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 assume_graph_lock(); /* FIXME */ 164 165 if (tran) { 166 BdrvRefreshLimitsState *s = g_new(BdrvRefreshLimitsState, 1); 167 *s = (BdrvRefreshLimitsState) { 168 .bs = bs, 169 .old_bl = bs->bl, 170 }; 171 tran_add(tran, &bdrv_refresh_limits_drv, s); 172 } 173 174 memset(&bs->bl, 0, sizeof(bs->bl)); 175 176 if (!drv) { 177 return; 178 } 179 180 /* Default alignment based on whether driver has byte interface */ 181 bs->bl.request_alignment = (drv->bdrv_co_preadv || 182 drv->bdrv_aio_preadv || 183 drv->bdrv_co_preadv_part) ? 1 : 512; 184 185 /* Take some limits from the children as a default */ 186 have_limits = false; 187 QLIST_FOREACH(c, &bs->children, next) { 188 if (c->role & (BDRV_CHILD_DATA | BDRV_CHILD_FILTERED | BDRV_CHILD_COW)) 189 { 190 bdrv_merge_limits(&bs->bl, &c->bs->bl); 191 have_limits = true; 192 } 193 } 194 195 if (!have_limits) { 196 bs->bl.min_mem_alignment = 512; 197 bs->bl.opt_mem_alignment = qemu_real_host_page_size(); 198 199 /* Safe default since most protocols use readv()/writev()/etc */ 200 bs->bl.max_iov = IOV_MAX; 201 } 202 203 /* Then let the driver override it */ 204 if (drv->bdrv_refresh_limits) { 205 drv->bdrv_refresh_limits(bs, errp); 206 if (*errp) { 207 return; 208 } 209 } 210 211 if (bs->bl.request_alignment > BDRV_MAX_ALIGNMENT) { 212 error_setg(errp, "Driver requires too large request alignment"); 213 } 214 } 215 216 /** 217 * The copy-on-read flag is actually a reference count so multiple users may 218 * use the feature without worrying about clobbering its previous state. 219 * Copy-on-read stays enabled until all users have called to disable it. 220 */ 221 void bdrv_enable_copy_on_read(BlockDriverState *bs) 222 { 223 IO_CODE(); 224 qatomic_inc(&bs->copy_on_read); 225 } 226 227 void bdrv_disable_copy_on_read(BlockDriverState *bs) 228 { 229 int old = qatomic_fetch_dec(&bs->copy_on_read); 230 IO_CODE(); 231 assert(old >= 1); 232 } 233 234 typedef struct { 235 Coroutine *co; 236 BlockDriverState *bs; 237 bool done; 238 bool begin; 239 bool poll; 240 BdrvChild *parent; 241 } BdrvCoDrainData; 242 243 /* Returns true if BDRV_POLL_WHILE() should go into a blocking aio_poll() */ 244 bool bdrv_drain_poll(BlockDriverState *bs, BdrvChild *ignore_parent, 245 bool ignore_bds_parents) 246 { 247 IO_OR_GS_CODE(); 248 249 if (bdrv_parent_drained_poll(bs, ignore_parent, ignore_bds_parents)) { 250 return true; 251 } 252 253 if (qatomic_read(&bs->in_flight)) { 254 return true; 255 } 256 257 return false; 258 } 259 260 static bool bdrv_drain_poll_top_level(BlockDriverState *bs, 261 BdrvChild *ignore_parent) 262 { 263 return bdrv_drain_poll(bs, ignore_parent, false); 264 } 265 266 static void bdrv_do_drained_begin(BlockDriverState *bs, BdrvChild *parent, 267 bool poll); 268 static void bdrv_do_drained_end(BlockDriverState *bs, BdrvChild *parent); 269 270 static void bdrv_co_drain_bh_cb(void *opaque) 271 { 272 BdrvCoDrainData *data = opaque; 273 Coroutine *co = data->co; 274 BlockDriverState *bs = data->bs; 275 276 if (bs) { 277 AioContext *ctx = bdrv_get_aio_context(bs); 278 aio_context_acquire(ctx); 279 bdrv_dec_in_flight(bs); 280 if (data->begin) { 281 bdrv_do_drained_begin(bs, data->parent, data->poll); 282 } else { 283 assert(!data->poll); 284 bdrv_do_drained_end(bs, data->parent); 285 } 286 aio_context_release(ctx); 287 } else { 288 assert(data->begin); 289 bdrv_drain_all_begin(); 290 } 291 292 data->done = true; 293 aio_co_wake(co); 294 } 295 296 static void coroutine_fn bdrv_co_yield_to_drain(BlockDriverState *bs, 297 bool begin, 298 BdrvChild *parent, 299 bool poll) 300 { 301 BdrvCoDrainData data; 302 Coroutine *self = qemu_coroutine_self(); 303 AioContext *ctx = bdrv_get_aio_context(bs); 304 AioContext *co_ctx = qemu_coroutine_get_aio_context(self); 305 306 /* Calling bdrv_drain() from a BH ensures the current coroutine yields and 307 * other coroutines run if they were queued by aio_co_enter(). */ 308 309 assert(qemu_in_coroutine()); 310 data = (BdrvCoDrainData) { 311 .co = self, 312 .bs = bs, 313 .done = false, 314 .begin = begin, 315 .parent = parent, 316 .poll = poll, 317 }; 318 319 if (bs) { 320 bdrv_inc_in_flight(bs); 321 } 322 323 /* 324 * Temporarily drop the lock across yield or we would get deadlocks. 325 * bdrv_co_drain_bh_cb() reaquires the lock as needed. 326 * 327 * When we yield below, the lock for the current context will be 328 * released, so if this is actually the lock that protects bs, don't drop 329 * it a second time. 330 */ 331 if (ctx != co_ctx) { 332 aio_context_release(ctx); 333 } 334 replay_bh_schedule_oneshot_event(ctx, bdrv_co_drain_bh_cb, &data); 335 336 qemu_coroutine_yield(); 337 /* If we are resumed from some other event (such as an aio completion or a 338 * timer callback), it is a bug in the caller that should be fixed. */ 339 assert(data.done); 340 341 /* Reaquire the AioContext of bs if we dropped it */ 342 if (ctx != co_ctx) { 343 aio_context_acquire(ctx); 344 } 345 } 346 347 static void bdrv_do_drained_begin(BlockDriverState *bs, BdrvChild *parent, 348 bool poll) 349 { 350 IO_OR_GS_CODE(); 351 352 if (qemu_in_coroutine()) { 353 bdrv_co_yield_to_drain(bs, true, parent, poll); 354 return; 355 } 356 357 /* Stop things in parent-to-child order */ 358 if (qatomic_fetch_inc(&bs->quiesce_counter) == 0) { 359 aio_disable_external(bdrv_get_aio_context(bs)); 360 bdrv_parent_drained_begin(bs, parent); 361 if (bs->drv && bs->drv->bdrv_drain_begin) { 362 bs->drv->bdrv_drain_begin(bs); 363 } 364 } 365 366 /* 367 * Wait for drained requests to finish. 368 * 369 * Calling BDRV_POLL_WHILE() only once for the top-level node is okay: The 370 * call is needed so things in this AioContext can make progress even 371 * though we don't return to the main AioContext loop - this automatically 372 * includes other nodes in the same AioContext and therefore all child 373 * nodes. 374 */ 375 if (poll) { 376 BDRV_POLL_WHILE(bs, bdrv_drain_poll_top_level(bs, parent)); 377 } 378 } 379 380 void bdrv_do_drained_begin_quiesce(BlockDriverState *bs, BdrvChild *parent) 381 { 382 bdrv_do_drained_begin(bs, parent, false); 383 } 384 385 void bdrv_drained_begin(BlockDriverState *bs) 386 { 387 IO_OR_GS_CODE(); 388 bdrv_do_drained_begin(bs, NULL, true); 389 } 390 391 /** 392 * This function does not poll, nor must any of its recursively called 393 * functions. 394 */ 395 static void bdrv_do_drained_end(BlockDriverState *bs, BdrvChild *parent) 396 { 397 int old_quiesce_counter; 398 399 if (qemu_in_coroutine()) { 400 bdrv_co_yield_to_drain(bs, false, parent, false); 401 return; 402 } 403 assert(bs->quiesce_counter > 0); 404 405 /* Re-enable things in child-to-parent order */ 406 old_quiesce_counter = qatomic_fetch_dec(&bs->quiesce_counter); 407 if (old_quiesce_counter == 1) { 408 if (bs->drv && bs->drv->bdrv_drain_end) { 409 bs->drv->bdrv_drain_end(bs); 410 } 411 bdrv_parent_drained_end(bs, parent); 412 aio_enable_external(bdrv_get_aio_context(bs)); 413 } 414 } 415 416 void bdrv_drained_end(BlockDriverState *bs) 417 { 418 IO_OR_GS_CODE(); 419 bdrv_do_drained_end(bs, NULL); 420 } 421 422 void bdrv_drain(BlockDriverState *bs) 423 { 424 IO_OR_GS_CODE(); 425 bdrv_drained_begin(bs); 426 bdrv_drained_end(bs); 427 } 428 429 static void bdrv_drain_assert_idle(BlockDriverState *bs) 430 { 431 BdrvChild *child, *next; 432 433 assert(qatomic_read(&bs->in_flight) == 0); 434 QLIST_FOREACH_SAFE(child, &bs->children, next, next) { 435 bdrv_drain_assert_idle(child->bs); 436 } 437 } 438 439 unsigned int bdrv_drain_all_count = 0; 440 441 static bool bdrv_drain_all_poll(void) 442 { 443 BlockDriverState *bs = NULL; 444 bool result = false; 445 GLOBAL_STATE_CODE(); 446 447 /* bdrv_drain_poll() can't make changes to the graph and we are holding the 448 * main AioContext lock, so iterating bdrv_next_all_states() is safe. */ 449 while ((bs = bdrv_next_all_states(bs))) { 450 AioContext *aio_context = bdrv_get_aio_context(bs); 451 aio_context_acquire(aio_context); 452 result |= bdrv_drain_poll(bs, NULL, true); 453 aio_context_release(aio_context); 454 } 455 456 return result; 457 } 458 459 /* 460 * Wait for pending requests to complete across all BlockDriverStates 461 * 462 * This function does not flush data to disk, use bdrv_flush_all() for that 463 * after calling this function. 464 * 465 * This pauses all block jobs and disables external clients. It must 466 * be paired with bdrv_drain_all_end(). 467 * 468 * NOTE: no new block jobs or BlockDriverStates can be created between 469 * the bdrv_drain_all_begin() and bdrv_drain_all_end() calls. 470 */ 471 void bdrv_drain_all_begin_nopoll(void) 472 { 473 BlockDriverState *bs = NULL; 474 GLOBAL_STATE_CODE(); 475 476 /* 477 * bdrv queue is managed by record/replay, 478 * waiting for finishing the I/O requests may 479 * be infinite 480 */ 481 if (replay_events_enabled()) { 482 return; 483 } 484 485 /* AIO_WAIT_WHILE() with a NULL context can only be called from the main 486 * loop AioContext, so make sure we're in the main context. */ 487 assert(qemu_get_current_aio_context() == qemu_get_aio_context()); 488 assert(bdrv_drain_all_count < INT_MAX); 489 bdrv_drain_all_count++; 490 491 /* Quiesce all nodes, without polling in-flight requests yet. The graph 492 * cannot change during this loop. */ 493 while ((bs = bdrv_next_all_states(bs))) { 494 AioContext *aio_context = bdrv_get_aio_context(bs); 495 496 aio_context_acquire(aio_context); 497 bdrv_do_drained_begin(bs, NULL, false); 498 aio_context_release(aio_context); 499 } 500 } 501 502 void bdrv_drain_all_begin(void) 503 { 504 BlockDriverState *bs = NULL; 505 506 if (qemu_in_coroutine()) { 507 bdrv_co_yield_to_drain(NULL, true, NULL, true); 508 return; 509 } 510 511 /* 512 * bdrv queue is managed by record/replay, 513 * waiting for finishing the I/O requests may 514 * be infinite 515 */ 516 if (replay_events_enabled()) { 517 return; 518 } 519 520 bdrv_drain_all_begin_nopoll(); 521 522 /* Now poll the in-flight requests */ 523 AIO_WAIT_WHILE(NULL, bdrv_drain_all_poll()); 524 525 while ((bs = bdrv_next_all_states(bs))) { 526 bdrv_drain_assert_idle(bs); 527 } 528 } 529 530 void bdrv_drain_all_end_quiesce(BlockDriverState *bs) 531 { 532 GLOBAL_STATE_CODE(); 533 534 g_assert(bs->quiesce_counter > 0); 535 g_assert(!bs->refcnt); 536 537 while (bs->quiesce_counter) { 538 bdrv_do_drained_end(bs, NULL); 539 } 540 } 541 542 void bdrv_drain_all_end(void) 543 { 544 BlockDriverState *bs = NULL; 545 GLOBAL_STATE_CODE(); 546 547 /* 548 * bdrv queue is managed by record/replay, 549 * waiting for finishing the I/O requests may 550 * be endless 551 */ 552 if (replay_events_enabled()) { 553 return; 554 } 555 556 while ((bs = bdrv_next_all_states(bs))) { 557 AioContext *aio_context = bdrv_get_aio_context(bs); 558 559 aio_context_acquire(aio_context); 560 bdrv_do_drained_end(bs, NULL); 561 aio_context_release(aio_context); 562 } 563 564 assert(qemu_get_current_aio_context() == qemu_get_aio_context()); 565 assert(bdrv_drain_all_count > 0); 566 bdrv_drain_all_count--; 567 } 568 569 void bdrv_drain_all(void) 570 { 571 GLOBAL_STATE_CODE(); 572 bdrv_drain_all_begin(); 573 bdrv_drain_all_end(); 574 } 575 576 /** 577 * Remove an active request from the tracked requests list 578 * 579 * This function should be called when a tracked request is completing. 