xref: /openbmc/qemu/block/file-posix.c (revision 7200fb21)
1 /*
2  * Block driver for RAW files (posix)
3  *
4  * Copyright (c) 2006 Fabrice Bellard
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to deal
8  * in the Software without restriction, including without limitation the rights
9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10  * copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22  * THE SOFTWARE.
23  */
24 
25 #include "qemu/osdep.h"
26 #include "qapi/error.h"
27 #include "qemu/cutils.h"
28 #include "qemu/error-report.h"
29 #include "block/block-io.h"
30 #include "block/block_int.h"
31 #include "qemu/module.h"
32 #include "qemu/option.h"
33 #include "qemu/units.h"
34 #include "qemu/memalign.h"
35 #include "trace.h"
36 #include "block/thread-pool.h"
37 #include "qemu/iov.h"
38 #include "block/raw-aio.h"
39 #include "qapi/qmp/qdict.h"
40 #include "qapi/qmp/qstring.h"
41 
42 #include "scsi/pr-manager.h"
43 #include "scsi/constants.h"
44 
45 #if defined(__APPLE__) && (__MACH__)
46 #include <sys/ioctl.h>
47 #if defined(HAVE_HOST_BLOCK_DEVICE)
48 #include <paths.h>
49 #include <sys/param.h>
50 #include <sys/mount.h>
51 #include <IOKit/IOKitLib.h>
52 #include <IOKit/IOBSD.h>
53 #include <IOKit/storage/IOMediaBSDClient.h>
54 #include <IOKit/storage/IOMedia.h>
55 #include <IOKit/storage/IOCDMedia.h>
56 //#include <IOKit/storage/IOCDTypes.h>
57 #include <IOKit/storage/IODVDMedia.h>
58 #include <CoreFoundation/CoreFoundation.h>
59 #endif /* defined(HAVE_HOST_BLOCK_DEVICE) */
60 #endif
61 
62 #ifdef __sun__
63 #define _POSIX_PTHREAD_SEMANTICS 1
64 #include <sys/dkio.h>
65 #endif
66 #ifdef __linux__
67 #include <sys/ioctl.h>
68 #include <sys/param.h>
69 #include <sys/syscall.h>
70 #include <sys/vfs.h>
71 #if defined(CONFIG_BLKZONED)
72 #include <linux/blkzoned.h>
73 #endif
74 #include <linux/cdrom.h>
75 #include <linux/fd.h>
76 #include <linux/fs.h>
77 #include <linux/hdreg.h>
78 #include <linux/magic.h>
79 #include <scsi/sg.h>
80 #ifdef __s390__
81 #include <asm/dasd.h>
82 #endif
83 #ifndef FS_NOCOW_FL
84 #define FS_NOCOW_FL                     0x00800000 /* Do not cow file */
85 #endif
86 #endif
87 #if defined(CONFIG_FALLOCATE_PUNCH_HOLE) || defined(CONFIG_FALLOCATE_ZERO_RANGE)
88 #include <linux/falloc.h>
89 #endif
90 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
91 #include <sys/disk.h>
92 #include <sys/cdio.h>
93 #endif
94 
95 #ifdef __OpenBSD__
96 #include <sys/ioctl.h>
97 #include <sys/disklabel.h>
98 #include <sys/dkio.h>
99 #endif
100 
101 #ifdef __NetBSD__
102 #include <sys/ioctl.h>
103 #include <sys/disklabel.h>
104 #include <sys/dkio.h>
105 #include <sys/disk.h>
106 #endif
107 
108 #ifdef __DragonFly__
109 #include <sys/ioctl.h>
110 #include <sys/diskslice.h>
111 #endif
112 
113 /* OS X does not have O_DSYNC */
114 #ifndef O_DSYNC
115 #ifdef O_SYNC
116 #define O_DSYNC O_SYNC
117 #elif defined(O_FSYNC)
118 #define O_DSYNC O_FSYNC
119 #endif
120 #endif
121 
122 /* Approximate O_DIRECT with O_DSYNC if O_DIRECT isn't available */
123 #ifndef O_DIRECT
124 #define O_DIRECT O_DSYNC
125 #endif
126 
127 #define FTYPE_FILE   0
128 #define FTYPE_CD     1
129 
130 #define MAX_BLOCKSIZE	4096
131 
132 /* Posix file locking bytes. Libvirt takes byte 0, we start from higher bytes,
133  * leaving a few more bytes for its future use. */
134 #define RAW_LOCK_PERM_BASE             100
135 #define RAW_LOCK_SHARED_BASE           200
136 
137 typedef struct BDRVRawState {
138     int fd;
139     bool use_lock;
140     int type;
141     int open_flags;
142     size_t buf_align;
143 
144     /* The current permissions. */
145     uint64_t perm;
146     uint64_t shared_perm;
147 
148     /* The perms bits whose corresponding bytes are already locked in
149      * s->fd. */
150     uint64_t locked_perm;
151     uint64_t locked_shared_perm;
152 
153     uint64_t aio_max_batch;
154 
155     int perm_change_fd;
156     int perm_change_flags;
157     BDRVReopenState *reopen_state;
158 
159     bool has_discard:1;
160     bool has_write_zeroes:1;
161     bool use_linux_aio:1;
162     bool use_linux_io_uring:1;
163     int page_cache_inconsistent; /* errno from fdatasync failure */
164     bool has_fallocate;
165     bool needs_alignment;
166     bool force_alignment;
167     bool drop_cache;
168     bool check_cache_dropped;
169     struct {
170         uint64_t discard_nb_ok;
171         uint64_t discard_nb_failed;
172         uint64_t discard_bytes_ok;
173     } stats;
174 
175     PRManager *pr_mgr;
176 } BDRVRawState;
177 
178 typedef struct BDRVRawReopenState {
179     int open_flags;
180     bool drop_cache;
181     bool check_cache_dropped;
182 } BDRVRawReopenState;
183 
184 static int fd_open(BlockDriverState *bs)
185 {
186     BDRVRawState *s = bs->opaque;
187 
188     /* this is just to ensure s->fd is sane (its called by io ops) */
189     if (s->fd >= 0) {
190         return 0;
191     }
192     return -EIO;
193 }
194 
195 static int64_t raw_getlength(BlockDriverState *bs);
196 
197 typedef struct RawPosixAIOData {
198     BlockDriverState *bs;
199     int aio_type;
200     int aio_fildes;
201 
202     off_t aio_offset;
203     uint64_t aio_nbytes;
204 
205     union {
206         struct {
207             struct iovec *iov;
208             int niov;
209         } io;
210         struct {
211             uint64_t cmd;
212             void *buf;
213         } ioctl;
214         struct {
215             int aio_fd2;
216             off_t aio_offset2;
217         } copy_range;
218         struct {
219             PreallocMode prealloc;
220             Error **errp;
221         } truncate;
222         struct {
223             unsigned int *nr_zones;
224             BlockZoneDescriptor *zones;
225         } zone_report;
226         struct {
227             unsigned long op;
228         } zone_mgmt;
229     };
230 } RawPosixAIOData;
231 
232 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
233 static int cdrom_reopen(BlockDriverState *bs);
234 #endif
235 
236 /*
237  * Elide EAGAIN and EACCES details when failing to lock, as this
238  * indicates that the specified file region is already locked by
239  * another process, which is considered a common scenario.
240  */
241 #define raw_lock_error_setg_errno(errp, err, fmt, ...)                  \
242     do {                                                                \
243         if ((err) == EAGAIN || (err) == EACCES) {                       \
244             error_setg((errp), (fmt), ## __VA_ARGS__);                  \
245         } else {                                                        \
246             error_setg_errno((errp), (err), (fmt), ## __VA_ARGS__);     \
247         }                                                               \
248     } while (0)
249 
250 #if defined(__NetBSD__)
251 static int raw_normalize_devicepath(const char **filename, Error **errp)
252 {
253     static char namebuf[PATH_MAX];
254     const char *dp, *fname;
255     struct stat sb;
256 
257     fname = *filename;
258     dp = strrchr(fname, '/');
259     if (lstat(fname, &sb) < 0) {
260         error_setg_file_open(errp, errno, fname);
261         return -errno;
262     }
263 
264     if (!S_ISBLK(sb.st_mode)) {
265         return 0;
266     }
267 
268     if (dp == NULL) {
269         snprintf(namebuf, PATH_MAX, "r%s", fname);
270     } else {
271         snprintf(namebuf, PATH_MAX, "%.*s/r%s",
272             (int)(dp - fname), fname, dp + 1);
273     }
274     *filename = namebuf;
275     warn_report("%s is a block device, using %s", fname, *filename);
276 
277     return 0;
278 }
279 #else
280 static int raw_normalize_devicepath(const char **filename, Error **errp)
281 {
282     return 0;
283 }
284 #endif
285 
286 /*
287  * Get logical block size via ioctl. On success store it in @sector_size_p.
288  */
289 static int probe_logical_blocksize(int fd, unsigned int *sector_size_p)
290 {
291     unsigned int sector_size;
292     bool success = false;
293     int i;
294 
295     errno = ENOTSUP;
296     static const unsigned long ioctl_list[] = {
297 #ifdef BLKSSZGET
298         BLKSSZGET,
299 #endif
300 #ifdef DKIOCGETBLOCKSIZE
301         DKIOCGETBLOCKSIZE,
302 #endif
303 #ifdef DIOCGSECTORSIZE
304         DIOCGSECTORSIZE,
305 #endif
306     };
307 
308     /* Try a few ioctls to get the right size */
309     for (i = 0; i < (int)ARRAY_SIZE(ioctl_list); i++) {
310         if (ioctl(fd, ioctl_list[i], &sector_size) >= 0) {
311             *sector_size_p = sector_size;
312             success = true;
313         }
314     }
315 
316     return success ? 0 : -errno;
317 }
318 
319 /**
320  * Get physical block size of @fd.
321  * On success, store it in @blk_size and return 0.
322  * On failure, return -errno.
323  */
324 static int probe_physical_blocksize(int fd, unsigned int *blk_size)
325 {
326 #ifdef BLKPBSZGET
327     if (ioctl(fd, BLKPBSZGET, blk_size) < 0) {
328         return -errno;
329     }
330     return 0;
331 #else
332     return -ENOTSUP;
333 #endif
334 }
335 
336 /*
337  * Returns true if no alignment restrictions are necessary even for files
338  * opened with O_DIRECT.
339  *
340  * raw_probe_alignment() probes the required alignment and assume that 1 means
341  * the probing failed, so it falls back to a safe default of 4k. This can be
342  * avoided if we know that byte alignment is okay for the file.
343  */
344 static bool dio_byte_aligned(int fd)
345 {
346 #ifdef __linux__
347     struct statfs buf;
348     int ret;
349 
350     ret = fstatfs(fd, &buf);
351     if (ret == 0 && buf.f_type == NFS_SUPER_MAGIC) {
352         return true;
353     }
354 #endif
355     return false;
356 }
357 
358 static bool raw_needs_alignment(BlockDriverState *bs)
359 {
360     BDRVRawState *s = bs->opaque;
361 
362     if ((bs->open_flags & BDRV_O_NOCACHE) != 0 && !dio_byte_aligned(s->fd)) {
363         return true;
364     }
365 
366     return s->force_alignment;
367 }
368 
369 /* Check if read is allowed with given memory buffer and length.
370  *
371  * This function is used to check O_DIRECT memory buffer and request alignment.
372  */
373 static bool raw_is_io_aligned(int fd, void *buf, size_t len)
374 {
375     ssize_t ret = pread(fd, buf, len, 0);
376 
377     if (ret >= 0) {
378         return true;
379     }
380 
381 #ifdef __linux__
382     /* The Linux kernel returns EINVAL for misaligned O_DIRECT reads.  Ignore
383      * other errors (e.g. real I/O error), which could happen on a failed
384      * drive, since we only care about probing alignment.
385      */
386     if (errno != EINVAL) {
387         return true;
388     }
389 #endif
390 
391     return false;
392 }
393 
394 static void raw_probe_alignment(BlockDriverState *bs, int fd, Error **errp)
395 {
396     BDRVRawState *s = bs->opaque;
397     char *buf;
398     size_t max_align = MAX(MAX_BLOCKSIZE, qemu_real_host_page_size());
399     size_t alignments[] = {1, 512, 1024, 2048, 4096};
400 
401     /* For SCSI generic devices the alignment is not really used.
402        With buffered I/O, we don't have any restrictions. */
403     if (bdrv_is_sg(bs) || !s->needs_alignment) {
404         bs->bl.request_alignment = 1;
405         s->buf_align = 1;
406         return;
407     }
408 
409     bs->bl.request_alignment = 0;
410     s->buf_align = 0;
411     /* Let's try to use the logical blocksize for the alignment. */
412     if (probe_logical_blocksize(fd, &bs->bl.request_alignment) < 0) {
413         bs->bl.request_alignment = 0;
414     }
415 
416 #ifdef __linux__
417     /*
418      * The XFS ioctl definitions are shipped in extra packages that might
419      * not always be available. Since we just need the XFS_IOC_DIOINFO ioctl
420      * here, we simply use our own definition instead:
421      */
422     struct xfs_dioattr {
423         uint32_t d_mem;
424         uint32_t d_miniosz;
425         uint32_t d_maxiosz;
426     } da;
427     if (ioctl(fd, _IOR('X', 30, struct xfs_dioattr), &da) >= 0) {
428         bs->bl.request_alignment = da.d_miniosz;
429         /* The kernel returns wrong information for d_mem */
430         /* s->buf_align = da.d_mem; */
431     }
432 #endif
433 
434     /*
435      * If we could not get the sizes so far, we can only guess them. First try
436      * to detect request alignment, since it is more likely to succeed. Then
437      * try to detect buf_align, which cannot be detected in some cases (e.g.
438      * Gluster). If buf_align cannot be detected, we fallback to the value of
439      * request_alignment.
440      */
441 
442     if (!bs->bl.request_alignment) {
443         int i;
444         size_t align;
445         buf = qemu_memalign(max_align, max_align);
446         for (i = 0; i < ARRAY_SIZE(alignments); i++) {
447             align = alignments[i];
448             if (raw_is_io_aligned(fd, buf, align)) {
449                 /* Fallback to safe value. */
450                 bs->bl.request_alignment = (align != 1) ? align : max_align;
451                 break;
452             }
453         }
454         qemu_vfree(buf);
455     }
456 
457     if (!s->buf_align) {
458         int i;
459         size_t align;
460         buf = qemu_memalign(max_align, 2 * max_align);
461         for (i = 0; i < ARRAY_SIZE(alignments); i++) {
462             align = alignments[i];
463             if (raw_is_io_aligned(fd, buf + align, max_align)) {
464                 /* Fallback to request_alignment. */
465                 s->buf_align = (align != 1) ? align : bs->bl.request_alignment;
466                 break;
467             }
468         }
469         qemu_vfree(buf);
470     }
471 
472     if (!s->buf_align || !bs->bl.request_alignment) {
473         error_setg(errp, "Could not find working O_DIRECT alignment");
474         error_append_hint(errp, "Try cache.direct=off\n");
475     }
476 }
477 
478 static int check_hdev_writable(int fd)
479 {
480 #if defined(BLKROGET)
481     /* Linux block devices can be configured "read-only" using blockdev(8).
482      * This is independent of device node permissions and therefore open(2)
483      * with O_RDWR succeeds.  Actual writes fail with EPERM.
484      *
485      * bdrv_open() is supposed to fail if the disk is read-only.  Explicitly
486      * check for read-only block devices so that Linux block devices behave
487      * properly.
488      */
489     struct stat st;
490     int readonly = 0;
491 
492     if (fstat(fd, &st)) {
493         return -errno;
494     }
495 
496     if (!S_ISBLK(st.st_mode)) {
497         return 0;
498     }
499 
500     if (ioctl(fd, BLKROGET, &readonly) < 0) {
501         return -errno;
502     }
503 
504     if (readonly) {
505         return -EACCES;
506     }
507 #endif /* defined(BLKROGET) */
508     return 0;
509 }
510 
511 static void raw_parse_flags(int bdrv_flags, int *open_flags, bool has_writers)
512 {
513     bool read_write = false;
514     assert(open_flags != NULL);
515 
516     *open_flags |= O_BINARY;
517     *open_flags &= ~O_ACCMODE;
518 
519     if (bdrv_flags & BDRV_O_AUTO_RDONLY) {
520         read_write = has_writers;
521     } else if (bdrv_flags & BDRV_O_RDWR) {
522         read_write = true;
523     }
524 
525     if (read_write) {
526         *open_flags |= O_RDWR;
527     } else {
528         *open_flags |= O_RDONLY;
529     }
530 
531     /* Use O_DSYNC for write-through caching, no flags for write-back caching,
532      * and O_DIRECT for no caching. */
533     if ((bdrv_flags & BDRV_O_NOCACHE)) {
534         *open_flags |= O_DIRECT;
535     }
536 }
537 
538 static void raw_parse_filename(const char *filename, QDict *options,
539                                Error **errp)
540 {
541     bdrv_parse_filename_strip_prefix(filename, "file:", options);
542 }
543 
544 static QemuOptsList raw_runtime_opts = {
545     .name = "raw",
546     .head = QTAILQ_HEAD_INITIALIZER(raw_runtime_opts.head),
547     .desc = {
548         {
549             .name = "filename",
550             .type = QEMU_OPT_STRING,
551             .help = "File name of the image",
552         },
553         {
554             .name = "aio",
555             .type = QEMU_OPT_STRING,
556             .help = "host AIO implementation (threads, native, io_uring)",
557         },
558         {
559             .name = "aio-max-batch",
560             .type = QEMU_OPT_NUMBER,
561             .help = "AIO max batch size (0 = auto handled by AIO backend, default: 0)",
562         },
563         {
564             .name = "locking",
565             .type = QEMU_OPT_STRING,
566             .help = "file locking mode (on/off/auto, default: auto)",
567         },
568         {
569             .name = "pr-manager",
570             .type = QEMU_OPT_STRING,
571             .help = "id of persistent reservation manager object (default: none)",
572         },
573 #if defined(__linux__)
574         {
575             .name = "drop-cache",
576             .type = QEMU_OPT_BOOL,
577             .help = "invalidate page cache during live migration (default: on)",
578         },
579 #endif
580         {
581             .name = "x-check-cache-dropped",
582             .type = QEMU_OPT_BOOL,
583             .help = "check that page cache was dropped on live migration (default: off)"
584         },
585         { /* end of list */ }
586     },
587 };
588 
589 static const char *const mutable_opts[] = { "x-check-cache-dropped", NULL };
590 
591 static int raw_open_common(BlockDriverState *bs, QDict *options,
592                            int bdrv_flags, int open_flags,
593                            bool device, Error **errp)
594 {
595     BDRVRawState *s = bs->opaque;
596     QemuOpts *opts;
597     Error *local_err = NULL;
598     const char *filename = NULL;
599     const char *str;
600     BlockdevAioOptions aio, aio_default;
601     int fd, ret;
602     struct stat st;
603     OnOffAuto locking;
604 
605     opts = qemu_opts_create(&raw_runtime_opts, NULL, 0, &error_abort);
606     if (!qemu_opts_absorb_qdict(opts, options, errp)) {
607         ret = -EINVAL;
608         goto fail;
609     }
610 
611     filename = qemu_opt_get(opts, "filename");
612 
613     ret = raw_normalize_devicepath(&filename, errp);
614     if (ret != 0) {
615         goto fail;
616     }
617 
618     if (bdrv_flags & BDRV_O_NATIVE_AIO) {
619         aio_default = BLOCKDEV_AIO_OPTIONS_NATIVE;
620 #ifdef CONFIG_LINUX_IO_URING
621     } else if (bdrv_flags & BDRV_O_IO_URING) {
622         aio_default = BLOCKDEV_AIO_OPTIONS_IO_URING;
623 #endif
624     } else {
625         aio_default = BLOCKDEV_AIO_OPTIONS_THREADS;
626     }
627 
628     aio = qapi_enum_parse(&BlockdevAioOptions_lookup,
629                           qemu_opt_get(opts, "aio"),
630                           aio_default, &local_err);
631     if (local_err) {
632         error_propagate(errp, local_err);
633         ret = -EINVAL;
634         goto fail;
635     }
636 
637     s->use_linux_aio = (aio == BLOCKDEV_AIO_OPTIONS_NATIVE);
638 #ifdef CONFIG_LINUX_IO_URING
639     s->use_linux_io_uring = (aio == BLOCKDEV_AIO_OPTIONS_IO_URING);
640 #endif
641 
642     s->aio_max_batch = qemu_opt_get_number(opts, "aio-max-batch", 0);
643 
644     locking = qapi_enum_parse(&OnOffAuto_lookup,
645                               qemu_opt_get(opts, "locking"),
646                               ON_OFF_AUTO_AUTO, &local_err);
647     if (local_err) {
648         error_propagate(errp, local_err);
649         ret = -EINVAL;
650         goto fail;
651     }
652     switch (locking) {
653     case ON_OFF_AUTO_ON:
654         s->use_lock = true;
655         if (!qemu_has_ofd_lock()) {
656             warn_report("File lock requested but OFD locking syscall is "
657                         "unavailable, falling back to POSIX file locks");
658             error_printf("Due to the implementation, locks can be lost "
659                          "unexpectedly.\n");
660         }
661         break;
662     case ON_OFF_AUTO_OFF:
663         s->use_lock = false;
664         break;
665     case ON_OFF_AUTO_AUTO:
666         s->use_lock = qemu_has_ofd_lock();
667         break;
668     default:
669         abort();
670     }
671 
672     str = qemu_opt_get(opts, "pr-manager");
673     if (str) {
674         s->pr_mgr = pr_manager_lookup(str, &local_err);
675         if (local_err) {
676             error_propagate(errp, local_err);
677             ret = -EINVAL;
678             goto fail;
679         }
680     }
681 
682     s->drop_cache = qemu_opt_get_bool(opts, "drop-cache", true);
683     s->check_cache_dropped = qemu_opt_get_bool(opts, "x-check-cache-dropped",
684                                                false);
685 
686     s->open_flags = open_flags;
687     raw_parse_flags(bdrv_flags, &s->open_flags, false);
688 
689     s->fd = -1;
690     fd = qemu_open(filename, s->open_flags, errp);
691     ret = fd < 0 ? -errno : 0;
692 
693     if (ret < 0) {
694         if (ret == -EROFS) {
695             ret = -EACCES;
696         }
697         goto fail;
698     }
699     s->fd = fd;
700 
701     /* Check s->open_flags rather than bdrv_flags due to auto-read-only */
702     if (s->open_flags & O_RDWR) {
703         ret = check_hdev_writable(s->fd);
704         if (ret < 0) {
705             error_setg_errno(errp, -ret, "The device is not writable");
706             goto fail;
707         }
708     }
709 
710     s->perm = 0;
711     s->shared_perm = BLK_PERM_ALL;
712 
713 #ifdef CONFIG_LINUX_AIO
714      /* Currently Linux does AIO only for files opened with O_DIRECT */
715     if (s->use_linux_aio) {
716         if (!(s->open_flags & O_DIRECT)) {
717             error_setg(errp, "aio=native was specified, but it requires "
718                              "cache.direct=on, which was not specified.");
719             ret = -EINVAL;
720             goto fail;
721         }
722         if (!aio_setup_linux_aio(bdrv_get_aio_context(bs), errp)) {
723             error_prepend(errp, "Unable to use native AIO: ");
724             goto fail;
725         }
726     }
727 #else
728     if (s->use_linux_aio) {
729         error_setg(errp, "aio=native was specified, but is not supported "
730                          "in this build.");
731         ret = -EINVAL;
732         goto fail;
733     }
734 #endif /* !defined(CONFIG_LINUX_AIO) */
735 
736 #ifdef CONFIG_LINUX_IO_URING
737     if (s->use_linux_io_uring) {
738         if (!aio_setup_linux_io_uring(bdrv_get_aio_context(bs), errp)) {
739             error_prepend(errp, "Unable to use io_uring: ");
740             goto fail;
741         }
742     }
743 #else
744     if (s->use_linux_io_uring) {
745         error_setg(errp, "aio=io_uring was specified, but is not supported "
746                          "in this build.");
747         ret = -EINVAL;
748         goto fail;
749     }
750 #endif /* !defined(CONFIG_LINUX_IO_URING) */
751 
752     s->has_discard = true;
753     s->has_write_zeroes = true;
754 
755     if (fstat(s->fd, &st) < 0) {
756         ret = -errno;
757         error_setg_errno(errp, errno, "Could not stat file");
758         goto fail;
759     }
760 
761     if (!device) {
762         if (!S_ISREG(st.st_mode)) {
763             error_setg(errp, "'%s' driver requires '%s' to be a regular file",
764                        bs->drv->format_name, bs->filename);
765             ret = -EINVAL;
766             goto fail;
767         } else {
768             s->has_fallocate = true;
769         }
770     } else {
771         if (!(S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
772             error_setg(errp, "'%s' driver requires '%s' to be either "
773                        "a character or block device",
774                        bs->drv->format_name, bs->filename);
775             ret = -EINVAL;
776             goto fail;
777         }
778     }
779 #ifdef CONFIG_BLKZONED
780     /*
781      * The kernel page cache does not reliably work for writes to SWR zones
782      * of zoned block device because it can not guarantee the order of writes.
