Merge remote-tracking branch 'remotes/kraxel/tags/seabios-20220118-pull-request'...
[qemu.git] / block.c
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/fuse.h"
31 #include "block/nbd.h"
32 #include "block/qdict.h"
33 #include "qemu/error-report.h"
34 #include "block/module_block.h"
35 #include "qemu/main-loop.h"
36 #include "qemu/module.h"
37 #include "qapi/error.h"
38 #include "qapi/qmp/qdict.h"
39 #include "qapi/qmp/qjson.h"
40 #include "qapi/qmp/qnull.h"
41 #include "qapi/qmp/qstring.h"
42 #include "qapi/qobject-output-visitor.h"
43 #include "qapi/qapi-visit-block-core.h"
44 #include "sysemu/block-backend.h"
45 #include "qemu/notify.h"
46 #include "qemu/option.h"
47 #include "qemu/coroutine.h"
48 #include "block/qapi.h"
49 #include "qemu/timer.h"
50 #include "qemu/cutils.h"
51 #include "qemu/id.h"
52 #include "qemu/range.h"
53 #include "qemu/rcu.h"
54 #include "block/coroutines.h"
55
56 #ifdef CONFIG_BSD
57 #include <sys/ioctl.h>
58 #include <sys/queue.h>
59 #if defined(HAVE_SYS_DISK_H)
60 #include <sys/disk.h>
61 #endif
62 #endif
63
64 #ifdef _WIN32
65 #include <windows.h>
66 #endif
67
68 #define NOT_DONE 0x7fffffff /* used while emulated sync operation in progress */
69
70 static QTAILQ_HEAD(, BlockDriverState) graph_bdrv_states =
71 QTAILQ_HEAD_INITIALIZER(graph_bdrv_states);
72
73 static QTAILQ_HEAD(, BlockDriverState) all_bdrv_states =
74 QTAILQ_HEAD_INITIALIZER(all_bdrv_states);
75
76 static QLIST_HEAD(, BlockDriver) bdrv_drivers =
77 QLIST_HEAD_INITIALIZER(bdrv_drivers);
78
79 static BlockDriverState *bdrv_open_inherit(const char *filename,
80 const char *reference,
81 QDict *options, int flags,
82 BlockDriverState *parent,
83 const BdrvChildClass *child_class,
84 BdrvChildRole child_role,
85 Error **errp);
86
87 static void bdrv_replace_child_noperm(BdrvChild *child,
88 BlockDriverState *new_bs);
89 static void bdrv_remove_file_or_backing_child(BlockDriverState *bs,
90 BdrvChild *child,
91 Transaction *tran);
92 static void bdrv_remove_filter_or_cow_child(BlockDriverState *bs,
93 Transaction *tran);
94
95 static int bdrv_reopen_prepare(BDRVReopenState *reopen_state,
96 BlockReopenQueue *queue,
97 Transaction *change_child_tran, Error **errp);
98 static void bdrv_reopen_commit(BDRVReopenState *reopen_state);
99 static void bdrv_reopen_abort(BDRVReopenState *reopen_state);
100
101 /* If non-zero, use only whitelisted block drivers */
102 static int use_bdrv_whitelist;
103
104 #ifdef _WIN32
105 static int is_windows_drive_prefix(const char *filename)
106 {
107 return (((filename[0] >= 'a' && filename[0] <= 'z') ||
108 (filename[0] >= 'A' && filename[0] <= 'Z')) &&
109 filename[1] == ':');
110 }
111
112 int is_windows_drive(const char *filename)
113 {
114 if (is_windows_drive_prefix(filename) &&
115 filename[2] == '\0')
116 return 1;
117 if (strstart(filename, "\\\\.\\", NULL) ||
118 strstart(filename, "//./", NULL))
119 return 1;
120 return 0;
121 }
122 #endif
123
124 size_t bdrv_opt_mem_align(BlockDriverState *bs)
125 {
126 if (!bs || !bs->drv) {
127 /* page size or 4k (hdd sector size) should be on the safe side */
128 return MAX(4096, qemu_real_host_page_size);
129 }
130
131 return bs->bl.opt_mem_alignment;
132 }
133
134 size_t bdrv_min_mem_align(BlockDriverState *bs)
135 {
136 if (!bs || !bs->drv) {
137 /* page size or 4k (hdd sector size) should be on the safe side */
138 return MAX(4096, qemu_real_host_page_size);
139 }
140
141 return bs->bl.min_mem_alignment;
142 }
143
144 /* check if the path starts with "<protocol>:" */
145 int path_has_protocol(const char *path)
146 {
147 const char *p;
148
149 #ifdef _WIN32
150 if (is_windows_drive(path) ||
151 is_windows_drive_prefix(path)) {
152 return 0;
153 }
154 p = path + strcspn(path, ":/\\");
155 #else
156 p = path + strcspn(path, ":/");
157 #endif
158
159 return *p == ':';
160 }
161
162 int path_is_absolute(const char *path)
163 {
164 #ifdef _WIN32
165 /* specific case for names like: "\\.\d:" */
166 if (is_windows_drive(path) || is_windows_drive_prefix(path)) {
167 return 1;
168 }
169 return (*path == '/' || *path == '\\');
170 #else
171 return (*path == '/');
172 #endif
173 }
174
175 /* if filename is absolute, just return its duplicate. Otherwise, build a
176 path to it by considering it is relative to base_path. URL are
177 supported. */
178 char *path_combine(const char *base_path, const char *filename)
179 {
180 const char *protocol_stripped = NULL;
181 const char *p, *p1;
182 char *result;
183 int len;
184
185 if (path_is_absolute(filename)) {
186 return g_strdup(filename);
187 }
188
189 if (path_has_protocol(base_path)) {
190 protocol_stripped = strchr(base_path, ':');
191 if (protocol_stripped) {
192 protocol_stripped++;
193 }
194 }
195 p = protocol_stripped ?: base_path;
196
197 p1 = strrchr(base_path, '/');
198 #ifdef _WIN32
199 {
200 const char *p2;
201 p2 = strrchr(base_path, '\\');
202 if (!p1 || p2 > p1) {
203 p1 = p2;
204 }
205 }
206 #endif
207 if (p1) {
208 p1++;
209 } else {
210 p1 = base_path;
211 }
212 if (p1 > p) {
213 p = p1;
214 }
215 len = p - base_path;
216
217 result = g_malloc(len + strlen(filename) + 1);
218 memcpy(result, base_path, len);
219 strcpy(result + len, filename);
220
221 return result;
222 }
223
224 /*
225 * Helper function for bdrv_parse_filename() implementations to remove optional
226 * protocol prefixes (especially "file:") from a filename and for putting the
227 * stripped filename into the options QDict if there is such a prefix.
228 */
229 void bdrv_parse_filename_strip_prefix(const char *filename, const char *prefix,
230 QDict *options)
231 {
232 if (strstart(filename, prefix, &filename)) {
233 /* Stripping the explicit protocol prefix may result in a protocol
234 * prefix being (wrongly) detected (if the filename contains a colon) */
235 if (path_has_protocol(filename)) {
236 GString *fat_filename;
237
238 /* This means there is some colon before the first slash; therefore,
239 * this cannot be an absolute path */
240 assert(!path_is_absolute(filename));
241
242 /* And we can thus fix the protocol detection issue by prefixing it
243 * by "./" */
244 fat_filename = g_string_new("./");
245 g_string_append(fat_filename, filename);
246
247 assert(!path_has_protocol(fat_filename->str));
248
249 qdict_put(options, "filename",
250 qstring_from_gstring(fat_filename));
251 } else {
252 /* If no protocol prefix was detected, we can use the shortened
253 * filename as-is */
254 qdict_put_str(options, "filename", filename);
255 }
256 }
257 }
258
259
260 /* Returns whether the image file is opened as read-only. Note that this can
261 * return false and writing to the image file is still not possible because the
262 * image is inactivated. */
263 bool bdrv_is_read_only(BlockDriverState *bs)
264 {
265 return !(bs->open_flags & BDRV_O_RDWR);
266 }
267
268 int bdrv_can_set_read_only(BlockDriverState *bs, bool read_only,
269 bool ignore_allow_rdw, Error **errp)
270 {
271 /* Do not set read_only if copy_on_read is enabled */
272 if (bs->copy_on_read && read_only) {
273 error_setg(errp, "Can't set node '%s' to r/o with copy-on-read enabled",
274 bdrv_get_device_or_node_name(bs));
275 return -EINVAL;
276 }
277
278 /* Do not clear read_only if it is prohibited */
279 if (!read_only && !(bs->open_flags & BDRV_O_ALLOW_RDWR) &&
280 !ignore_allow_rdw)
281 {
282 error_setg(errp, "Node '%s' is read only",
283 bdrv_get_device_or_node_name(bs));
284 return -EPERM;
285 }
286
287 return 0;
288 }
289
290 /*
291 * Called by a driver that can only provide a read-only image.
292 *
293 * Returns 0 if the node is already read-only or it could switch the node to
294 * read-only because BDRV_O_AUTO_RDONLY is set.
295 *
296 * Returns -EACCES if the node is read-write and BDRV_O_AUTO_RDONLY is not set
297 * or bdrv_can_set_read_only() forbids making the node read-only. If @errmsg
298 * is not NULL, it is used as the error message for the Error object.
299 */
300 int bdrv_apply_auto_read_only(BlockDriverState *bs, const char *errmsg,
301 Error **errp)
302 {
303 int ret = 0;
304
305 if (!(bs->open_flags & BDRV_O_RDWR)) {
306 return 0;
307 }
308 if (!(bs->open_flags & BDRV_O_AUTO_RDONLY)) {
309 goto fail;
310 }
311
312 ret = bdrv_can_set_read_only(bs, true, false, NULL);
313 if (ret < 0) {
314 goto fail;
315 }
316
317 bs->open_flags &= ~BDRV_O_RDWR;
318
319 return 0;
320
321 fail:
322 error_setg(errp, "%s", errmsg ?: "Image is read-only");
323 return -EACCES;
324 }
325
326 /*
327 * If @backing is empty, this function returns NULL without setting
328 * @errp. In all other cases, NULL will only be returned with @errp
329 * set.
330 *
331 * Therefore, a return value of NULL without @errp set means that
332 * there is no backing file; if @errp is set, there is one but its
333 * absolute filename cannot be generated.
334 */
335 char *bdrv_get_full_backing_filename_from_filename(const char *backed,
336 const char *backing,
337 Error **errp)
338 {
339 if (backing[0] == '\0') {
340 return NULL;
341 } else if (path_has_protocol(backing) || path_is_absolute(backing)) {
342 return g_strdup(backing);
343 } else if (backed[0] == '\0' || strstart(backed, "json:", NULL)) {
344 error_setg(errp, "Cannot use relative backing file names for '%s'",
345 backed);
346 return NULL;
347 } else {
348 return path_combine(backed, backing);
349 }
350 }
351
352 /*
353 * If @filename is empty or NULL, this function returns NULL without
354 * setting @errp. In all other cases, NULL will only be returned with
355 * @errp set.
356 */
357 static char *bdrv_make_absolute_filename(BlockDriverState *relative_to,
358 const char *filename, Error **errp)
359 {
360 char *dir, *full_name;
361
362 if (!filename || filename[0] == '\0') {
363 return NULL;
364 } else if (path_has_protocol(filename) || path_is_absolute(filename)) {
365 return g_strdup(filename);
366 }
367
368 dir = bdrv_dirname(relative_to, errp);
369 if (!dir) {
370 return NULL;
371 }
372
373 full_name = g_strconcat(dir, filename, NULL);
374 g_free(dir);
375 return full_name;
376 }
377
378 char *bdrv_get_full_backing_filename(BlockDriverState *bs, Error **errp)
379 {
380 return bdrv_make_absolute_filename(bs, bs->backing_file, errp);
381 }
382
383 void bdrv_register(BlockDriver *bdrv)
384 {
385 assert(bdrv->format_name);
386 QLIST_INSERT_HEAD(&bdrv_drivers, bdrv, list);
387 }
388
389 BlockDriverState *bdrv_new(void)
390 {
391 BlockDriverState *bs;
392 int i;
393
394 bs = g_new0(BlockDriverState, 1);
395 QLIST_INIT(&bs->dirty_bitmaps);
396 for (i = 0; i < BLOCK_OP_TYPE_MAX; i++) {
397 QLIST_INIT(&bs->op_blockers[i]);
398 }
399 qemu_co_mutex_init(&bs->reqs_lock);
400 qemu_mutex_init(&bs->dirty_bitmap_mutex);
401 bs->refcnt = 1;
402 bs->aio_context = qemu_get_aio_context();
403
404 qemu_co_queue_init(&bs->flush_queue);
405
406 qemu_co_mutex_init(&bs->bsc_modify_lock);
407 bs->block_status_cache = g_new0(BdrvBlockStatusCache, 1);
408
409 for (i = 0; i < bdrv_drain_all_count; i++) {
410 bdrv_drained_begin(bs);
411 }
412
413 QTAILQ_INSERT_TAIL(&all_bdrv_states, bs, bs_list);
414
415 return bs;
416 }
417
418 static BlockDriver *bdrv_do_find_format(const char *format_name)
419 {
420 BlockDriver *drv1;
421
422 QLIST_FOREACH(drv1, &bdrv_drivers, list) {
423 if (!strcmp(drv1->format_name, format_name)) {
424 return drv1;
425 }
426 }
427
428 return NULL;
429 }
430
431 BlockDriver *bdrv_find_format(const char *format_name)
432 {
433 BlockDriver *drv1;
434 int i;
435
436 drv1 = bdrv_do_find_format(format_name);
437 if (drv1) {
438 return drv1;
439 }
440
441 /* The driver isn't registered, maybe we need to load a module */
442 for (i = 0; i < (int)ARRAY_SIZE(block_driver_modules); ++i) {
443 if (!strcmp(block_driver_modules[i].format_name, format_name)) {
444 block_module_load_one(block_driver_modules[i].library_name);
445 break;
446 }
447 }
448
449 return bdrv_do_find_format(format_name);
450 }
451
452 static int bdrv_format_is_whitelisted(const char *format_name, bool read_only)
453 {
454 static const char *whitelist_rw[] = {
455 CONFIG_BDRV_RW_WHITELIST
456 NULL
457 };
458 static const char *whitelist_ro[] = {
459 CONFIG_BDRV_RO_WHITELIST
460 NULL
461 };
462 const char **p;
463
464 if (!whitelist_rw[0] && !whitelist_ro[0]) {
465 return 1; /* no whitelist, anything goes */
466 }
467
468 for (p = whitelist_rw; *p; p++) {
469 if (!strcmp(format_name, *p)) {
470 return 1;
471 }
472 }
473 if (read_only) {
474 for (p = whitelist_ro; *p; p++) {
475 if (!strcmp(format_name, *p)) {
476 return 1;
477 }
478 }
479 }
480 return 0;
481 }
482
483 int bdrv_is_whitelisted(BlockDriver *drv, bool read_only)
484 {
485 return bdrv_format_is_whitelisted(drv->format_name, read_only);
486 }
487
488 bool bdrv_uses_whitelist(void)
489 {
490 return use_bdrv_whitelist;
491 }
492
493 typedef struct CreateCo {
494 BlockDriver *drv;
495 char *filename;
496 QemuOpts *opts;
497 int ret;
498 Error *err;
499 } CreateCo;
500
501 static void coroutine_fn bdrv_create_co_entry(void *opaque)
502 {
503 Error *local_err = NULL;
504 int ret;
505
506 CreateCo *cco = opaque;
507 assert(cco->drv);
508
509 ret = cco->drv->bdrv_co_create_opts(cco->drv,
510 cco->filename, cco->opts, &local_err);
511 error_propagate(&cco->err, local_err);
512 cco->ret = ret;
513 }
514
515 int bdrv_create(BlockDriver *drv, const char* filename,
516 QemuOpts *opts, Error **errp)
517 {
518 int ret;
519
520 Coroutine *co;
521 CreateCo cco = {
522 .drv = drv,
523 .filename = g_strdup(filename),
524 .opts = opts,
525 .ret = NOT_DONE,
526 .err = NULL,
527 };
528
529 if (!drv->bdrv_co_create_opts) {
530 error_setg(errp, "Driver '%s' does not support image creation", drv->format_name);
531 ret = -ENOTSUP;
532 goto out;
533 }
534
535 if (qemu_in_coroutine()) {
536 /* Fast-path if already in coroutine context */
537 bdrv_create_co_entry(&cco);
538 } else {
539 co = qemu_coroutine_create(bdrv_create_co_entry, &cco);
540 qemu_coroutine_enter(co);
541 while (cco.ret == NOT_DONE) {
542 aio_poll(qemu_get_aio_context(), true);
543 }
544 }
545
546 ret = cco.ret;
547 if (ret < 0) {
548 if (cco.err) {
549 error_propagate(errp, cco.err);
550 } else {
551 error_setg_errno(errp, -ret, "Could not create image");
552 }
553 }
554
555 out:
556 g_free(cco.filename);
557 return ret;
558 }
559
560 /**
561 * Helper function for bdrv_create_file_fallback(): Resize @blk to at
562 * least the given @minimum_size.
563 *
564 * On success, return @blk's actual length.
565 * Otherwise, return -errno.
566 */
567 static int64_t create_file_fallback_truncate(BlockBackend *blk,
568 int64_t minimum_size, Error **errp)
569 {
570 Error *local_err = NULL;
571 int64_t size;
572 int ret;
573
574 ret = blk_truncate(blk, minimum_size, false, PREALLOC_MODE_OFF, 0,
575 &local_err);
576 if (ret < 0 && ret != -ENOTSUP) {
577 error_propagate(errp, local_err);
578 return ret;
579 }
580
581 size = blk_getlength(blk);
582 if (size < 0) {
583 error_free(local_err);
584 error_setg_errno(errp, -size,
585 "Failed to inquire the new image file's length");
586 return size;
587 }
588
589 if (size < minimum_size) {
590 /* Need to grow the image, but we failed to do that */
591 error_propagate(errp, local_err);
592 return -ENOTSUP;
593 }
594
595 error_free(local_err);
596 local_err = NULL;
597
598 return size;
599 }
600
601 /**
602 * Helper function for bdrv_create_file_fallback(): Zero the first
603 * sector to remove any potentially pre-existing image header.