580 */ 581 static void coroutine_fn tracked_request_end(BdrvTrackedRequest *req) 582 { 583 if (req->serialising) { 584 qatomic_dec(&req->bs->serialising_in_flight); 585 } 586 587 qemu_co_mutex_lock(&req->bs->reqs_lock); 588 QLIST_REMOVE(req, list); 589 qemu_co_queue_restart_all(&req->wait_queue); 590 qemu_co_mutex_unlock(&req->bs->reqs_lock); 591 } 592 593 /** 594 * Add an active request to the tracked requests list 595 */ 596 static void coroutine_fn tracked_request_begin(BdrvTrackedRequest *req, 597 BlockDriverState *bs, 598 int64_t offset, 599 int64_t bytes, 600 enum BdrvTrackedRequestType type) 601 { 602 bdrv_check_request(offset, bytes, &error_abort); 603 604 *req = (BdrvTrackedRequest){ 605 .bs = bs, 606 .offset = offset, 607 .bytes = bytes, 608 .type = type, 609 .co = qemu_coroutine_self(), 610 .serialising = false, 611 .overlap_offset = offset, 612 .overlap_bytes = bytes, 613 }; 614 615 qemu_co_queue_init(&req->wait_queue); 616 617 qemu_co_mutex_lock(&bs->reqs_lock); 618 QLIST_INSERT_HEAD(&bs->tracked_requests, req, list); 619 qemu_co_mutex_unlock(&bs->reqs_lock); 620 } 621 622 static bool tracked_request_overlaps(BdrvTrackedRequest *req, 623 int64_t offset, int64_t bytes) 624 { 625 bdrv_check_request(offset, bytes, &error_abort); 626 627 /* aaaa bbbb */ 628 if (offset >= req->overlap_offset + req->overlap_bytes) { 629 return false; 630 } 631 /* bbbb aaaa */ 632 if (req->overlap_offset >= offset + bytes) { 633 return false; 634 } 635 return true; 636 } 637 638 /* Called with self->bs->reqs_lock held */ 639 static coroutine_fn BdrvTrackedRequest * 640 bdrv_find_conflicting_request(BdrvTrackedRequest *self) 641 { 642 BdrvTrackedRequest *req; 643 644 QLIST_FOREACH(req, &self->bs->tracked_requests, list) { 645 if (req == self || (!req->serialising && !self->serialising)) { 646 continue; 647 } 648 if (tracked_request_overlaps(req, self->overlap_offset, 649 self->overlap_bytes)) 650 { 651 /* 652 * Hitting this means there was a reentrant request, for 653 * example, a block driver issuing nested requests. This must 654 * never happen since it means deadlock. 655 */ 656 assert(qemu_coroutine_self() != req->co); 657 658 /* 659 * If the request is already (indirectly) waiting for us, or 660 * will wait for us as soon as it wakes up, then just go on 661 * (instead of producing a deadlock in the former case). 662 */ 663 if (!req->waiting_for) { 664 return req; 665 } 666 } 667 } 668 669 return NULL; 670 } 671 672 /* Called with self->bs->reqs_lock held */ 673 static void coroutine_fn 674 bdrv_wait_serialising_requests_locked(BdrvTrackedRequest *self) 675 { 676 BdrvTrackedRequest *req; 677 678 while ((req = bdrv_find_conflicting_request(self))) { 679 self->waiting_for = req; 680 qemu_co_queue_wait(&req->wait_queue, &self->bs->reqs_lock); 681 self->waiting_for = NULL; 682 } 683 } 684 685 /* Called with req->bs->reqs_lock held */ 686 static void tracked_request_set_serialising(BdrvTrackedRequest *req, 687 uint64_t align) 688 { 689 int64_t overlap_offset = req->offset & ~(align - 1); 690 int64_t overlap_bytes = 691 ROUND_UP(req->offset + req->bytes, align) - overlap_offset; 692 693 bdrv_check_request(req->offset, req->bytes, &error_abort); 694 695 if (!req->serialising) { 696 qatomic_inc(&req->bs->serialising_in_flight); 697 req->serialising = true; 698 } 699 700 req->overlap_offset = MIN(req->overlap_offset, overlap_offset); 701 req->overlap_bytes = MAX(req->overlap_bytes, overlap_bytes); 702 } 703 704 /** 705 * Return the tracked request on @bs for the current coroutine, or 706 * NULL if there is none. 707 */ 708 BdrvTrackedRequest *coroutine_fn bdrv_co_get_self_request(BlockDriverState *bs) 709 { 710 BdrvTrackedRequest *req; 711 Coroutine *self = qemu_coroutine_self(); 712 IO_CODE(); 713 714 QLIST_FOREACH(req, &bs->tracked_requests, list) { 715 if (req->co == self) { 716 return req; 717 } 718 } 719 720 return NULL; 721 } 722 723 /** 724 * Round a region to cluster boundaries 725 */ 726 void coroutine_fn bdrv_round_to_clusters(BlockDriverState *bs, 727 int64_t offset, int64_t bytes, 728 int64_t *cluster_offset, 729 int64_t *cluster_bytes) 730 { 731 BlockDriverInfo bdi; 732 IO_CODE(); 733 if (bdrv_co_get_info(bs, &bdi) < 0 || bdi.cluster_size == 0) { 734 *cluster_offset = offset; 735 *cluster_bytes = bytes; 736 } else { 737 int64_t c = bdi.cluster_size; 738 *cluster_offset = QEMU_ALIGN_DOWN(offset, c); 739 *cluster_bytes = QEMU_ALIGN_UP(offset - *cluster_offset + bytes, c); 740 } 741 } 742 743 static coroutine_fn int bdrv_get_cluster_size(BlockDriverState *bs) 744 { 745 BlockDriverInfo bdi; 746 int ret; 747 748 ret = bdrv_co_get_info(bs, &bdi); 749 if (ret < 0 || bdi.cluster_size == 0) { 750 return bs->bl.request_alignment; 751 } else { 752 return bdi.cluster_size; 753 } 754 } 755 756 void bdrv_inc_in_flight(BlockDriverState *bs) 757 { 758 IO_CODE(); 759 qatomic_inc(&bs->in_flight); 760 } 761 762 void bdrv_wakeup(BlockDriverState *bs) 763 { 764 IO_CODE(); 765 aio_wait_kick(); 766 } 767 768 void bdrv_dec_in_flight(BlockDriverState *bs) 769 { 770 IO_CODE(); 771 qatomic_dec(&bs->in_flight); 772 bdrv_wakeup(bs); 773 } 774 775 static void coroutine_fn 776 bdrv_wait_serialising_requests(BdrvTrackedRequest *self) 777 { 778 BlockDriverState *bs = self->bs; 779 780 if (!qatomic_read(&bs->serialising_in_flight)) { 781 return; 782 } 783 784 qemu_co_mutex_lock(&bs->reqs_lock); 785 bdrv_wait_serialising_requests_locked(self); 786 qemu_co_mutex_unlock(&bs->reqs_lock); 787 } 788 789 void coroutine_fn bdrv_make_request_serialising(BdrvTrackedRequest *req, 790 uint64_t align) 791 { 792 IO_CODE(); 793 794 qemu_co_mutex_lock(&req->bs->reqs_lock); 795 796 tracked_request_set_serialising(req, align); 797 bdrv_wait_serialising_requests_locked(req); 798 799 qemu_co_mutex_unlock(&req->bs->reqs_lock); 800 } 801 802 int bdrv_check_qiov_request(int64_t offset, int64_t bytes, 803 QEMUIOVector *qiov, size_t qiov_offset, 804 Error **errp) 805 { 806 /* 807 * Check generic offset/bytes correctness 808 */ 809 810 if (offset < 0) { 811 error_setg(errp, "offset is negative: %" PRIi64, offset); 812 return -EIO; 813 } 814 815 if (bytes < 0) { 816 error_setg(errp, "bytes is negative: %" PRIi64, bytes); 817 return -EIO; 818 } 819 820 if (bytes > BDRV_MAX_LENGTH) { 821 error_setg(errp, "bytes(%" PRIi64 ") exceeds maximum(%" PRIi64 ")", 822 bytes, BDRV_MAX_LENGTH); 823 return -EIO; 824 } 825 826 if (offset > BDRV_MAX_LENGTH) { 827 error_setg(errp, "offset(%" PRIi64 ") exceeds maximum(%" PRIi64 ")", 828 offset, BDRV_MAX_LENGTH); 829 return -EIO; 830 } 831 832 if (offset > BDRV_MAX_LENGTH - bytes) { 833 error_setg(errp, "sum of offset(%" PRIi64 ") and bytes(%" PRIi64 ") " 834 "exceeds maximum(%" PRIi64 ")", offset, bytes, 835 BDRV_MAX_LENGTH); 836 return -EIO; 837 } 838 839 if (!qiov) { 840 return 0; 841 } 842 843 /* 844 * Check qiov and qiov_offset 845 */ 846 847 if (qiov_offset > qiov->size) { 848 error_setg(errp, "qiov_offset(%zu) overflow io vector size(%zu)", 849 qiov_offset, qiov->size); 850 return -EIO; 851 } 852 853 if (bytes > qiov->size - qiov_offset) { 854 error_setg(errp, "bytes(%" PRIi64 ") + qiov_offset(%zu) overflow io " 855 "vector size(%zu)", bytes, qiov_offset, qiov->size); 856 return -EIO; 857 } 858 859 return 0; 860 } 861 862 int bdrv_check_request(int64_t offset, int64_t bytes, Error **errp) 863 { 864 return bdrv_check_qiov_request(offset, bytes, NULL, 0, errp); 865 } 866 867 static int bdrv_check_request32(int64_t offset, int64_t bytes, 868 QEMUIOVector *qiov, size_t qiov_offset) 869 { 870 int ret = bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, NULL); 871 if (ret < 0) { 872 return ret; 873 } 874 875 if (bytes > BDRV_REQUEST_MAX_BYTES) { 876 return -EIO; 877 } 878 879 return 0; 880 } 881 882 /* 883 * Completely zero out a block device with the help of bdrv_pwrite_zeroes. 884 * The operation is sped up by checking the block status and only writing 885 * zeroes to the device if they currently do not return zeroes. Optional 886 * flags are passed through to bdrv_pwrite_zeroes (e.g. BDRV_REQ_MAY_UNMAP, 887 * BDRV_REQ_FUA). 888 * 889 * Returns < 0 on error, 0 on success. For error codes see bdrv_pwrite(). 890 */ 891 int bdrv_make_zero(BdrvChild *child, BdrvRequestFlags flags) 892 { 893 int ret; 894 int64_t target_size, bytes, offset = 0; 895 BlockDriverState *bs = child->bs; 896 IO_CODE(); 897 898 target_size = bdrv_getlength(bs); 899 if (target_size < 0) { 900 return target_size; 901 } 902 903 for (;;) { 904 bytes = MIN(target_size - offset, BDRV_REQUEST_MAX_BYTES); 905 if (bytes <= 0) { 906 return 0; 907 } 908 ret = bdrv_block_status(bs, offset, bytes, &bytes, NULL, NULL); 909 if (ret < 0) { 910 return ret; 911 } 912 if (ret & BDRV_BLOCK_ZERO) { 913 offset += bytes; 914 continue; 915 } 916 ret = bdrv_pwrite_zeroes(child, offset, bytes, flags); 917 if (ret < 0) { 918 return ret; 919 } 920 offset += bytes; 921 } 922 } 923 924 /* 925 * Writes to the file and ensures that no writes are reordered across this 926 * request (acts as a barrier) 927 * 928 * Returns 0 on success, -errno in error cases. 