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(BdrvChild *child, 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); 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 BdrvChild *child; 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->child, rwco->offset, 570 rwco->qiov->size, rwco->qiov, 571 rwco->flags); 572 } else { 573 rwco->ret = bdrv_co_pwritev(rwco->child, 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(BdrvChild *child, int64_t offset, 583 QEMUIOVector *qiov, bool is_write, 584 BdrvRequestFlags flags) 585 { 586 Coroutine *co; 587 RwCo rwco = { 588 .child = child, 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(child->bs); 601 602 co = qemu_coroutine_create(bdrv_rw_co_entry, &rwco); 603 qemu_coroutine_enter(co); 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(BdrvChild *child, 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(child, 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(BdrvChild *child, int64_t sector_num, 634 uint8_t *buf, int nb_sectors) 635 { 636 return bdrv_rw_co(child, 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(BdrvChild *child, int64_t sector_num, 646 const uint8_t *buf, int nb_sectors) 647 { 648 return bdrv_rw_co(child, sector_num, (uint8_t *)buf, nb_sectors, true, 0); 649 } 650 651 int bdrv_pwrite_zeroes(BdrvChild *child, 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(child, 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(BdrvChild *child, BdrvRequestFlags flags) 675 { 676 int64_t target_sectors, ret, nb_sectors, sector_num = 0; 677 BlockDriverState *bs = child->bs; 678 BlockDriverState *file; 679 int n; 680 681 target_sectors = bdrv_nb_sectors(bs); 682 if (target_sectors < 0) { 683 return target_sectors; 684 } 685 686 for (;;) { 687 nb_sectors = MIN(target_sectors - sector_num, BDRV_REQUEST_MAX_SECTORS); 688 if (nb_sectors <= 0) { 689 return 0; 690 } 691 ret = bdrv_get_block_status(bs, sector_num, nb_sectors, &n, &file); 692 if (ret < 0) { 693 error_report("error getting block status at sector %" PRId64 ": %s", 694 sector_num, strerror(-ret)); 695 return ret; 696 } 697 if (ret & BDRV_BLOCK_ZERO) { 698 sector_num += n; 699 continue; 700 } 701 ret = bdrv_pwrite_zeroes(child, sector_num << BDRV_SECTOR_BITS, 702 n << BDRV_SECTOR_BITS, flags); 703 if (ret < 0) { 704 error_report("error writing zeroes at sector %" PRId64 ": %s", 705 sector_num, strerror(-ret)); 706 return ret; 707 } 708 sector_num += n; 709 } 710 } 711 712 int bdrv_preadv(BdrvChild *child, int64_t offset, QEMUIOVector *qiov) 713 { 714 int ret; 715 716 ret = bdrv_prwv_co(child, offset, qiov, false, 0); 717 if (ret < 0) { 718 return ret; 719 } 720 721 return qiov->size; 722 } 723 724 int bdrv_pread(BdrvChild *child, int64_t offset, void *buf, int bytes) 725 { 726 QEMUIOVector qiov; 727 struct iovec iov = { 728 .iov_base = (void *)buf, 729 .iov_len = bytes, 730 }; 731 732 if (bytes < 0) { 733 return -EINVAL; 734 } 735 736 qemu_iovec_init_external(&qiov, &iov, 1); 737 return bdrv_preadv(child, offset, &qiov); 738 } 739 740 int bdrv_pwritev(BdrvChild *child, int64_t offset, QEMUIOVector *qiov) 741 { 742 int ret; 743 744 ret = bdrv_prwv_co(child, offset, qiov, true, 0); 745 if (ret < 0) { 746 return ret; 747 } 748 749 return qiov->size; 750 } 751 752 int bdrv_pwrite(BdrvChild *child, int64_t offset, 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(child, 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(BdrvChild *child, int64_t offset, 775 const void *buf, int count) 776 { 777 int ret; 778 779 ret = bdrv_pwrite(child, offset, buf, count); 780 if (ret < 0) { 781 return ret; 782 } 783 784 ret = bdrv_flush(child->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); 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(BdrvChild *child, 1065 int64_t offset, unsigned int bytes, QEMUIOVector *qiov, 1066 BdrvRequestFlags flags) 1067 { 1068 BlockDriverState *bs = child->bs; 1069 BlockDriver *drv = bs->drv; 1070 BdrvTrackedRequest req; 1071 1072 uint64_t align = bs->bl.request_alignment; 1073 uint8_t *head_buf = NULL; 1074 uint8_t *tail_buf = NULL; 1075 QEMUIOVector local_qiov; 1076 bool use_local_qiov = false; 1077 int ret; 1078 1079 if (!drv) { 1080 return -ENOMEDIUM; 1081 } 1082 1083 ret = bdrv_check_byte_request(bs, offset, bytes); 1084 if (ret < 0) { 1085 return ret; 1086 } 1087 1088 /* Don't do copy-on-read if we read data before write operation */ 1089 if (bs->copy_on_read && !(flags & BDRV_REQ_NO_SERIALISING)) { 1090 flags |= BDRV_REQ_COPY_ON_READ; 1091 } 1092 1093 /* Align read if necessary by padding qiov */ 1094 if (offset & (align - 1)) { 1095 head_buf = qemu_blockalign(bs, align); 1096 qemu_iovec_init(&local_qiov, qiov->niov + 2); 1097 qemu_iovec_add(&local_qiov, head_buf, offset & (align - 1)); 1098 qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size); 1099 use_local_qiov = true; 1100 1101 bytes += offset & (align - 1); 1102 offset = offset & ~(align - 1); 1103 } 1104 1105 if ((offset + bytes) & (align - 1)) { 1106 if (!use_local_qiov) { 1107 qemu_iovec_init(&local_qiov, qiov->niov + 1); 1108 qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size); 1109 use_local_qiov = true; 1110 } 1111 tail_buf = qemu_blockalign(bs, align); 1112 qemu_iovec_add(&local_qiov, tail_buf, 1113 align - ((offset + bytes) & (align - 1))); 1114 1115 bytes = ROUND_UP(bytes, align); 1116 } 1117 1118 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_READ); 1119 ret = bdrv_aligned_preadv(bs, &req, offset, bytes, align, 1120 use_local_qiov ? &local_qiov : qiov, 1121 flags); 1122 tracked_request_end(&req); 1123 1124 if (use_local_qiov) { 1125 qemu_iovec_destroy(&local_qiov); 1126 qemu_vfree(head_buf); 1127 qemu_vfree(tail_buf); 1128 } 1129 1130 return ret; 1131 } 1132 1133 static int coroutine_fn bdrv_co_do_readv(BdrvChild *child, 1134 int64_t sector_num, int nb_sectors, QEMUIOVector *qiov, 1135 BdrvRequestFlags flags) 1136 { 1137 if (nb_sectors < 0 || nb_sectors > BDRV_REQUEST_MAX_SECTORS) { 1138 return -EINVAL; 1139 } 1140 1141 return bdrv_co_preadv(child, sector_num << BDRV_SECTOR_BITS, 1142 nb_sectors << BDRV_SECTOR_BITS, qiov, flags); 1143 } 1144 1145 int coroutine_fn bdrv_co_readv(BdrvChild *child, int64_t sector_num, 1146 int nb_sectors, QEMUIOVector *qiov) 1147 { 1148 trace_bdrv_co_readv(child->bs, sector_num, nb_sectors); 1149 1150 return bdrv_co_do_readv(child, sector_num, nb_sectors, qiov, 0); 1151 } 1152 1153 /* Maximum buffer for write zeroes fallback, in bytes */ 1154 #define MAX_WRITE_ZEROES_BOUNCE_BUFFER (32768 << BDRV_SECTOR_BITS) 1155 1156 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs, 1157 int64_t offset, int count, BdrvRequestFlags flags) 1158 { 1159 BlockDriver *drv = bs->drv; 1160 QEMUIOVector qiov; 1161 struct iovec iov = {0}; 1162 int ret = 0; 1163 bool need_flush = false; 1164 int head = 0; 1165 int tail = 0; 1166 1167 int max_write_zeroes = MIN_NON_ZERO(bs->bl.max_pwrite_zeroes, INT_MAX); 1168 int alignment = MAX(bs->bl.pwrite_zeroes_alignment, 1169 bs->bl.request_alignment); 1170 1171 assert(is_power_of_2(alignment)); 1172 head = offset & (alignment - 1); 1173 tail = (offset + count) & (alignment - 1); 1174 max_write_zeroes &= ~(alignment - 1); 1175 1176 while (count > 0 && !ret) { 1177 int num = count; 1178 1179 /* Align request. Block drivers can expect the "bulk" of the request 1180 * to be aligned, and that unaligned requests do not cross cluster 1181 * boundaries. 1182 */ 1183 if (head) { 1184 /* Make a small request up to the first aligned sector. */ 1185 num = MIN(count, alignment - head); 1186 head = 0; 1187 } else if (tail && num > alignment) { 1188 /* Shorten the request to the last aligned sector. */ 1189 num -= tail; 1190 } 1191 1192 /* limit request size */ 1193 if (num > max_write_zeroes) { 1194 num = max_write_zeroes; 1195 } 1196 1197 ret = -ENOTSUP; 1198 /* First try the efficient write zeroes operation */ 1199 if (drv->bdrv_co_pwrite_zeroes) { 1200 ret = drv->bdrv_co_pwrite_zeroes(bs, offset, num, 1201 flags & bs->supported_zero_flags); 1202 if (ret != -ENOTSUP && (flags & BDRV_REQ_FUA) && 1203 !(bs->supported_zero_flags & BDRV_REQ_FUA)) { 1204 need_flush = true; 1205 } 1206 } else { 1207 assert(!bs->supported_zero_flags); 1208 } 1209 1210 if (ret == -ENOTSUP) { 1211 /* Fall back to bounce buffer if write zeroes is unsupported */ 1212 int max_transfer = MIN_NON_ZERO(bs->bl.max_transfer, 1213 MAX_WRITE_ZEROES_BOUNCE_BUFFER); 1214 BdrvRequestFlags write_flags = flags & ~BDRV_REQ_ZERO_WRITE; 1215 1216 if ((flags & BDRV_REQ_FUA) && 1217 !(bs->supported_write_flags & BDRV_REQ_FUA)) { 1218 /* No need for bdrv_driver_pwrite() to do a fallback 1219 * flush on each chunk; use just one at the end */ 1220 write_flags &= ~BDRV_REQ_FUA; 1221 need_flush = true; 1222 } 1223 num = MIN(num, max_transfer); 1224 iov.iov_len = num; 1225 if (iov.iov_base == NULL) { 1226 iov.iov_base = qemu_try_blockalign(bs, num); 1227 if (iov.iov_base == NULL) { 1228 ret = -ENOMEM; 1229 goto fail; 1230 } 1231 memset(iov.iov_base, 0, num); 1232 } 1233 qemu_iovec_init_external(&qiov, &iov, 1); 1234 1235 ret = bdrv_driver_pwritev(bs, offset, num, &qiov, write_flags); 1236 1237 /* Keep bounce buffer around if it is big enough for all 1238 * all future requests. 1239 */ 1240 if (num < max_transfer) { 1241 qemu_vfree(iov.iov_base); 1242 iov.iov_base = NULL; 1243 } 1244 } 1245 1246 offset += num; 1247 count -= num; 1248 } 1249 1250 fail: 1251 if (ret == 0 && need_flush) { 1252 ret = bdrv_co_flush(bs); 1253 } 1254 qemu_vfree(iov.iov_base); 1255 return ret; 1256 } 1257 1258 /* 1259 * Forwards an already correctly aligned write request to the BlockDriver. 1260 */ 1261 static int coroutine_fn bdrv_aligned_pwritev(BlockDriverState *bs, 1262 BdrvTrackedRequest *req, int64_t offset, unsigned int bytes, 1263 int64_t align, QEMUIOVector *qiov, int flags) 1264 { 1265 BlockDriver *drv = bs->drv; 1266 bool waited; 1267 int ret; 1268 1269 int64_t start_sector = offset >> BDRV_SECTOR_BITS; 1270 int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE); 1271 1272 assert(is_power_of_2(align)); 1273 assert((offset & (align - 1)) == 0); 1274 assert((bytes & (align - 1)) == 0); 1275 assert(!qiov || bytes == qiov->size); 1276 assert((bs->open_flags & BDRV_O_NO_IO) == 0); 1277 assert(!(flags & ~BDRV_REQ_MASK)); 1278 1279 waited = wait_serialising_requests(req); 1280 assert(!waited || !req->serialising); 1281 assert(req->overlap_offset <= offset); 1282 assert(offset + bytes <= req->overlap_offset + req->overlap_bytes); 1283 1284 ret = notifier_with_return_list_notify(&bs->before_write_notifiers, req); 1285 1286 if (!ret && bs->detect_zeroes != BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF && 1287 !(flags & BDRV_REQ_ZERO_WRITE) && drv->bdrv_co_pwrite_zeroes && 1288 qemu_iovec_is_zero(qiov)) { 1289 flags |= BDRV_REQ_ZERO_WRITE; 1290 if (bs->detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP) { 1291 flags |= BDRV_REQ_MAY_UNMAP; 1292 } 1293 } 1294 1295 if (ret < 0) { 1296 /* Do nothing, write notifier decided to fail this request */ 1297 } else if (flags & BDRV_REQ_ZERO_WRITE) { 1298 bdrv_debug_event(bs, BLKDBG_PWRITEV_ZERO); 1299 ret = bdrv_co_do_pwrite_zeroes(bs, offset, bytes, flags); 1300 } else { 1301 bdrv_debug_event(bs, BLKDBG_PWRITEV); 1302 ret = bdrv_driver_pwritev(bs, offset, bytes, qiov, flags); 1303 } 1304 bdrv_debug_event(bs, BLKDBG_PWRITEV_DONE); 1305 1306 bdrv_set_dirty(bs, start_sector, end_sector - start_sector); 1307 1308 if (bs->wr_highest_offset < offset + bytes) { 1309 bs->wr_highest_offset = offset + bytes; 1310 } 1311 1312 if (ret >= 0) { 1313 bs->total_sectors = MAX(bs->total_sectors, end_sector); 1314 } 1315 1316 return ret; 1317 } 1318 1319 static int coroutine_fn bdrv_co_do_zero_pwritev(BlockDriverState *bs, 1320 int64_t offset, 1321 unsigned int bytes, 1322 BdrvRequestFlags flags, 1323 BdrvTrackedRequest *req) 1324 { 1325 uint8_t *buf = NULL; 1326 QEMUIOVector local_qiov; 1327 struct iovec iov; 1328 uint64_t align = bs->bl.request_alignment; 1329 unsigned int head_padding_bytes, tail_padding_bytes; 1330 int ret = 0; 1331 1332 head_padding_bytes = offset & (align - 1); 1333 tail_padding_bytes = align - ((offset + bytes) & (align - 1)); 1334 1335 1336 assert(flags & BDRV_REQ_ZERO_WRITE); 1337 if (head_padding_bytes || tail_padding_bytes) { 1338 buf = qemu_blockalign(bs, align); 1339 iov = (struct iovec) { 1340 .iov_base = buf, 1341 .iov_len = align, 1342 }; 1343 qemu_iovec_init_external(&local_qiov, &iov, 1); 1344 } 1345 if (head_padding_bytes) { 1346 uint64_t zero_bytes = MIN(bytes, align - head_padding_bytes); 1347 1348 /* RMW the unaligned part before head. */ 1349 mark_request_serialising(req, align); 1350 wait_serialising_requests(req); 1351 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD); 1352 ret = bdrv_aligned_preadv(bs, req, offset & ~(align - 1), align, 1353 align, &local_qiov, 0); 1354 if (ret < 0) { 1355 goto fail; 1356 } 1357 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD); 1358 1359 memset(buf + head_padding_bytes, 0, zero_bytes); 1360 ret = bdrv_aligned_pwritev(bs, req, offset & ~(align - 1), align, 1361 align, &local_qiov, 1362 flags & ~BDRV_REQ_ZERO_WRITE); 1363 if (ret < 0) { 1364 goto fail; 1365 } 1366 offset += zero_bytes; 1367 bytes -= zero_bytes; 1368 } 1369 1370 assert(!bytes || (offset & (align - 1)) == 0); 1371 if (bytes >= align) { 1372 /* Write the aligned part in the middle. */ 1373 uint64_t aligned_bytes = bytes & ~(align - 1); 1374 ret = bdrv_aligned_pwritev(bs, req, offset, aligned_bytes, align, 1375 NULL, flags); 1376 if (ret < 0) { 1377 goto fail; 1378 } 1379 bytes -= aligned_bytes; 1380 offset += aligned_bytes; 1381 } 1382 1383 assert(!bytes || (offset & (align - 1)) == 0); 1384 if (bytes) { 1385 assert(align == tail_padding_bytes + bytes); 1386 /* RMW the unaligned part after tail. */ 1387 mark_request_serialising(req, align); 1388 wait_serialising_requests(req); 1389 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL); 1390 ret = bdrv_aligned_preadv(bs, req, offset, align, 1391 align, &local_qiov, 0); 1392 if (ret < 0) { 1393 goto fail; 1394 } 1395 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL); 1396 1397 memset(buf, 0, bytes); 1398 ret = bdrv_aligned_pwritev(bs, req, offset, align, align, 1399 &local_qiov, flags & ~BDRV_REQ_ZERO_WRITE); 1400 } 1401 fail: 1402 qemu_vfree(buf); 1403 return ret; 1404 1405 } 1406 1407 /* 1408 * Handle a write request in coroutine context 1409 */ 1410 int coroutine_fn bdrv_co_pwritev(BdrvChild *child, 1411 int64_t offset, unsigned int bytes, QEMUIOVector *qiov, 1412 BdrvRequestFlags flags) 1413 { 1414 BlockDriverState *bs = child->bs; 1415 BdrvTrackedRequest req; 1416 uint64_t align = bs->bl.request_alignment; 1417 uint8_t *head_buf = NULL; 1418 uint8_t *tail_buf = NULL; 1419 QEMUIOVector local_qiov; 1420 bool use_local_qiov = false; 1421 int ret; 1422 1423 if (!bs->drv) { 1424 return -ENOMEDIUM; 1425 } 1426 if (bs->read_only) { 1427 return -EPERM; 1428 } 1429 assert(!(bs->open_flags & BDRV_O_INACTIVE)); 1430 1431 ret = bdrv_check_byte_request(bs, offset, bytes); 1432 if (ret < 0) { 1433 return ret; 1434 } 1435 1436 /* 1437 * Align write if necessary by performing a read-modify-write cycle. 1438 * Pad qiov with the read parts and be sure to have a tracked request not 1439 * only for bdrv_aligned_pwritev, but also for the reads of the RMW cycle. 1440 */ 1441 tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_WRITE); 1442 1443 if (!qiov) { 1444 ret = bdrv_co_do_zero_pwritev(bs, offset, bytes, flags, &req); 1445 goto out; 1446 } 1447 1448 if (offset & (align - 1)) { 1449 QEMUIOVector head_qiov; 1450 struct iovec head_iov; 1451 1452 mark_request_serialising(&req, align); 1453 wait_serialising_requests(&req); 1454 1455 head_buf = qemu_blockalign(bs, align); 1456 head_iov = (struct iovec) { 1457 .iov_base = head_buf, 1458 .iov_len = align, 1459 }; 1460 qemu_iovec_init_external(&head_qiov, &head_iov, 1); 1461 1462 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD); 1463 ret = bdrv_aligned_preadv(bs, &req, offset & ~(align - 1), align, 1464 align, &head_qiov, 0); 1465 if (ret < 0) { 1466 goto fail; 1467 } 1468 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD); 1469 1470 qemu_iovec_init(&local_qiov, qiov->niov + 2); 1471 qemu_iovec_add(&local_qiov, head_buf, offset & (align - 1)); 1472 qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size); 1473 use_local_qiov = true; 1474 1475 bytes += offset & (align - 1); 1476 offset = offset & ~(align - 1); 1477 1478 /* We have read the tail already if the request is smaller 1479 * than one aligned block. 1480 */ 1481 if (bytes < align) { 1482 qemu_iovec_add(&local_qiov, head_buf + bytes, align - bytes); 1483 bytes = align; 1484 } 1485 } 1486 1487 if ((offset + bytes) & (align - 1)) { 1488 QEMUIOVector tail_qiov; 1489 struct iovec tail_iov; 1490 size_t tail_bytes; 1491 bool waited; 1492 1493 mark_request_serialising(&req, align); 1494 waited = wait_serialising_requests(&req); 1495 assert(!waited || !use_local_qiov); 1496 1497 tail_buf = qemu_blockalign(bs, align); 1498 tail_iov = (struct iovec) { 1499 .iov_base = tail_buf, 1500 .iov_len = align, 1501 }; 1502 qemu_iovec_init_external(&tail_qiov, &tail_iov, 1); 1503 1504 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL); 1505 ret = bdrv_aligned_preadv(bs, &req, (offset + bytes) & ~(align - 1), align, 1506 align, &tail_qiov, 0); 1507 if (ret < 0) { 1508 goto fail; 1509 } 1510 bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL); 1511 1512 if (!use_local_qiov) { 1513 qemu_iovec_init(&local_qiov, qiov->niov + 1); 1514 qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size); 1515 use_local_qiov = true; 1516 } 1517 1518 tail_bytes = (offset + bytes) & (align - 1); 1519 qemu_iovec_add(&local_qiov, tail_buf + tail_bytes, align - tail_bytes); 1520 1521 bytes = ROUND_UP(bytes, align); 1522 } 1523 1524 ret = bdrv_aligned_pwritev(bs, &req, offset, bytes, align, 1525 use_local_qiov ? &local_qiov : qiov, 1526 flags); 1527 1528 fail: 1529 1530 if (use_local_qiov) { 1531 qemu_iovec_destroy(&local_qiov); 1532 } 1533 qemu_vfree(head_buf); 1534 qemu_vfree(tail_buf); 1535 out: 1536 tracked_request_end(&req); 1537 return ret; 1538 } 1539 1540 static int coroutine_fn bdrv_co_do_writev(BdrvChild *child, 1541 int64_t sector_num, int nb_sectors, QEMUIOVector *qiov, 1542 BdrvRequestFlags flags) 1543 { 1544 if (nb_sectors < 0 || nb_sectors > BDRV_REQUEST_MAX_SECTORS) { 1545 return -EINVAL; 1546 } 1547 1548 return bdrv_co_pwritev(child, sector_num << BDRV_SECTOR_BITS, 1549 nb_sectors << BDRV_SECTOR_BITS, qiov, flags); 1550 } 1551 1552 int coroutine_fn bdrv_co_writev(BdrvChild *child, int64_t sector_num, 1553 int nb_sectors, QEMUIOVector *qiov) 1554 { 1555 trace_bdrv_co_writev(child->bs, sector_num, nb_sectors); 1556 1557 return bdrv_co_do_writev(child, sector_num, nb_sectors, qiov, 0); 1558 } 1559 1560 int coroutine_fn bdrv_co_pwrite_zeroes(BdrvChild *child, int64_t offset, 1561 int count, BdrvRequestFlags flags) 1562 { 1563 trace_bdrv_co_pwrite_zeroes(child->bs, offset, count, flags); 1564 1565 if (!