1 /* 2 * QEMU System Emulator block driver 3 * 4 * Copyright (c) 2003 Fabrice Bellard 5 * Copyright (c) 2020 Virtuozzo International GmbH. 6 * 7 * Permission is hereby granted, free of charge, to any person obtaining a copy 8 * of this software and associated documentation files (the "Software"), to deal 9 * in the Software without restriction, including without limitation the rights 10 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 11 * copies of the Software, and to permit persons to whom the Software is 12 * furnished to do so, subject to the following conditions: 13 * 14 * The above copyright notice and this permission notice shall be included in 15 * all copies or substantial portions of the Software. 16 * 17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 23 * THE SOFTWARE. 24 */ 25 26 #include "qemu/osdep.h" 27 #include "block/trace.h" 28 #include "block/block_int.h" 29 #include "block/blockjob.h" 30 #include "block/dirty-bitmap.h" 31 #include "block/fuse.h" 32 #include "block/nbd.h" 33 #include "block/qdict.h" 34 #include "qemu/error-report.h" 35 #include "block/module_block.h" 36 #include "qemu/main-loop.h" 37 #include "qemu/module.h" 38 #include "qapi/error.h" 39 #include "qapi/qmp/qdict.h" 40 #include "qapi/qmp/qjson.h" 41 #include "qapi/qmp/qnull.h" 42 #include "qapi/qmp/qstring.h" 43 #include "qapi/qobject-output-visitor.h" 44 #include "qapi/qapi-visit-block-core.h" 45 #include "sysemu/block-backend.h" 46 #include "qemu/notify.h" 47 #include "qemu/option.h" 48 #include "qemu/coroutine.h" 49 #include "block/qapi.h" 50 #include "qemu/timer.h" 51 #include "qemu/cutils.h" 52 #include "qemu/id.h" 53 #include "qemu/range.h" 54 #include "qemu/rcu.h" 55 #include "block/coroutines.h" 56 57 #ifdef CONFIG_BSD 58 #include <sys/ioctl.h> 59 #include <sys/queue.h> 60 #if defined(HAVE_SYS_DISK_H) 61 #include <sys/disk.h> 62 #endif 63 #endif 64 65 #ifdef _WIN32 66 #include <windows.h> 67 #endif 68 69 #define NOT_DONE 0x7fffffff /* used while emulated sync operation in progress */ 70 71 /* Protected by BQL */ 72 static QTAILQ_HEAD(, BlockDriverState) graph_bdrv_states = 73 QTAILQ_HEAD_INITIALIZER(graph_bdrv_states); 74 75 /* Protected by BQL */ 76 static QTAILQ_HEAD(, BlockDriverState) all_bdrv_states = 77 QTAILQ_HEAD_INITIALIZER(all_bdrv_states); 78 79 /* Protected by BQL */ 80 static QLIST_HEAD(, BlockDriver) bdrv_drivers = 81 QLIST_HEAD_INITIALIZER(bdrv_drivers); 82 83 static BlockDriverState *bdrv_open_inherit(const char *filename, 84 const char *reference, 85 QDict *options, int flags, 86 BlockDriverState *parent, 87 const BdrvChildClass *child_class, 88 BdrvChildRole child_role, 89 Error **errp); 90 91 static bool bdrv_recurse_has_child(BlockDriverState *bs, 92 BlockDriverState *child); 93 94 static void GRAPH_WRLOCK 95 bdrv_replace_child_noperm(BdrvChild *child, BlockDriverState *new_bs); 96 97 static void GRAPH_WRLOCK 98 bdrv_remove_child(BdrvChild *child, Transaction *tran); 99 100 static int bdrv_reopen_prepare(BDRVReopenState *reopen_state, 101 BlockReopenQueue *queue, 102 Transaction *change_child_tran, Error **errp); 103 static void bdrv_reopen_commit(BDRVReopenState *reopen_state); 104 static void bdrv_reopen_abort(BDRVReopenState *reopen_state); 105 106 static bool bdrv_backing_overridden(BlockDriverState *bs); 107 108 static bool bdrv_change_aio_context(BlockDriverState *bs, AioContext *ctx, 109 GHashTable *visited, Transaction *tran, 110 Error **errp); 111 112 /* If non-zero, use only whitelisted block drivers */ 113 static int use_bdrv_whitelist; 114 115 #ifdef _WIN32 116 static int is_windows_drive_prefix(const char *filename) 117 { 118 return (((filename[0] >= 'a' && filename[0] <= 'z') || 119 (filename[0] >= 'A' && filename[0] <= 'Z')) && 120 filename[1] == ':'); 121 } 122 123 int is_windows_drive(const char *filename) 124 { 125 if (is_windows_drive_prefix(filename) && 126 filename[2] == '\0') 127 return 1; 128 if (strstart(filename, "\\\\.\\", NULL) || 129 strstart(filename, "//./", NULL)) 130 return 1; 131 return 0; 132 } 133 #endif 134 135 size_t bdrv_opt_mem_align(BlockDriverState *bs) 136 { 137 if (!bs || !bs->drv) { 138 /* page size or 4k (hdd sector size) should be on the safe side */ 139 return MAX(4096, qemu_real_host_page_size()); 140 } 141 IO_CODE(); 142 143 return bs->bl.opt_mem_alignment; 144 } 145 146 size_t bdrv_min_mem_align(BlockDriverState *bs) 147 { 148 if (!bs || !bs->drv) { 149 /* page size or 4k (hdd sector size) should be on the safe side */ 150 return MAX(4096, qemu_real_host_page_size()); 151 } 152 IO_CODE(); 153 154 return bs->bl.min_mem_alignment; 155 } 156 157 /* check if the path starts with "<protocol>:" */ 158 int path_has_protocol(const char *path) 159 { 160 const char *p; 161 162 #ifdef _WIN32 163 if (is_windows_drive(path) || 164 is_windows_drive_prefix(path)) { 165 return 0; 166 } 167 p = path + strcspn(path, ":/\\"); 168 #else 169 p = path + strcspn(path, ":/"); 170 #endif 171 172 return *p == ':'; 173 } 174 175 int path_is_absolute(const char *path) 176 { 177 #ifdef _WIN32 178 /* specific case for names like: "\\.\d:" */ 179 if (is_windows_drive(path) || is_windows_drive_prefix(path)) { 180 return 1; 181 } 182 return (*path == '/' || *path == '\\'); 183 #else 184 return (*path == '/'); 185 #endif 186 } 187 188 /* if filename is absolute, just return its duplicate. Otherwise, build a 189 path to it by considering it is relative to base_path. URL are 190 supported. */ 191 char *path_combine(const char *base_path, const char *filename) 192 { 193 const char *protocol_stripped = NULL; 194 const char *p, *p1; 195 char *result; 196 int len; 197 198 if (path_is_absolute(filename)) { 199 return g_strdup(filename); 200 } 201 202 if (path_has_protocol(base_path)) { 203 protocol_stripped = strchr(base_path, ':'); 204 if (protocol_stripped) { 205 protocol_stripped++; 206 } 207 } 208 p = protocol_stripped ?: base_path; 209 210 p1 = strrchr(base_path, '/'); 211 #ifdef _WIN32 212 { 213 const char *p2; 214 p2 = strrchr(base_path, '\\'); 215 if (!p1 || p2 > p1) { 216 p1 = p2; 217 } 218 } 219 #endif 220 if (p1) { 221 p1++; 222 } else { 223 p1 = base_path; 224 } 225 if (p1 > p) { 226 p = p1; 227 } 228 len = p - base_path; 229 230 result = g_malloc(len + strlen(filename) + 1); 231 memcpy(result, base_path, len); 232 strcpy(result + len, filename); 233 234 return result; 235 } 236 237 /* 238 * Helper function for bdrv_parse_filename() implementations to remove optional 239 * protocol prefixes (especially "file:") from a filename and for putting the 240 * stripped filename into the options QDict if there is such a prefix. 241 */ 242 void bdrv_parse_filename_strip_prefix(const char *filename, const char *prefix, 243 QDict *options) 244 { 245 if (strstart(filename, prefix, &filename)) { 246 /* Stripping the explicit protocol prefix may result in a protocol 247 * prefix being (wrongly) detected (if the filename contains a colon) */ 248 if (path_has_protocol(filename)) { 249 GString *fat_filename; 250 251 /* This means there is some colon before the first slash; therefore, 252 * this cannot be an absolute path */ 253 assert(!path_is_absolute(filename)); 254 255 /* And we can thus fix the protocol detection issue by prefixing it 256 * by "./" */ 257 fat_filename = g_string_new("./"); 258 g_string_append(fat_filename, filename); 259 260 assert(!path_has_protocol(fat_filename->str)); 261 262 qdict_put(options, "filename", 263 qstring_from_gstring(fat_filename)); 264 } else { 265 /* If no protocol prefix was detected, we can use the shortened 266 * filename as-is */ 267 qdict_put_str(options, "filename", filename); 268 } 269 } 270 } 271 272 273 /* Returns whether the image file is opened as read-only. Note that this can 274 * return false and writing to the image file is still not possible because the 275 * image is inactivated. */ 276 bool bdrv_is_read_only(BlockDriverState *bs) 277 { 278 IO_CODE(); 279 return !(bs->open_flags & BDRV_O_RDWR); 280 } 281 282 static int GRAPH_RDLOCK 283 bdrv_can_set_read_only(BlockDriverState *bs, bool read_only, 284 bool ignore_allow_rdw, Error **errp) 285 { 286 IO_CODE(); 287 288 /* Do not set read_only if copy_on_read is enabled */ 289 if (bs->copy_on_read && read_only) { 290 error_setg(errp, "Can't set node '%s' to r/o with copy-on-read enabled", 291 bdrv_get_device_or_node_name(bs)); 292 return -EINVAL; 293 } 294 295 /* Do not clear read_only if it is prohibited */ 296 if (!read_only && !(bs->open_flags & BDRV_O_ALLOW_RDWR) && 297 !ignore_allow_rdw) 298 { 299 error_setg(errp, "Node '%s' is read only", 300 bdrv_get_device_or_node_name(bs)); 301 return -EPERM; 302 } 303 304 return 0; 305 } 306 307 /* 308 * Called by a driver that can only provide a read-only image. 309 * 310 * Returns 0 if the node is already read-only or it could switch the node to 311 * read-only because BDRV_O_AUTO_RDONLY is set. 312 * 313 * Returns -EACCES if the node is read-write and BDRV_O_AUTO_RDONLY is not set 314 * or bdrv_can_set_read_only() forbids making the node read-only. If @errmsg 315 * is not NULL, it is used as the error message for the Error object. 316 */ 317 int bdrv_apply_auto_read_only(BlockDriverState *bs, const char *errmsg, 318 Error **errp) 319 { 320 int ret = 0; 321 IO_CODE(); 322 323 if (!(bs->open_flags & BDRV_O_RDWR)) { 324 return 0; 325 } 326 if (!(bs->open_flags & BDRV_O_AUTO_RDONLY)) { 327 goto fail; 328 } 329 330 ret = bdrv_can_set_read_only(bs, true, false, NULL); 331 if (ret < 0) { 332 goto fail; 333 } 334 335 bs->open_flags &= ~BDRV_O_RDWR; 336 337 return 0; 338 339 fail: 340 error_setg(errp, "%s", errmsg ?: "Image is read-only"); 341 return -EACCES; 342 } 343 344 /* 345 * If @backing is empty, this function returns NULL without setting 346 * @errp. In all other cases, NULL will only be returned with @errp 347 * set. 348 * 349 * Therefore, a return value of NULL without @errp set means that 350 * there is no backing file; if @errp is set, there is one but its 351 * absolute filename cannot be generated. 352 */ 353 char *bdrv_get_full_backing_filename_from_filename(const char *backed, 354 const char *backing, 355 Error **errp) 356 { 357 if (backing[0] == '\0') { 358 return NULL; 359 } else if (path_has_protocol(backing) || path_is_absolute(backing)) { 360 return g_strdup(backing); 361 } else if (backed[0] == '\0' || strstart(backed, "json:", NULL)) { 362 error_setg(errp, "Cannot use relative backing file names for '%s'", 363 backed); 364 return NULL; 365 } else { 366 return path_combine(backed, backing); 367 } 368 } 369 370 /* 371 * If @filename is empty or NULL, this function returns NULL without 372 * setting @errp. In all other cases, NULL will only be returned with 373 * @errp set. 374 */ 375 static char * GRAPH_RDLOCK 376 bdrv_make_absolute_filename(BlockDriverState *relative_to, 377 const char *filename, Error **errp) 378 { 379 char *dir, *full_name; 380 381 if (!filename || filename[0] == '\0') { 382 return NULL; 383 } else if (path_has_protocol(filename) || path_is_absolute(filename)) { 384 return g_strdup(filename); 385 } 386 387 dir = bdrv_dirname(relative_to, errp); 388 if (!dir) { 389 return NULL; 390 } 391 392 full_name = g_strconcat(dir, filename, NULL); 393 g_free(dir); 394 return full_name; 395 } 396 397 char *bdrv_get_full_backing_filename(BlockDriverState *bs, Error **errp) 398 { 399 GLOBAL_STATE_CODE(); 400 return bdrv_make_absolute_filename(bs, bs->backing_file, errp); 401 } 402 403 void bdrv_register(BlockDriver *bdrv) 404 { 405 assert(bdrv->format_name); 406 GLOBAL_STATE_CODE(); 407 QLIST_INSERT_HEAD(&bdrv_drivers, bdrv, list); 408 } 409 410 BlockDriverState *bdrv_new(void) 411 { 412 BlockDriverState *bs; 413 int i; 414 415 GLOBAL_STATE_CODE(); 416 417 bs = g_new0(BlockDriverState, 1); 418 QLIST_INIT(&bs->dirty_bitmaps); 419 for (i = 0; i < BLOCK_OP_TYPE_MAX; i++) { 420 QLIST_INIT(&bs->op_blockers[i]); 421 } 422 qemu_mutex_init(&bs->reqs_lock); 423 qemu_mutex_init(&bs->dirty_bitmap_mutex); 424 bs->refcnt = 1; 425 bs->aio_context = qemu_get_aio_context(); 426 427 qemu_co_queue_init(&bs->flush_queue); 428 429 qemu_co_mutex_init(&bs->bsc_modify_lock); 430 bs->block_status_cache = g_new0(BdrvBlockStatusCache, 1); 431 432 for (i = 0; i < bdrv_drain_all_count; i++) { 433 bdrv_drained_begin(bs); 434 } 435 436 QTAILQ_INSERT_TAIL(&all_bdrv_states, bs, bs_list); 437 438 return bs; 439 } 440 441 static BlockDriver *bdrv_do_find_format(const char *format_name) 442 { 443 BlockDriver *drv1; 444 GLOBAL_STATE_CODE(); 445 446 QLIST_FOREACH(drv1, &bdrv_drivers, list) { 447 if (!strcmp(drv1->format_name, format_name)) { 448 return drv1; 449 } 450 } 451 452 return NULL; 453 } 454 455 BlockDriver *bdrv_find_format(const char *format_name) 456 { 457 BlockDriver *drv1; 458 int i; 459 460 GLOBAL_STATE_CODE(); 461 462 drv1 = bdrv_do_find_format(format_name); 463 if (drv1) { 464 return drv1; 465 } 466 467 /* The driver isn't registered, maybe we need to load a module */ 468 for (i = 0; i < (int)ARRAY_SIZE(block_driver_modules); ++i) { 469 if (!strcmp(block_driver_modules[i].format_name, format_name)) { 470 Error *local_err = NULL; 471 int rv = block_module_load(block_driver_modules[i].library_name, 472 &local_err); 473 if (rv > 0) { 474 return bdrv_do_find_format(format_name); 475 } else if (rv < 0) { 476 error_report_err(local_err); 477 } 478 break; 479 } 480 } 481 return NULL; 482 } 483 484 static int bdrv_format_is_whitelisted(const char *format_name, bool read_only) 485 { 486 static const char *whitelist_rw[] = { 487 CONFIG_BDRV_RW_WHITELIST 488 NULL 489 }; 490 static const char *whitelist_ro[] = { 491 CONFIG_BDRV_RO_WHITELIST 492 NULL 493 }; 494 const char **p; 495 496 if (!whitelist_rw[0] && !whitelist_ro[0]) { 497 return 1; /* no whitelist, anything goes */ 498 } 499 500 for (p = whitelist_rw; *p; p++) { 501 if (!strcmp(format_name, *p)) { 502 return 1; 503 } 504 } 505 if (read_only) { 506 for (p = whitelist_ro; *p; p++) { 507 if (!strcmp(format_name, *p)) { 508 return 1; 509 } 510 } 511 } 512 return 0; 513 } 514 515 int bdrv_is_whitelisted(BlockDriver *drv, bool read_only) 516 { 517 GLOBAL_STATE_CODE(); 518 return bdrv_format_is_whitelisted(drv->format_name, read_only); 519 } 520 521 bool bdrv_uses_whitelist(void) 522 { 523 return use_bdrv_whitelist; 524 } 525 526 typedef struct CreateCo { 527 BlockDriver *drv; 528 char *filename; 529 QemuOpts *opts; 530 int ret; 531 Error *err; 532 } CreateCo; 533 534 int coroutine_fn bdrv_co_create(BlockDriver *drv, const char *filename, 535 QemuOpts *opts, Error **errp) 536 { 537 int ret; 538 GLOBAL_STATE_CODE(); 539 ERRP_GUARD(); 540 541 if (!drv->bdrv_co_create_opts) { 542 error_setg(errp, "Driver '%s' does not support image creation", 543 drv->format_name); 544 return -ENOTSUP; 545 } 546 547 ret = drv->bdrv_co_create_opts(drv, filename, opts, errp); 548 if (ret < 0 && !*errp) { 549 error_setg_errno(errp, -ret, "Could not create image"); 550 } 551 552 return ret; 553 } 554 555 /** 556 * Helper function for bdrv_create_file_fallback(): Resize @blk to at 557 * least the given @minimum_size. 558 * 559 * On success, return @blk's actual length. 560 * Otherwise, return -errno. 561 */ 562 static int64_t coroutine_fn GRAPH_UNLOCKED 563 create_file_fallback_truncate(BlockBackend *blk, int64_t minimum_size, 564 Error **errp) 565 { 566 Error *local_err = NULL; 567 int64_t size; 568 int ret; 569 570 GLOBAL_STATE_CODE(); 571 572 ret = blk_co_truncate(blk, minimum_size, false, PREALLOC_MODE_OFF, 0, 573 &local_err); 574 if (ret < 0 && ret != -ENOTSUP) { 575 error_propagate(errp, local_err); 576 return ret; 577 } 578 579 size = blk_co_getlength(blk); 580 if (size < 0) { 581 error_free(local_err); 582 error_setg_errno(errp, -size, 583 "Failed to inquire the new image file's length"); 584 return size; 585 } 586 587 if (size < minimum_size) { 588 /* Need to grow the image, but we failed to do that */ 589 error_propagate(errp, local_err); 590 return -ENOTSUP; 591 } 592 593 error_free(local_err); 594 local_err = NULL; 595 596 return size; 597 } 598 599 /** 600 * Helper function for bdrv_create_file_fallback(): Zero the first 601 * sector to remove any potentially pre-existing image header. 602 */ 603 static int coroutine_fn 604 create_file_fallback_zero_first_sector(BlockBackend *blk, 605 int64_t current_size, 606 Error **errp) 607 { 608 int64_t bytes_to_clear; 609 int ret; 610 611 GLOBAL_STATE_CODE(); 612 613 bytes_to_clear = MIN(current_size, BDRV_SECTOR_SIZE); 614 if (bytes_to_clear) { 615 ret = blk_co_pwrite_zeroes(blk, 0, bytes_to_clear, BDRV_REQ_MAY_UNMAP); 616 if (ret < 0) { 617 error_setg_errno(errp, -ret, 618 "Failed to clear the new image's first sector"); 619 return ret; 620 } 621 } 622 623 return 0; 624 } 625 626 /** 627 * Simple implementation of bdrv_co_create_opts for protocol drivers 628 * which only support creation via opening a file 629 * (usually existing raw storage device) 630 */ 631 int coroutine_fn bdrv_co_create_opts_simple(BlockDriver *drv, 632 const char *filename, 633 QemuOpts *opts, 634 Error **errp) 635 { 636 BlockBackend *blk; 637 QDict *options; 638 int64_t size = 0; 639 char *buf = NULL; 640 PreallocMode prealloc; 641 Error *local_err = NULL; 642 int ret; 643 644 GLOBAL_STATE_CODE(); 645 646 size = qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0); 647 buf = qemu_opt_get_del(opts, BLOCK_OPT_PREALLOC); 648 prealloc = qapi_enum_parse(&PreallocMode_lookup, buf, 649 PREALLOC_MODE_OFF, &local_err); 650 g_free(buf); 651 if (local_err) { 652 error_propagate(errp, local_err); 653 return -EINVAL; 654 } 655 656 if (prealloc != PREALLOC_MODE_OFF) { 657 error_setg(errp, "Unsupported preallocation mode '%s'", 658 PreallocMode_str(prealloc)); 659 return -ENOTSUP; 660 } 661 662 options = qdict_new(); 663 qdict_put_str(options, "driver", drv->format_name); 664 665 blk = blk_co_new_open(filename, NULL, options, 666 BDRV_O_RDWR | BDRV_O_RESIZE, errp); 667 if (!blk) { 668 error_prepend(errp, "Protocol driver '%s' does not support creating " 669 "new images, so an existing image must be selected as " 670 "the target; however, opening the given target as an " 671 "existing image failed: ", 672 drv->format_name); 673 return -EINVAL; 674 } 675 676 size = create_file_fallback_truncate(blk, size, errp); 677 if (size < 0) { 678 ret = size; 679 goto out; 680 } 681 682 ret = create_file_fallback_zero_first_sector(blk, size, errp); 683 if (ret < 0) { 684 goto out; 685 } 686 687 ret = 0; 688 out: 689 blk_co_unref(blk); 690 return ret; 691 } 692 693 int coroutine_fn bdrv_co_create_file(const char *filename, QemuOpts *opts, 694 Error **errp) 695 { 696 QemuOpts *protocol_opts; 697 BlockDriver *drv; 698 QDict *qdict; 699 int ret; 700 701 GLOBAL_STATE_CODE(); 702 703 drv = bdrv_find_protocol(filename, true, errp); 704 if (drv == NULL) { 705 return -ENOENT; 706 } 707 708 if (!drv->create_opts) { 709 error_setg(errp, "Driver '%s' does not support image creation", 710 drv->format_name); 711 return -ENOTSUP; 712 } 713 714 /* 715 * 'opts' contains a QemuOptsList with a combination of format and protocol 716 * default values. 717 * 718 * The format properly removes its options, but the default values remain 719 * in 'opts->list'. So if the protocol has options with the same name 720 * (e.g. rbd has 'cluster_size' as qcow2), it will see the default values 721 * of the format, since for overlapping options, the format wins. 722 * 723 * To avoid this issue, lets convert QemuOpts to QDict, in this way we take 724 * only the set options, and then convert it back to QemuOpts, using the 725 * create_opts of the protocol. So the new QemuOpts, will contain only the 726 * protocol defaults. 727 */ 728 qdict = qemu_opts_to_qdict(opts, NULL); 729 protocol_opts = qemu_opts_from_qdict(drv->create_opts, qdict, errp); 730 if (protocol_opts == NULL) { 731 ret = -EINVAL; 732 goto out; 733 } 734 735 ret = bdrv_co_create(drv, filename, protocol_opts, errp); 736 out: 737 qemu_opts_del(protocol_opts); 738 qobject_unref(qdict); 739 return ret; 740 } 741 742 int coroutine_fn bdrv_co_delete_file(BlockDriverState *bs, Error **errp) 743 { 744 Error *local_err = NULL; 745 int ret; 746 747 IO_CODE(); 748 assert(bs != NULL); 749 assert_bdrv_graph_readable(); 750 751 if (!bs->drv) { 752 error_setg(errp, "Block node '%s' is not opened", bs->filename); 753 return -ENOMEDIUM; 754 } 755 756 if (!bs->drv->bdrv_co_delete_file) { 757 error_setg(errp, "Driver '%s' does not support image deletion", 758 bs->drv->format_name); 759 return -ENOTSUP; 760 } 761 762 ret = bs->drv->bdrv_co_delete_file(bs, &local_err); 763 if (ret < 0) { 764 error_propagate(errp, local_err); 765 } 766 767 return ret; 768 } 769 770 void coroutine_fn bdrv_co_delete_file_noerr(BlockDriverState *bs) 771 { 772 Error *local_err = NULL; 773 int ret; 774 IO_CODE(); 775 776 if (!bs) { 777 return; 778 } 779 780 ret = bdrv_co_delete_file(bs, &local_err); 781 /* 782 * ENOTSUP will happen if the block driver doesn't support 783 * the 'bdrv_co_delete_file' interface. This is a predictable 784 * scenario and shouldn't be reported back to the user. 785 */ 786 if (ret == -ENOTSUP) { 787 error_free(local_err); 788 } else if (ret < 0) { 789 error_report_err(local_err); 790 } 791 } 792 793 /** 794 * Try to get @bs's logical and physical block size. 795 * On success, store them in @bsz struct and return 0. 796 * On failure return -errno. 797 * @bs must not be empty. 798 */ 799 int bdrv_probe_blocksizes(BlockDriverState *bs, BlockSizes *bsz) 800 { 801 BlockDriver *drv = bs->drv; 802 BlockDriverState *filtered = bdrv_filter_bs(bs); 803 GLOBAL_STATE_CODE(); 804 805 if (drv && drv->bdrv_probe_blocksizes) { 806 return drv->bdrv_probe_blocksizes(bs, bsz); 807 } else if (filtered) { 808 return bdrv_probe_blocksizes(filtered, bsz); 809 } 810 811 return -ENOTSUP; 812 } 813 814 /** 815 * Try to get @bs's geometry (cyls, heads, sectors). 816 * On success, store them in @geo struct and return 0. 817 * On failure return -errno. 818 * @bs must not be empty. 819 */ 820 int bdrv_probe_geometry(BlockDriverState *bs, HDGeometry *geo) 821 { 822 BlockDriver *drv = bs->drv; 823 BlockDriverState *filtered; 824 825 GLOBAL_STATE_CODE(); 826 GRAPH_RDLOCK_GUARD_MAINLOOP(); 827 828 if (drv && drv->bdrv_probe_geometry) { 829 return drv->bdrv_probe_geometry(bs, geo); 830 } 831 832 filtered = bdrv_filter_bs(bs); 833 if (filtered) { 834 return bdrv_probe_geometry(filtered, geo); 835 } 836 837 return -ENOTSUP; 838 } 839 840 /* 841 * Create a uniquely-named empty temporary file. 842 * Return the actual file name used upon success, otherwise NULL. 843 * This string should be freed with g_free() when not needed any longer. 844 * 845 * Note: creating a temporary file for the caller to (re)open is 846 * inherently racy. Use g_file_open_tmp() instead whenever practical. 847 */ 848 char *create_tmp_file(Error **errp) 849 { 850 int fd; 851 const char *tmpdir; 852 g_autofree char *filename = NULL; 853 854 tmpdir = g_get_tmp_dir(); 855 #ifndef _WIN32 856 /* 857 * See commit 69bef79 ("block: use /var/tmp instead of /tmp for -snapshot") 858 * 859 * This function is used to create temporary disk images (like -snapshot), 860 * so the files can become very large. /tmp is often a tmpfs where as 861 * /var/tmp is usually on a disk, so more appropriate for disk images. 862 */ 863 if (!g_strcmp0(tmpdir, "/tmp")) { 864 tmpdir = "/var/tmp"; 865 } 866 #endif 867 868 filename = g_strdup_printf("%s/vl.XXXXXX", tmpdir); 869 fd = g_mkstemp(filename); 870 if (fd < 0) { 871 error_setg_errno(errp, errno, "Could not open temporary file '%s'", 872 filename); 873 return NULL; 874 } 875 close(fd); 876 877 return g_steal_pointer(&filename); 878 } 879 880 /* 881 * Detect host devices. By convention, /dev/cdrom[N] is always 882 * recognized as a host CDROM. 883 */ 884 static BlockDriver *find_hdev_driver(const char *filename) 885 { 886 int score_max = 0, score; 887 BlockDriver *drv = NULL, *d; 888 GLOBAL_STATE_CODE(); 889 890 QLIST_FOREACH(d, &bdrv_drivers, list) { 891 if (d->bdrv_probe_device) { 892 score = d->bdrv_probe_device(filename); 893 if (score > score_max) { 894 score_max = score; 895 drv = d; 896 } 897 } 898 } 899 900 return drv; 901 } 902 903 static BlockDriver *bdrv_do_find_protocol(const char *protocol) 904 { 905 BlockDriver *drv1; 906 GLOBAL_STATE_CODE(); 907 908 QLIST_FOREACH(drv1, &bdrv_drivers, list) { 909 if (drv1->protocol_name && !strcmp(drv1->protocol_name, protocol)) { 910 return drv1; 911 } 912 } 913 914 return NULL; 915 } 916 917 BlockDriver *bdrv_find_protocol(const char *filename, 918 bool allow_protocol_prefix, 919 Error **errp) 920 { 921 BlockDriver *drv1; 922 char protocol[128]; 923 int len; 924 const char *p; 925 int i; 926 927 GLOBAL_STATE_CODE(); 928 /* TODO Drivers without bdrv_file_open must be specified explicitly */ 929 930 /* 931 * XXX(hch): we really should not let host device detection 932 * override an explicit protocol specification, but moving this 933 * later breaks access to device names with colons in them. 934 * Thanks to the brain-dead persistent naming schemes on udev- 935 * based Linux systems those actually are quite common. 936 */ 937 drv1 = find_hdev_driver(filename); 938 if (drv1) { 939 return drv1; 940 } 941 942 if (!path_has_protocol(filename) || !allow_protocol_prefix) { 943 return &bdrv_file; 944 } 945 946 p = strchr(filename, ':'); 947 assert(p != NULL); 948 len = p - filename; 949 if (len > sizeof(protocol) - 1) 950 len = sizeof(protocol) - 1; 951 memcpy(protocol, filename, len); 952 protocol[len] = '\0'; 953 954 drv1 = bdrv_do_find_protocol(protocol); 955 if (drv1) { 956 return drv1; 957 } 958 959 for (i = 0; i < (int)ARRAY_SIZE(block_driver_modules); ++i) { 960 if (block_driver_modules[i].protocol_name && 961 !strcmp(block_driver_modules[i].protocol_name, protocol)) { 962 int rv = block_module_load(block_driver_modules[i].library_name, errp); 963 if (rv > 0) { 964 drv1 = bdrv_do_find_protocol(protocol); 965 } else if (rv < 0) { 966 return NULL; 967 } 968 break; 969 } 970 } 971 972 if (!drv1) { 973 error_setg(errp, "Unknown protocol '%s'", protocol); 974 } 975 return drv1; 976 } 977 978 /* 979 * Guess image format by probing its contents. 980 * This is not a good idea when your image is raw (CVE-2008-2004), but 981 * we do it anyway for backward compatibility. 982 * 983 * @buf contains the image's first @buf_size bytes. 984 * @buf_size is the buffer size in bytes (generally BLOCK_PROBE_BUF_SIZE, 985 * but can be smaller if the image file is smaller) 986 * @filename is its filename. 987 * 988 * For all block drivers, call the bdrv_probe() method to get its 989 * probing score. 990 * Return the first block driver with the highest probing score. 991 */ 992 BlockDriver *bdrv_probe_all(const uint8_t *buf, int buf_size, 993 const char *filename) 994 { 995 int score_max = 0, score; 996 BlockDriver *drv = NULL, *d; 997 IO_CODE(); 998 999 QLIST_FOREACH(d, &bdrv_drivers, list) { 1000 if (d->bdrv_probe) { 1001 score = d->bdrv_probe(buf, buf_size, filename); 1002 if (score > score_max) { 1003 score_max = score; 1004 drv = d; 1005 } 1006 } 1007 } 1008 1009 return drv; 1010 } 1011 1012 static int find_image_format(BlockBackend *file, const char *filename, 1013 BlockDriver **pdrv, Error **errp) 1014 { 1015 BlockDriver *drv; 1016 uint8_t buf[BLOCK_PROBE_BUF_SIZE]; 1017 int ret = 0; 1018 1019 GLOBAL_STATE_CODE(); 1020 1021 /* Return the raw BlockDriver * to scsi-generic devices or empty drives */ 1022 if (blk_is_sg(file) || !blk_is_inserted(file) || blk_getlength(file) == 0) { 1023 *pdrv = &bdrv_raw; 1024 return ret; 1025 } 1026 1027 ret = blk_pread(file, 0, sizeof(buf), buf, 0); 1028 if (ret < 0) { 1029 error_setg_errno(errp, -ret, "Could not read image for determining its " 1030 "format"); 1031 *pdrv = NULL; 1032 return ret; 1033 } 1034 1035 drv = bdrv_probe_all(buf, sizeof(buf), filename); 1036 if (!drv) { 1037 error_setg(errp, "Could not determine image format: No compatible " 1038 "driver found"); 1039 *pdrv = NULL; 1040 return -ENOENT; 1041 } 1042 1043 *pdrv = drv; 1044 return 0; 1045 } 1046 1047 /** 1048 * Set the current 'total_sectors' value 1049 * Return 0 on success, -errno on error. 1050 */ 1051 int coroutine_fn bdrv_co_refresh_total_sectors(BlockDriverState *bs, 1052 int64_t hint) 1053 { 1054 BlockDriver *drv = bs->drv; 1055 IO_CODE(); 1056 assert_bdrv_graph_readable(); 1057 1058 if (!drv) { 1059 return -ENOMEDIUM; 1060 } 1061 1062 /* Do not attempt drv->bdrv_co_getlength() on scsi-generic devices */ 1063 if (bdrv_is_sg(bs)) 1064 return 0; 1065 1066 /* query actual device if possible, otherwise just trust the hint */ 1067 if (drv->bdrv_co_getlength) { 1068 int64_t length = drv->bdrv_co_getlength(bs); 1069 if (length < 0) { 1070 return length; 1071 } 1072 hint = DIV_ROUND_UP(length, BDRV_SECTOR_SIZE); 1073 } 1074 1075 bs->total_sectors = hint; 1076 1077 if (bs->total_sectors * BDRV_SECTOR_SIZE > BDRV_MAX_LENGTH) { 1078 return -EFBIG; 1079 } 1080 1081 return 0; 1082 } 1083 1084 /** 1085 * Combines a QDict of new block driver @options with any missing options taken 1086 * from @old_options, so that leaving out an option defaults to its old value. 1087 */ 1088 static void bdrv_join_options(BlockDriverState *bs, QDict *options, 1089 QDict *old_options) 1090 { 1091 GLOBAL_STATE_CODE(); 1092 if (bs->drv && bs->drv->bdrv_join_options) { 1093 bs->drv->bdrv_join_options(options, old_options); 1094 } else { 1095 qdict_join(options, old_options, false); 1096 } 1097 } 1098 1099 static BlockdevDetectZeroesOptions bdrv_parse_detect_zeroes(QemuOpts *opts, 1100 int open_flags, 1101 Error **errp) 1102 { 1103 Error *local_err = NULL; 1104 char *value = qemu_opt_get_del(opts, "detect-zeroes"); 1105 BlockdevDetectZeroesOptions detect_zeroes = 1106 qapi_enum_parse(&BlockdevDetectZeroesOptions_lookup, value, 1107 BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF, &local_err); 1108 GLOBAL_STATE_CODE(); 1109 g_free(value); 1110 if (local_err) { 1111 error_propagate(errp, local_err); 1112 return detect_zeroes; 1113 } 1114 1115 if (detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP && 1116 !(open_flags & BDRV_O_UNMAP)) 1117 { 1118 error_setg(errp, "setting detect-zeroes to unmap is not allowed " 1119 "without setting discard operation to unmap"); 1120 } 1121 1122 return detect_zeroes; 1123 } 1124 1125 /** 1126 * Set open flags for aio engine 1127 * 1128 * Return 0 on success, -1 if the engine specified is invalid 1129 */ 1130 int bdrv_parse_aio(const char *mode, int *flags) 1131 { 1132 if (!strcmp(mode, "threads")) { 1133 /* do nothing, default */ 1134 } else if (!strcmp(mode, "native")) { 1135 *flags |= BDRV_O_NATIVE_AIO; 1136 #ifdef CONFIG_LINUX_IO_URING 1137 } else if (!strcmp(mode, "io_uring")) { 1138 *flags |= BDRV_O_IO_URING; 1139 #endif 1140 } else { 1141 return -1; 1142 } 1143 1144 return 0; 1145 } 1146 1147 /** 1148 * Set open flags for a given discard mode 1149 * 1150 * Return 0 on success, -1 if the discard mode was invalid. 1151 */ 1152 int bdrv_parse_discard_flags(const char *mode, int *flags) 1153 { 1154 *flags &= ~BDRV_O_UNMAP; 1155 1156 if (!strcmp(mode, "off") || !strcmp(mode, "ignore")) { 1157 /* do nothing */ 1158 } else if (!strcmp(mode, "on") || !strcmp(mode, "unmap")) { 1159 *flags |= BDRV_O_UNMAP; 1160 } else { 1161 return -1; 1162 } 1163 1164 return 0; 1165 } 1166 1167 /** 1168 * Set open flags for a given cache mode 1169 * 1170 * Return 0 on success, -1 if the cache mode was invalid. 