783      */
784     if ((bs->bl.zoned != BLK_Z_NONE) &&
785         (!(s->open_flags & O_DIRECT))) {
786         error_setg(errp, "The driver supports zoned devices, and it requires "
787                          "cache.direct=on, which was not specified.");
788         return -EINVAL; /* No host kernel page cache */
789     }
790 #endif
791 
792     if (S_ISBLK(st.st_mode)) {
793 #ifdef __linux__
794         /* On Linux 3.10, BLKDISCARD leaves stale data in the page cache.  Do
795          * not rely on the contents of discarded blocks unless using O_DIRECT.
796          * Same for BLKZEROOUT.
797          */
798         if (!(bs->open_flags & BDRV_O_NOCACHE)) {
799             s->has_write_zeroes = false;
800         }
801 #endif
802     }
803 #ifdef __FreeBSD__
804     if (S_ISCHR(st.st_mode)) {
805         /*
806          * The file is a char device (disk), which on FreeBSD isn't behind
807          * a pager, so force all requests to be aligned. This is needed
808          * so QEMU makes sure all IO operations on the device are aligned
809          * to sector size, or else FreeBSD will reject them with EINVAL.
810          */
811         s->force_alignment = true;
812     }
813 #endif
814     s->needs_alignment = raw_needs_alignment(bs);
815 
816     bs->supported_zero_flags = BDRV_REQ_MAY_UNMAP | BDRV_REQ_NO_FALLBACK;
817     if (S_ISREG(st.st_mode)) {
818         /* When extending regular files, we get zeros from the OS */
819         bs->supported_truncate_flags = BDRV_REQ_ZERO_WRITE;
820     }
821     ret = 0;
822 fail:
823     if (ret < 0 && s->fd != -1) {
824         qemu_close(s->fd);
825     }
826     if (filename && (bdrv_flags & BDRV_O_TEMPORARY)) {
827         unlink(filename);
828     }
829     qemu_opts_del(opts);
830     return ret;
831 }
832 
833 static int raw_open(BlockDriverState *bs, QDict *options, int flags,
834                     Error **errp)
835 {
836     BDRVRawState *s = bs->opaque;
837 
838     s->type = FTYPE_FILE;
839     return raw_open_common(bs, options, flags, 0, false, errp);
840 }
841 
842 typedef enum {
843     RAW_PL_PREPARE,
844     RAW_PL_COMMIT,
845     RAW_PL_ABORT,
846 } RawPermLockOp;
847 
848 #define PERM_FOREACH(i) \
849     for ((i) = 0; (1ULL << (i)) <= BLK_PERM_ALL; i++)
850 
851 /* Lock bytes indicated by @perm_lock_bits and @shared_perm_lock_bits in the
852  * file; if @unlock == true, also unlock the unneeded bytes.
853  * @shared_perm_lock_bits is the mask of all permissions that are NOT shared.
854  */
855 static int raw_apply_lock_bytes(BDRVRawState *s, int fd,
856                                 uint64_t perm_lock_bits,
857                                 uint64_t shared_perm_lock_bits,
858                                 bool unlock, Error **errp)
859 {
860     int ret;
861     int i;
862     uint64_t locked_perm, locked_shared_perm;
863 
864     if (s) {
865         locked_perm = s->locked_perm;
866         locked_shared_perm = s->locked_shared_perm;
867     } else {
868         /*
869          * We don't have the previous bits, just lock/unlock for each of the
870          * requested bits.
871          */
872         if (unlock) {
873             locked_perm = BLK_PERM_ALL;
874             locked_shared_perm = BLK_PERM_ALL;
875         } else {
876             locked_perm = 0;
877             locked_shared_perm = 0;
878         }
879     }
880 
881     PERM_FOREACH(i) {
882         int off = RAW_LOCK_PERM_BASE + i;
883         uint64_t bit = (1ULL << i);
884         if ((perm_lock_bits & bit) && !(locked_perm & bit)) {
885             ret = qemu_lock_fd(fd, off, 1, false);
886             if (ret) {
887                 raw_lock_error_setg_errno(errp, -ret, "Failed to lock byte %d",
888                                           off);
889                 return ret;
890             } else if (s) {
891                 s->locked_perm |= bit;
892             }
893         } else if (unlock && (locked_perm & bit) && !(perm_lock_bits & bit)) {
894             ret = qemu_unlock_fd(fd, off, 1);
895             if (ret) {
896                 error_setg_errno(errp, -ret, "Failed to unlock byte %d", off);
897                 return ret;
898             } else if (s) {
899                 s->locked_perm &= ~bit;
900             }
901         }
902     }
903     PERM_FOREACH(i) {
904         int off = RAW_LOCK_SHARED_BASE + i;
905         uint64_t bit = (1ULL << i);
906         if ((shared_perm_lock_bits & bit) && !(locked_shared_perm & bit)) {
907             ret = qemu_lock_fd(fd, off, 1, false);
908             if (ret) {
909                 raw_lock_error_setg_errno(errp, -ret, "Failed to lock byte %d",
910                                           off);
911                 return ret;
912             } else if (s) {
913                 s->locked_shared_perm |= bit;
914             }
915         } else if (unlock && (locked_shared_perm & bit) &&
916                    !(shared_perm_lock_bits & bit)) {
917             ret = qemu_unlock_fd(fd, off, 1);
918             if (ret) {
919                 error_setg_errno(errp, -ret, "Failed to unlock byte %d", off);
920                 return ret;
921             } else if (s) {
922                 s->locked_shared_perm &= ~bit;
923             }
924         }
925     }
926     return 0;
927 }
928 
929 /* Check "unshared" bytes implied by @perm and ~@shared_perm in the file. */
930 static int raw_check_lock_bytes(int fd, uint64_t perm, uint64_t shared_perm,
931                                 Error **errp)
932 {
933     int ret;
934     int i;
935 
936     PERM_FOREACH(i) {
937         int off = RAW_LOCK_SHARED_BASE + i;
938         uint64_t p = 1ULL << i;
939         if (perm & p) {
940             ret = qemu_lock_fd_test(fd, off, 1, true);
941             if (ret) {
942                 char *perm_name = bdrv_perm_names(p);
943 
944                 raw_lock_error_setg_errno(errp, -ret,
945                                           "Failed to get \"%s\" lock",
946                                           perm_name);
947                 g_free(perm_name);
948                 return ret;
949             }
950         }
951     }
952     PERM_FOREACH(i) {
953         int off = RAW_LOCK_PERM_BASE + i;
954         uint64_t p = 1ULL << i;
955         if (!(shared_perm & p)) {
956             ret = qemu_lock_fd_test(fd, off, 1, true);
957             if (ret) {
958                 char *perm_name = bdrv_perm_names(p);
959 
960                 raw_lock_error_setg_errno(errp, -ret,
961                                           "Failed to get shared \"%s\" lock",
962                                           perm_name);
963                 g_free(perm_name);
964                 return ret;
965             }
966         }
967     }
968     return 0;
969 }
970 
971 static int raw_handle_perm_lock(BlockDriverState *bs,
972                                 RawPermLockOp op,
973                                 uint64_t new_perm, uint64_t new_shared,
974                                 Error **errp)
975 {
976     BDRVRawState *s = bs->opaque;
977     int ret = 0;
978     Error *local_err = NULL;
979 
980     if (!s->use_lock) {
981         return 0;
982     }
983 
984     if (bdrv_get_flags(bs) & BDRV_O_INACTIVE) {
985         return 0;
986     }
987 
988     switch (op) {
989     case RAW_PL_PREPARE:
990         if ((s->perm | new_perm) == s->perm &&
991             (s->shared_perm & new_shared) == s->shared_perm)
992         {
993             /*
994              * We are going to unlock bytes, it should not fail. If it fail due
995              * to some fs-dependent permission-unrelated reasons (which occurs
996              * sometimes on NFS and leads to abort in bdrv_replace_child) we
997              * can't prevent such errors by any check here. And we ignore them
998              * anyway in ABORT and COMMIT.
999              */
1000             return 0;
1001         }
1002         ret = raw_apply_lock_bytes(s, s->fd, s->perm | new_perm,
1003                                    ~s->shared_perm | ~new_shared,
1004                                    false, errp);
1005         if (!ret) {
1006             ret = raw_check_lock_bytes(s->fd, new_perm, new_shared, errp);
1007             if (!ret) {
1008                 return 0;
1009             }
1010             error_append_hint(errp,
1011                               "Is another process using the image [%s]?\n",
1012                               bs->filename);
1013         }
1014         /* fall through to unlock bytes. */
1015     case RAW_PL_ABORT:
1016         raw_apply_lock_bytes(s, s->fd, s->perm, ~s->shared_perm,
1017                              true, &local_err);
1018         if (local_err) {
1019             /* Theoretically the above call only unlocks bytes and it cannot
1020              * fail. Something weird happened, report it.
1021              */
1022             warn_report_err(local_err);
1023         }
1024         break;
1025     case RAW_PL_COMMIT:
1026         raw_apply_lock_bytes(s, s->fd, new_perm, ~new_shared,
1027                              true, &local_err);
1028         if (local_err) {
1029             /* Theoretically the above call only unlocks bytes and it cannot
1030              * fail. Something weird happened, report it.
1031              */
1032             warn_report_err(local_err);
1033         }
1034         break;
1035     }
1036     return ret;
1037 }
1038 
1039 /* Sets a specific flag */
1040 static int fcntl_setfl(int fd, int flag)
1041 {
1042     int flags;
1043 
1044     flags = fcntl(fd, F_GETFL);
1045     if (flags == -1) {
1046         return -errno;
1047     }
1048     if (fcntl(fd, F_SETFL, flags | flag) == -1) {
1049         return -errno;
1050     }
1051     return 0;
1052 }
1053 
1054 static int raw_reconfigure_getfd(BlockDriverState *bs, int flags,
1055                                  int *open_flags, uint64_t perm, bool force_dup,
1056                                  Error **errp)
1057 {
1058     BDRVRawState *s = bs->opaque;
1059     int fd = -1;
1060     int ret;
1061     bool has_writers = perm &
1062         (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED | BLK_PERM_RESIZE);
1063     int fcntl_flags = O_APPEND | O_NONBLOCK;
1064 #ifdef O_NOATIME
1065     fcntl_flags |= O_NOATIME;
1066 #endif
1067 
1068     *open_flags = 0;
1069     if (s->type == FTYPE_CD) {
1070         *open_flags |= O_NONBLOCK;
1071     }
1072 
1073     raw_parse_flags(flags, open_flags, has_writers);
1074 
1075 #ifdef O_ASYNC
1076     /* Not all operating systems have O_ASYNC, and those that don't
1077      * will not let us track the state into rs->open_flags (typically
1078      * you achieve the same effect with an ioctl, for example I_SETSIG
1079      * on Solaris). But we do not use O_ASYNC, so that's fine.
1080      */
1081     assert((s->open_flags & O_ASYNC) == 0);
1082 #endif
1083 
1084     if (!force_dup && *open_flags == s->open_flags) {
1085         /* We're lucky, the existing fd is fine */
1086         return s->fd;
1087     }
1088 
1089     if ((*open_flags & ~fcntl_flags) == (s->open_flags & ~fcntl_flags)) {
1090         /* dup the original fd */
1091         fd = qemu_dup(s->fd);
1092         if (fd >= 0) {
1093             ret = fcntl_setfl(fd, *open_flags);
1094             if (ret) {
1095                 qemu_close(fd);
1096                 fd = -1;
1097             }
1098         }
1099     }
1100 
1101     /* If we cannot use fcntl, or fcntl failed, fall back to qemu_open() */
1102     if (fd == -1) {
1103         const char *normalized_filename = bs->filename;
1104         ret = raw_normalize_devicepath(&normalized_filename, errp);
1105         if (ret >= 0) {
1106             fd = qemu_open(normalized_filename, *open_flags, errp);
1107             if (fd == -1) {
1108                 return -1;
1109             }
1110         }
1111     }
1112 
1113     if (fd != -1 && (*open_flags & O_RDWR)) {
1114         ret = check_hdev_writable(fd);
1115         if (ret < 0) {
1116             qemu_close(fd);
1117             error_setg_errno(errp, -ret, "The device is not writable");
1118             return -1;
1119         }
1120     }
1121 
1122     return fd;
1123 }
1124 
1125 static int raw_reopen_prepare(BDRVReopenState *state,
1126                               BlockReopenQueue *queue, Error **errp)
1127 {
1128     BDRVRawState *s;
1129     BDRVRawReopenState *rs;
1130     QemuOpts *opts;
1131     int ret;
1132 
1133     assert(state != NULL);
1134     assert(state->bs != NULL);
1135 
1136     s = state->bs->opaque;
1137 
1138     state->opaque = g_new0(BDRVRawReopenState, 1);
1139     rs = state->opaque;
1140 
1141     /* Handle options changes */
1142     opts = qemu_opts_create(&raw_runtime_opts, NULL, 0, &error_abort);
1143     if (!qemu_opts_absorb_qdict(opts, state->options, errp)) {
1144         ret = -EINVAL;
1145         goto out;
1146     }
1147 
1148     rs->drop_cache = qemu_opt_get_bool_del(opts, "drop-cache", true);
1149     rs->check_cache_dropped =
1150         qemu_opt_get_bool_del(opts, "x-check-cache-dropped", false);
1151 
1152     /* This driver's reopen function doesn't currently allow changing
1153      * other options, so let's put them back in the original QDict and
1154      * bdrv_reopen_prepare() will detect changes and complain. */
1155     qemu_opts_to_qdict(opts, state->options);
1156 
1157     /*
1158      * As part of reopen prepare we also want to create new fd by
1159      * raw_reconfigure_getfd(). But it wants updated "perm", when in
1160      * bdrv_reopen_multiple() .bdrv_reopen_prepare() callback called prior to
1161      * permission update. Happily, permission update is always a part
1162      * (a separate stage) of bdrv_reopen_multiple() so we can rely on this
1163      * fact and reconfigure fd in raw_check_perm().
1164      */
1165 
1166     s->reopen_state = state;
1167     ret = 0;
1168 
1169 out:
1170     qemu_opts_del(opts);
1171     return ret;
1172 }
1173 
1174 static void raw_reopen_commit(BDRVReopenState *state)
1175 {
1176     BDRVRawReopenState *rs = state->opaque;
1177     BDRVRawState *s = state->bs->opaque;
1178 
1179     s->drop_cache = rs->drop_cache;
1180     s->check_cache_dropped = rs->check_cache_dropped;
1181     s->open_flags = rs->open_flags;
1182     g_free(state->opaque);
1183     state->opaque = NULL;
1184 
1185     assert(s->reopen_state == state);
1186     s->reopen_state = NULL;
1187 }
1188 
1189 
1190 static void raw_reopen_abort(BDRVReopenState *state)
1191 {
1192     BDRVRawReopenState *rs = state->opaque;
1193     BDRVRawState *s = state->bs->opaque;
1194 
1195      /* nothing to do if NULL, we didn't get far enough */
1196     if (rs == NULL) {
1197         return;
1198     }
1199 
1200     g_free(state->opaque);
1201     state->opaque = NULL;
1202 
1203     assert(s->reopen_state == state);
1204     s->reopen_state = NULL;
1205 }
1206 
1207 static int hdev_get_max_hw_transfer(int fd, struct stat *st)
1208 {
1209 #ifdef BLKSECTGET
1210     if (S_ISBLK(st->st_mode)) {
1211         unsigned short max_sectors = 0;
1212         if (ioctl(fd, BLKSECTGET, &max_sectors) == 0) {
1213             return max_sectors * 512;
1214         }
1215     } else {
1216         int max_bytes = 0;
1217         if (ioctl(fd, BLKSECTGET, &max_bytes) == 0) {
1218             return max_bytes;
1219         }
1220     }
1221     return -errno;
1222 #else
1223     return -ENOSYS;
1224 #endif
1225 }
1226 
1227 /*
1228  * Get a sysfs attribute value as character string.