929 */ 930 int coroutine_fn bdrv_co_pwrite_sync(BdrvChild *child, int64_t offset, 931 int64_t bytes, const void *buf, 932 BdrvRequestFlags flags) 933 { 934 int ret; 935 IO_CODE(); 936 937 assume_graph_lock(); /* FIXME */ 938 939 ret = bdrv_co_pwrite(child, offset, bytes, buf, flags); 940 if (ret < 0) { 941 return ret; 942 } 943 944 ret = bdrv_co_flush(child->bs); 945 if (ret < 0) { 946 return ret; 947 } 948 949 return 0; 950 } 951 952 typedef struct CoroutineIOCompletion { 953 Coroutine *coroutine; 954 int ret; 955 } CoroutineIOCompletion; 956 957 static void bdrv_co_io_em_complete(void *opaque, int ret) 958 { 959 CoroutineIOCompletion *co = opaque; 960 961 co->ret = ret; 962 aio_co_wake(co->coroutine); 963 } 964 965 static int coroutine_fn GRAPH_RDLOCK 966 bdrv_driver_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes, 967 QEMUIOVector *qiov, 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 GRAPH_RDLOCK 1034 bdrv_driver_pwritev(BlockDriverState *bs, int64_t offset, int64_t bytes, 1035 QEMUIOVector *qiov, size_t qiov_offset, 1036 BdrvRequestFlags flags) 1037 { 1038 BlockDriver *drv = bs->drv; 1039 bool emulate_fua = false; 1040 int64_t sector_num; 1041 unsigned int nb_sectors; 1042 QEMUIOVector local_qiov; 1043 int ret; 1044 1045 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort); 1046 1047 if (!drv) { 1048 return -ENOMEDIUM; 1049 } 1050 1051 if ((flags & BDRV_REQ_FUA) && 1052 (~bs->supported_write_flags & BDRV_REQ_FUA)) { 1053 flags &= ~BDRV_REQ_FUA; 1054 emulate_fua = true; 1055 } 1056 1057 flags &= bs->supported_write_flags; 1058 1059 if (drv->bdrv_co_pwritev_part) { 1060 ret = drv->bdrv_co_pwritev_part(bs, offset, bytes, qiov, qiov_offset, 1061 flags); 1062 goto emulate_flags; 1063 } 1064 1065 if (qiov_offset > 0 || bytes != qiov->size) { 1066 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes); 1067 qiov = &local_qiov; 1068 } 1069 1070 if (drv->bdrv_co_pwritev) { 1071 ret = drv->bdrv_co_pwritev(bs, offset, bytes, qiov, flags); 1072 goto emulate_flags; 1073 } 1074 1075 if (drv->bdrv_aio_pwritev) { 1076 BlockAIOCB *acb; 1077 CoroutineIOCompletion co = { 1078 .coroutine = qemu_coroutine_self(), 1079 }; 1080 1081 acb = drv->bdrv_aio_pwritev(bs, offset, bytes, qiov, flags, 1082 bdrv_co_io_em_complete, &co); 1083 if (acb == NULL) { 1084 ret = -EIO; 1085 } else { 1086 qemu_coroutine_yield(); 1087 ret = co.ret; 1088 } 1089 goto emulate_flags; 1090 } 1091 1092 sector_num = offset >> BDRV_SECTOR_BITS; 1093 nb_sectors = bytes >> BDRV_SECTOR_BITS; 1094 1095 assert(QEMU_IS_ALIGNED(offset, BDRV_SECTOR_SIZE)); 1096 assert(QEMU_IS_ALIGNED(bytes, BDRV_SECTOR_SIZE)); 1097 assert(bytes <= BDRV_REQUEST_MAX_BYTES); 1098 1099 assert(drv->bdrv_co_writev); 1100 ret = drv->bdrv_co_writev(bs, sector_num, nb_sectors, qiov, flags); 1101 1102 emulate_flags: 1103 if (ret == 0 && emulate_fua) { 1104 ret = bdrv_co_flush(bs); 1105 } 1106 1107 if (qiov == &local_qiov) { 1108 qemu_iovec_destroy(&local_qiov); 1109 } 1110 1111 return ret; 1112 } 1113 1114 static int coroutine_fn GRAPH_RDLOCK 1115 bdrv_driver_pwritev_compressed(BlockDriverState *bs, int64_t offset, 1116 int64_t bytes, QEMUIOVector *qiov, 1117 size_t qiov_offset) 1118 { 1119 BlockDriver *drv = bs->drv; 1120 QEMUIOVector local_qiov; 1121 int ret; 1122 1123 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort); 1124 1125 if (!drv) { 1126 return -ENOMEDIUM; 1127 } 1128 1129 if (!block_driver_can_compress(drv)) { 1130 return -ENOTSUP; 1131 } 1132 1133 if (drv->bdrv_co_pwritev_compressed_part) { 1134 return drv->bdrv_co_pwritev_compressed_part(bs, offset, bytes, 1135 qiov, qiov_offset); 1136 } 1137 1138 if (qiov_offset == 0) { 1139 return drv->bdrv_co_pwritev_compressed(bs, offset, bytes, qiov); 1140 } 1141 1142 qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes); 1143 ret = drv->bdrv_co_pwritev_compressed(bs, offset, bytes, &local_qiov); 1144 qemu_iovec_destroy(&local_qiov); 1145 1146 return ret; 1147 } 1148 1149 static int coroutine_fn GRAPH_RDLOCK 1150 bdrv_co_do_copy_on_readv(BdrvChild *child, int64_t offset, int64_t bytes, 1151 QEMUIOVector *qiov, size_t qiov_offset, int flags) 1152 { 1153 BlockDriverState *bs = child->bs; 1154 1155 /* Perform I/O through a temporary buffer so that users who scribble over 1156 * their read buffer while the operation is in progress do not end up 1157 * modifying the image file. This is critical for zero-copy guest I/O 1158 * where anything might happen inside guest memory. 1159 */ 1160 void *bounce_buffer = NULL; 1161 1162 BlockDriver *drv = bs->drv; 1163 int64_t cluster_offset; 1164 int64_t cluster_bytes; 1165 int64_t skip_bytes; 1166 int ret; 1167 int max_transfer = MIN_NON_ZERO(bs->bl.max_transfer, 1168 BDRV_REQUEST_MAX_BYTES); 1169 int64_t progress = 0; 1170 bool skip_write; 1171 1172 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort); 1173 1174 if (!drv) { 1175 return -ENOMEDIUM; 1176 } 1177 1178 /* 1179 * Do not write anything when the BDS is inactive. That is not 1180 * allowed, and it would not help. 1181 */ 1182 skip_write = (bs->open_flags & BDRV_O_INACTIVE); 1183 1184 /* FIXME We cannot require callers to have write permissions when all they 1185 * are doing is a read request. If we did things right, write permissions 1186 * would be obtained anyway, but internally by the copy-on-read code. As 1187 * long as it is implemented here rather than in a separate filter driver, 1188 * the copy-on-read code doesn't have its own BdrvChild, however, for which 1189 * it could request permissions. Therefore we have to bypass the permission 1190 * system for the moment. */ 1191 // assert(child->perm & (BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE)); 1192 1193 /* Cover entire cluster so no additional backing file I/O is required when 1194 * allocating cluster in the image file. Note that this value may exceed 1195 * BDRV_REQUEST_MAX_BYTES (even when the original read did not), which 1196 * is one reason we loop rather than doing it all at once. 1197 */ 1198 bdrv_round_to_clusters(bs, offset, bytes, &cluster_offset, &cluster_bytes); 1199 skip_bytes = offset - cluster_offset; 1200 1201 trace_bdrv_co_do_copy_on_readv(bs, offset, bytes, 1202 cluster_offset, cluster_bytes); 1203 1204 while (cluster_bytes) { 1205 int64_t pnum; 1206 1207 if (skip_write) { 1208 ret = 1; /* "already allocated", so nothing will be copied */ 1209 pnum = MIN(cluster_bytes, max_transfer); 1210 } else { 1211 ret = bdrv_is_allocated(bs, cluster_offset, 1212 MIN(cluster_bytes, max_transfer), &pnum); 1213 if (ret < 0) { 1214 /* 1215 * Safe to treat errors in querying allocation as if 1216 * unallocated; we'll probably fail again soon on the 1217 * read, but at least that will set a decent errno. 1218 */ 1219 pnum = MIN(cluster_bytes, max_transfer); 1220 } 1221 1222 /* Stop at EOF if the image ends in the middle of the cluster */ 1223 if (ret == 0 && pnum == 0) { 1224 assert(progress >= bytes); 1225 break; 1226 } 1227 1228 assert(skip_bytes < pnum); 1229 } 1230 1231 if (ret <= 0) { 1232 QEMUIOVector local_qiov; 1233 1234 /* Must copy-on-read; use the bounce buffer */ 1235 pnum = MIN(pnum, MAX_BOUNCE_BUFFER); 1236 if (!bounce_buffer) { 1237 int64_t max_we_need = MAX(pnum, cluster_bytes - pnum); 1238 int64_t max_allowed = MIN(max_transfer, MAX_BOUNCE_BUFFER); 1239 int64_t bounce_buffer_len = MIN(max_we_need, max_allowed); 1240 1241 bounce_buffer = qemu_try_blockalign(bs, bounce_buffer_len); 1242 if (!bounce_buffer) { 1243 ret = -ENOMEM; 1244 goto err; 1245 } 1246 } 1247 qemu_iovec_init_buf(&local_qiov, bounce_buffer, pnum); 1248 1249 ret = bdrv_driver_preadv(bs, cluster_offset, pnum, 1250 &local_qiov, 0, 0); 1251 if (ret < 0) { 1252 goto err; 1253 } 1254 1255 bdrv_co_debug_event(bs, BLKDBG_COR_WRITE); 1256 if (drv->bdrv_co_pwrite_zeroes && 1257 buffer_is_zero(bounce_buffer, pnum)) { 1258 /* FIXME: Should we (perhaps conditionally) be setting 1259 * BDRV_REQ_MAY_UNMAP, if it will allow for a sparser copy 1260 * that still correctly reads as zero? */ 1261 ret = bdrv_co_do_pwrite_zeroes(bs, cluster_offset, pnum, 1262 BDRV_REQ_WRITE_UNCHANGED); 1263 } else { 1264 /* This does not change the data on the disk, it is not 1265 * necessary to flush even in cache=writethrough mode. 1266 */ 1267 ret = bdrv_driver_pwritev(bs, cluster_offset, pnum, 1268 &local_qiov, 0, 1269 BDRV_REQ_WRITE_UNCHANGED); 1270 } 1271 1272 if (ret < 0) { 1273 /* It might be okay to ignore write errors for guest 1274 * requests. If this is a deliberate copy-on-read 1275 * then we don't want to ignore the error. Simply 1276 * report it in all cases. 1277 */ 1278 goto err; 1279 } 1280 1281 if (!(flags & BDRV_REQ_PREFETCH)) { 1282 qemu_iovec_from_buf(qiov, qiov_offset + progress, 1283 bounce_buffer + skip_bytes, 1284 MIN(pnum - skip_bytes, bytes - progress)); 1285 } 1286 } else if (!(flags & BDRV_REQ_PREFETCH)) { 1287 /* Read directly into the destination */ 1288 ret = bdrv_driver_preadv(bs, offset + progress, 1289 MIN(pnum - skip_bytes, bytes - progress), 1290 qiov, qiov_offset + progress, 0); 1291 if (ret < 0) { 1292 goto err; 1293 } 1294 } 1295 1296 cluster_offset += pnum; 1297 cluster_bytes -= pnum; 1298 progress += pnum - skip_bytes; 1299 skip_bytes = 0; 1300 } 1301 ret = 0; 1302 1303 err: 1304 qemu_vfree(bounce_buffer); 1305 return ret; 1306 } 1307 1308 /* 1309 * Forwards an already correctly aligned request to the BlockDriver. This 1310 * handles copy on read, zeroing after EOF, and fragmentation of large 1311 * reads; any other features must be implemented by the caller. 