1 /* 2 * Block layer I/O functions 3 * 4 * Copyright (c) 2003 Fabrice Bellard 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a copy 7 * of this software and associated documentation files (the "Software"), to deal 8 * in the Software without restriction, including without limitation the rights 9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 10 * copies of the Software, and to permit persons to whom the Software is 11 * furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice shall be included in 14 * all copies or substantial portions of the Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 22 * THE SOFTWARE. 23 */ 24 25 #include "qemu/osdep.h" 26 #include "trace.h" 27 #include "sysemu/block-backend.h" 28 #include "block/blockjob.h" 29 #include "block/block_int.h" 30 #include "qemu/cutils.h" 31 #include "qapi/error.h" 32 #include "qemu/error-report.h" 33 34 #define NOT_DONE 0x7fffffff /* used while emulated sync operation in progress */ 35 36 static BlockAIOCB *bdrv_co_aio_rw_vector(BlockDriverState *bs, 37 int64_t sector_num, 38 QEMUIOVector *qiov, 39 int nb_sectors, 40 BdrvRequestFlags flags, 41 BlockCompletionFunc *cb, 42 void *opaque, 43 bool is_write); 44 static void coroutine_fn bdrv_co_do_rw(void *opaque); 45 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs, 46 int64_t offset, int count, BdrvRequestFlags flags); 47 48 static void bdrv_parent_drained_begin(BlockDriverState *bs) 49 { 50 BdrvChild *c; 51 52 QLIST_FOREACH(c, &bs->parents, next_parent) { 53 if (c->role->drained_begin) { 54 c->role->drained_begin(c); 55 } 56 } 57 } 58 59 static void bdrv_parent_drained_end(BlockDriverState *bs) 60 { 61 BdrvChild *c; 62 63 QLIST_FOREACH(c, &bs->parents, next_parent) { 64 if (c->role->drained_end) { 65 c->role->drained_end(c); 66 } 67 } 68 } 69 70 static void bdrv_merge_limits(BlockLimits *dst, const BlockLimits *src) 71 { 72 dst->opt_transfer = MAX(dst->opt_transfer, src->opt_transfer); 73 dst->max_transfer = MIN_NON_ZERO(dst->max_transfer, src->max_transfer); 74 dst->opt_mem_alignment = MAX(dst->opt_mem_alignment, 75 src->opt_mem_alignment); 76 dst->min_mem_alignment = MAX(dst->min_mem_alignment, 77 src->min_mem_alignment); 78 dst->max_iov = MIN_NON_ZERO(dst->max_iov, src->max_iov); 79 } 80 81 void bdrv_refresh_limits(BlockDriverState *bs, Error **errp) 82 { 83 BlockDriver *drv = bs->drv; 84 Error *local_err = NULL; 85 86 memset(&bs->bl, 0, sizeof(bs->bl)); 87 88 if (!drv) { 89 return; 90 } 91 92 /* Default alignment based on whether driver has byte interface */ 93 bs->bl.request_alignment = drv->bdrv_co_preadv ? 1 : 512; 94 95 /* Take some limits from the children as a default */ 96 if (bs->file) { 97 bdrv_refresh_limits(bs->file->bs, &local_err); 98 if (local_err) { 99 error_propagate(errp, local_err); 100 return; 101 } 102 bdrv_merge_limits(&bs->bl, &bs->file->bs->bl); 103 } else { 104 bs->bl.min_mem_alignment = 512; 105 bs->bl.opt_mem_alignment = getpagesize(); 106 107 /* Safe default since most protocols use readv()/writev()/etc */ 108 bs->bl.max_iov = IOV_MAX; 109 } 110 111 if (bs->backing) { 112 bdrv_refresh_limits(bs->backing->bs, &local_err); 113 if (local_err) { 114 error_propagate(errp, local_err); 115 return; 116 } 117 bdrv_merge_limits(&bs->bl, &bs->backing->bs->bl); 118 } 119 120 /* Then let the driver override it */ 121 if (drv->bdrv_refresh_limits) { 122 drv->bdrv_refresh_limits(bs, errp); 123 } 124 } 125 126 /** 127 * The copy-on-read flag is actually a reference count so multiple users may 128 * use the feature without worrying about clobbering its previous state. 129 * Copy-on-read stays enabled until all users have called to disable it. 130 */ 131 void bdrv_enable_copy_on_read(BlockDriverState *bs) 132 { 133 bs->copy_on_read++; 134 } 135 136 void bdrv_disable_copy_on_read(BlockDriverState *bs) 137 { 138 assert(bs->copy_on_read > 0); 139 bs->copy_on_read--; 140 } 141 142 /* Check if any requests are in-flight (including throttled requests) */ 143 bool bdrv_requests_pending(BlockDriverState *bs) 144 { 145 BdrvChild *child; 146 147 if (!QLIST_EMPTY(&bs->tracked_requests)) { 148 return true; 149 } 150 151 QLIST_FOREACH(child, &bs->children, next) { 152 if (bdrv_requests_pending(child->bs)) { 153 return true; 154 } 155 } 156 157 return false; 158 } 159 160 static void bdrv_drain_recurse(BlockDriverState *bs) 161 { 162 BdrvChild *child; 163 164 if (bs->drv && bs->drv->bdrv_drain) { 165 bs->drv->bdrv_drain(bs); 166 } 167 QLIST_FOREACH(child, &bs->children, next) { 168 bdrv_drain_recurse(child->bs); 169 } 170 } 171 172 typedef struct { 173 Coroutine *co; 174 BlockDriverState *bs; 175 QEMUBH *bh; 176 bool done; 177 } BdrvCoDrainData; 178 179 static void bdrv_drain_poll(BlockDriverState *bs) 180 { 181 bool busy = true; 182 183 while (busy) { 184 /* Keep iterating */ 185 busy = bdrv_requests_pending(bs); 186 busy |= aio_poll(bdrv_get_aio_context(bs), busy); 187 } 188 } 189 190 static void bdrv_co_drain_bh_cb(void *opaque) 191 { 192 BdrvCoDrainData *data = opaque; 193 Coroutine *co = data->co; 194 195 qemu_bh_delete(data->bh); 196 bdrv_drain_poll(data->bs); 197 data->done = true; 198 qemu_coroutine_enter(co, NULL); 199 } 200 201 static void coroutine_fn bdrv_co_yield_to_drain(BlockDriverState *bs) 202 { 203 BdrvCoDrainData data; 204 205 /* Calling bdrv_drain() from a BH ensures the current coroutine yields and 206 * other coroutines run if they were queued from 207 * qemu_co_queue_run_restart(). */ 208 209 assert(qemu_in_coroutine()); 210 data = (BdrvCoDrainData) { 211 .co = qemu_coroutine_self(), 212 .bs = bs, 213 .done = false, 214 .bh = aio_bh_new(bdrv_get_aio_context(bs), bdrv_co_drain_bh_cb, &data), 215 }; 216 qemu_bh_schedule(data.bh); 217 218 qemu_coroutine_yield(); 219 /* If we are resumed from some other event (such as an aio completion or a 220 * timer callback), it is a bug in the caller that should be fixed. */ 221 assert(data.done); 222 } 223 224 void bdrv_drained_begin(BlockDriverState *bs) 225 { 226 if (!bs->quiesce_counter++) { 227 aio_disable_external(bdrv_get_aio_context(bs)); 228 bdrv_parent_drained_begin(bs); 229 } 230 231 bdrv_io_unplugged_begin(bs); 232 bdrv_drain_recurse(bs); 233 if (qemu_in_coroutine()) { 234 bdrv_co_yield_to_drain(bs); 235 } else { 236 bdrv_drain_poll(bs); 237 } 238 bdrv_io_unplugged_end(bs); 239 } 240 241 void bdrv_drained_end(BlockDriverState *bs) 242 { 243 assert(bs->quiesce_counter > 0); 244 if (--bs->quiesce_counter > 0) { 245 return; 246 } 247 248 bdrv_parent_drained_end(bs); 249 aio_enable_external(bdrv_get_aio_context(bs)); 250 } 251 252 /* 253 * Wait for pending requests to complete on a single BlockDriverState subtree, 254 * and suspend block driver's internal I/O until next request arrives. 255 * 256 * Note that unlike bdrv_drain_all(), the caller must hold the BlockDriverState 257 * AioContext. 258 * 259 * Only this BlockDriverState's AioContext is run, so in-flight requests must 260 * not depend on events in other AioContexts. In that case, use 261 * bdrv_drain_all() instead. 262 */ 263 void coroutine_fn bdrv_co_drain(BlockDriverState *bs) 264 { 265 assert(qemu_in_coroutine()); 266 bdrv_drained_begin(bs); 267 bdrv_drained_end(bs); 268 } 269 270 void bdrv_drain(BlockDriverState *bs) 271 { 272 bdrv_drained_begin(bs); 273 bdrv_drained_end(bs); 274 } 275 276 /* 277 * Wait for pending requests to complete across all BlockDriverStates 278 * 279 * This function does not flush data to disk, use bdrv_flush_all() for that 280 * after calling this function. 281 */ 282 void bdrv_drain_all(void) 283 { 284 /* Always run first iteration so any pending completion BHs run */ 285 bool busy = true; 286 BlockDriverState *bs; 287 BdrvNextIterator it; 288 BlockJob *job = NULL; 289 GSList *aio_ctxs = NULL, *ctx; 290 291 while ((job = block_job_next(job))) { 292 AioContext *aio_context = blk_get_aio_context(job->blk); 293 294 aio_context_acquire(aio_context); 295 block_job_pause(job); 296 aio_context_release(aio_context); 297 } 298 299 for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) { 300 AioContext *aio_context = bdrv_get_aio_context(bs); 301 302 aio_context_acquire(aio_context); 303 bdrv_parent_drained_begin(bs); 304 bdrv_io_unplugged_begin(bs); 305 bdrv_drain_recurse(bs); 306 aio_context_release(aio_context); 307 308 if (!g_slist_find(aio_ctxs, aio_context)) { 309 aio_ctxs = g_slist_prepend(aio_ctxs, aio_context); 310 } 311 } 312 313 /* Note that completion of an asynchronous I/O operation can trigger any 314 * number of other I/O operations on other devices---for example a 315 * coroutine can submit an I/O request to another device in response to 316 * request completion. Therefore we must keep looping until there was no 317 * more activity rather than simply draining each device independently. 