1229  */
1230 #ifdef CONFIG_LINUX
1231 static int get_sysfs_str_val(struct stat *st, const char *attribute,
1232                              char **val) {
1233     g_autofree char *sysfspath = NULL;
1234     size_t len;
1235 
1236     if (!S_ISBLK(st->st_mode)) {
1237         return -ENOTSUP;
1238     }
1239 
1240     sysfspath = g_strdup_printf("/sys/dev/block/%u:%u/queue/%s",
1241                                 major(st->st_rdev), minor(st->st_rdev),
1242                                 attribute);
1243     if (!g_file_get_contents(sysfspath, val, &len, NULL)) {
1244         return -ENOENT;
1245     }
1246 
1247     /* The file is ended with '\n' */
1248     char *p;
1249     p = *val;
1250     if (*(p + len - 1) == '\n') {
1251         *(p + len - 1) = '\0';
1252     }
1253     return 0;
1254 }
1255 #endif
1256 
1257 #if defined(CONFIG_BLKZONED)
1258 static int get_sysfs_zoned_model(struct stat *st, BlockZoneModel *zoned)
1259 {
1260     g_autofree char *val = NULL;
1261     int ret;
1262 
1263     ret = get_sysfs_str_val(st, "zoned", &val);
1264     if (ret < 0) {
1265         return ret;
1266     }
1267 
1268     if (strcmp(val, "host-managed") == 0) {
1269         *zoned = BLK_Z_HM;
1270     } else if (strcmp(val, "host-aware") == 0) {
1271         *zoned = BLK_Z_HA;
1272     } else if (strcmp(val, "none") == 0) {
1273         *zoned = BLK_Z_NONE;
1274     } else {
1275         return -ENOTSUP;
1276     }
1277     return 0;
1278 }
1279 #endif /* defined(CONFIG_BLKZONED) */
1280 
1281 /*
1282  * Get a sysfs attribute value as a long integer.
1283  */
1284 #ifdef CONFIG_LINUX
1285 static long get_sysfs_long_val(struct stat *st, const char *attribute)
1286 {
1287     g_autofree char *str = NULL;
1288     const char *end;
1289     long val;
1290     int ret;
1291 
1292     ret = get_sysfs_str_val(st, attribute, &str);
1293     if (ret < 0) {
1294         return ret;
1295     }
1296 
1297     /* The file is ended with '\n', pass 'end' to accept that. */
1298     ret = qemu_strtol(str, &end, 10, &val);
1299     if (ret == 0 && end && *end == '\0') {
1300         ret = val;
1301     }
1302     return ret;
1303 }
1304 #endif
1305 
1306 static int hdev_get_max_segments(int fd, struct stat *st)
1307 {
1308 #ifdef CONFIG_LINUX
1309     int ret;
1310 
1311     if (S_ISCHR(st->st_mode)) {
1312         if (ioctl(fd, SG_GET_SG_TABLESIZE, &ret) == 0) {
1313             return ret;
1314         }
1315         return -ENOTSUP;
1316     }
1317     return get_sysfs_long_val(st, "max_segments");
1318 #else
1319     return -ENOTSUP;
1320 #endif
1321 }
1322 
1323 #if defined(CONFIG_BLKZONED)
1324 /*
1325  * If the reset_all flag is true, then the wps of zone whose state is
1326  * not readonly or offline should be all reset to the start sector.
1327  * Else, take the real wp of the device.
1328  */
1329 static int get_zones_wp(BlockDriverState *bs, int fd, int64_t offset,
1330                         unsigned int nrz, bool reset_all)
1331 {
1332     struct blk_zone *blkz;
1333     size_t rep_size;
1334     uint64_t sector = offset >> BDRV_SECTOR_BITS;
1335     BlockZoneWps *wps = bs->wps;
1336     unsigned int j = offset / bs->bl.zone_size;
1337     unsigned int n = 0, i = 0;
1338     int ret;
1339     rep_size = sizeof(struct blk_zone_report) + nrz * sizeof(struct blk_zone);
1340     g_autofree struct blk_zone_report *rep = NULL;
1341 
1342     rep = g_malloc(rep_size);
1343     blkz = (struct blk_zone *)(rep + 1);
1344     while (n < nrz) {
1345         memset(rep, 0, rep_size);
1346         rep->sector = sector;
1347         rep->nr_zones = nrz - n;
1348 
1349         do {
1350             ret = ioctl(fd, BLKREPORTZONE, rep);
1351         } while (ret != 0 && errno == EINTR);
1352         if (ret != 0) {
1353             error_report("%d: ioctl BLKREPORTZONE at %" PRId64 " failed %d",
1354                     fd, offset, errno);
1355             return -errno;
1356         }
1357 
1358         if (!rep->nr_zones) {
1359             break;
1360         }
1361 
1362         for (i = 0; i < rep->nr_zones; ++i, ++n, ++j) {
1363             /*
1364              * The wp tracking cares only about sequential writes required and
1365              * sequential write preferred zones so that the wp can advance to
1366              * the right location.
1367              * Use the most significant bit of the wp location to indicate the
1368              * zone type: 0 for SWR/SWP zones and 1 for conventional zones.
1369              */
1370             if (blkz[i].type == BLK_ZONE_TYPE_CONVENTIONAL) {
1371                 wps->wp[j] |= 1ULL << 63;
1372             } else {
1373                 switch(blkz[i].cond) {
1374                 case BLK_ZONE_COND_FULL:
1375                 case BLK_ZONE_COND_READONLY:
1376                     /* Zone not writable */
1377                     wps->wp[j] = (blkz[i].start + blkz[i].len) << BDRV_SECTOR_BITS;
1378                     break;
1379                 case BLK_ZONE_COND_OFFLINE:
1380                     /* Zone not writable nor readable */
1381                     wps->wp[j] = (blkz[i].start) << BDRV_SECTOR_BITS;
1382                     break;
1383                 default:
1384                     if (reset_all) {
1385                         wps->wp[j] = blkz[i].start << BDRV_SECTOR_BITS;
1386                     } else {
1387                         wps->wp[j] = blkz[i].wp << BDRV_SECTOR_BITS;
1388                     }
1389                     break;
1390                 }
1391             }
1392         }
1393         sector = blkz[i - 1].start + blkz[i - 1].len;
1394     }
1395 
1396     return 0;
1397 }
1398 
1399 static void update_zones_wp(BlockDriverState *bs, int fd, int64_t offset,
1400                             unsigned int nrz)
1401 {
1402     if (get_zones_wp(bs, fd, offset, nrz, 0) < 0) {
1403         error_report("update zone wp failed");
1404     }
1405 }
1406 
1407 static void raw_refresh_zoned_limits(BlockDriverState *bs, struct stat *st,
1408                                      Error **errp)
1409 {
1410     BDRVRawState *s = bs->opaque;
1411     BlockZoneModel zoned;
1412     int ret;
1413 
1414     ret = get_sysfs_zoned_model(st, &zoned);
1415     if (ret < 0 || zoned == BLK_Z_NONE) {
1416         goto no_zoned;
1417     }
1418     bs->bl.zoned = zoned;
1419 
1420     ret = get_sysfs_long_val(st, "max_open_zones");
1421     if (ret >= 0) {
1422         bs->bl.max_open_zones = ret;
1423     }
1424 
1425     ret = get_sysfs_long_val(st, "max_active_zones");
1426     if (ret >= 0) {
1427         bs->bl.max_active_zones = ret;
1428     }
1429 
1430     /*
1431      * The zoned device must at least have zone size and nr_zones fields.
1432      */
1433     ret = get_sysfs_long_val(st, "chunk_sectors");
1434     if (ret < 0) {
1435         error_setg_errno(errp, -ret, "Unable to read chunk_sectors "
1436                                      "sysfs attribute");
1437         goto no_zoned;
1438     } else if (!ret) {
1439         error_setg(errp, "Read 0 from chunk_sectors sysfs attribute");
1440         goto no_zoned;
1441     }
1442     bs->bl.zone_size = ret << BDRV_SECTOR_BITS;
1443 
1444     ret = get_sysfs_long_val(st, "nr_zones");
1445     if (ret < 0) {
1446         error_setg_errno(errp, -ret, "Unable to read nr_zones "
1447                                      "sysfs attribute");
1448         goto no_zoned;
1449     } else if (!ret) {
1450         error_setg(errp, "Read 0 from nr_zones sysfs attribute");
1451         goto no_zoned;
1452     }
1453     bs->bl.nr_zones = ret;
1454 
1455     ret = get_sysfs_long_val(st, "zone_append_max_bytes");
1456     if (ret > 0) {
1457         bs->bl.max_append_sectors = ret >> BDRV_SECTOR_BITS;
1458     }
1459 
1460     ret = get_sysfs_long_val(st, "physical_block_size");
1461     if (ret >= 0) {
1462         bs->bl.write_granularity = ret;
1463     }
1464 
1465     /* The refresh_limits() function can be called multiple times. */
1466     g_free(bs->wps);
1467     bs->wps = g_malloc(sizeof(BlockZoneWps) +
1468             sizeof(int64_t) * bs->bl.nr_zones);
1469     ret = get_zones_wp(bs, s->fd, 0, bs->bl.nr_zones, 0);
1470     if (ret < 0) {
1471         error_setg_errno(errp, -ret, "report wps failed");
1472         goto no_zoned;
1473     }
1474     qemu_co_mutex_init(&bs->wps->colock);
1475     return;
1476 
1477 no_zoned:
1478     bs->bl.zoned = BLK_Z_NONE;
1479     g_free(bs->wps);
1480     bs->wps = NULL;
1481 }
1482 #else /* !defined(CONFIG_BLKZONED) */
1483 static void raw_refresh_zoned_limits(BlockDriverState *bs, struct stat *st,
1484                                      Error **errp)
1485 {
1486     bs->bl.zoned = BLK_Z_NONE;
1487 }
1488 #endif /* !defined(CONFIG_BLKZONED) */
1489 
1490 static void raw_refresh_limits(BlockDriverState *bs, Error **errp)
1491 {
1492     BDRVRawState *s = bs->opaque;
1493     struct stat st;
1494 
1495     s->needs_alignment = raw_needs_alignment(bs);
1496     raw_probe_alignment(bs, s->fd, errp);
1497 
1498     bs->bl.min_mem_alignment = s->buf_align;
1499     bs->bl.opt_mem_alignment = MAX(s->buf_align, qemu_real_host_page_size());
1500 
1501     /*
1502      * Maximum transfers are best effort, so it is okay to ignore any
1503      * errors.  That said, based on the man page errors in fstat would be
1504      * very much unexpected; the only possible case seems to be ENOMEM.
1505      */
1506     if (fstat(s->fd, &st)) {
1507         return;
1508     }
1509 
1510 #if defined(__APPLE__) && (__MACH__)
1511     struct statfs buf;
1512 
1513     if (!fstatfs(s->fd, &buf)) {
1514         bs->bl.opt_transfer = buf.f_iosize;
1515         bs->bl.pdiscard_alignment = buf.f_bsize;
1516     }
1517 #endif
1518 
1519     if (bdrv_is_sg(bs) || S_ISBLK(st.st_mode)) {
1520         int ret = hdev_get_max_hw_transfer(s->fd, &st);
1521 
1522         if (ret > 0 && ret <= BDRV_REQUEST_MAX_BYTES) {
1523             bs->bl.max_hw_transfer = ret;
1524         }
1525 
1526         ret = hdev_get_max_segments(s->fd, &st);
1527         if (ret > 0) {
1528             bs->bl.max_hw_iov = ret;
1529         }
1530     }
1531 
1532     raw_refresh_zoned_limits(bs, &st, errp);
1533 }
1534 
1535 static int check_for_dasd(int fd)
1536 {
1537 #ifdef BIODASDINFO2
1538     struct dasd_information2_t info = {0};
1539 
1540     return ioctl(fd, BIODASDINFO2, &info);
1541 #else
1542     return -1;
1543 #endif
1544 }
1545 
1546 /**
1547  * Try to get @bs's logical and physical block size.
1548  * On success, store them in @bsz and return zero.
1549  * On failure, return negative errno.
1550  */
1551 static int hdev_probe_blocksizes(BlockDriverState *bs, BlockSizes *bsz)
1552 {
1553     BDRVRawState *s = bs->opaque;
1554     int ret;
1555 
1556     /* If DASD or zoned devices, get blocksizes */
1557     if (check_for_dasd(s->fd) < 0) {
1558         /* zoned devices are not DASD */
1559         if (bs->bl.zoned == BLK_Z_NONE) {
1560             return -ENOTSUP;
1561         }
1562     }
1563     ret = probe_logical_blocksize(s->fd, &bsz->log);
1564     if (ret < 0) {
1565         return ret;
1566     }
1567     return probe_physical_blocksize(s->fd, &bsz->phys);
1568 }
1569 
1570 /**
1571  * Try to get @bs's geometry: cyls, heads, sectors.
1572  * On success, store them in @geo and return 0.
1573  * On failure return -errno.
1574  * (Allows block driver to assign default geometry values that guest sees)
1575  */
1576 #ifdef __linux__
1577 static int hdev_probe_geometry(BlockDriverState *bs, HDGeometry *geo)
1578 {
1579     BDRVRawState *s = bs->opaque;
1580     struct hd_geometry ioctl_geo = {0};
1581 
1582     /* If DASD, get its geometry */
1583     if (check_for_dasd(s->fd) < 0) {
1584         return -ENOTSUP;
1585     }
1586     if (ioctl(s->fd, HDIO_GETGEO, &ioctl_geo) < 0) {
1587         return -errno;
1588     }
1589     /* HDIO_GETGEO may return success even though geo contains zeros
1590        (e.g. certain multipath setups) */
1591     if (!ioctl_geo.heads || !ioctl_geo.sectors || !ioctl_geo.cylinders) {
1592         return -ENOTSUP;
1593     }
1594     /* Do not return a geometry for partition */
1595     if (ioctl_geo.start != 0) {
1596         return -ENOTSUP;
1597     }
1598     geo->heads = ioctl_geo.heads;
1599     geo->sectors = ioctl_geo.sectors;
1600     geo->cylinders = ioctl_geo.cylinders;
1601 
1602     return 0;
1603 }
1604 #else /* __linux__ */
1605 static int hdev_probe_geometry(BlockDriverState *bs, HDGeometry *geo)
1606 {
1607     return -ENOTSUP;
1608 }
1609 #endif
1610 
1611 #if defined(__linux__)
1612 static int handle_aiocb_ioctl(void *opaque)
1613 {
1614     RawPosixAIOData *aiocb = opaque;
1615     int ret;
1616 
1617     ret = RETRY_ON_EINTR(
1618         ioctl(aiocb->aio_fildes, aiocb->ioctl.cmd, aiocb->ioctl.buf)
1619     );
1620     if (ret == -1) {
1621         return -errno;
1622     }
1623 
1624     return 0;
1625 }
1626 #endif /* linux */
1627 
1628 static int handle_aiocb_flush(void *opaque)
1629 {
1630     RawPosixAIOData *aiocb = opaque;
1631     BDRVRawState *s = aiocb->bs->opaque;
1632     int ret;
1633 
1634     if (s->page_cache_inconsistent) {
1635         return -s->page_cache_inconsistent;
1636     }
1637 
1638     ret = qemu_fdatasync(aiocb->aio_fildes);
1639     if (ret == -1) {
1640         trace_file_flush_fdatasync_failed(errno);
1641 
1642         /* There is no clear definition of the semantics of a failing fsync(),
1643          * so we may have to assume the worst. The sad truth is that this
1644          * assumption is correct for Linux. Some pages are now probably marked
1645          * clean in the page cache even though they are inconsistent with the
1646          * on-disk contents. The next fdatasync() call would succeed, but no
1647          * further writeback attempt will be made. We can't get back to a state
1648          * in which we know what is on disk (we would have to rewrite
1649          * everything that was touched since the last fdatasync() at least), so
1650          * make bdrv_flush() fail permanently. Given that the behaviour isn't
1651          * really defined, I have little hope that other OSes are doing better.
1652          *
1653          * Obviously, this doesn't affect O_DIRECT, which bypasses the page
1654          * cache. */
1655         if ((s->open_flags & O_DIRECT) == 0) {
1656             s->page_cache_inconsistent = errno;
1657         }
1658         return -errno;
1659     }
1660     return 0;
1661 }
1662 
1663 #ifdef CONFIG_PREADV
1664 
1665 static bool preadv_present = true;
1666 
1667 static ssize_t
1668 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1669 {
1670     return preadv(fd, iov, nr_iov, offset);
1671 }
1672 
1673 static ssize_t
1674 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1675 {
1676     return pwritev(fd, iov, nr_iov, offset);
1677 }
1678 
1679 #else
1680 
1681 static bool preadv_present = false;
1682 
1683 static ssize_t
1684 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1685 {
1686     return -ENOSYS;
1687 }
1688 
1689 static ssize_t
1690 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1691 {
1692     return -ENOSYS;
1693 }
1694 
1695 #endif
1696 
1697 static ssize_t handle_aiocb_rw_vector(RawPosixAIOData *aiocb)
1698 {
1699     ssize_t len;
1700 
1701     len = RETRY_ON_EINTR(
1702         (aiocb->aio_type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) ?
1703             qemu_pwritev(aiocb->aio_fildes,
1704                            aiocb->io.iov,
1705                            aiocb->io.niov,
1706                            aiocb->aio_offset) :
1707             qemu_preadv(aiocb->aio_fildes,
1708                           aiocb->io.iov,
1709                           aiocb->io.niov,
1710                           aiocb->aio_offset)
1711     );
1712 
1713     if (len == -1) {
1714         return -errno;
1715     }
1716     return len;
1717 }
1718 
1719 /*
1720  * Read/writes the data to/from a given linear buffer.
1721  *
1722  * Returns the number of bytes handles or -errno in case of an error. Short
1723  * reads are only returned if the end of the file is reached.
1724  */
1725 static ssize_t handle_aiocb_rw_linear(RawPosixAIOData *aiocb, char *buf)
1726 {
1727     ssize_t offset = 0;
1728     ssize_t len;
1729 
1730     while (offset < aiocb->aio_nbytes) {
1731         if (aiocb->aio_type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) {
1732             len = pwrite(aiocb->aio_fildes,
1733                          (const char *)buf + offset,
1734                          aiocb->aio_nbytes - offset,
1735                          aiocb->aio_offset + offset);
1736         } else {
1737             len = pread(aiocb->aio_fildes,
1738                         buf + offset,
1739                         aiocb->aio_nbytes - offset,
1740                         aiocb->aio_offset + offset);
1741         }
1742         if (len == -1 && errno == EINTR) {
1743             continue;
1744         } else if (len == -1 && errno == EINVAL &&
1745                    (aiocb->bs->open_flags & BDRV_O_NOCACHE) &&
1746                    !(aiocb->aio_type & QEMU_AIO_WRITE) &&
1747                    offset > 0) {
1748             /* O_DIRECT pread() may fail with EINVAL when offset is unaligned
1749              * after a short read.  Assume that O_DIRECT short reads only occur
1750              * at EOF.  Therefore this is a short read, not an I/O error.
1751              */
1752             break;
1753         } else if (len == -1) {
1754             offset = -errno;
1755             break;
1756         } else if (len == 0) {
1757             break;
1758         }
1759         offset += len;
1760     }
1761 
1762     return offset;
1763 }
1764 
1765 static int handle_aiocb_rw(void *opaque)
1766 {
1767     RawPosixAIOData *aiocb = opaque;
1768     ssize_t nbytes;
1769     char *buf;
1770 
1771     if (!(aiocb->aio_type & QEMU_AIO_MISALIGNED)) {
1772         /*
1773          * If there is just a single buffer, and it is properly aligned
1774          * we can just use plain pread/pwrite without any problems.
1775          */
1776         if (aiocb->io.niov == 1) {
1777             nbytes = handle_aiocb_rw_linear(aiocb, aiocb->io.iov->iov_base);
1778             goto out;
1779         }
1780         /*
1781          * We have more than one iovec, and all are properly aligned.
1782          *
1783          * Try preadv/pwritev first and fall back to linearizing the
1784          * buffer if it's not supported.
1785          */
1786         if (preadv_present) {
1787             nbytes = handle_aiocb_rw_vector(aiocb);
1788             if (nbytes == aiocb->aio_nbytes ||
1789                 (nbytes < 0 && nbytes != -ENOSYS)) {
1790                 goto out;
1791             }
1792             preadv_present = false;
1793         }
1794 
1795         /*
1796          * XXX(hch): short read/write.  no easy way to handle the reminder
1797          * using these interfaces.  For now retry using plain
1798          * pread/pwrite?
1799          */
1800     }
1801 
1802     /*
1803      * Ok, we have to do it the hard way, copy all segments into
1804      * a single aligned buffer.
