xref: /openbmc/qemu/block.c (revision 9f8d2fdcce78587b7861ac6b9d2c8bdc2bbb9ad4)
1 /*
2  * QEMU System Emulator block driver
3  *
4  * Copyright (c) 2003 Fabrice Bellard
5  * Copyright (c) 2020 Virtuozzo International GmbH.
6  *
7  * Permission is hereby granted, free of charge, to any person obtaining a copy
8  * of this software and associated documentation files (the "Software"), to deal
9  * in the Software without restriction, including without limitation the rights
10  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
11  * copies of the Software, and to permit persons to whom the Software is
12  * furnished to do so, subject to the following conditions:
13  *
14  * The above copyright notice and this permission notice shall be included in
15  * all copies or substantial portions of the Software.
16  *
17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
23  * THE SOFTWARE.
24  */
25 
26 #include "qemu/osdep.h"
27 #include "block/trace.h"
28 #include "block/block_int.h"
29 #include "block/blockjob.h"
30 #include "block/dirty-bitmap.h"
31 #include "block/fuse.h"
32 #include "block/nbd.h"
33 #include "block/qdict.h"
34 #include "qemu/error-report.h"
35 #include "block/module_block.h"
36 #include "qemu/main-loop.h"
37 #include "qemu/module.h"
38 #include "qapi/error.h"
39 #include "qapi/qmp/qdict.h"
40 #include "qapi/qmp/qjson.h"
41 #include "qapi/qmp/qnull.h"
42 #include "qapi/qmp/qstring.h"
43 #include "qapi/qobject-output-visitor.h"
44 #include "qapi/qapi-visit-block-core.h"
45 #include "sysemu/block-backend.h"
46 #include "qemu/notify.h"
47 #include "qemu/option.h"
48 #include "qemu/coroutine.h"
49 #include "block/qapi.h"
50 #include "qemu/timer.h"
51 #include "qemu/cutils.h"
52 #include "qemu/id.h"
53 #include "qemu/range.h"
54 #include "qemu/rcu.h"
55 #include "block/coroutines.h"
56 
57 #ifdef CONFIG_BSD
58 #include <sys/ioctl.h>
59 #include <sys/queue.h>
60 #if defined(HAVE_SYS_DISK_H)
61 #include <sys/disk.h>
62 #endif
63 #endif
64 
65 #ifdef _WIN32
66 #include <windows.h>
67 #endif
68 
69 #define NOT_DONE 0x7fffffff /* used while emulated sync operation in progress */
70 
71 /* Protected by BQL */
72 static QTAILQ_HEAD(, BlockDriverState) graph_bdrv_states =
73     QTAILQ_HEAD_INITIALIZER(graph_bdrv_states);
74 
75 /* Protected by BQL */
76 static QTAILQ_HEAD(, BlockDriverState) all_bdrv_states =
77     QTAILQ_HEAD_INITIALIZER(all_bdrv_states);
78 
79 /* Protected by BQL */
80 static QLIST_HEAD(, BlockDriver) bdrv_drivers =
81     QLIST_HEAD_INITIALIZER(bdrv_drivers);
82 
83 static BlockDriverState *bdrv_open_inherit(const char *filename,
84                                            const char *reference,
85                                            QDict *options, int flags,
86                                            BlockDriverState *parent,
87                                            const BdrvChildClass *child_class,
88                                            BdrvChildRole child_role,
89                                            Error **errp);
90 
91 static bool bdrv_recurse_has_child(BlockDriverState *bs,
92                                    BlockDriverState *child);
93 
94 static void GRAPH_WRLOCK
95 bdrv_replace_child_noperm(BdrvChild *child, BlockDriverState *new_bs);
96 
97 static void GRAPH_WRLOCK
98 bdrv_remove_child(BdrvChild *child, Transaction *tran);
99 
100 static int bdrv_reopen_prepare(BDRVReopenState *reopen_state,
101                                BlockReopenQueue *queue,
102                                Transaction *change_child_tran, Error **errp);
103 static void bdrv_reopen_commit(BDRVReopenState *reopen_state);
104 static void bdrv_reopen_abort(BDRVReopenState *reopen_state);
105 
106 static bool bdrv_backing_overridden(BlockDriverState *bs);
107 
108 static bool bdrv_change_aio_context(BlockDriverState *bs, AioContext *ctx,
109                                     GHashTable *visited, Transaction *tran,
110                                     Error **errp);
111 
112 /* If non-zero, use only whitelisted block drivers */
113 static int use_bdrv_whitelist;
114 
115 #ifdef _WIN32
116 static int is_windows_drive_prefix(const char *filename)
117 {
118     return (((filename[0] >= 'a' && filename[0] <= 'z') ||
119              (filename[0] >= 'A' && filename[0] <= 'Z')) &&
120             filename[1] == ':');
121 }
122 
123 int is_windows_drive(const char *filename)
124 {
125     if (is_windows_drive_prefix(filename) &&
126         filename[2] == '\0')
127         return 1;
128     if (strstart(filename, "\\\\.\\", NULL) ||
129         strstart(filename, "//./", NULL))
130         return 1;
131     return 0;
132 }
133 #endif
134 
135 size_t bdrv_opt_mem_align(BlockDriverState *bs)
136 {
137     if (!bs || !bs->drv) {
138         /* page size or 4k (hdd sector size) should be on the safe side */
139         return MAX(4096, qemu_real_host_page_size());
140     }
141     IO_CODE();
142 
143     return bs->bl.opt_mem_alignment;
144 }
145 
146 size_t bdrv_min_mem_align(BlockDriverState *bs)
147 {
148     if (!bs || !bs->drv) {
149         /* page size or 4k (hdd sector size) should be on the safe side */
150         return MAX(4096, qemu_real_host_page_size());
151     }
152     IO_CODE();
153 
154     return bs->bl.min_mem_alignment;
155 }
156 
157 /* check if the path starts with "<protocol>:" */
158 int path_has_protocol(const char *path)
159 {
160     const char *p;
161 
162 #ifdef _WIN32
163     if (is_windows_drive(path) ||
164         is_windows_drive_prefix(path)) {
165         return 0;
166     }
167     p = path + strcspn(path, ":/\\");
168 #else
169     p = path + strcspn(path, ":/");
170 #endif
171 
172     return *p == ':';
173 }
174 
175 int path_is_absolute(const char *path)
176 {
177 #ifdef _WIN32
178     /* specific case for names like: "\\.\d:" */
179     if (is_windows_drive(path) || is_windows_drive_prefix(path)) {
180         return 1;
181     }
182     return (*path == '/' || *path == '\\');
183 #else
184     return (*path == '/');
185 #endif
186 }
187 
188 /* if filename is absolute, just return its duplicate. Otherwise, build a
189    path to it by considering it is relative to base_path. URL are
190    supported. */
191 char *path_combine(const char *base_path, const char *filename)
192 {
193     const char *protocol_stripped = NULL;
194     const char *p, *p1;
195     char *result;
196     int len;
197 
198     if (path_is_absolute(filename)) {
199         return g_strdup(filename);
200     }
201 
202     if (path_has_protocol(base_path)) {
203         protocol_stripped = strchr(base_path, ':');
204         if (protocol_stripped) {
205             protocol_stripped++;
206         }
207     }
208     p = protocol_stripped ?: base_path;
209 
210     p1 = strrchr(base_path, '/');
211 #ifdef _WIN32
212     {
213         const char *p2;
214         p2 = strrchr(base_path, '\\');
215         if (!p1 || p2 > p1) {
216             p1 = p2;
217         }
218     }
219 #endif
220     if (p1) {
221         p1++;
222     } else {
223         p1 = base_path;
224     }
225     if (p1 > p) {
226         p = p1;
227     }
228     len = p - base_path;
229 
230     result = g_malloc(len + strlen(filename) + 1);
231     memcpy(result, base_path, len);
232     strcpy(result + len, filename);
233 
234     return result;
235 }
236 
237 /*
238  * Helper function for bdrv_parse_filename() implementations to remove optional
239  * protocol prefixes (especially "file:") from a filename and for putting the
240  * stripped filename into the options QDict if there is such a prefix.
241  */
242 void bdrv_parse_filename_strip_prefix(const char *filename, const char *prefix,
243                                       QDict *options)
244 {
245     if (strstart(filename, prefix, &filename)) {
246         /* Stripping the explicit protocol prefix may result in a protocol
247          * prefix being (wrongly) detected (if the filename contains a colon) */
248         if (path_has_protocol(filename)) {
249             GString *fat_filename;
250 
251             /* This means there is some colon before the first slash; therefore,
252              * this cannot be an absolute path */
253             assert(!path_is_absolute(filename));
254 
255             /* And we can thus fix the protocol detection issue by prefixing it
256              * by "./" */
257             fat_filename = g_string_new("./");
258             g_string_append(fat_filename, filename);
259 
260             assert(!path_has_protocol(fat_filename->str));
261 
262             qdict_put(options, "filename",
263                       qstring_from_gstring(fat_filename));
264         } else {
265             /* If no protocol prefix was detected, we can use the shortened
266              * filename as-is */
267             qdict_put_str(options, "filename", filename);
268         }
269     }
270 }
271 
272 
273 /* Returns whether the image file is opened as read-only. Note that this can
274  * return false and writing to the image file is still not possible because the
275  * image is inactivated. */
276 bool bdrv_is_read_only(BlockDriverState *bs)
277 {
278     IO_CODE();
279     return !(bs->open_flags & BDRV_O_RDWR);
280 }
281 
282 static int GRAPH_RDLOCK
283 bdrv_can_set_read_only(BlockDriverState *bs, bool read_only,
284                        bool ignore_allow_rdw, Error **errp)
285 {
286     IO_CODE();
287 
288     /* Do not set read_only if copy_on_read is enabled */
289     if (bs->copy_on_read && read_only) {
290         error_setg(errp, "Can't set node '%s' to r/o with copy-on-read enabled",
291                    bdrv_get_device_or_node_name(bs));
292         return -EINVAL;
293     }
294 
295     /* Do not clear read_only if it is prohibited */
296     if (!read_only && !(bs->open_flags & BDRV_O_ALLOW_RDWR) &&
297         !ignore_allow_rdw)
298     {
299         error_setg(errp, "Node '%s' is read only",
300                    bdrv_get_device_or_node_name(bs));
301         return -EPERM;
302     }
303 
304     return 0;
305 }
306 
307 /*
308  * Called by a driver that can only provide a read-only image.
309  *
310  * Returns 0 if the node is already read-only or it could switch the node to
311  * read-only because BDRV_O_AUTO_RDONLY is set.
312  *
313  * Returns -EACCES if the node is read-write and BDRV_O_AUTO_RDONLY is not set
314  * or bdrv_can_set_read_only() forbids making the node read-only. If @errmsg
315  * is not NULL, it is used as the error message for the Error object.
316  */
317 int bdrv_apply_auto_read_only(BlockDriverState *bs, const char *errmsg,
318                               Error **errp)
319 {
320     int ret = 0;
321     IO_CODE();
322 
323     if (!(bs->open_flags & BDRV_O_RDWR)) {
324         return 0;
325     }
326     if (!(bs->open_flags & BDRV_O_AUTO_RDONLY)) {
327         goto fail;
328     }
329 
330     ret = bdrv_can_set_read_only(bs, true, false, NULL);
331     if (ret < 0) {
332         goto fail;
333     }
334 
335     bs->open_flags &= ~BDRV_O_RDWR;
336 
337     return 0;
338 
339 fail:
340     error_setg(errp, "%s", errmsg ?: "Image is read-only");
341     return -EACCES;
342 }
343 
344 /*
345  * If @backing is empty, this function returns NULL without setting
346  * @errp.  In all other cases, NULL will only be returned with @errp
347  * set.
348  *
349  * Therefore, a return value of NULL without @errp set means that
350  * there is no backing file; if @errp is set, there is one but its
351  * absolute filename cannot be generated.
352  */
353 char *bdrv_get_full_backing_filename_from_filename(const char *backed,
354                                                    const char *backing,
355                                                    Error **errp)
356 {
357     if (backing[0] == '\0') {
358         return NULL;
359     } else if (path_has_protocol(backing) || path_is_absolute(backing)) {
360         return g_strdup(backing);
361     } else if (backed[0] == '\0' || strstart(backed, "json:", NULL)) {
362         error_setg(errp, "Cannot use relative backing file names for '%s'",
363                    backed);
364         return NULL;
365     } else {
366         return path_combine(backed, backing);
367     }
368 }
369 
370 /*
371  * If @filename is empty or NULL, this function returns NULL without
372  * setting @errp.  In all other cases, NULL will only be returned with
373  * @errp set.
374  */
375 static char * GRAPH_RDLOCK
376 bdrv_make_absolute_filename(BlockDriverState *relative_to,
377                             const char *filename, Error **errp)
378 {
379     char *dir, *full_name;
380 
381     if (!filename || filename[0] == '\0') {
382         return NULL;
383     } else if (path_has_protocol(filename) || path_is_absolute(filename)) {
384         return g_strdup(filename);
385     }
386 
387     dir = bdrv_dirname(relative_to, errp);
388     if (!dir) {
389         return NULL;
390     }
391 
392     full_name = g_strconcat(dir, filename, NULL);
393     g_free(dir);
394     return full_name;
395 }
396 
397 char *bdrv_get_full_backing_filename(BlockDriverState *bs, Error **errp)
398 {
399     GLOBAL_STATE_CODE();
400     return bdrv_make_absolute_filename(bs, bs->backing_file, errp);
401 }
402 
403 void bdrv_register(BlockDriver *bdrv)
404 {
405     assert(bdrv->format_name);
406     GLOBAL_STATE_CODE();
407     QLIST_INSERT_HEAD(&bdrv_drivers, bdrv, list);
408 }
409 
410 BlockDriverState *bdrv_new(void)
411 {
412     BlockDriverState *bs;
413     int i;
414 
415     GLOBAL_STATE_CODE();
416 
417     bs = g_new0(BlockDriverState, 1);
418     QLIST_INIT(&bs->dirty_bitmaps);
419     for (i = 0; i < BLOCK_OP_TYPE_MAX; i++) {
420         QLIST_INIT(&bs->op_blockers[i]);
421     }
422     qemu_mutex_init(&bs->reqs_lock);
423     qemu_mutex_init(&bs->dirty_bitmap_mutex);
424     bs->refcnt = 1;
425     bs->aio_context = qemu_get_aio_context();
426 
427     qemu_co_queue_init(&bs->flush_queue);
428 
429     qemu_co_mutex_init(&bs->bsc_modify_lock);
430     bs->block_status_cache = g_new0(BdrvBlockStatusCache, 1);
431 
432     for (i = 0; i < bdrv_drain_all_count; i++) {
433         bdrv_drained_begin(bs);
434     }
435 
436     QTAILQ_INSERT_TAIL(&all_bdrv_states, bs, bs_list);
437 
438     return bs;
439 }
440 
441 static BlockDriver *bdrv_do_find_format(const char *format_name)
442 {
443     BlockDriver *drv1;
444     GLOBAL_STATE_CODE();
445 
446     QLIST_FOREACH(drv1, &bdrv_drivers, list) {
447         if (!strcmp(drv1->format_name, format_name)) {
448             return drv1;
449         }
450     }
451 
452     return NULL;
453 }
454 
455 BlockDriver *bdrv_find_format(const char *format_name)
456 {
457     BlockDriver *drv1;
458     int i;
459 
460     GLOBAL_STATE_CODE();
461 
462     drv1 = bdrv_do_find_format(format_name);
463     if (drv1) {
464         return drv1;
465     }
466 
467     /* The driver isn't registered, maybe we need to load a module */
468     for (i = 0; i < (int)ARRAY_SIZE(block_driver_modules); ++i) {
469         if (!strcmp(block_driver_modules[i].format_name, format_name)) {
470             Error *local_err = NULL;
471             int rv = block_module_load(block_driver_modules[i].library_name,
472                                        &local_err);
473             if (rv > 0) {
474                 return bdrv_do_find_format(format_name);
475             } else if (rv < 0) {
476                 error_report_err(local_err);
477             }
478             break;
479         }
480     }
481     return NULL;
482 }
483 
484 static int bdrv_format_is_whitelisted(const char *format_name, bool read_only)
485 {
486     static const char *whitelist_rw[] = {
487         CONFIG_BDRV_RW_WHITELIST
488         NULL
489     };
490     static const char *whitelist_ro[] = {
491         CONFIG_BDRV_RO_WHITELIST
492         NULL
493     };
494     const char **p;
495 
496     if (!whitelist_rw[0] && !whitelist_ro[0]) {
497         return 1;               /* no whitelist, anything goes */
498     }
499 
500     for (p = whitelist_rw; *p; p++) {
501         if (!strcmp(format_name, *p)) {
502             return 1;
503         }
504     }
505     if (read_only) {
506         for (p = whitelist_ro; *p; p++) {
507             if (!strcmp(format_name, *p)) {
508                 return 1;
509             }
510         }
511     }
512     return 0;
513 }
514 
515 int bdrv_is_whitelisted(BlockDriver *drv, bool read_only)
516 {
517     GLOBAL_STATE_CODE();
518     return bdrv_format_is_whitelisted(drv->format_name, read_only);
519 }
520 
521 bool bdrv_uses_whitelist(void)
522 {
523     return use_bdrv_whitelist;
524 }
525 
526 typedef struct CreateCo {
527     BlockDriver *drv;
528     char *filename;
529     QemuOpts *opts;
530     int ret;
531     Error *err;
532 } CreateCo;
533 
534 int coroutine_fn bdrv_co_create(BlockDriver *drv, const char *filename,
535                                 QemuOpts *opts, Error **errp)
536 {
537     int ret;
538     GLOBAL_STATE_CODE();
539     ERRP_GUARD();
540 
541     if (!drv->bdrv_co_create_opts) {
542         error_setg(errp, "Driver '%s' does not support image creation",
543                    drv->format_name);
544         return -ENOTSUP;
545     }
546 
547     ret = drv->bdrv_co_create_opts(drv, filename, opts, errp);
548     if (ret < 0 && !*errp) {
549         error_setg_errno(errp, -ret, "Could not create image");
550     }
551 
552     return ret;
553 }
554 
555 /**
556  * Helper function for bdrv_create_file_fallback(): Resize @blk to at
557  * least the given @minimum_size.
558  *
559  * On success, return @blk's actual length.
560  * Otherwise, return -errno.
561  */
562 static int64_t coroutine_fn GRAPH_UNLOCKED
563 create_file_fallback_truncate(BlockBackend *blk, int64_t minimum_size,
564                               Error **errp)
565 {
566     Error *local_err = NULL;
567     int64_t size;
568     int ret;
569 
570     GLOBAL_STATE_CODE();
571 
572     ret = blk_co_truncate(blk, minimum_size, false, PREALLOC_MODE_OFF, 0,
573                           &local_err);
574     if (ret < 0 && ret != -ENOTSUP) {
575         error_propagate(errp, local_err);
576         return ret;
577     }
578 
579     size = blk_co_getlength(blk);
580     if (size < 0) {
581         error_free(local_err);
582         error_setg_errno(errp, -size,
583                          "Failed to inquire the new image file's length");
584         return size;
585     }
586 
587     if (size < minimum_size) {
588         /* Need to grow the image, but we failed to do that */
589         error_propagate(errp, local_err);
590         return -ENOTSUP;
591     }
592 
593     error_free(local_err);
594     local_err = NULL;
595 
596     return size;
597 }
598 
599 /**
600  * Helper function for bdrv_create_file_fallback(): Zero the first
601  * sector to remove any potentially pre-existing image header.
602  */
603 static int coroutine_fn
604 create_file_fallback_zero_first_sector(BlockBackend *blk,
605                                        int64_t current_size,
606                                        Error **errp)
607 {
608     int64_t bytes_to_clear;
609     int ret;
610 
611     GLOBAL_STATE_CODE();
612 
613     bytes_to_clear = MIN(current_size, BDRV_SECTOR_SIZE);
614     if (bytes_to_clear) {
615         ret = blk_co_pwrite_zeroes(blk, 0, bytes_to_clear, BDRV_REQ_MAY_UNMAP);
616         if (ret < 0) {
617             error_setg_errno(errp, -ret,
618                              "Failed to clear the new image's first sector");
619             return ret;
620         }
621     }
622 
623     return 0;
624 }
625 
626 /**
627  * Simple implementation of bdrv_co_create_opts for protocol drivers
628  * which only support creation via opening a file
629  * (usually existing raw storage device)
630  */
631 int coroutine_fn bdrv_co_create_opts_simple(BlockDriver *drv,
632                                             const char *filename,
633                                             QemuOpts *opts,
634                                             Error **errp)
635 {
636     BlockBackend *blk;
637     QDict *options;
638     int64_t size = 0;
639     char *buf = NULL;
640     PreallocMode prealloc;
641     Error *local_err = NULL;
642     int ret;
643 
644     GLOBAL_STATE_CODE();
645 
646     size = qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0);
647     buf = qemu_opt_get_del(opts, BLOCK_OPT_PREALLOC);
648     prealloc = qapi_enum_parse(&PreallocMode_lookup, buf,
649                                PREALLOC_MODE_OFF, &local_err);
650     g_free(buf);
651     if (local_err) {
652         error_propagate(errp, local_err);
653         return -EINVAL;
654     }
655 
656     if (prealloc != PREALLOC_MODE_OFF) {
657         error_setg(errp, "Unsupported preallocation mode '%s'",
658                    PreallocMode_str(prealloc));
659         return -ENOTSUP;
660     }
661 
662     options = qdict_new();
663     qdict_put_str(options, "driver", drv->format_name);
664 
665     blk = blk_co_new_open(filename, NULL, options,
666                           BDRV_O_RDWR | BDRV_O_RESIZE, errp);
667     if (!blk) {
668         error_prepend(errp, "Protocol driver '%s' does not support creating "
669                       "new images, so an existing image must be selected as "
670                       "the target; however, opening the given target as an "
671                       "existing image failed: ",
672                       drv->format_name);
673         return -EINVAL;
674     }
675 
676     size = create_file_fallback_truncate(blk, size, errp);
677     if (size < 0) {
678         ret = size;
679         goto out;
680     }
681 
682     ret = create_file_fallback_zero_first_sector(blk, size, errp);
683     if (ret < 0) {
684         goto out;
685     }
686 
687     ret = 0;
688 out:
689     blk_co_unref(blk);
690     return ret;
691 }
692 
693 int coroutine_fn bdrv_co_create_file(const char *filename, QemuOpts *opts,
694                                      Error **errp)
695 {
696     QemuOpts *protocol_opts;
697     BlockDriver *drv;
698     QDict *qdict;
699     int ret;
700 
701     GLOBAL_STATE_CODE();
702 
703     drv = bdrv_find_protocol(filename, true, errp);
704     if (drv == NULL) {
705         return -ENOENT;
706     }
707 
708     if (!drv->create_opts) {
709         error_setg(errp, "Driver '%s' does not support image creation",
710                    drv->format_name);
711         return -ENOTSUP;
712     }
713 
714     /*
715      * 'opts' contains a QemuOptsList with a combination of format and protocol
716      * default values.
717      *
718      * The format properly removes its options, but the default values remain
719      * in 'opts->list'.  So if the protocol has options with the same name
720      * (e.g. rbd has 'cluster_size' as qcow2), it will see the default values
721      * of the format, since for overlapping options, the format wins.
722      *
723      * To avoid this issue, lets convert QemuOpts to QDict, in this way we take
724      * only the set options, and then convert it back to QemuOpts, using the
725      * create_opts of the protocol. So the new QemuOpts, will contain only the
726      * protocol defaults.
727      */
728     qdict = qemu_opts_to_qdict(opts, NULL);
729     protocol_opts = qemu_opts_from_qdict(drv->create_opts, qdict, errp);
730     if (protocol_opts == NULL) {
731         ret = -EINVAL;
732         goto out;
733     }
734 
735     ret = bdrv_co_create(drv, filename, protocol_opts, errp);
736 out:
737     qemu_opts_del(protocol_opts);
738     qobject_unref(qdict);
739     return ret;
740 }
741 
742 int coroutine_fn bdrv_co_delete_file(BlockDriverState *bs, Error **errp)
743 {
744     Error *local_err = NULL;
745     int ret;
746 
747     IO_CODE();
748     assert(bs != NULL);
749     assert_bdrv_graph_readable();
750 
751     if (!bs->drv) {
752         error_setg(errp, "Block node '%s' is not opened", bs->filename);
753         return -ENOMEDIUM;
754     }
755 
756     if (!bs->drv->bdrv_co_delete_file) {
757         error_setg(errp, "Driver '%s' does not support image deletion",
758                    bs->drv->format_name);
759         return -ENOTSUP;
760     }
761 
762     ret = bs->drv->bdrv_co_delete_file(bs, &local_err);
763     if (ret < 0) {
764         error_propagate(errp, local_err);
765     }
766 
767     return ret;
768 }
769 
770 void coroutine_fn bdrv_co_delete_file_noerr(BlockDriverState *bs)
771 {
772     Error *local_err = NULL;
773     int ret;
774     IO_CODE();
775 
776     if (!bs) {
777         return;
778     }
779 
780     ret = bdrv_co_delete_file(bs, &local_err);
781     /*
782      * ENOTSUP will happen if the block driver doesn't support
783      * the 'bdrv_co_delete_file' interface. This is a predictable
784      * scenario and shouldn't be reported back to the user.
785      */
786     if (ret == -ENOTSUP) {
787         error_free(local_err);
788     } else if (ret < 0) {
789         error_report_err(local_err);
790     }
791 }
792 
793 /**
794  * Try to get @bs's logical and physical block size.
795  * On success, store them in @bsz struct and return 0.
796  * On failure return -errno.
797  * @bs must not be empty.
798  */
799 int bdrv_probe_blocksizes(BlockDriverState *bs, BlockSizes *bsz)
800 {
801     BlockDriver *drv = bs->drv;
802     BlockDriverState *filtered = bdrv_filter_bs(bs);
803     GLOBAL_STATE_CODE();
804 
805     if (drv && drv->bdrv_probe_blocksizes) {
806         return drv->bdrv_probe_blocksizes(bs, bsz);
807     } else if (filtered) {
808         return bdrv_probe_blocksizes(filtered, bsz);
809     }
810 
811     return -ENOTSUP;
812 }
813 
814 /**
815  * Try to get @bs's geometry (cyls, heads, sectors).
816  * On success, store them in @geo struct and return 0.
817  * On failure return -errno.
818  * @bs must not be empty.
819  */
820 int bdrv_probe_geometry(BlockDriverState *bs, HDGeometry *geo)
821 {
822     BlockDriver *drv = bs->drv;
823     BlockDriverState *filtered;
824 
825     GLOBAL_STATE_CODE();
826     GRAPH_RDLOCK_GUARD_MAINLOOP();
827 
828     if (drv && drv->bdrv_probe_geometry) {
829         return drv->bdrv_probe_geometry(bs, geo);
830     }
831 
832     filtered = bdrv_filter_bs(bs);
833     if (filtered) {
834         return bdrv_probe_geometry(filtered, geo);
835     }
836 
837     return -ENOTSUP;
838 }
839 
840 /*
841  * Create a uniquely-named empty temporary file.
842  * Return the actual file name used upon success, otherwise NULL.
843  * This string should be freed with g_free() when not needed any longer.
844  *
845  * Note: creating a temporary file for the caller to (re)open is
846  * inherently racy. Use g_file_open_tmp() instead whenever practical.
847  */
848 char *create_tmp_file(Error **errp)
849 {
850     int fd;
851     const char *tmpdir;
852     g_autofree char *filename = NULL;
853 
854     tmpdir = g_get_tmp_dir();
855 #ifndef _WIN32
856     /*
857      * See commit 69bef79 ("block: use /var/tmp instead of /tmp for -snapshot")
858      *
859      * This function is used to create temporary disk images (like -snapshot),
860      * so the files can become very large. /tmp is often a tmpfs where as
861      * /var/tmp is usually on a disk, so more appropriate for disk images.
862      */
863     if (!g_strcmp0(tmpdir, "/tmp")) {
864         tmpdir = "/var/tmp";
865     }
866 #endif
867 
868     filename = g_strdup_printf("%s/vl.XXXXXX", tmpdir);
869     fd = g_mkstemp(filename);
870     if (fd < 0) {
871         error_setg_errno(errp, errno, "Could not open temporary file '%s'",
872                          filename);
873         return NULL;
874     }
875     close(fd);
876 
877     return g_steal_pointer(&filename);
878 }
879 
880 /*
881  * Detect host devices. By convention, /dev/cdrom[N] is always
882  * recognized as a host CDROM.
883  */
884 static BlockDriver *find_hdev_driver(const char *filename)
885 {
886     int score_max = 0, score;
887     BlockDriver *drv = NULL, *d;
888     GLOBAL_STATE_CODE();
889 
890     QLIST_FOREACH(d, &bdrv_drivers, list) {
891         if (d->bdrv_probe_device) {
892             score = d->bdrv_probe_device(filename);
893             if (score > score_max) {
894                 score_max = score;
895                 drv = d;
896             }
897         }
898     }
899 
900     return drv;
901 }
902 
903 static BlockDriver *bdrv_do_find_protocol(const char *protocol)
904 {
905     BlockDriver *drv1;
906     GLOBAL_STATE_CODE();
907 
908     QLIST_FOREACH(drv1, &bdrv_drivers, list) {
909         if (drv1->protocol_name && !strcmp(drv1->protocol_name, protocol)) {
910             return drv1;
911         }
912     }
913 
914     return NULL;
915 }
916 
917 BlockDriver *bdrv_find_protocol(const char *filename,
918                                 bool allow_protocol_prefix,
919                                 Error **errp)
920 {
921     BlockDriver *drv1;
922     char protocol[128];
923     int len;
924     const char *p;
925     int i;
926 
927     GLOBAL_STATE_CODE();
928     /* TODO Drivers without bdrv_file_open must be specified explicitly */
929 
930     /*
931      * XXX(hch): we really should not let host device detection
932      * override an explicit protocol specification, but moving this
933      * later breaks access to device names with colons in them.
934      * Thanks to the brain-dead persistent naming schemes on udev-
935      * based Linux systems those actually are quite common.
936      */
937     drv1 = find_hdev_driver(filename);
938     if (drv1) {
939         return drv1;
940     }
941 
942     if (!path_has_protocol(filename) || !allow_protocol_prefix) {
943         return &bdrv_file;
944     }
945 
946     p = strchr(filename, ':');
947     assert(p != NULL);
948     len = p - filename;
949     if (len > sizeof(protocol) - 1)
950         len = sizeof(protocol) - 1;
951     memcpy(protocol, filename, len);
952     protocol[len] = '\0';
953 
954     drv1 = bdrv_do_find_protocol(protocol);
955     if (drv1) {
956         return drv1;
957     }
958 
959     for (i = 0; i < (int)ARRAY_SIZE(block_driver_modules); ++i) {
960         if (block_driver_modules[i].protocol_name &&
961             !strcmp(block_driver_modules[i].protocol_name, protocol)) {
962             int rv = block_module_load(block_driver_modules[i].library_name, errp);
963             if (rv > 0) {
964                 drv1 = bdrv_do_find_protocol(protocol);
965             } else if (rv < 0) {
966                 return NULL;
967             }
968             break;
969         }
970     }
971 
972     if (!drv1) {
973         error_setg(errp, "Unknown protocol '%s'", protocol);
974     }
975     return drv1;
976 }
977 
978 /*
979  * Guess image format by probing its contents.
980  * This is not a good idea when your image is raw (CVE-2008-2004), but
981  * we do it anyway for backward compatibility.
982  *
983  * @buf         contains the image's first @buf_size bytes.
984  * @buf_size    is the buffer size in bytes (generally BLOCK_PROBE_BUF_SIZE,
985  *              but can be smaller if the image file is smaller)
986  * @filename    is its filename.
987  *
988  * For all block drivers, call the bdrv_probe() method to get its
989  * probing score.
990  * Return the first block driver with the highest probing score.
991  */
992 BlockDriver *bdrv_probe_all(const uint8_t *buf, int buf_size,
993                             const char *filename)
994 {
995     int score_max = 0, score;
996     BlockDriver *drv = NULL, *d;
997     IO_CODE();
998 
999     QLIST_FOREACH(d, &bdrv_drivers, list) {
1000         if (d->bdrv_probe) {
1001             score = d->bdrv_probe(buf, buf_size, filename);
1002             if (score > score_max) {
1003                 score_max = score;
1004                 drv = d;
1005             }
1006         }
1007     }
1008 
1009     return drv;
1010 }
1011 
1012 static int find_image_format(BlockBackend *file, const char *filename,
1013                              BlockDriver **pdrv, Error **errp)
1014 {
1015     BlockDriver *drv;
1016     uint8_t buf[BLOCK_PROBE_BUF_SIZE];
1017     int ret = 0;
1018 
1019     GLOBAL_STATE_CODE();
1020 
1021     /* Return the raw BlockDriver * to scsi-generic devices or empty drives */
1022     if (blk_is_sg(file) || !blk_is_inserted(file) || blk_getlength(file) == 0) {
1023         *pdrv = &bdrv_raw;
1024         return ret;
1025     }
1026 
1027     ret = blk_pread(file, 0, sizeof(buf), buf, 0);
1028     if (ret < 0) {
1029         error_setg_errno(errp, -ret, "Could not read image for determining its "
1030                          "format");
1031         *pdrv = NULL;
1032         return ret;
1033     }
1034 
1035     drv = bdrv_probe_all(buf, sizeof(buf), filename);
1036     if (!drv) {
1037         error_setg(errp, "Could not determine image format: No compatible "
1038                    "driver found");
1039         *pdrv = NULL;
1040         return -ENOENT;
1041     }
1042 
1043     *pdrv = drv;
1044     return 0;
1045 }
1046 
1047 /**
1048  * Set the current 'total_sectors' value
1049  * Return 0 on success, -errno on error.
1050  */
1051 int coroutine_fn bdrv_co_refresh_total_sectors(BlockDriverState *bs,
1052                                                int64_t hint)
1053 {
1054     BlockDriver *drv = bs->drv;
1055     IO_CODE();
1056     assert_bdrv_graph_readable();
1057 
1058     if (!drv) {
1059         return -ENOMEDIUM;
1060     }
1061 
1062     /* Do not attempt drv->bdrv_co_getlength() on scsi-generic devices */
1063     if (bdrv_is_sg(bs))
1064         return 0;
1065 
1066     /* query actual device if possible, otherwise just trust the hint */
1067     if (drv->bdrv_co_getlength) {
1068         int64_t length = drv->bdrv_co_getlength(bs);
1069         if (length < 0) {
1070             return length;
1071         }
1072         hint = DIV_ROUND_UP(length, BDRV_SECTOR_SIZE);
1073     }
1074 
1075     bs->total_sectors = hint;
1076 
1077     if (bs->total_sectors * BDRV_SECTOR_SIZE > BDRV_MAX_LENGTH) {
1078         return -EFBIG;
1079     }
1080 
1081     return 0;
1082 }
1083 
1084 /**
1085  * Combines a QDict of new block driver @options with any missing options taken
1086  * from @old_options, so that leaving out an option defaults to its old value.
1087  */
1088 static void bdrv_join_options(BlockDriverState *bs, QDict *options,
1089                               QDict *old_options)
1090 {
1091     GLOBAL_STATE_CODE();
1092     if (bs->drv && bs->drv->bdrv_join_options) {
1093         bs->drv->bdrv_join_options(options, old_options);
1094     } else {
1095         qdict_join(options, old_options, false);
1096     }
1097 }
1098 
1099 static BlockdevDetectZeroesOptions bdrv_parse_detect_zeroes(QemuOpts *opts,
1100                                                             int open_flags,
1101                                                             Error **errp)
1102 {
1103     Error *local_err = NULL;
1104     char *value = qemu_opt_get_del(opts, "detect-zeroes");
1105     BlockdevDetectZeroesOptions detect_zeroes =
1106         qapi_enum_parse(&BlockdevDetectZeroesOptions_lookup, value,
1107                         BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF, &local_err);
1108     GLOBAL_STATE_CODE();
1109     g_free(value);
1110     if (local_err) {
1111         error_propagate(errp, local_err);
1112         return detect_zeroes;
1113     }
1114 
1115     if (detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP &&
1116         !(open_flags & BDRV_O_UNMAP))
1117     {
1118         error_setg(errp, "setting detect-zeroes to unmap is not allowed "
1119                    "without setting discard operation to unmap");
1120     }
1121 
1122     return detect_zeroes;
1123 }
1124 
1125 /**
1126  * Set open flags for aio engine
1127  *
1128  * Return 0 on success, -1 if the engine specified is invalid
1129  */
1130 int bdrv_parse_aio(const char *mode, int *flags)
1131 {
1132     if (!strcmp(mode, "threads")) {
1133         /* do nothing, default */
1134     } else if (!strcmp(mode, "native")) {
1135         *flags |= BDRV_O_NATIVE_AIO;
1136 #ifdef CONFIG_LINUX_IO_URING
1137     } else if (!strcmp(mode, "io_uring")) {
1138         *flags |= BDRV_O_IO_URING;
1139 #endif
1140     } else {
1141         return -1;
1142     }
1143 
1144     return 0;
1145 }
1146 
1147 /**
1148  * Set open flags for a given discard mode
1149  *
1150  * Return 0 on success, -1 if the discard mode was invalid.
1151  */
1152 int bdrv_parse_discard_flags(const char *mode, int *flags)
1153 {
1154     *flags &= ~BDRV_O_UNMAP;
1155 
1156     if (!strcmp(mode, "off") || !strcmp(mode, "ignore")) {
1157         /* do nothing */
1158     } else if (!strcmp(mode, "on") || !strcmp(mode, "unmap")) {
1159         *flags |= BDRV_O_UNMAP;
1160     } else {
1161         return -1;
1162     }
1163 
1164     return 0;
1165 }
1166 
1167 /**
1168  * Set open flags for a given cache mode
1169  *
1170  * Return 0 on success, -1 if the cache mode was invalid.