1312 */ 1313 static int coroutine_fn GRAPH_RDLOCK 1314 bdrv_aligned_preadv(BdrvChild *child, BdrvTrackedRequest *req, 1315 int64_t offset, int64_t bytes, int64_t align, 1316 QEMUIOVector *qiov, size_t qiov_offset, int flags) 1317 { 1318 BlockDriverState *bs = child->bs; 1319 int64_t total_bytes, max_bytes; 1320 int ret = 0; 1321 int64_t bytes_remaining = bytes; 1322 int max_transfer; 1323 1324 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort); 1325 assert(is_power_of_2(align)); 1326 assert((offset & (align - 1)) == 0); 1327 assert((bytes & (align - 1)) == 0); 1328 assert((bs->open_flags & BDRV_O_NO_IO) == 0); 1329 max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX), 1330 align); 1331 1332 /* 1333 * TODO: We would need a per-BDS .supported_read_flags and 1334 * potential fallback support, if we ever implement any read flags 1335 * to pass through to drivers. For now, there aren't any 1336 * passthrough flags except the BDRV_REQ_REGISTERED_BUF optimization hint. 1337 */ 1338 assert(!(flags & ~(BDRV_REQ_COPY_ON_READ | BDRV_REQ_PREFETCH | 1339 BDRV_REQ_REGISTERED_BUF))); 1340 1341 /* Handle Copy on Read and associated serialisation */ 1342 if (flags & BDRV_REQ_COPY_ON_READ) { 1343 /* If we touch the same cluster it counts as an overlap. This 1344 * guarantees that allocating writes will be serialized and not race 1345 * with each other for the same cluster. For example, in copy-on-read 1346 * it ensures that the CoR read and write operations are atomic and 1347 * guest writes cannot interleave between them. */ 1348 bdrv_make_request_serialising(req, bdrv_get_cluster_size(bs)); 1349 } else { 1350 bdrv_wait_serialising_requests(req); 1351 } 1352 1353 if (flags & BDRV_REQ_COPY_ON_READ) { 1354 int64_t pnum; 1355 1356 /* The flag BDRV_REQ_COPY_ON_READ has reached its addressee */ 1357 flags &= ~BDRV_REQ_COPY_ON_READ; 1358 1359 ret = bdrv_is_allocated(bs, offset, bytes, &pnum); 1360 if (ret < 0) { 1361 goto out; 1362 } 1363 1364 if (!ret || pnum != bytes) { 1365 ret = bdrv_co_do_copy_on_readv(child, offset, bytes, 1366 qiov, qiov_offset, flags); 1367 goto out; 1368 } else if (flags & BDRV_REQ_PREFETCH) { 1369 goto out; 1370 } 1371 } 1372 1373 /* Forward the request to the BlockDriver, possibly fragmenting it */ 1374 total_bytes = bdrv_getlength(bs); 1375 if (total_bytes < 0) { 1376 ret = total_bytes; 1377 goto out; 1378 } 1379 1380 assert(!(flags & ~(bs->supported_read_flags | BDRV_REQ_REGISTERED_BUF))); 1381 1382 max_bytes = ROUND_UP(MAX(0, total_bytes - offset), align); 1383 if (bytes <= max_bytes && bytes <= max_transfer) { 1384 ret = bdrv_driver_preadv(bs, offset, bytes, qiov, qiov_offset, flags); 1385 goto out; 1386 } 1387 1388 while (bytes_remaining) { 1389 int64_t num; 1390 1391 if (max_bytes) { 1392 num = MIN(bytes_remaining, MIN(max_bytes, max_transfer)); 1393 assert(num); 1394 1395 ret = bdrv_driver_preadv(bs, offset + bytes - bytes_remaining, 1396 num, qiov, 1397 qiov_offset + bytes - bytes_remaining, 1398 flags); 1399 max_bytes -= num; 1400 } else { 1401 num = bytes_remaining; 1402 ret = qemu_iovec_memset(qiov, qiov_offset + bytes - bytes_remaining, 1403 0, bytes_remaining); 1404 } 1405 if (ret < 0) { 1406 goto out; 1407 } 1408 bytes_remaining -= num; 1409 } 1410 1411 out: 1412 return ret < 0 ? ret : 0; 1413 } 1414 1415 /* 1416 * Request padding 1417 * 1418 * |<---- align ----->| |<----- align ---->| 1419 * |<- head ->|<------------- bytes ------------->|<-- tail -->| 1420 * | | | | | | 1421 * -*----------$-------*-------- ... --------*-----$------------*--- 1422 * | | | | | | 1423 * | offset | | end | 1424 * ALIGN_DOWN(offset) ALIGN_UP(offset) ALIGN_DOWN(end) ALIGN_UP(end) 1425 * [buf ... ) [tail_buf ) 1426 * 1427 * @buf is an aligned allocation needed to store @head and @tail paddings. @head 1428 * is placed at the beginning of @buf and @tail at the @end. 1429 * 1430 * @tail_buf is a pointer to sub-buffer, corresponding to align-sized chunk 1431 * around tail, if tail exists. 1432 * 1433 * @merge_reads is true for small requests, 1434 * if @buf_len == @head + bytes + @tail. In this case it is possible that both 1435 * head and tail exist but @buf_len == align and @tail_buf == @buf. 1436 */ 1437 typedef struct BdrvRequestPadding { 1438 uint8_t *buf; 1439 size_t buf_len; 1440 uint8_t *tail_buf; 1441 size_t head; 1442 size_t tail; 1443 bool merge_reads; 1444 QEMUIOVector local_qiov; 1445 } BdrvRequestPadding; 1446 1447 static bool bdrv_init_padding(BlockDriverState *bs, 1448 int64_t offset, int64_t bytes, 1449 BdrvRequestPadding *pad) 1450 { 1451 int64_t align = bs->bl.request_alignment; 1452 int64_t sum; 1453 1454 bdrv_check_request(offset, bytes, &error_abort); 1455 assert(align <= INT_MAX); /* documented in block/block_int.h */ 1456 assert(align <= SIZE_MAX / 2); /* so we can allocate the buffer */ 1457 1458 memset(pad, 0, sizeof(*pad)); 1459 1460 pad->head = offset & (align - 1); 1461 pad->tail = ((offset + bytes) & (align - 1)); 1462 if (pad->tail) { 1463 pad->tail = align - pad->tail; 1464 } 1465 1466 if (!pad->head && !pad->tail) { 1467 return false; 1468 } 1469 1470 assert(bytes); /* Nothing good in aligning zero-length requests */ 1471 1472 sum = pad->head + bytes + pad->tail; 1473 pad->buf_len = (sum > align && pad->head && pad->tail) ? 2 * align : align; 1474 pad->buf = qemu_blockalign(bs, pad->buf_len); 1475 pad->merge_reads = sum == pad->buf_len; 1476 if (pad->tail) { 1477 pad->tail_buf = pad->buf + pad->buf_len - align; 1478 } 1479 1480 return true; 1481 } 1482 1483 static int coroutine_fn GRAPH_RDLOCK 1484 bdrv_padding_rmw_read(BdrvChild *child, BdrvTrackedRequest *req, 1485 BdrvRequestPadding *pad, bool zero_middle) 1486 { 1487 QEMUIOVector local_qiov; 1488 BlockDriverState *bs = child->bs; 1489 uint64_t align = bs->bl.request_alignment; 1490 int ret; 1491 1492 assert(req->serialising && pad->buf); 1493 1494 if (pad->head || pad->merge_reads) { 1495 int64_t bytes = pad->merge_reads ? pad->buf_len : align; 1496 1497 qemu_iovec_init_buf(&local_qiov, pad->buf, bytes); 1498 1499 if (pad->head) { 1500 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD); 1501 } 1502 if (pad->merge_reads && pad->tail) { 1503 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL); 1504 } 1505 ret = bdrv_aligned_preadv(child, req, req->overlap_offset, bytes, 1506 align, &local_qiov, 0, 0); 1507 if (ret < 0) { 1508 return ret; 1509 } 1510 if (pad->head) { 1511 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD); 1512 } 1513 if (pad->merge_reads && pad->tail) { 1514 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL); 1515 } 1516 1517 if (pad->merge_reads) { 1518 goto zero_mem; 1519 } 1520 } 1521 1522 if (pad->tail) { 1523 qemu_iovec_init_buf(&local_qiov, pad->tail_buf, align); 1524 1525 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL); 1526 ret = bdrv_aligned_preadv( 1527 child, req, 1528 req->overlap_offset + req->overlap_bytes - align, 1529 align, align, &local_qiov, 0, 0); 1530 if (ret < 0) { 1531 return ret; 1532 } 1533 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL); 1534 } 1535 1536 zero_mem: 1537 if (zero_middle) { 1538 memset(pad->buf + pad->head, 0, pad->buf_len - pad->head - pad->tail); 1539 } 1540 1541 return 0; 1542 } 1543 1544 static void bdrv_padding_destroy(BdrvRequestPadding *pad) 1545 { 1546 if (pad->buf) { 1547 qemu_vfree(pad->buf); 1548 qemu_iovec_destroy(&pad->local_qiov); 1549 } 1550 memset(pad, 0, sizeof(*pad)); 1551 } 1552 1553 /* 1554 * bdrv_pad_request 1555 * 1556 * Exchange request parameters with padded request if needed. Don't include RMW 1557 * read of padding, bdrv_padding_rmw_read() should be called separately if 1558 * needed. 1559 * 1560 * Request parameters (@qiov, &qiov_offset, &offset, &bytes) are in-out: 1561 * - on function start they represent original request 1562 * - on failure or when padding is not needed they are unchanged 1563 * - on success when padding is needed they represent padded request 1564 */ 1565 static int bdrv_pad_request(BlockDriverState *bs, 1566 QEMUIOVector **qiov, size_t *qiov_offset, 1567 int64_t *offset, int64_t *bytes, 1568 BdrvRequestPadding *pad, bool *padded, 1569 BdrvRequestFlags *flags) 1570 { 1571 int ret; 1572 1573 bdrv_check_qiov_request(*offset, *bytes, *qiov, *qiov_offset, &error_abort); 1574 1575 if (!bdrv_init_padding(bs, *offset, *bytes, pad)) { 1576 if (padded) { 1577 *padded = false; 1578 } 1579 return 0; 1580 } 1581 1582 ret = qemu_iovec_init_extended(&pad->local_qiov, pad->buf, pad->head, 1583 *qiov, *qiov_offset, *bytes, 1584 pad->buf + pad->buf_len - pad->tail, 1585 pad->tail); 1586 if (ret < 0) { 1587 bdrv_padding_destroy(pad); 1588 return ret; 1589 } 1590 *bytes += pad->head + pad->tail; 1591 *offset -= pad->head; 1592 *qiov = &pad->local_qiov; 1593 *qiov_offset = 0; 1594 if (padded) { 1595 *padded = true; 1596 } 1597 if (flags) { 1598 /* Can't use optimization hint with bounce buffer */ 1599 *flags &= ~BDRV_REQ_REGISTERED_BUF; 1600 } 1601 1602 return 0; 1603 } 1604 1605 int coroutine_fn bdrv_co_preadv(BdrvChild *child, 1606 int64_t offset, int64_t bytes, QEMUIOVector *qiov, 1607 BdrvRequestFlags flags) 1608 { 1609 IO_CODE(); 1610 return bdrv_co_preadv_part(child, offset, bytes, qiov, 0, flags); 1611 } 1612 1613 int coroutine_fn bdrv_co_preadv_part(BdrvChild *child, 1614 int64_t offset, int64_t bytes, 1615 QEMUIOVector *qiov, size_t qiov_offset, 1616 BdrvRequestFlags flags) 1617 { 1618 BlockDriverState *bs = child->bs; 1619 BdrvTrackedRequest req; 1620 BdrvRequestPadding pad; 1621 int ret; 1622 IO_CODE(); 1623 1624 assume_graph_lock(); /* FIXME */ 1625 1626 trace_bdrv_co_preadv_part(bs, offset, bytes, flags); 1627 1628 if (!bdrv_co_is_inserted(bs)) { 1629 return -ENOMEDIUM; 1630 } 1631 1632 ret = bdrv_check_request32(offset, bytes, qiov, qiov_offset); 1633 if (ret < 0) { 1634 return ret; 1635 } 1636 1637 if (bytes == 0 && !QEMU_IS_ALIGNED(offset, bs->bl.request_alignment)) { 1638 /* 1639 * Aligning zero request is nonsense. Even if driver has special meaning 1640 * of zero-length (like qcow2_co_pwritev_compressed_part), we can't pass 1641 * it to driver due to request_alignment. 1642 * 1643 * Still, no reason to return an error if someone do unaligned 1644 * zero-length read occasionally. 