318 */ 319 while (busy) { 320 busy = false; 321 322 for (ctx = aio_ctxs; ctx != NULL; ctx = ctx->next) { 323 AioContext *aio_context = ctx->data; 324 325 aio_context_acquire(aio_context); 326 for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) { 327 if (aio_context == bdrv_get_aio_context(bs)) { 328 if (bdrv_requests_pending(bs)) { 329 busy = true; 330 aio_poll(aio_context, busy); 331 } 332 } 333 } 334 busy |= aio_poll(aio_context, false); 335 aio_context_release(aio_context); 336 } 337 } 338 339 for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) { 340 AioContext *aio_context = bdrv_get_aio_context(bs); 341 342 aio_context_acquire(aio_context); 343 bdrv_io_unplugged_end(bs); 344 bdrv_parent_drained_end(bs); 345 aio_context_release(aio_context); 346 } 347 g_slist_free(aio_ctxs); 348 349 job = NULL; 350 while ((job = block_job_next(job))) { 351 AioContext *aio_context = blk_get_aio_context(job->blk); 352 353 aio_context_acquire(aio_context); 354 block_job_resume(job); 355 aio_context_release(aio_context); 356 } 357 } 358 359 /** 360 * Remove an active request from the tracked requests list 361 * 362 * This function should be called when a tracked request is completing. 363 */ 364 static void tracked_request_end(BdrvTrackedRequest *req) 365 { 366 if (req->serialising) { 367 req->bs->serialising_in_flight--; 368 } 369 370 QLIST_REMOVE(req, list); 371 qemu_co_queue_restart_all(&req->wait_queue); 372 } 373 374 /** 375 * Add an active request to the tracked requests list 376 */ 377 static void tracked_request_begin(BdrvTrackedRequest *req, 378 BlockDriverState *bs, 379 int64_t offset, 380 unsigned int bytes, 381 enum BdrvTrackedRequestType type) 382 { 383 *req = (BdrvTrackedRequest){ 384 .bs = bs, 385 .offset = offset, 386 .bytes = bytes, 387 .type = type, 388 .co = qemu_coroutine_self(), 389 .serialising = false, 390 .overlap_offset = offset, 391 .overlap_bytes = bytes, 392 }; 393 394 qemu_co_queue_init(&req->wait_queue); 395 396 QLIST_INSERT_HEAD(&bs->tracked_requests, req, list); 397 } 398 399 static void mark_request_serialising(BdrvTrackedRequest *req, uint64_t align) 400 { 401 int64_t overlap_offset = req->offset & ~(align - 1); 402 unsigned int overlap_bytes = ROUND_UP(req->offset + req->bytes, align) 403 - overlap_offset; 404 405 if (!req->serialising) { 406 req->bs->serialising_in_flight++; 407 req->serialising = true; 408 } 409 410 req->overlap_offset = MIN(req->overlap_offset, overlap_offset); 411 req->overlap_bytes = MAX(req->overlap_bytes, overlap_bytes); 412 } 413 414 /** 415 * Round a region to cluster boundaries (sector-based) 416 */ 417 void bdrv_round_sectors_to_clusters(BlockDriverState *bs, 418 int64_t sector_num, int nb_sectors, 419 int64_t *cluster_sector_num, 420 int *cluster_nb_sectors) 421 { 422 BlockDriverInfo bdi; 423 424 if (bdrv_get_info(bs, &bdi) < 0 || bdi.cluster_size == 0) { 425 *cluster_sector_num = sector_num; 426 *cluster_nb_sectors = nb_sectors; 427 } else { 428 int64_t c = bdi.cluster_size / BDRV_SECTOR_SIZE; 429 *cluster_sector_num = QEMU_ALIGN_DOWN(sector_num, c); 430 *cluster_nb_sectors = QEMU_ALIGN_UP(sector_num - *cluster_sector_num + 431 nb_sectors, c); 432 } 433 } 434 435 /** 436 * Round a region to cluster boundaries 437 */ 438 void bdrv_round_to_clusters(BlockDriverState *bs, 439 int64_t offset, unsigned int bytes, 440 int64_t *cluster_offset, 441 unsigned int *cluster_bytes) 442 { 443 BlockDriverInfo bdi; 444 445 if (bdrv_get_info(bs, &bdi) < 0 || bdi.cluster_size == 0) { 446 *cluster_offset = offset; 447 *cluster_bytes = bytes; 448 } else { 449 int64_t c = bdi.cluster_size; 450 *cluster_offset = QEMU_ALIGN_DOWN(offset, c); 451 *cluster_bytes = QEMU_ALIGN_UP(offset - *cluster_offset + bytes, c); 452 } 453 } 454 455 static int bdrv_get_cluster_size(BlockDriverState *bs) 456 { 457 BlockDriverInfo bdi; 458 int ret; 459 460 ret = bdrv_get_info(bs, &bdi); 461 if (ret < 0 || bdi.cluster_size == 0) { 462 return bs->bl.request_alignment; 463 } else { 464 return bdi.cluster_size; 465 } 466 } 467 468 static bool tracked_request_overlaps(BdrvTrackedRequest *req, 469 int64_t offset, unsigned int bytes) 470 { 471 /* aaaa bbbb */ 472 if (offset >= req->overlap_offset + req->overlap_bytes) { 473 return false; 474 } 475 /* bbbb aaaa */ 476 if (req->overlap_offset >= offset + bytes) { 477 return false; 478 } 479 return true; 480 } 481 482 static bool coroutine_fn wait_serialising_requests(BdrvTrackedRequest *self) 483 { 484 BlockDriverState *bs = self->bs; 485 BdrvTrackedRequest *req; 486 bool retry; 487 bool waited = false; 488 489 if (!bs->serialising_in_flight) { 490 return false; 491 } 492 493 do { 494 retry = false; 495 QLIST_FOREACH(req, &bs->tracked_requests, list) { 496 if (req == self || (!req->serialising && !self->serialising)) { 497 continue; 498 } 499 if (tracked_request_overlaps(req, self->overlap_offset, 500 self->overlap_bytes)) 501 { 502 /* Hitting this means there was a reentrant request, for 503 * example, a block driver issuing nested requests. This must 504 * never happen since it means deadlock. 505 */ 506 assert(qemu_coroutine_self() != req->co); 507 508 /* If the request is already (indirectly) waiting for us, or 509 * will wait for us as soon as it wakes up, then just go on 510 * (instead of producing a deadlock in the former case). */ 511 if (!req->waiting_for) { 512 self->waiting_for = req; 513 qemu_co_queue_wait(&req->wait_queue); 514 self->waiting_for = NULL; 515 retry = true; 516 waited = true; 517 break; 518 } 519 } 520 } 521 } while (retry); 522 523 return waited; 524 } 525 526 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset, 527 size_t size) 528 { 529 if (size > BDRV_REQUEST_MAX_SECTORS << BDRV_SECTOR_BITS) { 530 return -EIO; 531 } 532 533 if (!bdrv_is_inserted(bs)) { 534 return -ENOMEDIUM; 535 } 536 537 if (offset < 0) { 538 return -EIO; 539 } 540 541 return 0; 542 } 543 544 static int bdrv_check_request(BlockDriverState *bs, int64_t sector_num, 545 int nb_sectors) 546 { 547 if (nb_sectors < 0 || nb_sectors > BDRV_REQUEST_MAX_SECTORS) { 548 return -EIO; 549 } 550 551 return bdrv_check_byte_request(bs, sector_num * BDRV_SECTOR_SIZE, 552 nb_sectors * BDRV_SECTOR_SIZE); 553 } 554 555 typedef struct RwCo { 556 BlockDriverState *bs; 557 int64_t offset; 558 QEMUIOVector *qiov; 559 bool is_write; 560 int ret; 561 BdrvRequestFlags flags; 562 } RwCo; 563 564 static void coroutine_fn bdrv_rw_co_entry(void *opaque) 565 { 566 RwCo *rwco = opaque; 567 568 if (!rwco->is_write) { 569 rwco->ret = bdrv_co_preadv(rwco->bs, rwco->offset, 570 rwco->qiov->size, rwco->qiov, 571 rwco->flags); 572 } else { 573 rwco->ret = bdrv_co_pwritev(rwco->bs, rwco->offset, 574 rwco->qiov->size, rwco->qiov, 575 rwco->flags); 576 } 577 } 578 579 /* 580 * Process a vectored synchronous request using coroutines 581 */ 582 static int bdrv_prwv_co(BlockDriverState *bs, int64_t offset, 583 QEMUIOVector *qiov, bool is_write, 584 BdrvRequestFlags flags) 585 { 586 Coroutine *co; 587 RwCo rwco = { 588 .bs = bs, 589 .offset = offset, 590 .qiov = qiov, 591 .is_write = is_write, 592 .ret = NOT_DONE, 593 .flags = flags, 594 }; 595 596 if (qemu_in_coroutine()) { 597 /* Fast-path if already in coroutine context */ 598 bdrv_rw_co_entry(&rwco); 599 } else { 600 AioContext *aio_context = bdrv_get_aio_context(bs); 601 602 co = qemu_coroutine_create(bdrv_rw_co_entry); 603 qemu_coroutine_enter(co, &rwco); 604 while (rwco.ret == NOT_DONE) { 605 aio_poll(aio_context, true); 606 } 607 } 608 return rwco.ret; 609 } 610 611 /* 612 * Process a synchronous request using coroutines 613 */ 614 static int bdrv_rw_co(BlockDriverState *bs, int64_t sector_num, uint8_t *buf, 615 int nb_sectors, bool is_write, BdrvRequestFlags flags) 616 { 617 QEMUIOVector qiov; 618 struct iovec iov = { 619 .iov_base = (void *)buf, 620 .iov_len = nb_sectors * BDRV_SECTOR_SIZE, 621 }; 622 623 if (nb_sectors < 0 || nb_sectors > BDRV_REQUEST_MAX_SECTORS) { 624 return -EINVAL; 625 } 626 627 qemu_iovec_init_external(&qiov, &iov, 1); 628 return bdrv_prwv_co(bs, sector_num << BDRV_SECTOR_BITS, 629 &qiov, is_write, flags); 630 } 631 632 /* return < 0 if error. See bdrv_write() for the return codes */ 633 int bdrv_read(BlockDriverState *bs, int64_t sector_num, 634 uint8_t *buf, int nb_sectors) 635 { 636 return bdrv_rw_co(bs, sector_num, buf, nb_sectors, false, 0); 637 } 638 639 /* Return < 0 if error. Important errors are: 640 -EIO generic I/O error (may happen for all errors) 641 -ENOMEDIUM No media inserted. 642 -EINVAL Invalid sector number or nb_sectors 643 -EACCES Trying to write a read-only device 644 */ 645 int bdrv_write(BlockDriverState *bs, int64_t sector_num, 646 const uint8_t *buf, int nb_sectors) 647 { 648 return bdrv_rw_co(bs, sector_num, (uint8_t *)buf, nb_sectors, true, 0); 649 } 650 651 int bdrv_pwrite_zeroes(BlockDriverState *bs, int64_t offset, 652 int count, BdrvRequestFlags flags) 653 { 654 QEMUIOVector qiov; 655 struct iovec iov = { 656 .iov_base = NULL, 657 .iov_len = count, 658 }; 659 660 qemu_iovec_init_external(&qiov, &iov, 1); 661 return bdrv_prwv_co(bs, offset, &qiov, true, 662 BDRV_REQ_ZERO_WRITE | flags); 663 } 664 665 /* 666 * Completely zero out a block device with the help of bdrv_pwrite_zeroes. 667 * The operation is sped up by checking the block status and only writing 668 * zeroes to the device if they currently do not return zeroes. Optional 669 * flags are passed through to bdrv_pwrite_zeroes (e.g. BDRV_REQ_MAY_UNMAP, 670 * BDRV_REQ_FUA). 671 * 672 * Returns < 0 on error, 0 on success. For error codes see bdrv_write(). 