1171  */
1172 int bdrv_parse_cache_mode(const char *mode, int *flags, bool *writethrough)
1173 {
1174     *flags &= ~BDRV_O_CACHE_MASK;
1175 
1176     if (!strcmp(mode, "off") || !strcmp(mode, "none")) {
1177         *writethrough = false;
1178         *flags |= BDRV_O_NOCACHE;
1179     } else if (!strcmp(mode, "directsync")) {
1180         *writethrough = true;
1181         *flags |= BDRV_O_NOCACHE;
1182     } else if (!strcmp(mode, "writeback")) {
1183         *writethrough = false;
1184     } else if (!strcmp(mode, "unsafe")) {
1185         *writethrough = false;
1186         *flags |= BDRV_O_NO_FLUSH;
1187     } else if (!strcmp(mode, "writethrough")) {
1188         *writethrough = true;
1189     } else {
1190         return -1;
1191     }
1192 
1193     return 0;
1194 }
1195 
1196 static char *bdrv_child_get_parent_desc(BdrvChild *c)
1197 {
1198     BlockDriverState *parent = c->opaque;
1199     return g_strdup_printf("node '%s'", bdrv_get_node_name(parent));
1200 }
1201 
1202 static void GRAPH_RDLOCK bdrv_child_cb_drained_begin(BdrvChild *child)
1203 {
1204     BlockDriverState *bs = child->opaque;
1205     bdrv_do_drained_begin_quiesce(bs, NULL);
1206 }
1207 
1208 static bool GRAPH_RDLOCK bdrv_child_cb_drained_poll(BdrvChild *child)
1209 {
1210     BlockDriverState *bs = child->opaque;
1211     return bdrv_drain_poll(bs, NULL, false);
1212 }
1213 
1214 static void GRAPH_RDLOCK bdrv_child_cb_drained_end(BdrvChild *child)
1215 {
1216     BlockDriverState *bs = child->opaque;
1217     bdrv_drained_end(bs);
1218 }
1219 
1220 static int bdrv_child_cb_inactivate(BdrvChild *child)
1221 {
1222     BlockDriverState *bs = child->opaque;
1223     GLOBAL_STATE_CODE();
1224     assert(bs->open_flags & BDRV_O_INACTIVE);
1225     return 0;
1226 }
1227 
1228 static bool bdrv_child_cb_change_aio_ctx(BdrvChild *child, AioContext *ctx,
1229                                          GHashTable *visited, Transaction *tran,
1230                                          Error **errp)
1231 {
1232     BlockDriverState *bs = child->opaque;
1233     return bdrv_change_aio_context(bs, ctx, visited, tran, errp);
1234 }
1235 
1236 /*
1237  * Returns the options and flags that a temporary snapshot should get, based on
1238  * the originally requested flags (the originally requested image will have
1239  * flags like a backing file)
1240  */
1241 static void bdrv_temp_snapshot_options(int *child_flags, QDict *child_options,
1242                                        int parent_flags, QDict *parent_options)
1243 {
1244     GLOBAL_STATE_CODE();
1245     *child_flags = (parent_flags & ~BDRV_O_SNAPSHOT) | BDRV_O_TEMPORARY;
1246 
1247     /* For temporary files, unconditional cache=unsafe is fine */
1248     qdict_set_default_str(child_options, BDRV_OPT_CACHE_DIRECT, "off");
1249     qdict_set_default_str(child_options, BDRV_OPT_CACHE_NO_FLUSH, "on");
1250 
1251     /* Copy the read-only and discard options from the parent */
1252     qdict_copy_default(child_options, parent_options, BDRV_OPT_READ_ONLY);
1253     qdict_copy_default(child_options, parent_options, BDRV_OPT_DISCARD);
1254 
1255     /* aio=native doesn't work for cache.direct=off, so disable it for the
1256      * temporary snapshot */
1257     *child_flags &= ~BDRV_O_NATIVE_AIO;
1258 }
1259 
1260 static void GRAPH_WRLOCK bdrv_backing_attach(BdrvChild *c)
1261 {
1262     BlockDriverState *parent = c->opaque;
1263     BlockDriverState *backing_hd = c->bs;
1264 
1265     GLOBAL_STATE_CODE();
1266     assert(!parent->backing_blocker);
1267     error_setg(&parent->backing_blocker,
1268                "node is used as backing hd of '%s'",
1269                bdrv_get_device_or_node_name(parent));
1270 
1271     bdrv_refresh_filename(backing_hd);
1272 
1273     parent->open_flags &= ~BDRV_O_NO_BACKING;
1274 
1275     bdrv_op_block_all(backing_hd, parent->backing_blocker);
1276     /* Otherwise we won't be able to commit or stream */
1277     bdrv_op_unblock(backing_hd, BLOCK_OP_TYPE_COMMIT_TARGET,
1278                     parent->backing_blocker);
1279     bdrv_op_unblock(backing_hd, BLOCK_OP_TYPE_STREAM,
1280                     parent->backing_blocker);
1281     /*
1282      * We do backup in 3 ways:
1283      * 1. drive backup
1284      *    The target bs is new opened, and the source is top BDS
1285      * 2. blockdev backup
1286      *    Both the source and the target are top BDSes.
1287      * 3. internal backup(used for block replication)
1288      *    Both the source and the target are backing file
1289      *
1290      * In case 1 and 2, neither the source nor the target is the backing file.
1291      * In case 3, we will block the top BDS, so there is only one block job
1292      * for the top BDS and its backing chain.
1293      */
1294     bdrv_op_unblock(backing_hd, BLOCK_OP_TYPE_BACKUP_SOURCE,
1295                     parent->backing_blocker);
1296     bdrv_op_unblock(backing_hd, BLOCK_OP_TYPE_BACKUP_TARGET,
1297                     parent->backing_blocker);
1298 }
1299 
1300 static void bdrv_backing_detach(BdrvChild *c)
1301 {
1302     BlockDriverState *parent = c->opaque;
1303 
1304     GLOBAL_STATE_CODE();
1305     assert(parent->backing_blocker);
1306     bdrv_op_unblock_all(c->bs, parent->backing_blocker);
1307     error_free(parent->backing_blocker);
1308     parent->backing_blocker = NULL;
1309 }
1310 
1311 static int bdrv_backing_update_filename(BdrvChild *c, BlockDriverState *base,
1312                                         const char *filename, Error **errp)
1313 {
1314     BlockDriverState *parent = c->opaque;
1315     bool read_only = bdrv_is_read_only(parent);
1316     int ret;
1317     GLOBAL_STATE_CODE();
1318 
1319     if (read_only) {
1320         ret = bdrv_reopen_set_read_only(parent, false, errp);
1321         if (ret < 0) {
1322             return ret;
1323         }
1324     }
1325 
1326     ret = bdrv_change_backing_file(parent, filename,
1327                                    base->drv ? base->drv->format_name : "",
1328                                    false);
1329     if (ret < 0) {
1330         error_setg_errno(errp, -ret, "Could not update backing file link");
1331     }
1332 
1333     if (read_only) {
1334         bdrv_reopen_set_read_only(parent, true, NULL);
1335     }
1336 
1337     return ret;
1338 }
1339 
1340 /*
1341  * Returns the options and flags that a generic child of a BDS should
1342  * get, based on the given options and flags for the parent BDS.
1343  */
1344 static void bdrv_inherited_options(BdrvChildRole role, bool parent_is_format,
1345                                    int *child_flags, QDict *child_options,
1346                                    int parent_flags, QDict *parent_options)
1347 {
1348     int flags = parent_flags;
1349     GLOBAL_STATE_CODE();
1350 
1351     /*
1352      * First, decide whether to set, clear, or leave BDRV_O_PROTOCOL.
1353      * Generally, the question to answer is: Should this child be
1354      * format-probed by default?
1355      */
1356 
1357     /*
1358      * Pure and non-filtered data children of non-format nodes should
1359      * be probed by default (even when the node itself has BDRV_O_PROTOCOL
1360      * set).  This only affects a very limited set of drivers (namely
1361      * quorum and blkverify when this comment was written).
1362      * Force-clear BDRV_O_PROTOCOL then.
1363      */
1364     if (!parent_is_format &&
1365         (role & BDRV_CHILD_DATA) &&
1366         !(role & (BDRV_CHILD_METADATA | BDRV_CHILD_FILTERED)))
1367     {
1368         flags &= ~BDRV_O_PROTOCOL;
1369     }
1370 
1371     /*
1372      * All children of format nodes (except for COW children) and all
1373      * metadata children in general should never be format-probed.
1374      * Force-set BDRV_O_PROTOCOL then.
1375      */
1376     if ((parent_is_format && !(role & BDRV_CHILD_COW)) ||
1377         (role & BDRV_CHILD_METADATA))
1378     {
1379         flags |= BDRV_O_PROTOCOL;
1380     }
1381 
1382     /*
1383      * If the cache mode isn't explicitly set, inherit direct and no-flush from
1384      * the parent.
1385      */
1386     qdict_copy_default(child_options, parent_options, BDRV_OPT_CACHE_DIRECT);
1387     qdict_copy_default(child_options, parent_options, BDRV_OPT_CACHE_NO_FLUSH);
1388     qdict_copy_default(child_options, parent_options, BDRV_OPT_FORCE_SHARE);
1389 
1390     if (role & BDRV_CHILD_COW) {
1391         /* backing files are opened read-only by default */
1392         qdict_set_default_str(child_options, BDRV_OPT_READ_ONLY, "on");
1393         qdict_set_default_str(child_options, BDRV_OPT_AUTO_READ_ONLY, "off");
1394     } else {
1395         /* Inherit the read-only option from the parent if it's not set */
1396         qdict_copy_default(child_options, parent_options, BDRV_OPT_READ_ONLY);
1397         qdict_copy_default(child_options, parent_options,
1398                            BDRV_OPT_AUTO_READ_ONLY);
1399     }
1400 
1401     /*
1402      * bdrv_co_pdiscard() respects unmap policy for the parent, so we
1403      * can default to enable it on lower layers regardless of the
1404      * parent option.
1405      */
1406     qdict_set_default_str(child_options, BDRV_OPT_DISCARD, "unmap");
1407 
1408     /* Clear flags that only apply to the top layer */
1409     flags &= ~(BDRV_O_SNAPSHOT | BDRV_O_NO_BACKING | BDRV_O_COPY_ON_READ);
1410 
1411     if (role & BDRV_CHILD_METADATA) {
1412         flags &= ~BDRV_O_NO_IO;
1413     }
1414     if (role & BDRV_CHILD_COW) {
1415         flags &= ~BDRV_O_TEMPORARY;
1416     }
1417 
1418     *child_flags = flags;
1419 }
1420 
1421 static void GRAPH_WRLOCK bdrv_child_cb_attach(BdrvChild *child)
1422 {
1423     BlockDriverState *bs = child->opaque;
1424 
1425     assert_bdrv_graph_writable();
1426     QLIST_INSERT_HEAD(&bs->children, child, next);
1427     if (bs->drv->is_filter || (child->role & BDRV_CHILD_FILTERED)) {
1428         /*
1429          * Here we handle filters and block/raw-format.c when it behave like
1430          * filter. They generally have a single PRIMARY child, which is also the
1431          * FILTERED child, and that they may have multiple more children, which
1432          * are neither PRIMARY nor FILTERED. And never we have a COW child here.
1433          * So bs->file will be the PRIMARY child, unless the PRIMARY child goes
1434          * into bs->backing on exceptional cases; and bs->backing will be
1435          * nothing else.
1436          */
1437         assert(!(child->role & BDRV_CHILD_COW));
1438         if (child->role & BDRV_CHILD_PRIMARY) {
1439             assert(child->role & BDRV_CHILD_FILTERED);
1440             assert(!bs->backing);
1441             assert(!bs->file);
1442 
1443             if (bs->drv->filtered_child_is_backing) {
1444                 bs->backing = child;
1445             } else {
1446                 bs->file = child;
1447             }
1448         } else {
1449             assert(!(child->role & BDRV_CHILD_FILTERED));
1450         }
1451     } else if (child->role & BDRV_CHILD_COW) {
1452         assert(bs->drv->supports_backing);
1453         assert(!(child->role & BDRV_CHILD_PRIMARY));
1454         assert(!bs->backing);
1455         bs->backing = child;
1456         bdrv_backing_attach(child);
1457     } else if (child->role & BDRV_CHILD_PRIMARY) {
1458         assert(!bs->file);
1459         bs->file = child;
1460     }
1461 }
1462 
1463 static void GRAPH_WRLOCK bdrv_child_cb_detach(BdrvChild *child)
1464 {
1465     BlockDriverState *bs = child->opaque;
1466 
1467     if (child->role & BDRV_CHILD_COW) {
1468         bdrv_backing_detach(child);
1469     }
1470 
1471     assert_bdrv_graph_writable();
1472     QLIST_REMOVE(child, next);
1473     if (child == bs->backing) {
1474         assert(child != bs->file);
1475         bs->backing = NULL;
1476     } else if (child == bs->file) {
1477         bs->file = NULL;
1478     }
1479 }
1480 
1481 static int bdrv_child_cb_update_filename(BdrvChild *c, BlockDriverState *base,
1482                                          const char *filename, Error **errp)
1483 {
1484     if (c->role & BDRV_CHILD_COW) {
1485         return bdrv_backing_update_filename(c, base, filename, errp);
1486     }
1487     return 0;
1488 }
1489 
1490 AioContext *child_of_bds_get_parent_aio_context(BdrvChild *c)
1491 {
1492     BlockDriverState *bs = c->opaque;
1493     IO_CODE();
1494 
1495     return bdrv_get_aio_context(bs);
1496 }
1497 
1498 const BdrvChildClass child_of_bds = {
1499     .parent_is_bds   = true,
1500     .get_parent_desc = bdrv_child_get_parent_desc,
1501     .inherit_options = bdrv_inherited_options,
1502     .drained_begin   = bdrv_child_cb_drained_begin,
1503     .drained_poll    = bdrv_child_cb_drained_poll,
1504     .drained_end     = bdrv_child_cb_drained_end,
1505     .attach          = bdrv_child_cb_attach,
1506     .detach          = bdrv_child_cb_detach,
1507     .inactivate      = bdrv_child_cb_inactivate,
1508     .change_aio_ctx  = bdrv_child_cb_change_aio_ctx,
1509     .update_filename = bdrv_child_cb_update_filename,
1510     .get_parent_aio_context = child_of_bds_get_parent_aio_context,
1511 };
1512 
1513 AioContext *bdrv_child_get_parent_aio_context(BdrvChild *c)
1514 {
1515     IO_CODE();
1516     return c->klass->get_parent_aio_context(c);
1517 }
1518 
1519 static int bdrv_open_flags(BlockDriverState *bs, int flags)
1520 {
1521     int open_flags = flags;
1522     GLOBAL_STATE_CODE();
1523 
1524     /*
1525      * Clear flags that are internal to the block layer before opening the
1526      * image.
1527      */
1528     open_flags &= ~(BDRV_O_SNAPSHOT | BDRV_O_NO_BACKING | BDRV_O_PROTOCOL);
1529 
1530     return open_flags;
1531 }
1532 
1533 static void update_flags_from_options(int *flags, QemuOpts *opts)
1534 {
1535     GLOBAL_STATE_CODE();
1536 
1537     *flags &= ~(BDRV_O_CACHE_MASK | BDRV_O_RDWR | BDRV_O_AUTO_RDONLY);
1538 
1539     if (qemu_opt_get_bool_del(opts, BDRV_OPT_CACHE_NO_FLUSH, false)) {
1540         *flags |= BDRV_O_NO_FLUSH;
1541     }
1542 
1543     if (qemu_opt_get_bool_del(opts, BDRV_OPT_CACHE_DIRECT, false)) {
1544         *flags |= BDRV_O_NOCACHE;
1545     }
1546 
1547     if (!qemu_opt_get_bool_del(opts, BDRV_OPT_READ_ONLY, false)) {
1548         *flags |= BDRV_O_RDWR;
1549     }
1550 
1551     if (qemu_opt_get_bool_del(opts, BDRV_OPT_AUTO_READ_ONLY, false)) {
1552         *flags |= BDRV_O_AUTO_RDONLY;
1553     }
1554 }
1555 
1556 static void update_options_from_flags(QDict *options, int flags)
1557 {
1558     GLOBAL_STATE_CODE();
1559     if (!qdict_haskey(options, BDRV_OPT_CACHE_DIRECT)) {
1560         qdict_put_bool(options, BDRV_OPT_CACHE_DIRECT, flags & BDRV_O_NOCACHE);
1561     }
1562     if (!qdict_haskey(options, BDRV_OPT_CACHE_NO_FLUSH)) {
1563         qdict_put_bool(options, BDRV_OPT_CACHE_NO_FLUSH,
1564                        flags & BDRV_O_NO_FLUSH);
1565     }
1566     if (!qdict_haskey(options, BDRV_OPT_READ_ONLY)) {
1567         qdict_put_bool(options, BDRV_OPT_READ_ONLY, !(flags & BDRV_O_RDWR));
1568     }
1569     if (!qdict_haskey(options, BDRV_OPT_AUTO_READ_ONLY)) {
1570         qdict_put_bool(options, BDRV_OPT_AUTO_READ_ONLY,
1571                        flags & BDRV_O_AUTO_RDONLY);
1572     }
1573 }
1574 
1575 static void bdrv_assign_node_name(BlockDriverState *bs,
1576                                   const char *node_name,
1577                                   Error **errp)
1578 {
1579     char *gen_node_name = NULL;
1580     GLOBAL_STATE_CODE();
1581 
1582     if (!node_name) {
1583         node_name = gen_node_name = id_generate(ID_BLOCK);
1584     } else if (!id_wellformed(node_name)) {
1585         /*
1586          * Check for empty string or invalid characters, but not if it is
1587          * generated (generated names use characters not available to the user)
1588          */
1589         error_setg(errp, "Invalid node-name: '%s'", node_name);
1590         return;
1591     }
1592 
1593     /* takes care of avoiding namespaces collisions */
1594     if (blk_by_name(node_name)) {
1595         error_setg(errp, "node-name=%s is conflicting with a device id",
1596                    node_name);
1597         goto out;
1598     }
1599 
1600     /* takes care of avoiding duplicates node names */
1601     if (bdrv_find_node(node_name)) {
1602         error_setg(errp, "Duplicate nodes with node-name='%s'", node_name);
1603         goto out;
1604     }
1605 
1606     /* Make sure that the node name isn't truncated */
1607     if (strlen(node_name) >= sizeof(bs->node_name)) {
1608         error_setg(errp, "Node name too long");
1609         goto out;
1610     }
1611 
1612     /* copy node name into the bs and insert it into the graph list */
1613     pstrcpy(bs->node_name, sizeof(bs->node_name), node_name);
1614     QTAILQ_INSERT_TAIL(&graph_bdrv_states, bs, node_list);
1615 out:
1616     g_free(gen_node_name);
1617 }
1618 
1619 /*
1620  * The caller must always hold @bs AioContext lock, because this function calls
1621  * bdrv_refresh_total_sectors() which polls when called from non-coroutine
1622  * context.
1623  */
1624 static int no_coroutine_fn GRAPH_UNLOCKED
1625 bdrv_open_driver(BlockDriverState *bs, BlockDriver *drv, const char *node_name,
1626                  QDict *options, int open_flags, Error **errp)
1627 {
1628     Error *local_err = NULL;
1629     int i, ret;
1630     GLOBAL_STATE_CODE();
1631 
1632     bdrv_assign_node_name(bs, node_name, &local_err);
1633     if (local_err) {
1634         error_propagate(errp, local_err);
1635         return -EINVAL;
1636     }
1637 
1638     bs->drv = drv;
1639     bs->opaque = g_malloc0(drv->instance_size);
1640 
1641     if (drv->bdrv_file_open) {
1642         assert(!drv->bdrv_needs_filename || bs->filename[0]);
1643         ret = drv->bdrv_file_open(bs, options, open_flags, &local_err);
1644     } else if (drv->bdrv_open) {
1645         ret = drv->bdrv_open(bs, options, open_flags, &local_err);
1646     } else {
1647         ret = 0;
1648     }
1649 
1650     if (ret < 0) {
1651         if (local_err) {
1652             error_propagate(errp, local_err);
1653         } else if (bs->filename[0]) {
1654             error_setg_errno(errp, -ret, "Could not open '%s'", bs->filename);
1655         } else {
1656             error_setg_errno(errp, -ret, "Could not open image");
1657         }
1658         goto open_failed;
1659     }
1660 
1661     assert(!(bs->supported_read_flags & ~BDRV_REQ_MASK));
1662     assert(!(bs->supported_write_flags & ~BDRV_REQ_MASK));
1663 
1664     /*
1665      * Always allow the BDRV_REQ_REGISTERED_BUF optimization hint. This saves
1666      * drivers that pass read/write requests through to a child the trouble of
1667      * declaring support explicitly.
1668      *
1669      * Drivers must not propagate this flag accidentally when they initiate I/O
1670      * to a bounce buffer. That case should be rare though.
1671      */
1672     bs->supported_read_flags |= BDRV_REQ_REGISTERED_BUF;
1673     bs->supported_write_flags |= BDRV_REQ_REGISTERED_BUF;
1674 
1675     ret = bdrv_refresh_total_sectors(bs, bs->total_sectors);
1676     if (ret < 0) {
1677         error_setg_errno(errp, -ret, "Could not refresh total sector count");
1678         return ret;
1679     }
1680 
1681     bdrv_graph_rdlock_main_loop();
1682     bdrv_refresh_limits(bs, NULL, &local_err);
1683     bdrv_graph_rdunlock_main_loop();
1684 
1685     if (local_err) {
1686         error_propagate(errp, local_err);
1687         return -EINVAL;
1688     }
1689 
1690     assert(bdrv_opt_mem_align(bs) != 0);
1691     assert(bdrv_min_mem_align(bs) != 0);
1692     assert(is_power_of_2(bs->bl.request_alignment));
1693 
1694     for (i = 0; i < bs->quiesce_counter; i++) {
1695         if (drv->bdrv_drain_begin) {
1696             drv->bdrv_drain_begin(bs);
1697         }
1698     }
1699 
1700     return 0;
1701 open_failed:
1702     bs->drv = NULL;
1703 
1704     bdrv_graph_wrlock();
1705     if (bs->file != NULL) {
1706         bdrv_unref_child(bs, bs->file);
1707         assert(!bs->file);
1708     }
1709     bdrv_graph_wrunlock();
1710 
1711     g_free(bs->opaque);
1712     bs->opaque = NULL;
1713     return ret;
1714 }
1715 
1716 /*
1717  * Create and open a block node.
1718  *
1719  * @options is a QDict of options to pass to the block drivers, or NULL for an
1720  * empty set of options. The reference to the QDict belongs to the block layer
1721  * after the call (even on failure), so if the caller intends to reuse the
1722  * dictionary, it needs to use qobject_ref() before calling bdrv_open.
1723  */
1724 BlockDriverState *bdrv_new_open_driver_opts(BlockDriver *drv,
1725                                             const char *node_name,
1726                                             QDict *options, int flags,
1727                                             Error **errp)
1728 {
1729     BlockDriverState *bs;
1730     int ret;
1731 
1732     GLOBAL_STATE_CODE();
1733 
1734     bs = bdrv_new();
1735     bs->open_flags = flags;
1736     bs->options = options ?: qdict_new();
1737     bs->explicit_options = qdict_clone_shallow(bs->options);
1738     bs->opaque = NULL;
1739 
1740     update_options_from_flags(bs->options, flags);
1741 
1742     ret = bdrv_open_driver(bs, drv, node_name, bs->options, flags, errp);
1743     if (ret < 0) {
1744         qobject_unref(bs->explicit_options);
1745         bs->explicit_options = NULL;
1746         qobject_unref(bs->options);
1747         bs->options = NULL;
1748         bdrv_unref(bs);
1749         return NULL;
1750     }
1751 
1752     return bs;
1753 }
1754 
1755 /* Create and open a block node. */
1756 BlockDriverState *bdrv_new_open_driver(BlockDriver *drv, const char *node_name,
1757                                        int flags, Error **errp)
1758 {
1759     GLOBAL_STATE_CODE();
1760     return bdrv_new_open_driver_opts(drv, node_name, NULL, flags, errp);
1761 }
1762 
1763 QemuOptsList bdrv_runtime_opts = {
1764     .name = "bdrv_common",
1765     .head = QTAILQ_HEAD_INITIALIZER(bdrv_runtime_opts.head),
1766     .desc = {
1767         {
1768             .name = "node-name",
1769             .type = QEMU_OPT_STRING,
1770             .help = "Node name of the block device node",
1771         },
1772         {
1773             .name = "driver",
1774             .type = QEMU_OPT_STRING,
1775             .help = "Block driver to use for the node",
1776         },
1777         {
1778             .name = BDRV_OPT_CACHE_DIRECT,
1779             .type = QEMU_OPT_BOOL,
1780             .help = "Bypass software writeback cache on the host",
1781         },
1782         {
1783             .name = BDRV_OPT_CACHE_NO_FLUSH,
1784             .type = QEMU_OPT_BOOL,
1785             .help = "Ignore flush requests",
1786         },
1787         {
1788             .name = BDRV_OPT_READ_ONLY,
1789             .type = QEMU_OPT_BOOL,
1790             .help = "Node is opened in read-only mode",
1791         },
1792         {
1793             .name = BDRV_OPT_AUTO_READ_ONLY,
1794             .type = QEMU_OPT_BOOL,
1795             .help = "Node can become read-only if opening read-write fails",
1796         },
1797         {
1798             .name = "detect-zeroes",
1799             .type = QEMU_OPT_STRING,
1800             .help = "try to optimize zero writes (off, on, unmap)",
1801         },
1802         {
1803             .name = BDRV_OPT_DISCARD,
1804             .type = QEMU_OPT_STRING,
1805             .help = "discard operation (ignore/off, unmap/on)",
1806         },
1807         {
1808             .name = BDRV_OPT_FORCE_SHARE,
1809             .type = QEMU_OPT_BOOL,
1810             .help = "always accept other writers (default: off)",
1811         },
1812         { /* end of list */ }
1813     },
1814 };
1815 
1816 QemuOptsList bdrv_create_opts_simple = {
1817     .name = "simple-create-opts",
1818     .head = QTAILQ_HEAD_INITIALIZER(bdrv_create_opts_simple.head),
1819     .desc = {
1820         {
1821             .name = BLOCK_OPT_SIZE,
1822             .type = QEMU_OPT_SIZE,
1823             .help = "Virtual disk size"
1824         },
1825         {
1826             .name = BLOCK_OPT_PREALLOC,
1827             .type = QEMU_OPT_STRING,
1828             .help = "Preallocation mode (allowed values: off)"
1829         },
1830         { /* end of list */ }
1831     }
1832 };
1833 
1834 /*
1835  * Common part for opening disk images and files
1836  *
1837  * Removes all processed options from *options.
1838  */
1839 static int bdrv_open_common(BlockDriverState *bs, BlockBackend *file,
1840                             QDict *options, Error **errp)
1841 {
1842     int ret, open_flags;
1843     const char *filename;
1844     const char *driver_name = NULL;
1845     const char *node_name = NULL;
1846     const char *discard;
1847     QemuOpts *opts;
1848     BlockDriver *drv;
1849     Error *local_err = NULL;
1850     bool ro;
1851 
1852     GLOBAL_STATE_CODE();
1853 
1854     bdrv_graph_rdlock_main_loop();
1855     assert(bs->file == NULL);
1856     assert(options != NULL && bs->options != options);
1857     bdrv_graph_rdunlock_main_loop();
1858 
1859     opts = qemu_opts_create(&bdrv_runtime_opts, NULL, 0, &error_abort);
1860     if (!qemu_opts_absorb_qdict(opts, options, errp)) {
1861         ret = -EINVAL;
1862         goto fail_opts;
1863     }
1864 
1865     update_flags_from_options(&bs->open_flags, opts);
1866 
1867     driver_name = qemu_opt_get(opts, "driver");
1868     drv = bdrv_find_format(driver_name);
1869     assert(drv != NULL);
1870 
1871     bs->force_share = qemu_opt_get_bool(opts, BDRV_OPT_FORCE_SHARE, false);
1872 
1873     if (bs->force_share && (bs->open_flags & BDRV_O_RDWR)) {
1874         error_setg(errp,
1875                    BDRV_OPT_FORCE_SHARE
1876                    "=on can only be used with read-only images");
1877         ret = -EINVAL;
1878         goto fail_opts;
1879     }
1880 
1881     if (file != NULL) {
1882         bdrv_graph_rdlock_main_loop();
1883         bdrv_refresh_filename(blk_bs(file));
1884         bdrv_graph_rdunlock_main_loop();
1885 
1886         filename = blk_bs(file)->filename;
1887     } else {
1888         /*
1889          * Caution: while qdict_get_try_str() is fine, getting
1890          * non-string types would require more care.  When @options
1891          * come from -blockdev or blockdev_add, its members are typed
1892          * according to the QAPI schema, but when they come from
1893          * -drive, they're all QString.
1894          */
1895         filename = qdict_get_try_str(options, "filename");
1896     }
1897 
1898     if (drv->bdrv_needs_filename && (!filename || !filename[0])) {
1899         error_setg(errp, "The '%s' block driver requires a file name",
1900                    drv->format_name);
1901         ret = -EINVAL;
1902         goto fail_opts;
1903     }
1904 
1905     trace_bdrv_open_common(bs, filename ?: "", bs->open_flags,
1906                            drv->format_name);
1907 
1908     ro = bdrv_is_read_only(bs);
1909 
1910     if (use_bdrv_whitelist && !bdrv_is_whitelisted(drv, ro)) {
1911         if (!ro && bdrv_is_whitelisted(drv, true)) {
1912             bdrv_graph_rdlock_main_loop();
1913             ret = bdrv_apply_auto_read_only(bs, NULL, NULL);
1914             bdrv_graph_rdunlock_main_loop();
1915         } else {
1916             ret = -ENOTSUP;
1917         }
1918         if (ret < 0) {
1919             error_setg(errp,
1920                        !ro && bdrv_is_whitelisted(drv, true)
1921                        ? "Driver '%s' can only be used for read-only devices"
1922                        : "Driver '%s' is not whitelisted",
1923                        drv->format_name);
1924             goto fail_opts;
1925         }
1926     }
1927 
1928     /* bdrv_new() and bdrv_close() make it so */
1929     assert(qatomic_read(&bs->copy_on_read) == 0);
1930 
1931     if (bs->open_flags & BDRV_O_COPY_ON_READ) {
1932         if (!ro) {
1933             bdrv_enable_copy_on_read(bs);
1934         } else {
1935             error_setg(errp, "Can't use copy-on-read on read-only device");
1936             ret = -EINVAL;
1937             goto fail_opts;
1938         }
1939     }
1940 
1941     discard = qemu_opt_get(opts, BDRV_OPT_DISCARD);
1942     if (discard != NULL) {
1943         if (bdrv_parse_discard_flags(discard, &bs->open_flags) != 0) {
1944             error_setg(errp, "Invalid discard option");
1945             ret = -EINVAL;
1946             goto fail_opts;
1947         }
1948     }
1949 
1950     bs->detect_zeroes =
1951         bdrv_parse_detect_zeroes(opts, bs->open_flags, &local_err);
1952     if (local_err) {
1953         error_propagate(errp, local_err);
1954         ret = -EINVAL;
1955         goto fail_opts;
1956     }
1957 
1958     if (filename != NULL) {
1959         pstrcpy(bs->filename, sizeof(bs->filename), filename);
1960     } else {
1961         bs->filename[0] = '\0';
1962     }
1963     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename), bs->filename);
1964 
1965     /* Open the image, either directly or using a protocol */
1966     open_flags = bdrv_open_flags(bs, bs->open_flags);
1967     node_name = qemu_opt_get(opts, "node-name");
1968 
1969     assert(!drv->bdrv_file_open || file == NULL);
1970     ret = bdrv_open_driver(bs, drv, node_name, options, open_flags, errp);
1971     if (ret < 0) {
1972         goto fail_opts;
1973     }
1974 
1975     qemu_opts_del(opts);
1976     return 0;
1977 
1978 fail_opts:
1979     qemu_opts_del(opts);
1980     return ret;
1981 }
1982 
1983 static QDict *parse_json_filename(const char *filename, Error **errp)
1984 {
1985     QObject *options_obj;
1986     QDict *options;
1987     int ret;
1988     GLOBAL_STATE_CODE();
1989 
1990     ret = strstart(filename, "json:", &filename);
1991     assert(ret);
1992 
1993     options_obj = qobject_from_json(filename, errp);
1994     if (!options_obj) {
1995         error_prepend(errp, "Could not parse the JSON options: ");
1996         return NULL;
1997     }
1998 
1999     options = qobject_to(QDict, options_obj);
2000     if (!options) {
2001         qobject_unref(options_obj);
2002         error_setg(errp, "Invalid JSON object given");
2003         return NULL;
2004     }
2005 
2006     qdict_flatten(options);
2007 
2008     return options;
2009 }
2010 
2011 static void parse_json_protocol(QDict *options, const char **pfilename,
2012                                 Error **errp)
2013 {
2014     QDict *json_options;
2015     Error *local_err = NULL;
2016     GLOBAL_STATE_CODE();
2017 
2018     /* Parse json: pseudo-protocol */
2019     if (!*pfilename || !g_str_has_prefix(*pfilename, "json:")) {
2020         return;
2021     }
2022 
2023     json_options = parse_json_filename(*pfilename, &local_err);
2024     if (local_err) {
2025         error_propagate(errp, local_err);
2026         return;
2027     }
2028 
2029     /* Options given in the filename have lower priority than options
2030      * specified directly */
2031     qdict_join(options, json_options, false);
2032     qobject_unref(json_options);
2033     *pfilename = NULL;
2034 }
2035 
2036 /*
2037  * Fills in default options for opening images and converts the legacy
2038  * filename/flags pair to option QDict entries.
2039  * The BDRV_O_PROTOCOL flag in *flags will be set or cleared accordingly if a
2040  * block driver has been specified explicitly.
2041  */
2042 static int bdrv_fill_options(QDict **options, const char *filename,
2043                              int *flags, Error **errp)
2044 {
2045     const char *drvname;
2046     bool protocol = *flags & BDRV_O_PROTOCOL;
2047     bool parse_filename = false;
2048     BlockDriver *drv = NULL;
2049     Error *local_err = NULL;
2050 
2051     GLOBAL_STATE_CODE();
2052 
2053     /*
2054      * Caution: while qdict_get_try_str() is fine, getting non-string
2055      * types would require more care.  When @options come from
2056      * -blockdev or blockdev_add, its members are typed according to
2057      * the QAPI schema, but when they come from -drive, they're all
2058      * QString.
2059      */
2060     drvname = qdict_get_try_str(*options, "driver");
2061     if (drvname) {
2062         drv = bdrv_find_format(drvname);
2063         if (!drv) {
2064             error_setg(errp, "Unknown driver '%s'", drvname);
2065             return -ENOENT;
2066         }
2067         /* If the user has explicitly specified the driver, this choice should
2068          * override the BDRV_O_PROTOCOL flag */
2069         protocol = drv->bdrv_file_open;
2070     }
2071 
2072     if (protocol) {
2073         *flags |= BDRV_O_PROTOCOL;
2074     } else {
2075         *flags &= ~BDRV_O_PROTOCOL;
2076     }
2077 
2078     /* Translate cache options from flags into options */
2079     update_options_from_flags(*options, *flags);
2080 
2081     /* Fetch the file name from the options QDict if necessary */
2082     if (protocol && filename) {
2083         if (!qdict_haskey(*options, "filename")) {
2084             qdict_put_str(*options, "filename", filename);
2085             parse_filename = true;
2086         } else {
2087             error_setg(errp, "Can't specify 'file' and 'filename' options at "
2088                              "the same time");
2089             return -EINVAL;
2090         }
2091     }
2092 
2093     /* Find the right block driver */
2094     /* See cautionary note on accessing @options above */
2095     filename = qdict_get_try_str(*options, "filename");
2096 
2097     if (!drvname && protocol) {
2098         if (filename) {
2099             drv = bdrv_find_protocol(filename, parse_filename, errp);
2100             if (!drv) {
2101                 return -EINVAL;
2102             }
2103 
2104             drvname = drv->format_name;
2105             qdict_put_str(*options, "driver", drvname);
2106         } else {
2107             error_setg(errp, "Must specify either driver or file");
2108             return -EINVAL;
2109         }
2110     }
2111 
2112     assert(drv || !protocol);
2113 
2114     /* Driver-specific filename parsing */
2115     if (drv && drv->bdrv_parse_filename && parse_filename) {
2116         drv->bdrv_parse_filename(filename, *options, &local_err);
2117         if (local_err) {
2118             error_propagate(errp, local_err);
2119             return -EINVAL;
2120         }
2121 
2122         if (!drv->bdrv_needs_filename) {
2123             qdict_del(*options, "filename");
2124         }
2125     }
2126 
2127     return 0;
2128 }
2129 
2130 typedef struct BlockReopenQueueEntry {
2131      bool prepared;
2132      BDRVReopenState state;
2133      QTAILQ_ENTRY(BlockReopenQueueEntry) entry;
2134 } BlockReopenQueueEntry;
2135 
2136 /*
2137  * Return the flags that @bs will have after the reopens in @q have
2138  * successfully completed. If @q is NULL (or @bs is not contained in @q),
2139  * return the current flags.
2140  */
2141 static int bdrv_reopen_get_flags(BlockReopenQueue *q, BlockDriverState *bs)
2142 {
2143     BlockReopenQueueEntry *entry;
2144 
2145     if (q != NULL) {
2146         QTAILQ_FOREACH(entry, q, entry) {
2147             if (entry->state.bs == bs) {
2148                 return entry->state.flags;
2149             }
2150         }
2151     }
2152 
2153     return bs->open_flags;
2154 }
2155 
2156 /* Returns whether the image file can be written to after the reopen queue @q
2157  * has been successfully applied, or right now if @q is NULL. */
2158 static bool bdrv_is_writable_after_reopen(BlockDriverState *bs,
2159                                           BlockReopenQueue *q)
2160 {
2161     int flags = bdrv_reopen_get_flags(q, bs);
2162 
2163     return (flags & (BDRV_O_RDWR | BDRV_O_INACTIVE)) == BDRV_O_RDWR;
2164 }
2165 
2166 /*
2167  * Return whether the BDS can be written to.  This is not necessarily
2168  * the same as !bdrv_is_read_only(bs), as inactivated images may not
2169  * be written to but do not count as read-only images.