1645 */ 1646 return 0; 1647 } 1648 1649 bdrv_inc_in_flight(bs); 1650 1651 /* Don't do copy-on-read if we read data before write operation */ 1652 if (qatomic_read(&bs->copy_on_read)) { 1653 flags |= BDRV_REQ_COPY_ON_READ; 1654 } 1655 1656 ret = bdrv_pad_request(bs, &qiov, &qiov_offset, &offset, &bytes, &pad, 1657 NULL, &flags); 1658 if (ret < 0) { 1659 goto fail; 1660 } 1661 1662 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_READ); 1663 ret = bdrv_aligned_preadv(child, &req, offset, bytes, 1664 bs->bl.request_alignment, 1665 qiov, qiov_offset, flags); 1666 tracked_request_end(&req); 1667 bdrv_padding_destroy(&pad); 1668 1669 fail: 1670 bdrv_dec_in_flight(bs); 1671 1672 return ret; 1673 } 1674 1675 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs, 1676 int64_t offset, int64_t bytes, BdrvRequestFlags flags) 1677 { 1678 BlockDriver *drv = bs->drv; 1679 QEMUIOVector qiov; 1680 void *buf = NULL; 1681 int ret = 0; 1682 bool need_flush = false; 1683 int head = 0; 1684 int tail = 0; 1685 1686 assume_graph_lock(); /* FIXME */ 1687 1688 int64_t max_write_zeroes = MIN_NON_ZERO(bs->bl.max_pwrite_zeroes, 1689 INT64_MAX); 1690 int alignment = MAX(bs->bl.pwrite_zeroes_alignment, 1691 bs->bl.request_alignment); 1692 int max_transfer = MIN_NON_ZERO(bs->bl.max_transfer, MAX_BOUNCE_BUFFER); 1693 1694 assert_bdrv_graph_readable(); 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 GRAPH_RDLOCK 1899 bdrv_aligned_pwritev(BdrvChild *child, BdrvTrackedRequest *req, 1900 int64_t offset, int64_t bytes, int64_t align, 1901 QEMUIOVector *qiov, size_t qiov_offset, 1902 BdrvRequestFlags flags) 1903 { 1904 BlockDriverState *bs = child->bs; 1905 BlockDriver *drv = bs->drv; 1906 int ret; 1907 1908 int64_t bytes_remaining = bytes; 1909 int max_transfer; 1910 1911 bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, &error_abort); 1912 1913 if (!drv) { 1914 return -ENOMEDIUM; 1915 } 1916 1917 if (bdrv_has_readonly_bitmaps(bs)) { 1918 return -EPERM; 1919 } 1920 1921 assert(is_power_of_2(align)); 1922 assert((offset & (align - 1)) == 0); 1923 assert((bytes & (align - 1)) == 0); 1924 max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX), 1925 align); 1926 1927 ret = bdrv_co_write_req_prepare(child, offset, bytes, req, flags); 1928 1929 if (!ret && bs->detect_zeroes != BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF && 1930 !(flags & BDRV_REQ_ZERO_WRITE) && drv->bdrv_co_pwrite_zeroes && 1931 qemu_iovec_is_zero(qiov, qiov_offset, bytes)) { 1932 flags |= BDRV_REQ_ZERO_WRITE; 1933 if (bs->detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP) { 1934 flags |= BDRV_REQ_MAY_UNMAP; 1935 } 1936 1937 /* Can't use optimization hint with bufferless zero write */ 1938 flags &= ~BDRV_REQ_REGISTERED_BUF; 1939 } 1940 1941 if (ret < 0) { 1942 /* Do nothing, write notifier decided to fail this request */ 1943 } else if (flags & BDRV_REQ_ZERO_WRITE) { 1944 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_ZERO); 1945 ret = bdrv_co_do_pwrite_zeroes(bs, offset, bytes, flags); 1946 } else if (flags & BDRV_REQ_WRITE_COMPRESSED) { 1947 ret = bdrv_driver_pwritev_compressed(bs, offset, bytes, 1948 qiov, qiov_offset); 1949 } else if (bytes <= max_transfer) { 1950 bdrv_co_debug_event(bs, BLKDBG_PWRITEV); 1951 ret = bdrv_driver_pwritev(bs, offset, bytes, qiov, qiov_offset, flags); 1952 } else { 1953 bdrv_co_debug_event(bs, BLKDBG_PWRITEV); 1954 while (bytes_remaining) { 1955 int num = MIN(bytes_remaining, max_transfer); 1956 int local_flags = flags; 1957 1958 assert(num); 1959 if (num < bytes_remaining && (flags & BDRV_REQ_FUA) && 1960 !(bs->supported_write_flags & BDRV_REQ_FUA)) { 1961 /* If FUA is going to be emulated by flush, we only 1962 * need to flush on the last iteration */ 1963 local_flags &= ~BDRV_REQ_FUA; 1964 } 1965 1966 ret = bdrv_driver_pwritev(bs, offset + bytes - bytes_remaining, 1967 num, qiov, 1968 qiov_offset + bytes - bytes_remaining, 1969 local_flags); 1970 if (ret < 0) { 1971 break; 1972 } 1973 bytes_remaining -= num; 1974 } 1975 } 1976 bdrv_co_debug_event(bs, BLKDBG_PWRITEV_DONE); 1977 1978 if (ret >= 0) { 1979 ret = 0; 1980 } 1981 bdrv_co_write_req_finish(child, offset, bytes, req, ret); 1982 1983 return ret; 1984 } 1985 1986 static int coroutine_fn GRAPH_RDLOCK 1987 bdrv_co_do_zero_pwritev(BdrvChild *child, int64_t offset, int64_t bytes, 1988 BdrvRequestFlags flags, BdrvTrackedRequest *req) 1989 { 1990 BlockDriverState *bs = child->bs; 1991 QEMUIOVector local_qiov; 1992 uint64_t align = bs->bl.request_alignment; 1993 int ret = 0; 1994 bool padding; 1995 BdrvRequestPadding pad; 1996 1997 /* This flag doesn't make sense for padding or zero writes */ 1998 flags &= ~BDRV_REQ_REGISTERED_BUF; 1999 2000 padding = bdrv_init_padding(bs, offset, bytes, &pad); 2001 if (padding) { 2002 assert(!(flags & BDRV_REQ_NO_WAIT)); 2003 bdrv_make_request_serialising(req, align); 2004 2005 bdrv_padding_rmw_read(child, req, &pad, true); 2006 2007 if (pad.head || pad.merge_reads) { 2008 int64_t aligned_offset = offset & ~(align - 1); 2009 int64_t write_bytes = pad.merge_reads ? pad.buf_len : align; 2010 2011 qemu_iovec_init_buf(&local_qiov, pad.buf, write_bytes); 2012 ret = bdrv_aligned_pwritev(child, req, aligned_offset, write_bytes, 2013 align, &local_qiov, 0, 2014 flags & ~BDRV_REQ_ZERO_WRITE); 2015 if (ret < 0 || pad.merge_reads) { 2016 /* Error or all work is done */ 2017 goto out; 2018 } 2019 offset += write_bytes - pad.head; 2020 bytes -= write_bytes - pad.head; 2021 } 2022 } 2023 2024 assert(!bytes || (offset & (align - 1)) == 0); 2025 if (bytes >= align) { 2026 /* Write the aligned part in the middle. */ 2027 int64_t aligned_bytes = bytes & ~(align - 1); 2028 ret = bdrv_aligned_pwritev(child, req, offset, aligned_bytes, align, 2029 NULL, 0, flags); 2030 if (ret < 0) { 2031 goto out; 2032 } 2033 bytes -= aligned_bytes; 2034 offset += aligned_bytes; 2035 } 2036 2037 assert(!bytes || (offset & (align - 1)) == 0); 2038 if (bytes) { 2039 assert(align == pad.tail + bytes); 2040 2041 qemu_iovec_init_buf(&local_qiov, pad.tail_buf, align); 2042 ret = bdrv_aligned_pwritev(child, req, offset, align, align, 2043 &local_qiov, 0, 2044 flags & ~BDRV_REQ_ZERO_WRITE); 2045 } 2046 2047 out: 2048 bdrv_padding_destroy(&pad); 2049 2050 return ret; 2051 } 2052 2053 /* 2054 * Handle a write request in coroutine context 2055 */ 2056 int coroutine_fn bdrv_co_pwritev(BdrvChild *child, 2057 int64_t offset, int64_t bytes, QEMUIOVector *qiov, 2058 BdrvRequestFlags flags) 2059 { 2060 IO_CODE(); 2061 return bdrv_co_pwritev_part(child, offset, bytes, qiov, 0, flags); 2062 } 2063 2064 int coroutine_fn bdrv_co_pwritev_part(BdrvChild *child, 2065 int64_t offset, int64_t bytes, QEMUIOVector *qiov, size_t qiov_offset, 2066 BdrvRequestFlags flags) 2067 { 2068 BlockDriverState *bs = child->bs; 2069 BdrvTrackedRequest req; 2070 uint64_t align = bs->bl.request_alignment; 2071 BdrvRequestPadding pad; 2072 int ret; 2073 bool padded = false; 2074 IO_CODE(); 2075 2076 assume_graph_lock(); /* FIXME */ 2077 2078 trace_bdrv_co_pwritev_part(child->bs, offset, bytes, flags); 2079 2080 if (!bdrv_co_is_inserted(bs)) { 2081 return -ENOMEDIUM; 2082 } 2083 2084 if (flags & BDRV_REQ_ZERO_WRITE) { 2085 ret = bdrv_check_qiov_request(offset, bytes, qiov, qiov_offset, NULL); 2086 } else { 2087 ret = bdrv_check_request32(offset, bytes, qiov, qiov_offset); 2088 } 2089 if (ret < 0) { 2090 return ret; 2091 } 2092 2093 /* If the request is misaligned then we can't make it efficient */ 2094 if ((flags & BDRV_REQ_NO_FALLBACK) && 2095 !QEMU_IS_ALIGNED(offset | bytes, align)) 2096 { 2097 return -ENOTSUP; 2098 } 2099 2100 if (bytes == 0 && !QEMU_IS_ALIGNED(offset, bs->bl.request_alignment)) { 2101 /* 2102 * Aligning zero request is nonsense. Even if driver has special meaning 2103 * of zero-length (like qcow2_co_pwritev_compressed_part), we can't pass 2104 * it to driver due to request_alignment. 2105 * 2106 * Still, no reason to return an error if someone do unaligned 2107 * zero-length write occasionally. 2108 */ 2109 return 0; 2110 } 2111 2112 if (!(flags & BDRV_REQ_ZERO_WRITE)) { 2113 /* 2114 * Pad request for following read-modify-write cycle. 2115 * bdrv_co_do_zero_pwritev() does aligning by itself, so, we do 2116 * alignment only if there is no ZERO flag. 2117 */ 2118 ret = bdrv_pad_request(bs, &qiov, &qiov_offset, &offset, &bytes, &pad, 2119 &padded, &flags); 2120 if (ret < 0) { 2121 return ret; 2122 } 2123 } 2124 2125 bdrv_inc_in_flight(bs); 2126 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_WRITE); 2127 2128 if (flags & BDRV_REQ_ZERO_WRITE) { 2129 assert(!padded); 2130 ret = bdrv_co_do_zero_pwritev(child, offset, bytes, flags, &req); 2131 goto out; 2132 } 2133 2134 if (padded) { 2135 /* 2136 * Request was unaligned to request_alignment and therefore 2137 * padded. We are going to do read-modify-write, and must 2138 * serialize the request to prevent interactions of the 2139 * widened region with other transactions. 2140 */ 2141 assert(!(flags & BDRV_REQ_NO_WAIT)); 2142 bdrv_make_request_serialising(&req, align); 2143 bdrv_padding_rmw_read(child, &req, &pad, false); 2144 } 2145 2146 ret = bdrv_aligned_pwritev(child, &req, offset, bytes, align, 2147 qiov, qiov_offset, flags); 2148 2149 bdrv_padding_destroy(&pad); 2150 2151 out: 2152 tracked_request_end(&req); 2153 bdrv_dec_in_flight(bs); 2154 2155 return ret; 2156 } 2157 2158 int coroutine_fn bdrv_co_pwrite_zeroes(BdrvChild *child, int64_t offset, 2159 int64_t bytes, BdrvRequestFlags flags) 2160 { 2161 IO_CODE(); 2162 trace_bdrv_co_pwrite_zeroes(child->bs, offset, bytes, flags); 2163 assert_bdrv_graph_readable(); 2164 2165 if (!(child->bs->open_flags & BDRV_O_UNMAP)) { 2166 flags &= ~BDRV_REQ_MAY_UNMAP; 2167 } 2168 2169 return bdrv_co_pwritev(child, offset, bytes, NULL, 2170 BDRV_REQ_ZERO_WRITE | flags); 2171 } 2172 2173 /* 2174 * Flush ALL BDSes regardless of if they are reachable via a BlkBackend or not. 2175 */ 2176 int bdrv_flush_all(void) 2177 { 2178 BdrvNextIterator it; 2179 BlockDriverState *bs = NULL; 2180 int result = 0; 2181 2182 GLOBAL_STATE_CODE(); 2183 2184 /* 2185 * bdrv queue is managed by record/replay, 2186 * creating new flush request for stopping 2187 * the VM may break the determinism 2188 */ 2189 if (replay_events_enabled()) { 2190 return result; 2191 } 2192 2193 for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) { 2194 AioContext *aio_context = bdrv_get_aio_context(bs); 2195 int ret; 2196 2197 aio_context_acquire(aio_context); 2198 ret = bdrv_flush(bs); 2199 if (ret < 0 && !result) { 2200 result = ret; 2201 } 2202 aio_context_release(aio_context); 2203 } 2204 2205 return result; 2206 } 2207 2208 /* 2209 * Returns the allocation status of the specified sectors. 