_drained_end(BlockDriverState *bs) 416 { 417 IO_OR_GS_CODE(); 418 bdrv_do_drained_end(bs, NULL); 419 } 420 421 void bdrv_drain(BlockDriverState *bs) 422 { 423 IO_OR_GS_CODE(); 424 bdrv_drained_begin(bs); 425 bdrv_drained_end(bs); 426 } 427 428 static void bdrv_drain_assert_idle(BlockDriverState *bs) 429 { 430 BdrvChild *child, *next; 431 432 assert(qatomic_read(&bs->in_flight) == 0); 433 QLIST_FOREACH_SAFE(child, &bs->children, next, next) { 434 bdrv_drain_assert_idle(child->bs); 435 } 436 } 437 438 unsigned int bdrv_drain_all_count = 0; 439 440 static bool bdrv_drain_all_poll(void) 441 { 442 BlockDriverState *bs = NULL; 443 bool result = false; 444 GLOBAL_STATE_CODE(); 445 446 /* bdrv_drain_poll() can't make changes to the graph and we are holding the 447 * main AioContext lock, so iterating bdrv_next_all_states() is safe. */ 448 while ((bs = bdrv_next_all_states(bs))) { 449 AioContext *aio_context = bdrv_get_aio_context(bs); 450 aio_context_acquire(aio_context); 451 result |= bdrv_drain_poll(bs, NULL, true); 452 aio_context_release(aio_context); 453 } 454 455 return result; 456 } 457 458 /* 459 * Wait for pending requests to complete across all BlockDriverStates 460 * 461 * This function does not flush data to disk, use bdrv_flush_all() for that 462 * after calling this function. 463 * 464 * This pauses all block jobs and disables external clients. It must 465 * be paired with bdrv_drain_all_end(). 466 * 467 * NOTE: no new block jobs or BlockDriverStates can be created between 468 * the bdrv_drain_all_begin() and bdrv_drain_all_end() calls. 469 */ 470 void bdrv_drain_all_begin_nopoll(void) 471 { 472 BlockDriverState *bs = NULL; 473 GLOBAL_STATE_CODE(); 474 475 /* 476 * bdrv queue is managed by record/replay, 477 * waiting for finishing the I/O requests may 478 * be infinite 479 */ 480 if (replay_events_enabled()) { 481 return; 482 } 483 484 /* AIO_WAIT_WHILE() with a NULL context can only be called from the main 485 * loop AioContext, so make sure we're in the main context. */ 486 assert(qemu_get_current_aio_context() == qemu_get_aio_context()); 487 assert(bdrv_drain_all_count < INT_MAX); 488 bdrv_drain_all_count++; 489 490 /* Quiesce all nodes, without polling in-flight requests yet. The graph 491 * cannot change during this loop. */ 492 while ((bs = bdrv_next_all_states(bs))) { 493 AioContext *aio_context = bdrv_get_aio_context(bs); 494 495 aio_context_acquire(aio_context); 496 bdrv_do_drained_begin(bs, NULL, false); 497 aio_context_release(aio_context); 498 } 499 } 500 501 void bdrv_drain_all_begin(void) 502 { 503 BlockDriverState *bs = NULL; 504 505 if (qemu_in_coroutine()) { 506 bdrv_co_yield_to_drain(NULL, true, NULL, true); 507 return; 508 } 509 510 /* 511 * bdrv queue is managed by record/replay, 512 * waiting for finishing the I/O requests may 513 * be infinite 514 */ 515 if (replay_events_enabled()) { 516 return; 517 } 518 519 bdrv_drain_all_begin_nopoll(); 520 521 /* Now poll the in-flight requests */ 522 AIO_WAIT_WHILE(NULL, bdrv_drain_all_poll()); 523 524 while ((bs = bdrv_next_all_states(bs))) { 525 bdrv_drain_assert_idle(bs); 526 } 527 } 528 529 void bdrv_drain_all_end_quiesce(BlockDriverState *bs) 530 { 531 GLOBAL_STATE_CODE(); 532 533 g_assert(bs->quiesce_counter > 0); 534 g_assert(!bs->refcnt); 535 536 while (bs->quiesce_counter) { 537 bdrv_do_drained_end(bs, NULL); 538 } 539 } 540 541 void bdrv_drain_all_end(void) 542 { 543 BlockDriverState *bs = NULL; 544 GLOBAL_STATE_CODE(); 545 546 /* 547 * bdrv queue is managed by record/replay, 548 * waiting for finishing the I/O requests may 549 * be endless 550 */ 551 if (replay_events_enabled()) { 552 return; 553 } 554 555 while ((bs = bdrv_next_all_states(bs))) { 556 AioContext *aio_context = bdrv_get_aio_context(bs); 557 558 aio_context_acquire(aio_context); 559 bdrv_do_drained_end(bs, NULL); 560 aio_context_release(aio_context); 561 } 562 563 assert(qemu_get_current_aio_context() == qemu_get_aio_context()); 564 assert(bdrv_drain_all_count > 0); 565 bdrv_drain_all_count--; 566 } 567 568 void bdrv_drain_all(void) 569 { 570 GLOBAL_STATE_CODE(); 571 bdrv_drain_all_begin(); 572 bdrv_drain_all_end(); 573 } 574 575 /** 576 * Remove an active request from the tracked requests list 577 * 578 * This function should be called when a tracked request is completing. 579 */ 580 static void coroutine_fn tracked_request_end(BdrvTrackedRequest *req) 581 { 582 if (req->serialising) { 583 qatomic_dec(&req->bs->serialising_in_flight); 584 } 585 586 qemu_co_mutex_lock(&req->bs->reqs_lock); 587 QLIST_REMOVE(req, list); 588 qemu_co_queue_restart_all(&req->wait_queue); 589 qemu_co_mutex_unlock(&req->bs->reqs_lock); 590 } 591 592 /** 593 * Add an active request to the tracked requests list 594 */ 595 static void coroutine_fn tracked_request_begin(BdrvTrackedRequest *req, 596 BlockDriverState *bs, 597 int64_t offset, 598 int64_t bytes, 599 enum BdrvTrackedRequestType type) 600 { 601 bdrv_check_request(offset, bytes, &error_abort); 602 603 *req = (BdrvTrackedRequest){ 604 .bs = bs, 605 .offset = offset, 606 .bytes = bytes, 607 .type = type, 608 .co = qemu_coroutine_self(), 609 .serialising = false, 610 .overlap_offset = offset, 611 .overlap_bytes = bytes, 612 }; 613 614 qemu_co_queue_init(&req->wait_queue); 615 616 qemu_co_mutex_lock(&bs->reqs_lock); 617 QLIST_INSERT_HEAD(&bs->tracked_requests, req, list); 618 qemu_co_mutex_unlock(&bs->reqs_lock); 619 } 620 621 static bool tracked_request_overlaps(BdrvTrackedRequest *req, 622 int64_t offset, int64_t bytes) 623 { 624 bdrv_check_request(offset, bytes, &error_abort); 625 626 /* aaaa bbbb */ 627 if (offset >= req->overlap_offset + req->overlap_bytes) { 628 return false; 629 } 630 /* bbbb aaaa */ 631 if (req->overlap_offset >= offset + bytes) { 632 return false; 633 } 634 return true; 635 } 636 637 /* Called with self->bs->reqs_lock held */ 638 static coroutine_fn BdrvTrackedRequest * 639 bdrv_find_conflicting_request(BdrvTrackedRequest *self) 640 { 641 BdrvTrackedRequest *req; 642 643 QLIST_FOREACH(req, &self->bs->tracked_requests, list) { 644 if (req == self || (!req->serialising && !self->serialising)) { 645 continue; 646 } 647 if (tracked_request_overlaps(req, self->overlap_offset, 648 self->overlap_bytes)) 649 { 650 /* 651 * Hitting this means there was a reentrant request, for 652 * example, a block driver issuing nested requests. This must 653 * never happen since it means deadlock. 654 */ 655 assert(qemu_coroutine_self() != req->co); 656 657 /* 658 * If the request is already (indirectly) waiting for us, or 659 * will wait for us as soon as it wakes up, then just go on 660 * (instead of producing a deadlock in the former case). 661 */ 662 if (!req->waiting_for) { 663 return req; 664 } 665 } 666 } 667 668 return NULL; 669 } 670 671 /* Called with self->bs->reqs_lock held */ 672 static void coroutine_fn 673 bdrv_wait_serialising_requests_locked(BdrvTrackedRequest *self) 674 { 675 BdrvTrackedRequest *req; 676 677 while ((req = bdrv_find_conflicting_request(self))) { 678 self->waiting_for = req; 679 qemu_co_queue_wait(&req->wait_queue, &self->bs->reqs_lock); 680 self->waiting_for = NULL; 681 } 682 } 683 684 /* Called with req->bs->reqs_lock held */ 685 static void tracked_request_set_serialising(BdrvTrackedRequest *req, 686 uint64_t align) 687 { 688 int64_t overlap_offset = req->offset & ~(align - 1); 689 int64_t overlap_bytes = 690 ROUND_UP(req->offset + req->bytes, align) - overlap_offset; 691 692 bdrv_check_request(req->offset, req->bytes, &error_abort); 693 694 if (!req->serialising) { 695 qatomic_inc(&req->bs->serialising_in_flight); 696 req->serialising = true; 697 } 698 699 req->overlap_offset = MIN(req->overlap_offset, overlap_offset); 700 req->overlap_bytes = MAX(req->overlap_bytes, overlap_bytes); 701 } 702 703 /** 704 * Return the tracked request on @bs for the current coroutine, or 705 * NULL if there is none. 706 */ 707 BdrvTrackedRequest *coroutine_fn bdrv_co_get_self_request(BlockDriverState *bs) 708 { 709 BdrvTrackedRequest *req; 710 Coroutine *self = qemu_coroutine_self(); 711 IO_CODE(); 712 713 QLIST_FOREACH(req, &bs->tracked_requests, list) { 714 if (req->co == self) { 715 return req; 716 } 717 } 718 719 return NULL; 720 } 721 722 /** 723 * Round a region to cluster boundaries 724 */ 725 void coroutine_fn bdrv_round_to_clusters(BlockDriverState *bs, 726 int64_t offset, int64_t bytes, 727 int64_t *cluster_offset, 728 int64_t *cluster_bytes) 729 { 730 BlockDriverInfo bdi; 731 IO_CODE(); 732 if (bdrv_co_get_info(bs, &bdi) < 0 || bdi.cluster_size == 0) { 733 *cluster_offset = offset; 734 *cluster_bytes = bytes; 735 } else { 736 int64_t c = bdi.cluster_size; 737 *cluster_offset = QEMU_ALIGN_DOWN(offset, c); 738 *cluster_bytes = QEMU_ALIGN_UP(offset - *cluster_offset + bytes, c); 739 } 740 } 741 742 static coroutine_fn int bdrv_get_cluster_size(BlockDriverState *bs) 743 { 744 BlockDriverInfo bdi; 745 int ret; 746 747 ret = bdrv_co_get_info(bs, &bdi); 748 if (ret < 0 || bdi.cluster_size == 0) { 749 return bs->bl.request_alignment; 750 } else { 751 return bdi.cluster_size; 752 } 753 } 754 755 void bdrv_inc_in_flight(BlockDriverState *bs) 756 { 757 IO_CODE(); 758 qatomic_inc(&bs->in_flight); 759 } 760 761 void bdrv_wakeup(BlockDriverState *bs) 762 { 763 IO_CODE(); 764 aio_wait_kick(); 765 } 766 767 void bdrv_dec_in_flight(BlockDriverState *bs) 768 { 769 IO_CODE(); 770 qatomic_dec(&bs->in_flight); 771 bdrv_wakeup(bs); 772 } 773 774 static void coroutine_fn 775 bdrv_wait_serialising_requests(BdrvTrackedRequest *self) 776 { 777 BlockDriverState *bs = self->bs; 778 779 if (!qatomic_read(&bs->serialising_in_flight)) { 780 return; 781 } 782 783 qemu_co_mutex_lock(&bs->reqs_lock); 784 bdrv_wait_serialising_requests_locked(self); 785 qemu_co_mutex_unlock(&bs->reqs_lock); 786 } 787 788 void coroutine_fn bdrv_make_request_serialising(BdrvTrackedRequest *req, 789 uint64_t align) 790 { 791 IO_CODE(); 792 793 qemu_co_mutex_lock(&req->bs->reqs_lock); 794 795 tracked_request_set_serialising(req, align); 796 bdrv_wait_serialising_requests_locked(req); 797 798 qemu_co_mutex_unlock(&req->bs->reqs_lock); 799 } 800 801 int bdrv_check_qiov_request(int64_t offset, int64_t bytes, 802 QEMUIOVector *qiov, size_t qiov_offset, 803 Error **errp) 804 { 805 /* 806 * Check generic offset/bytes correctness 807 */ 808 809 if (offset < 0) { 810 error_setg(errp, "offset is negative: %" PRIi64, offset); 811 return -EIO; 812 } 813 814 if (bytes < 0) { 815 error_setg(errp, "bytes is negative: %" PRIi64, bytes); 816 return -EIO; 817 } 818 819 if (bytes > BDRV_MAX_LENGTH) { 820 error_setg(errp, "bytes(%" PRIi64 ") exceeds maximum(%" PRIi64 ")", 821 bytes, BDRV_MAX_LENGTH); 822 return -EIO; 823 } 824 825 if (offset > BDRV_MAX_LENGTH) { 826 error_setg(errp, "offset(%" PRIi64 ") exceeds maximum(%" PRIi64 ")", 827 offset, BDRV_MAX_LENGTH); 828 return -EIO; 829 } 830 831 if (offset > BDRV_MAX_LENGTH - bytes) { 832 error_setg(errp, "sum of offset(%" PRIi64 ") and bytes(%" PRIi64 ") " 833 "exceeds maximum(%" PRIi64 ")", offset, bytes, 834 BDRV_MAX_LENGTH); 835 return -EIO; 836 } 837 838 if (!qiov) { 839 return 0; 840 } 841 842 /* 843 * Check qiov and qiov_offset 844 */ 845 846 if (qiov_offset > qiov->size) { 847 error_setg(errp, "qiov_offset(%zu) overflow io vector size(%zu)", 848 qiov_offset, qiov->size); 849 return -EIO; 850 } 851 852 if (bytes > qiov->size - qiov_offset) { 853 error_setg(errp, "bytes(%" PRIi64 ") + qiov_offset(%zu) overflow io " 854 "vector size(%zu)", bytes, qiov_offset, qiov->size); 855 return -EIO; 856 } 857 858 return 0; 859 } 860 861 int bdrv_check_request(int64_t offset, int64_t bytes, Error **errp) 862 { 863 return bdrv_check_qiov_request(offset, bytes, NULL, 0, errp); 864 } 865 866 static int bdrv_check_request32(int64_t offset, int64_t bytes, 867 QEMUIOVector *qiov, size_t qiov_offset) 868 { 869 int ret = bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, NULL); 870 if (ret < 0) { 871 return ret; 872 } 873 874 if (bytes > BDRV_REQUEST_MAX_BYTES) { 875 return -EIO; 876 } 877 878 return 0; 879 } 880 881 /* 882 * Completely zero out a block device with the help of bdrv_pwrite_zeroes. 