2170  */
2171 bool bdrv_is_writable(BlockDriverState *bs)
2172 {
2173     IO_CODE();
2174     return bdrv_is_writable_after_reopen(bs, NULL);
2175 }
2176 
2177 static char *bdrv_child_user_desc(BdrvChild *c)
2178 {
2179     GLOBAL_STATE_CODE();
2180     return c->klass->get_parent_desc(c);
2181 }
2182 
2183 /*
2184  * Check that @a allows everything that @b needs. @a and @b must reference same
2185  * child node.
2186  */
2187 static bool bdrv_a_allow_b(BdrvChild *a, BdrvChild *b, Error **errp)
2188 {
2189     const char *child_bs_name;
2190     g_autofree char *a_user = NULL;
2191     g_autofree char *b_user = NULL;
2192     g_autofree char *perms = NULL;
2193 
2194     assert(a->bs);
2195     assert(a->bs == b->bs);
2196     GLOBAL_STATE_CODE();
2197 
2198     if ((b->perm & a->shared_perm) == b->perm) {
2199         return true;
2200     }
2201 
2202     child_bs_name = bdrv_get_node_name(b->bs);
2203     a_user = bdrv_child_user_desc(a);
2204     b_user = bdrv_child_user_desc(b);
2205     perms = bdrv_perm_names(b->perm & ~a->shared_perm);
2206 
2207     error_setg(errp, "Permission conflict on node '%s': permissions '%s' are "
2208                "both required by %s (uses node '%s' as '%s' child) and "
2209                "unshared by %s (uses node '%s' as '%s' child).",
2210                child_bs_name, perms,
2211                b_user, child_bs_name, b->name,
2212                a_user, child_bs_name, a->name);
2213 
2214     return false;
2215 }
2216 
2217 static bool GRAPH_RDLOCK
2218 bdrv_parent_perms_conflict(BlockDriverState *bs, Error **errp)
2219 {
2220     BdrvChild *a, *b;
2221     GLOBAL_STATE_CODE();
2222 
2223     /*
2224      * During the loop we'll look at each pair twice. That's correct because
2225      * bdrv_a_allow_b() is asymmetric and we should check each pair in both
2226      * directions.
2227      */
2228     QLIST_FOREACH(a, &bs->parents, next_parent) {
2229         QLIST_FOREACH(b, &bs->parents, next_parent) {
2230             if (a == b) {
2231                 continue;
2232             }
2233 
2234             if (!bdrv_a_allow_b(a, b, errp)) {
2235                 return true;
2236             }
2237         }
2238     }
2239 
2240     return false;
2241 }
2242 
2243 static void GRAPH_RDLOCK
2244 bdrv_child_perm(BlockDriverState *bs, BlockDriverState *child_bs,
2245                 BdrvChild *c, BdrvChildRole role,
2246                 BlockReopenQueue *reopen_queue,
2247                 uint64_t parent_perm, uint64_t parent_shared,
2248                 uint64_t *nperm, uint64_t *nshared)
2249 {
2250     assert(bs->drv && bs->drv->bdrv_child_perm);
2251     GLOBAL_STATE_CODE();
2252     bs->drv->bdrv_child_perm(bs, c, role, reopen_queue,
2253                              parent_perm, parent_shared,
2254                              nperm, nshared);
2255     /* TODO Take force_share from reopen_queue */
2256     if (child_bs && child_bs->force_share) {
2257         *nshared = BLK_PERM_ALL;
2258     }
2259 }
2260 
2261 /*
2262  * Adds the whole subtree of @bs (including @bs itself) to the @list (except for
2263  * nodes that are already in the @list, of course) so that final list is
2264  * topologically sorted. Return the result (GSList @list object is updated, so
2265  * don't use old reference after function call).
2266  *
2267  * On function start @list must be already topologically sorted and for any node
2268  * in the @list the whole subtree of the node must be in the @list as well. The
2269  * simplest way to satisfy this criteria: use only result of
2270  * bdrv_topological_dfs() or NULL as @list parameter.
2271  */
2272 static GSList * GRAPH_RDLOCK
2273 bdrv_topological_dfs(GSList *list, GHashTable *found, BlockDriverState *bs)
2274 {
2275     BdrvChild *child;
2276     g_autoptr(GHashTable) local_found = NULL;
2277 
2278     GLOBAL_STATE_CODE();
2279 
2280     if (!found) {
2281         assert(!list);
2282         found = local_found = g_hash_table_new(NULL, NULL);
2283     }
2284 
2285     if (g_hash_table_contains(found, bs)) {
2286         return list;
2287     }
2288     g_hash_table_add(found, bs);
2289 
2290     QLIST_FOREACH(child, &bs->children, next) {
2291         list = bdrv_topological_dfs(list, found, child->bs);
2292     }
2293 
2294     return g_slist_prepend(list, bs);
2295 }
2296 
2297 typedef struct BdrvChildSetPermState {
2298     BdrvChild *child;
2299     uint64_t old_perm;
2300     uint64_t old_shared_perm;
2301 } BdrvChildSetPermState;
2302 
2303 static void bdrv_child_set_perm_abort(void *opaque)
2304 {
2305     BdrvChildSetPermState *s = opaque;
2306 
2307     GLOBAL_STATE_CODE();
2308 
2309     s->child->perm = s->old_perm;
2310     s->child->shared_perm = s->old_shared_perm;
2311 }
2312 
2313 static TransactionActionDrv bdrv_child_set_pem_drv = {
2314     .abort = bdrv_child_set_perm_abort,
2315     .clean = g_free,
2316 };
2317 
2318 static void bdrv_child_set_perm(BdrvChild *c, uint64_t perm,
2319                                 uint64_t shared, Transaction *tran)
2320 {
2321     BdrvChildSetPermState *s = g_new(BdrvChildSetPermState, 1);
2322     GLOBAL_STATE_CODE();
2323 
2324     *s = (BdrvChildSetPermState) {
2325         .child = c,
2326         .old_perm = c->perm,
2327         .old_shared_perm = c->shared_perm,
2328     };
2329 
2330     c->perm = perm;
2331     c->shared_perm = shared;
2332 
2333     tran_add(tran, &bdrv_child_set_pem_drv, s);
2334 }
2335 
2336 static void GRAPH_RDLOCK bdrv_drv_set_perm_commit(void *opaque)
2337 {
2338     BlockDriverState *bs = opaque;
2339     uint64_t cumulative_perms, cumulative_shared_perms;
2340     GLOBAL_STATE_CODE();
2341 
2342     if (bs->drv->bdrv_set_perm) {
2343         bdrv_get_cumulative_perm(bs, &cumulative_perms,
2344                                  &cumulative_shared_perms);
2345         bs->drv->bdrv_set_perm(bs, cumulative_perms, cumulative_shared_perms);
2346     }
2347 }
2348 
2349 static void GRAPH_RDLOCK bdrv_drv_set_perm_abort(void *opaque)
2350 {
2351     BlockDriverState *bs = opaque;
2352     GLOBAL_STATE_CODE();
2353 
2354     if (bs->drv->bdrv_abort_perm_update) {
2355         bs->drv->bdrv_abort_perm_update(bs);
2356     }
2357 }
2358 
2359 TransactionActionDrv bdrv_drv_set_perm_drv = {
2360     .abort = bdrv_drv_set_perm_abort,
2361     .commit = bdrv_drv_set_perm_commit,
2362 };
2363 
2364 /*
2365  * After calling this function, the transaction @tran may only be completed
2366  * while holding a reader lock for the graph.
2367  */
2368 static int GRAPH_RDLOCK
2369 bdrv_drv_set_perm(BlockDriverState *bs, uint64_t perm, uint64_t shared_perm,
2370                   Transaction *tran, Error **errp)
2371 {
2372     GLOBAL_STATE_CODE();
2373     if (!bs->drv) {
2374         return 0;
2375     }
2376 
2377     if (bs->drv->bdrv_check_perm) {
2378         int ret = bs->drv->bdrv_check_perm(bs, perm, shared_perm, errp);
2379         if (ret < 0) {
2380             return ret;
2381         }
2382     }
2383 
2384     if (tran) {
2385         tran_add(tran, &bdrv_drv_set_perm_drv, bs);
2386     }
2387 
2388     return 0;
2389 }
2390 
2391 typedef struct BdrvReplaceChildState {
2392     BdrvChild *child;
2393     BlockDriverState *old_bs;
2394 } BdrvReplaceChildState;
2395 
2396 static void GRAPH_WRLOCK bdrv_replace_child_commit(void *opaque)
2397 {
2398     BdrvReplaceChildState *s = opaque;
2399     GLOBAL_STATE_CODE();
2400 
2401     bdrv_schedule_unref(s->old_bs);
2402 }
2403 
2404 static void GRAPH_WRLOCK bdrv_replace_child_abort(void *opaque)
2405 {
2406     BdrvReplaceChildState *s = opaque;
2407     BlockDriverState *new_bs = s->child->bs;
2408 
2409     GLOBAL_STATE_CODE();
2410     assert_bdrv_graph_writable();
2411 
2412     /* old_bs reference is transparently moved from @s to @s->child */
2413     if (!s->child->bs) {
2414         /*
2415          * The parents were undrained when removing old_bs from the child. New
2416          * requests can't have been made, though, because the child was empty.
2417          *
2418          * TODO Make bdrv_replace_child_noperm() transactionable to avoid
2419          * undraining the parent in the first place. Once this is done, having
2420          * new_bs drained when calling bdrv_replace_child_tran() is not a
2421          * requirement any more.
2422          */
2423         bdrv_parent_drained_begin_single(s->child);
2424         assert(!bdrv_parent_drained_poll_single(s->child));
2425     }
2426     assert(s->child->quiesced_parent);
2427     bdrv_replace_child_noperm(s->child, s->old_bs);
2428 
2429     bdrv_unref(new_bs);
2430 }
2431 
2432 static TransactionActionDrv bdrv_replace_child_drv = {
2433     .commit = bdrv_replace_child_commit,
2434     .abort = bdrv_replace_child_abort,
2435     .clean = g_free,
2436 };
2437 
2438 /*
2439  * bdrv_replace_child_tran
2440  *
2441  * Note: real unref of old_bs is done only on commit.
2442  *
2443  * Both @child->bs and @new_bs (if non-NULL) must be drained. @new_bs must be
2444  * kept drained until the transaction is completed.
2445  *
2446  * After calling this function, the transaction @tran may only be completed
2447  * while holding a writer lock for the graph.
2448  *
2449  * The function doesn't update permissions, caller is responsible for this.
2450  */
2451 static void GRAPH_WRLOCK
2452 bdrv_replace_child_tran(BdrvChild *child, BlockDriverState *new_bs,
2453                         Transaction *tran)
2454 {
2455     BdrvReplaceChildState *s = g_new(BdrvReplaceChildState, 1);
2456 
2457     assert(child->quiesced_parent);
2458     assert(!new_bs || new_bs->quiesce_counter);
2459 
2460     *s = (BdrvReplaceChildState) {
2461         .child = child,
2462         .old_bs = child->bs,
2463     };
2464     tran_add(tran, &bdrv_replace_child_drv, s);
2465 
2466     if (new_bs) {
2467         bdrv_ref(new_bs);
2468     }
2469 
2470     bdrv_replace_child_noperm(child, new_bs);
2471     /* old_bs reference is transparently moved from @child to @s */
2472 }
2473 
2474 /*
2475  * Refresh permissions in @bs subtree. The function is intended to be called
2476  * after some graph modification that was done without permission update.
2477  *
2478  * After calling this function, the transaction @tran may only be completed
2479  * while holding a reader lock for the graph.
2480  */
2481 static int GRAPH_RDLOCK
2482 bdrv_node_refresh_perm(BlockDriverState *bs, BlockReopenQueue *q,
2483                        Transaction *tran, Error **errp)
2484 {
2485     BlockDriver *drv = bs->drv;
2486     BdrvChild *c;
2487     int ret;
2488     uint64_t cumulative_perms, cumulative_shared_perms;
2489     GLOBAL_STATE_CODE();
2490 
2491     bdrv_get_cumulative_perm(bs, &cumulative_perms, &cumulative_shared_perms);
2492 
2493     /* Write permissions never work with read-only images */
2494     if ((cumulative_perms & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED)) &&
2495         !bdrv_is_writable_after_reopen(bs, q))
2496     {
2497         if (!bdrv_is_writable_after_reopen(bs, NULL)) {
2498             error_setg(errp, "Block node is read-only");
2499         } else {
2500             error_setg(errp, "Read-only block node '%s' cannot support "
2501                        "read-write users", bdrv_get_node_name(bs));
2502         }
2503 
2504         return -EPERM;
2505     }
2506 
2507     /*
2508      * Unaligned requests will automatically be aligned to bl.request_alignment
2509      * and without RESIZE we can't extend requests to write to space beyond the
2510      * end of the image, so it's required that the image size is aligned.
2511      */
2512     if ((cumulative_perms & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED)) &&
2513         !(cumulative_perms & BLK_PERM_RESIZE))
2514     {
2515         if ((bs->total_sectors * BDRV_SECTOR_SIZE) % bs->bl.request_alignment) {
2516             error_setg(errp, "Cannot get 'write' permission without 'resize': "
2517                              "Image size is not a multiple of request "
2518                              "alignment");
2519             return -EPERM;
2520         }
2521     }
2522 
2523     /* Check this node */
2524     if (!drv) {
2525         return 0;
2526     }
2527 
2528     ret = bdrv_drv_set_perm(bs, cumulative_perms, cumulative_shared_perms, tran,
2529                             errp);
2530     if (ret < 0) {
2531         return ret;
2532     }
2533 
2534     /* Drivers that never have children can omit .bdrv_child_perm() */
2535     if (!drv->bdrv_child_perm) {
2536         assert(QLIST_EMPTY(&bs->children));
2537         return 0;
2538     }
2539 
2540     /* Check all children */
2541     QLIST_FOREACH(c, &bs->children, next) {
2542         uint64_t cur_perm, cur_shared;
2543 
2544         bdrv_child_perm(bs, c->bs, c, c->role, q,
2545                         cumulative_perms, cumulative_shared_perms,
2546                         &cur_perm, &cur_shared);
2547         bdrv_child_set_perm(c, cur_perm, cur_shared, tran);
2548     }
2549 
2550     return 0;
2551 }
2552 
2553 /*
2554  * @list is a product of bdrv_topological_dfs() (may be called several times) -
2555  * a topologically sorted subgraph.
2556  *
2557  * After calling this function, the transaction @tran may only be completed
2558  * while holding a reader lock for the graph.
2559  */
2560 static int GRAPH_RDLOCK
2561 bdrv_do_refresh_perms(GSList *list, BlockReopenQueue *q, Transaction *tran,
2562                       Error **errp)
2563 {
2564     int ret;
2565     BlockDriverState *bs;
2566     GLOBAL_STATE_CODE();
2567 
2568     for ( ; list; list = list->next) {
2569         bs = list->data;
2570 
2571         if (bdrv_parent_perms_conflict(bs, errp)) {
2572             return -EINVAL;
2573         }
2574 
2575         ret = bdrv_node_refresh_perm(bs, q, tran, errp);
2576         if (ret < 0) {
2577             return ret;
2578         }
2579     }
2580 
2581     return 0;
2582 }
2583 
2584 /*
2585  * @list is any list of nodes. List is completed by all subtrees and
2586  * topologically sorted. It's not a problem if some node occurs in the @list
2587  * several times.
2588  *
2589  * After calling this function, the transaction @tran may only be completed
2590  * while holding a reader lock for the graph.
2591  */
2592 static int GRAPH_RDLOCK
2593 bdrv_list_refresh_perms(GSList *list, BlockReopenQueue *q, Transaction *tran,
2594                         Error **errp)
2595 {
2596     g_autoptr(GHashTable) found = g_hash_table_new(NULL, NULL);
2597     g_autoptr(GSList) refresh_list = NULL;
2598 
2599     for ( ; list; list = list->next) {
2600         refresh_list = bdrv_topological_dfs(refresh_list, found, list->data);
2601     }
2602 
2603     return bdrv_do_refresh_perms(refresh_list, q, tran, errp);
2604 }
2605 
2606 void bdrv_get_cumulative_perm(BlockDriverState *bs, uint64_t *perm,
2607                               uint64_t *shared_perm)
2608 {
2609     BdrvChild *c;
2610     uint64_t cumulative_perms = 0;
2611     uint64_t cumulative_shared_perms = BLK_PERM_ALL;
2612 
2613     GLOBAL_STATE_CODE();
2614 
2615     QLIST_FOREACH(c, &bs->parents, next_parent) {
2616         cumulative_perms |= c->perm;
2617         cumulative_shared_perms &= c->shared_perm;
2618     }
2619 
2620     *perm = cumulative_perms;
2621     *shared_perm = cumulative_shared_perms;
2622 }
2623 
2624 char *bdrv_perm_names(uint64_t perm)
2625 {
2626     struct perm_name {
2627         uint64_t perm;
2628         const char *name;
2629     } permissions[] = {
2630         { BLK_PERM_CONSISTENT_READ, "consistent read" },
2631         { BLK_PERM_WRITE,           "write" },
2632         { BLK_PERM_WRITE_UNCHANGED, "write unchanged" },
2633         { BLK_PERM_RESIZE,          "resize" },
2634         { 0, NULL }
2635     };
2636 
2637     GString *result = g_string_sized_new(30);
2638     struct perm_name *p;
2639 
2640     for (p = permissions; p->name; p++) {
2641         if (perm & p->perm) {
2642             if (result->len > 0) {
2643                 g_string_append(result, ", ");
2644             }
2645             g_string_append(result, p->name);
2646         }
2647     }
2648 
2649     return g_string_free(result, FALSE);
2650 }
2651 
2652 
2653 /*
2654  * @tran is allowed to be NULL. In this case no rollback is possible.
2655  *
2656  * After calling this function, the transaction @tran may only be completed
2657  * while holding a reader lock for the graph.
2658  */
2659 static int GRAPH_RDLOCK
2660 bdrv_refresh_perms(BlockDriverState *bs, Transaction *tran, Error **errp)
2661 {
2662     int ret;
2663     Transaction *local_tran = NULL;
2664     g_autoptr(GSList) list = bdrv_topological_dfs(NULL, NULL, bs);
2665     GLOBAL_STATE_CODE();
2666 
2667     if (!tran) {
2668         tran = local_tran = tran_new();
2669     }
2670 
2671     ret = bdrv_do_refresh_perms(list, NULL, tran, errp);
2672 
2673     if (local_tran) {
2674         tran_finalize(local_tran, ret);
2675     }
2676 
2677     return ret;
2678 }
2679 
2680 int bdrv_child_try_set_perm(BdrvChild *c, uint64_t perm, uint64_t shared,
2681                             Error **errp)
2682 {
2683     Error *local_err = NULL;
2684     Transaction *tran = tran_new();
2685     int ret;
2686 
2687     GLOBAL_STATE_CODE();
2688 
2689     bdrv_child_set_perm(c, perm, shared, tran);
2690 
2691     ret = bdrv_refresh_perms(c->bs, tran, &local_err);
2692 
2693     tran_finalize(tran, ret);
2694 
2695     if (ret < 0) {
2696         if ((perm & ~c->perm) || (c->shared_perm & ~shared)) {
2697             /* tighten permissions */
2698             error_propagate(errp, local_err);
2699         } else {
2700             /*
2701              * Our caller may intend to only loosen restrictions and
2702              * does not expect this function to fail.  Errors are not
2703              * fatal in such a case, so we can just hide them from our
2704              * caller.
2705              */
2706             error_free(local_err);
2707             ret = 0;
2708         }
2709     }
2710 
2711     return ret;
2712 }
2713 
2714 int bdrv_child_refresh_perms(BlockDriverState *bs, BdrvChild *c, Error **errp)
2715 {
2716     uint64_t parent_perms, parent_shared;
2717     uint64_t perms, shared;
2718 
2719     GLOBAL_STATE_CODE();
2720 
2721     bdrv_get_cumulative_perm(bs, &parent_perms, &parent_shared);
2722     bdrv_child_perm(bs, c->bs, c, c->role, NULL,
2723                     parent_perms, parent_shared, &perms, &shared);
2724 
2725     return bdrv_child_try_set_perm(c, perms, shared, errp);
2726 }
2727 
2728 /*
2729  * Default implementation for .bdrv_child_perm() for block filters:
2730  * Forward CONSISTENT_READ, WRITE, WRITE_UNCHANGED, and RESIZE to the
2731  * filtered child.
2732  */
2733 static void bdrv_filter_default_perms(BlockDriverState *bs, BdrvChild *c,
2734                                       BdrvChildRole role,
2735                                       BlockReopenQueue *reopen_queue,
2736                                       uint64_t perm, uint64_t shared,
2737                                       uint64_t *nperm, uint64_t *nshared)
2738 {
2739     GLOBAL_STATE_CODE();
2740     *nperm = perm & DEFAULT_PERM_PASSTHROUGH;
2741     *nshared = (shared & DEFAULT_PERM_PASSTHROUGH) | DEFAULT_PERM_UNCHANGED;
2742 }
2743 
2744 static void bdrv_default_perms_for_cow(BlockDriverState *bs, BdrvChild *c,
2745                                        BdrvChildRole role,
2746                                        BlockReopenQueue *reopen_queue,
2747                                        uint64_t perm, uint64_t shared,
2748                                        uint64_t *nperm, uint64_t *nshared)
2749 {
2750     assert(role & BDRV_CHILD_COW);
2751     GLOBAL_STATE_CODE();
2752 
2753     /*
2754      * We want consistent read from backing files if the parent needs it.
2755      * No other operations are performed on backing files.
2756      */
2757     perm &= BLK_PERM_CONSISTENT_READ;
2758 
2759     /*
2760      * If the parent can deal with changing data, we're okay with a
2761      * writable and resizable backing file.
2762      * TODO Require !(perm & BLK_PERM_CONSISTENT_READ), too?
2763      */
2764     if (shared & BLK_PERM_WRITE) {
2765         shared = BLK_PERM_WRITE | BLK_PERM_RESIZE;
2766     } else {
2767         shared = 0;
2768     }
2769 
2770     shared |= BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED;
2771 
2772     if (bs->open_flags & BDRV_O_INACTIVE) {
2773         shared |= BLK_PERM_WRITE | BLK_PERM_RESIZE;
2774     }
2775 
2776     *nperm = perm;
2777     *nshared = shared;
2778 }
2779 
2780 static void bdrv_default_perms_for_storage(BlockDriverState *bs, BdrvChild *c,
2781                                            BdrvChildRole role,
2782                                            BlockReopenQueue *reopen_queue,
2783                                            uint64_t perm, uint64_t shared,
2784                                            uint64_t *nperm, uint64_t *nshared)
2785 {
2786     int flags;
2787 
2788     GLOBAL_STATE_CODE();
2789     assert(role & (BDRV_CHILD_METADATA | BDRV_CHILD_DATA));
2790 
2791     flags = bdrv_reopen_get_flags(reopen_queue, bs);
2792 
2793     /*
2794      * Apart from the modifications below, the same permissions are
2795      * forwarded and left alone as for filters
2796      */
2797     bdrv_filter_default_perms(bs, c, role, reopen_queue,
2798                               perm, shared, &perm, &shared);
2799 
2800     if (role & BDRV_CHILD_METADATA) {
2801         /* Format drivers may touch metadata even if the guest doesn't write */
2802         if (bdrv_is_writable_after_reopen(bs, reopen_queue)) {
2803             perm |= BLK_PERM_WRITE | BLK_PERM_RESIZE;
2804         }
2805 
2806         /*
2807          * bs->file always needs to be consistent because of the
2808          * metadata. We can never allow other users to resize or write
2809          * to it.
2810          */
2811         if (!(flags & BDRV_O_NO_IO)) {
2812             perm |= BLK_PERM_CONSISTENT_READ;
2813         }
2814         shared &= ~(BLK_PERM_WRITE | BLK_PERM_RESIZE);
2815     }
2816 
2817     if (role & BDRV_CHILD_DATA) {
2818         /*
2819          * Technically, everything in this block is a subset of the
2820          * BDRV_CHILD_METADATA path taken above, and so this could
2821          * be an "else if" branch.  However, that is not obvious, and
2822          * this function is not performance critical, therefore we let
2823          * this be an independent "if".
2824          */
2825 
2826         /*
2827          * We cannot allow other users to resize the file because the
2828          * format driver might have some assumptions about the size
2829          * (e.g. because it is stored in metadata, or because the file
2830          * is split into fixed-size data files).
2831          */
2832         shared &= ~BLK_PERM_RESIZE;
2833 
2834         /*
2835          * WRITE_UNCHANGED often cannot be performed as such on the
2836          * data file.  For example, the qcow2 driver may still need to
2837          * write copied clusters on copy-on-read.
2838          */
2839         if (perm & BLK_PERM_WRITE_UNCHANGED) {
2840             perm |= BLK_PERM_WRITE;
2841         }
2842 
2843         /*
2844          * If the data file is written to, the format driver may
2845          * expect to be able to resize it by writing beyond the EOF.
2846          */
2847         if (perm & BLK_PERM_WRITE) {
2848             perm |= BLK_PERM_RESIZE;
2849         }
2850     }
2851 
2852     if (bs->open_flags & BDRV_O_INACTIVE) {
2853         shared |= BLK_PERM_WRITE | BLK_PERM_RESIZE;
2854     }
2855 
2856     *nperm = perm;
2857     *nshared = shared;
2858 }
2859 
2860 void bdrv_default_perms(BlockDriverState *bs, BdrvChild *c,
2861                         BdrvChildRole role, BlockReopenQueue *reopen_queue,
2862                         uint64_t perm, uint64_t shared,
2863                         uint64_t *nperm, uint64_t *nshared)
2864 {
2865     GLOBAL_STATE_CODE();
2866     if (role & BDRV_CHILD_FILTERED) {
2867         assert(!(role & (BDRV_CHILD_DATA | BDRV_CHILD_METADATA |
2868                          BDRV_CHILD_COW)));
2869         bdrv_filter_default_perms(bs, c, role, reopen_queue,
2870                                   perm, shared, nperm, nshared);
2871     } else if (role & BDRV_CHILD_COW) {
2872         assert(!(role & (BDRV_CHILD_DATA | BDRV_CHILD_METADATA)));
2873         bdrv_default_perms_for_cow(bs, c, role, reopen_queue,
2874                                    perm, shared, nperm, nshared);
2875     } else if (role & (BDRV_CHILD_METADATA | BDRV_CHILD_DATA)) {
2876         bdrv_default_perms_for_storage(bs, c, role, reopen_queue,
2877                                        perm, shared, nperm, nshared);
2878     } else {
2879         g_assert_not_reached();
2880     }
2881 }
2882 
2883 uint64_t bdrv_qapi_perm_to_blk_perm(BlockPermission qapi_perm)
2884 {
2885     static const uint64_t permissions[] = {
2886         [BLOCK_PERMISSION_CONSISTENT_READ]  = BLK_PERM_CONSISTENT_READ,
2887         [BLOCK_PERMISSION_WRITE]            = BLK_PERM_WRITE,
2888         [BLOCK_PERMISSION_WRITE_UNCHANGED]  = BLK_PERM_WRITE_UNCHANGED,
2889         [BLOCK_PERMISSION_RESIZE]           = BLK_PERM_RESIZE,
2890     };
2891 
2892     QEMU_BUILD_BUG_ON(ARRAY_SIZE(permissions) != BLOCK_PERMISSION__MAX);
2893     QEMU_BUILD_BUG_ON(1UL << ARRAY_SIZE(permissions) != BLK_PERM_ALL + 1);
2894 
2895     assert(qapi_perm < BLOCK_PERMISSION__MAX);
2896 
2897     return permissions[qapi_perm];
2898 }
2899 
2900 /*
2901  * Replaces the node that a BdrvChild points to without updating permissions.
2902  *
2903  * If @new_bs is non-NULL, the parent of @child must already be drained through
2904  * @child and the caller must hold the AioContext lock for @new_bs.
2905  */
2906 static void GRAPH_WRLOCK
2907 bdrv_replace_child_noperm(BdrvChild *child, BlockDriverState *new_bs)
2908 {
2909     BlockDriverState *old_bs = child->bs;
2910     int new_bs_quiesce_counter;
2911 
2912     assert(!child->frozen);
2913 
2914     /*
2915      * If we want to change the BdrvChild to point to a drained node as its new
2916      * child->bs, we need to make sure that its new parent is drained, too. In
2917      * other words, either child->quiesce_parent must already be true or we must
2918      * be able to set it and keep the parent's quiesce_counter consistent with
2919      * that, but without polling or starting new requests (this function
2920      * guarantees that it doesn't poll, and starting new requests would be
2921      * against the invariants of drain sections).
2922      *
2923      * To keep things simple, we pick the first option (child->quiesce_parent
2924      * must already be true). We also generalise the rule a bit to make it
2925      * easier to verify in callers and more likely to be covered in test cases:
2926      * The parent must be quiesced through this child even if new_bs isn't
2927      * currently drained.
2928      *
2929      * The only exception is for callers that always pass new_bs == NULL. In
2930      * this case, we obviously never need to consider the case of a drained
2931      * new_bs, so we can keep the callers simpler by allowing them not to drain
2932      * the parent.
2933      */
2934     assert(!new_bs || child->quiesced_parent);
2935     assert(old_bs != new_bs);
2936     GLOBAL_STATE_CODE();
2937 
2938     if (old_bs && new_bs) {
2939         assert(bdrv_get_aio_context(old_bs) == bdrv_get_aio_context(new_bs));
2940     }
2941 
2942     if (old_bs) {
2943         if (child->klass->detach) {
2944             child->klass->detach(child);
2945         }
2946         QLIST_REMOVE(child, next_parent);
2947     }
2948 
2949     child->bs = new_bs;
2950 
2951     if (new_bs) {
2952         QLIST_INSERT_HEAD(&new_bs->parents, child, next_parent);
2953         if (child->klass->attach) {
2954             child->klass->attach(child);
2955         }
2956     }
2957 
2958     /*
2959      * If the parent was drained through this BdrvChild previously, but new_bs
2960      * is not drained, allow requests to come in only after the new node has
2961      * been attached.
2962      */
2963     new_bs_quiesce_counter = (new_bs ? new_bs->quiesce_counter : 0);
2964     if (!new_bs_quiesce_counter && child->quiesced_parent) {
2965         bdrv_parent_drained_end_single(child);
2966     }
2967 }
2968 
2969 /**
2970  * Free the given @child.
2971  *
2972  * The child must be empty (i.e. `child->bs == NULL`) and it must be
2973  * unused (i.e. not in a children list).
2974  */
2975 static void bdrv_child_free(BdrvChild *child)
2976 {
2977     assert(!child->bs);
2978     GLOBAL_STATE_CODE();
2979     GRAPH_RDLOCK_GUARD_MAINLOOP();
2980 
2981     assert(!child->next.le_prev); /* not in children list */
2982 
2983     g_free(child->name);
2984     g_free(child);
2985 }
2986 
2987 typedef struct BdrvAttachChildCommonState {
2988     BdrvChild *child;
2989     AioContext *old_parent_ctx;
2990     AioContext *old_child_ctx;
2991 } BdrvAttachChildCommonState;
2992 
2993 static void GRAPH_WRLOCK bdrv_attach_child_common_abort(void *opaque)
2994 {
2995     BdrvAttachChildCommonState *s = opaque;
2996     BlockDriverState *bs = s->child->bs;
2997 
2998     GLOBAL_STATE_CODE();
2999     assert_bdrv_graph_writable();
3000 
3001     bdrv_replace_child_noperm(s->child, NULL);
3002 
3003     if (bdrv_get_aio_context(bs) != s->old_child_ctx) {
3004         bdrv_try_change_aio_context(bs, s->old_child_ctx, NULL, &error_abort);
3005     }
3006 
3007     if (bdrv_child_get_parent_aio_context(s->child) != s->old_parent_ctx) {
3008         Transaction *tran;
3009         GHashTable *visited;
3010         bool ret;
3011 
3012         tran = tran_new();
3013 
3014         /* No need to visit `child`, because it has been detached already */
3015         visited = g_hash_table_new(NULL, NULL);
3016         ret = s->child->klass->change_aio_ctx(s->child, s->old_parent_ctx,
3017                                               visited, tran, &error_abort);
3018         g_hash_table_destroy(visited);
3019 
3020         /* transaction is supposed to always succeed */
3021         assert(ret == true);
3022         tran_commit(tran);
3023     }
3024 
3025     bdrv_schedule_unref(bs);
3026     bdrv_child_free(s->child);
3027 }
3028 
3029 static TransactionActionDrv bdrv_attach_child_common_drv = {
3030     .abort = bdrv_attach_child_common_abort,
3031     .clean = g_free,
3032 };
3033 
3034 /*
3035  * Common part of attaching bdrv child to bs or to blk or to job
3036  *
3037  * Function doesn't update permissions, caller is responsible for this.
3038  *
3039  * After calling this function, the transaction @tran may only be completed
3040  * while holding a writer lock for the graph.
3041  *
3042  * Returns new created child.
3043  *
3044  * The caller must hold the AioContext lock for @child_bs. Both @parent_bs and
3045  * @child_bs can move to a different AioContext in this function. Callers must
3046  * make sure that their AioContext locking is still correct after this.
3047  */
3048 static BdrvChild * GRAPH_WRLOCK
3049 bdrv_attach_child_common(BlockDriverState *child_bs,
3050                          const char *child_name,
3051                          const BdrvChildClass *child_class,
3052                          BdrvChildRole child_role,
3053                          uint64_t perm, uint64_t shared_perm,
3054                          void *opaque,
3055                          Transaction *tran, Error **errp)
3056 {
3057     BdrvChild *new_child;
3058     AioContext *parent_ctx;
3059     AioContext *child_ctx = bdrv_get_aio_context(child_bs);
3060 
3061     assert(child_class->get_parent_desc);
3062     GLOBAL_STATE_CODE();
3063 
3064     new_child = g_new(BdrvChild, 1);
3065     *new_child = (BdrvChild) {
3066         .bs             = NULL,
3067         .name           = g_strdup(child_name),
3068         .klass          = child_class,
3069         .role           = child_role,
3070         .perm           = perm,
3071         .shared_perm    = shared_perm,
3072         .opaque         = opaque,
3073     };
3074 
3075     /*
3076      * If the AioContexts don't match, first try to move the subtree of
3077      * child_bs into the AioContext of the new parent. If this doesn't work,
3078      * try moving the parent into the AioContext of child_bs instead.
3079      */
3080     parent_ctx = bdrv_child_get_parent_aio_context(new_child);
3081     if (child_ctx != parent_ctx) {
3082         Error *local_err = NULL;
3083         int ret = bdrv_try_change_aio_context(child_bs, parent_ctx, NULL,
3084                                               &local_err);
3085 
3086         if (ret < 0 && child_class->change_aio_ctx) {
3087             Transaction *aio_ctx_tran = tran_new();
3088             GHashTable *visited = g_hash_table_new(NULL, NULL);
3089             bool ret_child;
3090 
3091             g_hash_table_add(visited, new_child);
3092             ret_child = child_class->change_aio_ctx(new_child, child_ctx,
3093                                                     visited, aio_ctx_tran,
3094                                                     NULL);
3095             if (ret_child == true) {
3096                 error_free(local_err);
3097                 ret = 0;
3098             }
3099             tran_finalize(aio_ctx_tran, ret_child == true ? 0 : -1);
3100             g_hash_table_destroy(visited);
3101         }
3102 
3103         if (ret < 0) {
3104             error_propagate(errp, local_err);
3105             bdrv_child_free(new_child);
3106             return NULL;
3107         }
3108     }
3109 
3110     bdrv_ref(child_bs);
3111     /*
3112      * Let every new BdrvChild start with a drained parent. Inserting the child
3113      * in the graph with bdrv_replace_child_noperm() will undrain it if
3114      * @child_bs is not drained.
3115      *
3116      * The child was only just created and is not yet visible in global state
3117      * until bdrv_replace_child_noperm() inserts it into the graph, so nobody
3118      * could have sent requests and polling is not necessary.
3119      *
3120      * Note that this means that the parent isn't fully drained yet, we only
3121      * stop new requests from coming in. This is fine, we don't care about the
3122      * old requests here, they are not for this child. If another place enters a
3123      * drain section for the same parent, but wants it to be fully quiesced, it
3124      * will not run most of the the code in .drained_begin() again (which is not
3125      * a problem, we already did this), but it will still poll until the parent
3126      * is fully quiesced, so it will not be negatively affected either.
3127      */
3128     bdrv_parent_drained_begin_single(new_child);
3129     bdrv_replace_child_noperm(new_child, child_bs);
3130 
3131     BdrvAttachChildCommonState *s = g_new(BdrvAttachChildCommonState, 1);
3132     *s = (BdrvAttachChildCommonState) {
3133         .child = new_child,
3134         .old_parent_ctx = parent_ctx,
3135         .old_child_ctx = child_ctx,
3136     };
3137     tran_add(tran, &bdrv_attach_child_common_drv, s);
3138 
3139     return new_child;
3140 }
3141 
3142 /*
3143  * Function doesn't update permissions, caller is responsible for this.
3144  *
3145  * The caller must hold the AioContext lock for @child_bs. Both @parent_bs and
3146  * @child_bs can move to a different AioContext in this function. Callers must
3147  * make sure that their AioContext locking is still correct after this.
3148  *
3149  * After calling this function, the transaction @tran may only be completed
3150  * while holding a writer lock for the graph.
