xref: /openbmc/qemu/block.c (revision 96796fae6f22931d91223d086e9fa56d0f3e6720)
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             /*
4084              * The current child must not be reopened if the child
4085              * reference is null or points to a different node.
4086              */
4087             if (g_strcmp0(childref, child->bs->node_name)) {
4088                 continue;
4089             }
4090             /*
4091              * If the child reference points to the current child then
4092              * reopen it with its existing set of options (note that
4093              * it can still inherit new options from the parent).
4094              */
4095             child_keep_old = true;
4096         } else {
4097             /* Extract child options ("child-name.*") */
4098             char *child_key_dot = g_strdup_printf("%s.", child->name);
4099             qdict_extract_subqdict(explicit_options, NULL, child_key_dot);
4100             qdict_extract_subqdict(options, &new_child_options, child_key_dot);
4101             g_free(child_key_dot);
4102         }
4103 
4104         bdrv_reopen_queue_child(bs_queue, child->bs, new_child_options,
4105                                 child->klass, child->role, bs->drv->is_format,
4106                                 options, flags, child_keep_old);
4107     }
4108 
4109     return bs_queue;
4110 }
4111 
4112 BlockReopenQueue *bdrv_reopen_queue(BlockReopenQueue *bs_queue,
4113                                     BlockDriverState *bs,
4114                                     QDict *options, bool keep_old_opts)
4115 {
4116     return bdrv_reopen_queue_child(bs_queue, bs, options, NULL, 0, false,
4117                                    NULL, 0, keep_old_opts);
4118 }
4119 
4120 void bdrv_reopen_queue_free(BlockReopenQueue *bs_queue)
4121 {
4122     if (bs_queue) {
4123         BlockReopenQueueEntry *bs_entry, *next;
4124         QTAILQ_FOREACH_SAFE(bs_entry, bs_queue, entry, next) {
4125             qobject_unref(bs_entry->state.explicit_options);
4126             qobject_unref(bs_entry->state.options);
4127             g_free(bs_entry);
4128         }
4129         g_free(bs_queue);
4130     }
4131 }
4132 
4133 /*
4134  * Reopen multiple BlockDriverStates atomically & transactionally.
4135  *
4136  * The queue passed in (bs_queue) must have been built up previous
4137  * via bdrv_reopen_queue().
4138  *
4139  * Reopens all BDS specified in the queue, with the appropriate
4140  * flags.  All devices are prepared for reopen, and failure of any
4141  * device will cause all device changes to be abandoned, and intermediate
4142  * data cleaned up.
4143  *
4144  * If all devices prepare successfully, then the changes are committed
4145  * to all devices.
4146  *
4147  * All affected nodes must be drained between bdrv_reopen_queue() and
4148  * bdrv_reopen_multiple().
4149  *
4150  * To be called from the main thread, with all other AioContexts unlocked.
4151  */
4152 int bdrv_reopen_multiple(BlockReopenQueue *bs_queue, Error **errp)
4153 {
4154     int ret = -1;
4155     BlockReopenQueueEntry *bs_entry, *next;
4156     AioContext *ctx;
4157     Transaction *tran = tran_new();
4158     g_autoptr(GHashTable) found = NULL;
4159     g_autoptr(GSList) refresh_list = NULL;
4160 
4161     assert(qemu_get_current_aio_context() == qemu_get_aio_context());
4162     assert(bs_queue != NULL);
4163 
4164     QTAILQ_FOREACH(bs_entry, bs_queue, entry) {
4165         ctx = bdrv_get_aio_context(bs_entry->state.bs);
4166         aio_context_acquire(ctx);
4167         ret = bdrv_flush(bs_entry->state.bs);
4168         aio_context_release(ctx);
4169         if (ret < 0) {
4170             error_setg_errno(errp, -ret, "Error flushing drive");
4171             goto abort;
4172         }
4173     }
4174 
4175     QTAILQ_FOREACH(bs_entry, bs_queue, entry) {
4176         assert(bs_entry->state.bs->quiesce_counter > 0);
4177         ctx = bdrv_get_aio_context(bs_entry->state.bs);
4178         aio_context_acquire(ctx);
4179         ret = bdrv_reopen_prepare(&bs_entry->state, bs_queue, tran, errp);
4180         aio_context_release(ctx);
4181         if (ret < 0) {
4182             goto abort;
4183         }
4184         bs_entry->prepared = true;
4185     }
4186 
4187     found = g_hash_table_new(NULL, NULL);
4188     QTAILQ_FOREACH(bs_entry, bs_queue, entry) {
4189         BDRVReopenState *state = &bs_entry->state;
4190 
4191         refresh_list = bdrv_topological_dfs(refresh_list, found, state->bs);
4192         if (state->old_backing_bs) {
4193             refresh_list = bdrv_topological_dfs(refresh_list, found,
4194                                                 state->old_backing_bs);
4195         }
4196         if (state->old_file_bs) {
4197             refresh_list = bdrv_topological_dfs(refresh_list, found,
4198                                                 state->old_file_bs);
4199         }
4200     }
4201 
4202     /*
4203      * Note that file-posix driver rely on permission update done during reopen
4204      * (even if no permission changed), because it wants "new" permissions for
4205      * reconfiguring the fd and that's why it does it in raw_check_perm(), not
4206      * in raw_reopen_prepare() which is called with "old" permissions.
4207      */
4208     ret = bdrv_list_refresh_perms(refresh_list, bs_queue, tran, errp);
4209     if (ret < 0) {
4210         goto abort;
4211     }
4212 
4213     /*
4214      * If we reach this point, we have success and just need to apply the
4215      * changes.
4216      *
4217      * Reverse order is used to comfort qcow2 driver: on commit it need to write
4218      * IN_USE flag to the image, to mark bitmaps in the image as invalid. But
4219      * children are usually goes after parents in reopen-queue, so go from last
4220      * to first element.
4221      */
4222     QTAILQ_FOREACH_REVERSE(bs_entry, bs_queue, entry) {
4223         ctx = bdrv_get_aio_context(bs_entry->state.bs);
4224         aio_context_acquire(ctx);
4225         bdrv_reopen_commit(&bs_entry->state);
4226         aio_context_release(ctx);
4227     }
4228 
4229     tran_commit(tran);
4230 
4231     QTAILQ_FOREACH_REVERSE(bs_entry, bs_queue, entry) {
4232         BlockDriverState *bs = bs_entry->state.bs;
4233 
4234         if (bs->drv->bdrv_reopen_commit_post) {
4235             ctx = bdrv_get_aio_context(bs);
4236             aio_context_acquire(ctx);
4237             bs->drv->bdrv_reopen_commit_post(&bs_entry->state);
4238             aio_context_release(ctx);
4239         }
4240     }
4241 
4242     ret = 0;
4243     goto cleanup;
4244 
4245 abort:
4246     tran_abort(tran);
4247     QTAILQ_FOREACH_SAFE(bs_entry, bs_queue, entry, next) {
4248         if (bs_entry->prepared) {
4249             ctx = bdrv_get_aio_context(bs_entry->state.bs);
4250             aio_context_acquire(ctx);
4251             bdrv_reopen_abort(&bs_entry->state);
4252             aio_context_release(ctx);
4253         }
4254     }
4255 
4256 cleanup:
4257     bdrv_reopen_queue_free(bs_queue);
4258 
4259     return ret;
4260 }
4261 
4262 int bdrv_reopen(BlockDriverState *bs, QDict *opts, bool keep_old_opts,
4263                 Error **errp)
4264 {
4265     AioContext *ctx = bdrv_get_aio_context(bs);
4266     BlockReopenQueue *queue;
4267     int ret;
4268 
4269     bdrv_subtree_drained_begin(bs);
4270     if (ctx != qemu_get_aio_context()) {
4271         aio_context_release(ctx);
4272     }
4273 
4274     queue = bdrv_reopen_queue(NULL, bs, opts, keep_old_opts);
4275     ret = bdrv_reopen_multiple(queue, errp);
4276 
4277     if (ctx != qemu_get_aio_context()) {
4278         aio_context_acquire(ctx);
4279     }
4280     bdrv_subtree_drained_end(bs);
4281 
4282     return ret;
4283 }
4284 
4285 int bdrv_reopen_set_read_only(BlockDriverState *bs, bool read_only,
4286                               Error **errp)
4287 {
4288     QDict *opts = qdict_new();
4289 
4290     qdict_put_bool(opts, BDRV_OPT_READ_ONLY, read_only);
4291 
4292     return bdrv_reopen(bs, opts, true, errp);
4293 }
4294 
4295 /*
4296  * Take a BDRVReopenState and check if the value of 'backing' in the
4297  * reopen_state->options QDict is valid or not.
4298  *
4299  * If 'backing' is missing from the QDict then return 0.
4300  *
4301  * If 'backing' contains the node name of the backing file of
4302  * reopen_state->bs then return 0.
4303  *
4304  * If 'backing' contains a different node name (or is null) then check
4305  * whether the current backing file can be replaced with the new one.
4306  * If that's the case then reopen_state->replace_backing_bs is set to
4307  * true and reopen_state->new_backing_bs contains a pointer to the new
4308  * backing BlockDriverState (or NULL).
4309  *
4310  * Return 0 on success, otherwise return < 0 and set @errp.
4311  */
4312 static int bdrv_reopen_parse_file_or_backing(BDRVReopenState *reopen_state,
4313                                              bool is_backing, Transaction *tran,
4314                                              Error **errp)
4315 {
4316     BlockDriverState *bs = reopen_state->bs;
4317     BlockDriverState *new_child_bs;
4318     BlockDriverState *old_child_bs = is_backing ? child_bs(bs->backing) :
4319                                                   child_bs(bs->file);
4320     const char *child_name = is_backing ? "backing" : "file";
4321     QObject *value;
4322     const char *str;
4323 
4324     value = qdict_get(reopen_state->options, child_name);
4325     if (value == NULL) {
4326         return 0;
4327     }
4328 
4329     switch (qobject_type(value)) {
4330     case QTYPE_QNULL:
4331         assert(is_backing); /* The 'file' option does not allow a null value */
4332         new_child_bs = NULL;
4333         break;
4334     case QTYPE_QSTRING:
4335         str = qstring_get_str(qobject_to(QString, value));
4336         new_child_bs = bdrv_lookup_bs(NULL, str, errp);
4337         if (new_child_bs == NULL) {
4338             return -EINVAL;
4339         } else if (bdrv_recurse_has_child(new_child_bs, bs)) {
4340             error_setg(errp, "Making '%s' a %s child of '%s' would create a "
4341                        "cycle", str, child_name, bs->node_name);
4342             return -EINVAL;
4343         }
4344         break;
4345     default:
4346         /*
4347          * The options QDict has been flattened, so 'backing' and 'file'
4348          * do not allow any other data type here.
4349          */
4350         g_assert_not_reached();
4351     }
4352 
4353     if (old_child_bs == new_child_bs) {
4354         return 0;
4355     }
4356 
4357     if (old_child_bs) {
4358         if (bdrv_skip_implicit_filters(old_child_bs) == new_child_bs) {
4359             return 0;
4360         }
4361 
4362         if (old_child_bs->implicit) {
4363             error_setg(errp, "Cannot replace implicit %s child of %s",
4364                        child_name, bs->node_name);
4365             return -EPERM;
4366         }
4367     }
4368 
4369     if (bs->drv->is_filter && !old_child_bs) {
4370         /*
4371          * Filters always have a file or a backing child, so we are trying to
4372          * change wrong child
4373          */
4374         error_setg(errp, "'%s' is a %s filter node that does not support a "
4375                    "%s child", bs->node_name, bs->drv->format_name, child_name);
4376         return -EINVAL;
4377     }
4378 
4379     if (is_backing) {
4380         reopen_state->old_backing_bs = old_child_bs;
4381     } else {
4382         reopen_state->old_file_bs = old_child_bs;
4383     }
4384 
4385     return bdrv_set_file_or_backing_noperm(bs, new_child_bs, is_backing,
4386                                            tran, errp);
4387 }
4388 
4389 /*
4390  * Prepares a BlockDriverState for reopen. All changes are staged in the
4391  * 'opaque' field of the BDRVReopenState, which is used and allocated by
4392  * the block driver layer .bdrv_reopen_prepare()
4393  *
4394  * bs is the BlockDriverState to reopen
4395  * flags are the new open flags
4396  * queue is the reopen queue
4397  *
4398  * Returns 0 on success, non-zero on error.  On error errp will be set
4399  * as well.
4400  *
4401  * On failure, bdrv_reopen_abort() will be called to clean up any data.
4402  * It is the responsibility of the caller to then call the abort() or
4403  * commit() for any other BDS that have been left in a prepare() state
4404  *
4405  */
4406 static int bdrv_reopen_prepare(BDRVReopenState *reopen_state,
4407                                BlockReopenQueue *queue,
4408                                Transaction *change_child_tran, Error **errp)
4409 {
4410     int ret = -1;
4411     int old_flags;
4412     Error *local_err = NULL;
4413     BlockDriver *drv;
4414     QemuOpts *opts;
4415     QDict *orig_reopen_opts;
4416     char *discard = NULL;
4417     bool read_only;
4418     bool drv_prepared = false;
4419 
4420     assert(reopen_state != NULL);
4421     assert(reopen_state->bs->drv != NULL);
4422     drv = reopen_state->bs->drv;
4423 
4424     /* This function and each driver's bdrv_reopen_prepare() remove
4425      * entries from reopen_state->options as they are processed, so
4426      * we need to make a copy of the original QDict. */
4427     orig_reopen_opts = qdict_clone_shallow(reopen_state->options);
4428 
4429     /* Process generic block layer options */
4430     opts = qemu_opts_create(&bdrv_runtime_opts, NULL, 0, &error_abort);
4431     if (!qemu_opts_absorb_qdict(opts, reopen_state->options, errp)) {
4432         ret = -EINVAL;
4433         goto error;
4434     }
4435 
4436     /* This was already called in bdrv_reopen_queue_child() so the flags
4437      * are up-to-date. This time we simply want to remove the options from
4438      * QemuOpts in order to indicate that they have been processed. */
4439     old_flags = reopen_state->flags;
4440     update_flags_from_options(&reopen_state->flags, opts);
4441     assert(old_flags == reopen_state->flags);
4442 
4443     discard = qemu_opt_get_del(opts, BDRV_OPT_DISCARD);
4444     if (discard != NULL) {
4445         if (bdrv_parse_discard_flags(discard, &reopen_state->flags) != 0) {
4446             error_setg(errp, "Invalid discard option");
4447             ret = -EINVAL;
4448             goto error;
4449         }
4450     }
4451 
4452     reopen_state->detect_zeroes =
4453         bdrv_parse_detect_zeroes(opts, reopen_state->flags, &local_err);
4454     if (local_err) {
4455         error_propagate(errp, local_err);
4456         ret = -EINVAL;
4457         goto error;
4458     }
4459 
4460     /* All other options (including node-name and driver) must be unchanged.
4461      * Put them back into the QDict, so that they are checked at the end
4462      * of this function. */
4463     qemu_opts_to_qdict(opts, reopen_state->options);
4464 
4465     /* If we are to stay read-only, do not allow permission change
4466      * to r/w. Attempting to set to r/w may fail if either BDRV_O_ALLOW_RDWR is
4467      * not set, or if the BDS still has copy_on_read enabled */
4468     read_only = !(reopen_state->flags & BDRV_O_RDWR);
4469     ret = bdrv_can_set_read_only(reopen_state->bs, read_only, true, &local_err);
4470     if (local_err) {
4471         error_propagate(errp, local_err);
4472         goto error;
4473     }
4474 
4475     if (drv->bdrv_reopen_prepare) {
4476         /*
4477          * If a driver-specific option is missing, it means that we
4478          * should reset it to its default value.
4479          * But not all options allow that, so we need to check it first.
4480          */
4481         ret = bdrv_reset_options_allowed(reopen_state->bs,
4482                                          reopen_state->options, errp);
4483         if (ret) {
4484             goto error;
4485         }
4486 
4487         ret = drv->bdrv_reopen_prepare(reopen_state, queue, &local_err);
4488         if (ret) {
4489             if (local_err != NULL) {
4490                 error_propagate(errp, local_err);
4491             } else {
4492                 bdrv_refresh_filename(reopen_state->bs);
4493                 error_setg(errp, "failed while preparing to reopen image '%s'",
4494                            reopen_state->bs->filename);
4495             }
4496             goto error;
4497         }
4498     } else {
4499         /* It is currently mandatory to have a bdrv_reopen_prepare()
4500          * handler for each supported drv. */
4501         error_setg(errp, "Block format '%s' used by node '%s' "
4502                    "does not support reopening files", drv->format_name,
4503                    bdrv_get_device_or_node_name(reopen_state->bs));
4504         ret = -1;
4505         goto error;
4506     }
4507 
4508     drv_prepared = true;
4509 
4510     /*
4511      * We must provide the 'backing' option if the BDS has a backing
4512      * file or if the image file has a backing file name as part of
4513      * its metadata. Otherwise the 'backing' option can be omitted.