673 */ 674 int bdrv_make_zero(BlockDriverState *bs, BdrvRequestFlags flags) 675 { 676 int64_t target_sectors, ret, nb_sectors, sector_num = 0; 677 BlockDriverState *file; 678 int n; 679 680 target_sectors = bdrv_nb_sectors(bs); 681 if (target_sectors < 0) { 682 return target_sectors; 683 } 684 685 for (;;) { 686 nb_sectors = MIN(target_sectors - sector_num, BDRV_REQUEST_MAX_SECTORS); 687 if (nb_sectors <= 0) { 688 return 0; 689 } 690 ret = bdrv_get_block_status(bs, sector_num, nb_sectors, &n, &file); 691 if (ret < 0) { 692 error_report("error getting block status at sector %" PRId64 ": %s", 693 sector_num, strerror(-ret)); 694 return ret; 695 } 696 if (ret & BDRV_BLOCK_ZERO) { 697 sector_num += n; 698 continue; 699 } 700 ret = bdrv_pwrite_zeroes(bs, sector_num << BDRV_SECTOR_BITS, 701 n << BDRV_SECTOR_BITS, flags); 702 if (ret < 0) { 703 error_report("error writing zeroes at sector %" PRId64 ": %s", 704 sector_num, strerror(-ret)); 705 return ret; 706 } 707 sector_num += n; 708 } 709 } 710 711 int bdrv_preadv(BlockDriverState *bs, int64_t offset, QEMUIOVector *qiov) 712 { 713 int ret; 714 715 ret = bdrv_prwv_co(bs, offset, qiov, false, 0); 716 if (ret < 0) { 717 return ret; 718 } 719 720 return qiov->size; 721 } 722 723 int bdrv_pread(BlockDriverState *bs, int64_t offset, void *buf, int bytes) 724 { 725 QEMUIOVector qiov; 726 struct iovec iov = { 727 .iov_base = (void *)buf, 728 .iov_len = bytes, 729 }; 730 731 if (bytes < 0) { 732 return -EINVAL; 733 } 734 735 qemu_iovec_init_external(&qiov, &iov, 1); 736 return bdrv_preadv(bs, offset, &qiov); 737 } 738 739 int bdrv_pwritev(BlockDriverState *bs, int64_t offset, QEMUIOVector *qiov) 740 { 741 int ret; 742 743 ret = bdrv_prwv_co(bs, offset, qiov, true, 0); 744 if (ret < 0) { 745 return ret; 746 } 747 748 return qiov->size; 749 } 750 751 int bdrv_pwrite(BlockDriverState *bs, int64_t offset, 752 const void *buf, int bytes) 753 { 754 QEMUIOVector qiov; 755 struct iovec iov = { 756 .iov_base = (void *) buf, 757 .iov_len = bytes, 758 }; 759 760 if (bytes < 0) { 761 return -EINVAL; 762 } 763 764 qemu_iovec_init_external(&qiov, &iov, 1); 765 return bdrv_pwritev(bs, offset, &qiov); 766 } 767 768 /* 769 * Writes to the file and ensures that no writes are reordered across this 770 * request (acts as a barrier) 771 * 772 * Returns 0 on success, -errno in error cases. 773 */ 774 int bdrv_pwrite_sync(BlockDriverState *bs, int64_t offset, 775 const void *buf, int count) 776 { 777 int ret; 778 779 ret = bdrv_pwrite(bs, offset, buf, count); 780 if (ret < 0) { 781 return ret; 782 } 783 784 ret = bdrv_flush(bs); 785 if (ret < 0) { 786 return ret; 787 } 788 789 return 0; 790 } 791 792 typedef struct CoroutineIOCompletion { 793 Coroutine *coroutine; 794 int ret; 795 } CoroutineIOCompletion; 796 797 static void bdrv_co_io_em_complete(void *opaque, int ret) 798 { 799 CoroutineIOCompletion *co = opaque; 800 801 co->ret = ret; 802 qemu_coroutine_enter(co->coroutine, NULL); 803 } 804 805 static int coroutine_fn bdrv_driver_preadv(BlockDriverState *bs, 806 uint64_t offset, uint64_t bytes, 807 QEMUIOVector *qiov, int flags) 808 { 809 BlockDriver *drv = bs->drv; 810 int64_t sector_num; 811 unsigned int nb_sectors; 812 813 assert(!(flags & ~BDRV_REQ_MASK)); 814 815 if (drv->bdrv_co_preadv) { 816 return drv->bdrv_co_preadv(bs, offset, bytes, qiov, flags); 817 } 818 819 sector_num = offset >> BDRV_SECTOR_BITS; 820 nb_sectors = bytes >> BDRV_SECTOR_BITS; 821 822 assert((offset & (BDRV_SECTOR_SIZE - 1)) == 0); 823 assert((bytes & (BDRV_SECTOR_SIZE - 1)) == 0); 824 assert((bytes >> BDRV_SECTOR_BITS) <= BDRV_REQUEST_MAX_SECTORS); 825 826 if (drv->bdrv_co_readv) { 827 return drv->bdrv_co_readv(bs, sector_num, nb_sectors, qiov); 828 } else { 829 BlockAIOCB *acb; 830 CoroutineIOCompletion co = { 831 .coroutine = qemu_coroutine_self(), 832 }; 833 834 acb = bs->drv->bdrv_aio_readv(bs, sector_num, qiov, nb_sectors, 835 bdrv_co_io_em_complete, &co); 836 if (acb == NULL) { 837 return -EIO; 838 } else { 839 qemu_coroutine_yield(); 840 return co.ret; 841 } 842 } 843 } 844 845 static int coroutine_fn bdrv_driver_pwritev(BlockDriverState *bs, 846 uint64_t offset, uint64_t bytes, 847 QEMUIOVector *qiov, int flags) 848 { 849 BlockDriver *drv = bs->drv; 850 int64_t sector_num; 851 unsigned int nb_sectors; 852 int ret; 853 854 assert(!(flags & ~BDRV_REQ_MASK)); 855 856 if (drv->bdrv_co_pwritev) { 857 ret = drv->bdrv_co_pwritev(bs, offset, bytes, qiov, 858 flags & bs->supported_write_flags); 859 flags &= ~bs->supported_write_flags; 860 goto emulate_flags; 861 } 862 863 sector_num = offset >> BDRV_SECTOR_BITS; 864 nb_sectors = bytes >> BDRV_SECTOR_BITS; 865 866 assert((offset & (BDRV_SECTOR_SIZE - 1)) == 0); 867 assert((bytes & (BDRV_SECTOR_SIZE - 1)) == 0); 868 assert((bytes >> BDRV_SECTOR_BITS) <= BDRV_REQUEST_MAX_SECTORS); 869 870 if (drv->bdrv_co_writev_flags) { 871 ret = drv->bdrv_co_writev_flags(bs, sector_num, nb_sectors, qiov, 872 flags & bs->supported_write_flags); 873 flags &= ~bs->supported_write_flags; 874 } else if (drv->bdrv_co_writev) { 875 assert(!bs->supported_write_flags); 876 ret = drv->bdrv_co_writev(bs, sector_num, nb_sectors, qiov); 877 } else { 878 BlockAIOCB *acb; 879 CoroutineIOCompletion co = { 880 .coroutine = qemu_coroutine_self(), 881 }; 882 883 acb = bs->drv->bdrv_aio_writev(bs, sector_num, qiov, nb_sectors, 884 bdrv_co_io_em_complete, &co); 885 if (acb == NULL) { 886 ret = -EIO; 887 } else { 888 qemu_coroutine_yield(); 889 ret = co.ret; 890 } 891 } 892 893 emulate_flags: 894 if (ret == 0 && (flags & BDRV_REQ_FUA)) { 895 ret = bdrv_co_flush(bs); 896 } 897 898 return ret; 899 } 900 901 static int coroutine_fn bdrv_co_do_copy_on_readv(BlockDriverState *bs, 902 int64_t offset, unsigned int bytes, QEMUIOVector *qiov) 903 { 904 /* Perform I/O through a temporary buffer so that users who scribble over 905 * their read buffer while the operation is in progress do not end up 906 * modifying the image file. This is critical for zero-copy guest I/O 907 * where anything might happen inside guest memory. 908 */ 909 void *bounce_buffer; 910 911 BlockDriver *drv = bs->drv; 912 struct iovec iov; 913 QEMUIOVector bounce_qiov; 914 int64_t cluster_offset; 915 unsigned int cluster_bytes; 916 size_t skip_bytes; 917 int ret; 918 919 /* Cover entire cluster so no additional backing file I/O is required when 920 * allocating cluster in the image file. 921 */ 922 bdrv_round_to_clusters(bs, offset, bytes, &cluster_offset, &cluster_bytes); 923 924 trace_bdrv_co_do_copy_on_readv(bs, offset, bytes, 925 cluster_offset, cluster_bytes); 926 927 iov.iov_len = cluster_bytes; 928 iov.iov_base = bounce_buffer = qemu_try_blockalign(bs, iov.iov_len); 929 if (bounce_buffer == NULL) { 930 ret = -ENOMEM; 931 goto err; 932 } 933 934 qemu_iovec_init_external(&bounce_qiov, &iov, 1); 935 936 ret = bdrv_driver_preadv(bs, cluster_offset, cluster_bytes, 937 &bounce_qiov, 0); 938 if (ret < 0) { 939 goto err; 940 } 941 942 if (drv->bdrv_co_pwrite_zeroes && 943 buffer_is_zero(bounce_buffer, iov.iov_len)) { 944 /* FIXME: Should we (perhaps conditionally) be setting 945 * BDRV_REQ_MAY_UNMAP, if it will allow for a sparser copy 946 * that still correctly reads as zero? */ 947 ret = bdrv_co_do_pwrite_zeroes(bs, cluster_offset, cluster_bytes, 0); 948 } else { 949 /* This does not change the data on the disk, it is not necessary 950 * to flush even in cache=writethrough mode. 951 */ 952 ret = bdrv_driver_pwritev(bs, cluster_offset, cluster_bytes, 953 &bounce_qiov, 0); 954 } 955 956 if (ret < 0) { 957 /* It might be okay to ignore write errors for guest requests. If this 958 * is a deliberate copy-on-read then we don't want to ignore the error. 959 * Simply report it in all cases. 960 */ 961 goto err; 962 } 963 964 skip_bytes = offset - cluster_offset; 965 qemu_iovec_from_buf(qiov, 0, bounce_buffer + skip_bytes, bytes); 966 967 err: 968 qemu_vfree(bounce_buffer); 969 return ret; 970 } 971 972 /* 973 * Forwards an already correctly aligned request to the BlockDriver. This 974 * handles copy on read and zeroing after EOF; any other features must be 975 * implemented by the caller. 976 */ 977 static int coroutine_fn bdrv_aligned_preadv(BlockDriverState *bs, 978 BdrvTrackedRequest *req, int64_t offset, unsigned int bytes, 979 int64_t align, QEMUIOVector *qiov, int flags) 980 { 981 int64_t total_bytes, max_bytes; 982 int ret; 983 984 assert(is_power_of_2(align)); 985 assert((offset & (align - 1)) == 0); 986 assert((bytes & (align - 1)) == 0); 987 assert(!qiov || bytes == qiov->size); 988 assert((bs->open_flags & BDRV_O_NO_IO) == 0); 989 990 /* TODO: We would need a per-BDS .supported_read_flags and 991 * potential fallback support, if we ever implement any read flags 992 * to pass through to drivers. For now, there aren't any 993 * passthrough flags. */ 994 assert(!(flags & ~(BDRV_REQ_NO_SERIALISING | BDRV_REQ_COPY_ON_READ))); 995 996 /* Handle Copy on Read and associated serialisation */ 997 if (flags & BDRV_REQ_COPY_ON_READ) { 998 /* If we touch the same cluster it counts as an overlap. This 999 * guarantees that allocating writes will be serialized and not race 1000 * with each other for the same cluster. For example, in copy-on-read 1001 * it ensures that the CoR read and write operations are atomic and 1002 * guest writes cannot interleave between them. */ 1003 mark_request_serialising(req, bdrv_get_cluster_size(bs)); 1004 } 1005 1006 if (!