883 * The operation is sped up by checking the block status and only writing 884 * zeroes to the device if they currently do not return zeroes. Optional 885 * flags are passed through to bdrv_pwrite_zeroes (e.g. BDRV_REQ_MAY_UNMAP, 886 * BDRV_REQ_FUA). 887 * 888 * Returns < 0 on error, 0 on success. For error codes see bdrv_pwrite(). 889 */ 890 int bdrv_make_zero(BdrvChild *child, BdrvRequestFlags flags) 891 { 892 int ret; 893 int64_t target_size, bytes, offset = 0; 894 BlockDriverState *bs = child->bs; 895 IO_CODE(); 896 897 target_size = bdrv_getlength(bs); 898 if (target_size < 0) { 899 return target_size; 900 } 901 902 for (;;) { 903 bytes = MIN(target_size - offset, BDRV_REQUEST_MAX_BYTES); 904 if (bytes <= 0) { 905 return 0; 906 } 907 ret = bdrv_block_status(bs, offset, bytes, &bytes, NULL, NULL); 908 if (ret < 0) { 909 return ret; 910 } 911 if (ret & BDRV_BLOCK_ZERO) { 912 offset += bytes; 913 continue; 914 } 915 ret = bdrv_pwrite_zeroes(child, offset, bytes, flags); 916 if (ret < 0) { 917 return ret; 918 } 919 offset += bytes; 920 } 921 } 922 923 /* 924 * Writes to the file and ensures that no writes are reordered across this 925 * request (acts as a barrier) 926 * 927 * Returns 0 on success, -errno in error cases. 928 */ 929 int coroutine_fn bdrv_co_pwrite_sync(BdrvChild *child, int64_t offset, 930 int64_t bytes, const void *buf, 931 BdrvRequestFlags flags) 932 { 933 int ret; 934 IO_CODE(); 935 936 assume_graph_lock(); /* FIXME */ 937 938 ret = bdrv_co_pwrite(child, offset, bytes, buf, flags); 939 if (ret < 0) { 940 return ret; 941 } 942 943 ret = bdrv_co_flush(child->bs); 944 if (ret < 0) { 945 return ret; 946 } 947 948 return 0; 949 } 950 951 typedef struct CoroutineIOCompletion { 952 Coroutine *coroutine; 953 int ret; 954 } CoroutineIOCompletion; 955 956 static void bdrv_co_io_em_complete(void *opaque, int ret) 957 { 958 CoroutineIOCompletion *co = opaque; 959 960 co->ret = ret; 961 aio_co_wake(co->coroutine); 962 } 963 964 static int coroutine_fn bdrv_driver_preadv(BlockDriverState *bs, 965 int64_t offset, int64_t bytes, 966 QEMUIOVector *qiov, 967 size_t qiov_offset, int flags) 968 { 969 BlockDriver *drv = bs->drv; 970 int64_t sector_num; 971 unsigned int nb_sectors; 972 QEMUIOVector local_qiov; 973 int ret; 974 975 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort); 976 assert(!(flags & ~bs->supported_read_flags)); 977 978 if (!drv) { 979 return -ENOMEDIUM; 980 } 981 982 if (drv->bdrv_co_preadv_part) { 983 return drv->bdrv_co_preadv_part(bs, offset, bytes, qiov, qiov_offset, 984 flags); 985 } 986 987 if (qiov_offset > 0 || bytes != qiov->size) { 988 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes); 989 qiov = &local_qiov; 990 } 991 992 if (drv->bdrv_co_preadv) { 993 ret = drv->bdrv_co_preadv(bs, offset, bytes, qiov, flags); 994 goto out; 995 } 996 997 if (drv->bdrv_aio_preadv) { 998 BlockAIOCB *acb; 999 CoroutineIOCompletion co = { 1000 .coroutine = qemu_coroutine_self(), 1001 }; 1002 1003 acb = drv->bdrv_aio_preadv(bs, offset, bytes, qiov, flags, 1004 bdrv_co_io_em_complete, &co); 1005 if (acb == NULL) { 1006 ret = -EIO; 1007 goto out; 1008 } else { 1009 qemu_coroutine_yield(); 1010 ret = co.ret; 1011 goto out; 1012 } 1013 } 1014 1015 sector_num = offset >> BDRV_SECTOR_BITS; 1016 nb_sectors = bytes >> BDRV_SECTOR_BITS; 1017 1018 assert(QEMU_IS_ALIGNED(offset, BDRV_SECTOR_SIZE)); 1019 assert(QEMU_IS_ALIGNED(bytes, BDRV_SECTOR_SIZE)); 1020 assert(bytes <= BDRV_REQUEST_MAX_BYTES); 1021 assert(drv->bdrv_co_readv); 1022 1023 ret = drv->bdrv_co_readv(bs, sector_num, nb_sectors, qiov); 1024 1025 out: 1026 if (qiov == &local_qiov) { 1027 qemu_iovec_destroy(&local_qiov); 1028 } 1029 1030 return ret; 1031 } 1032 1033 static int coroutine_fn bdrv_driver_pwritev(BlockDriverState *bs, 1034 int64_t offset, int64_t bytes, 1035 QEMUIOVector *qiov, 1036 size_t qiov_offset, 1037 BdrvRequestFlags flags) 1038 { 1039 BlockDriver *drv = bs->drv; 1040 bool emulate_fua = false; 1041 int64_t sector_num; 1042 unsigned int nb_sectors; 1043 QEMUIOVector local_qiov; 1044 int ret; 1045 1046 assume_graph_lock(); /* FIXME */ 1047 1048 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort); 1049 1050 if (!drv) { 1051 return -ENOMEDIUM; 1052 } 1053 1054 if ((flags & BDRV_REQ_FUA) && 1055 (~bs->supported_write_flags & BDRV_REQ_FUA)) { 1056 flags &= ~BDRV_REQ_FUA; 1057 emulate_fua = true; 1058 } 1059 1060 flags &= bs->supported_write_flags; 1061 1062 if (drv->bdrv_co_pwritev_part) { 1063 ret = drv->bdrv_co_pwritev_part(bs, offset, bytes, qiov, qiov_offset, 1064 flags); 1065 goto emulate_flags; 1066 } 1067 1068 if (qiov_offset > 0 || bytes != qiov->size) { 1069 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes); 1070 qiov = &local_qiov; 1071 } 1072 1073 if (drv->bdrv_co_pwritev) { 1074 ret = drv->bdrv_co_pwritev(bs, offset, bytes, qiov, flags); 1075 goto emulate_flags; 1076 } 1077 1078 if (drv->bdrv_aio_pwritev) { 1079 BlockAIOCB *acb; 1080 CoroutineIOCompletion co = { 1081 .coroutine = qemu_coroutine_self(), 1082 }; 1083 1084 acb = drv->bdrv_aio_pwritev(bs, offset, bytes, qiov, flags, 1085 bdrv_co_io_em_complete, &co); 1086 if (acb == NULL) { 1087 ret = -EIO; 1088 } else { 1089 qemu_coroutine_yield(); 1090 ret = co.ret; 1091 } 1092 goto emulate_flags; 1093 } 1094 1095 sector_num = offset >> BDRV_SECTOR_BITS; 1096 nb_sectors = bytes >> BDRV_SECTOR_BITS; 1097 1098 assert(QEMU_IS_ALIGNED(offset, BDRV_SECTOR_SIZE)); 1099 assert(QEMU_IS_ALIGNED(bytes, BDRV_SECTOR_SIZE)); 1100 assert(bytes <= BDRV_REQUEST_MAX_BYTES); 1101 1102 assert(drv->bdrv_co_writev); 1103 ret = drv->bdrv_co_writev(bs, sector_num, nb_sectors, qiov, flags); 1104 1105 emulate_flags: 1106 if (ret == 0 && emulate_fua) { 1107 ret = bdrv_co_flush(bs); 1108 } 1109 1110 if (qiov == &local_qiov) { 1111 qemu_iovec_destroy(&local_qiov); 1112 } 1113 1114 return ret; 1115 } 1116 1117 static int coroutine_fn 1118 bdrv_driver_pwritev_compressed(BlockDriverState *bs, int64_t offset, 1119 int64_t bytes, QEMUIOVector *qiov, 1120 size_t qiov_offset) 1121 { 1122 BlockDriver *drv = bs->drv; 1123 QEMUIOVector local_qiov; 1124 int ret; 1125 1126 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort); 1127 1128 if (!drv) { 1129 return -ENOMEDIUM; 1130 } 1131 1132 if (!block_driver_can_compress(drv)) { 1133 return -ENOTSUP; 1134 } 1135 1136 if (drv->bdrv_co_pwritev_compressed_part) { 1137 return drv->bdrv_co_pwritev_compressed_part(bs, offset, bytes, 1138 qiov, qiov_offset); 1139 } 1140 1141 if (qiov_offset == 0) { 1142 return drv->bdrv_co_pwritev_compressed(bs, offset, bytes, qiov); 1143 } 1144 1145 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes); 1146 ret = drv->bdrv_co_pwritev_compressed(bs, offset, bytes, &local_qiov); 1147 qemu_iovec_destroy(&local_qiov); 1148 1149 return ret; 1150 } 1151 1152 static int coroutine_fn bdrv_co_do_copy_on_readv(BdrvChild *child, 1153 int64_t offset, int64_t bytes, QEMUIOVector *qiov, 1154 size_t qiov_offset, int flags) 1155 { 1156 BlockDriverState *bs = child->bs; 1157 1158 /* Perform I/O through a temporary buffer so that users who scribble over 1159 * their read buffer while the operation is in progress do not end up 1160 * modifying the image file. This is critical for zero-copy guest I/O 1161 * where anything might happen inside guest memory. 1162 */ 1163 void *bounce_buffer = NULL; 1164 1165 BlockDriver *drv = bs->drv; 1166 int64_t cluster_offset; 1167 int64_t cluster_bytes; 1168 int64_t skip_bytes; 1169 int ret; 1170 int max_transfer = MIN_NON_ZERO(bs->bl.max_transfer, 1171 BDRV_REQUEST_MAX_BYTES); 1172 int64_t progress = 0; 1173 bool skip_write; 1174 1175 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort); 1176 1177 if (!drv) { 1178 return -ENOMEDIUM; 1179 } 1180 1181 /* 1182 * Do not write anything when the BDS is inactive. That is not 1183 * allowed, and it would not help. 1184 */ 1185 skip_write = (bs->open_flags & BDRV_O_INACTIVE); 1186 1187 /* FIXME We cannot require callers to have write permissions when all they 1188 * are doing is a read request. If we did things right, write permissions 1189 * would be obtained anyway, but internally by the copy-on-read code. As 1190 * long as it is implemented here rather than in a separate filter driver, 1191 * the copy-on-read code doesn't have its own BdrvChild, however, for which 1192 * it could request permissions. Therefore we have to bypass the permission 1193 * system for the moment. */ 1194 // assert(child->perm & (BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE)); 1195 1196 /* Cover entire cluster so no additional backing file I/O is required when 1197 * allocating cluster in the image file. Note that this value may exceed 1198 * BDRV_REQUEST_MAX_BYTES (even when the original read did not), which 1199 * is one reason we loop rather than doing it all at once. 1200 */ 1201 bdrv_round_to_clusters(bs, offset, bytes, &cluster_offset, &cluster_bytes); 1202 skip_bytes = offset - cluster_offset; 1203 1204 trace_bdrv_co_do_copy_on_readv(bs, offset, bytes, 1205 cluster_offset, cluster_bytes); 1206 1207 while (cluster_bytes) { 1208 int64_t pnum; 1209 1210 if (skip_write) { 1211 ret = 1; /* "already allocated", so nothing will be copied */ 1212 pnum = MIN(cluster_bytes, max_transfer); 1213 } else { 1214 ret = bdrv_is_allocated(bs, cluster_offset, 1215 MIN(cluster_bytes, max_transfer), &pnum); 1216 if (ret < 0) { 1217 /* 1218 * Safe to treat errors in querying allocation as if 1219 * unallocated; we'll probably fail again soon on the 1220 * read, but at least that will set a decent errno. 1221 */ 1222 pnum = MIN(cluster_bytes, max_transfer); 1223 } 1224 1225 /* Stop at EOF if the image ends in the middle of the cluster */ 1226 if (ret == 0 && pnum == 0) { 1227 assert(progress >= bytes); 1228 break; 1229 } 1230 1231 assert(skip_bytes < pnum); 1232 } 1233 1234 if (ret <= 0) { 1235 QEMUIOVector local_qiov; 1236 1237 /* Must copy-on-read; use the bounce buffer */ 1238 pnum = MIN(pnum, MAX_BOUNCE_BUFFER); 1239 if (!bounce_buffer) { 1240 int64_t max_we_need = MAX(pnum, cluster_bytes - pnum); 1241 int64_t max_allowed = MIN(max_transfer, MAX_BOUNCE_BUFFER); 1242 int64_t bounce_buffer_len = MIN(max_we_need, max_allowed); 1243 1244 bounce_buffer = qemu_try_blockalign(bs, bounce_buffer_len); 1245 if (!bounce_buffer) { 1246 ret = -ENOMEM; 1247 goto err; 1248 } 1249 } 1250 qemu_iovec_init_buf(&local_qiov, bounce_buffer, pnum); 1251 1252 ret = bdrv_driver_preadv(bs, cluster_offset, pnum, 1253 &local_qiov, 0, 0); 1254 if (ret < 0) { 1255 goto err; 1256 } 1257 1258 bdrv_co_debug_event(bs, BLKDBG_COR_WRITE); 1259 if (drv->bdrv_co_pwrite_zeroes && 1260 buffer_is_zero(bounce_buffer, pnum)) { 1261 /* FIXME: Should we (perhaps conditionally) be setting 1262 * BDRV_REQ_MAY_UNMAP, if it will allow for a sparser copy 1263 * that still correctly reads as zero? */ 1264 ret = bdrv_co_do_pwrite_zeroes(bs, cluster_offset, pnum, 1265 BDRV_REQ_WRITE_UNCHANGED); 1266 } else { 1267 /* This does not change the data on the disk, it is not 1268 * necessary to flush even in cache=writethrough mode. 1269 */ 1270 ret = bdrv_driver_pwritev(bs, cluster_offset, pnum, 1271 &local_qiov, 0, 1272 BDRV_REQ_WRITE_UNCHANGED); 1273 } 1274 1275 if (ret < 0) { 1276 /* It might be okay to ignore write errors for guest 1277 * requests. If this is a deliberate copy-on-read 1278 * then we don't want to ignore the error. Simply 1279 * report it in all cases. 