3151  */
3152 static BdrvChild * GRAPH_WRLOCK
3153 bdrv_attach_child_noperm(BlockDriverState *parent_bs,
3154                          BlockDriverState *child_bs,
3155                          const char *child_name,
3156                          const BdrvChildClass *child_class,
3157                          BdrvChildRole child_role,
3158                          Transaction *tran,
3159                          Error **errp)
3160 {
3161     uint64_t perm, shared_perm;
3162 
3163     assert(parent_bs->drv);
3164     GLOBAL_STATE_CODE();
3165 
3166     if (bdrv_recurse_has_child(child_bs, parent_bs)) {
3167         error_setg(errp, "Making '%s' a %s child of '%s' would create a cycle",
3168                    child_bs->node_name, child_name, parent_bs->node_name);
3169         return NULL;
3170     }
3171 
3172     bdrv_get_cumulative_perm(parent_bs, &perm, &shared_perm);
3173     bdrv_child_perm(parent_bs, child_bs, NULL, child_role, NULL,
3174                     perm, shared_perm, &perm, &shared_perm);
3175 
3176     return bdrv_attach_child_common(child_bs, child_name, child_class,
3177                                     child_role, perm, shared_perm, parent_bs,
3178                                     tran, errp);
3179 }
3180 
3181 /*
3182  * This function steals the reference to child_bs from the caller.
3183  * That reference is later dropped by bdrv_root_unref_child().
3184  *
3185  * On failure NULL is returned, errp is set and the reference to
3186  * child_bs is also dropped.
3187  *
3188  * The caller must hold the AioContext lock @child_bs, but not that of @ctx
3189  * (unless @child_bs is already in @ctx).
3190  */
3191 BdrvChild *bdrv_root_attach_child(BlockDriverState *child_bs,
3192                                   const char *child_name,
3193                                   const BdrvChildClass *child_class,
3194                                   BdrvChildRole child_role,
3195                                   uint64_t perm, uint64_t shared_perm,
3196                                   void *opaque, Error **errp)
3197 {
3198     int ret;
3199     BdrvChild *child;
3200     Transaction *tran = tran_new();
3201 
3202     GLOBAL_STATE_CODE();
3203 
3204     child = bdrv_attach_child_common(child_bs, child_name, child_class,
3205                                    child_role, perm, shared_perm, opaque,
3206                                    tran, errp);
3207     if (!child) {
3208         ret = -EINVAL;
3209         goto out;
3210     }
3211 
3212     ret = bdrv_refresh_perms(child_bs, tran, errp);
3213 
3214 out:
3215     tran_finalize(tran, ret);
3216 
3217     bdrv_schedule_unref(child_bs);
3218 
3219     return ret < 0 ? NULL : child;
3220 }
3221 
3222 /*
3223  * This function transfers the reference to child_bs from the caller
3224  * to parent_bs. That reference is later dropped by parent_bs on
3225  * bdrv_close() or if someone calls bdrv_unref_child().
3226  *
3227  * On failure NULL is returned, errp is set and the reference to
3228  * child_bs is also dropped.
3229  *
3230  * If @parent_bs and @child_bs are in different AioContexts, the caller must
3231  * hold the AioContext lock for @child_bs, but not for @parent_bs.
3232  */
3233 BdrvChild *bdrv_attach_child(BlockDriverState *parent_bs,
3234                              BlockDriverState *child_bs,
3235                              const char *child_name,
3236                              const BdrvChildClass *child_class,
3237                              BdrvChildRole child_role,
3238                              Error **errp)
3239 {
3240     int ret;
3241     BdrvChild *child;
3242     Transaction *tran = tran_new();
3243 
3244     GLOBAL_STATE_CODE();
3245 
3246     child = bdrv_attach_child_noperm(parent_bs, child_bs, child_name,
3247                                      child_class, child_role, tran, errp);
3248     if (!child) {
3249         ret = -EINVAL;
3250         goto out;
3251     }
3252 
3253     ret = bdrv_refresh_perms(parent_bs, tran, errp);
3254     if (ret < 0) {
3255         goto out;
3256     }
3257 
3258 out:
3259     tran_finalize(tran, ret);
3260 
3261     bdrv_schedule_unref(child_bs);
3262 
3263     return ret < 0 ? NULL : child;
3264 }
3265 
3266 /* Callers must ensure that child->frozen is false. */
3267 void bdrv_root_unref_child(BdrvChild *child)
3268 {
3269     BlockDriverState *child_bs = child->bs;
3270 
3271     GLOBAL_STATE_CODE();
3272     bdrv_replace_child_noperm(child, NULL);
3273     bdrv_child_free(child);
3274 
3275     if (child_bs) {
3276         /*
3277          * Update permissions for old node. We're just taking a parent away, so
3278          * we're loosening restrictions. Errors of permission update are not
3279          * fatal in this case, ignore them.
3280          */
3281         bdrv_refresh_perms(child_bs, NULL, NULL);
3282 
3283         /*
3284          * When the parent requiring a non-default AioContext is removed, the
3285          * node moves back to the main AioContext
3286          */
3287         bdrv_try_change_aio_context(child_bs, qemu_get_aio_context(), NULL,
3288                                     NULL);
3289     }
3290 
3291     bdrv_schedule_unref(child_bs);
3292 }
3293 
3294 typedef struct BdrvSetInheritsFrom {
3295     BlockDriverState *bs;
3296     BlockDriverState *old_inherits_from;
3297 } BdrvSetInheritsFrom;
3298 
3299 static void bdrv_set_inherits_from_abort(void *opaque)
3300 {
3301     BdrvSetInheritsFrom *s = opaque;
3302 
3303     s->bs->inherits_from = s->old_inherits_from;
3304 }
3305 
3306 static TransactionActionDrv bdrv_set_inherits_from_drv = {
3307     .abort = bdrv_set_inherits_from_abort,
3308     .clean = g_free,
3309 };
3310 
3311 /* @tran is allowed to be NULL. In this case no rollback is possible */
3312 static void bdrv_set_inherits_from(BlockDriverState *bs,
3313                                    BlockDriverState *new_inherits_from,
3314                                    Transaction *tran)
3315 {
3316     if (tran) {
3317         BdrvSetInheritsFrom *s = g_new(BdrvSetInheritsFrom, 1);
3318 
3319         *s = (BdrvSetInheritsFrom) {
3320             .bs = bs,
3321             .old_inherits_from = bs->inherits_from,
3322         };
3323 
3324         tran_add(tran, &bdrv_set_inherits_from_drv, s);
3325     }
3326 
3327     bs->inherits_from = new_inherits_from;
3328 }
3329 
3330 /**
3331  * Clear all inherits_from pointers from children and grandchildren of
3332  * @root that point to @root, where necessary.
3333  * @tran is allowed to be NULL. In this case no rollback is possible
3334  */
3335 static void GRAPH_WRLOCK
3336 bdrv_unset_inherits_from(BlockDriverState *root, BdrvChild *child,
3337                          Transaction *tran)
3338 {
3339     BdrvChild *c;
3340 
3341     if (child->bs->inherits_from == root) {
3342         /*
3343          * Remove inherits_from only when the last reference between root and
3344          * child->bs goes away.
3345          */
3346         QLIST_FOREACH(c, &root->children, next) {
3347             if (c != child && c->bs == child->bs) {
3348                 break;
3349             }
3350         }
3351         if (c == NULL) {
3352             bdrv_set_inherits_from(child->bs, NULL, tran);
3353         }
3354     }
3355 
3356     QLIST_FOREACH(c, &child->bs->children, next) {
3357         bdrv_unset_inherits_from(root, c, tran);
3358     }
3359 }
3360 
3361 /* Callers must ensure that child->frozen is false. */
3362 void bdrv_unref_child(BlockDriverState *parent, BdrvChild *child)
3363 {
3364     GLOBAL_STATE_CODE();
3365     if (child == NULL) {
3366         return;
3367     }
3368 
3369     bdrv_unset_inherits_from(parent, child, NULL);
3370     bdrv_root_unref_child(child);
3371 }
3372 
3373 
3374 static void GRAPH_RDLOCK
3375 bdrv_parent_cb_change_media(BlockDriverState *bs, bool load)
3376 {
3377     BdrvChild *c;
3378     GLOBAL_STATE_CODE();
3379     QLIST_FOREACH(c, &bs->parents, next_parent) {
3380         if (c->klass->change_media) {
3381             c->klass->change_media(c, load);
3382         }
3383     }
3384 }
3385 
3386 /* Return true if you can reach parent going through child->inherits_from
3387  * recursively. If parent or child are NULL, return false */
3388 static bool bdrv_inherits_from_recursive(BlockDriverState *child,
3389                                          BlockDriverState *parent)
3390 {
3391     while (child && child != parent) {
3392         child = child->inherits_from;
3393     }
3394 
3395     return child != NULL;
3396 }
3397 
3398 /*
3399  * Return the BdrvChildRole for @bs's backing child.  bs->backing is
3400  * mostly used for COW backing children (role = COW), but also for
3401  * filtered children (role = FILTERED | PRIMARY).
3402  */
3403 static BdrvChildRole bdrv_backing_role(BlockDriverState *bs)
3404 {
3405     if (bs->drv && bs->drv->is_filter) {
3406         return BDRV_CHILD_FILTERED | BDRV_CHILD_PRIMARY;
3407     } else {
3408         return BDRV_CHILD_COW;
3409     }
3410 }
3411 
3412 /*
3413  * Sets the bs->backing or bs->file link of a BDS. A new reference is created;
3414  * callers which don't need their own reference any more must call bdrv_unref().
3415  *
3416  * If the respective child is already present (i.e. we're detaching a node),
3417  * that child node must be drained.
3418  *
3419  * Function doesn't update permissions, caller is responsible for this.
3420  *
3421  * The caller must hold the AioContext lock for @child_bs. Both @parent_bs and
3422  * @child_bs can move to a different AioContext in this function. Callers must
3423  * make sure that their AioContext locking is still correct after this.
3424  *
3425  * After calling this function, the transaction @tran may only be completed
3426  * while holding a writer lock for the graph.
3427  */
3428 static int GRAPH_WRLOCK
3429 bdrv_set_file_or_backing_noperm(BlockDriverState *parent_bs,
3430                                 BlockDriverState *child_bs,
3431                                 bool is_backing,
3432                                 Transaction *tran, Error **errp)
3433 {
3434     bool update_inherits_from =
3435         bdrv_inherits_from_recursive(child_bs, parent_bs);
3436     BdrvChild *child = is_backing ? parent_bs->backing : parent_bs->file;
3437     BdrvChildRole role;
3438 
3439     GLOBAL_STATE_CODE();
3440 
3441     if (!parent_bs->drv) {
3442         /*
3443          * Node without drv is an object without a class :/. TODO: finally fix
3444          * qcow2 driver to never clear bs->drv and implement format corruption
3445          * handling in other way.
3446          */
3447         error_setg(errp, "Node corrupted");
3448         return -EINVAL;
3449     }
3450 
3451     if (child && child->frozen) {
3452         error_setg(errp, "Cannot change frozen '%s' link from '%s' to '%s'",
3453                    child->name, parent_bs->node_name, child->bs->node_name);
3454         return -EPERM;
3455     }
3456 
3457     if (is_backing && !parent_bs->drv->is_filter &&
3458         !parent_bs->drv->supports_backing)
3459     {
3460         error_setg(errp, "Driver '%s' of node '%s' does not support backing "
3461                    "files", parent_bs->drv->format_name, parent_bs->node_name);
3462         return -EINVAL;
3463     }
3464 
3465     if (parent_bs->drv->is_filter) {
3466         role = BDRV_CHILD_FILTERED | BDRV_CHILD_PRIMARY;
3467     } else if (is_backing) {
3468         role = BDRV_CHILD_COW;
3469     } else {
3470         /*
3471          * We only can use same role as it is in existing child. We don't have
3472          * infrastructure to determine role of file child in generic way
3473          */
3474         if (!child) {
3475             error_setg(errp, "Cannot set file child to format node without "
3476                        "file child");
3477             return -EINVAL;
3478         }
3479         role = child->role;
3480     }
3481 
3482     if (child) {
3483         assert(child->bs->quiesce_counter);
3484         bdrv_unset_inherits_from(parent_bs, child, tran);
3485         bdrv_remove_child(child, tran);
3486     }
3487 
3488     if (!child_bs) {
3489         goto out;
3490     }
3491 
3492     child = bdrv_attach_child_noperm(parent_bs, child_bs,
3493                                      is_backing ? "backing" : "file",
3494                                      &child_of_bds, role,
3495                                      tran, errp);
3496     if (!child) {
3497         return -EINVAL;
3498     }
3499 
3500 
3501     /*
3502      * If inherits_from pointed recursively to bs then let's update it to
3503      * point directly to bs (else it will become NULL).
3504      */
3505     if (update_inherits_from) {
3506         bdrv_set_inherits_from(child_bs, parent_bs, tran);
3507     }
3508 
3509 out:
3510     bdrv_refresh_limits(parent_bs, tran, NULL);
3511 
3512     return 0;
3513 }
3514 
3515 /*
3516  * The caller must hold the AioContext lock for @backing_hd. Both @bs and
3517  * @backing_hd can move to a different AioContext in this function. Callers must
3518  * make sure that their AioContext locking is still correct after this.
3519  *
3520  * If a backing child is already present (i.e. we're detaching a node), that
3521  * child node must be drained.
3522  */
3523 int bdrv_set_backing_hd_drained(BlockDriverState *bs,
3524                                 BlockDriverState *backing_hd,
3525                                 Error **errp)
3526 {
3527     int ret;
3528     Transaction *tran = tran_new();
3529 
3530     GLOBAL_STATE_CODE();
3531     assert(bs->quiesce_counter > 0);
3532     if (bs->backing) {
3533         assert(bs->backing->bs->quiesce_counter > 0);
3534     }
3535 
3536     ret = bdrv_set_file_or_backing_noperm(bs, backing_hd, true, tran, errp);
3537     if (ret < 0) {
3538         goto out;
3539     }
3540 
3541     ret = bdrv_refresh_perms(bs, tran, errp);
3542 out:
3543     tran_finalize(tran, ret);
3544     return ret;
3545 }
3546 
3547 int bdrv_set_backing_hd(BlockDriverState *bs, BlockDriverState *backing_hd,
3548                         Error **errp)
3549 {
3550     BlockDriverState *drain_bs;
3551     int ret;
3552     GLOBAL_STATE_CODE();
3553 
3554     bdrv_graph_rdlock_main_loop();
3555     drain_bs = bs->backing ? bs->backing->bs : bs;
3556     bdrv_graph_rdunlock_main_loop();
3557 
3558     bdrv_ref(drain_bs);
3559     bdrv_drained_begin(drain_bs);
3560     bdrv_graph_wrlock();
3561     ret = bdrv_set_backing_hd_drained(bs, backing_hd, errp);
3562     bdrv_graph_wrunlock();
3563     bdrv_drained_end(drain_bs);
3564     bdrv_unref(drain_bs);
3565 
3566     return ret;
3567 }
3568 
3569 /*
3570  * Opens the backing file for a BlockDriverState if not yet open
3571  *
3572  * bdref_key specifies the key for the image's BlockdevRef in the options QDict.
3573  * That QDict has to be flattened; therefore, if the BlockdevRef is a QDict
3574  * itself, all options starting with "${bdref_key}." are considered part of the
3575  * BlockdevRef.
3576  *
3577  * The caller must hold the main AioContext lock.
3578  *
3579  * TODO Can this be unified with bdrv_open_image()?
3580  */
3581 int bdrv_open_backing_file(BlockDriverState *bs, QDict *parent_options,
3582                            const char *bdref_key, Error **errp)
3583 {
3584     char *backing_filename = NULL;
3585     char *bdref_key_dot;
3586     const char *reference = NULL;
3587     int ret = 0;
3588     bool implicit_backing = false;
3589     BlockDriverState *backing_hd;
3590     QDict *options;
3591     QDict *tmp_parent_options = NULL;
3592     Error *local_err = NULL;
3593 
3594     GLOBAL_STATE_CODE();
3595     GRAPH_RDLOCK_GUARD_MAINLOOP();
3596 
3597     if (bs->backing != NULL) {
3598         goto free_exit;
3599     }
3600 
3601     /* NULL means an empty set of options */
3602     if (parent_options == NULL) {
3603         tmp_parent_options = qdict_new();
3604         parent_options = tmp_parent_options;
3605     }
3606 
3607     bs->open_flags &= ~BDRV_O_NO_BACKING;
3608 
3609     bdref_key_dot = g_strdup_printf("%s.", bdref_key);
3610     qdict_extract_subqdict(parent_options, &options, bdref_key_dot);
3611     g_free(bdref_key_dot);
3612 
3613     /*
3614      * Caution: while qdict_get_try_str() is fine, getting non-string
3615      * types would require more care.  When @parent_options come from
3616      * -blockdev or blockdev_add, its members are typed according to
3617      * the QAPI schema, but when they come from -drive, they're all
3618      * QString.
3619      */
3620     reference = qdict_get_try_str(parent_options, bdref_key);
3621     if (reference || qdict_haskey(options, "file.filename")) {
3622         /* keep backing_filename NULL */
3623     } else if (bs->backing_file[0] == '\0' && qdict_size(options) == 0) {
3624         qobject_unref(options);
3625         goto free_exit;
3626     } else {
3627         if (qdict_size(options) == 0) {
3628             /* If the user specifies options that do not modify the
3629              * backing file's behavior, we might still consider it the
3630              * implicit backing file.  But it's easier this way, and
3631              * just specifying some of the backing BDS's options is
3632              * only possible with -drive anyway (otherwise the QAPI
3633              * schema forces the user to specify everything). */
3634             implicit_backing = !strcmp(bs->auto_backing_file, bs->backing_file);
3635         }
3636 
3637         backing_filename = bdrv_get_full_backing_filename(bs, &local_err);
3638         if (local_err) {
3639             ret = -EINVAL;
3640             error_propagate(errp, local_err);
3641             qobject_unref(options);
3642             goto free_exit;
3643         }
3644     }
3645 
3646     if (!bs->drv || !bs->drv->supports_backing) {
3647         ret = -EINVAL;
3648         error_setg(errp, "Driver doesn't support backing files");
3649         qobject_unref(options);
3650         goto free_exit;
3651     }
3652 
3653     if (!reference &&
3654         bs->backing_format[0] != '\0' && !qdict_haskey(options, "driver")) {
3655         qdict_put_str(options, "driver", bs->backing_format);
3656     }
3657 
3658     backing_hd = bdrv_open_inherit(backing_filename, reference, options, 0, bs,
3659                                    &child_of_bds, bdrv_backing_role(bs), errp);
3660     if (!backing_hd) {
3661         bs->open_flags |= BDRV_O_NO_BACKING;
3662         error_prepend(errp, "Could not open backing file: ");
3663         ret = -EINVAL;
3664         goto free_exit;
3665     }
3666 
3667     if (implicit_backing) {
3668         bdrv_refresh_filename(backing_hd);
3669         pstrcpy(bs->auto_backing_file, sizeof(bs->auto_backing_file),
3670                 backing_hd->filename);
3671     }
3672 
3673     /* Hook up the backing file link; drop our reference, bs owns the
3674      * backing_hd reference now */
3675     ret = bdrv_set_backing_hd(bs, backing_hd, errp);
3676     bdrv_unref(backing_hd);
3677 
3678     if (ret < 0) {
3679         goto free_exit;
3680     }
3681 
3682     qdict_del(parent_options, bdref_key);
3683 
3684 free_exit:
3685     g_free(backing_filename);
3686     qobject_unref(tmp_parent_options);
3687     return ret;
3688 }
3689 
3690 static BlockDriverState *
3691 bdrv_open_child_bs(const char *filename, QDict *options, const char *bdref_key,
3692                    BlockDriverState *parent, const BdrvChildClass *child_class,
3693                    BdrvChildRole child_role, bool allow_none, Error **errp)
3694 {
3695     BlockDriverState *bs = NULL;
3696     QDict *image_options;
3697     char *bdref_key_dot;
3698     const char *reference;
3699 
3700     assert(child_class != NULL);
3701 
3702     bdref_key_dot = g_strdup_printf("%s.", bdref_key);
3703     qdict_extract_subqdict(options, &image_options, bdref_key_dot);
3704     g_free(bdref_key_dot);
3705 
3706     /*
3707      * Caution: while qdict_get_try_str() is fine, getting non-string
3708      * types would require more care.  When @options come from
3709      * -blockdev or blockdev_add, its members are typed according to
3710      * the QAPI schema, but when they come from -drive, they're all
3711      * QString.
3712      */
3713     reference = qdict_get_try_str(options, bdref_key);
3714     if (!filename && !reference && !qdict_size(image_options)) {
3715         if (!allow_none) {
3716             error_setg(errp, "A block device must be specified for \"%s\"",
3717                        bdref_key);
3718         }
3719         qobject_unref(image_options);
3720         goto done;
3721     }
3722 
3723     bs = bdrv_open_inherit(filename, reference, image_options, 0,
3724                            parent, child_class, child_role, errp);
3725     if (!bs) {
3726         goto done;
3727     }
3728 
3729 done:
3730     qdict_del(options, bdref_key);
3731     return bs;
3732 }
3733 
3734 /*
3735  * Opens a disk image whose options are given as BlockdevRef in another block
3736  * device's options.
3737  *
3738  * If allow_none is true, no image will be opened if filename is false and no
3739  * BlockdevRef is given. NULL will be returned, but errp remains unset.
3740  *
3741  * bdrev_key specifies the key for the image's BlockdevRef in the options QDict.
3742  * That QDict has to be flattened; therefore, if the BlockdevRef is a QDict
3743  * itself, all options starting with "${bdref_key}." are considered part of the
3744  * BlockdevRef.
3745  *
3746  * The BlockdevRef will be removed from the options QDict.
3747  *
3748  * The caller must hold the lock of the main AioContext and no other AioContext.
3749  * @parent can move to a different AioContext in this function. Callers must
3750  * make sure that their AioContext locking is still correct after this.
3751  */
3752 BdrvChild *bdrv_open_child(const char *filename,
3753                            QDict *options, const char *bdref_key,
3754                            BlockDriverState *parent,
3755                            const BdrvChildClass *child_class,
3756                            BdrvChildRole child_role,
3757                            bool allow_none, Error **errp)
3758 {
3759     BlockDriverState *bs;
3760     BdrvChild *child;
3761 
3762     GLOBAL_STATE_CODE();
3763 
3764     bs = bdrv_open_child_bs(filename, options, bdref_key, parent, child_class,
3765                             child_role, allow_none, errp);
3766     if (bs == NULL) {
3767         return NULL;
3768     }
3769 
3770     bdrv_graph_wrlock();
3771     child = bdrv_attach_child(parent, bs, bdref_key, child_class, child_role,
3772                               errp);
3773     bdrv_graph_wrunlock();
3774 
3775     return child;
3776 }
3777 
3778 /*
3779  * Wrapper on bdrv_open_child() for most popular case: open primary child of bs.
3780  *
3781  * The caller must hold the lock of the main AioContext and no other AioContext.
3782  * @parent can move to a different AioContext in this function. Callers must
3783  * make sure that their AioContext locking is still correct after this.
3784  */
3785 int bdrv_open_file_child(const char *filename,
3786                          QDict *options, const char *bdref_key,
3787                          BlockDriverState *parent, Error **errp)
3788 {
3789     BdrvChildRole role;
3790 
3791     /* commit_top and mirror_top don't use this function */
3792     assert(!parent->drv->filtered_child_is_backing);
3793     role = parent->drv->is_filter ?
3794         (BDRV_CHILD_FILTERED | BDRV_CHILD_PRIMARY) : BDRV_CHILD_IMAGE;
3795 
3796     if (!bdrv_open_child(filename, options, bdref_key, parent,
3797                          &child_of_bds, role, false, errp))
3798     {
3799         return -EINVAL;
3800     }
3801 
3802     return 0;
3803 }
3804 
3805 /*
3806  * TODO Future callers may need to specify parent/child_class in order for
3807  * option inheritance to work. Existing callers use it for the root node.
3808  */
3809 BlockDriverState *bdrv_open_blockdev_ref(BlockdevRef *ref, Error **errp)
3810 {
3811     BlockDriverState *bs = NULL;
3812     QObject *obj = NULL;
3813     QDict *qdict = NULL;
3814     const char *reference = NULL;
3815     Visitor *v = NULL;
3816 
3817     GLOBAL_STATE_CODE();
3818 
3819     if (ref->type == QTYPE_QSTRING) {
3820         reference = ref->u.reference;
3821     } else {
3822         BlockdevOptions *options = &ref->u.definition;
3823         assert(ref->type == QTYPE_QDICT);
3824 
3825         v = qobject_output_visitor_new(&obj);
3826         visit_type_BlockdevOptions(v, NULL, &options, &error_abort);
3827         visit_complete(v, &obj);
3828 
3829         qdict = qobject_to(QDict, obj);
3830         qdict_flatten(qdict);
3831 
3832         /* bdrv_open_inherit() defaults to the values in bdrv_flags (for
3833          * compatibility with other callers) rather than what we want as the
3834          * real defaults. Apply the defaults here instead. */
3835         qdict_set_default_str(qdict, BDRV_OPT_CACHE_DIRECT, "off");
3836         qdict_set_default_str(qdict, BDRV_OPT_CACHE_NO_FLUSH, "off");
3837         qdict_set_default_str(qdict, BDRV_OPT_READ_ONLY, "off");
3838         qdict_set_default_str(qdict, BDRV_OPT_AUTO_READ_ONLY, "off");
3839 
3840     }
3841 
3842     bs = bdrv_open_inherit(NULL, reference, qdict, 0, NULL, NULL, 0, errp);
3843     obj = NULL;
3844     qobject_unref(obj);
3845     visit_free(v);
3846     return bs;
3847 }
3848 
3849 static BlockDriverState *bdrv_append_temp_snapshot(BlockDriverState *bs,
3850                                                    int flags,
3851                                                    QDict *snapshot_options,
3852                                                    Error **errp)
3853 {
3854     g_autofree char *tmp_filename = NULL;
3855     int64_t total_size;
3856     QemuOpts *opts = NULL;
3857     BlockDriverState *bs_snapshot = NULL;
3858     int ret;
3859 
3860     GLOBAL_STATE_CODE();
3861 
3862     /* if snapshot, we create a temporary backing file and open it
3863        instead of opening 'filename' directly */
3864 
3865     /* Get the required size from the image */
3866     total_size = bdrv_getlength(bs);
3867 
3868     if (total_size < 0) {
3869         error_setg_errno(errp, -total_size, "Could not get image size");
3870         goto out;
3871     }
3872 
3873     /* Create the temporary image */
3874     tmp_filename = create_tmp_file(errp);
3875     if (!tmp_filename) {
3876         goto out;
3877     }
3878 
3879     opts = qemu_opts_create(bdrv_qcow2.create_opts, NULL, 0,
3880                             &error_abort);
3881     qemu_opt_set_number(opts, BLOCK_OPT_SIZE, total_size, &error_abort);
3882     ret = bdrv_create(&bdrv_qcow2, tmp_filename, opts, errp);
3883     qemu_opts_del(opts);
3884     if (ret < 0) {
3885         error_prepend(errp, "Could not create temporary overlay '%s': ",
3886                       tmp_filename);
3887         goto out;
3888     }
3889 
3890     /* Prepare options QDict for the temporary file */
3891     qdict_put_str(snapshot_options, "file.driver", "file");
3892     qdict_put_str(snapshot_options, "file.filename", tmp_filename);
3893     qdict_put_str(snapshot_options, "driver", "qcow2");
3894 
3895     bs_snapshot = bdrv_open(NULL, NULL, snapshot_options, flags, errp);
3896     snapshot_options = NULL;
3897     if (!bs_snapshot) {
3898         goto out;
3899     }
3900 
3901     ret = bdrv_append(bs_snapshot, bs, errp);
3902     if (ret < 0) {
3903         bs_snapshot = NULL;
3904         goto out;
3905     }
3906 
3907 out:
3908     qobject_unref(snapshot_options);
3909     return bs_snapshot;
3910 }
3911 
3912 /*
3913  * Opens a disk image (raw, qcow2, vmdk, ...)
3914  *
3915  * options is a QDict of options to pass to the block drivers, or NULL for an
3916  * empty set of options. The reference to the QDict belongs to the block layer
3917  * after the call (even on failure), so if the caller intends to reuse the
3918  * dictionary, it needs to use qobject_ref() before calling bdrv_open.
3919  *
3920  * If *pbs is NULL, a new BDS will be created with a pointer to it stored there.
3921  * If it is not NULL, the referenced BDS will be reused.
3922  *
3923  * The reference parameter may be used to specify an existing block device which
3924  * should be opened. If specified, neither options nor a filename may be given,
3925  * nor can an existing BDS be reused (that is, *pbs has to be NULL).
3926  *
3927  * The caller must always hold the main AioContext lock.
3928  */
3929 static BlockDriverState * no_coroutine_fn
3930 bdrv_open_inherit(const char *filename, const char *reference, QDict *options,
3931                   int flags, BlockDriverState *parent,
3932                   const BdrvChildClass *child_class, BdrvChildRole child_role,
3933                   Error **errp)
3934 {
3935     int ret;
3936     BlockBackend *file = NULL;
3937     BlockDriverState *bs;
3938     BlockDriver *drv = NULL;
3939     BdrvChild *child;
3940     const char *drvname;
3941     const char *backing;
3942     Error *local_err = NULL;
3943     QDict *snapshot_options = NULL;
3944     int snapshot_flags = 0;
3945 
3946     assert(!child_class || !flags);
3947     assert(!child_class == !parent);
3948     GLOBAL_STATE_CODE();
3949     assert(!qemu_in_coroutine());
3950 
3951     /* TODO We'll eventually have to take a writer lock in this function */
3952     GRAPH_RDLOCK_GUARD_MAINLOOP();
3953 
3954     if (reference) {
3955         bool options_non_empty = options ? qdict_size(options) : false;
3956         qobject_unref(options);
3957 
3958         if (filename || options_non_empty) {
3959             error_setg(errp, "Cannot reference an existing block device with "
3960                        "additional options or a new filename");
3961             return NULL;
3962         }
3963 
3964         bs = bdrv_lookup_bs(reference, reference, errp);
3965         if (!bs) {
3966             return NULL;
3967         }
3968 
3969         bdrv_ref(bs);
3970         return bs;
3971     }
3972 
3973     bs = bdrv_new();
3974 
3975     /* NULL means an empty set of options */
3976     if (options == NULL) {
3977         options = qdict_new();
3978     }
3979 
3980     /* json: syntax counts as explicit options, as if in the QDict */
3981     parse_json_protocol(options, &filename, &local_err);
3982     if (local_err) {
3983         goto fail;
3984     }
3985 
3986     bs->explicit_options = qdict_clone_shallow(options);
3987 
3988     if (child_class) {
3989         bool parent_is_format;
3990 
3991         if (parent->drv) {
3992             parent_is_format = parent->drv->is_format;
3993         } else {
3994             /*
3995              * parent->drv is not set yet because this node is opened for
3996              * (potential) format probing.  That means that @parent is going
3997              * to be a format node.
3998              */
3999             parent_is_format = true;
4000         }
4001 
4002         bs->inherits_from = parent;
4003         child_class->inherit_options(child_role, parent_is_format,
4004                                      &flags, options,
4005                                      parent->open_flags, parent->options);
4006     }
4007 
4008     ret = bdrv_fill_options(&options, filename, &flags, &local_err);
4009     if (ret < 0) {
4010         goto fail;
4011     }
4012 
4013     /*
4014      * Set the BDRV_O_RDWR and BDRV_O_ALLOW_RDWR flags.
4015      * Caution: getting a boolean member of @options requires care.
4016      * When @options come from -blockdev or blockdev_add, members are
4017      * typed according to the QAPI schema, but when they come from
4018      * -drive, they're all QString.
4019      */
4020     if (g_strcmp0(qdict_get_try_str(options, BDRV_OPT_READ_ONLY), "on") &&
4021         !qdict_get_try_bool(options, BDRV_OPT_READ_ONLY, false)) {
4022         flags |= (BDRV_O_RDWR | BDRV_O_ALLOW_RDWR);
4023     } else {
4024         flags &= ~BDRV_O_RDWR;
4025     }
4026 
4027     if (flags & BDRV_O_SNAPSHOT) {
4028         snapshot_options = qdict_new();
4029         bdrv_temp_snapshot_options(&snapshot_flags, snapshot_options,
4030                                    flags, options);
4031         /* Let bdrv_backing_options() override "read-only" */
4032         qdict_del(options, BDRV_OPT_READ_ONLY);
4033         bdrv_inherited_options(BDRV_CHILD_COW, true,
4034                                &flags, options, flags, options);
4035     }
4036 
4037     bs->open_flags = flags;
4038     bs->options = options;
4039     options = qdict_clone_shallow(options);
4040 
4041     /* Find the right image format driver */
4042     /* See cautionary note on accessing @options above */
4043     drvname = qdict_get_try_str(options, "driver");
4044     if (drvname) {
4045         drv = bdrv_find_format(drvname);
4046         if (!drv) {
4047             error_setg(errp, "Unknown driver: '%s'", drvname);
4048             goto fail;
4049         }
4050     }
4051 
4052     assert(drvname || !(flags & BDRV_O_PROTOCOL));
4053 
4054     /* See cautionary note on accessing @options above */
4055     backing = qdict_get_try_str(options, "backing");
4056     if (qobject_to(QNull, qdict_get(options, "backing")) != NULL ||
4057         (backing && *backing == '\0'))
4058     {
4059         if (backing) {
4060             warn_report("Use of \"backing\": \"\" is deprecated; "
4061                         "use \"backing\": null instead");
4062         }
4063         flags |= BDRV_O_NO_BACKING;
4064         qdict_del(bs->explicit_options, "backing");
4065         qdict_del(bs->options, "backing");
4066         qdict_del(options, "backing");
4067     }
4068 
4069     /* Open image file without format layer. This BlockBackend is only used for
4070      * probing, the block drivers will do their own bdrv_open_child() for the
4071      * same BDS, which is why we put the node name back into options. */
4072     if ((flags & BDRV_O_PROTOCOL) == 0) {
4073         BlockDriverState *file_bs;
4074 
4075         file_bs = bdrv_open_child_bs(filename, options, "file", bs,
4076                                      &child_of_bds, BDRV_CHILD_IMAGE,
4077                                      true, &local_err);
4078         if (local_err) {
4079             goto fail;
4080         }
4081         if (file_bs != NULL) {
4082             /* Not requesting BLK_PERM_CONSISTENT_READ because we're only
4083              * looking at the header to guess the image format. This works even
4084              * in cases where a guest would not see a consistent state. */
4085             AioContext *ctx = bdrv_get_aio_context(file_bs);
4086             file = blk_new(ctx, 0, BLK_PERM_ALL);
4087             blk_insert_bs(file, file_bs, &local_err);
4088             bdrv_unref(file_bs);
4089 
4090             if (local_err) {
4091                 goto fail;
4092             }
4093 
4094             qdict_put_str(options, "file", bdrv_get_node_name(file_bs));
4095         }
4096     }
4097 
4098     /* Image format probing */
4099     bs->probed = !drv;
4100     if (!drv && file) {
4101         ret = find_image_format(file, filename, &drv, &local_err);
4102         if (ret < 0) {
4103             goto fail;
4104         }
4105         /*
4106          * This option update would logically belong in bdrv_fill_options(),
4107          * but we first need to open bs->file for the probing to work, while
4108          * opening bs->file already requires the (mostly) final set of options
4109          * so that cache mode etc. can be inherited.
4110          *
4111          * Adding the driver later is somewhat ugly, but it's not an option
4112          * that would ever be inherited, so it's correct. We just need to make
4113          * sure to update both bs->options (which has the full effective
4114          * options for bs) and options (which has file.* already removed).
4115          */
4116         qdict_put_str(bs->options, "driver", drv->format_name);
4117         qdict_put_str(options, "driver", drv->format_name);
4118     } else if (!drv) {
4119         error_setg(errp, "Must specify either driver or file");
4120         goto fail;
4121     }
4122 
4123     /* BDRV_O_PROTOCOL must be set iff a protocol BDS is about to be created */
4124     assert(!!(flags & BDRV_O_PROTOCOL) == !!drv->bdrv_file_open);
4125     /* file must be NULL if a protocol BDS is about to be created
4126      * (the inverse results in an error message from bdrv_open_common()) */
4127     assert(!(flags & BDRV_O_PROTOCOL) || !file);
4128 
4129     /* Open the image */
4130     ret = bdrv_open_common(bs, file, options, &local_err);
4131     if (ret < 0) {
4132         goto fail;
4133     }
4134 
4135     if (file) {
4136         blk_unref(file);
4137         file = NULL;
4138     }
4139 
4140     /* If there is a backing file, use it */
4141     if ((flags & BDRV_O_NO_BACKING) == 0) {
4142         ret = bdrv_open_backing_file(bs, options, "backing", &local_err);
4143         if (ret < 0) {
4144             goto close_and_fail;
4145         }
4146     }
4147 
4148     /* Remove all children options and references
4149      * from bs->options and bs->explicit_options */
4150     QLIST_FOREACH(child, &bs->children, next) {
4151         char *child_key_dot;
4152         child_key_dot = g_strdup_printf("%s.", child->name);
4153         qdict_extract_subqdict(bs->explicit_options, NULL, child_key_dot);
4154         qdict_extract_subqdict(bs->options, NULL, child_key_dot);
4155         qdict_del(bs->explicit_options, child->name);
4156         qdict_del(bs->options, child->name);
4157         g_free(child_key_dot);
4158     }
4159 
4160     /* Check if any unknown options were used */
4161     if (qdict_size(options) != 0) {
4162         const QDictEntry *entry = qdict_first(options);
4163         if (flags & BDRV_O_PROTOCOL) {
4164             error_setg(errp, "Block protocol '%s' doesn't support the option "
4165                        "'%s'", drv->format_name, entry->key);
4166         } else {
4167             error_setg(errp,
4168                        "Block format '%s' does not support the option '%s'",
4169                        drv->format_name, entry->key);
4170         }
4171 
4172         goto close_and_fail;
4173     }
4174 
4175     bdrv_parent_cb_change_media(bs, true);
4176 
4177     qobject_unref(options);
4178     options = NULL;
4179 
4180     /* For snapshot=on, create a temporary qcow2 overlay. bs points to the
4181      * temporary snapshot afterwards. */
4182     if (snapshot_flags) {
4183         BlockDriverState *snapshot_bs;
4184         snapshot_bs = bdrv_append_temp_snapshot(bs, snapshot_flags,
4185                                                 snapshot_options, &local_err);
4186         snapshot_options = NULL;
4187         if (local_err) {
4188             goto close_and_fail;
4189         }
4190         /* We are not going to return bs but the overlay on top of it
4191          * (snapshot_bs); thus, we have to drop the strong reference to bs
4192          * (which we obtained by calling bdrv_new()). bs will not be deleted,
4193          * though, because the overlay still has a reference to it. */
4194         bdrv_unref(bs);
4195         bs = snapshot_bs;
4196     }
4197 
4198     return bs;
4199 
4200 fail:
4201     blk_unref(file);
4202     qobject_unref(snapshot_options);
4203     qobject_unref(bs->explicit_options);
4204     qobject_unref(bs->options);
4205     qobject_unref(options);
4206     bs->options = NULL;
4207     bs->explicit_options = NULL;
4208     bdrv_unref(bs);
4209     error_propagate(errp, local_err);
4210     return NULL;
4211 
4212 close_and_fail:
4213     bdrv_unref(bs);
4214     qobject_unref(snapshot_options);
4215     qobject_unref(options);
4216     error_propagate(errp, local_err);
4217     return NULL;
4218 }
4219 
4220 /* The caller must always hold the main AioContext lock. */
4221 BlockDriverState *bdrv_open(const char *filename, const char *reference,
4222                             QDict *options, int flags, Error **errp)
4223 {
4224     GLOBAL_STATE_CODE();
4225 
4226     return bdrv_open_inherit(filename, reference, options, flags, NULL,
4227                              NULL, 0, errp);
4228 }
4229 
4230 /* Return true if the NULL-terminated @list contains @str */
4231 static bool is_str_in_list(const char *str, const char *const *list)
4232 {
4233     if (str && list) {
4234         int i;
4235         for (i = 0; list[i] != NULL; i++) {
4236             if (!strcmp(str, list[i])) {
4237                 return true;
4238             }
4239         }
4240     }
4241     return false;
4242 }
4243 
4244 /*
4245  * Check that every option set in @bs->options is also set in
4246  * @new_opts.