1805      */
1806     buf = qemu_try_blockalign(aiocb->bs, aiocb->aio_nbytes);
1807     if (buf == NULL) {
1808         nbytes = -ENOMEM;
1809         goto out;
1810     }
1811 
1812     if (aiocb->aio_type & QEMU_AIO_WRITE) {
1813         char *p = buf;
1814         int i;
1815 
1816         for (i = 0; i < aiocb->io.niov; ++i) {
1817             memcpy(p, aiocb->io.iov[i].iov_base, aiocb->io.iov[i].iov_len);
1818             p += aiocb->io.iov[i].iov_len;
1819         }
1820         assert(p - buf == aiocb->aio_nbytes);
1821     }
1822 
1823     nbytes = handle_aiocb_rw_linear(aiocb, buf);
1824     if (!(aiocb->aio_type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND))) {
1825         char *p = buf;
1826         size_t count = aiocb->aio_nbytes, copy;
1827         int i;
1828 
1829         for (i = 0; i < aiocb->io.niov && count; ++i) {
1830             copy = count;
1831             if (copy > aiocb->io.iov[i].iov_len) {
1832                 copy = aiocb->io.iov[i].iov_len;
1833             }
1834             memcpy(aiocb->io.iov[i].iov_base, p, copy);
1835             assert(count >= copy);
1836             p     += copy;
1837             count -= copy;
1838         }
1839         assert(count == 0);
1840     }
1841     qemu_vfree(buf);
1842 
1843 out:
1844     if (nbytes == aiocb->aio_nbytes) {
1845         return 0;
1846     } else if (nbytes >= 0 && nbytes < aiocb->aio_nbytes) {
1847         if (aiocb->aio_type & QEMU_AIO_WRITE) {
1848             return -EINVAL;
1849         } else {
1850             iov_memset(aiocb->io.iov, aiocb->io.niov, nbytes,
1851                       0, aiocb->aio_nbytes - nbytes);
1852             return 0;
1853         }
1854     } else {
1855         assert(nbytes < 0);
1856         return nbytes;
1857     }
1858 }
1859 
1860 #if defined(CONFIG_FALLOCATE) || defined(BLKZEROOUT) || defined(BLKDISCARD)
1861 static int translate_err(int err)
1862 {
1863     if (err == -ENODEV || err == -ENOSYS || err == -EOPNOTSUPP ||
1864         err == -ENOTTY) {
1865         err = -ENOTSUP;
1866     }
1867     return err;
1868 }
1869 #endif
1870 
1871 #ifdef CONFIG_FALLOCATE
1872 static int do_fallocate(int fd, int mode, off_t offset, off_t len)
1873 {
1874     do {
1875         if (fallocate(fd, mode, offset, len) == 0) {
1876             return 0;
1877         }
1878     } while (errno == EINTR);
1879     return translate_err(-errno);
1880 }
1881 #endif
1882 
1883 static ssize_t handle_aiocb_write_zeroes_block(RawPosixAIOData *aiocb)
1884 {
1885     int ret = -ENOTSUP;
1886     BDRVRawState *s = aiocb->bs->opaque;
1887 
1888     if (!s->has_write_zeroes) {
1889         return -ENOTSUP;
1890     }
1891 
1892 #ifdef BLKZEROOUT
1893     /* The BLKZEROOUT implementation in the kernel doesn't set
1894      * BLKDEV_ZERO_NOFALLBACK, so we can't call this if we have to avoid slow
1895      * fallbacks. */
1896     if (!(aiocb->aio_type & QEMU_AIO_NO_FALLBACK)) {
1897         do {
1898             uint64_t range[2] = { aiocb->aio_offset, aiocb->aio_nbytes };
1899             if (ioctl(aiocb->aio_fildes, BLKZEROOUT, range) == 0) {
1900                 return 0;
1901             }
1902         } while (errno == EINTR);
1903 
1904         ret = translate_err(-errno);
1905         if (ret == -ENOTSUP) {
1906             s->has_write_zeroes = false;
1907         }
1908     }
1909 #endif
1910 
1911     return ret;
1912 }
1913 
1914 static int handle_aiocb_write_zeroes(void *opaque)
1915 {
1916     RawPosixAIOData *aiocb = opaque;
1917 #ifdef CONFIG_FALLOCATE
1918     BDRVRawState *s = aiocb->bs->opaque;
1919     int64_t len;
1920 #endif
1921 
1922     if (aiocb->aio_type & QEMU_AIO_BLKDEV) {
1923         return handle_aiocb_write_zeroes_block(aiocb);
1924     }
1925 
1926 #ifdef CONFIG_FALLOCATE_ZERO_RANGE
1927     if (s->has_write_zeroes) {
1928         int ret = do_fallocate(s->fd, FALLOC_FL_ZERO_RANGE,
1929                                aiocb->aio_offset, aiocb->aio_nbytes);
1930         if (ret == -ENOTSUP) {
1931             s->has_write_zeroes = false;
1932         } else if (ret == 0 || ret != -EINVAL) {
1933             return ret;
1934         }
1935         /*
1936          * Note: Some file systems do not like unaligned byte ranges, and
1937          * return EINVAL in such a case, though they should not do it according
1938          * to the man-page of fallocate(). Thus we simply ignore this return
1939          * value and try the other fallbacks instead.
1940          */
1941     }
1942 #endif
1943 
1944 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
1945     if (s->has_discard && s->has_fallocate) {
1946         int ret = do_fallocate(s->fd,
1947                                FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1948                                aiocb->aio_offset, aiocb->aio_nbytes);
1949         if (ret == 0) {
1950             ret = do_fallocate(s->fd, 0, aiocb->aio_offset, aiocb->aio_nbytes);
1951             if (ret == 0 || ret != -ENOTSUP) {
1952                 return ret;
1953             }
1954             s->has_fallocate = false;
1955         } else if (ret == -EINVAL) {
1956             /*
1957              * Some file systems like older versions of GPFS do not like un-
1958              * aligned byte ranges, and return EINVAL in such a case, though
1959              * they should not do it according to the man-page of fallocate().
1960              * Warn about the bad filesystem and try the final fallback instead.
1961              */
1962             warn_report_once("Your file system is misbehaving: "
1963                              "fallocate(FALLOC_FL_PUNCH_HOLE) returned EINVAL. "
1964                              "Please report this bug to your file system "
1965                              "vendor.");
1966         } else if (ret != -ENOTSUP) {
1967             return ret;
1968         } else {
1969             s->has_discard = false;
1970         }
1971     }
1972 #endif
1973 
1974 #ifdef CONFIG_FALLOCATE
1975     /* Last resort: we are trying to extend the file with zeroed data. This
1976      * can be done via fallocate(fd, 0) */
1977     len = raw_getlength(aiocb->bs);
1978     if (s->has_fallocate && len >= 0 && aiocb->aio_offset >= len) {
1979         int ret = do_fallocate(s->fd, 0, aiocb->aio_offset, aiocb->aio_nbytes);
1980         if (ret == 0 || ret != -ENOTSUP) {
1981             return ret;
1982         }
1983         s->has_fallocate = false;
1984     }
1985 #endif
1986 
1987     return -ENOTSUP;
1988 }
1989 
1990 static int handle_aiocb_write_zeroes_unmap(void *opaque)
1991 {
1992     RawPosixAIOData *aiocb = opaque;
1993     BDRVRawState *s G_GNUC_UNUSED = aiocb->bs->opaque;
1994 
1995     /* First try to write zeros and unmap at the same time */
1996 
1997 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
1998     int ret = do_fallocate(s->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1999                            aiocb->aio_offset, aiocb->aio_nbytes);
2000     switch (ret) {
2001     case -ENOTSUP:
2002     case -EINVAL:
2003     case -EBUSY:
2004         break;
2005     default:
2006         return ret;
2007     }
2008 #endif
2009 
2010     /* If we couldn't manage to unmap while guaranteed that the area reads as
2011      * all-zero afterwards, just write zeroes without unmapping */
2012     return handle_aiocb_write_zeroes(aiocb);
2013 }
2014 
2015 #ifndef HAVE_COPY_FILE_RANGE
2016 static off_t copy_file_range(int in_fd, off_t *in_off, int out_fd,
2017                              off_t *out_off, size_t len, unsigned int flags)
2018 {
2019 #ifdef __NR_copy_file_range
2020     return syscall(__NR_copy_file_range, in_fd, in_off, out_fd,
2021                    out_off, len, flags);
2022 #else
2023     errno = ENOSYS;
2024     return -1;
2025 #endif
2026 }
2027 #endif
2028 
2029 /*
2030  * parse_zone - Fill a zone descriptor
2031  */
2032 #if defined(CONFIG_BLKZONED)
2033 static inline int parse_zone(struct BlockZoneDescriptor *zone,
2034                               const struct blk_zone *blkz) {
2035     zone->start = blkz->start << BDRV_SECTOR_BITS;
2036     zone->length = blkz->len << BDRV_SECTOR_BITS;
2037     zone->wp = blkz->wp << BDRV_SECTOR_BITS;
2038 
2039 #ifdef HAVE_BLK_ZONE_REP_CAPACITY
2040     zone->cap = blkz->capacity << BDRV_SECTOR_BITS;
2041 #else
2042     zone->cap = blkz->len << BDRV_SECTOR_BITS;
2043 #endif
2044 
2045     switch (blkz->type) {
2046     case BLK_ZONE_TYPE_SEQWRITE_REQ:
2047         zone->type = BLK_ZT_SWR;
2048         break;
2049     case BLK_ZONE_TYPE_SEQWRITE_PREF:
2050         zone->type = BLK_ZT_SWP;
2051         break;
2052     case BLK_ZONE_TYPE_CONVENTIONAL:
2053         zone->type = BLK_ZT_CONV;
2054         break;
2055     default:
2056         error_report("Unsupported zone type: 0x%x", blkz->type);
2057         return -ENOTSUP;
2058     }
2059 
2060     switch (blkz->cond) {
2061     case BLK_ZONE_COND_NOT_WP:
2062         zone->state = BLK_ZS_NOT_WP;
2063         break;
2064     case BLK_ZONE_COND_EMPTY:
2065         zone->state = BLK_ZS_EMPTY;
2066         break;
2067     case BLK_ZONE_COND_IMP_OPEN:
2068         zone->state = BLK_ZS_IOPEN;
2069         break;
2070     case BLK_ZONE_COND_EXP_OPEN:
2071         zone->state = BLK_ZS_EOPEN;
2072         break;
2073     case BLK_ZONE_COND_CLOSED:
2074         zone->state = BLK_ZS_CLOSED;
2075         break;
2076     case BLK_ZONE_COND_READONLY:
2077         zone->state = BLK_ZS_RDONLY;
2078         break;
2079     case BLK_ZONE_COND_FULL:
2080         zone->state = BLK_ZS_FULL;
2081         break;
2082     case BLK_ZONE_COND_OFFLINE:
2083         zone->state = BLK_ZS_OFFLINE;
2084         break;
2085     default:
2086         error_report("Unsupported zone state: 0x%x", blkz->cond);
2087         return -ENOTSUP;
2088     }
2089     return 0;
2090 }
2091 #endif
2092 
2093 #if defined(CONFIG_BLKZONED)
2094 static int handle_aiocb_zone_report(void *opaque)
2095 {
2096     RawPosixAIOData *aiocb = opaque;
2097     int fd = aiocb->aio_fildes;
2098     unsigned int *nr_zones = aiocb->zone_report.nr_zones;
2099     BlockZoneDescriptor *zones = aiocb->zone_report.zones;
2100     /* zoned block devices use 512-byte sectors */
2101     uint64_t sector = aiocb->aio_offset / 512;
2102 
2103     struct blk_zone *blkz;
2104     size_t rep_size;
2105     unsigned int nrz;
2106     int ret;
2107     unsigned int n = 0, i = 0;
2108 
2109     nrz = *nr_zones;
2110     rep_size = sizeof(struct blk_zone_report) + nrz * sizeof(struct blk_zone);
2111     g_autofree struct blk_zone_report *rep = NULL;
2112     rep = g_malloc(rep_size);
2113 
2114     blkz = (struct blk_zone *)(rep + 1);
2115     while (n < nrz) {
2116         memset(rep, 0, rep_size);
2117         rep->sector = sector;
2118         rep->nr_zones = nrz - n;
2119 
2120         do {
2121             ret = ioctl(fd, BLKREPORTZONE, rep);
2122         } while (ret != 0 && errno == EINTR);
2123         if (ret != 0) {
2124             error_report("%d: ioctl BLKREPORTZONE at %" PRId64 " failed %d",
2125                          fd, sector, errno);
2126             return -errno;
2127         }
2128 
2129         if (!rep->nr_zones) {
2130             break;
2131         }
2132 
2133         for (i = 0; i < rep->nr_zones; i++, n++) {
2134             ret = parse_zone(&zones[n], &blkz[i]);
2135             if (ret != 0) {
2136                 return ret;
2137             }
2138 
2139             /* The next report should start after the last zone reported */
2140             sector = blkz[i].start + blkz[i].len;
2141         }
2142     }
2143 
2144     *nr_zones = n;
2145     return 0;
2146 }
2147 #endif
2148 
2149 #if defined(CONFIG_BLKZONED)
2150 static int handle_aiocb_zone_mgmt(void *opaque)
2151 {
2152     RawPosixAIOData *aiocb = opaque;
2153     int fd = aiocb->aio_fildes;
2154     uint64_t sector = aiocb->aio_offset / 512;
2155     int64_t nr_sectors = aiocb->aio_nbytes / 512;
2156     struct blk_zone_range range;
2157     int ret;
2158 
2159     /* Execute the operation */
2160     range.sector = sector;
2161     range.nr_sectors = nr_sectors;
2162     do {
2163         ret = ioctl(fd, aiocb->zone_mgmt.op, &range);
2164     } while (ret != 0 && errno == EINTR);
2165 
2166     return ret < 0 ? -errno : ret;
2167 }
2168 #endif
2169 
2170 static int handle_aiocb_copy_range(void *opaque)
2171 {
2172     RawPosixAIOData *aiocb = opaque;
2173     uint64_t bytes = aiocb->aio_nbytes;
2174     off_t in_off = aiocb->aio_offset;
2175     off_t out_off = aiocb->copy_range.aio_offset2;
2176 
2177     while (bytes) {
2178         ssize_t ret = copy_file_range(aiocb->aio_fildes, &in_off,
2179                                       aiocb->copy_range.aio_fd2, &out_off,
2180                                       bytes, 0);
2181         trace_file_copy_file_range(aiocb->bs, aiocb->aio_fildes, in_off,
2182                                    aiocb->copy_range.aio_fd2, out_off, bytes,
2183                                    0, ret);
2184         if (ret == 0) {
2185             /* No progress (e.g. when beyond EOF), let the caller fall back to
2186              * buffer I/O. */
2187             return -ENOSPC;
2188         }
2189         if (ret < 0) {
2190             switch (errno) {
2191             case ENOSYS:
2192                 return -ENOTSUP;
2193             case EINTR:
2194                 continue;
2195             default:
2196                 return -errno;
2197             }
2198         }
2199         bytes -= ret;
2200     }
2201     return 0;
2202 }
2203 
2204 static int handle_aiocb_discard(void *opaque)
2205 {
2206     RawPosixAIOData *aiocb = opaque;
2207     int ret = -ENOTSUP;
2208     BDRVRawState *s = aiocb->bs->opaque;
2209 
2210     if (!s->has_discard) {
2211         return -ENOTSUP;
2212     }
2213 
2214     if (aiocb->aio_type & QEMU_AIO_BLKDEV) {
2215 #ifdef BLKDISCARD
2216         do {
2217             uint64_t range[2] = { aiocb->aio_offset, aiocb->aio_nbytes };
2218             if (ioctl(aiocb->aio_fildes, BLKDISCARD, range) == 0) {
2219                 return 0;
2220             }
2221         } while (errno == EINTR);
2222 
2223         ret = translate_err(-errno);
2224 #endif
2225     } else {
2226 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
2227         ret = do_fallocate(s->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
2228                            aiocb->aio_offset, aiocb->aio_nbytes);
2229         ret = translate_err(ret);
2230 #elif defined(__APPLE__) && (__MACH__)
2231         fpunchhole_t fpunchhole;
2232         fpunchhole.fp_flags = 0;
2233         fpunchhole.reserved = 0;
2234         fpunchhole.fp_offset = aiocb->aio_offset;
2235         fpunchhole.fp_length = aiocb->aio_nbytes;
2236         if (fcntl(s->fd, F_PUNCHHOLE, &fpunchhole) == -1) {
2237             ret = errno == ENODEV ? -ENOTSUP : -errno;
2238         } else {
2239             ret = 0;
2240         }
2241 #endif
2242     }
2243 
2244     if (ret == -ENOTSUP) {
2245         s->has_discard = false;
2246     }
2247     return ret;
2248 }
2249 
2250 /*
2251  * Help alignment probing by allocating the first block.
2252  *
2253  * When reading with direct I/O from unallocated area on Gluster backed by XFS,
2254  * reading succeeds regardless of request length. In this case we fallback to
2255  * safe alignment which is not optimal. Allocating the first block avoids this
2256  * fallback.
2257  *
2258  * fd may be opened with O_DIRECT, but we don't know the buffer alignment or
2259  * request alignment, so we use safe values.
2260  *
2261  * Returns: 0 on success, -errno on failure. Since this is an optimization,
2262  * caller may ignore failures.
2263  */
2264 static int allocate_first_block(int fd, size_t max_size)
2265 {
2266     size_t write_size = (max_size < MAX_BLOCKSIZE)
2267         ? BDRV_SECTOR_SIZE
2268         : MAX_BLOCKSIZE;
2269     size_t max_align = MAX(MAX_BLOCKSIZE, qemu_real_host_page_size());
2270     void *buf;
2271     ssize_t n;
2272     int ret;
2273 
2274     buf = qemu_memalign(max_align, write_size);
2275     memset(buf, 0, write_size);
2276 
2277     n = RETRY_ON_EINTR(pwrite(fd, buf, write_size, 0));
2278 
2279     ret = (n == -1) ? -errno : 0;
2280 
2281     qemu_vfree(buf);
2282     return ret;
2283 }
2284 
2285 static int handle_aiocb_truncate(void *opaque)
2286 {
2287     RawPosixAIOData *aiocb = opaque;
2288     int result = 0;
2289     int64_t current_length = 0;
2290     char *buf = NULL;
2291     struct stat st;
2292     int fd = aiocb->aio_fildes;
2293     int64_t offset = aiocb->aio_offset;
2294     PreallocMode prealloc = aiocb->truncate.prealloc;
2295     Error **errp = aiocb->truncate.errp;
2296 
2297     if (fstat(fd, &st) < 0) {
2298         result = -errno;
2299         error_setg_errno(errp, -result, "Could not stat file");
2300         return result;
2301     }
2302 
2303     current_length = st.st_size;
2304     if (current_length > offset && prealloc != PREALLOC_MODE_OFF) {
2305         error_setg(errp, "Cannot use preallocation for shrinking files");
2306         return -ENOTSUP;
2307     }
2308 
2309     switch (prealloc) {
2310 #ifdef CONFIG_POSIX_FALLOCATE
2311     case PREALLOC_MODE_FALLOC:
2312         /*
2313          * Truncating before posix_fallocate() makes it about twice slower on
2314          * file systems that do not support fallocate(), trying to check if a
2315          * block is allocated before allocating it, so don't do that here.
2316          */
2317         if (offset != current_length) {
2318             result = -posix_fallocate(fd, current_length,
2319                                       offset - current_length);
2320             if (result != 0) {
2321                 /* posix_fallocate() doesn't set errno. */
2322                 error_setg_errno(errp, -result,
2323                                  "Could not preallocate new data");
2324             } else if (current_length == 0) {
2325                 /*
2326                  * posix_fallocate() uses fallocate() if the filesystem
2327                  * supports it, or fallback to manually writing zeroes. If
2328                  * fallocate() was used, unaligned reads from the fallocated
2329                  * area in raw_probe_alignment() will succeed, hence we need to
2330                  * allocate the first block.
2331                  *
2332                  * Optimize future alignment probing; ignore failures.
2333                  */
2334                 allocate_first_block(fd, offset);
2335             }
2336         } else {
2337             result = 0;
2338         }
2339         goto out;
2340 #endif
2341     case PREALLOC_MODE_FULL:
2342     {
2343         int64_t num = 0, left = offset - current_length;
2344         off_t seek_result;
2345 
2346         /*
2347          * Knowing the final size from the beginning could allow the file
2348          * system driver to do less allocations and possibly avoid
2349          * fragmentation of the file.