1280 */ 1281 goto err; 1282 } 1283 1284 if (!(flags & BDRV_REQ_PREFETCH)) { 1285 qemu_iovec_from_buf(qiov, qiov_offset + progress, 1286 bounce_buffer + skip_bytes, 1287 MIN(pnum - skip_bytes, bytes - progress)); 1288 } 1289 } else if (!(flags & BDRV_REQ_PREFETCH)) { 1290 /* Read directly into the destination */ 1291 ret = bdrv_driver_preadv(bs, offset + progress, 1292 MIN(pnum - skip_bytes, bytes - progress), 1293 qiov, qiov_offset + progress, 0); 1294 if (ret < 0) { 1295 goto err; 1296 } 1297 } 1298 1299 cluster_offset += pnum; 1300 cluster_bytes -= pnum; 1301 progress += pnum - skip_bytes; 1302 skip_bytes = 0; 1303 } 1304 ret = 0; 1305 1306 err: 1307 qemu_vfree(bounce_buffer); 1308 return ret; 1309 } 1310 1311 /* 1312 * Forwards an already correctly aligned request to the BlockDriver. This 1313 * handles copy on read, zeroing after EOF, and fragmentation of large 1314 * reads; any other features must be implemented by the caller. 1315 */ 1316 static int coroutine_fn bdrv_aligned_preadv(BdrvChild *child, 1317 BdrvTrackedRequest *req, int64_t offset, int64_t bytes, 1318 int64_t align, QEMUIOVector *qiov, size_t qiov_offset, int flags) 1319 { 1320 BlockDriverState *bs = child->bs; 1321 int64_t total_bytes, max_bytes; 1322 int ret = 0; 1323 int64_t bytes_remaining = bytes; 1324 int max_transfer; 1325 1326 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort); 1327 assert(is_power_of_2(align)); 1328 assert((offset & (align - 1)) == 0); 1329 assert((bytes & (align - 1)) == 0); 1330 assert((bs->open_flags & BDRV_O_NO_IO) == 0); 1331 max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX), 1332 align); 1333 1334 /* 1335 * TODO: We would need a per-BDS .supported_read_flags and 1336 * potential fallback support, if we ever implement any read flags 1337 * to pass through to drivers. For now, there aren't any 1338 * passthrough flags except the BDRV_REQ_REGISTERED_BUF optimization hint. 1339 */ 1340 assert(!(flags & ~(BDRV_REQ_COPY_ON_READ | BDRV_REQ_PREFETCH | 1341 BDRV_REQ_REGISTERED_BUF))); 1342 1343 /* Handle Copy on Read and associated serialisation */ 1344 if (flags & BDRV_REQ_COPY_ON_READ) { 1345 /* If we touch the same cluster it counts as an overlap. This 1346 * guarantees that allocating writes will be serialized and not race 1347 * with each other for the same cluster. For example, in copy-on-read 1348 * it ensures that the CoR read and write operations are atomic and 1349 * guest writes cannot interleave between them. */ 1350 bdrv_make_request_serialising(req, bdrv_get_cluster_size(bs)); 1351 } else { 1352 bdrv_wait_serialising_requests(req); 1353 } 1354 1355 if (flags & BDRV_REQ_COPY_ON_READ) { 1356 int64_t pnum; 1357 1358 /* The flag BDRV_REQ_COPY_ON_READ has reached its addressee */ 1359 flags &= ~BDRV_REQ_COPY_ON_READ; 1360 1361 ret = bdrv_is_allocated(bs, offset, bytes, &pnum); 1362 if (ret < 0) { 1363 goto out; 1364 } 1365 1366 if (!ret || pnum != bytes) { 1367 ret = bdrv_co_do_copy_on_readv(child, offset, bytes, 1368 qiov, qiov_offset, flags); 1369 goto out; 1370 } else if (flags & BDRV_REQ_PREFETCH) { 1371 goto out; 1372 } 1373 } 1374 1375 /* Forward the request to the BlockDriver, possibly fragmenting it */ 1376 total_bytes = bdrv_getlength(bs); 1377 if (total_bytes < 0) { 1378 ret = total_bytes; 1379 goto out; 1380 } 1381 1382 assert(!(flags & ~(bs->supported_read_flags | BDRV_REQ_REGISTERED_BUF))); 1383 1384 max_bytes = ROUND_UP(MAX(0, total_bytes - offset), align); 1385 if (bytes <= max_bytes && bytes <= max_transfer) { 1386 ret = bdrv_driver_preadv(bs, offset, bytes, qiov, qiov_offset, flags); 1387 goto out; 1388 } 1389 1390 while (bytes_remaining) { 1391 int64_t num; 1392 1393 if (max_bytes) { 1394 num = MIN(bytes_remaining, MIN(max_bytes, max_transfer)); 1395 assert(num); 1396 1397 ret = bdrv_driver_preadv(bs, offset + bytes - bytes_remaining, 1398 num, qiov, 1399 qiov_offset + bytes - bytes_remaining, 1400 flags); 1401 max_bytes -= num; 1402 } else { 1403 num = bytes_remaining; 1404 ret = qemu_iovec_memset(qiov, qiov_offset + bytes - bytes_remaining, 1405 0, bytes_remaining); 1406 } 1407 if (ret < 0) { 1408 goto out; 1409 } 1410 bytes_remaining -= num; 1411 } 1412 1413 out: 1414 return ret < 0 ? ret : 0; 1415 } 1416 1417 /* 1418 * Request padding 1419 * 1420 * |<---- align ----->| |<----- align ---->| 1421 * |<- head ->|<------------- bytes ------------->|<-- tail -->| 1422 * | | | | | | 1423 * -*----------$-------*-------- ... --------*-----$------------*--- 1424 * | | | | | | 1425 * | offset | | end | 1426 * ALIGN_DOWN(offset) ALIGN_UP(offset) ALIGN_DOWN(end) ALIGN_UP(end) 1427 * [buf ... ) [tail_buf ) 1428 * 1429 * @buf is an aligned allocation needed to store @head and @tail paddings. @head 1430 * is placed at the beginning of @buf and @tail at the @end. 1431 * 1432 * @tail_buf is a pointer to sub-buffer, corresponding to align-sized chunk 1433 * around tail, if tail exists. 1434 * 1435 * @merge_reads is true for small requests, 1436 * if @buf_len == @head + bytes + @tail. In this case it is possible that both 1437 * head and tail exist but @buf_len == align and @tail_buf == @buf. 1438 */ 1439 typedef struct BdrvRequestPadding { 1440 uint8_t *buf; 1441 size_t buf_len; 1442 uint8_t *tail_buf; 1443 size_t head; 1444 size_t tail; 1445 bool merge_reads; 1446 QEMUIOVector local_qiov; 1447 } BdrvRequestPadding; 1448 1449 static bool bdrv_init_padding(BlockDriverState *bs, 1450 int64_t offset, int64_t bytes, 1451 BdrvRequestPadding *pad) 1452 { 1453 int64_t align = bs->bl.request_alignment; 1454 int64_t sum; 1455 1456 bdrv_check_request(offset, bytes, &error_abort); 1457 assert(align <= INT_MAX); /* documented in block/block_int.h */ 1458 assert(align <= SIZE_MAX / 2); /* so we can allocate the buffer */ 1459 1460 memset(pad, 0, sizeof(*pad)); 1461 1462 pad->head = offset & (align - 1); 1463 pad->tail = ((offset + bytes) & (align - 1)); 1464 if (pad->tail) { 1465 pad->tail = align - pad->tail; 1466 } 1467 1468 if (!pad->head && !pad->tail) { 1469 return false; 1470 } 1471 1472 assert(bytes); /* Nothing good in aligning zero-length requests */ 1473 1474 sum = pad->head + bytes + pad->tail; 1475 pad->buf_len = (sum > align && pad->head && pad->tail) ? 2 * align : align; 1476 pad->buf = qemu_blockalign(bs, pad->buf_len); 1477 pad->merge_reads = sum == pad->buf_len; 1478 if (pad->tail) { 1479 pad->tail_buf = pad->buf + pad->buf_len - align; 1480 } 1481 1482 return true; 1483 } 1484 1485 static coroutine_fn int bdrv_padding_rmw_read(BdrvChild *child, 1486 BdrvTrackedRequest *req, 1487 BdrvRequestPadding *pad, 1488 bool zero_middle) 1489 { 1490 QEMUIOVector local_qiov; 1491 BlockDriverState *bs = child->bs; 1492 uint64_t align = bs->bl.request_alignment; 1493 int ret; 1494 1495 assert(req->serialising && pad->buf); 1496 1497 if (pad->head || pad->merge_reads) { 1498 int64_t bytes = pad->merge_reads ? pad->buf_len : align; 1499 1500 qemu_iovec_init_buf(&local_qiov, pad->buf, bytes); 1501 1502 if (pad->head) { 1503 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD); 1504 } 1505 if (pad->merge_reads && pad->tail) { 1506 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL); 1507 } 1508 ret = bdrv_aligned_preadv(child, req, req->overlap_offset, bytes, 1509 align, &local_qiov, 0, 0); 1510 if (ret < 0) { 1511 return ret; 1512 } 1513 if (pad->head) { 1514 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD); 1515 } 1516 if (pad->merge_reads && pad->tail) { 1517 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL); 1518 } 1519 1520 if (pad->merge_reads) { 1521 goto zero_mem; 1522 } 1523 } 1524 1525 if (pad->tail) { 1526 qemu_iovec_init_buf(&local_qiov, pad->tail_buf, align); 1527 1528 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL); 1529 ret = bdrv_aligned_preadv( 1530 child, req, 1531 req->overlap_offset + req->overlap_bytes - align, 1532 align, align, &local_qiov, 0, 0); 1533 if (ret < 0) { 1534 return ret; 1535 } 1536 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL); 1537 } 1538 1539 zero_mem: 1540 if (zero_middle) { 1541 memset(pad->buf + pad->head, 0, pad->buf_len - pad->head - pad->tail); 1542 } 1543 1544 return 0; 1545 } 1546 1547 static void bdrv_padding_destroy(BdrvRequestPadding *pad) 1548 { 1549 if (pad->buf) { 1550 qemu_vfree(pad->buf); 1551 qemu_iovec_destroy(&pad->local_qiov); 1552 } 1553 memset(pad, 0, sizeof(*pad)); 1554 } 1555 1556 /* 1557 * bdrv_pad_request 1558 * 1559 * Exchange request parameters with padded request if needed. Don't include RMW 1560 * read of padding, bdrv_padding_rmw_read() should be called separately if 1561 * needed. 1562 * 1563 * Request parameters (@qiov, &qiov_offset, &offset, &bytes) are in-out: 1564 * - on function start they represent original request 1565 * - on failure or when padding is not needed they are unchanged 1566 * - on success when padding is needed they represent padded request 1567 */ 1568 static int bdrv_pad_request(BlockDriverState *bs, 1569 QEMUIOVector **qiov, size_t *qiov_offset, 1570 int64_t *offset, int64_t *bytes, 1571 BdrvRequestPadding *pad, bool *padded, 1572 BdrvRequestFlags *flags) 1573 { 1574 int ret; 1575 1576 bdrv_check_qiov_request(*offset, *bytes, *qiov, *qiov_offset, &error_abort); 1577 1578 if (!bdrv_init_padding(bs, *offset, *bytes, pad)) { 1579 if (padded) { 1580 *padded = false; 1581 } 1582 return 0; 1583 } 1584 1585 ret = qemu_iovec_init_extended(&pad->local_qiov, pad->buf, pad->head, 1586 *qiov, *qiov_offset, *bytes, 1587 pad->buf + pad->buf_len - pad->tail, 1588 pad->tail); 1589 if (ret < 0) { 1590 bdrv_padding_destroy(pad); 1591 return ret; 1592 } 1593 *bytes += pad->head + pad->tail; 1594 *offset -= pad->head; 1595 *qiov = &pad->local_qiov; 1596 *qiov_offset = 0; 1597 if (padded) { 1598 *padded = true; 1599 } 1600 if (flags) { 1601 /* Can't use optimization hint with bounce buffer */ 1602 *flags &= ~BDRV_REQ_REGISTERED_BUF; 1603 } 1604 1605 return 0; 1606 } 1607 1608 int coroutine_fn bdrv_co_preadv(BdrvChild *child, 1609 int64_t offset, int64_t bytes, QEMUIOVector *qiov, 1610 BdrvRequestFlags flags) 1611 { 1612 IO_CODE(); 1613 return bdrv_co_preadv_part(child, offset, bytes, qiov, 0, flags); 1614 } 1615 1616 int coroutine_fn bdrv_co_preadv_part(BdrvChild *child, 1617 int64_t offset, int64_t bytes, 1618 QEMUIOVector *qiov, size_t qiov_offset, 1619 BdrvRequestFlags flags) 1620 { 1621 BlockDriverState *bs = child->bs; 1622 BdrvTrackedRequest req; 1623 BdrvRequestPadding pad; 1624 int ret; 1625 IO_CODE(); 1626 1627 trace_bdrv_co_preadv_part(bs, offset, bytes, flags); 1628 1629 if (!bdrv_co_is_inserted(bs)) { 1630 return -ENOMEDIUM; 1631 } 1632 1633 ret = bdrv_check_request32(offset, bytes, qiov, qiov_offset); 1634 if (ret < 0) { 1635 return ret; 1636 } 1637 1638 if (bytes == 0 && !QEMU_IS_ALIGNED(offset, bs->bl.request_alignment)) { 1639 /* 1640 * Aligning zero request is nonsense. Even if driver has special meaning 1641 * of zero-length (like qcow2_co_pwritev_compressed_part), we can't pass 1642 * it to driver due to request_alignment. 