604 */
605 static int create_file_fallback_zero_first_sector(BlockBackend *blk,
606 int64_t current_size,
607 Error **errp)
608 {
609 int64_t bytes_to_clear;
610 int ret;
611
612 bytes_to_clear = MIN(current_size, BDRV_SECTOR_SIZE);
613 if (bytes_to_clear) {
614 ret = blk_pwrite_zeroes(blk, 0, bytes_to_clear, BDRV_REQ_MAY_UNMAP);
615 if (ret < 0) {
616 error_setg_errno(errp, -ret,
617 "Failed to clear the new image's first sector");
618 return ret;
619 }
620 }
621
622 return 0;
623 }
624
625 /**
626 * Simple implementation of bdrv_co_create_opts for protocol drivers
627 * which only support creation via opening a file
628 * (usually existing raw storage device)
629 */
630 int coroutine_fn bdrv_co_create_opts_simple(BlockDriver *drv,
631 const char *filename,
632 QemuOpts *opts,
633 Error **errp)
634 {
635 BlockBackend *blk;
636 QDict *options;
637 int64_t size = 0;
638 char *buf = NULL;
639 PreallocMode prealloc;
640 Error *local_err = NULL;
641 int ret;
642
643 size = qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0);
644 buf = qemu_opt_get_del(opts, BLOCK_OPT_PREALLOC);
645 prealloc = qapi_enum_parse(&PreallocMode_lookup, buf,
646 PREALLOC_MODE_OFF, &local_err);
647 g_free(buf);
648 if (local_err) {
649 error_propagate(errp, local_err);
650 return -EINVAL;
651 }
652
653 if (prealloc != PREALLOC_MODE_OFF) {
654 error_setg(errp, "Unsupported preallocation mode '%s'",
655 PreallocMode_str(prealloc));
656 return -ENOTSUP;
657 }
658
659 options = qdict_new();
660 qdict_put_str(options, "driver", drv->format_name);
661
662 blk = blk_new_open(filename, NULL, options,
663 BDRV_O_RDWR | BDRV_O_RESIZE, errp);
664 if (!blk) {
665 error_prepend(errp, "Protocol driver '%s' does not support image "
666 "creation, and opening the image failed: ",
667 drv->format_name);
668 return -EINVAL;
669 }
670
671 size = create_file_fallback_truncate(blk, size, errp);
672 if (size < 0) {
673 ret = size;
674 goto out;
675 }
676
677 ret = create_file_fallback_zero_first_sector(blk, size, errp);
678 if (ret < 0) {
679 goto out;
680 }
681
682 ret = 0;
683 out:
684 blk_unref(blk);
685 return ret;
686 }
687
688 int bdrv_create_file(const char *filename, QemuOpts *opts, Error **errp)
689 {
690 QemuOpts *protocol_opts;
691 BlockDriver *drv;
692 QDict *qdict;
693 int ret;
694
695 drv = bdrv_find_protocol(filename, true, errp);
696 if (drv == NULL) {
697 return -ENOENT;
698 }
699
700 if (!drv->create_opts) {
701 error_setg(errp, "Driver '%s' does not support image creation",
702 drv->format_name);
703 return -ENOTSUP;
704 }
705
706 /*
707 * 'opts' contains a QemuOptsList with a combination of format and protocol
708 * default values.
709 *
710 * The format properly removes its options, but the default values remain
711 * in 'opts->list'. So if the protocol has options with the same name
712 * (e.g. rbd has 'cluster_size' as qcow2), it will see the default values
713 * of the format, since for overlapping options, the format wins.
714 *
715 * To avoid this issue, lets convert QemuOpts to QDict, in this way we take
716 * only the set options, and then convert it back to QemuOpts, using the
717 * create_opts of the protocol. So the new QemuOpts, will contain only the
718 * protocol defaults.
719 */
720 qdict = qemu_opts_to_qdict(opts, NULL);
721 protocol_opts = qemu_opts_from_qdict(drv->create_opts, qdict, errp);
722 if (protocol_opts == NULL) {
723 ret = -EINVAL;
724 goto out;
725 }
726
727 ret = bdrv_create(drv, filename, protocol_opts, errp);
728 out:
729 qemu_opts_del(protocol_opts);
730 qobject_unref(qdict);
731 return ret;
732 }
733
734 int coroutine_fn bdrv_co_delete_file(BlockDriverState *bs, Error **errp)
735 {
736 Error *local_err = NULL;
737 int ret;
738
739 assert(bs != NULL);
740
741 if (!bs->drv) {
742 error_setg(errp, "Block node '%s' is not opened", bs->filename);
743 return -ENOMEDIUM;
744 }
745
746 if (!bs->drv->bdrv_co_delete_file) {
747 error_setg(errp, "Driver '%s' does not support image deletion",
748 bs->drv->format_name);
749 return -ENOTSUP;
750 }
751
752 ret = bs->drv->bdrv_co_delete_file(bs, &local_err);
753 if (ret < 0) {
754 error_propagate(errp, local_err);
755 }
756
757 return ret;
758 }
759
760 void coroutine_fn bdrv_co_delete_file_noerr(BlockDriverState *bs)
761 {
762 Error *local_err = NULL;
763 int ret;
764
765 if (!bs) {
766 return;
767 }
768
769 ret = bdrv_co_delete_file(bs, &local_err);
770 /*
771 * ENOTSUP will happen if the block driver doesn't support
772 * the 'bdrv_co_delete_file' interface. This is a predictable
773 * scenario and shouldn't be reported back to the user.
774 */
775 if (ret == -ENOTSUP) {
776 error_free(local_err);
777 } else if (ret < 0) {
778 error_report_err(local_err);
779 }
780 }
781
782 /**
783 * Try to get @bs's logical and physical block size.
784 * On success, store them in @bsz struct and return 0.
785 * On failure return -errno.
786 * @bs must not be empty.
787 */
788 int bdrv_probe_blocksizes(BlockDriverState *bs, BlockSizes *bsz)
789 {
790 BlockDriver *drv = bs->drv;
791 BlockDriverState *filtered = bdrv_filter_bs(bs);
792
793 if (drv && drv->bdrv_probe_blocksizes) {
794 return drv->bdrv_probe_blocksizes(bs, bsz);
795 } else if (filtered) {
796 return bdrv_probe_blocksizes(filtered, bsz);
797 }
798
799 return -ENOTSUP;
800 }
801
802 /**
803 * Try to get @bs's geometry (cyls, heads, sectors).
804 * On success, store them in @geo struct and return 0.
805 * On failure return -errno.
806 * @bs must not be empty.
807 */
808 int bdrv_probe_geometry(BlockDriverState *bs, HDGeometry *geo)
809 {
810 BlockDriver *drv = bs->drv;
811 BlockDriverState *filtered = bdrv_filter_bs(bs);
812
813 if (drv && drv->bdrv_probe_geometry) {
814 return drv->bdrv_probe_geometry(bs, geo);
815 } else if (filtered) {
816 return bdrv_probe_geometry(filtered, geo);
817 }
818
819 return -ENOTSUP;
820 }
821
822 /*
823 * Create a uniquely-named empty temporary file.
824 * Return 0 upon success, otherwise a negative errno value.
825 */
826 int get_tmp_filename(char *filename, int size)
827 {
828 #ifdef _WIN32
829 char temp_dir[MAX_PATH];
830 /* GetTempFileName requires that its output buffer (4th param)
831 have length MAX_PATH or greater. */
832 assert(size >= MAX_PATH);
833 return (GetTempPath(MAX_PATH, temp_dir)
834 && GetTempFileName(temp_dir, "qem", 0, filename)
835 ? 0 : -GetLastError());
836 #else
837 int fd;
838 const char *tmpdir;
839 tmpdir = getenv("TMPDIR");
840 if (!tmpdir) {
841 tmpdir = "/var/tmp";
842 }
843 if (snprintf(filename, size, "%s/vl.XXXXXX", tmpdir) >= size) {
844 return -EOVERFLOW;
845 }
846 fd = mkstemp(filename);
847 if (fd < 0) {
848 return -errno;
849 }
850 if (close(fd) != 0) {
851 unlink(filename);
852 return -errno;
853 }
854 return 0;
855 #endif
856 }
857
858 /*
859 * Detect host devices. By convention, /dev/cdrom[N] is always
860 * recognized as a host CDROM.
861 */
862 static BlockDriver *find_hdev_driver(const char *filename)
863 {
864 int score_max = 0, score;
865 BlockDriver *drv = NULL, *d;
866
867 QLIST_FOREACH(d, &bdrv_drivers, list) {
868 if (d->bdrv_probe_device) {
869 score = d->bdrv_probe_device(filename);
870 if (score > score_max) {
871 score_max = score;
872 drv = d;
873 }
874 }
875 }
876
877 return drv;
878 }
879
880 static BlockDriver *bdrv_do_find_protocol(const char *protocol)
881 {
882 BlockDriver *drv1;
883
884 QLIST_FOREACH(drv1, &bdrv_drivers, list) {
885 if (drv1->protocol_name && !strcmp(drv1->protocol_name, protocol)) {
886 return drv1;
887 }
888 }
889
890 return NULL;
891 }
892
893 BlockDriver *bdrv_find_protocol(const char *filename,
894 bool allow_protocol_prefix,
895 Error **errp)
896 {
897 BlockDriver *drv1;
898 char protocol[128];
899 int len;
900 const char *p;
901 int i;
902
903 /* TODO Drivers without bdrv_file_open must be specified explicitly */
904
905 /*
906 * XXX(hch): we really should not let host device detection
907 * override an explicit protocol specification, but moving this
908 * later breaks access to device names with colons in them.
909 * Thanks to the brain-dead persistent naming schemes on udev-
910 * based Linux systems those actually are quite common.
911 */
912 drv1 = find_hdev_driver(filename);
913 if (drv1) {
914 return drv1;
915 }
916
917 if (!path_has_protocol(filename) || !allow_protocol_prefix) {
918 return &bdrv_file;
919 }
920
921 p = strchr(filename, ':');
922 assert(p != NULL);
923 len = p - filename;
924 if (len > sizeof(protocol) - 1)
925 len = sizeof(protocol) - 1;
926 memcpy(protocol, filename, len);
927 protocol[len] = '\0';
928
929 drv1 = bdrv_do_find_protocol(protocol);
930 if (drv1) {
931 return drv1;
932 }
933
934 for (i = 0; i < (int)ARRAY_SIZE(block_driver_modules); ++i) {
935 if (block_driver_modules[i].protocol_name &&
936 !strcmp(block_driver_modules[i].protocol_name, protocol)) {
937 block_module_load_one(block_driver_modules[i].library_name);
938 break;
939 }
940 }
941
942 drv1 = bdrv_do_find_protocol(protocol);
943 if (!drv1) {
944 error_setg(errp, "Unknown protocol '%s'", protocol);
945 }
946 return drv1;
947 }
948
949 /*
950 * Guess image format by probing its contents.
951 * This is not a good idea when your image is raw (CVE-2008-2004), but
952 * we do it anyway for backward compatibility.
953 *
954 * @buf contains the image's first @buf_size bytes.
955 * @buf_size is the buffer size in bytes (generally BLOCK_PROBE_BUF_SIZE,
956 * but can be smaller if the image file is smaller)
957 * @filename is its filename.
958 *
959 * For all block drivers, call the bdrv_probe() method to get its
960 * probing score.
961 * Return the first block driver with the highest probing score.
962 */
963 BlockDriver *bdrv_probe_all(const uint8_t *buf, int buf_size,
964 const char *filename)
965 {
966 int score_max = 0, score;
967 BlockDriver *drv = NULL, *d;
968
969 QLIST_FOREACH(d, &bdrv_drivers, list) {
970 if (d->bdrv_probe) {
971 score = d->bdrv_probe(buf, buf_size, filename);
972 if (score > score_max) {
973 score_max = score;
974 drv = d;
975 }
976 }
977 }
978
979 return drv;
980 }
981
982 static int find_image_format(BlockBackend *file, const char *filename,
983 BlockDriver **pdrv, Error **errp)
984 {
985 BlockDriver *drv;
986 uint8_t buf[BLOCK_PROBE_BUF_SIZE];
987 int ret = 0;
988
989 /* Return the raw BlockDriver * to scsi-generic devices or empty drives */
990 if (blk_is_sg(file) || !blk_is_inserted(file) || blk_getlength(file) == 0) {
991 *pdrv = &bdrv_raw;
992 return ret;
993 }
994
995 ret = blk_pread(file, 0, buf, sizeof(buf));
996 if (ret < 0) {
997 error_setg_errno(errp, -ret, "Could not read image for determining its "
998 "format");
999 *pdrv = NULL;
1000 return ret;
1001 }
1002
1003 drv = bdrv_probe_all(buf, ret, filename);
1004 if (!drv) {
1005 error_setg(errp, "Could not determine image format: No compatible "
1006 "driver found");
1007 ret = -ENOENT;
1008 }
1009 *pdrv = drv;
1010 return ret;
1011 }
1012
1013 /**
1014 * Set the current 'total_sectors' value
1015 * Return 0 on success, -errno on error.
1016 */
1017 int refresh_total_sectors(BlockDriverState *bs, int64_t hint)
1018 {
1019 BlockDriver *drv = bs->drv;
1020
1021 if (!drv) {
1022 return -ENOMEDIUM;
1023 }
1024
1025 /* Do not attempt drv->bdrv_getlength() on scsi-generic devices */
1026 if (bdrv_is_sg(bs))
1027 return 0;
1028
1029 /* query actual device if possible, otherwise just trust the hint */
1030 if (drv->bdrv_getlength) {
1031 int64_t length = drv->bdrv_getlength(bs);
1032 if (length < 0) {
1033 return length;
1034 }
1035 hint = DIV_ROUND_UP(length, BDRV_SECTOR_SIZE);
1036 }
1037
1038 bs->total_sectors = hint;
1039
1040 if (bs->total_sectors * BDRV_SECTOR_SIZE > BDRV_MAX_LENGTH) {
1041 return -EFBIG;
1042 }
1043
1044 return 0;
1045 }
1046
1047 /**
1048 * Combines a QDict of new block driver @options with any missing options taken
1049 * from @old_options, so that leaving out an option defaults to its old value.
1050 */
1051 static void bdrv_join_options(BlockDriverState *bs, QDict *options,
1052 QDict *old_options)
1053 {
1054 if (bs->drv && bs->drv->bdrv_join_options) {
1055 bs->drv->bdrv_join_options(options, old_options);
1056 } else {
1057 qdict_join(options, old_options, false);
1058 }
1059 }
1060
1061 static BlockdevDetectZeroesOptions bdrv_parse_detect_zeroes(QemuOpts *opts,
1062 int open_flags,
1063 Error **errp)
1064 {
1065 Error *local_err = NULL;
1066 char *value = qemu_opt_get_del(opts, "detect-zeroes");
1067 BlockdevDetectZeroesOptions detect_zeroes =
1068 qapi_enum_parse(&BlockdevDetectZeroesOptions_lookup, value,
1069 BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF, &local_err);
1070 g_free(value);
1071 if (local_err) {
1072 error_propagate(errp, local_err);
1073 return detect_zeroes;
1074 }
1075
1076 if (detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP &&
1077 !(open_flags & BDRV_O_UNMAP))
1078 {
1079 error_setg(errp, "setting detect-zeroes to unmap is not allowed "
1080 "without setting discard operation to unmap");
1081 }
1082
1083 return detect_zeroes;
1084 }
1085
1086 /**
1087 * Set open flags for aio engine
1088 *
1089 * Return 0 on success, -1 if the engine specified is invalid
1090 */
1091 int bdrv_parse_aio(const char *mode, int *flags)
1092 {
1093 if (!strcmp(mode, "threads")) {
1094 /* do nothing, default */
1095 } else if (!strcmp(mode, "native")) {
1096 *flags |= BDRV_O_NATIVE_AIO;
1097 #ifdef CONFIG_LINUX_IO_URING
1098 } else if (!strcmp(mode, "io_uring")) {
1099 *flags |= BDRV_O_IO_URING;
1100 #endif
1101 } else {
1102 return -1;
1103 }
1104
1105 return 0;
1106 }
1107
1108 /**
1109 * Set open flags for a given discard mode
1110 *
1111 * Return 0 on success, -1 if the discard mode was invalid.
1112 */
1113 int bdrv_parse_discard_flags(const char *mode, int *flags)
1114 {
1115 *flags &= ~BDRV_O_UNMAP;
1116
1117 if (!strcmp(mode, "off") || !strcmp(mode, "ignore")) {
1118 /* do nothing */
1119 } else if (!strcmp(mode, "on") || !strcmp(mode, "unmap")) {
1120 *flags |= BDRV_O_UNMAP;
1121 } else {
1122 return -1;
1123 }
1124
1125 return 0;
1126 }
1127
1128 /**
1129 * Set open flags for a given cache mode
1130 *
1131 * Return 0 on success, -1 if the cache mode was invalid.