4247  *
4248  * Options listed in the common_options list and in
4249  * @bs->drv->mutable_opts are skipped.
4250  *
4251  * Return 0 on success, otherwise return -EINVAL and set @errp.
4252  */
4253 static int bdrv_reset_options_allowed(BlockDriverState *bs,
4254                                       const QDict *new_opts, Error **errp)
4255 {
4256     const QDictEntry *e;
4257     /* These options are common to all block drivers and are handled
4258      * in bdrv_reopen_prepare() so they can be left out of @new_opts */
4259     const char *const common_options[] = {
4260         "node-name", "discard", "cache.direct", "cache.no-flush",
4261         "read-only", "auto-read-only", "detect-zeroes", NULL
4262     };
4263 
4264     for (e = qdict_first(bs->options); e; e = qdict_next(bs->options, e)) {
4265         if (!qdict_haskey(new_opts, e->key) &&
4266             !is_str_in_list(e->key, common_options) &&
4267             !is_str_in_list(e->key, bs->drv->mutable_opts)) {
4268             error_setg(errp, "Option '%s' cannot be reset "
4269                        "to its default value", e->key);
4270             return -EINVAL;
4271         }
4272     }
4273 
4274     return 0;
4275 }
4276 
4277 /*
4278  * Returns true if @child can be reached recursively from @bs
4279  */
4280 static bool GRAPH_RDLOCK
4281 bdrv_recurse_has_child(BlockDriverState *bs, BlockDriverState *child)
4282 {
4283     BdrvChild *c;
4284 
4285     if (bs == child) {
4286         return true;
4287     }
4288 
4289     QLIST_FOREACH(c, &bs->children, next) {
4290         if (bdrv_recurse_has_child(c->bs, child)) {
4291             return true;
4292         }
4293     }
4294 
4295     return false;
4296 }
4297 
4298 /*
4299  * Adds a BlockDriverState to a simple queue for an atomic, transactional
4300  * reopen of multiple devices.
4301  *
4302  * bs_queue can either be an existing BlockReopenQueue that has had QTAILQ_INIT
4303  * already performed, or alternatively may be NULL a new BlockReopenQueue will
4304  * be created and initialized. This newly created BlockReopenQueue should be
4305  * passed back in for subsequent calls that are intended to be of the same
4306  * atomic 'set'.
4307  *
4308  * bs is the BlockDriverState to add to the reopen queue.
4309  *
4310  * options contains the changed options for the associated bs
4311  * (the BlockReopenQueue takes ownership)
4312  *
4313  * flags contains the open flags for the associated bs
4314  *
4315  * returns a pointer to bs_queue, which is either the newly allocated
4316  * bs_queue, or the existing bs_queue being used.
4317  *
4318  * bs is drained here and undrained by bdrv_reopen_queue_free().
4319  *
4320  * To be called with bs->aio_context locked.
4321  */
4322 static BlockReopenQueue * GRAPH_RDLOCK
4323 bdrv_reopen_queue_child(BlockReopenQueue *bs_queue, BlockDriverState *bs,
4324                         QDict *options, const BdrvChildClass *klass,
4325                         BdrvChildRole role, bool parent_is_format,
4326                         QDict *parent_options, int parent_flags,
4327                         bool keep_old_opts)
4328 {
4329     assert(bs != NULL);
4330 
4331     BlockReopenQueueEntry *bs_entry;
4332     BdrvChild *child;
4333     QDict *old_options, *explicit_options, *options_copy;
4334     int flags;
4335     QemuOpts *opts;
4336 
4337     GLOBAL_STATE_CODE();
4338 
4339     /*
4340      * Strictly speaking, draining is illegal under GRAPH_RDLOCK. We know that
4341      * we've been called with bdrv_graph_rdlock_main_loop(), though, so it's ok
4342      * in practice.
4343      */
4344     bdrv_drained_begin(bs);
4345 
4346     if (bs_queue == NULL) {
4347         bs_queue = g_new0(BlockReopenQueue, 1);
4348         QTAILQ_INIT(bs_queue);
4349     }
4350 
4351     if (!options) {
4352         options = qdict_new();
4353     }
4354 
4355     /* Check if this BlockDriverState is already in the queue */
4356     QTAILQ_FOREACH(bs_entry, bs_queue, entry) {
4357         if (bs == bs_entry->state.bs) {
4358             break;
4359         }
4360     }
4361 
4362     /*
4363      * Precedence of options:
4364      * 1. Explicitly passed in options (highest)
4365      * 2. Retained from explicitly set options of bs
4366      * 3. Inherited from parent node
4367      * 4. Retained from effective options of bs
4368      */
4369 
4370     /* Old explicitly set values (don't overwrite by inherited value) */
4371     if (bs_entry || keep_old_opts) {
4372         old_options = qdict_clone_shallow(bs_entry ?
4373                                           bs_entry->state.explicit_options :
4374                                           bs->explicit_options);
4375         bdrv_join_options(bs, options, old_options);
4376         qobject_unref(old_options);
4377     }
4378 
4379     explicit_options = qdict_clone_shallow(options);
4380 
4381     /* Inherit from parent node */
4382     if (parent_options) {
4383         flags = 0;
4384         klass->inherit_options(role, parent_is_format, &flags, options,
4385                                parent_flags, parent_options);
4386     } else {
4387         flags = bdrv_get_flags(bs);
4388     }
4389 
4390     if (keep_old_opts) {
4391         /* Old values are used for options that aren't set yet */
4392         old_options = qdict_clone_shallow(bs->options);
4393         bdrv_join_options(bs, options, old_options);
4394         qobject_unref(old_options);
4395     }
4396 
4397     /* We have the final set of options so let's update the flags */
4398     options_copy = qdict_clone_shallow(options);
4399     opts = qemu_opts_create(&bdrv_runtime_opts, NULL, 0, &error_abort);
4400     qemu_opts_absorb_qdict(opts, options_copy, NULL);
4401     update_flags_from_options(&flags, opts);
4402     qemu_opts_del(opts);
4403     qobject_unref(options_copy);
4404 
4405     /* bdrv_open_inherit() sets and clears some additional flags internally */
4406     flags &= ~BDRV_O_PROTOCOL;
4407     if (flags & BDRV_O_RDWR) {
4408         flags |= BDRV_O_ALLOW_RDWR;
4409     }
4410 
4411     if (!bs_entry) {
4412         bs_entry = g_new0(BlockReopenQueueEntry, 1);
4413         QTAILQ_INSERT_TAIL(bs_queue, bs_entry, entry);
4414     } else {
4415         qobject_unref(bs_entry->state.options);
4416         qobject_unref(bs_entry->state.explicit_options);
4417     }
4418 
4419     bs_entry->state.bs = bs;
4420     bs_entry->state.options = options;
4421     bs_entry->state.explicit_options = explicit_options;
4422     bs_entry->state.flags = flags;
4423 
4424     /*
4425      * If keep_old_opts is false then it means that unspecified
4426      * options must be reset to their original value. We don't allow
4427      * resetting 'backing' but we need to know if the option is
4428      * missing in order to decide if we have to return an error.
4429      */
4430     if (!keep_old_opts) {
4431         bs_entry->state.backing_missing =
4432             !qdict_haskey(options, "backing") &&
4433             !qdict_haskey(options, "backing.driver");
4434     }
4435 
4436     QLIST_FOREACH(child, &bs->children, next) {
4437         QDict *new_child_options = NULL;
4438         bool child_keep_old = keep_old_opts;
4439 
4440         /* reopen can only change the options of block devices that were
4441          * implicitly created and inherited options. For other (referenced)
4442          * block devices, a syntax like "backing.foo" results in an error. */
4443         if (child->bs->inherits_from != bs) {
4444             continue;
4445         }
4446 
4447         /* Check if the options contain a child reference */
4448         if (qdict_haskey(options, child->name)) {
4449             const char *childref = qdict_get_try_str(options, child->name);
4450             /*
4451              * The current child must not be reopened if the child
4452              * reference is null or points to a different node.
4453              */
4454             if (g_strcmp0(childref, child->bs->node_name)) {
4455                 continue;
4456             }
4457             /*
4458              * If the child reference points to the current child then
4459              * reopen it with its existing set of options (note that
4460              * it can still inherit new options from the parent).
4461              */
4462             child_keep_old = true;
4463         } else {
4464             /* Extract child options ("child-name.*") */
4465             char *child_key_dot = g_strdup_printf("%s.", child->name);
4466             qdict_extract_subqdict(explicit_options, NULL, child_key_dot);
4467             qdict_extract_subqdict(options, &new_child_options, child_key_dot);
4468             g_free(child_key_dot);
4469         }
4470 
4471         bdrv_reopen_queue_child(bs_queue, child->bs, new_child_options,
4472                                 child->klass, child->role, bs->drv->is_format,
4473                                 options, flags, child_keep_old);
4474     }
4475 
4476     return bs_queue;
4477 }
4478 
4479 /* To be called with bs->aio_context locked */
4480 BlockReopenQueue *bdrv_reopen_queue(BlockReopenQueue *bs_queue,
4481                                     BlockDriverState *bs,
4482                                     QDict *options, bool keep_old_opts)
4483 {
4484     GLOBAL_STATE_CODE();
4485     GRAPH_RDLOCK_GUARD_MAINLOOP();
4486 
4487     return bdrv_reopen_queue_child(bs_queue, bs, options, NULL, 0, false,
4488                                    NULL, 0, keep_old_opts);
4489 }
4490 
4491 void bdrv_reopen_queue_free(BlockReopenQueue *bs_queue)
4492 {
4493     GLOBAL_STATE_CODE();
4494     if (bs_queue) {
4495         BlockReopenQueueEntry *bs_entry, *next;
4496         QTAILQ_FOREACH_SAFE(bs_entry, bs_queue, entry, next) {
4497             bdrv_drained_end(bs_entry->state.bs);
4498             qobject_unref(bs_entry->state.explicit_options);
4499             qobject_unref(bs_entry->state.options);
4500             g_free(bs_entry);
4501         }
4502         g_free(bs_queue);
4503     }
4504 }
4505 
4506 /*
4507  * Reopen multiple BlockDriverStates atomically & transactionally.
4508  *
4509  * The queue passed in (bs_queue) must have been built up previous
4510  * via bdrv_reopen_queue().
4511  *
4512  * Reopens all BDS specified in the queue, with the appropriate
4513  * flags.  All devices are prepared for reopen, and failure of any
4514  * device will cause all device changes to be abandoned, and intermediate
4515  * data cleaned up.
4516  *
4517  * If all devices prepare successfully, then the changes are committed
4518  * to all devices.
4519  *
4520  * All affected nodes must be drained between bdrv_reopen_queue() and
4521  * bdrv_reopen_multiple().
4522  *
4523  * To be called from the main thread, with all other AioContexts unlocked.
4524  */
4525 int bdrv_reopen_multiple(BlockReopenQueue *bs_queue, Error **errp)
4526 {
4527     int ret = -1;
4528     BlockReopenQueueEntry *bs_entry, *next;
4529     Transaction *tran = tran_new();
4530     g_autoptr(GSList) refresh_list = NULL;
4531 
4532     assert(qemu_get_current_aio_context() == qemu_get_aio_context());
4533     assert(bs_queue != NULL);
4534     GLOBAL_STATE_CODE();
4535 
4536     QTAILQ_FOREACH(bs_entry, bs_queue, entry) {
4537         ret = bdrv_flush(bs_entry->state.bs);
4538         if (ret < 0) {
4539             error_setg_errno(errp, -ret, "Error flushing drive");
4540             goto abort;
4541         }
4542     }
4543 
4544     QTAILQ_FOREACH(bs_entry, bs_queue, entry) {
4545         assert(bs_entry->state.bs->quiesce_counter > 0);
4546         ret = bdrv_reopen_prepare(&bs_entry->state, bs_queue, tran, errp);
4547         if (ret < 0) {
4548             goto abort;
4549         }
4550         bs_entry->prepared = true;
4551     }
4552 
4553     QTAILQ_FOREACH(bs_entry, bs_queue, entry) {
4554         BDRVReopenState *state = &bs_entry->state;
4555 
4556         refresh_list = g_slist_prepend(refresh_list, state->bs);
4557         if (state->old_backing_bs) {
4558             refresh_list = g_slist_prepend(refresh_list, state->old_backing_bs);
4559         }
4560         if (state->old_file_bs) {
4561             refresh_list = g_slist_prepend(refresh_list, state->old_file_bs);
4562         }
4563     }
4564 
4565     /*
4566      * Note that file-posix driver rely on permission update done during reopen
4567      * (even if no permission changed), because it wants "new" permissions for
4568      * reconfiguring the fd and that's why it does it in raw_check_perm(), not
4569      * in raw_reopen_prepare() which is called with "old" permissions.
4570      */
4571     bdrv_graph_rdlock_main_loop();
4572     ret = bdrv_list_refresh_perms(refresh_list, bs_queue, tran, errp);
4573     bdrv_graph_rdunlock_main_loop();
4574 
4575     if (ret < 0) {
4576         goto abort;
4577     }
4578 
4579     /*
4580      * If we reach this point, we have success and just need to apply the
4581      * changes.
4582      *
4583      * Reverse order is used to comfort qcow2 driver: on commit it need to write
4584      * IN_USE flag to the image, to mark bitmaps in the image as invalid. But
4585      * children are usually goes after parents in reopen-queue, so go from last
4586      * to first element.
4587      */
4588     QTAILQ_FOREACH_REVERSE(bs_entry, bs_queue, entry) {
4589         bdrv_reopen_commit(&bs_entry->state);
4590     }
4591 
4592     bdrv_graph_wrlock();
4593     tran_commit(tran);
4594     bdrv_graph_wrunlock();
4595 
4596     QTAILQ_FOREACH_REVERSE(bs_entry, bs_queue, entry) {
4597         BlockDriverState *bs = bs_entry->state.bs;
4598 
4599         if (bs->drv->bdrv_reopen_commit_post) {
4600             bs->drv->bdrv_reopen_commit_post(&bs_entry->state);
4601         }
4602     }
4603 
4604     ret = 0;
4605     goto cleanup;
4606 
4607 abort:
4608     bdrv_graph_wrlock();
4609     tran_abort(tran);
4610     bdrv_graph_wrunlock();
4611 
4612     QTAILQ_FOREACH_SAFE(bs_entry, bs_queue, entry, next) {
4613         if (bs_entry->prepared) {
4614             bdrv_reopen_abort(&bs_entry->state);
4615         }
4616     }
4617 
4618 cleanup:
4619     bdrv_reopen_queue_free(bs_queue);
4620 
4621     return ret;
4622 }
4623 
4624 int bdrv_reopen(BlockDriverState *bs, QDict *opts, bool keep_old_opts,
4625                 Error **errp)
4626 {
4627     BlockReopenQueue *queue;
4628 
4629     GLOBAL_STATE_CODE();
4630 
4631     queue = bdrv_reopen_queue(NULL, bs, opts, keep_old_opts);
4632 
4633     return bdrv_reopen_multiple(queue, errp);
4634 }
4635 
4636 int bdrv_reopen_set_read_only(BlockDriverState *bs, bool read_only,
4637                               Error **errp)
4638 {
4639     QDict *opts = qdict_new();
4640 
4641     GLOBAL_STATE_CODE();
4642 
4643     qdict_put_bool(opts, BDRV_OPT_READ_ONLY, read_only);
4644 
4645     return bdrv_reopen(bs, opts, true, errp);
4646 }
4647 
4648 /*
4649  * Take a BDRVReopenState and check if the value of 'backing' in the
4650  * reopen_state->options QDict is valid or not.
4651  *
4652  * If 'backing' is missing from the QDict then return 0.
4653  *
4654  * If 'backing' contains the node name of the backing file of
4655  * reopen_state->bs then return 0.
4656  *
4657  * If 'backing' contains a different node name (or is null) then check
4658  * whether the current backing file can be replaced with the new one.
4659  * If that's the case then reopen_state->replace_backing_bs is set to
4660  * true and reopen_state->new_backing_bs contains a pointer to the new
4661  * backing BlockDriverState (or NULL).
4662  *
4663  * After calling this function, the transaction @tran may only be completed
4664  * while holding a writer lock for the graph.
4665  *
4666  * Return 0 on success, otherwise return < 0 and set @errp.
4667  *
4668  * The caller must hold the AioContext lock of @reopen_state->bs.
4669  * @reopen_state->bs can move to a different AioContext in this function.
4670  * Callers must make sure that their AioContext locking is still correct after
4671  * this.
4672  */
4673 static int GRAPH_UNLOCKED
4674 bdrv_reopen_parse_file_or_backing(BDRVReopenState *reopen_state,
4675                                   bool is_backing, Transaction *tran,
4676                                   Error **errp)
4677 {
4678     BlockDriverState *bs = reopen_state->bs;
4679     BlockDriverState *new_child_bs;
4680     BlockDriverState *old_child_bs;
4681 
4682     const char *child_name = is_backing ? "backing" : "file";
4683     QObject *value;
4684     const char *str;
4685     bool has_child;
4686     int ret;
4687 
4688     GLOBAL_STATE_CODE();
4689 
4690     value = qdict_get(reopen_state->options, child_name);
4691     if (value == NULL) {
4692         return 0;
4693     }
4694 
4695     bdrv_graph_rdlock_main_loop();
4696 
4697     switch (qobject_type(value)) {
4698     case QTYPE_QNULL:
4699         assert(is_backing); /* The 'file' option does not allow a null value */
4700         new_child_bs = NULL;
4701         break;
4702     case QTYPE_QSTRING:
4703         str = qstring_get_str(qobject_to(QString, value));
4704         new_child_bs = bdrv_lookup_bs(NULL, str, errp);
4705         if (new_child_bs == NULL) {
4706             ret = -EINVAL;
4707             goto out_rdlock;
4708         }
4709 
4710         has_child = bdrv_recurse_has_child(new_child_bs, bs);
4711         if (has_child) {
4712             error_setg(errp, "Making '%s' a %s child of '%s' would create a "
4713                        "cycle", str, child_name, bs->node_name);
4714             ret = -EINVAL;
4715             goto out_rdlock;
4716         }
4717         break;
4718     default:
4719         /*
4720          * The options QDict has been flattened, so 'backing' and 'file'
4721          * do not allow any other data type here.
4722          */
4723         g_assert_not_reached();
4724     }
4725 
4726     old_child_bs = is_backing ? child_bs(bs->backing) : child_bs(bs->file);
4727     if (old_child_bs == new_child_bs) {
4728         ret = 0;
4729         goto out_rdlock;
4730     }
4731 
4732     if (old_child_bs) {
4733         if (bdrv_skip_implicit_filters(old_child_bs) == new_child_bs) {
4734             ret = 0;
4735             goto out_rdlock;
4736         }
4737 
4738         if (old_child_bs->implicit) {
4739             error_setg(errp, "Cannot replace implicit %s child of %s",
4740                        child_name, bs->node_name);
4741             ret = -EPERM;
4742             goto out_rdlock;
4743         }
4744     }
4745 
4746     if (bs->drv->is_filter && !old_child_bs) {
4747         /*
4748          * Filters always have a file or a backing child, so we are trying to
4749          * change wrong child
4750          */
4751         error_setg(errp, "'%s' is a %s filter node that does not support a "
4752                    "%s child", bs->node_name, bs->drv->format_name, child_name);
4753         ret = -EINVAL;
4754         goto out_rdlock;
4755     }
4756 
4757     if (is_backing) {
4758         reopen_state->old_backing_bs = old_child_bs;
4759     } else {
4760         reopen_state->old_file_bs = old_child_bs;
4761     }
4762 
4763     if (old_child_bs) {
4764         bdrv_ref(old_child_bs);
4765         bdrv_drained_begin(old_child_bs);
4766     }
4767 
4768     bdrv_graph_rdunlock_main_loop();
4769     bdrv_graph_wrlock();
4770 
4771     ret = bdrv_set_file_or_backing_noperm(bs, new_child_bs, is_backing,
4772                                           tran, errp);
4773 
4774     bdrv_graph_wrunlock();
4775 
4776     if (old_child_bs) {
4777         bdrv_drained_end(old_child_bs);
4778         bdrv_unref(old_child_bs);
4779     }
4780 
4781     return ret;
4782 
4783 out_rdlock:
4784     bdrv_graph_rdunlock_main_loop();
4785     return ret;
4786 }
4787 
4788 /*
4789  * Prepares a BlockDriverState for reopen. All changes are staged in the
4790  * 'opaque' field of the BDRVReopenState, which is used and allocated by
4791  * the block driver layer .bdrv_reopen_prepare()
4792  *
4793  * bs is the BlockDriverState to reopen
4794  * flags are the new open flags
4795  * queue is the reopen queue
4796  *
4797  * Returns 0 on success, non-zero on error.  On error errp will be set
4798  * as well.
4799  *
4800  * On failure, bdrv_reopen_abort() will be called to clean up any data.
4801  * It is the responsibility of the caller to then call the abort() or
4802  * commit() for any other BDS that have been left in a prepare() state
4803  *
4804  * The caller must hold the AioContext lock of @reopen_state->bs.
4805  *
4806  * After calling this function, the transaction @change_child_tran may only be
4807  * completed while holding a writer lock for the graph.
4808  */
4809 static int GRAPH_UNLOCKED
4810 bdrv_reopen_prepare(BDRVReopenState *reopen_state, BlockReopenQueue *queue,
4811                     Transaction *change_child_tran, Error **errp)
4812 {
4813     int ret = -1;
4814     int old_flags;
4815     Error *local_err = NULL;
4816     BlockDriver *drv;
4817     QemuOpts *opts;
4818     QDict *orig_reopen_opts;
4819     char *discard = NULL;
4820     bool read_only;
4821     bool drv_prepared = false;
4822 
4823     assert(reopen_state != NULL);
4824     assert(reopen_state->bs->drv != NULL);
4825     GLOBAL_STATE_CODE();
4826     drv = reopen_state->bs->drv;
4827 
4828     /* This function and each driver's bdrv_reopen_prepare() remove
4829      * entries from reopen_state->options as they are processed, so
4830      * we need to make a copy of the original QDict. */
4831     orig_reopen_opts = qdict_clone_shallow(reopen_state->options);
4832 
4833     /* Process generic block layer options */
4834     opts = qemu_opts_create(&bdrv_runtime_opts, NULL, 0, &error_abort);
4835     if (!qemu_opts_absorb_qdict(opts, reopen_state->options, errp)) {
4836         ret = -EINVAL;
4837         goto error;
4838     }
4839 
4840     /* This was already called in bdrv_reopen_queue_child() so the flags
4841      * are up-to-date. This time we simply want to remove the options from
4842      * QemuOpts in order to indicate that they have been processed. */
4843     old_flags = reopen_state->flags;
4844     update_flags_from_options(&reopen_state->flags, opts);
4845     assert(old_flags == reopen_state->flags);
4846 
4847     discard = qemu_opt_get_del(opts, BDRV_OPT_DISCARD);
4848     if (discard != NULL) {
4849         if (bdrv_parse_discard_flags(discard, &reopen_state->flags) != 0) {
4850             error_setg(errp, "Invalid discard option");
4851             ret = -EINVAL;
4852             goto error;
4853         }
4854     }
4855 
4856     reopen_state->detect_zeroes =
4857         bdrv_parse_detect_zeroes(opts, reopen_state->flags, &local_err);
4858     if (local_err) {
4859         error_propagate(errp, local_err);
4860         ret = -EINVAL;
4861         goto error;
4862     }
4863 
4864     /* All other options (including node-name and driver) must be unchanged.
4865      * Put them back into the QDict, so that they are checked at the end
4866      * of this function. */
4867     qemu_opts_to_qdict(opts, reopen_state->options);
4868 
4869     /* If we are to stay read-only, do not allow permission change
4870      * to r/w. Attempting to set to r/w may fail if either BDRV_O_ALLOW_RDWR is
4871      * not set, or if the BDS still has copy_on_read enabled */
4872     read_only = !(reopen_state->flags & BDRV_O_RDWR);
4873 
4874     bdrv_graph_rdlock_main_loop();
4875     ret = bdrv_can_set_read_only(reopen_state->bs, read_only, true, &local_err);
4876     bdrv_graph_rdunlock_main_loop();
4877     if (local_err) {
4878         error_propagate(errp, local_err);
4879         goto error;
4880     }
4881 
4882     if (drv->bdrv_reopen_prepare) {
4883         /*
4884          * If a driver-specific option is missing, it means that we
4885          * should reset it to its default value.
4886          * But not all options allow that, so we need to check it first.
4887          */
4888         ret = bdrv_reset_options_allowed(reopen_state->bs,
4889                                          reopen_state->options, errp);
4890         if (ret) {
4891             goto error;
4892         }
4893 
4894         ret = drv->bdrv_reopen_prepare(reopen_state, queue, &local_err);
4895         if (ret) {
4896             if (local_err != NULL) {
4897                 error_propagate(errp, local_err);
4898             } else {
4899                 bdrv_graph_rdlock_main_loop();
4900                 bdrv_refresh_filename(reopen_state->bs);
4901                 bdrv_graph_rdunlock_main_loop();
4902                 error_setg(errp, "failed while preparing to reopen image '%s'",
4903                            reopen_state->bs->filename);
4904             }
4905             goto error;
4906         }
4907     } else {
4908         /* It is currently mandatory to have a bdrv_reopen_prepare()
4909          * handler for each supported drv. */
4910         bdrv_graph_rdlock_main_loop();
4911         error_setg(errp, "Block format '%s' used by node '%s' "
4912                    "does not support reopening files", drv->format_name,
4913                    bdrv_get_device_or_node_name(reopen_state->bs));
4914         bdrv_graph_rdunlock_main_loop();
4915         ret = -1;
4916         goto error;
4917     }
4918 
4919     drv_prepared = true;
4920 
4921     /*
4922      * We must provide the 'backing' option if the BDS has a backing
4923      * file or if the image file has a backing file name as part of
4924      * its metadata. Otherwise the 'backing' option can be omitted.
4925      */
4926     bdrv_graph_rdlock_main_loop();
4927     if (drv->supports_backing && reopen_state->backing_missing &&
4928         (reopen_state->bs->backing || reopen_state->bs->backing_file[0])) {
4929         error_setg(errp, "backing is missing for '%s'",
4930                    reopen_state->bs->node_name);
4931         bdrv_graph_rdunlock_main_loop();
4932         ret = -EINVAL;
4933         goto error;
4934     }
4935     bdrv_graph_rdunlock_main_loop();
4936 
4937     /*
4938      * Allow changing the 'backing' option. The new value can be
4939      * either a reference to an existing node (using its node name)
4940      * or NULL to simply detach the current backing file.
4941      */
4942     ret = bdrv_reopen_parse_file_or_backing(reopen_state, true,
4943                                             change_child_tran, errp);
4944     if (ret < 0) {
4945         goto error;
4946     }
4947     qdict_del(reopen_state->options, "backing");
4948 
4949     /* Allow changing the 'file' option. In this case NULL is not allowed */
4950     ret = bdrv_reopen_parse_file_or_backing(reopen_state, false,
4951                                             change_child_tran, errp);
4952     if (ret < 0) {
4953         goto error;
4954     }
4955     qdict_del(reopen_state->options, "file");
4956 
4957     /* Options that are not handled are only okay if they are unchanged
4958      * compared to the old state. It is expected that some options are only
4959      * used for the initial open, but not reopen (e.g. filename) */
4960     if (qdict_size(reopen_state->options)) {
4961         const QDictEntry *entry = qdict_first(reopen_state->options);
4962 
4963         GRAPH_RDLOCK_GUARD_MAINLOOP();
4964 
4965         do {
4966             QObject *new = entry->value;
4967             QObject *old = qdict_get(reopen_state->bs->options, entry->key);
4968 
4969             /* Allow child references (child_name=node_name) as long as they
4970              * point to the current child (i.e. everything stays the same). */
4971             if (qobject_type(new) == QTYPE_QSTRING) {
4972                 BdrvChild *child;
4973                 QLIST_FOREACH(child, &reopen_state->bs->children, next) {
4974                     if (!strcmp(child->name, entry->key)) {
4975                         break;
4976                     }
4977                 }
4978 
4979                 if (child) {
4980                     if (!strcmp(child->bs->node_name,
4981                                 qstring_get_str(qobject_to(QString, new)))) {
4982                         continue; /* Found child with this name, skip option */
4983                     }
4984                 }
4985             }
4986 
4987             /*
4988              * TODO: When using -drive to specify blockdev options, all values
4989              * will be strings; however, when using -blockdev, blockdev-add or
4990              * filenames using the json:{} pseudo-protocol, they will be
4991              * correctly typed.
4992              * In contrast, reopening options are (currently) always strings
4993              * (because you can only specify them through qemu-io; all other
4994              * callers do not specify any options).
4995              * Therefore, when using anything other than -drive to create a BDS,
4996              * this cannot detect non-string options as unchanged, because
4997              * qobject_is_equal() always returns false for objects of different
4998              * type.  In the future, this should be remedied by correctly typing
4999              * all options.  For now, this is not too big of an issue because
5000              * the user can simply omit options which cannot be changed anyway,
5001              * so they will stay unchanged.
5002              */
5003             if (!qobject_is_equal(new, old)) {
5004                 error_setg(errp, "Cannot change the option '%s'", entry->key);
5005                 ret = -EINVAL;
5006                 goto error;
5007             }
5008         } while ((entry = qdict_next(reopen_state->options, entry)));
5009     }
5010 
5011     ret = 0;
5012 
5013     /* Restore the original reopen_state->options QDict */
5014     qobject_unref(reopen_state->options);
5015     reopen_state->options = qobject_ref(orig_reopen_opts);
5016 
5017 error:
5018     if (ret < 0 && drv_prepared) {
5019         /* drv->bdrv_reopen_prepare() has succeeded, so we need to
5020          * call drv->bdrv_reopen_abort() before signaling an error
5021          * (bdrv_reopen_multiple() will not call bdrv_reopen_abort()
5022          * when the respective bdrv_reopen_prepare() has failed) */
5023         if (drv->bdrv_reopen_abort) {
5024             drv->bdrv_reopen_abort(reopen_state);
5025         }
5026     }
5027     qemu_opts_del(opts);
5028     qobject_unref(orig_reopen_opts);
5029     g_free(discard);
5030     return ret;
5031 }
5032 
5033 /*
5034  * Takes the staged changes for the reopen from bdrv_reopen_prepare(), and
5035  * makes them final by swapping the staging BlockDriverState contents into
5036  * the active BlockDriverState contents.
5037  */
5038 static void GRAPH_UNLOCKED bdrv_reopen_commit(BDRVReopenState *reopen_state)
5039 {
5040     BlockDriver *drv;
5041     BlockDriverState *bs;
5042     BdrvChild *child;
5043 
5044     assert(reopen_state != NULL);
5045     bs = reopen_state->bs;
5046     drv = bs->drv;
5047     assert(drv != NULL);
5048     GLOBAL_STATE_CODE();
5049 
5050     /* If there are any driver level actions to take */
5051     if (drv->bdrv_reopen_commit) {
5052         drv->bdrv_reopen_commit(reopen_state);
5053     }
5054 
5055     GRAPH_RDLOCK_GUARD_MAINLOOP();
5056 
5057     /* set BDS specific flags now */
5058     qobject_unref(bs->explicit_options);
5059     qobject_unref(bs->options);
5060     qobject_ref(reopen_state->explicit_options);
5061     qobject_ref(reopen_state->options);
5062 
5063     bs->explicit_options   = reopen_state->explicit_options;
5064     bs->options            = reopen_state->options;
5065     bs->open_flags         = reopen_state->flags;
5066     bs->detect_zeroes      = reopen_state->detect_zeroes;
5067 
5068     /* Remove child references from bs->options and bs->explicit_options.
5069      * Child options were already removed in bdrv_reopen_queue_child() */
5070     QLIST_FOREACH(child, &bs->children, next) {
5071         qdict_del(bs->explicit_options, child->name);
5072         qdict_del(bs->options, child->name);
5073     }
5074     /* backing is probably removed, so it's not handled by previous loop */
5075     qdict_del(bs->explicit_options, "backing");
5076     qdict_del(bs->options, "backing");
5077 
5078     bdrv_refresh_limits(bs, NULL, NULL);
5079     bdrv_refresh_total_sectors(bs, bs->total_sectors);
5080 }
5081 
5082 /*
5083  * Abort the reopen, and delete and free the staged changes in
5084  * reopen_state
5085  */
5086 static void GRAPH_UNLOCKED bdrv_reopen_abort(BDRVReopenState *reopen_state)
5087 {
5088     BlockDriver *drv;
5089 
5090     assert(reopen_state != NULL);
5091     drv = reopen_state->bs->drv;
5092     assert(drv != NULL);
5093     GLOBAL_STATE_CODE();
5094 
5095     if (drv->bdrv_reopen_abort) {
5096         drv->bdrv_reopen_abort(reopen_state);
5097     }
5098 }
5099 
5100 
5101 static void bdrv_close(BlockDriverState *bs)
5102 {
5103     BdrvAioNotifier *ban, *ban_next;
5104     BdrvChild *child, *next;
5105 
5106     GLOBAL_STATE_CODE();
5107     assert(!bs->refcnt);
5108 
5109     bdrv_drained_begin(bs); /* complete I/O */
5110     bdrv_flush(bs);
5111     bdrv_drain(bs); /* in case flush left pending I/O */
5112 
5113     if (bs->drv) {
5114         if (bs->drv->bdrv_close) {
5115             /* Must unfreeze all children, so bdrv_unref_child() works */
5116             bs->drv->bdrv_close(bs);
5117         }
5118         bs->drv = NULL;
5119     }
5120 
5121     bdrv_graph_wrlock();
5122     QLIST_FOREACH_SAFE(child, &bs->children, next, next) {
5123         bdrv_unref_child(bs, child);
5124     }
5125 
5126     assert(!bs->backing);
5127     assert(!bs->file);
5128     bdrv_graph_wrunlock();
5129 
5130     g_free(bs->opaque);
5131     bs->opaque = NULL;
5132     qatomic_set(&bs->copy_on_read, 0);
5133     bs->backing_file[0] = '\0';
5134     bs->backing_format[0] = '\0';
5135     bs->total_sectors = 0;
5136     bs->encrypted = false;
5137     bs->sg = false;
5138     qobject_unref(bs->options);
5139     qobject_unref(bs->explicit_options);
5140     bs->options = NULL;
5141     bs->explicit_options = NULL;
5142     qobject_unref(bs->full_open_options);
5143     bs->full_open_options = NULL;
5144     g_free(bs->block_status_cache);
5145     bs->block_status_cache = NULL;
5146 
5147     bdrv_release_named_dirty_bitmaps(bs);
5148     assert(QLIST_EMPTY(&bs->dirty_bitmaps));
5149 
5150     QLIST_FOREACH_SAFE(ban, &bs->aio_notifiers, list, ban_next) {
5151         g_free(ban);
5152     }
5153     QLIST_INIT(&bs->aio_notifiers);
5154     bdrv_drained_end(bs);
5155 
5156     /*
5157      * If we're still inside some bdrv_drain_all_begin()/end() sections, end
5158      * them now since this BDS won't exist anymore when bdrv_drain_all_end()
5159      * gets called.
5160      */
5161     if (bs->quiesce_counter) {
5162         bdrv_drain_all_end_quiesce(bs);
5163     }
5164 }
5165 
5166 void bdrv_close_all(void)
5167 {
5168     GLOBAL_STATE_CODE();
5169     assert(job_next(NULL) == NULL);
5170 
5171     /* Drop references from requests still in flight, such as canceled block
5172      * jobs whose AIO context has not been polled yet */
5173     bdrv_drain_all();
5174 
5175     blk_remove_all_bs();
5176     blockdev_close_all_bdrv_states();
5177 
5178     assert(QTAILQ_EMPTY(&all_bdrv_states));
5179 }
5180 
5181 static bool GRAPH_RDLOCK should_update_child(BdrvChild *c, BlockDriverState *to)
5182 {
5183     GQueue *queue;
5184     GHashTable *found;
5185     bool ret;
5186 
5187     if (c->klass->stay_at_node) {
5188         return false;
5189     }
5190 
5191     /* If the child @c belongs to the BDS @to, replacing the current
5192      * c->bs by @to would mean to create a loop.