4514      */
4515     if (drv->supports_backing && reopen_state->backing_missing &&
4516         (reopen_state->bs->backing || reopen_state->bs->backing_file[0])) {
4517         error_setg(errp, "backing is missing for '%s'",
4518                    reopen_state->bs->node_name);
4519         ret = -EINVAL;
4520         goto error;
4521     }
4522 
4523     /*
4524      * Allow changing the 'backing' option. The new value can be
4525      * either a reference to an existing node (using its node name)
4526      * or NULL to simply detach the current backing file.
4527      */
4528     ret = bdrv_reopen_parse_file_or_backing(reopen_state, true,
4529                                             change_child_tran, errp);
4530     if (ret < 0) {
4531         goto error;
4532     }
4533     qdict_del(reopen_state->options, "backing");
4534 
4535     /* Allow changing the 'file' option. In this case NULL is not allowed */
4536     ret = bdrv_reopen_parse_file_or_backing(reopen_state, false,
4537                                             change_child_tran, errp);
4538     if (ret < 0) {
4539         goto error;
4540     }
4541     qdict_del(reopen_state->options, "file");
4542 
4543     /* Options that are not handled are only okay if they are unchanged
4544      * compared to the old state. It is expected that some options are only
4545      * used for the initial open, but not reopen (e.g. filename) */
4546     if (qdict_size(reopen_state->options)) {
4547         const QDictEntry *entry = qdict_first(reopen_state->options);
4548 
4549         do {
4550             QObject *new = entry->value;
4551             QObject *old = qdict_get(reopen_state->bs->options, entry->key);
4552 
4553             /* Allow child references (child_name=node_name) as long as they
4554              * point to the current child (i.e. everything stays the same). */
4555             if (qobject_type(new) == QTYPE_QSTRING) {
4556                 BdrvChild *child;
4557                 QLIST_FOREACH(child, &reopen_state->bs->children, next) {
4558                     if (!strcmp(child->name, entry->key)) {
4559                         break;
4560                     }
4561                 }
4562 
4563                 if (child) {
4564                     if (!strcmp(child->bs->node_name,
4565                                 qstring_get_str(qobject_to(QString, new)))) {
4566                         continue; /* Found child with this name, skip option */
4567                     }
4568                 }
4569             }
4570 
4571             /*
4572              * TODO: When using -drive to specify blockdev options, all values
4573              * will be strings; however, when using -blockdev, blockdev-add or
4574              * filenames using the json:{} pseudo-protocol, they will be
4575              * correctly typed.
4576              * In contrast, reopening options are (currently) always strings
4577              * (because you can only specify them through qemu-io; all other
4578              * callers do not specify any options).
4579              * Therefore, when using anything other than -drive to create a BDS,
4580              * this cannot detect non-string options as unchanged, because
4581              * qobject_is_equal() always returns false for objects of different
4582              * type.  In the future, this should be remedied by correctly typing
4583              * all options.  For now, this is not too big of an issue because
4584              * the user can simply omit options which cannot be changed anyway,
4585              * so they will stay unchanged.
4586              */
4587             if (!qobject_is_equal(new, old)) {
4588                 error_setg(errp, "Cannot change the option '%s'", entry->key);
4589                 ret = -EINVAL;
4590                 goto error;
4591             }
4592         } while ((entry = qdict_next(reopen_state->options, entry)));
4593     }
4594 
4595     ret = 0;
4596 
4597     /* Restore the original reopen_state->options QDict */
4598     qobject_unref(reopen_state->options);
4599     reopen_state->options = qobject_ref(orig_reopen_opts);
4600 
4601 error:
4602     if (ret < 0 && drv_prepared) {
4603         /* drv->bdrv_reopen_prepare() has succeeded, so we need to
4604          * call drv->bdrv_reopen_abort() before signaling an error
4605          * (bdrv_reopen_multiple() will not call bdrv_reopen_abort()
4606          * when the respective bdrv_reopen_prepare() has failed) */
4607         if (drv->bdrv_reopen_abort) {
4608             drv->bdrv_reopen_abort(reopen_state);
4609         }
4610     }
4611     qemu_opts_del(opts);
4612     qobject_unref(orig_reopen_opts);
4613     g_free(discard);
4614     return ret;
4615 }
4616 
4617 /*
4618  * Takes the staged changes for the reopen from bdrv_reopen_prepare(), and
4619  * makes them final by swapping the staging BlockDriverState contents into
4620  * the active BlockDriverState contents.
4621  */
4622 static void bdrv_reopen_commit(BDRVReopenState *reopen_state)
4623 {
4624     BlockDriver *drv;
4625     BlockDriverState *bs;
4626     BdrvChild *child;
4627 
4628     assert(reopen_state != NULL);
4629     bs = reopen_state->bs;
4630     drv = bs->drv;
4631     assert(drv != NULL);
4632 
4633     /* If there are any driver level actions to take */
4634     if (drv->bdrv_reopen_commit) {
4635         drv->bdrv_reopen_commit(reopen_state);
4636     }
4637 
4638     /* set BDS specific flags now */
4639     qobject_unref(bs->explicit_options);
4640     qobject_unref(bs->options);
4641     qobject_ref(reopen_state->explicit_options);
4642     qobject_ref(reopen_state->options);
4643 
4644     bs->explicit_options   = reopen_state->explicit_options;
4645     bs->options            = reopen_state->options;
4646     bs->open_flags         = reopen_state->flags;
4647     bs->detect_zeroes      = reopen_state->detect_zeroes;
4648 
4649     /* Remove child references from bs->options and bs->explicit_options.
4650      * Child options were already removed in bdrv_reopen_queue_child() */
4651     QLIST_FOREACH(child, &bs->children, next) {
4652         qdict_del(bs->explicit_options, child->name);
4653         qdict_del(bs->options, child->name);
4654     }
4655     /* backing is probably removed, so it's not handled by previous loop */
4656     qdict_del(bs->explicit_options, "backing");
4657     qdict_del(bs->options, "backing");
4658 
4659     bdrv_refresh_limits(bs, NULL, NULL);
4660 }
4661 
4662 /*
4663  * Abort the reopen, and delete and free the staged changes in
4664  * reopen_state
4665  */
4666 static void bdrv_reopen_abort(BDRVReopenState *reopen_state)
4667 {
4668     BlockDriver *drv;
4669 
4670     assert(reopen_state != NULL);
4671     drv = reopen_state->bs->drv;
4672     assert(drv != NULL);
4673 
4674     if (drv->bdrv_reopen_abort) {
4675         drv->bdrv_reopen_abort(reopen_state);
4676     }
4677 }
4678 
4679 
4680 static void bdrv_close(BlockDriverState *bs)
4681 {
4682     BdrvAioNotifier *ban, *ban_next;
4683     BdrvChild *child, *next;
4684 
4685     assert(!bs->refcnt);
4686 
4687     bdrv_drained_begin(bs); /* complete I/O */
4688     bdrv_flush(bs);
4689     bdrv_drain(bs); /* in case flush left pending I/O */
4690 
4691     if (bs->drv) {
4692         if (bs->drv->bdrv_close) {
4693             /* Must unfreeze all children, so bdrv_unref_child() works */
4694             bs->drv->bdrv_close(bs);
4695         }
4696         bs->drv = NULL;
4697     }
4698 
4699     QLIST_FOREACH_SAFE(child, &bs->children, next, next) {
4700         bdrv_unref_child(bs, child);
4701     }
4702 
4703     bs->backing = NULL;
4704     bs->file = NULL;
4705     g_free(bs->opaque);
4706     bs->opaque = NULL;
4707     qatomic_set(&bs->copy_on_read, 0);
4708     bs->backing_file[0] = '\0';
4709     bs->backing_format[0] = '\0';
4710     bs->total_sectors = 0;
4711     bs->encrypted = false;
4712     bs->sg = false;
4713     qobject_unref(bs->options);
4714     qobject_unref(bs->explicit_options);
4715     bs->options = NULL;
4716     bs->explicit_options = NULL;
4717     qobject_unref(bs->full_open_options);
4718     bs->full_open_options = NULL;
4719     g_free(bs->block_status_cache);
4720     bs->block_status_cache = NULL;
4721 
4722     bdrv_release_named_dirty_bitmaps(bs);
4723     assert(QLIST_EMPTY(&bs->dirty_bitmaps));
4724 
4725     QLIST_FOREACH_SAFE(ban, &bs->aio_notifiers, list, ban_next) {
4726         g_free(ban);
4727     }
4728     QLIST_INIT(&bs->aio_notifiers);
4729     bdrv_drained_end(bs);
4730 
4731     /*
4732      * If we're still inside some bdrv_drain_all_begin()/end() sections, end
4733      * them now since this BDS won't exist anymore when bdrv_drain_all_end()
4734      * gets called.
4735      */
4736     if (bs->quiesce_counter) {
4737         bdrv_drain_all_end_quiesce(bs);
4738     }
4739 }
4740 
4741 void bdrv_close_all(void)
4742 {
4743     assert(job_next(NULL) == NULL);
4744 
4745     /* Drop references from requests still in flight, such as canceled block
4746      * jobs whose AIO context has not been polled yet */
4747     bdrv_drain_all();
4748 
4749     blk_remove_all_bs();
4750     blockdev_close_all_bdrv_states();
4751 
4752     assert(QTAILQ_EMPTY(&all_bdrv_states));
4753 }
4754 
4755 static bool should_update_child(BdrvChild *c, BlockDriverState *to)
4756 {
4757     GQueue *queue;
4758     GHashTable *found;
4759     bool ret;
4760 
4761     if (c->klass->stay_at_node) {
4762         return false;
4763     }
4764 
4765     /* If the child @c belongs to the BDS @to, replacing the current
4766      * c->bs by @to would mean to create a loop.
4767      *
4768      * Such a case occurs when appending a BDS to a backing chain.
4769      * For instance, imagine the following chain:
4770      *
4771      *   guest device -> node A -> further backing chain...
4772      *
4773      * Now we create a new BDS B which we want to put on top of this
4774      * chain, so we first attach A as its backing node:
4775      *
4776      *                   node B
4777      *                     |
4778      *                     v
4779      *   guest device -> node A -> further backing chain...
4780      *
4781      * Finally we want to replace A by B.  When doing that, we want to
4782      * replace all pointers to A by pointers to B -- except for the
4783      * pointer from B because (1) that would create a loop, and (2)
4784      * that pointer should simply stay intact:
4785      *
4786      *   guest device -> node B
4787      *                     |
4788      *                     v
4789      *                   node A -> further backing chain...
4790      *
4791      * In general, when replacing a node A (c->bs) by a node B (@to),
4792      * if A is a child of B, that means we cannot replace A by B there
4793      * because that would create a loop.  Silently detaching A from B
4794      * is also not really an option.  So overall just leaving A in
4795      * place there is the most sensible choice.
4796      *
4797      * We would also create a loop in any cases where @c is only
4798      * indirectly referenced by @to. Prevent this by returning false
4799      * if @c is found (by breadth-first search) anywhere in the whole
4800      * subtree of @to.
4801      */
4802 
4803     ret = true;
4804     found = g_hash_table_new(NULL, NULL);
4805     g_hash_table_add(found, to);
4806     queue = g_queue_new();
4807     g_queue_push_tail(queue, to);
4808 
4809     while (!g_queue_is_empty(queue)) {
4810         BlockDriverState *v = g_queue_pop_head(queue);
4811         BdrvChild *c2;
4812 
4813         QLIST_FOREACH(c2, &v->children, next) {
4814             if (c2 == c) {
4815                 ret = false;
4816                 break;
4817             }
4818 
4819             if (g_hash_table_contains(found, c2->bs)) {
4820                 continue;
4821             }
4822 
4823             g_queue_push_tail(queue, c2->bs);
4824             g_hash_table_add(found, c2->bs);
4825         }
4826     }
4827 
4828     g_queue_free(queue);
4829     g_hash_table_destroy(found);
4830 
4831     return ret;
4832 }
4833 
4834 typedef struct BdrvRemoveFilterOrCowChild {
4835     BdrvChild *child;
4836     bool is_backing;
4837 } BdrvRemoveFilterOrCowChild;
4838 
4839 static void bdrv_remove_filter_or_cow_child_abort(void *opaque)
4840 {
4841     BdrvRemoveFilterOrCowChild *s = opaque;
4842     BlockDriverState *parent_bs = s->child->opaque;
4843 
4844     QLIST_INSERT_HEAD(&parent_bs->children, s->child, next);
4845     if (s->is_backing) {
4846         parent_bs->backing = s->child;
4847     } else {
4848         parent_bs->file = s->child;
4849     }
4850 
4851     /*
4852      * We don't have to restore child->bs here to undo bdrv_replace_child_tran()
4853      * because that function is transactionable and it registered own completion
4854      * entries in @tran, so .abort() for bdrv_replace_child_safe() will be
4855      * called automatically.
4856      */
4857 }
4858 
4859 static void bdrv_remove_filter_or_cow_child_commit(void *opaque)
4860 {
4861     BdrvRemoveFilterOrCowChild *s = opaque;
4862 
4863     bdrv_child_free(s->child);
4864 }
4865 
4866 static TransactionActionDrv bdrv_remove_filter_or_cow_child_drv = {
4867     .abort = bdrv_remove_filter_or_cow_child_abort,
4868     .commit = bdrv_remove_filter_or_cow_child_commit,
4869     .clean = g_free,
4870 };
4871 
4872 /*
4873  * A function to remove backing or file child of @bs.
4874  * Function doesn't update permissions, caller is responsible for this.
4875  */
4876 static void bdrv_remove_file_or_backing_child(BlockDriverState *bs,
4877                                               BdrvChild *child,
4878                                               Transaction *tran)
4879 {
4880     BdrvRemoveFilterOrCowChild *s;
4881 
4882     assert(child == bs->backing || child == bs->file);
4883 
4884     if (!child) {
4885         return;
4886     }
4887 
4888     if (child->bs) {
4889         bdrv_replace_child_tran(child, NULL, tran);
4890     }
4891 
4892     s = g_new(BdrvRemoveFilterOrCowChild, 1);
4893     *s = (BdrvRemoveFilterOrCowChild) {
4894         .child = child,
4895         .is_backing = (child == bs->backing),
4896     };
4897     tran_add(tran, &bdrv_remove_filter_or_cow_child_drv, s);
4898 
4899     QLIST_SAFE_REMOVE(child, next);
4900     if (s->is_backing) {
4901         bs->backing = NULL;
4902     } else {
4903         bs->file = NULL;
4904     }
4905 }
4906 
4907 /*
4908  * A function to remove backing-chain child of @bs if exists: cow child for
4909  * format nodes (always .backing) and filter child for filters (may be .file or
4910  * .backing)
4911  */
4912 static void bdrv_remove_filter_or_cow_child(BlockDriverState *bs,
4913                                             Transaction *tran)
4914 {
4915     bdrv_remove_file_or_backing_child(bs, bdrv_filter_or_cow_child(bs), tran);
4916 }
4917 
4918 static int bdrv_replace_node_noperm(BlockDriverState *from,
4919                                     BlockDriverState *to,
4920                                     bool auto_skip, Transaction *tran,
4921                                     Error **errp)
4922 {
4923     BdrvChild *c, *next;
4924 
4925     QLIST_FOREACH_SAFE(c, &from->parents, next_parent, next) {
4926         assert(c->bs == from);
4927         if (!should_update_child(c, to)) {
4928             if (auto_skip) {
4929                 continue;
4930             }
4931             error_setg(errp, "Should not change '%s' link to '%s'",
4932                        c->name, from->node_name);
4933             return -EINVAL;
4934         }
4935         if (c->frozen) {
4936             error_setg(errp, "Cannot change '%s' link to '%s'",
4937                        c->name, from->node_name);
4938             return -EPERM;
4939         }
4940         bdrv_replace_child_tran(c, to, tran);
4941     }
4942 
4943     return 0;
4944 }
4945 
4946 /*
4947  * With auto_skip=true bdrv_replace_node_common skips updating from parents
4948  * if it creates a parent-child relation loop or if parent is block-job.
4949  *
4950  * With auto_skip=false the error is returned if from has a parent which should
4951  * not be updated.
4952  *
4953  * With @detach_subchain=true @to must be in a backing chain of @from. In this
4954  * case backing link of the cow-parent of @to is removed.