1132 */
1133 int bdrv_parse_cache_mode(const char *mode, int *flags, bool *writethrough)
1134 {
1135 *flags &= ~BDRV_O_CACHE_MASK;
1136
1137 if (!strcmp(mode, "off") || !strcmp(mode, "none")) {
1138 *writethrough = false;
1139 *flags |= BDRV_O_NOCACHE;
1140 } else if (!strcmp(mode, "directsync")) {
1141 *writethrough = true;
1142 *flags |= BDRV_O_NOCACHE;
1143 } else if (!strcmp(mode, "writeback")) {
1144 *writethrough = false;
1145 } else if (!strcmp(mode, "unsafe")) {
1146 *writethrough = false;
1147 *flags |= BDRV_O_NO_FLUSH;
1148 } else if (!strcmp(mode, "writethrough")) {
1149 *writethrough = true;
1150 } else {
1151 return -1;
1152 }
1153
1154 return 0;
1155 }
1156
1157 static char *bdrv_child_get_parent_desc(BdrvChild *c)
1158 {
1159 BlockDriverState *parent = c->opaque;
1160 return g_strdup_printf("node '%s'", bdrv_get_node_name(parent));
1161 }
1162
1163 static void bdrv_child_cb_drained_begin(BdrvChild *child)
1164 {
1165 BlockDriverState *bs = child->opaque;
1166 bdrv_do_drained_begin_quiesce(bs, NULL, false);
1167 }
1168
1169 static bool bdrv_child_cb_drained_poll(BdrvChild *child)
1170 {
1171 BlockDriverState *bs = child->opaque;
1172 return bdrv_drain_poll(bs, false, NULL, false);
1173 }
1174
1175 static void bdrv_child_cb_drained_end(BdrvChild *child,
1176 int *drained_end_counter)
1177 {
1178 BlockDriverState *bs = child->opaque;
1179 bdrv_drained_end_no_poll(bs, drained_end_counter);
1180 }
1181
1182 static int bdrv_child_cb_inactivate(BdrvChild *child)
1183 {
1184 BlockDriverState *bs = child->opaque;
1185 assert(bs->open_flags & BDRV_O_INACTIVE);
1186 return 0;
1187 }
1188
1189 static bool bdrv_child_cb_can_set_aio_ctx(BdrvChild *child, AioContext *ctx,
1190 GSList **ignore, Error **errp)
1191 {
1192 BlockDriverState *bs = child->opaque;
1193 return bdrv_can_set_aio_context(bs, ctx, ignore, errp);
1194 }
1195
1196 static void bdrv_child_cb_set_aio_ctx(BdrvChild *child, AioContext *ctx,
1197 GSList **ignore)
1198 {
1199 BlockDriverState *bs = child->opaque;
1200 return bdrv_set_aio_context_ignore(bs, ctx, ignore);
1201 }
1202
1203 /*
1204 * Returns the options and flags that a temporary snapshot should get, based on
1205 * the originally requested flags (the originally requested image will have
1206 * flags like a backing file)
1207 */
1208 static void bdrv_temp_snapshot_options(int *child_flags, QDict *child_options,
1209 int parent_flags, QDict *parent_options)
1210 {
1211 *child_flags = (parent_flags & ~BDRV_O_SNAPSHOT) | BDRV_O_TEMPORARY;
1212
1213 /* For temporary files, unconditional cache=unsafe is fine */
1214 qdict_set_default_str(child_options, BDRV_OPT_CACHE_DIRECT, "off");
1215 qdict_set_default_str(child_options, BDRV_OPT_CACHE_NO_FLUSH, "on");
1216
1217 /* Copy the read-only and discard options from the parent */
1218 qdict_copy_default(child_options, parent_options, BDRV_OPT_READ_ONLY);
1219 qdict_copy_default(child_options, parent_options, BDRV_OPT_DISCARD);
1220
1221 /* aio=native doesn't work for cache.direct=off, so disable it for the
1222 * temporary snapshot */
1223 *child_flags &= ~BDRV_O_NATIVE_AIO;
1224 }
1225
1226 static void bdrv_backing_attach(BdrvChild *c)
1227 {
1228 BlockDriverState *parent = c->opaque;
1229 BlockDriverState *backing_hd = c->bs;
1230
1231 assert(!parent->backing_blocker);
1232 error_setg(&parent->backing_blocker,
1233 "node is used as backing hd of '%s'",
1234 bdrv_get_device_or_node_name(parent));
1235
1236 bdrv_refresh_filename(backing_hd);
1237
1238 parent->open_flags &= ~BDRV_O_NO_BACKING;
1239
1240 bdrv_op_block_all(backing_hd, parent->backing_blocker);
1241 /* Otherwise we won't be able to commit or stream */
1242 bdrv_op_unblock(backing_hd, BLOCK_OP_TYPE_COMMIT_TARGET,
1243 parent->backing_blocker);
1244 bdrv_op_unblock(backing_hd, BLOCK_OP_TYPE_STREAM,
1245 parent->backing_blocker);
1246 /*
1247 * We do backup in 3 ways:
1248 * 1. drive backup
1249 * The target bs is new opened, and the source is top BDS
1250 * 2. blockdev backup
1251 * Both the source and the target are top BDSes.
1252 * 3. internal backup(used for block replication)
1253 * Both the source and the target are backing file
1254 *
1255 * In case 1 and 2, neither the source nor the target is the backing file.
1256 * In case 3, we will block the top BDS, so there is only one block job
1257 * for the top BDS and its backing chain.
1258 */
1259 bdrv_op_unblock(backing_hd, BLOCK_OP_TYPE_BACKUP_SOURCE,
1260 parent->backing_blocker);
1261 bdrv_op_unblock(backing_hd, BLOCK_OP_TYPE_BACKUP_TARGET,
1262 parent->backing_blocker);
1263 }
1264
1265 static void bdrv_backing_detach(BdrvChild *c)
1266 {
1267 BlockDriverState *parent = c->opaque;
1268
1269 assert(parent->backing_blocker);
1270 bdrv_op_unblock_all(c->bs, parent->backing_blocker);
1271 error_free(parent->backing_blocker);
1272 parent->backing_blocker = NULL;
1273 }
1274
1275 static int bdrv_backing_update_filename(BdrvChild *c, BlockDriverState *base,
1276 const char *filename, Error **errp)
1277 {
1278 BlockDriverState *parent = c->opaque;
1279 bool read_only = bdrv_is_read_only(parent);
1280 int ret;
1281
1282 if (read_only) {
1283 ret = bdrv_reopen_set_read_only(parent, false, errp);
1284 if (ret < 0) {
1285 return ret;
1286 }
1287 }
1288
1289 ret = bdrv_change_backing_file(parent, filename,
1290 base->drv ? base->drv->format_name : "",
1291 false);
1292 if (ret < 0) {
1293 error_setg_errno(errp, -ret, "Could not update backing file link");
1294 }
1295
1296 if (read_only) {
1297 bdrv_reopen_set_read_only(parent, true, NULL);
1298 }
1299
1300 return ret;
1301 }
1302
1303 /*
1304 * Returns the options and flags that a generic child of a BDS should
1305 * get, based on the given options and flags for the parent BDS.
1306 */
1307 static void bdrv_inherited_options(BdrvChildRole role, bool parent_is_format,
1308 int *child_flags, QDict *child_options,
1309 int parent_flags, QDict *parent_options)
1310 {
1311 int flags = parent_flags;
1312
1313 /*
1314 * First, decide whether to set, clear, or leave BDRV_O_PROTOCOL.
1315 * Generally, the question to answer is: Should this child be
1316 * format-probed by default?
1317 */
1318
1319 /*
1320 * Pure and non-filtered data children of non-format nodes should
1321 * be probed by default (even when the node itself has BDRV_O_PROTOCOL
1322 * set). This only affects a very limited set of drivers (namely
1323 * quorum and blkverify when this comment was written).
1324 * Force-clear BDRV_O_PROTOCOL then.
1325 */
1326 if (!parent_is_format &&
1327 (role & BDRV_CHILD_DATA) &&
1328 !(role & (BDRV_CHILD_METADATA | BDRV_CHILD_FILTERED)))
1329 {
1330 flags &= ~BDRV_O_PROTOCOL;
1331 }
1332
1333 /*
1334 * All children of format nodes (except for COW children) and all
1335 * metadata children in general should never be format-probed.
1336 * Force-set BDRV_O_PROTOCOL then.
1337 */
1338 if ((parent_is_format && !(role & BDRV_CHILD_COW)) ||
1339 (role & BDRV_CHILD_METADATA))
1340 {
1341 flags |= BDRV_O_PROTOCOL;
1342 }
1343
1344 /*
1345 * If the cache mode isn't explicitly set, inherit direct and no-flush from
1346 * the parent.
1347 */
1348 qdict_copy_default(child_options, parent_options, BDRV_OPT_CACHE_DIRECT);
1349 qdict_copy_default(child_options, parent_options, BDRV_OPT_CACHE_NO_FLUSH);
1350 qdict_copy_default(child_options, parent_options, BDRV_OPT_FORCE_SHARE);
1351
1352 if (role & BDRV_CHILD_COW) {
1353 /* backing files are opened read-only by default */
1354 qdict_set_default_str(child_options, BDRV_OPT_READ_ONLY, "on");
1355 qdict_set_default_str(child_options, BDRV_OPT_AUTO_READ_ONLY, "off");
1356 } else {
1357 /* Inherit the read-only option from the parent if it's not set */
1358 qdict_copy_default(child_options, parent_options, BDRV_OPT_READ_ONLY);
1359 qdict_copy_default(child_options, parent_options,
1360 BDRV_OPT_AUTO_READ_ONLY);
1361 }
1362
1363 /*
1364 * bdrv_co_pdiscard() respects unmap policy for the parent, so we
1365 * can default to enable it on lower layers regardless of the
1366 * parent option.
1367 */
1368 qdict_set_default_str(child_options, BDRV_OPT_DISCARD, "unmap");
1369
1370 /* Clear flags that only apply to the top layer */
1371 flags &= ~(BDRV_O_SNAPSHOT | BDRV_O_NO_BACKING | BDRV_O_COPY_ON_READ);
1372
1373 if (role & BDRV_CHILD_METADATA) {
1374 flags &= ~BDRV_O_NO_IO;
1375 }
1376 if (role & BDRV_CHILD_COW) {
1377 flags &= ~BDRV_O_TEMPORARY;
1378 }
1379
1380 *child_flags = flags;
1381 }
1382
1383 static void bdrv_child_cb_attach(BdrvChild *child)
1384 {
1385 BlockDriverState *bs = child->opaque;
1386
1387 if (child->role & BDRV_CHILD_COW) {
1388 bdrv_backing_attach(child);
1389 }
1390
1391 bdrv_apply_subtree_drain(child, bs);
1392 }
1393
1394 static void bdrv_child_cb_detach(BdrvChild *child)
1395 {
1396 BlockDriverState *bs = child->opaque;
1397
1398 if (child->role & BDRV_CHILD_COW) {
1399 bdrv_backing_detach(child);
1400 }
1401
1402 bdrv_unapply_subtree_drain(child, bs);
1403 }
1404
1405 static int bdrv_child_cb_update_filename(BdrvChild *c, BlockDriverState *base,
1406 const char *filename, Error **errp)
1407 {
1408 if (c->role & BDRV_CHILD_COW) {
1409 return bdrv_backing_update_filename(c, base, filename, errp);
1410 }
1411 return 0;
1412 }
1413
1414 AioContext *child_of_bds_get_parent_aio_context(BdrvChild *c)
1415 {
1416 BlockDriverState *bs = c->opaque;
1417
1418 return bdrv_get_aio_context(bs);
1419 }
1420
1421 const BdrvChildClass child_of_bds = {
1422 .parent_is_bds = true,
1423 .get_parent_desc = bdrv_child_get_parent_desc,
1424 .inherit_options = bdrv_inherited_options,
1425 .drained_begin = bdrv_child_cb_drained_begin,
1426 .drained_poll = bdrv_child_cb_drained_poll,
1427 .drained_end = bdrv_child_cb_drained_end,
1428 .attach = bdrv_child_cb_attach,
1429 .detach = bdrv_child_cb_detach,
1430 .inactivate = bdrv_child_cb_inactivate,
1431 .can_set_aio_ctx = bdrv_child_cb_can_set_aio_ctx,
1432 .set_aio_ctx = bdrv_child_cb_set_aio_ctx,
1433 .update_filename = bdrv_child_cb_update_filename,
1434 .get_parent_aio_context = child_of_bds_get_parent_aio_context,
1435 };
1436
1437 AioContext *bdrv_child_get_parent_aio_context(BdrvChild *c)
1438 {
1439 return c->klass->get_parent_aio_context(c);
1440 }
1441
1442 static int bdrv_open_flags(BlockDriverState *bs, int flags)
1443 {
1444 int open_flags = flags;
1445
1446 /*
1447 * Clear flags that are internal to the block layer before opening the
1448 * image.
1449 */
1450 open_flags &= ~(BDRV_O_SNAPSHOT | BDRV_O_NO_BACKING | BDRV_O_PROTOCOL);
1451
1452 return open_flags;
1453 }
1454
1455 static void update_flags_from_options(int *flags, QemuOpts *opts)
1456 {
1457 *flags &= ~(BDRV_O_CACHE_MASK | BDRV_O_RDWR | BDRV_O_AUTO_RDONLY);
1458
1459 if (qemu_opt_get_bool_del(opts, BDRV_OPT_CACHE_NO_FLUSH, false)) {
1460 *flags |= BDRV_O_NO_FLUSH;
1461 }
1462
1463 if (qemu_opt_get_bool_del(opts, BDRV_OPT_CACHE_DIRECT, false)) {
1464 *flags |= BDRV_O_NOCACHE;
1465 }
1466
1467 if (!qemu_opt_get_bool_del(opts, BDRV_OPT_READ_ONLY, false)) {
1468 *flags |= BDRV_O_RDWR;
1469 }
1470
1471 if (qemu_opt_get_bool_del(opts, BDRV_OPT_AUTO_READ_ONLY, false)) {
1472 *flags |= BDRV_O_AUTO_RDONLY;
1473 }
1474 }
1475
1476 static void update_options_from_flags(QDict *options, int flags)
1477 {
1478 if (!qdict_haskey(options, BDRV_OPT_CACHE_DIRECT)) {
1479 qdict_put_bool(options, BDRV_OPT_CACHE_DIRECT, flags & BDRV_O_NOCACHE);
1480 }
1481 if (!qdict_haskey(options, BDRV_OPT_CACHE_NO_FLUSH)) {
1482 qdict_put_bool(options, BDRV_OPT_CACHE_NO_FLUSH,
1483 flags & BDRV_O_NO_FLUSH);
1484 }
1485 if (!qdict_haskey(options, BDRV_OPT_READ_ONLY)) {
1486 qdict_put_bool(options, BDRV_OPT_READ_ONLY, !(flags & BDRV_O_RDWR));
1487 }
1488 if (!qdict_haskey(options, BDRV_OPT_AUTO_READ_ONLY)) {
1489 qdict_put_bool(options, BDRV_OPT_AUTO_READ_ONLY,
1490 flags & BDRV_O_AUTO_RDONLY);
1491 }
1492 }
1493
1494 static void bdrv_assign_node_name(BlockDriverState *bs,
1495 const char *node_name,
1496 Error **errp)
1497 {
1498 char *gen_node_name = NULL;
1499
1500 if (!node_name) {
1501 node_name = gen_node_name = id_generate(ID_BLOCK);
1502 } else if (!id_wellformed(node_name)) {
1503 /*
1504 * Check for empty string or invalid characters, but not if it is
1505 * generated (generated names use characters not available to the user)
1506 */
1507 error_setg(errp, "Invalid node-name: '%s'", node_name);
1508 return;
1509 }
1510
1511 /* takes care of avoiding namespaces collisions */
1512 if (blk_by_name(node_name)) {
1513 error_setg(errp, "node-name=%s is conflicting with a device id",
1514 node_name);
1515 goto out;
1516 }
1517
1518 /* takes care of avoiding duplicates node names */
1519 if (bdrv_find_node(node_name)) {
1520 error_setg(errp, "Duplicate nodes with node-name='%s'", node_name);
1521 goto out;
1522 }
1523
1524 /* Make sure that the node name isn't truncated */
1525 if (strlen(node_name) >= sizeof(bs->node_name)) {
1526 error_setg(errp, "Node name too long");
1527 goto out;
1528 }
1529
1530 /* copy node name into the bs and insert it into the graph list */
1531 pstrcpy(bs->node_name, sizeof(bs->node_name), node_name);
1532 QTAILQ_INSERT_TAIL(&graph_bdrv_states, bs, node_list);
1533 out:
1534 g_free(gen_node_name);
1535 }
1536
1537 static int bdrv_open_driver(BlockDriverState *bs, BlockDriver *drv,
1538 const char *node_name, QDict *options,
1539 int open_flags, Error **errp)
1540 {
1541 Error *local_err = NULL;
1542 int i, ret;
1543
1544 bdrv_assign_node_name(bs, node_name, &local_err);
1545 if (local_err) {
1546 error_propagate(errp, local_err);
1547 return -EINVAL;
1548 }
1549
1550 bs->drv = drv;
1551 bs->opaque = g_malloc0(drv->instance_size);
1552
1553 if (drv->bdrv_file_open) {
1554 assert(!drv->bdrv_needs_filename || bs->filename[0]);
1555 ret = drv->bdrv_file_open(bs, options, open_flags, &local_err);
1556 } else if (drv->bdrv_open) {
1557 ret = drv->bdrv_open(bs, options, open_flags, &local_err);
1558 } else {
1559 ret = 0;
1560 }
1561
1562 if (ret < 0) {
1563 if (local_err) {
1564 error_propagate(errp, local_err);
1565 } else if (bs->filename[0]) {
1566 error_setg_errno(errp, -ret, "Could not open '%s'", bs->filename);
1567 } else {
1568 error_setg_errno(errp, -ret, "Could not open image");
1569 }
1570 goto open_failed;
1571 }
1572
1573 ret = refresh_total_sectors(bs, bs->total_sectors);
1574 if (ret < 0) {
1575 error_setg_errno(errp, -ret, "Could not refresh total sector count");
1576 return ret;
1577 }
1578
1579 bdrv_refresh_limits(bs, NULL, &local_err);
1580 if (local_err) {
1581 error_propagate(errp, local_err);
1582 return -EINVAL;
1583 }
1584
1585 assert(bdrv_opt_mem_align(bs) != 0);
1586 assert(bdrv_min_mem_align(bs) != 0);
1587 assert(is_power_of_2(bs->bl.request_alignment));
1588
1589 for (i = 0; i < bs->quiesce_counter; i++) {
1590 if (drv->bdrv_co_drain_begin) {
1591 drv->bdrv_co_drain_begin(bs);
1592 }
1593 }
1594
1595 return 0;
1596 open_failed:
1597 bs->drv = NULL;
1598 if (bs->file != NULL) {
1599 bdrv_unref_child(bs, bs->file);
1600 bs->file = NULL;
1601 }
1602 g_free(bs->opaque);
1603 bs->opaque = NULL;
1604 return ret;
1605 }
1606
1607 /*
1608 * Create and open a block node.