5193      *
5194      * Such a case occurs when appending a BDS to a backing chain.
5195      * For instance, imagine the following chain:
5196      *
5197      *   guest device -> node A -> further backing chain...
5198      *
5199      * Now we create a new BDS B which we want to put on top of this
5200      * chain, so we first attach A as its backing node:
5201      *
5202      *                   node B
5203      *                     |
5204      *                     v
5205      *   guest device -> node A -> further backing chain...
5206      *
5207      * Finally we want to replace A by B.  When doing that, we want to
5208      * replace all pointers to A by pointers to B -- except for the
5209      * pointer from B because (1) that would create a loop, and (2)
5210      * that pointer should simply stay intact:
5211      *
5212      *   guest device -> node B
5213      *                     |
5214      *                     v
5215      *                   node A -> further backing chain...
5216      *
5217      * In general, when replacing a node A (c->bs) by a node B (@to),
5218      * if A is a child of B, that means we cannot replace A by B there
5219      * because that would create a loop.  Silently detaching A from B
5220      * is also not really an option.  So overall just leaving A in
5221      * place there is the most sensible choice.
5222      *
5223      * We would also create a loop in any cases where @c is only
5224      * indirectly referenced by @to. Prevent this by returning false
5225      * if @c is found (by breadth-first search) anywhere in the whole
5226      * subtree of @to.
5227      */
5228 
5229     ret = true;
5230     found = g_hash_table_new(NULL, NULL);
5231     g_hash_table_add(found, to);
5232     queue = g_queue_new();
5233     g_queue_push_tail(queue, to);
5234 
5235     while (!g_queue_is_empty(queue)) {
5236         BlockDriverState *v = g_queue_pop_head(queue);
5237         BdrvChild *c2;
5238 
5239         QLIST_FOREACH(c2, &v->children, next) {
5240             if (c2 == c) {
5241                 ret = false;
5242                 break;
5243             }
5244 
5245             if (g_hash_table_contains(found, c2->bs)) {
5246                 continue;
5247             }
5248 
5249             g_queue_push_tail(queue, c2->bs);
5250             g_hash_table_add(found, c2->bs);
5251         }
5252     }
5253 
5254     g_queue_free(queue);
5255     g_hash_table_destroy(found);
5256 
5257     return ret;
5258 }
5259 
5260 static void bdrv_remove_child_commit(void *opaque)
5261 {
5262     GLOBAL_STATE_CODE();
5263     bdrv_child_free(opaque);
5264 }
5265 
5266 static TransactionActionDrv bdrv_remove_child_drv = {
5267     .commit = bdrv_remove_child_commit,
5268 };
5269 
5270 /*
5271  * Function doesn't update permissions, caller is responsible for this.
5272  *
5273  * @child->bs (if non-NULL) must be drained.
5274  *
5275  * After calling this function, the transaction @tran may only be completed
5276  * while holding a writer lock for the graph.
5277  */
5278 static void GRAPH_WRLOCK bdrv_remove_child(BdrvChild *child, Transaction *tran)
5279 {
5280     if (!child) {
5281         return;
5282     }
5283 
5284     if (child->bs) {
5285         assert(child->quiesced_parent);
5286         bdrv_replace_child_tran(child, NULL, tran);
5287     }
5288 
5289     tran_add(tran, &bdrv_remove_child_drv, child);
5290 }
5291 
5292 /*
5293  * Both @from and @to (if non-NULL) must be drained. @to must be kept drained
5294  * until the transaction is completed.
5295  *
5296  * After calling this function, the transaction @tran may only be completed
5297  * while holding a writer lock for the graph.
5298  */
5299 static int GRAPH_WRLOCK
5300 bdrv_replace_node_noperm(BlockDriverState *from,
5301                          BlockDriverState *to,
5302                          bool auto_skip, Transaction *tran,
5303                          Error **errp)
5304 {
5305     BdrvChild *c, *next;
5306 
5307     GLOBAL_STATE_CODE();
5308 
5309     assert(from->quiesce_counter);
5310     assert(to->quiesce_counter);
5311 
5312     QLIST_FOREACH_SAFE(c, &from->parents, next_parent, next) {
5313         assert(c->bs == from);
5314         if (!should_update_child(c, to)) {
5315             if (auto_skip) {
5316                 continue;
5317             }
5318             error_setg(errp, "Should not change '%s' link to '%s'",
5319                        c->name, from->node_name);
5320             return -EINVAL;
5321         }
5322         if (c->frozen) {
5323             error_setg(errp, "Cannot change '%s' link to '%s'",
5324                        c->name, from->node_name);
5325             return -EPERM;
5326         }
5327         bdrv_replace_child_tran(c, to, tran);
5328     }
5329 
5330     return 0;
5331 }
5332 
5333 /*
5334  * Switch all parents of @from to point to @to instead. @from and @to must be in
5335  * the same AioContext and both must be drained.
5336  *
5337  * With auto_skip=true bdrv_replace_node_common skips updating from parents
5338  * if it creates a parent-child relation loop or if parent is block-job.
5339  *
5340  * With auto_skip=false the error is returned if from has a parent which should
5341  * not be updated.
5342  *
5343  * With @detach_subchain=true @to must be in a backing chain of @from. In this
5344  * case backing link of the cow-parent of @to is removed.
5345  */
5346 static int GRAPH_WRLOCK
5347 bdrv_replace_node_common(BlockDriverState *from, BlockDriverState *to,
5348                          bool auto_skip, bool detach_subchain, Error **errp)
5349 {
5350     Transaction *tran = tran_new();
5351     g_autoptr(GSList) refresh_list = NULL;
5352     BlockDriverState *to_cow_parent = NULL;
5353     int ret;
5354 
5355     GLOBAL_STATE_CODE();
5356 
5357     assert(from->quiesce_counter);
5358     assert(to->quiesce_counter);
5359     assert(bdrv_get_aio_context(from) == bdrv_get_aio_context(to));
5360 
5361     if (detach_subchain) {
5362         assert(bdrv_chain_contains(from, to));
5363         assert(from != to);
5364         for (to_cow_parent = from;
5365              bdrv_filter_or_cow_bs(to_cow_parent) != to;
5366              to_cow_parent = bdrv_filter_or_cow_bs(to_cow_parent))
5367         {
5368             ;
5369         }
5370     }
5371 
5372     /*
5373      * Do the replacement without permission update.
5374      * Replacement may influence the permissions, we should calculate new
5375      * permissions based on new graph. If we fail, we'll roll-back the
5376      * replacement.
5377      */
5378     ret = bdrv_replace_node_noperm(from, to, auto_skip, tran, errp);
5379     if (ret < 0) {
5380         goto out;
5381     }
5382 
5383     if (detach_subchain) {
5384         /* to_cow_parent is already drained because from is drained */
5385         bdrv_remove_child(bdrv_filter_or_cow_child(to_cow_parent), tran);
5386     }
5387 
5388     refresh_list = g_slist_prepend(refresh_list, to);
5389     refresh_list = g_slist_prepend(refresh_list, from);
5390 
5391     ret = bdrv_list_refresh_perms(refresh_list, NULL, tran, errp);
5392     if (ret < 0) {
5393         goto out;
5394     }
5395 
5396     ret = 0;
5397 
5398 out:
5399     tran_finalize(tran, ret);
5400     return ret;
5401 }
5402 
5403 int bdrv_replace_node(BlockDriverState *from, BlockDriverState *to,
5404                       Error **errp)
5405 {
5406     return bdrv_replace_node_common(from, to, true, false, errp);
5407 }
5408 
5409 int bdrv_drop_filter(BlockDriverState *bs, Error **errp)
5410 {
5411     BlockDriverState *child_bs;
5412     int ret;
5413 
5414     GLOBAL_STATE_CODE();
5415 
5416     bdrv_graph_rdlock_main_loop();
5417     child_bs = bdrv_filter_or_cow_bs(bs);
5418     bdrv_graph_rdunlock_main_loop();
5419 
5420     bdrv_drained_begin(child_bs);
5421     bdrv_graph_wrlock();
5422     ret = bdrv_replace_node_common(bs, child_bs, true, true, errp);
5423     bdrv_graph_wrunlock();
5424     bdrv_drained_end(child_bs);
5425 
5426     return ret;
5427 }
5428 
5429 /*
5430  * Add new bs contents at the top of an image chain while the chain is
5431  * live, while keeping required fields on the top layer.
5432  *
5433  * This will modify the BlockDriverState fields, and swap contents
5434  * between bs_new and bs_top. Both bs_new and bs_top are modified.
5435  *
5436  * bs_new must not be attached to a BlockBackend and must not have backing
5437  * child.
5438  *
5439  * This function does not create any image files.
5440  *
5441  * The caller must hold the AioContext lock for @bs_top.
5442  */
5443 int bdrv_append(BlockDriverState *bs_new, BlockDriverState *bs_top,
5444                 Error **errp)
5445 {
5446     int ret;
5447     BdrvChild *child;
5448     Transaction *tran = tran_new();
5449 
5450     GLOBAL_STATE_CODE();
5451 
5452     bdrv_graph_rdlock_main_loop();
5453     assert(!bs_new->backing);
5454     bdrv_graph_rdunlock_main_loop();
5455 
5456     bdrv_drained_begin(bs_top);
5457     bdrv_drained_begin(bs_new);
5458 
5459     bdrv_graph_wrlock();
5460 
5461     child = bdrv_attach_child_noperm(bs_new, bs_top, "backing",
5462                                      &child_of_bds, bdrv_backing_role(bs_new),
5463                                      tran, errp);
5464     if (!child) {
5465         ret = -EINVAL;
5466         goto out;
5467     }
5468 
5469     ret = bdrv_replace_node_noperm(bs_top, bs_new, true, tran, errp);
5470     if (ret < 0) {
5471         goto out;
5472     }
5473 
5474     ret = bdrv_refresh_perms(bs_new, tran, errp);
5475 out:
5476     tran_finalize(tran, ret);
5477 
5478     bdrv_refresh_limits(bs_top, NULL, NULL);
5479     bdrv_graph_wrunlock();
5480 
5481     bdrv_drained_end(bs_top);
5482     bdrv_drained_end(bs_new);
5483 
5484     return ret;
5485 }
5486 
5487 /* Not for empty child */
5488 int bdrv_replace_child_bs(BdrvChild *child, BlockDriverState *new_bs,
5489                           Error **errp)
5490 {
5491     int ret;
5492     Transaction *tran = tran_new();
5493     g_autoptr(GSList) refresh_list = NULL;
5494     BlockDriverState *old_bs = child->bs;
5495 
5496     GLOBAL_STATE_CODE();
5497 
5498     bdrv_ref(old_bs);
5499     bdrv_drained_begin(old_bs);
5500     bdrv_drained_begin(new_bs);
5501     bdrv_graph_wrlock();
5502 
5503     bdrv_replace_child_tran(child, new_bs, tran);
5504 
5505     refresh_list = g_slist_prepend(refresh_list, old_bs);
5506     refresh_list = g_slist_prepend(refresh_list, new_bs);
5507 
5508     ret = bdrv_list_refresh_perms(refresh_list, NULL, tran, errp);
5509 
5510     tran_finalize(tran, ret);
5511 
5512     bdrv_graph_wrunlock();
5513     bdrv_drained_end(old_bs);
5514     bdrv_drained_end(new_bs);
5515     bdrv_unref(old_bs);
5516 
5517     return ret;
5518 }
5519 
5520 static void bdrv_delete(BlockDriverState *bs)
5521 {
5522     assert(bdrv_op_blocker_is_empty(bs));
5523     assert(!bs->refcnt);
5524     GLOBAL_STATE_CODE();
5525 
5526     /* remove from list, if necessary */
5527     if (bs->node_name[0] != '\0') {
5528         QTAILQ_REMOVE(&graph_bdrv_states, bs, node_list);
5529     }
5530     QTAILQ_REMOVE(&all_bdrv_states, bs, bs_list);
5531 
5532     bdrv_close(bs);
5533 
5534     qemu_mutex_destroy(&bs->reqs_lock);
5535 
5536     g_free(bs);
5537 }
5538 
5539 
5540 /*
5541  * Replace @bs by newly created block node.
5542  *
5543  * @options is a QDict of options to pass to the block drivers, or NULL for an
5544  * empty set of options. The reference to the QDict belongs to the block layer
5545  * after the call (even on failure), so if the caller intends to reuse the
5546  * dictionary, it needs to use qobject_ref() before calling bdrv_open.
5547  *
5548  * The caller holds the AioContext lock for @bs. It must make sure that @bs
5549  * stays in the same AioContext, i.e. @options must not refer to nodes in a
5550  * different AioContext.
5551  */
5552 BlockDriverState *bdrv_insert_node(BlockDriverState *bs, QDict *options,
5553                                    int flags, Error **errp)
5554 {
5555     ERRP_GUARD();
5556     int ret;
5557     AioContext *ctx = bdrv_get_aio_context(bs);
5558     BlockDriverState *new_node_bs = NULL;
5559     const char *drvname, *node_name;
5560     BlockDriver *drv;
5561 
5562     drvname = qdict_get_try_str(options, "driver");
5563     if (!drvname) {
5564         error_setg(errp, "driver is not specified");
5565         goto fail;
5566     }
5567 
5568     drv = bdrv_find_format(drvname);
5569     if (!drv) {
5570         error_setg(errp, "Unknown driver: '%s'", drvname);
5571         goto fail;
5572     }
5573 
5574     node_name = qdict_get_try_str(options, "node-name");
5575 
5576     GLOBAL_STATE_CODE();
5577 
5578     new_node_bs = bdrv_new_open_driver_opts(drv, node_name, options, flags,
5579                                             errp);
5580     assert(bdrv_get_aio_context(bs) == ctx);
5581 
5582     options = NULL; /* bdrv_new_open_driver() eats options */
5583     if (!new_node_bs) {
5584         error_prepend(errp, "Could not create node: ");
5585         goto fail;
5586     }
5587 
5588     /*
5589      * Make sure that @bs doesn't go away until we have successfully attached
5590      * all of its parents to @new_node_bs and undrained it again.
5591      */
5592     bdrv_ref(bs);
5593     bdrv_drained_begin(bs);
5594     bdrv_drained_begin(new_node_bs);
5595     bdrv_graph_wrlock();
5596     ret = bdrv_replace_node(bs, new_node_bs, errp);
5597     bdrv_graph_wrunlock();
5598     bdrv_drained_end(new_node_bs);
5599     bdrv_drained_end(bs);
5600     bdrv_unref(bs);
5601 
5602     if (ret < 0) {
5603         error_prepend(errp, "Could not replace node: ");
5604         goto fail;
5605     }
5606 
5607     return new_node_bs;
5608 
5609 fail:
5610     qobject_unref(options);
5611     bdrv_unref(new_node_bs);
5612     return NULL;
5613 }
5614 
5615 /*
5616  * Run consistency checks on an image
5617  *
5618  * Returns 0 if the check could be completed (it doesn't mean that the image is
5619  * free of errors) or -errno when an internal error occurred. The results of the
5620  * check are stored in res.
5621  */
5622 int coroutine_fn bdrv_co_check(BlockDriverState *bs,
5623                                BdrvCheckResult *res, BdrvCheckMode fix)
5624 {
5625     IO_CODE();
5626     assert_bdrv_graph_readable();
5627     if (bs->drv == NULL) {
5628         return -ENOMEDIUM;
5629     }
5630     if (bs->drv->bdrv_co_check == NULL) {
5631         return -ENOTSUP;
5632     }
5633 
5634     memset(res, 0, sizeof(*res));
5635     return bs->drv->bdrv_co_check(bs, res, fix);
5636 }
5637 
5638 /*
5639  * Return values:
5640  * 0        - success
5641  * -EINVAL  - backing format specified, but no file
5642  * -ENOSPC  - can't update the backing file because no space is left in the
5643  *            image file header
5644  * -ENOTSUP - format driver doesn't support changing the backing file
5645  */
5646 int coroutine_fn
5647 bdrv_co_change_backing_file(BlockDriverState *bs, const char *backing_file,
5648                             const char *backing_fmt, bool require)
5649 {
5650     BlockDriver *drv = bs->drv;
5651     int ret;
5652 
5653     IO_CODE();
5654 
5655     if (!drv) {
5656         return -ENOMEDIUM;
5657     }
5658 
5659     /* Backing file format doesn't make sense without a backing file */
5660     if (backing_fmt && !backing_file) {
5661         return -EINVAL;
5662     }
5663 
5664     if (require && backing_file && !backing_fmt) {
5665         return -EINVAL;
5666     }
5667 
5668     if (drv->bdrv_co_change_backing_file != NULL) {
5669         ret = drv->bdrv_co_change_backing_file(bs, backing_file, backing_fmt);
5670     } else {
5671         ret = -ENOTSUP;
5672     }
5673 
5674     if (ret == 0) {
5675         pstrcpy(bs->backing_file, sizeof(bs->backing_file), backing_file ?: "");
5676         pstrcpy(bs->backing_format, sizeof(bs->backing_format), backing_fmt ?: "");
5677         pstrcpy(bs->auto_backing_file, sizeof(bs->auto_backing_file),
5678                 backing_file ?: "");
5679     }
5680     return ret;
5681 }
5682 
5683 /*
5684  * Finds the first non-filter node above bs in the chain between
5685  * active and bs.  The returned node is either an immediate parent of
5686  * bs, or there are only filter nodes between the two.
5687  *
5688  * Returns NULL if bs is not found in active's image chain,
5689  * or if active == bs.
5690  *
5691  * Returns the bottommost base image if bs == NULL.
5692  */
5693 BlockDriverState *bdrv_find_overlay(BlockDriverState *active,
5694                                     BlockDriverState *bs)
5695 {
5696 
5697     GLOBAL_STATE_CODE();
5698 
5699     bs = bdrv_skip_filters(bs);
5700     active = bdrv_skip_filters(active);
5701 
5702     while (active) {
5703         BlockDriverState *next = bdrv_backing_chain_next(active);
5704         if (bs == next) {
5705             return active;
5706         }
5707         active = next;
5708     }
5709 
5710     return NULL;
5711 }
5712 
5713 /* Given a BDS, searches for the base layer. */
5714 BlockDriverState *bdrv_find_base(BlockDriverState *bs)
5715 {
5716     GLOBAL_STATE_CODE();
5717 
5718     return bdrv_find_overlay(bs, NULL);
5719 }
5720 
5721 /*
5722  * Return true if at least one of the COW (backing) and filter links
5723  * between @bs and @base is frozen. @errp is set if that's the case.
5724  * @base must be reachable from @bs, or NULL.
5725  */
5726 static bool GRAPH_RDLOCK
5727 bdrv_is_backing_chain_frozen(BlockDriverState *bs, BlockDriverState *base,
5728                              Error **errp)
5729 {
5730     BlockDriverState *i;
5731     BdrvChild *child;
5732 
5733     GLOBAL_STATE_CODE();
5734 
5735     for (i = bs; i != base; i = child_bs(child)) {
5736         child = bdrv_filter_or_cow_child(i);
5737 
5738         if (child && child->frozen) {
5739             error_setg(errp, "Cannot change '%s' link from '%s' to '%s'",
5740                        child->name, i->node_name, child->bs->node_name);
5741             return true;
5742         }
5743     }
5744 
5745     return false;
5746 }
5747 
5748 /*
5749  * Freeze all COW (backing) and filter links between @bs and @base.
5750  * If any of the links is already frozen the operation is aborted and
5751  * none of the links are modified.
5752  * @base must be reachable from @bs, or NULL.
5753  * Returns 0 on success. On failure returns < 0 and sets @errp.
5754  */
5755 int bdrv_freeze_backing_chain(BlockDriverState *bs, BlockDriverState *base,
5756                               Error **errp)
5757 {
5758     BlockDriverState *i;
5759     BdrvChild *child;
5760 
5761     GLOBAL_STATE_CODE();
5762 
5763     if (bdrv_is_backing_chain_frozen(bs, base, errp)) {
5764         return -EPERM;
5765     }
5766 
5767     for (i = bs; i != base; i = child_bs(child)) {
5768         child = bdrv_filter_or_cow_child(i);
5769         if (child && child->bs->never_freeze) {
5770             error_setg(errp, "Cannot freeze '%s' link to '%s'",
5771                        child->name, child->bs->node_name);
5772             return -EPERM;
5773         }
5774     }
5775 
5776     for (i = bs; i != base; i = child_bs(child)) {
5777         child = bdrv_filter_or_cow_child(i);
5778         if (child) {
5779             child->frozen = true;
5780         }
5781     }
5782 
5783     return 0;
5784 }
5785 
5786 /*
5787  * Unfreeze all COW (backing) and filter links between @bs and @base.
5788  * The caller must ensure that all links are frozen before using this
5789  * function.
5790  * @base must be reachable from @bs, or NULL.
5791  */
5792 void bdrv_unfreeze_backing_chain(BlockDriverState *bs, BlockDriverState *base)
5793 {
5794     BlockDriverState *i;
5795     BdrvChild *child;
5796 
5797     GLOBAL_STATE_CODE();
5798 
5799     for (i = bs; i != base; i = child_bs(child)) {
5800         child = bdrv_filter_or_cow_child(i);
5801         if (child) {
5802             assert(child->frozen);
5803             child->frozen = false;
5804         }
5805     }
5806 }
5807 
5808 /*
5809  * Drops images above 'base' up to and including 'top', and sets the image
5810  * above 'top' to have base as its backing file.
5811  *
5812  * Requires that the overlay to 'top' is opened r/w, so that the backing file
5813  * information in 'bs' can be properly updated.
5814  *
5815  * E.g., this will convert the following chain:
5816  * bottom <- base <- intermediate <- top <- active
5817  *
5818  * to
5819  *
5820  * bottom <- base <- active
5821  *
5822  * It is allowed for bottom==base, in which case it converts:
5823  *
5824  * base <- intermediate <- top <- active
5825  *
5826  * to
5827  *
5828  * base <- active
5829  *
5830  * If backing_file_str is non-NULL, it will be used when modifying top's
5831  * overlay image metadata.
5832  *
5833  * Error conditions:
5834  *  if active == top, that is considered an error
5835  *
5836  */
5837 int bdrv_drop_intermediate(BlockDriverState *top, BlockDriverState *base,
5838                            const char *backing_file_str)
5839 {
5840     BlockDriverState *explicit_top = top;
5841     bool update_inherits_from;
5842     BdrvChild *c;
5843     Error *local_err = NULL;
5844     int ret = -EIO;
5845     g_autoptr(GSList) updated_children = NULL;
5846     GSList *p;
5847 
5848     GLOBAL_STATE_CODE();
5849 
5850     bdrv_ref(top);
5851     bdrv_drained_begin(base);
5852     bdrv_graph_wrlock();
5853 
5854     if (!top->drv || !base->drv) {
5855         goto exit_wrlock;
5856     }
5857 
5858     /* Make sure that base is in the backing chain of top */
5859     if (!bdrv_chain_contains(top, base)) {
5860         goto exit_wrlock;
5861     }
5862 
5863     /* If 'base' recursively inherits from 'top' then we should set
5864      * base->inherits_from to top->inherits_from after 'top' and all
5865      * other intermediate nodes have been dropped.
5866      * If 'top' is an implicit node (e.g. "commit_top") we should skip
5867      * it because no one inherits from it. We use explicit_top for that. */
5868     explicit_top = bdrv_skip_implicit_filters(explicit_top);
5869     update_inherits_from = bdrv_inherits_from_recursive(base, explicit_top);
5870 
5871     /* success - we can delete the intermediate states, and link top->base */
5872     if (!backing_file_str) {
5873         bdrv_refresh_filename(base);
5874         backing_file_str = base->filename;
5875     }
5876 
5877     QLIST_FOREACH(c, &top->parents, next_parent) {
5878         updated_children = g_slist_prepend(updated_children, c);
5879     }
5880 
5881     /*
5882      * It seems correct to pass detach_subchain=true here, but it triggers
5883      * one more yet not fixed bug, when due to nested aio_poll loop we switch to
5884      * another drained section, which modify the graph (for example, removing
5885      * the child, which we keep in updated_children list). So, it's a TODO.
5886      *
5887      * Note, bug triggered if pass detach_subchain=true here and run
5888      * test-bdrv-drain. test_drop_intermediate_poll() test-case will crash.
5889      * That's a FIXME.
5890      */
5891     bdrv_replace_node_common(top, base, false, false, &local_err);
5892     bdrv_graph_wrunlock();
5893 
5894     if (local_err) {
5895         error_report_err(local_err);
5896         goto exit;
5897     }
5898 
5899     for (p = updated_children; p; p = p->next) {
5900         c = p->data;
5901 
5902         if (c->klass->update_filename) {
5903             ret = c->klass->update_filename(c, base, backing_file_str,
5904                                             &local_err);
5905             if (ret < 0) {
5906                 /*
5907                  * TODO: Actually, we want to rollback all previous iterations
5908                  * of this loop, and (which is almost impossible) previous
5909                  * bdrv_replace_node()...
5910                  *
5911                  * Note, that c->klass->update_filename may lead to permission
5912                  * update, so it's a bad idea to call it inside permission
5913                  * update transaction of bdrv_replace_node.
5914                  */
5915                 error_report_err(local_err);
5916                 goto exit;
5917             }
5918         }
5919     }
5920 
5921     if (update_inherits_from) {
5922         base->inherits_from = explicit_top->inherits_from;
5923     }
5924 
5925     ret = 0;
5926     goto exit;
5927 
5928 exit_wrlock:
5929     bdrv_graph_wrunlock();
5930 exit:
5931     bdrv_drained_end(base);
5932     bdrv_unref(top);
5933     return ret;
5934 }
5935 
5936 /**
5937  * Implementation of BlockDriver.bdrv_co_get_allocated_file_size() that
5938  * sums the size of all data-bearing children.  (This excludes backing
5939  * children.)
5940  */
5941 static int64_t coroutine_fn GRAPH_RDLOCK
5942 bdrv_sum_allocated_file_size(BlockDriverState *bs)
5943 {
5944     BdrvChild *child;
5945     int64_t child_size, sum = 0;
5946 
5947     QLIST_FOREACH(child, &bs->children, next) {
5948         if (child->role & (BDRV_CHILD_DATA | BDRV_CHILD_METADATA |
5949                            BDRV_CHILD_FILTERED))
5950         {
5951             child_size = bdrv_co_get_allocated_file_size(child->bs);
5952             if (child_size < 0) {
5953                 return child_size;
5954             }
5955             sum += child_size;
5956         }
5957     }
5958 
5959     return sum;
5960 }
5961 
5962 /**
5963  * Length of a allocated file in bytes. Sparse files are counted by actual
5964  * allocated space. Return < 0 if error or unknown.
5965  */
5966 int64_t coroutine_fn bdrv_co_get_allocated_file_size(BlockDriverState *bs)
5967 {
5968     BlockDriver *drv = bs->drv;
5969     IO_CODE();
5970     assert_bdrv_graph_readable();
5971 
5972     if (!drv) {
5973         return -ENOMEDIUM;
5974     }
5975     if (drv->bdrv_co_get_allocated_file_size) {
5976         return drv->bdrv_co_get_allocated_file_size(bs);
5977     }
5978 
5979     if (drv->bdrv_file_open) {
5980         /*
5981          * Protocol drivers default to -ENOTSUP (most of their data is
5982          * not stored in any of their children (if they even have any),
5983          * so there is no generic way to figure it out).
5984          */
5985         return -ENOTSUP;
5986     } else if (drv->is_filter) {
5987         /* Filter drivers default to the size of their filtered child */
5988         return bdrv_co_get_allocated_file_size(bdrv_filter_bs(bs));
5989     } else {
5990         /* Other drivers default to summing their children's sizes */
5991         return bdrv_sum_allocated_file_size(bs);
5992     }
5993 }
5994 
5995 /*
5996  * bdrv_measure:
5997  * @drv: Format driver
5998  * @opts: Creation options for new image
5999  * @in_bs: Existing image containing data for new image (may be NULL)
6000  * @errp: Error object
6001  * Returns: A #BlockMeasureInfo (free using qapi_free_BlockMeasureInfo())
6002  *          or NULL on error
6003  *
6004  * Calculate file size required to create a new image.
6005  *
6006  * If @in_bs is given then space for allocated clusters and zero clusters
6007  * from that image are included in the calculation.  If @opts contains a
6008  * backing file that is shared by @in_bs then backing clusters may be omitted
6009  * from the calculation.
6010  *
6011  * If @in_bs is NULL then the calculation includes no allocated clusters
6012  * unless a preallocation option is given in @opts.
6013  *
6014  * Note that @in_bs may use a different BlockDriver from @drv.
6015  *
6016  * If an error occurs the @errp pointer is set.
6017  */
6018 BlockMeasureInfo *bdrv_measure(BlockDriver *drv, QemuOpts *opts,
6019                                BlockDriverState *in_bs, Error **errp)
6020 {
6021     IO_CODE();
6022     if (!drv->bdrv_measure) {
6023         error_setg(errp, "Block driver '%s' does not support size measurement",
6024                    drv->format_name);
6025         return NULL;
6026     }
6027 
6028     return drv->bdrv_measure(opts, in_bs, errp);
6029 }
6030 
6031 /**
6032  * Return number of sectors on success, -errno on error.
6033  */
6034 int64_t coroutine_fn bdrv_co_nb_sectors(BlockDriverState *bs)
6035 {
6036     BlockDriver *drv = bs->drv;
6037     IO_CODE();
6038     assert_bdrv_graph_readable();
6039 
6040     if (!drv)
6041         return -ENOMEDIUM;
6042 
6043     if (bs->bl.has_variable_length) {
6044         int ret = bdrv_co_refresh_total_sectors(bs, bs->total_sectors);
6045         if (ret < 0) {
6046             return ret;
6047         }
6048     }
6049     return bs->total_sectors;
6050 }
6051 
6052 /*
6053  * This wrapper is written by hand because this function is in the hot I/O path,
6054  * via blk_get_geometry.
6055  */
6056 int64_t coroutine_mixed_fn bdrv_nb_sectors(BlockDriverState *bs)
6057 {
6058     BlockDriver *drv = bs->drv;
6059     IO_CODE();
6060 
6061     if (!drv)
6062         return -ENOMEDIUM;
6063 
6064     if (bs->bl.has_variable_length) {
6065         int ret = bdrv_refresh_total_sectors(bs, bs->total_sectors);
6066         if (ret < 0) {
6067             return ret;
6068         }
6069     }
6070 
6071     return bs->total_sectors;
6072 }
6073 
6074 /**
6075  * Return length in bytes on success, -errno on error.
6076  * The length is always a multiple of BDRV_SECTOR_SIZE.
6077  */
6078 int64_t coroutine_fn bdrv_co_getlength(BlockDriverState *bs)
6079 {
6080     int64_t ret;
6081     IO_CODE();
6082     assert_bdrv_graph_readable();
6083 
6084     ret = bdrv_co_nb_sectors(bs);
6085     if (ret < 0) {
6086         return ret;
6087     }
6088     if (ret > INT64_MAX / BDRV_SECTOR_SIZE) {
6089         return -EFBIG;
6090     }
6091     return ret * BDRV_SECTOR_SIZE;
6092 }
6093 
6094 bool bdrv_is_sg(BlockDriverState *bs)
6095 {
6096     IO_CODE();
6097     return bs->sg;
6098 }
6099 
6100 /**
6101  * Return whether the given node supports compressed writes.
6102  */
6103 bool bdrv_supports_compressed_writes(BlockDriverState *bs)
6104 {
6105     BlockDriverState *filtered;
6106     IO_CODE();
6107 
6108     if (!bs->drv || !block_driver_can_compress(bs->drv)) {
6109         return false;
6110     }
6111 
6112     filtered = bdrv_filter_bs(bs);
6113     if (filtered) {
6114         /*
6115          * Filters can only forward compressed writes, so we have to
6116          * check the child.
6117          */
6118         return bdrv_supports_compressed_writes(filtered);
6119     }
6120 
6121     return true;
6122 }
6123 
6124 const char *bdrv_get_format_name(BlockDriverState *bs)
6125 {
6126     IO_CODE();
6127     return bs->drv ? bs->drv->format_name : NULL;
6128 }
6129 
6130 static int qsort_strcmp(const void *a, const void *b)
6131 {
6132     return strcmp(*(char *const *)a, *(char *const *)b);
6133 }
6134 
6135 void bdrv_iterate_format(void (*it)(void *opaque, const char *name),
6136                          void *opaque, bool read_only)
6137 {
6138     BlockDriver *drv;
6139     int count = 0;
6140     int i;
6141     const char **formats = NULL;
6142 
6143     GLOBAL_STATE_CODE();
6144 
6145     QLIST_FOREACH(drv, &bdrv_drivers, list) {
6146         if (drv->format_name) {
6147             bool found = false;
6148 
6149             if (use_bdrv_whitelist && !bdrv_is_whitelisted(drv, read_only)) {
6150                 continue;
6151             }
6152 
6153             i = count;
6154             while (formats && i && !found) {
6155                 found = !strcmp(formats[--i], drv->format_name);
6156             }
6157 
6158             if (!found) {
6159                 formats = g_renew(const char *, formats, count + 1);
6160                 formats[count++] = drv->format_name;
6161             }
6162         }
6163     }
6164 
6165     for (i = 0; i < (int)ARRAY_SIZE(block_driver_modules); i++) {
6166         const char *format_name = block_driver_modules[i].format_name;
6167 
6168         if (format_name) {
6169             bool found = false;
6170             int j = count;
6171 
6172             if (use_bdrv_whitelist &&
6173                 !bdrv_format_is_whitelisted(format_name, read_only)) {
6174                 continue;
6175             }
6176 
6177             while (formats && j && !found) {
6178                 found = !strcmp(formats[--j], format_name);
6179             }
6180 
6181             if (!found) {
6182                 formats = g_renew(const char *, formats, count + 1);
6183                 formats[count++] = format_name;
6184             }
6185         }
6186     }
6187 
6188     qsort(formats, count, sizeof(formats[0]), qsort_strcmp);
6189 
6190     for (i = 0; i < count; i++) {
6191         it(opaque, formats[i]);
6192     }
6193 
6194     g_free(formats);
6195 }
6196 
6197 /* This function is to find a node in the bs graph */
6198 BlockDriverState *bdrv_find_node(const char *node_name)
6199 {
6200     BlockDriverState *bs;
6201 
6202     assert(node_name);
6203     GLOBAL_STATE_CODE();
6204 
6205     QTAILQ_FOREACH(bs, &graph_bdrv_states, node_list) {
6206         if (!strcmp(node_name, bs->node_name)) {
6207             return bs;
6208         }
6209     }
6210     return NULL;
6211 }
6212 
6213 /* Put this QMP function here so it can access the static graph_bdrv_states. */
6214 BlockDeviceInfoList *bdrv_named_nodes_list(bool flat,
6215                                            Error **errp)
6216 {
6217     BlockDeviceInfoList *list;
6218     BlockDriverState *bs;
6219 
6220     GLOBAL_STATE_CODE();
6221     GRAPH_RDLOCK_GUARD_MAINLOOP();
6222 
6223     list = NULL;
6224     QTAILQ_FOREACH(bs, &graph_bdrv_states, node_list) {
6225         BlockDeviceInfo *info = bdrv_block_device_info(NULL, bs, flat, errp);
6226         if (!info) {
6227             qapi_free_BlockDeviceInfoList(list);
6228             return NULL;
6229         }
6230         QAPI_LIST_PREPEND(list, info);
6231     }
6232 
6233     return list;
6234 }
6235 
6236 typedef struct XDbgBlockGraphConstructor {
6237     XDbgBlockGraph *graph;
6238     GHashTable *graph_nodes;
6239 } XDbgBlockGraphConstructor;
6240 
6241 static XDbgBlockGraphConstructor *xdbg_graph_new(void)
6242 {
6243     XDbgBlockGraphConstructor *gr = g_new(XDbgBlockGraphConstructor, 1);
6244 
6245     gr->graph = g_new0(XDbgBlockGraph, 1);
6246     gr->graph_nodes = g_hash_table_new(NULL, NULL);
6247 
6248     return gr;
6249 }
6250 
6251 static XDbgBlockGraph *xdbg_graph_finalize(XDbgBlockGraphConstructor *gr)
6252 {
6253     XDbgBlockGraph *graph = gr->graph;
6254 
6255     g_hash_table_destroy(gr->graph_nodes);
6256     g_free(gr);
6257 
6258     return graph;
6259 }
6260 
6261 static uintptr_t xdbg_graph_node_num(XDbgBlockGraphConstructor *gr, void *node)
6262 {
6263     uintptr_t ret = (uintptr_t)g_hash_table_lookup(gr->graph_nodes, node);
6264 
6265     if (ret != 0) {
6266         return ret;
6267     }
6268 
6269     /*
6270      * Start counting from 1, not 0, because 0 interferes with not-found (NULL)
6271      * answer of g_hash_table_lookup.