4955  */
4956 static int bdrv_replace_node_common(BlockDriverState *from,
4957                                     BlockDriverState *to,
4958                                     bool auto_skip, bool detach_subchain,
4959                                     Error **errp)
4960 {
4961     Transaction *tran = tran_new();
4962     g_autoptr(GHashTable) found = NULL;
4963     g_autoptr(GSList) refresh_list = NULL;
4964     BlockDriverState *to_cow_parent = NULL;
4965     int ret;
4966 
4967     if (detach_subchain) {
4968         assert(bdrv_chain_contains(from, to));
4969         assert(from != to);
4970         for (to_cow_parent = from;
4971              bdrv_filter_or_cow_bs(to_cow_parent) != to;
4972              to_cow_parent = bdrv_filter_or_cow_bs(to_cow_parent))
4973         {
4974             ;
4975         }
4976     }
4977 
4978     /* Make sure that @from doesn't go away until we have successfully attached
4979      * all of its parents to @to. */
4980     bdrv_ref(from);
4981 
4982     assert(qemu_get_current_aio_context() == qemu_get_aio_context());
4983     assert(bdrv_get_aio_context(from) == bdrv_get_aio_context(to));
4984     bdrv_drained_begin(from);
4985 
4986     /*
4987      * Do the replacement without permission update.
4988      * Replacement may influence the permissions, we should calculate new
4989      * permissions based on new graph. If we fail, we'll roll-back the
4990      * replacement.
4991      */
4992     ret = bdrv_replace_node_noperm(from, to, auto_skip, tran, errp);
4993     if (ret < 0) {
4994         goto out;
4995     }
4996 
4997     if (detach_subchain) {
4998         bdrv_remove_filter_or_cow_child(to_cow_parent, tran);
4999     }
5000 
5001     found = g_hash_table_new(NULL, NULL);
5002 
5003     refresh_list = bdrv_topological_dfs(refresh_list, found, to);
5004     refresh_list = bdrv_topological_dfs(refresh_list, found, from);
5005 
5006     ret = bdrv_list_refresh_perms(refresh_list, NULL, tran, errp);
5007     if (ret < 0) {
5008         goto out;
5009     }
5010 
5011     ret = 0;
5012 
5013 out:
5014     tran_finalize(tran, ret);
5015 
5016     bdrv_drained_end(from);
5017     bdrv_unref(from);
5018 
5019     return ret;
5020 }
5021 
5022 int bdrv_replace_node(BlockDriverState *from, BlockDriverState *to,
5023                       Error **errp)
5024 {
5025     return bdrv_replace_node_common(from, to, true, false, errp);
5026 }
5027 
5028 int bdrv_drop_filter(BlockDriverState *bs, Error **errp)
5029 {
5030     return bdrv_replace_node_common(bs, bdrv_filter_or_cow_bs(bs), true, true,
5031                                     errp);
5032 }
5033 
5034 /*
5035  * Add new bs contents at the top of an image chain while the chain is
5036  * live, while keeping required fields on the top layer.
5037  *
5038  * This will modify the BlockDriverState fields, and swap contents
5039  * between bs_new and bs_top. Both bs_new and bs_top are modified.
5040  *
5041  * bs_new must not be attached to a BlockBackend and must not have backing
5042  * child.
5043  *
5044  * This function does not create any image files.
5045  */
5046 int bdrv_append(BlockDriverState *bs_new, BlockDriverState *bs_top,
5047                 Error **errp)
5048 {
5049     int ret;
5050     Transaction *tran = tran_new();
5051 
5052     assert(!bs_new->backing);
5053 
5054     ret = bdrv_attach_child_noperm(bs_new, bs_top, "backing",
5055                                    &child_of_bds, bdrv_backing_role(bs_new),
5056                                    &bs_new->backing, tran, errp);
5057     if (ret < 0) {
5058         goto out;
5059     }
5060 
5061     ret = bdrv_replace_node_noperm(bs_top, bs_new, true, tran, errp);
5062     if (ret < 0) {
5063         goto out;
5064     }
5065 
5066     ret = bdrv_refresh_perms(bs_new, errp);
5067 out:
5068     tran_finalize(tran, ret);
5069 
5070     bdrv_refresh_limits(bs_top, NULL, NULL);
5071 
5072     return ret;
5073 }
5074 
5075 /* Not for empty child */
5076 int bdrv_replace_child_bs(BdrvChild *child, BlockDriverState *new_bs,
5077                           Error **errp)
5078 {
5079     int ret;
5080     Transaction *tran = tran_new();
5081     g_autoptr(GHashTable) found = NULL;
5082     g_autoptr(GSList) refresh_list = NULL;
5083     BlockDriverState *old_bs = child->bs;
5084 
5085     bdrv_ref(old_bs);
5086     bdrv_drained_begin(old_bs);
5087     bdrv_drained_begin(new_bs);
5088 
5089     bdrv_replace_child_tran(child, new_bs, tran);
5090 
5091     found = g_hash_table_new(NULL, NULL);
5092     refresh_list = bdrv_topological_dfs(refresh_list, found, old_bs);
5093     refresh_list = bdrv_topological_dfs(refresh_list, found, new_bs);
5094 
5095     ret = bdrv_list_refresh_perms(refresh_list, NULL, tran, errp);
5096 
5097     tran_finalize(tran, ret);
5098 
5099     bdrv_drained_end(old_bs);
5100     bdrv_drained_end(new_bs);
5101     bdrv_unref(old_bs);
5102 
5103     return ret;
5104 }
5105 
5106 static void bdrv_delete(BlockDriverState *bs)
5107 {
5108     assert(bdrv_op_blocker_is_empty(bs));
5109     assert(!bs->refcnt);
5110 
5111     /* remove from list, if necessary */
5112     if (bs->node_name[0] != '\0') {
5113         QTAILQ_REMOVE(&graph_bdrv_states, bs, node_list);
5114     }
5115     QTAILQ_REMOVE(&all_bdrv_states, bs, bs_list);
5116 
5117     bdrv_close(bs);
5118 
5119     g_free(bs);
5120 }
5121 
5122 
5123 /*
5124  * Replace @bs by newly created block node.
5125  *
5126  * @options is a QDict of options to pass to the block drivers, or NULL for an
5127  * empty set of options. The reference to the QDict belongs to the block layer
5128  * after the call (even on failure), so if the caller intends to reuse the
5129  * dictionary, it needs to use qobject_ref() before calling bdrv_open.
5130  */
5131 BlockDriverState *bdrv_insert_node(BlockDriverState *bs, QDict *options,
5132                                    int flags, Error **errp)
5133 {
5134     ERRP_GUARD();
5135     int ret;
5136     BlockDriverState *new_node_bs;
5137 
5138     new_node_bs = bdrv_open(NULL, NULL, options, flags, errp);
5139     if (new_node_bs == NULL) {
5140         error_prepend(errp, "Could not create node: ");
5141         return NULL;
5142     }
5143 
5144     bdrv_drained_begin(bs);
5145     ret = bdrv_replace_node(bs, new_node_bs, errp);
5146     bdrv_drained_end(bs);
5147 
5148     if (ret < 0) {
5149         error_prepend(errp, "Could not replace node: ");
5150         bdrv_unref(new_node_bs);
5151         return NULL;
5152     }
5153 
5154     return new_node_bs;
5155 }
5156 
5157 /*
5158  * Run consistency checks on an image
5159  *
5160  * Returns 0 if the check could be completed (it doesn't mean that the image is
5161  * free of errors) or -errno when an internal error occurred. The results of the
5162  * check are stored in res.
5163  */
5164 int coroutine_fn bdrv_co_check(BlockDriverState *bs,
5165                                BdrvCheckResult *res, BdrvCheckMode fix)
5166 {
5167     if (bs->drv == NULL) {
5168         return -ENOMEDIUM;
5169     }
5170     if (bs->drv->bdrv_co_check == NULL) {
5171         return -ENOTSUP;
5172     }
5173 
5174     memset(res, 0, sizeof(*res));
5175     return bs->drv->bdrv_co_check(bs, res, fix);
5176 }
5177 
5178 /*
5179  * Return values:
5180  * 0        - success
5181  * -EINVAL  - backing format specified, but no file
5182  * -ENOSPC  - can't update the backing file because no space is left in the
5183  *            image file header
5184  * -ENOTSUP - format driver doesn't support changing the backing file
5185  */
5186 int bdrv_change_backing_file(BlockDriverState *bs, const char *backing_file,
5187                              const char *backing_fmt, bool require)
5188 {
5189     BlockDriver *drv = bs->drv;
5190     int ret;
5191 
5192     if (!drv) {
5193         return -ENOMEDIUM;
5194     }
5195 
5196     /* Backing file format doesn't make sense without a backing file */
5197     if (backing_fmt && !backing_file) {
5198         return -EINVAL;
5199     }
5200 
5201     if (require && backing_file && !backing_fmt) {
5202         return -EINVAL;
5203     }
5204 
5205     if (drv->bdrv_change_backing_file != NULL) {
5206         ret = drv->bdrv_change_backing_file(bs, backing_file, backing_fmt);
5207     } else {
5208         ret = -ENOTSUP;
5209     }
5210 
5211     if (ret == 0) {
5212         pstrcpy(bs->backing_file, sizeof(bs->backing_file), backing_file ?: "");
5213         pstrcpy(bs->backing_format, sizeof(bs->backing_format), backing_fmt ?: "");
5214         pstrcpy(bs->auto_backing_file, sizeof(bs->auto_backing_file),
5215                 backing_file ?: "");
5216     }
5217     return ret;
5218 }
5219 
5220 /*
5221  * Finds the first non-filter node above bs in the chain between
5222  * active and bs.  The returned node is either an immediate parent of
5223  * bs, or there are only filter nodes between the two.
5224  *
5225  * Returns NULL if bs is not found in active's image chain,
5226  * or if active == bs.
5227  *
5228  * Returns the bottommost base image if bs == NULL.
5229  */
5230 BlockDriverState *bdrv_find_overlay(BlockDriverState *active,
5231                                     BlockDriverState *bs)
5232 {
5233     bs = bdrv_skip_filters(bs);
5234     active = bdrv_skip_filters(active);
5235 
5236     while (active) {
5237         BlockDriverState *next = bdrv_backing_chain_next(active);
5238         if (bs == next) {
5239             return active;
5240         }
5241         active = next;
5242     }
5243 
5244     return NULL;
5245 }
5246 
5247 /* Given a BDS, searches for the base layer. */
5248 BlockDriverState *bdrv_find_base(BlockDriverState *bs)
5249 {
5250     return bdrv_find_overlay(bs, NULL);
5251 }
5252 
5253 /*
5254  * Return true if at least one of the COW (backing) and filter links
5255  * between @bs and @base is frozen. @errp is set if that's the case.
5256  * @base must be reachable from @bs, or NULL.
5257  */
5258 bool bdrv_is_backing_chain_frozen(BlockDriverState *bs, BlockDriverState *base,
5259                                   Error **errp)
5260 {
5261     BlockDriverState *i;
5262     BdrvChild *child;
5263 
5264     for (i = bs; i != base; i = child_bs(child)) {
5265         child = bdrv_filter_or_cow_child(i);
5266 
5267         if (child && child->frozen) {
5268             error_setg(errp, "Cannot change '%s' link from '%s' to '%s'",
5269                        child->name, i->node_name, child->bs->node_name);
5270             return true;
5271         }
5272     }
5273 
5274     return false;
5275 }
5276 
5277 /*
5278  * Freeze all COW (backing) and filter links between @bs and @base.
5279  * If any of the links is already frozen the operation is aborted and
5280  * none of the links are modified.
5281  * @base must be reachable from @bs, or NULL.
5282  * Returns 0 on success. On failure returns < 0 and sets @errp.
5283  */
5284 int bdrv_freeze_backing_chain(BlockDriverState *bs, BlockDriverState *base,
5285                               Error **errp)
5286 {
5287     BlockDriverState *i;
5288     BdrvChild *child;
5289 
5290     if (bdrv_is_backing_chain_frozen(bs, base, errp)) {
5291         return -EPERM;
5292     }
5293 
5294     for (i = bs; i != base; i = child_bs(child)) {
5295         child = bdrv_filter_or_cow_child(i);
5296         if (child && child->bs->never_freeze) {
5297             error_setg(errp, "Cannot freeze '%s' link to '%s'",
5298                        child->name, child->bs->node_name);
5299             return -EPERM;
5300         }
5301     }
5302 
5303     for (i = bs; i != base; i = child_bs(child)) {
5304         child = bdrv_filter_or_cow_child(i);
5305         if (child) {
5306             child->frozen = true;
5307         }
5308     }
5309 
5310     return 0;
5311 }
5312 
5313 /*
5314  * Unfreeze all COW (backing) and filter links between @bs and @base.
5315  * The caller must ensure that all links are frozen before using this
5316  * function.
5317  * @base must be reachable from @bs, or NULL.
5318  */
5319 void bdrv_unfreeze_backing_chain(BlockDriverState *bs, BlockDriverState *base)
5320 {
5321     BlockDriverState *i;
5322     BdrvChild *child;
5323 
5324     for (i = bs; i != base; i = child_bs(child)) {
5325         child = bdrv_filter_or_cow_child(i);
5326         if (child) {
5327             assert(child->frozen);
5328             child->frozen = false;
5329         }
5330     }
5331 }
5332 
5333 /*
5334  * Drops images above 'base' up to and including 'top', and sets the image
5335  * above 'top' to have base as its backing file.
5336  *
5337  * Requires that the overlay to 'top' is opened r/w, so that the backing file
5338  * information in 'bs' can be properly updated.
5339  *
5340  * E.g., this will convert the following chain:
5341  * bottom <- base <- intermediate <- top <- active
5342  *
5343  * to
5344  *
5345  * bottom <- base <- active
5346  *
5347  * It is allowed for bottom==base, in which case it converts:
5348  *
5349  * base <- intermediate <- top <- active
5350  *
5351  * to
5352  *
5353  * base <- active
5354  *
5355  * If backing_file_str is non-NULL, it will be used when modifying top's
5356  * overlay image metadata.
5357  *
5358  * Error conditions:
5359  *  if active == top, that is considered an error
5360  *
5361  */
5362 int bdrv_drop_intermediate(BlockDriverState *top, BlockDriverState *base,
5363                            const char *backing_file_str)
5364 {
5365     BlockDriverState *explicit_top = top;
5366     bool update_inherits_from;
5367     BdrvChild *c;
5368     Error *local_err = NULL;
5369     int ret = -EIO;
5370     g_autoptr(GSList) updated_children = NULL;
5371     GSList *p;
5372 
5373     bdrv_ref(top);
5374     bdrv_subtree_drained_begin(top);
5375 
5376     if (!top->drv || !base->drv) {
5377         goto exit;
5378     }
5379 
5380     /* Make sure that base is in the backing chain of top */
5381     if (!bdrv_chain_contains(top, base)) {
5382         goto exit;
5383     }
5384 
5385     /* If 'base' recursively inherits from 'top' then we should set
5386      * base->inherits_from to top->inherits_from after 'top' and all
5387      * other intermediate nodes have been dropped.
5388      * If 'top' is an implicit node (e.g. "commit_top") we should skip
5389      * it because no one inherits from it. We use explicit_top for that. */
5390     explicit_top = bdrv_skip_implicit_filters(explicit_top);
5391     update_inherits_from = bdrv_inherits_from_recursive(base, explicit_top);
5392 
5393     /* success - we can delete the intermediate states, and link top->base */
5394     /* TODO Check graph modification op blockers (BLK_PERM_GRAPH_MOD) once
5395      * we've figured out how they should work. */
5396     if (!backing_file_str) {
5397         bdrv_refresh_filename(base);
5398         backing_file_str = base->filename;
5399     }
5400 
5401     QLIST_FOREACH(c, &top->parents, next_parent) {
5402         updated_children = g_slist_prepend(updated_children, c);
5403     }
5404 
5405     /*
5406      * It seems correct to pass detach_subchain=true here, but it triggers
5407      * one more yet not fixed bug, when due to nested aio_poll loop we switch to
5408      * another drained section, which modify the graph (for example, removing
5409      * the child, which we keep in updated_children list). So, it's a TODO.
5410      *
5411      * Note, bug triggered if pass detach_subchain=true here and run
5412      * test-bdrv-drain. test_drop_intermediate_poll() test-case will crash.
5413      * That's a FIXME.
5414      */
5415     bdrv_replace_node_common(top, base, false, false, &local_err);
5416     if (local_err) {
5417         error_report_err(local_err);
5418         goto exit;
5419     }
5420 
5421     for (p = updated_children; p; p = p->next) {
5422         c = p->data;
5423 
5424         if (c->klass->update_filename) {
5425             ret = c->klass->update_filename(c, base, backing_file_str,
5426                                             &local_err);
5427             if (ret < 0) {
5428                 /*
5429                  * TODO: Actually, we want to rollback all previous iterations
5430                  * of this loop, and (which is almost impossible) previous
5431                  * bdrv_replace_node()...
5432                  *
5433                  * Note, that c->klass->update_filename may lead to permission
5434                  * update, so it's a bad idea to call it inside permission
5435                  * update transaction of bdrv_replace_node.