1609 *
1610 * @options is a QDict of options to pass to the block drivers, or NULL for an
1611 * empty set of options. The reference to the QDict belongs to the block layer
1612 * after the call (even on failure), so if the caller intends to reuse the
1613 * dictionary, it needs to use qobject_ref() before calling bdrv_open.
1614 */
1615 BlockDriverState *bdrv_new_open_driver_opts(BlockDriver *drv,
1616 const char *node_name,
1617 QDict *options, int flags,
1618 Error **errp)
1619 {
1620 BlockDriverState *bs;
1621 int ret;
1622
1623 bs = bdrv_new();
1624 bs->open_flags = flags;
1625 bs->options = options ?: qdict_new();
1626 bs->explicit_options = qdict_clone_shallow(bs->options);
1627 bs->opaque = NULL;
1628
1629 update_options_from_flags(bs->options, flags);
1630
1631 ret = bdrv_open_driver(bs, drv, node_name, bs->options, flags, errp);
1632 if (ret < 0) {
1633 qobject_unref(bs->explicit_options);
1634 bs->explicit_options = NULL;
1635 qobject_unref(bs->options);
1636 bs->options = NULL;
1637 bdrv_unref(bs);
1638 return NULL;
1639 }
1640
1641 return bs;
1642 }
1643
1644 /* Create and open a block node. */
1645 BlockDriverState *bdrv_new_open_driver(BlockDriver *drv, const char *node_name,
1646 int flags, Error **errp)
1647 {
1648 return bdrv_new_open_driver_opts(drv, node_name, NULL, flags, errp);
1649 }
1650
1651 QemuOptsList bdrv_runtime_opts = {
1652 .name = "bdrv_common",
1653 .head = QTAILQ_HEAD_INITIALIZER(bdrv_runtime_opts.head),
1654 .desc = {
1655 {
1656 .name = "node-name",
1657 .type = QEMU_OPT_STRING,
1658 .help = "Node name of the block device node",
1659 },
1660 {
1661 .name = "driver",
1662 .type = QEMU_OPT_STRING,
1663 .help = "Block driver to use for the node",
1664 },
1665 {
1666 .name = BDRV_OPT_CACHE_DIRECT,
1667 .type = QEMU_OPT_BOOL,
1668 .help = "Bypass software writeback cache on the host",
1669 },
1670 {
1671 .name = BDRV_OPT_CACHE_NO_FLUSH,
1672 .type = QEMU_OPT_BOOL,
1673 .help = "Ignore flush requests",
1674 },
1675 {
1676 .name = BDRV_OPT_READ_ONLY,
1677 .type = QEMU_OPT_BOOL,
1678 .help = "Node is opened in read-only mode",
1679 },
1680 {
1681 .name = BDRV_OPT_AUTO_READ_ONLY,
1682 .type = QEMU_OPT_BOOL,
1683 .help = "Node can become read-only if opening read-write fails",
1684 },
1685 {
1686 .name = "detect-zeroes",
1687 .type = QEMU_OPT_STRING,
1688 .help = "try to optimize zero writes (off, on, unmap)",
1689 },
1690 {
1691 .name = BDRV_OPT_DISCARD,
1692 .type = QEMU_OPT_STRING,
1693 .help = "discard operation (ignore/off, unmap/on)",
1694 },
1695 {
1696 .name = BDRV_OPT_FORCE_SHARE,
1697 .type = QEMU_OPT_BOOL,
1698 .help = "always accept other writers (default: off)",
1699 },
1700 { /* end of list */ }
1701 },
1702 };
1703
1704 QemuOptsList bdrv_create_opts_simple = {
1705 .name = "simple-create-opts",
1706 .head = QTAILQ_HEAD_INITIALIZER(bdrv_create_opts_simple.head),
1707 .desc = {
1708 {
1709 .name = BLOCK_OPT_SIZE,
1710 .type = QEMU_OPT_SIZE,
1711 .help = "Virtual disk size"
1712 },
1713 {
1714 .name = BLOCK_OPT_PREALLOC,
1715 .type = QEMU_OPT_STRING,
1716 .help = "Preallocation mode (allowed values: off)"
1717 },
1718 { /* end of list */ }
1719 }
1720 };
1721
1722 /*
1723 * Common part for opening disk images and files
1724 *
1725 * Removes all processed options from *options.
1726 */
1727 static int bdrv_open_common(BlockDriverState *bs, BlockBackend *file,
1728 QDict *options, Error **errp)
1729 {
1730 int ret, open_flags;
1731 const char *filename;
1732 const char *driver_name = NULL;
1733 const char *node_name = NULL;
1734 const char *discard;
1735 QemuOpts *opts;
1736 BlockDriver *drv;
1737 Error *local_err = NULL;
1738 bool ro;
1739
1740 assert(bs->file == NULL);
1741 assert(options != NULL && bs->options != options);
1742
1743 opts = qemu_opts_create(&bdrv_runtime_opts, NULL, 0, &error_abort);
1744 if (!qemu_opts_absorb_qdict(opts, options, errp)) {
1745 ret = -EINVAL;
1746 goto fail_opts;
1747 }
1748
1749 update_flags_from_options(&bs->open_flags, opts);
1750
1751 driver_name = qemu_opt_get(opts, "driver");
1752 drv = bdrv_find_format(driver_name);
1753 assert(drv != NULL);
1754
1755 bs->force_share = qemu_opt_get_bool(opts, BDRV_OPT_FORCE_SHARE, false);
1756
1757 if (bs->force_share && (bs->open_flags & BDRV_O_RDWR)) {
1758 error_setg(errp,
1759 BDRV_OPT_FORCE_SHARE
1760 "=on can only be used with read-only images");
1761 ret = -EINVAL;
1762 goto fail_opts;
1763 }
1764
1765 if (file != NULL) {
1766 bdrv_refresh_filename(blk_bs(file));
1767 filename = blk_bs(file)->filename;
1768 } else {
1769 /*
1770 * Caution: while qdict_get_try_str() is fine, getting
1771 * non-string types would require more care. When @options
1772 * come from -blockdev or blockdev_add, its members are typed
1773 * according to the QAPI schema, but when they come from
1774 * -drive, they're all QString.
1775 */
1776 filename = qdict_get_try_str(options, "filename");
1777 }
1778
1779 if (drv->bdrv_needs_filename && (!filename || !filename[0])) {
1780 error_setg(errp, "The '%s' block driver requires a file name",
1781 drv->format_name);
1782 ret = -EINVAL;
1783 goto fail_opts;
1784 }
1785
1786 trace_bdrv_open_common(bs, filename ?: "", bs->open_flags,
1787 drv->format_name);
1788
1789 ro = bdrv_is_read_only(bs);
1790
1791 if (use_bdrv_whitelist && !bdrv_is_whitelisted(drv, ro)) {
1792 if (!ro && bdrv_is_whitelisted(drv, true)) {
1793 ret = bdrv_apply_auto_read_only(bs, NULL, NULL);
1794 } else {
1795 ret = -ENOTSUP;
1796 }
1797 if (ret < 0) {
1798 error_setg(errp,
1799 !ro && bdrv_is_whitelisted(drv, true)
1800 ? "Driver '%s' can only be used for read-only devices"
1801 : "Driver '%s' is not whitelisted",
1802 drv->format_name);
1803 goto fail_opts;
1804 }
1805 }
1806
1807 /* bdrv_new() and bdrv_close() make it so */
1808 assert(qatomic_read(&bs->copy_on_read) == 0);
1809
1810 if (bs->open_flags & BDRV_O_COPY_ON_READ) {
1811 if (!ro) {
1812 bdrv_enable_copy_on_read(bs);
1813 } else {
1814 error_setg(errp, "Can't use copy-on-read on read-only device");
1815 ret = -EINVAL;
1816 goto fail_opts;
1817 }
1818 }
1819
1820 discard = qemu_opt_get(opts, BDRV_OPT_DISCARD);
1821 if (discard != NULL) {
1822 if (bdrv_parse_discard_flags(discard, &bs->open_flags) != 0) {
1823 error_setg(errp, "Invalid discard option");
1824 ret = -EINVAL;
1825 goto fail_opts;
1826 }
1827 }
1828
1829 bs->detect_zeroes =
1830 bdrv_parse_detect_zeroes(opts, bs->open_flags, &local_err);
1831 if (local_err) {
1832 error_propagate(errp, local_err);
1833 ret = -EINVAL;
1834 goto fail_opts;
1835 }
1836
1837 if (filename != NULL) {
1838 pstrcpy(bs->filename, sizeof(bs->filename), filename);
1839 } else {
1840 bs->filename[0] = '\0';
1841 }
1842 pstrcpy(bs->exact_filename, sizeof(bs->exact_filename), bs->filename);
1843
1844 /* Open the image, either directly or using a protocol */
1845 open_flags = bdrv_open_flags(bs, bs->open_flags);
1846 node_name = qemu_opt_get(opts, "node-name");
1847
1848 assert(!drv->bdrv_file_open || file == NULL);
1849 ret = bdrv_open_driver(bs, drv, node_name, options, open_flags, errp);
1850 if (ret < 0) {
1851 goto fail_opts;
1852 }
1853
1854 qemu_opts_del(opts);
1855 return 0;
1856
1857 fail_opts:
1858 qemu_opts_del(opts);
1859 return ret;
1860 }
1861
1862 static QDict *parse_json_filename(const char *filename, Error **errp)
1863 {
1864 QObject *options_obj;
1865 QDict *options;
1866 int ret;
1867
1868 ret = strstart(filename, "json:", &filename);
1869 assert(ret);
1870
1871 options_obj = qobject_from_json(filename, errp);
1872 if (!options_obj) {
1873 error_prepend(errp, "Could not parse the JSON options: ");
1874 return NULL;
1875 }
1876
1877 options = qobject_to(QDict, options_obj);
1878 if (!options) {
1879 qobject_unref(options_obj);
1880 error_setg(errp, "Invalid JSON object given");
1881 return NULL;
1882 }
1883
1884 qdict_flatten(options);
1885
1886 return options;
1887 }
1888
1889 static void parse_json_protocol(QDict *options, const char **pfilename,
1890 Error **errp)
1891 {
1892 QDict *json_options;
1893 Error *local_err = NULL;
1894
1895 /* Parse json: pseudo-protocol */
1896 if (!*pfilename || !g_str_has_prefix(*pfilename, "json:")) {
1897 return;
1898 }
1899
1900 json_options = parse_json_filename(*pfilename, &local_err);
1901 if (local_err) {
1902 error_propagate(errp, local_err);
1903 return;
1904 }
1905
1906 /* Options given in the filename have lower priority than options
1907 * specified directly */
1908 qdict_join(options, json_options, false);
1909 qobject_unref(json_options);
1910 *pfilename = NULL;
1911 }
1912
1913 /*
1914 * Fills in default options for opening images and converts the legacy
1915 * filename/flags pair to option QDict entries.
1916 * The BDRV_O_PROTOCOL flag in *flags will be set or cleared accordingly if a
1917 * block driver has been specified explicitly.
1918 */
1919 static int bdrv_fill_options(QDict **options, const char *filename,
1920 int *flags, Error **errp)
1921 {
1922 const char *drvname;
1923 bool protocol = *flags & BDRV_O_PROTOCOL;
1924 bool parse_filename = false;
1925 BlockDriver *drv = NULL;
1926 Error *local_err = NULL;
1927
1928 /*
1929 * Caution: while qdict_get_try_str() is fine, getting non-string
1930 * types would require more care. When @options come from
1931 * -blockdev or blockdev_add, its members are typed according to
1932 * the QAPI schema, but when they come from -drive, they're all
1933 * QString.
1934 */
1935 drvname = qdict_get_try_str(*options, "driver");
1936 if (drvname) {
1937 drv = bdrv_find_format(drvname);
1938 if (!drv) {
1939 error_setg(errp, "Unknown driver '%s'", drvname);
1940 return -ENOENT;
1941 }
1942 /* If the user has explicitly specified the driver, this choice should
1943 * override the BDRV_O_PROTOCOL flag */
1944 protocol = drv->bdrv_file_open;
1945 }
1946
1947 if (protocol) {
1948 *flags |= BDRV_O_PROTOCOL;
1949 } else {
1950 *flags &= ~BDRV_O_PROTOCOL;
1951 }
1952
1953 /* Translate cache options from flags into options */
1954 update_options_from_flags(*options, *flags);
1955
1956 /* Fetch the file name from the options QDict if necessary */
1957 if (protocol && filename) {
1958 if (!qdict_haskey(*options, "filename")) {
1959 qdict_put_str(*options, "filename", filename);
1960 parse_filename = true;
1961 } else {
1962 error_setg(errp, "Can't specify 'file' and 'filename' options at "
1963 "the same time");
1964 return -EINVAL;
1965 }
1966 }
1967
1968 /* Find the right block driver */
1969 /* See cautionary note on accessing @options above */
1970 filename = qdict_get_try_str(*options, "filename");
1971
1972 if (!drvname && protocol) {
1973 if (filename) {
1974 drv = bdrv_find_protocol(filename, parse_filename, errp);
1975 if (!drv) {
1976 return -EINVAL;
1977 }
1978
1979 drvname = drv->format_name;
1980 qdict_put_str(*options, "driver", drvname);
1981 } else {
1982 error_setg(errp, "Must specify either driver or file");
1983 return -EINVAL;
1984 }
1985 }
1986
1987 assert(drv || !protocol);
1988
1989 /* Driver-specific filename parsing */
1990 if (drv && drv->bdrv_parse_filename && parse_filename) {
1991 drv->bdrv_parse_filename(filename, *options, &local_err);
1992 if (local_err) {
1993 error_propagate(errp, local_err);
1994 return -EINVAL;
1995 }
1996
1997 if (!drv->bdrv_needs_filename) {
1998 qdict_del(*options, "filename");
1999 }
2000 }
2001
2002 return 0;
2003 }
2004
2005 typedef struct BlockReopenQueueEntry {
2006 bool prepared;
2007 bool perms_checked;
2008 BDRVReopenState state;
2009 QTAILQ_ENTRY(BlockReopenQueueEntry) entry;
2010 } BlockReopenQueueEntry;
2011
2012 /*
2013 * Return the flags that @bs will have after the reopens in @q have
2014 * successfully completed. If @q is NULL (or @bs is not contained in @q),
2015 * return the current flags.
2016 */
2017 static int bdrv_reopen_get_flags(BlockReopenQueue *q, BlockDriverState *bs)
2018 {
2019 BlockReopenQueueEntry *entry;
2020
2021 if (q != NULL) {
2022 QTAILQ_FOREACH(entry, q, entry) {
2023 if (entry->state.bs == bs) {
2024 return entry->state.flags;
2025 }
2026 }
2027 }
2028
2029 return bs->open_flags;
2030 }
2031
2032 /* Returns whether the image file can be written to after the reopen queue @q
2033 * has been successfully applied, or right now if @q is NULL. */
2034 static bool bdrv_is_writable_after_reopen(BlockDriverState *bs,
2035 BlockReopenQueue *q)
2036 {
2037 int flags = bdrv_reopen_get_flags(q, bs);
2038
2039 return (flags & (BDRV_O_RDWR | BDRV_O_INACTIVE)) == BDRV_O_RDWR;
2040 }
2041
2042 /*
2043 * Return whether the BDS can be written to. This is not necessarily
2044 * the same as !bdrv_is_read_only(bs), as inactivated images may not
2045 * be written to but do not count as read-only images.
2046 */
2047 bool bdrv_is_writable(BlockDriverState *bs)
2048 {
2049 return bdrv_is_writable_after_reopen(bs, NULL);
2050 }
2051
2052 static char *bdrv_child_user_desc(BdrvChild *c)
2053 {
2054 return c->klass->get_parent_desc(c);
2055 }
2056
2057 /*
2058 * Check that @a allows everything that @b needs. @a and @b must reference same
2059 * child node.
2060 */
2061 static bool bdrv_a_allow_b(BdrvChild *a, BdrvChild *b, Error **errp)
2062 {
2063 const char *child_bs_name;
2064 g_autofree char *a_user = NULL;
2065 g_autofree char *b_user = NULL;
2066 g_autofree char *perms = NULL;
2067
2068 assert(a->bs);
2069 assert(a->bs == b->bs);
2070
2071 if ((b->perm & a->shared_perm) == b->perm) {
2072 return true;
2073 }
2074
2075 child_bs_name = bdrv_get_node_name(b->bs);
2076 a_user = bdrv_child_user_desc(a);
2077 b_user = bdrv_child_user_desc(b);
2078 perms = bdrv_perm_names(b->perm & ~a->shared_perm);
2079
2080 error_setg(errp, "Permission conflict on node '%s': permissions '%s' are "
2081 "both required by %s (uses node '%s' as '%s' child) and "
2082 "unshared by %s (uses node '%s' as '%s' child).",
2083 child_bs_name, perms,
2084 b_user, child_bs_name, b->name,
2085 a_user, child_bs_name, a->name);
2086
2087 return false;
2088 }
2089
2090 static bool bdrv_parent_perms_conflict(BlockDriverState *bs, Error **errp)
2091 {
2092 BdrvChild *a, *b;
2093
2094 /*
2095 * During the loop we'll look at each pair twice. That's correct because
2096 * bdrv_a_allow_b() is asymmetric and we should check each pair in both
2097 * directions.
2098 */
2099 QLIST_FOREACH(a, &bs->parents, next_parent) {
2100 QLIST_FOREACH(b, &bs->parents, next_parent) {
2101 if (a == b) {
2102 continue;
2103 }
2104
2105 if (!bdrv_a_allow_b(a, b, errp)) {
2106 return true;
2107 }
2108 }
2109 }
2110
2111 return false;
2112 }
2113
2114 static void bdrv_child_perm(BlockDriverState *bs, BlockDriverState *child_bs,
2115 BdrvChild *c, BdrvChildRole role,
2116 BlockReopenQueue *reopen_queue,
2117 uint64_t parent_perm, uint64_t parent_shared,
2118 uint64_t *nperm, uint64_t *nshared)
2119 {
2120 assert(bs->drv && bs->drv->bdrv_child_perm);
2121 bs->drv->bdrv_child_perm(bs, c, role, reopen_queue,
2122 parent_perm, parent_shared,
2123 nperm, nshared);
2124 /* TODO Take force_share from reopen_queue */
2125 if (child_bs && child_bs->force_share) {
2126 *nshared = BLK_PERM_ALL;
2127 }
2128 }
2129
2130 /*
2131 * Adds the whole subtree of @bs (including @bs itself) to the @list (except for
2132 * nodes that are already in the @list, of course) so that final list is
2133 * topologically sorted. Return the result (GSList @list object is updated, so
2134 * don't use old reference after function call).