(flags & BDRV_REQ_NO_SERIALISING)) { 1007 wait_serialising_requests(req); 1008 } 1009 1010 if (flags & BDRV_REQ_COPY_ON_READ) { 1011 int64_t start_sector = offset >> BDRV_SECTOR_BITS; 1012 int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE); 1013 unsigned int nb_sectors = end_sector - start_sector; 1014 int pnum; 1015 1016 ret = bdrv_is_allocated(bs, start_sector, nb_sectors, &pnum); 1017 if (ret < 0) { 1018 goto out; 1019 } 1020 1021 if (!ret || pnum != nb_sectors) { 1022 ret = bdrv_co_do_copy_on_readv(bs, offset, bytes, qiov); 1023 goto out; 1024 } 1025 } 1026 1027 /* Forward the request to the BlockDriver */ 1028 total_bytes = bdrv_getlength(bs); 1029 if (total_bytes < 0) { 1030 ret = total_bytes; 1031 goto out; 1032 } 1033 1034 max_bytes = ROUND_UP(MAX(0, total_bytes - offset), align); 1035 if (bytes <= max_bytes) { 1036 ret = bdrv_driver_preadv(bs, offset, bytes, qiov, 0); 1037 } else if (max_bytes > 0) { 1038 QEMUIOVector local_qiov; 1039 1040 qemu_iovec_init(&local_qiov, qiov->niov); 1041 qemu_iovec_concat(&local_qiov, qiov, 0, max_bytes); 1042 1043 ret = bdrv_driver_preadv(bs, offset, max_bytes, &local_qiov, 0); 1044 1045 qemu_iovec_destroy(&local_qiov); 1046 } else { 1047 ret = 0; 1048 } 1049 1050 /* Reading beyond end of file is supposed to produce zeroes */ 1051 if (ret == 0 && total_bytes < offset + bytes) { 1052 uint64_t zero_offset = MAX(0, total_bytes - offset); 1053 uint64_t zero_bytes = offset + bytes - zero_offset; 1054 qemu_iovec_memset(qiov, zero_offset, 0, zero_bytes); 1055 } 1056 1057 out: 1058 return ret; 1059 } 1060 1061 /* 1062 * Handle a read request in coroutine context 1063 */ 1064 int coroutine_fn bdrv_co_preadv(BlockDriverState *bs, 1065 int64_t offset, unsigned int bytes, QEMUIOVector *qiov, 1066 BdrvRequestFlags flags) 1067 { 1068 BlockDriver *drv = bs->drv; 1069 BdrvTrackedRequest req; 1070 1071 uint64_t align = bs->bl.request_alignment; 1072 uint8_t *head_buf = NULL; 1073 uint8_t *tail_buf = NULL; 1074 QEMUIOVector local_qiov; 1075 bool use_local_qiov = false; 1076 int ret; 1077 1078 if (!drv) { 1079 return -ENOMEDIUM; 1080 } 1081 1082 ret = bdrv_check_byte_request(bs, offset, bytes); 1083 if (ret < 0) { 1084 return ret; 1085 } 1086 1087 /* Don't do copy-on-read if we read data before write operation */ 1088 if (bs->copy_on_read && !(flags & BDRV_REQ_NO_SERIALISING)) { 1089 flags |= BDRV_REQ_COPY_ON_READ; 1090 } 1091 1092 /* Align read if necessary by padding qiov */ 1093 if (offset & (align - 1)) { 1094 head_buf = qemu_blockalign(bs, align); 1095 qemu_iovec_init(&local_qiov, qiov->niov + 2); 1096 qemu_iovec_add(&local_qiov, head_buf, offset & (align - 1)); 1097 qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size); 1098 use_local_qiov = true; 1099 1100 bytes += offset & (align - 1); 1101 offset = offset & ~(align - 1); 1102 } 1103 1104 if ((offset + bytes) & (align - 1)) { 1105 if (!use_local_qiov) { 1106 qemu_iovec_init(&local_qiov, qiov->niov + 1); 1107 qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size); 1108 use_local_qiov = true; 1109 } 1110 tail_buf = qemu_blockalign(bs, align); 1111 qemu_iovec_add(&local_qiov, tail_buf, 1112 align - ((offset + bytes) & (align - 1))); 1113 1114 bytes = ROUND_UP(bytes, align); 1115 } 1116 1117 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_READ); 1118 ret = bdrv_aligned_preadv(bs, &req, offset, bytes, align, 1119 use_local_qiov ? &local_qiov : qiov, 1120 flags); 1121 tracked_request_end(&req); 1122 1123 if (use_local_qiov) { 1124 qemu_iovec_destroy(&local_qiov); 1125 qemu_vfree(head_buf); 1126 qemu_vfree(tail_buf); 1127 } 1128 1129 return ret; 1130 } 1131 1132 static int coroutine_fn bdrv_co_do_readv(BlockDriverState *bs, 1133 int64_t sector_num, int nb_sectors, QEMUIOVector *qiov, 1134 BdrvRequestFlags flags) 1135 { 1136 if (nb_sectors < 0 || nb_sectors > BDRV_REQUEST_MAX_SECTORS) { 1137 return -EINVAL; 1138 } 1139 1140 return bdrv_co_preadv(bs, sector_num << BDRV_SECTOR_BITS, 1141 nb_sectors << BDRV_SECTOR_BITS, qiov, flags); 1142 } 1143 1144 int coroutine_fn bdrv_co_readv(BlockDriverState *bs, int64_t sector_num, 1145 int nb_sectors, QEMUIOVector *qiov) 1146 { 1147 trace_bdrv_co_readv(bs, sector_num, nb_sectors); 1148 1149 return bdrv_co_do_readv(bs, sector_num, nb_sectors, qiov, 0); 1150 } 1151 1152 /* Maximum buffer for write zeroes fallback, in bytes */ 1153 #define MAX_WRITE_ZEROES_BOUNCE_BUFFER (32768 << BDRV_SECTOR_BITS) 1154 1155 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs, 1156 int64_t offset, int count, BdrvRequestFlags flags) 1157 { 1158 BlockDriver *drv = bs->drv; 1159 QEMUIOVector qiov; 1160 struct iovec iov = {0}; 1161 int ret = 0; 1162 bool need_flush = false; 1163 int head = 0; 1164 int tail = 0; 1165 1166 int max_write_zeroes = MIN_NON_ZERO(bs->bl.max_pwrite_zeroes, INT_MAX); 1167 int alignment = MAX(bs->bl.pwrite_zeroes_alignment, 1168 bs->bl.request_alignment); 1169 1170 assert(is_power_of_2(alignment)); 1171 head = offset & (alignment - 1); 1172 tail = (offset + count) & (alignment - 1); 1173 max_write_zeroes &= ~(alignment - 1); 1174 1175 while (count > 0 && !ret) { 1176 int num = count; 1177 1178 /* Align request. Block drivers can expect the "bulk" of the request 1179 * to be aligned, and that unaligned requests do not cross cluster 1180 * boundaries. 1181 */ 1182 if (head) { 1183 /* Make a small request up to the first aligned sector. */ 1184 num = MIN(count, alignment - head); 1185 head = 0; 1186 } else if (tail && num > alignment) { 1187 /* Shorten the request to the last aligned sector. */ 1188 num -= tail; 1189 } 1190 1191 /* limit request size */ 1192 if (num > max_write_zeroes) { 1193 num = max_write_zeroes; 1194 } 1195 1196 ret = -ENOTSUP; 1197 /* First try the efficient write zeroes operation */ 1198 if (drv->bdrv_co_pwrite_zeroes) { 1199 ret = drv->bdrv_co_pwrite_zeroes(bs, offset, num, 1200 flags & bs->supported_zero_flags); 1201 if (ret != -ENOTSUP && (flags & BDRV_REQ_FUA) && 1202 !(bs->supported_zero_flags & BDRV_REQ_FUA)) { 1203 need_flush = true; 1204 } 1205 } else { 1206 assert(!bs->supported_zero_flags); 1207 } 1208 1209 if (ret == -ENOTSUP) { 1210 /* Fall back to bounce buffer if write zeroes is unsupported */ 1211 int max_transfer = MIN_NON_ZERO(bs->bl.max_transfer, 1212 MAX_WRITE_ZEROES_BOUNCE_BUFFER); 1213 BdrvRequestFlags write_flags = flags & ~BDRV_REQ_ZERO_WRITE; 1214 1215 if ((flags & BDRV_REQ_FUA) && 1216 !(bs->supported_write_flags & BDRV_REQ_FUA)) { 1217 /* No need for bdrv_driver_pwrite() to do a fallback 1218 * flush on each chunk; use just one at the end */ 1219 write_flags &= ~BDRV_REQ_FUA; 1220 need_flush = true; 1221 } 1222 num = MIN(num, max_transfer); 1223 iov.iov_len = num; 1224 if (iov.iov_base == NULL) { 1225 iov.iov_base = qemu_try_blockalign(bs, num); 1226 if (iov.iov_base == NULL) { 1227 ret = -ENOMEM; 1228 goto fail; 1229 } 1230 memset(iov.iov_base, 0, num); 1231 } 1232 qemu_iovec_init_external(&qiov, &iov, 1); 1233 1234 ret = bdrv_driver_pwritev(bs, offset, num, &qiov, write_flags); 1235 1236 /* Keep bounce buffer around if it is big enough for all 1237 * all future requests. 1238 */ 1239 if (num < max_transfer) { 1240 qemu_vfree(iov.iov_base); 1241 iov.iov_base = NULL; 1242 } 1243 } 1244 1245 offset += num; 1246 count -= num; 1247 } 1248 1249 fail: 1250 if (ret == 0 && need_flush) { 1251 ret = bdrv_co_flush(bs); 1252 } 1253 qemu_vfree(iov.iov_base); 1254 return ret; 1255 } 1256 1257 /* 1258 * Forwards an already correctly aligned write request to the BlockDriver. 1259 */ 1260 static int coroutine_fn bdrv_aligned_pwritev(BlockDriverState *bs, 1261 BdrvTrackedRequest *req, int64_t offset, unsigned int bytes, 1262 int64_t align, QEMUIOVector *qiov, int flags) 1263 { 1264 BlockDriver *drv = bs->drv; 1265 bool waited; 1266 int ret; 1267 1268 int64_t start_sector = offset >> BDRV_SECTOR_BITS; 1269 int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE); 1270 1271 assert(is_power_of_2(align)); 1272 assert((offset & (align - 1)) == 0); 1273 assert((bytes & (align - 1)) == 0); 1274 assert(!qiov || bytes == qiov->size); 1275 assert((bs->open_flags & BDRV_O_NO_IO) == 0); 1276 assert(!(flags & ~BDRV_REQ_MASK)); 1277 1278 waited = wait_serialising_requests(req); 1279 assert(!waited || !req->serialising); 1280 assert(req->overlap_offset <= offset); 1281 assert(offset + bytes <= req->overlap_offset + req->overlap_bytes); 1282 1283 ret = notifier_with_return_list_notify(&bs->before_write_notifiers, req); 1284 1285 if (!ret && bs->detect_zeroes != BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF && 1286 !