(child->bs->open_flags & BDRV_O_UNMAP)) { 1566 flags &= ~BDRV_REQ_MAY_UNMAP; 1567 } 1568 1569 return bdrv_co_pwritev(child, offset, count, NULL, 1570 BDRV_REQ_ZERO_WRITE | flags); 1571 } 1572 1573 typedef struct BdrvCoGetBlockStatusData { 1574 BlockDriverState *bs; 1575 BlockDriverState *base; 1576 BlockDriverState **file; 1577 int64_t sector_num; 1578 int nb_sectors; 1579 int *pnum; 1580 int64_t ret; 1581 bool done; 1582 } BdrvCoGetBlockStatusData; 1583 1584 /* 1585 * Returns the allocation status of the specified sectors. 1586 * Drivers not implementing the functionality are assumed to not support 1587 * backing files, hence all their sectors are reported as allocated. 1588 * 1589 * If 'sector_num' is beyond the end of the disk image the return value is 0 1590 * and 'pnum' is set to 0. 1591 * 1592 * 'pnum' is set to the number of sectors (including and immediately following 1593 * the specified sector) that are known to be in the same 1594 * allocated/unallocated state. 1595 * 1596 * 'nb_sectors' is the max value 'pnum' should be set to. If nb_sectors goes 1597 * beyond the end of the disk image it will be clamped. 1598 * 1599 * If returned value is positive and BDRV_BLOCK_OFFSET_VALID bit is set, 'file' 1600 * points to the BDS which the sector range is allocated in. 1601 */ 1602 static int64_t coroutine_fn bdrv_co_get_block_status(BlockDriverState *bs, 1603 int64_t sector_num, 1604 int nb_sectors, int *pnum, 1605 BlockDriverState **file) 1606 { 1607 int64_t total_sectors; 1608 int64_t n; 1609 int64_t ret, ret2; 1610 1611 total_sectors = bdrv_nb_sectors(bs); 1612 if (total_sectors < 0) { 1613 return total_sectors; 1614 } 1615 1616 if (sector_num >= total_sectors) { 1617 *pnum = 0; 1618 return 0; 1619 } 1620 1621 n = total_sectors - sector_num; 1622 if (n < nb_sectors) { 1623 nb_sectors = n; 1624 } 1625 1626 if (!bs->drv->bdrv_co_get_block_status) { 1627 *pnum = nb_sectors; 1628 ret = BDRV_BLOCK_DATA | BDRV_BLOCK_ALLOCATED; 1629 if (bs->drv->protocol_name) { 1630 ret |= BDRV_BLOCK_OFFSET_VALID | (sector_num * BDRV_SECTOR_SIZE); 1631 } 1632 return ret; 1633 } 1634 1635 *file = NULL; 1636 ret = bs->drv->bdrv_co_get_block_status(bs, sector_num, nb_sectors, pnum, 1637 file); 1638 if (ret < 0) { 1639 *pnum = 0; 1640 return ret; 1641 } 1642 1643 if (ret & BDRV_BLOCK_RAW) { 1644 assert(ret & BDRV_BLOCK_OFFSET_VALID); 1645 return bdrv_get_block_status(bs->file->bs, ret >> BDRV_SECTOR_BITS, 1646 *pnum, pnum, file); 1647 } 1648 1649 if (ret & (BDRV_BLOCK_DATA | BDRV_BLOCK_ZERO)) { 1650 ret |= BDRV_BLOCK_ALLOCATED; 1651 } else { 1652 if (bdrv_unallocated_blocks_are_zero(bs)) { 1653 ret |= BDRV_BLOCK_ZERO; 1654 } else if (bs->backing) { 1655 BlockDriverState *bs2 = bs->backing->bs; 1656 int64_t nb_sectors2 = bdrv_nb_sectors(bs2); 1657 if (nb_sectors2 >= 0 && sector_num >= nb_sectors2) { 1658 ret |= BDRV_BLOCK_ZERO; 1659 } 1660 } 1661 } 1662 1663 if (*file && *file != bs && 1664 (ret & BDRV_BLOCK_DATA) && !(ret & BDRV_BLOCK_ZERO) && 1665 (ret & BDRV_BLOCK_OFFSET_VALID)) { 1666 BlockDriverState *file2; 1667 int file_pnum; 1668 1669 ret2 = bdrv_co_get_block_status(*file, ret >> BDRV_SECTOR_BITS, 1670 *pnum, &file_pnum, &file2); 1671 if (ret2 >= 0) { 1672 /* Ignore errors. This is just providing extra information, it 1673 * is useful but not necessary. 1674 */ 1675 if (!file_pnum) { 1676 /* !file_pnum indicates an offset at or beyond the EOF; it is 1677 * perfectly valid for the format block driver to point to such 1678 * offsets, so catch it and mark everything as zero */ 1679 ret |= BDRV_BLOCK_ZERO; 1680 } else { 1681 /* Limit request to the range reported by the protocol driver */ 1682 *pnum = file_pnum; 1683 ret |= (ret2 & BDRV_BLOCK_ZERO); 1684 } 1685 } 1686 } 1687 1688 return ret; 1689 } 1690 1691 static int64_t coroutine_fn bdrv_co_get_block_status_above(BlockDriverState *bs, 1692 BlockDriverState *base, 1693 int64_t sector_num, 1694 int nb_sectors, 1695 int *pnum, 1696 BlockDriverState **file) 1697 { 1698 BlockDriverState *p; 1699 int64_t ret = 0; 1700 1701 assert(bs != base); 1702 for (p = bs; p != base; p = backing_bs(p)) { 1703 ret = bdrv_co_get_block_status(p, sector_num, nb_sectors, pnum, file); 1704 if (ret < 0 || ret & BDRV_BLOCK_ALLOCATED) { 1705 break; 1706 } 1707 /* [sector_num, pnum] unallocated on this layer, which could be only 1708 * the first part of [sector_num, nb_sectors]. */ 1709 nb_sectors = MIN(nb_sectors, *pnum); 1710 } 1711 return ret; 1712 } 1713 1714 /* Coroutine wrapper for bdrv_get_block_status_above() */ 1715 static void coroutine_fn bdrv_get_block_status_above_co_entry(void *opaque) 1716 { 1717 BdrvCoGetBlockStatusData *data = opaque; 1718 1719 data->ret = bdrv_co_get_block_status_above(data->bs, data->base, 1720 data->sector_num, 1721 data->nb_sectors, 1722 data->pnum, 1723 data->file); 1724 data->done = true; 1725 } 1726 1727 /* 1728 * Synchronous wrapper around bdrv_co_get_block_status_above(). 1729 * 1730 * See bdrv_co_get_block_status_above() for details. 