2210 * Drivers not implementing the functionality are assumed to not support 2211 * backing files, hence all their sectors are reported as allocated. 2212 * 2213 * If 'want_zero' is true, the caller is querying for mapping 2214 * purposes, with a focus on valid BDRV_BLOCK_OFFSET_VALID, _DATA, and 2215 * _ZERO where possible; otherwise, the result favors larger 'pnum', 2216 * with a focus on accurate BDRV_BLOCK_ALLOCATED. 2217 * 2218 * If 'offset' is beyond the end of the disk image the return value is 2219 * BDRV_BLOCK_EOF and 'pnum' is set to 0. 2220 * 2221 * 'bytes' is the max value 'pnum' should be set to. If bytes goes 2222 * beyond the end of the disk image it will be clamped; if 'pnum' is set to 2223 * the end of the image, then the returned value will include BDRV_BLOCK_EOF. 2224 * 2225 * 'pnum' is set to the number of bytes (including and immediately 2226 * following the specified offset) that are easily known to be in the 2227 * same allocated/unallocated state. Note that a second call starting 2228 * at the original offset plus returned pnum may have the same status. 2229 * The returned value is non-zero on success except at end-of-file. 2230 * 2231 * Returns negative errno on failure. Otherwise, if the 2232 * BDRV_BLOCK_OFFSET_VALID bit is set, 'map' and 'file' (if non-NULL) are 2233 * set to the host mapping and BDS corresponding to the guest offset. 2234 */ 2235 static int coroutine_fn GRAPH_RDLOCK 2236 bdrv_co_block_status(BlockDriverState *bs, bool want_zero, 2237 int64_t offset, int64_t bytes, 2238 int64_t *pnum, int64_t *map, BlockDriverState **file) 2239 { 2240 int64_t total_size; 2241 int64_t n; /* bytes */ 2242 int ret; 2243 int64_t local_map = 0; 2244 BlockDriverState *local_file = NULL; 2245 int64_t aligned_offset, aligned_bytes; 2246 uint32_t align; 2247 bool has_filtered_child; 2248 2249 assert(pnum); 2250 assert_bdrv_graph_readable(); 2251 *pnum = 0; 2252 total_size = bdrv_getlength(bs); 2253 if (total_size < 0) { 2254 ret = total_size; 2255 goto early_out; 2256 } 2257 2258 if (offset >= total_size) { 2259 ret = BDRV_BLOCK_EOF; 2260 goto early_out; 2261 } 2262 if (!bytes) { 2263 ret = 0; 2264 goto early_out; 2265 } 2266 2267 n = total_size - offset; 2268 if (n < bytes) { 2269 bytes = n; 2270 } 2271 2272 /* Must be non-NULL or bdrv_getlength() would have failed */ 2273 assert(bs->drv); 2274 has_filtered_child = bdrv_filter_child(bs); 2275 if (!bs->drv->bdrv_co_block_status && !has_filtered_child) { 2276 *pnum = bytes; 2277 ret = BDRV_BLOCK_DATA | BDRV_BLOCK_ALLOCATED; 2278 if (offset + bytes == total_size) { 2279 ret |= BDRV_BLOCK_EOF; 2280 } 2281 if (bs->drv->protocol_name) { 2282 ret |= BDRV_BLOCK_OFFSET_VALID; 2283 local_map = offset; 2284 local_file = bs; 2285 } 2286 goto early_out; 2287 } 2288 2289 bdrv_inc_in_flight(bs); 2290 2291 /* Round out to request_alignment boundaries */ 2292 align = bs->bl.request_alignment; 2293 aligned_offset = QEMU_ALIGN_DOWN(offset, align); 2294 aligned_bytes = ROUND_UP(offset + bytes, align) - aligned_offset; 2295 2296 if (bs->drv->bdrv_co_block_status) { 2297 /* 2298 * Use the block-status cache only for protocol nodes: Format 2299 * drivers are generally quick to inquire the status, but protocol 2300 * drivers often need to get information from outside of qemu, so 2301 * we do not have control over the actual implementation. There 2302 * have been cases where inquiring the status took an unreasonably 2303 * long time, and we can do nothing in qemu to fix it. 2304 * This is especially problematic for images with large data areas, 2305 * because finding the few holes in them and giving them special 2306 * treatment does not gain much performance. Therefore, we try to 2307 * cache the last-identified data region. 2308 * 2309 * Second, limiting ourselves to protocol nodes allows us to assume 2310 * the block status for data regions to be DATA | OFFSET_VALID, and 2311 * that the host offset is the same as the guest offset. 2312 * 2313 * Note that it is possible that external writers zero parts of 2314 * the cached regions without the cache being invalidated, and so 2315 * we may report zeroes as data. This is not catastrophic, 2316 * however, because reporting zeroes as data is fine. 2317 */ 2318 if (QLIST_EMPTY(&bs->children) && 2319 bdrv_bsc_is_data(bs, aligned_offset, pnum)) 2320 { 2321 ret = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID; 2322 local_file = bs; 2323 local_map = aligned_offset; 2324 } else { 2325 ret = bs->drv->bdrv_co_block_status(bs, want_zero, aligned_offset, 2326 aligned_bytes, pnum, &local_map, 2327 &local_file); 2328 2329 /* 2330 * Note that checking QLIST_EMPTY(&bs->children) is also done when 2331 * the cache is queried above. Technically, we do not need to check 2332 * it here; the worst that can happen is that we fill the cache for 2333 * non-protocol nodes, and then it is never used. However, filling 2334 * the cache requires an RCU update, so double check here to avoid 2335 * such an update if possible. 2336 * 2337 * Check want_zero, because we only want to update the cache when we 2338 * have accurate information about what is zero and what is data. 2339 */ 2340 if (want_zero && 2341 ret == (BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID) && 2342 QLIST_EMPTY(&bs->children)) 2343 { 2344 /* 2345 * When a protocol driver reports BLOCK_OFFSET_VALID, the 2346 * returned local_map value must be the same as the offset we 2347 * have passed (aligned_offset), and local_bs must be the node 2348 * itself. 2349 * Assert this, because we follow this rule when reading from 2350 * the cache (see the `local_file = bs` and 2351 * `local_map = aligned_offset` assignments above), and the 2352 * result the cache delivers must be the same as the driver 2353 * would deliver. 2354 */ 2355 assert(local_file == bs); 2356 assert(local_map == aligned_offset); 2357 bdrv_bsc_fill(bs, aligned_offset, *pnum); 2358 } 2359 } 2360 } else { 2361 /* Default code for filters */ 2362 2363 local_file = bdrv_filter_bs(bs); 2364 assert(local_file); 2365 2366 *pnum = aligned_bytes; 2367 local_map = aligned_offset; 2368 ret = BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID; 2369 } 2370 if (ret < 0) { 2371 *pnum = 0; 2372 goto out; 2373 } 2374 2375 /* 2376 * The driver's result must be a non-zero multiple of request_alignment. 2377 * Clamp pnum and adjust map to original request. 2378 */ 2379 assert(*pnum && QEMU_IS_ALIGNED(*pnum, align) && 2380 align > offset - aligned_offset); 2381 if (ret & BDRV_BLOCK_RECURSE) { 2382 assert(ret & BDRV_BLOCK_DATA); 2383 assert(ret & BDRV_BLOCK_OFFSET_VALID); 2384 assert(!(ret & BDRV_BLOCK_ZERO)); 2385 } 2386 2387 *pnum -= offset - aligned_offset; 2388 if (*pnum > bytes) { 2389 *pnum = bytes; 2390 } 2391 if (ret & BDRV_BLOCK_OFFSET_VALID) { 2392 local_map += offset - aligned_offset; 2393 } 2394 2395 if (ret & BDRV_BLOCK_RAW) { 2396 assert(ret & BDRV_BLOCK_OFFSET_VALID && local_file); 2397 ret = bdrv_co_block_status(local_file, want_zero, local_map, 2398 *pnum, pnum, &local_map, &local_file); 2399 goto out; 2400 } 2401 2402 if (ret & (BDRV_BLOCK_DATA | BDRV_BLOCK_ZERO)) { 2403 ret |= BDRV_BLOCK_ALLOCATED; 2404 } else if (bs->drv->supports_backing) { 2405 BlockDriverState *cow_bs = bdrv_cow_bs(bs); 2406 2407 if (!cow_bs) { 2408 ret |= BDRV_BLOCK_ZERO; 2409 } else if (want_zero) { 2410 int64_t size2 = bdrv_getlength(cow_bs); 2411 2412 if (size2 >= 0 && offset >= size2) { 2413 ret |= BDRV_BLOCK_ZERO; 2414 } 2415 } 2416 } 2417 2418 if (want_zero && ret & BDRV_BLOCK_RECURSE && 2419 local_file && local_file != bs && 2420 (ret & BDRV_BLOCK_DATA) && !(ret & BDRV_BLOCK_ZERO) && 2421 (ret & BDRV_BLOCK_OFFSET_VALID)) { 2422 int64_t file_pnum; 2423 int ret2; 2424 2425 ret2 = bdrv_co_block_status(local_file, want_zero, local_map, 2426 *pnum, &file_pnum, NULL, NULL); 2427 if (ret2 >= 0) { 2428 /* Ignore errors. This is just providing extra information, it 2429 * is useful but not necessary. 2430 */ 2431 if (ret2 & BDRV_BLOCK_EOF && 2432 (!file_pnum || ret2 & BDRV_BLOCK_ZERO)) { 2433 /* 2434 * It is valid for the format block driver to read 2435 * beyond the end of the underlying file's current 2436 * size; such areas read as zero. 2437 */ 2438 ret |= BDRV_BLOCK_ZERO; 2439 } else { 2440 /* Limit request to the range reported by the protocol driver */ 2441 *pnum = file_pnum; 2442 ret |= (ret2 & BDRV_BLOCK_ZERO); 2443 } 2444 } 2445 } 2446 2447 out: 2448 bdrv_dec_in_flight(bs); 2449 if (ret >= 0 && offset + *pnum == total_size) { 2450 ret |= BDRV_BLOCK_EOF; 2451 } 2452 early_out: 2453 if (file) { 2454 *file = local_file; 2455 } 2456 if (map) { 2457 *map = local_map; 2458 } 2459 return ret; 2460 } 2461 2462 int coroutine_fn 2463 bdrv_co_common_block_status_above(BlockDriverState *bs, 2464 BlockDriverState *base, 2465 bool include_base, 2466 bool want_zero, 2467 int64_t offset, 2468 int64_t bytes, 2469 int64_t *pnum, 2470 int64_t *map, 2471 BlockDriverState **file, 2472 int *depth) 2473 { 2474 int ret; 2475 BlockDriverState *p; 2476 int64_t eof = 0; 2477 int dummy; 2478 IO_CODE(); 2479 2480 assert(!include_base || base); /* Can't include NULL base */ 2481 assert_bdrv_graph_readable(); 2482 2483 if (!depth) { 2484 depth = &dummy; 2485 } 2486 *depth = 0; 2487 2488 if (!include_base && bs == base) { 2489 *pnum = bytes; 2490 return 0; 2491 } 2492 2493 ret = bdrv_co_block_status(bs, want_zero, offset, bytes, pnum, map, file); 2494 ++*depth; 2495 if (ret < 0 || *pnum == 0 || ret & BDRV_BLOCK_ALLOCATED || bs == base) { 2496 return ret; 2497 } 2498 2499 if (ret & BDRV_BLOCK_EOF) { 2500 eof = offset + *pnum; 2501 } 2502 2503 assert(*pnum <= bytes); 2504 bytes = *pnum; 2505 2506 for (p = bdrv_filter_or_cow_bs(bs); include_base || p != base; 2507 p = bdrv_filter_or_cow_bs(p)) 2508 { 2509 ret = bdrv_co_block_status(p, want_zero, offset, bytes, pnum, map, 2510 file); 2511 ++*depth; 2512 if (ret < 0) { 2513 return ret; 2514 } 2515 if (*pnum == 0) { 2516 /* 2517 * The top layer deferred to this layer, and because this layer is 2518 * short, any zeroes that we synthesize beyond EOF behave as if they 2519 * were allocated at this layer. 