1171 */ 1172 int bdrv_parse_cache_mode(const char *mode, int *flags, bool *writethrough) 1173 { 1174 *flags &= ~BDRV_O_CACHE_MASK; 1175 1176 if (!strcmp(mode, "off") || !strcmp(mode, "none")) { 1177 *writethrough = false; 1178 *flags |= BDRV_O_NOCACHE; 1179 } else if (!strcmp(mode, "directsync")) { 1180 *writethrough = true; 1181 *flags |= BDRV_O_NOCACHE; 1182 } else if (!strcmp(mode, "writeback")) { 1183 *writethrough = false; 1184 } else if (!strcmp(mode, "unsafe")) { 1185 *writethrough = false; 1186 *flags |= BDRV_O_NO_FLUSH; 1187 } else if (!strcmp(mode, "writethrough")) { 1188 *writethrough = true; 1189 } else { 1190 return -1; 1191 } 1192 1193 return 0; 1194 } 1195 1196 static char *bdrv_child_get_parent_desc(BdrvChild *c) 1197 { 1198 BlockDriverState *parent = c->opaque; 1199 return g_strdup_printf("node '%s'", bdrv_get_node_name(parent)); 1200 } 1201 1202 static void GRAPH_RDLOCK bdrv_child_cb_drained_begin(BdrvChild *child) 1203 { 1204 BlockDriverState *bs = child->opaque; 1205 bdrv_do_drained_begin_quiesce(bs, NULL); 1206 } 1207 1208 static bool GRAPH_RDLOCK bdrv_child_cb_drained_poll(BdrvChild *child) 1209 { 1210 BlockDriverState *bs = child->opaque; 1211 return bdrv_drain_poll(bs, NULL, false); 1212 } 1213 1214 static void GRAPH_RDLOCK bdrv_child_cb_drained_end(BdrvChild *child) 1215 { 1216 BlockDriverState *bs = child->opaque; 1217 bdrv_drained_end(bs); 1218 } 1219 1220 static int bdrv_child_cb_inactivate(BdrvChild *child) 1221 { 1222 BlockDriverState *bs = child->opaque; 1223 GLOBAL_STATE_CODE(); 1224 assert(bs->open_flags & BDRV_O_INACTIVE); 1225 return 0; 1226 } 1227 1228 static bool bdrv_child_cb_change_aio_ctx(BdrvChild *child, AioContext *ctx, 1229 GHashTable *visited, Transaction *tran, 1230 Error **errp) 1231 { 1232 BlockDriverState *bs = child->opaque; 1233 return bdrv_change_aio_context(bs, ctx, visited, tran, errp); 1234 } 1235 1236 /* 1237 * Returns the options and flags that a temporary snapshot should get, based on 1238 * the originally requested flags (the originally requested image will have 1239 * flags like a backing file) 1240 */ 1241 static void bdrv_temp_snapshot_options(int *child_flags, QDict *child_options, 1242 int parent_flags, QDict *parent_options) 1243 { 1244 GLOBAL_STATE_CODE(); 1245 *child_flags = (parent_flags & ~BDRV_O_SNAPSHOT) | BDRV_O_TEMPORARY; 1246 1247 /* For temporary files, unconditional cache=unsafe is fine */ 1248 qdict_set_default_str(child_options, BDRV_OPT_CACHE_DIRECT, "off"); 1249 qdict_set_default_str(child_options, BDRV_OPT_CACHE_NO_FLUSH, "on"); 1250 1251 /* Copy the read-only and discard options from the parent */ 1252 qdict_copy_default(child_options, parent_options, BDRV_OPT_READ_ONLY); 1253 qdict_copy_default(child_options, parent_options, BDRV_OPT_DISCARD); 1254 1255 /* aio=native doesn't work for cache.direct=off, so disable it for the 1256 * temporary snapshot */ 1257 *child_flags &= ~BDRV_O_NATIVE_AIO; 1258 } 1259 1260 static void GRAPH_WRLOCK bdrv_backing_attach(BdrvChild *c) 1261 { 1262 BlockDriverState *parent = c->opaque; 1263 BlockDriverState *backing_hd = c->bs; 1264 1265 GLOBAL_STATE_CODE(); 1266 assert(!parent->backing_blocker); 1267 error_setg(&parent->backing_blocker, 1268 "node is used as backing hd of '%s'", 1269 bdrv_get_device_or_node_name(parent)); 1270 1271 bdrv_refresh_filename(backing_hd); 1272 1273 parent->open_flags &= ~BDRV_O_NO_BACKING; 1274 1275 bdrv_op_block_all(backing_hd, parent->backing_blocker); 1276 /* Otherwise we won't be able to commit or stream */ 1277 bdrv_op_unblock(backing_hd, BLOCK_OP_TYPE_COMMIT_TARGET, 1278 parent->backing_blocker); 1279 bdrv_op_unblock(backing_hd, BLOCK_OP_TYPE_STREAM, 1280 parent->backing_blocker); 1281 /* 1282 * We do backup in 3 ways: 1283 * 1. drive backup 1284 * The target bs is new opened, and the source is top BDS 1285 * 2. blockdev backup 1286 * Both the source and the target are top BDSes. 1287 * 3. internal backup(used for block replication) 1288 * Both the source and the target are backing file 1289 * 1290 * In case 1 and 2, neither the source nor the target is the backing file. 1291 * In case 3, we will block the top BDS, so there is only one block job 1292 * for the top BDS and its backing chain. 1293 */ 1294 bdrv_op_unblock(backing_hd, BLOCK_OP_TYPE_BACKUP_SOURCE, 1295 parent->backing_blocker); 1296 bdrv_op_unblock(backing_hd, BLOCK_OP_TYPE_BACKUP_TARGET, 1297 parent->backing_blocker); 1298 } 1299 1300 static void bdrv_backing_detach(BdrvChild *c) 1301 { 1302 BlockDriverState *parent = c->opaque; 1303 1304 GLOBAL_STATE_CODE(); 1305 assert(parent->backing_blocker); 1306 bdrv_op_unblock_all(c->bs, parent->backing_blocker); 1307 error_free(parent->backing_blocker); 1308 parent->backing_blocker = NULL; 1309 } 1310 1311 static int bdrv_backing_update_filename(BdrvChild *c, BlockDriverState *base, 1312 const char *filename, Error **errp) 1313 { 1314 BlockDriverState *parent = c->opaque; 1315 bool read_only = bdrv_is_read_only(parent); 1316 int ret; 1317 GLOBAL_STATE_CODE(); 1318 1319 if (read_only) { 1320 ret = bdrv_reopen_set_read_only(parent, false, errp); 1321 if (ret < 0) { 1322 return ret; 1323 } 1324 } 1325 1326 ret = bdrv_change_backing_file(parent, filename, 1327 base->drv ? base->drv->format_name : "", 1328 false); 1329 if (ret < 0) { 1330 error_setg_errno(errp, -ret, "Could not update backing file link"); 1331 } 1332 1333 if (read_only) { 1334 bdrv_reopen_set_read_only(parent, true, NULL); 1335 } 1336 1337 return ret; 1338 } 1339 1340 /* 1341 * Returns the options and flags that a generic child of a BDS should 1342 * get, based on the given options and flags for the parent BDS. 1343 */ 1344 static void bdrv_inherited_options(BdrvChildRole role, bool parent_is_format, 1345 int *child_flags, QDict *child_options, 1346 int parent_flags, QDict *parent_options) 1347 { 1348 int flags = parent_flags; 1349 GLOBAL_STATE_CODE(); 1350 1351 /* 1352 * First, decide whether to set, clear, or leave BDRV_O_PROTOCOL. 1353 * Generally, the question to answer is: Should this child be 1354 * format-probed by default? 1355 */ 1356 1357 /* 1358 * Pure and non-filtered data children of non-format nodes should 1359 * be probed by default (even when the node itself has BDRV_O_PROTOCOL 1360 * set). This only affects a very limited set of drivers (namely 1361 * quorum and blkverify when this comment was written). 1362 * Force-clear BDRV_O_PROTOCOL then. 1363 */ 1364 if (!parent_is_format && 1365 (role & BDRV_CHILD_DATA) && 1366 !(role & (BDRV_CHILD_METADATA | BDRV_CHILD_FILTERED))) 1367 { 1368 flags &= ~BDRV_O_PROTOCOL; 1369 } 1370 1371 /* 1372 * All children of format nodes (except for COW children) and all 1373 * metadata children in general should never be format-probed. 1374 * Force-set BDRV_O_PROTOCOL then. 1375 */ 1376 if ((parent_is_format && !(role & BDRV_CHILD_COW)) || 1377 (role & BDRV_CHILD_METADATA)) 1378 { 1379 flags |= BDRV_O_PROTOCOL; 1380 } 1381 1382 /* 1383 * If the cache mode isn't explicitly set, inherit direct and no-flush from 1384 * the parent. 1385 */ 1386 qdict_copy_default(child_options, parent_options, BDRV_OPT_CACHE_DIRECT); 1387 qdict_copy_default(child_options, parent_options, BDRV_OPT_CACHE_NO_FLUSH); 1388 qdict_copy_default(child_options, parent_options, BDRV_OPT_FORCE_SHARE); 1389 1390 if (role & BDRV_CHILD_COW) { 1391 /* backing files are opened read-only by default */ 1392 qdict_set_default_str(child_options, BDRV_OPT_READ_ONLY, "on"); 1393 qdict_set_default_str(child_options, BDRV_OPT_AUTO_READ_ONLY, "off"); 1394 } else { 1395 /* Inherit the read-only option from the parent if it's not set */ 1396 qdict_copy_default(child_options, parent_options, BDRV_OPT_READ_ONLY); 1397 qdict_copy_default(child_options, parent_options, 1398 BDRV_OPT_AUTO_READ_ONLY); 1399 } 1400 1401 /* 1402 * bdrv_co_pdiscard() respects unmap policy for the parent, so we 1403 * can default to enable it on lower layers regardless of the 1404 * parent option. 1405 */ 1406 qdict_set_default_str(child_options, BDRV_OPT_DISCARD, "unmap"); 1407 1408 /* Clear flags that only apply to the top layer */ 1409 flags &= ~(BDRV_O_SNAPSHOT | BDRV_O_NO_BACKING | BDRV_O_COPY_ON_READ); 1410 1411 if (role & BDRV_CHILD_METADATA) { 1412 flags &= ~BDRV_O_NO_IO; 1413 } 1414 if (role & BDRV_CHILD_COW) { 1415 flags &= ~BDRV_O_TEMPORARY; 1416 } 1417 1418 *child_flags = flags; 1419 } 1420 1421 static void GRAPH_WRLOCK bdrv_child_cb_attach(BdrvChild *child) 1422 { 1423 BlockDriverState *bs = child->opaque; 1424 1425 assert_bdrv_graph_writable(); 1426 QLIST_INSERT_HEAD(&bs->children, child, next); 1427 if (bs->drv->is_filter || (child->role & BDRV_CHILD_FILTERED)) { 1428 /* 1429 * Here we handle filters and block/raw-format.c when it behave like 1430 * filter. They generally have a single PRIMARY child, which is also the 1431 * FILTERED child, and that they may have multiple more children, which 1432 * are neither PRIMARY nor FILTERED. And never we have a COW child here. 1433 * So bs->file will be the PRIMARY child, unless the PRIMARY child goes 1434 * into bs->backing on exceptional cases; and bs->backing will be 1435 * nothing else. 1436 */ 1437 assert(!(child->role & BDRV_CHILD_COW)); 1438 if (child->role & BDRV_CHILD_PRIMARY) { 1439 assert(child->role & BDRV_CHILD_FILTERED); 1440 assert(!bs->backing); 1441 assert(!bs->file); 1442 1443 if (bs->drv->filtered_child_is_backing) { 1444 bs->backing = child; 1445 } else { 1446 bs->file = child; 1447 } 1448 } else { 1449 assert(!(child->role & BDRV_CHILD_FILTERED)); 1450 } 1451 } else if (child->role & BDRV_CHILD_COW) { 1452 assert(bs->drv->supports_backing); 1453 assert(!(child->role & BDRV_CHILD_PRIMARY)); 1454 assert(!bs->backing); 1455 bs->backing = child; 1456 bdrv_backing_attach(child); 1457 } else if (child->role & BDRV_CHILD_PRIMARY) { 1458 assert(!bs->file); 1459 bs->file = child; 1460 } 1461 } 1462 1463 static void GRAPH_WRLOCK bdrv_child_cb_detach(BdrvChild *child) 1464 { 1465 BlockDriverState *bs = child->opaque; 1466 1467 if (child->role & BDRV_CHILD_COW) { 1468 bdrv_backing_detach(child); 1469 } 1470 1471 assert_bdrv_graph_writable(); 1472 QLIST_REMOVE(child, next); 1473 if (child == bs->backing) { 1474 assert(child != bs->file); 1475 bs->backing = NULL; 1476 } else if (child == bs->file) { 1477 bs->file = NULL; 1478 } 1479 } 1480 1481 static int bdrv_child_cb_update_filename(BdrvChild *c, BlockDriverState *base, 1482 const char *filename, Error **errp) 1483 { 1484 if (c->role & BDRV_CHILD_COW) { 1485 return bdrv_backing_update_filename(c, base, filename, errp); 1486 } 1487 return 0; 1488 } 1489 1490 AioContext *child_of_bds_get_parent_aio_context(BdrvChild *c) 1491 { 1492 BlockDriverState *bs = c->opaque; 1493 IO_CODE(); 1494 1495 return bdrv_get_aio_context(bs); 1496 } 1497 1498 const BdrvChildClass child_of_bds = { 1499 .parent_is_bds = true, 1500 .get_parent_desc = bdrv_child_get_parent_desc, 1501 .inherit_options = bdrv_inherited_options, 1502 .drained_begin = bdrv_child_cb_drained_begin, 1503 .drained_poll = bdrv_child_cb_drained_poll, 1504 .drained_end = bdrv_child_cb_drained_end, 1505 .attach = bdrv_child_cb_attach, 1506 .detach = bdrv_child_cb_detach, 1507 .inactivate = bdrv_child_cb_inactivate, 1508 .change_aio_ctx = bdrv_child_cb_change_aio_ctx, 1509 .update_filename = bdrv_child_cb_update_filename, 1510 .get_parent_aio_context = child_of_bds_get_parent_aio_context, 1511 }; 1512 1513 AioContext *bdrv_child_get_parent_aio_context(BdrvChild *c) 1514 { 1515 IO_CODE(); 1516 return c->klass->get_parent_aio_context(c); 1517 } 1518 1519 static int bdrv_open_flags(BlockDriverState *bs, int flags) 1520 { 1521 int open_flags = flags; 1522 GLOBAL_STATE_CODE(); 1523 1524 /* 1525 * Clear flags that are internal to the block layer before opening the 1526 * image. 1527 */ 1528 open_flags &= ~(BDRV_O_SNAPSHOT | BDRV_O_NO_BACKING | BDRV_O_PROTOCOL); 1529 1530 return open_flags; 1531 } 1532 1533 static void update_flags_from_options(int *flags, QemuOpts *opts) 1534 { 1535 GLOBAL_STATE_CODE(); 1536 1537 *flags &= ~(BDRV_O_CACHE_MASK | BDRV_O_RDWR | BDRV_O_AUTO_RDONLY); 1538 1539 if (qemu_opt_get_bool_del(opts, BDRV_OPT_CACHE_NO_FLUSH, false)) { 1540 *flags |= BDRV_O_NO_FLUSH; 1541 } 1542 1543 if (qemu_opt_get_bool_del(opts, BDRV_OPT_CACHE_DIRECT, false)) { 1544 *flags |= BDRV_O_NOCACHE; 1545 } 1546 1547 if (!qemu_opt_get_bool_del(opts, BDRV_OPT_READ_ONLY, false)) { 1548 *flags |= BDRV_O_RDWR; 1549 } 1550 1551 if (qemu_opt_get_bool_del(opts, BDRV_OPT_AUTO_READ_ONLY, false)) { 1552 *flags |= BDRV_O_AUTO_RDONLY; 1553 } 1554 } 1555 1556 static void update_options_from_flags(QDict *options, int flags) 1557 { 1558 GLOBAL_STATE_CODE(); 1559 if (!qdict_haskey(options, BDRV_OPT_CACHE_DIRECT)) { 1560 qdict_put_bool(options, BDRV_OPT_CACHE_DIRECT, flags & BDRV_O_NOCACHE); 1561 } 1562 if (!qdict_haskey(options, BDRV_OPT_CACHE_NO_FLUSH)) { 1563 qdict_put_bool(options, BDRV_OPT_CACHE_NO_FLUSH, 1564 flags & BDRV_O_NO_FLUSH); 1565 } 1566 if (!qdict_haskey(options, BDRV_OPT_READ_ONLY)) { 1567 qdict_put_bool(options, BDRV_OPT_READ_ONLY, !(flags & BDRV_O_RDWR)); 1568 } 1569 if (!qdict_haskey(options, BDRV_OPT_AUTO_READ_ONLY)) { 1570 qdict_put_bool(options, BDRV_OPT_AUTO_READ_ONLY, 1571 flags & BDRV_O_AUTO_RDONLY); 1572 } 1573 } 1574 1575 static void bdrv_assign_node_name(BlockDriverState *bs, 1576 const char *node_name, 1577 Error **errp) 1578 { 1579 char *gen_node_name = NULL; 1580 GLOBAL_STATE_CODE(); 1581 1582 if (!node_name) { 1583 node_name = gen_node_name = id_generate(ID_BLOCK); 1584 } else if (!id_wellformed(node_name)) { 1585 /* 1586 * Check for empty string or invalid characters, but not if it is 1587 * generated (generated names use characters not available to the user) 1588 */ 1589 error_setg(errp, "Invalid node-name: '%s'", node_name); 1590 return; 1591 } 1592 1593 /* takes care of avoiding namespaces collisions */ 1594 if (blk_by_name(node_name)) { 1595 error_setg(errp, "node-name=%s is conflicting with a device id", 1596 node_name); 1597 goto out; 1598 } 1599 1600 /* takes care of avoiding duplicates node names */ 1601 if (bdrv_find_node(node_name)) { 1602 error_setg(errp, "Duplicate nodes with node-name='%s'", node_name); 1603 goto out; 1604 } 1605 1606 /* Make sure that the node name isn't truncated */ 1607 if (strlen(node_name) >= sizeof(bs->node_name)) { 1608 error_setg(errp, "Node name too long"); 1609 goto out; 1610 } 1611 1612 /* copy node name into the bs and insert it into the graph list */ 1613 pstrcpy(bs->node_name, sizeof(bs->node_name), node_name); 1614 QTAILQ_INSERT_TAIL(&graph_bdrv_states, bs, node_list); 1615 out: 1616 g_free(gen_node_name); 1617 } 1618 1619 /* 1620 * The caller must always hold @bs AioContext lock, because this function calls 1621 * bdrv_refresh_total_sectors() which polls when called from non-coroutine 1622 * context. 1623 */ 1624 static int no_coroutine_fn GRAPH_UNLOCKED 1625 bdrv_open_driver(BlockDriverState *bs, BlockDriver *drv, const char *node_name, 1626 QDict *options, int open_flags, Error **errp) 1627 { 1628 AioContext *ctx; 1629 Error *local_err = NULL; 1630 int i, ret; 1631 GLOBAL_STATE_CODE(); 1632 1633 bdrv_assign_node_name(bs, node_name, &local_err); 1634 if (local_err) { 1635 error_propagate(errp, local_err); 1636 return -EINVAL; 1637 } 1638 1639 bs->drv = drv; 1640 bs->opaque = g_malloc0(drv->instance_size); 1641 1642 if (drv->bdrv_file_open) { 1643 assert(!drv->bdrv_needs_filename || bs->filename[0]); 1644 ret = drv->bdrv_file_open(bs, options, open_flags, &local_err); 1645 } else if (drv->bdrv_open) { 1646 ret = drv->bdrv_open(bs, options, open_flags, &local_err); 1647 } else { 1648 ret = 0; 1649 } 1650 1651 if (ret < 0) { 1652 if (local_err) { 1653 error_propagate(errp, local_err); 1654 } else if (bs->filename[0]) { 1655 error_setg_errno(errp, -ret, "Could not open '%s'", bs->filename); 1656 } else { 1657 error_setg_errno(errp, -ret, "Could not open image"); 1658 } 1659 goto open_failed; 1660 } 1661 1662 assert(!(bs->supported_read_flags & ~BDRV_REQ_MASK)); 1663 assert(!(bs->supported_write_flags & ~BDRV_REQ_MASK)); 1664 1665 /* 1666 * Always allow the BDRV_REQ_REGISTERED_BUF optimization hint. This saves 1667 * drivers that pass read/write requests through to a child the trouble of 1668 * declaring support explicitly. 1669 * 1670 * Drivers must not propagate this flag accidentally when they initiate I/O 1671 * to a bounce buffer. That case should be rare though. 1672 */ 1673 bs->supported_read_flags |= BDRV_REQ_REGISTERED_BUF; 1674 bs->supported_write_flags |= BDRV_REQ_REGISTERED_BUF; 1675 1676 /* Get the context after .bdrv_open, it can change the context */ 1677 ctx = bdrv_get_aio_context(bs); 1678 aio_context_acquire(ctx); 1679 1680 ret = bdrv_refresh_total_sectors(bs, bs->total_sectors); 1681 if (ret < 0) { 1682 error_setg_errno(errp, -ret, "Could not refresh total sector count"); 1683 aio_context_release(ctx); 1684 return ret; 1685 } 1686 1687 bdrv_graph_rdlock_main_loop(); 1688 bdrv_refresh_limits(bs, NULL, &local_err); 1689 bdrv_graph_rdunlock_main_loop(); 1690 aio_context_release(ctx); 1691 1692 if (local_err) { 1693 error_propagate(errp, local_err); 1694 return -EINVAL; 1695 } 1696 1697 assert(bdrv_opt_mem_align(bs) != 0); 1698 assert(bdrv_min_mem_align(bs) != 0); 1699 assert(is_power_of_2(bs->bl.request_alignment)); 1700 1701 for (i = 0; i < bs->quiesce_counter; i++) { 1702 if (drv->bdrv_drain_begin) { 1703 drv->bdrv_drain_begin(bs); 1704 } 1705 } 1706 1707 return 0; 1708 open_failed: 1709 bs->drv = NULL; 1710 if (bs->file != NULL) { 1711 bdrv_graph_wrlock(NULL); 1712 bdrv_unref_child(bs, bs->file); 1713 bdrv_graph_wrunlock(); 1714 assert(!bs->file); 1715 } 1716 g_free(bs->opaque); 1717 bs->opaque = NULL; 1718 return ret; 1719 } 1720 1721 /* 1722 * Create and open a block node. 1723 * 1724 * @options is a QDict of options to pass to the block drivers, or NULL for an 1725 * empty set of options. The reference to the QDict belongs to the block layer 1726 * after the call (even on failure), so if the caller intends to reuse the 1727 * dictionary, it needs to use qobject_ref() before calling bdrv_open. 1728 */ 1729 BlockDriverState *bdrv_new_open_driver_opts(BlockDriver *drv, 1730 const char *node_name, 1731 QDict *options, int flags, 1732 Error **errp) 1733 { 1734 BlockDriverState *bs; 1735 int ret; 1736 1737 GLOBAL_STATE_CODE(); 1738 1739 bs = bdrv_new(); 1740 bs->open_flags = flags; 1741 bs->options = options ?: qdict_new(); 1742 bs->explicit_options = qdict_clone_shallow(bs->options); 1743 bs->opaque = NULL; 1744 1745 update_options_from_flags(bs->options, flags); 1746 1747 ret = bdrv_open_driver(bs, drv, node_name, bs->options, flags, errp); 1748 if (ret < 0) { 1749 qobject_unref(bs->explicit_options); 1750 bs->explicit_options = NULL; 1751 qobject_unref(bs->options); 1752 bs->options = NULL; 1753 bdrv_unref(bs); 1754 return NULL; 1755 } 1756 1757 return bs; 1758 } 1759 1760 /* Create and open a block node. */ 1761 BlockDriverState *bdrv_new_open_driver(BlockDriver *drv, const char *node_name, 1762 int flags, Error **errp) 1763 { 1764 GLOBAL_STATE_CODE(); 1765 return bdrv_new_open_driver_opts(drv, node_name, NULL, flags, errp); 1766 } 1767 1768 QemuOptsList bdrv_runtime_opts = { 1769 .name = "bdrv_common", 1770 .head = QTAILQ_HEAD_INITIALIZER(bdrv_runtime_opts.head), 1771 .desc = { 1772 { 1773 .name = "node-name", 1774 .type = QEMU_OPT_STRING, 1775 .help = "Node name of the block device node", 1776 }, 1777 { 1778 .name = "driver", 1779 .type = QEMU_OPT_STRING, 1780 .help = "Block driver to use for the node", 1781 }, 1782 { 1783 .name = BDRV_OPT_CACHE_DIRECT, 1784 .type = QEMU_OPT_BOOL, 1785 .help = "Bypass software writeback cache on the host", 1786 }, 1787 { 1788 .name = BDRV_OPT_CACHE_NO_FLUSH, 1789 .type = QEMU_OPT_BOOL, 1790 .help = "Ignore flush requests", 1791 }, 1792 { 1793 .name = BDRV_OPT_READ_ONLY, 1794 .type = QEMU_OPT_BOOL, 1795 .help = "Node is opened in read-only mode", 1796 }, 1797 { 1798 .name = BDRV_OPT_AUTO_READ_ONLY, 1799 .type = QEMU_OPT_BOOL, 1800 .help = "Node can become read-only if opening read-write fails", 1801 }, 1802 { 1803 .name = "detect-zeroes", 1804 .type = QEMU_OPT_STRING, 1805 .help = "try to optimize zero writes (off, on, unmap)", 1806 }, 1807 { 1808 .name = BDRV_OPT_DISCARD, 1809 .type = QEMU_OPT_STRING, 1810 .help = "discard operation (ignore/off, unmap/on)", 1811 }, 1812 { 1813 .name = BDRV_OPT_FORCE_SHARE, 1814 .type = QEMU_OPT_BOOL, 1815 .help = "always accept other writers (default: off)", 1816 }, 1817 { /* end of list */ } 1818 }, 1819 }; 1820 1821 QemuOptsList bdrv_create_opts_simple = { 1822 .name = "simple-create-opts", 1823 .head = QTAILQ_HEAD_INITIALIZER(bdrv_create_opts_simple.head), 1824 .desc = { 1825 { 1826 .name = BLOCK_OPT_SIZE, 1827 .type = QEMU_OPT_SIZE, 1828 .help = "Virtual disk size" 1829 }, 1830 { 1831 .name = BLOCK_OPT_PREALLOC, 1832 .type = QEMU_OPT_STRING, 1833 .help = "Preallocation mode (allowed values: off)" 1834 }, 1835 { /* end of list */ } 1836 } 1837 }; 1838 1839 /* 1840 * Common part for opening disk images and files 1841 * 1842 * Removes all processed options from *options. 1843 */ 1844 static int bdrv_open_common(BlockDriverState *bs, BlockBackend *file, 1845 QDict *options, Error **errp) 1846 { 1847 int ret, open_flags; 1848 const char *filename; 1849 const char *driver_name = NULL; 1850 const char *node_name = NULL; 1851 const char *discard; 1852 QemuOpts *opts; 1853 BlockDriver *drv; 1854 Error *local_err = NULL; 1855 bool ro; 1856 1857 assert(bs->file == NULL); 1858 assert(options != NULL && bs->options != options); 1859 GLOBAL_STATE_CODE(); 1860 1861 opts = qemu_opts_create(&bdrv_runtime_opts, NULL, 0, &error_abort); 1862 if (!qemu_opts_absorb_qdict(opts, options, errp)) { 1863 ret = -EINVAL; 1864 goto fail_opts; 1865 } 1866 1867 update_flags_from_options(&bs->open_flags, opts); 1868 1869 driver_name = qemu_opt_get(opts, "driver"); 1870 drv = bdrv_find_format(driver_name); 1871 assert(drv != NULL); 1872 1873 bs->force_share = qemu_opt_get_bool(opts, BDRV_OPT_FORCE_SHARE, false); 1874 1875 if (bs->force_share && (bs->open_flags & BDRV_O_RDWR)) { 1876 error_setg(errp, 1877 BDRV_OPT_FORCE_SHARE 1878 "=on can only be used with read-only images"); 1879 ret = -EINVAL; 1880 goto fail_opts; 1881 } 1882 1883 if (file != NULL) { 1884 bdrv_graph_rdlock_main_loop(); 1885 bdrv_refresh_filename(blk_bs(file)); 1886 bdrv_graph_rdunlock_main_loop(); 1887 1888 filename = blk_bs(file)->filename; 1889 } else { 1890 /* 1891 * Caution: while qdict_get_try_str() is fine, getting 1892 * non-string types would require more care. When @options 1893 * come from -blockdev or blockdev_add, its members are typed 1894 * according to the QAPI schema, but when they come from 1895 * -drive, they're all QString. 1896 */ 1897 filename = qdict_get_try_str(options, "filename"); 1898 } 1899 1900 if (drv->bdrv_needs_filename && (!filename || !filename[0])) { 1901 error_setg(errp, "The '%s' block driver requires a file name", 1902 drv->format_name); 1903 ret = -EINVAL; 1904 goto fail_opts; 1905 } 1906 1907 trace_bdrv_open_common(bs, filename ?: "", bs->open_flags, 1908 drv->format_name); 1909 1910 ro = bdrv_is_read_only(bs); 1911 1912 if (use_bdrv_whitelist && !bdrv_is_whitelisted(drv, ro)) { 1913 if (!ro && bdrv_is_whitelisted(drv, true)) { 1914 bdrv_graph_rdlock_main_loop(); 1915 ret = bdrv_apply_auto_read_only(bs, NULL, NULL); 1916 bdrv_graph_rdunlock_main_loop(); 1917 } else { 1918 ret = -ENOTSUP; 1919 } 1920 if (ret < 0) { 1921 error_setg(errp, 1922 !ro && bdrv_is_whitelisted(drv, true) 1923 ? "Driver '%s' can only be used for read-only devices" 1924 : "Driver '%s' is not whitelisted", 1925 drv->format_name); 1926 goto fail_opts; 1927 } 1928 } 1929 1930 /* bdrv_new() and bdrv_close() make it so */ 1931 assert(qatomic_read(&bs->copy_on_read) == 0); 1932 1933 if (bs->open_flags & BDRV_O_COPY_ON_READ) { 1934 if (!ro) { 1935 bdrv_enable_copy_on_read(bs); 1936 } else { 1937 error_setg(errp, "Can't use copy-on-read on read-only device"); 1938 ret = -EINVAL; 1939 goto fail_opts; 1940 } 1941 } 1942 1943 discard = qemu_opt_get(opts, BDRV_OPT_DISCARD); 1944 if (discard != NULL) { 1945 if (bdrv_parse_discard_flags(discard, &bs->open_flags) != 0) { 1946 error_setg(errp, "Invalid discard option"); 1947 ret = -EINVAL; 1948 goto fail_opts; 1949 } 1950 } 1951 1952 bs->detect_zeroes = 1953 bdrv_parse_detect_zeroes(opts, bs->open_flags, &local_err); 1954 if (local_err) { 1955 error_propagate(errp, local_err); 1956 ret = -EINVAL; 1957 goto fail_opts; 1958 } 1959 1960 if (filename != NULL) { 1961 pstrcpy(bs->filename, sizeof(bs->filename), filename); 1962 } else { 1963 bs->filename[0] = '\0'; 1964 } 1965 pstrcpy(bs->exact_filename, sizeof(bs->exact_filename), bs->filename); 1966 1967 /* Open the image, either directly or using a protocol */ 1968 open_flags = bdrv_open_flags(bs, bs->open_flags); 1969 node_name = qemu_opt_get(opts, "node-name"); 1970 1971 assert(!drv->bdrv_file_open || file == NULL); 1972 ret = bdrv_open_driver(bs, drv, node_name, options, open_flags, errp); 1973 if (ret < 0) { 1974 goto fail_opts; 1975 } 1976 1977 qemu_opts_del(opts); 1978 return 0; 1979 1980 fail_opts: 1981 qemu_opts_del(opts); 1982 return ret; 1983 } 1984 1985 static QDict *parse_json_filename(const char *filename, Error **errp) 1986 { 1987 QObject *options_obj; 1988 QDict *options; 1989 int ret; 1990 GLOBAL_STATE_CODE(); 1991 1992 ret = strstart(filename, "json:", &filename); 1993 assert(ret); 1994 1995 options_obj = qobject_from_json(filename, errp); 1996 if (!options_obj) { 1997 error_prepend(errp, "Could not parse the JSON options: "); 1998 return NULL; 1999 } 2000 2001 options = qobject_to(QDict, options_obj); 2002 if (!options) { 2003 qobject_unref(options_obj); 2004 error_setg(errp, "Invalid JSON object given"); 2005 return NULL; 2006 } 2007 2008 qdict_flatten(options); 2009 2010 return options; 2011 } 2012 2013 static void parse_json_protocol(QDict *options, const char **pfilename, 2014 Error **errp) 2015 { 2016 QDict *json_options; 2017 Error *local_err = NULL; 2018 GLOBAL_STATE_CODE(); 2019 2020 /* Parse json: pseudo-protocol */ 2021 if (!*pfilename || !g_str_has_prefix(*pfilename, "json:")) { 2022 return; 2023 } 2024 2025 json_options = parse_json_filename(*pfilename, &local_err); 2026 if (local_err) { 2027 error_propagate(errp, local_err); 2028 return; 2029 } 2030 2031 /* Options given in the filename have lower priority than options 2032 * specified directly */ 2033 qdict_join(options, json_options, false); 2034 qobject_unref(json_options); 2035 *pfilename = NULL; 2036 } 2037 2038 /* 2039 * Fills in default options for opening images and converts the legacy 2040 * filename/flags pair to option QDict entries. 2041 * The BDRV_O_PROTOCOL flag in *flags will be set or cleared accordingly if a 2042 * block driver has been specified explicitly. 2043 */ 2044 static int bdrv_fill_options(QDict **options, const char *filename, 2045 int *flags, Error **errp) 2046 { 2047 const char *drvname; 2048 bool protocol = *flags & BDRV_O_PROTOCOL; 2049 bool parse_filename = false; 2050 BlockDriver *drv = NULL; 2051 Error *local_err = NULL; 2052 2053 GLOBAL_STATE_CODE(); 2054 2055 /* 2056 * Caution: while qdict_get_try_str() is fine, getting non-string 2057 * types would require more care. When @options come from 2058 * -blockdev or blockdev_add, its members are typed according to 2059 * the QAPI schema, but when they come from -drive, they're all 2060 * QString. 2061 */ 2062 drvname = qdict_get_try_str(*options, "driver"); 2063 if (drvname) { 2064 drv = bdrv_find_format(drvname); 2065 if (!drv) { 2066 error_setg(errp, "Unknown driver '%s'", drvname); 2067 return -ENOENT; 2068 } 2069 /* If the user has explicitly specified the driver, this choice should 2070 * override the BDRV_O_PROTOCOL flag */ 2071 protocol = drv->bdrv_file_open; 2072 } 2073 2074 if (protocol) { 2075 *flags |= BDRV_O_PROTOCOL; 2076 } else { 2077 *flags &= ~BDRV_O_PROTOCOL; 2078 } 2079 2080 /* Translate cache options from flags into options */ 2081 update_options_from_flags(*options, *flags); 2082 2083 /* Fetch the file name from the options QDict if necessary */ 2084 if (protocol && filename) { 2085 if (!qdict_haskey(*options, "filename")) { 2086 qdict_put_str(*options, "filename", filename); 2087 parse_filename = true; 2088 } else { 2089 error_setg(errp, "Can't specify 'file' and 'filename' options at " 2090 "the same time"); 2091 return -EINVAL; 2092 } 2093 } 2094 2095 /* Find the right block driver */ 2096 /* See cautionary note on accessing @options above */ 2097 filename = qdict_get_try_str(*options, "filename"); 2098 2099 if (!drvname && protocol) { 2100 if (filename) { 2101 drv = bdrv_find_protocol(filename, parse_filename, errp); 2102 if (!drv) { 2103 return -EINVAL; 2104 } 2105 2106 drvname = drv->format_name; 2107 qdict_put_str(*options, "driver", drvname); 2108 } else { 2109 error_setg(errp, "Must specify either driver or file"); 2110 return -EINVAL; 2111 } 2112 } 2113 2114 assert(drv || !protocol); 2115 2116 /* Driver-specific filename parsing */ 2117 if (drv && drv->bdrv_parse_filename && parse_filename) { 2118 drv->bdrv_parse_filename(filename, *options, &local_err); 2119 if (local_err) { 2120 error_propagate(errp, local_err); 2121 return -EINVAL; 2122 } 2123 2124 if (!drv->bdrv_needs_filename) { 2125 qdict_del(*options, "filename"); 2126 } 2127 } 2128 2129 return 0; 2130 } 2131 2132 typedef struct BlockReopenQueueEntry { 2133 bool prepared; 2134 BDRVReopenState state; 2135 QTAILQ_ENTRY(BlockReopenQueueEntry) entry; 2136 } BlockReopenQueueEntry; 2137 2138 /* 2139 * Return the flags that @bs will have after the reopens in @q have 2140 * successfully completed. If @q is NULL (or @bs is not contained in @q), 2141 * return the current flags. 2142 */ 2143 static int bdrv_reopen_get_flags(BlockReopenQueue *q, BlockDriverState *bs) 2144 { 2145 BlockReopenQueueEntry *entry; 2146 2147 if (q != NULL) { 2148 QTAILQ_FOREACH(entry, q, entry) { 2149 if (entry->state.bs == bs) { 2150 return entry->state.flags; 2151 } 2152 } 2153 } 2154 2155 return bs->open_flags; 2156 } 2157 2158 /* Returns whether the image file can be written to after the reopen queue @q 2159 * has been successfully applied, or right now if @q is NULL. */ 2160 static bool bdrv_is_writable_after_reopen(BlockDriverState *bs, 2161 BlockReopenQueue *q) 2162 { 2163 int flags = bdrv_reopen_get_flags(q, bs); 2164 2165 return (flags & (BDRV_O_RDWR | BDRV_O_INACTIVE)) == BDRV_O_RDWR; 2166 } 2167 2168 /* 2169 * Return whether the BDS can be written to. This is not necessarily 2170 * the same as !bdrv_is_read_only(bs), as inactivated images may not 2171 * be written to but do not count as read-only images. 2172 */ 2173 bool bdrv_is_writable(BlockDriverState *bs) 2174 { 2175 IO_CODE(); 2176 return bdrv_is_writable_after_reopen(bs, NULL); 2177 } 2178 2179 static char *bdrv_child_user_desc(BdrvChild *c) 2180 { 2181 GLOBAL_STATE_CODE(); 2182 return c->klass->get_parent_desc(c); 2183 } 2184 2185 /* 2186 * Check that @a allows everything that @b needs. @a and @b must reference same 2187 * child node. 2188 */ 2189 static bool bdrv_a_allow_b(BdrvChild *a, BdrvChild *b, Error **errp) 2190 { 2191 const char *child_bs_name; 2192 g_autofree char *a_user = NULL; 2193 g_autofree char *b_user = NULL; 2194 g_autofree char *perms = NULL; 2195 2196 assert(a->bs); 2197 assert(a->bs == b->bs); 2198 GLOBAL_STATE_CODE(); 2199 2200 if ((b->perm & a->shared_perm) == b->perm) { 2201 return true; 2202 } 2203 2204 child_bs_name = bdrv_get_node_name(b->bs); 2205 a_user = bdrv_child_user_desc(a); 2206 b_user = bdrv_child_user_desc(b); 2207 perms = bdrv_perm_names(b->perm & ~a->shared_perm); 2208 2209 error_setg(errp, "Permission conflict on node '%s': permissions '%s' are " 2210 "both required by %s (uses node '%s' as '%s' child) and " 2211 "unshared by %s (uses node '%s' as '%s' child).", 2212 child_bs_name, perms, 2213 b_user, child_bs_name, b->name, 2214 a_user, child_bs_name, a->name); 2215 2216 return false; 2217 } 2218 2219 static bool GRAPH_RDLOCK 2220 bdrv_parent_perms_conflict(BlockDriverState *bs, Error **errp) 2221 { 2222 BdrvChild *a, *b; 2223 GLOBAL_STATE_CODE(); 2224 2225 /* 2226 * During the loop we'll look at each pair twice. That's correct because 2227 * bdrv_a_allow_b() is asymmetric and we should check each pair in both 2228 * directions. 