1731 */ 1732 int64_t bdrv_get_block_status_above(BlockDriverState *bs, 1733 BlockDriverState *base, 1734 int64_t sector_num, 1735 int nb_sectors, int *pnum, 1736 BlockDriverState **file) 1737 { 1738 Coroutine *co; 1739 BdrvCoGetBlockStatusData data = { 1740 .bs = bs, 1741 .base = base, 1742 .file = file, 1743 .sector_num = sector_num, 1744 .nb_sectors = nb_sectors, 1745 .pnum = pnum, 1746 .done = false, 1747 }; 1748 1749 if (qemu_in_coroutine()) { 1750 /* Fast-path if already in coroutine context */ 1751 bdrv_get_block_status_above_co_entry(&data); 1752 } else { 1753 AioContext *aio_context = bdrv_get_aio_context(bs); 1754 1755 co = qemu_coroutine_create(bdrv_get_block_status_above_co_entry, 1756 &data); 1757 qemu_coroutine_enter(co); 1758 while (!data.done) { 1759 aio_poll(aio_context, true); 1760 } 1761 } 1762 return data.ret; 1763 } 1764 1765 int64_t bdrv_get_block_status(BlockDriverState *bs, 1766 int64_t sector_num, 1767 int nb_sectors, int *pnum, 1768 BlockDriverState **file) 1769 { 1770 return bdrv_get_block_status_above(bs, backing_bs(bs), 1771 sector_num, nb_sectors, pnum, file); 1772 } 1773 1774 int coroutine_fn bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, 1775 int nb_sectors, int *pnum) 1776 { 1777 BlockDriverState *file; 1778 int64_t ret = bdrv_get_block_status(bs, sector_num, nb_sectors, pnum, 1779 &file); 1780 if (ret < 0) { 1781 return ret; 1782 } 1783 return !!(ret & BDRV_BLOCK_ALLOCATED); 1784 } 1785 1786 /* 1787 * Given an image chain: ... -> [BASE] -> [INTER1] -> [INTER2] -> [TOP] 1788 * 1789 * Return true if the given sector is allocated in any image between 1790 * BASE and TOP (inclusive). BASE can be NULL to check if the given 1791 * sector is allocated in any image of the chain. Return false otherwise. 1792 * 1793 * 'pnum' is set to the number of sectors (including and immediately following 1794 * the specified sector) that are known to be in the same 1795 * allocated/unallocated state. 1796 * 1797 */ 1798 int bdrv_is_allocated_above(BlockDriverState *top, 1799 BlockDriverState *base, 1800 int64_t sector_num, 1801 int nb_sectors, int *pnum) 1802 { 1803 BlockDriverState *intermediate; 1804 int ret, n = nb_sectors; 1805 1806 intermediate = top; 1807 while (intermediate && intermediate != base) { 1808 int pnum_inter; 1809 ret = bdrv_is_allocated(intermediate, sector_num, nb_sectors, 1810 &pnum_inter); 1811 if (ret < 0) { 1812 return ret; 1813 } else if (ret) { 1814 *pnum = pnum_inter; 1815 return 1; 1816 } 1817 1818 /* 1819 * [sector_num, nb_sectors] is unallocated on top but intermediate 1820 * might have 1821 * 1822 * [sector_num+x, nr_sectors] allocated. 1823 */ 1824 if (n > pnum_inter && 1825 (intermediate == top || 1826 sector_num + pnum_inter < intermediate->total_sectors)) { 1827 n = pnum_inter; 1828 } 1829 1830 intermediate = backing_bs(intermediate); 1831 } 1832 1833 *pnum = n; 1834 return 0; 1835 } 1836 1837 int bdrv_write_compressed(BlockDriverState *bs, int64_t sector_num, 1838 const uint8_t *buf, int nb_sectors) 1839 { 1840 BlockDriver *drv = bs->drv; 1841 int ret; 1842 1843 if (!drv) { 1844 return -ENOMEDIUM; 1845 } 1846 if (!drv->bdrv_write_compressed) { 1847 return -ENOTSUP; 1848 } 1849 ret = bdrv_check_request(bs, sector_num, nb_sectors); 1850 if (ret < 0) { 1851 return ret; 1852 } 1853 1854 assert(QLIST_EMPTY(&bs->dirty_bitmaps)); 1855 1856 return drv->bdrv_write_compressed(bs, sector_num, buf, nb_sectors); 1857 } 1858 1859 typedef struct BdrvVmstateCo { 1860 BlockDriverState *bs; 1861 QEMUIOVector *qiov; 1862 int64_t pos; 1863 bool is_read; 1864 int ret; 1865 } BdrvVmstateCo; 1866 1867 static int coroutine_fn 1868 bdrv_co_rw_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos, 1869 bool is_read) 1870 { 1871 BlockDriver *drv = bs->drv; 1872 1873 if (!drv) { 1874 return -ENOMEDIUM; 1875 } else if (drv->bdrv_load_vmstate) { 1876 return is_read ? drv->bdrv_load_vmstate(bs, qiov, pos) 1877 : drv->bdrv_save_vmstate(bs, qiov, pos); 1878 } else if (bs->file) { 1879 return bdrv_co_rw_vmstate(bs->file->bs, qiov, pos, is_read); 1880 } 1881 1882 return -ENOTSUP; 1883 } 1884 1885 static void coroutine_fn bdrv_co_rw_vmstate_entry(void *opaque) 1886 { 1887 BdrvVmstateCo *co = opaque; 1888 co->ret = bdrv_co_rw_vmstate(co->bs, co->qiov, co->pos, co->is_read); 1889 } 1890 1891 static inline int 1892 bdrv_rw_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos, 1893 bool is_read) 1894 { 1895 if (qemu_in_coroutine()) { 1896 return bdrv_co_rw_vmstate(bs, qiov, pos, is_read); 1897 } else { 1898 BdrvVmstateCo data = { 1899 .bs = bs, 1900 .qiov = qiov, 1901 .pos = pos, 1902 .is_read = is_read, 1903 .ret = -EINPROGRESS, 1904 }; 1905 Coroutine *co = qemu_coroutine_create(bdrv_co_rw_vmstate_entry, &data); 1906 1907 qemu_coroutine_enter(co); 1908 while (data.ret == -EINPROGRESS) { 1909 aio_poll(bdrv_get_aio_context(bs), true); 1910 } 1911 return data.ret; 1912 } 1913 } 1914 1915 int bdrv_save_vmstate(BlockDriverState *bs, const uint8_t *buf, 1916 int64_t pos, int size) 1917 { 1918 QEMUIOVector qiov; 1919 struct iovec iov = { 1920 .iov_base = (void *) buf, 1921 .iov_len = size, 1922 }; 1923 int ret; 1924 1925 qemu_iovec_init_external(&qiov, &iov, 1); 1926 1927 ret = bdrv_writev_vmstate(bs, &qiov, pos); 1928 if (ret < 0) { 1929 return ret; 1930 } 1931 1932 return size; 1933 } 1934 1935 int bdrv_writev_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos) 1936 { 1937 return bdrv_rw_vmstate(bs, qiov, pos, false); 1938 } 1939 1940 int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf, 1941 int64_t pos, int size) 1942 { 1943 QEMUIOVector qiov; 1944 struct iovec iov = { 1945 .iov_base = buf, 1946 .iov_len = size, 1947 }; 1948 int ret; 1949 1950 qemu_iovec_init_external(&qiov, &iov, 1); 1951 ret = bdrv_readv_vmstate(bs, &qiov, pos); 1952 if (ret < 0) { 1953 return ret; 1954 } 1955 1956 return size; 1957 } 1958 1959 int bdrv_readv_vmstate(BlockDriverState *bs, QEMUIOVector *qiov, int64_t pos) 1960 { 1961 return bdrv_rw_vmstate(bs, qiov, pos, true); 1962 } 1963 1964 /**************************************************************/ 1965 /* async I/Os */ 1966 1967 BlockAIOCB *bdrv_aio_readv(BdrvChild *child, int64_t sector_num, 1968 QEMUIOVector *qiov, int nb_sectors, 1969 BlockCompletionFunc *cb, void *opaque) 1970 { 1971 trace_bdrv_aio_readv(child->bs, sector_num, nb_sectors, opaque); 1972 1973 return