2350          */
2351         if (ftruncate(fd, offset) != 0) {
2352             result = -errno;
2353             error_setg_errno(errp, -result, "Could not resize file");
2354             goto out;
2355         }
2356 
2357         buf = g_malloc0(65536);
2358 
2359         seek_result = lseek(fd, current_length, SEEK_SET);
2360         if (seek_result < 0) {
2361             result = -errno;
2362             error_setg_errno(errp, -result,
2363                              "Failed to seek to the old end of file");
2364             goto out;
2365         }
2366 
2367         while (left > 0) {
2368             num = MIN(left, 65536);
2369             result = write(fd, buf, num);
2370             if (result < 0) {
2371                 if (errno == EINTR) {
2372                     continue;
2373                 }
2374                 result = -errno;
2375                 error_setg_errno(errp, -result,
2376                                  "Could not write zeros for preallocation");
2377                 goto out;
2378             }
2379             left -= result;
2380         }
2381         if (result >= 0) {
2382             result = fsync(fd);
2383             if (result < 0) {
2384                 result = -errno;
2385                 error_setg_errno(errp, -result,
2386                                  "Could not flush file to disk");
2387                 goto out;
2388             }
2389         }
2390         goto out;
2391     }
2392     case PREALLOC_MODE_OFF:
2393         if (ftruncate(fd, offset) != 0) {
2394             result = -errno;
2395             error_setg_errno(errp, -result, "Could not resize file");
2396         } else if (current_length == 0 && offset > current_length) {
2397             /* Optimize future alignment probing; ignore failures. */
2398             allocate_first_block(fd, offset);
2399         }
2400         return result;
2401     default:
2402         result = -ENOTSUP;
2403         error_setg(errp, "Unsupported preallocation mode: %s",
2404                    PreallocMode_str(prealloc));
2405         return result;
2406     }
2407 
2408 out:
2409     if (result < 0) {
2410         if (ftruncate(fd, current_length) < 0) {
2411             error_report("Failed to restore old file length: %s",
2412                          strerror(errno));
2413         }
2414     }
2415 
2416     g_free(buf);
2417     return result;
2418 }
2419 
2420 static int coroutine_fn raw_thread_pool_submit(ThreadPoolFunc func, void *arg)
2421 {
2422     return thread_pool_submit_co(func, arg);
2423 }
2424 
2425 /*
2426  * Check if all memory in this vector is sector aligned.
2427  */
2428 static bool bdrv_qiov_is_aligned(BlockDriverState *bs, QEMUIOVector *qiov)
2429 {
2430     int i;
2431     size_t alignment = bdrv_min_mem_align(bs);
2432     size_t len = bs->bl.request_alignment;
2433     IO_CODE();
2434 
2435     for (i = 0; i < qiov->niov; i++) {
2436         if ((uintptr_t) qiov->iov[i].iov_base % alignment) {
2437             return false;
2438         }
2439         if (qiov->iov[i].iov_len % len) {
2440             return false;
2441         }
2442     }
2443 
2444     return true;
2445 }
2446 
2447 static int coroutine_fn raw_co_prw(BlockDriverState *bs, int64_t *offset_ptr,
2448                                    uint64_t bytes, QEMUIOVector *qiov, int type)
2449 {
2450     BDRVRawState *s = bs->opaque;
2451     RawPosixAIOData acb;
2452     int ret;
2453     uint64_t offset = *offset_ptr;
2454 
2455     if (fd_open(bs) < 0)
2456         return -EIO;
2457 #if defined(CONFIG_BLKZONED)
2458     if ((type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) &&
2459         bs->bl.zoned != BLK_Z_NONE) {
2460         qemu_co_mutex_lock(&bs->wps->colock);
2461         if (type & QEMU_AIO_ZONE_APPEND) {
2462             int index = offset / bs->bl.zone_size;
2463             offset = bs->wps->wp[index];
2464         }
2465     }
2466 #endif
2467 
2468     /*
2469      * When using O_DIRECT, the request must be aligned to be able to use
2470      * either libaio or io_uring interface. If not fail back to regular thread
2471      * pool read/write code which emulates this for us if we
2472      * set QEMU_AIO_MISALIGNED.
2473      */
2474     if (s->needs_alignment && !bdrv_qiov_is_aligned(bs, qiov)) {
2475         type |= QEMU_AIO_MISALIGNED;
2476 #ifdef CONFIG_LINUX_IO_URING
2477     } else if (s->use_linux_io_uring) {
2478         assert(qiov->size == bytes);
2479         ret = luring_co_submit(bs, s->fd, offset, qiov, type);
2480         goto out;
2481 #endif
2482 #ifdef CONFIG_LINUX_AIO
2483     } else if (s->use_linux_aio) {
2484         assert(qiov->size == bytes);
2485         ret = laio_co_submit(s->fd, offset, qiov, type,
2486                               s->aio_max_batch);
2487         goto out;
2488 #endif
2489     }
2490 
2491     acb = (RawPosixAIOData) {
2492         .bs             = bs,
2493         .aio_fildes     = s->fd,
2494         .aio_type       = type,
2495         .aio_offset     = offset,
2496         .aio_nbytes     = bytes,
2497         .io             = {
2498             .iov            = qiov->iov,
2499             .niov           = qiov->niov,
2500         },
2501     };
2502 
2503     assert(qiov->size == bytes);
2504     ret = raw_thread_pool_submit(handle_aiocb_rw, &acb);
2505     goto out; /* Avoid the compiler err of unused label */
2506 
2507 out:
2508 #if defined(CONFIG_BLKZONED)
2509     if ((type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) &&
2510         bs->bl.zoned != BLK_Z_NONE) {
2511         BlockZoneWps *wps = bs->wps;
2512         if (ret == 0) {
2513             uint64_t *wp = &wps->wp[offset / bs->bl.zone_size];
2514             if (!BDRV_ZT_IS_CONV(*wp)) {
2515                 if (type & QEMU_AIO_ZONE_APPEND) {
2516                     *offset_ptr = *wp;
2517                     trace_zbd_zone_append_complete(bs, *offset_ptr
2518                         >> BDRV_SECTOR_BITS);
2519                 }
2520                 /* Advance the wp if needed */
2521                 if (offset + bytes > *wp) {
2522                     *wp = offset + bytes;
2523                 }
2524             }
2525         } else {
2526             /*
2527              * write and append write are not allowed to cross zone boundaries
2528              */
2529             update_zones_wp(bs, s->fd, offset, 1);
2530         }
2531 
2532         qemu_co_mutex_unlock(&wps->colock);
2533     }
2534 #endif
2535     return ret;
2536 }
2537 
2538 static int coroutine_fn raw_co_preadv(BlockDriverState *bs, int64_t offset,
2539                                       int64_t bytes, QEMUIOVector *qiov,
2540                                       BdrvRequestFlags flags)
2541 {
2542     return raw_co_prw(bs, &offset, bytes, qiov, QEMU_AIO_READ);
2543 }
2544 
2545 static int coroutine_fn raw_co_pwritev(BlockDriverState *bs, int64_t offset,
2546                                        int64_t bytes, QEMUIOVector *qiov,
2547                                        BdrvRequestFlags flags)
2548 {
2549     return raw_co_prw(bs, &offset, bytes, qiov, QEMU_AIO_WRITE);
2550 }
2551 
2552 static int coroutine_fn raw_co_flush_to_disk(BlockDriverState *bs)
2553 {
2554     BDRVRawState *s = bs->opaque;
2555     RawPosixAIOData acb;
2556     int ret;
2557 
2558     ret = fd_open(bs);
2559     if (ret < 0) {
2560         return ret;
2561     }
2562 
2563     acb = (RawPosixAIOData) {
2564         .bs             = bs,
2565         .aio_fildes     = s->fd,
2566         .aio_type       = QEMU_AIO_FLUSH,
2567     };
2568 
2569 #ifdef CONFIG_LINUX_IO_URING
2570     if (s->use_linux_io_uring) {
2571         return luring_co_submit(bs, s->fd, 0, NULL, QEMU_AIO_FLUSH);
2572     }
2573 #endif
2574     return raw_thread_pool_submit(handle_aiocb_flush, &acb);
2575 }
2576 
2577 static void raw_aio_attach_aio_context(BlockDriverState *bs,
2578                                        AioContext *new_context)
2579 {
2580     BDRVRawState __attribute__((unused)) *s = bs->opaque;
2581 #ifdef CONFIG_LINUX_AIO
2582     if (s->use_linux_aio) {
2583         Error *local_err = NULL;
2584         if (!aio_setup_linux_aio(new_context, &local_err)) {
2585             error_reportf_err(local_err, "Unable to use native AIO, "
2586                                          "falling back to thread pool: ");
2587             s->use_linux_aio = false;
2588         }
2589     }
2590 #endif
2591 #ifdef CONFIG_LINUX_IO_URING
2592     if (s->use_linux_io_uring) {
2593         Error *local_err = NULL;
2594         if (!aio_setup_linux_io_uring(new_context, &local_err)) {
2595             error_reportf_err(local_err, "Unable to use linux io_uring, "
2596                                          "falling back to thread pool: ");
2597             s->use_linux_io_uring = false;
2598         }
2599     }
2600 #endif
2601 }
2602 
2603 static void raw_close(BlockDriverState *bs)
2604 {
2605     BDRVRawState *s = bs->opaque;
2606 
2607     if (s->fd >= 0) {
2608 #if defined(CONFIG_BLKZONED)
2609         g_free(bs->wps);
2610 #endif
2611         qemu_close(s->fd);
2612         s->fd = -1;
2613     }
2614 }
2615 
2616 /**
2617  * Truncates the given regular file @fd to @offset and, when growing, fills the
2618  * new space according to @prealloc.
2619  *
2620  * Returns: 0 on success, -errno on failure.
2621  */
2622 static int coroutine_fn
2623 raw_regular_truncate(BlockDriverState *bs, int fd, int64_t offset,
2624                      PreallocMode prealloc, Error **errp)
2625 {
2626     RawPosixAIOData acb;
2627 
2628     acb = (RawPosixAIOData) {
2629         .bs             = bs,
2630         .aio_fildes     = fd,
2631         .aio_type       = QEMU_AIO_TRUNCATE,
2632         .aio_offset     = offset,
2633         .truncate       = {
2634             .prealloc       = prealloc,
2635             .errp           = errp,
2636         },
2637     };
2638 
2639     return raw_thread_pool_submit(handle_aiocb_truncate, &acb);
2640 }
2641 
2642 static int coroutine_fn raw_co_truncate(BlockDriverState *bs, int64_t offset,
2643                                         bool exact, PreallocMode prealloc,
2644                                         BdrvRequestFlags flags, Error **errp)
2645 {
2646     BDRVRawState *s = bs->opaque;
2647     struct stat st;
2648     int ret;
2649 
2650     if (fstat(s->fd, &st)) {
2651         ret = -errno;
2652         error_setg_errno(errp, -ret, "Failed to fstat() the file");
2653         return ret;
2654     }
2655 
2656     if (S_ISREG(st.st_mode)) {
2657         /* Always resizes to the exact @offset */
2658         return raw_regular_truncate(bs, s->fd, offset, prealloc, errp);
2659     }
2660 
2661     if (prealloc != PREALLOC_MODE_OFF) {
2662         error_setg(errp, "Preallocation mode '%s' unsupported for this "
2663                    "non-regular file", PreallocMode_str(prealloc));
2664         return -ENOTSUP;
2665     }
2666 
2667     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
2668         int64_t cur_length = raw_getlength(bs);
2669 
2670         if (offset != cur_length && exact) {
2671             error_setg(errp, "Cannot resize device files");
2672             return -ENOTSUP;
2673         } else if (offset > cur_length) {
2674             error_setg(errp, "Cannot grow device files");
2675             return -EINVAL;
2676         }
2677     } else {
2678         error_setg(errp, "Resizing this file is not supported");
2679         return -ENOTSUP;
2680     }
2681 
2682     return 0;
2683 }
2684 
2685 #ifdef __OpenBSD__
2686 static int64_t raw_getlength(BlockDriverState *bs)
2687 {
2688     BDRVRawState *s = bs->opaque;
2689     int fd = s->fd;
2690     struct stat st;
2691 
2692     if (fstat(fd, &st))
2693         return -errno;
2694     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
2695         struct disklabel dl;
2696 
2697         if (ioctl(fd, DIOCGDINFO, &dl))
2698             return -errno;
2699         return (uint64_t)dl.d_secsize *
2700             dl.d_partitions[DISKPART(st.st_rdev)].p_size;
2701     } else
2702         return st.st_size;
2703 }
2704 #elif defined(__NetBSD__)
2705 static int64_t raw_getlength(BlockDriverState *bs)
2706 {
2707     BDRVRawState *s = bs->opaque;
2708     int fd = s->fd;
2709     struct stat st;
2710 
2711     if (fstat(fd, &st))
2712         return -errno;
2713     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
2714         struct dkwedge_info dkw;
2715 
2716         if (ioctl(fd, DIOCGWEDGEINFO, &dkw) != -1) {
2717             return dkw.dkw_size * 512;
2718         } else {
2719             struct disklabel dl;
2720 
2721             if (ioctl(fd, DIOCGDINFO, &dl))
2722                 return -errno;
2723             return (uint64_t)dl.d_secsize *
2724                 dl.d_partitions[DISKPART(st.st_rdev)].p_size;
2725         }
2726     } else
2727         return st.st_size;
2728 }
2729 #elif defined(__sun__)
2730 static int64_t raw_getlength(BlockDriverState *bs)
2731 {
2732     BDRVRawState *s = bs->opaque;
2733     struct dk_minfo minfo;
2734     int ret;
2735     int64_t size;
2736 
2737     ret = fd_open(bs);
2738     if (ret < 0) {
2739         return ret;
2740     }
2741 
2742     /*
2743      * Use the DKIOCGMEDIAINFO ioctl to read the size.
2744      */
2745     ret = ioctl(s->fd, DKIOCGMEDIAINFO, &minfo);
2746     if (ret != -1) {
2747         return minfo.dki_lbsize * minfo.dki_capacity;
2748     }
2749 
2750     /*
2751      * There are reports that lseek on some devices fails, but
2752      * irc discussion said that contingency on contingency was overkill.
2753      */
2754     size = lseek(s->fd, 0, SEEK_END);
2755     if (size < 0) {
2756         return -errno;
2757     }
2758     return size;
2759 }
2760 #elif defined(CONFIG_BSD)
2761 static int64_t raw_getlength(BlockDriverState *bs)
2762 {
2763     BDRVRawState *s = bs->opaque;
2764     int fd = s->fd;
2765     int64_t size;
2766     struct stat sb;
2767 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
2768     int reopened = 0;
2769 #endif
2770     int ret;
2771 
2772     ret = fd_open(bs);
2773     if (ret < 0)
2774         return ret;
2775 
2776 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
2777 again:
2778 #endif
2779     if (!fstat(fd, &sb) && (S_IFCHR & sb.st_mode)) {
2780         size = 0;
2781 #ifdef DIOCGMEDIASIZE
2782         if (ioctl(fd, DIOCGMEDIASIZE, (off_t *)&size)) {
2783             size = 0;
2784         }
2785 #endif
2786 #ifdef DIOCGPART
2787         if (size == 0) {
2788             struct partinfo pi;
2789             if (ioctl(fd, DIOCGPART, &pi) == 0) {
2790                 size = pi.media_size;
2791             }
2792         }
2793 #endif
2794 #if defined(DKIOCGETBLOCKCOUNT) && defined(DKIOCGETBLOCKSIZE)
2795         if (size == 0) {
2796             uint64_t sectors = 0;
2797             uint32_t sector_size = 0;
2798 
2799             if (ioctl(fd, DKIOCGETBLOCKCOUNT, &sectors) == 0
2800                && ioctl(fd, DKIOCGETBLOCKSIZE, &sector_size) == 0) {
2801                 size = sectors * sector_size;
2802             }
2803         }
2804 #endif
2805         if (size == 0) {
2806             size = lseek(fd, 0LL, SEEK_END);
2807         }
2808         if (size < 0) {
2809             return -errno;
2810         }
2811 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
2812         switch(s->type) {
2813         case FTYPE_CD:
2814             /* XXX FreeBSD acd returns UINT_MAX sectors for an empty drive */
2815             if (size == 2048LL * (unsigned)-1)
2816                 size = 0;
2817             /* XXX no disc?  maybe we need to reopen... */
2818             if (size <= 0 && !reopened && cdrom_reopen(bs) >= 0) {
2819                 reopened = 1;
2820                 goto again;
2821             }
2822         }
2823 #endif
2824     } else {
2825         size = lseek(fd, 0, SEEK_END);
2826         if (size < 0) {
2827             return -errno;
2828         }
2829     }
2830     return size;
2831 }
2832 #else
2833 static int64_t raw_getlength(BlockDriverState *bs)
2834 {
2835     BDRVRawState *s = bs->opaque;
2836     int ret;
2837     int64_t size;
2838 
2839     ret = fd_open(bs);
2840     if (ret < 0) {
2841         return ret;
2842     }
2843 
2844     size = lseek(s->fd, 0, SEEK_END);
2845     if (size < 0) {
2846         return -errno;
2847     }
2848     return size;
2849 }
2850 #endif
2851 
2852 static int64_t coroutine_fn raw_co_getlength(BlockDriverState *bs)
2853 {
2854     return raw_getlength(bs);
2855 }
2856 
2857 static int64_t coroutine_fn raw_co_get_allocated_file_size(BlockDriverState *bs)
2858 {
2859     struct stat st;
2860     BDRVRawState *s = bs->opaque;
2861 
2862     if (fstat(s->fd, &st) < 0) {
2863         return -errno;
2864     }
2865     return (int64_t)st.st_blocks * 512;
2866 }
2867 
2868 static int coroutine_fn
2869 raw_co_create(BlockdevCreateOptions *options, Error **errp)
2870 {
2871     BlockdevCreateOptionsFile *file_opts;
2872     Error *local_err = NULL;
2873     int fd;
2874     uint64_t perm, shared;
2875     int result = 0;
2876 
2877     /* Validate options and set default values */
2878     assert(options->driver == BLOCKDEV_DRIVER_FILE);
2879     file_opts = &options->u.file;
2880 
2881     if (!file_opts->has_nocow) {
2882         file_opts->nocow = false;
2883     }
2884     if (!file_opts->has_preallocation) {
2885         file_opts->preallocation = PREALLOC_MODE_OFF;
2886     }
2887     if (!file_opts->has_extent_size_hint) {
2888         file_opts->extent_size_hint = 1 * MiB;
2889     }
2890     if (file_opts->extent_size_hint > UINT32_MAX) {
2891         result = -EINVAL;
2892         error_setg(errp, "Extent size hint is too large");
2893         goto out;
2894     }
2895 
2896     /* Create file */
2897     fd = qemu_create(file_opts->filename, O_RDWR | O_BINARY, 0644, errp);
2898     if (fd < 0) {
2899         result = -errno;
2900         goto out;
2901     }
2902 
2903     /* Take permissions: We want to discard everything, so we need
2904      * BLK_PERM_WRITE; and truncation to the desired size requires
2905      * BLK_PERM_RESIZE.
2906      * On the other hand, we cannot share the RESIZE permission
2907      * because we promise that after this function, the file has the
2908      * size given in the options.  If someone else were to resize it
2909      * concurrently, we could not guarantee that.
2910      * Note that after this function, we can no longer guarantee that
2911      * the file is not touched by a third party, so it may be resized
2912      * then. */
2913     perm = BLK_PERM_WRITE | BLK_PERM_RESIZE;
2914     shared = BLK_PERM_ALL & ~BLK_PERM_RESIZE;
2915 
2916     /* Step one: Take locks */
2917     result = raw_apply_lock_bytes(NULL, fd, perm, ~shared, false, errp);
2918     if (result < 0) {
2919         goto out_close;
2920     }
2921 
2922     /* Step two: Check that nobody else has taken conflicting locks */
2923     result = raw_check_lock_bytes(fd, perm, shared, errp);
2924     if (result < 0) {
2925         error_append_hint(errp,
2926                           "Is another process using the image [%s]?\n",
2927                           file_opts->filename);
2928         goto out_unlock;
2929     }
2930 
2931     /* Clear the file by truncating it to 0 */
2932     result = raw_regular_truncate(NULL, fd, 0, PREALLOC_MODE_OFF, errp);
2933     if (result < 0) {
2934         goto out_unlock;
2935     }
2936 
2937     if (file_opts->nocow) {
2938 #ifdef __linux__
2939         /* Set NOCOW flag to solve performance issue on fs like btrfs.
2940          * This is an optimisation. The FS_IOC_SETFLAGS ioctl return value
2941          * will be ignored since any failure of this operation should not
2942          * block the left work.
2943          */
2944         int attr;
2945         if (ioctl(fd, FS_IOC_GETFLAGS, &attr) == 0) {
2946             attr |= FS_NOCOW_FL;
2947             ioctl(fd, FS_IOC_SETFLAGS, &attr);
2948         }
2949 #endif
2950     }
2951 #ifdef FS_IOC_FSSETXATTR
2952     /*
2953      * Try to set the extent size hint. Failure is not fatal, and a warning is
2954      * only printed if the option was explicitly specified.