2229 */ 2230 QLIST_FOREACH(a, &bs->parents, next_parent) { 2231 QLIST_FOREACH(b, &bs->parents, next_parent) { 2232 if (a == b) { 2233 continue; 2234 } 2235 2236 if (!bdrv_a_allow_b(a, b, errp)) { 2237 return true; 2238 } 2239 } 2240 } 2241 2242 return false; 2243 } 2244 2245 static void GRAPH_RDLOCK 2246 bdrv_child_perm(BlockDriverState *bs, BlockDriverState *child_bs, 2247 BdrvChild *c, BdrvChildRole role, 2248 BlockReopenQueue *reopen_queue, 2249 uint64_t parent_perm, uint64_t parent_shared, 2250 uint64_t *nperm, uint64_t *nshared) 2251 { 2252 assert(bs->drv && bs->drv->bdrv_child_perm); 2253 GLOBAL_STATE_CODE(); 2254 bs->drv->bdrv_child_perm(bs, c, role, reopen_queue, 2255 parent_perm, parent_shared, 2256 nperm, nshared); 2257 /* TODO Take force_share from reopen_queue */ 2258 if (child_bs && child_bs->force_share) { 2259 *nshared = BLK_PERM_ALL; 2260 } 2261 } 2262 2263 /* 2264 * Adds the whole subtree of @bs (including @bs itself) to the @list (except for 2265 * nodes that are already in the @list, of course) so that final list is 2266 * topologically sorted. Return the result (GSList @list object is updated, so 2267 * don't use old reference after function call). 2268 * 2269 * On function start @list must be already topologically sorted and for any node 2270 * in the @list the whole subtree of the node must be in the @list as well. The 2271 * simplest way to satisfy this criteria: use only result of 2272 * bdrv_topological_dfs() or NULL as @list parameter. 2273 */ 2274 static GSList * GRAPH_RDLOCK 2275 bdrv_topological_dfs(GSList *list, GHashTable *found, BlockDriverState *bs) 2276 { 2277 BdrvChild *child; 2278 g_autoptr(GHashTable) local_found = NULL; 2279 2280 GLOBAL_STATE_CODE(); 2281 2282 if (!found) { 2283 assert(!list); 2284 found = local_found = g_hash_table_new(NULL, NULL); 2285 } 2286 2287 if (g_hash_table_contains(found, bs)) { 2288 return list; 2289 } 2290 g_hash_table_add(found, bs); 2291 2292 QLIST_FOREACH(child, &bs->children, next) { 2293 list = bdrv_topological_dfs(list, found, child->bs); 2294 } 2295 2296 return g_slist_prepend(list, bs); 2297 } 2298 2299 typedef struct BdrvChildSetPermState { 2300 BdrvChild *child; 2301 uint64_t old_perm; 2302 uint64_t old_shared_perm; 2303 } BdrvChildSetPermState; 2304 2305 static void bdrv_child_set_perm_abort(void *opaque) 2306 { 2307 BdrvChildSetPermState *s = opaque; 2308 2309 GLOBAL_STATE_CODE(); 2310 2311 s->child->perm = s->old_perm; 2312 s->child->shared_perm = s->old_shared_perm; 2313 } 2314 2315 static TransactionActionDrv bdrv_child_set_pem_drv = { 2316 .abort = bdrv_child_set_perm_abort, 2317 .clean = g_free, 2318 }; 2319 2320 static void bdrv_child_set_perm(BdrvChild *c, uint64_t perm, 2321 uint64_t shared, Transaction *tran) 2322 { 2323 BdrvChildSetPermState *s = g_new(BdrvChildSetPermState, 1); 2324 GLOBAL_STATE_CODE(); 2325 2326 *s = (BdrvChildSetPermState) { 2327 .child = c, 2328 .old_perm = c->perm, 2329 .old_shared_perm = c->shared_perm, 2330 }; 2331 2332 c->perm = perm; 2333 c->shared_perm = shared; 2334 2335 tran_add(tran, &bdrv_child_set_pem_drv, s); 2336 } 2337 2338 static void GRAPH_RDLOCK bdrv_drv_set_perm_commit(void *opaque) 2339 { 2340 BlockDriverState *bs = opaque; 2341 uint64_t cumulative_perms, cumulative_shared_perms; 2342 GLOBAL_STATE_CODE(); 2343 2344 if (bs->drv->bdrv_set_perm) { 2345 bdrv_get_cumulative_perm(bs, &cumulative_perms, 2346 &cumulative_shared_perms); 2347 bs->drv->bdrv_set_perm(bs, cumulative_perms, cumulative_shared_perms); 2348 } 2349 } 2350 2351 static void GRAPH_RDLOCK bdrv_drv_set_perm_abort(void *opaque) 2352 { 2353 BlockDriverState *bs = opaque; 2354 GLOBAL_STATE_CODE(); 2355 2356 if (bs->drv->bdrv_abort_perm_update) { 2357 bs->drv->bdrv_abort_perm_update(bs); 2358 } 2359 } 2360 2361 TransactionActionDrv bdrv_drv_set_perm_drv = { 2362 .abort = bdrv_drv_set_perm_abort, 2363 .commit = bdrv_drv_set_perm_commit, 2364 }; 2365 2366 /* 2367 * After calling this function, the transaction @tran may only be completed 2368 * while holding a reader lock for the graph. 2369 */ 2370 static int GRAPH_RDLOCK 2371 bdrv_drv_set_perm(BlockDriverState *bs, uint64_t perm, uint64_t shared_perm, 2372 Transaction *tran, Error **errp) 2373 { 2374 GLOBAL_STATE_CODE(); 2375 if (!bs->drv) { 2376 return 0; 2377 } 2378 2379 if (bs->drv->bdrv_check_perm) { 2380 int ret = bs->drv->bdrv_check_perm(bs, perm, shared_perm, errp); 2381 if (ret < 0) { 2382 return ret; 2383 } 2384 } 2385 2386 if (tran) { 2387 tran_add(tran, &bdrv_drv_set_perm_drv, bs); 2388 } 2389 2390 return 0; 2391 } 2392 2393 typedef struct BdrvReplaceChildState { 2394 BdrvChild *child; 2395 BlockDriverState *old_bs; 2396 } BdrvReplaceChildState; 2397 2398 static void GRAPH_WRLOCK bdrv_replace_child_commit(void *opaque) 2399 { 2400 BdrvReplaceChildState *s = opaque; 2401 GLOBAL_STATE_CODE(); 2402 2403 bdrv_schedule_unref(s->old_bs); 2404 } 2405 2406 static void GRAPH_WRLOCK bdrv_replace_child_abort(void *opaque) 2407 { 2408 BdrvReplaceChildState *s = opaque; 2409 BlockDriverState *new_bs = s->child->bs; 2410 2411 GLOBAL_STATE_CODE(); 2412 assert_bdrv_graph_writable(); 2413 2414 /* old_bs reference is transparently moved from @s to @s->child */ 2415 if (!s->child->bs) { 2416 /* 2417 * The parents were undrained when removing old_bs from the child. New 2418 * requests can't have been made, though, because the child was empty. 2419 * 2420 * TODO Make bdrv_replace_child_noperm() transactionable to avoid 2421 * undraining the parent in the first place. Once this is done, having 2422 * new_bs drained when calling bdrv_replace_child_tran() is not a 2423 * requirement any more. 2424 */ 2425 bdrv_parent_drained_begin_single(s->child); 2426 assert(!bdrv_parent_drained_poll_single(s->child)); 2427 } 2428 assert(s->child->quiesced_parent); 2429 bdrv_replace_child_noperm(s->child, s->old_bs); 2430 2431 bdrv_unref(new_bs); 2432 } 2433 2434 static TransactionActionDrv bdrv_replace_child_drv = { 2435 .commit = bdrv_replace_child_commit, 2436 .abort = bdrv_replace_child_abort, 2437 .clean = g_free, 2438 }; 2439 2440 /* 2441 * bdrv_replace_child_tran 2442 * 2443 * Note: real unref of old_bs is done only on commit. 2444 * 2445 * Both @child->bs and @new_bs (if non-NULL) must be drained. @new_bs must be 2446 * kept drained until the transaction is completed. 2447 * 2448 * After calling this function, the transaction @tran may only be completed 2449 * while holding a writer lock for the graph. 2450 * 2451 * The function doesn't update permissions, caller is responsible for this. 2452 */ 2453 static void GRAPH_WRLOCK 2454 bdrv_replace_child_tran(BdrvChild *child, BlockDriverState *new_bs, 2455 Transaction *tran) 2456 { 2457 BdrvReplaceChildState *s = g_new(BdrvReplaceChildState, 1); 2458 2459 assert(child->quiesced_parent); 2460 assert(!new_bs || new_bs->quiesce_counter); 2461 2462 *s = (BdrvReplaceChildState) { 2463 .child = child, 2464 .old_bs = child->bs, 2465 }; 2466 tran_add(tran, &bdrv_replace_child_drv, s); 2467 2468 if (new_bs) { 2469 bdrv_ref(new_bs); 2470 } 2471 2472 bdrv_replace_child_noperm(child, new_bs); 2473 /* old_bs reference is transparently moved from @child to @s */ 2474 } 2475 2476 /* 2477 * Refresh permissions in @bs subtree. The function is intended to be called 2478 * after some graph modification that was done without permission update. 2479 * 2480 * After calling this function, the transaction @tran may only be completed 2481 * while holding a reader lock for the graph. 2482 */ 2483 static int GRAPH_RDLOCK 2484 bdrv_node_refresh_perm(BlockDriverState *bs, BlockReopenQueue *q, 2485 Transaction *tran, Error **errp) 2486 { 2487 BlockDriver *drv = bs->drv; 2488 BdrvChild *c; 2489 int ret; 2490 uint64_t cumulative_perms, cumulative_shared_perms; 2491 GLOBAL_STATE_CODE(); 2492 2493 bdrv_get_cumulative_perm(bs, &cumulative_perms, &cumulative_shared_perms); 2494 2495 /* Write permissions never work with read-only images */ 2496 if ((cumulative_perms & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED)) && 2497 !bdrv_is_writable_after_reopen(bs, q)) 2498 { 2499 if (!bdrv_is_writable_after_reopen(bs, NULL)) { 2500 error_setg(errp, "Block node is read-only"); 2501 } else { 2502 error_setg(errp, "Read-only block node '%s' cannot support " 2503 "read-write users", bdrv_get_node_name(bs)); 2504 } 2505 2506 return -EPERM; 2507 } 2508 2509 /* 2510 * Unaligned requests will automatically be aligned to bl.request_alignment 2511 * and without RESIZE we can't extend requests to write to space beyond the 2512 * end of the image, so it's required that the image size is aligned. 2513 */ 2514 if ((cumulative_perms & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED)) && 2515 !(cumulative_perms & BLK_PERM_RESIZE)) 2516 { 2517 if ((bs->total_sectors * BDRV_SECTOR_SIZE) % bs->bl.request_alignment) { 2518 error_setg(errp, "Cannot get 'write' permission without 'resize': " 2519 "Image size is not a multiple of request " 2520 "alignment"); 2521 return -EPERM; 2522 } 2523 } 2524 2525 /* Check this node */ 2526 if (!drv) { 2527 return 0; 2528 } 2529 2530 ret = bdrv_drv_set_perm(bs, cumulative_perms, cumulative_shared_perms, tran, 2531 errp); 2532 if (ret < 0) { 2533 return ret; 2534 } 2535 2536 /* Drivers that never have children can omit .bdrv_child_perm() */ 2537 if (!drv->bdrv_child_perm) { 2538 assert(QLIST_EMPTY(&bs->children)); 2539 return 0; 2540 } 2541 2542 /* Check all children */ 2543 QLIST_FOREACH(c, &bs->children, next) { 2544 uint64_t cur_perm, cur_shared; 2545 2546 bdrv_child_perm(bs, c->bs, c, c->role, q, 2547 cumulative_perms, cumulative_shared_perms, 2548 &cur_perm, &cur_shared); 2549 bdrv_child_set_perm(c, cur_perm, cur_shared, tran); 2550 } 2551 2552 return 0; 2553 } 2554 2555 /* 2556 * @list is a product of bdrv_topological_dfs() (may be called several times) - 2557 * a topologically sorted subgraph. 2558 * 2559 * After calling this function, the transaction @tran may only be completed 2560 * while holding a reader lock for the graph. 2561 */ 2562 static int GRAPH_RDLOCK 2563 bdrv_do_refresh_perms(GSList *list, BlockReopenQueue *q, Transaction *tran, 2564 Error **errp) 2565 { 2566 int ret; 2567 BlockDriverState *bs; 2568 GLOBAL_STATE_CODE(); 2569 2570 for ( ; list; list = list->next) { 2571 bs = list->data; 2572 2573 if (bdrv_parent_perms_conflict(bs, errp)) { 2574 return -EINVAL; 2575 } 2576 2577 ret = bdrv_node_refresh_perm(bs, q, tran, errp); 2578 if (ret < 0) { 2579 return ret; 2580 } 2581 } 2582 2583 return 0; 2584 } 2585 2586 /* 2587 * @list is any list of nodes. List is completed by all subtrees and 2588 * topologically sorted. It's not a problem if some node occurs in the @list 2589 * several times. 2590 * 2591 * After calling this function, the transaction @tran may only be completed 2592 * while holding a reader lock for the graph. 2593 */ 2594 static int GRAPH_RDLOCK 2595 bdrv_list_refresh_perms(GSList *list, BlockReopenQueue *q, Transaction *tran, 2596 Error **errp) 2597 { 2598 g_autoptr(GHashTable) found = g_hash_table_new(NULL, NULL); 2599 g_autoptr(GSList) refresh_list = NULL; 2600 2601 for ( ; list; list = list->next) { 2602 refresh_list = bdrv_topological_dfs(refresh_list, found, list->data); 2603 } 2604 2605 return bdrv_do_refresh_perms(refresh_list, q, tran, errp); 2606 } 2607 2608 void bdrv_get_cumulative_perm(BlockDriverState *bs, uint64_t *perm, 2609 uint64_t *shared_perm) 2610 { 2611 BdrvChild *c; 2612 uint64_t cumulative_perms = 0; 2613 uint64_t cumulative_shared_perms = BLK_PERM_ALL; 2614 2615 GLOBAL_STATE_CODE(); 2616 2617 QLIST_FOREACH(c, &bs->parents, next_parent) { 2618 cumulative_perms |= c->perm; 2619 cumulative_shared_perms &= c->shared_perm; 2620 } 2621 2622 *perm = cumulative_perms; 2623 *shared_perm = cumulative_shared_perms; 2624 } 2625 2626 char *bdrv_perm_names(uint64_t perm) 2627 { 2628 struct perm_name { 2629 uint64_t perm; 2630 const char *name; 2631 } permissions[] = { 2632 { BLK_PERM_CONSISTENT_READ, "consistent read" }, 2633 { BLK_PERM_WRITE, "write" }, 2634 { BLK_PERM_WRITE_UNCHANGED, "write unchanged" }, 2635 { BLK_PERM_RESIZE, "resize" }, 2636 { 0, NULL } 2637 }; 2638 2639 GString *result = g_string_sized_new(30); 2640 struct perm_name *p; 2641 2642 for (p = permissions; p->name; p++) { 2643 if (perm & p->perm) { 2644 if (result->len > 0) { 2645 g_string_append(result, ", "); 2646 } 2647 g_string_append(result, p->name); 2648 } 2649 } 2650 2651 return g_string_free(result, FALSE); 2652 } 2653 2654 2655 /* 2656 * @tran is allowed to be NULL. In this case no rollback is possible. 2657 * 2658 * After calling this function, the transaction @tran may only be completed 2659 * while holding a reader lock for the graph. 2660 */ 2661 static int GRAPH_RDLOCK 2662 bdrv_refresh_perms(BlockDriverState *bs, Transaction *tran, Error **errp) 2663 { 2664 int ret; 2665 Transaction *local_tran = NULL; 2666 g_autoptr(GSList) list = bdrv_topological_dfs(NULL, NULL, bs); 2667 GLOBAL_STATE_CODE(); 2668 2669 if (!tran) { 2670 tran = local_tran = tran_new(); 2671 } 2672 2673 ret = bdrv_do_refresh_perms(list, NULL, tran, errp); 2674 2675 if (local_tran) { 2676 tran_finalize(local_tran, ret); 2677 } 2678 2679 return ret; 2680 } 2681 2682 int bdrv_child_try_set_perm(BdrvChild *c, uint64_t perm, uint64_t shared, 2683 Error **errp) 2684 { 2685 Error *local_err = NULL; 2686 Transaction *tran = tran_new(); 2687 int ret; 2688 2689 GLOBAL_STATE_CODE(); 2690 2691 bdrv_child_set_perm(c, perm, shared, tran); 2692 2693 ret = bdrv_refresh_perms(c->bs, tran, &local_err); 2694 2695 tran_finalize(tran, ret); 2696 2697 if (ret < 0) { 2698 if ((perm & ~c->perm) || (c->shared_perm & ~shared)) { 2699 /* tighten permissions */ 2700 error_propagate(errp, local_err); 2701 } else { 2702 /* 2703 * Our caller may intend to only loosen restrictions and 2704 * does not expect this function to fail. Errors are not 2705 * fatal in such a case, so we can just hide them from our 2706 * caller. 2707 */ 2708 error_free(local_err); 2709 ret = 0; 2710 } 2711 } 2712 2713 return ret; 2714 } 2715 2716 int bdrv_child_refresh_perms(BlockDriverState *bs, BdrvChild *c, Error **errp) 2717 { 2718 uint64_t parent_perms, parent_shared; 2719 uint64_t perms, shared; 2720 2721 GLOBAL_STATE_CODE(); 2722 2723 bdrv_get_cumulative_perm(bs, &parent_perms, &parent_shared); 2724 bdrv_child_perm(bs, c->bs, c, c->role, NULL, 2725 parent_perms, parent_shared, &perms, &shared); 2726 2727 return bdrv_child_try_set_perm(c, perms, shared, errp); 2728 } 2729 2730 /* 2731 * Default implementation for .bdrv_child_perm() for block filters: 2732 * Forward CONSISTENT_READ, WRITE, WRITE_UNCHANGED, and RESIZE to the 2733 * filtered child. 2734 */ 2735 static void bdrv_filter_default_perms(BlockDriverState *bs, BdrvChild *c, 2736 BdrvChildRole role, 2737 BlockReopenQueue *reopen_queue, 2738 uint64_t perm, uint64_t shared, 2739 uint64_t *nperm, uint64_t *nshared) 2740 { 2741 GLOBAL_STATE_CODE(); 2742 *nperm = perm & DEFAULT_PERM_PASSTHROUGH; 2743 *nshared = (shared & DEFAULT_PERM_PASSTHROUGH) | DEFAULT_PERM_UNCHANGED; 2744 } 2745 2746 static void bdrv_default_perms_for_cow(BlockDriverState *bs, BdrvChild *c, 2747 BdrvChildRole role, 2748 BlockReopenQueue *reopen_queue, 2749 uint64_t perm, uint64_t shared, 2750 uint64_t *nperm, uint64_t *nshared) 2751 { 2752 assert(role & BDRV_CHILD_COW); 2753 GLOBAL_STATE_CODE(); 2754 2755 /* 2756 * We want consistent read from backing files if the parent needs it. 2757 * No other operations are performed on backing files. 2758 */ 2759 perm &= BLK_PERM_CONSISTENT_READ; 2760 2761 /* 2762 * If the parent can deal with changing data, we're okay with a 2763 * writable and resizable backing file. 2764 * TODO Require !(perm & BLK_PERM_CONSISTENT_READ), too? 2765 */ 2766 if (shared & BLK_PERM_WRITE) { 2767 shared = BLK_PERM_WRITE | BLK_PERM_RESIZE; 2768 } else { 2769 shared = 0; 2770 } 2771 2772 shared |= BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED; 2773 2774 if (bs->open_flags & BDRV_O_INACTIVE) { 2775 shared |= BLK_PERM_WRITE | BLK_PERM_RESIZE; 2776 } 2777 2778 *nperm = perm; 2779 *nshared = shared; 2780 } 2781 2782 static void bdrv_default_perms_for_storage(BlockDriverState *bs, BdrvChild *c, 2783 BdrvChildRole role, 2784 BlockReopenQueue *reopen_queue, 2785 uint64_t perm, uint64_t shared, 2786 uint64_t *nperm, uint64_t *nshared) 2787 { 2788 int flags; 2789 2790 GLOBAL_STATE_CODE(); 2791 assert(role & (BDRV_CHILD_METADATA | BDRV_CHILD_DATA)); 2792 2793 flags = bdrv_reopen_get_flags(reopen_queue, bs); 2794 2795 /* 2796 * Apart from the modifications below, the same permissions are 2797 * forwarded and left alone as for filters 2798 */ 2799 bdrv_filter_default_perms(bs, c, role, reopen_queue, 2800 perm, shared, &perm, &shared); 2801 2802 if (role & BDRV_CHILD_METADATA) { 2803 /* Format drivers may touch metadata even if the guest doesn't write */ 2804 if (bdrv_is_writable_after_reopen(bs, reopen_queue)) { 2805 perm |= BLK_PERM_WRITE | BLK_PERM_RESIZE; 2806 } 2807 2808 /* 2809 * bs->file always needs to be consistent because of the 2810 * metadata. We can never allow other users to resize or write 2811 * to it. 2812 */ 2813 if (!(flags & BDRV_O_NO_IO)) { 2814 perm |= BLK_PERM_CONSISTENT_READ; 2815 } 2816 shared &= ~(BLK_PERM_WRITE | BLK_PERM_RESIZE); 2817 } 2818 2819 if (role & BDRV_CHILD_DATA) { 2820 /* 2821 * Technically, everything in this block is a subset of the 2822 * BDRV_CHILD_METADATA path taken above, and so this could 2823 * be an "else if" branch. However, that is not obvious, and 2824 * this function is not performance critical, therefore we let 2825 * this be an independent "if". 2826 */ 2827 2828 /* 2829 * We cannot allow other users to resize the file because the 2830 * format driver might have some assumptions about the size 2831 * (e.g. because it is stored in metadata, or because the file 2832 * is split into fixed-size data files). 2833 */ 2834 shared &= ~BLK_PERM_RESIZE; 2835 2836 /* 2837 * WRITE_UNCHANGED often cannot be performed as such on the 2838 * data file. For example, the qcow2 driver may still need to 2839 * write copied clusters on copy-on-read. 2840 */ 2841 if (perm & BLK_PERM_WRITE_UNCHANGED) { 2842 perm |= BLK_PERM_WRITE; 2843 } 2844 2845 /* 2846 * If the data file is written to, the format driver may 2847 * expect to be able to resize it by writing beyond the EOF. 2848 */ 2849 if (perm & BLK_PERM_WRITE) { 2850 perm |= BLK_PERM_RESIZE; 2851 } 2852 } 2853 2854 if (bs->open_flags & BDRV_O_INACTIVE) { 2855 shared |= BLK_PERM_WRITE | BLK_PERM_RESIZE; 2856 } 2857 2858 *nperm = perm; 2859 *nshared = shared; 2860 } 2861 2862 void bdrv_default_perms(BlockDriverState *bs, BdrvChild *c, 2863 BdrvChildRole role, BlockReopenQueue *reopen_queue, 2864 uint64_t perm, uint64_t shared, 2865 uint64_t *nperm, uint64_t *nshared) 2866 { 2867 GLOBAL_STATE_CODE(); 2868 if (role & BDRV_CHILD_FILTERED) { 2869 assert(!(role & (BDRV_CHILD_DATA | BDRV_CHILD_METADATA | 2870 BDRV_CHILD_COW))); 2871 bdrv_filter_default_perms(bs, c, role, reopen_queue, 2872 perm, shared, nperm, nshared); 2873 } else if (role & BDRV_CHILD_COW) { 2874 assert(!(role & (BDRV_CHILD_DATA | BDRV_CHILD_METADATA))); 2875 bdrv_default_perms_for_cow(bs, c, role, reopen_queue, 2876 perm, shared, nperm, nshared); 2877 } else if (role & (BDRV_CHILD_METADATA | BDRV_CHILD_DATA)) { 2878 bdrv_default_perms_for_storage(bs, c, role, reopen_queue, 2879 perm, shared, nperm, nshared); 2880 } else { 2881 g_assert_not_reached(); 2882 } 2883 } 2884 2885 uint64_t bdrv_qapi_perm_to_blk_perm(BlockPermission qapi_perm) 2886 { 2887 static const uint64_t permissions[] = { 2888 [BLOCK_PERMISSION_CONSISTENT_READ] = BLK_PERM_CONSISTENT_READ, 2889 [BLOCK_PERMISSION_WRITE] = BLK_PERM_WRITE, 2890 [BLOCK_PERMISSION_WRITE_UNCHANGED] = BLK_PERM_WRITE_UNCHANGED, 2891 [BLOCK_PERMISSION_RESIZE] = BLK_PERM_RESIZE, 2892 }; 2893 2894 QEMU_BUILD_BUG_ON(ARRAY_SIZE(permissions) != BLOCK_PERMISSION__MAX); 2895 QEMU_BUILD_BUG_ON(1UL << ARRAY_SIZE(permissions) != BLK_PERM_ALL + 1); 2896 2897 assert(qapi_perm < BLOCK_PERMISSION__MAX); 2898 2899 return permissions[qapi_perm]; 2900 } 2901 2902 /* 2903 * Replaces the node that a BdrvChild points to without updating permissions. 2904 * 2905 * If @new_bs is non-NULL, the parent of @child must already be drained through 2906 * @child and the caller must hold the AioContext lock for @new_bs. 2907 */ 2908 static void GRAPH_WRLOCK 2909 bdrv_replace_child_noperm(BdrvChild *child, BlockDriverState *new_bs) 2910 { 2911 BlockDriverState *old_bs = child->bs; 2912 int new_bs_quiesce_counter; 2913 2914 assert(!child->frozen); 2915 2916 /* 2917 * If we want to change the BdrvChild to point to a drained node as its new 2918 * child->bs, we need to make sure that its new parent is drained, too. In 2919 * other words, either child->quiesce_parent must already be true or we must 2920 * be able to set it and keep the parent's quiesce_counter consistent with 2921 * that, but without polling or starting new requests (this function 2922 * guarantees that it doesn't poll, and starting new requests would be 2923 * against the invariants of drain sections). 2924 * 2925 * To keep things simple, we pick the first option (child->quiesce_parent 2926 * must already be true). We also generalise the rule a bit to make it 2927 * easier to verify in callers and more likely to be covered in test cases: 2928 * The parent must be quiesced through this child even if new_bs isn't 2929 * currently drained. 2930 * 2931 * The only exception is for callers that always pass new_bs == NULL. In 2932 * this case, we obviously never need to consider the case of a drained 2933 * new_bs, so we can keep the callers simpler by allowing them not to drain 2934 * the parent. 2935 */ 2936 assert(!new_bs || child->quiesced_parent); 2937 assert(old_bs != new_bs); 2938 GLOBAL_STATE_CODE(); 2939 2940 if (old_bs && new_bs) { 2941 assert(bdrv_get_aio_context(old_bs) == bdrv_get_aio_context(new_bs)); 2942 } 2943 2944 if (old_bs) { 2945 if (child->klass->detach) { 2946 child->klass->detach(child); 2947 } 2948 QLIST_REMOVE(child, next_parent); 2949 } 2950 2951 child->bs = new_bs; 2952 2953 if (new_bs) { 2954 QLIST_INSERT_HEAD(&new_bs->parents, child, next_parent); 2955 if (child->klass->attach) { 2956 child->klass->attach(child); 2957 } 2958 } 2959 2960 /* 2961 * If the parent was drained through this BdrvChild previously, but new_bs 2962 * is not drained, allow requests to come in only after the new node has 2963 * been attached. 2964 */ 2965 new_bs_quiesce_counter = (new_bs ? new_bs->quiesce_counter : 0); 2966 if (!new_bs_quiesce_counter && child->quiesced_parent) { 2967 bdrv_parent_drained_end_single(child); 2968 } 2969 } 2970 2971 /** 2972 * Free the given @child. 2973 * 2974 * The child must be empty (i.e. `child->bs == NULL`) and it must be 2975 * unused (i.e. not in a children list). 2976 */ 2977 static void bdrv_child_free(BdrvChild *child) 2978 { 2979 assert(!child->bs); 2980 GLOBAL_STATE_CODE(); 2981 GRAPH_RDLOCK_GUARD_MAINLOOP(); 2982 2983 assert(!child->next.le_prev); /* not in children list */ 2984 2985 g_free(child->name); 2986 g_free(child); 2987 } 2988 2989 typedef struct BdrvAttachChildCommonState { 2990 BdrvChild *child; 2991 AioContext *old_parent_ctx; 2992 AioContext *old_child_ctx; 2993 } BdrvAttachChildCommonState; 2994 2995 static void GRAPH_WRLOCK bdrv_attach_child_common_abort(void *opaque) 2996 { 2997 BdrvAttachChildCommonState *s = opaque; 2998 BlockDriverState *bs = s->child->bs; 2999 3000 GLOBAL_STATE_CODE(); 3001 assert_bdrv_graph_writable(); 3002 3003 bdrv_replace_child_noperm(s->child, NULL); 3004 3005 if (bdrv_get_aio_context(bs) != s->old_child_ctx) { 3006 bdrv_try_change_aio_context(bs, s->old_child_ctx, NULL, &error_abort); 3007 } 3008 3009 if (bdrv_child_get_parent_aio_context(s->child) != s->old_parent_ctx) { 3010 Transaction *tran; 3011 GHashTable *visited; 3012 bool ret; 3013 3014 tran = tran_new(); 3015 3016 /* No need to visit `child`, because it has been detached already */ 3017 visited = g_hash_table_new(NULL, NULL); 3018 ret = s->child->klass->change_aio_ctx(s->child, s->old_parent_ctx, 3019 visited, tran, &error_abort); 3020 g_hash_table_destroy(visited); 3021 3022 /* transaction is supposed to always succeed */ 3023 assert(ret == true); 3024 tran_commit(tran); 3025 } 3026 3027 bdrv_schedule_unref(bs); 3028 bdrv_child_free(s->child); 3029 } 3030 3031 static TransactionActionDrv bdrv_attach_child_common_drv = { 3032 .abort = bdrv_attach_child_common_abort, 3033 .clean = g_free, 3034 }; 3035 3036 /* 3037 * Common part of attaching bdrv child to bs or to blk or to job 3038 * 3039 * Function doesn't update permissions, caller is responsible for this. 3040 * 3041 * After calling this function, the transaction @tran may only be completed 3042 * while holding a writer lock for the graph. 3043 * 3044 * Returns new created child. 3045 * 3046 * The caller must hold the AioContext lock for @child_bs. Both @parent_bs and 3047 * @child_bs can move to a different AioContext in this function. Callers must 3048 * make sure that their AioContext locking is still correct after this. 3049 */ 3050 static BdrvChild * GRAPH_WRLOCK 3051 bdrv_attach_child_common(BlockDriverState *child_bs, 3052 const char *child_name, 3053 const BdrvChildClass *child_class, 3054 BdrvChildRole child_role, 3055 uint64_t perm, uint64_t shared_perm, 3056 void *opaque, 3057 Transaction *tran, Error **errp) 3058 { 3059 BdrvChild *new_child; 3060 AioContext *parent_ctx, *new_child_ctx; 3061 AioContext *child_ctx = bdrv_get_aio_context(child_bs); 3062 3063 assert(child_class->get_parent_desc); 3064 GLOBAL_STATE_CODE(); 3065 3066 new_child = g_new(BdrvChild, 1); 3067 *new_child = (BdrvChild) { 3068 .bs = NULL, 3069 .name = g_strdup(child_name), 3070 .klass = child_class, 3071 .role = child_role, 3072 .perm = perm, 3073 .shared_perm = shared_perm, 3074 .opaque = opaque, 3075 }; 3076 3077 /* 3078 * If the AioContexts don't match, first try to move the subtree of 3079 * child_bs into the AioContext of the new parent. If this doesn't work, 3080 * try moving the parent into the AioContext of child_bs instead. 3081 */ 3082 parent_ctx = bdrv_child_get_parent_aio_context(new_child); 3083 if (child_ctx != parent_ctx) { 3084 Error *local_err = NULL; 3085 int ret = bdrv_try_change_aio_context(child_bs, parent_ctx, NULL, 3086 &local_err); 3087 3088 if (ret < 0 && child_class->change_aio_ctx) { 3089 Transaction *aio_ctx_tran = tran_new(); 3090 GHashTable *visited = g_hash_table_new(NULL, NULL); 3091 bool ret_child; 3092 3093 g_hash_table_add(visited, new_child); 3094 ret_child = child_class->change_aio_ctx(new_child, child_ctx, 3095 visited, aio_ctx_tran, 3096 NULL); 3097 if (ret_child == true) { 3098 error_free(local_err); 3099 ret = 0; 3100 } 3101 tran_finalize(aio_ctx_tran, ret_child == true ? 0 : -1); 3102 g_hash_table_destroy(visited); 3103 } 3104 3105 if (ret < 0) { 3106 error_propagate(errp, local_err); 3107 bdrv_child_free(new_child); 3108 return NULL; 3109 } 3110 } 3111 3112 new_child_ctx = bdrv_get_aio_context(child_bs); 3113 if (new_child_ctx != child_ctx) { 3114 aio_context_release(child_ctx); 3115 aio_context_acquire(new_child_ctx); 3116 } 3117 3118 bdrv_ref(child_bs); 3119 /* 3120 * Let every new BdrvChild start with a drained parent. Inserting the child 3121 * in the graph with bdrv_replace_child_noperm() will undrain it if 3122 * @child_bs is not drained. 3123 * 3124 * The child was only just created and is not yet visible in global state 3125 * until bdrv_replace_child_noperm() inserts it into the graph, so nobody 3126 * could have sent requests and polling is not necessary. 3127 * 3128 * Note that this means that the parent isn't fully drained yet, we only 3129 * stop new requests from coming in. This is fine, we don't care about the 3130 * old requests here, they are not for this child. If another place enters a 3131 * drain section for the same parent, but wants it to be fully quiesced, it 3132 * will not run most of the the code in .drained_begin() again (which is not 3133 * a problem, we already did this), but it will still poll until the parent 3134 * is fully quiesced, so it will not be negatively affected either. 3135 */ 3136 bdrv_parent_drained_begin_single(new_child); 3137 bdrv_replace_child_noperm(new_child, child_bs); 3138 3139 BdrvAttachChildCommonState *s = g_new(BdrvAttachChildCommonState, 1); 3140 *s = (BdrvAttachChildCommonState) { 3141 .child = new_child, 3142 .old_parent_ctx = parent_ctx, 3143 .old_child_ctx = child_ctx, 3144 }; 3145 tran_add(tran, &bdrv_attach_child_common_drv, s); 3146 3147 if (new_child_ctx != child_ctx) { 3148 aio_context_release(new_child_ctx); 3149 aio_context_acquire(child_ctx); 3150 } 3151 3152 return new_child; 3153 } 3154 3155 /* 3156 * Function doesn't update permissions, caller is responsible for this. 3157 * 3158 * The caller must hold the AioContext lock for @child_bs. Both @parent_bs and 3159 * @child_bs can move to a different AioContext in this function. Callers must 3160 * make sure that their AioContext locking is still correct after this. 3161 * 3162 * After calling this function, the transaction @tran may only be completed 3163 * while holding a writer lock for the graph. 3164 */ 3165 static BdrvChild * GRAPH_WRLOCK 3166 bdrv_attach_child_noperm(BlockDriverState *parent_bs, 3167 BlockDriverState *child_bs, 3168 const char *child_name, 3169 const BdrvChildClass *child_class, 3170 BdrvChildRole child_role, 3171 Transaction *tran, 3172 Error **errp) 3173 { 3174 uint64_t perm, shared_perm; 3175 3176 assert(parent_bs->drv); 3177 GLOBAL_STATE_CODE(); 3178 3179 if (bdrv_recurse_has_child(child_bs, parent_bs)) { 3180 error_setg(errp, "Making '%s' a %s child of '%s' would create a cycle", 3181 child_bs->node_name, child_name, parent_bs->node_name); 3182 return NULL; 3183 } 3184 3185 bdrv_get_cumulative_perm(parent_bs, &perm, &shared_perm); 3186 bdrv_child_perm(parent_bs, child_bs, NULL, child_role, NULL, 3187 perm, shared_perm, &perm, &shared_perm); 3188 3189 return bdrv_attach_child_common(child_bs, child_name, child_class, 3190 child_role, perm, shared_perm, parent_bs, 3191 tran, errp); 3192 } 3193 3194 /* 3195 * This function steals the reference to child_bs from the caller. 3196 * That reference is later dropped by bdrv_root_unref_child(). 3197 * 3198 * On failure NULL is returned, errp is set and the reference to 3199 * child_bs is also dropped. 3200 * 3201 * The caller must hold the AioContext lock @child_bs, but not that of @ctx 3202 * (unless @child_bs is already in @ctx). 3203 */ 3204 BdrvChild *bdrv_root_attach_child(BlockDriverState *child_bs, 3205 const char *child_name, 3206 const BdrvChildClass *child_class, 3207 BdrvChildRole child_role, 3208 uint64_t perm, uint64_t shared_perm, 3209 void *opaque, Error **errp) 3210 { 3211 int ret; 3212 BdrvChild *child; 3213 Transaction *tran = tran_new(); 3214 3215 GLOBAL_STATE_CODE(); 3216 3217 child = bdrv_attach_child_common(child_bs, child_name, child_class, 3218 child_role, perm, shared_perm, opaque, 3219 tran, errp); 3220 if (!child) { 3221 ret = -EINVAL; 3222 goto out; 3223 } 3224 3225 ret = bdrv_refresh_perms(child_bs, tran, errp); 3226 3227 out: 3228 tran_finalize(tran, ret); 3229 3230 bdrv_schedule_unref(child_bs); 3231 3232 return ret < 0 ? NULL : child; 3233 } 3234 3235 /* 3236 * This function transfers the reference to child_bs from the caller 3237 * to parent_bs. That reference is later dropped by parent_bs on 3238 * bdrv_close() or if someone calls bdrv_unref_child(). 3239 * 3240 * On failure NULL is returned, errp is set and the reference to 3241 * child_bs is also dropped. 