1643 * 1644 * Still, no reason to return an error if someone do unaligned 1645 * zero-length read occasionally. 1646 */ 1647 return 0; 1648 } 1649 1650 bdrv_inc_in_flight(bs); 1651 1652 /* Don't do copy-on-read if we read data before write operation */ 1653 if (qatomic_read(&bs->copy_on_read)) { 1654 flags |= BDRV_REQ_COPY_ON_READ; 1655 } 1656 1657 ret = bdrv_pad_request(bs, &qiov, &qiov_offset, &offset, &bytes, &pad, 1658 NULL, &flags); 1659 if (ret < 0) { 1660 goto fail; 1661 } 1662 1663 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_READ); 1664 ret = bdrv_aligned_preadv(child, &req, offset, bytes, 1665 bs->bl.request_alignment, 1666 qiov, qiov_offset, flags); 1667 tracked_request_end(&req); 1668 bdrv_padding_destroy(&pad); 1669 1670 fail: 1671 bdrv_dec_in_flight(bs); 1672 1673 return ret; 1674 } 1675 1676 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs, 1677 int64_t offset, int64_t bytes, BdrvRequestFlags flags) 1678 { 1679 BlockDriver *drv = bs->drv; 1680 QEMUIOVector qiov; 1681 void *buf = NULL; 1682 int ret = 0; 1683 bool need_flush = false; 1684 int head = 0; 1685 int tail = 0; 1686 1687 assume_graph_lock(); /* FIXME */ 1688 1689 int64_t max_write_zeroes = MIN_NON_ZERO(bs->bl.max_pwrite_zeroes, 1690 INT64_MAX); 1691 int alignment = MAX(bs->bl.pwrite_zeroes_alignment, 1692 bs->bl.request_alignment); 1693 int max_transfer = MIN_NON_ZERO(bs->bl.max_transfer, MAX_BOUNCE_BUFFER); 1694 1695 bdrv_check_request(offset, bytes, &error_abort); 1696 1697 if (!drv) { 1698 return -ENOMEDIUM; 1699 } 1700 1701 if ((flags & ~bs->supported_zero_flags) & BDRV_REQ_NO_FALLBACK) { 1702 return -ENOTSUP; 1703 } 1704 1705 /* By definition there is no user buffer so this flag doesn't make sense */ 1706 if (flags & BDRV_REQ_REGISTERED_BUF) { 1707 return -EINVAL; 1708 } 1709 1710 /* Invalidate the cached block-status data range if this write overlaps */ 1711 bdrv_bsc_invalidate_range(bs, offset, bytes); 1712 1713 assert(alignment % bs->bl.request_alignment == 0); 1714 head = offset % alignment; 1715 tail = (offset + bytes) % alignment; 1716 max_write_zeroes = QEMU_ALIGN_DOWN(max_write_zeroes, alignment); 1717 assert(max_write_zeroes >= bs->bl.request_alignment); 1718 1719 while (bytes > 0 && !ret) { 1720 int64_t num = bytes; 1721 1722 /* Align request. Block drivers can expect the "bulk" of the request 1723 * to be aligned, and that unaligned requests do not cross cluster 1724 * boundaries. 1725 */ 1726 if (head) { 1727 /* Make a small request up to the first aligned sector. For 1728 * convenience, limit this request to max_transfer even if 1729 * we don't need to fall back to writes. */ 1730 num = MIN(MIN(bytes, max_transfer), alignment - head); 1731 head = (head + num) % alignment; 1732 assert(num < max_write_zeroes); 1733 } else if (tail && num > alignment) { 1734 /* Shorten the request to the last aligned sector. */ 1735 num -= tail; 1736 } 1737 1738 /* limit request size */ 1739 if (num > max_write_zeroes) { 1740 num = max_write_zeroes; 1741 } 1742 1743 ret = -ENOTSUP; 1744 /* First try the efficient write zeroes operation */ 1745 if (drv->bdrv_co_pwrite_zeroes) { 1746 ret = drv->bdrv_co_pwrite_zeroes(bs, offset, num, 1747 flags & bs->supported_zero_flags); 1748 if (ret != -ENOTSUP && (flags & BDRV_REQ_FUA) && 1749 !(bs->supported_zero_flags & BDRV_REQ_FUA)) { 1750 need_flush = true; 1751 } 1752 } else { 1753 assert(!bs->supported_zero_flags); 1754 } 1755 1756 if (ret == -ENOTSUP && !(flags & BDRV_REQ_NO_FALLBACK)) { 1757 /* Fall back to bounce buffer if write zeroes is unsupported */ 1758 BdrvRequestFlags write_flags = flags & ~BDRV_REQ_ZERO_WRITE; 1759 1760 if ((flags & BDRV_REQ_FUA) && 1761 !(bs->supported_write_flags & BDRV_REQ_FUA)) { 1762 /* No need for bdrv_driver_pwrite() to do a fallback 1763 * flush on each chunk; use just one at the end */ 1764 write_flags &= ~BDRV_REQ_FUA; 1765 need_flush = true; 1766 } 1767 num = MIN(num, max_transfer); 1768 if (buf == NULL) { 1769 buf = qemu_try_blockalign0(bs, num); 1770 if (buf == NULL) { 1771 ret = -ENOMEM; 1772 goto fail; 1773 } 1774 } 1775 qemu_iovec_init_buf(&qiov, buf, num); 1776 1777 ret = bdrv_driver_pwritev(bs, offset, num, &qiov, 0, write_flags); 1778 1779 /* Keep bounce buffer around if it is big enough for all 1780 * all future requests. 1781 */ 1782 if (num < max_transfer) { 1783 qemu_vfree(buf); 1784 buf = NULL; 1785 } 1786 } 1787 1788 offset += num; 1789 bytes -= num; 1790 } 1791 1792 fail: 1793 if (ret == 0 && need_flush) { 1794 ret = bdrv_co_flush(bs); 1795 } 1796 qemu_vfree(buf); 1797 return ret; 1798 } 1799 1800 static inline int coroutine_fn 1801 bdrv_co_write_req_prepare(BdrvChild *child, int64_t offset, int64_t bytes, 1802 BdrvTrackedRequest *req, int flags) 1803 { 1804 BlockDriverState *bs = child->bs; 1805 1806 bdrv_check_request(offset, bytes, &error_abort); 1807 1808 if (bdrv_is_read_only(bs)) { 1809 return -EPERM; 1810 } 1811 1812 assert(!(bs->open_flags & BDRV_O_INACTIVE)); 1813 assert((bs->open_flags & BDRV_O_NO_IO) == 0); 1814 assert(!(flags & ~BDRV_REQ_MASK)); 1815 assert(!((flags & BDRV_REQ_NO_WAIT) && !(flags & BDRV_REQ_SERIALISING))); 1816 1817 if (flags & BDRV_REQ_SERIALISING) { 1818 QEMU_LOCK_GUARD(&bs->reqs_lock); 1819 1820 tracked_request_set_serialising(req, bdrv_get_cluster_size(bs)); 1821 1822 if ((flags & BDRV_REQ_NO_WAIT) && bdrv_find_conflicting_request(req)) { 1823 return -EBUSY; 1824 } 1825 1826 bdrv_wait_serialising_requests_locked(req); 1827 } else { 1828 bdrv_wait_serialising_requests(req); 1829 } 1830 1831 assert(req->overlap_offset <= offset); 1832 assert(offset + bytes <= req->overlap_offset + req->overlap_bytes); 1833 assert(offset + bytes <= bs->total_sectors * BDRV_SECTOR_SIZE || 1834 child->perm & BLK_PERM_RESIZE); 1835 1836 switch (req->type) { 1837 case BDRV_TRACKED_WRITE: 1838 case BDRV_TRACKED_DISCARD: 1839 if (flags & BDRV_REQ_WRITE_UNCHANGED) { 1840 assert(child->perm & (BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE)); 1841 } else { 1842 assert(child->perm & BLK_PERM_WRITE); 1843 } 1844 bdrv_write_threshold_check_write(bs, offset, bytes); 1845 return 0; 1846 case BDRV_TRACKED_TRUNCATE: 1847 assert(child->perm & BLK_PERM_RESIZE); 1848 return 0; 1849 default: 1850 abort(); 1851 } 1852 } 1853 1854 static inline void coroutine_fn 1855 bdrv_co_write_req_finish(BdrvChild *child, int64_t offset, int64_t bytes, 1856 BdrvTrackedRequest *req, int ret) 1857 { 1858 int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE); 1859 BlockDriverState *bs = child->bs; 1860 1861 bdrv_check_request(offset, bytes, &error_abort); 1862 1863 qatomic_inc(&bs->write_gen); 1864 1865 /* 1866 * Discard cannot extend the image, but in error handling cases, such as 1867 * when reverting a qcow2 cluster allocation, the discarded range can pass 1868 * the end of image file, so we cannot assert about BDRV_TRACKED_DISCARD 1869 * here. Instead, just skip it, since semantically a discard request 1870 * beyond EOF cannot expand the image anyway. 1871 */ 1872 if (ret == 0 && 1873 (req->type == BDRV_TRACKED_TRUNCATE || 1874 end_sector > bs->total_sectors) && 1875 req->type != BDRV_TRACKED_DISCARD) { 1876 bs->total_sectors = end_sector; 1877 bdrv_parent_cb_resize(bs); 1878 bdrv_dirty_bitmap_truncate(bs, end_sector << BDRV_SECTOR_BITS); 1879 } 1880 if (req->bytes) { 1881 switch (req->type) { 1882 case BDRV_TRACKED_WRITE: 1883 stat64_max(&bs->wr_highest_offset, offset + bytes); 1884 /* fall through, to set dirty bits */ 1885 case BDRV_TRACKED_DISCARD: 1886 bdrv_set_dirty(bs, offset, bytes); 1887 break; 1888 default: 1889 break; 1890 } 1891 } 1892 } 1893 1894 /* 1895 * Forwards an already correctly aligned write request to the BlockDriver, 1896 * after possibly fragmenting it. 1897 */ 1898 static int coroutine_fn bdrv_aligned_pwritev(BdrvChild *child, 1899 BdrvTrackedRequest *req, int64_t offset, int64_t bytes, 1900 int64_t align, QEMUIOVector *qiov, size_t qiov_offset, 1901 BdrvRequestFlags flags) 1902 { 1903 BlockDriverState *bs = child->bs; 1904 BlockDriver *drv = bs->drv; 1905 int ret; 1906 1907 int64_t bytes_remaining = bytes; 1908 int max_transfer; 1909 1910 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort); 1911 1912 if (!drv) { 1913 return -ENOMEDIUM; 1914 } 1915 1916 if (bdrv_has_readonly_bitmaps(bs)) { 1917 return -EPERM; 1918 } 1919 1920 assert(is_power_of_2(align)); 1921 assert((offset & (align - 1)) == 0); 1922 assert((bytes & (align - 1)) == 0); 1923 max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX), 1924 align); 1925 1926 ret = bdrv_co_write_req_prepare(child, offset, bytes, req, flags); 1927 1928 if (!ret && bs->detect_zeroes != BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF && 1929 !(flags & BDRV_REQ_ZERO_WRITE) && drv->bdrv_co_pwrite_zeroes && 1930 qemu_iovec_is_zero(qiov, qiov_offset, bytes)) { 1931 flags |= BDRV_REQ_ZERO_WRITE; 1932 if (bs->detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP) { 1933 flags |= BDRV_REQ_MAY_UNMAP; 1934 } 1935 1936 /* Can't use optimization hint with bufferless zero write */ 1937 flags &= ~BDRV_REQ_REGISTERED_BUF; 1938 } 1939 1940 if (ret < 0) { 1941 /* Do nothing, write notifier decided to fail this request */ 1942 } else if (flags & BDRV_REQ_ZERO_WRITE) { 1943 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_ZERO); 1944 ret = bdrv_co_do_pwrite_zeroes(bs, offset, bytes, flags); 1945 } else if (flags & BDRV_REQ_WRITE_COMPRESSED) { 1946 ret = bdrv_driver_pwritev_compressed(bs, offset, bytes, 1947 qiov, qiov_offset); 1948 } else if (bytes <= max_transfer) { 1949 bdrv_co_debug_event(bs, BLKDBG_PWRITEV); 1950 ret = bdrv_driver_pwritev(bs, offset, bytes, qiov, qiov_offset, flags); 1951 } else { 1952 bdrv_co_debug_event(bs, BLKDBG_PWRITEV); 1953 while (bytes_remaining) { 1954 int num = MIN(bytes_remaining, max_transfer); 1955 int local_flags = flags; 1956 1957 assert(num); 1958 if (num < bytes_remaining && (flags & BDRV_REQ_FUA) && 1959 !(bs->supported_write_flags & BDRV_REQ_FUA)) { 1960 /* If FUA is going to be emulated by flush, we only 1961 * need to flush on the last iteration */ 1962 local_flags &= ~BDRV_REQ_FUA; 1963 } 1964 1965 ret = bdrv_driver_pwritev(bs, offset + bytes - bytes_remaining, 1966 num, qiov, 1967 qiov_offset + bytes - bytes_remaining, 1968 local_flags); 1969 if (ret < 0) { 1970 break; 1971 } 1972 bytes_remaining -= num; 1973 } 1974 } 1975 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_DONE); 1976 1977 if (ret >= 0) { 1978 ret = 0; 1979 } 1980 bdrv_co_write_req_finish(child, offset, bytes, req, ret); 1981 1982 return ret; 1983 } 1984 1985 static int coroutine_fn bdrv_co_do_zero_pwritev(BdrvChild *child, 1986 int64_t offset, 1987 int64_t bytes, 1988 BdrvRequestFlags flags, 1989 BdrvTrackedRequest *req) 1990 { 1991 BlockDriverState *bs = child->bs; 1992 QEMUIOVector local_qiov; 1993 uint64_t align = bs->bl.request_alignment; 1994 int ret = 0; 1995 bool padding; 1996 BdrvRequestPadding pad; 1997 1998 /* This flag doesn't make sense for padding or zero writes */ 1999 flags &= ~BDRV_REQ_REGISTERED_BUF; 2000 2001 padding = bdrv_init_padding(bs, offset, bytes, &pad); 2002 if (padding) { 2003 assert(!