(flags & BDRV_REQ_ZERO_WRITE) && drv->bdrv_co_pwrite_zeroes && 1287 qemu_iovec_is_zero(qiov)) { 1288 flags |= BDRV_REQ_ZERO_WRITE; 1289 if (bs->detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP) { 1290 flags |= BDRV_REQ_MAY_UNMAP; 1291 } 1292 } 1293 1294 if (ret < 0) { 1295 /* Do nothing, write notifier decided to fail this request */ 1296 } else if (flags & BDRV_REQ_ZERO_WRITE) { 1297 bdrv_debug_event(bs, BLKDBG_PWRITEV_ZERO); 1298 ret = bdrv_co_do_pwrite_zeroes(bs, offset, bytes, flags); 1299 } else { 1300 bdrv_debug_event(bs, BLKDBG_PWRITEV); 1301 ret = bdrv_driver_pwritev(bs, offset, bytes, qiov, flags); 1302 } 1303 bdrv_debug_event(bs, BLKDBG_PWRITEV_DONE); 1304 1305 bdrv_set_dirty(bs, start_sector, end_sector - start_sector); 1306 1307 if (bs->wr_highest_offset < offset + bytes) { 1308 bs->wr_highest_offset = offset + bytes; 1309 } 1310 1311 if (ret >= 0) { 1312 bs->total_sectors = MAX(bs->total_sectors, end_sector); 1313 } 1314 1315 return ret; 1316 } 1317 1318 static int coroutine_fn bdrv_co_do_zero_pwritev(BlockDriverState *bs, 1319 int64_t offset, 1320 unsigned int bytes, 1321 BdrvRequestFlags flags, 1322 BdrvTrackedRequest *req) 1323 { 1324 uint8_t *buf = NULL; 1325 QEMUIOVector local_qiov; 1326 struct iovec iov; 1327 uint64_t align = bs->bl.request_alignment; 1328 unsigned int head_padding_bytes, tail_padding_bytes; 1329 int ret = 0; 1330 1331 head_padding_bytes = offset & (align - 1); 1332 tail_padding_bytes = align - ((offset + bytes) & (align - 1)); 1333 1334 1335 assert(flags & BDRV_REQ_ZERO_WRITE); 1336 if (head_padding_bytes || tail_padding_bytes) { 1337 buf = qemu_blockalign(bs, align); 1338 iov = (struct iovec) { 1339 .iov_base = buf, 1340 .iov_len = align, 1341 }; 1342 qemu_iovec_init_external(&local_qiov, &iov, 1); 1343 } 1344 if (head_padding_bytes) { 1345 uint64_t zero_bytes = MIN(bytes, align - head_padding_bytes); 1346 1347 /* RMW the unaligned part before head. */ 1348 mark_request_serialising(req, align); 1349 wait_serialising_requests(req); 1350 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD); 1351 ret = bdrv_aligned_preadv(bs, req, offset & ~(align - 1), align, 1352 align, &local_qiov, 0); 1353 if (ret < 0) { 1354 goto fail; 1355 } 1356 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD); 1357 1358 memset(buf + head_padding_bytes, 0, zero_bytes); 1359 ret = bdrv_aligned_pwritev(bs, req, offset & ~(align - 1), align, 1360 align, &local_qiov, 1361 flags & ~BDRV_REQ_ZERO_WRITE); 1362 if (ret < 0) { 1363 goto fail; 1364 } 1365 offset += zero_bytes; 1366 bytes -= zero_bytes; 1367 } 1368 1369 assert(!bytes || (offset & (align - 1)) == 0); 1370 if (bytes >= align) { 1371 /* Write the aligned part in the middle. */ 1372 uint64_t aligned_bytes = bytes & ~(align - 1); 1373 ret = bdrv_aligned_pwritev(bs, req, offset, aligned_bytes, align, 1374 NULL, flags); 1375 if (ret < 0) { 1376 goto fail; 1377 } 1378 bytes -= aligned_bytes; 1379 offset += aligned_bytes; 1380 } 1381 1382 assert(!bytes || (offset & (align - 1)) == 0); 1383 if (bytes) { 1384 assert(align == tail_padding_bytes + bytes); 1385 /* RMW the unaligned part after tail. */ 1386 mark_request_serialising(req, align); 1387 wait_serialising_requests(req); 1388 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL); 1389 ret = bdrv_aligned_preadv(bs, req, offset, align, 1390 align, &local_qiov, 0); 1391 if (ret < 0) { 1392 goto fail; 1393 } 1394 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL); 1395 1396 memset(buf, 0, bytes); 1397 ret = bdrv_aligned_pwritev(bs, req, offset, align, align, 1398 &local_qiov, flags & ~BDRV_REQ_ZERO_WRITE); 1399 } 1400 fail: 1401 qemu_vfree(buf); 1402 return ret; 1403 1404 } 1405 1406 /* 1407 * Handle a write request in coroutine context 1408 */ 1409 int coroutine_fn bdrv_co_pwritev(BlockDriverState *bs, 1410 int64_t offset, unsigned int bytes, QEMUIOVector *qiov, 1411 BdrvRequestFlags flags) 1412 { 1413 BdrvTrackedRequest req; 1414 uint64_t align = bs->bl.request_alignment; 1415 uint8_t *head_buf = NULL; 1416 uint8_t *tail_buf = NULL; 1417 QEMUIOVector local_qiov; 1418 bool use_local_qiov = false; 1419 int ret; 1420 1421 if (!bs->drv) { 1422 return -ENOMEDIUM; 1423 } 1424 if (bs->read_only) { 1425 return -EPERM; 1426 } 1427 assert(!(bs->open_flags & BDRV_O_INACTIVE)); 1428 1429 ret = bdrv_check_byte_request(bs, offset, bytes); 1430 if (ret < 0) { 1431 return ret; 1432 } 1433 1434 /* 1435 * Align write if necessary by performing a read-modify-write cycle. 1436 * Pad qiov with the read parts and be sure to have a tracked request not 1437 * only for bdrv_aligned_pwritev, but also for the reads of the RMW cycle. 1438 */ 1439 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_WRITE); 1440 1441 if (!qiov) { 1442 ret = bdrv_co_do_zero_pwritev(bs, offset, bytes, flags, &req); 1443 goto out; 1444 } 1445 1446 if (offset & (align - 1)) { 1447 QEMUIOVector head_qiov; 1448 struct iovec head_iov; 1449 1450 mark_request_serialising(&req, align); 1451 wait_serialising_requests(&req); 1452 1453 head_buf = qemu_blockalign(bs, align); 1454 head_iov = (struct iovec) { 1455 .iov_base = head_buf, 1456 .iov_len = align, 1457 }; 1458 qemu_iovec_init_external(&head_qiov, &head_iov, 1); 1459 1460 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD); 1461 ret = bdrv_aligned_preadv(bs, &req, offset & ~(align - 1), align, 1462 align, &head_qiov, 0); 1463 if (ret < 0) { 1464 goto fail; 1465 } 1466 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD); 1467 1468 qemu_iovec_init(&local_qiov, qiov->niov + 2); 1469 qemu_iovec_add(&local_qiov, head_buf, offset & (align - 1)); 1470 qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size); 1471 use_local_qiov = true; 1472 1473 bytes += offset & (align - 1); 1474 offset = offset & ~(align - 1); 1475 1476 /* We have read the tail already if the request is smaller 1477 * than one aligned block. 1478 */ 1479 if (bytes < align) { 1480 qemu_iovec_add(&local_qiov, head_buf + bytes, align - bytes); 1481 bytes = align; 1482 } 1483 } 1484 1485 if ((offset + bytes) & (align - 1)) { 1486 QEMUIOVector tail_qiov; 1487 struct iovec tail_iov; 1488 size_t tail_bytes; 1489 bool waited; 1490 1491 mark_request_serialising(&req, align); 1492 waited = wait_serialising_requests(&req); 1493 assert(!waited || !use_local_qiov); 1494 1495 tail_buf = qemu_blockalign(bs, align); 1496 tail_iov = (struct iovec) { 1497 .iov_base = tail_buf, 1498 .iov_len = align, 1499 }; 1500 qemu_iovec_init_external(&tail_qiov, &tail_iov, 1); 1501 1502 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL); 1503 ret = bdrv_aligned_preadv(bs, &req, (offset + bytes) & ~(align - 1), align, 1504 align, &tail_qiov, 0); 1505 if (ret < 0) { 1506 goto fail; 1507 } 1508 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL); 1509 1510 if (!use_local_qiov) { 1511 qemu_iovec_init(&local_qiov, qiov->niov + 1); 1512 qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size); 1513 use_local_qiov = true; 1514 } 1515 1516 tail_bytes = (offset + bytes) & (align - 1); 1517 qemu_iovec_add(&local_qiov, tail_buf + tail_bytes, align - tail_bytes); 1518 1519 bytes = ROUND_UP(bytes, align); 1520 } 1521 1522 ret = bdrv_aligned_pwritev(bs, &req, offset, bytes, align, 1523 use_local_qiov ? &local_qiov : qiov, 1524 flags); 1525 1526 fail: 1527 1528 if (use_local_qiov) { 1529 qemu_iovec_destroy(&local_qiov); 1530 } 1531 qemu_vfree(head_buf); 1532 qemu_vfree(tail_buf); 1533 out: 1534 tracked_request_end(&req); 1535 return ret; 1536 } 1537 1538 static int coroutine_fn bdrv_co_do_writev(BlockDriverState *bs, 1539 int64_t sector_num, int nb_sectors, QEMUIOVector *qiov, 1540 BdrvRequestFlags flags) 1541 { 1542 if (nb_sectors < 0 || nb_sectors > BDRV_REQUEST_MAX_SECTORS) { 1543 return -EINVAL; 1544 } 1545 1546 return bdrv_co_pwritev(bs, sector_num << BDRV_SECTOR_BITS, 1547 nb_sectors << BDRV_SECTOR_BITS, qiov, flags); 1548 } 1549 1550 int coroutine_fn bdrv_co_writev(BlockDriverState *bs, int64_t sector_num, 1551 int nb_sectors, QEMUIOVector *qiov) 1552 { 1553 trace_bdrv_co_writev(bs, sector_num, nb_sectors); 1554 1555 return bdrv_co_do_writev(bs, sector_num, nb_sectors, qiov, 0); 1556 } 1557 1558 int coroutine_fn bdrv_co_pwrite_zeroes(BlockDriverState *bs, 1559 int64_t offset, int count, 1560 BdrvRequestFlags flags) 1561 { 1562 trace_bdrv_co_pwrite_zeroes(bs, offset, count, flags); 1563 1564 if (!(bs->open_flags & BDRV_O_UNMAP)) { 1565 flags &= ~BDRV_REQ_MAY_UNMAP; 1566 } 1567 1568 return bdrv_co_pwritev(bs, offset, count, NULL, 1569 BDRV_REQ_ZERO_WRITE | flags); 1570 } 1571 1572 typedef struct BdrvCoGetBlockStatusData { 1573 BlockDriverState *bs; 1574 BlockDriverState *base; 1575 BlockDriverState **file; 1576 int64_t sector_num; 1577 int nb_sectors; 1578 int *pnum; 1579 int64_t ret; 1580 bool done; 1581 } BdrvCoGetBlockStatusData; 1582 1583 /* 1584 * Returns the allocation status of the specified sectors. 1585 * Drivers not implementing the functionality are assumed to not support 1586 * backing files, hence all their sectors are reported as allocated. 1587 * 1588 * If 'sector_num' is beyond the end of the disk image the return value is 0 1589 * and 'pnum' is set to 0. 1590 * 1591 * 'pnum' is set to the number of sectors (including and immediately following 1592 * the specified sector) that are known to be in the same 1593 * allocated/unallocated state. 1594 * 1595 * 'nb_sectors' is the max value 'pnum' should be set to. If nb_sectors goes 1596 * beyond the end of the disk image it will be clamped. 