3242 * 3243 * If @parent_bs and @child_bs are in different AioContexts, the caller must 3244 * hold the AioContext lock for @child_bs, but not for @parent_bs. 3245 */ 3246 BdrvChild *bdrv_attach_child(BlockDriverState *parent_bs, 3247 BlockDriverState *child_bs, 3248 const char *child_name, 3249 const BdrvChildClass *child_class, 3250 BdrvChildRole child_role, 3251 Error **errp) 3252 { 3253 int ret; 3254 BdrvChild *child; 3255 Transaction *tran = tran_new(); 3256 3257 GLOBAL_STATE_CODE(); 3258 3259 child = bdrv_attach_child_noperm(parent_bs, child_bs, child_name, 3260 child_class, child_role, tran, errp); 3261 if (!child) { 3262 ret = -EINVAL; 3263 goto out; 3264 } 3265 3266 ret = bdrv_refresh_perms(parent_bs, tran, errp); 3267 if (ret < 0) { 3268 goto out; 3269 } 3270 3271 out: 3272 tran_finalize(tran, ret); 3273 3274 bdrv_schedule_unref(child_bs); 3275 3276 return ret < 0 ? NULL : child; 3277 } 3278 3279 /* Callers must ensure that child->frozen is false. */ 3280 void bdrv_root_unref_child(BdrvChild *child) 3281 { 3282 BlockDriverState *child_bs = child->bs; 3283 3284 GLOBAL_STATE_CODE(); 3285 bdrv_replace_child_noperm(child, NULL); 3286 bdrv_child_free(child); 3287 3288 if (child_bs) { 3289 /* 3290 * Update permissions for old node. We're just taking a parent away, so 3291 * we're loosening restrictions. Errors of permission update are not 3292 * fatal in this case, ignore them. 3293 */ 3294 bdrv_refresh_perms(child_bs, NULL, NULL); 3295 3296 /* 3297 * When the parent requiring a non-default AioContext is removed, the 3298 * node moves back to the main AioContext 3299 */ 3300 bdrv_try_change_aio_context(child_bs, qemu_get_aio_context(), NULL, 3301 NULL); 3302 } 3303 3304 bdrv_schedule_unref(child_bs); 3305 } 3306 3307 typedef struct BdrvSetInheritsFrom { 3308 BlockDriverState *bs; 3309 BlockDriverState *old_inherits_from; 3310 } BdrvSetInheritsFrom; 3311 3312 static void bdrv_set_inherits_from_abort(void *opaque) 3313 { 3314 BdrvSetInheritsFrom *s = opaque; 3315 3316 s->bs->inherits_from = s->old_inherits_from; 3317 } 3318 3319 static TransactionActionDrv bdrv_set_inherits_from_drv = { 3320 .abort = bdrv_set_inherits_from_abort, 3321 .clean = g_free, 3322 }; 3323 3324 /* @tran is allowed to be NULL. In this case no rollback is possible */ 3325 static void bdrv_set_inherits_from(BlockDriverState *bs, 3326 BlockDriverState *new_inherits_from, 3327 Transaction *tran) 3328 { 3329 if (tran) { 3330 BdrvSetInheritsFrom *s = g_new(BdrvSetInheritsFrom, 1); 3331 3332 *s = (BdrvSetInheritsFrom) { 3333 .bs = bs, 3334 .old_inherits_from = bs->inherits_from, 3335 }; 3336 3337 tran_add(tran, &bdrv_set_inherits_from_drv, s); 3338 } 3339 3340 bs->inherits_from = new_inherits_from; 3341 } 3342 3343 /** 3344 * Clear all inherits_from pointers from children and grandchildren of 3345 * @root that point to @root, where necessary. 3346 * @tran is allowed to be NULL. In this case no rollback is possible 3347 */ 3348 static void GRAPH_WRLOCK 3349 bdrv_unset_inherits_from(BlockDriverState *root, BdrvChild *child, 3350 Transaction *tran) 3351 { 3352 BdrvChild *c; 3353 3354 if (child->bs->inherits_from == root) { 3355 /* 3356 * Remove inherits_from only when the last reference between root and 3357 * child->bs goes away. 3358 */ 3359 QLIST_FOREACH(c, &root->children, next) { 3360 if (c != child && c->bs == child->bs) { 3361 break; 3362 } 3363 } 3364 if (c == NULL) { 3365 bdrv_set_inherits_from(child->bs, NULL, tran); 3366 } 3367 } 3368 3369 QLIST_FOREACH(c, &child->bs->children, next) { 3370 bdrv_unset_inherits_from(root, c, tran); 3371 } 3372 } 3373 3374 /* Callers must ensure that child->frozen is false. */ 3375 void bdrv_unref_child(BlockDriverState *parent, BdrvChild *child) 3376 { 3377 GLOBAL_STATE_CODE(); 3378 if (child == NULL) { 3379 return; 3380 } 3381 3382 bdrv_unset_inherits_from(parent, child, NULL); 3383 bdrv_root_unref_child(child); 3384 } 3385 3386 3387 static void GRAPH_RDLOCK 3388 bdrv_parent_cb_change_media(BlockDriverState *bs, bool load) 3389 { 3390 BdrvChild *c; 3391 GLOBAL_STATE_CODE(); 3392 QLIST_FOREACH(c, &bs->parents, next_parent) { 3393 if (c->klass->change_media) { 3394 c->klass->change_media(c, load); 3395 } 3396 } 3397 } 3398 3399 /* Return true if you can reach parent going through child->inherits_from 3400 * recursively. If parent or child are NULL, return false */ 3401 static bool bdrv_inherits_from_recursive(BlockDriverState *child, 3402 BlockDriverState *parent) 3403 { 3404 while (child && child != parent) { 3405 child = child->inherits_from; 3406 } 3407 3408 return child != NULL; 3409 } 3410 3411 /* 3412 * Return the BdrvChildRole for @bs's backing child. bs->backing is 3413 * mostly used for COW backing children (role = COW), but also for 3414 * filtered children (role = FILTERED | PRIMARY). 3415 */ 3416 static BdrvChildRole bdrv_backing_role(BlockDriverState *bs) 3417 { 3418 if (bs->drv && bs->drv->is_filter) { 3419 return BDRV_CHILD_FILTERED | BDRV_CHILD_PRIMARY; 3420 } else { 3421 return BDRV_CHILD_COW; 3422 } 3423 } 3424 3425 /* 3426 * Sets the bs->backing or bs->file link of a BDS. A new reference is created; 3427 * callers which don't need their own reference any more must call bdrv_unref(). 3428 * 3429 * If the respective child is already present (i.e. we're detaching a node), 3430 * that child node must be drained. 3431 * 3432 * Function doesn't update permissions, caller is responsible for this. 3433 * 3434 * The caller must hold the AioContext lock for @child_bs. Both @parent_bs and 3435 * @child_bs can move to a different AioContext in this function. Callers must 3436 * make sure that their AioContext locking is still correct after this. 3437 * 3438 * After calling this function, the transaction @tran may only be completed 3439 * while holding a writer lock for the graph. 3440 */ 3441 static int GRAPH_WRLOCK 3442 bdrv_set_file_or_backing_noperm(BlockDriverState *parent_bs, 3443 BlockDriverState *child_bs, 3444 bool is_backing, 3445 Transaction *tran, Error **errp) 3446 { 3447 bool update_inherits_from = 3448 bdrv_inherits_from_recursive(child_bs, parent_bs); 3449 BdrvChild *child = is_backing ? parent_bs->backing : parent_bs->file; 3450 BdrvChildRole role; 3451 3452 GLOBAL_STATE_CODE(); 3453 3454 if (!parent_bs->drv) { 3455 /* 3456 * Node without drv is an object without a class :/. TODO: finally fix 3457 * qcow2 driver to never clear bs->drv and implement format corruption 3458 * handling in other way. 3459 */ 3460 error_setg(errp, "Node corrupted"); 3461 return -EINVAL; 3462 } 3463 3464 if (child && child->frozen) { 3465 error_setg(errp, "Cannot change frozen '%s' link from '%s' to '%s'", 3466 child->name, parent_bs->node_name, child->bs->node_name); 3467 return -EPERM; 3468 } 3469 3470 if (is_backing && !parent_bs->drv->is_filter && 3471 !parent_bs->drv->supports_backing) 3472 { 3473 error_setg(errp, "Driver '%s' of node '%s' does not support backing " 3474 "files", parent_bs->drv->format_name, parent_bs->node_name); 3475 return -EINVAL; 3476 } 3477 3478 if (parent_bs->drv->is_filter) { 3479 role = BDRV_CHILD_FILTERED | BDRV_CHILD_PRIMARY; 3480 } else if (is_backing) { 3481 role = BDRV_CHILD_COW; 3482 } else { 3483 /* 3484 * We only can use same role as it is in existing child. We don't have 3485 * infrastructure to determine role of file child in generic way 3486 */ 3487 if (!child) { 3488 error_setg(errp, "Cannot set file child to format node without " 3489 "file child"); 3490 return -EINVAL; 3491 } 3492 role = child->role; 3493 } 3494 3495 if (child) { 3496 assert(child->bs->quiesce_counter); 3497 bdrv_unset_inherits_from(parent_bs, child, tran); 3498 bdrv_remove_child(child, tran); 3499 } 3500 3501 if (!child_bs) { 3502 goto out; 3503 } 3504 3505 child = bdrv_attach_child_noperm(parent_bs, child_bs, 3506 is_backing ? "backing" : "file", 3507 &child_of_bds, role, 3508 tran, errp); 3509 if (!child) { 3510 return -EINVAL; 3511 } 3512 3513 3514 /* 3515 * If inherits_from pointed recursively to bs then let's update it to 3516 * point directly to bs (else it will become NULL). 3517 */ 3518 if (update_inherits_from) { 3519 bdrv_set_inherits_from(child_bs, parent_bs, tran); 3520 } 3521 3522 out: 3523 bdrv_refresh_limits(parent_bs, tran, NULL); 3524 3525 return 0; 3526 } 3527 3528 /* 3529 * The caller must hold the AioContext lock for @backing_hd. Both @bs and 3530 * @backing_hd can move to a different AioContext in this function. Callers must 3531 * make sure that their AioContext locking is still correct after this. 3532 * 3533 * If a backing child is already present (i.e. we're detaching a node), that 3534 * child node must be drained. 3535 */ 3536 int bdrv_set_backing_hd_drained(BlockDriverState *bs, 3537 BlockDriverState *backing_hd, 3538 Error **errp) 3539 { 3540 int ret; 3541 Transaction *tran = tran_new(); 3542 3543 GLOBAL_STATE_CODE(); 3544 assert(bs->quiesce_counter > 0); 3545 if (bs->backing) { 3546 assert(bs->backing->bs->quiesce_counter > 0); 3547 } 3548 3549 ret = bdrv_set_file_or_backing_noperm(bs, backing_hd, true, tran, errp); 3550 if (ret < 0) { 3551 goto out; 3552 } 3553 3554 ret = bdrv_refresh_perms(bs, tran, errp); 3555 out: 3556 tran_finalize(tran, ret); 3557 return ret; 3558 } 3559 3560 int bdrv_set_backing_hd(BlockDriverState *bs, BlockDriverState *backing_hd, 3561 Error **errp) 3562 { 3563 BlockDriverState *drain_bs = bs->backing ? bs->backing->bs : bs; 3564 int ret; 3565 GLOBAL_STATE_CODE(); 3566 3567 bdrv_ref(drain_bs); 3568 bdrv_drained_begin(drain_bs); 3569 bdrv_graph_wrlock(backing_hd); 3570 ret = bdrv_set_backing_hd_drained(bs, backing_hd, errp); 3571 bdrv_graph_wrunlock(); 3572 bdrv_drained_end(drain_bs); 3573 bdrv_unref(drain_bs); 3574 3575 return ret; 3576 } 3577 3578 /* 3579 * Opens the backing file for a BlockDriverState if not yet open 3580 * 3581 * bdref_key specifies the key for the image's BlockdevRef in the options QDict. 3582 * That QDict has to be flattened; therefore, if the BlockdevRef is a QDict 3583 * itself, all options starting with "${bdref_key}." are considered part of the 3584 * BlockdevRef. 3585 * 3586 * The caller must hold the main AioContext lock. 3587 * 3588 * TODO Can this be unified with bdrv_open_image()? 3589 */ 3590 int bdrv_open_backing_file(BlockDriverState *bs, QDict *parent_options, 3591 const char *bdref_key, Error **errp) 3592 { 3593 char *backing_filename = NULL; 3594 char *bdref_key_dot; 3595 const char *reference = NULL; 3596 int ret = 0; 3597 bool implicit_backing = false; 3598 BlockDriverState *backing_hd; 3599 AioContext *backing_hd_ctx; 3600 QDict *options; 3601 QDict *tmp_parent_options = NULL; 3602 Error *local_err = NULL; 3603 3604 GLOBAL_STATE_CODE(); 3605 3606 if (bs->backing != NULL) { 3607 goto free_exit; 3608 } 3609 3610 /* NULL means an empty set of options */ 3611 if (parent_options == NULL) { 3612 tmp_parent_options = qdict_new(); 3613 parent_options = tmp_parent_options; 3614 } 3615 3616 bs->open_flags &= ~BDRV_O_NO_BACKING; 3617 3618 bdref_key_dot = g_strdup_printf("%s.", bdref_key); 3619 qdict_extract_subqdict(parent_options, &options, bdref_key_dot); 3620 g_free(bdref_key_dot); 3621 3622 /* 3623 * Caution: while qdict_get_try_str() is fine, getting non-string 3624 * types would require more care. When @parent_options come from 3625 * -blockdev or blockdev_add, its members are typed according to 3626 * the QAPI schema, but when they come from -drive, they're all 3627 * QString. 3628 */ 3629 reference = qdict_get_try_str(parent_options, bdref_key); 3630 if (reference || qdict_haskey(options, "file.filename")) { 3631 /* keep backing_filename NULL */ 3632 } else if (bs->backing_file[0] == '\0' && qdict_size(options) == 0) { 3633 qobject_unref(options); 3634 goto free_exit; 3635 } else { 3636 if (qdict_size(options) == 0) { 3637 /* If the user specifies options that do not modify the 3638 * backing file's behavior, we might still consider it the 3639 * implicit backing file. But it's easier this way, and 3640 * just specifying some of the backing BDS's options is 3641 * only possible with -drive anyway (otherwise the QAPI 3642 * schema forces the user to specify everything). */ 3643 implicit_backing = !strcmp(bs->auto_backing_file, bs->backing_file); 3644 } 3645 3646 bdrv_graph_rdlock_main_loop(); 3647 backing_filename = bdrv_get_full_backing_filename(bs, &local_err); 3648 bdrv_graph_rdunlock_main_loop(); 3649 3650 if (local_err) { 3651 ret = -EINVAL; 3652 error_propagate(errp, local_err); 3653 qobject_unref(options); 3654 goto free_exit; 3655 } 3656 } 3657 3658 if (!bs->drv || !bs->drv->supports_backing) { 3659 ret = -EINVAL; 3660 error_setg(errp, "Driver doesn't support backing files"); 3661 qobject_unref(options); 3662 goto free_exit; 3663 } 3664 3665 if (!reference && 3666 bs->backing_format[0] != '\0' && !qdict_haskey(options, "driver")) { 3667 qdict_put_str(options, "driver", bs->backing_format); 3668 } 3669 3670 backing_hd = bdrv_open_inherit(backing_filename, reference, options, 0, bs, 3671 &child_of_bds, bdrv_backing_role(bs), errp); 3672 if (!backing_hd) { 3673 bs->open_flags |= BDRV_O_NO_BACKING; 3674 error_prepend(errp, "Could not open backing file: "); 3675 ret = -EINVAL; 3676 goto free_exit; 3677 } 3678 3679 if (implicit_backing) { 3680 bdrv_graph_rdlock_main_loop(); 3681 bdrv_refresh_filename(backing_hd); 3682 bdrv_graph_rdunlock_main_loop(); 3683 pstrcpy(bs->auto_backing_file, sizeof(bs->auto_backing_file), 3684 backing_hd->filename); 3685 } 3686 3687 /* Hook up the backing file link; drop our reference, bs owns the 3688 * backing_hd reference now */ 3689 backing_hd_ctx = bdrv_get_aio_context(backing_hd); 3690 aio_context_acquire(backing_hd_ctx); 3691 ret = bdrv_set_backing_hd(bs, backing_hd, errp); 3692 bdrv_unref(backing_hd); 3693 aio_context_release(backing_hd_ctx); 3694 3695 if (ret < 0) { 3696 goto free_exit; 3697 } 3698 3699 qdict_del(parent_options, bdref_key); 3700 3701 free_exit: 3702 g_free(backing_filename); 3703 qobject_unref(tmp_parent_options); 3704 return ret; 3705 } 3706 3707 static BlockDriverState * 3708 bdrv_open_child_bs(const char *filename, QDict *options, const char *bdref_key, 3709 BlockDriverState *parent, const BdrvChildClass *child_class, 3710 BdrvChildRole child_role, bool allow_none, Error **errp) 3711 { 3712 BlockDriverState *bs = NULL; 3713 QDict *image_options; 3714 char *bdref_key_dot; 3715 const char *reference; 3716 3717 assert(child_class != NULL); 3718 3719 bdref_key_dot = g_strdup_printf("%s.", bdref_key); 3720 qdict_extract_subqdict(options, &image_options, bdref_key_dot); 3721 g_free(bdref_key_dot); 3722 3723 /* 3724 * Caution: while qdict_get_try_str() is fine, getting non-string 3725 * types would require more care. When @options come from 3726 * -blockdev or blockdev_add, its members are typed according to 3727 * the QAPI schema, but when they come from -drive, they're all 3728 * QString. 3729 */ 3730 reference = qdict_get_try_str(options, bdref_key); 3731 if (!filename && !reference && !qdict_size(image_options)) { 3732 if (!allow_none) { 3733 error_setg(errp, "A block device must be specified for \"%s\"", 3734 bdref_key); 3735 } 3736 qobject_unref(image_options); 3737 goto done; 3738 } 3739 3740 bs = bdrv_open_inherit(filename, reference, image_options, 0, 3741 parent, child_class, child_role, errp); 3742 if (!bs) { 3743 goto done; 3744 } 3745 3746 done: 3747 qdict_del(options, bdref_key); 3748 return bs; 3749 } 3750 3751 /* 3752 * Opens a disk image whose options are given as BlockdevRef in another block 3753 * device's options. 3754 * 3755 * If allow_none is true, no image will be opened if filename is false and no 3756 * BlockdevRef is given. NULL will be returned, but errp remains unset. 3757 * 3758 * bdrev_key specifies the key for the image's BlockdevRef in the options QDict. 3759 * That QDict has to be flattened; therefore, if the BlockdevRef is a QDict 3760 * itself, all options starting with "${bdref_key}." are considered part of the 3761 * BlockdevRef. 3762 * 3763 * The BlockdevRef will be removed from the options QDict. 3764 * 3765 * The caller must hold the lock of the main AioContext and no other AioContext. 3766 * @parent can move to a different AioContext in this function. Callers must 3767 * make sure that their AioContext locking is still correct after this. 3768 */ 3769 BdrvChild *bdrv_open_child(const char *filename, 3770 QDict *options, const char *bdref_key, 3771 BlockDriverState *parent, 3772 const BdrvChildClass *child_class, 3773 BdrvChildRole child_role, 3774 bool allow_none, Error **errp) 3775 { 3776 BlockDriverState *bs; 3777 BdrvChild *child; 3778 AioContext *ctx; 3779 3780 GLOBAL_STATE_CODE(); 3781 3782 bs = bdrv_open_child_bs(filename, options, bdref_key, parent, child_class, 3783 child_role, allow_none, errp); 3784 if (bs == NULL) { 3785 return NULL; 3786 } 3787 3788 bdrv_graph_wrlock(NULL); 3789 ctx = bdrv_get_aio_context(bs); 3790 aio_context_acquire(ctx); 3791 child = bdrv_attach_child(parent, bs, bdref_key, child_class, child_role, 3792 errp); 3793 aio_context_release(ctx); 3794 bdrv_graph_wrunlock(); 3795 3796 return child; 3797 } 3798 3799 /* 3800 * Wrapper on bdrv_open_child() for most popular case: open primary child of bs. 3801 * 3802 * The caller must hold the lock of the main AioContext and no other AioContext. 3803 * @parent can move to a different AioContext in this function. Callers must 3804 * make sure that their AioContext locking is still correct after this. 3805 */ 3806 int bdrv_open_file_child(const char *filename, 3807 QDict *options, const char *bdref_key, 3808 BlockDriverState *parent, Error **errp) 3809 { 3810 BdrvChildRole role; 3811 3812 /* commit_top and mirror_top don't use this function */ 3813 assert(!parent->drv->filtered_child_is_backing); 3814 role = parent->drv->is_filter ? 3815 (BDRV_CHILD_FILTERED | BDRV_CHILD_PRIMARY) : BDRV_CHILD_IMAGE; 3816 3817 if (!bdrv_open_child(filename, options, bdref_key, parent, 3818 &child_of_bds, role, false, errp)) 3819 { 3820 return -EINVAL; 3821 } 3822 3823 return 0; 3824 } 3825 3826 /* 3827 * TODO Future callers may need to specify parent/child_class in order for 3828 * option inheritance to work. Existing callers use it for the root node. 3829 */ 3830 BlockDriverState *bdrv_open_blockdev_ref(BlockdevRef *ref, Error **errp) 3831 { 3832 BlockDriverState *bs = NULL; 3833 QObject *obj = NULL; 3834 QDict *qdict = NULL; 3835 const char *reference = NULL; 3836 Visitor *v = NULL; 3837 3838 GLOBAL_STATE_CODE(); 3839 3840 if (ref->type == QTYPE_QSTRING) { 3841 reference = ref->u.reference; 3842 } else { 3843 BlockdevOptions *options = &ref->u.definition; 3844 assert(ref->type == QTYPE_QDICT); 3845 3846 v = qobject_output_visitor_new(&obj); 3847 visit_type_BlockdevOptions(v, NULL, &options, &error_abort); 3848 visit_complete(v, &obj); 3849 3850 qdict = qobject_to(QDict, obj); 3851 qdict_flatten(qdict); 3852 3853 /* bdrv_open_inherit() defaults to the values in bdrv_flags (for 3854 * compatibility with other callers) rather than what we want as the 3855 * real defaults. Apply the defaults here instead. */ 3856 qdict_set_default_str(qdict, BDRV_OPT_CACHE_DIRECT, "off"); 3857 qdict_set_default_str(qdict, BDRV_OPT_CACHE_NO_FLUSH, "off"); 3858 qdict_set_default_str(qdict, BDRV_OPT_READ_ONLY, "off"); 3859 qdict_set_default_str(qdict, BDRV_OPT_AUTO_READ_ONLY, "off"); 3860 3861 } 3862 3863 bs = bdrv_open_inherit(NULL, reference, qdict, 0, NULL, NULL, 0, errp); 3864 obj = NULL; 3865 qobject_unref(obj); 3866 visit_free(v); 3867 return bs; 3868 } 3869 3870 static BlockDriverState *bdrv_append_temp_snapshot(BlockDriverState *bs, 3871 int flags, 3872 QDict *snapshot_options, 3873 Error **errp) 3874 { 3875 g_autofree char *tmp_filename = NULL; 3876 int64_t total_size; 3877 QemuOpts *opts = NULL; 3878 BlockDriverState *bs_snapshot = NULL; 3879 AioContext *ctx = bdrv_get_aio_context(bs); 3880 int ret; 3881 3882 GLOBAL_STATE_CODE(); 3883 3884 /* if snapshot, we create a temporary backing file and open it 3885 instead of opening 'filename' directly */ 3886 3887 /* Get the required size from the image */ 3888 aio_context_acquire(ctx); 3889 total_size = bdrv_getlength(bs); 3890 aio_context_release(ctx); 3891 3892 if (total_size < 0) { 3893 error_setg_errno(errp, -total_size, "Could not get image size"); 3894 goto out; 3895 } 3896 3897 /* Create the temporary image */ 3898 tmp_filename = create_tmp_file(errp); 3899 if (!tmp_filename) { 3900 goto out; 3901 } 3902 3903 opts = qemu_opts_create(bdrv_qcow2.create_opts, NULL, 0, 3904 &error_abort); 3905 qemu_opt_set_number(opts, BLOCK_OPT_SIZE, total_size, &error_abort); 3906 ret = bdrv_create(&bdrv_qcow2, tmp_filename, opts, errp); 3907 qemu_opts_del(opts); 3908 if (ret < 0) { 3909 error_prepend(errp, "Could not create temporary overlay '%s': ", 3910 tmp_filename); 3911 goto out; 3912 } 3913 3914 /* Prepare options QDict for the temporary file */ 3915 qdict_put_str(snapshot_options, "file.driver", "file"); 3916 qdict_put_str(snapshot_options, "file.filename", tmp_filename); 3917 qdict_put_str(snapshot_options, "driver", "qcow2"); 3918 3919 bs_snapshot = bdrv_open(NULL, NULL, snapshot_options, flags, errp); 3920 snapshot_options = NULL; 3921 if (!bs_snapshot) { 3922 goto out; 3923 } 3924 3925 aio_context_acquire(ctx); 3926 ret = bdrv_append(bs_snapshot, bs, errp); 3927 aio_context_release(ctx); 3928 3929 if (ret < 0) { 3930 bs_snapshot = NULL; 3931 goto out; 3932 } 3933 3934 out: 3935 qobject_unref(snapshot_options); 3936 return bs_snapshot; 3937 } 3938 3939 /* 3940 * Opens a disk image (raw, qcow2, vmdk, ...) 3941 * 3942 * options is a QDict of options to pass to the block drivers, or NULL for an 3943 * empty set of options. The reference to the QDict belongs to the block layer 3944 * after the call (even on failure), so if the caller intends to reuse the 3945 * dictionary, it needs to use qobject_ref() before calling bdrv_open. 3946 * 3947 * If *pbs is NULL, a new BDS will be created with a pointer to it stored there. 3948 * If it is not NULL, the referenced BDS will be reused. 3949 * 3950 * The reference parameter may be used to specify an existing block device which 3951 * should be opened. If specified, neither options nor a filename may be given, 3952 * nor can an existing BDS be reused (that is, *pbs has to be NULL). 3953 * 3954 * The caller must always hold the main AioContext lock. 3955 */ 3956 static BlockDriverState * no_coroutine_fn 3957 bdrv_open_inherit(const char *filename, const char *reference, QDict *options, 3958 int flags, BlockDriverState *parent, 3959 const BdrvChildClass *child_class, BdrvChildRole child_role, 3960 Error **errp) 3961 { 3962 int ret; 3963 BlockBackend *file = NULL; 3964 BlockDriverState *bs; 3965 BlockDriver *drv = NULL; 3966 BdrvChild *child; 3967 const char *drvname; 3968 const char *backing; 3969 Error *local_err = NULL; 3970 QDict *snapshot_options = NULL; 3971 int snapshot_flags = 0; 3972 AioContext *ctx = qemu_get_aio_context(); 3973 3974 assert(!child_class || !flags); 3975 assert(!child_class == !parent); 3976 GLOBAL_STATE_CODE(); 3977 assert(!qemu_in_coroutine()); 3978 3979 /* TODO We'll eventually have to take a writer lock in this function */ 3980 GRAPH_RDLOCK_GUARD_MAINLOOP(); 3981 3982 if (reference) { 3983 bool options_non_empty = options ? qdict_size(options) : false; 3984 qobject_unref(options); 3985 3986 if (filename || options_non_empty) { 3987 error_setg(errp, "Cannot reference an existing block device with " 3988 "additional options or a new filename"); 3989 return NULL; 3990 } 3991 3992 bs = bdrv_lookup_bs(reference, reference, errp); 3993 if (!bs) { 3994 return NULL; 3995 } 3996 3997 bdrv_ref(bs); 3998 return bs; 3999 } 4000 4001 bs = bdrv_new(); 4002 4003 /* NULL means an empty set of options */ 4004 if (options == NULL) { 4005 options = qdict_new(); 4006 } 4007 4008 /* json: syntax counts as explicit options, as if in the QDict */ 4009 parse_json_protocol(options, &filename, &local_err); 4010 if (local_err) { 4011 goto fail; 4012 } 4013 4014 bs->explicit_options = qdict_clone_shallow(options); 4015 4016 if (child_class) { 4017 bool parent_is_format; 4018 4019 if (parent->drv) { 4020 parent_is_format = parent->drv->is_format; 4021 } else { 4022 /* 4023 * parent->drv is not set yet because this node is opened for 4024 * (potential) format probing. That means that @parent is going 4025 * to be a format node. 4026 */ 4027 parent_is_format = true; 4028 } 4029 4030 bs->inherits_from = parent; 4031 child_class->inherit_options(child_role, parent_is_format, 4032 &flags, options, 4033 parent->open_flags, parent->options); 4034 } 4035 4036 ret = bdrv_fill_options(&options, filename, &flags, &local_err); 4037 if (ret < 0) { 4038 goto fail; 4039 } 4040 4041 /* 4042 * Set the BDRV_O_RDWR and BDRV_O_ALLOW_RDWR flags. 4043 * Caution: getting a boolean member of @options requires care. 4044 * When @options come from -blockdev or blockdev_add, members are 4045 * typed according to the QAPI schema, but when they come from 4046 * -drive, they're all QString. 4047 */ 4048 if (g_strcmp0(qdict_get_try_str(options, BDRV_OPT_READ_ONLY), "on") && 4049 !qdict_get_try_bool(options, BDRV_OPT_READ_ONLY, false)) { 4050 flags |= (BDRV_O_RDWR | BDRV_O_ALLOW_RDWR); 4051 } else { 4052 flags &= ~BDRV_O_RDWR; 4053 } 4054 4055 if (flags & BDRV_O_SNAPSHOT) { 4056 snapshot_options = qdict_new(); 4057 bdrv_temp_snapshot_options(&snapshot_flags, snapshot_options, 4058 flags, options); 4059 /* Let bdrv_backing_options() override "read-only" */ 4060 qdict_del(options, BDRV_OPT_READ_ONLY); 4061 bdrv_inherited_options(BDRV_CHILD_COW, true, 4062 &flags, options, flags, options); 4063 } 4064 4065 bs->open_flags = flags; 4066 bs->options = options; 4067 options = qdict_clone_shallow(options); 4068 4069 /* Find the right image format driver */ 4070 /* See cautionary note on accessing @options above */ 4071 drvname = qdict_get_try_str(options, "driver"); 4072 if (drvname) { 4073 drv = bdrv_find_format(drvname); 4074 if (!drv) { 4075 error_setg(errp, "Unknown driver: '%s'", drvname); 4076 goto fail; 4077 } 4078 } 4079 4080 assert(drvname || !(flags & BDRV_O_PROTOCOL)); 4081 4082 /* See cautionary note on accessing @options above */ 4083 backing = qdict_get_try_str(options, "backing"); 4084 if (qobject_to(QNull, qdict_get(options, "backing")) != NULL || 4085 (backing && *backing == '\0')) 4086 { 4087 if (backing) { 4088 warn_report("Use of \"backing\": \"\" is deprecated; " 4089 "use \"backing\": null instead"); 4090 } 4091 flags |= BDRV_O_NO_BACKING; 4092 qdict_del(bs->explicit_options, "backing"); 4093 qdict_del(bs->options, "backing"); 4094 qdict_del(options, "backing"); 4095 } 4096 4097 /* Open image file without format layer. This BlockBackend is only used for 4098 * probing, the block drivers will do their own bdrv_open_child() for the 4099 * same BDS, which is why we put the node name back into options. */ 4100 if ((flags & BDRV_O_PROTOCOL) == 0) { 4101 BlockDriverState *file_bs; 4102 4103 file_bs = bdrv_open_child_bs(filename, options, "file", bs, 4104 &child_of_bds, BDRV_CHILD_IMAGE, 4105 true, &local_err); 4106 if (local_err) { 4107 goto fail; 4108 } 4109 if (file_bs != NULL) { 4110 /* Not requesting BLK_PERM_CONSISTENT_READ because we're only 4111 * looking at the header to guess the image format. This works even 4112 * in cases where a guest would not see a consistent state. */ 4113 ctx = bdrv_get_aio_context(file_bs); 4114 aio_context_acquire(ctx); 4115 file = blk_new(ctx, 0, BLK_PERM_ALL); 4116 blk_insert_bs(file, file_bs, &local_err); 4117 bdrv_unref(file_bs); 4118 aio_context_release(ctx); 4119 4120 if (local_err) { 4121 goto fail; 4122 } 4123 4124 qdict_put_str(options, "file", bdrv_get_node_name(file_bs)); 4125 } 4126 } 4127 4128 /* Image format probing */ 4129 bs->probed = !drv; 4130 if (!drv && file) { 4131 ret = find_image_format(file, filename, &drv, &local_err); 4132 if (ret < 0) { 4133 goto fail; 4134 } 4135 /* 4136 * This option update would logically belong in bdrv_fill_options(), 4137 * but we first need to open bs->file for the probing to work, while 4138 * opening bs->file already requires the (mostly) final set of options 4139 * so that cache mode etc. can be inherited. 4140 * 4141 * Adding the driver later is somewhat ugly, but it's not an option 4142 * that would ever be inherited, so it's correct. We just need to make 4143 * sure to update both bs->options (which has the full effective 4144 * options for bs) and options (which has file.* already removed). 4145 */ 4146 qdict_put_str(bs->options, "driver", drv->format_name); 4147 qdict_put_str(options, "driver", drv->format_name); 4148 } else if (!drv) { 4149 error_setg(errp, "Must specify either driver or file"); 4150 goto fail; 4151 } 4152 4153 /* BDRV_O_PROTOCOL must be set iff a protocol BDS is about to be created */ 4154 assert(!!(flags & BDRV_O_PROTOCOL) == !!drv->bdrv_file_open); 4155 /* file must be NULL if a protocol BDS is about to be created 4156 * (the inverse results in an error message from bdrv_open_common()) */ 4157 assert(!(flags & BDRV_O_PROTOCOL) || !file); 4158 4159 /* Open the image */ 4160 ret = bdrv_open_common(bs, file, options, &local_err); 4161 if (ret < 0) { 4162 goto fail; 4163 } 4164 4165 /* The AioContext could have changed during bdrv_open_common() */ 4166 ctx = bdrv_get_aio_context(bs); 4167 4168 if (file) { 4169 aio_context_acquire(ctx); 4170 blk_unref(file); 4171 aio_context_release(ctx); 4172 file = NULL; 4173 } 4174 4175 /* If there is a backing file, use it */ 4176 if ((flags & BDRV_O_NO_BACKING) == 0) { 4177 ret = bdrv_open_backing_file(bs, options, "backing", &local_err); 4178 if (ret < 0) { 4179 goto close_and_fail; 4180 } 4181 } 4182 4183 /* Remove all children options and references 4184 * from bs->options and bs->explicit_options */ 4185 QLIST_FOREACH(child, &bs->children, next) { 4186 char *child_key_dot; 4187 child_key_dot = g_strdup_printf("%s.", child->name); 4188 qdict_extract_subqdict(bs->explicit_options, NULL, child_key_dot); 4189 qdict_extract_subqdict(bs->options, NULL, child_key_dot); 4190 qdict_del(bs->explicit_options, child->name); 4191 qdict_del(bs->options, child->name); 4192 g_free(child_key_dot); 4193 } 4194 4195 /* Check if any unknown options were used */ 4196 if (qdict_size(options) != 0) { 4197 const QDictEntry *entry = qdict_first(options); 4198 if (flags & BDRV_O_PROTOCOL) { 4199 error_setg(errp, "Block protocol '%s' doesn't support the option " 4200 "'%s'", drv->format_name, entry->key); 4201 } else { 4202 error_setg(errp, 4203 "Block format '%s' does not support the option '%s'", 4204 drv->format_name, entry->key); 4205 } 4206 4207 goto close_and_fail; 4208 } 4209 4210 bdrv_parent_cb_change_media(bs, true); 4211 4212 qobject_unref(options); 4213 options = NULL; 4214 4215 /* For snapshot=on, create a temporary qcow2 overlay. bs points to the 4216 * temporary snapshot afterwards. */ 4217 if (snapshot_flags) { 4218 BlockDriverState *snapshot_bs; 4219 snapshot_bs = bdrv_append_temp_snapshot(bs, snapshot_flags, 4220 snapshot_options, &local_err); 4221 snapshot_options = NULL; 4222 if (local_err) { 4223 goto close_and_fail; 4224 } 4225 /* We are not going to return bs but the overlay on top of it 4226 * (snapshot_bs); thus, we have to drop the strong reference to bs 4227 * (which we obtained by calling bdrv_new()). bs will not be deleted, 4228 * though, because the overlay still has a reference to it. */ 4229 aio_context_acquire(ctx); 4230 bdrv_unref(bs); 4231 aio_context_release(ctx); 4232 bs = snapshot_bs; 4233 } 4234 4235 return bs; 4236 4237 fail: 4238 aio_context_acquire(ctx); 4239 blk_unref(file); 4240 qobject_unref(snapshot_options); 4241 qobject_unref(bs->explicit_options); 4242 qobject_unref(bs->options); 4243 qobject_unref(options); 4244 bs->options = NULL; 4245 bs->explicit_options = NULL; 4246 bdrv_unref(bs); 4247 aio_context_release(ctx); 4248 error_propagate(errp, local_err); 4249 return NULL; 4250 4251 close_and_fail: 4252 aio_context_acquire(ctx); 4253 bdrv_unref(bs); 4254 aio_context_release(ctx); 4255 qobject_unref(snapshot_options); 4256 qobject_unref(options); 4257 error_propagate(errp, local_err); 4258 return NULL; 4259 } 4260 4261 /* The caller must always hold the main AioContext lock. */ 4262 BlockDriverState *bdrv_open(const char *filename, const char *reference, 4263 QDict *options, int flags, Error **errp) 4264 { 4265 GLOBAL_STATE_CODE(); 4266 4267 return bdrv_open_inherit(filename, reference, options, flags, NULL, 4268 NULL, 0, errp); 4269 } 4270 4271 /* Return true if the NULL-terminated @list contains @str */ 4272 static bool is_str_in_list(const char *str, const char *const *list) 4273 { 4274 if (str && list) { 4275 int i; 4276 for (i = 0; list[i] != NULL; i++) { 4277 if (!strcmp(str, list[i])) { 4278 return true; 4279 } 4280 } 4281 } 4282 return false; 4283 } 4284 4285 /* 4286 * Check that every option set in @bs->options is also set in 4287 * @new_opts. 4288 * 4289 * Options listed in the common_options list and in 4290 * @bs->drv->mutable_opts are skipped. 