(flags & BDRV_REQ_NO_WAIT)); 2004 bdrv_make_request_serialising(req, align); 2005 2006 bdrv_padding_rmw_read(child, req, &pad, true); 2007 2008 if (pad.head || pad.merge_reads) { 2009 int64_t aligned_offset = offset & ~(align - 1); 2010 int64_t write_bytes = pad.merge_reads ? pad.buf_len : align; 2011 2012 qemu_iovec_init_buf(&local_qiov, pad.buf, write_bytes); 2013 ret = bdrv_aligned_pwritev(child, req, aligned_offset, write_bytes, 2014 align, &local_qiov, 0, 2015 flags & ~BDRV_REQ_ZERO_WRITE); 2016 if (ret < 0 || pad.merge_reads) { 2017 /* Error or all work is done */ 2018 goto out; 2019 } 2020 offset += write_bytes - pad.head; 2021 bytes -= write_bytes - pad.head; 2022 } 2023 } 2024 2025 assert(!bytes || (offset & (align - 1)) == 0); 2026 if (bytes >= align) { 2027 /* Write the aligned part in the middle. */ 2028 int64_t aligned_bytes = bytes & ~(align - 1); 2029 ret = bdrv_aligned_pwritev(child, req, offset, aligned_bytes, align, 2030 NULL, 0, flags); 2031 if (ret < 0) { 2032 goto out; 2033 } 2034 bytes -= aligned_bytes; 2035 offset += aligned_bytes; 2036 } 2037 2038 assert(!bytes || (offset & (align - 1)) == 0); 2039 if (bytes) { 2040 assert(align == pad.tail + bytes); 2041 2042 qemu_iovec_init_buf(&local_qiov, pad.tail_buf, align); 2043 ret = bdrv_aligned_pwritev(child, req, offset, align, align, 2044 &local_qiov, 0, 2045 flags & ~BDRV_REQ_ZERO_WRITE); 2046 } 2047 2048 out: 2049 bdrv_padding_destroy(&pad); 2050 2051 return ret; 2052 } 2053 2054 /* 2055 * Handle a write request in coroutine context 2056 */ 2057 int coroutine_fn bdrv_co_pwritev(BdrvChild *child, 2058 int64_t offset, int64_t bytes, QEMUIOVector *qiov, 2059 BdrvRequestFlags flags) 2060 { 2061 IO_CODE(); 2062 return bdrv_co_pwritev_part(child, offset, bytes, qiov, 0, flags); 2063 } 2064 2065 int coroutine_fn bdrv_co_pwritev_part(BdrvChild *child, 2066 int64_t offset, int64_t bytes, QEMUIOVector *qiov, size_t qiov_offset, 2067 BdrvRequestFlags flags) 2068 { 2069 BlockDriverState *bs = child->bs; 2070 BdrvTrackedRequest req; 2071 uint64_t align = bs->bl.request_alignment; 2072 BdrvRequestPadding pad; 2073 int ret; 2074 bool padded = false; 2075 IO_CODE(); 2076 2077 trace_bdrv_co_pwritev_part(child->bs, offset, bytes, flags); 2078 2079 if (!bdrv_co_is_inserted(bs)) { 2080 return -ENOMEDIUM; 2081 } 2082 2083 if (flags & BDRV_REQ_ZERO_WRITE) { 2084 ret = bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, NULL); 2085 } else { 2086 ret = bdrv_check_request32(offset, bytes, qiov, qiov_offset); 2087 } 2088 if (ret < 0) { 2089 return ret; 2090 } 2091 2092 /* If the request is misaligned then we can't make it efficient */ 2093 if ((flags & BDRV_REQ_NO_FALLBACK) && 2094 !QEMU_IS_ALIGNED(offset | bytes, align)) 2095 { 2096 return -ENOTSUP; 2097 } 2098 2099 if (bytes == 0 && !QEMU_IS_ALIGNED(offset, bs->bl.request_alignment)) { 2100 /* 2101 * Aligning zero request is nonsense. Even if driver has special meaning 2102 * of zero-length (like qcow2_co_pwritev_compressed_part), we can't pass 2103 * it to driver due to request_alignment. 2104 * 2105 * Still, no reason to return an error if someone do unaligned 2106 * zero-length write occasionally. 2107 */ 2108 return 0; 2109 } 2110 2111 if (!(flags & BDRV_REQ_ZERO_WRITE)) { 2112 /* 2113 * Pad request for following read-modify-write cycle. 2114 * bdrv_co_do_zero_pwritev() does aligning by itself, so, we do 2115 * alignment only if there is no ZERO flag. 2116 */ 2117 ret = bdrv_pad_request(bs, &qiov, &qiov_offset, &offset, &bytes, &pad, 2118 &padded, &flags); 2119 if (ret < 0) { 2120 return ret; 2121 } 2122 } 2123 2124 bdrv_inc_in_flight(bs); 2125 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_WRITE); 2126 2127 if (flags & BDRV_REQ_ZERO_WRITE) { 2128 assert(!padded); 2129 ret = bdrv_co_do_zero_pwritev(child, offset, bytes, flags, &req); 2130 goto out; 2131 } 2132 2133 if (padded) { 2134 /* 2135 * Request was unaligned to request_alignment and therefore 2136 * padded. We are going to do read-modify-write, and must 2137 * serialize the request to prevent interactions of the 2138 * widened region with other transactions. 2139 */ 2140 assert(!(flags & BDRV_REQ_NO_WAIT)); 2141 bdrv_make_request_serialising(&req, align); 2142 bdrv_padding_rmw_read(child, &req, &pad, false); 2143 } 2144 2145 ret = bdrv_aligned_pwritev(child, &req, offset, bytes, align, 2146 qiov, qiov_offset, flags); 2147 2148 bdrv_padding_destroy(&pad); 2149 2150 out: 2151 tracked_request_end(&req); 2152 bdrv_dec_in_flight(bs); 2153 2154 return ret; 2155 } 2156 2157 int coroutine_fn bdrv_co_pwrite_zeroes(BdrvChild *child, int64_t offset, 2158 int64_t bytes, BdrvRequestFlags flags) 2159 { 2160 IO_CODE(); 2161 trace_bdrv_co_pwrite_zeroes(child->bs, offset, bytes, flags); 2162 2163 if (!(child->bs->open_flags & BDRV_O_UNMAP)) { 2164 flags &= ~BDRV_REQ_MAY_UNMAP; 2165 } 2166 2167 return bdrv_co_pwritev(child, offset, bytes, NULL, 2168 BDRV_REQ_ZERO_WRITE | flags); 2169 } 2170 2171 /* 2172 * Flush ALL BDSes regardless of if they are reachable via a BlkBackend or not. 2173 */ 2174 int bdrv_flush_all(void) 2175 { 2176 BdrvNextIterator it; 2177 BlockDriverState *bs = NULL; 2178 int result = 0; 2179 2180 GLOBAL_STATE_CODE(); 2181 2182 /* 2183 * bdrv queue is managed by record/replay, 2184 * creating new flush request for stopping 2185 * the VM may break the determinism 2186 */ 2187 if (replay_events_enabled()) { 2188 return result; 2189 } 2190 2191 for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) { 2192 AioContext *aio_context = bdrv_get_aio_context(bs); 2193 int ret; 2194 2195 aio_context_acquire(aio_context); 2196 ret = bdrv_flush(bs); 2197 if (ret < 0 && !result) { 2198 result = ret; 2199 } 2200 aio_context_release(aio_context); 2201 } 2202 2203 return result; 2204 } 2205 2206 /* 2207 * Returns the allocation status of the specified sectors. 2208 * Drivers not implementing the functionality are assumed to not support 2209 * backing files, hence all their sectors are reported as allocated. 2210 * 2211 * If 'want_zero' is true, the caller is querying for mapping 2212 * purposes, with a focus on valid BDRV_BLOCK_OFFSET_VALID, _DATA, and 2213 * _ZERO where possible; otherwise, the result favors larger 'pnum', 2214 * with a focus on accurate BDRV_BLOCK_ALLOCATED. 2215 * 2216 * If 'offset' is beyond the end of the disk image the return value is 2217 * BDRV_BLOCK_EOF and 'pnum' is set to 0. 2218 * 2219 * 'bytes' is the max value 'pnum' should be set to. If bytes goes 2220 * beyond the end of the disk image it will be clamped; if 'pnum' is set to 2221 * the end of the image, then the returned value will include BDRV_BLOCK_EOF. 2222 * 2223 * 'pnum' is set to the number of bytes (including and immediately 2224 * following the specified offset) that are easily known to be in the 2225 * same allocated/unallocated state. Note that a second call starting 2226 * at the original offset plus returned pnum may have the same status. 2227 * The returned value is non-zero on success except at end-of-file. 2228 * 2229 * Returns negative errno on failure. Otherwise, if the 2230 * BDRV_BLOCK_OFFSET_VALID bit is set, 'map' and 'file' (if non-NULL) are 2231 * set to the host mapping and BDS corresponding to the guest offset. 2232 */ 2233 static int coroutine_fn GRAPH_RDLOCK 2234 bdrv_co_block_status(BlockDriverState *bs, bool want_zero, 2235 int64_t offset, int64_t bytes, 2236 int64_t *pnum, int64_t *map, BlockDriverState **file) 2237 { 2238 int64_t total_size; 2239 int64_t n; /* bytes */ 2240 int ret; 2241 int64_t local_map = 0; 2242 BlockDriverState *local_file = NULL; 2243 int64_t aligned_offset, aligned_bytes; 2244 uint32_t align; 2245 bool has_filtered_child; 2246 2247 assert(pnum); 2248 assert_bdrv_graph_readable(); 2249 *pnum = 0; 2250 total_size = bdrv_getlength(bs); 2251 if (total_size < 0) { 2252 ret = total_size; 2253 goto early_out; 2254 } 2255 2256 if (offset >= total_size) { 2257 ret = BDRV_BLOCK_EOF; 2258 goto early_out; 2259 } 2260 if (!bytes) { 2261 ret = 0; 2262 goto early_out; 2263 } 2264 2265 n = total_size - offset; 2266 if (n < bytes) { 2267 bytes = n; 2268 } 2269 2270 /* Must be non-NULL or bdrv_getlength() would have failed */ 2271 assert(bs->drv); 2272 has_filtered_child = bdrv_filter_child(bs); 2273 if (!bs->drv->bdrv_co_block_status && !has_filtered_child) { 2274 *pnum = bytes; 2275 ret = BDRV_BLOCK_DATA | BDRV_BLOCK_ALLOCATED; 2276 if (offset + bytes == total_size) { 2277 ret |= BDRV_BLOCK_EOF; 2278 } 2279 if (bs->drv->protocol_name) { 2280 ret |= BDRV_BLOCK_OFFSET_VALID; 2281 local_map = offset; 2282 local_file = bs; 2283 } 2284 goto early_out; 2285 } 2286 2287 bdrv_inc_in_flight(bs); 2288 2289 /* Round out to request_alignment boundaries */ 2290 align = bs->bl.request_alignment; 2291 aligned_offset = QEMU_ALIGN_DOWN(offset, align); 2292 aligned_bytes = ROUND_UP(offset + bytes, align) - aligned_offset; 2293 2294 if (bs->drv->bdrv_co_block_status) { 2295 /* 2296 * Use the block-status cache only for protocol nodes: Format 2297 * drivers are generally quick to inquire the status, but protocol 2298 * drivers often need to get information from outside of qemu, so 2299 * we do not have control over the actual implementation. There 2300 * have been cases where inquiring the status took an unreasonably 2301 * long time, and we can do nothing in qemu to fix it. 2302 * This is especially problematic for images with large data areas, 2303 * because finding the few holes in them and giving them special 2304 * treatment does not gain much performance. Therefore, we try to 2305 * cache the last-identified data region. 2306 * 2307 * Second, limiting ourselves to protocol nodes allows us to assume 2308 * the block status for data regions to be DATA | OFFSET_VALID, and 2309 * that the host offset is the same as the guest offset. 2310 * 2311 * Note that it is possible that external writers zero parts of 2312 * the cached regions without the cache being invalidated, and so 2313 * we may report zeroes as data. This is not catastrophic, 2314 * however, because reporting zeroes as data is fine. 2315 */ 2316 if (QLIST_EMPTY(&bs->children) && 2317 bdrv_bsc_is_data(bs, aligned_offset, pnum)) 2318 { 2319 ret = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID; 2320 local_file = bs; 2321 local_map = aligned_offset; 2322 } else { 2323 ret = bs->drv->bdrv_co_block_status(bs, want_zero, aligned_offset, 2324 aligned_bytes, pnum, &local_map, 2325 &local_file); 2326 2327 /* 2328 * Note that checking QLIST_EMPTY(&bs->children) is also done when 2329 * the cache is queried above. Technically, we do not need to check 2330 * it here; the worst that can happen is that we fill the cache for 2331 * non-protocol nodes, and then it is never used. However, filling 2332 * the cache requires an RCU update, so double check here to avoid 2333 * such an update if possible. 2334 * 2335 * Check want_zero, because we only want to update the cache when we 2336 * have accurate information about what is zero and what is data. 2337 */ 2338 if (want_zero && 2339 ret == (BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID) && 2340 QLIST_EMPTY(&bs->children)) 2341 { 2342 /* 2343 * When a protocol driver reports BLOCK_OFFSET_VALID, the 2344 * returned local_map value must be the same as the offset we 2345 * have passed (aligned_offset), and local_bs must be the node 2346 * itself. 2347 * Assert this, because we follow this rule when reading from 2348 * the cache (see the `local_file = bs` and 2349 * `local_map = aligned_offset` assignments above), and the 2350 * result the cache delivers must be the same as the driver 2351 * would deliver. 