bdrv_co_aio_rw_vector(child, sector_num, qiov, nb_sectors, 0, 1974 cb, opaque, false); 1975 } 1976 1977 BlockAIOCB *bdrv_aio_writev(BdrvChild *child, int64_t sector_num, 1978 QEMUIOVector *qiov, int nb_sectors, 1979 BlockCompletionFunc *cb, void *opaque) 1980 { 1981 trace_bdrv_aio_writev(child->bs, sector_num, nb_sectors, opaque); 1982 1983 return bdrv_co_aio_rw_vector(child, sector_num, qiov, nb_sectors, 0, 1984 cb, opaque, true); 1985 } 1986 1987 void bdrv_aio_cancel(BlockAIOCB *acb) 1988 { 1989 qemu_aio_ref(acb); 1990 bdrv_aio_cancel_async(acb); 1991 while (acb->refcnt > 1) { 1992 if (acb->aiocb_info->get_aio_context) { 1993 aio_poll(acb->aiocb_info->get_aio_context(acb), true); 1994 } else if (acb->bs) { 1995 aio_poll(bdrv_get_aio_context(acb->bs), true); 1996 } else { 1997 abort(); 1998 } 1999 } 2000 qemu_aio_unref(acb); 2001 } 2002 2003 /* Async version of aio cancel. The caller is not blocked if the acb implements 2004 * cancel_async, otherwise we do nothing and let the request normally complete. 2005 * In either case the completion callback must be called. */ 2006 void bdrv_aio_cancel_async(BlockAIOCB *acb) 2007 { 2008 if (acb->aiocb_info->cancel_async) { 2009 acb->aiocb_info->cancel_async(acb); 2010 } 2011 } 2012 2013 /**************************************************************/ 2014 /* async block device emulation */ 2015 2016 typedef struct BlockRequest { 2017 union { 2018 /* Used during read, write, trim */ 2019 struct { 2020 int64_t sector; 2021 int nb_sectors; 2022 int flags; 2023 QEMUIOVector *qiov; 2024 }; 2025 /* Used during ioctl */ 2026 struct { 2027 int req; 2028 void *buf; 2029 }; 2030 }; 2031 BlockCompletionFunc *cb; 2032 void *opaque; 2033 2034 int error; 2035 } BlockRequest; 2036 2037 typedef struct BlockAIOCBCoroutine { 2038 BlockAIOCB common; 2039 BdrvChild *child; 2040 BlockRequest req; 2041 bool is_write; 2042 bool need_bh; 2043 bool *done; 2044 QEMUBH* bh; 2045 } BlockAIOCBCoroutine; 2046 2047 static const AIOCBInfo bdrv_em_co_aiocb_info = { 2048 .aiocb_size = sizeof(BlockAIOCBCoroutine), 2049 }; 2050 2051 static void bdrv_co_complete(BlockAIOCBCoroutine *acb) 2052 { 2053 if (!acb->need_bh) { 2054 acb->common.cb(acb->common.opaque, acb->req.error); 2055 qemu_aio_unref(acb); 2056 } 2057 } 2058 2059 static void bdrv_co_em_bh(void *opaque) 2060 { 2061 BlockAIOCBCoroutine *acb = opaque; 2062 2063 assert(!acb->need_bh); 2064 qemu_bh_delete(acb->bh); 2065 bdrv_co_complete(acb); 2066 } 2067 2068 static void bdrv_co_maybe_schedule_bh(BlockAIOCBCoroutine *acb) 2069 { 2070 acb->need_bh = false; 2071 if (acb->req.error != -EINPROGRESS) { 2072 BlockDriverState *bs = acb->common.bs; 2073 2074 acb->bh = aio_bh_new(bdrv_get_aio_context(bs), bdrv_co_em_bh, acb); 2075 qemu_bh_schedule(acb->bh); 2076 } 2077 } 2078 2079 /* Invoke bdrv_co_do_readv/bdrv_co_do_writev */ 2080 static void coroutine_fn bdrv_co_do_rw(void *opaque) 2081 { 2082 BlockAIOCBCoroutine *acb = opaque; 2083 2084 if (!acb->is_write) { 2085 acb->req.error = bdrv_co_do_readv(acb->child, acb->req.sector, 2086 acb->req.nb_sectors, acb->req.qiov, acb->req.flags); 2087 } else { 2088 acb->req.error = bdrv_co_do_writev(acb->child, acb->req.sector, 2089 acb->req.nb_sectors, acb->req.qiov, acb->req.flags); 2090 } 2091 2092 bdrv_co_complete(acb); 2093 } 2094 2095 static BlockAIOCB *bdrv_co_aio_rw_vector(BdrvChild *child, 2096 int64_t sector_num, 2097 QEMUIOVector *qiov, 2098 int nb_sectors, 2099 BdrvRequestFlags flags, 2100 BlockCompletionFunc *cb, 2101 void *opaque, 2102 bool is_write) 2103 { 2104 Coroutine *co; 2105 BlockAIOCBCoroutine *acb; 2106 2107 acb = qemu_aio_get(&bdrv_em_co_aiocb_info, child->bs, cb, opaque); 2108 acb->child = child; 2109 acb->need_bh = true; 2110 acb->req.error = -EINPROGRESS; 2111 acb->req.sector = sector_num; 2112 acb->req.nb_sectors = nb_sectors; 2113 acb->req.qiov = qiov; 2114 acb->req.flags = flags; 2115 acb->is_write = is_write; 2116 2117 co = qemu_coroutine_create(bdrv_co_do_rw, acb); 2118 qemu_coroutine_enter(co); 2119 2120 bdrv_co_maybe_schedule_bh(acb); 2121 return &acb->common; 2122 } 2123 2124 static void coroutine_fn bdrv_aio_flush_co_entry(void *opaque) 2125 { 2126 BlockAIOCBCoroutine *acb = opaque; 2127 BlockDriverState *bs = acb->common.bs; 2128 2129 acb->req.error = bdrv_co_flush(bs); 2130 bdrv_co_complete(acb); 2131 } 2132 2133 BlockAIOCB *bdrv_aio_flush(BlockDriverState *bs, 2134 BlockCompletionFunc *cb, void *opaque) 2135 { 2136 trace_bdrv_aio_flush(bs, opaque); 2137 2138 Coroutine *co; 2139 BlockAIOCBCoroutine *acb; 2140 2141 acb = qemu_aio_get(&bdrv_em_co_aiocb_info, bs, cb, opaque); 2142 acb->need_bh = true; 2143 acb->req.error = -EINPROGRESS; 2144 2145 co = qemu_coroutine_create(bdrv_aio_flush_co_entry, acb); 2146 qemu_coroutine_enter(co); 2147 2148 bdrv_co_maybe_schedule_bh(acb); 2149 return &acb->common; 2150 } 2151 2152 static void coroutine_fn bdrv_aio_discard_co_entry(void *opaque) 2153 { 2154 BlockAIOCBCoroutine *acb = opaque; 2155 BlockDriverState *bs = acb->common.bs; 2156 2157 acb->req.error = bdrv_co_discard(bs, acb->req.sector, acb->req.nb_sectors); 2158 bdrv_co_complete(acb); 2159 } 2160 2161 BlockAIOCB *bdrv_aio_discard(BlockDriverState *bs, 2162 int64_t sector_num, int nb_sectors, 2163 BlockCompletionFunc *cb, void *opaque) 2164 { 2165 Coroutine *co; 2166 BlockAIOCBCoroutine *acb; 2167 2168 trace_bdrv_aio_discard(bs, sector_num, nb_sectors, opaque); 2169 2170 acb = qemu_aio_get(&bdrv_em_co_aiocb_info, bs, cb, opaque); 2171 acb->need_bh = true; 2172 acb->req.error = -EINPROGRESS; 2173 acb->req.sector = sector_num; 2174 acb->req.nb_sectors = nb_sectors; 2175 co = qemu_coroutine_create(bdrv_aio_discard_co_entry, acb); 2176 qemu_coroutine_enter(co); 2177 2178 bdrv_co_maybe_schedule_bh(acb); 2179 return &acb->common; 2180 } 2181 2182 void *qemu_aio_get(const AIOCBInfo *aiocb_info, BlockDriverState *bs, 2183 BlockCompletionFunc *cb, void *opaque) 2184 { 2185 BlockAIOCB *acb; 2186 2187 acb = g_malloc(aiocb_info->aiocb_size); 2188 acb->aiocb_info = aiocb_info; 2189 acb->bs = bs; 2190 acb->cb = cb; 2191 acb->opaque = opaque; 2192 