4291 * 4292 * Return 0 on success, otherwise return -EINVAL and set @errp. 4293 */ 4294 static int bdrv_reset_options_allowed(BlockDriverState *bs, 4295 const QDict *new_opts, Error **errp) 4296 { 4297 const QDictEntry *e; 4298 /* These options are common to all block drivers and are handled 4299 * in bdrv_reopen_prepare() so they can be left out of @new_opts */ 4300 const char *const common_options[] = { 4301 "node-name", "discard", "cache.direct", "cache.no-flush", 4302 "read-only", "auto-read-only", "detect-zeroes", NULL 4303 }; 4304 4305 for (e = qdict_first(bs->options); e; e = qdict_next(bs->options, e)) { 4306 if (!qdict_haskey(new_opts, e->key) && 4307 !is_str_in_list(e->key, common_options) && 4308 !is_str_in_list(e->key, bs->drv->mutable_opts)) { 4309 error_setg(errp, "Option '%s' cannot be reset " 4310 "to its default value", e->key); 4311 return -EINVAL; 4312 } 4313 } 4314 4315 return 0; 4316 } 4317 4318 /* 4319 * Returns true if @child can be reached recursively from @bs 4320 */ 4321 static bool GRAPH_RDLOCK 4322 bdrv_recurse_has_child(BlockDriverState *bs, BlockDriverState *child) 4323 { 4324 BdrvChild *c; 4325 4326 if (bs == child) { 4327 return true; 4328 } 4329 4330 QLIST_FOREACH(c, &bs->children, next) { 4331 if (bdrv_recurse_has_child(c->bs, child)) { 4332 return true; 4333 } 4334 } 4335 4336 return false; 4337 } 4338 4339 /* 4340 * Adds a BlockDriverState to a simple queue for an atomic, transactional 4341 * reopen of multiple devices. 4342 * 4343 * bs_queue can either be an existing BlockReopenQueue that has had QTAILQ_INIT 4344 * already performed, or alternatively may be NULL a new BlockReopenQueue will 4345 * be created and initialized. This newly created BlockReopenQueue should be 4346 * passed back in for subsequent calls that are intended to be of the same 4347 * atomic 'set'. 4348 * 4349 * bs is the BlockDriverState to add to the reopen queue. 4350 * 4351 * options contains the changed options for the associated bs 4352 * (the BlockReopenQueue takes ownership) 4353 * 4354 * flags contains the open flags for the associated bs 4355 * 4356 * returns a pointer to bs_queue, which is either the newly allocated 4357 * bs_queue, or the existing bs_queue being used. 4358 * 4359 * bs is drained here and undrained by bdrv_reopen_queue_free(). 4360 * 4361 * To be called with bs->aio_context locked. 4362 */ 4363 static BlockReopenQueue * GRAPH_RDLOCK 4364 bdrv_reopen_queue_child(BlockReopenQueue *bs_queue, BlockDriverState *bs, 4365 QDict *options, const BdrvChildClass *klass, 4366 BdrvChildRole role, bool parent_is_format, 4367 QDict *parent_options, int parent_flags, 4368 bool keep_old_opts) 4369 { 4370 assert(bs != NULL); 4371 4372 BlockReopenQueueEntry *bs_entry; 4373 BdrvChild *child; 4374 QDict *old_options, *explicit_options, *options_copy; 4375 int flags; 4376 QemuOpts *opts; 4377 4378 GLOBAL_STATE_CODE(); 4379 4380 /* 4381 * Strictly speaking, draining is illegal under GRAPH_RDLOCK. We know that 4382 * we've been called with bdrv_graph_rdlock_main_loop(), though, so it's ok 4383 * in practice. 4384 */ 4385 bdrv_drained_begin(bs); 4386 4387 if (bs_queue == NULL) { 4388 bs_queue = g_new0(BlockReopenQueue, 1); 4389 QTAILQ_INIT(bs_queue); 4390 } 4391 4392 if (!options) { 4393 options = qdict_new(); 4394 } 4395 4396 /* Check if this BlockDriverState is already in the queue */ 4397 QTAILQ_FOREACH(bs_entry, bs_queue, entry) { 4398 if (bs == bs_entry->state.bs) { 4399 break; 4400 } 4401 } 4402 4403 /* 4404 * Precedence of options: 4405 * 1. Explicitly passed in options (highest) 4406 * 2. Retained from explicitly set options of bs 4407 * 3. Inherited from parent node 4408 * 4. Retained from effective options of bs 4409 */ 4410 4411 /* Old explicitly set values (don't overwrite by inherited value) */ 4412 if (bs_entry || keep_old_opts) { 4413 old_options = qdict_clone_shallow(bs_entry ? 4414 bs_entry->state.explicit_options : 4415 bs->explicit_options); 4416 bdrv_join_options(bs, options, old_options); 4417 qobject_unref(old_options); 4418 } 4419 4420 explicit_options = qdict_clone_shallow(options); 4421 4422 /* Inherit from parent node */ 4423 if (parent_options) { 4424 flags = 0; 4425 klass->inherit_options(role, parent_is_format, &flags, options, 4426 parent_flags, parent_options); 4427 } else { 4428 flags = bdrv_get_flags(bs); 4429 } 4430 4431 if (keep_old_opts) { 4432 /* Old values are used for options that aren't set yet */ 4433 old_options = qdict_clone_shallow(bs->options); 4434 bdrv_join_options(bs, options, old_options); 4435 qobject_unref(old_options); 4436 } 4437 4438 /* We have the final set of options so let's update the flags */ 4439 options_copy = qdict_clone_shallow(options); 4440 opts = qemu_opts_create(&bdrv_runtime_opts, NULL, 0, &error_abort); 4441 qemu_opts_absorb_qdict(opts, options_copy, NULL); 4442 update_flags_from_options(&flags, opts); 4443 qemu_opts_del(opts); 4444 qobject_unref(options_copy); 4445 4446 /* bdrv_open_inherit() sets and clears some additional flags internally */ 4447 flags &= ~BDRV_O_PROTOCOL; 4448 if (flags & BDRV_O_RDWR) { 4449 flags |= BDRV_O_ALLOW_RDWR; 4450 } 4451 4452 if (!bs_entry) { 4453 bs_entry = g_new0(BlockReopenQueueEntry, 1); 4454 QTAILQ_INSERT_TAIL(bs_queue, bs_entry, entry); 4455 } else { 4456 qobject_unref(bs_entry->state.options); 4457 qobject_unref(bs_entry->state.explicit_options); 4458 } 4459 4460 bs_entry->state.bs = bs; 4461 bs_entry->state.options = options; 4462 bs_entry->state.explicit_options = explicit_options; 4463 bs_entry->state.flags = flags; 4464 4465 /* 4466 * If keep_old_opts is false then it means that unspecified 4467 * options must be reset to their original value. We don't allow 4468 * resetting 'backing' but we need to know if the option is 4469 * missing in order to decide if we have to return an error. 4470 */ 4471 if (!keep_old_opts) { 4472 bs_entry->state.backing_missing = 4473 !qdict_haskey(options, "backing") && 4474 !qdict_haskey(options, "backing.driver"); 4475 } 4476 4477 QLIST_FOREACH(child, &bs->children, next) { 4478 QDict *new_child_options = NULL; 4479 bool child_keep_old = keep_old_opts; 4480 4481 /* reopen can only change the options of block devices that were 4482 * implicitly created and inherited options. For other (referenced) 4483 * block devices, a syntax like "backing.foo" results in an error. */ 4484 if (child->bs->inherits_from != bs) { 4485 continue; 4486 } 4487 4488 /* Check if the options contain a child reference */ 4489 if (qdict_haskey(options, child->name)) { 4490 const char *childref = qdict_get_try_str(options, child->name); 4491 /* 4492 * The current child must not be reopened if the child 4493 * reference is null or points to a different node. 4494 */ 4495 if (g_strcmp0(childref, child->bs->node_name)) { 4496 continue; 4497 } 4498 /* 4499 * If the child reference points to the current child then 4500 * reopen it with its existing set of options (note that 4501 * it can still inherit new options from the parent). 4502 */ 4503 child_keep_old = true; 4504 } else { 4505 /* Extract child options ("child-name.*") */ 4506 char *child_key_dot = g_strdup_printf("%s.", child->name); 4507 qdict_extract_subqdict(explicit_options, NULL, child_key_dot); 4508 qdict_extract_subqdict(options, &new_child_options, child_key_dot); 4509 g_free(child_key_dot); 4510 } 4511 4512 bdrv_reopen_queue_child(bs_queue, child->bs, new_child_options, 4513 child->klass, child->role, bs->drv->is_format, 4514 options, flags, child_keep_old); 4515 } 4516 4517 return bs_queue; 4518 } 4519 4520 /* To be called with bs->aio_context locked */ 4521 BlockReopenQueue *bdrv_reopen_queue(BlockReopenQueue *bs_queue, 4522 BlockDriverState *bs, 4523 QDict *options, bool keep_old_opts) 4524 { 4525 GLOBAL_STATE_CODE(); 4526 GRAPH_RDLOCK_GUARD_MAINLOOP(); 4527 4528 return bdrv_reopen_queue_child(bs_queue, bs, options, NULL, 0, false, 4529 NULL, 0, keep_old_opts); 4530 } 4531 4532 void bdrv_reopen_queue_free(BlockReopenQueue *bs_queue) 4533 { 4534 GLOBAL_STATE_CODE(); 4535 if (bs_queue) { 4536 BlockReopenQueueEntry *bs_entry, *next; 4537 QTAILQ_FOREACH_SAFE(bs_entry, bs_queue, entry, next) { 4538 AioContext *ctx = bdrv_get_aio_context(bs_entry->state.bs); 4539 4540 aio_context_acquire(ctx); 4541 bdrv_drained_end(bs_entry->state.bs); 4542 aio_context_release(ctx); 4543 4544 qobject_unref(bs_entry->state.explicit_options); 4545 qobject_unref(bs_entry->state.options); 4546 g_free(bs_entry); 4547 } 4548 g_free(bs_queue); 4549 } 4550 } 4551 4552 /* 4553 * Reopen multiple BlockDriverStates atomically & transactionally. 4554 * 4555 * The queue passed in (bs_queue) must have been built up previous 4556 * via bdrv_reopen_queue(). 4557 * 4558 * Reopens all BDS specified in the queue, with the appropriate 4559 * flags. All devices are prepared for reopen, and failure of any 4560 * device will cause all device changes to be abandoned, and intermediate 4561 * data cleaned up. 4562 * 4563 * If all devices prepare successfully, then the changes are committed 4564 * to all devices. 4565 * 4566 * All affected nodes must be drained between bdrv_reopen_queue() and 4567 * bdrv_reopen_multiple(). 4568 * 4569 * To be called from the main thread, with all other AioContexts unlocked. 4570 */ 4571 int bdrv_reopen_multiple(BlockReopenQueue *bs_queue, Error **errp) 4572 { 4573 int ret = -1; 4574 BlockReopenQueueEntry *bs_entry, *next; 4575 AioContext *ctx; 4576 Transaction *tran = tran_new(); 4577 g_autoptr(GSList) refresh_list = NULL; 4578 4579 assert(qemu_get_current_aio_context() == qemu_get_aio_context()); 4580 assert(bs_queue != NULL); 4581 GLOBAL_STATE_CODE(); 4582 4583 QTAILQ_FOREACH(bs_entry, bs_queue, entry) { 4584 ctx = bdrv_get_aio_context(bs_entry->state.bs); 4585 aio_context_acquire(ctx); 4586 ret = bdrv_flush(bs_entry->state.bs); 4587 aio_context_release(ctx); 4588 if (ret < 0) { 4589 error_setg_errno(errp, -ret, "Error flushing drive"); 4590 goto abort; 4591 } 4592 } 4593 4594 QTAILQ_FOREACH(bs_entry, bs_queue, entry) { 4595 assert(bs_entry->state.bs->quiesce_counter > 0); 4596 ctx = bdrv_get_aio_context(bs_entry->state.bs); 4597 aio_context_acquire(ctx); 4598 ret = bdrv_reopen_prepare(&bs_entry->state, bs_queue, tran, errp); 4599 aio_context_release(ctx); 4600 if (ret < 0) { 4601 goto abort; 4602 } 4603 bs_entry->prepared = true; 4604 } 4605 4606 QTAILQ_FOREACH(bs_entry, bs_queue, entry) { 4607 BDRVReopenState *state = &bs_entry->state; 4608 4609 refresh_list = g_slist_prepend(refresh_list, state->bs); 4610 if (state->old_backing_bs) { 4611 refresh_list = g_slist_prepend(refresh_list, state->old_backing_bs); 4612 } 4613 if (state->old_file_bs) { 4614 refresh_list = g_slist_prepend(refresh_list, state->old_file_bs); 4615 } 4616 } 4617 4618 /* 4619 * Note that file-posix driver rely on permission update done during reopen 4620 * (even if no permission changed), because it wants "new" permissions for 4621 * reconfiguring the fd and that's why it does it in raw_check_perm(), not 4622 * in raw_reopen_prepare() which is called with "old" permissions. 4623 */ 4624 bdrv_graph_rdlock_main_loop(); 4625 ret = bdrv_list_refresh_perms(refresh_list, bs_queue, tran, errp); 4626 bdrv_graph_rdunlock_main_loop(); 4627 4628 if (ret < 0) { 4629 goto abort; 4630 } 4631 4632 /* 4633 * If we reach this point, we have success and just need to apply the 4634 * changes. 4635 * 4636 * Reverse order is used to comfort qcow2 driver: on commit it need to write 4637 * IN_USE flag to the image, to mark bitmaps in the image as invalid. But 4638 * children are usually goes after parents in reopen-queue, so go from last 4639 * to first element. 4640 */ 4641 QTAILQ_FOREACH_REVERSE(bs_entry, bs_queue, entry) { 4642 ctx = bdrv_get_aio_context(bs_entry->state.bs); 4643 aio_context_acquire(ctx); 4644 bdrv_reopen_commit(&bs_entry->state); 4645 aio_context_release(ctx); 4646 } 4647 4648 bdrv_graph_wrlock(NULL); 4649 tran_commit(tran); 4650 bdrv_graph_wrunlock(); 4651 4652 QTAILQ_FOREACH_REVERSE(bs_entry, bs_queue, entry) { 4653 BlockDriverState *bs = bs_entry->state.bs; 4654 4655 if (bs->drv->bdrv_reopen_commit_post) { 4656 ctx = bdrv_get_aio_context(bs); 4657 aio_context_acquire(ctx); 4658 bs->drv->bdrv_reopen_commit_post(&bs_entry->state); 4659 aio_context_release(ctx); 4660 } 4661 } 4662 4663 ret = 0; 4664 goto cleanup; 4665 4666 abort: 4667 bdrv_graph_wrlock(NULL); 4668 tran_abort(tran); 4669 bdrv_graph_wrunlock(); 4670 4671 QTAILQ_FOREACH_SAFE(bs_entry, bs_queue, entry, next) { 4672 if (bs_entry->prepared) { 4673 ctx = bdrv_get_aio_context(bs_entry->state.bs); 4674 aio_context_acquire(ctx); 4675 bdrv_reopen_abort(&bs_entry->state); 4676 aio_context_release(ctx); 4677 } 4678 } 4679 4680 cleanup: 4681 bdrv_reopen_queue_free(bs_queue); 4682 4683 return ret; 4684 } 4685 4686 int bdrv_reopen(BlockDriverState *bs, QDict *opts, bool keep_old_opts, 4687 Error **errp) 4688 { 4689 AioContext *ctx = bdrv_get_aio_context(bs); 4690 BlockReopenQueue *queue; 4691 int ret; 4692 4693 GLOBAL_STATE_CODE(); 4694 4695 queue = bdrv_reopen_queue(NULL, bs, opts, keep_old_opts); 4696 4697 if (ctx != qemu_get_aio_context()) { 4698 aio_context_release(ctx); 4699 } 4700 ret = bdrv_reopen_multiple(queue, errp); 4701 4702 if (ctx != qemu_get_aio_context()) { 4703 aio_context_acquire(ctx); 4704 } 4705 4706 return ret; 4707 } 4708 4709 int bdrv_reopen_set_read_only(BlockDriverState *bs, bool read_only, 4710 Error **errp) 4711 { 4712 QDict *opts = qdict_new(); 4713 4714 GLOBAL_STATE_CODE(); 4715 4716 qdict_put_bool(opts, BDRV_OPT_READ_ONLY, read_only); 4717 4718 return bdrv_reopen(bs, opts, true, errp); 4719 } 4720 4721 /* 4722 * Take a BDRVReopenState and check if the value of 'backing' in the 4723 * reopen_state->options QDict is valid or not. 4724 * 4725 * If 'backing' is missing from the QDict then return 0. 4726 * 4727 * If 'backing' contains the node name of the backing file of 4728 * reopen_state->bs then return 0. 4729 * 4730 * If 'backing' contains a different node name (or is null) then check 4731 * whether the current backing file can be replaced with the new one. 4732 * If that's the case then reopen_state->replace_backing_bs is set to 4733 * true and reopen_state->new_backing_bs contains a pointer to the new 4734 * backing BlockDriverState (or NULL). 4735 * 4736 * After calling this function, the transaction @tran may only be completed 4737 * while holding a writer lock for the graph. 4738 * 4739 * Return 0 on success, otherwise return < 0 and set @errp. 4740 * 4741 * The caller must hold the AioContext lock of @reopen_state->bs. 4742 * @reopen_state->bs can move to a different AioContext in this function. 4743 * Callers must make sure that their AioContext locking is still correct after 4744 * this. 4745 */ 4746 static int GRAPH_UNLOCKED 4747 bdrv_reopen_parse_file_or_backing(BDRVReopenState *reopen_state, 4748 bool is_backing, Transaction *tran, 4749 Error **errp) 4750 { 4751 BlockDriverState *bs = reopen_state->bs; 4752 BlockDriverState *new_child_bs; 4753 BlockDriverState *old_child_bs = is_backing ? child_bs(bs->backing) : 4754 child_bs(bs->file); 4755 const char *child_name = is_backing ? "backing" : "file"; 4756 QObject *value; 4757 const char *str; 4758 AioContext *ctx, *old_ctx; 4759 bool has_child; 4760 int ret; 4761 4762 GLOBAL_STATE_CODE(); 4763 4764 value = qdict_get(reopen_state->options, child_name); 4765 if (value == NULL) { 4766 return 0; 4767 } 4768 4769 bdrv_graph_rdlock_main_loop(); 4770 4771 switch (qobject_type(value)) { 4772 case QTYPE_QNULL: 4773 assert(is_backing); /* The 'file' option does not allow a null value */ 4774 new_child_bs = NULL; 4775 break; 4776 case QTYPE_QSTRING: 4777 str = qstring_get_str(qobject_to(QString, value)); 4778 new_child_bs = bdrv_lookup_bs(NULL, str, errp); 4779 if (new_child_bs == NULL) { 4780 ret = -EINVAL; 4781 goto out_rdlock; 4782 } 4783 4784 has_child = bdrv_recurse_has_child(new_child_bs, bs); 4785 if (has_child) { 4786 error_setg(errp, "Making '%s' a %s child of '%s' would create a " 4787 "cycle", str, child_name, bs->node_name); 4788 ret = -EINVAL; 4789 goto out_rdlock; 4790 } 4791 break; 4792 default: 4793 /* 4794 * The options QDict has been flattened, so 'backing' and 'file' 4795 * do not allow any other data type here. 4796 */ 4797 g_assert_not_reached(); 4798 } 4799 4800 if (old_child_bs == new_child_bs) { 4801 ret = 0; 4802 goto out_rdlock; 4803 } 4804 4805 if (old_child_bs) { 4806 if (bdrv_skip_implicit_filters(old_child_bs) == new_child_bs) { 4807 ret = 0; 4808 goto out_rdlock; 4809 } 4810 4811 if (old_child_bs->implicit) { 4812 error_setg(errp, "Cannot replace implicit %s child of %s", 4813 child_name, bs->node_name); 4814 ret = -EPERM; 4815 goto out_rdlock; 4816 } 4817 } 4818 4819 if (bs->drv->is_filter && !old_child_bs) { 4820 /* 4821 * Filters always have a file or a backing child, so we are trying to 4822 * change wrong child 4823 */ 4824 error_setg(errp, "'%s' is a %s filter node that does not support a " 4825 "%s child", bs->node_name, bs->drv->format_name, child_name); 4826 ret = -EINVAL; 4827 goto out_rdlock; 4828 } 4829 4830 if (is_backing) { 4831 reopen_state->old_backing_bs = old_child_bs; 4832 } else { 4833 reopen_state->old_file_bs = old_child_bs; 4834 } 4835 4836 if (old_child_bs) { 4837 bdrv_ref(old_child_bs); 4838 bdrv_drained_begin(old_child_bs); 4839 } 4840 4841 old_ctx = bdrv_get_aio_context(bs); 4842 ctx = bdrv_get_aio_context(new_child_bs); 4843 if (old_ctx != ctx) { 4844 aio_context_release(old_ctx); 4845 aio_context_acquire(ctx); 4846 } 4847 4848 bdrv_graph_rdunlock_main_loop(); 4849 bdrv_graph_wrlock(new_child_bs); 4850 4851 ret = bdrv_set_file_or_backing_noperm(bs, new_child_bs, is_backing, 4852 tran, errp); 4853 4854 bdrv_graph_wrunlock(); 4855 4856 if (old_ctx != ctx) { 4857 aio_context_release(ctx); 4858 aio_context_acquire(old_ctx); 4859 } 4860 4861 if (old_child_bs) { 4862 bdrv_drained_end(old_child_bs); 4863 bdrv_unref(old_child_bs); 4864 } 4865 4866 return ret; 4867 4868 out_rdlock: 4869 bdrv_graph_rdunlock_main_loop(); 4870 return ret; 4871 } 4872 4873 /* 4874 * Prepares a BlockDriverState for reopen. All changes are staged in the 4875 * 'opaque' field of the BDRVReopenState, which is used and allocated by 4876 * the block driver layer .bdrv_reopen_prepare() 4877 * 4878 * bs is the BlockDriverState to reopen 4879 * flags are the new open flags 4880 * queue is the reopen queue 4881 * 4882 * Returns 0 on success, non-zero on error. On error errp will be set 4883 * as well. 4884 * 4885 * On failure, bdrv_reopen_abort() will be called to clean up any data. 4886 * It is the responsibility of the caller to then call the abort() or 4887 * commit() for any other BDS that have been left in a prepare() state 4888 * 4889 * The caller must hold the AioContext lock of @reopen_state->bs. 4890 * 4891 * After calling this function, the transaction @change_child_tran may only be 4892 * completed while holding a writer lock for the graph. 4893 */ 4894 static int GRAPH_UNLOCKED 4895 bdrv_reopen_prepare(BDRVReopenState *reopen_state, BlockReopenQueue *queue, 4896 Transaction *change_child_tran, Error **errp) 4897 { 4898 int ret = -1; 4899 int old_flags; 4900 Error *local_err = NULL; 4901 BlockDriver *drv; 4902 QemuOpts *opts; 4903 QDict *orig_reopen_opts; 4904 char *discard = NULL; 4905 bool read_only; 4906 bool drv_prepared = false; 4907 4908 assert(reopen_state != NULL); 4909 assert(reopen_state->bs->drv != NULL); 4910 GLOBAL_STATE_CODE(); 4911 drv = reopen_state->bs->drv; 4912 4913 /* This function and each driver's bdrv_reopen_prepare() remove 4914 * entries from reopen_state->options as they are processed, so 4915 * we need to make a copy of the original QDict. */ 4916 orig_reopen_opts = qdict_clone_shallow(reopen_state->options); 4917 4918 /* Process generic block layer options */ 4919 opts = qemu_opts_create(&bdrv_runtime_opts, NULL, 0, &error_abort); 4920 if (!qemu_opts_absorb_qdict(opts, reopen_state->options, errp)) { 4921 ret = -EINVAL; 4922 goto error; 4923 } 4924 4925 /* This was already called in bdrv_reopen_queue_child() so the flags 4926 * are up-to-date. This time we simply want to remove the options from 4927 * QemuOpts in order to indicate that they have been processed. */ 4928 old_flags = reopen_state->flags; 4929 update_flags_from_options(&reopen_state->flags, opts); 4930 assert(old_flags == reopen_state->flags); 4931 4932 discard = qemu_opt_get_del(opts, BDRV_OPT_DISCARD); 4933 if (discard != NULL) { 4934 if (bdrv_parse_discard_flags(discard, &reopen_state->flags) != 0) { 4935 error_setg(errp, "Invalid discard option"); 4936 ret = -EINVAL; 4937 goto error; 4938 } 4939 } 4940 4941 reopen_state->detect_zeroes = 4942 bdrv_parse_detect_zeroes(opts, reopen_state->flags, &local_err); 4943 if (local_err) { 4944 error_propagate(errp, local_err); 4945 ret = -EINVAL; 4946 goto error; 4947 } 4948 4949 /* All other options (including node-name and driver) must be unchanged. 4950 * Put them back into the QDict, so that they are checked at the end 4951 * of this function. */ 4952 qemu_opts_to_qdict(opts, reopen_state->options); 4953 4954 /* If we are to stay read-only, do not allow permission change 4955 * to r/w. Attempting to set to r/w may fail if either BDRV_O_ALLOW_RDWR is 4956 * not set, or if the BDS still has copy_on_read enabled */ 4957 read_only = !(reopen_state->flags & BDRV_O_RDWR); 4958 4959 bdrv_graph_rdlock_main_loop(); 4960 ret = bdrv_can_set_read_only(reopen_state->bs, read_only, true, &local_err); 4961 bdrv_graph_rdunlock_main_loop(); 4962 if (local_err) { 4963 error_propagate(errp, local_err); 4964 goto error; 4965 } 4966 4967 if (drv->bdrv_reopen_prepare) { 4968 /* 4969 * If a driver-specific option is missing, it means that we 4970 * should reset it to its default value. 4971 * But not all options allow that, so we need to check it first. 4972 */ 4973 ret = bdrv_reset_options_allowed(reopen_state->bs, 4974 reopen_state->options, errp); 4975 if (ret) { 4976 goto error; 4977 } 4978 4979 ret = drv->bdrv_reopen_prepare(reopen_state, queue, &local_err); 4980 if (ret) { 4981 if (local_err != NULL) { 4982 error_propagate(errp, local_err); 4983 } else { 4984 bdrv_graph_rdlock_main_loop(); 4985 bdrv_refresh_filename(reopen_state->bs); 4986 bdrv_graph_rdunlock_main_loop(); 4987 error_setg(errp, "failed while preparing to reopen image '%s'", 4988 reopen_state->bs->filename); 4989 } 4990 goto error; 4991 } 4992 } else { 4993 /* It is currently mandatory to have a bdrv_reopen_prepare() 4994 * handler for each supported drv. */ 4995 bdrv_graph_rdlock_main_loop(); 4996 error_setg(errp, "Block format '%s' used by node '%s' " 4997 "does not support reopening files", drv->format_name, 4998 bdrv_get_device_or_node_name(reopen_state->bs)); 4999 bdrv_graph_rdunlock_main_loop(); 5000 ret = -1; 5001 goto error; 5002 } 5003 5004 drv_prepared = true; 5005 5006 /* 5007 * We must provide the 'backing' option if the BDS has a backing 5008 * file or if the image file has a backing file name as part of 5009 * its metadata. Otherwise the 'backing' option can be omitted. 5010 */ 5011 if (drv->supports_backing && reopen_state->backing_missing && 5012 (reopen_state->bs->backing || reopen_state->bs->backing_file[0])) { 5013 error_setg(errp, "backing is missing for '%s'", 5014 reopen_state->bs->node_name); 5015 ret = -EINVAL; 5016 goto error; 5017 } 5018 5019 /* 5020 * Allow changing the 'backing' option. The new value can be 5021 * either a reference to an existing node (using its node name) 5022 * or NULL to simply detach the current backing file. 5023 */ 5024 ret = bdrv_reopen_parse_file_or_backing(reopen_state, true, 5025 change_child_tran, errp); 5026 if (ret < 0) { 5027 goto error; 5028 } 5029 qdict_del(reopen_state->options, "backing"); 5030 5031 /* Allow changing the 'file' option. In this case NULL is not allowed */ 5032 ret = bdrv_reopen_parse_file_or_backing(reopen_state, false, 5033 change_child_tran, errp); 5034 if (ret < 0) { 5035 goto error; 5036 } 5037 qdict_del(reopen_state->options, "file"); 5038 5039 /* Options that are not handled are only okay if they are unchanged 5040 * compared to the old state. It is expected that some options are only 5041 * used for the initial open, but not reopen (e.g. filename) */ 5042 if (qdict_size(reopen_state->options)) { 5043 const QDictEntry *entry = qdict_first(reopen_state->options); 5044 5045 GRAPH_RDLOCK_GUARD_MAINLOOP(); 5046 5047 do { 5048 QObject *new = entry->value; 5049 QObject *old = qdict_get(reopen_state->bs->options, entry->key); 5050 5051 /* Allow child references (child_name=node_name) as long as they 5052 * point to the current child (i.e. everything stays the same). */ 5053 if (qobject_type(new) == QTYPE_QSTRING) { 5054 BdrvChild *child; 5055 QLIST_FOREACH(child, &reopen_state->bs->children, next) { 5056 if (!strcmp(child->name, entry->key)) { 5057 break; 5058 } 5059 } 5060 5061 if (child) { 5062 if (!strcmp(child->bs->node_name, 5063 qstring_get_str(qobject_to(QString, new)))) { 5064 continue; /* Found child with this name, skip option */ 5065 } 5066 } 5067 } 5068 5069 /* 5070 * TODO: When using -drive to specify blockdev options, all values 5071 * will be strings; however, when using -blockdev, blockdev-add or 5072 * filenames using the json:{} pseudo-protocol, they will be 5073 * correctly typed. 5074 * In contrast, reopening options are (currently) always strings 5075 * (because you can only specify them through qemu-io; all other 5076 * callers do not specify any options). 5077 * Therefore, when using anything other than -drive to create a BDS, 5078 * this cannot detect non-string options as unchanged, because 5079 * qobject_is_equal() always returns false for objects of different 5080 * type. In the future, this should be remedied by correctly typing 5081 * all options. For now, this is not too big of an issue because 5082 * the user can simply omit options which cannot be changed anyway, 5083 * so they will stay unchanged. 5084 */ 5085 if (!qobject_is_equal(new, old)) { 5086 error_setg(errp, "Cannot change the option '%s'", entry->key); 5087 ret = -EINVAL; 5088 goto error; 5089 } 5090 } while ((entry = qdict_next(reopen_state->options, entry))); 5091 } 5092 5093 ret = 0; 5094 5095 /* Restore the original reopen_state->options QDict */ 5096 qobject_unref(reopen_state->options); 5097 reopen_state->options = qobject_ref(orig_reopen_opts); 5098 5099 error: 5100 if (ret < 0 && drv_prepared) { 5101 /* drv->bdrv_reopen_prepare() has succeeded, so we need to 5102 * call drv->bdrv_reopen_abort() before signaling an error 5103 * (bdrv_reopen_multiple() will not call bdrv_reopen_abort() 5104 * when the respective bdrv_reopen_prepare() has failed) */ 5105 if (drv->bdrv_reopen_abort) { 5106 drv->bdrv_reopen_abort(reopen_state); 5107 } 5108 } 5109 qemu_opts_del(opts); 5110 qobject_unref(orig_reopen_opts); 5111 g_free(discard); 5112 return ret; 5113 } 5114 5115 /* 5116 * Takes the staged changes for the reopen from bdrv_reopen_prepare(), and 5117 * makes them final by swapping the staging BlockDriverState contents into 5118 * the active BlockDriverState contents. 5119 */ 5120 static void GRAPH_UNLOCKED bdrv_reopen_commit(BDRVReopenState *reopen_state) 5121 { 5122 BlockDriver *drv; 5123 BlockDriverState *bs; 5124 BdrvChild *child; 5125 5126 assert(reopen_state != NULL); 5127 bs = reopen_state->bs; 5128 drv = bs->drv; 5129 assert(drv != NULL); 5130 GLOBAL_STATE_CODE(); 5131 5132 /* If there are any driver level actions to take */ 5133 if (drv->bdrv_reopen_commit) { 5134 drv->bdrv_reopen_commit(reopen_state); 5135 } 5136 5137 GRAPH_RDLOCK_GUARD_MAINLOOP(); 5138 5139 /* set BDS specific flags now */ 5140 qobject_unref(bs->explicit_options); 5141 qobject_unref(bs->options); 5142 qobject_ref(reopen_state->explicit_options); 5143 qobject_ref(reopen_state->options); 5144 5145 bs->explicit_options = reopen_state->explicit_options; 5146 bs->options = reopen_state->options; 5147 bs->open_flags = reopen_state->flags; 5148 bs->detect_zeroes = reopen_state->detect_zeroes; 5149 5150 /* Remove child references from bs->options and bs->explicit_options. 5151 * Child options were already removed in bdrv_reopen_queue_child() */ 5152 QLIST_FOREACH(child, &bs->children, next) { 5153 qdict_del(bs->explicit_options, child->name); 5154 qdict_del(bs->options, child->name); 5155 } 5156 /* backing is probably removed, so it's not handled by previous loop */ 5157 qdict_del(bs->explicit_options, "backing"); 5158 qdict_del(bs->options, "backing"); 5159 5160 bdrv_refresh_limits(bs, NULL, NULL); 5161 bdrv_refresh_total_sectors(bs, bs->total_sectors); 5162 } 5163 5164 /* 5165 * Abort the reopen, and delete and free the staged changes in 5166 * reopen_state 5167 */ 5168 static void GRAPH_UNLOCKED bdrv_reopen_abort(BDRVReopenState *reopen_state) 5169 { 5170 BlockDriver *drv; 5171 5172 assert(reopen_state != NULL); 5173 drv = reopen_state->bs->drv; 5174 assert(drv != NULL); 5175 GLOBAL_STATE_CODE(); 5176 5177 if (drv->bdrv_reopen_abort) { 5178 drv->bdrv_reopen_abort(reopen_state); 5179 } 5180 } 5181 5182 5183 static void bdrv_close(BlockDriverState *bs) 5184 { 5185 BdrvAioNotifier *ban, *ban_next; 5186 BdrvChild *child, *next; 5187 5188 GLOBAL_STATE_CODE(); 5189 assert(!bs->refcnt); 5190 5191 bdrv_drained_begin(bs); /* complete I/O */ 5192 bdrv_flush(bs); 5193 bdrv_drain(bs); /* in case flush left pending I/O */ 5194 5195 if (bs->drv) { 5196 if (bs->drv->bdrv_close) { 5197 /* Must unfreeze all children, so bdrv_unref_child() works */ 5198 bs->drv->bdrv_close(bs); 5199 } 5200 bs->drv = NULL; 5201 } 5202 5203 bdrv_graph_wrlock(bs); 5204 QLIST_FOREACH_SAFE(child, &bs->children, next, next) { 5205 bdrv_unref_child(bs, child); 5206 } 5207 bdrv_graph_wrunlock(); 5208 5209 assert(!bs->backing); 5210 assert(!bs->file); 5211 g_free(bs->opaque); 5212 bs->opaque = NULL; 5213 qatomic_set(&bs->copy_on_read, 0); 5214 bs->backing_file[0] = '\0'; 5215 bs->backing_format[0] = '\0'; 5216 bs->total_sectors = 0; 5217 bs->encrypted = false; 5218 bs->sg = false; 5219 qobject_unref(bs->options); 5220 qobject_unref(bs->explicit_options); 5221 bs->options = NULL; 5222 bs->explicit_options = NULL; 5223 qobject_unref(bs->full_open_options); 5224 bs->full_open_options = NULL; 5225 g_free(bs->block_status_cache); 5226 bs->block_status_cache = NULL; 5227 5228 bdrv_release_named_dirty_bitmaps(bs); 5229 assert(QLIST_EMPTY(&bs->dirty_bitmaps)); 5230 5231 QLIST_FOREACH_SAFE(ban, &bs->aio_notifiers, list, ban_next) { 5232 g_free(ban); 5233 } 5234 QLIST_INIT(&bs->aio_notifiers); 5235 bdrv_drained_end(bs); 5236 5237 /* 5238 * If we're still inside some bdrv_drain_all_begin()/end() sections, end 5239 * them now since this BDS won't exist anymore when bdrv_drain_all_end() 5240 * gets called. 5241 */ 5242 if (bs->quiesce_counter) { 5243 bdrv_drain_all_end_quiesce(bs); 5244 } 5245 } 5246 5247 void bdrv_close_all(void) 5248 { 5249 GLOBAL_STATE_CODE(); 5250 assert(job_next(NULL) == NULL); 5251 5252 /* Drop references from requests still in flight, such as canceled block 5253 * jobs whose AIO context has not been polled yet */ 5254 bdrv_drain_all(); 5255 5256 blk_remove_all_bs(); 5257 blockdev_close_all_bdrv_states(); 5258 5259 assert(QTAILQ_EMPTY(&all_bdrv_states)); 5260 } 5261 5262 static bool GRAPH_RDLOCK should_update_child(BdrvChild *c, BlockDriverState *to) 5263 { 5264 GQueue *queue; 5265 GHashTable *found; 5266 bool ret; 5267 5268 if (c->klass->stay_at_node) { 5269 return false; 5270 } 5271 5272 /* If the child @c belongs to the BDS @to, replacing the current 5273 * c->bs by @to would mean to create a loop. 5274 * 5275 * Such a case occurs when appending a BDS to a backing chain. 5276 * For instance, imagine the following chain: 5277 * 5278 * guest device -> node A -> further backing chain... 5279 * 5280 * Now we create a new BDS B which we want to put on top of this 5281 * chain, so we first attach A as its backing node: 5282 * 5283 * node B 5284 * | 5285 * v 5286 * guest device -> node A -> further backing chain... 