2352 */ 2353 assert(local_file == bs); 2354 assert(local_map == aligned_offset); 2355 bdrv_bsc_fill(bs, aligned_offset, *pnum); 2356 } 2357 } 2358 } else { 2359 /* Default code for filters */ 2360 2361 local_file = bdrv_filter_bs(bs); 2362 assert(local_file); 2363 2364 *pnum = aligned_bytes; 2365 local_map = aligned_offset; 2366 ret = BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID; 2367 } 2368 if (ret < 0) { 2369 *pnum = 0; 2370 goto out; 2371 } 2372 2373 /* 2374 * The driver's result must be a non-zero multiple of request_alignment. 2375 * Clamp pnum and adjust map to original request. 2376 */ 2377 assert(*pnum && QEMU_IS_ALIGNED(*pnum, align) && 2378 align > offset - aligned_offset); 2379 if (ret & BDRV_BLOCK_RECURSE) { 2380 assert(ret & BDRV_BLOCK_DATA); 2381 assert(ret & BDRV_BLOCK_OFFSET_VALID); 2382 assert(!(ret & BDRV_BLOCK_ZERO)); 2383 } 2384 2385 *pnum -= offset - aligned_offset; 2386 if (*pnum > bytes) { 2387 *pnum = bytes; 2388 } 2389 if (ret & BDRV_BLOCK_OFFSET_VALID) { 2390 local_map += offset - aligned_offset; 2391 } 2392 2393 if (ret & BDRV_BLOCK_RAW) { 2394 assert(ret & BDRV_BLOCK_OFFSET_VALID && local_file); 2395 ret = bdrv_co_block_status(local_file, want_zero, local_map, 2396 *pnum, pnum, &local_map, &local_file); 2397 goto out; 2398 } 2399 2400 if (ret & (BDRV_BLOCK_DATA | BDRV_BLOCK_ZERO)) { 2401 ret |= BDRV_BLOCK_ALLOCATED; 2402 } else if (bs->drv->supports_backing) { 2403 BlockDriverState *cow_bs = bdrv_cow_bs(bs); 2404 2405 if (!cow_bs) { 2406 ret |= BDRV_BLOCK_ZERO; 2407 } else if (want_zero) { 2408 int64_t size2 = bdrv_getlength(cow_bs); 2409 2410 if (size2 >= 0 && offset >= size2) { 2411 ret |= BDRV_BLOCK_ZERO; 2412 } 2413 } 2414 } 2415 2416 if (want_zero && ret & BDRV_BLOCK_RECURSE && 2417 local_file && local_file != bs && 2418 (ret & BDRV_BLOCK_DATA) && !(ret & BDRV_BLOCK_ZERO) && 2419 (ret & BDRV_BLOCK_OFFSET_VALID)) { 2420 int64_t file_pnum; 2421 int ret2; 2422 2423 ret2 = bdrv_co_block_status(local_file, want_zero, local_map, 2424 *pnum, &file_pnum, NULL, NULL); 2425 if (ret2 >= 0) { 2426 /* Ignore errors. This is just providing extra information, it 2427 * is useful but not necessary. 2428 */ 2429 if (ret2 & BDRV_BLOCK_EOF && 2430 (!file_pnum || ret2 & BDRV_BLOCK_ZERO)) { 2431 /* 2432 * It is valid for the format block driver to read 2433 * beyond the end of the underlying file's current 2434 * size; such areas read as zero. 2435 */ 2436 ret |= BDRV_BLOCK_ZERO; 2437 } else { 2438 /* Limit request to the range reported by the protocol driver */ 2439 *pnum = file_pnum; 2440 ret |= (ret2 & BDRV_BLOCK_ZERO); 2441 } 2442 } 2443 } 2444 2445 out: 2446 bdrv_dec_in_flight(bs); 2447 if (ret >= 0 && offset + *pnum == total_size) { 2448 ret |= BDRV_BLOCK_EOF; 2449 } 2450 early_out: 2451 if (file) { 2452 *file = local_file; 2453 } 2454 if (map) { 2455 *map = local_map; 2456 } 2457 return ret; 2458 } 2459 2460 int coroutine_fn 2461 bdrv_co_common_block_status_above(BlockDriverState *bs, 2462 BlockDriverState *base, 2463 bool include_base, 2464 bool want_zero, 2465 int64_t offset, 2466 int64_t bytes, 2467 int64_t *pnum, 2468 int64_t *map, 2469 BlockDriverState **file, 2470 int *depth) 2471 { 2472 int ret; 2473 BlockDriverState *p; 2474 int64_t eof = 0; 2475 int dummy; 2476 IO_CODE(); 2477 2478 assert(!include_base || base); /* Can't include NULL base */ 2479 assert_bdrv_graph_readable(); 2480 2481 if (!depth) { 2482 depth = &dummy; 2483 } 2484 *depth = 0; 2485 2486 if (!include_base && bs == base) { 2487 *pnum = bytes; 2488 return 0; 2489 } 2490 2491 ret = bdrv_co_block_status(bs, want_zero, offset, bytes, pnum, map, file); 2492 ++*depth; 2493 if (ret < 0 || *pnum == 0 || ret & BDRV_BLOCK_ALLOCATED || bs == base) { 2494 return ret; 2495 } 2496 2497 if (ret & BDRV_BLOCK_EOF) { 2498 eof = offset + *pnum; 2499 } 2500 2501 assert(*pnum <= bytes); 2502 bytes = *pnum; 2503 2504 for (p = bdrv_filter_or_cow_bs(bs); include_base || p != base; 2505 p = bdrv_filter_or_cow_bs(p)) 2506 { 2507 ret = bdrv_co_block_status(p, want_zero, offset, bytes, pnum, map, 2508 file); 2509 ++*depth; 2510 if (ret < 0) { 2511 return ret; 2512 } 2513 if (*pnum == 0) { 2514 /* 2515 * The top layer deferred to this layer, and because this layer is 2516 * short, any zeroes that we synthesize beyond EOF behave as if they 2517 * were allocated at this layer. 2518 * 2519 * We don't include BDRV_BLOCK_EOF into ret, as upper layer may be 2520 * larger. We'll add BDRV_BLOCK_EOF if needed at function end, see 2521 * below. 2522 */ 2523 assert(ret & BDRV_BLOCK_EOF); 2524 *pnum = bytes; 2525 if (file) { 2526 *file = p; 2527 } 2528 ret = BDRV_BLOCK_ZERO | BDRV_BLOCK_ALLOCATED; 2529 break; 2530 } 2531 if (ret & BDRV_BLOCK_ALLOCATED) { 2532 /* 2533 * We've found the node and the status, we must break. 2534 * 2535 * Drop BDRV_BLOCK_EOF, as it's not for upper layer, which may be 2536 * larger. We'll add BDRV_BLOCK_EOF if needed at function end, see 2537 * below. 2538 */ 2539 ret &= ~BDRV_BLOCK_EOF; 2540 break; 2541 } 2542 2543 if (p == base) { 2544 assert(include_base); 2545 break; 2546 } 2547 2548 /* 2549 * OK, [offset, offset + *pnum) region is unallocated on this layer, 2550 * let's continue the diving. 2551 */ 2552 assert(*pnum <= bytes); 2553 bytes = *pnum; 2554 } 2555 2556 if (offset + *pnum == eof) { 2557 ret |= BDRV_BLOCK_EOF; 2558 } 2559 2560 return ret; 2561 } 2562 2563 int coroutine_fn bdrv_co_block_status_above(BlockDriverState *bs, 2564 BlockDriverState *base, 2565 int64_t offset, int64_t bytes, 2566 int64_t *pnum, int64_t *map, 2567 BlockDriverState **file) 2568 { 2569 IO_CODE(); 2570 return bdrv_co_common_block_status_above(bs, base, false, true, offset, 2571 bytes, pnum, map, file, NULL); 2572 } 2573 2574 int bdrv_block_status_above(BlockDriverState *bs, BlockDriverState *base, 2575 int64_t offset, int64_t bytes, int64_t *pnum, 2576 int64_t *map, BlockDriverState **file) 2577 { 2578 IO_CODE(); 2579 return bdrv_common_block_status_above(bs, base, false, true, offset, bytes, 2580 pnum, map, file, NULL); 2581 } 2582 2583 int bdrv_block_status(BlockDriverState *bs, int64_t offset, int64_t bytes, 2584 int64_t *pnum, int64_t *map, BlockDriverState **file) 2585 { 2586 IO_CODE(); 2587 return bdrv_block_status_above(bs, bdrv_filter_or_cow_bs(bs), 2588 offset, bytes, pnum, map, file); 2589 } 2590 2591 /* 2592 * Check @bs (and its backing chain) to see if the range defined 2593 * by @offset and @bytes is known to read as zeroes. 2594 * Return 1 if that is the case, 0 otherwise and -errno on error. 2595 * This test is meant to be fast rather than accurate so returning 0 2596 * does not guarantee non-zero data. 2597 */ 2598 int coroutine_fn bdrv_co_is_zero_fast(BlockDriverState *bs, int64_t offset, 2599 int64_t bytes) 2600 { 2601 int ret; 2602 int64_t pnum = bytes; 2603 IO_CODE(); 2604 2605 assume_graph_lock(); /* FIXME */ 2606 2607 if (!bytes) { 2608 return 1; 2609 } 2610 2611 ret = bdrv_co_common_block_status_above(bs, NULL, false, false, offset, 2612 bytes, &pnum, NULL, NULL, NULL); 2613 2614 if (ret < 0) { 2615 return ret; 2616 } 2617 2618 return (pnum == bytes) && (ret & BDRV_BLOCK_ZERO); 2619 } 2620 2621 int coroutine_fn bdrv_co_is_allocated(BlockDriverState *bs, int64_t offset, 2622 int64_t bytes, int64_t *pnum) 2623 { 2624 int ret; 2625 int64_t dummy; 2626 IO_CODE(); 2627 2628 ret = bdrv_co_common_block_status_above(bs, bs, true, false, offset, 2629 bytes, pnum ? pnum : &dummy, NULL, 2630 NULL, NULL); 2631 if (ret < 0) { 2632 return ret; 2633 } 2634 return !!(ret & BDRV_BLOCK_ALLOCATED); 2635 } 2636 2637 int bdrv_is_allocated(BlockDriverState *bs, int64_t offset, int64_t bytes, 2638 int64_t *pnum) 2639 { 2640 int ret; 2641 int64_t dummy; 2642 IO_CODE(); 2643 2644 ret = bdrv_common_block_status_above(bs, bs, true, false, offset, 2645 bytes, pnum ? pnum : &dummy, NULL, 2646 NULL, NULL); 2647 if (ret < 0) { 2648 return ret; 2649 } 2650 return !!(ret & BDRV_BLOCK_ALLOCATED); 2651 } 2652 2653 /* See bdrv_is_allocated_above for documentation */ 2654 int coroutine_fn bdrv_co_is_allocated_above(BlockDriverState *top, 2655 BlockDriverState *base, 2656 bool include_base, int64_t offset, 2657 int64_t bytes, int64_t *pnum) 2658 { 2659 int depth; 2660 int ret; 2661 IO_CODE(); 2662 2663 ret = bdrv_co_common_block_status_above(top, base, include_base, false, 2664 offset, bytes, pnum, NULL, NULL, 2665 &depth); 2666 if (ret < 0) { 2667 return ret; 2668 } 2669 2670 if (ret & BDRV_BLOCK_ALLOCATED) { 2671 return depth; 2672 } 2673 return 0; 2674 } 2675 2676 /* 2677 * Given an image chain: ... -> [BASE] -> [INTER1] -> [INTER2] -> [TOP] 2678 * 2679 * Return a positive depth if (a prefix of) the given range is allocated 2680 * in any image between BASE and TOP (BASE is only included if include_base 2681 * is set). Depth 1 is TOP, 2 is the first backing layer, and so forth. 2682 * BASE can be NULL to check if the given offset is allocated in any 2683 * image of the chain. Return 0 otherwise, or negative errno on 2684 * failure. 2685 * 2686 * 'pnum' is set to the number of bytes (including and immediately 2687 * following the specified offset) that are known to be in the same 2688 * allocated/unallocated state. Note that a subsequent call starting 2689 * at 'offset + *pnum' may return the same allocation status (in other 2690 * words, the result is not necessarily the maximum possible range); 2691 * but 'pnum' will only be 0 when end of file is reached. 2692 */ 2693 int bdrv_is_allocated_above(BlockDriverState *top, 2694 BlockDriverState *base, 2695 bool include_base, int64_t offset, 2696 int64_t bytes, int64_t *pnum) 2697 { 2698 int depth; 2699 int ret; 2700 IO_CODE(); 2701 2702 ret = bdrv_common_block_status_above(top, base, include_base, false, 2703 offset, bytes, pnum, NULL, NULL, 2704 &depth); 2705 if (ret < 0) { 2706 return ret; 2707 } 2708 2709 if (ret & BDRV_BLOCK_ALLOCATED) { 2710 return depth; 2711 } 2712 return 0; 2713 } 2714 2715 int coroutine_fn 2716 bdrv_co_readv_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos) 2717 { 2718 BlockDriver *drv = bs->drv; 2719 BlockDriverState *child_bs = bdrv_primary_bs(bs); 2720 int ret; 2721 IO_CODE(); 2722 assert_bdrv_graph_readable(); 2723 2724 ret = bdrv_check_qiov_request(pos, qiov->size, qiov, 0, NULL); 2725 if (ret < 0) { 2726 return ret; 2727 } 2728 2729 if (!drv) { 2730 return -ENOMEDIUM; 2731 } 2732 2733 bdrv_inc_in_flight(bs); 2734 2735 if (drv->bdrv_co_load_vmstate) { 2736 ret = drv->bdrv_co_load_vmstate(bs, qiov, pos); 2737 } else if (child_bs) { 2738 ret = bdrv_co_readv_vmstate(child_bs, qiov, pos); 2739 } else { 2740 ret = -ENOTSUP; 2741 } 2742 2743 bdrv_dec_in_flight(bs); 2744 2745 return ret; 2746 } 2747 2748 int coroutine_fn 2749 bdrv_co_writev_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos) 2750 { 2751 BlockDriver *drv = bs->drv; 2752 BlockDriverState *child_bs = bdrv_primary_bs(bs); 2753 int ret; 2754 IO_CODE(); 2755 assert_bdrv_graph_readable(); 2756 2757 ret = bdrv_check_qiov_request(pos, qiov->size, qiov, 0, NULL); 2758 if (ret < 0) { 2759 return ret; 2760 } 2761 2762 if (!drv) { 2763 return -ENOMEDIUM; 2764 } 2765 2766 bdrv_inc_in_flight(bs); 2767 2768 if (drv->bdrv_co_save_vmstate) { 2769 ret = drv->bdrv_co_save_vmstate(bs, qiov, pos); 2770 } else if (child_bs) { 2771 ret = bdrv_co_writev_vmstate(child_bs, qiov, pos); 2772 } else { 2773 ret = -ENOTSUP; 2774 } 2775 2776 bdrv_dec_in_flight(bs); 2777 2778 return ret; 2779 } 2780 2781 int bdrv_save_vmstate(BlockDriverState *bs, const uint8_t *buf, 2782 int64_t pos, int size) 2783 { 2784 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, size); 2785 int ret = bdrv_writev_vmstate(bs, &qiov, pos); 2786 IO_CODE(); 2787 2788 return ret < 0 ? ret : size; 2789 } 2790 2791 int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf, 2792 int64_t pos, int size) 2793 { 2794 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, size); 2795 int ret = bdrv_readv_vmstate(bs, &qiov, pos); 2796 IO_CODE(); 2797 2798 return ret < 0 ? ret : size; 2799 } 2800 2801 /**************************************************************/ 2802 /* async I/Os */ 2803 2804 void bdrv_aio_cancel(BlockAIOCB *acb) 2805 { 2806 IO_CODE(); 2807 qemu_aio_ref(acb); 2808 bdrv_aio_cancel_async(acb); 2809 while (acb->refcnt > 1) { 2810 if (acb->aiocb_info->get_aio_context) { 2811 aio_poll(acb->aiocb_info->get_aio_context(acb), true); 2812 } else if (acb->bs) { 2813 /* qemu_aio_ref and qemu_aio_unref are not thread-safe, so 2814 * assert that we're not using an I/O thread. Thread-safe 2815 * code should use bdrv_aio_cancel_async exclusively. 