acb->refcnt = 1; 2193 return acb; 2194 } 2195 2196 void qemu_aio_ref(void *p) 2197 { 2198 BlockAIOCB *acb = p; 2199 acb->refcnt++; 2200 } 2201 2202 void qemu_aio_unref(void *p) 2203 { 2204 BlockAIOCB *acb = p; 2205 assert(acb->refcnt > 0); 2206 if (--acb->refcnt == 0) { 2207 g_free(acb); 2208 } 2209 } 2210 2211 /**************************************************************/ 2212 /* Coroutine block device emulation */ 2213 2214 typedef struct FlushCo { 2215 BlockDriverState *bs; 2216 int ret; 2217 } FlushCo; 2218 2219 2220 static void coroutine_fn bdrv_flush_co_entry(void *opaque) 2221 { 2222 FlushCo *rwco = opaque; 2223 2224 rwco->ret = bdrv_co_flush(rwco->bs); 2225 } 2226 2227 int coroutine_fn bdrv_co_flush(BlockDriverState *bs) 2228 { 2229 int ret; 2230 BdrvTrackedRequest req; 2231 2232 if (!bs || !bdrv_is_inserted(bs) || bdrv_is_read_only(bs) || 2233 bdrv_is_sg(bs)) { 2234 return 0; 2235 } 2236 2237 tracked_request_begin(&req, bs, 0, 0, BDRV_TRACKED_FLUSH); 2238 2239 /* Write back all layers by calling one driver function */ 2240 if (bs->drv->bdrv_co_flush) { 2241 ret = bs->drv->bdrv_co_flush(bs); 2242 goto out; 2243 } 2244 2245 /* Write back cached data to the OS even with cache=unsafe */ 2246 BLKDBG_EVENT(bs->file, BLKDBG_FLUSH_TO_OS); 2247 if (bs->drv->bdrv_co_flush_to_os) { 2248 ret = bs->drv->bdrv_co_flush_to_os(bs); 2249 if (ret < 0) { 2250 goto out; 2251 } 2252 } 2253 2254 /* But don't actually force it to the disk with cache=unsafe */ 2255 if (bs->open_flags & BDRV_O_NO_FLUSH) { 2256 goto flush_parent; 2257 } 2258 2259 BLKDBG_EVENT(bs->file, BLKDBG_FLUSH_TO_DISK); 2260 if (bs->drv->bdrv_co_flush_to_disk) { 2261 ret = bs->drv->bdrv_co_flush_to_disk(bs); 2262 } else if (bs->drv->bdrv_aio_flush) { 2263 BlockAIOCB *acb; 2264 CoroutineIOCompletion co = { 2265 .coroutine = qemu_coroutine_self(), 2266 }; 2267 2268 acb = bs->drv->bdrv_aio_flush(bs, bdrv_co_io_em_complete, &co); 2269 if (acb == NULL) { 2270 ret = -EIO; 2271 } else { 2272 qemu_coroutine_yield(); 2273 ret = co.ret; 2274 } 2275 } else { 2276 /* 2277 * Some block drivers always operate in either writethrough or unsafe 2278 * mode and don't support bdrv_flush therefore. Usually qemu doesn't 2279 * know how the server works (because the behaviour is hardcoded or 2280 * depends on server-side configuration), so we can't ensure that 2281 * everything is safe on disk. Returning an error doesn't work because 2282 * that would break guests even if the server operates in writethrough 2283 * mode. 2284 * 2285 * Let's hope the user knows what he's doing. 2286 */ 2287 ret = 0; 2288 } 2289 if (ret < 0) { 2290 goto out; 2291 } 2292 2293 /* Now flush the underlying protocol. It will also have BDRV_O_NO_FLUSH 2294 * in the case of cache=unsafe, so there are no useless flushes. 2295 */ 2296 flush_parent: 2297 ret = bs->file ? bdrv_co_flush(bs->file->bs) : 0; 2298 out: 2299 tracked_request_end(&req); 2300 return ret; 2301 } 2302 2303 int bdrv_flush(BlockDriverState *bs) 2304 { 2305 Coroutine *co; 2306 FlushCo flush_co = { 2307 .bs = bs, 2308 .ret = NOT_DONE, 2309 }; 2310 2311 if (qemu_in_coroutine()) { 2312 /* Fast-path if already in coroutine context */ 2313 bdrv_flush_co_entry(&flush_co); 2314 } else { 2315 AioContext *aio_context = bdrv_get_aio_context(bs); 2316 2317 co = qemu_coroutine_create(bdrv_flush_co_entry, &flush_co); 2318 qemu_coroutine_enter(co); 2319 while (flush_co.ret == NOT_DONE) { 2320 aio_poll(aio_context, true); 2321 } 2322 } 2323 2324 return flush_co.ret; 2325 } 2326 2327 typedef struct DiscardCo { 2328 BlockDriverState *bs; 2329 int64_t sector_num; 2330 int nb_sectors; 2331 int ret; 2332 } DiscardCo; 2333 static void coroutine_fn bdrv_discard_co_entry(void *opaque) 2334 { 2335 DiscardCo *rwco = opaque; 2336 2337 rwco->ret = bdrv_co_discard(rwco->bs, rwco->sector_num, rwco->nb_sectors); 2338 } 2339 2340 int coroutine_fn bdrv_co_discard(BlockDriverState *bs, int64_t sector_num, 2341 int nb_sectors) 2342 { 2343 BdrvTrackedRequest req; 2344 int max_discard, ret; 2345 2346 if (!bs->drv) { 2347 return -ENOMEDIUM; 2348 } 2349 2350 ret = bdrv_check_request(bs, sector_num, nb_sectors); 2351 if (ret < 0) { 2352 return ret; 2353 } else if (bs->read_only) { 2354 return -EPERM; 2355 } 2356 assert(!(bs->open_flags & BDRV_O_INACTIVE)); 2357 2358 /* Do nothing if disabled. */ 2359 if (!(bs->open_flags & BDRV_O_UNMAP)) { 2360 return 0; 2361 } 2362 2363 if (!bs->drv->bdrv_co_discard && !bs->drv->bdrv_aio_discard) { 2364 return 0; 2365 } 2366 2367 tracked_request_begin(&req, bs, sector_num << BDRV_SECTOR_BITS, 2368 nb_sectors << BDRV_SECTOR_BITS, BDRV_TRACKED_DISCARD); 2369 2370 ret = notifier_with_return_list_notify(&bs->before_write_notifiers, &req); 2371 if (ret < 0) { 2372 goto out; 2373 } 2374 2375 max_discard = MIN_NON_ZERO(bs->bl.max_pdiscard >> BDRV_SECTOR_BITS, 2376 BDRV_REQUEST_MAX_SECTORS); 2377 while (nb_sectors > 0) { 2378 int ret; 2379 int num = nb_sectors; 2380 int discard_alignment = bs->bl.pdiscard_alignment >> BDRV_SECTOR_BITS; 2381 2382 /* align request */ 2383 if (discard_alignment && 2384 num >= discard_alignment && 2385 sector_num % discard_alignment) { 2386 if (num > discard_alignment) { 2387 num = discard_alignment; 2388 } 2389 num -= sector_num % discard_alignment; 2390 } 2391 2392 /* limit request size */ 2393 if (num > max_discard) { 2394 num = max_discard; 2395 } 2396 2397 if (bs->drv->bdrv_co_discard) { 2398 ret = bs->drv->bdrv_co_discard(bs, sector_num, num); 2399 } else { 2400 BlockAIOCB *acb; 2401 CoroutineIOCompletion co = { 2402 .coroutine = qemu_coroutine_self(), 2403 }; 2404 2405 acb = bs->drv->bdrv_aio_discard(bs, sector_num, nb_sectors, 2406 bdrv_co_io_em_complete, &co); 2407 if (acb == NULL) { 2408 ret = -EIO; 2409 goto out; 2410 } else { 2411 qemu_coroutine_yield(); 2412 ret = co.ret; 2413 } 2414 } 2415 if (ret && ret != -ENOTSUP) { 2416 goto out; 2417 } 2418 2419 sector_num += num; 2420 nb_sectors -= num; 2421 } 2422 ret = 0; 2423 out: 2424 bdrv_set_dirty(bs, req.offset >> BDRV_SECTOR_BITS, 2425 req.bytes >> BDRV_SECTOR_BITS); 2426 tracked_request_end(&req); 2427 