1597 * 1598 * If returned value is positive and BDRV_BLOCK_OFFSET_VALID bit is set, 'file' 1599 * points to the BDS which the sector range is allocated in. 1600 */ 1601 static int64_t coroutine_fn bdrv_co_get_block_status(BlockDriverState *bs, 1602 int64_t sector_num, 1603 int nb_sectors, int *pnum, 1604 BlockDriverState **file) 1605 { 1606 int64_t total_sectors; 1607 int64_t n; 1608 int64_t ret, ret2; 1609 1610 total_sectors = bdrv_nb_sectors(bs); 1611 if (total_sectors < 0) { 1612 return total_sectors; 1613 } 1614 1615 if (sector_num >= total_sectors) { 1616 *pnum = 0; 1617 return 0; 1618 } 1619 1620 n = total_sectors - sector_num; 1621 if (n < nb_sectors) { 1622 nb_sectors = n; 1623 } 1624 1625 if (!bs->drv->bdrv_co_get_block_status) { 1626 *pnum = nb_sectors; 1627 ret = BDRV_BLOCK_DATA | BDRV_BLOCK_ALLOCATED; 1628 if (bs->drv->protocol_name) { 1629 ret |= BDRV_BLOCK_OFFSET_VALID | (sector_num * BDRV_SECTOR_SIZE); 1630 } 1631 return ret; 1632 } 1633 1634 *file = NULL; 1635 ret = bs->drv->bdrv_co_get_block_status(bs, sector_num, nb_sectors, pnum, 1636 file); 1637 if (ret < 0) { 1638 *pnum = 0; 1639 return ret; 1640 } 1641 1642 if (ret & BDRV_BLOCK_RAW) { 1643 assert(ret & BDRV_BLOCK_OFFSET_VALID); 1644 return bdrv_get_block_status(bs->file->bs, ret >> BDRV_SECTOR_BITS, 1645 *pnum, pnum, file); 1646 } 1647 1648 if (ret & (BDRV_BLOCK_DATA | BDRV_BLOCK_ZERO)) { 1649 ret |= BDRV_BLOCK_ALLOCATED; 1650 } else { 1651 if (bdrv_unallocated_blocks_are_zero(bs)) { 1652 ret |= BDRV_BLOCK_ZERO; 1653 } else if (bs->backing) { 1654 BlockDriverState *bs2 = bs->backing->bs; 1655 int64_t nb_sectors2 = bdrv_nb_sectors(bs2); 1656 if (nb_sectors2 >= 0 && sector_num >= nb_sectors2) { 1657 ret |= BDRV_BLOCK_ZERO; 1658 } 1659 } 1660 } 1661 1662 if (*file && *file != bs && 1663 (ret & BDRV_BLOCK_DATA) && !(ret & BDRV_BLOCK_ZERO) && 1664 (ret & BDRV_BLOCK_OFFSET_VALID)) { 1665 BlockDriverState *file2; 1666 int file_pnum; 1667 1668 ret2 = bdrv_co_get_block_status(*file, ret >> BDRV_SECTOR_BITS, 1669 *pnum, &file_pnum, &file2); 1670 if (ret2 >= 0) { 1671 /* Ignore errors. This is just providing extra information, it 1672 * is useful but not necessary. 1673 */ 1674 if (!file_pnum) { 1675 /* !file_pnum indicates an offset at or beyond the EOF; it is 1676 * perfectly valid for the format block driver to point to such 1677 * offsets, so catch it and mark everything as zero */ 1678 ret |= BDRV_BLOCK_ZERO; 1679 } else { 1680 /* Limit request to the range reported by the protocol driver */ 1681 *pnum = file_pnum; 1682 ret |= (ret2 & BDRV_BLOCK_ZERO); 1683 } 1684 } 1685 } 1686 1687 return ret; 1688 } 1689 1690 static int64_t coroutine_fn bdrv_co_get_block_status_above(BlockDriverState *bs, 1691 BlockDriverState *base, 1692 int64_t sector_num, 1693 int nb_sectors, 1694 int *pnum, 1695 BlockDriverState **file) 1696 { 1697 BlockDriverState *p; 1698 int64_t ret = 0; 1699 1700 assert(bs != base); 1701 for (p = bs; p != base; p = backing_bs(p)) { 1702 ret = bdrv_co_get_block_status(p, sector_num, nb_sectors, pnum, file); 1703 if (ret < 0 || ret & BDRV_BLOCK_ALLOCATED) { 1704 break; 1705 } 1706 /* [sector_num, pnum] unallocated on this layer, which could be only 1707 * the first part of [sector_num, nb_sectors]. */ 1708 nb_sectors = MIN(nb_sectors, *pnum); 1709 } 1710 return ret; 1711 } 1712 1713 /* Coroutine wrapper for bdrv_get_block_status_above() */ 1714 static void coroutine_fn bdrv_get_block_status_above_co_entry(void *opaque) 1715 { 1716 BdrvCoGetBlockStatusData *data = opaque; 1717 1718 data->ret = bdrv_co_get_block_status_above(data->bs, data->base, 1719 data->sector_num, 1720 data->nb_sectors, 1721 data->pnum, 1722 data->file); 1723 data->done = true; 1724 } 1725 1726 /* 1727 * Synchronous wrapper around bdrv_co_get_block_status_above(). 1728 * 1729 * See bdrv_co_get_block_status_above() for details. 1730 */ 1731 int64_t bdrv_get_block_status_above(BlockDriverState *bs, 1732 BlockDriverState *base, 1733 int64_t sector_num, 1734 int nb_sectors, int *pnum, 1735 BlockDriverState **file) 1736 { 1737 Coroutine *co; 1738 BdrvCoGetBlockStatusData data = { 1739 .bs = bs, 1740 .base = base, 1741 .file = file, 1742 .sector_num = sector_num, 1743 .nb_sectors = nb_sectors, 1744 .pnum = pnum, 1745 .done = false, 1746 }; 1747 1748 if (qemu_in_coroutine()) { 1749 /* Fast-path if already in coroutine context */ 1750 bdrv_get_block_status_above_co_entry(&data); 1751 } else { 1752 AioContext *aio_context = bdrv_get_aio_context(bs); 1753 1754 co = qemu_coroutine_create(bdrv_get_block_status_above_co_entry); 1755 qemu_coroutine_enter(co, &data); 1756 while (!data.done) { 1757 aio_poll(aio_context, true); 1758 } 1759 } 1760 return data.ret; 1761 } 1762 1763 int64_t bdrv_get_block_status(BlockDriverState *bs, 1764 int64_t sector_num, 1765 int nb_sectors, int *pnum, 1766 BlockDriverState **file) 1767 { 1768 return bdrv_get_block_status_above(bs, backing_bs(bs), 1769 sector_num, nb_sectors, pnum, file); 1770 } 1771 1772 int coroutine_fn bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, 1773 int nb_sectors, int *pnum) 1774 { 1775 BlockDriverState *file; 1776 int64_t ret = bdrv_get_block_status(bs, sector_num, nb_sectors, pnum, 1777 &file); 1778 if (ret < 0) { 1779 return ret; 1780 } 1781 return !!(ret & BDRV_BLOCK_ALLOCATED); 1782 } 1783 1784 /* 1785 * Given an image chain: ... -> [BASE] -> [INTER1] -> [INTER2] -> [TOP] 1786 * 1787 * Return true if the given sector is allocated in any image between 1788 * BASE and TOP (inclusive). BASE can be NULL to check if the given 1789 * sector is allocated in any image of the chain. Return false otherwise. 1790 * 1791 * 'pnum' is set to the number of sectors (including and immediately following 1792 * the specified sector) that are known to be in the same 1793 * allocated/unallocated state. 1794 * 1795 */ 1796 int bdrv_is_allocated_above(BlockDriverState *top, 1797 BlockDriverState *base, 1798 int64_t sector_num, 1799 int nb_sectors, int *pnum) 1800 { 1801 BlockDriverState *intermediate; 1802 int ret, n = nb_sectors; 1803 1804 intermediate = top; 1805 while (intermediate && intermediate != base) { 1806 int pnum_inter; 1807 ret = bdrv_is_allocated(intermediate, sector_num, nb_sectors, 1808 &pnum_inter); 1809 if (ret < 0) { 1810 return ret; 1811 } else if (ret) { 1812 *pnum = pnum_inter; 1813 return 1; 1814 } 1815 1816 /* 1817 * [sector_num, nb_sectors] is unallocated on top but intermediate 1818 * might have 1819 * 1820 * [sector_num+x, nr_sectors] allocated. 1821 */ 1822 if (n > pnum_inter && 1823 (intermediate == top || 1824 sector_num + pnum_inter < intermediate->total_sectors)) { 1825 n = pnum_inter; 1826 } 1827 1828 intermediate = backing_bs(intermediate); 1829 } 1830 1831 *pnum = n; 1832 return 0; 1833 } 1834 1835 int bdrv_write_compressed(BlockDriverState *bs, int64_t sector_num, 1836 const uint8_t *buf, int nb_sectors) 1837 { 1838 BlockDriver *drv = bs->drv; 1839 int ret; 1840 1841 if (!drv) { 1842 return -ENOMEDIUM; 1843 } 1844 if (!drv->bdrv_write_compressed) { 1845 return -ENOTSUP; 1846 } 1847 ret = bdrv_check_request(bs, sector_num, nb_sectors); 1848 if (ret < 0) { 1849 return ret; 1850 } 1851 1852 assert(QLIST_EMPTY(&bs->dirty_bitmaps)); 1853 1854 return drv->bdrv_write_compressed(bs, sector_num, buf, nb_sectors); 1855 } 1856 1857 typedef struct BdrvVmstateCo { 1858 BlockDriverState *bs; 1859 QEMUIOVector *qiov; 1860 int64_t pos; 1861 bool is_read; 1862 int ret; 1863 } BdrvVmstateCo; 1864 1865 static int coroutine_fn 1866 bdrv_co_rw_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos, 1867 bool is_read) 1868 { 1869 BlockDriver *drv = bs->drv; 1870 1871 if (!drv) { 1872 return -ENOMEDIUM; 1873 } else if (drv->bdrv_load_vmstate) { 1874 return is_read ? drv->bdrv_load_vmstate(bs, qiov, pos) 1875 : drv->bdrv_save_vmstate(bs, qiov, pos); 1876 } else if (bs->file) { 1877 return bdrv_co_rw_vmstate(bs->file->bs, qiov, pos, is_read); 1878 } 1879 1880 return -ENOTSUP; 1881 } 1882 1883 static void coroutine_fn bdrv_co_rw_vmstate_entry(void *opaque) 1884 { 1885 BdrvVmstateCo *co = opaque; 1886 co->ret = bdrv_co_rw_vmstate(co->bs, co->qiov, co->pos, co->is_read); 1887 } 1888 1889 static inline int 1890 bdrv_rw_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos, 1891 bool is_read) 1892 { 1893 if (qemu_in_coroutine()) { 1894 return bdrv_co_rw_vmstate(bs, qiov, pos, is_read); 1895 } else { 1896 BdrvVmstateCo data = { 1897 .bs = bs, 1898 .qiov = qiov, 1899 .pos = pos, 1900 .is_read = is_read, 1901 .ret = -EINPROGRESS, 1902 }; 1903 Coroutine *co = qemu_coroutine_create(bdrv_co_rw_vmstate_entry); 1904 1905 qemu_coroutine_enter(co, &data); 1906 while (data.ret == -EINPROGRESS) { 1907 aio_poll(bdrv_get_aio_context(bs), true); 1908 } 1909 return data.ret; 1910 } 1911 } 1912 1913 int bdrv_save_vmstate(BlockDriverState *bs, const uint8_t *buf, 1914 int64_t pos, int size) 1915 { 1916 QEMUIOVector qiov; 1917 struct iovec iov = { 1918 .iov_base = (void *) buf, 1919 .iov_len = size, 1920 }; 1921 int ret; 1922 1923 qemu_iovec_init_external(&qiov, &iov, 1); 1924 1925 ret = bdrv_writev_vmstate(bs, &qiov, pos); 1926 if (ret < 0) { 1927 return ret; 1928 } 1929 1930 return size; 1931 } 1932 1933 int bdrv_writev_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos) 1934 { 1935 return bdrv_rw_vmstate(bs, qiov, pos, false); 1936 } 1937 1938 int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf, 1939 int64_t pos, int size) 1940 { 1941 QEMUIOVector qiov; 1942 struct iovec iov = { 1943 .iov_base = buf, 1944 .iov_len = size, 1945 }; 1946 int ret; 1947 1948 qemu_iovec_init_external(&qiov, &iov, 1); 1949 ret = bdrv_readv_vmstate(bs, &qiov, pos); 1950 if (ret < 0) { 1951 return ret; 1952 } 1953 1954 return size; 1955 } 1956 1957 int bdrv_readv_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos) 