2816 */ 2817 assert(bdrv_get_aio_context(acb->bs) == qemu_get_aio_context()); 2818 aio_poll(bdrv_get_aio_context(acb->bs), true); 2819 } else { 2820 abort(); 2821 } 2822 } 2823 qemu_aio_unref(acb); 2824 } 2825 2826 /* Async version of aio cancel. The caller is not blocked if the acb implements 2827 * cancel_async, otherwise we do nothing and let the request normally complete. 2828 * In either case the completion callback must be called. */ 2829 void bdrv_aio_cancel_async(BlockAIOCB *acb) 2830 { 2831 IO_CODE(); 2832 if (acb->aiocb_info->cancel_async) { 2833 acb->aiocb_info->cancel_async(acb); 2834 } 2835 } 2836 2837 /**************************************************************/ 2838 /* Coroutine block device emulation */ 2839 2840 int coroutine_fn bdrv_co_flush(BlockDriverState *bs) 2841 { 2842 BdrvChild *primary_child = bdrv_primary_child(bs); 2843 BdrvChild *child; 2844 int current_gen; 2845 int ret = 0; 2846 IO_CODE(); 2847 2848 assert_bdrv_graph_readable(); 2849 bdrv_inc_in_flight(bs); 2850 2851 if (!bdrv_co_is_inserted(bs) || bdrv_is_read_only(bs) || 2852 bdrv_is_sg(bs)) { 2853 goto early_exit; 2854 } 2855 2856 qemu_co_mutex_lock(&bs->reqs_lock); 2857 current_gen = qatomic_read(&bs->write_gen); 2858 2859 /* Wait until any previous flushes are completed */ 2860 while (bs->active_flush_req) { 2861 qemu_co_queue_wait(&bs->flush_queue, &bs->reqs_lock); 2862 } 2863 2864 /* Flushes reach this point in nondecreasing current_gen order. */ 2865 bs->active_flush_req = true; 2866 qemu_co_mutex_unlock(&bs->reqs_lock); 2867 2868 /* Write back all layers by calling one driver function */ 2869 if (bs->drv->bdrv_co_flush) { 2870 ret = bs->drv->bdrv_co_flush(bs); 2871 goto out; 2872 } 2873 2874 /* Write back cached data to the OS even with cache=unsafe */ 2875 BLKDBG_EVENT(primary_child, BLKDBG_FLUSH_TO_OS); 2876 if (bs->drv->bdrv_co_flush_to_os) { 2877 ret = bs->drv->bdrv_co_flush_to_os(bs); 2878 if (ret < 0) { 2879 goto out; 2880 } 2881 } 2882 2883 /* But don't actually force it to the disk with cache=unsafe */ 2884 if (bs->open_flags & BDRV_O_NO_FLUSH) { 2885 goto flush_children; 2886 } 2887 2888 /* Check if we really need to flush anything */ 2889 if (bs->flushed_gen == current_gen) { 2890 goto flush_children; 2891 } 2892 2893 BLKDBG_EVENT(primary_child, BLKDBG_FLUSH_TO_DISK); 2894 if (!bs->drv) { 2895 /* bs->drv->bdrv_co_flush() might have ejected the BDS 2896 * (even in case of apparent success) */ 2897 ret = -ENOMEDIUM; 2898 goto out; 2899 } 2900 if (bs->drv->bdrv_co_flush_to_disk) { 2901 ret = bs->drv->bdrv_co_flush_to_disk(bs); 2902 } else if (bs->drv->bdrv_aio_flush) { 2903 BlockAIOCB *acb; 2904 CoroutineIOCompletion co = { 2905 .coroutine = qemu_coroutine_self(), 2906 }; 2907 2908 acb = bs->drv->bdrv_aio_flush(bs, bdrv_co_io_em_complete, &co); 2909 if (acb == NULL) { 2910 ret = -EIO; 2911 } else { 2912 qemu_coroutine_yield(); 2913 ret = co.ret; 2914 } 2915 } else { 2916 /* 2917 * Some block drivers always operate in either writethrough or unsafe 2918 * mode and don't support bdrv_flush therefore. Usually qemu doesn't 2919 * know how the server works (because the behaviour is hardcoded or 2920 * depends on server-side configuration), so we can't ensure that 2921 * everything is safe on disk. Returning an error doesn't work because 2922 * that would break guests even if the server operates in writethrough 2923 * mode. 2924 * 2925 * Let's hope the user knows what he's doing. 2926 */ 2927 ret = 0; 2928 } 2929 2930 if (ret < 0) { 2931 goto out; 2932 } 2933 2934 /* Now flush the underlying protocol. It will also have BDRV_O_NO_FLUSH 2935 * in the case of cache=unsafe, so there are no useless flushes. 2936 */ 2937 flush_children: 2938 ret = 0; 2939 QLIST_FOREACH(child, &bs->children, next) { 2940 if (child->perm & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED)) { 2941 int this_child_ret = bdrv_co_flush(child->bs); 2942 if (!ret) { 2943 ret = this_child_ret; 2944 } 2945 } 2946 } 2947 2948 out: 2949 /* Notify any pending flushes that we have completed */ 2950 if (ret == 0) { 2951 bs->flushed_gen = current_gen; 2952 } 2953 2954 qemu_co_mutex_lock(&bs->reqs_lock); 2955 bs->active_flush_req = false; 2956 /* Return value is ignored - it's ok if wait queue is empty */ 2957 qemu_co_queue_next(&bs->flush_queue); 2958 qemu_co_mutex_unlock(&bs->reqs_lock); 2959 2960 early_exit: 2961 bdrv_dec_in_flight(bs); 2962 return ret; 2963 } 2964 2965 int coroutine_fn bdrv_co_pdiscard(BdrvChild *child, int64_t offset, 2966 int64_t bytes) 2967 { 2968 BdrvTrackedRequest req; 2969 int ret; 2970 int64_t max_pdiscard; 2971 int head, tail, align; 2972 BlockDriverState *bs = child->bs; 2973 IO_CODE(); 2974 2975 if (!bs || !bs->drv || !bdrv_co_is_inserted(bs)) { 2976 return -ENOMEDIUM; 2977 } 2978 2979 if (bdrv_has_readonly_bitmaps(bs)) { 2980 return -EPERM; 2981 } 2982 2983 ret = bdrv_check_request(offset, bytes, NULL); 2984 if (ret < 0) { 2985 return ret; 2986 } 2987 2988 /* Do nothing if disabled. */ 2989 if (!(bs->open_flags & BDRV_O_UNMAP)) { 2990 return 0; 2991 } 2992 2993 if (!bs->drv->bdrv_co_pdiscard && !bs->drv->bdrv_aio_pdiscard) { 2994 return 0; 2995 } 2996 2997 /* Invalidate the cached block-status data range if this discard overlaps */ 2998 bdrv_bsc_invalidate_range(bs, offset, bytes); 2999 3000 /* Discard is advisory, but some devices track and coalesce 3001 * unaligned requests, so we must pass everything down rather than 3002 * round here. Still, most devices will just silently ignore 3003 * unaligned requests (by returning -ENOTSUP), so we must fragment 3004 * the request accordingly. */ 3005 align = MAX(bs->bl.pdiscard_alignment, bs->bl.request_alignment); 3006 assert(align % bs->bl.request_alignment == 0); 3007 head = offset % align; 3008 tail = (offset + bytes) % align; 3009 3010 bdrv_inc_in_flight(bs); 3011 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_DISCARD); 3012 3013 ret = bdrv_co_write_req_prepare(child, offset, bytes, &req, 0); 3014 if (ret < 0) { 3015 goto out; 3016 } 3017 3018 max_pdiscard = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_pdiscard, INT64_MAX), 3019 align); 3020 assert(max_pdiscard >= bs->bl.request_alignment); 3021 3022 while (bytes > 0) { 3023 int64_t num = bytes; 3024 3025 if (head) { 3026 /* Make small requests to get to alignment boundaries. */ 3027 num = MIN(bytes, align - head); 3028 if (!QEMU_IS_ALIGNED(num, bs->bl.request_alignment)) { 3029 num %= bs->bl.request_alignment; 3030 } 3031 head = (head + num) % align; 3032 assert(num < max_pdiscard); 3033 } else if (tail) { 3034 if (num > align) { 3035 /* Shorten the request to the last aligned cluster. */ 3036 num -= tail; 3037 } else if (!QEMU_IS_ALIGNED(tail, bs->bl.request_alignment) && 3038 tail > bs->bl.request_alignment) { 3039 tail %= bs->bl.request_alignment; 3040 num -= tail; 3041 } 3042 } 3043 /* limit request size */ 3044 if (num > max_pdiscard) { 3045 num = max_pdiscard; 3046 } 3047 3048 if (!bs->drv) { 3049 ret = -ENOMEDIUM; 3050 goto out; 3051 } 3052 if (bs->drv->bdrv_co_pdiscard) { 3053 ret = bs->drv->bdrv_co_pdiscard(bs, offset, num); 3054 } else { 3055 BlockAIOCB *acb; 3056 CoroutineIOCompletion co = { 3057 .coroutine = qemu_coroutine_self(), 3058 }; 3059 3060 acb = bs->drv->bdrv_aio_pdiscard(bs, offset, num, 3061 bdrv_co_io_em_complete, &co); 3062 if (acb == NULL) { 3063 ret = -EIO; 3064 goto out; 3065 } else { 3066 qemu_coroutine_yield(); 3067 ret = co.ret; 3068 } 3069 } 3070 if (ret && ret != -ENOTSUP) { 3071 goto out; 3072 } 3073 3074 offset += num; 3075 bytes -= num; 3076 } 3077 ret = 0; 3078 out: 3079 bdrv_co_write_req_finish(child, req.offset, req.bytes, &req, ret); 3080 tracked_request_end(&req); 3081 bdrv_dec_in_flight(bs); 3082 return ret; 3083 } 3084 3085 int coroutine_fn bdrv_co_ioctl(BlockDriverState *bs, int req, void *buf) 3086 { 3087 BlockDriver *drv = bs->drv; 3088 CoroutineIOCompletion co = { 3089 .coroutine = qemu_coroutine_self(), 3090 }; 3091 BlockAIOCB *acb; 3092 IO_CODE(); 3093 assert_bdrv_graph_readable(); 3094 3095 bdrv_inc_in_flight(bs); 3096 if (!drv || (!drv->bdrv_aio_ioctl && !drv->bdrv_co_ioctl)) { 3097 co.ret = -ENOTSUP; 3098 goto out; 3099 } 3100 3101 if (drv->bdrv_co_ioctl) { 3102 co.ret = drv->bdrv_co_ioctl(bs, req, buf); 3103 } else { 3104 acb = drv->bdrv_aio_ioctl(bs, req, buf, bdrv_co_io_em_complete, &co); 3105 if (!acb) { 3106 co.ret = -ENOTSUP; 3107 goto out; 3108 } 3109 qemu_coroutine_yield(); 3110 } 3111 out: 3112 bdrv_dec_in_flight(bs); 3113 return co.ret; 3114 } 3115 3116 void *qemu_blockalign(BlockDriverState *bs, size_t size) 3117 { 3118 IO_CODE(); 3119 return qemu_memalign(bdrv_opt_mem_align(bs), size); 3120 } 3121 3122 void *qemu_blockalign0(BlockDriverState *bs, size_t size) 3123 { 3124 IO_CODE(); 3125 return memset(qemu_blockalign(bs, size), 0, size); 3126 } 3127 3128 void *qemu_try_blockalign(BlockDriverState *bs, size_t size) 3129 { 3130 size_t align = bdrv_opt_mem_align(bs); 3131 IO_CODE(); 3132 3133 /* Ensure that NULL is never returned on success */ 3134 assert(align > 0); 3135 if (size == 0) { 3136 size = align; 3137 } 3138 3139 return qemu_try_memalign(align, size); 3140 } 3141 3142 void *qemu_try_blockalign0(BlockDriverState *bs, size_t size) 3143 { 3144 void *mem = qemu_try_blockalign(bs, size); 3145 IO_CODE(); 3146 3147 if (mem) { 3148 memset(mem, 0, size); 3149 } 3150 3151 return mem; 3152 } 3153 3154 void coroutine_fn bdrv_co_io_plug(BlockDriverState *bs) 3155 { 3156 BdrvChild *child; 3157 IO_CODE(); 3158 3159 QLIST_FOREACH(child, &bs->children, next) { 3160 bdrv_co_io_plug(child->bs); 3161 } 3162 3163 if (qatomic_fetch_inc(&bs->io_plugged) == 0) { 3164 BlockDriver *drv = bs->drv; 3165 if (drv && drv->bdrv_co_io_plug) { 3166 drv->bdrv_co_io_plug(bs); 3167 } 3168 } 3169 } 3170 3171 void coroutine_fn bdrv_co_io_unplug(BlockDriverState *bs) 3172 { 3173 BdrvChild *child; 3174 IO_CODE(); 3175 3176 assert(bs->io_plugged); 3177 if (qatomic_fetch_dec(&bs->io_plugged) == 1) { 3178 BlockDriver *drv = bs->drv; 3179 if (drv && drv->bdrv_co_io_unplug) { 3180 drv->bdrv_co_io_unplug(bs); 3181 } 3182 } 3183 3184 QLIST_FOREACH(child, &bs->children, next) { 3185 bdrv_co_io_unplug(child->bs); 3186 } 3187 } 3188 3189 /* Helper that undoes bdrv_register_buf() when it fails partway through */ 3190 static void bdrv_register_buf_rollback(BlockDriverState *bs, 3191 void *host, 3192 size_t size, 3193 BdrvChild *final_child) 3194 { 3195 BdrvChild *child; 3196 3197 QLIST_FOREACH(child, &bs->children, next) { 3198 if (child == final_child) { 3199 break; 3200 } 3201 3202 bdrv_unregister_buf(child->bs, host, size); 3203 } 3204 3205 if (bs->drv && bs->drv->bdrv_unregister_buf) { 3206 bs->drv->bdrv_unregister_buf(bs, host, size); 3207 } 3208 } 3209 3210 bool bdrv_register_buf(BlockDriverState *bs, void *host, size_t size, 3211 Error **errp) 3212 { 3213 BdrvChild *child; 3214 3215 GLOBAL_STATE_CODE(); 3216 if (bs->drv && bs->drv->bdrv_register_buf) { 3217 if (!bs->drv->bdrv_register_buf(bs, host, size, errp)) { 3218 return false; 3219 } 3220 } 3221 QLIST_FOREACH(child, &bs->children, next) { 3222 if (!bdrv_register_buf(child->bs, host, size, errp)) { 3223 bdrv_register_buf_rollback(bs, host, size, child); 3224 return false; 3225 } 3226 } 3227 return true; 3228 } 3229 3230 void bdrv_unregister_buf(BlockDriverState *bs, void *host, size_t size) 3231 { 3232 BdrvChild *child; 3233 3234 GLOBAL_STATE_CODE(); 3235 if (bs->drv && bs->drv->bdrv_unregister_buf) { 3236 bs->drv->bdrv_unregister_buf(bs, host, size); 3237 } 3238 QLIST_FOREACH(child, &bs->children, next) { 3239 bdrv_unregister_buf(child->bs, host, size); 3240 } 3241 } 3242 3243 static int coroutine_fn bdrv_co_copy_range_internal( 3244 BdrvChild *src, int64_t src_offset, BdrvChild *dst, 3245 int64_t dst_offset, int64_t bytes, 3246 BdrvRequestFlags read_flags, BdrvRequestFlags write_flags, 3247 bool recurse_src) 3248 { 3249 BdrvTrackedRequest req; 3250 int ret; 3251 3252 /* TODO We can support BDRV_REQ_NO_FALLBACK here */ 3253 assert(!