5436                  */
5437                 error_report_err(local_err);
5438                 goto exit;
5439             }
5440         }
5441     }
5442 
5443     if (update_inherits_from) {
5444         base->inherits_from = explicit_top->inherits_from;
5445     }
5446 
5447     ret = 0;
5448 exit:
5449     bdrv_subtree_drained_end(top);
5450     bdrv_unref(top);
5451     return ret;
5452 }
5453 
5454 /**
5455  * Implementation of BlockDriver.bdrv_get_allocated_file_size() that
5456  * sums the size of all data-bearing children.  (This excludes backing
5457  * children.)
5458  */
5459 static int64_t bdrv_sum_allocated_file_size(BlockDriverState *bs)
5460 {
5461     BdrvChild *child;
5462     int64_t child_size, sum = 0;
5463 
5464     QLIST_FOREACH(child, &bs->children, next) {
5465         if (child->role & (BDRV_CHILD_DATA | BDRV_CHILD_METADATA |
5466                            BDRV_CHILD_FILTERED))
5467         {
5468             child_size = bdrv_get_allocated_file_size(child->bs);
5469             if (child_size < 0) {
5470                 return child_size;
5471             }
5472             sum += child_size;
5473         }
5474     }
5475 
5476     return sum;
5477 }
5478 
5479 /**
5480  * Length of a allocated file in bytes. Sparse files are counted by actual
5481  * allocated space. Return < 0 if error or unknown.
5482  */
5483 int64_t bdrv_get_allocated_file_size(BlockDriverState *bs)
5484 {
5485     BlockDriver *drv = bs->drv;
5486     if (!drv) {
5487         return -ENOMEDIUM;
5488     }
5489     if (drv->bdrv_get_allocated_file_size) {
5490         return drv->bdrv_get_allocated_file_size(bs);
5491     }
5492 
5493     if (drv->bdrv_file_open) {
5494         /*
5495          * Protocol drivers default to -ENOTSUP (most of their data is
5496          * not stored in any of their children (if they even have any),
5497          * so there is no generic way to figure it out).
5498          */
5499         return -ENOTSUP;
5500     } else if (drv->is_filter) {
5501         /* Filter drivers default to the size of their filtered child */
5502         return bdrv_get_allocated_file_size(bdrv_filter_bs(bs));
5503     } else {
5504         /* Other drivers default to summing their children's sizes */
5505         return bdrv_sum_allocated_file_size(bs);
5506     }
5507 }
5508 
5509 /*
5510  * bdrv_measure:
5511  * @drv: Format driver
5512  * @opts: Creation options for new image
5513  * @in_bs: Existing image containing data for new image (may be NULL)
5514  * @errp: Error object
5515  * Returns: A #BlockMeasureInfo (free using qapi_free_BlockMeasureInfo())
5516  *          or NULL on error
5517  *
5518  * Calculate file size required to create a new image.
5519  *
5520  * If @in_bs is given then space for allocated clusters and zero clusters
5521  * from that image are included in the calculation.  If @opts contains a
5522  * backing file that is shared by @in_bs then backing clusters may be omitted
5523  * from the calculation.
5524  *
5525  * If @in_bs is NULL then the calculation includes no allocated clusters
5526  * unless a preallocation option is given in @opts.
5527  *
5528  * Note that @in_bs may use a different BlockDriver from @drv.
5529  *
5530  * If an error occurs the @errp pointer is set.
5531  */
5532 BlockMeasureInfo *bdrv_measure(BlockDriver *drv, QemuOpts *opts,
5533                                BlockDriverState *in_bs, Error **errp)
5534 {
5535     if (!drv->bdrv_measure) {
5536         error_setg(errp, "Block driver '%s' does not support size measurement",
5537                    drv->format_name);
5538         return NULL;
5539     }
5540 
5541     return drv->bdrv_measure(opts, in_bs, errp);
5542 }
5543 
5544 /**
5545  * Return number of sectors on success, -errno on error.
5546  */
5547 int64_t bdrv_nb_sectors(BlockDriverState *bs)
5548 {
5549     BlockDriver *drv = bs->drv;
5550 
5551     if (!drv)
5552         return -ENOMEDIUM;
5553 
5554     if (drv->has_variable_length) {
5555         int ret = refresh_total_sectors(bs, bs->total_sectors);
5556         if (ret < 0) {
5557             return ret;
5558         }
5559     }
5560     return bs->total_sectors;
5561 }
5562 
5563 /**
5564  * Return length in bytes on success, -errno on error.
5565  * The length is always a multiple of BDRV_SECTOR_SIZE.
5566  */
5567 int64_t bdrv_getlength(BlockDriverState *bs)
5568 {
5569     int64_t ret = bdrv_nb_sectors(bs);
5570 
5571     if (ret < 0) {
5572         return ret;
5573     }
5574     if (ret > INT64_MAX / BDRV_SECTOR_SIZE) {
5575         return -EFBIG;
5576     }
5577     return ret * BDRV_SECTOR_SIZE;
5578 }
5579 
5580 /* return 0 as number of sectors if no device present or error */
5581 void bdrv_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr)
5582 {
5583     int64_t nb_sectors = bdrv_nb_sectors(bs);
5584 
5585     *nb_sectors_ptr = nb_sectors < 0 ? 0 : nb_sectors;
5586 }
5587 
5588 bool bdrv_is_sg(BlockDriverState *bs)
5589 {
5590     return bs->sg;
5591 }
5592 
5593 /**
5594  * Return whether the given node supports compressed writes.
5595  */
5596 bool bdrv_supports_compressed_writes(BlockDriverState *bs)
5597 {
5598     BlockDriverState *filtered;
5599 
5600     if (!bs->drv || !block_driver_can_compress(bs->drv)) {
5601         return false;
5602     }
5603 
5604     filtered = bdrv_filter_bs(bs);
5605     if (filtered) {
5606         /*
5607          * Filters can only forward compressed writes, so we have to
5608          * check the child.
5609          */
5610         return bdrv_supports_compressed_writes(filtered);
5611     }
5612 
5613     return true;
5614 }
5615 
5616 const char *bdrv_get_format_name(BlockDriverState *bs)
5617 {
5618     return bs->drv ? bs->drv->format_name : NULL;
5619 }
5620 
5621 static int qsort_strcmp(const void *a, const void *b)
5622 {
5623     return strcmp(*(char *const *)a, *(char *const *)b);
5624 }
5625 
5626 void bdrv_iterate_format(void (*it)(void *opaque, const char *name),
5627                          void *opaque, bool read_only)
5628 {
5629     BlockDriver *drv;
5630     int count = 0;
5631     int i;
5632     const char **formats = NULL;
5633 
5634     QLIST_FOREACH(drv, &bdrv_drivers, list) {
5635         if (drv->format_name) {
5636             bool found = false;
5637             int i = count;
5638 
5639             if (use_bdrv_whitelist && !bdrv_is_whitelisted(drv, read_only)) {
5640                 continue;
5641             }
5642 
5643             while (formats && i && !found) {
5644                 found = !strcmp(formats[--i], drv->format_name);
5645             }
5646 
5647             if (!found) {
5648                 formats = g_renew(const char *, formats, count + 1);
5649                 formats[count++] = drv->format_name;
5650             }
5651         }
5652     }
5653 
5654     for (i = 0; i < (int)ARRAY_SIZE(block_driver_modules); i++) {
5655         const char *format_name = block_driver_modules[i].format_name;
5656 
5657         if (format_name) {
5658             bool found = false;
5659             int j = count;
5660 
5661             if (use_bdrv_whitelist &&
5662                 !bdrv_format_is_whitelisted(format_name, read_only)) {
5663                 continue;
5664             }
5665 
5666             while (formats && j && !found) {
5667                 found = !strcmp(formats[--j], format_name);
5668             }
5669 
5670             if (!found) {
5671                 formats = g_renew(const char *, formats, count + 1);
5672                 formats[count++] = format_name;
5673             }
5674         }
5675     }
5676 
5677     qsort(formats, count, sizeof(formats[0]), qsort_strcmp);
5678 
5679     for (i = 0; i < count; i++) {
5680         it(opaque, formats[i]);
5681     }
5682 
5683     g_free(formats);
5684 }
5685 
5686 /* This function is to find a node in the bs graph */
5687 BlockDriverState *bdrv_find_node(const char *node_name)
5688 {
5689     BlockDriverState *bs;
5690 
5691     assert(node_name);
5692 
5693     QTAILQ_FOREACH(bs, &graph_bdrv_states, node_list) {
5694         if (!strcmp(node_name, bs->node_name)) {
5695             return bs;
5696         }
5697     }
5698     return NULL;
5699 }
5700 
5701 /* Put this QMP function here so it can access the static graph_bdrv_states. */
5702 BlockDeviceInfoList *bdrv_named_nodes_list(bool flat,
5703                                            Error **errp)
5704 {
5705     BlockDeviceInfoList *list;
5706     BlockDriverState *bs;
5707 
5708     list = NULL;
5709     QTAILQ_FOREACH(bs, &graph_bdrv_states, node_list) {
5710         BlockDeviceInfo *info = bdrv_block_device_info(NULL, bs, flat, errp);
5711         if (!info) {
5712             qapi_free_BlockDeviceInfoList(list);
5713             return NULL;
5714         }
5715         QAPI_LIST_PREPEND(list, info);
5716     }
5717 
5718     return list;
5719 }
5720 
5721 typedef struct XDbgBlockGraphConstructor {
5722     XDbgBlockGraph *graph;
5723     GHashTable *graph_nodes;
5724 } XDbgBlockGraphConstructor;
5725 
5726 static XDbgBlockGraphConstructor *xdbg_graph_new(void)
5727 {
5728     XDbgBlockGraphConstructor *gr = g_new(XDbgBlockGraphConstructor, 1);
5729 
5730     gr->graph = g_new0(XDbgBlockGraph, 1);
5731     gr->graph_nodes = g_hash_table_new(NULL, NULL);
5732 
5733     return gr;
5734 }
5735 
5736 static XDbgBlockGraph *xdbg_graph_finalize(XDbgBlockGraphConstructor *gr)
5737 {
5738     XDbgBlockGraph *graph = gr->graph;
5739 
5740     g_hash_table_destroy(gr->graph_nodes);
5741     g_free(gr);
5742 
5743     return graph;
5744 }
5745 
5746 static uintptr_t xdbg_graph_node_num(XDbgBlockGraphConstructor *gr, void *node)
5747 {
5748     uintptr_t ret = (uintptr_t)g_hash_table_lookup(gr->graph_nodes, node);
5749 
5750     if (ret != 0) {
5751         return ret;
5752     }
5753 
5754     /*
5755      * Start counting from 1, not 0, because 0 interferes with not-found (NULL)
5756      * answer of g_hash_table_lookup.
5757      */
5758     ret = g_hash_table_size(gr->graph_nodes) + 1;
5759     g_hash_table_insert(gr->graph_nodes, node, (void *)ret);
5760 
5761     return ret;
5762 }
5763 
5764 static void xdbg_graph_add_node(XDbgBlockGraphConstructor *gr, void *node,
5765                                 XDbgBlockGraphNodeType type, const char *name)
5766 {
5767     XDbgBlockGraphNode *n;
5768 
5769     n = g_new0(XDbgBlockGraphNode, 1);
5770 
5771     n->id = xdbg_graph_node_num(gr, node);
5772     n->type = type;
5773     n->name = g_strdup(name);
5774 
5775     QAPI_LIST_PREPEND(gr->graph->nodes, n);
5776 }
5777 
5778 static void xdbg_graph_add_edge(XDbgBlockGraphConstructor *gr, void *parent,
5779                                 const BdrvChild *child)
5780 {
5781     BlockPermission qapi_perm;
5782     XDbgBlockGraphEdge *edge;
5783 
5784     edge = g_new0(XDbgBlockGraphEdge, 1);
5785 
5786     edge->parent = xdbg_graph_node_num(gr, parent);
5787     edge->child = xdbg_graph_node_num(gr, child->bs);
5788     edge->name = g_strdup(child->name);
5789 
5790     for (qapi_perm = 0; qapi_perm < BLOCK_PERMISSION__MAX; qapi_perm++) {
5791         uint64_t flag = bdrv_qapi_perm_to_blk_perm(qapi_perm);
5792 
5793         if (flag & child->perm) {
5794             QAPI_LIST_PREPEND(edge->perm, qapi_perm);
5795         }
5796         if (flag & child->shared_perm) {
5797             QAPI_LIST_PREPEND(edge->shared_perm, qapi_perm);
5798         }
5799     }
5800 
5801     QAPI_LIST_PREPEND(gr->graph->edges, edge);
5802 }
5803 
5804 
5805 XDbgBlockGraph *bdrv_get_xdbg_block_graph(Error **errp)
5806 {
5807     BlockBackend *blk;
5808     BlockJob *job;
5809     BlockDriverState *bs;
5810     BdrvChild *child;
5811     XDbgBlockGraphConstructor *gr = xdbg_graph_new();
5812 
5813     for (blk = blk_all_next(NULL); blk; blk = blk_all_next(blk)) {
5814         char *allocated_name = NULL;
5815         const char *name = blk_name(blk);
5816 
5817         if (!*name) {
5818             name = allocated_name = blk_get_attached_dev_id(blk);
5819         }
5820         xdbg_graph_add_node(gr, blk, X_DBG_BLOCK_GRAPH_NODE_TYPE_BLOCK_BACKEND,
5821                            name);
5822         g_free(allocated_name);
5823         if (blk_root(blk)) {
5824             xdbg_graph_add_edge(gr, blk, blk_root(blk));
5825         }
5826     }
5827 
5828     for (job = block_job_next(NULL); job; job = block_job_next(job)) {
5829         GSList *el;
5830 
5831         xdbg_graph_add_node(gr, job, X_DBG_BLOCK_GRAPH_NODE_TYPE_BLOCK_JOB,
5832                            job->job.id);
5833         for (el = job->nodes; el; el = el->next) {
5834             xdbg_graph_add_edge(gr, job, (BdrvChild *)el->data);
5835         }
5836     }
5837 
5838     QTAILQ_FOREACH(bs, &graph_bdrv_states, node_list) {
5839         xdbg_graph_add_node(gr, bs, X_DBG_BLOCK_GRAPH_NODE_TYPE_BLOCK_DRIVER,
5840                            bs->node_name);
5841         QLIST_FOREACH(child, &bs->children, next) {
5842             xdbg_graph_add_edge(gr, bs, child);
5843         }
5844     }
5845 
5846     return xdbg_graph_finalize(gr);
5847 }
5848 
5849 BlockDriverState *bdrv_lookup_bs(const char *device,
5850                                  const char *node_name,
5851                                  Error **errp)
5852 {
5853     BlockBackend *blk;
5854     BlockDriverState *bs;
5855 
5856     if (device) {
5857         blk = blk_by_name(device);
5858 
5859         if (blk) {
5860             bs = blk_bs(blk);
5861             if (!bs) {
5862                 error_setg(errp, "Device '%s' has no medium", device);
5863             }
5864 
5865             return bs;
5866         }
5867     }
5868 
5869     if (node_name) {
5870         bs = bdrv_find_node(node_name);
5871 
5872         if (bs) {
5873             return bs;
5874         }
5875     }
5876 
5877     error_setg(errp, "Cannot find device=\'%s\' nor node-name=\'%s\'",
5878                      device ? device : "",
5879                      node_name ? node_name : "");
5880     return NULL;
5881 }
5882 
5883 /* If 'base' is in the same chain as 'top', return true. Otherwise,
5884  * return false.  If either argument is NULL, return false. */
5885 bool bdrv_chain_contains(BlockDriverState *top, BlockDriverState *base)
5886 {
5887     while (top && top != base) {
5888         top = bdrv_filter_or_cow_bs(top);
5889     }
5890 
5891     return top != NULL;
5892 }
5893 
5894 BlockDriverState *bdrv_next_node(BlockDriverState *bs)
5895 {
5896     if (!bs) {
5897         return QTAILQ_FIRST(&graph_bdrv_states);
5898     }
5899     return QTAILQ_NEXT(bs, node_list);
5900 }
5901 
5902 BlockDriverState *bdrv_next_all_states(BlockDriverState *bs)
5903 {
5904     if (!bs) {
5905         return QTAILQ_FIRST(&all_bdrv_states);
5906     }
5907     return QTAILQ_NEXT(bs, bs_list);
5908 }
5909 
5910 const char *bdrv_get_node_name(const BlockDriverState *bs)
5911 {
5912     return bs->node_name;
5913 }
5914 
5915 const char *bdrv_get_parent_name(const BlockDriverState *bs)
5916 {
5917     BdrvChild *c;
5918     const char *name;
5919 
5920     /* If multiple parents have a name, just pick the first one. */
5921     QLIST_FOREACH(c, &bs->parents, next_parent) {
5922         if (c->klass->get_name) {
5923             name = c->klass->get_name(c);
5924             if (name && *name) {
5925                 return name;
5926             }
5927         }
5928     }
5929 
5930     return NULL;
5931 }
5932 
5933 /* TODO check what callers really want: bs->node_name or blk_name() */
5934 const char *bdrv_get_device_name(const BlockDriverState *bs)