2135 *
2136 * On function start @list must be already topologically sorted and for any node
2137 * in the @list the whole subtree of the node must be in the @list as well. The
2138 * simplest way to satisfy this criteria: use only result of
2139 * bdrv_topological_dfs() or NULL as @list parameter.
2140 */
2141 static GSList *bdrv_topological_dfs(GSList *list, GHashTable *found,
2142 BlockDriverState *bs)
2143 {
2144 BdrvChild *child;
2145 g_autoptr(GHashTable) local_found = NULL;
2146
2147 if (!found) {
2148 assert(!list);
2149 found = local_found = g_hash_table_new(NULL, NULL);
2150 }
2151
2152 if (g_hash_table_contains(found, bs)) {
2153 return list;
2154 }
2155 g_hash_table_add(found, bs);
2156
2157 QLIST_FOREACH(child, &bs->children, next) {
2158 list = bdrv_topological_dfs(list, found, child->bs);
2159 }
2160
2161 return g_slist_prepend(list, bs);
2162 }
2163
2164 typedef struct BdrvChildSetPermState {
2165 BdrvChild *child;
2166 uint64_t old_perm;
2167 uint64_t old_shared_perm;
2168 } BdrvChildSetPermState;
2169
2170 static void bdrv_child_set_perm_abort(void *opaque)
2171 {
2172 BdrvChildSetPermState *s = opaque;
2173
2174 s->child->perm = s->old_perm;
2175 s->child->shared_perm = s->old_shared_perm;
2176 }
2177
2178 static TransactionActionDrv bdrv_child_set_pem_drv = {
2179 .abort = bdrv_child_set_perm_abort,
2180 .clean = g_free,
2181 };
2182
2183 static void bdrv_child_set_perm(BdrvChild *c, uint64_t perm,
2184 uint64_t shared, Transaction *tran)
2185 {
2186 BdrvChildSetPermState *s = g_new(BdrvChildSetPermState, 1);
2187
2188 *s = (BdrvChildSetPermState) {
2189 .child = c,
2190 .old_perm = c->perm,
2191 .old_shared_perm = c->shared_perm,
2192 };
2193
2194 c->perm = perm;
2195 c->shared_perm = shared;
2196
2197 tran_add(tran, &bdrv_child_set_pem_drv, s);
2198 }
2199
2200 static void bdrv_drv_set_perm_commit(void *opaque)
2201 {
2202 BlockDriverState *bs = opaque;
2203 uint64_t cumulative_perms, cumulative_shared_perms;
2204
2205 if (bs->drv->bdrv_set_perm) {
2206 bdrv_get_cumulative_perm(bs, &cumulative_perms,
2207 &cumulative_shared_perms);
2208 bs->drv->bdrv_set_perm(bs, cumulative_perms, cumulative_shared_perms);
2209 }
2210 }
2211
2212 static void bdrv_drv_set_perm_abort(void *opaque)
2213 {
2214 BlockDriverState *bs = opaque;
2215
2216 if (bs->drv->bdrv_abort_perm_update) {
2217 bs->drv->bdrv_abort_perm_update(bs);
2218 }
2219 }
2220
2221 TransactionActionDrv bdrv_drv_set_perm_drv = {
2222 .abort = bdrv_drv_set_perm_abort,
2223 .commit = bdrv_drv_set_perm_commit,
2224 };
2225
2226 static int bdrv_drv_set_perm(BlockDriverState *bs, uint64_t perm,
2227 uint64_t shared_perm, Transaction *tran,
2228 Error **errp)
2229 {
2230 if (!bs->drv) {
2231 return 0;
2232 }
2233
2234 if (bs->drv->bdrv_check_perm) {
2235 int ret = bs->drv->bdrv_check_perm(bs, perm, shared_perm, errp);
2236 if (ret < 0) {
2237 return ret;
2238 }
2239 }
2240
2241 if (tran) {
2242 tran_add(tran, &bdrv_drv_set_perm_drv, bs);
2243 }
2244
2245 return 0;
2246 }
2247
2248 typedef struct BdrvReplaceChildState {
2249 BdrvChild *child;
2250 BlockDriverState *old_bs;
2251 } BdrvReplaceChildState;
2252
2253 static void bdrv_replace_child_commit(void *opaque)
2254 {
2255 BdrvReplaceChildState *s = opaque;
2256
2257 bdrv_unref(s->old_bs);
2258 }
2259
2260 static void bdrv_replace_child_abort(void *opaque)
2261 {
2262 BdrvReplaceChildState *s = opaque;
2263 BlockDriverState *new_bs = s->child->bs;
2264
2265 /* old_bs reference is transparently moved from @s to @s->child */
2266 bdrv_replace_child_noperm(s->child, s->old_bs);
2267 bdrv_unref(new_bs);
2268 }
2269
2270 static TransactionActionDrv bdrv_replace_child_drv = {
2271 .commit = bdrv_replace_child_commit,
2272 .abort = bdrv_replace_child_abort,
2273 .clean = g_free,
2274 };
2275
2276 /*
2277 * bdrv_replace_child_tran
2278 *
2279 * Note: real unref of old_bs is done only on commit.
2280 *
2281 * The function doesn't update permissions, caller is responsible for this.
2282 */
2283 static void bdrv_replace_child_tran(BdrvChild *child, BlockDriverState *new_bs,
2284 Transaction *tran)
2285 {
2286 BdrvReplaceChildState *s = g_new(BdrvReplaceChildState, 1);
2287 *s = (BdrvReplaceChildState) {
2288 .child = child,
2289 .old_bs = child->bs,
2290 };
2291 tran_add(tran, &bdrv_replace_child_drv, s);
2292
2293 if (new_bs) {
2294 bdrv_ref(new_bs);
2295 }
2296 bdrv_replace_child_noperm(child, new_bs);
2297 /* old_bs reference is transparently moved from @child to @s */
2298 }
2299
2300 /*
2301 * Refresh permissions in @bs subtree. The function is intended to be called
2302 * after some graph modification that was done without permission update.
2303 */
2304 static int bdrv_node_refresh_perm(BlockDriverState *bs, BlockReopenQueue *q,
2305 Transaction *tran, Error **errp)
2306 {
2307 BlockDriver *drv = bs->drv;
2308 BdrvChild *c;
2309 int ret;
2310 uint64_t cumulative_perms, cumulative_shared_perms;
2311
2312 bdrv_get_cumulative_perm(bs, &cumulative_perms, &cumulative_shared_perms);
2313
2314 /* Write permissions never work with read-only images */
2315 if ((cumulative_perms & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED)) &&
2316 !bdrv_is_writable_after_reopen(bs, q))
2317 {
2318 if (!bdrv_is_writable_after_reopen(bs, NULL)) {
2319 error_setg(errp, "Block node is read-only");
2320 } else {
2321 error_setg(errp, "Read-only block node '%s' cannot support "
2322 "read-write users", bdrv_get_node_name(bs));
2323 }
2324
2325 return -EPERM;
2326 }
2327
2328 /*
2329 * Unaligned requests will automatically be aligned to bl.request_alignment
2330 * and without RESIZE we can't extend requests to write to space beyond the
2331 * end of the image, so it's required that the image size is aligned.
2332 */
2333 if ((cumulative_perms & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED)) &&
2334 !(cumulative_perms & BLK_PERM_RESIZE))
2335 {
2336 if ((bs->total_sectors * BDRV_SECTOR_SIZE) % bs->bl.request_alignment) {
2337 error_setg(errp, "Cannot get 'write' permission without 'resize': "
2338 "Image size is not a multiple of request "
2339 "alignment");
2340 return -EPERM;
2341 }
2342 }
2343
2344 /* Check this node */
2345 if (!drv) {
2346 return 0;
2347 }
2348
2349 ret = bdrv_drv_set_perm(bs, cumulative_perms, cumulative_shared_perms, tran,
2350 errp);
2351 if (ret < 0) {
2352 return ret;
2353 }
2354
2355 /* Drivers that never have children can omit .bdrv_child_perm() */
2356 if (!drv->bdrv_child_perm) {
2357 assert(QLIST_EMPTY(&bs->children));
2358 return 0;
2359 }
2360
2361 /* Check all children */
2362 QLIST_FOREACH(c, &bs->children, next) {
2363 uint64_t cur_perm, cur_shared;
2364
2365 bdrv_child_perm(bs, c->bs, c, c->role, q,
2366 cumulative_perms, cumulative_shared_perms,
2367 &cur_perm, &cur_shared);
2368 bdrv_child_set_perm(c, cur_perm, cur_shared, tran);
2369 }
2370
2371 return 0;
2372 }
2373
2374 static int bdrv_list_refresh_perms(GSList *list, BlockReopenQueue *q,
2375 Transaction *tran, Error **errp)
2376 {
2377 int ret;
2378 BlockDriverState *bs;
2379
2380 for ( ; list; list = list->next) {
2381 bs = list->data;
2382
2383 if (bdrv_parent_perms_conflict(bs, errp)) {
2384 return -EINVAL;
2385 }
2386
2387 ret = bdrv_node_refresh_perm(bs, q, tran, errp);
2388 if (ret < 0) {
2389 return ret;
2390 }
2391 }
2392
2393 return 0;
2394 }
2395
2396 void bdrv_get_cumulative_perm(BlockDriverState *bs, uint64_t *perm,
2397 uint64_t *shared_perm)
2398 {
2399 BdrvChild *c;
2400 uint64_t cumulative_perms = 0;
2401 uint64_t cumulative_shared_perms = BLK_PERM_ALL;
2402
2403 QLIST_FOREACH(c, &bs->parents, next_parent) {
2404 cumulative_perms |= c->perm;
2405 cumulative_shared_perms &= c->shared_perm;
2406 }
2407
2408 *perm = cumulative_perms;
2409 *shared_perm = cumulative_shared_perms;
2410 }
2411
2412 char *bdrv_perm_names(uint64_t perm)
2413 {
2414 struct perm_name {
2415 uint64_t perm;
2416 const char *name;
2417 } permissions[] = {
2418 { BLK_PERM_CONSISTENT_READ, "consistent read" },
2419 { BLK_PERM_WRITE, "write" },
2420 { BLK_PERM_WRITE_UNCHANGED, "write unchanged" },
2421 { BLK_PERM_RESIZE, "resize" },
2422 { BLK_PERM_GRAPH_MOD, "change children" },
2423 { 0, NULL }
2424 };
2425
2426 GString *result = g_string_sized_new(30);
2427 struct perm_name *p;
2428
2429 for (p = permissions; p->name; p++) {
2430 if (perm & p->perm) {
2431 if (result->len > 0) {
2432 g_string_append(result, ", ");
2433 }
2434 g_string_append(result, p->name);
2435 }
2436 }
2437
2438 return g_string_free(result, FALSE);
2439 }
2440
2441
2442 static int bdrv_refresh_perms(BlockDriverState *bs, Error **errp)
2443 {
2444 int ret;
2445 Transaction *tran = tran_new();
2446 g_autoptr(GSList) list = bdrv_topological_dfs(NULL, NULL, bs);
2447
2448 ret = bdrv_list_refresh_perms(list, NULL, tran, errp);
2449 tran_finalize(tran, ret);
2450
2451 return ret;
2452 }
2453
2454 int bdrv_child_try_set_perm(BdrvChild *c, uint64_t perm, uint64_t shared,
2455 Error **errp)
2456 {
2457 Error *local_err = NULL;
2458 Transaction *tran = tran_new();
2459 int ret;
2460
2461 bdrv_child_set_perm(c, perm, shared, tran);
2462
2463 ret = bdrv_refresh_perms(c->bs, &local_err);
2464
2465 tran_finalize(tran, ret);
2466
2467 if (ret < 0) {
2468 if ((perm & ~c->perm) || (c->shared_perm & ~shared)) {
2469 /* tighten permissions */
2470 error_propagate(errp, local_err);
2471 } else {
2472 /*
2473 * Our caller may intend to only loosen restrictions and
2474 * does not expect this function to fail. Errors are not
2475 * fatal in such a case, so we can just hide them from our
2476 * caller.
2477 */
2478 error_free(local_err);
2479 ret = 0;
2480 }
2481 }
2482
2483 return ret;
2484 }
2485
2486 int bdrv_child_refresh_perms(BlockDriverState *bs, BdrvChild *c, Error **errp)
2487 {
2488 uint64_t parent_perms, parent_shared;
2489 uint64_t perms, shared;
2490
2491 bdrv_get_cumulative_perm(bs, &parent_perms, &parent_shared);
2492 bdrv_child_perm(bs, c->bs, c, c->role, NULL,
2493 parent_perms, parent_shared, &perms, &shared);
2494
2495 return bdrv_child_try_set_perm(c, perms, shared, errp);
2496 }
2497
2498 /*
2499 * Default implementation for .bdrv_child_perm() for block filters:
2500 * Forward CONSISTENT_READ, WRITE, WRITE_UNCHANGED, and RESIZE to the
2501 * filtered child.
2502 */
2503 static void bdrv_filter_default_perms(BlockDriverState *bs, BdrvChild *c,
2504 BdrvChildRole role,
2505 BlockReopenQueue *reopen_queue,
2506 uint64_t perm, uint64_t shared,
2507 uint64_t *nperm, uint64_t *nshared)
2508 {
2509 *nperm = perm & DEFAULT_PERM_PASSTHROUGH;
2510 *nshared = (shared & DEFAULT_PERM_PASSTHROUGH) | DEFAULT_PERM_UNCHANGED;
2511 }
2512
2513 static void bdrv_default_perms_for_cow(BlockDriverState *bs, BdrvChild *c,
2514 BdrvChildRole role,
2515 BlockReopenQueue *reopen_queue,
2516 uint64_t perm, uint64_t shared,
2517 uint64_t *nperm, uint64_t *nshared)
2518 {
2519 assert(role & BDRV_CHILD_COW);
2520
2521 /*
2522 * We want consistent read from backing files if the parent needs it.
2523 * No other operations are performed on backing files.
2524 */
2525 perm &= BLK_PERM_CONSISTENT_READ;
2526
2527 /*
2528 * If the parent can deal with changing data, we're okay with a
2529 * writable and resizable backing file.
2530 * TODO Require !(perm & BLK_PERM_CONSISTENT_READ), too?
2531 */
2532 if (shared & BLK_PERM_WRITE) {
2533 shared = BLK_PERM_WRITE | BLK_PERM_RESIZE;
2534 } else {
2535 shared = 0;
2536 }
2537
2538 shared |= BLK_PERM_CONSISTENT_READ | BLK_PERM_GRAPH_MOD |
2539 BLK_PERM_WRITE_UNCHANGED;
2540
2541 if (bs->open_flags & BDRV_O_INACTIVE) {
2542 shared |= BLK_PERM_WRITE | BLK_PERM_RESIZE;
2543 }
2544
2545 *nperm = perm;
2546 *nshared = shared;
2547 }
2548
2549 static void bdrv_default_perms_for_storage(BlockDriverState *bs, BdrvChild *c,
2550 BdrvChildRole role,
2551 BlockReopenQueue *reopen_queue,
2552 uint64_t perm, uint64_t shared,
2553 uint64_t *nperm, uint64_t *nshared)
2554 {
2555 int flags;
2556
2557 assert(role & (BDRV_CHILD_METADATA | BDRV_CHILD_DATA));
2558
2559 flags = bdrv_reopen_get_flags(reopen_queue, bs);
2560
2561 /*
2562 * Apart from the modifications below, the same permissions are
2563 * forwarded and left alone as for filters
2564 */
2565 bdrv_filter_default_perms(bs, c, role, reopen_queue,
2566 perm, shared, &perm, &shared);
2567
2568 if (role & BDRV_CHILD_METADATA) {
2569 /* Format drivers may touch metadata even if the guest doesn't write */
2570 if (bdrv_is_writable_after_reopen(bs, reopen_queue)) {
2571 perm |= BLK_PERM_WRITE | BLK_PERM_RESIZE;
2572 }
2573
2574 /*
2575 * bs->file always needs to be consistent because of the
2576 * metadata. We can never allow other users to resize or write
2577 * to it.
2578 */
2579 if (!(flags & BDRV_O_NO_IO)) {
2580 perm |= BLK_PERM_CONSISTENT_READ;
2581 }
2582 shared &= ~(BLK_PERM_WRITE | BLK_PERM_RESIZE);
2583 }
2584
2585 if (role & BDRV_CHILD_DATA) {
2586 /*
2587 * Technically, everything in this block is a subset of the
2588 * BDRV_CHILD_METADATA path taken above, and so this could
2589 * be an "else if" branch. However, that is not obvious, and
2590 * this function is not performance critical, therefore we let
2591 * this be an independent "if".
2592 */
2593
2594 /*
2595 * We cannot allow other users to resize the file because the
2596 * format driver might have some assumptions about the size
2597 * (e.g. because it is stored in metadata, or because the file
2598 * is split into fixed-size data files).
2599 */
2600 shared &= ~BLK_PERM_RESIZE;
2601
2602 /*
2603 * WRITE_UNCHANGED often cannot be performed as such on the
2604 * data file. For example, the qcow2 driver may still need to
2605 * write copied clusters on copy-on-read.
2606 */
2607 if (perm & BLK_PERM_WRITE_UNCHANGED) {
2608 perm |= BLK_PERM_WRITE;
2609 }
2610
2611 /*
2612 * If the data file is written to, the format driver may
2613 * expect to be able to resize it by writing beyond the EOF.