1958 { 1959 return bdrv_rw_vmstate(bs, qiov, pos, true); 1960 } 1961 1962 /**************************************************************/ 1963 /* async I/Os */ 1964 1965 BlockAIOCB *bdrv_aio_readv(BlockDriverState *bs, int64_t sector_num, 1966 QEMUIOVector *qiov, int nb_sectors, 1967 BlockCompletionFunc *cb, void *opaque) 1968 { 1969 trace_bdrv_aio_readv(bs, sector_num, nb_sectors, opaque); 1970 1971 return bdrv_co_aio_rw_vector(bs, sector_num, qiov, nb_sectors, 0, 1972 cb, opaque, false); 1973 } 1974 1975 BlockAIOCB *bdrv_aio_writev(BlockDriverState *bs, int64_t sector_num, 1976 QEMUIOVector *qiov, int nb_sectors, 1977 BlockCompletionFunc *cb, void *opaque) 1978 { 1979 trace_bdrv_aio_writev(bs, sector_num, nb_sectors, opaque); 1980 1981 return bdrv_co_aio_rw_vector(bs, sector_num, qiov, nb_sectors, 0, 1982 cb, opaque, true); 1983 } 1984 1985 void bdrv_aio_cancel(BlockAIOCB *acb) 1986 { 1987 qemu_aio_ref(acb); 1988 bdrv_aio_cancel_async(acb); 1989 while (acb->refcnt > 1) { 1990 if (acb->aiocb_info->get_aio_context) { 1991 aio_poll(acb->aiocb_info->get_aio_context(acb), true); 1992 } else if (acb->bs) { 1993 aio_poll(bdrv_get_aio_context(acb->bs), true); 1994 } else { 1995 abort(); 1996 } 1997 } 1998 qemu_aio_unref(acb); 1999 } 2000 2001 /* Async version of aio cancel. The caller is not blocked if the acb implements 2002 * cancel_async, otherwise we do nothing and let the request normally complete. 2003 * In either case the completion callback must be called. */ 2004 void bdrv_aio_cancel_async(BlockAIOCB *acb) 2005 { 2006 if (acb->aiocb_info->cancel_async) { 2007 acb->aiocb_info->cancel_async(acb); 2008 } 2009 } 2010 2011 /**************************************************************/ 2012 /* async block device emulation */ 2013 2014 typedef struct BlockRequest { 2015 union { 2016 /* Used during read, write, trim */ 2017 struct { 2018 int64_t sector; 2019 int nb_sectors; 2020 int flags; 2021 QEMUIOVector *qiov; 2022 }; 2023 /* Used during ioctl */ 2024 struct { 2025 int req; 2026 void *buf; 2027 }; 2028 }; 2029 BlockCompletionFunc *cb; 2030 void *opaque; 2031 2032 int error; 2033 } BlockRequest; 2034 2035 typedef struct BlockAIOCBCoroutine { 2036 BlockAIOCB common; 2037 BlockRequest req; 2038 bool is_write; 2039 bool need_bh; 2040 bool *done; 2041 QEMUBH* bh; 2042 } BlockAIOCBCoroutine; 2043 2044 static const AIOCBInfo bdrv_em_co_aiocb_info = { 2045 .aiocb_size = sizeof(BlockAIOCBCoroutine), 2046 }; 2047 2048 static void bdrv_co_complete(BlockAIOCBCoroutine *acb) 2049 { 2050 if (!acb->need_bh) { 2051 acb->common.cb(acb->common.opaque, acb->req.error); 2052 qemu_aio_unref(acb); 2053 } 2054 } 2055 2056 static void bdrv_co_em_bh(void *opaque) 2057 { 2058 BlockAIOCBCoroutine *acb = opaque; 2059 2060 assert(!acb->need_bh); 2061 qemu_bh_delete(acb->bh); 2062 bdrv_co_complete(acb); 2063 } 2064 2065 static void bdrv_co_maybe_schedule_bh(BlockAIOCBCoroutine *acb) 2066 { 2067 acb->need_bh = false; 2068 if (acb->req.error != -EINPROGRESS) { 2069 BlockDriverState *bs = acb->common.bs; 2070 2071 acb->bh = aio_bh_new(bdrv_get_aio_context(bs), bdrv_co_em_bh, acb); 2072 qemu_bh_schedule(acb->bh); 2073 } 2074 } 2075 2076 /* Invoke bdrv_co_do_readv/bdrv_co_do_writev */ 2077 static void coroutine_fn bdrv_co_do_rw(void *opaque) 2078 { 2079 BlockAIOCBCoroutine *acb = opaque; 2080 BlockDriverState *bs = acb->common.bs; 2081 2082 if (!acb->is_write) { 2083 acb->req.error = bdrv_co_do_readv(bs, acb->req.sector, 2084 acb->req.nb_sectors, acb->req.qiov, acb->req.flags); 2085 } else { 2086 acb->req.error = bdrv_co_do_writev(bs, acb->req.sector, 2087 acb->req.nb_sectors, acb->req.qiov, acb->req.flags); 2088 } 2089 2090 bdrv_co_complete(acb); 2091 } 2092 2093 static BlockAIOCB *bdrv_co_aio_rw_vector(BlockDriverState *bs, 2094 int64_t sector_num, 2095 QEMUIOVector *qiov, 2096 int nb_sectors, 2097 BdrvRequestFlags flags, 2098 BlockCompletionFunc *cb, 2099 void *opaque, 2100 bool is_write) 2101 { 2102 Coroutine *co; 2103 BlockAIOCBCoroutine *acb; 2104 2105 acb = qemu_aio_get(&bdrv_em_co_aiocb_info, bs, cb, opaque); 2106 acb->need_bh = true; 2107 acb->req.error = -EINPROGRESS; 2108 acb->req.sector = sector_num; 2109 acb->req.nb_sectors = nb_sectors; 2110 acb->req.qiov = qiov; 2111 acb->req.flags = flags; 2112 acb->is_write = is_write; 2113 2114 co = qemu_coroutine_create(bdrv_co_do_rw); 2115 qemu_coroutine_enter(co, acb); 2116 2117 bdrv_co_maybe_schedule_bh(acb); 2118 return &acb->common; 2119 } 2120 2121 static void coroutine_fn bdrv_aio_flush_co_entry(void *opaque) 2122 { 2123 BlockAIOCBCoroutine *acb = opaque; 2124 BlockDriverState *bs = acb->common.bs; 2125 2126 acb->req.error = bdrv_co_flush(bs); 2127 bdrv_co_complete(acb); 2128 } 2129 2130 BlockAIOCB *bdrv_aio_flush(BlockDriverState *bs, 2131 BlockCompletionFunc *cb, void *opaque) 2132 { 2133 trace_bdrv_aio_flush(bs, opaque); 2134 2135 Coroutine *co; 2136 BlockAIOCBCoroutine *acb; 2137 2138 acb = qemu_aio_get(&bdrv_em_co_aiocb_info, bs, cb, opaque); 2139 acb->need_bh = true; 2140 acb->req.error = -EINPROGRESS; 2141 2142 co = qemu_coroutine_create(bdrv_aio_flush_co_entry); 2143 qemu_coroutine_enter(co, acb); 2144 2145 bdrv_co_maybe_schedule_bh(acb); 2146 return &acb->common; 2147 } 2148 2149 static void coroutine_fn bdrv_aio_discard_co_entry(void *opaque) 2150 { 2151 BlockAIOCBCoroutine *acb = opaque; 2152 BlockDriverState *bs = acb->common.bs; 2153 2154 acb->req.error = bdrv_co_discard(bs, acb->req.sector, acb->req.nb_sectors); 2155 bdrv_co_complete(acb); 2156 } 2157 2158 BlockAIOCB *bdrv_aio_discard(BlockDriverState *bs, 2159 int64_t sector_num, int nb_sectors, 2160 BlockCompletionFunc *cb, void *opaque) 2161 { 2162 Coroutine *co; 2163 BlockAIOCBCoroutine *acb; 2164 2165 trace_bdrv_aio_discard(bs, sector_num, nb_sectors, opaque); 2166 2167 acb = qemu_aio_get(&bdrv_em_co_aiocb_info, bs, cb, opaque); 2168 acb->need_bh = true; 2169 acb->req.error = -EINPROGRESS; 2170 acb->req.sector = sector_num; 2171 acb->req.nb_sectors = nb_sectors; 2172 co = qemu_coroutine_create(bdrv_aio_discard_co_entry); 2173 qemu_coroutine_enter(co, acb); 2174 2175 bdrv_co_maybe_schedule_bh(acb); 2176 return &acb->common; 2177 } 2178 2179 void *qemu_aio_get(const AIOCBInfo *aiocb_info, BlockDriverState *bs, 2180 BlockCompletionFunc *cb, void *opaque) 2181 { 2182 BlockAIOCB *acb; 2183 2184 acb = g_malloc(aiocb_info->aiocb_size); 2185 acb->aiocb_info = aiocb_info; 2186 acb->bs = bs; 2187 acb->cb = cb; 2188 acb->opaque = opaque; 2189 acb->refcnt = 1; 2190 return acb; 2191 } 2192 2193 void qemu_aio_ref(void *p) 2194 { 2195 BlockAIOCB *acb = p; 2196 acb->refcnt++; 2197 } 2198 2199 void qemu_aio_unref(void *p) 2200 { 2201 BlockAIOCB *acb = p; 2202 assert(acb->refcnt > 0); 2203 if (--acb->refcnt == 0) { 2204 g_free(acb); 2205 } 2206 } 2207 2208 /**************************************************************/ 2209 /* Coroutine block device emulation */ 2210 2211 static void coroutine_fn bdrv_flush_co_entry(void *opaque) 2212 { 2213 RwCo *rwco = opaque; 2214 2215 rwco->ret = bdrv_co_flush(rwco->bs); 2216 } 2217 2218 int coroutine_fn bdrv_co_flush(BlockDriverState *bs) 2219 { 2220 int ret; 2221 BdrvTrackedRequest req; 2222 2223 if (!bs || !bdrv_is_inserted(bs) || bdrv_is_read_only(bs) || 2224 bdrv_is_sg(bs)) { 2225 return 0; 2226 } 2227 2228 tracked_request_begin(&req, bs, 0, 0, BDRV_TRACKED_FLUSH); 2229 2230 /* Write back all layers by calling one driver function */ 2231 if (bs->drv->bdrv_co_flush) { 2232 ret = bs->drv->bdrv_co_flush(bs); 2233 goto out; 2234 } 2235 2236 /* Write back cached data to the OS even with cache=unsafe */ 2237 BLKDBG_EVENT(bs->file, BLKDBG_FLUSH_TO_OS); 2238 if (bs->drv->bdrv_co_flush_to_os) { 2239 ret = bs->drv->bdrv_co_flush_to_os(bs); 2240 if (ret < 0) { 2241 goto out; 2242 } 2243 } 2244 2245 /* But don't actually force it to the disk with cache=unsafe */ 2246 if (bs->open_flags & BDRV_O_NO_FLUSH) { 2247 goto flush_parent; 2248 } 2249 2250 BLKDBG_EVENT(bs->file, BLKDBG_FLUSH_TO_DISK); 2251 if (bs->drv->bdrv_co_flush_to_disk) { 2252 ret = bs->drv->bdrv_co_flush_to_disk(bs); 2253 } else if (bs->drv->bdrv_aio_flush) { 2254 BlockAIOCB *acb; 2255 CoroutineIOCompletion co = { 2256 .coroutine = qemu_coroutine_self(), 2257 }; 2258 2259 acb = bs->drv->bdrv_aio_flush(bs, bdrv_co_io_em_complete, &co); 2260 if (acb == NULL) { 2261 ret = -EIO; 2262 } else { 2263 qemu_coroutine_yield(); 2264 ret = co.ret; 2265 } 2266 } else { 2267 /* 2268 * Some block drivers always operate in either writethrough or unsafe 2269 * mode and don't support bdrv_flush therefore. Usually qemu doesn't 2270 * know how the server works (because the behaviour is hardcoded or 2271 * depends on server-side configuration), so we can't ensure that 2272 * everything is safe on disk. Returning an error doesn't work because 2273 * that would break guests even if the server operates in writethrough 2274 * mode. 2275 * 2276 * Let's hope the user knows what he's doing. 2277 */ 2278 ret = 0; 2279 } 2280 if (ret < 0) { 2281 goto out; 2282 } 2283 2284 /* Now flush the underlying protocol. It will also have BDRV_O_NO_FLUSH 2285 * in the case of cache=unsafe, so there are no useless flushes. 