5935 {
5936     return bdrv_get_parent_name(bs) ?: "";
5937 }
5938 
5939 /* This can be used to identify nodes that might not have a device
5940  * name associated. Since node and device names live in the same
5941  * namespace, the result is unambiguous. The exception is if both are
5942  * absent, then this returns an empty (non-null) string. */
5943 const char *bdrv_get_device_or_node_name(const BlockDriverState *bs)
5944 {
5945     return bdrv_get_parent_name(bs) ?: bs->node_name;
5946 }
5947 
5948 int bdrv_get_flags(BlockDriverState *bs)
5949 {
5950     return bs->open_flags;
5951 }
5952 
5953 int bdrv_has_zero_init_1(BlockDriverState *bs)
5954 {
5955     return 1;
5956 }
5957 
5958 int bdrv_has_zero_init(BlockDriverState *bs)
5959 {
5960     BlockDriverState *filtered;
5961 
5962     if (!bs->drv) {
5963         return 0;
5964     }
5965 
5966     /* If BS is a copy on write image, it is initialized to
5967        the contents of the base image, which may not be zeroes.  */
5968     if (bdrv_cow_child(bs)) {
5969         return 0;
5970     }
5971     if (bs->drv->bdrv_has_zero_init) {
5972         return bs->drv->bdrv_has_zero_init(bs);
5973     }
5974 
5975     filtered = bdrv_filter_bs(bs);
5976     if (filtered) {
5977         return bdrv_has_zero_init(filtered);
5978     }
5979 
5980     /* safe default */
5981     return 0;
5982 }
5983 
5984 bool bdrv_can_write_zeroes_with_unmap(BlockDriverState *bs)
5985 {
5986     if (!(bs->open_flags & BDRV_O_UNMAP)) {
5987         return false;
5988     }
5989 
5990     return bs->supported_zero_flags & BDRV_REQ_MAY_UNMAP;
5991 }
5992 
5993 void bdrv_get_backing_filename(BlockDriverState *bs,
5994                                char *filename, int filename_size)
5995 {
5996     pstrcpy(filename, filename_size, bs->backing_file);
5997 }
5998 
5999 int bdrv_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
6000 {
6001     int ret;
6002     BlockDriver *drv = bs->drv;
6003     /* if bs->drv == NULL, bs is closed, so there's nothing to do here */
6004     if (!drv) {
6005         return -ENOMEDIUM;
6006     }
6007     if (!drv->bdrv_get_info) {
6008         BlockDriverState *filtered = bdrv_filter_bs(bs);
6009         if (filtered) {
6010             return bdrv_get_info(filtered, bdi);
6011         }
6012         return -ENOTSUP;
6013     }
6014     memset(bdi, 0, sizeof(*bdi));
6015     ret = drv->bdrv_get_info(bs, bdi);
6016     if (ret < 0) {
6017         return ret;
6018     }
6019 
6020     if (bdi->cluster_size > BDRV_MAX_ALIGNMENT) {
6021         return -EINVAL;
6022     }
6023 
6024     return 0;
6025 }
6026 
6027 ImageInfoSpecific *bdrv_get_specific_info(BlockDriverState *bs,
6028                                           Error **errp)
6029 {
6030     BlockDriver *drv = bs->drv;
6031     if (drv && drv->bdrv_get_specific_info) {
6032         return drv->bdrv_get_specific_info(bs, errp);
6033     }
6034     return NULL;
6035 }
6036 
6037 BlockStatsSpecific *bdrv_get_specific_stats(BlockDriverState *bs)
6038 {
6039     BlockDriver *drv = bs->drv;
6040     if (!drv || !drv->bdrv_get_specific_stats) {
6041         return NULL;
6042     }
6043     return drv->bdrv_get_specific_stats(bs);
6044 }
6045 
6046 void bdrv_debug_event(BlockDriverState *bs, BlkdebugEvent event)
6047 {
6048     if (!bs || !bs->drv || !bs->drv->bdrv_debug_event) {
6049         return;
6050     }
6051 
6052     bs->drv->bdrv_debug_event(bs, event);
6053 }
6054 
6055 static BlockDriverState *bdrv_find_debug_node(BlockDriverState *bs)
6056 {
6057     while (bs && bs->drv && !bs->drv->bdrv_debug_breakpoint) {
6058         bs = bdrv_primary_bs(bs);
6059     }
6060 
6061     if (bs && bs->drv && bs->drv->bdrv_debug_breakpoint) {
6062         assert(bs->drv->bdrv_debug_remove_breakpoint);
6063         return bs;
6064     }
6065 
6066     return NULL;
6067 }
6068 
6069 int bdrv_debug_breakpoint(BlockDriverState *bs, const char *event,
6070                           const char *tag)
6071 {
6072     bs = bdrv_find_debug_node(bs);
6073     if (bs) {
6074         return bs->drv->bdrv_debug_breakpoint(bs, event, tag);
6075     }
6076 
6077     return -ENOTSUP;
6078 }
6079 
6080 int bdrv_debug_remove_breakpoint(BlockDriverState *bs, const char *tag)
6081 {
6082     bs = bdrv_find_debug_node(bs);
6083     if (bs) {
6084         return bs->drv->bdrv_debug_remove_breakpoint(bs, tag);
6085     }
6086 
6087     return -ENOTSUP;
6088 }
6089 
6090 int bdrv_debug_resume(BlockDriverState *bs, const char *tag)
6091 {
6092     while (bs && (!bs->drv || !bs->drv->bdrv_debug_resume)) {
6093         bs = bdrv_primary_bs(bs);
6094     }
6095 
6096     if (bs && bs->drv && bs->drv->bdrv_debug_resume) {
6097         return bs->drv->bdrv_debug_resume(bs, tag);
6098     }
6099 
6100     return -ENOTSUP;
6101 }
6102 
6103 bool bdrv_debug_is_suspended(BlockDriverState *bs, const char *tag)
6104 {
6105     while (bs && bs->drv && !bs->drv->bdrv_debug_is_suspended) {
6106         bs = bdrv_primary_bs(bs);
6107     }
6108 
6109     if (bs && bs->drv && bs->drv->bdrv_debug_is_suspended) {
6110         return bs->drv->bdrv_debug_is_suspended(bs, tag);
6111     }
6112 
6113     return false;
6114 }
6115 
6116 /* backing_file can either be relative, or absolute, or a protocol.  If it is
6117  * relative, it must be relative to the chain.  So, passing in bs->filename
6118  * from a BDS as backing_file should not be done, as that may be relative to
6119  * the CWD rather than the chain. */
6120 BlockDriverState *bdrv_find_backing_image(BlockDriverState *bs,
6121         const char *backing_file)
6122 {
6123     char *filename_full = NULL;
6124     char *backing_file_full = NULL;
6125     char *filename_tmp = NULL;
6126     int is_protocol = 0;
6127     bool filenames_refreshed = false;
6128     BlockDriverState *curr_bs = NULL;
6129     BlockDriverState *retval = NULL;
6130     BlockDriverState *bs_below;
6131 
6132     if (!bs || !bs->drv || !backing_file) {
6133         return NULL;
6134     }
6135 
6136     filename_full     = g_malloc(PATH_MAX);
6137     backing_file_full = g_malloc(PATH_MAX);
6138 
6139     is_protocol = path_has_protocol(backing_file);
6140 
6141     /*
6142      * Being largely a legacy function, skip any filters here
6143      * (because filters do not have normal filenames, so they cannot
6144      * match anyway; and allowing json:{} filenames is a bit out of
6145      * scope).
6146      */
6147     for (curr_bs = bdrv_skip_filters(bs);
6148          bdrv_cow_child(curr_bs) != NULL;
6149          curr_bs = bs_below)
6150     {
6151         bs_below = bdrv_backing_chain_next(curr_bs);
6152 
6153         if (bdrv_backing_overridden(curr_bs)) {
6154             /*
6155              * If the backing file was overridden, we can only compare
6156              * directly against the backing node's filename.
6157              */
6158 
6159             if (!filenames_refreshed) {
6160                 /*
6161                  * This will automatically refresh all of the
6162                  * filenames in the rest of the backing chain, so we
6163                  * only need to do this once.
6164                  */
6165                 bdrv_refresh_filename(bs_below);
6166                 filenames_refreshed = true;
6167             }
6168 
6169             if (strcmp(backing_file, bs_below->filename) == 0) {
6170                 retval = bs_below;
6171                 break;
6172             }
6173         } else if (is_protocol || path_has_protocol(curr_bs->backing_file)) {
6174             /*
6175              * If either of the filename paths is actually a protocol, then
6176              * compare unmodified paths; otherwise make paths relative.
6177              */
6178             char *backing_file_full_ret;
6179 
6180             if (strcmp(backing_file, curr_bs->backing_file) == 0) {
6181                 retval = bs_below;
6182                 break;
6183             }
6184             /* Also check against the full backing filename for the image */
6185             backing_file_full_ret = bdrv_get_full_backing_filename(curr_bs,
6186                                                                    NULL);
6187             if (backing_file_full_ret) {
6188                 bool equal = strcmp(backing_file, backing_file_full_ret) == 0;
6189                 g_free(backing_file_full_ret);
6190                 if (equal) {
6191                     retval = bs_below;
6192                     break;
6193                 }
6194             }
6195         } else {
6196             /* If not an absolute filename path, make it relative to the current
6197              * image's filename path */
6198             filename_tmp = bdrv_make_absolute_filename(curr_bs, backing_file,
6199                                                        NULL);
6200             /* We are going to compare canonicalized absolute pathnames */
6201             if (!filename_tmp || !realpath(filename_tmp, filename_full)) {
6202                 g_free(filename_tmp);
6203                 continue;
6204             }
6205             g_free(filename_tmp);
6206 
6207             /* We need to make sure the backing filename we are comparing against
6208              * is relative to the current image filename (or absolute) */
6209             filename_tmp = bdrv_get_full_backing_filename(curr_bs, NULL);
6210             if (!filename_tmp || !realpath(filename_tmp, backing_file_full)) {
6211                 g_free(filename_tmp);
6212                 continue;
6213             }
6214             g_free(filename_tmp);
6215 
6216             if (strcmp(backing_file_full, filename_full) == 0) {
6217                 retval = bs_below;
6218                 break;
6219             }
6220         }
6221     }
6222 
6223     g_free(filename_full);
6224     g_free(backing_file_full);
6225     return retval;
6226 }
6227 
6228 void bdrv_init(void)
6229 {
6230 #ifdef CONFIG_BDRV_WHITELIST_TOOLS
6231     use_bdrv_whitelist = 1;
6232 #endif
6233     module_call_init(MODULE_INIT_BLOCK);
6234 }
6235 
6236 void bdrv_init_with_whitelist(void)
6237 {
6238     use_bdrv_whitelist = 1;
6239     bdrv_init();
6240 }
6241 
6242 int coroutine_fn bdrv_co_invalidate_cache(BlockDriverState *bs, Error **errp)
6243 {
6244     BdrvChild *child, *parent;
6245     Error *local_err = NULL;
6246     int ret;
6247     BdrvDirtyBitmap *bm;
6248 
6249     if (!bs->drv)  {
6250         return -ENOMEDIUM;
6251     }
6252 
6253     QLIST_FOREACH(child, &bs->children, next) {
6254         bdrv_co_invalidate_cache(child->bs, &local_err);
6255         if (local_err) {
6256             error_propagate(errp, local_err);
6257             return -EINVAL;
6258         }
6259     }
6260 
6261     /*
6262      * Update permissions, they may differ for inactive nodes.
6263      *
6264      * Note that the required permissions of inactive images are always a
6265      * subset of the permissions required after activating the image. This
6266      * allows us to just get the permissions upfront without restricting
6267      * drv->bdrv_invalidate_cache().
6268      *
6269      * It also means that in error cases, we don't have to try and revert to
6270      * the old permissions (which is an operation that could fail, too). We can
6271      * just keep the extended permissions for the next time that an activation
6272      * of the image is tried.
6273      */
6274     if (bs->open_flags & BDRV_O_INACTIVE) {
6275         bs->open_flags &= ~BDRV_O_INACTIVE;
6276         ret = bdrv_refresh_perms(bs, errp);
6277         if (ret < 0) {
6278             bs->open_flags |= BDRV_O_INACTIVE;
6279             return ret;
6280         }
6281 
6282         if (bs->drv->bdrv_co_invalidate_cache) {
6283             bs->drv->bdrv_co_invalidate_cache(bs, &local_err);
6284             if (local_err) {
6285                 bs->open_flags |= BDRV_O_INACTIVE;
6286                 error_propagate(errp, local_err);
6287                 return -EINVAL;
6288             }
6289         }
6290 
6291         FOR_EACH_DIRTY_BITMAP(bs, bm) {
6292             bdrv_dirty_bitmap_skip_store(bm, false);
6293         }
6294 
6295         ret = refresh_total_sectors(bs, bs->total_sectors);
6296         if (ret < 0) {
6297             bs->open_flags |= BDRV_O_INACTIVE;
6298             error_setg_errno(errp, -ret, "Could not refresh total sector count");
6299             return ret;
6300         }
6301     }
6302 
6303     QLIST_FOREACH(parent, &bs->parents, next_parent) {
6304         if (parent->klass->activate) {
6305             parent->klass->activate(parent, &local_err);
6306             if (local_err) {
6307                 bs->open_flags |= BDRV_O_INACTIVE;
6308                 error_propagate(errp, local_err);
6309                 return -EINVAL;
6310             }
6311         }
6312     }
6313 
6314     return 0;
6315 }
6316 
6317 void bdrv_invalidate_cache_all(Error **errp)
6318 {
6319     BlockDriverState *bs;
6320     BdrvNextIterator it;
6321 
6322     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
6323         AioContext *aio_context = bdrv_get_aio_context(bs);
6324         int ret;
6325 
6326         aio_context_acquire(aio_context);
6327         ret = bdrv_invalidate_cache(bs, errp);
6328         aio_context_release(aio_context);
6329         if (ret < 0) {
6330             bdrv_next_cleanup(&it);
6331             return;
6332         }
6333     }
6334 }
6335 
6336 static bool bdrv_has_bds_parent(BlockDriverState *bs, bool only_active)
6337 {
6338     BdrvChild *parent;
6339 
6340     QLIST_FOREACH(parent, &bs->parents, next_parent) {
6341         if (parent->klass->parent_is_bds) {
6342             BlockDriverState *parent_bs = parent->opaque;
6343             if (!only_active || !(parent_bs->open_flags & BDRV_O_INACTIVE)) {
6344                 return true;
6345             }
6346         }
6347     }
6348 
6349     return false;
6350 }
6351 
6352 static int bdrv_inactivate_recurse(BlockDriverState *bs)
6353 {
6354     BdrvChild *child, *parent;
6355     int ret;
6356     uint64_t cumulative_perms, cumulative_shared_perms;
6357 
6358     if (!bs->drv) {
6359         return -ENOMEDIUM;
6360     }
6361 
6362     /* Make sure that we don't inactivate a child before its parent.
6363      * It will be covered by recursion from the yet active parent. */
6364     if (bdrv_has_bds_parent(bs, true)) {
6365         return 0;
6366     }
6367 
6368     assert(!(bs->open_flags & BDRV_O_INACTIVE));
6369 
6370     /* Inactivate this node */
6371     if (bs->drv->bdrv_inactivate) {
6372         ret = bs->drv->bdrv_inactivate(bs);
6373         if (ret < 0) {
6374             return ret;
6375         }
6376     }
6377 
6378     QLIST_FOREACH(parent, &bs->parents, next_parent) {
6379         if (parent->klass->inactivate) {
6380             ret = parent->klass->inactivate(parent);
6381             if (ret < 0) {
6382                 return ret;
6383             }
6384         }
6385     }
6386 
6387     bdrv_get_cumulative_perm(bs, &cumulative_perms,
6388                              &cumulative_shared_perms);
6389     if (cumulative_perms & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED)) {
6390         /* Our inactive parents still need write access. Inactivation failed. */
6391         return -EPERM;
6392     }
6393 
6394     bs->open_flags |= BDRV_O_INACTIVE;
6395 
6396     /*
6397      * Update permissions, they may differ for inactive nodes.
6398      * We only tried to loosen restrictions, so errors are not fatal, ignore
6399      * them.