(read_flags & BDRV_REQ_NO_FALLBACK)); 3254 assert(!(write_flags & BDRV_REQ_NO_FALLBACK)); 3255 assert(!(read_flags & BDRV_REQ_NO_WAIT)); 3256 assert(!(write_flags & BDRV_REQ_NO_WAIT)); 3257 3258 if (!dst || !dst->bs || !bdrv_co_is_inserted(dst->bs)) { 3259 return -ENOMEDIUM; 3260 } 3261 ret = bdrv_check_request32(dst_offset, bytes, NULL, 0); 3262 if (ret) { 3263 return ret; 3264 } 3265 if (write_flags & BDRV_REQ_ZERO_WRITE) { 3266 return bdrv_co_pwrite_zeroes(dst, dst_offset, bytes, write_flags); 3267 } 3268 3269 if (!src || !src->bs || !bdrv_co_is_inserted(src->bs)) { 3270 return -ENOMEDIUM; 3271 } 3272 ret = bdrv_check_request32(src_offset, bytes, NULL, 0); 3273 if (ret) { 3274 return ret; 3275 } 3276 3277 if (!src->bs->drv->bdrv_co_copy_range_from 3278 || !dst->bs->drv->bdrv_co_copy_range_to 3279 || src->bs->encrypted || dst->bs->encrypted) { 3280 return -ENOTSUP; 3281 } 3282 3283 if (recurse_src) { 3284 bdrv_inc_in_flight(src->bs); 3285 tracked_request_begin(&req, src->bs, src_offset, bytes, 3286 BDRV_TRACKED_READ); 3287 3288 /* BDRV_REQ_SERIALISING is only for write operation */ 3289 assert(!(read_flags & BDRV_REQ_SERIALISING)); 3290 bdrv_wait_serialising_requests(&req); 3291 3292 ret = src->bs->drv->bdrv_co_copy_range_from(src->bs, 3293 src, src_offset, 3294 dst, dst_offset, 3295 bytes, 3296 read_flags, write_flags); 3297 3298 tracked_request_end(&req); 3299 bdrv_dec_in_flight(src->bs); 3300 } else { 3301 bdrv_inc_in_flight(dst->bs); 3302 tracked_request_begin(&req, dst->bs, dst_offset, bytes, 3303 BDRV_TRACKED_WRITE); 3304 ret = bdrv_co_write_req_prepare(dst, dst_offset, bytes, &req, 3305 write_flags); 3306 if (!ret) { 3307 ret = dst->bs->drv->bdrv_co_copy_range_to(dst->bs, 3308 src, src_offset, 3309 dst, dst_offset, 3310 bytes, 3311 read_flags, write_flags); 3312 } 3313 bdrv_co_write_req_finish(dst, dst_offset, bytes, &req, ret); 3314 tracked_request_end(&req); 3315 bdrv_dec_in_flight(dst->bs); 3316 } 3317 3318 return ret; 3319 } 3320 3321 /* Copy range from @src to @dst. 3322 * 3323 * See the comment of bdrv_co_copy_range for the parameter and return value 3324 * semantics. */ 3325 int coroutine_fn bdrv_co_copy_range_from(BdrvChild *src, int64_t src_offset, 3326 BdrvChild *dst, int64_t dst_offset, 3327 int64_t bytes, 3328 BdrvRequestFlags read_flags, 3329 BdrvRequestFlags write_flags) 3330 { 3331 IO_CODE(); 3332 trace_bdrv_co_copy_range_from(src, src_offset, dst, dst_offset, bytes, 3333 read_flags, write_flags); 3334 return bdrv_co_copy_range_internal(src, src_offset, dst, dst_offset, 3335 bytes, read_flags, write_flags, true); 3336 } 3337 3338 /* Copy range from @src to @dst. 3339 * 3340 * See the comment of bdrv_co_copy_range for the parameter and return value 3341 * semantics. */ 3342 int coroutine_fn bdrv_co_copy_range_to(BdrvChild *src, int64_t src_offset, 3343 BdrvChild *dst, int64_t dst_offset, 3344 int64_t bytes, 3345 BdrvRequestFlags read_flags, 3346 BdrvRequestFlags write_flags) 3347 { 3348 IO_CODE(); 3349 trace_bdrv_co_copy_range_to(src, src_offset, dst, dst_offset, bytes, 3350 read_flags, write_flags); 3351 return bdrv_co_copy_range_internal(src, src_offset, dst, dst_offset, 3352 bytes, read_flags, write_flags, false); 3353 } 3354 3355 int coroutine_fn bdrv_co_copy_range(BdrvChild *src, int64_t src_offset, 3356 BdrvChild *dst, int64_t dst_offset, 3357 int64_t bytes, BdrvRequestFlags read_flags, 3358 BdrvRequestFlags write_flags) 3359 { 3360 IO_CODE(); 3361 return bdrv_co_copy_range_from(src, src_offset, 3362 dst, dst_offset, 3363 bytes, read_flags, write_flags); 3364 } 3365 3366 static void bdrv_parent_cb_resize(BlockDriverState *bs) 3367 { 3368 BdrvChild *c; 3369 QLIST_FOREACH(c, &bs->parents, next_parent) { 3370 if (c->klass->resize) { 3371 c->klass->resize(c); 3372 } 3373 } 3374 } 3375 3376 /** 3377 * Truncate file to 'offset' bytes (needed only for file protocols) 3378 * 3379 * If 'exact' is true, the file must be resized to exactly the given 3380 * 'offset'. Otherwise, it is sufficient for the node to be at least 3381 * 'offset' bytes in length. 3382 */ 3383 int coroutine_fn bdrv_co_truncate(BdrvChild *child, int64_t offset, bool exact, 3384 PreallocMode prealloc, BdrvRequestFlags flags, 3385 Error **errp) 3386 { 3387 BlockDriverState *bs = child->bs; 3388 BdrvChild *filtered, *backing; 3389 BlockDriver *drv = bs->drv; 3390 BdrvTrackedRequest req; 3391 int64_t old_size, new_bytes; 3392 int ret; 3393 IO_CODE(); 3394 assert_bdrv_graph_readable(); 3395 3396 /* if bs->drv == NULL, bs is closed, so there's nothing to do here */ 3397 if (!drv) { 3398 error_setg(errp, "No medium inserted"); 3399 return -ENOMEDIUM; 3400 } 3401 if (offset < 0) { 3402 error_setg(errp, "Image size cannot be negative"); 3403 return -EINVAL; 3404 } 3405 3406 ret = bdrv_check_request(offset, 0, errp); 3407 if (ret < 0) { 3408 return ret; 3409 } 3410 3411 old_size = bdrv_getlength(bs); 3412 if (old_size < 0) { 3413 error_setg_errno(errp, -old_size, "Failed to get old image size"); 3414 return old_size; 3415 } 3416 3417 if (bdrv_is_read_only(bs)) { 3418 error_setg(errp, "Image is read-only"); 3419 return -EACCES; 3420 } 3421 3422 if (offset > old_size) { 3423 new_bytes = offset - old_size; 3424 } else { 3425 new_bytes = 0; 3426 } 3427 3428 bdrv_inc_in_flight(bs); 3429 tracked_request_begin(&req, bs, offset - new_bytes, new_bytes, 3430 BDRV_TRACKED_TRUNCATE); 3431 3432 /* If we are growing the image and potentially using preallocation for the 3433 * new area, we need to make sure that no write requests are made to it 3434 * concurrently or they might be overwritten by preallocation. */ 3435 if (new_bytes) { 3436 bdrv_make_request_serialising(&req, 1); 3437 } 3438 ret = bdrv_co_write_req_prepare(child, offset - new_bytes, new_bytes, &req, 3439 0); 3440 if (ret < 0) { 3441 error_setg_errno(errp, -ret, 3442 "Failed to prepare request for truncation"); 3443 goto out; 3444 } 3445 3446 filtered = bdrv_filter_child(bs); 3447 backing = bdrv_cow_child(bs); 3448 3449 /* 3450 * If the image has a backing file that is large enough that it would 3451 * provide data for the new area, we cannot leave it unallocated because 3452 * then the backing file content would become visible. Instead, zero-fill 3453 * the new area. 3454 * 3455 * Note that if the image has a backing file, but was opened without the 3456 * backing file, taking care of keeping things consistent with that backing 3457 * file is the user's responsibility. 3458 */ 3459 if (new_bytes && backing) { 3460 int64_t backing_len; 3461 3462 backing_len = bdrv_co_getlength(backing->bs); 3463 if (backing_len < 0) { 3464 ret = backing_len; 3465 error_setg_errno(errp, -ret, "Could not get backing file size"); 3466 goto out; 3467 } 3468 3469 if (backing_len > old_size) { 3470 flags |= BDRV_REQ_ZERO_WRITE; 3471 } 3472 } 3473 3474 if (drv->bdrv_co_truncate) { 3475 if (flags & ~bs->supported_truncate_flags) { 3476 error_setg(errp, "Block driver does not support requested flags"); 3477 ret = -ENOTSUP; 3478 goto out; 3479 } 3480 ret = drv->bdrv_co_truncate(bs, offset, exact, prealloc, flags, errp); 3481 } else if (filtered) { 3482 ret = bdrv_co_truncate(filtered, offset, exact, prealloc, flags, errp); 3483 } else { 3484 error_setg(errp, "Image format driver does not support resize"); 3485 ret = -ENOTSUP; 3486 goto out; 3487 } 3488 if (ret < 0) { 3489 goto out; 3490 } 3491 3492 ret = bdrv_co_refresh_total_sectors(bs, offset >> BDRV_SECTOR_BITS); 3493 if (ret < 0) { 3494 error_setg_errno(errp, -ret, "Could not refresh total sector count"); 3495 } else { 3496 offset = bs->total_sectors * BDRV_SECTOR_SIZE; 3497 } 3498 /* 3499 * It's possible that truncation succeeded but bdrv_refresh_total_sectors 3500 * failed, but the latter doesn't affect how we should finish the request. 3501 * Pass 0 as the last parameter so that dirty bitmaps etc. are handled. 3502 */ 3503 bdrv_co_write_req_finish(child, offset - new_bytes, new_bytes, &req, 0); 3504 3505 out: 3506 tracked_request_end(&req); 3507 bdrv_dec_in_flight(bs); 3508 3509 return ret; 3510 } 3511 3512 void bdrv_cancel_in_flight(BlockDriverState *bs) 3513 { 3514 GLOBAL_STATE_CODE(); 3515 if (!bs || !bs->drv) { 3516 return; 3517 } 3518 3519 if (bs->drv->bdrv_cancel_in_flight) { 3520 bs->drv->bdrv_cancel_in_flight(bs); 3521 } 3522 } 3523 3524 int coroutine_fn 3525 bdrv_co_preadv_snapshot(BdrvChild *child, int64_t offset, int64_t bytes, 3526 QEMUIOVector *qiov, size_t qiov_offset) 3527 { 3528 BlockDriverState *bs = child->bs; 3529 BlockDriver *drv = bs->drv; 3530 int ret; 3531 IO_CODE(); 3532 3533 if (!drv) { 3534 return -ENOMEDIUM; 3535 } 3536 3537 if (!drv->bdrv_co_preadv_snapshot) { 3538 return -ENOTSUP; 3539 } 3540 3541 bdrv_inc_in_flight(bs); 3542 ret = drv->bdrv_co_preadv_snapshot(bs, offset, bytes, qiov, qiov_offset); 3543 bdrv_dec_in_flight(bs); 3544 3545 return ret; 3546 } 3547 3548 int coroutine_fn 3549 bdrv_co_snapshot_block_status(BlockDriverState *bs, 3550 bool want_zero, int64_t offset, int64_t bytes, 3551 int64_t *pnum, int64_t *map, 3552 BlockDriverState **file) 3553 { 3554 BlockDriver *drv = bs->drv; 3555 int ret; 3556 IO_CODE(); 3557 3558 if (!drv) { 3559 return -ENOMEDIUM; 3560 } 3561 3562 if (!drv->bdrv_co_snapshot_block_status) { 3563 return -ENOTSUP; 3564 } 3565 3566 bdrv_inc_in_flight(bs); 3567 ret = drv->bdrv_co_snapshot_block_status(bs, want_zero, offset, bytes, 3568 pnum, map, file); 3569 bdrv_dec_in_flight(bs); 3570 3571 return ret; 3572 } 3573 3574 int coroutine_fn 3575 bdrv_co_pdiscard_snapshot(BlockDriverState *bs, int64_t offset, int64_t bytes) 3576 { 3577 BlockDriver *drv = bs->drv; 3578 int ret; 3579 IO_CODE(); 3580 3581 if (!drv) { 3582 return -ENOMEDIUM; 3583 } 3584 3585 if (!drv->bdrv_co_pdiscard_snapshot) { 3586 return -ENOTSUP; 3587 } 3588 3589 bdrv_inc_in_flight(bs); 3590 ret = drv->bdrv_co_pdiscard_snapshot(bs, offset, bytes); 3591 bdrv_dec_in_flight(bs); 3592 3593 return ret; 3594 } 3595