2520 * 2521 * We don't include BDRV_BLOCK_EOF into ret, as upper layer may be 2522 * larger. We'll add BDRV_BLOCK_EOF if needed at function end, see 2523 * below. 2524 */ 2525 assert(ret & BDRV_BLOCK_EOF); 2526 *pnum = bytes; 2527 if (file) { 2528 *file = p; 2529 } 2530 ret = BDRV_BLOCK_ZERO | BDRV_BLOCK_ALLOCATED; 2531 break; 2532 } 2533 if (ret & BDRV_BLOCK_ALLOCATED) { 2534 /* 2535 * We've found the node and the status, we must break. 2536 * 2537 * Drop BDRV_BLOCK_EOF, as it's not for upper layer, which may be 2538 * larger. We'll add BDRV_BLOCK_EOF if needed at function end, see 2539 * below. 2540 */ 2541 ret &= ~BDRV_BLOCK_EOF; 2542 break; 2543 } 2544 2545 if (p == base) { 2546 assert(include_base); 2547 break; 2548 } 2549 2550 /* 2551 * OK, [offset, offset + *pnum) region is unallocated on this layer, 2552 * let's continue the diving. 2553 */ 2554 assert(*pnum <= bytes); 2555 bytes = *pnum; 2556 } 2557 2558 if (offset + *pnum == eof) { 2559 ret |= BDRV_BLOCK_EOF; 2560 } 2561 2562 return ret; 2563 } 2564 2565 int coroutine_fn bdrv_co_block_status_above(BlockDriverState *bs, 2566 BlockDriverState *base, 2567 int64_t offset, int64_t bytes, 2568 int64_t *pnum, int64_t *map, 2569 BlockDriverState **file) 2570 { 2571 IO_CODE(); 2572 return bdrv_co_common_block_status_above(bs, base, false, true, offset, 2573 bytes, pnum, map, file, NULL); 2574 } 2575 2576 int bdrv_block_status_above(BlockDriverState *bs, BlockDriverState *base, 2577 int64_t offset, int64_t bytes, int64_t *pnum, 2578 int64_t *map, BlockDriverState **file) 2579 { 2580 IO_CODE(); 2581 return bdrv_common_block_status_above(bs, base, false, true, offset, bytes, 2582 pnum, map, file, NULL); 2583 } 2584 2585 int bdrv_block_status(BlockDriverState *bs, int64_t offset, int64_t bytes, 2586 int64_t *pnum, int64_t *map, BlockDriverState **file) 2587 { 2588 IO_CODE(); 2589 return bdrv_block_status_above(bs, bdrv_filter_or_cow_bs(bs), 2590 offset, bytes, pnum, map, file); 2591 } 2592 2593 /* 2594 * Check @bs (and its backing chain) to see if the range defined 2595 * by @offset and @bytes is known to read as zeroes. 2596 * Return 1 if that is the case, 0 otherwise and -errno on error. 2597 * This test is meant to be fast rather than accurate so returning 0 2598 * does not guarantee non-zero data. 2599 */ 2600 int coroutine_fn bdrv_co_is_zero_fast(BlockDriverState *bs, int64_t offset, 2601 int64_t bytes) 2602 { 2603 int ret; 2604 int64_t pnum = bytes; 2605 IO_CODE(); 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 assert_bdrv_graph_readable(); 2975 2976 if (!bs || !bs->drv || !bdrv_co_is_inserted(bs)) { 2977 return -ENOMEDIUM; 2978 } 2979 2980 if (bdrv_has_readonly_bitmaps(bs)) { 2981 return -EPERM; 2982 } 2983 2984 ret = bdrv_check_request(offset, bytes, NULL); 2985 if (ret < 0) { 2986 return ret; 2987 } 2988 2989 /* Do nothing if disabled. */ 2990 if (!(bs->open_flags & BDRV_O_UNMAP)) { 2991 return 0; 2992 } 2993 2994 if (!bs->drv->bdrv_co_pdiscard && !bs->drv->bdrv_aio_pdiscard) { 2995 return 0; 2996 } 2997 2998 /* Invalidate the cached block-status data range if this discard overlaps */ 2999 bdrv_bsc_invalidate_range(bs, offset, bytes); 3000 3001 /* Discard is advisory, but some devices track and coalesce 3002 * unaligned requests, so we must pass everything down rather than 3003 * round here. Still, most devices will just silently ignore 3004 * unaligned requests (by returning -ENOTSUP), so we must fragment 3005 * the request accordingly. */ 3006 align = MAX(bs->bl.pdiscard_alignment, bs->bl.request_alignment); 3007 assert(align % bs->bl.request_alignment == 0); 3008 head = offset % align; 3009 tail = (offset + bytes) % align; 3010 3011 bdrv_inc_in_flight(bs); 3012 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_DISCARD); 3013 3014 ret = bdrv_co_write_req_prepare(child, offset, bytes, &req, 0); 3015 if (ret < 0) { 3016 goto out; 3017 } 3018 3019 max_pdiscard = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_pdiscard, INT64_MAX), 3020 align); 3021 assert(max_pdiscard >= bs->bl.request_alignment); 3022 3023 while (bytes > 0) { 3024 int64_t num = bytes; 3025 3026 if (head) { 3027 /* Make small requests to get to alignment boundaries. */ 3028 num = MIN(bytes, align - head); 3029 if (!QEMU_IS_ALIGNED(num, bs->bl.request_alignment)) { 3030 num %= bs->bl.request_alignment; 3031 } 3032 head = (head + num) % align; 3033 assert(num < max_pdiscard); 3034 } else if (tail) { 3035 if (num > align) { 3036 /* Shorten the request to the last aligned cluster. */ 3037 num -= tail; 3038 } else if (!QEMU_IS_ALIGNED(tail, bs->bl.request_alignment) && 3039 tail > bs->bl.request_alignment) { 3040 tail %= bs->bl.request_alignment; 3041 num -= tail; 3042 } 3043 } 3044 /* limit request size */ 3045 if (num > max_pdiscard) { 3046 num = max_pdiscard; 3047 } 3048 3049 if (!bs->drv) { 3050 ret = -ENOMEDIUM; 3051 goto out; 3052 } 3053 if (bs->drv->bdrv_co_pdiscard) { 3054 ret = bs->drv->bdrv_co_pdiscard(bs, offset, num); 3055 } else { 3056 BlockAIOCB *acb; 3057 CoroutineIOCompletion co = { 3058 .coroutine = qemu_coroutine_self(), 3059 }; 3060 3061 acb = bs->drv->bdrv_aio_pdiscard(bs, offset, num, 3062 bdrv_co_io_em_complete, &co); 3063 if (acb == NULL) { 3064 ret = -EIO; 3065 goto out; 3066 } else { 3067 qemu_coroutine_yield(); 3068 ret = co.ret; 3069 } 3070 } 3071 if (ret && ret != -ENOTSUP) { 3072 goto out; 3073 } 3074 3075 offset += num; 3076 bytes -= num; 3077 } 3078 ret = 0; 3079 out: 3080 bdrv_co_write_req_finish(child, req.offset, req.bytes, &req, ret); 3081 tracked_request_end(&req); 3082 bdrv_dec_in_flight(bs); 3083 return ret; 3084 } 3085 3086 int coroutine_fn bdrv_co_ioctl(BlockDriverState *bs, int req, void *buf) 3087 { 3088 BlockDriver *drv = bs->drv; 3089 CoroutineIOCompletion co = { 3090 .coroutine = qemu_coroutine_self(), 3091 }; 3092 BlockAIOCB *acb; 3093 IO_CODE(); 3094 assert_bdrv_graph_readable(); 3095 3096 bdrv_inc_in_flight(bs); 3097 if (!drv || (!drv->bdrv_aio_ioctl && !drv->bdrv_co_ioctl)) { 3098 co.ret = -ENOTSUP; 3099 goto out; 3100 } 3101 3102 if (drv->bdrv_co_ioctl) { 3103 co.ret = drv->bdrv_co_ioctl(bs, req, buf); 3104 } else { 3105 acb = drv->bdrv_aio_ioctl(bs, req, buf, bdrv_co_io_em_complete, &co); 3106 if (!acb) { 3107 co.ret = -ENOTSUP; 3108 goto out; 3109 } 3110 qemu_coroutine_yield(); 3111 } 3112 out: 3113 bdrv_dec_in_flight(bs); 3114 return co.ret; 3115 } 3116 3117 void *qemu_blockalign(BlockDriverState *bs, size_t size) 3118 { 3119 IO_CODE(); 3120 return qemu_memalign(bdrv_opt_mem_align(bs), size); 3121 } 3122 3123 void *qemu_blockalign0(BlockDriverState *bs, size_t size) 3124 { 3125 IO_CODE(); 3126 return memset(qemu_blockalign(bs, size), 0, size); 3127 } 3128 3129 void *qemu_try_blockalign(BlockDriverState *bs, size_t size) 3130 { 3131 size_t align = bdrv_opt_mem_align(bs); 3132 IO_CODE(); 3133 3134 /* Ensure that NULL is never returned on success */ 3135 assert(align > 0); 3136 if (size == 0) { 3137 size = align; 3138 } 3139 3140 return qemu_try_memalign(align, size); 3141 } 3142 3143 void *qemu_try_blockalign0(BlockDriverState *bs, size_t size) 3144 { 3145 void *mem = qemu_try_blockalign(bs, size); 3146 IO_CODE(); 3147 3148 if (mem) { 3149 memset(mem, 0, size); 3150 } 3151 3152 return mem; 3153 } 3154 3155 void coroutine_fn bdrv_co_io_plug(BlockDriverState *bs) 3156 { 3157 BdrvChild *child; 3158 IO_CODE(); 3159 3160 QLIST_FOREACH(child, &bs->children, next) { 3161 bdrv_co_io_plug(child->bs); 3162 } 3163 3164 if (qatomic_fetch_inc(&bs->io_plugged) == 0) { 3165 BlockDriver *drv = bs->drv; 3166 if (drv && drv->bdrv_co_io_plug) { 3167 drv->bdrv_co_io_plug(bs); 3168 } 3169 } 3170 } 3171 3172 void coroutine_fn bdrv_co_io_unplug(BlockDriverState *bs) 3173 { 3174 BdrvChild *child; 3175 IO_CODE(); 3176 3177 assert(bs->io_plugged); 3178 if (qatomic_fetch_dec(&bs->io_plugged) == 1) { 3179 BlockDriver *drv = bs->drv; 3180 if (drv && drv->bdrv_co_io_unplug) { 3181 drv->bdrv_co_io_unplug(bs); 3182 } 3183 } 3184 3185 QLIST_FOREACH(child, &bs->children, next) { 3186 bdrv_co_io_unplug(child->bs); 3187 } 3188 } 3189 3190 /* Helper that undoes bdrv_register_buf() when it fails partway through */ 3191 static void bdrv_register_buf_rollback(BlockDriverState *bs, 3192 void *host, 3193 size_t size, 3194 BdrvChild *final_child) 3195 { 3196 BdrvChild *child; 3197 3198 QLIST_FOREACH(child, &bs->children, next) { 3199 if (child == final_child) { 3200 break; 3201 } 3202 3203 bdrv_unregister_buf(child->bs, host, size); 3204 } 3205 3206 if (bs->drv && bs->drv->bdrv_unregister_buf) { 3207 bs->drv->bdrv_unregister_buf(bs, host, size); 3208 } 3209 } 3210 3211 bool bdrv_register_buf(BlockDriverState *bs, void *host, size_t size, 3212 Error **errp) 3213 { 3214 BdrvChild *child; 3215 3216 GLOBAL_STATE_CODE(); 3217 if (bs->drv && bs->drv->bdrv_register_buf) { 3218 if (!bs->drv->bdrv_register_buf(bs, host, size, errp)) { 3219 return false; 3220 } 3221 } 3222 QLIST_FOREACH(child, &bs->children, next) { 3223 if (!bdrv_register_buf(child->bs, host, size, errp)) { 3224 bdrv_register_buf_rollback(bs, host, size, child); 3225 return false; 3226 } 3227 } 3228 return true; 3229 } 3230 3231 void bdrv_unregister_buf(BlockDriverState *bs, void *host, size_t size) 3232 { 3233 BdrvChild *child; 3234 3235 GLOBAL_STATE_CODE(); 3236 if (bs->drv && bs->drv->bdrv_unregister_buf) { 3237 bs->drv->bdrv_unregister_buf(bs, host, size); 3238 } 3239 QLIST_FOREACH(child, &bs->children, next) { 3240 bdrv_unregister_buf(child->bs, host, size); 3241 } 3242 } 3243 3244 static int coroutine_fn GRAPH_RDLOCK bdrv_co_copy_range_internal( 3245 BdrvChild *src, int64_t src_offset, BdrvChild *dst, 3246 int64_t dst_offset, int64_t bytes, 3247 BdrvRequestFlags read_flags, BdrvRequestFlags write_flags, 3248 bool recurse_src) 3249 { 3250 BdrvTrackedRequest req; 3251 int ret; 3252 3253 /* TODO We can support BDRV_REQ_NO_FALLBACK here */ 3254 assert(!