6400      */
6401     bdrv_refresh_perms(bs, NULL);
6402 
6403     /* Recursively inactivate children */
6404     QLIST_FOREACH(child, &bs->children, next) {
6405         ret = bdrv_inactivate_recurse(child->bs);
6406         if (ret < 0) {
6407             return ret;
6408         }
6409     }
6410 
6411     return 0;
6412 }
6413 
6414 int bdrv_inactivate_all(void)
6415 {
6416     BlockDriverState *bs = NULL;
6417     BdrvNextIterator it;
6418     int ret = 0;
6419     GSList *aio_ctxs = NULL, *ctx;
6420 
6421     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
6422         AioContext *aio_context = bdrv_get_aio_context(bs);
6423 
6424         if (!g_slist_find(aio_ctxs, aio_context)) {
6425             aio_ctxs = g_slist_prepend(aio_ctxs, aio_context);
6426             aio_context_acquire(aio_context);
6427         }
6428     }
6429 
6430     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
6431         /* Nodes with BDS parents are covered by recursion from the last
6432          * parent that gets inactivated. Don't inactivate them a second
6433          * time if that has already happened. */
6434         if (bdrv_has_bds_parent(bs, false)) {
6435             continue;
6436         }
6437         ret = bdrv_inactivate_recurse(bs);
6438         if (ret < 0) {
6439             bdrv_next_cleanup(&it);
6440             goto out;
6441         }
6442     }
6443 
6444 out:
6445     for (ctx = aio_ctxs; ctx != NULL; ctx = ctx->next) {
6446         AioContext *aio_context = ctx->data;
6447         aio_context_release(aio_context);
6448     }
6449     g_slist_free(aio_ctxs);
6450 
6451     return ret;
6452 }
6453 
6454 /**************************************************************/
6455 /* removable device support */
6456 
6457 /**
6458  * Return TRUE if the media is present
6459  */
6460 bool bdrv_is_inserted(BlockDriverState *bs)
6461 {
6462     BlockDriver *drv = bs->drv;
6463     BdrvChild *child;
6464 
6465     if (!drv) {
6466         return false;
6467     }
6468     if (drv->bdrv_is_inserted) {
6469         return drv->bdrv_is_inserted(bs);
6470     }
6471     QLIST_FOREACH(child, &bs->children, next) {
6472         if (!bdrv_is_inserted(child->bs)) {
6473             return false;
6474         }
6475     }
6476     return true;
6477 }
6478 
6479 /**
6480  * If eject_flag is TRUE, eject the media. Otherwise, close the tray
6481  */
6482 void bdrv_eject(BlockDriverState *bs, bool eject_flag)
6483 {
6484     BlockDriver *drv = bs->drv;
6485 
6486     if (drv && drv->bdrv_eject) {
6487         drv->bdrv_eject(bs, eject_flag);
6488     }
6489 }
6490 
6491 /**
6492  * Lock or unlock the media (if it is locked, the user won't be able
6493  * to eject it manually).
6494  */
6495 void bdrv_lock_medium(BlockDriverState *bs, bool locked)
6496 {
6497     BlockDriver *drv = bs->drv;
6498 
6499     trace_bdrv_lock_medium(bs, locked);
6500 
6501     if (drv && drv->bdrv_lock_medium) {
6502         drv->bdrv_lock_medium(bs, locked);
6503     }
6504 }
6505 
6506 /* Get a reference to bs */
6507 void bdrv_ref(BlockDriverState *bs)
6508 {
6509     bs->refcnt++;
6510 }
6511 
6512 /* Release a previously grabbed reference to bs.
6513  * If after releasing, reference count is zero, the BlockDriverState is
6514  * deleted. */
6515 void bdrv_unref(BlockDriverState *bs)
6516 {
6517     if (!bs) {
6518         return;
6519     }
6520     assert(bs->refcnt > 0);
6521     if (--bs->refcnt == 0) {
6522         bdrv_delete(bs);
6523     }
6524 }
6525 
6526 struct BdrvOpBlocker {
6527     Error *reason;
6528     QLIST_ENTRY(BdrvOpBlocker) list;
6529 };
6530 
6531 bool bdrv_op_is_blocked(BlockDriverState *bs, BlockOpType op, Error **errp)
6532 {
6533     BdrvOpBlocker *blocker;
6534     assert((int) op >= 0 && op < BLOCK_OP_TYPE_MAX);
6535     if (!QLIST_EMPTY(&bs->op_blockers[op])) {
6536         blocker = QLIST_FIRST(&bs->op_blockers[op]);
6537         error_propagate_prepend(errp, error_copy(blocker->reason),
6538                                 "Node '%s' is busy: ",
6539                                 bdrv_get_device_or_node_name(bs));
6540         return true;
6541     }
6542     return false;
6543 }
6544 
6545 void bdrv_op_block(BlockDriverState *bs, BlockOpType op, Error *reason)
6546 {
6547     BdrvOpBlocker *blocker;
6548     assert((int) op >= 0 && op < BLOCK_OP_TYPE_MAX);
6549 
6550     blocker = g_new0(BdrvOpBlocker, 1);
6551     blocker->reason = reason;
6552     QLIST_INSERT_HEAD(&bs->op_blockers[op], blocker, list);
6553 }
6554 
6555 void bdrv_op_unblock(BlockDriverState *bs, BlockOpType op, Error *reason)
6556 {
6557     BdrvOpBlocker *blocker, *next;
6558     assert((int) op >= 0 && op < BLOCK_OP_TYPE_MAX);
6559     QLIST_FOREACH_SAFE(blocker, &bs->op_blockers[op], list, next) {
6560         if (blocker->reason == reason) {
6561             QLIST_REMOVE(blocker, list);
6562             g_free(blocker);
6563         }
6564     }
6565 }
6566 
6567 void bdrv_op_block_all(BlockDriverState *bs, Error *reason)
6568 {
6569     int i;
6570     for (i = 0; i < BLOCK_OP_TYPE_MAX; i++) {
6571         bdrv_op_block(bs, i, reason);
6572     }
6573 }
6574 
6575 void bdrv_op_unblock_all(BlockDriverState *bs, Error *reason)
6576 {
6577     int i;
6578     for (i = 0; i < BLOCK_OP_TYPE_MAX; i++) {
6579         bdrv_op_unblock(bs, i, reason);
6580     }
6581 }
6582 
6583 bool bdrv_op_blocker_is_empty(BlockDriverState *bs)
6584 {
6585     int i;
6586 
6587     for (i = 0; i < BLOCK_OP_TYPE_MAX; i++) {
6588         if (!QLIST_EMPTY(&bs->op_blockers[i])) {
6589             return false;
6590         }
6591     }
6592     return true;
6593 }
6594 
6595 void bdrv_img_create(const char *filename, const char *fmt,
6596                      const char *base_filename, const char *base_fmt,
6597                      char *options, uint64_t img_size, int flags, bool quiet,
6598                      Error **errp)
6599 {
6600     QemuOptsList *create_opts = NULL;
6601     QemuOpts *opts = NULL;
6602     const char *backing_fmt, *backing_file;
6603     int64_t size;
6604     BlockDriver *drv, *proto_drv;
6605     Error *local_err = NULL;
6606     int ret = 0;
6607 
6608     /* Find driver and parse its options */
6609     drv = bdrv_find_format(fmt);
6610     if (!drv) {
6611         error_setg(errp, "Unknown file format '%s'", fmt);
6612         return;
6613     }
6614 
6615     proto_drv = bdrv_find_protocol(filename, true, errp);
6616     if (!proto_drv) {
6617         return;
6618     }
6619 
6620     if (!drv->create_opts) {
6621         error_setg(errp, "Format driver '%s' does not support image creation",
6622                    drv->format_name);
6623         return;
6624     }
6625 
6626     if (!proto_drv->create_opts) {
6627         error_setg(errp, "Protocol driver '%s' does not support image creation",
6628                    proto_drv->format_name);
6629         return;
6630     }
6631 
6632     /* Create parameter list */
6633     create_opts = qemu_opts_append(create_opts, drv->create_opts);
6634     create_opts = qemu_opts_append(create_opts, proto_drv->create_opts);
6635 
6636     opts = qemu_opts_create(create_opts, NULL, 0, &error_abort);
6637 
6638     /* Parse -o options */
6639     if (options) {
6640         if (!qemu_opts_do_parse(opts, options, NULL, errp)) {
6641             goto out;
6642         }
6643     }
6644 
6645     if (!qemu_opt_get(opts, BLOCK_OPT_SIZE)) {
6646         qemu_opt_set_number(opts, BLOCK_OPT_SIZE, img_size, &error_abort);
6647     } else if (img_size != UINT64_C(-1)) {
6648         error_setg(errp, "The image size must be specified only once");
6649         goto out;
6650     }
6651 
6652     if (base_filename) {
6653         if (!qemu_opt_set(opts, BLOCK_OPT_BACKING_FILE, base_filename,
6654                           NULL)) {
6655             error_setg(errp, "Backing file not supported for file format '%s'",
6656                        fmt);
6657             goto out;
6658         }
6659     }
6660 
6661     if (base_fmt) {
6662         if (!qemu_opt_set(opts, BLOCK_OPT_BACKING_FMT, base_fmt, NULL)) {
6663             error_setg(errp, "Backing file format not supported for file "
6664                              "format '%s'", fmt);
6665             goto out;
6666         }
6667     }
6668 
6669     backing_file = qemu_opt_get(opts, BLOCK_OPT_BACKING_FILE);
6670     if (backing_file) {
6671         if (!strcmp(filename, backing_file)) {
6672             error_setg(errp, "Error: Trying to create an image with the "
6673                              "same filename as the backing file");
6674             goto out;
6675         }
6676         if (backing_file[0] == '\0') {
6677             error_setg(errp, "Expected backing file name, got empty string");
6678             goto out;
6679         }
6680     }
6681 
6682     backing_fmt = qemu_opt_get(opts, BLOCK_OPT_BACKING_FMT);
6683 
6684     /* The size for the image must always be specified, unless we have a backing
6685      * file and we have not been forbidden from opening it. */
6686     size = qemu_opt_get_size(opts, BLOCK_OPT_SIZE, img_size);
6687     if (backing_file && !(flags & BDRV_O_NO_BACKING)) {
6688         BlockDriverState *bs;
6689         char *full_backing;
6690         int back_flags;
6691         QDict *backing_options = NULL;
6692 
6693         full_backing =
6694             bdrv_get_full_backing_filename_from_filename(filename, backing_file,
6695                                                          &local_err);
6696         if (local_err) {
6697             goto out;
6698         }
6699         assert(full_backing);
6700 
6701         /*
6702          * No need to do I/O here, which allows us to open encrypted
6703          * backing images without needing the secret
6704          */
6705         back_flags = flags;
6706         back_flags &= ~(BDRV_O_RDWR | BDRV_O_SNAPSHOT | BDRV_O_NO_BACKING);
6707         back_flags |= BDRV_O_NO_IO;
6708 
6709         backing_options = qdict_new();
6710         if (backing_fmt) {
6711             qdict_put_str(backing_options, "driver", backing_fmt);
6712         }
6713         qdict_put_bool(backing_options, BDRV_OPT_FORCE_SHARE, true);
6714 
6715         bs = bdrv_open(full_backing, NULL, backing_options, back_flags,
6716                        &local_err);
6717         g_free(full_backing);
6718         if (!bs) {
6719             error_append_hint(&local_err, "Could not open backing image.\n");
6720             goto out;
6721         } else {
6722             if (!backing_fmt) {
6723                 error_setg(&local_err,
6724                            "Backing file specified without backing format");
6725                 error_append_hint(&local_err, "Detected format of %s.",
6726                                   bs->drv->format_name);
6727                 goto out;
6728             }
6729             if (size == -1) {
6730                 /* Opened BS, have no size */
6731                 size = bdrv_getlength(bs);
6732                 if (size < 0) {
6733                     error_setg_errno(errp, -size, "Could not get size of '%s'",
6734                                      backing_file);
6735                     bdrv_unref(bs);
6736                     goto out;
6737                 }
6738                 qemu_opt_set_number(opts, BLOCK_OPT_SIZE, size, &error_abort);
6739             }
6740             bdrv_unref(bs);
6741         }
6742         /* (backing_file && !(flags & BDRV_O_NO_BACKING)) */
6743     } else if (backing_file && !backing_fmt) {
6744         error_setg(&local_err,
6745                    "Backing file specified without backing format");
6746         goto out;
6747     }
6748 
6749     if (size == -1) {
6750         error_setg(errp, "Image creation needs a size parameter");
6751         goto out;
6752     }
6753 
6754     if (!quiet) {
6755         printf("Formatting '%s', fmt=%s ", filename, fmt);
6756         qemu_opts_print(opts, " ");
6757         puts("");
6758         fflush(stdout);
6759     }
6760 
6761     ret = bdrv_create(drv, filename, opts, &local_err);
6762 
6763     if (ret == -EFBIG) {
6764         /* This is generally a better message than whatever the driver would
6765          * deliver (especially because of the cluster_size_hint), since that
6766          * is most probably not much different from "image too large". */
6767         const char *cluster_size_hint = "";
6768         if (qemu_opt_get_size(opts, BLOCK_OPT_CLUSTER_SIZE, 0)) {
6769             cluster_size_hint = " (try using a larger cluster size)";
6770         }
6771         error_setg(errp, "The image size is too large for file format '%s'"
6772                    "%s", fmt, cluster_size_hint);
6773         error_free(local_err);
6774         local_err = NULL;
6775     }
6776 
6777 out:
6778     qemu_opts_del(opts);
6779     qemu_opts_free(create_opts);
6780     error_propagate(errp, local_err);
6781 }
6782 
6783 AioContext *bdrv_get_aio_context(BlockDriverState *bs)
6784 {
6785     return bs ? bs->aio_context : qemu_get_aio_context();
6786 }
6787 
6788 AioContext *coroutine_fn bdrv_co_enter(BlockDriverState *bs)
6789 {
6790     Coroutine *self = qemu_coroutine_self();
6791     AioContext *old_ctx = qemu_coroutine_get_aio_context(self);
6792     AioContext *new_ctx;
6793 
6794     /*
6795      * Increase bs->in_flight to ensure that this operation is completed before
6796      * moving the node to a different AioContext. Read new_ctx only afterwards.
6797      */
6798     bdrv_inc_in_flight(bs);
6799 
6800     new_ctx = bdrv_get_aio_context(bs);
6801     aio_co_reschedule_self(new_ctx);
6802     return old_ctx;
6803 }
6804 
6805 void coroutine_fn bdrv_co_leave(BlockDriverState *bs, AioContext *old_ctx)
6806 {
6807     aio_co_reschedule_self(old_ctx);
6808     bdrv_dec_in_flight(bs);
6809 }
6810 
6811 void coroutine_fn bdrv_co_lock(BlockDriverState *bs)
6812 {
6813     AioContext *ctx = bdrv_get_aio_context(bs);
6814 
6815     /* In the main thread, bs->aio_context won't change concurrently */
6816     assert(qemu_get_current_aio_context() == qemu_get_aio_context());
6817 
6818     /*
6819      * We're in coroutine context, so we already hold the lock of the main
6820      * loop AioContext. Don't lock it twice to avoid deadlocks.
6821      */
6822     assert(qemu_in_coroutine());
6823     if (ctx != qemu_get_aio_context()) {
6824         aio_context_acquire(ctx);
6825     }
6826 }
6827 
6828 void coroutine_fn bdrv_co_unlock(BlockDriverState *bs)
6829 {
6830     AioContext *ctx = bdrv_get_aio_context(bs);
6831 
6832     assert(qemu_in_coroutine());
6833     if (ctx != qemu_get_aio_context()) {
6834         aio_context_release(ctx);
6835     }
6836 }
6837 
6838 void bdrv_coroutine_enter(BlockDriverState *bs, Coroutine *co)
6839 {
6840     aio_co_enter(bdrv_get_aio_context(bs), co);
6841 }
6842 
6843 static void bdrv_do_remove_aio_context_notifier(BdrvAioNotifier *ban)
6844 {
6845     QLIST_REMOVE(ban, list);
6846     g_free(ban);
6847 }
6848 
6849 static void bdrv_detach_aio_context(BlockDriverState *bs)
6850 {
6851     BdrvAioNotifier *baf, *baf_tmp;
6852 
6853     assert(!bs->walking_aio_notifiers);
6854     bs->walking_aio_notifiers = true;
6855     QLIST_FOREACH_SAFE(baf, &bs->aio_notifiers, list, baf_tmp) {
6856         if (baf->deleted) {
6857             bdrv_do_remove_aio_context_notifier(baf);
6858         } else {
6859             baf->detach_aio_context(baf->opaque);
6860         }
6861     }
6862     /* Never mind iterating again to check for ->deleted.  bdrv_close() will
6863      * remove remaining aio notifiers if we aren't called again.