2955      */
2956     {
2957         struct fsxattr attr;
2958         result = ioctl(fd, FS_IOC_FSGETXATTR, &attr);
2959         if (result == 0) {
2960             attr.fsx_xflags |= FS_XFLAG_EXTSIZE;
2961             attr.fsx_extsize = file_opts->extent_size_hint;
2962             result = ioctl(fd, FS_IOC_FSSETXATTR, &attr);
2963         }
2964         if (result < 0 && file_opts->has_extent_size_hint &&
2965             file_opts->extent_size_hint)
2966         {
2967             warn_report("Failed to set extent size hint: %s",
2968                         strerror(errno));
2969         }
2970     }
2971 #endif
2972 
2973     /* Resize and potentially preallocate the file to the desired
2974      * final size */
2975     result = raw_regular_truncate(NULL, fd, file_opts->size,
2976                                   file_opts->preallocation, errp);
2977     if (result < 0) {
2978         goto out_unlock;
2979     }
2980 
2981 out_unlock:
2982     raw_apply_lock_bytes(NULL, fd, 0, 0, true, &local_err);
2983     if (local_err) {
2984         /* The above call should not fail, and if it does, that does
2985          * not mean the whole creation operation has failed.  So
2986          * report it the user for their convenience, but do not report
2987          * it to the caller. */
2988         warn_report_err(local_err);
2989     }
2990 
2991 out_close:
2992     if (qemu_close(fd) != 0 && result == 0) {
2993         result = -errno;
2994         error_setg_errno(errp, -result, "Could not close the new file");
2995     }
2996 out:
2997     return result;
2998 }
2999 
3000 static int coroutine_fn GRAPH_RDLOCK
3001 raw_co_create_opts(BlockDriver *drv, const char *filename,
3002                    QemuOpts *opts, Error **errp)
3003 {
3004     BlockdevCreateOptions options;
3005     int64_t total_size = 0;
3006     int64_t extent_size_hint = 0;
3007     bool has_extent_size_hint = false;
3008     bool nocow = false;
3009     PreallocMode prealloc;
3010     char *buf = NULL;
3011     Error *local_err = NULL;
3012 
3013     /* Skip file: protocol prefix */
3014     strstart(filename, "file:", &filename);
3015 
3016     /* Read out options */
3017     total_size = ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
3018                           BDRV_SECTOR_SIZE);
3019     if (qemu_opt_get(opts, BLOCK_OPT_EXTENT_SIZE_HINT)) {
3020         has_extent_size_hint = true;
3021         extent_size_hint =
3022             qemu_opt_get_size_del(opts, BLOCK_OPT_EXTENT_SIZE_HINT, -1);
3023     }
3024     nocow = qemu_opt_get_bool(opts, BLOCK_OPT_NOCOW, false);
3025     buf = qemu_opt_get_del(opts, BLOCK_OPT_PREALLOC);
3026     prealloc = qapi_enum_parse(&PreallocMode_lookup, buf,
3027                                PREALLOC_MODE_OFF, &local_err);
3028     g_free(buf);
3029     if (local_err) {
3030         error_propagate(errp, local_err);
3031         return -EINVAL;
3032     }
3033 
3034     options = (BlockdevCreateOptions) {
3035         .driver     = BLOCKDEV_DRIVER_FILE,
3036         .u.file     = {
3037             .filename           = (char *) filename,
3038             .size               = total_size,
3039             .has_preallocation  = true,
3040             .preallocation      = prealloc,
3041             .has_nocow          = true,
3042             .nocow              = nocow,
3043             .has_extent_size_hint = has_extent_size_hint,
3044             .extent_size_hint   = extent_size_hint,
3045         },
3046     };
3047     return raw_co_create(&options, errp);
3048 }
3049 
3050 static int coroutine_fn raw_co_delete_file(BlockDriverState *bs,
3051                                            Error **errp)
3052 {
3053     struct stat st;
3054     int ret;
3055 
3056     if (!(stat(bs->filename, &st) == 0) || !S_ISREG(st.st_mode)) {
3057         error_setg_errno(errp, ENOENT, "%s is not a regular file",
3058                          bs->filename);
3059         return -ENOENT;
3060     }
3061 
3062     ret = unlink(bs->filename);
3063     if (ret < 0) {
3064         ret = -errno;
3065         error_setg_errno(errp, -ret, "Error when deleting file %s",
3066                          bs->filename);
3067     }
3068 
3069     return ret;
3070 }
3071 
3072 /*
3073  * Find allocation range in @bs around offset @start.
3074  * May change underlying file descriptor's file offset.
3075  * If @start is not in a hole, store @start in @data, and the
3076  * beginning of the next hole in @hole, and return 0.
3077  * If @start is in a non-trailing hole, store @start in @hole and the
3078  * beginning of the next non-hole in @data, and return 0.
3079  * If @start is in a trailing hole or beyond EOF, return -ENXIO.
3080  * If we can't find out, return a negative errno other than -ENXIO.
3081  */
3082 static int find_allocation(BlockDriverState *bs, off_t start,
3083                            off_t *data, off_t *hole)
3084 {
3085 #if defined SEEK_HOLE && defined SEEK_DATA
3086     BDRVRawState *s = bs->opaque;
3087     off_t offs;
3088 
3089     /*
3090      * SEEK_DATA cases:
3091      * D1. offs == start: start is in data
3092      * D2. offs > start: start is in a hole, next data at offs
3093      * D3. offs < 0, errno = ENXIO: either start is in a trailing hole
3094      *                              or start is beyond EOF
3095      *     If the latter happens, the file has been truncated behind
3096      *     our back since we opened it.  All bets are off then.
3097      *     Treating like a trailing hole is simplest.
3098      * D4. offs < 0, errno != ENXIO: we learned nothing
3099      */
3100     offs = lseek(s->fd, start, SEEK_DATA);
3101     if (offs < 0) {
3102         return -errno;          /* D3 or D4 */
3103     }
3104 
3105     if (offs < start) {
3106         /* This is not a valid return by lseek().  We are safe to just return
3107          * -EIO in this case, and we'll treat it like D4. */
3108         return -EIO;
3109     }
3110 
3111     if (offs > start) {
3112         /* D2: in hole, next data at offs */
3113         *hole = start;
3114         *data = offs;
3115         return 0;
3116     }
3117 
3118     /* D1: in data, end not yet known */
3119 
3120     /*
3121      * SEEK_HOLE cases:
3122      * H1. offs == start: start is in a hole
3123      *     If this happens here, a hole has been dug behind our back
3124      *     since the previous lseek().
3125      * H2. offs > start: either start is in data, next hole at offs,
3126      *                   or start is in trailing hole, EOF at offs
3127      *     Linux treats trailing holes like any other hole: offs ==
3128      *     start.  Solaris seeks to EOF instead: offs > start (blech).
3129      *     If that happens here, a hole has been dug behind our back
3130      *     since the previous lseek().
3131      * H3. offs < 0, errno = ENXIO: start is beyond EOF
3132      *     If this happens, the file has been truncated behind our
3133      *     back since we opened it.  Treat it like a trailing hole.
3134      * H4. offs < 0, errno != ENXIO: we learned nothing
3135      *     Pretend we know nothing at all, i.e. "forget" about D1.
3136      */
3137     offs = lseek(s->fd, start, SEEK_HOLE);
3138     if (offs < 0) {
3139         return -errno;          /* D1 and (H3 or H4) */
3140     }
3141 
3142     if (offs < start) {
3143         /* This is not a valid return by lseek().  We are safe to just return
3144          * -EIO in this case, and we'll treat it like H4. */
3145         return -EIO;
3146     }
3147 
3148     if (offs > start) {
3149         /*
3150          * D1 and H2: either in data, next hole at offs, or it was in
3151          * data but is now in a trailing hole.  In the latter case,
3152          * all bets are off.  Treating it as if it there was data all
3153          * the way to EOF is safe, so simply do that.
3154          */
3155         *data = start;
3156         *hole = offs;
3157         return 0;
3158     }
3159 
3160     /* D1 and H1 */
3161     return -EBUSY;
3162 #else
3163     return -ENOTSUP;
3164 #endif
3165 }
3166 
3167 /*
3168  * Returns the allocation status of the specified offset.
3169  *
3170  * The block layer guarantees 'offset' and 'bytes' are within bounds.
3171  *
3172  * 'pnum' is set to the number of bytes (including and immediately following
3173  * the specified offset) that are known to be in the same
3174  * allocated/unallocated state.
3175  *
3176  * 'bytes' is a soft cap for 'pnum'.  If the information is free, 'pnum' may
3177  * well exceed it.
3178  */
3179 static int coroutine_fn raw_co_block_status(BlockDriverState *bs,
3180                                             bool want_zero,
3181                                             int64_t offset,
3182                                             int64_t bytes, int64_t *pnum,
3183                                             int64_t *map,
3184                                             BlockDriverState **file)
3185 {
3186     off_t data = 0, hole = 0;
3187     int ret;
3188 
3189     assert(QEMU_IS_ALIGNED(offset | bytes, bs->bl.request_alignment));
3190 
3191     ret = fd_open(bs);
3192     if (ret < 0) {
3193         return ret;
3194     }
3195 
3196     if (!want_zero) {
3197         *pnum = bytes;
3198         *map = offset;
3199         *file = bs;
3200         return BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
3201     }
3202 
3203     ret = find_allocation(bs, offset, &data, &hole);
3204     if (ret == -ENXIO) {
3205         /* Trailing hole */
3206         *pnum = bytes;
3207         ret = BDRV_BLOCK_ZERO;
3208     } else if (ret < 0) {
3209         /* No info available, so pretend there are no holes */
3210         *pnum = bytes;
3211         ret = BDRV_BLOCK_DATA;
3212     } else if (data == offset) {
3213         /* On a data extent, compute bytes to the end of the extent,
3214          * possibly including a partial sector at EOF. */
3215         *pnum = hole - offset;
3216 
3217         /*
3218          * We are not allowed to return partial sectors, though, so
3219          * round up if necessary.
3220          */
3221         if (!QEMU_IS_ALIGNED(*pnum, bs->bl.request_alignment)) {
3222             int64_t file_length = raw_getlength(bs);
3223             if (file_length > 0) {
3224                 /* Ignore errors, this is just a safeguard */
3225                 assert(hole == file_length);
3226             }
3227             *pnum = ROUND_UP(*pnum, bs->bl.request_alignment);
3228         }
3229 
3230         ret = BDRV_BLOCK_DATA;
3231     } else {
3232         /* On a hole, compute bytes to the beginning of the next extent.  */
3233         assert(hole == offset);
3234         *pnum = data - offset;
3235         ret = BDRV_BLOCK_ZERO;
3236     }
3237     *map = offset;
3238     *file = bs;
3239     return ret | BDRV_BLOCK_OFFSET_VALID;
3240 }
3241 
3242 #if defined(__linux__)
3243 /* Verify that the file is not in the page cache */
3244 static void check_cache_dropped(BlockDriverState *bs, Error **errp)
3245 {
3246     const size_t window_size = 128 * 1024 * 1024;
3247     BDRVRawState *s = bs->opaque;
3248     void *window = NULL;
3249     size_t length = 0;
3250     unsigned char *vec;
3251     size_t page_size;
3252     off_t offset;
3253     off_t end;
3254 
3255     /* mincore(2) page status information requires 1 byte per page */
3256     page_size = sysconf(_SC_PAGESIZE);
3257     vec = g_malloc(DIV_ROUND_UP(window_size, page_size));
3258 
3259     end = raw_getlength(bs);
3260 
3261     for (offset = 0; offset < end; offset += window_size) {
3262         void *new_window;
3263         size_t new_length;
3264         size_t vec_end;
3265         size_t i;
3266         int ret;
3267 
3268         /* Unmap previous window if size has changed */
3269         new_length = MIN(end - offset, window_size);
3270         if (new_length != length) {
3271             munmap(window, length);
3272             window = NULL;
3273             length = 0;
3274         }
3275 
3276         new_window = mmap(window, new_length, PROT_NONE, MAP_PRIVATE,
3277                           s->fd, offset);
3278         if (new_window == MAP_FAILED) {
3279             error_setg_errno(errp, errno, "mmap failed");
3280             break;
3281         }
3282 
3283         window = new_window;
3284         length = new_length;
3285 
3286         ret = mincore(window, length, vec);
3287         if (ret < 0) {
3288             error_setg_errno(errp, errno, "mincore failed");
3289             break;
3290         }
3291 
3292         vec_end = DIV_ROUND_UP(length, page_size);
3293         for (i = 0; i < vec_end; i++) {
3294             if (vec[i] & 0x1) {
3295                 break;
3296             }
3297         }
3298         if (i < vec_end) {
3299             error_setg(errp, "page cache still in use!");
3300             break;
3301         }
3302     }
3303 
3304     if (window) {
3305         munmap(window, length);
3306     }
3307 
3308     g_free(vec);
3309 }
3310 #endif /* __linux__ */
3311 
3312 static void coroutine_fn GRAPH_RDLOCK
3313 raw_co_invalidate_cache(BlockDriverState *bs, Error **errp)
3314 {
3315     BDRVRawState *s = bs->opaque;
3316     int ret;
3317 
3318     ret = fd_open(bs);
3319     if (ret < 0) {
3320         error_setg_errno(errp, -ret, "The file descriptor is not open");
3321         return;
3322     }
3323 
3324     if (!s->drop_cache) {
3325         return;
3326     }
3327 
3328     if (s->open_flags & O_DIRECT) {
3329         return; /* No host kernel page cache */
3330     }
3331 
3332 #if defined(__linux__)
3333     /* This sets the scene for the next syscall... */
3334     ret = bdrv_co_flush(bs);
3335     if (ret < 0) {
3336         error_setg_errno(errp, -ret, "flush failed");
3337         return;
3338     }
3339 
3340     /* Linux does not invalidate pages that are dirty, locked, or mmapped by a
3341      * process.  These limitations are okay because we just fsynced the file,
3342      * we don't use mmap, and the file should not be in use by other processes.
3343      */
3344     ret = posix_fadvise(s->fd, 0, 0, POSIX_FADV_DONTNEED);
3345     if (ret != 0) { /* the return value is a positive errno */
3346         error_setg_errno(errp, ret, "fadvise failed");
3347         return;
3348     }
3349 
3350     if (s->check_cache_dropped) {
3351         check_cache_dropped(bs, errp);
3352     }
3353 #else /* __linux__ */
3354     /* Do nothing.  Live migration to a remote host with cache.direct=off is
3355      * unsupported on other host operating systems.  Cache consistency issues
3356      * may occur but no error is reported here, partly because that's the
3357      * historical behavior and partly because it's hard to differentiate valid
3358      * configurations that should not cause errors.
3359      */
3360 #endif /* !__linux__ */
3361 }
3362 
3363 static void raw_account_discard(BDRVRawState *s, uint64_t nbytes, int ret)
3364 {
3365     if (ret) {
3366         s->stats.discard_nb_failed++;
3367     } else {
3368         s->stats.discard_nb_ok++;
3369         s->stats.discard_bytes_ok += nbytes;
3370     }
3371 }
3372 
3373 /*
3374  * zone report - Get a zone block device's information in the form
3375  * of an array of zone descriptors.
3376  * zones is an array of zone descriptors to hold zone information on reply;
3377  * offset can be any byte within the entire size of the device;
3378  * nr_zones is the maximum number of sectors the command should operate on.
3379  */
3380 #if defined(CONFIG_BLKZONED)
3381 static int coroutine_fn raw_co_zone_report(BlockDriverState *bs, int64_t offset,
3382                                            unsigned int *nr_zones,
3383                                            BlockZoneDescriptor *zones) {
3384     BDRVRawState *s = bs->opaque;
3385     RawPosixAIOData acb = (RawPosixAIOData) {
3386         .bs         = bs,
3387         .aio_fildes = s->fd,
3388         .aio_type   = QEMU_AIO_ZONE_REPORT,
3389         .aio_offset = offset,
3390         .zone_report    = {
3391             .nr_zones       = nr_zones,
3392             .zones          = zones,
3393         },
3394     };
3395 
3396     trace_zbd_zone_report(bs, *nr_zones, offset >> BDRV_SECTOR_BITS);
3397     return raw_thread_pool_submit(handle_aiocb_zone_report, &acb);
3398 }
3399 #endif
3400 
3401 /*
3402  * zone management operations - Execute an operation on a zone
3403  */
3404 #if defined(CONFIG_BLKZONED)
3405 static int coroutine_fn raw_co_zone_mgmt(BlockDriverState *bs, BlockZoneOp op,
3406         int64_t offset, int64_t len) {
3407     BDRVRawState *s = bs->opaque;
3408     RawPosixAIOData acb;
3409     int64_t zone_size, zone_size_mask;
3410     const char *op_name;
3411     unsigned long zo;
3412     int ret;
3413     BlockZoneWps *wps = bs->wps;
3414     int64_t capacity = bs->total_sectors << BDRV_SECTOR_BITS;
3415 
3416     zone_size = bs->bl.zone_size;
3417     zone_size_mask = zone_size - 1;
3418     if (offset & zone_size_mask) {
3419         error_report("sector offset %" PRId64 " is not aligned to zone size "
3420                      "%" PRId64 "", offset / 512, zone_size / 512);
3421         return -EINVAL;
3422     }
3423 
3424     if (((offset + len) < capacity && len & zone_size_mask) ||
3425         offset + len > capacity) {
3426         error_report("number of sectors %" PRId64 " is not aligned to zone size"
3427                       " %" PRId64 "", len / 512, zone_size / 512);
3428         return -EINVAL;
3429     }
3430 
3431     uint32_t i = offset / bs->bl.zone_size;
3432     uint32_t nrz = len / bs->bl.zone_size;
3433     uint64_t *wp = &wps->wp[i];
3434     if (BDRV_ZT_IS_CONV(*wp) && len != capacity) {
3435         error_report("zone mgmt operations are not allowed for conventional zones");
3436         return -EIO;
3437     }
3438 
3439     switch (op) {
3440     case BLK_ZO_OPEN:
3441         op_name = "BLKOPENZONE";
3442         zo = BLKOPENZONE;
3443         break;
3444     case BLK_ZO_CLOSE:
3445         op_name = "BLKCLOSEZONE";
3446         zo = BLKCLOSEZONE;
3447         break;
3448     case BLK_ZO_FINISH:
3449         op_name = "BLKFINISHZONE";
3450         zo = BLKFINISHZONE;
3451         break;
3452     case BLK_ZO_RESET:
3453         op_name = "BLKRESETZONE";
3454         zo = BLKRESETZONE;
3455         break;
3456     default:
3457         error_report("Unsupported zone op: 0x%x", op);
3458         return -ENOTSUP;
3459     }
3460 
3461     acb = (RawPosixAIOData) {
3462         .bs             = bs,
3463         .aio_fildes     = s->fd,
3464         .aio_type       = QEMU_AIO_ZONE_MGMT,
3465         .aio_offset     = offset,
3466         .aio_nbytes     = len,
3467         .zone_mgmt  = {
3468             .op = zo,
3469         },
3470     };
3471 
3472     trace_zbd_zone_mgmt(bs, op_name, offset >> BDRV_SECTOR_BITS,
3473                         len >> BDRV_SECTOR_BITS);
3474     ret = raw_thread_pool_submit(handle_aiocb_zone_mgmt, &acb);
3475     if (ret != 0) {
3476         update_zones_wp(bs, s->fd, offset, nrz);
3477         error_report("ioctl %s failed %d", op_name, ret);
3478         return ret;
3479     }
3480 
3481     if (zo == BLKRESETZONE && len == capacity) {
3482         ret = get_zones_wp(bs, s->fd, 0, bs->bl.nr_zones, 1);
3483         if (ret < 0) {
3484             error_report("reporting single wp failed");
3485             return ret;
3486         }
3487     } else if (zo == BLKRESETZONE) {
3488         for (unsigned int j = 0; j < nrz; ++j) {
3489             wp[j] = offset + j * zone_size;
3490         }
3491     } else if (zo == BLKFINISHZONE) {
3492         for (unsigned int j = 0; j < nrz; ++j) {
3493             /* The zoned device allows the last zone smaller that the
3494              * zone size. */
3495             wp[j] = MIN(offset + (j + 1) * zone_size, offset + len);
3496         }
3497     }
3498 
3499     return ret;
3500 }
3501 #endif
3502 
3503 #if defined(CONFIG_BLKZONED)
3504 static int coroutine_fn raw_co_zone_append(BlockDriverState *bs,
3505                                            int64_t *offset,
3506                                            QEMUIOVector *qiov,
3507                                            BdrvRequestFlags flags) {
3508     assert(flags == 0);
3509     int64_t zone_size_mask = bs->bl.zone_size - 1;
3510     int64_t iov_len = 0;
3511     int64_t len = 0;
3512 
3513     if (*offset & zone_size_mask) {
3514         error_report("sector offset %" PRId64 " is not aligned to zone size "
3515                      "%" PRId32 "", *offset / 512, bs->bl.zone_size / 512);
3516         return -EINVAL;
3517     }
3518 
3519     int64_t wg = bs->bl.write_granularity;
3520     int64_t wg_mask = wg - 1;
3521     for (int i = 0; i < qiov->niov; i++) {
3522         iov_len = qiov->iov[i].iov_len;
3523         if (iov_len & wg_mask) {
3524             error_report("len of IOVector[%d] %" PRId64 " is not aligned to "
3525                          "block size %" PRId64 "", i, iov_len, wg);
3526             return -EINVAL;
3527         }
3528         len += iov_len;
3529     }
3530 
3531     trace_zbd_zone_append(bs, *offset >> BDRV_SECTOR_BITS);
3532     return raw_co_prw(bs, offset, len, qiov, QEMU_AIO_ZONE_APPEND);
3533 }
3534 #endif
3535 
3536 static coroutine_fn int
3537 raw_do_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes,
3538                 bool blkdev)
3539 {
3540     BDRVRawState *s = bs->opaque;
3541     RawPosixAIOData acb;
3542     int ret;
3543 
3544     acb = (RawPosixAIOData) {
3545         .bs             = bs,
3546         .aio_fildes     = s->fd,
3547         .aio_type       = QEMU_AIO_DISCARD,
3548         .aio_offset     = offset,
3549         .aio_nbytes     = bytes,
3550     };
3551 
3552     if (blkdev) {
3553         acb.aio_type |= QEMU_AIO_BLKDEV;
3554     }
3555 
3556     ret = raw_thread_pool_submit(handle_aiocb_discard, &acb);
3557     raw_account_discard(s, bytes, ret);
3558     return ret;
3559 }
3560 
3561 static coroutine_fn int
3562 raw_co_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes)
3563 {
3564     return raw_do_pdiscard(bs, offset, bytes, false);
3565 }
3566 
3567 static int coroutine_fn
3568 raw_do_pwrite_zeroes(BlockDriverState *bs, int64_t offset, int64_t bytes,
3569                      BdrvRequestFlags flags, bool blkdev)
3570 {
3571     BDRVRawState *s = bs->opaque;
3572     RawPosixAIOData acb;
3573     ThreadPoolFunc *handler;
3574 
3575 #ifdef CONFIG_FALLOCATE
3576     if (offset + bytes > bs->total_sectors * BDRV_SECTOR_SIZE) {
3577         BdrvTrackedRequest *req;
3578 
3579         /*
3580          * This is a workaround for a bug in the Linux XFS driver,
3581          * where writes submitted through the AIO interface will be
3582          * discarded if they happen beyond a concurrently running
3583          * fallocate() that increases the file length (i.e., both the
3584          * write and the fallocate() happen beyond the EOF).