5287 * 5288 * Finally we want to replace A by B. When doing that, we want to 5289 * replace all pointers to A by pointers to B -- except for the 5290 * pointer from B because (1) that would create a loop, and (2) 5291 * that pointer should simply stay intact: 5292 * 5293 * guest device -> node B 5294 * | 5295 * v 5296 * node A -> further backing chain... 5297 * 5298 * In general, when replacing a node A (c->bs) by a node B (@to), 5299 * if A is a child of B, that means we cannot replace A by B there 5300 * because that would create a loop. Silently detaching A from B 5301 * is also not really an option. So overall just leaving A in 5302 * place there is the most sensible choice. 5303 * 5304 * We would also create a loop in any cases where @c is only 5305 * indirectly referenced by @to. Prevent this by returning false 5306 * if @c is found (by breadth-first search) anywhere in the whole 5307 * subtree of @to. 5308 */ 5309 5310 ret = true; 5311 found = g_hash_table_new(NULL, NULL); 5312 g_hash_table_add(found, to); 5313 queue = g_queue_new(); 5314 g_queue_push_tail(queue, to); 5315 5316 while (!g_queue_is_empty(queue)) { 5317 BlockDriverState *v = g_queue_pop_head(queue); 5318 BdrvChild *c2; 5319 5320 QLIST_FOREACH(c2, &v->children, next) { 5321 if (c2 == c) { 5322 ret = false; 5323 break; 5324 } 5325 5326 if (g_hash_table_contains(found, c2->bs)) { 5327 continue; 5328 } 5329 5330 g_queue_push_tail(queue, c2->bs); 5331 g_hash_table_add(found, c2->bs); 5332 } 5333 } 5334 5335 g_queue_free(queue); 5336 g_hash_table_destroy(found); 5337 5338 return ret; 5339 } 5340 5341 static void bdrv_remove_child_commit(void *opaque) 5342 { 5343 GLOBAL_STATE_CODE(); 5344 bdrv_child_free(opaque); 5345 } 5346 5347 static TransactionActionDrv bdrv_remove_child_drv = { 5348 .commit = bdrv_remove_child_commit, 5349 }; 5350 5351 /* 5352 * Function doesn't update permissions, caller is responsible for this. 5353 * 5354 * @child->bs (if non-NULL) must be drained. 5355 * 5356 * After calling this function, the transaction @tran may only be completed 5357 * while holding a writer lock for the graph. 5358 */ 5359 static void GRAPH_WRLOCK bdrv_remove_child(BdrvChild *child, Transaction *tran) 5360 { 5361 if (!child) { 5362 return; 5363 } 5364 5365 if (child->bs) { 5366 assert(child->quiesced_parent); 5367 bdrv_replace_child_tran(child, NULL, tran); 5368 } 5369 5370 tran_add(tran, &bdrv_remove_child_drv, child); 5371 } 5372 5373 /* 5374 * Both @from and @to (if non-NULL) must be drained. @to must be kept drained 5375 * until the transaction is completed. 5376 * 5377 * After calling this function, the transaction @tran may only be completed 5378 * while holding a writer lock for the graph. 5379 */ 5380 static int GRAPH_WRLOCK 5381 bdrv_replace_node_noperm(BlockDriverState *from, 5382 BlockDriverState *to, 5383 bool auto_skip, Transaction *tran, 5384 Error **errp) 5385 { 5386 BdrvChild *c, *next; 5387 5388 GLOBAL_STATE_CODE(); 5389 5390 assert(from->quiesce_counter); 5391 assert(to->quiesce_counter); 5392 5393 QLIST_FOREACH_SAFE(c, &from->parents, next_parent, next) { 5394 assert(c->bs == from); 5395 if (!should_update_child(c, to)) { 5396 if (auto_skip) { 5397 continue; 5398 } 5399 error_setg(errp, "Should not change '%s' link to '%s'", 5400 c->name, from->node_name); 5401 return -EINVAL; 5402 } 5403 if (c->frozen) { 5404 error_setg(errp, "Cannot change '%s' link to '%s'", 5405 c->name, from->node_name); 5406 return -EPERM; 5407 } 5408 bdrv_replace_child_tran(c, to, tran); 5409 } 5410 5411 return 0; 5412 } 5413 5414 /* 5415 * Switch all parents of @from to point to @to instead. @from and @to must be in 5416 * the same AioContext and both must be drained. 5417 * 5418 * With auto_skip=true bdrv_replace_node_common skips updating from parents 5419 * if it creates a parent-child relation loop or if parent is block-job. 5420 * 5421 * With auto_skip=false the error is returned if from has a parent which should 5422 * not be updated. 5423 * 5424 * With @detach_subchain=true @to must be in a backing chain of @from. In this 5425 * case backing link of the cow-parent of @to is removed. 5426 */ 5427 static int GRAPH_WRLOCK 5428 bdrv_replace_node_common(BlockDriverState *from, BlockDriverState *to, 5429 bool auto_skip, bool detach_subchain, Error **errp) 5430 { 5431 Transaction *tran = tran_new(); 5432 g_autoptr(GSList) refresh_list = NULL; 5433 BlockDriverState *to_cow_parent = NULL; 5434 int ret; 5435 5436 GLOBAL_STATE_CODE(); 5437 5438 assert(from->quiesce_counter); 5439 assert(to->quiesce_counter); 5440 assert(bdrv_get_aio_context(from) == bdrv_get_aio_context(to)); 5441 5442 if (detach_subchain) { 5443 assert(bdrv_chain_contains(from, to)); 5444 assert(from != to); 5445 for (to_cow_parent = from; 5446 bdrv_filter_or_cow_bs(to_cow_parent) != to; 5447 to_cow_parent = bdrv_filter_or_cow_bs(to_cow_parent)) 5448 { 5449 ; 5450 } 5451 } 5452 5453 /* 5454 * Do the replacement without permission update. 5455 * Replacement may influence the permissions, we should calculate new 5456 * permissions based on new graph. If we fail, we'll roll-back the 5457 * replacement. 5458 */ 5459 ret = bdrv_replace_node_noperm(from, to, auto_skip, tran, errp); 5460 if (ret < 0) { 5461 goto out; 5462 } 5463 5464 if (detach_subchain) { 5465 /* to_cow_parent is already drained because from is drained */ 5466 bdrv_remove_child(bdrv_filter_or_cow_child(to_cow_parent), tran); 5467 } 5468 5469 refresh_list = g_slist_prepend(refresh_list, to); 5470 refresh_list = g_slist_prepend(refresh_list, from); 5471 5472 ret = bdrv_list_refresh_perms(refresh_list, NULL, tran, errp); 5473 if (ret < 0) { 5474 goto out; 5475 } 5476 5477 ret = 0; 5478 5479 out: 5480 tran_finalize(tran, ret); 5481 return ret; 5482 } 5483 5484 int bdrv_replace_node(BlockDriverState *from, BlockDriverState *to, 5485 Error **errp) 5486 { 5487 return bdrv_replace_node_common(from, to, true, false, errp); 5488 } 5489 5490 int bdrv_drop_filter(BlockDriverState *bs, Error **errp) 5491 { 5492 BlockDriverState *child_bs; 5493 int ret; 5494 5495 GLOBAL_STATE_CODE(); 5496 5497 bdrv_graph_rdlock_main_loop(); 5498 child_bs = bdrv_filter_or_cow_bs(bs); 5499 bdrv_graph_rdunlock_main_loop(); 5500 5501 bdrv_drained_begin(child_bs); 5502 bdrv_graph_wrlock(bs); 5503 ret = bdrv_replace_node_common(bs, child_bs, true, true, errp); 5504 bdrv_graph_wrunlock(); 5505 bdrv_drained_end(child_bs); 5506 5507 return ret; 5508 } 5509 5510 /* 5511 * Add new bs contents at the top of an image chain while the chain is 5512 * live, while keeping required fields on the top layer. 5513 * 5514 * This will modify the BlockDriverState fields, and swap contents 5515 * between bs_new and bs_top. Both bs_new and bs_top are modified. 5516 * 5517 * bs_new must not be attached to a BlockBackend and must not have backing 5518 * child. 5519 * 5520 * This function does not create any image files. 5521 * 5522 * The caller must hold the AioContext lock for @bs_top. 5523 */ 5524 int bdrv_append(BlockDriverState *bs_new, BlockDriverState *bs_top, 5525 Error **errp) 5526 { 5527 int ret; 5528 BdrvChild *child; 5529 Transaction *tran = tran_new(); 5530 AioContext *old_context, *new_context = NULL; 5531 5532 GLOBAL_STATE_CODE(); 5533 5534 assert(!bs_new->backing); 5535 5536 old_context = bdrv_get_aio_context(bs_top); 5537 bdrv_drained_begin(bs_top); 5538 5539 /* 5540 * bdrv_drained_begin() requires that only the AioContext of the drained 5541 * node is locked, and at this point it can still differ from the AioContext 5542 * of bs_top. 5543 */ 5544 new_context = bdrv_get_aio_context(bs_new); 5545 aio_context_release(old_context); 5546 aio_context_acquire(new_context); 5547 bdrv_drained_begin(bs_new); 5548 aio_context_release(new_context); 5549 aio_context_acquire(old_context); 5550 new_context = NULL; 5551 5552 bdrv_graph_wrlock(bs_top); 5553 5554 child = bdrv_attach_child_noperm(bs_new, bs_top, "backing", 5555 &child_of_bds, bdrv_backing_role(bs_new), 5556 tran, errp); 5557 if (!child) { 5558 ret = -EINVAL; 5559 goto out; 5560 } 5561 5562 /* 5563 * bdrv_attach_child_noperm could change the AioContext of bs_top and 5564 * bs_new, but at least they are in the same AioContext now. This is the 5565 * AioContext that we need to lock for the rest of the function. 5566 */ 5567 new_context = bdrv_get_aio_context(bs_top); 5568 5569 if (old_context != new_context) { 5570 aio_context_release(old_context); 5571 aio_context_acquire(new_context); 5572 } 5573 5574 ret = bdrv_replace_node_noperm(bs_top, bs_new, true, tran, errp); 5575 if (ret < 0) { 5576 goto out; 5577 } 5578 5579 ret = bdrv_refresh_perms(bs_new, tran, errp); 5580 out: 5581 tran_finalize(tran, ret); 5582 5583 bdrv_refresh_limits(bs_top, NULL, NULL); 5584 bdrv_graph_wrunlock(); 5585 5586 bdrv_drained_end(bs_top); 5587 bdrv_drained_end(bs_new); 5588 5589 if (new_context && old_context != new_context) { 5590 aio_context_release(new_context); 5591 aio_context_acquire(old_context); 5592 } 5593 5594 return ret; 5595 } 5596 5597 /* Not for empty child */ 5598 int bdrv_replace_child_bs(BdrvChild *child, BlockDriverState *new_bs, 5599 Error **errp) 5600 { 5601 int ret; 5602 Transaction *tran = tran_new(); 5603 g_autoptr(GSList) refresh_list = NULL; 5604 BlockDriverState *old_bs = child->bs; 5605 5606 GLOBAL_STATE_CODE(); 5607 5608 bdrv_ref(old_bs); 5609 bdrv_drained_begin(old_bs); 5610 bdrv_drained_begin(new_bs); 5611 bdrv_graph_wrlock(new_bs); 5612 5613 bdrv_replace_child_tran(child, new_bs, tran); 5614 5615 refresh_list = g_slist_prepend(refresh_list, old_bs); 5616 refresh_list = g_slist_prepend(refresh_list, new_bs); 5617 5618 ret = bdrv_list_refresh_perms(refresh_list, NULL, tran, errp); 5619 5620 tran_finalize(tran, ret); 5621 5622 bdrv_graph_wrunlock(); 5623 bdrv_drained_end(old_bs); 5624 bdrv_drained_end(new_bs); 5625 bdrv_unref(old_bs); 5626 5627 return ret; 5628 } 5629 5630 static void bdrv_delete(BlockDriverState *bs) 5631 { 5632 assert(bdrv_op_blocker_is_empty(bs)); 5633 assert(!bs->refcnt); 5634 GLOBAL_STATE_CODE(); 5635 5636 /* remove from list, if necessary */ 5637 if (bs->node_name[0] != '\0') { 5638 QTAILQ_REMOVE(&graph_bdrv_states, bs, node_list); 5639 } 5640 QTAILQ_REMOVE(&all_bdrv_states, bs, bs_list); 5641 5642 bdrv_close(bs); 5643 5644 qemu_mutex_destroy(&bs->reqs_lock); 5645 5646 g_free(bs); 5647 } 5648 5649 5650 /* 5651 * Replace @bs by newly created block node. 5652 * 5653 * @options is a QDict of options to pass to the block drivers, or NULL for an 5654 * empty set of options. The reference to the QDict belongs to the block layer 5655 * after the call (even on failure), so if the caller intends to reuse the 5656 * dictionary, it needs to use qobject_ref() before calling bdrv_open. 5657 * 5658 * The caller holds the AioContext lock for @bs. It must make sure that @bs 5659 * stays in the same AioContext, i.e. @options must not refer to nodes in a 5660 * different AioContext. 5661 */ 5662 BlockDriverState *bdrv_insert_node(BlockDriverState *bs, QDict *options, 5663 int flags, Error **errp) 5664 { 5665 ERRP_GUARD(); 5666 int ret; 5667 AioContext *ctx = bdrv_get_aio_context(bs); 5668 BlockDriverState *new_node_bs = NULL; 5669 const char *drvname, *node_name; 5670 BlockDriver *drv; 5671 5672 drvname = qdict_get_try_str(options, "driver"); 5673 if (!drvname) { 5674 error_setg(errp, "driver is not specified"); 5675 goto fail; 5676 } 5677 5678 drv = bdrv_find_format(drvname); 5679 if (!drv) { 5680 error_setg(errp, "Unknown driver: '%s'", drvname); 5681 goto fail; 5682 } 5683 5684 node_name = qdict_get_try_str(options, "node-name"); 5685 5686 GLOBAL_STATE_CODE(); 5687 5688 aio_context_release(ctx); 5689 aio_context_acquire(qemu_get_aio_context()); 5690 new_node_bs = bdrv_new_open_driver_opts(drv, node_name, options, flags, 5691 errp); 5692 aio_context_release(qemu_get_aio_context()); 5693 aio_context_acquire(ctx); 5694 assert(bdrv_get_aio_context(bs) == ctx); 5695 5696 options = NULL; /* bdrv_new_open_driver() eats options */ 5697 if (!new_node_bs) { 5698 error_prepend(errp, "Could not create node: "); 5699 goto fail; 5700 } 5701 5702 /* 5703 * Make sure that @bs doesn't go away until we have successfully attached 5704 * all of its parents to @new_node_bs and undrained it again. 5705 */ 5706 bdrv_ref(bs); 5707 bdrv_drained_begin(bs); 5708 bdrv_drained_begin(new_node_bs); 5709 bdrv_graph_wrlock(new_node_bs); 5710 ret = bdrv_replace_node(bs, new_node_bs, errp); 5711 bdrv_graph_wrunlock(); 5712 bdrv_drained_end(new_node_bs); 5713 bdrv_drained_end(bs); 5714 bdrv_unref(bs); 5715 5716 if (ret < 0) { 5717 error_prepend(errp, "Could not replace node: "); 5718 goto fail; 5719 } 5720 5721 return new_node_bs; 5722 5723 fail: 5724 qobject_unref(options); 5725 bdrv_unref(new_node_bs); 5726 return NULL; 5727 } 5728 5729 /* 5730 * Run consistency checks on an image 5731 * 5732 * Returns 0 if the check could be completed (it doesn't mean that the image is 5733 * free of errors) or -errno when an internal error occurred. The results of the 5734 * check are stored in res. 5735 */ 5736 int coroutine_fn bdrv_co_check(BlockDriverState *bs, 5737 BdrvCheckResult *res, BdrvCheckMode fix) 5738 { 5739 IO_CODE(); 5740 assert_bdrv_graph_readable(); 5741 if (bs->drv == NULL) { 5742 return -ENOMEDIUM; 5743 } 5744 if (bs->drv->bdrv_co_check == NULL) { 5745 return -ENOTSUP; 5746 } 5747 5748 memset(res, 0, sizeof(*res)); 5749 return bs->drv->bdrv_co_check(bs, res, fix); 5750 } 5751 5752 /* 5753 * Return values: 5754 * 0 - success 5755 * -EINVAL - backing format specified, but no file 5756 * -ENOSPC - can't update the backing file because no space is left in the 5757 * image file header 5758 * -ENOTSUP - format driver doesn't support changing the backing file 5759 */ 5760 int bdrv_change_backing_file(BlockDriverState *bs, const char *backing_file, 5761 const char *backing_fmt, bool require) 5762 { 5763 BlockDriver *drv = bs->drv; 5764 int ret; 5765 5766 GLOBAL_STATE_CODE(); 5767 5768 if (!drv) { 5769 return -ENOMEDIUM; 5770 } 5771 5772 /* Backing file format doesn't make sense without a backing file */ 5773 if (backing_fmt && !backing_file) { 5774 return -EINVAL; 5775 } 5776 5777 if (require && backing_file && !backing_fmt) { 5778 return -EINVAL; 5779 } 5780 5781 if (drv->bdrv_change_backing_file != NULL) { 5782 ret = drv->bdrv_change_backing_file(bs, backing_file, backing_fmt); 5783 } else { 5784 ret = -ENOTSUP; 5785 } 5786 5787 if (ret == 0) { 5788 pstrcpy(bs->backing_file, sizeof(bs->backing_file), backing_file ?: ""); 5789 pstrcpy(bs->backing_format, sizeof(bs->backing_format), backing_fmt ?: ""); 5790 pstrcpy(bs->auto_backing_file, sizeof(bs->auto_backing_file), 5791 backing_file ?: ""); 5792 } 5793 return ret; 5794 } 5795 5796 /* 5797 * Finds the first non-filter node above bs in the chain between 5798 * active and bs. The returned node is either an immediate parent of 5799 * bs, or there are only filter nodes between the two. 5800 * 5801 * Returns NULL if bs is not found in active's image chain, 5802 * or if active == bs. 5803 * 5804 * Returns the bottommost base image if bs == NULL. 5805 */ 5806 BlockDriverState *bdrv_find_overlay(BlockDriverState *active, 5807 BlockDriverState *bs) 5808 { 5809 5810 GLOBAL_STATE_CODE(); 5811 5812 bs = bdrv_skip_filters(bs); 5813 active = bdrv_skip_filters(active); 5814 5815 while (active) { 5816 BlockDriverState *next = bdrv_backing_chain_next(active); 5817 if (bs == next) { 5818 return active; 5819 } 5820 active = next; 5821 } 5822 5823 return NULL; 5824 } 5825 5826 /* Given a BDS, searches for the base layer. */ 5827 BlockDriverState *bdrv_find_base(BlockDriverState *bs) 5828 { 5829 GLOBAL_STATE_CODE(); 5830 5831 return bdrv_find_overlay(bs, NULL); 5832 } 5833 5834 /* 5835 * Return true if at least one of the COW (backing) and filter links 5836 * between @bs and @base is frozen. @errp is set if that's the case. 5837 * @base must be reachable from @bs, or NULL. 5838 */ 5839 static bool GRAPH_RDLOCK 5840 bdrv_is_backing_chain_frozen(BlockDriverState *bs, BlockDriverState *base, 5841 Error **errp) 5842 { 5843 BlockDriverState *i; 5844 BdrvChild *child; 5845 5846 GLOBAL_STATE_CODE(); 5847 5848 for (i = bs; i != base; i = child_bs(child)) { 5849 child = bdrv_filter_or_cow_child(i); 5850 5851 if (child && child->frozen) { 5852 error_setg(errp, "Cannot change '%s' link from '%s' to '%s'", 5853 child->name, i->node_name, child->bs->node_name); 5854 return true; 5855 } 5856 } 5857 5858 return false; 5859 } 5860 5861 /* 5862 * Freeze all COW (backing) and filter links between @bs and @base. 5863 * If any of the links is already frozen the operation is aborted and 5864 * none of the links are modified. 5865 * @base must be reachable from @bs, or NULL. 5866 * Returns 0 on success. On failure returns < 0 and sets @errp. 5867 */ 5868 int bdrv_freeze_backing_chain(BlockDriverState *bs, BlockDriverState *base, 5869 Error **errp) 5870 { 5871 BlockDriverState *i; 5872 BdrvChild *child; 5873 5874 GLOBAL_STATE_CODE(); 5875 5876 if (bdrv_is_backing_chain_frozen(bs, base, errp)) { 5877 return -EPERM; 5878 } 5879 5880 for (i = bs; i != base; i = child_bs(child)) { 5881 child = bdrv_filter_or_cow_child(i); 5882 if (child && child->bs->never_freeze) { 5883 error_setg(errp, "Cannot freeze '%s' link to '%s'", 5884 child->name, child->bs->node_name); 5885 return -EPERM; 5886 } 5887 } 5888 5889 for (i = bs; i != base; i = child_bs(child)) { 5890 child = bdrv_filter_or_cow_child(i); 5891 if (child) { 5892 child->frozen = true; 5893 } 5894 } 5895 5896 return 0; 5897 } 5898 5899 /* 5900 * Unfreeze all COW (backing) and filter links between @bs and @base. 5901 * The caller must ensure that all links are frozen before using this 5902 * function. 5903 * @base must be reachable from @bs, or NULL. 5904 */ 5905 void bdrv_unfreeze_backing_chain(BlockDriverState *bs, BlockDriverState *base) 5906 { 5907 BlockDriverState *i; 5908 BdrvChild *child; 5909 5910 GLOBAL_STATE_CODE(); 5911 5912 for (i = bs; i != base; i = child_bs(child)) { 5913 child = bdrv_filter_or_cow_child(i); 5914 if (child) { 5915 assert(child->frozen); 5916 child->frozen = false; 5917 } 5918 } 5919 } 5920 5921 /* 5922 * Drops images above 'base' up to and including 'top', and sets the image 5923 * above 'top' to have base as its backing file. 5924 * 5925 * Requires that the overlay to 'top' is opened r/w, so that the backing file 5926 * information in 'bs' can be properly updated. 5927 * 5928 * E.g., this will convert the following chain: 5929 * bottom <- base <- intermediate <- top <- active 5930 * 5931 * to 5932 * 5933 * bottom <- base <- active 5934 * 5935 * It is allowed for bottom==base, in which case it converts: 5936 * 5937 * base <- intermediate <- top <- active 5938 * 5939 * to 5940 * 5941 * base <- active 5942 * 5943 * If backing_file_str is non-NULL, it will be used when modifying top's 5944 * overlay image metadata. 5945 * 5946 * Error conditions: 5947 * if active == top, that is considered an error 5948 * 5949 */ 5950 int bdrv_drop_intermediate(BlockDriverState *top, BlockDriverState *base, 5951 const char *backing_file_str) 5952 { 5953 BlockDriverState *explicit_top = top; 5954 bool update_inherits_from; 5955 BdrvChild *c; 5956 Error *local_err = NULL; 5957 int ret = -EIO; 5958 g_autoptr(GSList) updated_children = NULL; 5959 GSList *p; 5960 5961 GLOBAL_STATE_CODE(); 5962 5963 bdrv_ref(top); 5964 bdrv_drained_begin(base); 5965 bdrv_graph_wrlock(base); 5966 5967 if (!top->drv || !base->drv) { 5968 goto exit_wrlock; 5969 } 5970 5971 /* Make sure that base is in the backing chain of top */ 5972 if (!bdrv_chain_contains(top, base)) { 5973 goto exit_wrlock; 5974 } 5975 5976 /* If 'base' recursively inherits from 'top' then we should set 5977 * base->inherits_from to top->inherits_from after 'top' and all 5978 * other intermediate nodes have been dropped. 5979 * If 'top' is an implicit node (e.g. "commit_top") we should skip 5980 * it because no one inherits from it. We use explicit_top for that. */ 5981 explicit_top = bdrv_skip_implicit_filters(explicit_top); 5982 update_inherits_from = bdrv_inherits_from_recursive(base, explicit_top); 5983 5984 /* success - we can delete the intermediate states, and link top->base */ 5985 if (!backing_file_str) { 5986 bdrv_refresh_filename(base); 5987 backing_file_str = base->filename; 5988 } 5989 5990 QLIST_FOREACH(c, &top->parents, next_parent) { 5991 updated_children = g_slist_prepend(updated_children, c); 5992 } 5993 5994 /* 5995 * It seems correct to pass detach_subchain=true here, but it triggers 5996 * one more yet not fixed bug, when due to nested aio_poll loop we switch to 5997 * another drained section, which modify the graph (for example, removing 5998 * the child, which we keep in updated_children list). So, it's a TODO. 5999 * 6000 * Note, bug triggered if pass detach_subchain=true here and run 6001 * test-bdrv-drain. test_drop_intermediate_poll() test-case will crash. 6002 * That's a FIXME. 6003 */ 6004 bdrv_replace_node_common(top, base, false, false, &local_err); 6005 bdrv_graph_wrunlock(); 6006 6007 if (local_err) { 6008 error_report_err(local_err); 6009 goto exit; 6010 } 6011 6012 for (p = updated_children; p; p = p->next) { 6013 c = p->data; 6014 6015 if (c->klass->update_filename) { 6016 ret = c->klass->update_filename(c, base, backing_file_str, 6017 &local_err); 6018 if (ret < 0) { 6019 /* 6020 * TODO: Actually, we want to rollback all previous iterations 6021 * of this loop, and (which is almost impossible) previous 6022 * bdrv_replace_node()... 6023 * 6024 * Note, that c->klass->update_filename may lead to permission 6025 * update, so it's a bad idea to call it inside permission 6026 * update transaction of bdrv_replace_node. 6027 */ 6028 error_report_err(local_err); 6029 goto exit; 6030 } 6031 } 6032 } 6033 6034 if (update_inherits_from) { 6035 base->inherits_from = explicit_top->inherits_from; 6036 } 6037 6038 ret = 0; 6039 goto exit; 6040 6041 exit_wrlock: 6042 bdrv_graph_wrunlock(); 6043 exit: 6044 bdrv_drained_end(base); 6045 bdrv_unref(top); 6046 return ret; 6047 } 6048 6049 /** 6050 * Implementation of BlockDriver.bdrv_co_get_allocated_file_size() that 6051 * sums the size of all data-bearing children. (This excludes backing 6052 * children.) 6053 */ 6054 static int64_t coroutine_fn GRAPH_RDLOCK 6055 bdrv_sum_allocated_file_size(BlockDriverState *bs) 6056 { 6057 BdrvChild *child; 6058 int64_t child_size, sum = 0; 6059 6060 QLIST_FOREACH(child, &bs->children, next) { 6061 if (child->role & (BDRV_CHILD_DATA | BDRV_CHILD_METADATA | 6062 BDRV_CHILD_FILTERED)) 6063 { 6064 child_size = bdrv_co_get_allocated_file_size(child->bs); 6065 if (child_size < 0) { 6066 return child_size; 6067 } 6068 sum += child_size; 6069 } 6070 } 6071 6072 return sum; 6073 } 6074 6075 /** 6076 * Length of a allocated file in bytes. Sparse files are counted by actual 6077 * allocated space. Return < 0 if error or unknown. 6078 */ 6079 int64_t coroutine_fn bdrv_co_get_allocated_file_size(BlockDriverState *bs) 6080 { 6081 BlockDriver *drv = bs->drv; 6082 IO_CODE(); 6083 assert_bdrv_graph_readable(); 6084 6085 if (!drv) { 6086 return -ENOMEDIUM; 6087 } 6088 if (drv->bdrv_co_get_allocated_file_size) { 6089 return drv->bdrv_co_get_allocated_file_size(bs); 6090 } 6091 6092 if (drv->bdrv_file_open) { 6093 /* 6094 * Protocol drivers default to -ENOTSUP (most of their data is 6095 * not stored in any of their children (if they even have any), 6096 * so there is no generic way to figure it out). 6097 */ 6098 return -ENOTSUP; 6099 } else if (drv->is_filter) { 6100 /* Filter drivers default to the size of their filtered child */ 6101 return bdrv_co_get_allocated_file_size(bdrv_filter_bs(bs)); 6102 } else { 6103 /* Other drivers default to summing their children's sizes */ 6104 return bdrv_sum_allocated_file_size(bs); 6105 } 6106 } 6107 6108 /* 6109 * bdrv_measure: 6110 * @drv: Format driver 6111 * @opts: Creation options for new image 6112 * @in_bs: Existing image containing data for new image (may be NULL) 6113 * @errp: Error object 6114 * Returns: A #BlockMeasureInfo (free using qapi_free_BlockMeasureInfo()) 6115 * or NULL on error 6116 * 6117 * Calculate file size required to create a new image. 6118 * 6119 * If @in_bs is given then space for allocated clusters and zero clusters 6120 * from that image are included in the calculation. If @opts contains a 6121 * backing file that is shared by @in_bs then backing clusters may be omitted 6122 * from the calculation. 6123 * 6124 * If @in_bs is NULL then the calculation includes no allocated clusters 6125 * unless a preallocation option is given in @opts. 6126 * 6127 * Note that @in_bs may use a different BlockDriver from @drv. 6128 * 6129 * If an error occurs the @errp pointer is set. 6130 */ 6131 BlockMeasureInfo *bdrv_measure(BlockDriver *drv, QemuOpts *opts, 6132 BlockDriverState *in_bs, Error **errp) 6133 { 6134 IO_CODE(); 6135 if (!drv->bdrv_measure) { 6136 error_setg(errp, "Block driver '%s' does not support size measurement", 6137 drv->format_name); 6138 return NULL; 6139 } 6140 6141 return drv->bdrv_measure(opts, in_bs, errp); 6142 } 6143 6144 /** 6145 * Return number of sectors on success, -errno on error. 6146 */ 6147 int64_t coroutine_fn bdrv_co_nb_sectors(BlockDriverState *bs) 6148 { 6149 BlockDriver *drv = bs->drv; 6150 IO_CODE(); 6151 assert_bdrv_graph_readable(); 6152 6153 if (!drv) 6154 return -ENOMEDIUM; 6155 6156 if (bs->bl.has_variable_length) { 6157 int ret = bdrv_co_refresh_total_sectors(bs, bs->total_sectors); 6158 if (ret < 0) { 6159 return ret; 6160 } 6161 } 6162 return bs->total_sectors; 6163 } 6164 6165 /* 6166 * This wrapper is written by hand because this function is in the hot I/O path, 6167 * via blk_get_geometry. 6168 */ 6169 int64_t coroutine_mixed_fn bdrv_nb_sectors(BlockDriverState *bs) 6170 { 6171 BlockDriver *drv = bs->drv; 6172 IO_CODE(); 6173 6174 if (!drv) 6175 return -ENOMEDIUM; 6176 6177 if (bs->bl.has_variable_length) { 6178 int ret = bdrv_refresh_total_sectors(bs, bs->total_sectors); 6179 if (ret < 0) { 6180 return ret; 6181 } 6182 } 6183 6184 return bs->total_sectors; 6185 } 6186 6187 /** 6188 * Return length in bytes on success, -errno on error. 6189 * The length is always a multiple of BDRV_SECTOR_SIZE. 6190 */ 6191 int64_t coroutine_fn bdrv_co_getlength(BlockDriverState *bs) 6192 { 6193 int64_t ret; 6194 IO_CODE(); 6195 assert_bdrv_graph_readable(); 6196 6197 ret = bdrv_co_nb_sectors(bs); 6198 if (ret < 0) { 6199 return ret; 6200 } 6201 if (ret > INT64_MAX / BDRV_SECTOR_SIZE) { 6202 return -EFBIG; 6203 } 6204 return ret * BDRV_SECTOR_SIZE; 6205 } 6206 6207 bool bdrv_is_sg(BlockDriverState *bs) 6208 { 6209 IO_CODE(); 6210 return bs->sg; 6211 } 6212 6213 /** 6214 * Return whether the given node supports compressed writes. 6215 */ 6216 bool bdrv_supports_compressed_writes(BlockDriverState *bs) 6217 { 6218 BlockDriverState *filtered; 6219 IO_CODE(); 6220 6221 if (!bs->drv || !block_driver_can_compress(bs->drv)) { 6222 return false; 6223 } 6224 6225 filtered = bdrv_filter_bs(bs); 6226 if (filtered) { 6227 /* 6228 * Filters can only forward compressed writes, so we have to 6229 * check the child. 6230 */ 6231 return bdrv_supports_compressed_writes(filtered); 6232 } 6233 6234 return true; 6235 } 6236 6237 const char *bdrv_get_format_name(BlockDriverState *bs) 6238 { 6239 IO_CODE(); 6240 return bs->drv ? bs->drv->format_name : NULL; 6241 } 6242 6243 static int qsort_strcmp(const void *a, const void *b) 6244 { 6245 return strcmp(*(char *const *)a, *(char *const *)b); 6246 } 6247 6248 void bdrv_iterate_format(void (*it)(void *opaque, const char *name), 6249 void *opaque, bool read_only) 6250 { 6251 BlockDriver *drv; 6252 int count = 0; 6253 int i; 6254 const char **formats = NULL; 6255 6256 GLOBAL_STATE_CODE(); 6257 6258 QLIST_FOREACH(drv, &bdrv_drivers, list) { 6259 if (drv->format_name) { 6260 bool found = false; 6261 6262 if (use_bdrv_whitelist && !bdrv_is_whitelisted(drv, read_only)) { 6263 continue; 6264 } 6265 6266 i = count; 6267 while (formats && i && !found) { 6268 found = !strcmp(formats[--i], drv->format_name); 6269 } 6270 6271 if (!found) { 6272 formats = g_renew(const char *, formats, count + 1); 6273 formats[count++] = drv->format_name; 6274 } 6275 } 6276 } 6277 6278 for (i = 0; i < (int)ARRAY_SIZE(block_driver_modules); i++) { 6279 const char *format_name = block_driver_modules[i].format_name; 6280 6281 if (format_name) { 6282 bool found = false; 6283 int j = count; 6284 6285 if (use_bdrv_whitelist && 6286 !bdrv_format_is_whitelisted(format_name, read_only)) { 6287 continue; 6288 } 6289 6290 while (formats && j && !found) { 6291 found = !strcmp(formats[--j], format_name); 6292 } 6293 6294 if (!found) { 6295 formats = g_renew(const char *, formats, count + 1); 6296 formats[count++] = format_name; 6297 } 6298 } 6299 } 6300 6301 qsort(formats, count, sizeof(formats[0]), qsort_strcmp); 6302 6303 for (i = 0; i < count; i++) { 6304 it(opaque, formats[i]); 6305 } 6306 6307 g_free(formats); 6308 } 6309 6310 /* This function is to find a node in the bs graph */ 6311 BlockDriverState *bdrv_find_node(const char *node_name) 6312 { 6313 BlockDriverState *bs; 6314 6315 assert(node_name); 6316 GLOBAL_STATE_CODE(); 6317 6318 QTAILQ_FOREACH(bs, &graph_bdrv_states, node_list) { 6319 if (!strcmp(node_name, bs->node_name)) { 6320 return bs; 6321 } 6322 } 6323 return NULL; 6324 } 6325 6326 /* Put this QMP function here so it can access the static graph_bdrv_states. */ 6327 BlockDeviceInfoList *bdrv_named_nodes_list(bool flat, 6328 Error **errp) 6329 { 6330 BlockDeviceInfoList *list; 6331 BlockDriverState *bs; 6332 6333 GLOBAL_STATE_CODE(); 6334 GRAPH_RDLOCK_GUARD_MAINLOOP(); 6335 6336 list = NULL; 6337 QTAILQ_FOREACH(bs, &graph_bdrv_states, node_list) { 6338 BlockDeviceInfo *info = bdrv_block_device_info(NULL, bs, flat, errp); 6339 if (!info) { 6340 qapi_free_BlockDeviceInfoList(list); 6341 return NULL; 6342 } 6343 QAPI_LIST_PREPEND(list, info); 6344 } 6345 6346 return list; 6347 } 6348 6349 typedef struct XDbgBlockGraphConstructor { 6350 XDbgBlockGraph *graph; 6351 GHashTable *graph_nodes; 6352 } XDbgBlockGraphConstructor; 6353 6354 static XDbgBlockGraphConstructor *xdbg_graph_new(void) 6355 { 6356 XDbgBlockGraphConstructor *gr = g_new(XDbgBlockGraphConstructor, 1); 6357 6358 gr->graph = g_new0(XDbgBlockGraph, 1); 6359 gr->graph_nodes = g_hash_table_new(NULL, NULL); 6360 6361 return gr; 6362 } 6363 6364 static XDbgBlockGraph *xdbg_graph_finalize(XDbgBlockGraphConstructor *gr) 6365 { 6366 XDbgBlockGraph *graph = gr->graph; 6367 6368 g_hash_table_destroy(gr->graph_nodes); 6369 g_free(gr); 6370 6371 return graph; 6372 } 6373 6374 static uintptr_t xdbg_graph_node_num(XDbgBlockGraphConstructor *gr, void *node) 6375 { 6376 uintptr_t ret = (uintptr_t)g_hash_table_lookup(gr->graph_nodes, node); 6377 6378 if (ret != 0) { 6379 return ret; 6380 } 6381 6382 /* 6383 * Start counting from 1, not 0, because 0 interferes with not-found (NULL) 6384 * answer of g_hash_table_lookup. 6385 */ 6386 ret = g_hash_table_size(gr->graph_nodes) + 1; 6387 g_hash_table_insert(gr->graph_nodes, node, (void *)ret); 6388 6389 return ret; 6390 } 6391 6392 static void xdbg_graph_add_node(XDbgBlockGraphConstructor *gr, void *node, 6393 XDbgBlockGraphNodeType type, const char *name) 6394 { 6395 XDbgBlockGraphNode *n; 6396 6397 n = g_new0(XDbgBlockGraphNode, 1); 6398 6399 n->id = xdbg_graph_node_num(gr, node); 6400 n->type = type; 6401 n->name = g_strdup(name); 6402 6403 QAPI_LIST_PREPEND(gr->graph->nodes, n); 6404 } 6405 6406 static void xdbg_graph_add_edge(XDbgBlockGraphConstructor *gr, void *parent, 6407 const BdrvChild *child) 6408 { 6409 BlockPermission qapi_perm; 6410 XDbgBlockGraphEdge *edge; 6411 GLOBAL_STATE_CODE(); 6412 6413 edge = g_new0(XDbgBlockGraphEdge, 1); 6414 6415 edge->parent = xdbg_graph_node_num(gr, parent); 6416 edge->child = xdbg_graph_node_num(gr, child->bs); 6417 edge->name = g_strdup(child->name); 6418 6419 for (qapi_perm = 0; qapi_perm < BLOCK_PERMISSION__MAX; qapi_perm++) { 6420 uint64_t flag = bdrv_qapi_perm_to_blk_perm(qapi_perm); 6421 6422 if (flag & child->perm) { 6423 QAPI_LIST_PREPEND(edge->perm, qapi_perm); 6424 } 6425 if (flag & child->shared_perm) { 6426 QAPI_LIST_PREPEND(edge->shared_perm, qapi_perm); 6427 } 6428 } 6429 6430 QAPI_LIST_PREPEND(gr->graph->edges, edge); 6431 } 6432 6433 6434 XDbgBlockGraph *bdrv_get_xdbg_block_graph(Error **errp) 6435 { 6436 BlockBackend *blk; 6437 BlockJob *job; 6438 BlockDriverState *bs; 6439 BdrvChild *child; 6440 XDbgBlockGraphConstructor *gr = xdbg_graph_new(); 6441 6442 GLOBAL_STATE_CODE(); 6443 6444 for (blk = blk_all_next(NULL); blk; blk = blk_all_next(blk)) { 6445 char *allocated_name = NULL; 6446 const char *name = blk_name(blk); 6447 6448 if (!