return ret; 2428 } 2429 2430 int bdrv_discard(BlockDriverState *bs, int64_t sector_num, int nb_sectors) 2431 { 2432 Coroutine *co; 2433 DiscardCo rwco = { 2434 .bs = bs, 2435 .sector_num = sector_num, 2436 .nb_sectors = nb_sectors, 2437 .ret = NOT_DONE, 2438 }; 2439 2440 if (qemu_in_coroutine()) { 2441 /* Fast-path if already in coroutine context */ 2442 bdrv_discard_co_entry(&rwco); 2443 } else { 2444 AioContext *aio_context = bdrv_get_aio_context(bs); 2445 2446 co = qemu_coroutine_create(bdrv_discard_co_entry, &rwco); 2447 qemu_coroutine_enter(co); 2448 while (rwco.ret == NOT_DONE) { 2449 aio_poll(aio_context, true); 2450 } 2451 } 2452 2453 return rwco.ret; 2454 } 2455 2456 static int bdrv_co_do_ioctl(BlockDriverState *bs, int req, void *buf) 2457 { 2458 BlockDriver *drv = bs->drv; 2459 BdrvTrackedRequest tracked_req; 2460 CoroutineIOCompletion co = { 2461 .coroutine = qemu_coroutine_self(), 2462 }; 2463 BlockAIOCB *acb; 2464 2465 tracked_request_begin(&tracked_req, bs, 0, 0, BDRV_TRACKED_IOCTL); 2466 if (!drv || !drv->bdrv_aio_ioctl) { 2467 co.ret = -ENOTSUP; 2468 goto out; 2469 } 2470 2471 acb = drv->bdrv_aio_ioctl(bs, req, buf, bdrv_co_io_em_complete, &co); 2472 if (!acb) { 2473 co.ret = -ENOTSUP; 2474 goto out; 2475 } 2476 qemu_coroutine_yield(); 2477 out: 2478 tracked_request_end(&tracked_req); 2479 return co.ret; 2480 } 2481 2482 typedef struct { 2483 BlockDriverState *bs; 2484 int req; 2485 void *buf; 2486 int ret; 2487 } BdrvIoctlCoData; 2488 2489 static void coroutine_fn bdrv_co_ioctl_entry(void *opaque) 2490 { 2491 BdrvIoctlCoData *data = opaque; 2492 data->ret = bdrv_co_do_ioctl(data->bs, data->req, data->buf); 2493 } 2494 2495 /* needed for generic scsi interface */ 2496 int bdrv_ioctl(BlockDriverState *bs, unsigned long int req, void *buf) 2497 { 2498 BdrvIoctlCoData data = { 2499 .bs = bs, 2500 .req = req, 2501 .buf = buf, 2502 .ret = -EINPROGRESS, 2503 }; 2504 2505 if (qemu_in_coroutine()) { 2506 /* Fast-path if already in coroutine context */ 2507 bdrv_co_ioctl_entry(&data); 2508 } else { 2509 Coroutine *co = qemu_coroutine_create(bdrv_co_ioctl_entry, &data); 2510 2511 qemu_coroutine_enter(co); 2512 while (data.ret == -EINPROGRESS) { 2513 aio_poll(bdrv_get_aio_context(bs), true); 2514 } 2515 } 2516 return data.ret; 2517 } 2518 2519 static void coroutine_fn bdrv_co_aio_ioctl_entry(void *opaque) 2520 { 2521 BlockAIOCBCoroutine *acb = opaque; 2522 acb->req.error = bdrv_co_do_ioctl(acb->common.bs, 2523 acb->req.req, acb->req.buf); 2524 bdrv_co_complete(acb); 2525 } 2526 2527 BlockAIOCB *bdrv_aio_ioctl(BlockDriverState *bs, 2528 unsigned long int req, void *buf, 2529 BlockCompletionFunc *cb, void *opaque) 2530 { 2531 BlockAIOCBCoroutine *acb = qemu_aio_get(&bdrv_em_co_aiocb_info, 2532 bs, cb, opaque); 2533 Coroutine *co; 2534 2535 acb->need_bh = true; 2536 acb->req.error = -EINPROGRESS; 2537 acb->req.req = req; 2538 acb->req.buf = buf; 2539 co = qemu_coroutine_create(bdrv_co_aio_ioctl_entry, acb); 2540 qemu_coroutine_enter(co); 2541 2542 bdrv_co_maybe_schedule_bh(acb); 2543 return &acb->common; 2544 } 2545 2546 void *qemu_blockalign(BlockDriverState *bs, size_t size) 2547 { 2548 return qemu_memalign(bdrv_opt_mem_align(bs), size); 2549 } 2550 2551 void *qemu_blockalign0(BlockDriverState *bs, size_t size) 2552 { 2553 return memset(qemu_blockalign(bs, size), 0, size); 2554 } 2555 2556 void *qemu_try_blockalign(BlockDriverState *bs, size_t size) 2557 { 2558 size_t align = bdrv_opt_mem_align(bs); 2559 2560 /* Ensure that NULL is never returned on success */ 2561 assert(align > 0); 2562 if (size == 0) { 2563 size = align; 2564 } 2565 2566 return qemu_try_memalign(align, size); 2567 } 2568 2569 void *qemu_try_blockalign0(BlockDriverState *bs, size_t size) 2570 { 2571 void *mem = qemu_try_blockalign(bs, size); 2572 2573 if (mem) { 2574 memset(mem, 0, size); 2575 } 2576 2577 return mem; 2578 } 2579 2580 /* 2581 * Check if all memory in this vector is sector aligned. 2582 */ 2583 bool bdrv_qiov_is_aligned(BlockDriverState *bs, QEMUIOVector *qiov) 2584 { 2585 int i; 2586 size_t alignment = bdrv_min_mem_align(bs); 2587 2588 for (i = 0; i < qiov->niov; i++) { 2589 if ((uintptr_t) qiov->iov[i].iov_base % alignment) { 2590 return false; 2591 } 2592 if (qiov->iov[i].iov_len % alignment) { 2593 return false; 2594 } 2595 } 2596 2597 return true; 2598 } 2599 2600 void bdrv_add_before_write_notifier(BlockDriverState *bs, 2601 NotifierWithReturn *notifier) 2602 { 2603 notifier_with_return_list_add(&bs->before_write_notifiers, notifier); 2604 } 2605 2606 void bdrv_io_plug(BlockDriverState *bs) 2607 { 2608 BdrvChild *child; 2609 2610 QLIST_FOREACH(child, &bs->children, next) { 2611 bdrv_io_plug(child->bs); 2612 } 2613 2614 if (bs->io_plugged++ == 0 && bs->io_plug_disabled == 0) { 2615 BlockDriver *drv = bs->drv; 2616 if (drv && drv->bdrv_io_plug) { 2617 drv->bdrv_io_plug(bs); 2618 } 2619 } 2620 } 2621 2622 void bdrv_io_unplug(BlockDriverState *bs) 2623 { 2624 BdrvChild *child; 2625 2626 assert(bs->io_plugged); 2627 if (--bs->io_plugged == 0 && bs->io_plug_disabled == 0) { 2628 BlockDriver *drv = bs->drv; 2629 if (drv && drv->bdrv_io_unplug) { 2630 drv->bdrv_io_unplug(bs); 2631 } 2632 } 2633 2634 QLIST_FOREACH(child, &bs->children, next) { 2635 bdrv_io_unplug(child->bs); 2636 } 2637 } 2638 2639 void bdrv_io_unplugged_begin(BlockDriverState *bs) 2640 { 2641 BdrvChild *child; 2642 2643 if (bs->io_plug_disabled++ == 0 && bs->io_plugged > 0) { 2644 BlockDriver *drv = bs->drv; 2645 if (drv && drv->bdrv_io_unplug) { 2646 drv->bdrv_io_unplug(bs); 2647 } 2648 } 2649 2650 QLIST_FOREACH(child, &bs->children, next) { 2651 bdrv_io_unplugged_begin(child->bs); 2652 } 2653 } 2654 2655 void bdrv_io_unplugged_end(BlockDriverState *bs) 2656 { 2657 BdrvChild *child; 2658 2659 assert(bs->io_plug_disabled); 2660 QLIST_FOREACH(child, &bs->children, next) { 2661 bdrv_io_unplugged_end(child->bs); 2662 } 2663 2664 if (--bs->io_plug_disabled == 0 && bs->io_plugged > 0) { 2665 BlockDriver *drv = bs->drv; 2666 if (drv && drv->bdrv_io_plug) { 2667 drv->bdrv_io_plug(bs); 2668 } 2669 } 2670 } 2671