6864      */
6865     bs->walking_aio_notifiers = false;
6866 
6867     if (bs->drv && bs->drv->bdrv_detach_aio_context) {
6868         bs->drv->bdrv_detach_aio_context(bs);
6869     }
6870 
6871     if (bs->quiesce_counter) {
6872         aio_enable_external(bs->aio_context);
6873     }
6874     bs->aio_context = NULL;
6875 }
6876 
6877 static void bdrv_attach_aio_context(BlockDriverState *bs,
6878                                     AioContext *new_context)
6879 {
6880     BdrvAioNotifier *ban, *ban_tmp;
6881 
6882     if (bs->quiesce_counter) {
6883         aio_disable_external(new_context);
6884     }
6885 
6886     bs->aio_context = new_context;
6887 
6888     if (bs->drv && bs->drv->bdrv_attach_aio_context) {
6889         bs->drv->bdrv_attach_aio_context(bs, new_context);
6890     }
6891 
6892     assert(!bs->walking_aio_notifiers);
6893     bs->walking_aio_notifiers = true;
6894     QLIST_FOREACH_SAFE(ban, &bs->aio_notifiers, list, ban_tmp) {
6895         if (ban->deleted) {
6896             bdrv_do_remove_aio_context_notifier(ban);
6897         } else {
6898             ban->attached_aio_context(new_context, ban->opaque);
6899         }
6900     }
6901     bs->walking_aio_notifiers = false;
6902 }
6903 
6904 /*
6905  * Changes the AioContext used for fd handlers, timers, and BHs by this
6906  * BlockDriverState and all its children and parents.
6907  *
6908  * Must be called from the main AioContext.
6909  *
6910  * The caller must own the AioContext lock for the old AioContext of bs, but it
6911  * must not own the AioContext lock for new_context (unless new_context is the
6912  * same as the current context of bs).
6913  *
6914  * @ignore will accumulate all visited BdrvChild object. The caller is
6915  * responsible for freeing the list afterwards.
6916  */
6917 void bdrv_set_aio_context_ignore(BlockDriverState *bs,
6918                                  AioContext *new_context, GSList **ignore)
6919 {
6920     AioContext *old_context = bdrv_get_aio_context(bs);
6921     GSList *children_to_process = NULL;
6922     GSList *parents_to_process = NULL;
6923     GSList *entry;
6924     BdrvChild *child, *parent;
6925 
6926     g_assert(qemu_get_current_aio_context() == qemu_get_aio_context());
6927 
6928     if (old_context == new_context) {
6929         return;
6930     }
6931 
6932     bdrv_drained_begin(bs);
6933 
6934     QLIST_FOREACH(child, &bs->children, next) {
6935         if (g_slist_find(*ignore, child)) {
6936             continue;
6937         }
6938         *ignore = g_slist_prepend(*ignore, child);
6939         children_to_process = g_slist_prepend(children_to_process, child);
6940     }
6941 
6942     QLIST_FOREACH(parent, &bs->parents, next_parent) {
6943         if (g_slist_find(*ignore, parent)) {
6944             continue;
6945         }
6946         *ignore = g_slist_prepend(*ignore, parent);
6947         parents_to_process = g_slist_prepend(parents_to_process, parent);
6948     }
6949 
6950     for (entry = children_to_process;
6951          entry != NULL;
6952          entry = g_slist_next(entry)) {
6953         child = entry->data;
6954         bdrv_set_aio_context_ignore(child->bs, new_context, ignore);
6955     }
6956     g_slist_free(children_to_process);
6957 
6958     for (entry = parents_to_process;
6959          entry != NULL;
6960          entry = g_slist_next(entry)) {
6961         parent = entry->data;
6962         assert(parent->klass->set_aio_ctx);
6963         parent->klass->set_aio_ctx(parent, new_context, ignore);
6964     }
6965     g_slist_free(parents_to_process);
6966 
6967     bdrv_detach_aio_context(bs);
6968 
6969     /* Acquire the new context, if necessary */
6970     if (qemu_get_aio_context() != new_context) {
6971         aio_context_acquire(new_context);
6972     }
6973 
6974     bdrv_attach_aio_context(bs, new_context);
6975 
6976     /*
6977      * If this function was recursively called from
6978      * bdrv_set_aio_context_ignore(), there may be nodes in the
6979      * subtree that have not yet been moved to the new AioContext.
6980      * Release the old one so bdrv_drained_end() can poll them.
6981      */
6982     if (qemu_get_aio_context() != old_context) {
6983         aio_context_release(old_context);
6984     }
6985 
6986     bdrv_drained_end(bs);
6987 
6988     if (qemu_get_aio_context() != old_context) {
6989         aio_context_acquire(old_context);
6990     }
6991     if (qemu_get_aio_context() != new_context) {
6992         aio_context_release(new_context);
6993     }
6994 }
6995 
6996 static bool bdrv_parent_can_set_aio_context(BdrvChild *c, AioContext *ctx,
6997                                             GSList **ignore, Error **errp)
6998 {
6999     if (g_slist_find(*ignore, c)) {
7000         return true;
7001     }
7002     *ignore = g_slist_prepend(*ignore, c);
7003 
7004     /*
7005      * A BdrvChildClass that doesn't handle AioContext changes cannot
7006      * tolerate any AioContext changes
7007      */
7008     if (!c->klass->can_set_aio_ctx) {
7009         char *user = bdrv_child_user_desc(c);
7010         error_setg(errp, "Changing iothreads is not supported by %s", user);
7011         g_free(user);
7012         return false;
7013     }
7014     if (!c->klass->can_set_aio_ctx(c, ctx, ignore, errp)) {
7015         assert(!errp || *errp);
7016         return false;
7017     }
7018     return true;
7019 }
7020 
7021 bool bdrv_child_can_set_aio_context(BdrvChild *c, AioContext *ctx,
7022                                     GSList **ignore, Error **errp)
7023 {
7024     if (g_slist_find(*ignore, c)) {
7025         return true;
7026     }
7027     *ignore = g_slist_prepend(*ignore, c);
7028     return bdrv_can_set_aio_context(c->bs, ctx, ignore, errp);
7029 }
7030 
7031 /* @ignore will accumulate all visited BdrvChild object. The caller is
7032  * responsible for freeing the list afterwards. */
7033 bool bdrv_can_set_aio_context(BlockDriverState *bs, AioContext *ctx,
7034                               GSList **ignore, Error **errp)
7035 {
7036     BdrvChild *c;
7037 
7038     if (bdrv_get_aio_context(bs) == ctx) {
7039         return true;
7040     }
7041 
7042     QLIST_FOREACH(c, &bs->parents, next_parent) {
7043         if (!bdrv_parent_can_set_aio_context(c, ctx, ignore, errp)) {
7044             return false;
7045         }
7046     }
7047     QLIST_FOREACH(c, &bs->children, next) {
7048         if (!bdrv_child_can_set_aio_context(c, ctx, ignore, errp)) {
7049             return false;
7050         }
7051     }
7052 
7053     return true;
7054 }
7055 
7056 int bdrv_child_try_set_aio_context(BlockDriverState *bs, AioContext *ctx,
7057                                    BdrvChild *ignore_child, Error **errp)
7058 {
7059     GSList *ignore;
7060     bool ret;
7061 
7062     ignore = ignore_child ? g_slist_prepend(NULL, ignore_child) : NULL;
7063     ret = bdrv_can_set_aio_context(bs, ctx, &ignore, errp);
7064     g_slist_free(ignore);
7065 
7066     if (!ret) {
7067         return -EPERM;
7068     }
7069 
7070     ignore = ignore_child ? g_slist_prepend(NULL, ignore_child) : NULL;
7071     bdrv_set_aio_context_ignore(bs, ctx, &ignore);
7072     g_slist_free(ignore);
7073 
7074     return 0;
7075 }
7076 
7077 int bdrv_try_set_aio_context(BlockDriverState *bs, AioContext *ctx,
7078                              Error **errp)
7079 {
7080     return bdrv_child_try_set_aio_context(bs, ctx, NULL, errp);
7081 }
7082 
7083 void bdrv_add_aio_context_notifier(BlockDriverState *bs,
7084         void (*attached_aio_context)(AioContext *new_context, void *opaque),
7085         void (*detach_aio_context)(void *opaque), void *opaque)
7086 {
7087     BdrvAioNotifier *ban = g_new(BdrvAioNotifier, 1);
7088     *ban = (BdrvAioNotifier){
7089         .attached_aio_context = attached_aio_context,
7090         .detach_aio_context   = detach_aio_context,
7091         .opaque               = opaque
7092     };
7093 
7094     QLIST_INSERT_HEAD(&bs->aio_notifiers, ban, list);
7095 }
7096 
7097 void bdrv_remove_aio_context_notifier(BlockDriverState *bs,
7098                                       void (*attached_aio_context)(AioContext *,
7099                                                                    void *),
7100                                       void (*detach_aio_context)(void *),
7101                                       void *opaque)
7102 {
7103     BdrvAioNotifier *ban, *ban_next;
7104 
7105     QLIST_FOREACH_SAFE(ban, &bs->aio_notifiers, list, ban_next) {
7106         if (ban->attached_aio_context == attached_aio_context &&
7107             ban->detach_aio_context   == detach_aio_context   &&
7108             ban->opaque               == opaque               &&
7109             ban->deleted              == false)
7110         {
7111             if (bs->walking_aio_notifiers) {
7112                 ban->deleted = true;
7113             } else {
7114                 bdrv_do_remove_aio_context_notifier(ban);
7115             }
7116             return;
7117         }
7118     }
7119 
7120     abort();
7121 }
7122 
7123 int bdrv_amend_options(BlockDriverState *bs, QemuOpts *opts,
7124                        BlockDriverAmendStatusCB *status_cb, void *cb_opaque,
7125                        bool force,
7126                        Error **errp)
7127 {
7128     if (!bs->drv) {
7129         error_setg(errp, "Node is ejected");
7130         return -ENOMEDIUM;
7131     }
7132     if (!bs->drv->bdrv_amend_options) {
7133         error_setg(errp, "Block driver '%s' does not support option amendment",
7134                    bs->drv->format_name);
7135         return -ENOTSUP;
7136     }
7137     return bs->drv->bdrv_amend_options(bs, opts, status_cb,
7138                                        cb_opaque, force, errp);
7139 }
7140 
7141 /*
7142  * This function checks whether the given @to_replace is allowed to be
7143  * replaced by a node that always shows the same data as @bs.  This is
7144  * used for example to verify whether the mirror job can replace
7145  * @to_replace by the target mirrored from @bs.
7146  * To be replaceable, @bs and @to_replace may either be guaranteed to
7147  * always show the same data (because they are only connected through
7148  * filters), or some driver may allow replacing one of its children
7149  * because it can guarantee that this child's data is not visible at
7150  * all (for example, for dissenting quorum children that have no other
7151  * parents).
7152  */
7153 bool bdrv_recurse_can_replace(BlockDriverState *bs,
7154                               BlockDriverState *to_replace)
7155 {
7156     BlockDriverState *filtered;
7157 
7158     if (!bs || !bs->drv) {
7159         return false;
7160     }
7161 
7162     if (bs == to_replace) {
7163         return true;
7164     }
7165 
7166     /* See what the driver can do */
7167     if (bs->drv->bdrv_recurse_can_replace) {
7168         return bs->drv->bdrv_recurse_can_replace(bs, to_replace);
7169     }
7170 
7171     /* For filters without an own implementation, we can recurse on our own */
7172     filtered = bdrv_filter_bs(bs);
7173     if (filtered) {
7174         return bdrv_recurse_can_replace(filtered, to_replace);
7175     }
7176 
7177     /* Safe default */
7178     return false;
7179 }
7180 
7181 /*
7182  * Check whether the given @node_name can be replaced by a node that
7183  * has the same data as @parent_bs.  If so, return @node_name's BDS;
7184  * NULL otherwise.
7185  *
7186  * @node_name must be a (recursive) *child of @parent_bs (or this
7187  * function will return NULL).
7188  *
7189  * The result (whether the node can be replaced or not) is only valid
7190  * for as long as no graph or permission changes occur.
7191  */
7192 BlockDriverState *check_to_replace_node(BlockDriverState *parent_bs,
7193                                         const char *node_name, Error **errp)
7194 {
7195     BlockDriverState *to_replace_bs = bdrv_find_node(node_name);
7196     AioContext *aio_context;
7197 
7198     if (!to_replace_bs) {
7199         error_setg(errp, "Failed to find node with node-name='%s'", node_name);
7200         return NULL;
7201     }
7202 
7203     aio_context = bdrv_get_aio_context(to_replace_bs);
7204     aio_context_acquire(aio_context);
7205 
7206     if (bdrv_op_is_blocked(to_replace_bs, BLOCK_OP_TYPE_REPLACE, errp)) {
7207         to_replace_bs = NULL;
7208         goto out;
7209     }
7210 
7211     /* We don't want arbitrary node of the BDS chain to be replaced only the top
7212      * most non filter in order to prevent data corruption.
7213      * Another benefit is that this tests exclude backing files which are
7214      * blocked by the backing blockers.
7215      */
7216     if (!bdrv_recurse_can_replace(parent_bs, to_replace_bs)) {
7217         error_setg(errp, "Cannot replace '%s' by a node mirrored from '%s', "
7218                    "because it cannot be guaranteed that doing so would not "
7219                    "lead to an abrupt change of visible data",
7220                    node_name, parent_bs->node_name);
7221         to_replace_bs = NULL;
7222         goto out;
7223     }
7224 
7225 out:
7226     aio_context_release(aio_context);
7227     return to_replace_bs;
7228 }
7229 
7230 /**
7231  * Iterates through the list of runtime option keys that are said to
7232  * be "strong" for a BDS.  An option is called "strong" if it changes
7233  * a BDS's data.  For example, the null block driver's "size" and
7234  * "read-zeroes" options are strong, but its "latency-ns" option is
7235  * not.
7236  *
7237  * If a key returned by this function ends with a dot, all options
7238  * starting with that prefix are strong.
7239  */
7240 static const char *const *strong_options(BlockDriverState *bs,
7241                                          const char *const *curopt)
7242 {
7243     static const char *const global_options[] = {
7244         "driver", "filename", NULL
7245     };
7246 
7247     if (!curopt) {
7248         return &global_options[0];
7249     }
7250 
7251     curopt++;
7252     if (curopt == &global_options[ARRAY_SIZE(global_options) - 1] && bs->drv) {
7253         curopt = bs->drv->strong_runtime_opts;
7254     }
7255 
7256     return (curopt && *curopt) ? curopt : NULL;
7257 }
7258 
7259 /**
7260  * Copies all strong runtime options from bs->options to the given
7261  * QDict.  The set of strong option keys is determined by invoking
7262  * strong_options().
7263  *
7264  * Returns true iff any strong option was present in bs->options (and
7265  * thus copied to the target QDict) with the exception of "filename"
7266  * and "driver".  The caller is expected to use this value to decide
7267  * whether the existence of strong options prevents the generation of
7268  * a plain filename.
7269  */
7270 static bool append_strong_runtime_options(QDict *d, BlockDriverState *bs)
7271 {
7272     bool found_any = false;
7273     const char *const *option_name = NULL;
7274 
7275     if (!bs->drv) {
7276         return false;
7277     }
7278 
7279     while ((option_name = strong_options(bs, option_name))) {
7280         bool option_given = false;
7281 
7282         assert(strlen(*option_name) > 0);
7283         if ((*option_name)[strlen(*option_name) - 1] != '.') {
7284             QObject *entry = qdict_get(bs->options, *option_name);
7285             if (!entry) {
7286                 continue;
7287             }
7288 
7289             qdict_put_obj(d, *option_name, qobject_ref(entry));
7290             option_given = true;
7291         } else {
7292             const QDictEntry *entry;
7293             for (entry = qdict_first(bs->options); entry;
7294                  entry = qdict_next(bs->options, entry))
7295             {
7296                 if (strstart(qdict_entry_key(entry), *option_name, NULL)) {
7297                     qdict_put_obj(d, qdict_entry_key(entry),
7298                                   qobject_ref(qdict_entry_value(entry)));
7299                     option_given = true;
7300                 }
7301             }
7302         }
7303 
7304         /* While "driver" and "filename" need to be included in a JSON filename,
7305          * their existence does not prohibit generation of a plain filename. */
7306         if (!found_any && option_given &&
7307             strcmp(*option_name, "driver") && strcmp(*option_name, "filename"))
7308         {
7309             found_any = true;
7310         }
7311     }
7312 
7313     if (!qdict_haskey(d, "driver")) {
7314         /* Drivers created with bdrv_new_open_driver() may not have a
7315          * @driver option.  Add it here. */
7316         qdict_put_str(d, "driver", bs->drv->format_name);
7317     }
7318 
7319     return found_any;
7320 }
7321 
7322 /* Note: This function may return false positives; it may return true
7323  * even if opening the backing file specified by bs's image header
7324  * would result in exactly bs->backing. */
7325 bool bdrv_backing_overridden(BlockDriverState *bs)
7326 {
7327     if (bs->backing) {
7328         return strcmp(bs->auto_backing_file,
7329                       bs->backing->bs->filename);
7330     } else {
7331         /* No backing BDS, so if the image header reports any backing
7332          * file, it must have been suppressed */
7333         return bs->auto_backing_file[0] != '\0';
7334     }
7335 }
7336 
7337 /* Updates the following BDS fields:
7338  *  - exact_filename: A filename which may be used for opening a block device
7339  *                    which (mostly) equals the given BDS (even without any
7340  *                    other options; so reading and writing must return the same
7341  *                    results, but caching etc. may be different)
7342  *  - full_open_options: Options which, when given when opening a block device
7343  *                       (without a filename), result in a BDS (mostly)
7344  *                       equalling the given one
7345  *  - filename: If exact_filename is set, it is copied here. Otherwise,
7346  *              full_open_options is converted to a JSON object, prefixed with
7347  *              "json:" (for use through the JSON pseudo protocol) and put here.