*name) { 6449 name = allocated_name = blk_get_attached_dev_id(blk); 6450 } 6451 xdbg_graph_add_node(gr, blk, X_DBG_BLOCK_GRAPH_NODE_TYPE_BLOCK_BACKEND, 6452 name); 6453 g_free(allocated_name); 6454 if (blk_root(blk)) { 6455 xdbg_graph_add_edge(gr, blk, blk_root(blk)); 6456 } 6457 } 6458 6459 WITH_JOB_LOCK_GUARD() { 6460 for (job = block_job_next_locked(NULL); job; 6461 job = block_job_next_locked(job)) { 6462 GSList *el; 6463 6464 xdbg_graph_add_node(gr, job, X_DBG_BLOCK_GRAPH_NODE_TYPE_BLOCK_JOB, 6465 job->job.id); 6466 for (el = job->nodes; el; el = el->next) { 6467 xdbg_graph_add_edge(gr, job, (BdrvChild *)el->data); 6468 } 6469 } 6470 } 6471 6472 QTAILQ_FOREACH(bs, &graph_bdrv_states, node_list) { 6473 xdbg_graph_add_node(gr, bs, X_DBG_BLOCK_GRAPH_NODE_TYPE_BLOCK_DRIVER, 6474 bs->node_name); 6475 QLIST_FOREACH(child, &bs->children, next) { 6476 xdbg_graph_add_edge(gr, bs, child); 6477 } 6478 } 6479 6480 return xdbg_graph_finalize(gr); 6481 } 6482 6483 BlockDriverState *bdrv_lookup_bs(const char *device, 6484 const char *node_name, 6485 Error **errp) 6486 { 6487 BlockBackend *blk; 6488 BlockDriverState *bs; 6489 6490 GLOBAL_STATE_CODE(); 6491 6492 if (device) { 6493 blk = blk_by_name(device); 6494 6495 if (blk) { 6496 bs = blk_bs(blk); 6497 if (!bs) { 6498 error_setg(errp, "Device '%s' has no medium", device); 6499 } 6500 6501 return bs; 6502 } 6503 } 6504 6505 if (node_name) { 6506 bs = bdrv_find_node(node_name); 6507 6508 if (bs) { 6509 return bs; 6510 } 6511 } 6512 6513 error_setg(errp, "Cannot find device=\'%s\' nor node-name=\'%s\'", 6514 device ? device : "", 6515 node_name ? node_name : ""); 6516 return NULL; 6517 } 6518 6519 /* If 'base' is in the same chain as 'top', return true. Otherwise, 6520 * return false. If either argument is NULL, return false. */ 6521 bool bdrv_chain_contains(BlockDriverState *top, BlockDriverState *base) 6522 { 6523 6524 GLOBAL_STATE_CODE(); 6525 6526 while (top && top != base) { 6527 top = bdrv_filter_or_cow_bs(top); 6528 } 6529 6530 return top != NULL; 6531 } 6532 6533 BlockDriverState *bdrv_next_node(BlockDriverState *bs) 6534 { 6535 GLOBAL_STATE_CODE(); 6536 if (!bs) { 6537 return QTAILQ_FIRST(&graph_bdrv_states); 6538 } 6539 return QTAILQ_NEXT(bs, node_list); 6540 } 6541 6542 BlockDriverState *bdrv_next_all_states(BlockDriverState *bs) 6543 { 6544 GLOBAL_STATE_CODE(); 6545 if (!bs) { 6546 return QTAILQ_FIRST(&all_bdrv_states); 6547 } 6548 return QTAILQ_NEXT(bs, bs_list); 6549 } 6550 6551 const char *bdrv_get_node_name(const BlockDriverState *bs) 6552 { 6553 IO_CODE(); 6554 return bs->node_name; 6555 } 6556 6557 const char *bdrv_get_parent_name(const BlockDriverState *bs) 6558 { 6559 BdrvChild *c; 6560 const char *name; 6561 IO_CODE(); 6562 6563 /* If multiple parents have a name, just pick the first one. */ 6564 QLIST_FOREACH(c, &bs->parents, next_parent) { 6565 if (c->klass->get_name) { 6566 name = c->klass->get_name(c); 6567 if (name && *name) { 6568 return name; 6569 } 6570 } 6571 } 6572 6573 return NULL; 6574 } 6575 6576 /* TODO check what callers really want: bs->node_name or blk_name() */ 6577 const char *bdrv_get_device_name(const BlockDriverState *bs) 6578 { 6579 IO_CODE(); 6580 return bdrv_get_parent_name(bs) ?: ""; 6581 } 6582 6583 /* This can be used to identify nodes that might not have a device 6584 * name associated. Since node and device names live in the same 6585 * namespace, the result is unambiguous. The exception is if both are 6586 * absent, then this returns an empty (non-null) string. */ 6587 const char *bdrv_get_device_or_node_name(const BlockDriverState *bs) 6588 { 6589 IO_CODE(); 6590 return bdrv_get_parent_name(bs) ?: bs->node_name; 6591 } 6592 6593 int bdrv_get_flags(BlockDriverState *bs) 6594 { 6595 IO_CODE(); 6596 return bs->open_flags; 6597 } 6598 6599 int bdrv_has_zero_init_1(BlockDriverState *bs) 6600 { 6601 GLOBAL_STATE_CODE(); 6602 return 1; 6603 } 6604 6605 int coroutine_mixed_fn bdrv_has_zero_init(BlockDriverState *bs) 6606 { 6607 BlockDriverState *filtered; 6608 GLOBAL_STATE_CODE(); 6609 6610 if (!bs->drv) { 6611 return 0; 6612 } 6613 6614 /* If BS is a copy on write image, it is initialized to 6615 the contents of the base image, which may not be zeroes. */ 6616 if (bdrv_cow_child(bs)) { 6617 return 0; 6618 } 6619 if (bs->drv->bdrv_has_zero_init) { 6620 return bs->drv->bdrv_has_zero_init(bs); 6621 } 6622 6623 filtered = bdrv_filter_bs(bs); 6624 if (filtered) { 6625 return bdrv_has_zero_init(filtered); 6626 } 6627 6628 /* safe default */ 6629 return 0; 6630 } 6631 6632 bool bdrv_can_write_zeroes_with_unmap(BlockDriverState *bs) 6633 { 6634 IO_CODE(); 6635 if (!(bs->open_flags & BDRV_O_UNMAP)) { 6636 return false; 6637 } 6638 6639 return bs->supported_zero_flags & BDRV_REQ_MAY_UNMAP; 6640 } 6641 6642 void bdrv_get_backing_filename(BlockDriverState *bs, 6643 char *filename, int filename_size) 6644 { 6645 IO_CODE(); 6646 pstrcpy(filename, filename_size, bs->backing_file); 6647 } 6648 6649 int coroutine_fn bdrv_co_get_info(BlockDriverState *bs, BlockDriverInfo *bdi) 6650 { 6651 int ret; 6652 BlockDriver *drv = bs->drv; 6653 IO_CODE(); 6654 assert_bdrv_graph_readable(); 6655 6656 /* if bs->drv == NULL, bs is closed, so there's nothing to do here */ 6657 if (!drv) { 6658 return -ENOMEDIUM; 6659 } 6660 if (!drv->bdrv_co_get_info) { 6661 BlockDriverState *filtered = bdrv_filter_bs(bs); 6662 if (filtered) { 6663 return bdrv_co_get_info(filtered, bdi); 6664 } 6665 return -ENOTSUP; 6666 } 6667 memset(bdi, 0, sizeof(*bdi)); 6668 ret = drv->bdrv_co_get_info(bs, bdi); 6669 if (bdi->subcluster_size == 0) { 6670 /* 6671 * If the driver left this unset, subclusters are not supported. 6672 * Then it is safe to treat each cluster as having only one subcluster. 6673 */ 6674 bdi->subcluster_size = bdi->cluster_size; 6675 } 6676 if (ret < 0) { 6677 return ret; 6678 } 6679 6680 if (bdi->cluster_size > BDRV_MAX_ALIGNMENT) { 6681 return -EINVAL; 6682 } 6683 6684 return 0; 6685 } 6686 6687 ImageInfoSpecific *bdrv_get_specific_info(BlockDriverState *bs, 6688 Error **errp) 6689 { 6690 BlockDriver *drv = bs->drv; 6691 IO_CODE(); 6692 if (drv && drv->bdrv_get_specific_info) { 6693 return drv->bdrv_get_specific_info(bs, errp); 6694 } 6695 return NULL; 6696 } 6697 6698 BlockStatsSpecific *bdrv_get_specific_stats(BlockDriverState *bs) 6699 { 6700 BlockDriver *drv = bs->drv; 6701 IO_CODE(); 6702 if (!drv || !drv->bdrv_get_specific_stats) { 6703 return NULL; 6704 } 6705 return drv->bdrv_get_specific_stats(bs); 6706 } 6707 6708 void coroutine_fn bdrv_co_debug_event(BlockDriverState *bs, BlkdebugEvent event) 6709 { 6710 IO_CODE(); 6711 assert_bdrv_graph_readable(); 6712 6713 if (!bs || !bs->drv || !bs->drv->bdrv_co_debug_event) { 6714 return; 6715 } 6716 6717 bs->drv->bdrv_co_debug_event(bs, event); 6718 } 6719 6720 static BlockDriverState * GRAPH_RDLOCK 6721 bdrv_find_debug_node(BlockDriverState *bs) 6722 { 6723 GLOBAL_STATE_CODE(); 6724 while (bs && bs->drv && !bs->drv->bdrv_debug_breakpoint) { 6725 bs = bdrv_primary_bs(bs); 6726 } 6727 6728 if (bs && bs->drv && bs->drv->bdrv_debug_breakpoint) { 6729 assert(bs->drv->bdrv_debug_remove_breakpoint); 6730 return bs; 6731 } 6732 6733 return NULL; 6734 } 6735 6736 int bdrv_debug_breakpoint(BlockDriverState *bs, const char *event, 6737 const char *tag) 6738 { 6739 GLOBAL_STATE_CODE(); 6740 GRAPH_RDLOCK_GUARD_MAINLOOP(); 6741 6742 bs = bdrv_find_debug_node(bs); 6743 if (bs) { 6744 return bs->drv->bdrv_debug_breakpoint(bs, event, tag); 6745 } 6746 6747 return -ENOTSUP; 6748 } 6749 6750 int bdrv_debug_remove_breakpoint(BlockDriverState *bs, const char *tag) 6751 { 6752 GLOBAL_STATE_CODE(); 6753 GRAPH_RDLOCK_GUARD_MAINLOOP(); 6754 6755 bs = bdrv_find_debug_node(bs); 6756 if (bs) { 6757 return bs->drv->bdrv_debug_remove_breakpoint(bs, tag); 6758 } 6759 6760 return -ENOTSUP; 6761 } 6762 6763 int bdrv_debug_resume(BlockDriverState *bs, const char *tag) 6764 { 6765 GLOBAL_STATE_CODE(); 6766 GRAPH_RDLOCK_GUARD_MAINLOOP(); 6767 6768 while (bs && (!bs->drv || !bs->drv->bdrv_debug_resume)) { 6769 bs = bdrv_primary_bs(bs); 6770 } 6771 6772 if (bs && bs->drv && bs->drv->bdrv_debug_resume) { 6773 return bs->drv->bdrv_debug_resume(bs, tag); 6774 } 6775 6776 return -ENOTSUP; 6777 } 6778 6779 bool bdrv_debug_is_suspended(BlockDriverState *bs, const char *tag) 6780 { 6781 GLOBAL_STATE_CODE(); 6782 GRAPH_RDLOCK_GUARD_MAINLOOP(); 6783 6784 while (bs && bs->drv && !bs->drv->bdrv_debug_is_suspended) { 6785 bs = bdrv_primary_bs(bs); 6786 } 6787 6788 if (bs && bs->drv && bs->drv->bdrv_debug_is_suspended) { 6789 return bs->drv->bdrv_debug_is_suspended(bs, tag); 6790 } 6791 6792 return false; 6793 } 6794 6795 /* backing_file can either be relative, or absolute, or a protocol. If it is 6796 * relative, it must be relative to the chain. So, passing in bs->filename 6797 * from a BDS as backing_file should not be done, as that may be relative to 6798 * the CWD rather than the chain. */ 6799 BlockDriverState *bdrv_find_backing_image(BlockDriverState *bs, 6800 const char *backing_file) 6801 { 6802 char *filename_full = NULL; 6803 char *backing_file_full = NULL; 6804 char *filename_tmp = NULL; 6805 int is_protocol = 0; 6806 bool filenames_refreshed = false; 6807 BlockDriverState *curr_bs = NULL; 6808 BlockDriverState *retval = NULL; 6809 BlockDriverState *bs_below; 6810 6811 GLOBAL_STATE_CODE(); 6812 GRAPH_RDLOCK_GUARD_MAINLOOP(); 6813 6814 if (!bs || !bs->drv || !backing_file) { 6815 return NULL; 6816 } 6817 6818 filename_full = g_malloc(PATH_MAX); 6819 backing_file_full = g_malloc(PATH_MAX); 6820 6821 is_protocol = path_has_protocol(backing_file); 6822 6823 /* 6824 * Being largely a legacy function, skip any filters here 6825 * (because filters do not have normal filenames, so they cannot 6826 * match anyway; and allowing json:{} filenames is a bit out of 6827 * scope). 6828 */ 6829 for (curr_bs = bdrv_skip_filters(bs); 6830 bdrv_cow_child(curr_bs) != NULL; 6831 curr_bs = bs_below) 6832 { 6833 bs_below = bdrv_backing_chain_next(curr_bs); 6834 6835 if (bdrv_backing_overridden(curr_bs)) { 6836 /* 6837 * If the backing file was overridden, we can only compare 6838 * directly against the backing node's filename. 6839 */ 6840 6841 if (!filenames_refreshed) { 6842 /* 6843 * This will automatically refresh all of the 6844 * filenames in the rest of the backing chain, so we 6845 * only need to do this once. 6846 */ 6847 bdrv_refresh_filename(bs_below); 6848 filenames_refreshed = true; 6849 } 6850 6851 if (strcmp(backing_file, bs_below->filename) == 0) { 6852 retval = bs_below; 6853 break; 6854 } 6855 } else if (is_protocol || path_has_protocol(curr_bs->backing_file)) { 6856 /* 6857 * If either of the filename paths is actually a protocol, then 6858 * compare unmodified paths; otherwise make paths relative. 6859 */ 6860 char *backing_file_full_ret; 6861 6862 if (strcmp(backing_file, curr_bs->backing_file) == 0) { 6863 retval = bs_below; 6864 break; 6865 } 6866 /* Also check against the full backing filename for the image */ 6867 backing_file_full_ret = bdrv_get_full_backing_filename(curr_bs, 6868 NULL); 6869 if (backing_file_full_ret) { 6870 bool equal = strcmp(backing_file, backing_file_full_ret) == 0; 6871 g_free(backing_file_full_ret); 6872 if (equal) { 6873 retval = bs_below; 6874 break; 6875 } 6876 } 6877 } else { 6878 /* If not an absolute filename path, make it relative to the current 6879 * image's filename path */ 6880 filename_tmp = bdrv_make_absolute_filename(curr_bs, backing_file, 6881 NULL); 6882 /* We are going to compare canonicalized absolute pathnames */ 6883 if (!filename_tmp || !realpath(filename_tmp, filename_full)) { 6884 g_free(filename_tmp); 6885 continue; 6886 } 6887 g_free(filename_tmp); 6888 6889 /* We need to make sure the backing filename we are comparing against 6890 * is relative to the current image filename (or absolute) */ 6891 filename_tmp = bdrv_get_full_backing_filename(curr_bs, NULL); 6892 if (!filename_tmp || !realpath(filename_tmp, backing_file_full)) { 6893 g_free(filename_tmp); 6894 continue; 6895 } 6896 g_free(filename_tmp); 6897 6898 if (strcmp(backing_file_full, filename_full) == 0) { 6899 retval = bs_below; 6900 break; 6901 } 6902 } 6903 } 6904 6905 g_free(filename_full); 6906 g_free(backing_file_full); 6907 return retval; 6908 } 6909 6910 void bdrv_init(void) 6911 { 6912 #ifdef CONFIG_BDRV_WHITELIST_TOOLS 6913 use_bdrv_whitelist = 1; 6914 #endif 6915 module_call_init(MODULE_INIT_BLOCK); 6916 } 6917 6918 void bdrv_init_with_whitelist(void) 6919 { 6920 use_bdrv_whitelist = 1; 6921 bdrv_init(); 6922 } 6923 6924 int bdrv_activate(BlockDriverState *bs, Error **errp) 6925 { 6926 BdrvChild *child, *parent; 6927 Error *local_err = NULL; 6928 int ret; 6929 BdrvDirtyBitmap *bm; 6930 6931 GLOBAL_STATE_CODE(); 6932 GRAPH_RDLOCK_GUARD_MAINLOOP(); 6933 6934 if (!bs->drv) { 6935 return -ENOMEDIUM; 6936 } 6937 6938 QLIST_FOREACH(child, &bs->children, next) { 6939 bdrv_activate(child->bs, &local_err); 6940 if (local_err) { 6941 error_propagate(errp, local_err); 6942 return -EINVAL; 6943 } 6944 } 6945 6946 /* 6947 * Update permissions, they may differ for inactive nodes. 6948 * 6949 * Note that the required permissions of inactive images are always a 6950 * subset of the permissions required after activating the image. This 6951 * allows us to just get the permissions upfront without restricting 6952 * bdrv_co_invalidate_cache(). 6953 * 6954 * It also means that in error cases, we don't have to try and revert to 6955 * the old permissions (which is an operation that could fail, too). We can 6956 * just keep the extended permissions for the next time that an activation 6957 * of the image is tried. 6958 */ 6959 if (bs->open_flags & BDRV_O_INACTIVE) { 6960 bs->open_flags &= ~BDRV_O_INACTIVE; 6961 ret = bdrv_refresh_perms(bs, NULL, errp); 6962 if (ret < 0) { 6963 bs->open_flags |= BDRV_O_INACTIVE; 6964 return ret; 6965 } 6966 6967 ret = bdrv_invalidate_cache(bs, errp); 6968 if (ret < 0) { 6969 bs->open_flags |= BDRV_O_INACTIVE; 6970 return ret; 6971 } 6972 6973 FOR_EACH_DIRTY_BITMAP(bs, bm) { 6974 bdrv_dirty_bitmap_skip_store(bm, false); 6975 } 6976 6977 ret = bdrv_refresh_total_sectors(bs, bs->total_sectors); 6978 if (ret < 0) { 6979 bs->open_flags |= BDRV_O_INACTIVE; 6980 error_setg_errno(errp, -ret, "Could not refresh total sector count"); 6981 return ret; 6982 } 6983 } 6984 6985 QLIST_FOREACH(parent, &bs->parents, next_parent) { 6986 if (parent->klass->activate) { 6987 parent->klass->activate(parent, &local_err); 6988 if (local_err) { 6989 bs->open_flags |= BDRV_O_INACTIVE; 6990 error_propagate(errp, local_err); 6991 return -EINVAL; 6992 } 6993 } 6994 } 6995 6996 return 0; 6997 } 6998 6999 int coroutine_fn bdrv_co_invalidate_cache(BlockDriverState *bs, Error **errp) 7000 { 7001 Error *local_err = NULL; 7002 IO_CODE(); 7003 7004 assert(!(bs->open_flags & BDRV_O_INACTIVE)); 7005 assert_bdrv_graph_readable(); 7006 7007 if (bs->drv->bdrv_co_invalidate_cache) { 7008 bs->drv->bdrv_co_invalidate_cache(bs, &local_err); 7009 if (local_err) { 7010 error_propagate(errp, local_err); 7011 return -EINVAL; 7012 } 7013 } 7014 7015 return 0; 7016 } 7017 7018 void bdrv_activate_all(Error **errp) 7019 { 7020 BlockDriverState *bs; 7021 BdrvNextIterator it; 7022 7023 GLOBAL_STATE_CODE(); 7024 GRAPH_RDLOCK_GUARD_MAINLOOP(); 7025 7026 for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) { 7027 AioContext *aio_context = bdrv_get_aio_context(bs); 7028 int ret; 7029 7030 aio_context_acquire(aio_context); 7031 ret = bdrv_activate(bs, errp); 7032 aio_context_release(aio_context); 7033 if (ret < 0) { 7034 bdrv_next_cleanup(&it); 7035 return; 7036 } 7037 } 7038 } 7039 7040 static bool GRAPH_RDLOCK 7041 bdrv_has_bds_parent(BlockDriverState *bs, bool only_active) 7042 { 7043 BdrvChild *parent; 7044 GLOBAL_STATE_CODE(); 7045 7046 QLIST_FOREACH(parent, &bs->parents, next_parent) { 7047 if (parent->klass->parent_is_bds) { 7048 BlockDriverState *parent_bs = parent->opaque; 7049 if (!only_active || !(parent_bs->open_flags & BDRV_O_INACTIVE)) { 7050 return true; 7051 } 7052 } 7053 } 7054 7055 return false; 7056 } 7057 7058 static int GRAPH_RDLOCK bdrv_inactivate_recurse(BlockDriverState *bs) 7059 { 7060 BdrvChild *child, *parent; 7061 int ret; 7062 uint64_t cumulative_perms, cumulative_shared_perms; 7063 7064 GLOBAL_STATE_CODE(); 7065 7066 if (!bs->drv) { 7067 return -ENOMEDIUM; 7068 } 7069 7070 /* Make sure that we don't inactivate a child before its parent. 7071 * It will be covered by recursion from the yet active parent. */ 7072 if (bdrv_has_bds_parent(bs, true)) { 7073 return 0; 7074 } 7075 7076 assert(!(bs->open_flags & BDRV_O_INACTIVE)); 7077 7078 /* Inactivate this node */ 7079 if (bs->drv->bdrv_inactivate) { 7080 ret = bs->drv->bdrv_inactivate(bs); 7081 if (ret < 0) { 7082 return ret; 7083 } 7084 } 7085 7086 QLIST_FOREACH(parent, &bs->parents, next_parent) { 7087 if (parent->klass->inactivate) { 7088 ret = parent->klass->inactivate(parent); 7089 if (ret < 0) { 7090 return ret; 7091 } 7092 } 7093 } 7094 7095 bdrv_get_cumulative_perm(bs, &cumulative_perms, 7096 &cumulative_shared_perms); 7097 if (cumulative_perms & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED)) { 7098 /* Our inactive parents still need write access. Inactivation failed. */ 7099 return -EPERM; 7100 } 7101 7102 bs->open_flags |= BDRV_O_INACTIVE; 7103 7104 /* 7105 * Update permissions, they may differ for inactive nodes. 7106 * We only tried to loosen restrictions, so errors are not fatal, ignore 7107 * them. 7108 */ 7109 bdrv_refresh_perms(bs, NULL, NULL); 7110 7111 /* Recursively inactivate children */ 7112 QLIST_FOREACH(child, &bs->children, next) { 7113 ret = bdrv_inactivate_recurse(child->bs); 7114 if (ret < 0) { 7115 return ret; 7116 } 7117 } 7118 7119 return 0; 7120 } 7121 7122 int bdrv_inactivate_all(void) 7123 { 7124 BlockDriverState *bs = NULL; 7125 BdrvNextIterator it; 7126 int ret = 0; 7127 GSList *aio_ctxs = NULL, *ctx; 7128 7129 GLOBAL_STATE_CODE(); 7130 GRAPH_RDLOCK_GUARD_MAINLOOP(); 7131 7132 for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) { 7133 AioContext *aio_context = bdrv_get_aio_context(bs); 7134 7135 if (!g_slist_find(aio_ctxs, aio_context)) { 7136 aio_ctxs = g_slist_prepend(aio_ctxs, aio_context); 7137 aio_context_acquire(aio_context); 7138 } 7139 } 7140 7141 for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) { 7142 /* Nodes with BDS parents are covered by recursion from the last 7143 * parent that gets inactivated. Don't inactivate them a second 7144 * time if that has already happened. */ 7145 if (bdrv_has_bds_parent(bs, false)) { 7146 continue; 7147 } 7148 ret = bdrv_inactivate_recurse(bs); 7149 if (ret < 0) { 7150 bdrv_next_cleanup(&it); 7151 goto out; 7152 } 7153 } 7154 7155 out: 7156 for (ctx = aio_ctxs; ctx != NULL; ctx = ctx->next) { 7157 AioContext *aio_context = ctx->data; 7158 aio_context_release(aio_context); 7159 } 7160 g_slist_free(aio_ctxs); 7161 7162 return ret; 7163 } 7164 7165 /**************************************************************/ 7166 /* removable device support */ 7167 7168 /** 7169 * Return TRUE if the media is present 7170 */ 7171 bool coroutine_fn bdrv_co_is_inserted(BlockDriverState *bs) 7172 { 7173 BlockDriver *drv = bs->drv; 7174 BdrvChild *child; 7175 IO_CODE(); 7176 assert_bdrv_graph_readable(); 7177 7178 if (!drv) { 7179 return false; 7180 } 7181 if (drv->bdrv_co_is_inserted) { 7182 return drv->bdrv_co_is_inserted(bs); 7183 } 7184 QLIST_FOREACH(child, &bs->children, next) { 7185 if (!bdrv_co_is_inserted(child->bs)) { 7186 return false; 7187 } 7188 } 7189 return true; 7190 } 7191 7192 /** 7193 * If eject_flag is TRUE, eject the media. Otherwise, close the tray 7194 */ 7195 void coroutine_fn bdrv_co_eject(BlockDriverState *bs, bool eject_flag) 7196 { 7197 BlockDriver *drv = bs->drv; 7198 IO_CODE(); 7199 assert_bdrv_graph_readable(); 7200 7201 if (drv && drv->bdrv_co_eject) { 7202 drv->bdrv_co_eject(bs, eject_flag); 7203 } 7204 } 7205 7206 /** 7207 * Lock or unlock the media (if it is locked, the user won't be able 7208 * to eject it manually). 7209 */ 7210 void coroutine_fn bdrv_co_lock_medium(BlockDriverState *bs, bool locked) 7211 { 7212 BlockDriver *drv = bs->drv; 7213 IO_CODE(); 7214 assert_bdrv_graph_readable(); 7215 trace_bdrv_lock_medium(bs, locked); 7216 7217 if (drv && drv->bdrv_co_lock_medium) { 7218 drv->bdrv_co_lock_medium(bs, locked); 7219 } 7220 } 7221 7222 /* Get a reference to bs */ 7223 void bdrv_ref(BlockDriverState *bs) 7224 { 7225 GLOBAL_STATE_CODE(); 7226 bs->refcnt++; 7227 } 7228 7229 /* Release a previously grabbed reference to bs. 7230 * If after releasing, reference count is zero, the BlockDriverState is 7231 * deleted. */ 7232 void bdrv_unref(BlockDriverState *bs) 7233 { 7234 GLOBAL_STATE_CODE(); 7235 if (!bs) { 7236 return; 7237 } 7238 assert(bs->refcnt > 0); 7239 if (--bs->refcnt == 0) { 7240 bdrv_delete(bs); 7241 } 7242 } 7243 7244 /* 7245 * Release a BlockDriverState reference while holding the graph write lock. 7246 * 7247 * Calling bdrv_unref() directly is forbidden while holding the graph lock 7248 * because bdrv_close() both involves polling and taking the graph lock 7249 * internally. bdrv_schedule_unref() instead delays decreasing the refcount and 7250 * possibly closing @bs until the graph lock is released. 7251 */ 7252 void bdrv_schedule_unref(BlockDriverState *bs) 7253 { 7254 if (!bs) { 7255 return; 7256 } 7257 aio_bh_schedule_oneshot(qemu_get_aio_context(), 7258 (QEMUBHFunc *) bdrv_unref, bs); 7259 } 7260 7261 struct BdrvOpBlocker { 7262 Error *reason; 7263 QLIST_ENTRY(BdrvOpBlocker) list; 7264 }; 7265 7266 bool bdrv_op_is_blocked(BlockDriverState *bs, BlockOpType op, Error **errp) 7267 { 7268 BdrvOpBlocker *blocker; 7269 GLOBAL_STATE_CODE(); 7270 7271 assert((int) op >= 0 && op < BLOCK_OP_TYPE_MAX); 7272 if (!QLIST_EMPTY(&bs->op_blockers[op])) { 7273 blocker = QLIST_FIRST(&bs->op_blockers[op]); 7274 error_propagate_prepend(errp, error_copy(blocker->reason), 7275 "Node '%s' is busy: ", 7276 bdrv_get_device_or_node_name(bs)); 7277 return true; 7278 } 7279 return false; 7280 } 7281 7282 void bdrv_op_block(BlockDriverState *bs, BlockOpType op, Error *reason) 7283 { 7284 BdrvOpBlocker *blocker; 7285 GLOBAL_STATE_CODE(); 7286 assert((int) op >= 0 && op < BLOCK_OP_TYPE_MAX); 7287 7288 blocker = g_new0(BdrvOpBlocker, 1); 7289 blocker->reason = reason; 7290 QLIST_INSERT_HEAD(&bs->op_blockers[op], blocker, list); 7291 } 7292 7293 void bdrv_op_unblock(BlockDriverState *bs, BlockOpType op, Error *reason) 7294 { 7295 BdrvOpBlocker *blocker, *next; 7296 GLOBAL_STATE_CODE(); 7297 assert((int) op >= 0 && op < BLOCK_OP_TYPE_MAX); 7298 QLIST_FOREACH_SAFE(blocker, &bs->op_blockers[op], list, next) { 7299 if (blocker->reason == reason) { 7300 QLIST_REMOVE(blocker, list); 7301 g_free(blocker); 7302 } 7303 } 7304 } 7305 7306 void bdrv_op_block_all(BlockDriverState *bs, Error *reason) 7307 { 7308 int i; 7309 GLOBAL_STATE_CODE(); 7310 for (i = 0; i < BLOCK_OP_TYPE_MAX; i++) { 7311 bdrv_op_block(bs, i, reason); 7312 } 7313 } 7314 7315 void bdrv_op_unblock_all(BlockDriverState *bs, Error *reason) 7316 { 7317 int i; 7318 GLOBAL_STATE_CODE(); 7319 for (i = 0; i < BLOCK_OP_TYPE_MAX; i++) { 7320 bdrv_op_unblock(bs, i, reason); 7321 } 7322 } 7323 7324 bool bdrv_op_blocker_is_empty(BlockDriverState *bs) 7325 { 7326 int i; 7327 GLOBAL_STATE_CODE(); 7328 for (i = 0; i < BLOCK_OP_TYPE_MAX; i++) { 7329 if (!QLIST_EMPTY(&bs->op_blockers[i])) { 7330 return false; 7331 } 7332 } 7333 return true; 7334 } 7335 7336 /* 7337 * Must not be called while holding the lock of an AioContext other than the 7338 * current one. 7339 */ 7340 void bdrv_img_create(const char *filename, const char *fmt, 7341 const char *base_filename, const char *base_fmt, 7342 char *options, uint64_t img_size, int flags, bool quiet, 7343 Error **errp) 7344 { 7345 QemuOptsList *create_opts = NULL; 7346 QemuOpts *opts = NULL; 7347 const char *backing_fmt, *backing_file; 7348 int64_t size; 7349 BlockDriver *drv, *proto_drv; 7350 Error *local_err = NULL; 7351 int ret = 0; 7352 7353 GLOBAL_STATE_CODE(); 7354 7355 /* Find driver and parse its options */ 7356 drv = bdrv_find_format(fmt); 7357 if (!drv) { 7358 error_setg(errp, "Unknown file format '%s'", fmt); 7359 return; 7360 } 7361 7362 proto_drv = bdrv_find_protocol(filename, true, errp); 7363 if (!proto_drv) { 7364 return; 7365 } 7366 7367 if (!drv->create_opts) { 7368 error_setg(errp, "Format driver '%s' does not support image creation", 7369 drv->format_name); 7370 return; 7371 } 7372 7373 if (!proto_drv->create_opts) { 7374 error_setg(errp, "Protocol driver '%s' does not support image creation", 7375 proto_drv->format_name); 7376 return; 7377 } 7378 7379 aio_context_acquire(qemu_get_aio_context()); 7380 7381 /* Create parameter list */ 7382 create_opts = qemu_opts_append(create_opts, drv->create_opts); 7383 create_opts = qemu_opts_append(create_opts, proto_drv->create_opts); 7384 7385 opts = qemu_opts_create(create_opts, NULL, 0, &error_abort); 7386 7387 /* Parse -o options */ 7388 if (options) { 7389 if (!qemu_opts_do_parse(opts, options, NULL, errp)) { 7390 goto out; 7391 } 7392 } 7393 7394 if (!qemu_opt_get(opts, BLOCK_OPT_SIZE)) { 7395 qemu_opt_set_number(opts, BLOCK_OPT_SIZE, img_size, &error_abort); 7396 } else if (img_size != UINT64_C(-1)) { 7397 error_setg(errp, "The image size must be specified only once"); 7398 goto out; 7399 } 7400 7401 if (base_filename) { 7402 if (!qemu_opt_set(opts, BLOCK_OPT_BACKING_FILE, base_filename, 7403 NULL)) { 7404 error_setg(errp, "Backing file not supported for file format '%s'", 7405 fmt); 7406 goto out; 7407 } 7408 } 7409 7410 if (base_fmt) { 7411 if (!qemu_opt_set(opts, BLOCK_OPT_BACKING_FMT, base_fmt, NULL)) { 7412 error_setg(errp, "Backing file format not supported for file " 7413 "format '%s'", fmt); 7414 goto out; 7415 } 7416 } 7417 7418 backing_file = qemu_opt_get(opts, BLOCK_OPT_BACKING_FILE); 7419 if (backing_file) { 7420 if (!strcmp(filename, backing_file)) { 7421 error_setg(errp, "Error: Trying to create an image with the " 7422 "same filename as the backing file"); 7423 goto out; 7424 } 7425 if (backing_file[0] == '\0') { 7426 error_setg(errp, "Expected backing file name, got empty string"); 7427 goto out; 7428 } 7429 } 7430 7431 backing_fmt = qemu_opt_get(opts, BLOCK_OPT_BACKING_FMT); 7432 7433 /* The size for the image must always be specified, unless we have a backing 7434 * file and we have not been forbidden from opening it. */ 7435 size = qemu_opt_get_size(opts, BLOCK_OPT_SIZE, img_size); 7436 if (backing_file && !(flags & BDRV_O_NO_BACKING)) { 7437 BlockDriverState *bs; 7438 char *full_backing; 7439 int back_flags; 7440 QDict *backing_options = NULL; 7441 7442 full_backing = 7443 bdrv_get_full_backing_filename_from_filename(filename, backing_file, 7444 &local_err); 7445 if (local_err) { 7446 goto out; 7447 } 7448 assert(full_backing); 7449 7450 /* 7451 * No need to do I/O here, which allows us to open encrypted 7452 * backing images without needing the secret 7453 */ 7454 back_flags = flags; 7455 back_flags &= ~(BDRV_O_RDWR | BDRV_O_SNAPSHOT | BDRV_O_NO_BACKING); 7456 back_flags |= BDRV_O_NO_IO; 7457 7458 backing_options = qdict_new(); 7459 if (backing_fmt) { 7460 qdict_put_str(backing_options, "driver", backing_fmt); 7461 } 7462 qdict_put_bool(backing_options, BDRV_OPT_FORCE_SHARE, true); 7463 7464 bs = bdrv_open(full_backing, NULL, backing_options, back_flags, 7465 &local_err); 7466 g_free(full_backing); 7467 if (!bs) { 7468 error_append_hint(&local_err, "Could not open backing image.\n"); 7469 goto out; 7470 } else { 7471 if (!backing_fmt) { 7472 error_setg(&local_err, 7473 "Backing file specified without backing format"); 7474 error_append_hint(&local_err, "Detected format of %s.\n", 7475 bs->drv->format_name); 7476 goto out; 7477 } 7478 if (size == -1) { 7479 /* Opened BS, have no size */ 7480 size = bdrv_getlength(bs); 7481 if (size < 0) { 7482 error_setg_errno(errp, -size, "Could not get size of '%s'", 7483 backing_file); 7484 bdrv_unref(bs); 7485 goto out; 7486 } 7487 qemu_opt_set_number(opts, BLOCK_OPT_SIZE, size, &error_abort); 7488 } 7489 bdrv_unref(bs); 7490 } 7491 /* (backing_file && !(flags & BDRV_O_NO_BACKING)) */ 7492 } else if (backing_file && !backing_fmt) { 7493 error_setg(&local_err, 7494 "Backing file specified without backing format"); 7495 goto out; 7496 } 7497 7498 if (size == -1) { 7499 error_setg(errp, "Image creation needs a size parameter"); 7500 goto out; 7501 } 7502 7503 if (!quiet) { 7504 printf("Formatting '%s', fmt=%s ", filename, fmt); 7505 qemu_opts_print(opts, " "); 7506 puts(""); 7507 fflush(stdout); 7508 } 7509 7510 ret = bdrv_create(drv, filename, opts, &local_err); 7511 7512 if (ret == -EFBIG) { 7513 /* This is generally a better message than whatever the driver would 7514 * deliver (especially because of the cluster_size_hint), since that 7515 * is most probably not much different from "image too large". */ 7516 const char *cluster_size_hint = ""; 7517 if (qemu_opt_get_size(opts, BLOCK_OPT_CLUSTER_SIZE, 0)) { 7518 cluster_size_hint = " (try using a larger cluster size)"; 7519 } 7520 error_setg(errp, "The image size is too large for file format '%s'" 7521 "%s", fmt, cluster_size_hint); 7522 error_free(local_err); 7523 local_err = NULL; 7524 } 7525 7526 out: 7527 qemu_opts_del(opts); 7528 qemu_opts_free(create_opts); 7529 error_propagate(errp, local_err); 7530 aio_context_release(qemu_get_aio_context()); 7531 } 7532 7533 AioContext *bdrv_get_aio_context(BlockDriverState *bs) 7534 { 7535 IO_CODE(); 7536 return bs ? bs->aio_context : qemu_get_aio_context(); 7537 } 7538 7539 AioContext *coroutine_fn bdrv_co_enter(BlockDriverState *bs) 7540 { 7541 Coroutine *self = qemu_coroutine_self(); 7542 AioContext *old_ctx = qemu_coroutine_get_aio_context(self); 7543 AioContext *new_ctx; 7544 IO_CODE(); 7545 7546 /* 7547 * Increase bs->in_flight to ensure that this operation is completed before 7548 * moving the node to a different AioContext. Read new_ctx only afterwards. 7549 */ 7550 bdrv_inc_in_flight(bs); 7551 7552 new_ctx = bdrv_get_aio_context(bs); 7553 aio_co_reschedule_self(new_ctx); 7554 return old_ctx; 7555 } 7556 7557 void coroutine_fn bdrv_co_leave(BlockDriverState *bs, AioContext *old_ctx) 7558 { 7559 IO_CODE(); 7560 aio_co_reschedule_self(old_ctx); 7561 bdrv_dec_in_flight(bs); 7562 } 7563 7564 void coroutine_fn bdrv_co_lock(BlockDriverState *bs) 7565 { 7566 AioContext *ctx = bdrv_get_aio_context(bs); 7567 7568 /* In the main thread, bs->aio_context won't change concurrently */ 7569 assert(qemu_get_current_aio_context() == qemu_get_aio_context()); 7570 7571 /* 7572 * We're in coroutine context, so we already hold the lock of the main 7573 * loop AioContext. Don't lock it twice to avoid deadlocks. 