2614 */
2615 if (perm & BLK_PERM_WRITE) {
2616 perm |= BLK_PERM_RESIZE;
2617 }
2618 }
2619
2620 if (bs->open_flags & BDRV_O_INACTIVE) {
2621 shared |= BLK_PERM_WRITE | BLK_PERM_RESIZE;
2622 }
2623
2624 *nperm = perm;
2625 *nshared = shared;
2626 }
2627
2628 void bdrv_default_perms(BlockDriverState *bs, BdrvChild *c,
2629 BdrvChildRole role, BlockReopenQueue *reopen_queue,
2630 uint64_t perm, uint64_t shared,
2631 uint64_t *nperm, uint64_t *nshared)
2632 {
2633 if (role & BDRV_CHILD_FILTERED) {
2634 assert(!(role & (BDRV_CHILD_DATA | BDRV_CHILD_METADATA |
2635 BDRV_CHILD_COW)));
2636 bdrv_filter_default_perms(bs, c, role, reopen_queue,
2637 perm, shared, nperm, nshared);
2638 } else if (role & BDRV_CHILD_COW) {
2639 assert(!(role & (BDRV_CHILD_DATA | BDRV_CHILD_METADATA)));
2640 bdrv_default_perms_for_cow(bs, c, role, reopen_queue,
2641 perm, shared, nperm, nshared);
2642 } else if (role & (BDRV_CHILD_METADATA | BDRV_CHILD_DATA)) {
2643 bdrv_default_perms_for_storage(bs, c, role, reopen_queue,
2644 perm, shared, nperm, nshared);
2645 } else {
2646 g_assert_not_reached();
2647 }
2648 }
2649
2650 uint64_t bdrv_qapi_perm_to_blk_perm(BlockPermission qapi_perm)
2651 {
2652 static const uint64_t permissions[] = {
2653 [BLOCK_PERMISSION_CONSISTENT_READ] = BLK_PERM_CONSISTENT_READ,
2654 [BLOCK_PERMISSION_WRITE] = BLK_PERM_WRITE,
2655 [BLOCK_PERMISSION_WRITE_UNCHANGED] = BLK_PERM_WRITE_UNCHANGED,
2656 [BLOCK_PERMISSION_RESIZE] = BLK_PERM_RESIZE,
2657 [BLOCK_PERMISSION_GRAPH_MOD] = BLK_PERM_GRAPH_MOD,
2658 };
2659
2660 QEMU_BUILD_BUG_ON(ARRAY_SIZE(permissions) != BLOCK_PERMISSION__MAX);
2661 QEMU_BUILD_BUG_ON(1UL << ARRAY_SIZE(permissions) != BLK_PERM_ALL + 1);
2662
2663 assert(qapi_perm < BLOCK_PERMISSION__MAX);
2664
2665 return permissions[qapi_perm];
2666 }
2667
2668 static void bdrv_replace_child_noperm(BdrvChild *child,
2669 BlockDriverState *new_bs)
2670 {
2671 BlockDriverState *old_bs = child->bs;
2672 int new_bs_quiesce_counter;
2673 int drain_saldo;
2674
2675 assert(!child->frozen);
2676
2677 if (old_bs && new_bs) {
2678 assert(bdrv_get_aio_context(old_bs) == bdrv_get_aio_context(new_bs));
2679 }
2680
2681 new_bs_quiesce_counter = (new_bs ? new_bs->quiesce_counter : 0);
2682 drain_saldo = new_bs_quiesce_counter - child->parent_quiesce_counter;
2683
2684 /*
2685 * If the new child node is drained but the old one was not, flush
2686 * all outstanding requests to the old child node.
2687 */
2688 while (drain_saldo > 0 && child->klass->drained_begin) {
2689 bdrv_parent_drained_begin_single(child, true);
2690 drain_saldo--;
2691 }
2692
2693 if (old_bs) {
2694 /* Detach first so that the recursive drain sections coming from @child
2695 * are already gone and we only end the drain sections that came from
2696 * elsewhere. */
2697 if (child->klass->detach) {
2698 child->klass->detach(child);
2699 }
2700 QLIST_REMOVE(child, next_parent);
2701 }
2702
2703 child->bs = new_bs;
2704
2705 if (new_bs) {
2706 QLIST_INSERT_HEAD(&new_bs->parents, child, next_parent);
2707
2708 /*
2709 * Detaching the old node may have led to the new node's
2710 * quiesce_counter having been decreased. Not a problem, we
2711 * just need to recognize this here and then invoke
2712 * drained_end appropriately more often.
2713 */
2714 assert(new_bs->quiesce_counter <= new_bs_quiesce_counter);
2715 drain_saldo += new_bs->quiesce_counter - new_bs_quiesce_counter;
2716
2717 /* Attach only after starting new drained sections, so that recursive
2718 * drain sections coming from @child don't get an extra .drained_begin
2719 * callback. */
2720 if (child->klass->attach) {
2721 child->klass->attach(child);
2722 }
2723 }
2724
2725 /*
2726 * If the old child node was drained but the new one is not, allow
2727 * requests to come in only after the new node has been attached.
2728 */
2729 while (drain_saldo < 0 && child->klass->drained_end) {
2730 bdrv_parent_drained_end_single(child);
2731 drain_saldo++;
2732 }
2733 }
2734
2735 static void bdrv_child_free(void *opaque)
2736 {
2737 BdrvChild *c = opaque;
2738
2739 g_free(c->name);
2740 g_free(c);
2741 }
2742
2743 static void bdrv_remove_empty_child(BdrvChild *child)
2744 {
2745 assert(!child->bs);
2746 QLIST_SAFE_REMOVE(child, next);
2747 bdrv_child_free(child);
2748 }
2749
2750 typedef struct BdrvAttachChildCommonState {
2751 BdrvChild **child;
2752 AioContext *old_parent_ctx;
2753 AioContext *old_child_ctx;
2754 } BdrvAttachChildCommonState;
2755
2756 static void bdrv_attach_child_common_abort(void *opaque)
2757 {
2758 BdrvAttachChildCommonState *s = opaque;
2759 BdrvChild *child = *s->child;
2760 BlockDriverState *bs = child->bs;
2761
2762 bdrv_replace_child_noperm(child, NULL);
2763
2764 if (bdrv_get_aio_context(bs) != s->old_child_ctx) {
2765 bdrv_try_set_aio_context(bs, s->old_child_ctx, &error_abort);
2766 }
2767
2768 if (bdrv_child_get_parent_aio_context(child) != s->old_parent_ctx) {
2769 GSList *ignore = g_slist_prepend(NULL, child);
2770
2771 child->klass->can_set_aio_ctx(child, s->old_parent_ctx, &ignore,
2772 &error_abort);
2773 g_slist_free(ignore);
2774 ignore = g_slist_prepend(NULL, child);
2775 child->klass->set_aio_ctx(child, s->old_parent_ctx, &ignore);
2776
2777 g_slist_free(ignore);
2778 }
2779
2780 bdrv_unref(bs);
2781 bdrv_remove_empty_child(child);
2782 *s->child = NULL;
2783 }
2784
2785 static TransactionActionDrv bdrv_attach_child_common_drv = {
2786 .abort = bdrv_attach_child_common_abort,
2787 .clean = g_free,
2788 };
2789
2790 /*
2791 * Common part of attaching bdrv child to bs or to blk or to job
2792 *
2793 * Resulting new child is returned through @child.
2794 * At start *@child must be NULL.
2795 * @child is saved to a new entry of @tran, so that *@child could be reverted to
2796 * NULL on abort(). So referenced variable must live at least until transaction
2797 * end.
2798 *
2799 * Function doesn't update permissions, caller is responsible for this.
2800 */
2801 static int bdrv_attach_child_common(BlockDriverState *child_bs,
2802 const char *child_name,
2803 const BdrvChildClass *child_class,
2804 BdrvChildRole child_role,
2805 uint64_t perm, uint64_t shared_perm,
2806 void *opaque, BdrvChild **child,
2807 Transaction *tran, Error **errp)
2808 {
2809 BdrvChild *new_child;
2810 AioContext *parent_ctx;
2811 AioContext *child_ctx = bdrv_get_aio_context(child_bs);
2812
2813 assert(child);
2814 assert(*child == NULL);
2815 assert(child_class->get_parent_desc);
2816
2817 new_child = g_new(BdrvChild, 1);
2818 *new_child = (BdrvChild) {
2819 .bs = NULL,
2820 .name = g_strdup(child_name),
2821 .klass = child_class,
2822 .role = child_role,
2823 .perm = perm,
2824 .shared_perm = shared_perm,
2825 .opaque = opaque,
2826 };
2827
2828 /*
2829 * If the AioContexts don't match, first try to move the subtree of
2830 * child_bs into the AioContext of the new parent. If this doesn't work,
2831 * try moving the parent into the AioContext of child_bs instead.
2832 */
2833 parent_ctx = bdrv_child_get_parent_aio_context(new_child);
2834 if (child_ctx != parent_ctx) {
2835 Error *local_err = NULL;
2836 int ret = bdrv_try_set_aio_context(child_bs, parent_ctx, &local_err);
2837
2838 if (ret < 0 && child_class->can_set_aio_ctx) {
2839 GSList *ignore = g_slist_prepend(NULL, new_child);
2840 if (child_class->can_set_aio_ctx(new_child, child_ctx, &ignore,
2841 NULL))
2842 {
2843 error_free(local_err);
2844 ret = 0;
2845 g_slist_free(ignore);
2846 ignore = g_slist_prepend(NULL, new_child);
2847 child_class->set_aio_ctx(new_child, child_ctx, &ignore);
2848 }
2849 g_slist_free(ignore);
2850 }
2851
2852 if (ret < 0) {
2853 error_propagate(errp, local_err);
2854 bdrv_remove_empty_child(new_child);
2855 return ret;
2856 }
2857 }
2858
2859 bdrv_ref(child_bs);
2860 bdrv_replace_child_noperm(new_child, child_bs);
2861
2862 *child = new_child;
2863
2864 BdrvAttachChildCommonState *s = g_new(BdrvAttachChildCommonState, 1);
2865 *s = (BdrvAttachChildCommonState) {
2866 .child = child,
2867 .old_parent_ctx = parent_ctx,
2868 .old_child_ctx = child_ctx,
2869 };
2870 tran_add(tran, &bdrv_attach_child_common_drv, s);
2871
2872 return 0;
2873 }
2874
2875 /*
2876 * Variable referenced by @child must live at least until transaction end.
2877 * (see bdrv_attach_child_common() doc for details)
2878 *
2879 * Function doesn't update permissions, caller is responsible for this.
2880 */
2881 static int bdrv_attach_child_noperm(BlockDriverState *parent_bs,
2882 BlockDriverState *child_bs,
2883 const char *child_name,
2884 const BdrvChildClass *child_class,
2885 BdrvChildRole child_role,
2886 BdrvChild **child,
2887 Transaction *tran,
2888 Error **errp)
2889 {
2890 int ret;
2891 uint64_t perm, shared_perm;
2892
2893 assert(parent_bs->drv);
2894
2895 bdrv_get_cumulative_perm(parent_bs, &perm, &shared_perm);
2896 bdrv_child_perm(parent_bs, child_bs, NULL, child_role, NULL,
2897 perm, shared_perm, &perm, &shared_perm);
2898
2899 ret = bdrv_attach_child_common(child_bs, child_name, child_class,
2900 child_role, perm, shared_perm, parent_bs,
2901 child, tran, errp);
2902 if (ret < 0) {
2903 return ret;
2904 }
2905
2906 QLIST_INSERT_HEAD(&parent_bs->children, *child, next);
2907 /*
2908 * child is removed in bdrv_attach_child_common_abort(), so don't care to
2909 * abort this change separately.
2910 */
2911
2912 return 0;
2913 }
2914
2915 static void bdrv_detach_child(BdrvChild *child)
2916 {
2917 BlockDriverState *old_bs = child->bs;
2918
2919 bdrv_replace_child_noperm(child, NULL);
2920 bdrv_remove_empty_child(child);
2921
2922 if (old_bs) {
2923 /*
2924 * Update permissions for old node. We're just taking a parent away, so
2925 * we're loosening restrictions. Errors of permission update are not
2926 * fatal in this case, ignore them.
2927 */
2928 bdrv_refresh_perms(old_bs, NULL);
2929
2930 /*
2931 * When the parent requiring a non-default AioContext is removed, the
2932 * node moves back to the main AioContext
2933 */
2934 bdrv_try_set_aio_context(old_bs, qemu_get_aio_context(), NULL);
2935 }
2936 }
2937
2938 /*
2939 * This function steals the reference to child_bs from the caller.
2940 * That reference is later dropped by bdrv_root_unref_child().
2941 *
2942 * On failure NULL is returned, errp is set and the reference to
2943 * child_bs is also dropped.
2944 *
2945 * The caller must hold the AioContext lock @child_bs, but not that of @ctx
2946 * (unless @child_bs is already in @ctx).
2947 */
2948 BdrvChild *bdrv_root_attach_child(BlockDriverState *child_bs,
2949 const char *child_name,
2950 const BdrvChildClass *child_class,
2951 BdrvChildRole child_role,
2952 uint64_t perm, uint64_t shared_perm,
2953 void *opaque, Error **errp)
2954 {
2955 int ret;
2956 BdrvChild *child = NULL;
2957 Transaction *tran = tran_new();
2958
2959 ret = bdrv_attach_child_common(child_bs, child_name, child_class,
2960 child_role, perm, shared_perm, opaque,
2961 &child, tran, errp);
2962 if (ret < 0) {
2963 goto out;
2964 }
2965
2966 ret = bdrv_refresh_perms(child_bs, errp);
2967
2968 out:
2969 tran_finalize(tran, ret);
2970 /* child is unset on failure by bdrv_attach_child_common_abort() */
2971 assert((ret < 0) == !child);
2972
2973 bdrv_unref(child_bs);
2974 return child;
2975 }
2976
2977 /*
2978 * This function transfers the reference to child_bs from the caller
2979 * to parent_bs. That reference is later dropped by parent_bs on
2980 * bdrv_close() or if someone calls bdrv_unref_child().
2981 *
2982 * On failure NULL is returned, errp is set and the reference to
2983 * child_bs is also dropped.
2984 *
2985 * If @parent_bs and @child_bs are in different AioContexts, the caller must
2986 * hold the AioContext lock for @child_bs, but not for @parent_bs.
2987 */
2988 BdrvChild *bdrv_attach_child(BlockDriverState *parent_bs,
2989 BlockDriverState *child_bs,
2990 const char *child_name,
2991 const BdrvChildClass *child_class,
2992 BdrvChildRole child_role,
2993 Error **errp)
2994 {
2995 int ret;
2996 BdrvChild *child = NULL;
2997 Transaction *tran = tran_new();
2998
2999 ret = bdrv_attach_child_noperm(parent_bs, child_bs, child_name, child_class,
3000 child_role, &child, tran, errp);
3001 if (ret < 0) {
3002 goto out;
3003 }
3004
3005 ret = bdrv_refresh_perms(parent_bs, errp);
3006 if (ret < 0) {
3007 goto out;
3008 }
3009
3010 out:
3011 tran_finalize(tran, ret);
3012 /* child is unset on failure by bdrv_attach_child_common_abort() */
3013 assert((ret < 0) == !child);
3014
3015 bdrv_unref(child_bs);
3016
3017 return child;
3018 }
3019
3020 /* Callers must ensure that child->frozen is false. */
3021 void bdrv_root_unref_child(BdrvChild *child)
3022 {
3023 BlockDriverState *child_bs;
3024
3025 child_bs = child->bs;
3026 bdrv_detach_child(child);
3027 bdrv_unref(child_bs);
3028 }
3029
3030 typedef struct BdrvSetInheritsFrom {
3031 BlockDriverState *bs;
3032 BlockDriverState *old_inherits_from;
3033 } BdrvSetInheritsFrom;
3034
3035 static void bdrv_set_inherits_from_abort(void *opaque)
3036 {
3037 BdrvSetInheritsFrom *s = opaque;
3038
3039 s->bs->inherits_from = s->old_inherits_from;
3040 }
3041
3042 static TransactionActionDrv bdrv_set_inherits_from_drv = {
3043 .abort = bdrv_set_inherits_from_abort,
3044 .clean = g_free,
3045 };
3046
3047 /* @tran is allowed to be NULL. In this case no rollback is possible */
3048 static void bdrv_set_inherits_from(BlockDriverState *bs,
3049 BlockDriverState *new_inherits_from,
3050 Transaction *tran)
3051 {
3052 if (tran) {
3053 BdrvSetInheritsFrom *s = g_new(BdrvSetInheritsFrom, 1);
3054
3055 *s = (BdrvSetInheritsFrom) {
3056 .bs = bs,
3057 .old_inherits_from = bs->inherits_from,
3058 };
3059
3060 tran_add(tran, &bdrv_set_inherits_from_drv, s);
3061 }
3062
3063 bs->inherits_from = new_inherits_from;
3064 }
3065
3066 /**
3067 * Clear all inherits_from pointers from children and grandchildren of
3068 * @root that point to @root, where necessary.
3069 * @tran is allowed to be NULL. In this case no rollback is possible
3070 */
3071 static void bdrv_unset_inherits_from(BlockDriverState *root, BdrvChild *child,
3072 Transaction *tran)
3073 {
3074 BdrvChild *c;
3075
3076 if (child->bs->inherits_from == root) {
3077 /*
3078 * Remove inherits_from only when the last reference between root and
3079 * child->bs goes away.
3080 */
3081 QLIST_FOREACH(c, &root->children, next) {
3082 if (c != child && c->bs == child->bs) {
3083 break;
3084 }
3085 }
3086 if (c == NULL) {
3087 bdrv_set_inherits_from(child->bs, NULL, tran);
3088 }
3089 }
3090
3091 QLIST_FOREACH(c, &child->bs->children, next) {
3092 bdrv_unset_inherits_from(root, c, tran);
3093 }
3094 }
3095
3096 /* Callers must ensure that child->frozen is false. */
3097 void bdrv_unref_child(BlockDriverState *parent, BdrvChild *child)
3098 {
3099 if (child == NULL) {
3100 return;
3101 }
3102
3103 bdrv_unset_inherits_from(parent, child, NULL);
3104 bdrv_root_unref_child(child);
3105 }
3106
3107
3108 static void bdrv_parent_cb_change_media(BlockDriverState *bs, bool load)
3109 {
3110 BdrvChild *c;
3111 QLIST_FOREACH(c, &bs->parents, next_parent) {
3112 if (c->klass->change_media) {
3113 c->klass->change_media(c, load);
3114 }
3115 }
3116 }
3117
3118 /* Return true if you can reach parent going through child->inherits_from
3119 * recursively. If parent or child are NULL, return false */
3120 static bool bdrv_inherits_from_recursive(BlockDriverState *child,
3121 BlockDriverState *parent)
3122 {
3123 while (child && child != parent) {
3124 child = child->inherits_from;
3125 }
3126
3127 return child != NULL;
3128 }
3129
3130 /*
3131 * Return the BdrvChildRole for @bs's backing child. bs->backing is
3132 * mostly used for COW backing children (role = COW), but also for
3133 * filtered children (role = FILTERED | PRIMARY).