(read_flags & BDRV_REQ_NO_FALLBACK)); 3255 assert(!(write_flags & BDRV_REQ_NO_FALLBACK)); 3256 assert(!(read_flags & BDRV_REQ_NO_WAIT)); 3257 assert(!(write_flags & BDRV_REQ_NO_WAIT)); 3258 3259 if (!dst || !dst->bs || !bdrv_co_is_inserted(dst->bs)) { 3260 return -ENOMEDIUM; 3261 } 3262 ret = bdrv_check_request32(dst_offset, bytes, NULL, 0); 3263 if (ret) { 3264 return ret; 3265 } 3266 if (write_flags & BDRV_REQ_ZERO_WRITE) { 3267 return bdrv_co_pwrite_zeroes(dst, dst_offset, bytes, write_flags); 3268 } 3269 3270 if (!src || !src->bs || !bdrv_co_is_inserted(src->bs)) { 3271 return -ENOMEDIUM; 3272 } 3273 ret = bdrv_check_request32(src_offset, bytes, NULL, 0); 3274 if (ret) { 3275 return ret; 3276 } 3277 3278 if (!src->bs->drv->bdrv_co_copy_range_from 3279 || !dst->bs->drv->bdrv_co_copy_range_to 3280 || src->bs->encrypted || dst->bs->encrypted) { 3281 return -ENOTSUP; 3282 } 3283 3284 if (recurse_src) { 3285 bdrv_inc_in_flight(src->bs); 3286 tracked_request_begin(&req, src->bs, src_offset, bytes, 3287 BDRV_TRACKED_READ); 3288 3289 /* BDRV_REQ_SERIALISING is only for write operation */ 3290 assert(!(read_flags & BDRV_REQ_SERIALISING)); 3291 bdrv_wait_serialising_requests(&req); 3292 3293 ret = src->bs->drv->bdrv_co_copy_range_from(src->bs, 3294 src, src_offset, 3295 dst, dst_offset, 3296 bytes, 3297 read_flags, write_flags); 3298 3299 tracked_request_end(&req); 3300 bdrv_dec_in_flight(src->bs); 3301 } else { 3302 bdrv_inc_in_flight(dst->bs); 3303 tracked_request_begin(&req, dst->bs, dst_offset, bytes, 3304 BDRV_TRACKED_WRITE); 3305 ret = bdrv_co_write_req_prepare(dst, dst_offset, bytes, &req, 3306 write_flags); 3307 if (!ret) { 3308 ret = dst->bs->drv->bdrv_co_copy_range_to(dst->bs, 3309 src, src_offset, 3310 dst, dst_offset, 3311 bytes, 3312 read_flags, write_flags); 3313 } 3314 bdrv_co_write_req_finish(dst, dst_offset, bytes, &req, ret); 3315 tracked_request_end(&req); 3316 bdrv_dec_in_flight(dst->bs); 3317 } 3318 3319 return ret; 3320 } 3321 3322 /* Copy range from @src to @dst. 3323 * 3324 * See the comment of bdrv_co_copy_range for the parameter and return value 3325 * semantics. */ 3326 int coroutine_fn bdrv_co_copy_range_from(BdrvChild *src, int64_t src_offset, 3327 BdrvChild *dst, int64_t dst_offset, 3328 int64_t bytes, 3329 BdrvRequestFlags read_flags, 3330 BdrvRequestFlags write_flags) 3331 { 3332 IO_CODE(); 3333 assume_graph_lock(); /* FIXME */ 3334 trace_bdrv_co_copy_range_from(src, src_offset, dst, dst_offset, bytes, 3335 read_flags, write_flags); 3336 return bdrv_co_copy_range_internal(src, src_offset, dst, dst_offset, 3337 bytes, read_flags, write_flags, true); 3338 } 3339 3340 /* Copy range from @src to @dst. 3341 * 3342 * See the comment of bdrv_co_copy_range for the parameter and return value 3343 * semantics. */ 3344 int coroutine_fn bdrv_co_copy_range_to(BdrvChild *src, int64_t src_offset, 3345 BdrvChild *dst, int64_t dst_offset, 3346 int64_t bytes, 3347 BdrvRequestFlags read_flags, 3348 BdrvRequestFlags write_flags) 3349 { 3350 IO_CODE(); 3351 assume_graph_lock(); /* FIXME */ 3352 trace_bdrv_co_copy_range_to(src, src_offset, dst, dst_offset, bytes, 3353 read_flags, write_flags); 3354 return bdrv_co_copy_range_internal(src, src_offset, dst, dst_offset, 3355 bytes, read_flags, write_flags, false); 3356 } 3357 3358 int coroutine_fn bdrv_co_copy_range(BdrvChild *src, int64_t src_offset, 3359 BdrvChild *dst, int64_t dst_offset, 3360 int64_t bytes, BdrvRequestFlags read_flags, 3361 BdrvRequestFlags write_flags) 3362 { 3363 IO_CODE(); 3364 return bdrv_co_copy_range_from(src, src_offset, 3365 dst, dst_offset, 3366 bytes, read_flags, write_flags); 3367 } 3368 3369 static void bdrv_parent_cb_resize(BlockDriverState *bs) 3370 { 3371 BdrvChild *c; 3372 QLIST_FOREACH(c, &bs->parents, next_parent) { 3373 if (c->klass->resize) { 3374 c->klass->resize(c); 3375 } 3376 } 3377 } 3378 3379 /** 3380 * Truncate file to 'offset' bytes (needed only for file protocols) 3381 * 3382 * If 'exact' is true, the file must be resized to exactly the given 3383 * 'offset'. Otherwise, it is sufficient for the node to be at least 3384 * 'offset' bytes in length. 3385 */ 3386 int coroutine_fn bdrv_co_truncate(BdrvChild *child, int64_t offset, bool exact, 3387 PreallocMode prealloc, BdrvRequestFlags flags, 3388 Error **errp) 3389 { 3390 BlockDriverState *bs = child->bs; 3391 BdrvChild *filtered, *backing; 3392 BlockDriver *drv = bs->drv; 3393 BdrvTrackedRequest req; 3394 int64_t old_size, new_bytes; 3395 int ret; 3396 IO_CODE(); 3397 assert_bdrv_graph_readable(); 3398 3399 /* if bs->drv == NULL, bs is closed, so there's nothing to do here */ 3400 if (!drv) { 3401 error_setg(errp, "No medium inserted"); 3402 return -ENOMEDIUM; 3403 } 3404 if (offset < 0) { 3405 error_setg(errp, "Image size cannot be negative"); 3406 return -EINVAL; 3407 } 3408 3409 ret = bdrv_check_request(offset, 0, errp); 3410 if (ret < 0) { 3411 return ret; 3412 } 3413 3414 old_size = bdrv_getlength(bs); 3415 if (old_size < 0) { 3416 error_setg_errno(errp, -old_size, "Failed to get old image size"); 3417 return old_size; 3418 } 3419 3420 if (bdrv_is_read_only(bs)) { 3421 error_setg(errp, "Image is read-only"); 3422 return -EACCES; 3423 } 3424 3425 if (offset > old_size) { 3426 new_bytes = offset - old_size; 3427 } else { 3428 new_bytes = 0; 3429 } 3430 3431 bdrv_inc_in_flight(bs); 3432 tracked_request_begin(&req, bs, offset - new_bytes, new_bytes, 3433 BDRV_TRACKED_TRUNCATE); 3434 3435 /* If we are growing the image and potentially using preallocation for the 3436 * new area, we need to make sure that no write requests are made to it 3437 * concurrently or they might be overwritten by preallocation. */ 3438 if (new_bytes) { 3439 bdrv_make_request_serialising(&req, 1); 3440 } 3441 ret = bdrv_co_write_req_prepare(child, offset - new_bytes, new_bytes, &req, 3442 0); 3443 if (ret < 0) { 3444 error_setg_errno(errp, -ret, 3445 "Failed to prepare request for truncation"); 3446 goto out; 3447 } 3448 3449 filtered = bdrv_filter_child(bs); 3450 backing = bdrv_cow_child(bs); 3451 3452 /* 3453 * If the image has a backing file that is large enough that it would 3454 * provide data for the new area, we cannot leave it unallocated because 3455 * then the backing file content would become visible. Instead, zero-fill 3456 * the new area. 3457 * 3458 * Note that if the image has a backing file, but was opened without the 3459 * backing file, taking care of keeping things consistent with that backing 3460 * file is the user's responsibility. 3461 */ 3462 if (new_bytes && backing) { 3463 int64_t backing_len; 3464 3465 backing_len = bdrv_co_getlength(backing->bs); 3466 if (backing_len < 0) { 3467 ret = backing_len; 3468 error_setg_errno(errp, -ret, "Could not get backing file size"); 3469 goto out; 3470 } 3471 3472 if (backing_len > old_size) { 3473 flags |= BDRV_REQ_ZERO_WRITE; 3474 } 3475 } 3476 3477 if (drv->bdrv_co_truncate) { 3478 if (flags & ~bs->supported_truncate_flags) { 3479 error_setg(errp, "Block driver does not support requested flags"); 3480 ret = -ENOTSUP; 3481 goto out; 3482 } 3483 ret = drv->bdrv_co_truncate(bs, offset, exact, prealloc, flags, errp); 3484 } else if (filtered) { 3485 ret = bdrv_co_truncate(filtered, offset, exact, prealloc, flags, errp); 3486 } else { 3487 error_setg(errp, "Image format driver does not support resize"); 3488 ret = -ENOTSUP; 3489 goto out; 3490 } 3491 if (ret < 0) { 3492 goto out; 3493 } 3494 3495 ret = bdrv_co_refresh_total_sectors(bs, offset >> BDRV_SECTOR_BITS); 3496 if (ret < 0) { 3497 error_setg_errno(errp, -ret, "Could not refresh total sector count"); 3498 } else { 3499 offset = bs->total_sectors * BDRV_SECTOR_SIZE; 3500 } 3501 /* 3502 * It's possible that truncation succeeded but bdrv_refresh_total_sectors 3503 * failed, but the latter doesn't affect how we should finish the request. 3504 * Pass 0 as the last parameter so that dirty bitmaps etc. are handled. 3505 */ 3506 bdrv_co_write_req_finish(child, offset - new_bytes, new_bytes, &req, 0); 3507 3508 out: 3509 tracked_request_end(&req); 3510 bdrv_dec_in_flight(bs); 3511 3512 return ret; 3513 } 3514 3515 void bdrv_cancel_in_flight(BlockDriverState *bs) 3516 { 3517 GLOBAL_STATE_CODE(); 3518 if (!bs || !bs->drv) { 3519 return; 3520 } 3521 3522 if (bs->drv->bdrv_cancel_in_flight) { 3523 bs->drv->bdrv_cancel_in_flight(bs); 3524 } 3525 } 3526 3527 int coroutine_fn 3528 bdrv_co_preadv_snapshot(BdrvChild *child, int64_t offset, int64_t bytes, 3529 QEMUIOVector *qiov, size_t qiov_offset) 3530 { 3531 BlockDriverState *bs = child->bs; 3532 BlockDriver *drv = bs->drv; 3533 int ret; 3534 IO_CODE(); 3535 3536 if (!drv) { 3537 return -ENOMEDIUM; 3538 } 3539 3540 if (!drv->bdrv_co_preadv_snapshot) { 3541 return -ENOTSUP; 3542 } 3543 3544 bdrv_inc_in_flight(bs); 3545 ret = drv->bdrv_co_preadv_snapshot(bs, offset, bytes, qiov, qiov_offset); 3546 bdrv_dec_in_flight(bs); 3547 3548 return ret; 3549 } 3550 3551 int coroutine_fn 3552 bdrv_co_snapshot_block_status(BlockDriverState *bs, 3553 bool want_zero, int64_t offset, int64_t bytes, 3554 int64_t *pnum, int64_t *map, 3555 BlockDriverState **file) 3556 { 3557 BlockDriver *drv = bs->drv; 3558 int ret; 3559 IO_CODE(); 3560 3561 if (!drv) { 3562 return -ENOMEDIUM; 3563 } 3564 3565 if (!drv->bdrv_co_snapshot_block_status) { 3566 return -ENOTSUP; 3567 } 3568 3569 bdrv_inc_in_flight(bs); 3570 ret = drv->bdrv_co_snapshot_block_status(bs, want_zero, offset, bytes, 3571 pnum, map, file); 3572 bdrv_dec_in_flight(bs); 3573 3574 return ret; 3575 } 3576 3577 int coroutine_fn 3578 bdrv_co_pdiscard_snapshot(BlockDriverState *bs, int64_t offset, int64_t bytes) 3579 { 3580 BlockDriver *drv = bs->drv; 3581 int ret; 3582 IO_CODE(); 3583 assert_bdrv_graph_readable(); 3584 3585 if (!drv) { 3586 return -ENOMEDIUM; 3587 } 3588 3589 if (!drv->bdrv_co_pdiscard_snapshot) { 3590 return -ENOTSUP; 3591 } 3592 3593 bdrv_inc_in_flight(bs); 3594 ret = drv->bdrv_co_pdiscard_snapshot(bs, offset, bytes); 3595 bdrv_dec_in_flight(bs); 3596 3597 return ret; 3598 } 3599