3585          *
3586          * To work around it, we extend the tracked request for this
3587          * zero write until INT64_MAX (effectively infinity), and mark
3588          * it as serializing.
3589          *
3590          * We have to enable this workaround for all filesystems and
3591          * AIO modes (not just XFS with aio=native), because for
3592          * remote filesystems we do not know the host configuration.
3593          */
3594 
3595         req = bdrv_co_get_self_request(bs);
3596         assert(req);
3597         assert(req->type == BDRV_TRACKED_WRITE);
3598         assert(req->offset <= offset);
3599         assert(req->offset + req->bytes >= offset + bytes);
3600 
3601         req->bytes = BDRV_MAX_LENGTH - req->offset;
3602 
3603         bdrv_check_request(req->offset, req->bytes, &error_abort);
3604 
3605         bdrv_make_request_serialising(req, bs->bl.request_alignment);
3606     }
3607 #endif
3608 
3609     acb = (RawPosixAIOData) {
3610         .bs             = bs,
3611         .aio_fildes     = s->fd,
3612         .aio_type       = QEMU_AIO_WRITE_ZEROES,
3613         .aio_offset     = offset,
3614         .aio_nbytes     = bytes,
3615     };
3616 
3617     if (blkdev) {
3618         acb.aio_type |= QEMU_AIO_BLKDEV;
3619     }
3620     if (flags & BDRV_REQ_NO_FALLBACK) {
3621         acb.aio_type |= QEMU_AIO_NO_FALLBACK;
3622     }
3623 
3624     if (flags & BDRV_REQ_MAY_UNMAP) {
3625         acb.aio_type |= QEMU_AIO_DISCARD;
3626         handler = handle_aiocb_write_zeroes_unmap;
3627     } else {
3628         handler = handle_aiocb_write_zeroes;
3629     }
3630 
3631     return raw_thread_pool_submit(handler, &acb);
3632 }
3633 
3634 static int coroutine_fn raw_co_pwrite_zeroes(
3635     BlockDriverState *bs, int64_t offset,
3636     int64_t bytes, BdrvRequestFlags flags)
3637 {
3638     return raw_do_pwrite_zeroes(bs, offset, bytes, flags, false);
3639 }
3640 
3641 static int coroutine_fn
3642 raw_co_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
3643 {
3644     return 0;
3645 }
3646 
3647 static ImageInfoSpecific *raw_get_specific_info(BlockDriverState *bs,
3648                                                 Error **errp)
3649 {
3650     ImageInfoSpecificFile *file_info = g_new0(ImageInfoSpecificFile, 1);
3651     ImageInfoSpecific *spec_info = g_new(ImageInfoSpecific, 1);
3652 
3653     *spec_info = (ImageInfoSpecific){
3654         .type = IMAGE_INFO_SPECIFIC_KIND_FILE,
3655         .u.file.data = file_info,
3656     };
3657 
3658 #ifdef FS_IOC_FSGETXATTR
3659     {
3660         BDRVRawState *s = bs->opaque;
3661         struct fsxattr attr;
3662         int ret;
3663 
3664         ret = ioctl(s->fd, FS_IOC_FSGETXATTR, &attr);
3665         if (!ret && attr.fsx_extsize != 0) {
3666             file_info->has_extent_size_hint = true;
3667             file_info->extent_size_hint = attr.fsx_extsize;
3668         }
3669     }
3670 #endif
3671 
3672     return spec_info;
3673 }
3674 
3675 static BlockStatsSpecificFile get_blockstats_specific_file(BlockDriverState *bs)
3676 {
3677     BDRVRawState *s = bs->opaque;
3678     return (BlockStatsSpecificFile) {
3679         .discard_nb_ok = s->stats.discard_nb_ok,
3680         .discard_nb_failed = s->stats.discard_nb_failed,
3681         .discard_bytes_ok = s->stats.discard_bytes_ok,
3682     };
3683 }
3684 
3685 static BlockStatsSpecific *raw_get_specific_stats(BlockDriverState *bs)
3686 {
3687     BlockStatsSpecific *stats = g_new(BlockStatsSpecific, 1);
3688 
3689     stats->driver = BLOCKDEV_DRIVER_FILE;
3690     stats->u.file = get_blockstats_specific_file(bs);
3691 
3692     return stats;
3693 }
3694 
3695 #if defined(HAVE_HOST_BLOCK_DEVICE)
3696 static BlockStatsSpecific *hdev_get_specific_stats(BlockDriverState *bs)
3697 {
3698     BlockStatsSpecific *stats = g_new(BlockStatsSpecific, 1);
3699 
3700     stats->driver = BLOCKDEV_DRIVER_HOST_DEVICE;
3701     stats->u.host_device = get_blockstats_specific_file(bs);
3702 
3703     return stats;
3704 }
3705 #endif /* HAVE_HOST_BLOCK_DEVICE */
3706 
3707 static QemuOptsList raw_create_opts = {
3708     .name = "raw-create-opts",
3709     .head = QTAILQ_HEAD_INITIALIZER(raw_create_opts.head),
3710     .desc = {
3711         {
3712             .name = BLOCK_OPT_SIZE,
3713             .type = QEMU_OPT_SIZE,
3714             .help = "Virtual disk size"
3715         },
3716         {
3717             .name = BLOCK_OPT_NOCOW,
3718             .type = QEMU_OPT_BOOL,
3719             .help = "Turn off copy-on-write (valid only on btrfs)"
3720         },
3721         {
3722             .name = BLOCK_OPT_PREALLOC,
3723             .type = QEMU_OPT_STRING,
3724             .help = "Preallocation mode (allowed values: off"
3725 #ifdef CONFIG_POSIX_FALLOCATE
3726                     ", falloc"
3727 #endif
3728                     ", full)"
3729         },
3730         {
3731             .name = BLOCK_OPT_EXTENT_SIZE_HINT,
3732             .type = QEMU_OPT_SIZE,
3733             .help = "Extent size hint for the image file, 0 to disable"
3734         },
3735         { /* end of list */ }
3736     }
3737 };
3738 
3739 static int raw_check_perm(BlockDriverState *bs, uint64_t perm, uint64_t shared,
3740                           Error **errp)
3741 {
3742     BDRVRawState *s = bs->opaque;
3743     int input_flags = s->reopen_state ? s->reopen_state->flags : bs->open_flags;
3744     int open_flags;
3745     int ret;
3746 
3747     /* We may need a new fd if auto-read-only switches the mode */
3748     ret = raw_reconfigure_getfd(bs, input_flags, &open_flags, perm,
3749                                 false, errp);
3750     if (ret < 0) {
3751         return ret;
3752     } else if (ret != s->fd) {
3753         Error *local_err = NULL;
3754 
3755         /*
3756          * Fail already check_perm() if we can't get a working O_DIRECT
3757          * alignment with the new fd.
3758          */
3759         raw_probe_alignment(bs, ret, &local_err);
3760         if (local_err) {
3761             error_propagate(errp, local_err);
3762             return -EINVAL;
3763         }
3764 
3765         s->perm_change_fd = ret;
3766         s->perm_change_flags = open_flags;
3767     }
3768 
3769     /* Prepare permissions on old fd to avoid conflicts between old and new,
3770      * but keep everything locked that new will need. */
3771     ret = raw_handle_perm_lock(bs, RAW_PL_PREPARE, perm, shared, errp);
3772     if (ret < 0) {
3773         goto fail;
3774     }
3775 
3776     /* Copy locks to the new fd */
3777     if (s->perm_change_fd && s->use_lock) {
3778         ret = raw_apply_lock_bytes(NULL, s->perm_change_fd, perm, ~shared,
3779                                    false, errp);
3780         if (ret < 0) {
3781             raw_handle_perm_lock(bs, RAW_PL_ABORT, 0, 0, NULL);
3782             goto fail;
3783         }
3784     }
3785     return 0;
3786 
3787 fail:
3788     if (s->perm_change_fd) {
3789         qemu_close(s->perm_change_fd);
3790     }
3791     s->perm_change_fd = 0;
3792     return ret;
3793 }
3794 
3795 static void raw_set_perm(BlockDriverState *bs, uint64_t perm, uint64_t shared)
3796 {
3797     BDRVRawState *s = bs->opaque;
3798 
3799     /* For reopen, we have already switched to the new fd (.bdrv_set_perm is
3800      * called after .bdrv_reopen_commit) */
3801     if (s->perm_change_fd && s->fd != s->perm_change_fd) {
3802         qemu_close(s->fd);
3803         s->fd = s->perm_change_fd;
3804         s->open_flags = s->perm_change_flags;
3805     }
3806     s->perm_change_fd = 0;
3807 
3808     raw_handle_perm_lock(bs, RAW_PL_COMMIT, perm, shared, NULL);
3809     s->perm = perm;
3810     s->shared_perm = shared;
3811 }
3812 
3813 static void raw_abort_perm_update(BlockDriverState *bs)
3814 {
3815     BDRVRawState *s = bs->opaque;
3816 
3817     /* For reopen, .bdrv_reopen_abort is called afterwards and will close
3818      * the file descriptor. */
3819     if (s->perm_change_fd) {
3820         qemu_close(s->perm_change_fd);
3821     }
3822     s->perm_change_fd = 0;
3823 
3824     raw_handle_perm_lock(bs, RAW_PL_ABORT, 0, 0, NULL);
3825 }
3826 
3827 static int coroutine_fn GRAPH_RDLOCK raw_co_copy_range_from(
3828         BlockDriverState *bs, BdrvChild *src, int64_t src_offset,
3829         BdrvChild *dst, int64_t dst_offset, int64_t bytes,
3830         BdrvRequestFlags read_flags, BdrvRequestFlags write_flags)
3831 {
3832     return bdrv_co_copy_range_to(src, src_offset, dst, dst_offset, bytes,
3833                                  read_flags, write_flags);
3834 }
3835 
3836 static int coroutine_fn GRAPH_RDLOCK
3837 raw_co_copy_range_to(BlockDriverState *bs,
3838                      BdrvChild *src, int64_t src_offset,
3839                      BdrvChild *dst, int64_t dst_offset,
3840                      int64_t bytes, BdrvRequestFlags read_flags,
3841                      BdrvRequestFlags write_flags)
3842 {
3843     RawPosixAIOData acb;
3844     BDRVRawState *s = bs->opaque;
3845     BDRVRawState *src_s;
3846 
3847     assert(dst->bs == bs);
3848     if (src->bs->drv->bdrv_co_copy_range_to != raw_co_copy_range_to) {
3849         return -ENOTSUP;
3850     }
3851 
3852     src_s = src->bs->opaque;
3853     if (fd_open(src->bs) < 0 || fd_open(dst->bs) < 0) {
3854         return -EIO;
3855     }
3856 
3857     acb = (RawPosixAIOData) {
3858         .bs             = bs,
3859         .aio_type       = QEMU_AIO_COPY_RANGE,
3860         .aio_fildes     = src_s->fd,
3861         .aio_offset     = src_offset,
3862         .aio_nbytes     = bytes,
3863         .copy_range     = {
3864             .aio_fd2        = s->fd,
3865             .aio_offset2    = dst_offset,
3866         },
3867     };
3868 
3869     return raw_thread_pool_submit(handle_aiocb_copy_range, &acb);
3870 }
3871 
3872 BlockDriver bdrv_file = {
3873     .format_name = "file",
3874     .protocol_name = "file",
3875     .instance_size = sizeof(BDRVRawState),
3876     .bdrv_needs_filename = true,
3877     .bdrv_probe = NULL, /* no probe for protocols */
3878     .bdrv_parse_filename = raw_parse_filename,
3879     .bdrv_file_open = raw_open,
3880     .bdrv_reopen_prepare = raw_reopen_prepare,
3881     .bdrv_reopen_commit = raw_reopen_commit,
3882     .bdrv_reopen_abort = raw_reopen_abort,
3883     .bdrv_close = raw_close,
3884     .bdrv_co_create = raw_co_create,
3885     .bdrv_co_create_opts = raw_co_create_opts,
3886     .bdrv_has_zero_init = bdrv_has_zero_init_1,
3887     .bdrv_co_block_status = raw_co_block_status,
3888     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
3889     .bdrv_co_pwrite_zeroes = raw_co_pwrite_zeroes,
3890     .bdrv_co_delete_file = raw_co_delete_file,
3891 
3892     .bdrv_co_preadv         = raw_co_preadv,
3893     .bdrv_co_pwritev        = raw_co_pwritev,
3894     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
3895     .bdrv_co_pdiscard       = raw_co_pdiscard,
3896     .bdrv_co_copy_range_from = raw_co_copy_range_from,
3897     .bdrv_co_copy_range_to  = raw_co_copy_range_to,
3898     .bdrv_refresh_limits = raw_refresh_limits,
3899     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
3900 
3901     .bdrv_co_truncate                   = raw_co_truncate,
3902     .bdrv_co_getlength                  = raw_co_getlength,
3903     .bdrv_co_get_info                   = raw_co_get_info,
3904     .bdrv_get_specific_info             = raw_get_specific_info,
3905     .bdrv_co_get_allocated_file_size    = raw_co_get_allocated_file_size,
3906     .bdrv_get_specific_stats = raw_get_specific_stats,
3907     .bdrv_check_perm = raw_check_perm,
3908     .bdrv_set_perm   = raw_set_perm,
3909     .bdrv_abort_perm_update = raw_abort_perm_update,
3910     .create_opts = &raw_create_opts,
3911     .mutable_opts = mutable_opts,
3912 };
3913 
3914 /***********************************************/
3915 /* host device */
3916 
3917 #if defined(HAVE_HOST_BLOCK_DEVICE)
3918 
3919 #if defined(__APPLE__) && defined(__MACH__)
3920 static kern_return_t GetBSDPath(io_iterator_t mediaIterator, char *bsdPath,
3921                                 CFIndex maxPathSize, int flags);
3922 
3923 #if !defined(MAC_OS_VERSION_12_0) \
3924     || (MAC_OS_X_VERSION_MIN_REQUIRED < MAC_OS_VERSION_12_0)
3925 #define IOMainPort IOMasterPort
3926 #endif
3927 
3928 static char *FindEjectableOpticalMedia(io_iterator_t *mediaIterator)
3929 {
3930     kern_return_t kernResult = KERN_FAILURE;
3931     mach_port_t mainPort;
3932     CFMutableDictionaryRef  classesToMatch;
3933     const char *matching_array[] = {kIODVDMediaClass, kIOCDMediaClass};
3934     char *mediaType = NULL;
3935 
3936     kernResult = IOMainPort(MACH_PORT_NULL, &mainPort);
3937     if ( KERN_SUCCESS != kernResult ) {
3938         printf("IOMainPort returned %d\n", kernResult);
3939     }
3940 
3941     int index;
3942     for (index = 0; index < ARRAY_SIZE(matching_array); index++) {
3943         classesToMatch = IOServiceMatching(matching_array[index]);
3944         if (classesToMatch == NULL) {
3945             error_report("IOServiceMatching returned NULL for %s",
3946                          matching_array[index]);
3947             continue;
3948         }
3949         CFDictionarySetValue(classesToMatch, CFSTR(kIOMediaEjectableKey),
3950                              kCFBooleanTrue);
3951         kernResult = IOServiceGetMatchingServices(mainPort, classesToMatch,
3952                                                   mediaIterator);
3953         if (kernResult != KERN_SUCCESS) {
3954             error_report("Note: IOServiceGetMatchingServices returned %d",
3955                          kernResult);
3956             continue;
3957         }
3958 
3959         /* If a match was found, leave the loop */
3960         if (*mediaIterator != 0) {
3961             trace_file_FindEjectableOpticalMedia(matching_array[index]);
3962             mediaType = g_strdup(matching_array[index]);
3963             break;
3964         }
3965     }
3966     return mediaType;
3967 }
3968 
3969 kern_return_t GetBSDPath(io_iterator_t mediaIterator, char *bsdPath,
3970                          CFIndex maxPathSize, int flags)
3971 {
3972     io_object_t     nextMedia;
3973     kern_return_t   kernResult = KERN_FAILURE;
3974     *bsdPath = '\0';
3975     nextMedia = IOIteratorNext( mediaIterator );
3976     if ( nextMedia )
3977     {
3978         CFTypeRef   bsdPathAsCFString;
3979     bsdPathAsCFString = IORegistryEntryCreateCFProperty( nextMedia, CFSTR( kIOBSDNameKey ), kCFAllocatorDefault, 0 );
3980         if ( bsdPathAsCFString ) {
3981             size_t devPathLength;
3982             strcpy( bsdPath, _PATH_DEV );
3983             if (flags & BDRV_O_NOCACHE) {
3984                 strcat(bsdPath, "r");
3985             }
3986             devPathLength = strlen( bsdPath );
3987             if ( CFStringGetCString( bsdPathAsCFString, bsdPath + devPathLength, maxPathSize - devPathLength, kCFStringEncodingASCII ) ) {
3988                 kernResult = KERN_SUCCESS;
3989             }
3990             CFRelease( bsdPathAsCFString );
3991         }
3992         IOObjectRelease( nextMedia );
3993     }
3994 
3995     return kernResult;
3996 }
3997 
3998 /* Sets up a real cdrom for use in QEMU */
3999 static bool setup_cdrom(char *bsd_path, Error **errp)
4000 {
4001     int index, num_of_test_partitions = 2, fd;
4002     char test_partition[MAXPATHLEN];
4003     bool partition_found = false;
4004 
4005     /* look for a working partition */
4006     for (index = 0; index < num_of_test_partitions; index++) {
4007         snprintf(test_partition, sizeof(test_partition), "%ss%d", bsd_path,
4008                  index);
4009         fd = qemu_open(test_partition, O_RDONLY | O_BINARY | O_LARGEFILE, NULL);
4010         if (fd >= 0) {
4011             partition_found = true;
4012             qemu_close(fd);
4013             break;
4014         }
4015     }
4016 
4017     /* if a working partition on the device was not found */
4018     if (partition_found == false) {
4019         error_setg(errp, "Failed to find a working partition on disc");
4020     } else {
4021         trace_file_setup_cdrom(test_partition);
4022         pstrcpy(bsd_path, MAXPATHLEN, test_partition);
4023     }
4024     return partition_found;
4025 }
4026 
4027 /* Prints directions on mounting and unmounting a device */
4028 static void print_unmounting_directions(const char *file_name)
4029 {
4030     error_report("If device %s is mounted on the desktop, unmount"
4031                  " it first before using it in QEMU", file_name);
4032     error_report("Command to unmount device: diskutil unmountDisk %s",
4033                  file_name);
4034     error_report("Command to mount device: diskutil mountDisk %s", file_name);
4035 }
4036 
4037 #endif /* defined(__APPLE__) && defined(__MACH__) */
4038 
4039 static int hdev_probe_device(const char *filename)
4040 {
4041     struct stat st;
4042 
4043     /* allow a dedicated CD-ROM driver to match with a higher priority */
4044     if (strstart(filename, "/dev/cdrom", NULL))
4045         return 50;
4046 
4047     if (stat(filename, &st) >= 0 &&
4048             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
4049         return 100;
4050     }
4051 
4052     return 0;
4053 }
4054 
4055 static void hdev_parse_filename(const char *filename, QDict *options,
4056                                 Error **errp)
4057 {
4058     bdrv_parse_filename_strip_prefix(filename, "host_device:", options);
4059 }
4060 
4061 static bool hdev_is_sg(BlockDriverState *bs)
4062 {
4063 
4064 #if defined(__linux__)
4065 
4066     BDRVRawState *s = bs->opaque;
4067     struct stat st;
4068     struct sg_scsi_id scsiid;
4069     int sg_version;
4070     int ret;
4071 
4072     if (stat(bs->filename, &st) < 0 || !S_ISCHR(st.st_mode)) {
4073         return false;
4074     }
4075 
4076     ret = ioctl(s->fd, SG_GET_VERSION_NUM, &sg_version);
4077     if (ret < 0) {
4078         return false;
4079     }
4080 
4081     ret = ioctl(s->fd, SG_GET_SCSI_ID, &scsiid);
4082     if (ret >= 0) {
4083         trace_file_hdev_is_sg(scsiid.scsi_type, sg_version);
4084         return true;
4085     }
4086 
4087 #endif
4088 
4089     return false;
4090 }
4091 
4092 static int hdev_open(BlockDriverState *bs, QDict *options, int flags,
4093                      Error **errp)
4094 {
4095     BDRVRawState *s = bs->opaque;
4096     int ret;
4097 
4098 #if defined(__APPLE__) && defined(__MACH__)
4099     /*
4100      * Caution: while qdict_get_str() is fine, getting non-string types
4101      * would require more care.  When @options come from -blockdev or
4102      * blockdev_add, its members are typed according to the QAPI
4103      * schema, but when they come from -drive, they're all QString.