6272      */
6273     ret = g_hash_table_size(gr->graph_nodes) + 1;
6274     g_hash_table_insert(gr->graph_nodes, node, (void *)ret);
6275 
6276     return ret;
6277 }
6278 
6279 static void xdbg_graph_add_node(XDbgBlockGraphConstructor *gr, void *node,
6280                                 XDbgBlockGraphNodeType type, const char *name)
6281 {
6282     XDbgBlockGraphNode *n;
6283 
6284     n = g_new0(XDbgBlockGraphNode, 1);
6285 
6286     n->id = xdbg_graph_node_num(gr, node);
6287     n->type = type;
6288     n->name = g_strdup(name);
6289 
6290     QAPI_LIST_PREPEND(gr->graph->nodes, n);
6291 }
6292 
6293 static void xdbg_graph_add_edge(XDbgBlockGraphConstructor *gr, void *parent,
6294                                 const BdrvChild *child)
6295 {
6296     BlockPermission qapi_perm;
6297     XDbgBlockGraphEdge *edge;
6298     GLOBAL_STATE_CODE();
6299 
6300     edge = g_new0(XDbgBlockGraphEdge, 1);
6301 
6302     edge->parent = xdbg_graph_node_num(gr, parent);
6303     edge->child = xdbg_graph_node_num(gr, child->bs);
6304     edge->name = g_strdup(child->name);
6305 
6306     for (qapi_perm = 0; qapi_perm < BLOCK_PERMISSION__MAX; qapi_perm++) {
6307         uint64_t flag = bdrv_qapi_perm_to_blk_perm(qapi_perm);
6308 
6309         if (flag & child->perm) {
6310             QAPI_LIST_PREPEND(edge->perm, qapi_perm);
6311         }
6312         if (flag & child->shared_perm) {
6313             QAPI_LIST_PREPEND(edge->shared_perm, qapi_perm);
6314         }
6315     }
6316 
6317     QAPI_LIST_PREPEND(gr->graph->edges, edge);
6318 }
6319 
6320 
6321 XDbgBlockGraph *bdrv_get_xdbg_block_graph(Error **errp)
6322 {
6323     BlockBackend *blk;
6324     BlockJob *job;
6325     BlockDriverState *bs;
6326     BdrvChild *child;
6327     XDbgBlockGraphConstructor *gr = xdbg_graph_new();
6328 
6329     GLOBAL_STATE_CODE();
6330 
6331     for (blk = blk_all_next(NULL); blk; blk = blk_all_next(blk)) {
6332         char *allocated_name = NULL;
6333         const char *name = blk_name(blk);
6334 
6335         if (!*name) {
6336             name = allocated_name = blk_get_attached_dev_id(blk);
6337         }
6338         xdbg_graph_add_node(gr, blk, X_DBG_BLOCK_GRAPH_NODE_TYPE_BLOCK_BACKEND,
6339                            name);
6340         g_free(allocated_name);
6341         if (blk_root(blk)) {
6342             xdbg_graph_add_edge(gr, blk, blk_root(blk));
6343         }
6344     }
6345 
6346     WITH_JOB_LOCK_GUARD() {
6347         for (job = block_job_next_locked(NULL); job;
6348              job = block_job_next_locked(job)) {
6349             GSList *el;
6350 
6351             xdbg_graph_add_node(gr, job, X_DBG_BLOCK_GRAPH_NODE_TYPE_BLOCK_JOB,
6352                                 job->job.id);
6353             for (el = job->nodes; el; el = el->next) {
6354                 xdbg_graph_add_edge(gr, job, (BdrvChild *)el->data);
6355             }
6356         }
6357     }
6358 
6359     QTAILQ_FOREACH(bs, &graph_bdrv_states, node_list) {
6360         xdbg_graph_add_node(gr, bs, X_DBG_BLOCK_GRAPH_NODE_TYPE_BLOCK_DRIVER,
6361                            bs->node_name);
6362         QLIST_FOREACH(child, &bs->children, next) {
6363             xdbg_graph_add_edge(gr, bs, child);
6364         }
6365     }
6366 
6367     return xdbg_graph_finalize(gr);
6368 }
6369 
6370 BlockDriverState *bdrv_lookup_bs(const char *device,
6371                                  const char *node_name,
6372                                  Error **errp)
6373 {
6374     BlockBackend *blk;
6375     BlockDriverState *bs;
6376 
6377     GLOBAL_STATE_CODE();
6378 
6379     if (device) {
6380         blk = blk_by_name(device);
6381 
6382         if (blk) {
6383             bs = blk_bs(blk);
6384             if (!bs) {
6385                 error_setg(errp, "Device '%s' has no medium", device);
6386             }
6387 
6388             return bs;
6389         }
6390     }
6391 
6392     if (node_name) {
6393         bs = bdrv_find_node(node_name);
6394 
6395         if (bs) {
6396             return bs;
6397         }
6398     }
6399 
6400     error_setg(errp, "Cannot find device=\'%s\' nor node-name=\'%s\'",
6401                      device ? device : "",
6402                      node_name ? node_name : "");
6403     return NULL;
6404 }
6405 
6406 /* If 'base' is in the same chain as 'top', return true. Otherwise,
6407  * return false.  If either argument is NULL, return false. */
6408 bool bdrv_chain_contains(BlockDriverState *top, BlockDriverState *base)
6409 {
6410 
6411     GLOBAL_STATE_CODE();
6412 
6413     while (top && top != base) {
6414         top = bdrv_filter_or_cow_bs(top);
6415     }
6416 
6417     return top != NULL;
6418 }
6419 
6420 BlockDriverState *bdrv_next_node(BlockDriverState *bs)
6421 {
6422     GLOBAL_STATE_CODE();
6423     if (!bs) {
6424         return QTAILQ_FIRST(&graph_bdrv_states);
6425     }
6426     return QTAILQ_NEXT(bs, node_list);
6427 }
6428 
6429 BlockDriverState *bdrv_next_all_states(BlockDriverState *bs)
6430 {
6431     GLOBAL_STATE_CODE();
6432     if (!bs) {
6433         return QTAILQ_FIRST(&all_bdrv_states);
6434     }
6435     return QTAILQ_NEXT(bs, bs_list);
6436 }
6437 
6438 const char *bdrv_get_node_name(const BlockDriverState *bs)
6439 {
6440     IO_CODE();
6441     return bs->node_name;
6442 }
6443 
6444 const char *bdrv_get_parent_name(const BlockDriverState *bs)
6445 {
6446     BdrvChild *c;
6447     const char *name;
6448     IO_CODE();
6449 
6450     /* If multiple parents have a name, just pick the first one. */
6451     QLIST_FOREACH(c, &bs->parents, next_parent) {
6452         if (c->klass->get_name) {
6453             name = c->klass->get_name(c);
6454             if (name && *name) {
6455                 return name;
6456             }
6457         }
6458     }
6459 
6460     return NULL;
6461 }
6462 
6463 /* TODO check what callers really want: bs->node_name or blk_name() */
6464 const char *bdrv_get_device_name(const BlockDriverState *bs)
6465 {
6466     IO_CODE();
6467     return bdrv_get_parent_name(bs) ?: "";
6468 }
6469 
6470 /* This can be used to identify nodes that might not have a device
6471  * name associated. Since node and device names live in the same
6472  * namespace, the result is unambiguous. The exception is if both are
6473  * absent, then this returns an empty (non-null) string. */
6474 const char *bdrv_get_device_or_node_name(const BlockDriverState *bs)
6475 {
6476     IO_CODE();
6477     return bdrv_get_parent_name(bs) ?: bs->node_name;
6478 }
6479 
6480 int bdrv_get_flags(BlockDriverState *bs)
6481 {
6482     IO_CODE();
6483     return bs->open_flags;
6484 }
6485 
6486 int bdrv_has_zero_init_1(BlockDriverState *bs)
6487 {
6488     GLOBAL_STATE_CODE();
6489     return 1;
6490 }
6491 
6492 int coroutine_mixed_fn bdrv_has_zero_init(BlockDriverState *bs)
6493 {
6494     BlockDriverState *filtered;
6495     GLOBAL_STATE_CODE();
6496 
6497     if (!bs->drv) {
6498         return 0;
6499     }
6500 
6501     /* If BS is a copy on write image, it is initialized to
6502        the contents of the base image, which may not be zeroes.  */
6503     if (bdrv_cow_child(bs)) {
6504         return 0;
6505     }
6506     if (bs->drv->bdrv_has_zero_init) {
6507         return bs->drv->bdrv_has_zero_init(bs);
6508     }
6509 
6510     filtered = bdrv_filter_bs(bs);
6511     if (filtered) {
6512         return bdrv_has_zero_init(filtered);
6513     }
6514 
6515     /* safe default */
6516     return 0;
6517 }
6518 
6519 bool bdrv_can_write_zeroes_with_unmap(BlockDriverState *bs)
6520 {
6521     IO_CODE();
6522     if (!(bs->open_flags & BDRV_O_UNMAP)) {
6523         return false;
6524     }
6525 
6526     return bs->supported_zero_flags & BDRV_REQ_MAY_UNMAP;
6527 }
6528 
6529 void bdrv_get_backing_filename(BlockDriverState *bs,
6530                                char *filename, int filename_size)
6531 {
6532     IO_CODE();
6533     pstrcpy(filename, filename_size, bs->backing_file);
6534 }
6535 
6536 int coroutine_fn bdrv_co_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
6537 {
6538     int ret;
6539     BlockDriver *drv = bs->drv;
6540     IO_CODE();
6541     assert_bdrv_graph_readable();
6542 
6543     /* if bs->drv == NULL, bs is closed, so there's nothing to do here */
6544     if (!drv) {
6545         return -ENOMEDIUM;
6546     }
6547     if (!drv->bdrv_co_get_info) {
6548         BlockDriverState *filtered = bdrv_filter_bs(bs);
6549         if (filtered) {
6550             return bdrv_co_get_info(filtered, bdi);
6551         }
6552         return -ENOTSUP;
6553     }
6554     memset(bdi, 0, sizeof(*bdi));
6555     ret = drv->bdrv_co_get_info(bs, bdi);
6556     if (bdi->subcluster_size == 0) {
6557         /*
6558          * If the driver left this unset, subclusters are not supported.
6559          * Then it is safe to treat each cluster as having only one subcluster.
6560          */
6561         bdi->subcluster_size = bdi->cluster_size;
6562     }
6563     if (ret < 0) {
6564         return ret;
6565     }
6566 
6567     if (bdi->cluster_size > BDRV_MAX_ALIGNMENT) {
6568         return -EINVAL;
6569     }
6570 
6571     return 0;
6572 }
6573 
6574 ImageInfoSpecific *bdrv_get_specific_info(BlockDriverState *bs,
6575                                           Error **errp)
6576 {
6577     BlockDriver *drv = bs->drv;
6578     IO_CODE();
6579     if (drv && drv->bdrv_get_specific_info) {
6580         return drv->bdrv_get_specific_info(bs, errp);
6581     }
6582     return NULL;
6583 }
6584 
6585 BlockStatsSpecific *bdrv_get_specific_stats(BlockDriverState *bs)
6586 {
6587     BlockDriver *drv = bs->drv;
6588     IO_CODE();
6589     if (!drv || !drv->bdrv_get_specific_stats) {
6590         return NULL;
6591     }
6592     return drv->bdrv_get_specific_stats(bs);
6593 }
6594 
6595 void coroutine_fn bdrv_co_debug_event(BlockDriverState *bs, BlkdebugEvent event)
6596 {
6597     IO_CODE();
6598     assert_bdrv_graph_readable();
6599 
6600     if (!bs || !bs->drv || !bs->drv->bdrv_co_debug_event) {
6601         return;
6602     }
6603 
6604     bs->drv->bdrv_co_debug_event(bs, event);
6605 }
6606 
6607 static BlockDriverState * GRAPH_RDLOCK
6608 bdrv_find_debug_node(BlockDriverState *bs)
6609 {
6610     GLOBAL_STATE_CODE();
6611     while (bs && bs->drv && !bs->drv->bdrv_debug_breakpoint) {
6612         bs = bdrv_primary_bs(bs);
6613     }
6614 
6615     if (bs && bs->drv && bs->drv->bdrv_debug_breakpoint) {
6616         assert(bs->drv->bdrv_debug_remove_breakpoint);
6617         return bs;
6618     }
6619 
6620     return NULL;
6621 }
6622 
6623 int bdrv_debug_breakpoint(BlockDriverState *bs, const char *event,
6624                           const char *tag)
6625 {
6626     GLOBAL_STATE_CODE();
6627     GRAPH_RDLOCK_GUARD_MAINLOOP();
6628 
6629     bs = bdrv_find_debug_node(bs);
6630     if (bs) {
6631         return bs->drv->bdrv_debug_breakpoint(bs, event, tag);
6632     }
6633 
6634     return -ENOTSUP;
6635 }
6636 
6637 int bdrv_debug_remove_breakpoint(BlockDriverState *bs, const char *tag)
6638 {
6639     GLOBAL_STATE_CODE();
6640     GRAPH_RDLOCK_GUARD_MAINLOOP();
6641 
6642     bs = bdrv_find_debug_node(bs);
6643     if (bs) {
6644         return bs->drv->bdrv_debug_remove_breakpoint(bs, tag);
6645     }
6646 
6647     return -ENOTSUP;
6648 }
6649 
6650 int bdrv_debug_resume(BlockDriverState *bs, const char *tag)
6651 {
6652     GLOBAL_STATE_CODE();
6653     GRAPH_RDLOCK_GUARD_MAINLOOP();
6654 
6655     while (bs && (!bs->drv || !bs->drv->bdrv_debug_resume)) {
6656         bs = bdrv_primary_bs(bs);
6657     }
6658 
6659     if (bs && bs->drv && bs->drv->bdrv_debug_resume) {
6660         return bs->drv->bdrv_debug_resume(bs, tag);
6661     }
6662 
6663     return -ENOTSUP;
6664 }
6665 
6666 bool bdrv_debug_is_suspended(BlockDriverState *bs, const char *tag)
6667 {
6668     GLOBAL_STATE_CODE();
6669     GRAPH_RDLOCK_GUARD_MAINLOOP();
6670 
6671     while (bs && bs->drv && !bs->drv->bdrv_debug_is_suspended) {
6672         bs = bdrv_primary_bs(bs);
6673     }
6674 
6675     if (bs && bs->drv && bs->drv->bdrv_debug_is_suspended) {
6676         return bs->drv->bdrv_debug_is_suspended(bs, tag);
6677     }
6678 
6679     return false;
6680 }
6681 
6682 /* backing_file can either be relative, or absolute, or a protocol.  If it is
6683  * relative, it must be relative to the chain.  So, passing in bs->filename
6684  * from a BDS as backing_file should not be done, as that may be relative to
6685  * the CWD rather than the chain. */
6686 BlockDriverState *bdrv_find_backing_image(BlockDriverState *bs,
6687         const char *backing_file)
6688 {
6689     char *filename_full = NULL;
6690     char *backing_file_full = NULL;
6691     char *filename_tmp = NULL;
6692     int is_protocol = 0;
6693     bool filenames_refreshed = false;
6694     BlockDriverState *curr_bs = NULL;
6695     BlockDriverState *retval = NULL;
6696     BlockDriverState *bs_below;
6697 
6698     GLOBAL_STATE_CODE();
6699     GRAPH_RDLOCK_GUARD_MAINLOOP();
6700 
6701     if (!bs || !bs->drv || !backing_file) {
6702         return NULL;
6703     }
6704 
6705     filename_full     = g_malloc(PATH_MAX);
6706     backing_file_full = g_malloc(PATH_MAX);
6707 
6708     is_protocol = path_has_protocol(backing_file);
6709 
6710     /*
6711      * Being largely a legacy function, skip any filters here
6712      * (because filters do not have normal filenames, so they cannot
6713      * match anyway; and allowing json:{} filenames is a bit out of
6714      * scope).
6715      */
6716     for (curr_bs = bdrv_skip_filters(bs);
6717          bdrv_cow_child(curr_bs) != NULL;
6718          curr_bs = bs_below)
6719     {
6720         bs_below = bdrv_backing_chain_next(curr_bs);
6721 
6722         if (bdrv_backing_overridden(curr_bs)) {
6723             /*
6724              * If the backing file was overridden, we can only compare
6725              * directly against the backing node's filename.
6726              */
6727 
6728             if (!filenames_refreshed) {
6729                 /*
6730                  * This will automatically refresh all of the
6731                  * filenames in the rest of the backing chain, so we
6732                  * only need to do this once.
6733                  */
6734                 bdrv_refresh_filename(bs_below);
6735                 filenames_refreshed = true;
6736             }
6737 
6738             if (strcmp(backing_file, bs_below->filename) == 0) {
6739                 retval = bs_below;
6740                 break;
6741             }
6742         } else if (is_protocol || path_has_protocol(curr_bs->backing_file)) {
6743             /*
6744              * If either of the filename paths is actually a protocol, then
6745              * compare unmodified paths; otherwise make paths relative.
6746              */
6747             char *backing_file_full_ret;
6748 
6749             if (strcmp(backing_file, curr_bs->backing_file) == 0) {
6750                 retval = bs_below;
6751                 break;
6752             }
6753             /* Also check against the full backing filename for the image */
6754             backing_file_full_ret = bdrv_get_full_backing_filename(curr_bs,
6755                                                                    NULL);
6756             if (backing_file_full_ret) {
6757                 bool equal = strcmp(backing_file, backing_file_full_ret) == 0;
6758                 g_free(backing_file_full_ret);
6759                 if (equal) {
6760                     retval = bs_below;
6761                     break;
6762                 }
6763             }
6764         } else {
6765             /* If not an absolute filename path, make it relative to the current
6766              * image's filename path */
6767             filename_tmp = bdrv_make_absolute_filename(curr_bs, backing_file,
6768                                                        NULL);
6769             /* We are going to compare canonicalized absolute pathnames */
6770             if (!filename_tmp || !realpath(filename_tmp, filename_full)) {
6771                 g_free(filename_tmp);
6772                 continue;
6773             }
6774             g_free(filename_tmp);
6775 
6776             /* We need to make sure the backing filename we are comparing against
6777              * is relative to the current image filename (or absolute) */
6778             filename_tmp = bdrv_get_full_backing_filename(curr_bs, NULL);
6779             if (!filename_tmp || !realpath(filename_tmp, backing_file_full)) {
6780                 g_free(filename_tmp);
6781                 continue;
6782             }
6783             g_free(filename_tmp);
6784 
6785             if (strcmp(backing_file_full, filename_full) == 0) {
6786                 retval = bs_below;
6787                 break;
6788             }
6789         }
6790     }
6791 
6792     g_free(filename_full);
6793     g_free(backing_file_full);
6794     return retval;
6795 }
6796 
6797 void bdrv_init(void)
6798 {
6799 #ifdef CONFIG_BDRV_WHITELIST_TOOLS
6800     use_bdrv_whitelist = 1;
6801 #endif
6802     module_call_init(MODULE_INIT_BLOCK);
6803 }
6804 
6805 void bdrv_init_with_whitelist(void)
6806 {
6807     use_bdrv_whitelist = 1;
6808     bdrv_init();
6809 }
6810 
6811 int bdrv_activate(BlockDriverState *bs, Error **errp)
6812 {
6813     BdrvChild *child, *parent;
6814     Error *local_err = NULL;
6815     int ret;
6816     BdrvDirtyBitmap *bm;
6817 
6818     GLOBAL_STATE_CODE();
6819     GRAPH_RDLOCK_GUARD_MAINLOOP();
6820 
6821     if (!bs->drv)  {
6822         return -ENOMEDIUM;
6823     }
6824 
6825     QLIST_FOREACH(child, &bs->children, next) {
6826         bdrv_activate(child->bs, &local_err);
6827         if (local_err) {
6828             error_propagate(errp, local_err);
6829             return -EINVAL;
6830         }
6831     }
6832 
6833     /*
6834      * Update permissions, they may differ for inactive nodes.
6835      *
6836      * Note that the required permissions of inactive images are always a
6837      * subset of the permissions required after activating the image. This
6838      * allows us to just get the permissions upfront without restricting
6839      * bdrv_co_invalidate_cache().
6840      *
6841      * It also means that in error cases, we don't have to try and revert to
6842      * the old permissions (which is an operation that could fail, too). We can
6843      * just keep the extended permissions for the next time that an activation
6844      * of the image is tried.
6845      */
6846     if (bs->open_flags & BDRV_O_INACTIVE) {
6847         bs->open_flags &= ~BDRV_O_INACTIVE;
6848         ret = bdrv_refresh_perms(bs, NULL, errp);
6849         if (ret < 0) {
6850             bs->open_flags |= BDRV_O_INACTIVE;
6851             return ret;
6852         }
6853 
6854         ret = bdrv_invalidate_cache(bs, errp);
6855         if (ret < 0) {
6856             bs->open_flags |= BDRV_O_INACTIVE;
6857             return ret;
6858         }
6859 
6860         FOR_EACH_DIRTY_BITMAP(bs, bm) {
6861             bdrv_dirty_bitmap_skip_store(bm, false);
6862         }
6863 
6864         ret = bdrv_refresh_total_sectors(bs, bs->total_sectors);
6865         if (ret < 0) {
6866             bs->open_flags |= BDRV_O_INACTIVE;
6867             error_setg_errno(errp, -ret, "Could not refresh total sector count");
6868             return ret;
6869         }
6870     }
6871 
6872     QLIST_FOREACH(parent, &bs->parents, next_parent) {
6873         if (parent->klass->activate) {
6874             parent->klass->activate(parent, &local_err);
6875             if (local_err) {
6876                 bs->open_flags |= BDRV_O_INACTIVE;
6877                 error_propagate(errp, local_err);
6878                 return -EINVAL;
6879             }
6880         }
6881     }
6882 
6883     return 0;
6884 }
6885 
6886 int coroutine_fn bdrv_co_invalidate_cache(BlockDriverState *bs, Error **errp)
6887 {
6888     Error *local_err = NULL;
6889     IO_CODE();
6890 
6891     assert(!(bs->open_flags & BDRV_O_INACTIVE));
6892     assert_bdrv_graph_readable();
6893 
6894     if (bs->drv->bdrv_co_invalidate_cache) {
6895         bs->drv->bdrv_co_invalidate_cache(bs, &local_err);
6896         if (local_err) {
6897             error_propagate(errp, local_err);
6898             return -EINVAL;
6899         }
6900     }
6901 
6902     return 0;
6903 }
6904 
6905 void bdrv_activate_all(Error **errp)
6906 {
6907     BlockDriverState *bs;
6908     BdrvNextIterator it;
6909 
6910     GLOBAL_STATE_CODE();
6911     GRAPH_RDLOCK_GUARD_MAINLOOP();
6912 
6913     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
6914         int ret;
6915 
6916         ret = bdrv_activate(bs, errp);
6917         if (ret < 0) {
6918             bdrv_next_cleanup(&it);
6919             return;
6920         }
6921     }
6922 }
6923 
6924 static bool GRAPH_RDLOCK
6925 bdrv_has_bds_parent(BlockDriverState *bs, bool only_active)
6926 {
6927     BdrvChild *parent;
6928     GLOBAL_STATE_CODE();
6929 
6930     QLIST_FOREACH(parent, &bs->parents, next_parent) {
6931         if (parent->klass->parent_is_bds) {
6932             BlockDriverState *parent_bs = parent->opaque;
6933             if (!only_active || !(parent_bs->open_flags & BDRV_O_INACTIVE)) {
6934                 return true;
6935             }
6936         }
6937     }
6938 
6939     return false;
6940 }
6941 
6942 static int GRAPH_RDLOCK bdrv_inactivate_recurse(BlockDriverState *bs)
6943 {
6944     BdrvChild *child, *parent;
6945     int ret;
6946     uint64_t cumulative_perms, cumulative_shared_perms;
6947 
6948     GLOBAL_STATE_CODE();
6949 
6950     if (!bs->drv) {
6951         return -ENOMEDIUM;
6952     }
6953 
6954     /* Make sure that we don't inactivate a child before its parent.
6955      * It will be covered by recursion from the yet active parent. */
6956     if (bdrv_has_bds_parent(bs, true)) {
6957         return 0;
6958     }
6959 
6960     assert(!(bs->open_flags & BDRV_O_INACTIVE));
6961 
6962     /* Inactivate this node */
6963     if (bs->drv->bdrv_inactivate) {
6964         ret = bs->drv->bdrv_inactivate(bs);
6965         if (ret < 0) {
6966             return ret;
6967         }
6968     }
6969 
6970     QLIST_FOREACH(parent, &bs->parents, next_parent) {
6971         if (parent->klass->inactivate) {
6972             ret = parent->klass->inactivate(parent);
6973             if (ret < 0) {
6974                 return ret;
6975             }
6976         }
6977     }
6978 
6979     bdrv_get_cumulative_perm(bs, &cumulative_perms,
6980                              &cumulative_shared_perms);
6981     if (cumulative_perms & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED)) {
6982         /* Our inactive parents still need write access. Inactivation failed. */
6983         return -EPERM;
6984     }
6985 
6986     bs->open_flags |= BDRV_O_INACTIVE;
6987 
6988     /*
6989      * Update permissions, they may differ for inactive nodes.
6990      * We only tried to loosen restrictions, so errors are not fatal, ignore
6991      * them.
6992      */
6993     bdrv_refresh_perms(bs, NULL, NULL);
6994 
6995     /* Recursively inactivate children */
6996     QLIST_FOREACH(child, &bs->children, next) {
6997         ret = bdrv_inactivate_recurse(child->bs);
6998         if (ret < 0) {
6999             return ret;
7000         }
7001     }
7002 
7003     return 0;
7004 }
7005 
7006 int bdrv_inactivate_all(void)
7007 {
7008     BlockDriverState *bs = NULL;
7009     BdrvNextIterator it;
7010     int ret = 0;
7011 
7012     GLOBAL_STATE_CODE();
7013     GRAPH_RDLOCK_GUARD_MAINLOOP();
7014 
7015     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
7016         /* Nodes with BDS parents are covered by recursion from the last
7017          * parent that gets inactivated. Don't inactivate them a second
7018          * time if that has already happened. */
7019         if (bdrv_has_bds_parent(bs, false)) {
7020             continue;
7021         }
7022         ret = bdrv_inactivate_recurse(bs);
7023         if (ret < 0) {
7024             bdrv_next_cleanup(&it);
7025             break;
7026         }
7027     }
7028 
7029     return ret;
7030 }
7031 
7032 /**************************************************************/
7033 /* removable device support */
7034 
7035 /**
7036  * Return TRUE if the media is present
7037  */
7038 bool coroutine_fn bdrv_co_is_inserted(BlockDriverState *bs)
7039 {
7040     BlockDriver *drv = bs->drv;
7041     BdrvChild *child;
7042     IO_CODE();
7043     assert_bdrv_graph_readable();
7044 
7045     if (!drv) {
7046         return false;
7047     }
7048     if (drv->bdrv_co_is_inserted) {
7049         return drv->bdrv_co_is_inserted(bs);
7050     }
7051     QLIST_FOREACH(child, &bs->children, next) {
7052         if (!bdrv_co_is_inserted(child->bs)) {
7053             return false;
7054         }
7055     }
7056     return true;
7057 }
7058 
7059 /**
7060  * If eject_flag is TRUE, eject the media. Otherwise, close the tray
7061  */
7062 void coroutine_fn bdrv_co_eject(BlockDriverState *bs, bool eject_flag)
7063 {
7064     BlockDriver *drv = bs->drv;
7065     IO_CODE();
7066     assert_bdrv_graph_readable();
7067 
7068     if (drv && drv->bdrv_co_eject) {
7069         drv->bdrv_co_eject(bs, eject_flag);
7070     }
7071 }
7072 
7073 /**
7074  * Lock or unlock the media (if it is locked, the user won't be able
7075  * to eject it manually).
7076  */
7077 void coroutine_fn bdrv_co_lock_medium(BlockDriverState *bs, bool locked)
7078 {
7079     BlockDriver *drv = bs->drv;
7080     IO_CODE();
7081     assert_bdrv_graph_readable();
7082     trace_bdrv_lock_medium(bs, locked);
7083 
7084     if (drv && drv->bdrv_co_lock_medium) {
7085         drv->bdrv_co_lock_medium(bs, locked);
7086     }
7087 }
7088 
7089 /* Get a reference to bs */
7090 void bdrv_ref(BlockDriverState *bs)
7091 {
7092     GLOBAL_STATE_CODE();
7093     bs->refcnt++;
7094 }
7095 
7096 /* Release a previously grabbed reference to bs.
7097  * If after releasing, reference count is zero, the BlockDriverState is
7098  * deleted. */
7099 void bdrv_unref(BlockDriverState *bs)
7100 {
7101     GLOBAL_STATE_CODE();
7102     if (!bs) {
7103         return;
7104     }
7105     assert(bs->refcnt > 0);
7106     if (--bs->refcnt == 0) {
7107         bdrv_delete(bs);
7108     }
7109 }
7110 
7111 static void bdrv_schedule_unref_bh(void *opaque)
7112 {
7113     BlockDriverState *bs = opaque;
7114 
7115     bdrv_unref(bs);
7116 }
7117 
7118 /*
7119  * Release a BlockDriverState reference while holding the graph write lock.
7120  *
7121  * Calling bdrv_unref() directly is forbidden while holding the graph lock
7122  * because bdrv_close() both involves polling and taking the graph lock
7123  * internally. bdrv_schedule_unref() instead delays decreasing the refcount and
7124  * possibly closing @bs until the graph lock is released.
7125  */
7126 void bdrv_schedule_unref(BlockDriverState *bs)
7127 {
7128     if (!bs) {
7129         return;
7130     }
7131     aio_bh_schedule_oneshot(qemu_get_aio_context(), bdrv_schedule_unref_bh, bs);
7132 }
7133 
7134 struct BdrvOpBlocker {
7135     Error *reason;
7136     QLIST_ENTRY(BdrvOpBlocker) list;
7137 };
7138 
7139 bool bdrv_op_is_blocked(BlockDriverState *bs, BlockOpType op, Error **errp)
7140 {
7141     BdrvOpBlocker *blocker;
7142     GLOBAL_STATE_CODE();
7143 
7144     assert((int) op >= 0 && op < BLOCK_OP_TYPE_MAX);
7145     if (!QLIST_EMPTY(&bs->op_blockers[op])) {
7146         blocker = QLIST_FIRST(&bs->op_blockers[op]);
7147         error_propagate_prepend(errp, error_copy(blocker->reason),
7148                                 "Node '%s' is busy: ",
7149                                 bdrv_get_device_or_node_name(bs));
7150         return true;
7151     }
7152     return false;
7153 }
7154 
7155 void bdrv_op_block(BlockDriverState *bs, BlockOpType op, Error *reason)
7156 {
7157     BdrvOpBlocker *blocker;
7158     GLOBAL_STATE_CODE();
7159     assert((int) op >= 0 && op < BLOCK_OP_TYPE_MAX);
7160 
7161     blocker = g_new0(BdrvOpBlocker, 1);
7162     blocker->reason = reason;
7163     QLIST_INSERT_HEAD(&bs->op_blockers[op], blocker, list);
7164 }
7165 
7166 void bdrv_op_unblock(BlockDriverState *bs, BlockOpType op, Error *reason)
7167 {
7168     BdrvOpBlocker *blocker, *next;
7169     GLOBAL_STATE_CODE();
7170     assert((int) op >= 0 && op < BLOCK_OP_TYPE_MAX);
7171     QLIST_FOREACH_SAFE(blocker, &bs->op_blockers[op], list, next) {
7172         if (blocker->reason == reason) {
7173             QLIST_REMOVE(blocker, list);
7174             g_free(blocker);
7175         }
7176     }
7177 }
7178 
7179 void bdrv_op_block_all(BlockDriverState *bs, Error *reason)
7180 {
7181     int i;
7182     GLOBAL_STATE_CODE();
7183     for (i = 0; i < BLOCK_OP_TYPE_MAX; i++) {
7184         bdrv_op_block(bs, i, reason);
7185     }
7186 }
7187 
7188 void bdrv_op_unblock_all(BlockDriverState *bs, Error *reason)
7189 {
7190     int i;
7191     GLOBAL_STATE_CODE();
7192     for (i = 0; i < BLOCK_OP_TYPE_MAX; i++) {
7193         bdrv_op_unblock(bs, i, reason);
7194     }
7195 }
7196 
7197 bool bdrv_op_blocker_is_empty(BlockDriverState *bs)
7198 {
7199     int i;
7200     GLOBAL_STATE_CODE();
7201     for (i = 0; i < BLOCK_OP_TYPE_MAX; i++) {
7202         if (!QLIST_EMPTY(&bs->op_blockers[i])) {
7203             return false;
7204         }
7205     }
7206     return true;
7207 }
7208 
7209 /*
7210  * Must not be called while holding the lock of an AioContext other than the
7211  * current one.
7212  */
7213 void bdrv_img_create(const char *filename, const char *fmt,
7214                      const char *base_filename, const char *base_fmt,
7215                      char *options, uint64_t img_size, int flags, bool quiet,
7216                      Error **errp)
7217 {
7218     QemuOptsList *create_opts = NULL;
7219     QemuOpts *opts = NULL;
7220     const char *backing_fmt, *backing_file;
7221     int64_t size;
7222     BlockDriver *drv, *proto_drv;
7223     Error *local_err = NULL;
7224     int ret = 0;
7225 
7226     GLOBAL_STATE_CODE();
7227 
7228     /* Find driver and parse its options */
7229     drv = bdrv_find_format(fmt);
7230     if (!drv) {
7231         error_setg(errp, "Unknown file format '%s'", fmt);
7232         return;
7233     }
7234 
7235     proto_drv = bdrv_find_protocol(filename, true, errp);
7236     if (!proto_drv) {
7237         return;
7238     }
7239 
7240     if (!drv->create_opts) {
7241         error_setg(errp, "Format driver '%s' does not support image creation",
7242                    drv->format_name);
7243         return;
7244     }
7245 
7246     if (!proto_drv->create_opts) {
7247         error_setg(errp, "Protocol driver '%s' does not support image creation",
7248                    proto_drv->format_name);
7249         return;
7250     }
7251 
7252     /* Create parameter list */
7253     create_opts = qemu_opts_append(create_opts, drv->create_opts);
7254     create_opts = qemu_opts_append(create_opts, proto_drv->create_opts);
7255 
7256     opts = qemu_opts_create(create_opts, NULL, 0, &error_abort);
7257 
7258     /* Parse -o options */
7259     if (options) {
7260         if (!qemu_opts_do_parse(opts, options, NULL, errp)) {
7261             goto out;
7262         }
7263     }
7264 
7265     if (!qemu_opt_get(opts, BLOCK_OPT_SIZE)) {
7266         qemu_opt_set_number(opts, BLOCK_OPT_SIZE, img_size, &error_abort);
7267     } else if (img_size != UINT64_C(-1)) {
7268         error_setg(errp, "The image size must be specified only once");
7269         goto out;
7270     }
7271 
7272     if (base_filename) {
7273         if (!qemu_opt_set(opts, BLOCK_OPT_BACKING_FILE, base_filename,
7274                           NULL)) {
7275             error_setg(errp, "Backing file not supported for file format '%s'",
7276                        fmt);
7277             goto out;
7278         }
7279     }
7280 
7281     if (base_fmt) {
7282         if (!qemu_opt_set(opts, BLOCK_OPT_BACKING_FMT, base_fmt, NULL)) {
7283             error_setg(errp, "Backing file format not supported for file "
7284                              "format '%s'", fmt);
7285             goto out;
7286         }
7287     }
7288 
7289     backing_file = qemu_opt_get(opts, BLOCK_OPT_BACKING_FILE);
7290     if (backing_file) {
7291         if (!strcmp(filename, backing_file)) {
7292             error_setg(errp, "Error: Trying to create an image with the "
7293                              "same filename as the backing file");
7294             goto out;
7295         }
7296         if (backing_file[0] == '\0') {
7297             error_setg(errp, "Expected backing file name, got empty string");
7298             goto out;
7299         }
7300     }
7301 
7302     backing_fmt = qemu_opt_get(opts, BLOCK_OPT_BACKING_FMT);
7303 
7304     /* The size for the image must always be specified, unless we have a backing
7305      * file and we have not been forbidden from opening it. */
7306     size = qemu_opt_get_size(opts, BLOCK_OPT_SIZE, img_size);
7307     if (backing_file && !(flags & BDRV_O_NO_BACKING)) {
7308         BlockDriverState *bs;
7309         char *full_backing;
7310         int back_flags;
7311         QDict *backing_options = NULL;
7312 
7313         full_backing =
7314             bdrv_get_full_backing_filename_from_filename(filename, backing_file,
7315                                                          &local_err);
7316         if (local_err) {
7317             goto out;
7318         }
7319         assert(full_backing);
7320 
7321         /*
7322          * No need to do I/O here, which allows us to open encrypted
7323          * backing images without needing the secret
7324          */
7325         back_flags = flags;
7326         back_flags &= ~(BDRV_O_RDWR | BDRV_O_SNAPSHOT | BDRV_O_NO_BACKING);
7327         back_flags |= BDRV_O_NO_IO;
7328 
7329         backing_options = qdict_new();
7330         if (backing_fmt) {
7331             qdict_put_str(backing_options, "driver", backing_fmt);
7332         }
7333         qdict_put_bool(backing_options, BDRV_OPT_FORCE_SHARE, true);
7334 
7335         bs = bdrv_open(full_backing, NULL, backing_options, back_flags,
7336                        &local_err);
7337         g_free(full_backing);
7338         if (!bs) {
7339             error_append_hint(&local_err, "Could not open backing image.\n");
7340             goto out;
7341         } else {
7342             if (!backing_fmt) {
7343                 error_setg(&local_err,
7344                            "Backing file specified without backing format");
7345                 error_append_hint(&local_err, "Detected format of %s.\n",
7346                                   bs->drv->format_name);
7347                 goto out;
7348             }
7349             if (size == -1) {
7350                 /* Opened BS, have no size */
7351                 size = bdrv_getlength(bs);
7352                 if (size < 0) {
7353                     error_setg_errno(errp, -size, "Could not get size of '%s'",
7354                                      backing_file);
7355                     bdrv_unref(bs);
7356                     goto out;
7357                 }
7358                 qemu_opt_set_number(opts, BLOCK_OPT_SIZE, size, &error_abort);
7359             }
7360             bdrv_unref(bs);
7361         }
7362         /* (backing_file && !(flags & BDRV_O_NO_BACKING)) */
7363     } else if (backing_file && !backing_fmt) {
7364         error_setg(&local_err,
7365                    "Backing file specified without backing format");
7366         goto out;
7367     }
7368 
7369     if (size == -1) {
7370         error_setg(errp, "Image creation needs a size parameter");
7371         goto out;
7372     }
7373 
7374     if (!quiet) {
7375         printf("Formatting '%s', fmt=%s ", filename, fmt);
7376         qemu_opts_print(opts, " ");
7377         puts("");
7378         fflush(stdout);
7379     }
7380 
7381     ret = bdrv_create(drv, filename, opts, &local_err);
7382 
7383     if (ret == -EFBIG) {
7384         /* This is generally a better message than whatever the driver would
7385          * deliver (especially because of the cluster_size_hint), since that
7386          * is most probably not much different from "image too large". */
7387         const char *cluster_size_hint = "";
7388         if (qemu_opt_get_size(opts, BLOCK_OPT_CLUSTER_SIZE, 0)) {
7389             cluster_size_hint = " (try using a larger cluster size)";
7390         }
7391         error_setg(errp, "The image size is too large for file format '%s'"
7392                    "%s", fmt, cluster_size_hint);
7393         error_free(local_err);
7394         local_err = NULL;
7395     }
7396 
7397 out:
7398     qemu_opts_del(opts);
7399     qemu_opts_free(create_opts);
7400     error_propagate(errp, local_err);
7401 }
7402 
7403 AioContext *bdrv_get_aio_context(BlockDriverState *bs)
7404 {
7405     IO_CODE();
7406     return bs ? bs->aio_context : qemu_get_aio_context();
7407 }
7408 
7409 AioContext *coroutine_fn bdrv_co_enter(BlockDriverState *bs)
7410 {
7411     Coroutine *self = qemu_coroutine_self();
7412     AioContext *old_ctx = qemu_coroutine_get_aio_context(self);
7413     AioContext *new_ctx;
7414     IO_CODE();
7415 
7416     /*
7417      * Increase bs->in_flight to ensure that this operation is completed before
7418      * moving the node to a different AioContext. Read new_ctx only afterwards.