3134 */
3135 static BdrvChildRole bdrv_backing_role(BlockDriverState *bs)
3136 {
3137 if (bs->drv && bs->drv->is_filter) {
3138 return BDRV_CHILD_FILTERED | BDRV_CHILD_PRIMARY;
3139 } else {
3140 return BDRV_CHILD_COW;
3141 }
3142 }
3143
3144 /*
3145 * Sets the bs->backing or bs->file link of a BDS. A new reference is created;
3146 * callers which don't need their own reference any more must call bdrv_unref().
3147 *
3148 * Function doesn't update permissions, caller is responsible for this.
3149 */
3150 static int bdrv_set_file_or_backing_noperm(BlockDriverState *parent_bs,
3151 BlockDriverState *child_bs,
3152 bool is_backing,
3153 Transaction *tran, Error **errp)
3154 {
3155 int ret = 0;
3156 bool update_inherits_from =
3157 bdrv_inherits_from_recursive(child_bs, parent_bs);
3158 BdrvChild *child = is_backing ? parent_bs->backing : parent_bs->file;
3159 BdrvChildRole role;
3160
3161 if (!parent_bs->drv) {
3162 /*
3163 * Node without drv is an object without a class :/. TODO: finally fix
3164 * qcow2 driver to never clear bs->drv and implement format corruption
3165 * handling in other way.
3166 */
3167 error_setg(errp, "Node corrupted");
3168 return -EINVAL;
3169 }
3170
3171 if (child && child->frozen) {
3172 error_setg(errp, "Cannot change frozen '%s' link from '%s' to '%s'",
3173 child->name, parent_bs->node_name, child->bs->node_name);
3174 return -EPERM;
3175 }
3176
3177 if (is_backing && !parent_bs->drv->is_filter &&
3178 !parent_bs->drv->supports_backing)
3179 {
3180 error_setg(errp, "Driver '%s' of node '%s' does not support backing "
3181 "files", parent_bs->drv->format_name, parent_bs->node_name);
3182 return -EINVAL;
3183 }
3184
3185 if (parent_bs->drv->is_filter) {
3186 role = BDRV_CHILD_FILTERED | BDRV_CHILD_PRIMARY;
3187 } else if (is_backing) {
3188 role = BDRV_CHILD_COW;
3189 } else {
3190 /*
3191 * We only can use same role as it is in existing child. We don't have
3192 * infrastructure to determine role of file child in generic way
3193 */
3194 if (!child) {
3195 error_setg(errp, "Cannot set file child to format node without "
3196 "file child");
3197 return -EINVAL;
3198 }
3199 role = child->role;
3200 }
3201
3202 if (child) {
3203 bdrv_unset_inherits_from(parent_bs, child, tran);
3204 bdrv_remove_file_or_backing_child(parent_bs, child, tran);
3205 }
3206
3207 if (!child_bs) {
3208 goto out;
3209 }
3210
3211 ret = bdrv_attach_child_noperm(parent_bs, child_bs,
3212 is_backing ? "backing" : "file",
3213 &child_of_bds, role,
3214 is_backing ? &parent_bs->backing :
3215 &parent_bs->file,
3216 tran, errp);
3217 if (ret < 0) {
3218 return ret;
3219 }
3220
3221
3222 /*
3223 * If inherits_from pointed recursively to bs then let's update it to
3224 * point directly to bs (else it will become NULL).
3225 */
3226 if (update_inherits_from) {
3227 bdrv_set_inherits_from(child_bs, parent_bs, tran);
3228 }
3229
3230 out:
3231 bdrv_refresh_limits(parent_bs, tran, NULL);
3232
3233 return 0;
3234 }
3235
3236 static int bdrv_set_backing_noperm(BlockDriverState *bs,
3237 BlockDriverState *backing_hd,
3238 Transaction *tran, Error **errp)
3239 {
3240 return bdrv_set_file_or_backing_noperm(bs, backing_hd, true, tran, errp);
3241 }
3242
3243 int bdrv_set_backing_hd(BlockDriverState *bs, BlockDriverState *backing_hd,
3244 Error **errp)
3245 {
3246 int ret;
3247 Transaction *tran = tran_new();
3248
3249 ret = bdrv_set_backing_noperm(bs, backing_hd, tran, errp);
3250 if (ret < 0) {
3251 goto out;
3252 }
3253
3254 ret = bdrv_refresh_perms(bs, errp);
3255 out:
3256 tran_finalize(tran, ret);
3257
3258 return ret;
3259 }
3260
3261 /*
3262 * Opens the backing file for a BlockDriverState if not yet open
3263 *
3264 * bdref_key specifies the key for the image's BlockdevRef in the options QDict.
3265 * That QDict has to be flattened; therefore, if the BlockdevRef is a QDict
3266 * itself, all options starting with "${bdref_key}." are considered part of the
3267 * BlockdevRef.
3268 *
3269 * TODO Can this be unified with bdrv_open_image()?
3270 */
3271 int bdrv_open_backing_file(BlockDriverState *bs, QDict *parent_options,
3272 const char *bdref_key, Error **errp)
3273 {
3274 char *backing_filename = NULL;
3275 char *bdref_key_dot;
3276 const char *reference = NULL;
3277 int ret = 0;
3278 bool implicit_backing = false;
3279 BlockDriverState *backing_hd;
3280 QDict *options;
3281 QDict *tmp_parent_options = NULL;
3282 Error *local_err = NULL;
3283
3284 if (bs->backing != NULL) {
3285 goto free_exit;
3286 }
3287
3288 /* NULL means an empty set of options */
3289 if (parent_options == NULL) {
3290 tmp_parent_options = qdict_new();
3291 parent_options = tmp_parent_options;
3292 }
3293
3294 bs->open_flags &= ~BDRV_O_NO_BACKING;
3295
3296 bdref_key_dot = g_strdup_printf("%s.", bdref_key);
3297 qdict_extract_subqdict(parent_options, &options, bdref_key_dot);
3298 g_free(bdref_key_dot);
3299
3300 /*
3301 * Caution: while qdict_get_try_str() is fine, getting non-string
3302 * types would require more care. When @parent_options come from
3303 * -blockdev or blockdev_add, its members are typed according to
3304 * the QAPI schema, but when they come from -drive, they're all
3305 * QString.
3306 */
3307 reference = qdict_get_try_str(parent_options, bdref_key);
3308 if (reference || qdict_haskey(options, "file.filename")) {
3309 /* keep backing_filename NULL */
3310 } else if (bs->backing_file[0] == '\0' && qdict_size(options) == 0) {
3311 qobject_unref(options);
3312 goto free_exit;
3313 } else {
3314 if (qdict_size(options) == 0) {
3315 /* If the user specifies options that do not modify the
3316 * backing file's behavior, we might still consider it the
3317 * implicit backing file. But it's easier this way, and
3318 * just specifying some of the backing BDS's options is
3319 * only possible with -drive anyway (otherwise the QAPI
3320 * schema forces the user to specify everything). */
3321 implicit_backing = !strcmp(bs->auto_backing_file, bs->backing_file);
3322 }
3323
3324 backing_filename = bdrv_get_full_backing_filename(bs, &local_err);
3325 if (local_err) {
3326 ret = -EINVAL;
3327 error_propagate(errp, local_err);
3328 qobject_unref(options);
3329 goto free_exit;
3330 }
3331 }
3332
3333 if (!bs->drv || !bs->drv->supports_backing) {
3334 ret = -EINVAL;
3335 error_setg(errp, "Driver doesn't support backing files");
3336 qobject_unref(options);
3337 goto free_exit;
3338 }
3339
3340 if (!reference &&
3341 bs->backing_format[0] != '\0' && !qdict_haskey(options, "driver")) {
3342 qdict_put_str(options, "driver", bs->backing_format);
3343 }
3344
3345 backing_hd = bdrv_open_inherit(backing_filename, reference, options, 0, bs,
3346 &child_of_bds, bdrv_backing_role(bs), errp);
3347 if (!backing_hd) {
3348 bs->open_flags |= BDRV_O_NO_BACKING;
3349 error_prepend(errp, "Could not open backing file: ");
3350 ret = -EINVAL;
3351 goto free_exit;
3352 }
3353
3354 if (implicit_backing) {
3355 bdrv_refresh_filename(backing_hd);
3356 pstrcpy(bs->auto_backing_file, sizeof(bs->auto_backing_file),
3357 backing_hd->filename);
3358 }
3359
3360 /* Hook up the backing file link; drop our reference, bs owns the
3361 * backing_hd reference now */
3362 ret = bdrv_set_backing_hd(bs, backing_hd, errp);
3363 bdrv_unref(backing_hd);
3364 if (ret < 0) {
3365 goto free_exit;
3366 }
3367
3368 qdict_del(parent_options, bdref_key);
3369
3370 free_exit:
3371 g_free(backing_filename);
3372 qobject_unref(tmp_parent_options);
3373 return ret;
3374 }
3375
3376 static BlockDriverState *
3377 bdrv_open_child_bs(const char *filename, QDict *options, const char *bdref_key,
3378 BlockDriverState *parent, const BdrvChildClass *child_class,
3379 BdrvChildRole child_role, bool allow_none, Error **errp)
3380 {
3381 BlockDriverState *bs = NULL;
3382 QDict *image_options;
3383 char *bdref_key_dot;
3384 const char *reference;
3385
3386 assert(child_class != NULL);
3387
3388 bdref_key_dot = g_strdup_printf("%s.", bdref_key);
3389 qdict_extract_subqdict(options, &image_options, bdref_key_dot);
3390 g_free(bdref_key_dot);
3391
3392 /*
3393 * Caution: while qdict_get_try_str() is fine, getting non-string
3394 * types would require more care. When @options come from
3395 * -blockdev or blockdev_add, its members are typed according to
3396 * the QAPI schema, but when they come from -drive, they're all
3397 * QString.
3398 */
3399 reference = qdict_get_try_str(options, bdref_key);
3400 if (!filename && !reference && !qdict_size(image_options)) {
3401 if (!allow_none) {
3402 error_setg(errp, "A block device must be specified for \"%s\"",
3403 bdref_key);
3404 }
3405 qobject_unref(image_options);
3406 goto done;
3407 }
3408
3409 bs = bdrv_open_inherit(filename, reference, image_options, 0,
3410 parent, child_class, child_role, errp);
3411 if (!bs) {
3412 goto done;
3413 }
3414
3415 done:
3416 qdict_del(options, bdref_key);
3417 return bs;
3418 }
3419
3420 /*
3421 * Opens a disk image whose options are given as BlockdevRef in another block
3422 * device's options.
3423 *
3424 * If allow_none is true, no image will be opened if filename is false and no
3425 * BlockdevRef is given. NULL will be returned, but errp remains unset.
3426 *
3427 * bdrev_key specifies the key for the image's BlockdevRef in the options QDict.
3428 * That QDict has to be flattened; therefore, if the BlockdevRef is a QDict
3429 * itself, all options starting with "${bdref_key}." are considered part of the
3430 * BlockdevRef.
3431 *
3432 * The BlockdevRef will be removed from the options QDict.
3433 */
3434 BdrvChild *bdrv_open_child(const char *filename,
3435 QDict *options, const char *bdref_key,
3436 BlockDriverState *parent,
3437 const BdrvChildClass *child_class,
3438 BdrvChildRole child_role,
3439 bool allow_none, Error **errp)
3440 {
3441 BlockDriverState *bs;
3442
3443 bs = bdrv_open_child_bs(filename, options, bdref_key, parent, child_class,
3444 child_role, allow_none, errp);
3445 if (bs == NULL) {
3446 return NULL;
3447 }
3448
3449 return bdrv_attach_child(parent, bs, bdref_key, child_class, child_role,
3450 errp);
3451 }
3452
3453 /*
3454 * TODO Future callers may need to specify parent/child_class in order for
3455 * option inheritance to work. Existing callers use it for the root node.
3456 */
3457 BlockDriverState *bdrv_open_blockdev_ref(BlockdevRef *ref, Error **errp)
3458 {
3459 BlockDriverState *bs = NULL;
3460 QObject *obj = NULL;
3461 QDict *qdict = NULL;
3462 const char *reference = NULL;
3463 Visitor *v = NULL;
3464
3465 if (ref->type == QTYPE_QSTRING) {
3466 reference = ref->u.reference;
3467 } else {
3468 BlockdevOptions *options = &ref->u.definition;
3469 assert(ref->type == QTYPE_QDICT);
3470
3471 v = qobject_output_visitor_new(&obj);
3472 visit_type_BlockdevOptions(v, NULL, &options, &error_abort);
3473 visit_complete(v, &obj);
3474
3475 qdict = qobject_to(QDict, obj);
3476 qdict_flatten(qdict);
3477
3478 /* bdrv_open_inherit() defaults to the values in bdrv_flags (for
3479 * compatibility with other callers) rather than what we want as the
3480 * real defaults. Apply the defaults here instead. */
3481 qdict_set_default_str(qdict, BDRV_OPT_CACHE_DIRECT, "off");
3482 qdict_set_default_str(qdict, BDRV_OPT_CACHE_NO_FLUSH, "off");
3483 qdict_set_default_str(qdict, BDRV_OPT_READ_ONLY, "off");
3484 qdict_set_default_str(qdict, BDRV_OPT_AUTO_READ_ONLY, "off");
3485
3486 }
3487
3488 bs = bdrv_open_inherit(NULL, reference, qdict, 0, NULL, NULL, 0, errp);
3489 obj = NULL;
3490 qobject_unref(obj);
3491 visit_free(v);
3492 return bs;
3493 }
3494
3495 static BlockDriverState *bdrv_append_temp_snapshot(BlockDriverState *bs,
3496 int flags,
3497 QDict *snapshot_options,
3498 Error **errp)
3499 {
3500 /* TODO: extra byte is a hack to ensure MAX_PATH space on Windows. */
3501 char *tmp_filename = g_malloc0(PATH_MAX + 1);
3502 int64_t total_size;
3503 QemuOpts *opts = NULL;
3504 BlockDriverState *bs_snapshot = NULL;
3505 int ret;
3506
3507 /* if snapshot, we create a temporary backing file and open it
3508 instead of opening 'filename' directly */
3509
3510 /* Get the required size from the image */
3511 total_size = bdrv_getlength(bs);
3512 if (total_size < 0) {
3513 error_setg_errno(errp, -total_size, "Could not get image size");
3514 goto out;
3515 }
3516
3517 /* Create the temporary image */
3518 ret = get_tmp_filename(tmp_filename, PATH_MAX + 1);
3519 if (ret < 0) {
3520 error_setg_errno(errp, -ret, "Could not get temporary filename");
3521 goto out;
3522 }
3523
3524 opts = qemu_opts_create(bdrv_qcow2.create_opts, NULL, 0,
3525 &error_abort);
3526 qemu_opt_set_number(opts, BLOCK_OPT_SIZE, total_size, &error_abort);
3527 ret = bdrv_create(&bdrv_qcow2, tmp_filename, opts, errp);
3528 qemu_opts_del(opts);
3529 if (ret < 0) {
3530 error_prepend(errp, "Could not create temporary overlay '%s': ",
3531 tmp_filename);
3532 goto out;
3533 }
3534
3535 /* Prepare options QDict for the temporary file */
3536 qdict_put_str(snapshot_options, "file.driver", "file");
3537 qdict_put_str(snapshot_options, "file.filename", tmp_filename);
3538 qdict_put_str(snapshot_options, "driver", "qcow2");
3539
3540 bs_snapshot = bdrv_open(NULL, NULL, snapshot_options, flags, errp);
3541 snapshot_options = NULL;
3542 if (!bs_snapshot) {
3543 goto out;
3544 }
3545
3546 ret = bdrv_append(bs_snapshot, bs, errp);
3547 if (ret < 0) {
3548 bs_snapshot = NULL;
3549 goto out;
3550 }
3551
3552 out:
3553 qobject_unref(snapshot_options);
3554 g_free(tmp_filename);
3555 return bs_snapshot;
3556 }
3557
3558 /*
3559 * Opens a disk image (raw, qcow2, vmdk, ...)
3560 *
3561 * options is a QDict of options to pass to the block drivers, or NULL for an
3562 * empty set of options. The reference to the QDict belongs to the block layer
3563 * after the call (even on failure), so if the caller intends to reuse the
3564 * dictionary, it needs to use qobject_ref() before calling bdrv_open.
3565 *
3566 * If *pbs is NULL, a new BDS will be created with a pointer to it stored there.
3567 * If it is not NULL, the referenced BDS will be reused.
3568 *
3569 * The reference parameter may be used to specify an existing block device which
3570 * should be opened. If specified, neither options nor a filename may be given,
3571 * nor can an existing BDS be reused (that is, *pbs has to be NULL).