4104      */
4105     const char *filename = qdict_get_str(options, "filename");
4106     char bsd_path[MAXPATHLEN] = "";
4107     bool error_occurred = false;
4108 
4109     /* If using a real cdrom */
4110     if (strcmp(filename, "/dev/cdrom") == 0) {
4111         char *mediaType = NULL;
4112         kern_return_t ret_val;
4113         io_iterator_t mediaIterator = 0;
4114 
4115         mediaType = FindEjectableOpticalMedia(&mediaIterator);
4116         if (mediaType == NULL) {
4117             error_setg(errp, "Please make sure your CD/DVD is in the optical"
4118                        " drive");
4119             error_occurred = true;
4120             goto hdev_open_Mac_error;
4121         }
4122 
4123         ret_val = GetBSDPath(mediaIterator, bsd_path, sizeof(bsd_path), flags);
4124         if (ret_val != KERN_SUCCESS) {
4125             error_setg(errp, "Could not get BSD path for optical drive");
4126             error_occurred = true;
4127             goto hdev_open_Mac_error;
4128         }
4129 
4130         /* If a real optical drive was not found */
4131         if (bsd_path[0] == '\0') {
4132             error_setg(errp, "Failed to obtain bsd path for optical drive");
4133             error_occurred = true;
4134             goto hdev_open_Mac_error;
4135         }
4136 
4137         /* If using a cdrom disc and finding a partition on the disc failed */
4138         if (strncmp(mediaType, kIOCDMediaClass, 9) == 0 &&
4139             setup_cdrom(bsd_path, errp) == false) {
4140             print_unmounting_directions(bsd_path);
4141             error_occurred = true;
4142             goto hdev_open_Mac_error;
4143         }
4144 
4145         qdict_put_str(options, "filename", bsd_path);
4146 
4147 hdev_open_Mac_error:
4148         g_free(mediaType);
4149         if (mediaIterator) {
4150             IOObjectRelease(mediaIterator);
4151         }
4152         if (error_occurred) {
4153             return -ENOENT;
4154         }
4155     }
4156 #endif /* defined(__APPLE__) && defined(__MACH__) */
4157 
4158     s->type = FTYPE_FILE;
4159 
4160     ret = raw_open_common(bs, options, flags, 0, true, errp);
4161     if (ret < 0) {
4162 #if defined(__APPLE__) && defined(__MACH__)
4163         if (*bsd_path) {
4164             filename = bsd_path;
4165         }
4166         /* if a physical device experienced an error while being opened */
4167         if (strncmp(filename, "/dev/", 5) == 0) {
4168             print_unmounting_directions(filename);
4169         }
4170 #endif /* defined(__APPLE__) && defined(__MACH__) */
4171         return ret;
4172     }
4173 
4174     /* Since this does ioctl the device must be already opened */
4175     bs->sg = hdev_is_sg(bs);
4176 
4177     return ret;
4178 }
4179 
4180 #if defined(__linux__)
4181 static int coroutine_fn
4182 hdev_co_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
4183 {
4184     BDRVRawState *s = bs->opaque;
4185     RawPosixAIOData acb;
4186     int ret;
4187 
4188     ret = fd_open(bs);
4189     if (ret < 0) {
4190         return ret;
4191     }
4192 
4193     if (req == SG_IO && s->pr_mgr) {
4194         struct sg_io_hdr *io_hdr = buf;
4195         if (io_hdr->cmdp[0] == PERSISTENT_RESERVE_OUT ||
4196             io_hdr->cmdp[0] == PERSISTENT_RESERVE_IN) {
4197             return pr_manager_execute(s->pr_mgr, qemu_get_current_aio_context(),
4198                                       s->fd, io_hdr);
4199         }
4200     }
4201 
4202     acb = (RawPosixAIOData) {
4203         .bs         = bs,
4204         .aio_type   = QEMU_AIO_IOCTL,
4205         .aio_fildes = s->fd,
4206         .aio_offset = 0,
4207         .ioctl      = {
4208             .buf        = buf,
4209             .cmd        = req,
4210         },
4211     };
4212 
4213     return raw_thread_pool_submit(handle_aiocb_ioctl, &acb);
4214 }
4215 #endif /* linux */
4216 
4217 static coroutine_fn int
4218 hdev_co_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes)
4219 {
4220     BDRVRawState *s = bs->opaque;
4221     int ret;
4222 
4223     ret = fd_open(bs);
4224     if (ret < 0) {
4225         raw_account_discard(s, bytes, ret);
4226         return ret;
4227     }
4228     return raw_do_pdiscard(bs, offset, bytes, true);
4229 }
4230 
4231 static coroutine_fn int hdev_co_pwrite_zeroes(BlockDriverState *bs,
4232     int64_t offset, int64_t bytes, BdrvRequestFlags flags)
4233 {
4234     int rc;
4235 
4236     rc = fd_open(bs);
4237     if (rc < 0) {
4238         return rc;
4239     }
4240 
4241     return raw_do_pwrite_zeroes(bs, offset, bytes, flags, true);
4242 }
4243 
4244 static BlockDriver bdrv_host_device = {
4245     .format_name        = "host_device",
4246     .protocol_name        = "host_device",
4247     .instance_size      = sizeof(BDRVRawState),
4248     .bdrv_needs_filename = true,
4249     .bdrv_probe_device  = hdev_probe_device,
4250     .bdrv_parse_filename = hdev_parse_filename,
4251     .bdrv_file_open     = hdev_open,
4252     .bdrv_close         = raw_close,
4253     .bdrv_reopen_prepare = raw_reopen_prepare,
4254     .bdrv_reopen_commit  = raw_reopen_commit,
4255     .bdrv_reopen_abort   = raw_reopen_abort,
4256     .bdrv_co_create_opts = bdrv_co_create_opts_simple,
4257     .create_opts         = &bdrv_create_opts_simple,
4258     .mutable_opts        = mutable_opts,
4259     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
4260     .bdrv_co_pwrite_zeroes = hdev_co_pwrite_zeroes,
4261 
4262     .bdrv_co_preadv         = raw_co_preadv,
4263     .bdrv_co_pwritev        = raw_co_pwritev,
4264     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
4265     .bdrv_co_pdiscard       = hdev_co_pdiscard,
4266     .bdrv_co_copy_range_from = raw_co_copy_range_from,
4267     .bdrv_co_copy_range_to  = raw_co_copy_range_to,
4268     .bdrv_refresh_limits = raw_refresh_limits,
4269     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
4270 
4271     .bdrv_co_truncate                   = raw_co_truncate,
4272     .bdrv_co_getlength                  = raw_co_getlength,
4273     .bdrv_co_get_info                   = raw_co_get_info,
4274     .bdrv_get_specific_info             = raw_get_specific_info,
4275     .bdrv_co_get_allocated_file_size    = raw_co_get_allocated_file_size,
4276     .bdrv_get_specific_stats = hdev_get_specific_stats,
4277     .bdrv_check_perm = raw_check_perm,
4278     .bdrv_set_perm   = raw_set_perm,
4279     .bdrv_abort_perm_update = raw_abort_perm_update,
4280     .bdrv_probe_blocksizes = hdev_probe_blocksizes,
4281     .bdrv_probe_geometry = hdev_probe_geometry,
4282 
4283     /* generic scsi device */
4284 #ifdef __linux__
4285     .bdrv_co_ioctl          = hdev_co_ioctl,
4286 #endif
4287 
4288     /* zoned device */
4289 #if defined(CONFIG_BLKZONED)
4290     /* zone management operations */
4291     .bdrv_co_zone_report = raw_co_zone_report,
4292     .bdrv_co_zone_mgmt = raw_co_zone_mgmt,
4293     .bdrv_co_zone_append = raw_co_zone_append,
4294 #endif
4295 };
4296 
4297 #if defined(__linux__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
4298 static void cdrom_parse_filename(const char *filename, QDict *options,
4299                                  Error **errp)
4300 {
4301     bdrv_parse_filename_strip_prefix(filename, "host_cdrom:", options);
4302 }
4303 
4304 static void cdrom_refresh_limits(BlockDriverState *bs, Error **errp)
4305 {
4306     bs->bl.has_variable_length = true;
4307     raw_refresh_limits(bs, errp);
4308 }
4309 #endif
4310 
4311 #ifdef __linux__
4312 static int cdrom_open(BlockDriverState *bs, QDict *options, int flags,
4313                       Error **errp)
4314 {
4315     BDRVRawState *s = bs->opaque;
4316 
4317     s->type = FTYPE_CD;
4318 
4319     /* open will not fail even if no CD is inserted, so add O_NONBLOCK */
4320     return raw_open_common(bs, options, flags, O_NONBLOCK, true, errp);
4321 }
4322 
4323 static int cdrom_probe_device(const char *filename)
4324 {
4325     int fd, ret;
4326     int prio = 0;
4327     struct stat st;
4328 
4329     fd = qemu_open(filename, O_RDONLY | O_NONBLOCK, NULL);
4330     if (fd < 0) {
4331         goto out;
4332     }
4333     ret = fstat(fd, &st);
4334     if (ret == -1 || !S_ISBLK(st.st_mode)) {
4335         goto outc;
4336     }
4337 
4338     /* Attempt to detect via a CDROM specific ioctl */
4339     ret = ioctl(fd, CDROM_DRIVE_STATUS, CDSL_CURRENT);
4340     if (ret >= 0)
4341         prio = 100;
4342 
4343 outc:
4344     qemu_close(fd);
4345 out:
4346     return prio;
4347 }
4348 
4349 static bool coroutine_fn cdrom_co_is_inserted(BlockDriverState *bs)
4350 {
4351     BDRVRawState *s = bs->opaque;
4352     int ret;
4353 
4354     ret = ioctl(s->fd, CDROM_DRIVE_STATUS, CDSL_CURRENT);
4355     return ret == CDS_DISC_OK;
4356 }
4357 
4358 static void coroutine_fn cdrom_co_eject(BlockDriverState *bs, bool eject_flag)
4359 {
4360     BDRVRawState *s = bs->opaque;
4361 
4362     if (eject_flag) {
4363         if (ioctl(s->fd, CDROMEJECT, NULL) < 0)
4364             perror("CDROMEJECT");
4365     } else {
4366         if (ioctl(s->fd, CDROMCLOSETRAY, NULL) < 0)
4367             perror("CDROMEJECT");
4368     }
4369 }
4370 
4371 static void coroutine_fn cdrom_co_lock_medium(BlockDriverState *bs, bool locked)
4372 {
4373     BDRVRawState *s = bs->opaque;
4374 
4375     if (ioctl(s->fd, CDROM_LOCKDOOR, locked) < 0) {
4376         /*
4377          * Note: an error can happen if the distribution automatically
4378          * mounts the CD-ROM
4379          */
4380         /* perror("CDROM_LOCKDOOR"); */
4381     }
4382 }
4383 
4384 static BlockDriver bdrv_host_cdrom = {
4385     .format_name        = "host_cdrom",
4386     .protocol_name      = "host_cdrom",
4387     .instance_size      = sizeof(BDRVRawState),
4388     .bdrv_needs_filename = true,
4389     .bdrv_probe_device	= cdrom_probe_device,
4390     .bdrv_parse_filename = cdrom_parse_filename,
4391     .bdrv_file_open     = cdrom_open,
4392     .bdrv_close         = raw_close,
4393     .bdrv_reopen_prepare = raw_reopen_prepare,
4394     .bdrv_reopen_commit  = raw_reopen_commit,
4395     .bdrv_reopen_abort   = raw_reopen_abort,
4396     .bdrv_co_create_opts = bdrv_co_create_opts_simple,
4397     .create_opts         = &bdrv_create_opts_simple,
4398     .mutable_opts        = mutable_opts,
4399     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
4400 
4401     .bdrv_co_preadv         = raw_co_preadv,
4402     .bdrv_co_pwritev        = raw_co_pwritev,
4403     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
4404     .bdrv_refresh_limits    = cdrom_refresh_limits,
4405     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
4406 
4407     .bdrv_co_truncate                   = raw_co_truncate,
4408     .bdrv_co_getlength                  = raw_co_getlength,
4409     .bdrv_co_get_allocated_file_size    = raw_co_get_allocated_file_size,
4410 
4411     /* removable device support */
4412     .bdrv_co_is_inserted    = cdrom_co_is_inserted,
4413     .bdrv_co_eject          = cdrom_co_eject,
4414     .bdrv_co_lock_medium    = cdrom_co_lock_medium,
4415 
4416     /* generic scsi device */
4417     .bdrv_co_ioctl      = hdev_co_ioctl,
4418 };
4419 #endif /* __linux__ */
4420 
4421 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
4422 static int cdrom_open(BlockDriverState *bs, QDict *options, int flags,
4423                       Error **errp)
4424 {
4425     BDRVRawState *s = bs->opaque;
4426     int ret;
4427 
4428     s->type = FTYPE_CD;
4429 
4430     ret = raw_open_common(bs, options, flags, 0, true, errp);
4431     if (ret) {
4432         return ret;
4433     }
4434 
4435     /* make sure the door isn't locked at this time */
4436     ioctl(s->fd, CDIOCALLOW);
4437     return 0;
4438 }
4439 
4440 static int cdrom_probe_device(const char *filename)
4441 {
4442     if (strstart(filename, "/dev/cd", NULL) ||
4443             strstart(filename, "/dev/acd", NULL))
4444         return 100;
4445     return 0;
4446 }
4447 
4448 static int cdrom_reopen(BlockDriverState *bs)
4449 {
4450     BDRVRawState *s = bs->opaque;
4451     int fd;
4452 
4453     /*
4454      * Force reread of possibly changed/newly loaded disc,
4455      * FreeBSD seems to not notice sometimes...
4456      */
4457     if (s->fd >= 0)
4458         qemu_close(s->fd);
4459     fd = qemu_open(bs->filename, s->open_flags, NULL);
4460     if (fd < 0) {
4461         s->fd = -1;
4462         return -EIO;
4463     }
4464     s->fd = fd;
4465 
4466     /* make sure the door isn't locked at this time */
4467     ioctl(s->fd, CDIOCALLOW);
4468     return 0;
4469 }
4470 
4471 static bool coroutine_fn cdrom_co_is_inserted(BlockDriverState *bs)
4472 {
4473     return raw_getlength(bs) > 0;
4474 }
4475 
4476 static void coroutine_fn cdrom_co_eject(BlockDriverState *bs, bool eject_flag)
4477 {
4478     BDRVRawState *s = bs->opaque;
4479 
4480     if (s->fd < 0)
4481         return;
4482 
4483     (void) ioctl(s->fd, CDIOCALLOW);
4484 
4485     if (eject_flag) {
4486         if (ioctl(s->fd, CDIOCEJECT) < 0)
4487             perror("CDIOCEJECT");
4488     } else {
4489         if (ioctl(s->fd, CDIOCCLOSE) < 0)
4490             perror("CDIOCCLOSE");
4491     }
4492 
4493     cdrom_reopen(bs);
4494 }
4495 
4496 static void coroutine_fn cdrom_co_lock_medium(BlockDriverState *bs, bool locked)
4497 {
4498     BDRVRawState *s = bs->opaque;
4499 
4500     if (s->fd < 0)
4501         return;
4502     if (ioctl(s->fd, (locked ? CDIOCPREVENT : CDIOCALLOW)) < 0) {
4503         /*
4504          * Note: an error can happen if the distribution automatically
4505          * mounts the CD-ROM
4506          */
4507         /* perror("CDROM_LOCKDOOR"); */
4508     }
4509 }
4510 
4511 static BlockDriver bdrv_host_cdrom = {
4512     .format_name        = "host_cdrom",
4513     .protocol_name      = "host_cdrom",
4514     .instance_size      = sizeof(BDRVRawState),
4515     .bdrv_needs_filename = true,
4516     .bdrv_probe_device	= cdrom_probe_device,
4517     .bdrv_parse_filename = cdrom_parse_filename,
4518     .bdrv_file_open     = cdrom_open,
4519     .bdrv_close         = raw_close,
4520     .bdrv_reopen_prepare = raw_reopen_prepare,
4521     .bdrv_reopen_commit  = raw_reopen_commit,
4522     .bdrv_reopen_abort   = raw_reopen_abort,
4523     .bdrv_co_create_opts = bdrv_co_create_opts_simple,
4524     .create_opts         = &bdrv_create_opts_simple,
4525     .mutable_opts       = mutable_opts,
4526 
4527     .bdrv_co_preadv         = raw_co_preadv,
4528     .bdrv_co_pwritev        = raw_co_pwritev,
4529     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
4530     .bdrv_refresh_limits    = cdrom_refresh_limits,
4531     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
4532 
4533     .bdrv_co_truncate                   = raw_co_truncate,
4534     .bdrv_co_getlength                  = raw_co_getlength,
4535     .bdrv_co_get_allocated_file_size    = raw_co_get_allocated_file_size,
4536 
4537     /* removable device support */
4538     .bdrv_co_is_inserted     = cdrom_co_is_inserted,
4539     .bdrv_co_eject           = cdrom_co_eject,
4540     .bdrv_co_lock_medium     = cdrom_co_lock_medium,
4541 };
4542 #endif /* __FreeBSD__ */
4543 
4544 #endif /* HAVE_HOST_BLOCK_DEVICE */
4545 
4546 static void bdrv_file_init(void)
4547 {
4548     /*
4549      * Register all the drivers.  Note that order is important, the driver
4550      * registered last will get probed first.
4551      */
4552     bdrv_register(&bdrv_file);
4553 #if defined(HAVE_HOST_BLOCK_DEVICE)
4554     bdrv_register(&bdrv_host_device);
4555 #ifdef __linux__
4556     bdrv_register(&bdrv_host_cdrom);
4557 #endif
4558 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
4559     bdrv_register(&bdrv_host_cdrom);
4560 #endif
4561 #endif /* HAVE_HOST_BLOCK_DEVICE */
4562 }
4563 
4564 block_init(bdrv_file_init);
4565