7348  */
7349 void bdrv_refresh_filename(BlockDriverState *bs)
7350 {
7351     BlockDriver *drv = bs->drv;
7352     BdrvChild *child;
7353     BlockDriverState *primary_child_bs;
7354     QDict *opts;
7355     bool backing_overridden;
7356     bool generate_json_filename; /* Whether our default implementation should
7357                                     fill exact_filename (false) or not (true) */
7358 
7359     if (!drv) {
7360         return;
7361     }
7362 
7363     /* This BDS's file name may depend on any of its children's file names, so
7364      * refresh those first */
7365     QLIST_FOREACH(child, &bs->children, next) {
7366         bdrv_refresh_filename(child->bs);
7367     }
7368 
7369     if (bs->implicit) {
7370         /* For implicit nodes, just copy everything from the single child */
7371         child = QLIST_FIRST(&bs->children);
7372         assert(QLIST_NEXT(child, next) == NULL);
7373 
7374         pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
7375                 child->bs->exact_filename);
7376         pstrcpy(bs->filename, sizeof(bs->filename), child->bs->filename);
7377 
7378         qobject_unref(bs->full_open_options);
7379         bs->full_open_options = qobject_ref(child->bs->full_open_options);
7380 
7381         return;
7382     }
7383 
7384     backing_overridden = bdrv_backing_overridden(bs);
7385 
7386     if (bs->open_flags & BDRV_O_NO_IO) {
7387         /* Without I/O, the backing file does not change anything.
7388          * Therefore, in such a case (primarily qemu-img), we can
7389          * pretend the backing file has not been overridden even if
7390          * it technically has been. */
7391         backing_overridden = false;
7392     }
7393 
7394     /* Gather the options QDict */
7395     opts = qdict_new();
7396     generate_json_filename = append_strong_runtime_options(opts, bs);
7397     generate_json_filename |= backing_overridden;
7398 
7399     if (drv->bdrv_gather_child_options) {
7400         /* Some block drivers may not want to present all of their children's
7401          * options, or name them differently from BdrvChild.name */
7402         drv->bdrv_gather_child_options(bs, opts, backing_overridden);
7403     } else {
7404         QLIST_FOREACH(child, &bs->children, next) {
7405             if (child == bs->backing && !backing_overridden) {
7406                 /* We can skip the backing BDS if it has not been overridden */
7407                 continue;
7408             }
7409 
7410             qdict_put(opts, child->name,
7411                       qobject_ref(child->bs->full_open_options));
7412         }
7413 
7414         if (backing_overridden && !bs->backing) {
7415             /* Force no backing file */
7416             qdict_put_null(opts, "backing");
7417         }
7418     }
7419 
7420     qobject_unref(bs->full_open_options);
7421     bs->full_open_options = opts;
7422 
7423     primary_child_bs = bdrv_primary_bs(bs);
7424 
7425     if (drv->bdrv_refresh_filename) {
7426         /* Obsolete information is of no use here, so drop the old file name
7427          * information before refreshing it */
7428         bs->exact_filename[0] = '\0';
7429 
7430         drv->bdrv_refresh_filename(bs);
7431     } else if (primary_child_bs) {
7432         /*
7433          * Try to reconstruct valid information from the underlying
7434          * file -- this only works for format nodes (filter nodes
7435          * cannot be probed and as such must be selected by the user
7436          * either through an options dict, or through a special
7437          * filename which the filter driver must construct in its
7438          * .bdrv_refresh_filename() implementation).
7439          */
7440 
7441         bs->exact_filename[0] = '\0';
7442 
7443         /*
7444          * We can use the underlying file's filename if:
7445          * - it has a filename,
7446          * - the current BDS is not a filter,
7447          * - the file is a protocol BDS, and
7448          * - opening that file (as this BDS's format) will automatically create
7449          *   the BDS tree we have right now, that is:
7450          *   - the user did not significantly change this BDS's behavior with
7451          *     some explicit (strong) options
7452          *   - no non-file child of this BDS has been overridden by the user
7453          *   Both of these conditions are represented by generate_json_filename.
7454          */
7455         if (primary_child_bs->exact_filename[0] &&
7456             primary_child_bs->drv->bdrv_file_open &&
7457             !drv->is_filter && !generate_json_filename)
7458         {
7459             strcpy(bs->exact_filename, primary_child_bs->exact_filename);
7460         }
7461     }
7462 
7463     if (bs->exact_filename[0]) {
7464         pstrcpy(bs->filename, sizeof(bs->filename), bs->exact_filename);
7465     } else {
7466         GString *json = qobject_to_json(QOBJECT(bs->full_open_options));
7467         if (snprintf(bs->filename, sizeof(bs->filename), "json:%s",
7468                      json->str) >= sizeof(bs->filename)) {
7469             /* Give user a hint if we truncated things. */
7470             strcpy(bs->filename + sizeof(bs->filename) - 4, "...");
7471         }
7472         g_string_free(json, true);
7473     }
7474 }
7475 
7476 char *bdrv_dirname(BlockDriverState *bs, Error **errp)
7477 {
7478     BlockDriver *drv = bs->drv;
7479     BlockDriverState *child_bs;
7480 
7481     if (!drv) {
7482         error_setg(errp, "Node '%s' is ejected", bs->node_name);
7483         return NULL;
7484     }
7485 
7486     if (drv->bdrv_dirname) {
7487         return drv->bdrv_dirname(bs, errp);
7488     }
7489 
7490     child_bs = bdrv_primary_bs(bs);
7491     if (child_bs) {
7492         return bdrv_dirname(child_bs, errp);
7493     }
7494 
7495     bdrv_refresh_filename(bs);
7496     if (bs->exact_filename[0] != '\0') {
7497         return path_combine(bs->exact_filename, "");
7498     }
7499 
7500     error_setg(errp, "Cannot generate a base directory for %s nodes",
7501                drv->format_name);
7502     return NULL;
7503 }
7504 
7505 /*
7506  * Hot add/remove a BDS's child. So the user can take a child offline when
7507  * it is broken and take a new child online
7508  */
7509 void bdrv_add_child(BlockDriverState *parent_bs, BlockDriverState *child_bs,
7510                     Error **errp)
7511 {
7512 
7513     if (!parent_bs->drv || !parent_bs->drv->bdrv_add_child) {
7514         error_setg(errp, "The node %s does not support adding a child",
7515                    bdrv_get_device_or_node_name(parent_bs));
7516         return;
7517     }
7518 
7519     if (!QLIST_EMPTY(&child_bs->parents)) {
7520         error_setg(errp, "The node %s already has a parent",
7521                    child_bs->node_name);
7522         return;
7523     }
7524 
7525     parent_bs->drv->bdrv_add_child(parent_bs, child_bs, errp);
7526 }
7527 
7528 void bdrv_del_child(BlockDriverState *parent_bs, BdrvChild *child, Error **errp)
7529 {
7530     BdrvChild *tmp;
7531 
7532     if (!parent_bs->drv || !parent_bs->drv->bdrv_del_child) {
7533         error_setg(errp, "The node %s does not support removing a child",
7534                    bdrv_get_device_or_node_name(parent_bs));
7535         return;
7536     }
7537 
7538     QLIST_FOREACH(tmp, &parent_bs->children, next) {
7539         if (tmp == child) {
7540             break;
7541         }
7542     }
7543 
7544     if (!tmp) {
7545         error_setg(errp, "The node %s does not have a child named %s",
7546                    bdrv_get_device_or_node_name(parent_bs),
7547                    bdrv_get_device_or_node_name(child->bs));
7548         return;
7549     }
7550 
7551     parent_bs->drv->bdrv_del_child(parent_bs, child, errp);
7552 }
7553 
7554 int bdrv_make_empty(BdrvChild *c, Error **errp)
7555 {
7556     BlockDriver *drv = c->bs->drv;
7557     int ret;
7558 
7559     assert(c->perm & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED));
7560 
7561     if (!drv->bdrv_make_empty) {
7562         error_setg(errp, "%s does not support emptying nodes",
7563                    drv->format_name);
7564         return -ENOTSUP;
7565     }
7566 
7567     ret = drv->bdrv_make_empty(c->bs);
7568     if (ret < 0) {
7569         error_setg_errno(errp, -ret, "Failed to empty %s",
7570                          c->bs->filename);
7571         return ret;
7572     }
7573 
7574     return 0;
7575 }
7576 
7577 /*
7578  * Return the child that @bs acts as an overlay for, and from which data may be
7579  * copied in COW or COR operations.  Usually this is the backing file.
7580  */
7581 BdrvChild *bdrv_cow_child(BlockDriverState *bs)
7582 {
7583     if (!bs || !bs->drv) {
7584         return NULL;
7585     }
7586 
7587     if (bs->drv->is_filter) {
7588         return NULL;
7589     }
7590 
7591     if (!bs->backing) {
7592         return NULL;
7593     }
7594 
7595     assert(bs->backing->role & BDRV_CHILD_COW);
7596     return bs->backing;
7597 }
7598 
7599 /*
7600  * If @bs acts as a filter for exactly one of its children, return
7601  * that child.
7602  */
7603 BdrvChild *bdrv_filter_child(BlockDriverState *bs)
7604 {
7605     BdrvChild *c;
7606 
7607     if (!bs || !bs->drv) {
7608         return NULL;
7609     }
7610 
7611     if (!bs->drv->is_filter) {
7612         return NULL;
7613     }
7614 
7615     /* Only one of @backing or @file may be used */
7616     assert(!(bs->backing && bs->file));
7617 
7618     c = bs->backing ?: bs->file;
7619     if (!c) {
7620         return NULL;
7621     }
7622 
7623     assert(c->role & BDRV_CHILD_FILTERED);
7624     return c;
7625 }
7626 
7627 /*
7628  * Return either the result of bdrv_cow_child() or bdrv_filter_child(),
7629  * whichever is non-NULL.
7630  *
7631  * Return NULL if both are NULL.
7632  */
7633 BdrvChild *bdrv_filter_or_cow_child(BlockDriverState *bs)
7634 {
7635     BdrvChild *cow_child = bdrv_cow_child(bs);
7636     BdrvChild *filter_child = bdrv_filter_child(bs);
7637 
7638     /* Filter nodes cannot have COW backing files */
7639     assert(!(cow_child && filter_child));
7640 
7641     return cow_child ?: filter_child;
7642 }
7643 
7644 /*
7645  * Return the primary child of this node: For filters, that is the
7646  * filtered child.  For other nodes, that is usually the child storing
7647  * metadata.
7648  * (A generally more helpful description is that this is (usually) the
7649  * child that has the same filename as @bs.)
7650  *
7651  * Drivers do not necessarily have a primary child; for example quorum
7652  * does not.
7653  */
7654 BdrvChild *bdrv_primary_child(BlockDriverState *bs)
7655 {
7656     BdrvChild *c, *found = NULL;
7657 
7658     QLIST_FOREACH(c, &bs->children, next) {
7659         if (c->role & BDRV_CHILD_PRIMARY) {
7660             assert(!found);
7661             found = c;
7662         }
7663     }
7664 
7665     return found;
7666 }
7667 
7668 static BlockDriverState *bdrv_do_skip_filters(BlockDriverState *bs,
7669                                               bool stop_on_explicit_filter)
7670 {
7671     BdrvChild *c;
7672 
7673     if (!bs) {
7674         return NULL;
7675     }
7676 
7677     while (!(stop_on_explicit_filter && !bs->implicit)) {
7678         c = bdrv_filter_child(bs);
7679         if (!c) {
7680             /*
7681              * A filter that is embedded in a working block graph must
7682              * have a child.  Assert this here so this function does
7683              * not return a filter node that is not expected by the
7684              * caller.
7685              */
7686             assert(!bs->drv || !bs->drv->is_filter);
7687             break;
7688         }
7689         bs = c->bs;
7690     }
7691     /*
7692      * Note that this treats nodes with bs->drv == NULL as not being
7693      * filters (bs->drv == NULL should be replaced by something else
7694      * anyway).
7695      * The advantage of this behavior is that this function will thus
7696      * always return a non-NULL value (given a non-NULL @bs).
7697      */
7698 
7699     return bs;
7700 }
7701 
7702 /*
7703  * Return the first BDS that has not been added implicitly or that
7704  * does not have a filtered child down the chain starting from @bs
7705  * (including @bs itself).
7706  */
7707 BlockDriverState *bdrv_skip_implicit_filters(BlockDriverState *bs)
7708 {
7709     return bdrv_do_skip_filters(bs, true);
7710 }
7711 
7712 /*
7713  * Return the first BDS that does not have a filtered child down the
7714  * chain starting from @bs (including @bs itself).
7715  */
7716 BlockDriverState *bdrv_skip_filters(BlockDriverState *bs)
7717 {
7718     return bdrv_do_skip_filters(bs, false);
7719 }
7720 
7721 /*
7722  * For a backing chain, return the first non-filter backing image of
7723  * the first non-filter image.
7724  */
7725 BlockDriverState *bdrv_backing_chain_next(BlockDriverState *bs)
7726 {
7727     return bdrv_skip_filters(bdrv_cow_bs(bdrv_skip_filters(bs)));
7728 }
7729 
7730 /**
7731  * Check whether [offset, offset + bytes) overlaps with the cached
7732  * block-status data region.
7733  *
7734  * If so, and @pnum is not NULL, set *pnum to `bsc.data_end - offset`,
7735  * which is what bdrv_bsc_is_data()'s interface needs.
7736  * Otherwise, *pnum is not touched.
7737  */
7738 static bool bdrv_bsc_range_overlaps_locked(BlockDriverState *bs,
7739                                            int64_t offset, int64_t bytes,
7740                                            int64_t *pnum)
7741 {
7742     BdrvBlockStatusCache *bsc = qatomic_rcu_read(&bs->block_status_cache);
7743     bool overlaps;
7744 
7745     overlaps =
7746         qatomic_read(&bsc->valid) &&
7747         ranges_overlap(offset, bytes, bsc->data_start,
7748                        bsc->data_end - bsc->data_start);
7749 
7750     if (overlaps && pnum) {
7751         *pnum = bsc->data_end - offset;
7752     }
7753 
7754     return overlaps;
7755 }
7756 
7757 /**
7758  * See block_int.h for this function's documentation.
7759  */
7760 bool bdrv_bsc_is_data(BlockDriverState *bs, int64_t offset, int64_t *pnum)
7761 {
7762     RCU_READ_LOCK_GUARD();
7763 
7764     return bdrv_bsc_range_overlaps_locked(bs, offset, 1, pnum);
7765 }
7766 
7767 /**
7768  * See block_int.h for this function's documentation.
7769  */
7770 void bdrv_bsc_invalidate_range(BlockDriverState *bs,
7771                                int64_t offset, int64_t bytes)
7772 {
7773     RCU_READ_LOCK_GUARD();
7774 
7775     if (bdrv_bsc_range_overlaps_locked(bs, offset, bytes, NULL)) {
7776         qatomic_set(&bs->block_status_cache->valid, false);
7777     }
7778 }
7779 
7780 /**
7781  * See block_int.h for this function's documentation.
7782  */
7783 void bdrv_bsc_fill(BlockDriverState *bs, int64_t offset, int64_t bytes)
7784 {
7785     BdrvBlockStatusCache *new_bsc = g_new(BdrvBlockStatusCache, 1);
7786     BdrvBlockStatusCache *old_bsc;
7787 
7788     *new_bsc = (BdrvBlockStatusCache) {
7789         .valid = true,
7790         .data_start = offset,
7791         .data_end = offset + bytes,
7792     };
7793 
7794     QEMU_LOCK_GUARD(&bs->bsc_modify_lock);
7795 
7796     old_bsc = qatomic_rcu_read(&bs->block_status_cache);
7797     qatomic_rcu_set(&bs->block_status_cache, new_bsc);
7798     if (old_bsc) {
7799         g_free_rcu(old_bsc, rcu);
7800     }
7801 }
7802