7574 */ 7575 assert(qemu_in_coroutine()); 7576 if (ctx != qemu_get_aio_context()) { 7577 aio_context_acquire(ctx); 7578 } 7579 } 7580 7581 void coroutine_fn bdrv_co_unlock(BlockDriverState *bs) 7582 { 7583 AioContext *ctx = bdrv_get_aio_context(bs); 7584 7585 assert(qemu_in_coroutine()); 7586 if (ctx != qemu_get_aio_context()) { 7587 aio_context_release(ctx); 7588 } 7589 } 7590 7591 static void bdrv_do_remove_aio_context_notifier(BdrvAioNotifier *ban) 7592 { 7593 GLOBAL_STATE_CODE(); 7594 QLIST_REMOVE(ban, list); 7595 g_free(ban); 7596 } 7597 7598 static void bdrv_detach_aio_context(BlockDriverState *bs) 7599 { 7600 BdrvAioNotifier *baf, *baf_tmp; 7601 7602 assert(!bs->walking_aio_notifiers); 7603 GLOBAL_STATE_CODE(); 7604 bs->walking_aio_notifiers = true; 7605 QLIST_FOREACH_SAFE(baf, &bs->aio_notifiers, list, baf_tmp) { 7606 if (baf->deleted) { 7607 bdrv_do_remove_aio_context_notifier(baf); 7608 } else { 7609 baf->detach_aio_context(baf->opaque); 7610 } 7611 } 7612 /* Never mind iterating again to check for ->deleted. bdrv_close() will 7613 * remove remaining aio notifiers if we aren't called again. 7614 */ 7615 bs->walking_aio_notifiers = false; 7616 7617 if (bs->drv && bs->drv->bdrv_detach_aio_context) { 7618 bs->drv->bdrv_detach_aio_context(bs); 7619 } 7620 7621 bs->aio_context = NULL; 7622 } 7623 7624 static void bdrv_attach_aio_context(BlockDriverState *bs, 7625 AioContext *new_context) 7626 { 7627 BdrvAioNotifier *ban, *ban_tmp; 7628 GLOBAL_STATE_CODE(); 7629 7630 bs->aio_context = new_context; 7631 7632 if (bs->drv && bs->drv->bdrv_attach_aio_context) { 7633 bs->drv->bdrv_attach_aio_context(bs, new_context); 7634 } 7635 7636 assert(!bs->walking_aio_notifiers); 7637 bs->walking_aio_notifiers = true; 7638 QLIST_FOREACH_SAFE(ban, &bs->aio_notifiers, list, ban_tmp) { 7639 if (ban->deleted) { 7640 bdrv_do_remove_aio_context_notifier(ban); 7641 } else { 7642 ban->attached_aio_context(new_context, ban->opaque); 7643 } 7644 } 7645 bs->walking_aio_notifiers = false; 7646 } 7647 7648 typedef struct BdrvStateSetAioContext { 7649 AioContext *new_ctx; 7650 BlockDriverState *bs; 7651 } BdrvStateSetAioContext; 7652 7653 static bool bdrv_parent_change_aio_context(BdrvChild *c, AioContext *ctx, 7654 GHashTable *visited, 7655 Transaction *tran, 7656 Error **errp) 7657 { 7658 GLOBAL_STATE_CODE(); 7659 if (g_hash_table_contains(visited, c)) { 7660 return true; 7661 } 7662 g_hash_table_add(visited, c); 7663 7664 /* 7665 * A BdrvChildClass that doesn't handle AioContext changes cannot 7666 * tolerate any AioContext changes 7667 */ 7668 if (!c->klass->change_aio_ctx) { 7669 char *user = bdrv_child_user_desc(c); 7670 error_setg(errp, "Changing iothreads is not supported by %s", user); 7671 g_free(user); 7672 return false; 7673 } 7674 if (!c->klass->change_aio_ctx(c, ctx, visited, tran, errp)) { 7675 assert(!errp || *errp); 7676 return false; 7677 } 7678 return true; 7679 } 7680 7681 bool bdrv_child_change_aio_context(BdrvChild *c, AioContext *ctx, 7682 GHashTable *visited, Transaction *tran, 7683 Error **errp) 7684 { 7685 GLOBAL_STATE_CODE(); 7686 if (g_hash_table_contains(visited, c)) { 7687 return true; 7688 } 7689 g_hash_table_add(visited, c); 7690 return bdrv_change_aio_context(c->bs, ctx, visited, tran, errp); 7691 } 7692 7693 static void bdrv_set_aio_context_clean(void *opaque) 7694 { 7695 BdrvStateSetAioContext *state = (BdrvStateSetAioContext *) opaque; 7696 BlockDriverState *bs = (BlockDriverState *) state->bs; 7697 7698 /* Paired with bdrv_drained_begin in bdrv_change_aio_context() */ 7699 bdrv_drained_end(bs); 7700 7701 g_free(state); 7702 } 7703 7704 static void bdrv_set_aio_context_commit(void *opaque) 7705 { 7706 BdrvStateSetAioContext *state = (BdrvStateSetAioContext *) opaque; 7707 BlockDriverState *bs = (BlockDriverState *) state->bs; 7708 AioContext *new_context = state->new_ctx; 7709 AioContext *old_context = bdrv_get_aio_context(bs); 7710 7711 /* 7712 * Take the old AioContex when detaching it from bs. 7713 * At this point, new_context lock is already acquired, and we are now 7714 * also taking old_context. This is safe as long as bdrv_detach_aio_context 7715 * does not call AIO_POLL_WHILE(). 7716 */ 7717 if (old_context != qemu_get_aio_context()) { 7718 aio_context_acquire(old_context); 7719 } 7720 bdrv_detach_aio_context(bs); 7721 if (old_context != qemu_get_aio_context()) { 7722 aio_context_release(old_context); 7723 } 7724 bdrv_attach_aio_context(bs, new_context); 7725 } 7726 7727 static TransactionActionDrv set_aio_context = { 7728 .commit = bdrv_set_aio_context_commit, 7729 .clean = bdrv_set_aio_context_clean, 7730 }; 7731 7732 /* 7733 * Changes the AioContext used for fd handlers, timers, and BHs by this 7734 * BlockDriverState and all its children and parents. 7735 * 7736 * Must be called from the main AioContext. 7737 * 7738 * The caller must own the AioContext lock for the old AioContext of bs, but it 7739 * must not own the AioContext lock for new_context (unless new_context is the 7740 * same as the current context of bs). 7741 * 7742 * @visited will accumulate all visited BdrvChild objects. The caller is 7743 * responsible for freeing the list afterwards. 7744 */ 7745 static bool bdrv_change_aio_context(BlockDriverState *bs, AioContext *ctx, 7746 GHashTable *visited, Transaction *tran, 7747 Error **errp) 7748 { 7749 BdrvChild *c; 7750 BdrvStateSetAioContext *state; 7751 7752 GLOBAL_STATE_CODE(); 7753 7754 if (bdrv_get_aio_context(bs) == ctx) { 7755 return true; 7756 } 7757 7758 bdrv_graph_rdlock_main_loop(); 7759 QLIST_FOREACH(c, &bs->parents, next_parent) { 7760 if (!bdrv_parent_change_aio_context(c, ctx, visited, tran, errp)) { 7761 bdrv_graph_rdunlock_main_loop(); 7762 return false; 7763 } 7764 } 7765 7766 QLIST_FOREACH(c, &bs->children, next) { 7767 if (!bdrv_child_change_aio_context(c, ctx, visited, tran, errp)) { 7768 bdrv_graph_rdunlock_main_loop(); 7769 return false; 7770 } 7771 } 7772 bdrv_graph_rdunlock_main_loop(); 7773 7774 state = g_new(BdrvStateSetAioContext, 1); 7775 *state = (BdrvStateSetAioContext) { 7776 .new_ctx = ctx, 7777 .bs = bs, 7778 }; 7779 7780 /* Paired with bdrv_drained_end in bdrv_set_aio_context_clean() */ 7781 bdrv_drained_begin(bs); 7782 7783 tran_add(tran, &set_aio_context, state); 7784 7785 return true; 7786 } 7787 7788 /* 7789 * Change bs's and recursively all of its parents' and children's AioContext 7790 * to the given new context, returning an error if that isn't possible. 7791 * 7792 * If ignore_child is not NULL, that child (and its subgraph) will not 7793 * be touched. 7794 * 7795 * This function still requires the caller to take the bs current 7796 * AioContext lock, otherwise draining will fail since AIO_WAIT_WHILE 7797 * assumes the lock is always held if bs is in another AioContext. 7798 * For the same reason, it temporarily also holds the new AioContext, since 7799 * bdrv_drained_end calls BDRV_POLL_WHILE that assumes the lock is taken too. 7800 * Therefore the new AioContext lock must not be taken by the caller. 7801 */ 7802 int bdrv_try_change_aio_context(BlockDriverState *bs, AioContext *ctx, 7803 BdrvChild *ignore_child, Error **errp) 7804 { 7805 Transaction *tran; 7806 GHashTable *visited; 7807 int ret; 7808 AioContext *old_context = bdrv_get_aio_context(bs); 7809 GLOBAL_STATE_CODE(); 7810 7811 /* 7812 * Recursion phase: go through all nodes of the graph. 7813 * Take care of checking that all nodes support changing AioContext 7814 * and drain them, building a linear list of callbacks to run if everything 7815 * is successful (the transaction itself). 7816 */ 7817 tran = tran_new(); 7818 visited = g_hash_table_new(NULL, NULL); 7819 if (ignore_child) { 7820 g_hash_table_add(visited, ignore_child); 7821 } 7822 ret = bdrv_change_aio_context(bs, ctx, visited, tran, errp); 7823 g_hash_table_destroy(visited); 7824 7825 /* 7826 * Linear phase: go through all callbacks collected in the transaction. 7827 * Run all callbacks collected in the recursion to switch all nodes 7828 * AioContext lock (transaction commit), or undo all changes done in the 7829 * recursion (transaction abort). 7830 */ 7831 7832 if (!ret) { 7833 /* Just run clean() callbacks. No AioContext changed. */ 7834 tran_abort(tran); 7835 return -EPERM; 7836 } 7837 7838 /* 7839 * Release old AioContext, it won't be needed anymore, as all 7840 * bdrv_drained_begin() have been called already. 7841 */ 7842 if (qemu_get_aio_context() != old_context) { 7843 aio_context_release(old_context); 7844 } 7845 7846 /* 7847 * Acquire new AioContext since bdrv_drained_end() is going to be called 7848 * after we switched all nodes in the new AioContext, and the function 7849 * assumes that the lock of the bs is always taken. 7850 */ 7851 if (qemu_get_aio_context() != ctx) { 7852 aio_context_acquire(ctx); 7853 } 7854 7855 tran_commit(tran); 7856 7857 if (qemu_get_aio_context() != ctx) { 7858 aio_context_release(ctx); 7859 } 7860 7861 /* Re-acquire the old AioContext, since the caller takes and releases it. */ 7862 if (qemu_get_aio_context() != old_context) { 7863 aio_context_acquire(old_context); 7864 } 7865 7866 return 0; 7867 } 7868 7869 void bdrv_add_aio_context_notifier(BlockDriverState *bs, 7870 void (*attached_aio_context)(AioContext *new_context, void *opaque), 7871 void (*detach_aio_context)(void *opaque), void *opaque) 7872 { 7873 BdrvAioNotifier *ban = g_new(BdrvAioNotifier, 1); 7874 *ban = (BdrvAioNotifier){ 7875 .attached_aio_context = attached_aio_context, 7876 .detach_aio_context = detach_aio_context, 7877 .opaque = opaque 7878 }; 7879 GLOBAL_STATE_CODE(); 7880 7881 QLIST_INSERT_HEAD(&bs->aio_notifiers, ban, list); 7882 } 7883 7884 void bdrv_remove_aio_context_notifier(BlockDriverState *bs, 7885 void (*attached_aio_context)(AioContext *, 7886 void *), 7887 void (*detach_aio_context)(void *), 7888 void *opaque) 7889 { 7890 BdrvAioNotifier *ban, *ban_next; 7891 GLOBAL_STATE_CODE(); 7892 7893 QLIST_FOREACH_SAFE(ban, &bs->aio_notifiers, list, ban_next) { 7894 if (ban->attached_aio_context == attached_aio_context && 7895 ban->detach_aio_context == detach_aio_context && 7896 ban->opaque == opaque && 7897 ban->deleted == false) 7898 { 7899 if (bs->walking_aio_notifiers) { 7900 ban->deleted = true; 7901 } else { 7902 bdrv_do_remove_aio_context_notifier(ban); 7903 } 7904 return; 7905 } 7906 } 7907 7908 abort(); 7909 } 7910 7911 int bdrv_amend_options(BlockDriverState *bs, QemuOpts *opts, 7912 BlockDriverAmendStatusCB *status_cb, void *cb_opaque, 7913 bool force, 7914 Error **errp) 7915 { 7916 GLOBAL_STATE_CODE(); 7917 if (!bs->drv) { 7918 error_setg(errp, "Node is ejected"); 7919 return -ENOMEDIUM; 7920 } 7921 if (!bs->drv->bdrv_amend_options) { 7922 error_setg(errp, "Block driver '%s' does not support option amendment", 7923 bs->drv->format_name); 7924 return -ENOTSUP; 7925 } 7926 return bs->drv->bdrv_amend_options(bs, opts, status_cb, 7927 cb_opaque, force, errp); 7928 } 7929 7930 /* 7931 * This function checks whether the given @to_replace is allowed to be 7932 * replaced by a node that always shows the same data as @bs. This is 7933 * used for example to verify whether the mirror job can replace 7934 * @to_replace by the target mirrored from @bs. 7935 * To be replaceable, @bs and @to_replace may either be guaranteed to 7936 * always show the same data (because they are only connected through 7937 * filters), or some driver may allow replacing one of its children 7938 * because it can guarantee that this child's data is not visible at 7939 * all (for example, for dissenting quorum children that have no other 7940 * parents). 7941 */ 7942 bool bdrv_recurse_can_replace(BlockDriverState *bs, 7943 BlockDriverState *to_replace) 7944 { 7945 BlockDriverState *filtered; 7946 7947 GLOBAL_STATE_CODE(); 7948 7949 if (!bs || !bs->drv) { 7950 return false; 7951 } 7952 7953 if (bs == to_replace) { 7954 return true; 7955 } 7956 7957 /* See what the driver can do */ 7958 if (bs->drv->bdrv_recurse_can_replace) { 7959 return bs->drv->bdrv_recurse_can_replace(bs, to_replace); 7960 } 7961 7962 /* For filters without an own implementation, we can recurse on our own */ 7963 filtered = bdrv_filter_bs(bs); 7964 if (filtered) { 7965 return bdrv_recurse_can_replace(filtered, to_replace); 7966 } 7967 7968 /* Safe default */ 7969 return false; 7970 } 7971 7972 /* 7973 * Check whether the given @node_name can be replaced by a node that 7974 * has the same data as @parent_bs. If so, return @node_name's BDS; 7975 * NULL otherwise. 7976 * 7977 * @node_name must be a (recursive) *child of @parent_bs (or this 7978 * function will return NULL). 7979 * 7980 * The result (whether the node can be replaced or not) is only valid 7981 * for as long as no graph or permission changes occur. 7982 */ 7983 BlockDriverState *check_to_replace_node(BlockDriverState *parent_bs, 7984 const char *node_name, Error **errp) 7985 { 7986 BlockDriverState *to_replace_bs = bdrv_find_node(node_name); 7987 AioContext *aio_context; 7988 7989 GLOBAL_STATE_CODE(); 7990 7991 if (!to_replace_bs) { 7992 error_setg(errp, "Failed to find node with node-name='%s'", node_name); 7993 return NULL; 7994 } 7995 7996 aio_context = bdrv_get_aio_context(to_replace_bs); 7997 aio_context_acquire(aio_context); 7998 7999 if (bdrv_op_is_blocked(to_replace_bs, BLOCK_OP_TYPE_REPLACE, errp)) { 8000 to_replace_bs = NULL; 8001 goto out; 8002 } 8003 8004 /* We don't want arbitrary node of the BDS chain to be replaced only the top 8005 * most non filter in order to prevent data corruption. 8006 * Another benefit is that this tests exclude backing files which are 8007 * blocked by the backing blockers. 8008 */ 8009 if (!bdrv_recurse_can_replace(parent_bs, to_replace_bs)) { 8010 error_setg(errp, "Cannot replace '%s' by a node mirrored from '%s', " 8011 "because it cannot be guaranteed that doing so would not " 8012 "lead to an abrupt change of visible data", 8013 node_name, parent_bs->node_name); 8014 to_replace_bs = NULL; 8015 goto out; 8016 } 8017 8018 out: 8019 aio_context_release(aio_context); 8020 return to_replace_bs; 8021 } 8022 8023 /** 8024 * Iterates through the list of runtime option keys that are said to 8025 * be "strong" for a BDS. An option is called "strong" if it changes 8026 * a BDS's data. For example, the null block driver's "size" and 8027 * "read-zeroes" options are strong, but its "latency-ns" option is 8028 * not. 8029 * 8030 * If a key returned by this function ends with a dot, all options 8031 * starting with that prefix are strong. 8032 */ 8033 static const char *const *strong_options(BlockDriverState *bs, 8034 const char *const *curopt) 8035 { 8036 static const char *const global_options[] = { 8037 "driver", "filename", NULL 8038 }; 8039 8040 if (!curopt) { 8041 return &global_options[0]; 8042 } 8043 8044 curopt++; 8045 if (curopt == &global_options[ARRAY_SIZE(global_options) - 1] && bs->drv) { 8046 curopt = bs->drv->strong_runtime_opts; 8047 } 8048 8049 return (curopt && *curopt) ? curopt : NULL; 8050 } 8051 8052 /** 8053 * Copies all strong runtime options from bs->options to the given 8054 * QDict. The set of strong option keys is determined by invoking 8055 * strong_options(). 8056 * 8057 * Returns true iff any strong option was present in bs->options (and 8058 * thus copied to the target QDict) with the exception of "filename" 8059 * and "driver". The caller is expected to use this value to decide 8060 * whether the existence of strong options prevents the generation of 8061 * a plain filename. 8062 */ 8063 static bool append_strong_runtime_options(QDict *d, BlockDriverState *bs) 8064 { 8065 bool found_any = false; 8066 const char *const *option_name = NULL; 8067 8068 if (!bs->drv) { 8069 return false; 8070 } 8071 8072 while ((option_name = strong_options(bs, option_name))) { 8073 bool option_given = false; 8074 8075 assert(strlen(*option_name) > 0); 8076 if ((*option_name)[strlen(*option_name) - 1] != '.') { 8077 QObject *entry = qdict_get(bs->options, *option_name); 8078 if (!entry) { 8079 continue; 8080 } 8081 8082 qdict_put_obj(d, *option_name, qobject_ref(entry)); 8083 option_given = true; 8084 } else { 8085 const QDictEntry *entry; 8086 for (entry = qdict_first(bs->options); entry; 8087 entry = qdict_next(bs->options, entry)) 8088 { 8089 if (strstart(qdict_entry_key(entry), *option_name, NULL)) { 8090 qdict_put_obj(d, qdict_entry_key(entry), 8091 qobject_ref(qdict_entry_value(entry))); 8092 option_given = true; 8093 } 8094 } 8095 } 8096 8097 /* While "driver" and "filename" need to be included in a JSON filename, 8098 * their existence does not prohibit generation of a plain filename. */ 8099 if (!found_any && option_given && 8100 strcmp(*option_name, "driver") && strcmp(*option_name, "filename")) 8101 { 8102 found_any = true; 8103 } 8104 } 8105 8106 if (!qdict_haskey(d, "driver")) { 8107 /* Drivers created with bdrv_new_open_driver() may not have a 8108 * @driver option. Add it here. */ 8109 qdict_put_str(d, "driver", bs->drv->format_name); 8110 } 8111 8112 return found_any; 8113 } 8114 8115 /* Note: This function may return false positives; it may return true 8116 * even if opening the backing file specified by bs's image header 8117 * would result in exactly bs->backing. */ 8118 static bool bdrv_backing_overridden(BlockDriverState *bs) 8119 { 8120 GLOBAL_STATE_CODE(); 8121 if (bs->backing) { 8122 return strcmp(bs->auto_backing_file, 8123 bs->backing->bs->filename); 8124 } else { 8125 /* No backing BDS, so if the image header reports any backing 8126 * file, it must have been suppressed */ 8127 return bs->auto_backing_file[0] != '\0'; 8128 } 8129 } 8130 8131 /* Updates the following BDS fields: 8132 * - exact_filename: A filename which may be used for opening a block device 8133 * which (mostly) equals the given BDS (even without any 8134 * other options; so reading and writing must return the same 8135 * results, but caching etc. may be different) 8136 * - full_open_options: Options which, when given when opening a block device 8137 * (without a filename), result in a BDS (mostly) 8138 * equalling the given one 8139 * - filename: If exact_filename is set, it is copied here. Otherwise, 8140 * full_open_options is converted to a JSON object, prefixed with 8141 * "json:" (for use through the JSON pseudo protocol) and put here. 8142 */ 8143 void bdrv_refresh_filename(BlockDriverState *bs) 8144 { 8145 BlockDriver *drv = bs->drv; 8146 BdrvChild *child; 8147 BlockDriverState *primary_child_bs; 8148 QDict *opts; 8149 bool backing_overridden; 8150 bool generate_json_filename; /* Whether our default implementation should 8151 fill exact_filename (false) or not (true) */ 8152 8153 GLOBAL_STATE_CODE(); 8154 8155 if (!drv) { 8156 return; 8157 } 8158 8159 /* This BDS's file name may depend on any of its children's file names, so 8160 * refresh those first */ 8161 QLIST_FOREACH(child, &bs->children, next) { 8162 bdrv_refresh_filename(child->bs); 8163 } 8164 8165 if (bs->implicit) { 8166 /* For implicit nodes, just copy everything from the single child */ 8167 child = QLIST_FIRST(&bs->children); 8168 assert(QLIST_NEXT(child, next) == NULL); 8169 8170 pstrcpy(bs->exact_filename, sizeof(bs->exact_filename), 8171 child->bs->exact_filename); 8172 pstrcpy(bs->filename, sizeof(bs->filename), child->bs->filename); 8173 8174 qobject_unref(bs->full_open_options); 8175 bs->full_open_options = qobject_ref(child->bs->full_open_options); 8176 8177 return; 8178 } 8179 8180 backing_overridden = bdrv_backing_overridden(bs); 8181 8182 if (bs->open_flags & BDRV_O_NO_IO) { 8183 /* Without I/O, the backing file does not change anything. 8184 * Therefore, in such a case (primarily qemu-img), we can 8185 * pretend the backing file has not been overridden even if 8186 * it technically has been. */ 8187 backing_overridden = false; 8188 } 8189 8190 /* Gather the options QDict */ 8191 opts = qdict_new(); 8192 generate_json_filename = append_strong_runtime_options(opts, bs); 8193 generate_json_filename |= backing_overridden; 8194 8195 if (drv->bdrv_gather_child_options) { 8196 /* Some block drivers may not want to present all of their children's 8197 * options, or name them differently from BdrvChild.name */ 8198 drv->bdrv_gather_child_options(bs, opts, backing_overridden); 8199 } else { 8200 QLIST_FOREACH(child, &bs->children, next) { 8201 if (child == bs->backing && !backing_overridden) { 8202 /* We can skip the backing BDS if it has not been overridden */ 8203 continue; 8204 } 8205 8206 qdict_put(opts, child->name, 8207 qobject_ref(child->bs->full_open_options)); 8208 } 8209 8210 if (backing_overridden && !bs->backing) { 8211 /* Force no backing file */ 8212 qdict_put_null(opts, "backing"); 8213 } 8214 } 8215 8216 qobject_unref(bs->full_open_options); 8217 bs->full_open_options = opts; 8218 8219 primary_child_bs = bdrv_primary_bs(bs); 8220 8221 if (drv->bdrv_refresh_filename) { 8222 /* Obsolete information is of no use here, so drop the old file name 8223 * information before refreshing it */ 8224 bs->exact_filename[0] = '\0'; 8225 8226 drv->bdrv_refresh_filename(bs); 8227 } else if (primary_child_bs) { 8228 /* 8229 * Try to reconstruct valid information from the underlying 8230 * file -- this only works for format nodes (filter nodes 8231 * cannot be probed and as such must be selected by the user 8232 * either through an options dict, or through a special 8233 * filename which the filter driver must construct in its 8234 * .bdrv_refresh_filename() implementation). 8235 */ 8236 8237 bs->exact_filename[0] = '\0'; 8238 8239 /* 8240 * We can use the underlying file's filename if: 8241 * - it has a filename, 8242 * - the current BDS is not a filter, 8243 * - the file is a protocol BDS, and 8244 * - opening that file (as this BDS's format) will automatically create 8245 * the BDS tree we have right now, that is: 8246 * - the user did not significantly change this BDS's behavior with 8247 * some explicit (strong) options 8248 * - no non-file child of this BDS has been overridden by the user 8249 * Both of these conditions are represented by generate_json_filename. 8250 */ 8251 if (primary_child_bs->exact_filename[0] && 8252 primary_child_bs->drv->bdrv_file_open && 8253 !drv->is_filter && !generate_json_filename) 8254 { 8255 strcpy(bs->exact_filename, primary_child_bs->exact_filename); 8256 } 8257 } 8258 8259 if (bs->exact_filename[0]) { 8260 pstrcpy(bs->filename, sizeof(bs->filename), bs->exact_filename); 8261 } else { 8262 GString *json = qobject_to_json(QOBJECT(bs->full_open_options)); 8263 if (snprintf(bs->filename, sizeof(bs->filename), "json:%s", 8264 json->str) >= sizeof(bs->filename)) { 8265 /* Give user a hint if we truncated things. */ 8266 strcpy(bs->filename + sizeof(bs->filename) - 4, "..."); 8267 } 8268 g_string_free(json, true); 8269 } 8270 } 8271 8272 char *bdrv_dirname(BlockDriverState *bs, Error **errp) 8273 { 8274 BlockDriver *drv = bs->drv; 8275 BlockDriverState *child_bs; 8276 8277 GLOBAL_STATE_CODE(); 8278 8279 if (!drv) { 8280 error_setg(errp, "Node '%s' is ejected", bs->node_name); 8281 return NULL; 8282 } 8283 8284 if (drv->bdrv_dirname) { 8285 return drv->bdrv_dirname(bs, errp); 8286 } 8287 8288 child_bs = bdrv_primary_bs(bs); 8289 if (child_bs) { 8290 return bdrv_dirname(child_bs, errp); 8291 } 8292 8293 bdrv_refresh_filename(bs); 8294 if (bs->exact_filename[0] != '\0') { 8295 return path_combine(bs->exact_filename, ""); 8296 } 8297 8298 error_setg(errp, "Cannot generate a base directory for %s nodes", 8299 drv->format_name); 8300 return NULL; 8301 } 8302 8303 /* 8304 * Hot add/remove a BDS's child. So the user can take a child offline when 8305 * it is broken and take a new child online 8306 */ 8307 void bdrv_add_child(BlockDriverState *parent_bs, BlockDriverState *child_bs, 8308 Error **errp) 8309 { 8310 GLOBAL_STATE_CODE(); 8311 if (!parent_bs->drv || !parent_bs->drv->bdrv_add_child) { 8312 error_setg(errp, "The node %s does not support adding a child", 8313 bdrv_get_device_or_node_name(parent_bs)); 8314 return; 8315 } 8316 8317 /* 8318 * Non-zoned block drivers do not follow zoned storage constraints 8319 * (i.e. sequential writes to zones). Refuse mixing zoned and non-zoned 8320 * drivers in a graph. 8321 */ 8322 if (!parent_bs->drv->supports_zoned_children && 8323 child_bs->bl.zoned == BLK_Z_HM) { 8324 /* 8325 * The host-aware model allows zoned storage constraints and random 8326 * write. Allow mixing host-aware and non-zoned drivers. Using 8327 * host-aware device as a regular device. 8328 */ 8329 error_setg(errp, "Cannot add a %s child to a %s parent", 8330 child_bs->bl.zoned == BLK_Z_HM ? "zoned" : "non-zoned", 8331 parent_bs->drv->supports_zoned_children ? 8332 "support zoned children" : "not support zoned children"); 8333 return; 8334 } 8335 8336 if (!QLIST_EMPTY(&child_bs->parents)) { 8337 error_setg(errp, "The node %s already has a parent", 8338 child_bs->node_name); 8339 return; 8340 } 8341 8342 parent_bs->drv->bdrv_add_child(parent_bs, child_bs, errp); 8343 } 8344 8345 void bdrv_del_child(BlockDriverState *parent_bs, BdrvChild *child, Error **errp) 8346 { 8347 BdrvChild *tmp; 8348 8349 GLOBAL_STATE_CODE(); 8350 if (!parent_bs->drv || !parent_bs->drv->bdrv_del_child) { 8351 error_setg(errp, "The node %s does not support removing a child", 8352 bdrv_get_device_or_node_name(parent_bs)); 8353 return; 8354 } 8355 8356 QLIST_FOREACH(tmp, &parent_bs->children, next) { 8357 if (tmp == child) { 8358 break; 8359 } 8360 } 8361 8362 if (!tmp) { 8363 error_setg(errp, "The node %s does not have a child named %s", 8364 bdrv_get_device_or_node_name(parent_bs), 8365 bdrv_get_device_or_node_name(child->bs)); 8366 return; 8367 } 8368 8369 parent_bs->drv->bdrv_del_child(parent_bs, child, errp); 8370 } 8371 8372 int bdrv_make_empty(BdrvChild *c, Error **errp) 8373 { 8374 BlockDriver *drv = c->bs->drv; 8375 int ret; 8376 8377 GLOBAL_STATE_CODE(); 8378 assert(c->perm & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED)); 8379 8380 if (!drv->bdrv_make_empty) { 8381 error_setg(errp, "%s does not support emptying nodes", 8382 drv->format_name); 8383 return -ENOTSUP; 8384 } 8385 8386 ret = drv->bdrv_make_empty(c->bs); 8387 if (ret < 0) { 8388 error_setg_errno(errp, -ret, "Failed to empty %s", 8389 c->bs->filename); 8390 return ret; 8391 } 8392 8393 return 0; 8394 } 8395 8396 /* 8397 * Return the child that @bs acts as an overlay for, and from which data may be 8398 * copied in COW or COR operations. Usually this is the backing file. 8399 */ 8400 BdrvChild *bdrv_cow_child(BlockDriverState *bs) 8401 { 8402 IO_CODE(); 8403 8404 if (!bs || !bs->drv) { 8405 return NULL; 8406 } 8407 8408 if (bs->drv->is_filter) { 8409 return NULL; 8410 } 8411 8412 if (!bs->backing) { 8413 return NULL; 8414 } 8415 8416 assert(bs->backing->role & BDRV_CHILD_COW); 8417 return bs->backing; 8418 } 8419 8420 /* 8421 * If @bs acts as a filter for exactly one of its children, return 8422 * that child. 8423 */ 8424 BdrvChild *bdrv_filter_child(BlockDriverState *bs) 8425 { 8426 BdrvChild *c; 8427 IO_CODE(); 8428 8429 if (!bs || !bs->drv) { 8430 return NULL; 8431 } 8432 8433 if (!bs->drv->is_filter) { 8434 return NULL; 8435 } 8436 8437 /* Only one of @backing or @file may be used */ 8438 assert(!(bs->backing && bs->file)); 8439 8440 c = bs->backing ?: bs->file; 8441 if (!c) { 8442 return NULL; 8443 } 8444 8445 assert(c->role & BDRV_CHILD_FILTERED); 8446 return c; 8447 } 8448 8449 /* 8450 * Return either the result of bdrv_cow_child() or bdrv_filter_child(), 8451 * whichever is non-NULL. 8452 * 8453 * Return NULL if both are NULL. 8454 */ 8455 BdrvChild *bdrv_filter_or_cow_child(BlockDriverState *bs) 8456 { 8457 BdrvChild *cow_child = bdrv_cow_child(bs); 8458 BdrvChild *filter_child = bdrv_filter_child(bs); 8459 IO_CODE(); 8460 8461 /* Filter nodes cannot have COW backing files */ 8462 assert(!(cow_child && filter_child)); 8463 8464 return cow_child ?: filter_child; 8465 } 8466 8467 /* 8468 * Return the primary child of this node: For filters, that is the 8469 * filtered child. For other nodes, that is usually the child storing 8470 * metadata. 8471 * (A generally more helpful description is that this is (usually) the 8472 * child that has the same filename as @bs.) 8473 * 8474 * Drivers do not necessarily have a primary child; for example quorum 8475 * does not. 8476 */ 8477 BdrvChild *bdrv_primary_child(BlockDriverState *bs) 8478 { 8479 BdrvChild *c, *found = NULL; 8480 IO_CODE(); 8481 8482 QLIST_FOREACH(c, &bs->children, next) { 8483 if (c->role & BDRV_CHILD_PRIMARY) { 8484 assert(!found); 8485 found = c; 8486 } 8487 } 8488 8489 return found; 8490 } 8491 8492 static BlockDriverState * GRAPH_RDLOCK 8493 bdrv_do_skip_filters(BlockDriverState *bs, bool stop_on_explicit_filter) 8494 { 8495 BdrvChild *c; 8496 8497 if (!bs) { 8498 return NULL; 8499 } 8500 8501 while (!(stop_on_explicit_filter && !bs->implicit)) { 8502 c = bdrv_filter_child(bs); 8503 if (!c) { 8504 /* 8505 * A filter that is embedded in a working block graph must 8506 * have a child. Assert this here so this function does 8507 * not return a filter node that is not expected by the 8508 * caller. 8509 */ 8510 assert(!bs->drv || !bs->drv->is_filter); 8511 break; 8512 } 8513 bs = c->bs; 8514 } 8515 /* 8516 * Note that this treats nodes with bs->drv == NULL as not being 8517 * filters (bs->drv == NULL should be replaced by something else 8518 * anyway). 8519 * The advantage of this behavior is that this function will thus 8520 * always return a non-NULL value (given a non-NULL @bs). 8521 */ 8522 8523 return bs; 8524 } 8525 8526 /* 8527 * Return the first BDS that has not been added implicitly or that 8528 * does not have a filtered child down the chain starting from @bs 8529 * (including @bs itself). 8530 */ 8531 BlockDriverState *bdrv_skip_implicit_filters(BlockDriverState *bs) 8532 { 8533 GLOBAL_STATE_CODE(); 8534 return bdrv_do_skip_filters(bs, true); 8535 } 8536 8537 /* 8538 * Return the first BDS that does not have a filtered child down the 8539 * chain starting from @bs (including @bs itself). 8540 */ 8541 BlockDriverState *bdrv_skip_filters(BlockDriverState *bs) 8542 { 8543 IO_CODE(); 8544 return bdrv_do_skip_filters(bs, false); 8545 } 8546 8547 /* 8548 * For a backing chain, return the first non-filter backing image of 8549 * the first non-filter image. 8550 */ 8551 BlockDriverState *bdrv_backing_chain_next(BlockDriverState *bs) 8552 { 8553 IO_CODE(); 8554 return bdrv_skip_filters(bdrv_cow_bs(bdrv_skip_filters(bs))); 8555 } 8556 8557 /** 8558 * Check whether [offset, offset + bytes) overlaps with the cached 8559 * block-status data region. 8560 * 8561 * If so, and @pnum is not NULL, set *pnum to `bsc.data_end - offset`, 8562 * which is what bdrv_bsc_is_data()'s interface needs. 8563 * Otherwise, *pnum is not touched. 8564 */ 8565 static bool bdrv_bsc_range_overlaps_locked(BlockDriverState *bs, 8566 int64_t offset, int64_t bytes, 8567 int64_t *pnum) 8568 { 8569 BdrvBlockStatusCache *bsc = qatomic_rcu_read(&bs->block_status_cache); 8570 bool overlaps; 8571 8572 overlaps = 8573 qatomic_read(&bsc->valid) && 8574 ranges_overlap(offset, bytes, bsc->data_start, 8575 bsc->data_end - bsc->data_start); 8576 8577 if (overlaps && pnum) { 8578 *pnum = bsc->data_end - offset; 8579 } 8580 8581 return overlaps; 8582 } 8583 8584 /** 8585 * See block_int.h for this function's documentation. 8586 */ 8587 bool bdrv_bsc_is_data(BlockDriverState *bs, int64_t offset, int64_t *pnum) 8588 { 8589 IO_CODE(); 8590 RCU_READ_LOCK_GUARD(); 8591 return bdrv_bsc_range_overlaps_locked(bs, offset, 1, pnum); 8592 } 8593 8594 /** 8595 * See block_int.h for this function's documentation. 8596 */ 8597 void bdrv_bsc_invalidate_range(BlockDriverState *bs, 8598 int64_t offset, int64_t bytes) 8599 { 8600 IO_CODE(); 8601 RCU_READ_LOCK_GUARD(); 8602 8603 if (bdrv_bsc_range_overlaps_locked(bs, offset, bytes, NULL)) { 8604 qatomic_set(&bs->block_status_cache->valid, false); 8605 } 8606 } 8607 8608 /** 8609 * See block_int.h for this function's documentation. 8610 */ 8611 void bdrv_bsc_fill(BlockDriverState *bs, int64_t offset, int64_t bytes) 8612 { 8613 BdrvBlockStatusCache *new_bsc = g_new(BdrvBlockStatusCache, 1); 8614 BdrvBlockStatusCache *old_bsc; 8615 IO_CODE(); 8616 8617 *new_bsc = (BdrvBlockStatusCache) { 8618 .valid = true, 8619 .data_start = offset, 8620 .data_end = offset + bytes, 8621 }; 8622 8623 QEMU_LOCK_GUARD(&bs->bsc_modify_lock); 8624 8625 old_bsc = qatomic_rcu_read(&bs->block_status_cache); 8626 qatomic_rcu_set(&bs->block_status_cache, new_bsc); 8627 if (old_bsc) { 8628 g_free_rcu(old_bsc, rcu); 8629 } 8630 } 8631