7419      */
7420     bdrv_inc_in_flight(bs);
7421 
7422     new_ctx = bdrv_get_aio_context(bs);
7423     aio_co_reschedule_self(new_ctx);
7424     return old_ctx;
7425 }
7426 
7427 void coroutine_fn bdrv_co_leave(BlockDriverState *bs, AioContext *old_ctx)
7428 {
7429     IO_CODE();
7430     aio_co_reschedule_self(old_ctx);
7431     bdrv_dec_in_flight(bs);
7432 }
7433 
7434 static void bdrv_do_remove_aio_context_notifier(BdrvAioNotifier *ban)
7435 {
7436     GLOBAL_STATE_CODE();
7437     QLIST_REMOVE(ban, list);
7438     g_free(ban);
7439 }
7440 
7441 static void bdrv_detach_aio_context(BlockDriverState *bs)
7442 {
7443     BdrvAioNotifier *baf, *baf_tmp;
7444 
7445     assert(!bs->walking_aio_notifiers);
7446     GLOBAL_STATE_CODE();
7447     bs->walking_aio_notifiers = true;
7448     QLIST_FOREACH_SAFE(baf, &bs->aio_notifiers, list, baf_tmp) {
7449         if (baf->deleted) {
7450             bdrv_do_remove_aio_context_notifier(baf);
7451         } else {
7452             baf->detach_aio_context(baf->opaque);
7453         }
7454     }
7455     /* Never mind iterating again to check for ->deleted.  bdrv_close() will
7456      * remove remaining aio notifiers if we aren't called again.
7457      */
7458     bs->walking_aio_notifiers = false;
7459 
7460     if (bs->drv && bs->drv->bdrv_detach_aio_context) {
7461         bs->drv->bdrv_detach_aio_context(bs);
7462     }
7463 
7464     bs->aio_context = NULL;
7465 }
7466 
7467 static void bdrv_attach_aio_context(BlockDriverState *bs,
7468                                     AioContext *new_context)
7469 {
7470     BdrvAioNotifier *ban, *ban_tmp;
7471     GLOBAL_STATE_CODE();
7472 
7473     bs->aio_context = new_context;
7474 
7475     if (bs->drv && bs->drv->bdrv_attach_aio_context) {
7476         bs->drv->bdrv_attach_aio_context(bs, new_context);
7477     }
7478 
7479     assert(!bs->walking_aio_notifiers);
7480     bs->walking_aio_notifiers = true;
7481     QLIST_FOREACH_SAFE(ban, &bs->aio_notifiers, list, ban_tmp) {
7482         if (ban->deleted) {
7483             bdrv_do_remove_aio_context_notifier(ban);
7484         } else {
7485             ban->attached_aio_context(new_context, ban->opaque);
7486         }
7487     }
7488     bs->walking_aio_notifiers = false;
7489 }
7490 
7491 typedef struct BdrvStateSetAioContext {
7492     AioContext *new_ctx;
7493     BlockDriverState *bs;
7494 } BdrvStateSetAioContext;
7495 
7496 static bool bdrv_parent_change_aio_context(BdrvChild *c, AioContext *ctx,
7497                                            GHashTable *visited,
7498                                            Transaction *tran,
7499                                            Error **errp)
7500 {
7501     GLOBAL_STATE_CODE();
7502     if (g_hash_table_contains(visited, c)) {
7503         return true;
7504     }
7505     g_hash_table_add(visited, c);
7506 
7507     /*
7508      * A BdrvChildClass that doesn't handle AioContext changes cannot
7509      * tolerate any AioContext changes
7510      */
7511     if (!c->klass->change_aio_ctx) {
7512         char *user = bdrv_child_user_desc(c);
7513         error_setg(errp, "Changing iothreads is not supported by %s", user);
7514         g_free(user);
7515         return false;
7516     }
7517     if (!c->klass->change_aio_ctx(c, ctx, visited, tran, errp)) {
7518         assert(!errp || *errp);
7519         return false;
7520     }
7521     return true;
7522 }
7523 
7524 bool bdrv_child_change_aio_context(BdrvChild *c, AioContext *ctx,
7525                                    GHashTable *visited, Transaction *tran,
7526                                    Error **errp)
7527 {
7528     GLOBAL_STATE_CODE();
7529     if (g_hash_table_contains(visited, c)) {
7530         return true;
7531     }
7532     g_hash_table_add(visited, c);
7533     return bdrv_change_aio_context(c->bs, ctx, visited, tran, errp);
7534 }
7535 
7536 static void bdrv_set_aio_context_clean(void *opaque)
7537 {
7538     BdrvStateSetAioContext *state = (BdrvStateSetAioContext *) opaque;
7539     BlockDriverState *bs = (BlockDriverState *) state->bs;
7540 
7541     /* Paired with bdrv_drained_begin in bdrv_change_aio_context() */
7542     bdrv_drained_end(bs);
7543 
7544     g_free(state);
7545 }
7546 
7547 static void bdrv_set_aio_context_commit(void *opaque)
7548 {
7549     BdrvStateSetAioContext *state = (BdrvStateSetAioContext *) opaque;
7550     BlockDriverState *bs = (BlockDriverState *) state->bs;
7551     AioContext *new_context = state->new_ctx;
7552 
7553     bdrv_detach_aio_context(bs);
7554     bdrv_attach_aio_context(bs, new_context);
7555 }
7556 
7557 static TransactionActionDrv set_aio_context = {
7558     .commit = bdrv_set_aio_context_commit,
7559     .clean = bdrv_set_aio_context_clean,
7560 };
7561 
7562 /*
7563  * Changes the AioContext used for fd handlers, timers, and BHs by this
7564  * BlockDriverState and all its children and parents.
7565  *
7566  * Must be called from the main AioContext.
7567  *
7568  * The caller must own the AioContext lock for the old AioContext of bs, but it
7569  * must not own the AioContext lock for new_context (unless new_context is the
7570  * same as the current context of bs).
7571  *
7572  * @visited will accumulate all visited BdrvChild objects. The caller is
7573  * responsible for freeing the list afterwards.
7574  */
7575 static bool bdrv_change_aio_context(BlockDriverState *bs, AioContext *ctx,
7576                                     GHashTable *visited, Transaction *tran,
7577                                     Error **errp)
7578 {
7579     BdrvChild *c;
7580     BdrvStateSetAioContext *state;
7581 
7582     GLOBAL_STATE_CODE();
7583 
7584     if (bdrv_get_aio_context(bs) == ctx) {
7585         return true;
7586     }
7587 
7588     bdrv_graph_rdlock_main_loop();
7589     QLIST_FOREACH(c, &bs->parents, next_parent) {
7590         if (!bdrv_parent_change_aio_context(c, ctx, visited, tran, errp)) {
7591             bdrv_graph_rdunlock_main_loop();
7592             return false;
7593         }
7594     }
7595 
7596     QLIST_FOREACH(c, &bs->children, next) {
7597         if (!bdrv_child_change_aio_context(c, ctx, visited, tran, errp)) {
7598             bdrv_graph_rdunlock_main_loop();
7599             return false;
7600         }
7601     }
7602     bdrv_graph_rdunlock_main_loop();
7603 
7604     state = g_new(BdrvStateSetAioContext, 1);
7605     *state = (BdrvStateSetAioContext) {
7606         .new_ctx = ctx,
7607         .bs = bs,
7608     };
7609 
7610     /* Paired with bdrv_drained_end in bdrv_set_aio_context_clean() */
7611     bdrv_drained_begin(bs);
7612 
7613     tran_add(tran, &set_aio_context, state);
7614 
7615     return true;
7616 }
7617 
7618 /*
7619  * Change bs's and recursively all of its parents' and children's AioContext
7620  * to the given new context, returning an error if that isn't possible.
7621  *
7622  * If ignore_child is not NULL, that child (and its subgraph) will not
7623  * be touched.
7624  *
7625  * This function still requires the caller to take the bs current
7626  * AioContext lock, otherwise draining will fail since AIO_WAIT_WHILE
7627  * assumes the lock is always held if bs is in another AioContext.
7628  * For the same reason, it temporarily also holds the new AioContext, since
7629  * bdrv_drained_end calls BDRV_POLL_WHILE that assumes the lock is taken too.
7630  * Therefore the new AioContext lock must not be taken by the caller.
7631  */
7632 int bdrv_try_change_aio_context(BlockDriverState *bs, AioContext *ctx,
7633                                 BdrvChild *ignore_child, Error **errp)
7634 {
7635     Transaction *tran;
7636     GHashTable *visited;
7637     int ret;
7638     GLOBAL_STATE_CODE();
7639 
7640     /*
7641      * Recursion phase: go through all nodes of the graph.
7642      * Take care of checking that all nodes support changing AioContext
7643      * and drain them, building a linear list of callbacks to run if everything
7644      * is successful (the transaction itself).
7645      */
7646     tran = tran_new();
7647     visited = g_hash_table_new(NULL, NULL);
7648     if (ignore_child) {
7649         g_hash_table_add(visited, ignore_child);
7650     }
7651     ret = bdrv_change_aio_context(bs, ctx, visited, tran, errp);
7652     g_hash_table_destroy(visited);
7653 
7654     /*
7655      * Linear phase: go through all callbacks collected in the transaction.
7656      * Run all callbacks collected in the recursion to switch all nodes
7657      * AioContext lock (transaction commit), or undo all changes done in the
7658      * recursion (transaction abort).
7659      */
7660 
7661     if (!ret) {
7662         /* Just run clean() callbacks. No AioContext changed. */
7663         tran_abort(tran);
7664         return -EPERM;
7665     }
7666 
7667     tran_commit(tran);
7668     return 0;
7669 }
7670 
7671 void bdrv_add_aio_context_notifier(BlockDriverState *bs,
7672         void (*attached_aio_context)(AioContext *new_context, void *opaque),
7673         void (*detach_aio_context)(void *opaque), void *opaque)
7674 {
7675     BdrvAioNotifier *ban = g_new(BdrvAioNotifier, 1);
7676     *ban = (BdrvAioNotifier){
7677         .attached_aio_context = attached_aio_context,
7678         .detach_aio_context   = detach_aio_context,
7679         .opaque               = opaque
7680     };
7681     GLOBAL_STATE_CODE();
7682 
7683     QLIST_INSERT_HEAD(&bs->aio_notifiers, ban, list);
7684 }
7685 
7686 void bdrv_remove_aio_context_notifier(BlockDriverState *bs,
7687                                       void (*attached_aio_context)(AioContext *,
7688                                                                    void *),
7689                                       void (*detach_aio_context)(void *),
7690                                       void *opaque)
7691 {
7692     BdrvAioNotifier *ban, *ban_next;
7693     GLOBAL_STATE_CODE();
7694 
7695     QLIST_FOREACH_SAFE(ban, &bs->aio_notifiers, list, ban_next) {
7696         if (ban->attached_aio_context == attached_aio_context &&
7697             ban->detach_aio_context   == detach_aio_context   &&
7698             ban->opaque               == opaque               &&
7699             ban->deleted              == false)
7700         {
7701             if (bs->walking_aio_notifiers) {
7702                 ban->deleted = true;
7703             } else {
7704                 bdrv_do_remove_aio_context_notifier(ban);
7705             }
7706             return;
7707         }
7708     }
7709 
7710     abort();
7711 }
7712 
7713 int bdrv_amend_options(BlockDriverState *bs, QemuOpts *opts,
7714                        BlockDriverAmendStatusCB *status_cb, void *cb_opaque,
7715                        bool force,
7716                        Error **errp)
7717 {
7718     GLOBAL_STATE_CODE();
7719     if (!bs->drv) {
7720         error_setg(errp, "Node is ejected");
7721         return -ENOMEDIUM;
7722     }
7723     if (!bs->drv->bdrv_amend_options) {
7724         error_setg(errp, "Block driver '%s' does not support option amendment",
7725                    bs->drv->format_name);
7726         return -ENOTSUP;
7727     }
7728     return bs->drv->bdrv_amend_options(bs, opts, status_cb,
7729                                        cb_opaque, force, errp);
7730 }
7731 
7732 /*
7733  * This function checks whether the given @to_replace is allowed to be
7734  * replaced by a node that always shows the same data as @bs.  This is
7735  * used for example to verify whether the mirror job can replace
7736  * @to_replace by the target mirrored from @bs.
7737  * To be replaceable, @bs and @to_replace may either be guaranteed to
7738  * always show the same data (because they are only connected through
7739  * filters), or some driver may allow replacing one of its children
7740  * because it can guarantee that this child's data is not visible at
7741  * all (for example, for dissenting quorum children that have no other
7742  * parents).
7743  */
7744 bool bdrv_recurse_can_replace(BlockDriverState *bs,
7745                               BlockDriverState *to_replace)
7746 {
7747     BlockDriverState *filtered;
7748 
7749     GLOBAL_STATE_CODE();
7750 
7751     if (!bs || !bs->drv) {
7752         return false;
7753     }
7754 
7755     if (bs == to_replace) {
7756         return true;
7757     }
7758 
7759     /* See what the driver can do */
7760     if (bs->drv->bdrv_recurse_can_replace) {
7761         return bs->drv->bdrv_recurse_can_replace(bs, to_replace);
7762     }
7763 
7764     /* For filters without an own implementation, we can recurse on our own */
7765     filtered = bdrv_filter_bs(bs);
7766     if (filtered) {
7767         return bdrv_recurse_can_replace(filtered, to_replace);
7768     }
7769 
7770     /* Safe default */
7771     return false;
7772 }
7773 
7774 /*
7775  * Check whether the given @node_name can be replaced by a node that
7776  * has the same data as @parent_bs.  If so, return @node_name's BDS;
7777  * NULL otherwise.
7778  *
7779  * @node_name must be a (recursive) *child of @parent_bs (or this
7780  * function will return NULL).
7781  *
7782  * The result (whether the node can be replaced or not) is only valid
7783  * for as long as no graph or permission changes occur.
7784  */
7785 BlockDriverState *check_to_replace_node(BlockDriverState *parent_bs,
7786                                         const char *node_name, Error **errp)
7787 {
7788     BlockDriverState *to_replace_bs = bdrv_find_node(node_name);
7789 
7790     GLOBAL_STATE_CODE();
7791 
7792     if (!to_replace_bs) {
7793         error_setg(errp, "Failed to find node with node-name='%s'", node_name);
7794         return NULL;
7795     }
7796 
7797     if (bdrv_op_is_blocked(to_replace_bs, BLOCK_OP_TYPE_REPLACE, errp)) {
7798         return NULL;
7799     }
7800 
7801     /* We don't want arbitrary node of the BDS chain to be replaced only the top
7802      * most non filter in order to prevent data corruption.
7803      * Another benefit is that this tests exclude backing files which are
7804      * blocked by the backing blockers.
7805      */
7806     if (!bdrv_recurse_can_replace(parent_bs, to_replace_bs)) {
7807         error_setg(errp, "Cannot replace '%s' by a node mirrored from '%s', "
7808                    "because it cannot be guaranteed that doing so would not "
7809                    "lead to an abrupt change of visible data",
7810                    node_name, parent_bs->node_name);
7811         return NULL;
7812     }
7813 
7814     return to_replace_bs;
7815 }
7816 
7817 /**
7818  * Iterates through the list of runtime option keys that are said to
7819  * be "strong" for a BDS.  An option is called "strong" if it changes
7820  * a BDS's data.  For example, the null block driver's "size" and
7821  * "read-zeroes" options are strong, but its "latency-ns" option is
7822  * not.
7823  *
7824  * If a key returned by this function ends with a dot, all options
7825  * starting with that prefix are strong.
7826  */
7827 static const char *const *strong_options(BlockDriverState *bs,
7828                                          const char *const *curopt)
7829 {
7830     static const char *const global_options[] = {
7831         "driver", "filename", NULL
7832     };
7833 
7834     if (!curopt) {
7835         return &global_options[0];
7836     }
7837 
7838     curopt++;
7839     if (curopt == &global_options[ARRAY_SIZE(global_options) - 1] && bs->drv) {
7840         curopt = bs->drv->strong_runtime_opts;
7841     }
7842 
7843     return (curopt && *curopt) ? curopt : NULL;
7844 }
7845 
7846 /**
7847  * Copies all strong runtime options from bs->options to the given
7848  * QDict.  The set of strong option keys is determined by invoking
7849  * strong_options().
7850  *
7851  * Returns true iff any strong option was present in bs->options (and
7852  * thus copied to the target QDict) with the exception of "filename"
7853  * and "driver".  The caller is expected to use this value to decide
7854  * whether the existence of strong options prevents the generation of
7855  * a plain filename.
7856  */
7857 static bool append_strong_runtime_options(QDict *d, BlockDriverState *bs)
7858 {
7859     bool found_any = false;
7860     const char *const *option_name = NULL;
7861 
7862     if (!bs->drv) {
7863         return false;
7864     }
7865 
7866     while ((option_name = strong_options(bs, option_name))) {
7867         bool option_given = false;
7868 
7869         assert(strlen(*option_name) > 0);
7870         if ((*option_name)[strlen(*option_name) - 1] != '.') {
7871             QObject *entry = qdict_get(bs->options, *option_name);
7872             if (!entry) {
7873                 continue;
7874             }
7875 
7876             qdict_put_obj(d, *option_name, qobject_ref(entry));
7877             option_given = true;
7878         } else {
7879             const QDictEntry *entry;
7880             for (entry = qdict_first(bs->options); entry;
7881                  entry = qdict_next(bs->options, entry))
7882             {
7883                 if (strstart(qdict_entry_key(entry), *option_name, NULL)) {
7884                     qdict_put_obj(d, qdict_entry_key(entry),
7885                                   qobject_ref(qdict_entry_value(entry)));
7886                     option_given = true;
7887                 }
7888             }
7889         }
7890 
7891         /* While "driver" and "filename" need to be included in a JSON filename,
7892          * their existence does not prohibit generation of a plain filename. */
7893         if (!found_any && option_given &&
7894             strcmp(*option_name, "driver") && strcmp(*option_name, "filename"))
7895         {
7896             found_any = true;
7897         }
7898     }
7899 
7900     if (!qdict_haskey(d, "driver")) {
7901         /* Drivers created with bdrv_new_open_driver() may not have a
7902          * @driver option.  Add it here. */
7903         qdict_put_str(d, "driver", bs->drv->format_name);
7904     }
7905 
7906     return found_any;
7907 }
7908 
7909 /* Note: This function may return false positives; it may return true
7910  * even if opening the backing file specified by bs's image header
7911  * would result in exactly bs->backing. */
7912 static bool GRAPH_RDLOCK bdrv_backing_overridden(BlockDriverState *bs)
7913 {
7914     GLOBAL_STATE_CODE();
7915     if (bs->backing) {
7916         return strcmp(bs->auto_backing_file,
7917                       bs->backing->bs->filename);
7918     } else {
7919         /* No backing BDS, so if the image header reports any backing
7920          * file, it must have been suppressed */
7921         return bs->auto_backing_file[0] != '\0';
7922     }
7923 }
7924 
7925 /* Updates the following BDS fields:
7926  *  - exact_filename: A filename which may be used for opening a block device
7927  *                    which (mostly) equals the given BDS (even without any
7928  *                    other options; so reading and writing must return the same
7929  *                    results, but caching etc. may be different)
7930  *  - full_open_options: Options which, when given when opening a block device
7931  *                       (without a filename), result in a BDS (mostly)
7932  *                       equalling the given one
7933  *  - filename: If exact_filename is set, it is copied here. Otherwise,
7934  *              full_open_options is converted to a JSON object, prefixed with
7935  *              "json:" (for use through the JSON pseudo protocol) and put here.
7936  */
7937 void bdrv_refresh_filename(BlockDriverState *bs)
7938 {
7939     BlockDriver *drv = bs->drv;
7940     BdrvChild *child;
7941     BlockDriverState *primary_child_bs;
7942     QDict *opts;
7943     bool backing_overridden;
7944     bool generate_json_filename; /* Whether our default implementation should
7945                                     fill exact_filename (false) or not (true) */
7946 
7947     GLOBAL_STATE_CODE();
7948 
7949     if (!drv) {
7950         return;
7951     }
7952 
7953     /* This BDS's file name may depend on any of its children's file names, so
7954      * refresh those first */
7955     QLIST_FOREACH(child, &bs->children, next) {
7956         bdrv_refresh_filename(child->bs);
7957     }
7958 
7959     if (bs->implicit) {
7960         /* For implicit nodes, just copy everything from the single child */
7961         child = QLIST_FIRST(&bs->children);
7962         assert(QLIST_NEXT(child, next) == NULL);
7963 
7964         pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
7965                 child->bs->exact_filename);
7966         pstrcpy(bs->filename, sizeof(bs->filename), child->bs->filename);
7967 
7968         qobject_unref(bs->full_open_options);
7969         bs->full_open_options = qobject_ref(child->bs->full_open_options);
7970 
7971         return;
7972     }
7973 
7974     backing_overridden = bdrv_backing_overridden(bs);
7975 
7976     if (bs->open_flags & BDRV_O_NO_IO) {
7977         /* Without I/O, the backing file does not change anything.
7978          * Therefore, in such a case (primarily qemu-img), we can
7979          * pretend the backing file has not been overridden even if
7980          * it technically has been. */
7981         backing_overridden = false;
7982     }
7983 
7984     /* Gather the options QDict */
7985     opts = qdict_new();
7986     generate_json_filename = append_strong_runtime_options(opts, bs);
7987     generate_json_filename |= backing_overridden;
7988 
7989     if (drv->bdrv_gather_child_options) {
7990         /* Some block drivers may not want to present all of their children's
7991          * options, or name them differently from BdrvChild.name */
7992         drv->bdrv_gather_child_options(bs, opts, backing_overridden);
7993     } else {
7994         QLIST_FOREACH(child, &bs->children, next) {
7995             if (child == bs->backing && !backing_overridden) {
7996                 /* We can skip the backing BDS if it has not been overridden */
7997                 continue;
7998             }
7999 
8000             qdict_put(opts, child->name,
8001                       qobject_ref(child->bs->full_open_options));
8002         }
8003 
8004         if (backing_overridden && !bs->backing) {
8005             /* Force no backing file */
8006             qdict_put_null(opts, "backing");
8007         }
8008     }
8009 
8010     qobject_unref(bs->full_open_options);
8011     bs->full_open_options = opts;
8012 
8013     primary_child_bs = bdrv_primary_bs(bs);
8014 
8015     if (drv->bdrv_refresh_filename) {
8016         /* Obsolete information is of no use here, so drop the old file name
8017          * information before refreshing it */
8018         bs->exact_filename[0] = '\0';
8019 
8020         drv->bdrv_refresh_filename(bs);
8021     } else if (primary_child_bs) {
8022         /*
8023          * Try to reconstruct valid information from the underlying
8024          * file -- this only works for format nodes (filter nodes
8025          * cannot be probed and as such must be selected by the user
8026          * either through an options dict, or through a special
8027          * filename which the filter driver must construct in its
8028          * .bdrv_refresh_filename() implementation).
8029          */
8030 
8031         bs->exact_filename[0] = '\0';
8032 
8033         /*
8034          * We can use the underlying file's filename if:
8035          * - it has a filename,
8036          * - the current BDS is not a filter,
8037          * - the file is a protocol BDS, and
8038          * - opening that file (as this BDS's format) will automatically create
8039          *   the BDS tree we have right now, that is:
8040          *   - the user did not significantly change this BDS's behavior with
8041          *     some explicit (strong) options
8042          *   - no non-file child of this BDS has been overridden by the user
8043          *   Both of these conditions are represented by generate_json_filename.
8044          */
8045         if (primary_child_bs->exact_filename[0] &&
8046             primary_child_bs->drv->bdrv_file_open &&
8047             !drv->is_filter && !generate_json_filename)
8048         {
8049             strcpy(bs->exact_filename, primary_child_bs->exact_filename);
8050         }
8051     }
8052 
8053     if (bs->exact_filename[0]) {
8054         pstrcpy(bs->filename, sizeof(bs->filename), bs->exact_filename);
8055     } else {
8056         GString *json = qobject_to_json(QOBJECT(bs->full_open_options));
8057         if (snprintf(bs->filename, sizeof(bs->filename), "json:%s",
8058                      json->str) >= sizeof(bs->filename)) {
8059             /* Give user a hint if we truncated things. */
8060             strcpy(bs->filename + sizeof(bs->filename) - 4, "...");
8061         }
8062         g_string_free(json, true);
8063     }
8064 }
8065 
8066 char *bdrv_dirname(BlockDriverState *bs, Error **errp)
8067 {
8068     BlockDriver *drv = bs->drv;
8069     BlockDriverState *child_bs;
8070 
8071     GLOBAL_STATE_CODE();
8072 
8073     if (!drv) {
8074         error_setg(errp, "Node '%s' is ejected", bs->node_name);
8075         return NULL;
8076     }
8077 
8078     if (drv->bdrv_dirname) {
8079         return drv->bdrv_dirname(bs, errp);
8080     }
8081 
8082     child_bs = bdrv_primary_bs(bs);
8083     if (child_bs) {
8084         return bdrv_dirname(child_bs, errp);
8085     }
8086 
8087     bdrv_refresh_filename(bs);
8088     if (bs->exact_filename[0] != '\0') {
8089         return path_combine(bs->exact_filename, "");
8090     }
8091 
8092     error_setg(errp, "Cannot generate a base directory for %s nodes",
8093                drv->format_name);
8094     return NULL;
8095 }
8096 
8097 /*
8098  * Hot add/remove a BDS's child. So the user can take a child offline when
8099  * it is broken and take a new child online
8100  */
8101 void bdrv_add_child(BlockDriverState *parent_bs, BlockDriverState *child_bs,
8102                     Error **errp)
8103 {
8104     GLOBAL_STATE_CODE();
8105     if (!parent_bs->drv || !parent_bs->drv->bdrv_add_child) {
8106         error_setg(errp, "The node %s does not support adding a child",
8107                    bdrv_get_device_or_node_name(parent_bs));
8108         return;
8109     }
8110 
8111     /*
8112      * Non-zoned block drivers do not follow zoned storage constraints
8113      * (i.e. sequential writes to zones). Refuse mixing zoned and non-zoned
8114      * drivers in a graph.
8115      */
8116     if (!parent_bs->drv->supports_zoned_children &&
8117         child_bs->bl.zoned == BLK_Z_HM) {
8118         /*
8119          * The host-aware model allows zoned storage constraints and random
8120          * write. Allow mixing host-aware and non-zoned drivers. Using
8121          * host-aware device as a regular device.
8122          */
8123         error_setg(errp, "Cannot add a %s child to a %s parent",
8124                    child_bs->bl.zoned == BLK_Z_HM ? "zoned" : "non-zoned",
8125                    parent_bs->drv->supports_zoned_children ?
8126                    "support zoned children" : "not support zoned children");
8127         return;
8128     }
8129 
8130     if (!QLIST_EMPTY(&child_bs->parents)) {
8131         error_setg(errp, "The node %s already has a parent",
8132                    child_bs->node_name);
8133         return;
8134     }
8135 
8136     parent_bs->drv->bdrv_add_child(parent_bs, child_bs, errp);
8137 }
8138 
8139 void bdrv_del_child(BlockDriverState *parent_bs, BdrvChild *child, Error **errp)
8140 {
8141     BdrvChild *tmp;
8142 
8143     GLOBAL_STATE_CODE();
8144     if (!parent_bs->drv || !parent_bs->drv->bdrv_del_child) {
8145         error_setg(errp, "The node %s does not support removing a child",
8146                    bdrv_get_device_or_node_name(parent_bs));
8147         return;
8148     }
8149 
8150     QLIST_FOREACH(tmp, &parent_bs->children, next) {
8151         if (tmp == child) {
8152             break;
8153         }
8154     }
8155 
8156     if (!tmp) {
8157         error_setg(errp, "The node %s does not have a child named %s",
8158                    bdrv_get_device_or_node_name(parent_bs),
8159                    bdrv_get_device_or_node_name(child->bs));
8160         return;
8161     }
8162 
8163     parent_bs->drv->bdrv_del_child(parent_bs, child, errp);
8164 }
8165 
8166 int bdrv_make_empty(BdrvChild *c, Error **errp)
8167 {
8168     BlockDriver *drv = c->bs->drv;
8169     int ret;
8170 
8171     GLOBAL_STATE_CODE();
8172     assert(c->perm & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED));
8173 
8174     if (!drv->bdrv_make_empty) {
8175         error_setg(errp, "%s does not support emptying nodes",
8176                    drv->format_name);
8177         return -ENOTSUP;
8178     }
8179 
8180     ret = drv->bdrv_make_empty(c->bs);
8181     if (ret < 0) {
8182         error_setg_errno(errp, -ret, "Failed to empty %s",
8183                          c->bs->filename);
8184         return ret;
8185     }
8186 
8187     return 0;
8188 }
8189 
8190 /*
8191  * Return the child that @bs acts as an overlay for, and from which data may be
8192  * copied in COW or COR operations.  Usually this is the backing file.
8193  */
8194 BdrvChild *bdrv_cow_child(BlockDriverState *bs)
8195 {
8196     IO_CODE();
8197 
8198     if (!bs || !bs->drv) {
8199         return NULL;
8200     }
8201 
8202     if (bs->drv->is_filter) {
8203         return NULL;
8204     }
8205 
8206     if (!bs->backing) {
8207         return NULL;
8208     }
8209 
8210     assert(bs->backing->role & BDRV_CHILD_COW);
8211     return bs->backing;
8212 }
8213 
8214 /*
8215  * If @bs acts as a filter for exactly one of its children, return
8216  * that child.
8217  */
8218 BdrvChild *bdrv_filter_child(BlockDriverState *bs)
8219 {
8220     BdrvChild *c;
8221     IO_CODE();
8222 
8223     if (!bs || !bs->drv) {
8224         return NULL;
8225     }
8226 
8227     if (!bs->drv->is_filter) {
8228         return NULL;
8229     }
8230 
8231     /* Only one of @backing or @file may be used */
8232     assert(!(bs->backing && bs->file));
8233 
8234     c = bs->backing ?: bs->file;
8235     if (!c) {
8236         return NULL;
8237     }
8238 
8239     assert(c->role & BDRV_CHILD_FILTERED);
8240     return c;
8241 }
8242 
8243 /*
8244  * Return either the result of bdrv_cow_child() or bdrv_filter_child(),
8245  * whichever is non-NULL.
8246  *
8247  * Return NULL if both are NULL.
8248  */
8249 BdrvChild *bdrv_filter_or_cow_child(BlockDriverState *bs)
8250 {
8251     BdrvChild *cow_child = bdrv_cow_child(bs);
8252     BdrvChild *filter_child = bdrv_filter_child(bs);
8253     IO_CODE();
8254 
8255     /* Filter nodes cannot have COW backing files */
8256     assert(!(cow_child && filter_child));
8257 
8258     return cow_child ?: filter_child;
8259 }
8260 
8261 /*
8262  * Return the primary child of this node: For filters, that is the
8263  * filtered child.  For other nodes, that is usually the child storing
8264  * metadata.
8265  * (A generally more helpful description is that this is (usually) the
8266  * child that has the same filename as @bs.)
8267  *
8268  * Drivers do not necessarily have a primary child; for example quorum
8269  * does not.
8270  */
8271 BdrvChild *bdrv_primary_child(BlockDriverState *bs)
8272 {
8273     BdrvChild *c, *found = NULL;
8274     IO_CODE();
8275 
8276     QLIST_FOREACH(c, &bs->children, next) {
8277         if (c->role & BDRV_CHILD_PRIMARY) {
8278             assert(!found);
8279             found = c;
8280         }
8281     }
8282 
8283     return found;
8284 }
8285 
8286 static BlockDriverState * GRAPH_RDLOCK
8287 bdrv_do_skip_filters(BlockDriverState *bs, bool stop_on_explicit_filter)
8288 {
8289     BdrvChild *c;
8290 
8291     if (!bs) {
8292         return NULL;
8293     }
8294 
8295     while (!(stop_on_explicit_filter && !bs->implicit)) {
8296         c = bdrv_filter_child(bs);
8297         if (!c) {
8298             /*
8299              * A filter that is embedded in a working block graph must
8300              * have a child.  Assert this here so this function does
8301              * not return a filter node that is not expected by the
8302              * caller.
8303              */
8304             assert(!bs->drv || !bs->drv->is_filter);
8305             break;
8306         }
8307         bs = c->bs;
8308     }
8309     /*
8310      * Note that this treats nodes with bs->drv == NULL as not being
8311      * filters (bs->drv == NULL should be replaced by something else
8312      * anyway).
8313      * The advantage of this behavior is that this function will thus
8314      * always return a non-NULL value (given a non-NULL @bs).
8315      */
8316 
8317     return bs;
8318 }
8319 
8320 /*
8321  * Return the first BDS that has not been added implicitly or that
8322  * does not have a filtered child down the chain starting from @bs
8323  * (including @bs itself).
8324  */
8325 BlockDriverState *bdrv_skip_implicit_filters(BlockDriverState *bs)
8326 {
8327     GLOBAL_STATE_CODE();
8328     return bdrv_do_skip_filters(bs, true);
8329 }
8330 
8331 /*
8332  * Return the first BDS that does not have a filtered child down the
8333  * chain starting from @bs (including @bs itself).
8334  */
8335 BlockDriverState *bdrv_skip_filters(BlockDriverState *bs)
8336 {
8337     IO_CODE();
8338     return bdrv_do_skip_filters(bs, false);
8339 }
8340 
8341 /*
8342  * For a backing chain, return the first non-filter backing image of
8343  * the first non-filter image.
8344  */
8345 BlockDriverState *bdrv_backing_chain_next(BlockDriverState *bs)
8346 {
8347     IO_CODE();
8348     return bdrv_skip_filters(bdrv_cow_bs(bdrv_skip_filters(bs)));
8349 }
8350 
8351 /**
8352  * Check whether [offset, offset + bytes) overlaps with the cached
8353  * block-status data region.
8354  *
8355  * If so, and @pnum is not NULL, set *pnum to `bsc.data_end - offset`,
8356  * which is what bdrv_bsc_is_data()'s interface needs.
8357  * Otherwise, *pnum is not touched.
8358  */
8359 static bool bdrv_bsc_range_overlaps_locked(BlockDriverState *bs,
8360                                            int64_t offset, int64_t bytes,
8361                                            int64_t *pnum)
8362 {
8363     BdrvBlockStatusCache *bsc = qatomic_rcu_read(&bs->block_status_cache);
8364     bool overlaps;
8365 
8366     overlaps =
8367         qatomic_read(&bsc->valid) &&
8368         ranges_overlap(offset, bytes, bsc->data_start,
8369                        bsc->data_end - bsc->data_start);
8370 
8371     if (overlaps && pnum) {
8372         *pnum = bsc->data_end - offset;
8373     }
8374 
8375     return overlaps;
8376 }
8377 
8378 /**
8379  * See block_int.h for this function's documentation.
8380  */
8381 bool bdrv_bsc_is_data(BlockDriverState *bs, int64_t offset, int64_t *pnum)
8382 {
8383     IO_CODE();
8384     RCU_READ_LOCK_GUARD();
8385     return bdrv_bsc_range_overlaps_locked(bs, offset, 1, pnum);
8386 }
8387 
8388 /**
8389  * See block_int.h for this function's documentation.
8390  */
8391 void bdrv_bsc_invalidate_range(BlockDriverState *bs,
8392                                int64_t offset, int64_t bytes)
8393 {
8394     IO_CODE();
8395     RCU_READ_LOCK_GUARD();
8396 
8397     if (bdrv_bsc_range_overlaps_locked(bs, offset, bytes, NULL)) {
8398         qatomic_set(&bs->block_status_cache->valid, false);
8399     }
8400 }
8401 
8402 /**
8403  * See block_int.h for this function's documentation.
8404  */
8405 void bdrv_bsc_fill(BlockDriverState *bs, int64_t offset, int64_t bytes)
8406 {
8407     BdrvBlockStatusCache *new_bsc = g_new(BdrvBlockStatusCache, 1);
8408     BdrvBlockStatusCache *old_bsc;
8409     IO_CODE();
8410 
8411     *new_bsc = (BdrvBlockStatusCache) {
8412         .valid = true,
8413         .data_start = offset,
8414         .data_end = offset + bytes,
8415     };
8416 
8417     QEMU_LOCK_GUARD(&bs->bsc_modify_lock);
8418 
8419     old_bsc = qatomic_rcu_read(&bs->block_status_cache);
8420     qatomic_rcu_set(&bs->block_status_cache, new_bsc);
8421     if (old_bsc) {
8422         g_free_rcu(old_bsc, rcu);
8423     }
8424 }
8425