3572 */
3573 static BlockDriverState *bdrv_open_inherit(const char *filename,
3574 const char *reference,
3575 QDict *options, int flags,
3576 BlockDriverState *parent,
3577 const BdrvChildClass *child_class,
3578 BdrvChildRole child_role,
3579 Error **errp)
3580 {
3581 int ret;
3582 BlockBackend *file = NULL;
3583 BlockDriverState *bs;
3584 BlockDriver *drv = NULL;
3585 BdrvChild *child;
3586 const char *drvname;
3587 const char *backing;
3588 Error *local_err = NULL;
3589 QDict *snapshot_options = NULL;
3590 int snapshot_flags = 0;
3591
3592 assert(!child_class || !flags);
3593 assert(!child_class == !parent);
3594
3595 if (reference) {
3596 bool options_non_empty = options ? qdict_size(options) : false;
3597 qobject_unref(options);
3598
3599 if (filename || options_non_empty) {
3600 error_setg(errp, "Cannot reference an existing block device with "
3601 "additional options or a new filename");
3602 return NULL;
3603 }
3604
3605 bs = bdrv_lookup_bs(reference, reference, errp);
3606 if (!bs) {
3607 return NULL;
3608 }
3609
3610 bdrv_ref(bs);
3611 return bs;
3612 }
3613
3614 bs = bdrv_new();
3615
3616 /* NULL means an empty set of options */
3617 if (options == NULL) {
3618 options = qdict_new();
3619 }
3620
3621 /* json: syntax counts as explicit options, as if in the QDict */
3622 parse_json_protocol(options, &filename, &local_err);
3623 if (local_err) {
3624 goto fail;
3625 }
3626
3627 bs->explicit_options = qdict_clone_shallow(options);
3628
3629 if (child_class) {
3630 bool parent_is_format;
3631
3632 if (parent->drv) {
3633 parent_is_format = parent->drv->is_format;
3634 } else {
3635 /*
3636 * parent->drv is not set yet because this node is opened for
3637 * (potential) format probing. That means that @parent is going
3638 * to be a format node.
3639 */
3640 parent_is_format = true;
3641 }
3642
3643 bs->inherits_from = parent;
3644 child_class->inherit_options(child_role, parent_is_format,
3645 &flags, options,
3646 parent->open_flags, parent->options);
3647 }
3648
3649 ret = bdrv_fill_options(&options, filename, &flags, &local_err);
3650 if (ret < 0) {
3651 goto fail;
3652 }
3653
3654 /*
3655 * Set the BDRV_O_RDWR and BDRV_O_ALLOW_RDWR flags.
3656 * Caution: getting a boolean member of @options requires care.
3657 * When @options come from -blockdev or blockdev_add, members are
3658 * typed according to the QAPI schema, but when they come from
3659 * -drive, they're all QString.
3660 */
3661 if (g_strcmp0(qdict_get_try_str(options, BDRV_OPT_READ_ONLY), "on") &&
3662 !qdict_get_try_bool(options, BDRV_OPT_READ_ONLY, false)) {
3663 flags |= (BDRV_O_RDWR | BDRV_O_ALLOW_RDWR);
3664 } else {
3665 flags &= ~BDRV_O_RDWR;
3666 }
3667
3668 if (flags & BDRV_O_SNAPSHOT) {
3669 snapshot_options = qdict_new();
3670 bdrv_temp_snapshot_options(&snapshot_flags, snapshot_options,
3671 flags, options);
3672 /* Let bdrv_backing_options() override "read-only" */
3673 qdict_del(options, BDRV_OPT_READ_ONLY);
3674 bdrv_inherited_options(BDRV_CHILD_COW, true,
3675 &flags, options, flags, options);
3676 }
3677
3678 bs->open_flags = flags;
3679 bs->options = options;
3680 options = qdict_clone_shallow(options);
3681
3682 /* Find the right image format driver */
3683 /* See cautionary note on accessing @options above */
3684 drvname = qdict_get_try_str(options, "driver");
3685 if (drvname) {
3686 drv = bdrv_find_format(drvname);
3687 if (!drv) {
3688 error_setg(errp, "Unknown driver: '%s'", drvname);
3689 goto fail;
3690 }
3691 }
3692
3693 assert(drvname || !(flags & BDRV_O_PROTOCOL));
3694
3695 /* See cautionary note on accessing @options above */
3696 backing = qdict_get_try_str(options, "backing");
3697 if (qobject_to(QNull, qdict_get(options, "backing")) != NULL ||
3698 (backing && *backing == '\0'))
3699 {
3700 if (backing) {
3701 warn_report("Use of \"backing\": \"\" is deprecated; "
3702 "use \"backing\": null instead");
3703 }
3704 flags |= BDRV_O_NO_BACKING;
3705 qdict_del(bs->explicit_options, "backing");
3706 qdict_del(bs->options, "backing");
3707 qdict_del(options, "backing");
3708 }
3709
3710 /* Open image file without format layer. This BlockBackend is only used for
3711 * probing, the block drivers will do their own bdrv_open_child() for the
3712 * same BDS, which is why we put the node name back into options. */
3713 if ((flags & BDRV_O_PROTOCOL) == 0) {
3714 BlockDriverState *file_bs;
3715
3716 file_bs = bdrv_open_child_bs(filename, options, "file", bs,
3717 &child_of_bds, BDRV_CHILD_IMAGE,
3718 true, &local_err);
3719 if (local_err) {
3720 goto fail;
3721 }
3722 if (file_bs != NULL) {
3723 /* Not requesting BLK_PERM_CONSISTENT_READ because we're only
3724 * looking at the header to guess the image format. This works even
3725 * in cases where a guest would not see a consistent state. */
3726 file = blk_new(bdrv_get_aio_context(file_bs), 0, BLK_PERM_ALL);
3727 blk_insert_bs(file, file_bs, &local_err);
3728 bdrv_unref(file_bs);
3729 if (local_err) {
3730 goto fail;
3731 }
3732
3733 qdict_put_str(options, "file", bdrv_get_node_name(file_bs));
3734 }
3735 }
3736
3737 /* Image format probing */
3738 bs->probed = !drv;
3739 if (!drv && file) {
3740 ret = find_image_format(file, filename, &drv, &local_err);
3741 if (ret < 0) {
3742 goto fail;
3743 }
3744 /*
3745 * This option update would logically belong in bdrv_fill_options(),
3746 * but we first need to open bs->file for the probing to work, while
3747 * opening bs->file already requires the (mostly) final set of options
3748 * so that cache mode etc. can be inherited.
3749 *
3750 * Adding the driver later is somewhat ugly, but it's not an option
3751 * that would ever be inherited, so it's correct. We just need to make
3752 * sure to update both bs->options (which has the full effective
3753 * options for bs) and options (which has file.* already removed).
3754 */
3755 qdict_put_str(bs->options, "driver", drv->format_name);
3756 qdict_put_str(options, "driver", drv->format_name);
3757 } else if (!drv) {
3758 error_setg(errp, "Must specify either driver or file");
3759 goto fail;
3760 }
3761
3762 /* BDRV_O_PROTOCOL must be set iff a protocol BDS is about to be created */
3763 assert(!!(flags & BDRV_O_PROTOCOL) == !!drv->bdrv_file_open);
3764 /* file must be NULL if a protocol BDS is about to be created
3765 * (the inverse results in an error message from bdrv_open_common()) */
3766 assert(!(flags & BDRV_O_PROTOCOL) || !file);
3767
3768 /* Open the image */
3769 ret = bdrv_open_common(bs, file, options, &local_err);
3770 if (ret < 0) {
3771 goto fail;
3772 }
3773
3774 if (file) {
3775 blk_unref(file);
3776 file = NULL;
3777 }
3778
3779 /* If there is a backing file, use it */
3780 if ((flags & BDRV_O_NO_BACKING) == 0) {
3781 ret = bdrv_open_backing_file(bs, options, "backing", &local_err);
3782 if (ret < 0) {
3783 goto close_and_fail;
3784 }
3785 }
3786
3787 /* Remove all children options and references
3788 * from bs->options and bs->explicit_options */
3789 QLIST_FOREACH(child, &bs->children, next) {
3790 char *child_key_dot;
3791 child_key_dot = g_strdup_printf("%s.", child->name);
3792 qdict_extract_subqdict(bs->explicit_options, NULL, child_key_dot);
3793 qdict_extract_subqdict(bs->options, NULL, child_key_dot);
3794 qdict_del(bs->explicit_options, child->name);
3795 qdict_del(bs->options, child->name);
3796 g_free(child_key_dot);
3797 }
3798
3799 /* Check if any unknown options were used */
3800 if (qdict_size(options) != 0) {
3801 const QDictEntry *entry = qdict_first(options);
3802 if (flags & BDRV_O_PROTOCOL) {
3803 error_setg(errp, "Block protocol '%s' doesn't support the option "
3804 "'%s'", drv->format_name, entry->key);
3805 } else {
3806 error_setg(errp,
3807 "Block format '%s' does not support the option '%s'",
3808 drv->format_name, entry->key);
3809 }
3810
3811 goto close_and_fail;
3812 }
3813
3814 bdrv_parent_cb_change_media(bs, true);
3815
3816 qobject_unref(options);
3817 options = NULL;
3818
3819 /* For snapshot=on, create a temporary qcow2 overlay. bs points to the
3820 * temporary snapshot afterwards. */
3821 if (snapshot_flags) {
3822 BlockDriverState *snapshot_bs;
3823 snapshot_bs = bdrv_append_temp_snapshot(bs, snapshot_flags,
3824 snapshot_options, &local_err);
3825 snapshot_options = NULL;
3826 if (local_err) {
3827 goto close_and_fail;
3828 }
3829 /* We are not going to return bs but the overlay on top of it
3830 * (snapshot_bs); thus, we have to drop the strong reference to bs
3831 * (which we obtained by calling bdrv_new()). bs will not be deleted,
3832 * though, because the overlay still has a reference to it. */
3833 bdrv_unref(bs);
3834 bs = snapshot_bs;
3835 }
3836
3837 return bs;
3838
3839 fail:
3840 blk_unref(file);
3841 qobject_unref(snapshot_options);
3842 qobject_unref(bs->explicit_options);
3843 qobject_unref(bs->options);
3844 qobject_unref(options);
3845 bs->options = NULL;
3846 bs->explicit_options = NULL;
3847 bdrv_unref(bs);
3848 error_propagate(errp, local_err);
3849 return NULL;
3850
3851 close_and_fail:
3852 bdrv_unref(bs);
3853 qobject_unref(snapshot_options);
3854 qobject_unref(options);
3855 error_propagate(errp, local_err);
3856 return NULL;
3857 }
3858
3859 BlockDriverState *bdrv_open(const char *filename, const char *reference,
3860 QDict *options, int flags, Error **errp)
3861 {
3862 return bdrv_open_inherit(filename, reference, options, flags, NULL,
3863 NULL, 0, errp);
3864 }
3865
3866 /* Return true if the NULL-terminated @list contains @str */
3867 static bool is_str_in_list(const char *str, const char *const *list)
3868 {
3869 if (str && list) {
3870 int i;
3871 for (i = 0; list[i] != NULL; i++) {
3872 if (!strcmp(str, list[i])) {
3873 return true;
3874 }
3875 }
3876 }
3877 return false;
3878 }
3879
3880 /*
3881 * Check that every option set in @bs->options is also set in
3882 * @new_opts.
3883 *
3884 * Options listed in the common_options list and in
3885 * @bs->drv->mutable_opts are skipped.
3886 *
3887 * Return 0 on success, otherwise return -EINVAL and set @errp.
3888 */
3889 static int bdrv_reset_options_allowed(BlockDriverState *bs,
3890 const QDict *new_opts, Error **errp)
3891 {
3892 const QDictEntry *e;
3893 /* These options are common to all block drivers and are handled
3894 * in bdrv_reopen_prepare() so they can be left out of @new_opts */
3895 const char *const common_options[] = {
3896 "node-name", "discard", "cache.direct", "cache.no-flush",
3897 "read-only", "auto-read-only", "detect-zeroes", NULL
3898 };
3899
3900 for (e = qdict_first(bs->options); e; e = qdict_next(bs->options, e)) {
3901 if (!qdict_haskey(new_opts, e->key) &&
3902 !is_str_in_list(e->key, common_options) &&
3903 !is_str_in_list(e->key, bs->drv->mutable_opts)) {
3904 error_setg(errp, "Option '%s' cannot be reset "
3905 "to its default value", e->key);
3906 return -EINVAL;
3907 }
3908 }
3909
3910 return 0;
3911 }
3912
3913 /*
3914 * Returns true if @child can be reached recursively from @bs
3915 */
3916 static bool bdrv_recurse_has_child(BlockDriverState *bs,
3917 BlockDriverState *child)
3918 {
3919 BdrvChild *c;
3920
3921 if (bs == child) {
3922 return true;
3923 }
3924
3925 QLIST_FOREACH(c, &bs->children, next) {
3926 if (bdrv_recurse_has_child(c->bs, child)) {
3927 return true;
3928 }
3929 }
3930
3931 return false;
3932 }
3933
3934 /*
3935 * Adds a BlockDriverState to a simple queue for an atomic, transactional
3936 * reopen of multiple devices.
3937 *
3938 * bs_queue can either be an existing BlockReopenQueue that has had QTAILQ_INIT
3939 * already performed, or alternatively may be NULL a new BlockReopenQueue will
3940 * be created and initialized. This newly created BlockReopenQueue should be
3941 * passed back in for subsequent calls that are intended to be of the same
3942 * atomic 'set'.
3943 *
3944 * bs is the BlockDriverState to add to the reopen queue.
3945 *
3946 * options contains the changed options for the associated bs
3947 * (the BlockReopenQueue takes ownership)
3948 *
3949 * flags contains the open flags for the associated bs
3950 *
3951 * returns a pointer to bs_queue, which is either the newly allocated
3952 * bs_queue, or the existing bs_queue being used.
3953 *
3954 * bs must be drained between bdrv_reopen_queue() and bdrv_reopen_multiple().
3955 */
3956 static BlockReopenQueue *bdrv_reopen_queue_child(BlockReopenQueue *bs_queue,
3957 BlockDriverState *bs,
3958 QDict *options,
3959 const BdrvChildClass *klass,
3960 BdrvChildRole role,
3961 bool parent_is_format,
3962 QDict *parent_options,
3963 int parent_flags,
3964 bool keep_old_opts)
3965 {
3966 assert(bs != NULL);
3967
3968 BlockReopenQueueEntry *bs_entry;
3969 BdrvChild *child;
3970 QDict *old_options, *explicit_options, *options_copy;
3971 int flags;
3972 QemuOpts *opts;
3973
3974 /* Make sure that the caller remembered to use a drained section. This is
3975 * important to avoid graph changes between the recursive queuing here and
3976 * bdrv_reopen_multiple(). */
3977 assert(bs->quiesce_counter > 0);
3978
3979 if (bs_queue == NULL) {
3980 bs_queue = g_new0(BlockReopenQueue, 1);
3981 QTAILQ_INIT(bs_queue);
3982 }
3983
3984 if (!options) {
3985 options = qdict_new();
3986 }
3987
3988 /* Check if this BlockDriverState is already in the queue */
3989 QTAILQ_FOREACH(bs_entry, bs_queue, entry) {
3990 if (bs == bs_entry->state.bs) {
3991 break;
3992 }
3993 }
3994
3995 /*
3996 * Precedence of options:
3997 * 1. Explicitly passed in options (highest)
3998 * 2. Retained from explicitly set options of bs
3999 * 3. Inherited from parent node
4000 * 4. Retained from effective options of bs
4001 */
4002
4003 /* Old explicitly set values (don't overwrite by inherited value) */
4004 if (bs_entry || keep_old_opts) {
4005 old_options = qdict_clone_shallow(bs_entry ?
4006 bs_entry->state.explicit_options :
4007 bs->explicit_options);
4008 bdrv_join_options(bs, options, old_options);
4009 qobject_unref(old_options);
4010 }
4011
4012 explicit_options = qdict_clone_shallow(options);
4013
4014 /* Inherit from parent node */
4015 if (parent_options) {
4016 flags = 0;
4017 klass->inherit_options(role, parent_is_format, &flags, options,
4018 parent_flags, parent_options);
4019 } else {
4020 flags = bdrv_get_flags(bs);
4021 }
4022
4023 if (keep_old_opts) {
4024 /* Old values are used for options that aren't set yet */
4025 old_options = qdict_clone_shallow(bs->options);
4026 bdrv_join_options(bs, options, old_options);
4027 qobject_unref(old_options);
4028 }
4029
4030 /* We have the final set of options so let's update the flags */
4031 options_copy = qdict_clone_shallow(options);
4032 opts = qemu_opts_create(&bdrv_runtime_opts, NULL, 0, &error_abort);
4033 qemu_opts_absorb_qdict(opts, options_copy, NULL);
4034 update_flags_from_options(&flags, opts);
4035 qemu_opts_del(opts);
4036 qobject_unref(options_copy);
4037
4038 /* bdrv_open_inherit() sets and clears some additional flags internally */
4039 flags &= ~BDRV_O_PROTOCOL;
4040 if (flags & BDRV_O_RDWR) {
4041 flags |= BDRV_O_ALLOW_RDWR;
4042 }
4043
4044 if (!bs_entry) {
4045 bs_entry = g_new0(BlockReopenQueueEntry, 1);
4046 QTAILQ_INSERT_TAIL(bs_queue, bs_entry, entry);
4047 } else {
4048 qobject_unref(bs_entry->state.options);
4049 qobject_unref(bs_entry->state.explicit_options);
4050 }
4051
4052 bs_entry->state.bs = bs;
4053 bs_entry->state.options = options;
4054 bs_entry->state.explicit_options = explicit_options;
4055 bs_entry->state.flags = flags;
4056
4057 /*
4058 * If keep_old_opts is false then it means that unspecified
4059 * options must be reset to their original value. We don't allow
4060 * resetting 'backing' but we need to know if the option is
4061 * missing in order to decide if we have to return an error.
4062 */
4063 if (!keep_old_opts) {
4064 bs_entry->state.backing_missing =
4065 !qdict_haskey(options, "backing") &&
4066 !qdict_haskey(options, "backing.driver");
4067 }
4068
4069 QLIST_FOREACH(child, &bs->children, next) {
4070 QDict *new_child_options = NULL;
4071 bool child_keep_old = keep_old_opts;
4072
4073 /* reopen can only change the options of block devices that were
4074 * implicitly created and inherited options. For other (referenced)
4075 * block devices, a syntax like "backing.foo" results in an error. */
4076 if (child->bs->inherits_from != bs) {
4077 continue;
4078 }
4079
4080 /* Check if the options contain a child reference */
4081 if (qdict_haskey(options, child->name)) {
4082 const char *childref = qdict_get_try_str(options, child->name);
4083