2286 */ 2287 flush_parent: 2288 ret = bs->file ? bdrv_co_flush(bs->file->bs) : 0; 2289 out: 2290 tracked_request_end(&req); 2291 return ret; 2292 } 2293 2294 int bdrv_flush(BlockDriverState *bs) 2295 { 2296 Coroutine *co; 2297 RwCo rwco = { 2298 .bs = bs, 2299 .ret = NOT_DONE, 2300 }; 2301 2302 if (qemu_in_coroutine()) { 2303 /* Fast-path if already in coroutine context */ 2304 bdrv_flush_co_entry(&rwco); 2305 } else { 2306 AioContext *aio_context = bdrv_get_aio_context(bs); 2307 2308 co = qemu_coroutine_create(bdrv_flush_co_entry); 2309 qemu_coroutine_enter(co, &rwco); 2310 while (rwco.ret == NOT_DONE) { 2311 aio_poll(aio_context, true); 2312 } 2313 } 2314 2315 return rwco.ret; 2316 } 2317 2318 typedef struct DiscardCo { 2319 BlockDriverState *bs; 2320 int64_t sector_num; 2321 int nb_sectors; 2322 int ret; 2323 } DiscardCo; 2324 static void coroutine_fn bdrv_discard_co_entry(void *opaque) 2325 { 2326 DiscardCo *rwco = opaque; 2327 2328 rwco->ret = bdrv_co_discard(rwco->bs, rwco->sector_num, rwco->nb_sectors); 2329 } 2330 2331 int coroutine_fn bdrv_co_discard(BlockDriverState *bs, int64_t sector_num, 2332 int nb_sectors) 2333 { 2334 BdrvTrackedRequest req; 2335 int max_discard, ret; 2336 2337 if (!bs->drv) { 2338 return -ENOMEDIUM; 2339 } 2340 2341 ret = bdrv_check_request(bs, sector_num, nb_sectors); 2342 if (ret < 0) { 2343 return ret; 2344 } else if (bs->read_only) { 2345 return -EPERM; 2346 } 2347 assert(!(bs->open_flags & BDRV_O_INACTIVE)); 2348 2349 /* Do nothing if disabled. */ 2350 if (!(bs->open_flags & BDRV_O_UNMAP)) { 2351 return 0; 2352 } 2353 2354 if (!bs->drv->bdrv_co_discard && !bs->drv->bdrv_aio_discard) { 2355 return 0; 2356 } 2357 2358 tracked_request_begin(&req, bs, sector_num << BDRV_SECTOR_BITS, 2359 nb_sectors << BDRV_SECTOR_BITS, BDRV_TRACKED_DISCARD); 2360 2361 ret = notifier_with_return_list_notify(&bs->before_write_notifiers, &req); 2362 if (ret < 0) { 2363 goto out; 2364 } 2365 2366 max_discard = MIN_NON_ZERO(bs->bl.max_pdiscard >> BDRV_SECTOR_BITS, 2367 BDRV_REQUEST_MAX_SECTORS); 2368 while (nb_sectors > 0) { 2369 int ret; 2370 int num = nb_sectors; 2371 int discard_alignment = bs->bl.pdiscard_alignment >> BDRV_SECTOR_BITS; 2372 2373 /* align request */ 2374 if (discard_alignment && 2375 num >= discard_alignment && 2376 sector_num % discard_alignment) { 2377 if (num > discard_alignment) { 2378 num = discard_alignment; 2379 } 2380 num -= sector_num % discard_alignment; 2381 } 2382 2383 /* limit request size */ 2384 if (num > max_discard) { 2385 num = max_discard; 2386 } 2387 2388 if (bs->drv->bdrv_co_discard) { 2389 ret = bs->drv->bdrv_co_discard(bs, sector_num, num); 2390 } else { 2391 BlockAIOCB *acb; 2392 CoroutineIOCompletion co = { 2393 .coroutine = qemu_coroutine_self(), 2394 }; 2395 2396 acb = bs->drv->bdrv_aio_discard(bs, sector_num, nb_sectors, 2397 bdrv_co_io_em_complete, &co); 2398 if (acb == NULL) { 2399 ret = -EIO; 2400 goto out; 2401 } else { 2402 qemu_coroutine_yield(); 2403 ret = co.ret; 2404 } 2405 } 2406 if (ret && ret != -ENOTSUP) { 2407 goto out; 2408 } 2409 2410 sector_num += num; 2411 nb_sectors -= num; 2412 } 2413 ret = 0; 2414 out: 2415 bdrv_set_dirty(bs, req.offset >> BDRV_SECTOR_BITS, 2416 req.bytes >> BDRV_SECTOR_BITS); 2417 tracked_request_end(&req); 2418 return ret; 2419 } 2420 2421 int bdrv_discard(BlockDriverState *bs, int64_t sector_num, int nb_sectors) 2422 { 2423 Coroutine *co; 2424 DiscardCo rwco = { 2425 .bs = bs, 2426 .sector_num = sector_num, 2427 .nb_sectors = nb_sectors, 2428 .ret = NOT_DONE, 2429 }; 2430 2431 if (qemu_in_coroutine()) { 2432 /* Fast-path if already in coroutine context */ 2433 bdrv_discard_co_entry(&rwco); 2434 } else { 2435 AioContext *aio_context = bdrv_get_aio_context(bs); 2436 2437 co = qemu_coroutine_create(bdrv_discard_co_entry); 2438 qemu_coroutine_enter(co, &rwco); 2439 while (rwco.ret == NOT_DONE) { 2440 aio_poll(aio_context, true); 2441 } 2442 } 2443 2444 return rwco.ret; 2445 } 2446 2447 static int bdrv_co_do_ioctl(BlockDriverState *bs, int req, void *buf) 2448 { 2449 BlockDriver *drv = bs->drv; 2450 BdrvTrackedRequest tracked_req; 2451 CoroutineIOCompletion co = { 2452 .coroutine = qemu_coroutine_self(), 2453 }; 2454 BlockAIOCB *acb; 2455 2456 tracked_request_begin(&tracked_req, bs, 0, 0, BDRV_TRACKED_IOCTL); 2457 if (!drv || !drv->bdrv_aio_ioctl) { 2458 co.ret = -ENOTSUP; 2459 goto out; 2460 } 2461 2462 acb = drv->bdrv_aio_ioctl(bs, req, buf, bdrv_co_io_em_complete, &co); 2463 if (!acb) { 2464 co.ret = -ENOTSUP; 2465 goto out; 2466 } 2467 qemu_coroutine_yield(); 2468 out: 2469 tracked_request_end(&tracked_req); 2470 return co.ret; 2471 } 2472 2473 typedef struct { 2474 BlockDriverState *bs; 2475 int req; 2476 void *buf; 2477 int ret; 2478 } BdrvIoctlCoData; 2479 2480 static void coroutine_fn bdrv_co_ioctl_entry(void *opaque) 2481 { 2482 BdrvIoctlCoData *data = opaque; 2483 data->ret = bdrv_co_do_ioctl(data->bs, data->req, data->buf); 2484 } 2485 2486 /* needed for generic scsi interface */ 2487 int bdrv_ioctl(BlockDriverState *bs, unsigned long int req, void *buf) 2488 { 2489 BdrvIoctlCoData data = { 2490 .bs = bs, 2491 .req = req, 2492 .buf = buf, 2493 .ret = -EINPROGRESS, 2494 }; 2495 2496 if (qemu_in_coroutine()) { 2497 /* Fast-path if already in coroutine context */ 2498 bdrv_co_ioctl_entry(&data); 2499 } else { 2500 Coroutine *co = qemu_coroutine_create(bdrv_co_ioctl_entry); 2501 2502 qemu_coroutine_enter(co, &data); 2503 while (data.ret == -EINPROGRESS) { 2504 aio_poll(bdrv_get_aio_context(bs), true); 2505 } 2506 } 2507 return data.ret; 2508 } 2509 2510 static void coroutine_fn bdrv_co_aio_ioctl_entry(void *opaque) 2511 { 2512 BlockAIOCBCoroutine *acb = opaque; 2513 acb->req.error = bdrv_co_do_ioctl(acb->common.bs, 2514 acb->req.req, acb->req.buf); 2515 bdrv_co_complete(acb); 2516 } 2517 2518 BlockAIOCB *bdrv_aio_ioctl(BlockDriverState *bs, 2519 unsigned long int req, void *buf, 2520 BlockCompletionFunc *cb, void *opaque) 2521 { 2522 BlockAIOCBCoroutine *acb = qemu_aio_get(&bdrv_em_co_aiocb_info, 2523 bs, cb, opaque); 2524 Coroutine *co; 2525 2526 acb->need_bh = true; 2527 acb->req.error = -EINPROGRESS; 2528 acb->req.req = req; 2529 acb->req.buf = buf; 2530 co = qemu_coroutine_create(bdrv_co_aio_ioctl_entry); 2531 qemu_coroutine_enter(co, acb); 2532 2533 bdrv_co_maybe_schedule_bh(acb); 2534 return &acb->common; 2535 } 2536 2537 void *qemu_blockalign(BlockDriverState *bs, size_t size) 2538 { 2539 return qemu_memalign(bdrv_opt_mem_align(bs), size); 2540 } 2541 2542 void *qemu_blockalign0(BlockDriverState *bs, size_t size) 2543 { 2544 return memset(qemu_blockalign(bs, size), 0, size); 2545 } 2546 2547 void *qemu_try_blockalign(BlockDriverState *bs, size_t size) 2548 { 2549 size_t align = bdrv_opt_mem_align(bs); 2550 2551 /* Ensure that NULL is never returned on success */ 2552 assert(align > 0); 2553 if (size == 0) { 2554 size = align; 2555 } 2556 2557 return qemu_try_memalign(align, size); 2558 } 2559 2560 void *qemu_try_blockalign0(BlockDriverState *bs, size_t size) 2561 { 2562 void *mem = qemu_try_blockalign(bs, size); 2563 2564 if (mem) { 2565 memset(mem, 0, size); 2566 } 2567 2568 return mem; 2569 } 2570 2571 /* 2572 * Check if all memory in this vector is sector aligned. 2573 */ 2574 bool bdrv_qiov_is_aligned(BlockDriverState *bs, QEMUIOVector *qiov) 2575 { 2576 int i; 2577 size_t alignment = bdrv_min_mem_align(bs); 2578 2579 for (i = 0; i < qiov->niov; i++) { 2580 if ((uintptr_t) qiov->iov[i].iov_base % alignment) { 2581 return false; 2582 } 2583 if (qiov->iov[i].iov_len % alignment) { 2584 return false; 2585 } 2586 } 2587 2588 return true; 2589 } 2590 2591 void bdrv_add_before_write_notifier(BlockDriverState *bs, 2592 NotifierWithReturn *notifier) 2593 { 2594 notifier_with_return_list_add(&bs->before_write_notifiers, notifier); 2595 } 2596 2597 void bdrv_io_plug(BlockDriverState *bs) 2598 { 2599 BdrvChild *child; 2600 2601 QLIST_FOREACH(child, &bs->children, next) { 2602 bdrv_io_plug(child->bs); 2603 } 2604 2605 if (bs->io_plugged++ == 0 && bs->io_plug_disabled == 0) { 2606 BlockDriver *drv = bs->drv; 2607 if (drv && drv->bdrv_io_plug) { 2608 drv->bdrv_io_plug(bs); 2609 } 2610 } 2611 } 2612 2613 void bdrv_io_unplug(BlockDriverState *bs) 2614 { 2615 BdrvChild *child; 2616 2617 assert(bs->io_plugged); 2618 if (--bs->io_plugged == 0 && bs->io_plug_disabled == 0) { 2619 BlockDriver *drv = bs->drv; 2620 if (drv && drv->bdrv_io_unplug) { 2621 drv->bdrv_io_unplug(bs); 2622 } 2623 } 2624 2625 QLIST_FOREACH(child, &bs->children, next) { 2626 bdrv_io_unplug(child->bs); 2627 } 2628 } 2629 2630 void bdrv_io_unplugged_begin(BlockDriverState *bs) 2631 { 2632 BdrvChild *child; 2633 2634 if (bs->io_plug_disabled++ == 0 && bs->io_plugged > 0) { 2635 BlockDriver *drv = bs->drv; 2636 if (drv && drv->bdrv_io_unplug) { 2637 drv->bdrv_io_unplug(bs); 2638 } 2639 } 2640 2641 QLIST_FOREACH(child, &bs->children, next) { 2642 bdrv_io_unplugged_begin(child->bs); 2643 } 2644 } 2645 2646 void bdrv_io_unplugged_end(BlockDriverState *bs) 2647 { 2648 BdrvChild *child; 2649 2650 assert(bs->io_plug_disabled); 2651 QLIST_FOREACH(child, &bs->children, next) { 2652 bdrv_io_unplugged_end(child->bs); 2653 } 2654 2655 if (--bs->io_plug_disabled == 0 && bs->io_plugged > 0) { 2656 BlockDriver *drv = bs->drv; 2657 if (drv && drv->bdrv_io_plug) { 2658 drv->bdrv_io_plug(bs); 2659 } 2660 } 2661 } 2662